Grundlehren der mathematischen Wissenschaften 359

A Series of Comprehensive Studies in Mathematics

François Treves

Analytic
Partial
Differential
Equations
Grundlehren der
mathematischen Wissenschaften

A Series of Comprehensive Studies in Mathematics

Volume 359

Editors-in-Chief
Alain Chenciner, Observatoire de Paris, Paris, France
S.R.S. Varadhan, New York University, New York, NY, USA

Series Editors
Henri Darmon, McGill University, Montréal, Canada
Pierre de la Harpe, University of Geneva, Geneva, Switzerland
Frank den Hollander, Leiden University, Leiden, The Netherlands
Nigel J. Hitchin, University of Oxford, Oxford, UK
Nalini Joshi, University of Sydney, Sydney, Australia
Antti Kupiainen, University of Helsinki, Helsinki, Finland
Gilles Lebeau, Côte d’Azur University, Nice, France
Jean-François Le Gall, Paris-Saclay University, Orsay, France
Fang-Hua Lin, New York University, New York, NY, USA
Shigefumi Mori, Kyoto University, Kyoto, Japan
Bào Châu Ngô, University of Chicago, Chicago, IL, USA
Denis Serre, École Normale Supérieure de Lyon, Lyon, France
Michel Waldschmidt, Sorbonne University, Paris, France
Grundlehren der mathematischen Wissenschaften (subtitled Comprehensive
Studies in Mathematics), Springer’s first series in higher mathematics, was founded
by Richard Courant in 1920. It was conceived as a series of modern textbooks.
A number of significant changes appear after World War II. Outwardly, the change
was in language: whereas most of the first 100 volumes were published in German,
the following volumes are almost all in English. A more important change concerns
the contents of the books. The original objective of the Grundlehren had been to
lead readers to the principal results and to recent research questions in a single
relatively elementary and accessible book. Good examples are van der Waerden’s
2-volume Introduction to Algebra or the two famous volumes of Courant and
Hilbert on Methods of Mathematical Physics.
Today, it is seldom possible to start at the basics and, in one volume or even
two, reach the frontiers of current research. Thus many later volumes are both more
specialized and more advanced. Nevertheless, most volumes of the series are meant
to be textbooks of a kind, with occasional reference works or pure research
monographs. Each book should lead up to current research, without over-
emphasizing the author’s own interests. Proofs of the major statements should be
enunciated, however the presentation should remain expository. Examples of books
that fit this description are Maclane’s Homology, Siegel & Moser on Celestial
Mechanics, Gilbarg & Trudinger on Elliptic PDE of Second Order, Dafermos’s
Hyperbolic Conservation Laws in Continuum Physics ... Longevity is an important
criterion: a GL volume should continue to have an impact over many years. Topics
should be of current mathematical relevance, and not too narrow.
The tastes of the editors play a pivotal role in the selection of topics.
Authors are encouraged to follow their individual style, but keep the interests of
the reader in mind when presenting their subject. The inclusion of exercises and
historical background is encouraged.
The GL series does not strive for systematic coverage of all of mathematics.
There are both overlaps between books and gaps. However, a systematic effort is
made to cover important areas of current interest in a GL volume when they become
ripe for GL-type treatment.
As far as the development of mathematics permits, the direction of GL remains
true to the original spirit of Courant. Many of the oldest volumes are popular to this
day and some have not been superseded. One should perhaps never advertise a
contemporary book as a classic but many recent volumes and many forthcoming
volumes will surely earn this attribute through their use by generations of
mathematicians.

More information about this series at https://link.springer.com/bookseries/138

François Treves

Analytic Partial Differential

Equations
François Treves
Golden, CO, USA

ISSN 0072-7830 ISSN 2196-9701 (electronic)

Grundlehren der mathematischen Wissenschaften
ISBN 978-3-030-94054-6 ISBN 978-3-030-94055-3 (eBook)
https://doi.org/10.1007/978-3-030-94055-3

Mathematics Subject Classification (2020): 35-02, 35A01, 35A09, 35A10, 44-02, 44F05, 58Jxx, 35Axx

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

There is no need for a lengthy summary of the material in this book; a look at the
table of contents should give a first idea of what it is about; details are provided in
the introductions to (or the table of contents of) the individual chapters. It seems
preferable to explain the perspective of the book, and its aims. This book is not a
treatise; its ambitions are mainly pedagogical: to introduce, and ideally to attract,
students and, possibly, mathematicians engaged in adjacent areas of research, to the
theory of partial differential equations (PDEs) with analytic coefficients developed
in the second half of the XXth century (or earlier). Important areas of the theory
have been insufficiently studied, or hardly explored at all (see below). History has
shown that results about analytic PDEs often provide guidance in the investigation
of PDEs whose coefficients are not analytic.
The book is divided into seven Parts. Each Part is devoted to an important aspect
of the theory, or to the introduction of concepts, methods and results needed in later
Parts. A good example of the latter is Part III, titled Geometric Background, whose
contents, with a few exceptions, are covered in much greater detail in innumerable,
but mostly separate, monographs or treatises. Generally speaking, the exposition is
based on the assumption that the reader knows very little on the subject, or needs a
refresher without having to go to other texts. Most of the time the exposition starts
from the basic definitions and propositions and tries to be self-sufficient as much
as possible. This entails frequent repetitions. It must also be said that the attempt
at self-sufficiency has its limits, mainly because of the need for a few deep results
from the theory of Several Complex Variables (SCV); in this text, these results are
stated without proofs. On the subject of SCV we encourage the student to go to the
masterful monograph [Hörmander, 1966]; another good text is [Gunning and Rossi,
1965].
Another expository choice was to give detailed proofs, even of quite elementary
(or well-known) results, and limit the (still too large) number of usual evasions of
the sort “it can easily be proved”, “we leave it to the reader to prove” and such. The
intent, here, is to avoid breaking the “flow” of the reading; also, of course, to allay
the possible doubts of the reader. A sensible approach for a first reading of any one
of the Parts of the book (past Chapters 1 and 2) is to skip as much of the proofs

v
vi Preface

as possible, especially those that are heavily technical and lengthy; and, instead, to
try to get a clear view of the mathematical architecture of the matter at hand, of its
foundations, its pillars and main features. In this respect it is hoped that the examples
provided will be helpful.
The only Part that is more than introductory, Part VII, is devoted to PDEs of
principal type (in older parlance, PDEs with complex coefficients and simple real
characteristics). Its first two chapters (Chs. 23, 24) provide a detailed, verging on
complete, presentation of the results on the existence (or lack thereof) of local
solutions of a single PDE of principal type, and of the analytic singularities of the
solutions, focusing on the propagation of these singularities. The solutions can be
distributions (introduced in Chapter 2) or hyperfunctions (introduced in Part II from
a variety of viewpoints). With hyperfunctions it is relatively easy to go from local
to semiglobal solutions under the hypothesis of nontrapping of bicharacteristics. It
should be said that Part VII relies on practically all the material that precedes it, in
particular on the so-called FBI (for Fourier–Brós–Iagolnitzer) transform (introduced
for distributions in Ch. 3, for hyperfunctions in Ch. 7) and, especially, the version
in [Sjöstrand, 1982] (explained in Ch. 22). Following Chapters 23, 24, the text
moves to the (microlocal) solvability Condition (Ψ) for pseudodifferential equations
of principal type and touches on its possible generalization to involutive systems
of such equations or, more accurately, to the differential complexes these systems
define.
Some topics that are not, but should have been, discussed in the book, must be
mentioned. Firstly, nothing will be found about hyperbolic equations and diffraction
theory, despite the fact that analyticity makes certain features of this subject matter
particularly elegant. Of course, hyperbolicity is a special case of principal type and
the propagation of analytic singularities of the solutions throws some light on the
hyperbolic case; but nothing is said about this important connection. Secondly, and
most regrettably, not even an allusion will be found to the promising links between
microlocal analysis and asymptotic expansions, whose theory is closely linked to
ordinary differential equations (ODEs) in the complex plane (of the type generalizing
Fuchsian ODEs). Thirdly, nothing will be found about higher microlocalizations. The
only excuses for these lacunæ are the limitations of the author, and the all too banal
observation that art is long and life is short.
It was mentioned at the beginning of this foreword that, in the land of analytic
PDEs, there remain vast territories open for further exploration. Four come directly
to mind:
(1) Nonlinear equations and, to start, nonlinear equations whose linearizations (at a
selected solution) are of principal type. Practically nothing, beyond the Cauchy–
Kovalevskaya Theorem, is known about the solvability of these (complex, C 𝜔 )
equations.
(2) Systems of linear PDEs, whether determined or not, and the properties (existence,
regularity, etc.) of their (say, hyperfunction) solutions.
(3) Linear PDEs with multiple real characteristics, at a respectable level of gener-
ality.
Preface vii

(4) PDEs which degenerate at a point, i.e., whose principal symbol vanishes identi-
cally on the cotangent space at the point. Little is known about these equations
(even of order 1) in dimension 𝑛 ≥ 2.

The writing of this book has taken a long time. It would not have reached comple-
tion without the constant encouragement of Antonio Bove and Paulo Cordaro. Bove
pointed out repeated errors and provided examples. Cordaro suggested the insertion
of significant results in various parts of the text, and that of an entire chapter (Ch. 8)
devoted to hyperdifferential operators, for which he provided most of the material.
The author wishes to thank the referees for their very constructive criticism. It is
a ritual, at this point, to admit that all remaining errors in the book are solely the
fault of the author; this is particularly appropriate in the case of this book and this
author.

Boulder, Colorado, François Treves

in the first year of the pandemic.
Contents

Part I Distributions and Analyticity in Euclidean Space

1 Functions and Differential Operators in Euclidean Space . . . . . . . . . . . 3

1.1 Basic Notation and Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Smooth, Real-analytic, Holomorphic Functions . . . . . . . . . . . . . . . . . 6
1.3 Differential Operators with Smooth Coefficients . . . . . . . . . . . . . . . . . 14

2 Distributions in Euclidean Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.1 Basics on Distributions in Euclidean Space . . . . . . . . . . . . . . . . . . . . . 18
2.2 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Distribution Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Fundamental Solutions, Parametrix, Hypoelliptic PDOs . . . . . . . . . . 31

3 Analytic Tools in Distribution Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.1 Analytic Parametrices, Analytic Hypoellipticity . . . . . . . . . . . . . . . . . 36
3.2 Ehrenpreis’ Cutoffs and Analytic Regularity of Distributions . . . . . . 38
3.3 Distribution Boundary Values of Holomorphic Functions . . . . . . . . . 42
3.4 The FBI Transform of Distributions. An Introduction . . . . . . . . . . . . . 51
3.5 The Analytic Wave-Front Set of a Distribution . . . . . . . . . . . . . . . . . . 57

4 Analyticity of Solutions of Linear PDEs. Basic Results . . . . . . . . . . . . . . 69

4.1 Analyticity of Solutions of Elliptic Linear PDEs . . . . . . . . . . . . . . . . . 70
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms . . . . 80
4.3 A Generalization of the Harmonic Oscillator . . . . . . . . . . . . . . . . . . . . 97
4.A Appendix: Hermite’s Functions and the Schwartz Space . . . . . . . . . . 102

5 The Cauchy–Kovalevskaya Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

5.1 A Nonlinear Ovsyannikov Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2 Application: the Nonlinear Cauchy–Kovalevskaya Theorem . . . . . . . 120
5.3 Applications to Linear PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4 Application to Integrodifferential Cauchy Problems . . . . . . . . . . . . . . 136

ix
x Contents

Part II Hyperfunctions in Euclidean Space

6 Analytic Functionals in Euclidean Space . . . . . . . . . . . . . . . . . . . . . . . . . . 141

6.1 Analytic Functionals in Complex Domains . . . . . . . . . . . . . . . . . . . . . 143
6.2 Analytic Functionals in C𝑛 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.3 Analytic Functionals in R𝑛 as Cohomology Classes . . . . . . . . . . . . . . 154

7 Hyperfunctions in Euclidean Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

7.1 The Sheaf of Hyperfunctions in Euclidean Space . . . . . . . . . . . . . . . . 167
7.2 Boundary values of holomorphic functions in wedges . . . . . . . . . . . . 180
7.3 The FBI Transform of Analytic Functionals . . . . . . . . . . . . . . . . . . . . . 184
7.4 Analytic Wave-front Set of a Hyperfunction . . . . . . . . . . . . . . . . . . . . 192
7.5 Edge of the Wedge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
7.6 Microfunctions in Euclidean space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

8 Hyperdifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

8.1 Action on Holomorphic Functions and on Hyperfunctions . . . . . . . . 222
8.2 Local Representation of Hyperfunctions . . . . . . . . . . . . . . . . . . . . . . . 237
8.3 Elliptic Hyperdifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
8.4 Solvability of Constant Coefficients Hyperdifferential Equations . . . 247

Part III Geometric Background

9 Elements of Differential Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

9.1 Regular Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
9.2 Fibre Bundles, Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
9.3 Tangent and Cotangent Bundles of a Manifold . . . . . . . . . . . . . . . . . . 271
9.4 Differential Complexes and Grassman Algebras . . . . . . . . . . . . . . . . . 280

10 A Primer on Sheaf Cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

10.1 Basics on Sheaf Cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
10.2 Fine Sheaves and Fine Resolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
10.3 Relative Sheaf Cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
10.4 Edge of the Wedge in (Co)homological Terms . . . . . . . . . . . . . . . . . . 322

11 Distributions and Hyperfunctions on a Manifold . . . . . . . . . . . . . . . . . . . 327

11.1 Distributions and Currents on a Manifold . . . . . . . . . . . . . . . . . . . . . . . 328
11.2 Plurisubharmonic functions and pseudoconvex domains . . . . . . . . . . 336
11.3 Hyperfunctions and Microfunctions in an Analytic Manifold . . . . . . 345

12 Lie Algebras of Vector Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

12.1 The Lie Algebra of Smooth Vector Fields . . . . . . . . . . . . . . . . . . . . . . 352
12.2 Integral Manifolds. Frobenius’ Theorem . . . . . . . . . . . . . . . . . . . . . . . 354
12.3 Local Flow of a Regular Vector Field . . . . . . . . . . . . . . . . . . . . . . . . . . 361
12.4 Foliations Defined By Analytic Vector Fields . . . . . . . . . . . . . . . . . . . 364
12.5 Systems of Vector Fields Generating Special Lie Algebras . . . . . . . . 368
Contents xi

13 Elements of Symplectic Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375

13.1 Elements of Symplectic Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
13.2 The Metaplectic Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
13.3 Symplectic Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
13.4 Involutive Systems of Functions of Principal Type . . . . . . . . . . . . . . . 414
13.5 Real and Imaginary Symplectic Structures in C2𝑛 . . . . . . . . . . . . . . . . 429
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds . 440

Part IV Stratification of Analytic Varieties and Division of Distributions by

Analytic Functions

14 Analytic Stratifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457

14.1 Analytic Stratifications and Stratifiable Sets . . . . . . . . . . . . . . . . . . . . . 458
14.2 Analytic Subvarieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
14.3 The Weierstrass Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
14.4 Local Partitions of a Complex Hypersurface . . . . . . . . . . . . . . . . . . . . 483
14.5 Local Stratifications of a Real-Analytic Variety . . . . . . . . . . . . . . . . . . 493
14.6 Semianalytic Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506

15 Division of Distributions by Analytic Functions . . . . . . . . . . . . . . . . . . . . 513

15.1 The Lojasiewicz Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
15.2 Division of Distributions by Analytic Functions . . . . . . . . . . . . . . . . . 526
15.3 Desingularization and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
15.A Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555

Part V Analytic Pseudodifferential Operators and Fourier Integral Operators

16 Elementary Pseudodifferential Calculus in the 𝑪 ∞ Class . . . . . . . . . . . . 565

16.1 Standard Pseudodifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . . 566
16.2 Symbolic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
16.3 Classical symbols and classical pseudodifferential operators . . . . . . . 595
16.4 The Weyl Calculus in Euclidean Space . . . . . . . . . . . . . . . . . . . . . . . . . 603

17 Analytic Pseudodifferential Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615

17.1 Analytic Pseudodifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . . 616
17.2 Symbolic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
17.3 Analytic Microlocalization In Distribution Theory . . . . . . . . . . . . . . . 657
17.4 Action on Singularity Hyperfunctions . . . . . . . . . . . . . . . . . . . . . . . . . 668
17.5 Microdifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683

18 Fourier Integral Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685

18.1 Fourier Distribution Kernels in Euclidean Space . . . . . . . . . . . . . . . . . 686
18.2 The Lagrangian Manifold Associated to a Phase-function . . . . . . . . . 698
18.3 Fourier Integral Operators. Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 706
18.4 Reduction of the Fiber Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
18.5 Composition and Continuity of Fourier Integral Operators . . . . . . . . 729
xii Contents

18.6 Globally Defined Fourier Integral Operators . . . . . . . . . . . . . . . . . . . . 746

18.7 Principles of Analytic Fourier Integral Operators . . . . . . . . . . . . . . . . 755
18.A Appendix: Stationary Phase Formal Expansion . . . . . . . . . . . . . . . . . . 757

Part VI Complex Microlocal Analysis

19 Classical Analytic Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763

19.1 Formal Analytic Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764
19.2 Classical Analytic Differential Operators of Infinite Order . . . . . . . . . 776
19.3 The Complex Stationary Phase Formula . . . . . . . . . . . . . . . . . . . . . . . . 788
19.4 Symbolic Calculus and the KdV Hierarchy . . . . . . . . . . . . . . . . . . . . . 802

20 Germ Fourier Integral Operators in Complex Space . . . . . . . . . . . . . . . 825

20.1 Analytic Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 827
20.2 Contours and Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 834
20.3 Sjöstrand Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 841
20.4 Germ Fourier-like Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855
20.5 Sjöstrand Triads and Germ Fourier Integral Operators . . . . . . . . . . . . 863

21 Germ Pseudodifferential Operators in Complex Space . . . . . . . . . . . . . 871

21.1 Germ Pseudodifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 872
21.2 Classical Germ Pseudodifferential Operators . . . . . . . . . . . . . . . . . . . . 891
21.3 Action on distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 902
21.4 Action on Hyperfunctions and Microfunctions . . . . . . . . . . . . . . . . . . 903

22 Germ FBI Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915

22.1 Germ FBI Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915
22.2 Germ FBI Transforms of Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 920
22.3 The Equivalence Theorem for Distributions . . . . . . . . . . . . . . . . . . . . . 925

Part VII Analytic Pseudodifferential Operators of Principal Type

23 Analytic PDEs of Principal Type. Local Solvability . . . . . . . . . . . . . . . . . 939

23.1 Pseudodifferential Operators of Principal Type . . . . . . . . . . . . . . . . . . 940
23.2 Local Solvability of Analytic PDEs of Principal Type . . . . . . . . . . . . 963

24 Analytic PDEs of Principal Type. Regularity of the Solutions . . . . . . . . 985

24.1 A New Concept: Subellipticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986
24.2 Statement of the Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 992
24.3 Hypoellipticity Implies (Q) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993
24.4 Property (Q) Implies Subellipticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 997
24.5 Analytic Hypoellipticity Implies (Q) . . . . . . . . . . . . . . . . . . . . . . . . . . 1030
24.6 Property (Q) Implies Analytic Hypoellipticity . . . . . . . . . . . . . . . . . . . 1033
24.7 The C ∞ Situation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044
24.8 Propagation of Analytic Singularities . . . . . . . . . . . . . . . . . . . . . . . . . . 1048
24.A Appendix: Properties of Real Polynomials in a Single Variable . . . . 1057
Contents xiii

24.B Appendix: Analytic Estimates of Exponential Amplitudes . . . . . . . . . 1061

25 Solvability of Constant Vector Fields of Type (1,0) . . . . . . . . . . . . . . . . . 1067

25.1 C-Convexity and Global Solvability . . . . . . . . . . . . . . . . . . . . . . . . . . . 1068
25.2 Local Solvability at the Boundary. First Steps . . . . . . . . . . . . . . . . . . . 1079
25.3 Local Solvability at the Boundary. Final Characterization . . . . . . . . . 1100
25.4 The Differential Complex. Generalities . . . . . . . . . . . . . . . . . . . . . . . . 1114
25.A Appendix: Minima of Families of Plurisubharmonic Functions . . . . 1120

26 Pseudodifferential Solvability and Property (𝚿) . . . . . . . . . . . . . . . . . . . . 1131

26.1 Solvability: the Difference between Differential and
Pseudodifferential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1132
26.2 Property (𝚿) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1133
26.3 Microlocal Solvability in Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 1143

27 Pseudodifferential Complexes in Tube Structures . . . . . . . . . . . . . . . . . . 1159

27.1 Pseudodifferential Complexes of Principal Type . . . . . . . . . . . . . . . . . 1160
27.2 Tube Pseudodifferential Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1166
27.3 Phase-function and Amplitude . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1170
27.4 Approximate Homotopy Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1174
27.5 Homotopy Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1184
27.6 Poincaré Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1190

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1199

Notation Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1211

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1217
Part I
Distributions and Analyticity in Euclidean
Space
Chapter 1
Functions and Differential Operators in
Euclidean Space

This chapter is devoted to introducing the most basic notation, terminology and
definitions related to the function classes constantly used in the book. These are
the real-analytic (also referred to as C 𝜔 ), complex-analytic (also referred to as
holomorphic) and infinitely differentiable (also referred to as smooth or C ∞ ) functions
defined in an open subset of Euclidean space R𝑛 or C𝑛 and valued in a vector space
where the scalars are real or complex (of course, always the latter if the function is
holomorphic). When the same statement is valid simultaneously in more than one of
these categories we have chosen to talk of regular functions valued in a vector space
over the field K = R or C. This material is introduced solely for the convenience of
the reader. A subsection gives a very rapid description of the basic concepts of germs
and sheaves, indispensable in dealing with complex-analytic objects and much used,
later in the book, in the overlaps with the theory of Several Complex Variables.
There are numerous textbooks presenting each of these topics in great detail; an
introductory monograph on sheaf theory is [Godement, 1964], an updated, and more
technical, text is [Bredon, 1997].

1.1 Basic Notation and Terminology

1.1.1 Multi-index notation, partial derivatives, Leibniz rules

We shall make frequent use of the multi-index notation: if 𝛼 = (𝛼1 , ..., 𝛼𝑛 ) ∈ Z+𝑛 we
write

𝑥 𝛼 = 𝑥 1𝛼1 · · · 𝑥 𝑛𝛼𝑛 ; (1.1.1)

|𝛼| = 𝛼1 + · · · + 𝛼𝑛 , 𝛼! = 𝛼1 ! · · · 𝛼𝑛 !;

𝛼 𝛼!
= ,𝛽⪯𝛼
𝛽 (𝛼 − 𝛽)!𝛽!

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 3

F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_1
4 1 Functions and Differential Operators in Euclidean Space

(𝛽 ⪯ 𝛼 shall always mean that 𝛽 𝑗 ≤ 𝛼 𝑗 for each 𝑗 = 1, ..., 𝑛; 𝛽 ≺ 𝛼 shall mean

𝛽 ⪯ 𝛼, 𝛽 ≠ 𝛼).
𝜕𝑓
The notation 𝜕𝑥 𝑗
or 𝜕𝑥 𝑗 𝑓 (or even, sometimes, 𝜕 𝑗 𝑓 ) will stand for the partial
derivative of the differentiable function 𝑓 with respect to 𝑥 𝑗 ; we write √1 𝜕𝑥
𝜕𝑓
= D𝑗 𝑓
−1 𝑗
or D 𝑥 𝑗 𝑓 if there is a risk of confusion.
We use the following notation for the partial derivatives of order 𝛼 of the (suffi-
ciently differentiable) function 𝑓 :

𝑓 ( 𝛼) = 𝜕𝑥𝛼 𝑓 = 𝜕𝑥𝛼11 · · · 𝜕𝑥𝛼𝑛𝑛 𝑓 , D 𝑥𝛼 𝑓 = D 𝛼 𝑓 = D1𝛼1 · · · D𝑛𝛼𝑛 𝑓 . (1.1.2)

We shall make use of the multivariable Leibniz rule (valid for pairs of sufficiently
differentiable functions 𝑓 , 𝑔):
∑︁ 𝛼
D 𝛼 ( 𝑓 𝑔) = D 𝛼−𝛽 𝑓 D𝛽 𝑔. (1.1.3)
𝛽⪯𝛼
𝛽

Let 𝑃 (𝜉) be a polynomial with complex coefficients in the variables 𝜉1 , ..., 𝜉 𝑛 ;

substituting D 𝑗 for 𝜉 𝑗 yields the differential polynomial 𝑃 (D). The following gen-
eralized Leibniz rule will be used in this text
∑︁ 1
𝛽 𝛽
𝑃 (D) ( 𝑓 𝑔) = D 𝑥 𝑔 𝜕 𝜉 𝑃 (D) 𝑓 . (1.1.4)
𝛽 ∈Z𝑛
𝛽!
+

To prove (1.1.4) (which is linear with respect to 𝑃) it suffices to deal with 𝑃 (𝜉) = 𝜉 𝛼 ,
in which case it reduces to (1.1.3).
The following transpose Leibniz rule of (1.1.3) is also often useful:

∑︁ 𝛼 𝛽 𝛼−𝛽
𝑓 D𝛼𝑔 = (−1) | 𝛼−𝛽 | D D 𝑓 𝑔 . (1.1.5)
𝛽⪯𝛼
𝛽

To prove (1.1.5) it suffices to apply (1.1.3) to each term in the sum on the right of
(1.1.5). A quick check of the “right” sign (−1) | 𝛼−𝛽 | is obtained by integrating both
sides of (1.1.5) over some domain against a third sufficiently differentiable function
ℎ (assuming the needed order of vanishing at the boundary) and applying (1.1.3):
∫ ∫
ℎ 𝑓 D 𝑥𝛼 𝑔d𝑥 = (−1) | 𝛼 | 𝑔D 𝑥𝛼 (ℎ 𝑓 ) d𝑥
∑︁ 𝛼 ∫
|𝛼| 𝛼−𝛽 𝛽
= (−1) 𝑔 D 𝑥 𝑓 D 𝑥 ℎd𝑥
𝛽⪯𝛼
𝛽
∫
| 𝛼−𝛽 | 𝛼
∑︁
𝛽 𝛼−𝛽
= (−1) ℎD 𝑥 D 𝑥 𝑓 𝑔 d𝑥.
𝛽⪯𝛼
𝛽
1.1 Basic Notation and Terminology 5

By the same token we derive from (1.1.5)

∫ ∫
ℎ 𝑓 𝑃 (D) 𝑔d𝑥 = 𝑔𝑃 (−D) (ℎ 𝑓 ) d𝑥
∑︁ (−1) | 𝛼 | ∫
= 𝑔 D 𝑥𝛼 𝑓 𝜕 𝜉𝛼 𝑃 (−D) ℎd𝑥
𝛼∈Z+𝑛
𝛼!
∑︁ (−1) | 𝛼 | ∫
= ℎ 𝜕 𝜉𝛼 𝑃 (D) 𝑔D 𝑥𝛼 𝑓 d𝑥,
𝛼∈Z𝑛
𝛼!
+

whence the transpose of (1.1.5),

∑︁ (−1) | 𝛼 |
𝑓 𝑃 (D) 𝑔 = 𝜕 𝜉𝛼 𝑃 (D) 𝑔D 𝑥𝛼 𝑓 . (1.1.6)
𝛼∈Z𝑛
𝛼!
+

Most frequently the variable point in C𝑛 will be 𝑧 = (𝑧1 , ..., 𝑧 𝑛 ) (often other
letters, such as 𝑤, 𝜁, etc., are substituted for 𝑧). Naturally we also write 𝑧 = 𝑥 + 𝑖𝑦,
with 𝑥 = Re 𝑧, 𝑦 = Im 𝑧, belonging to R𝑛 . We frequently use the notation
𝑛
∑︁ 𝑛
∑︁
𝑧·𝑤 = 𝑧 𝑗 𝑤 𝑗 , ⟨𝑧⟩ 2 = 𝑧 · 𝑧 = 𝑧2𝑗 , (1.1.7)
𝑗=1 𝑗=1

as well as ⟨𝑧⟩ for the main branch of the square-root of ⟨𝑧⟩ 2 when said main branch
makes sense.

Remark 1.1.1 The notation

√︃ ⟨𝑧⟩ runs against the common use of the so-called
Japanese norm ⟨𝜉⟩ = 1 + |𝜉 | 2 , 𝜉 ∈ R𝑛 . The latter meaning shall never be used in

the book. More importantly, we will need exp − ⟨𝐹⟩ 2 for complicated functions
2
√ as a factor in certain integrands; replacing ⟨𝐹⟩ by
𝐹 of several complex variables,
𝐹 · 𝐹 or, even worse, ⟨𝐹⟩ by 𝐹 · 𝐹 (when mathematically well-defined), would be
typographically unfeasible.
∫ 𝑝1
We will use the standard notation of Lebesgue theory: ∥ 𝑓 ∥ 𝐿 𝑝 = R𝑛 | 𝑓 (𝑥)| 𝑝 d𝑥
for the 𝐿 𝑝 norm (1 ≤ 𝑝 < +∞), ∥ 𝑓 ∥ 𝐿 ∞ for the “essential supremum”, etc.; the space
of locally (Lebesgue-)integrable (or locally 𝐿 𝑝 ) functions in Ω shall be denoted
𝑝 𝑝
by 𝐿 loc (Ω). By the support of 𝑓 ∈ 𝐿 loc (Ω) we mean the intersection, denoted by
supp 𝑓 , of all the closed subsets 𝐸 of Ω such that 𝑓 ≡ 0 in Ω\𝐸 (the complement of
𝐸 in Ω, an open set).
6 1 Functions and Differential Operators in Euclidean Space

1.2 Smooth, Real-analytic, Holomorphic Functions

1.2.1 Regular functions in open subsets of Euclidean space

In this section, unless specified otherwise, Ω shall be an open subset of R𝑛 . We

denote by C ∞ (Ω) the linear space of the complex-valued smooth functions in Ω,
i.e., functions having derivatives of all orders. A sequence {𝜑 𝑛 } 𝑛=0,1,2,... of such
functions converges to zero in C ∞ (Ω) if for every 𝛼 ∈ Z+𝑛 the functions 𝜕𝑥𝛼 𝜑 𝑛
converge uniformly to zero on every compact subset of Ω. Ø
A partition of unity in C ∞ (Ω) subordinate to a covering Ω = Ω 𝛼 of Ω by
𝛼
open sets Ω 𝛼 is a family of functions 𝜑 𝛽 ∈ C ∞ (Ω) with the following properties:
∀𝛽, ∃𝛼 such that supp 𝜑 𝛽 ⊂ Ω 𝛼 ; to each compact set 𝐾 ⊂ Í Ω there are only finitely
many functions 𝜑 𝛽 such that 𝐾 ∩ supp 𝜑 𝛽 ≠ ∅; and, lastly, 𝛽 𝜑 𝛽 ≡ 1 in Ω.
Let 𝐾 be an arbitrary compact subset of Ω. We denote by Cc∞ (𝐾) the linear
subspace of C ∞ (Ω) consisting of the functions 𝑓 such that supp 𝑓 ⊂ 𝐾 and by
Cc∞ (Ω) the union of the Cc∞ (𝐾) as 𝐾 ranges over the set of all compact subsets of
Ω. The elements of Cc∞ (Ω) are commonly referred to as test-functions. A sequence
{𝜑 𝑛 } 𝑛=0,1,2,... of test-functions converges to zero in Cc∞ (Ω) if the 𝜑 𝑛 converge to
zero in C ∞ (Ω) and if, moreover, there is a compact subset 𝐾 of Ω such that 𝜑 𝑛
∈ Cc∞ (𝐾) for all 𝑛. An important property is the existence of partitions of unity in
Cc∞ (Ω) subordinate to any open covering of Ω.
A function 𝑓 : Ω −→ C is said to be real-analytic or, equivalently, of class C 𝜔 ,
if 𝑓 ∈ C ∞ (Ω) and if to every compact set 𝐾 ⊂ Ω there is a constant 𝐶𝐾 > 0 such
that
∀𝛼 ∈ Z+𝑛 , max 𝜕𝑥𝛼 𝑓 ≤ 𝐶𝐾| 𝛼 |+1 𝛼! (1.2.1)
𝐾

in the notation (1.1.1)–(1.1.2). This is equivalent to saying that the Taylor expansion
of 𝑓 about an arbitrary point 𝑥 ◦ of Ω converges uniformly to 𝑓 in some neighborhood
of 𝑥 ◦ .
Proposition 1.2.1 For 𝑓 ∈ C ∞ (Ω) to be real-analytic in Ω it is necessary and
sufficient that to each compact set 𝐾 ⊂ Ω there be a constant 𝐶𝐾 > 0 such that
∫ 21
∀𝛼 ∈ Z+𝑛 , 𝛼 2
|D 𝑓 (𝑥)| d𝑥 ≲ 𝐶𝐾| 𝛼 | |𝛼|!. (1.2.2)
𝐾

Proof The necessity being evident we concentrate on the proof of the sufficiency.
We make use of the Fourier transform of test-functions:
∫
𝑢 (𝜉) =
b e−𝑖 𝑥· 𝜉 𝑢 (𝑥) d𝑥, 𝑢 ∈ Cc∞ (R𝑛 ) , (1.2.3)
R𝑛

of the inversion formula,

∫
𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 b
𝑢 (𝜉) d𝜉, (1.2.4)
R𝑛
1.2 Smooth, Real-analytic, Holomorphic Functions 7

and of the Parseval–Plancherel identity,

∫ ∫
|𝑢 (𝑥)| 2 d𝑥 = (2𝜋) −𝑛 𝑢 (𝜉)| 2 d𝜉.
|b (1.2.5)
R𝑛

We are going to make repeated use of the obvious consequence of (1.2.4):

∫
D 𝛼 𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 𝜉 𝛼 b
𝑢 (𝜉) d𝜉. (1.2.6)
R𝑛

We introduce the Laplacian of 𝑢, Δ𝑢 = − 𝑛𝑗=1 D2𝑗 𝑢. On the one hand we have,

Í
by (1.2.6),
𝑁 ∫ 𝑁
1 − Δ2𝑥 𝑢 (𝑥) = (2𝜋) −𝑛 𝑢 (𝜉) 1 + |𝜉 | 2 d𝜉
e𝑖 𝑥· 𝜉 b
R𝑛

whence, by (1.2.5),
∫ 𝑁 2 ∫ 2𝑁
1 − Δ2𝑥 𝑢 (𝑥) d𝑥 = (2𝜋) −𝑛 𝑢 (𝜉)| 2 1 + |𝜉 | 2
|b d𝜉. (1.2.7)
R𝑛

On the other hand, if 𝑁 ≥ 21 (𝑛 + 1) the Cauchy–Schwarz inequality and (1.2.4),

(1.2.5), (1.2.7) entail
∫ 𝑁 d𝜉
|𝑢 (𝑥)| ≲ 1 + |𝜉 | 2 |b 𝑢 (𝜉)| 12 (𝑛+1)
R𝑛
1 + |𝜉 | 2
21 ! 12
∫ 2𝑁 ∫ 𝑁 2
≲ 1 + |𝜉 | 2 𝑢 (𝜉)| 2 d𝜉 =
|b 1 − Δ2𝑦 𝑢 (𝑦) d𝑦
R𝑛 R𝑛

for every 𝑥 ∈ R𝑛 .
Now let 𝑥 ◦ ∈ Ω be arbitrary and 𝑈 ⊂⊂ Ω be a neighborhood of 𝑥 ◦ ; we take 𝐾
to be the closure of 𝑈. Let 𝜒 ∈ Cc∞ (𝐾), 0 ≤ 𝜒 ≤ 1 everywhere and 𝜒 ≡ 1 in a
neighborhood 𝑉 ⊂⊂ 𝑈 of 𝑥 ◦ . Applying the preceding inequality to 𝑢 = 𝜒D 𝛼 𝑓 yields
easily

∑︁ ∫ 21
𝛼 𝛼 𝛼+𝛽 2
sup |D 𝑓 | ≤ ∥ 𝜒D 𝑓 ∥ 𝐿 ∞ ≲ D 𝑓 (𝑥) d𝑥
𝑉 |𝛽 | ≤𝑛+2 𝐾

≲ 𝐶𝐾| 𝛼 |+𝑛+2 (|𝛼| + 𝑛 + 2)! ≲ 𝐵 | 𝛼 | 𝛼!

for 𝐵 > 0 sufficiently large. □

8 1 Functions and Differential Operators in Euclidean Space

Remark 1.2.2 Thus the inequalities in (1.2.1) and (1.2.2) tell us that

∥D 𝛼 𝑓 ∥ 𝐿 𝑝 (𝐾) ≲ 𝐶𝐾| 𝛼, |𝑝 𝛼! (1.2.8)

for 𝑝 = 2, ∞. It is a fairly simple exercise to show that they are equivalent to (1.2.8)
for 𝑝 = 1. It is then a consequence of the classical Riesz convexity theorem (see
[Stein-Weiss, 1971]) that (1.2.8) is valid for all 𝑝 ∈ [1, +∞].

The real-analytic functions in Ω form an algebra (with respect to ordinary addition

and multiplication) that will be denoted by C 𝜔 (Ω). If Ω is a domain, meaning that
Ω is open and connected, a function 𝑓 ∈ C 𝜔 (Ω) cannot vanish to infinite order at
any point of Ω unless 𝑓 ≡ 0. This precludes the availability of partitions of unity
in C 𝜔 (Ω) subordinate to nontrivial open coverings of Ω. Classical examples of C ∞
functions 𝑓 that are real-analytic in R𝑛 \ {0} but not at the origin are the exponentials
exp (− |𝑥| −𝑠 ), 𝑠 > 0. Of course, C 𝜔 (Ω) ∩ Cc∞ (Ω) = {0}.
We denote by 𝜕𝑧¯ (or simply 𝜕) the Cauchy–Riemann operator: 𝜕 𝑓 = 𝜕𝜕𝑧¯𝑓1 , ..., 𝜕𝜕𝑧¯𝑓𝑛

with 𝜕𝜕𝑧¯𝑓𝑗 = 12 𝜕𝑥𝜕𝑓
+ 𝑖 𝜕𝑓
𝜕𝑦 , 𝑥 𝑗 = Re 𝑧 𝑗 , 𝑦 𝑗 = Im 𝑧 𝑗 . We have 𝜕 𝑧 𝑓 = 𝜕𝑓
𝜕𝑧1
, ..., 𝜕𝑓
𝜕𝑧
𝑗 𝑗
𝑛
1 𝜕𝑓
𝜕𝑓 𝜕𝑓
with 𝜕𝑧 𝑗
= 2 𝜕𝑥 𝑗 − 𝑖 𝜕𝑦 𝑗 . If Ω C is an open subset of C𝑛 we shall denote by O ΩC

the space of holomorphic functions in ΩC : 𝑓 ∈ O ΩC if 𝑓 ∈ C ∞ ΩC and if

𝜕 𝑓 ≡ 0 in ΩC ; O (C𝑛 ) is the space of entire functions in C . The natural topology

𝑛

on the vector space O Ω is that of normal convergence

C
, i.e., uniform convergence
on each compact subset of ΩC . This makes O ΩC a complete metrizable, locally
convex topological vector space, which is to say, a Fréchet space.
The adjective “holomorphic” is synonymous with “complex-analytic”: 𝑓 ∈
C ∞ ΩC is complex-analytic if and only if every point 𝑧◦ ∈ ΩC is the center
of an open polydisk
n o
Δ𝑅(𝑛) (𝑧◦ ) = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 − 𝑧◦𝑗 < 𝑅 𝑗 , 𝑗 = 1, ..., 𝑛 ⊂ ΩC (1.2.9)

(𝑅 𝑗 > 0) in which a power series 𝛼∈Z+𝑛 𝑐 𝛼 (𝑧 − 𝑧 ◦ ) 𝛼 converges uniformly to 𝑓

Í

(𝑐 𝛼 ∈ C). This is equivalent to saying that to every compact subset 𝐾 of ΩC there is

a constant 𝐶𝐾 > 0 such that max 𝜕𝑧𝛼 𝑓 ≲ 𝐶𝐾| 𝛼 | 𝛼! for all 𝛼 ∈ Z+𝑛 [cf. (1.2.1)]. If this
𝐾
is the case, the multivariable Cauchy formula yields
1
𝑐 𝛼 = 𝜕𝑧𝛼 𝑓 (𝑧◦ ) (1.2.10)
∮ ∮ 𝛼!
1 𝑓 (𝑧)d𝑧1 · · · d𝑧 𝑛
= ··· 𝛼1 +1 ,
(2𝜋𝑖) 𝑛 ◦
𝑧1 − 𝑧1 · · · (𝑧 𝑛 − 𝑧 ◦𝑛 ) 𝛼𝑛 +1
◦
| 𝑧1 −𝑧1◦ | =𝑟1 |𝑧𝑛 −𝑧𝑛 |=𝑟𝑛

where 0 < 𝑟 𝑗 < 𝑅 𝑗 .

1.2 Smooth, Real-analytic, Holomorphic Functions 9

The link between real and complex-analyticity is well known. A continuous

function 𝑓 in an open subset Ω of R𝑛 is real-analytic in Ω if and only if there is
an open subset ΩC of C𝑛 , Ω ⊂ ΩC ∩ R𝑛 , and a holomorphic function in ΩC whose
restriction to Ω is equal to 𝑓 .

Remark 1.2.3 An important difference between real-analytic and complex-analytic

functions of several variables is that a continuous function 𝑓 of the complex variables
𝑧1 .., 𝑧 𝑛 (𝑛 ≥ 2) is holomorphic if it is holomorphic with respect to each variable 𝑧 𝑗
separately (see, e.g., [Hörmander, 1966], Theorem 2.2.8, [Stein-Shakarchi, 2011],
Vol. IV, p. 277). Nothing of the sort is true for real-analytic functions.

There is a natural topology on C 𝜔 (Ω); it suffices to describe it when

ÖΩ is con-
nected. When Ω is not connected, C 𝜔 (Ω) is isomorphic to the product C 𝜔 (Ω 𝜄 ),
𝜄
with Ω 𝜄 ranging over the connected components of Ω, and carries the product space
topology. Thus suppose that Ω is a domain in R 𝜔. If 𝑈 is a domain in C that con-
𝑛 C 𝑛

tains Ω, the restriction

mapping O 𝑈 −→ C (Ω) is injective; denote its range
C

by C 𝜔 Ω, 𝑈 C ; we equip C 𝜔 Ω, 𝑈 C with the Fréchet space structure transferred

from O 𝑈 C . Because Ω is connected, C 𝜔 (Ω) is equal to the union of the subspaces
C 𝜔 Ω, 𝑈 C . Formally, the topology of C 𝜔 (Ω) is locally convex: the convex open
sets form a basis of this topology. They are defined as follows: a convex subset ℭ
of C 𝜔 (Ω) is open if and only if ℭ ∩ C 𝜔 Ω, 𝑈 C is open in C 𝜔 Ω, 𝑈 C for every
domain 𝑈 C in C𝑛 that contains Ω (a nonconvex subset with this property is not
necessarily open). Practically, the following two properties are useful:
(1) A sequence { 𝑓 𝜈 } 𝜈=1,2,... converges in C 𝜔 (Ω) if and only if there is an open

subset 𝑈 C of C𝑛 , Ω ⊂ 𝑈 C , such that, whatever 𝜈, 𝑓 𝜈 ∈ C 𝜔 Ω, 𝑈 C and the
sequence { 𝑓 𝜈 } 𝜈=1,2,... converges in C 𝜔 Ω, 𝑈 C .
(2) Let 𝑬 be a Fréchet space. For a linear map 𝜙 : C 𝜔 (Ω) −→ 𝑬 to be continuous
it is necessary and sufficient that its restriction to each subspace C 𝜔 Ω, 𝑈 C be
continuous.
Among the function algebras intermediate between smooth and analytic, the most
frequently encountered (and perhaps the most significant in linear PDE theory) are
the Gevrey classes. They are defined by generalizing (1.2.1). A function 𝑓 : Ω −→ C
is of Gevrey class 𝑠 (𝑠 > 1) and said to belong to the space 𝐺 𝑠 (Ω) if 𝑓 ∈ C ∞ (Ω)
and if to every compact set 𝐾 ⊂ Ω there is a constant 𝐶𝐾 > 0 such that

∀𝛼 ∈ Z+𝑛 , max 𝜕𝑥𝛼 𝑓 ≲ 𝐶𝐾| 𝛼 | (𝛼!) 𝑠 . (1.2.11)

𝐾

We posit 𝐺 1 (Ω) = C 𝜔 (Ω). The topology of 𝐺 𝑠 (Ω) can be defined by us-

ing the properties (1.2.11): select an exhaustive sequence of compact subsets 𝐾 𝜈
(𝜈 = 1, 2, ...) of Ω such that 𝐾 𝜈 ⊂ 𝐾 𝜈+1 and Ω = ∞
Ð
𝜈=1 𝐾 𝜈 . To each monotone
increasing sequence of numbers 𝐶𝜈 > 0 we can associate the linear subspace
𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... of 𝐺 𝑠 (Ω) consisting of the functions 𝑓 satisfying (1.2.11)
with 𝐾 = 𝐾 𝜈 and 𝐶𝐾𝜈 = 𝐶𝜈 for every 𝜈. Assuming that (1.2.11) holds we can regard
10 1 Functions and Differential Operators in Euclidean Space

−|𝛼| −𝑠
𝑓 ↦→ sup 𝐶𝐾 (𝛼!) max 𝜕𝑥𝛼 𝑓
𝛼∈Z+𝑛 𝐾

as a continuous seminorm on 𝐺 𝑠 (Ω). These seminorms with 𝐾 = 𝐾 𝜈 and 𝐶 𝐾𝜈 = 𝐶𝜈 ,

𝜈 = 1, 2, ..., define a Fréchet space structure on 𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... ; 𝐺 𝑠 (Ω)
is the union of these Fréchet spaces as {𝐶𝜈 } 𝜈=1,2,... ranges over the set of all
possible increasing sequences of positive numbers. The locally convex “natural”
topology of 𝐺 𝑠 (Ω) is then the inductive limit of the topologies of the subspaces
𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... : a convex subset of 𝐺 𝑠 (Ω) is open if and only if its intersection
with each Fréchet subspace 𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... is open in the latter. A linear map of
𝐺 𝑠 (Ω) into an arbitrary locally convex
space 𝑬 is continuous if and only if its restric-
tion to every 𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... is continuous. A sequence 𝑓 𝑗 𝑗=1,2,... converges

to zero in 𝐺 𝑠 (Ω) if and only if 𝑓 𝑗 ∈ 𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... for some {𝐶𝜈 } 𝜈=1,2,... and
all 𝑗, and if 𝑓 𝑗 → 0 in 𝐺 𝑠 Ω; {𝐶𝜈 } 𝜈=1,2,... .
It is easily seen that the definition of the topology of 𝐺 𝑠 (Ω) is independent of
the choice of the (exhaustive) sequence of compact subsets 𝐾 𝜈 . It is also easily seen
that, when 𝑠 = 1, this topology is identical to that defined on C 𝜔 (Ω) through the
holomorphic extensions: the inequalities (1.2.11) determine a domain 𝑈 C ⊃ 𝐾 in C𝑛
to which 𝑓 can be extended holomorphically and, conversely, the Cauchy inequalities
in O 𝑈 C yield (1.2.11).

1.2.2 Germs and the language of Sheaf Theory

It is sometimes convenient, or even unavoidable, to deal with the germs of functions

rather than with the functions themselves. In this subsection we rapidly recall the
notion of germ and the closely related notions of presheaf and sheaf.
We define the terms at a very general level: let X be a topological space (often
referred to as the base) and let 𝔒 (X) be a basis of the topology of X, i.e., a family
of open subsets of X such that every open set belonging to 𝔒 (X) is the union of sets
belonging to 𝔒 (X). In Parts I and II of this book X will generally be an open subset
Ω of R𝑛 or of C𝑛 , in which case 𝔒 (X) might be the family of bounded open subsets
(or of balls, polydisks, etc.) of Ω. We suppose given a map 𝔒 (X) ∋ 𝑈 −→ 𝑭 (𝑈).
For the moment we allow 𝑭 (𝑈) to denote a set of mathematical objects of a very
general “nature”: belonging to a given category, loosely speaking, such as sets of a
certain kind, say, closed or compact set; or analytic subvarieties (introduced in Ch.
14); or functions, say smooth, analytic, Gevrey or of any other kind; or generalized
functions of the type considered in the sequel (e.g., distributions introduced in Ch.
2, hyperfunctions introduced in Ch. 7, etc.). We suppose also that for each pair 𝑈,
𝑉 of open sets belonging to 𝔒 (X) such that 𝑉 ⊂ 𝑈 we are given a restriction
map 𝜌𝑈 𝑉 : 𝑭 (𝑈) −→ 𝑭 (𝑉), with the natural properties of restrictions, namely: if
𝑊 ∈ 𝔒 (X) is such that 𝑊 ⊂ 𝑉 ⊂ 𝑈 then 𝜌𝑊 𝑉
◦ 𝜌𝑈
𝑉 = 𝜌𝑊 .
𝑈
1.2 Smooth, Real-analytic, Holomorphic Functions 11

Example 1.2.4 Take 𝑭 (𝑈) to be the family of all closed subsets of 𝑈 ∈ 𝔒 (X). If
𝑈, 𝑉 are open sets belonging to 𝔒 (X) such that 𝑉 ⊂ 𝑈 then, for every closed subset
𝐸 of 𝑈, we take 𝜌𝑈 𝑉 𝐸 = 𝐸 ∩ 𝑉, a closed subset of 𝑉 and therefore an element of
𝑭 (𝑉).

Example 1.2.5 Take X = Ω, an open subset of R𝑛 , and for any open subset 𝑈 of Ω,
let 𝑭 (𝑈) = C ∞ (𝑈) [resp., C 𝜔 (𝑈)]. If 𝑈, 𝑉 are open subsets of Ω such that 𝑉 ⊂ 𝑈
∞
𝑉 𝑓 = 𝑓 | 𝑉 for every 𝑓 ∈ C (𝑈) [resp., 𝑓 ∈ C (𝑈)].
we set 𝜌𝑈 𝜔

In the general case, the family of sets {𝑭 (𝑈)}𝑈 ∈𝔒(X) together with the family of

maps 𝜌𝑈 𝑉 𝑈,𝑉 ∈𝔒(X),𝑉 ⊂𝑈 is called a presheaf on (or over) the space X, which we

prefer to denote by 𝑭 (𝑈) , 𝜌𝑈 𝑉 .

Remark 1.2.6 In the sequel we will mostly deal with presheaves 𝑭 (𝑈) , 𝜌𝑈 𝑉 in
which the sets 𝑭 (𝑈) carry some algebraic or algebraic-topological structure: they
might be groups or rings, most often, vector spaces, and more particularly, topo-
logical vector spaces, topological algebras, etc.. The restriction maps 𝜌𝑈 𝑉 will be
morphisms for that structure. Likewise, the sheaf maps will be morphisms for the
structure: group or ring homomorphisms, continuous linear maps, continuous alge-
bra homomorphisms, etc.

We associate to every point 𝑥 ∈ X the set e 𝑭 𝑥 of pairs (𝑈, 𝒆𝑈 ), 𝑈 ∈ 𝔒 (X) such

that 𝑥 ∈ 𝑈, 𝒆𝑈 ∈ 𝑭 (𝑈). Two such pairs (𝑈, 𝒆𝑈 ) and (𝑉, 𝒆 𝑉 ) will be regarded as
representing the same germ at 𝑥 if there is a 𝑊 ∈ 𝔒 (X) such that 𝑥 ∈ 𝑊 ⊂ 𝑈 ∩ 𝑉
and 𝜌𝑈𝑊 𝒆𝑈 = 𝜌𝑊 𝒆 𝑉 . This is an equivalence relation 𝔊𝑥 between pairs (𝑈, 𝒆𝑈 ) ∈ 𝑭 𝑥 .
𝑉 e
We introduce the quotient set F 𝑥 = 𝑭 𝑥 /𝔊𝑥 . Given 𝑈 ∈ 𝔒 (X) such that 𝑥 ∈ 𝑈 the
e
quotient map e 𝑭 𝑥 −→ F 𝑥 induces a map 𝜌𝑈 𝑥 : 𝑭 (𝑈) −→ F 𝑥 ; for every 𝒆𝑈 ∈ 𝑭 (𝑈),
𝜌 𝑥 (𝒆𝑈 ) is the coset of 𝒆𝑈 mod 𝔊𝑥 .
𝑈

Definition 1.2.7 If 𝑥 ∈ 𝑈 ∈ 𝔒 (X) and 𝒆𝑈 ∈ 𝑭 (𝑈) the element 𝜌𝑈

𝑥 (𝒆𝑈 ) of F 𝑥 is
called the germ of 𝒆𝑈 at 𝑥.

Example 1.2.8 The germ of an analytic function at 𝑥 ◦ ∈ R𝑛 can be identified with its
Taylor expansion at 𝑥 ◦ since the sum of the series converges uniformly to the function
in a full neighborhood of 𝑥 ◦ . This is of course not the case for a smooth function
[even when its Taylor expansion converges at every point, as for exp −1/𝑥 2 ].

Back to the general case of a presheaf 𝑭 (𝑈) , 𝜌𝑈 𝑉 over X we now consider the
disjoint union
•
F=
+
𝑥 ∈X
F𝑥 .

There is a natural map 𝜋 : F −→ X (to which we refer as the base projection): if

𝒆 ∈ F 𝑥 then 𝜋 (𝒆) = 𝑥. A section of F over a subset 𝑆 of X is a map 𝜎 : 𝑆 −→ F
such that 𝜋 (𝜎 (𝑥)) = 𝑥 for each 𝑥 ∈ 𝑆. If 𝑈 ∈ 𝔒 (X) there is a natural map
𝜑𝑈 : 𝑭 (𝑈) −→ F: for each 𝒆𝑈 ∈ 𝑭 (𝑈) and each 𝑥 ∈ 𝑈 we define 𝜑𝑈 (𝒆𝑈 )| 𝑥 to be
the germ of 𝒆𝑈 at 𝑥; 𝑥 ↦→ 𝜑𝑈 (𝒆𝑈 )| 𝑥 is a section of F over 𝑈.
12 1 Functions and Differential Operators in Euclidean Space

Definition 1.2.9 By the sheaf topology on the set F we shall mean the weakest
topology that renders continuous every section 𝑈 ∋ 𝑥 ↦→ 𝜑𝑈 (𝒆𝑈 )| 𝑥 ∈ F, with
𝑈 ⊂ 𝔒 (X) and 𝒆𝑈 ∈ 𝑭 (𝑈). We refer to the topological space Fequipped with its
sheaf topology as the sheaf defined by the presheaf 𝑭 (𝑈) , 𝜌𝑈
𝑉 .

−1
The subset F 𝑥 = 𝜋 (𝑥) is called the stalk of the sheaf F at 𝑥 ∈ X.
Given two sheaves F ( 𝑗) ( 𝑗 = 1, 2) over the same base X, by a sheaf map (strictly
speaking, a sheaf morphism) of F (1) into F (2) we shall mean a continuous map
F (1) −→ F (2) which maps F 𝑥(1) into F 𝑥(2) for every 𝑥 ∈ X. Such a map is a sheaf
isomorphism if it is a homeomorphism.
In view of Definition 1.2.9 a basis of the sheaf topology of F consists of the
open sets 𝜑𝑈 (𝒆𝑈 )| 𝑥 ; 𝑥 ∈ 𝑈 as 𝑈 ranges over 𝔒 (X) and 𝒆𝑈 ranges over 𝑭 (𝑈).
A section 𝜎 of F over a subset 𝑆 of X is continuous at a point 𝑥 ◦ ∈ 𝑆 if there is a
𝑈 ∈ 𝔒 (X) such that 𝑥 ◦ ∈ 𝑈 and an element 𝒆𝑈 ∈ 𝑭 (𝑈) such that 𝜎 (𝑥) = 𝜑𝑈 (𝒆𝑈 )| 𝑥
for every 𝑥 ∈ 𝑈 ∩ 𝑆. As usual, 𝜎 is said to be continuous in 𝑆 if it is continuous at
every point of 𝑆; if this is so, there is necessarily an open subset Ω of X such that
𝑆 ⊂ Ω, to which the section 𝜎 extends (as a continuous section over Ω). Another
noteworthy difference between true, say continuous, functions in Ω and a continuous
section over Ω of the sheaf of their germs is that the subset of Ω in which a continuous
section vanishes is an open subset of Ω.

Example 1.2.10 Let E be an arbitrary set and for every open subset 𝑈 of X let E (𝑈)
be the set of constant maps 𝑈 −→ E. Take the maps 𝜌𝑈 𝑉 to be the true restrictions.
The continuous sections of the associated sheaf over an arbitrary open subset Ω of
X are the locally constant maps Ω −→ E.

Example 1.2.11 Let the set E be equipped with the discrete topology (meaning that
each point in E is an open set) and let E (𝑈) be the set of continuous maps 𝑈 −→ E,
with the natural restrictions as maps 𝜌𝑈 𝑉 . In this case also, the continuous sections
of the associated sheaf over an arbitrary open subset Ω of X are the locally constant
maps Ω −→ E.

In general, the sheaf topology is not Hausdorff.

Example 1.2.12 Let X = R𝑛 and for each open subset 𝑈 of R𝑛 let E (𝑈) = C ∞ (𝑈);
if 𝑉 ⊂ 𝑈 is also open we define 𝜌𝑈 𝑉 𝑓 = 𝑓 | 𝑉 for every 𝑓 ∈ C ∞ (𝑈). Let us denote
∞
∞
by C (R ) the sheaf associated to the presheaf C (𝑈) , 𝜌𝑈
𝑛
𝑉 . Consider then a
function 𝜑 ∈ C ∞ (R𝑛 ) that vanishes identically in a half-space, say 𝑥1 < 0, but
𝜑 (𝑥) ≠ 0 for all 𝑥 = (𝑥1 , ..., 𝑥 𝑛 ), 𝑥1 > 0. Let 𝜑 𝑥 denote the germ of 𝜑 at 𝑥. If 𝑈 ⊂ R𝑛
is an arbitrary open set then 𝑈 e𝜑 = {𝜑 𝑥 } 𝑥 ∈𝑈 is an open subset of C∞ (R𝑛 ). Note that
𝜑0 ≠ 0; indeed, 𝜑0 = 0 would mean that the function 𝜑 vanishes identically in some
neighborhood of 0. Note also that every neighborhood of 𝜑0 contains a set 𝑈 e𝜑 with 𝑈
open, 0 ∈ 𝑈. On the other hand, let 𝜓 ∈ C ∞ (R𝑛 ) be an arbitrary function vanishing
identically in some neighborhood of the origin and in the half-space 𝑥1 < 0. We
have 𝜓0 = 0; since 𝜑 𝑥 = 𝜓 𝑥 =0 if 𝑥1 < 0 we have 𝑈 e𝜑 ∩ 𝑉e𝜓 ≠ ∅ whatever the open
set 𝑉 ∋ 0.
1.2 Smooth, Real-analytic, Holomorphic Functions 13

Example 1.2.13 Let us duplicate the construction in Example 1.2.12, replacing C ∞

with C 𝜔 : we define thus the sheaf C 𝜔 (R𝑛 ) of germs of C 𝜔 functions in R𝑛 . If the
germs of two C 𝜔 functions 𝜑, 𝜓 differ at 𝑥 ◦ then there is a neighborhood 𝑈 of 𝑥 ◦
such that 𝜑 𝑥 ≠ 𝜓 𝑥 for every 𝑥 ∈ 𝑈. In this case the sheaf topology is Hausdorff.

In accordance with custom we denote by Γ (𝑈, F) the set of continuous sections

of the sheaf F over the open subset 𝑈 of X. Given another open subset of X, 𝑉 ⊂ 𝑈,
we can define 𝜌𝑈 𝑉 : Γ (𝑈, F) −→ Γ (𝑉, F) to be the natural
restriction mapping
Γ (𝑈, F) ∋ 𝜎 −→ 𝜎| 𝑉 . Thus we get a presheaf 𝑭 (𝑈) , 𝜌𝑈 𝑉 with special properties:
for one, the sets 𝑭 (𝑈) are defined for every open subset of X; but more importantly
the following property holds:
(Sh) Let be given a (possibly infinite) family {𝑈 𝛼 } 𝛼∈ 𝐴 of open sets belonging to
𝔒 (X) and for each index 𝛼 ∈ 𝐴 a section 𝜎𝛼 ∈ 𝑭 (𝑈 𝛼 ) such that, if 𝛽 ∈ 𝐴 and
𝑈𝛼 𝑈
𝑈 𝛼 ∩ 𝑈𝛽 ≠ ∅, then 𝜌𝑈 𝛼 ∩𝑈𝛽
𝜎𝛼 = 𝜌𝑈𝛽𝛼 ∩𝑈𝛽 𝜎𝛽 . When this condition is satisfied
Ø
there is a unique section 𝜎 ∈ 𝑭 (𝑈), 𝑈 = 𝑈 𝛼 , such that 𝜌𝑈
𝑈
𝛼
𝜎 = 𝜎𝛼 .
𝛼∈ 𝐴
(Loosely speaking, if they agree on overlaps the 𝜎𝛼 can be patched together
unambiguously).

When (Sh) holds it is common to say that the presheaf 𝑭 (𝑈) , 𝜌𝑈 𝑉 is a sheaf,
identified, of course, with F.
When the sets 𝑭 (𝑈) carry an algebraic-topological structure (cf. Remark 1.2.6)
the same will be true of the stalks F 𝑥 and of the sets of sections over a subset of X,
in which case we will talk of sheaves of groups, topological vector spaces, etc. The
usual concepts and operations of algebra make sense for such sheaves (over one and
the same base) and have self-evident definitions: “stalkwise” product, direct sum,
subsheaf, quotient sheaf, sheaf homomorphism, isomorphism, etc. Equally evident
are the notions of pullback or pushforward of sheaves under a map of one base space
into another.
Let us now assume that 𝔒 (X) is indeed the family of all open subsets of X. This
enables us to duplicate for an arbitrary subset 𝑆 ≠ ∅ of X the definition of germs as
done for single points (Definition 1.2.7). We introduce the disjoint union
•
𝑭 (𝑆) =
e
+
𝑆 ⊂𝑈 ∈𝔒(X)
𝑭 (𝑈) .

The elements of e 𝑭 (𝑆) are pairs (𝑈, 𝒆𝑈 ), 𝑈 ⊂ X open such that 𝑆 ⊂ 𝑈, 𝒆𝑈 ∈ 𝑭 (𝑈).
Two such pairs (𝑈, 𝒆𝑈 ) and (𝑉, 𝒆 𝑉 ) will be regarded as representing the same germ
at 𝑆 if there is 𝑊 ∈ 𝔒 (X) such that 𝑆 ⊂ 𝑊 ⊂ 𝑈 ∩ 𝑉 and 𝜌𝑈 𝑊 𝒆𝑈 = 𝜌𝑊 𝒆 𝑉 . This is an
𝑉

equivalence relation 𝔊𝑆 between pairs of the type (𝑈, 𝒆𝑈 ). We shall use the notation
F (𝑆) = e 𝑭 (𝑆) /𝔊𝑆 .

Definition 1.2.14 The elements of F (𝑆) are called germs at 𝑆 of the continuous
sections of the sheaf F [or of the objects that make up the sets 𝑭 (𝑈), 𝑈 open,
𝑆 ⊂ 𝑈].
14 1 Functions and Differential Operators in Euclidean Space

If 𝑥 ∈ X we shall continue to use the notation F 𝑥 rather than F ({𝑥}).

Example 1.2.15 Let 𝑆 ◦ be an arbitrary subset of the topological space X and 𝑥 ◦ be a
point of the closure 𝑆 ◦ of 𝑆 ◦ . The germ of 𝑆 ◦ at 𝑥 ◦ can be thought of as the family of
subsets 𝑆 of X such that 𝑆 ∩ 𝑈 = 𝑆 ◦ ∩ 𝑈 for some open subset 𝑈 of X containing 𝑥 ◦ .

1.3 Differential Operators with Smooth Coefficients

A linear partial differential operator with coefficients 𝑐 𝛼 ∈ C ∞ (Ω) (in the sequel,
often abbreviated to differential operator or even to PDO – we shall practically never
come across a linear partial differential equation, i.e., a PDE, whose coefficients are
not smooth) acts on functions 𝑓 ∈ C ∞ (Ω):
∑︁ ∑︁
𝑃 (𝑥, 𝜕𝑥 ) 𝑓 (𝑥) = 𝑐 𝛼 (𝑥) 𝜕𝑥𝛼 𝑓 (𝑥) or 𝑃 (𝑥, D 𝑥 ) 𝑓 = 𝑐 𝛼 (𝑥) D 𝑥𝛼 𝑓 (𝑥) .
𝛼 𝛼
(1.3.1)
The sums in (1.3.1) are locally finite: in every compact subset of Ω only finitely
many 𝑐 𝛼 do not vanish identically; 𝑓 ↦→ 𝑃 (𝑥, D) 𝑓 is a linear continuous endomor-
phism of C ∞ (Ω). It is not difficult to prove that the obvious inclusion

supp 𝑃 (𝑥, D) 𝑓 ⊂ supp 𝑓 (1.3.2)

characterizes linear PDOs among all continuous linear endomorphisms of C ∞ (Ω)

[actually among all linear endomorphisms of C ∞ (Ω), but this is more difficult to
prove; see [Peetre, 1960]]. Formula (1.3.2) implies that 𝑃 (𝑥, D) maps Cc∞ (Ω) into
itself.
The transpose of 𝑃 (𝑥, D) is the differential operator
∑︁
𝑃⊤ (𝑥, D 𝑥 ) 𝑓 (𝑥) = (−1) | 𝛼 | D 𝑥𝛼 (𝑐 𝛼 𝑓 ) (𝑥) (1.3.3)
𝛼

whereas the adjoint of 𝑃 (𝑥, D) is the differential operator

∑︁
𝑃∗ (𝑥, D 𝑥 ) 𝑓 (𝑥) = D 𝑥𝛼 (𝑐 𝛼 𝑓 ) (𝑥) (1.3.4)
𝛼

(𝑐 𝛼 : complex
√ conjugate of 𝑐 𝛼 ). Formula (1.3.3) is also valid with 𝜕𝑥 in the place of
D 𝑥 = − −1𝜕𝑥 . If we apply the Leibniz rule (1.1.3) we get

∑︁ ∑︁ 𝛼
𝑃⊤ (𝑥, D 𝑥 ) 𝑓 = (−1) | 𝛼 |
𝛽 𝛼−𝛽
D𝑥 𝑐 𝛼 D𝑥 𝑓 , (1.3.5)
𝛼 𝛽⪯𝛼
𝛽
∑︁ ∑︁ 𝛼
𝑃∗ (𝑥, D 𝑥 ) 𝑓 =
𝛽 𝛼−𝛽
D𝑥 𝑐 𝛼 D𝑥 𝑓 . (1.3.6)
𝛼 𝛽⪯𝛼
𝛽

The polynomial in the variables 𝜉 𝑗 , 𝑗 = 1, ..., 𝑛,

1.3 Differential Operators with Smooth Coefficients 15
∑︁
𝑃 (𝑥, 𝜉) = 𝑐 𝛼 (𝑥) 𝜉 𝛼 , (1.3.7)
| 𝛼 | ≤𝑚

is called the total symbol (or simply the symbol) of 𝑃 (𝑥, D).
The linear PDOs with C ∞ coefficients in Ω form a subring of the ring (with respect
to addition and composition) of continuous linear operators C ∞ (Ω) −→ C ∞ (Ω). If
𝑃 𝑘 (𝑥, D 𝑥 ) (𝑘 = 1, 2) are two linear PDOs with C ∞ coefficients in Ω then the symbol
of their composite 𝑃1 (𝑥, D 𝑥 ) 𝑃2 (𝑥, D 𝑥 ) is
∑︁ 1
(𝑃1 #𝑃2 ) (𝑥, 𝜉) = 𝜕 𝜉𝛼 𝑃1 (𝑥, 𝜉) D 𝑥𝛼 𝑃2 (𝑥, 𝜉) . (1.3.8)
𝛼∈Z𝑛
𝛼!
+

This is the same as saying that, whatever 𝑓 ∈ C ∞ (Ω),

𝑃1 (𝑥, D 𝑥 ) 𝑃2 (𝑥, D 𝑥 ) 𝑓 = (𝑃1 #𝑃2 ) (𝑥, D 𝑥 ) 𝑓 . (1.3.9)

Proof (of (1.3.9)) The formula is bilinear with respect to 𝑃1 and 𝑃2 ; it is therefore
(𝑘)
sufficient to prove it when 𝑃 𝑘 (𝑥, 𝜉) = 𝑔 𝑘 (𝑥) 𝜉 𝛼 , 𝛼 (𝑘) ∈ Z+𝑛 , 𝑔 𝑘 ∈ C ∞ (Ω),
𝑘 = 1, 2. We see at once that it suffices to take 𝑔1 ≡ 1, 𝛼 (2) = 0. In this case (1.3.9)
reduces to the multivariable Leibniz rule (1.1.3). □
In practice we shall be dealing with a restricted class of differential operators
𝑃 = 𝑃 (𝑥, D). We shall assume that the sum in (1.3.7) is globally finite: there will
always be a smallest integer 𝑚 ≥ 0 such that |𝛼| > 𝑚 =⇒ 𝑐 𝛼 ≡ 0; 𝑚 is called
the order of 𝑃 (𝑥, D 𝑥 ). If the coefficients of 𝑃 (𝑥, D) belong to C 𝜔 (Ω) and if Ω is
connected, the sum in (1.3.7) is always globally finite.
The following definitions are all important in PDE theory:
Definition 1.3.1 If 𝑃 (𝑥, D 𝑥 ) is a differential operator of order 𝑚 in Ω with total
symbol (1.3.7) the homogeneous polynomial 𝑃𝑚 (𝑥, 𝜉) is called the principal sym-
bol of 𝑃 (𝑥, D). The set of points {(𝑥, 𝜉) ∈ Ω × R𝑛 ; 𝑃𝑚 (𝑥, 𝜉) = 0, 𝜉 ≠ 0} is called
the characteristic set of 𝑃 (𝑥, D) and shall be denoted by Char 𝑃.
We identify the tangent space 𝑇𝑥 Ω to Ω at an (arbitrary) point 𝑥 with R𝑛 . In
(1.3.7) we regard 𝜉 = (𝜉1 , ..., 𝜉 𝑛 ) as a covector at 𝑥, which is to say: an element of
the cotangent space to Ω at 𝑥 , 𝑇𝑥∗ Ω, the dual of 𝑇𝑥 Ω (𝑇𝑥∗ Ω can also, evidently, be
identified with R𝑛 ). We can regard 𝑃𝑚 (𝑥, 𝜉) as a function on the cotangent bundle
𝑇 ∗ Ω, the disjoint union of all the linear spaces 𝑇𝑥∗ Ω as 𝑥 ranges over Ω (more on this
in the forthcoming chapters).
Definition 1.3.2 The differential operator 𝑃 (of order 𝑚) is said to be elliptic if
Char 𝑃 = ∅, i.e., if 𝑃𝑚 (𝑥, 𝜉) ≠ 0 for all 𝑥 ∈ Ω, 0 ≠ 𝜉 ∈ R𝑛 .
Let a C 1 hypersurface 𝑆 ⊂ Ω be defined in some neighborhood 𝑈 ⊂ Ω of one
of its points 𝑥 ◦ by an equation 𝑓 (𝑥) = 0 such that 𝑓 ∈ C 1 (𝑈) and d 𝑓 (𝑥) ≠ 0
everywhere in 𝑈. A covector 𝜉 is conormal to 𝑆 at 𝑥 ◦ if 𝜉 𝑗 = 𝑐 𝜕𝑥
𝜕𝑓
𝑗
(𝑥 ◦ ) for some
𝑐 ∈ R and every 𝑗 = 1, ..., 𝑛, or, equivalently, if ⟨𝜉, v⟩ = 0 whatever the vector
v ∈ 𝑇𝑥 ◦ Ω tangent to 𝑆 at 𝑥 ◦ .
16 1 Functions and Differential Operators in Euclidean Space

Definition 1.3.3 The hypersurface 𝑆 ⊂ Ω is said to be noncharacteristic for the

differential operator 𝑃 at 𝑥 ◦ if there is a covector 𝜉 ≠ 0 conormal to 𝑆 at 𝑥 ◦ such that
𝑃𝑚 (𝑥 ◦ , 𝜉) ≠ 0.

The covectors conormal to 𝑆 at 𝑥 ◦ form a one-dimensional vector subspace 𝑁 𝑥∗ ◦ Ω

of 𝑇𝑥∗◦ Ω. It follows that saying that 𝑆 ⊂ Ω is noncharacteristic for 𝑃 at 𝑥 ◦ is the same
as saying that 𝑁 𝑥∗ ◦ Ω ∩ Char 𝑃 = ∅ (recall that 0 ∉ Char 𝑃). If 𝑁 𝑥∗ ◦ Ω\ {0} ⊂ Char 𝑃
the hypersurface 𝑆 ⊂ Ω is said to be characteristic for the operator 𝑃 at 𝑥 ◦ .
Í
Example 1.3.4 Let 𝑋 = 𝑛𝑗=1 𝑎 𝑗 𝜕𝑥𝜕 𝑗 , 𝑎 𝑗 ∈ R, regarded as a differential operator in
R𝑛 (𝑛 ≥ 2). For a smooth hypersurface 𝑆 in R𝑛 to be noncharacteristic for 𝑋 at
𝑥 ◦ ∈ 𝑆 it is necessary and sufficient that 𝑋 not be tangent (i.e., be transverse) to 𝑆 at
𝑥◦.

This book will be mostly concerned with analytic linear PDOs: 𝑃 (𝑥, D) is analytic
if every one of its coefficients 𝑐 𝛼 belongs to C 𝜔 (Ω), in which case each 𝑐 𝛼 can
be extended as a holomorphic function in one and the same neighborhood ΩC of Ω
in C𝑛 . The operator
itself can be extended as the continuous endomorphism of the
space O ΩC of holomorphic functions in ΩC defined in the natural manner:
∑︁
𝑃 (𝑧, 𝜕𝑧 ) 𝑓 (𝑧) = 𝑐 𝛼 (𝑧) 𝜕𝑧𝛼 𝑓 (𝑧) , 𝑓 ∈ O ΩC . (1.3.10)
𝛼

The definitions 1.3.1, 1.3.3, extend naturally to differential operators (1.3.10) by

substituting the complex cotangent spaces to C𝑛 for the real ones:

Definition 1.3.5 Suppose 𝑃 (𝑧, 𝜕𝑧 ) is of order 𝑚 and 𝑃𝑚 (𝑧, 𝜁) is its principal symbol.
The complex characteristic set of 𝑃 (𝑧, 𝜕𝑧 ) is the subset of ΩC × C𝑛 ,

Char 𝑃 = (𝑧, 𝜁) ∈ ΩC × C𝑛 ; 𝜁 ≠ 0, 𝑃𝑚 (𝑧, 𝜁) = 0 . (1.3.11)

A complex-analytic hypersurface 𝑆 C ⊂ ΩC is said to be noncharacteristic for the

operator (1.3.10) at 𝑧 ◦ ∈ 𝑆 C if there is a covector 𝜁 ≠ 0 conormal to 𝑆 C at 𝑧 ◦ such
that 𝑃𝑚 (𝑧 ◦ , 𝜁) ≠ 0.
Chapter 2
Distributions in Euclidean Space

In the first section of this chapter we transition from functions to generalized functions
of the most commonly used kind, the distributions in an open subset Ω of R𝑛 , the
elements of the complex vector space D ′ (Ω). It is again (as in Ch. 1) a matter of
agreeing on notation and a few basic definitions (such as those of support and singular
support), introducing concepts that are well known to analysts but of fundamental
importance, such as the Schwartz space of smooth functions in R𝑛 rapidly decaying
at infinity and its dual, the space of tempered distributions in R𝑛 , and the Fourier
transform as a linear automorphism of each one of these spaces. The first application
of the Fourier transform is the definition of the wave-front set of a distribution 𝑢
in Ω, a subset of phase-space Ω × R𝑛 whose elements (𝑥, 𝜉) can be thought of as
a (very “naked”) particle with position 𝑥 and momentum 𝜉 or, alternatively, as an
element of a wave with frequency 𝜉 (with 𝑢 standing for some kind of measurement
of a scalar field in the neighborhood of the point 𝑥). This is the first step in lifting the
analysis to phase-space, usually referred to as microlocal analysis, the framework
for many of the more “advanced” parts of the book. Partial differential operators
(with C ∞ coefficients) are defined as acting on distributions by transposition in the
duality between distributions and test-functions, the smooth functions with compact
support in Ω that form the linear space Cc∞ (Ω).
Section 2.2 introduces the most frequently used distribution spaces [meaning
subspaces of D ′ (Ω)], namely the Sobolev spaces 𝐻 𝑚 ( Ω), with 𝑚 ∈ Z, and their
global version 𝐻 𝑠 (R𝑛 ), 𝑠 ∈ R, all based on the space 𝐿 2 of square integrable
functions and all Hilbert spaces for their natural inner product. The more general
Sobolev spaces based on 𝐿 𝑝 , 𝑝 ≠ 2, are not used in this book, nor are the standard
spaces of Harmonic Analysis. For an in-depth study of Sobolev spaces and their
applications to boundary value problems, see [Lions-Magenes, 1968].
The last two sections discuss distribution kernels, namely distributions in the
product Ω1 × Ω2 of two open subsets of Euclidean space (possibly of different
dimensions). These kernels define (bounded) linear maps Cc∞ (Ω1 ) −→ D ′ (Ω2 );
actually every such map is defined in this manner, in sharp contrast with bounded
linear maps between Banach spaces and their duals.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 17

F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_2
18 2 Distributions in Euclidean Space

For us the most interesting such kernels (here called semiregular) are those that
depend smoothly on each one of the variables when regarded as valued in distri-
butions with respect to the other variable. Semiregular kernel distributions define
continuous linear maps Cc∞ (Ω1 ) −→ C ∞ (Ω2 ) as well as E ′ (Ω1 ) −→ D ′ (Ω2 )
[here E ′ is the space of compactly supported distributions, the dual of C ∞ ].
Examples of semiregular kernels in the product Ω × Ω are the fundamental so-
lutions of linear PDEs. Among semiregular kernels, those of utmost importance in
establishing the smoothness of solutions in Ω of a linear partial differential equation
(PDE, with C ∞ coefficients) have the additional property that their restrictions to the
complement of the diagonal in Ω × Ω are smooth functions. If a PDE has a funda-
mental solution or an “approximate” fundamental solution (called a parametrix, a
naming of unknown origin) with the latter property, the PDE is said to be hypoelliptic.
To mention a link to the classical taxonomy of PDEs: all elliptic PDEs (prototype:
the Laplace equation) and all parabolic PDEs (prototype: the heat equation) are
hypoelliptic; no hyperbolic PDE (prototype: the wave equation) is hypoelliptic.
For an in-depth presentation of these and related topics there is a profusion of
texts, foremost [Schwartz, 1966], [Hörmander, 1983, I] (also, [Treves, 1967]).

2.1 Basics on Distributions in Euclidean Space

2.1.1 Definitions. Support and singular support

Let Ω be an open subset of R𝑛 , as before. If 𝑢 is a complex-valued linear functional

on the vector space Cc∞ (Ω), i.e., if 𝑢 is a linear map Cc∞ (Ω) −→ C, we denote by
⟨𝑢, 𝜑⟩ its evaluation at the test-function 𝜑 ∈ Cc∞ (Ω). The linear functional 𝑢 is a
distribution in Ω if ⟨𝑢, 𝜑 𝑗 ⟩ → 0 whenever the sequence 𝜑 𝑗 𝑗=0,1,2,... ⊂ Cc∞ (Ω)
converges to zero in the following sense:
(•) all derivatives 𝜕 𝛼 𝜑 𝑗 converge uniformly to zero and there is a compact set
𝐾 ⊂ Ω such that supp 𝜑 𝑗 ⊂ 𝐾 whatever 𝑗.
The space of distributions in Ω is denoted by D ′ (Ω). The restriction of a distri-
bution 𝑢 ∈ D ′ (Ω) to an open subset Ω′ of Ω is simply the restriction of the linear
functional 𝑢 to the linear subspace Cc∞ (Ω′) of Cc∞ (Ω). By using partitions of unity
in Cc∞ (Ω) it is readily proved that there is a smallest closed subset of Ω, called the
support of 𝑢 and denoted by supp 𝑢, such that 𝑢 vanishes (“identically”) in Ω\𝐹.
The subspace of distributions in Ω that have compact support (contained in Ω) is
denoted by E ′ (Ω); it can be identified with the dual of C ∞ (Ω).
The convergence of a sequence of distributions 𝑢 𝑗 ( 𝑗 ∈ Z+ ) is to be understood in
the “weak sense”: 𝑢 𝑗 → 0 if ⟨𝑢 𝑗 , 𝜑⟩ → 0 for each 𝜑 ∈ Cc∞ (Ω). For 𝑢 𝑗 ∈ E ′ (Ω) to
converge to zero in E ′ (Ω) it is moreover required that there be a compact set 𝐾 ⊂ Ω
such that supp 𝑢 𝑗 ⊂ 𝐾 for all 𝑗.
2.1 Basics on Distributions in Euclidean Space 19

Every continuous linear map of Cc∞ (Ω) into itself defines, by transposition,
a continuous linear map of D ′ (Ω) into itself. Most important among these are
multiplication by smooth functions in Ω and partial derivatives. If 𝑃 (𝑥, D 𝑥 ) is
a linear partial differential operator with smooth coefficients in Ω we define, for
arbitrary 𝑢 ∈ D ′ (Ω), 𝜑 ∈ Cc∞ (Ω),

⟨𝑃 (𝑥, D 𝑥 ) 𝑢, 𝜑⟩ = 𝑢, 𝑃 (𝑥, D 𝑥 ) ⊤ 𝜑 ,

where 𝑃 (𝑥, D 𝑥 ) ⊤ is the transpose of 𝑃 (𝑥, D 𝑥 ) [cf. (1.3.3)].

A distribution 𝑢 is said to be a function
∫ if there is a locally (Lebesgue-)integrable
function 𝑓 in Ω ∫ such that ⟨𝑢, 𝜑⟩ = 𝜑 𝑓 d𝑥 for every 𝜑 ∈ Cc∞ (Ω). It is not difficult
to prove that 𝜑 𝑓 d𝑥 = 0 for all 𝜑 ∈ Cc∞ (Ω) if and only if 𝑓 = 0 a.e. (almost
everywhere), and the value ⟨𝑢, 𝜑⟩ does not change if we modify 𝑓 on a set of
measure zero. Thus it is natural to regard 𝑓 as an equivalence class (a coset) of
a.e. equal locally integrable functions in Ω: 𝑓 ∈ 𝐿 loc 1 (Ω); and then to call 𝑓 the
1
distribution itself. This allows us to regard 𝐿 loc (Ω) as well as all its subspaces as
spaces of distributions, which is to say, linear subspaces of D ′ (Ω) equipped with a
topology finer than that inherited from D ′ (Ω). Thus the Lebesgue spaces 𝐿 loc 𝑝
(Ω)
(1 ≤ 𝑝 ≤ +∞) and the spaces C (Ω) of 𝑘 times differentiable functions in Ω (𝑘 ∈ Z+
𝑘

or 𝑘 = +∞) are all spaces of distributions. In particular, C ∞ (Ω) and Cc∞ (Ω) are
subspaces of D ′ (Ω).
A defining property of distributions (repeatedly used in the sequel) is their local
representation: given a compact subset 𝐾 of Ω and a distribution 𝑢 in Ω there is a
finite subset 𝑆 of Z+𝑛 and a corresponding set of functions 𝑓 𝛼 ∈ 𝐿 loc 1 (Ω) such that
Í
𝑢 = 𝛼∈𝑆 D 𝑓 𝑎 in a neighborhood of 𝐾 in Ω.
𝛼

There are distributions that are not functions, the primordial example being the
Dirac measure at a point 𝑥 ◦ ∈ R𝑛 :
∫
∞
Cc (Ω) ∋ 𝜑 ↦→ ⟨𝛿 𝑥 ◦ , 𝜑⟩ = 𝜑 (𝑥) 𝛿 (𝑥 − 𝑥 ◦ ) d𝑥 = 𝜑 (𝑥 ◦ ) ;

more generally, any Radon∫measure 𝜇 in Ω which is not a density defines the

distribution Cc∞ (Ω) ∋ 𝜑 ↦→ 𝜑d𝜇. Other classical examples are the principal value
distributions, the prototype being the distribution on R defined by
∫
1 d𝑥
⟨pv , 𝜑⟩ = lim 𝜑 (𝑥) ,
𝑥 𝜀↘0 | 𝑥 |> 𝜀 𝑥

which is not a Radon measure. Even when a distribution 𝑢 is not a function we

often use the notation 𝑢 (𝑥) to indicate its domain of definition, whose generic point
is 𝑥. A distribution 𝑢 in Ω is said to be of finite order if there is a finite set 𝑆 of
multi-indices 𝛼 ∈ Z+𝑛 and a corresponding set of functions 𝑓 𝛼 ∈ 𝐿 loc 1 (Ω) such that
Í 𝑝
𝑢 = 𝛼∈𝑆 D 𝑓 𝛼 . (In this definition one can also make use of 𝐿 loc , 𝑝 > 1, or of 𝐶 𝑘 ,
𝛼
1 , with possibly a different choice of 𝑆.) It is the essence of
𝑘 < +∞, in place of 𝐿 loc
the concept of a distribution that it be locally of finite order: if Ω′ is a bounded open
20 2 Distributions in Euclidean Space

set of R𝑛 whose closure is contained in Ω (i.e., Ω′ ⊂⊂ Ω) then the restriction to Ω′

of an arbitrary distribution in Ω is a distribution of finite order in Ω′. On all this we
refer to [Schwartz, 1966] (see also [Hörmander, 1983, I], [Treves, 1967]).
One of focal topics of this book is the regularity (in truth, mainly the analytic
regularity) of solutions of linear PDEs, making the following definitions important.

Definition 2.1.1 The intersection of all the closed subsets 𝐹 of Ω such that the
restriction of 𝑢 ∈ D ′ (Ω) to Ω\𝐹 is a C ∞ (resp., C 𝜔 ) function in Ω\𝐹 is called
the singular support (resp., the analytic singular support); it shall be denoted by
singsupp 𝑢 (singsuppa 𝑢).

Obviously, one can introduce singular supports in the sense of categories other
than C ∞ or C 𝜔 , for example Gevrey classes 𝐺 𝑠 [see (1.2.10)].

2.1.2 Tempered distributions and their Fourier transforms

As is customary, S (R𝑛 ) stands for the (Schwartz) space of functions 𝜑 ∈ C ∞ (R𝑛 )

rapidly decaying at infinity: given arbitrary 𝛼 ∈ Z+𝑛 and 𝑚 ∈ Z+ ,
21 𝑚
sup 1 + |𝑥| 2 𝜕𝑥𝛼 𝜑 (𝑥) < +∞. (2.1.1)
𝑥 ∈R𝑛

A sequence of functions 𝜑 ∈ S (R𝑛 ) converges to zero if the seminorms on the left

in (2.1.1) converge to zero for all choices of 𝑚 and 𝛼; S (R𝑛 ) is a Fréchet space and
thus its topology can be defined by (equivalent) metrics that turn it into a complete
metric space. The space S ′ (R𝑛 ) of tempered distributions in R𝑛 is the subspace
of D ′ (R𝑛 ) consisting of the distributions 𝑢 which can be written as finite sums of
distribution derivatives ∑︁
𝑢= D 𝛼 (𝑃 𝛼 𝑓 𝛼 ) (2.1.2)
| 𝛼 | ≤𝑚

in which the 𝑃 𝛼 are polynomials and the 𝑓 𝛼 belong, say, to 𝐿 1 (R𝑛 ). By transposing
the dense injection Cc∞ (R𝑛∫) ↩→ S (R𝑛 ) the dual of S (R𝑛 ) is identified with S ′ (R𝑛 ).
Below we often denote by 𝑢 (𝑥) 𝜑 (𝑥) d𝑥 (rather than by ⟨𝑢, 𝜑⟩) the duality bracket
between 𝑢 ∈ S ′ (R𝑛 ) and 𝜑 ∈ S (R𝑛 ).
The Fourier transform
∫
𝑢 (𝜉) =
b e−𝑖 𝑥· 𝜉 𝑢 (𝑥) d𝑥 (2.1.3)
R𝑛

defines a Fréchet space isomorphism of S R𝑛𝑥 onto S R𝑛𝜉 whose inverse is given
by ∫
𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 b
𝑢 (𝜉) d𝑥. (2.1.4)
R𝑛
2.1 Basics on Distributions in Euclidean Space 21

By transposition it defines the Fourier

transform of tempered distributions, an
′ ′

isomorphism of S R 𝑥 onto S R𝑛𝜉 also given by (2.1.3) with inverse also given
𝑛

by (2.1.4), with the integral signs now standing for duality brackets. On all this we
refer to textbooks in distribution theory ([Schwartz, 1966], also [Hörmander, 1983,
I], [Treves, 1967]).
The dense injection S (R𝑛 ) ↩→ C ∞ (R𝑛 ) yields, by transposition, an injection of
the space E ′ (R𝑛 ) of compactly supported distributions in R𝑛 into S ′ (R𝑛 ). It follows
from (1.3.1) that an arbitrary 𝑢 ∈ E ′ (R𝑛 ) admits representations of the type
∑︁
𝑢= D 𝑥𝛼 𝑓 𝛼 (2.1.5)
| 𝛼 | ≤𝑚

1 (R𝑛 ). With 𝜀 > 0 define Ω = {𝑥 ∈ R𝑛 ; dist (𝑥, supp 𝑢) < 𝜀} and

with 𝑓 𝛼 ∈ 𝐿 loc 𝜀
introduce a cut-off function 𝜒 ∈ Cc∞ (Ω 𝜀 ), 𝜒 = 1 in an open set containing supp 𝑢.
It follows immediately from the transpose Leibniz rule (1.1.5) that

| 𝛼−𝛽 | 𝛼
∑︁ ∑︁
𝛽 𝛼−𝛽
𝑢 = 𝜒𝑢 = (−1) D𝑥 D𝑥 𝜒 𝑓 𝛼 .
𝛽⪯𝛼
𝛽
| 𝛼 | ≤𝑚

In other words, (2.1.5) is valid with supp 𝑓 𝛼 ⊂⊂ Ω 𝜀 , which we state by writing

𝑓 𝛼 ∈ 𝐿 c1 (Ω 𝜀 ).
In connection with these last remarks we recall the Paley–Wiener–Schwartz The-
orem ([Schwartz, 1966], Ch. VII, Théorème XVI):

Theorem 2.1.2 For a tempered distribution 𝑢 in R𝑛 to have compact support it is

necessary and sufficient that its Fourier transform b
𝑢 be the restriction to real space
of an entire holomorphic function b 𝑢 (𝜁) in C𝑛 such that, for some positive constants
𝐵, 𝑚 and all 𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 ,

𝑢 (𝜉 + 𝑖𝜂)| ≲ (1 + |𝜉 |) 𝑚 exp 𝐵 |𝜂| .

|b

We also recall the link between the holomorphic extension of a function 𝑓 and
the decay of its Fourier transform b
𝑓.

Proposition 2.1.3 If 𝑓 ∈ 𝐿 1 (R𝑛 ) and if sup b𝑓 (𝜉) exp (𝑅 |𝜉 |) < +∞ for some
𝜉 ∈R𝑛
𝑅 > 0 then 𝑓 is the restriction to R𝑛 of a holomorphic function 𝐹 in the slab
{𝑧 ∈ C𝑛 ; |Im 𝑧| < 𝑅}.

Proof It suffices to note that 𝑓 (𝑥) is the restriction to 𝑦 = 0 of the function

∫
𝐹 (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 −𝑦· 𝜉 b
𝑓 (𝜉) d𝜉.
R𝑛

The hypothesis ensures that the integral on the right is absolutely convergent provided
|𝑦| < 𝑅 and, by differentiation under the integral sign, that 𝜕𝑧¯ 𝐹 ≡ 0 in that slab. □
22 2 Distributions in Euclidean Space

Under appropriate hypotheses on 𝑓 the converse of the entailment in Proposition

2.1.3 is also valid. At any rate the obvious defect of such results is that they are
global, meaning that they do not allow a detection of purely local properties. This
will be remedied (in Section 3.4) by substituting the FBI transform for the Fourier
transform.

2.1.3 The C ∞ wave-front set of a distribution

Let Ω ⊂ R𝑛 be an open set and let 𝑥 ◦ ∈ Ω, 𝜉 ◦ ∈ R𝑛 \ {0} be arbitrary. By a cone in

R𝑛 \ {0} we shall always mean a set invariant under all dilations 𝜉 ↦→ 𝜆𝜉, 𝜆 > 0 (i.e.,
a cone with vertex at the origin).

Lemma 2.1.4 Let 𝑢 ∈ D ′ (Ω) have the following property:

(NWF) There exist an open set 𝑈 ⊂⊂ Ω containing 𝑥 ◦ and 𝜑 ∈ Cc∞ (Ω), 𝜑 (𝑥) = 1
for every 𝑥 ∈ 𝑈, and an open cone Γ ⊂ R𝑛 \ {0} containing 𝜉 ◦ such that

∀𝑚 ∈ Z+ , sup (1 + |𝜉 |) 𝑚 š
(𝜑𝑢) (𝜉) < +∞.
𝜉 ∈Γ

Then, if Γ ′ ⊂ R𝑛 \ {0} is an open cone such that Γ ′ ∩ S𝑛−1 ⊂⊂ Γ, we have

∀𝑚 ∈ Z+ , sup (1 + |𝜉 |) 𝑚 (𝜓𝑢)
š (𝜉) < +∞
𝜉 ∈Γ′

for every 𝜓 ∈ Cc∞ (𝑈).

Proof Let 𝜑 and 𝜓 be as in the statement; we have 𝜓𝑢 = 𝜓𝜑𝑢 and therefore

∫
š (𝜉) = (2𝜋) −𝑛
(𝜓𝑢) b (𝜉 − 𝜂) š
𝜓 (𝜑𝑢) (𝜂) d𝜂.

Here we shall use the notation, for 𝑘 ∈ Z+ ,

|𝜓| 𝑘 = sup (1 + |𝜉 |) 𝑘 𝜓
b (𝜉) ,
𝜉 ∈R𝑛

as well as
|𝜑𝑢| 𝑘,Γ = sup (1 + |𝜉 |) 𝑘 š
(𝜑𝑢) (𝜉) .
𝜉 ∈Γ

Using the self-evident inequality (1 + |𝜉 |) 𝑚 ≤ (1 + |𝜂|) 𝑚 (1 + |𝜉 − 𝜂|) 𝑚 we get, for

𝜉 ∈ Γ ′,
2.1 Basics on Distributions in Euclidean Space 23
∫
(1 + |𝜉 |) 𝑚 b (𝜉 − 𝜂) š
𝜓 (𝜑𝑢) (𝜂) d𝜂
∫
d𝜂
≤ |𝜓| 𝑚+𝑛+1 (1 + |𝜂|) 𝑚 š
(𝜑𝑢) (𝜂)
Γ (1 + |𝜉 − 𝜂|) 𝑛+1
∫ 𝑚
(1 + |𝜂|)
+ |𝜓| 𝑚+𝑛+1 (𝜑𝑢) (𝜂) d𝜂.
š
R𝑛 \Γ (1 + |𝜉 − 𝜂|) 𝑁

There is an 𝜀 > 0 such that |𝜉 − 𝜂| ≥ 𝜀 |𝜂| for all 𝜉 ∈ Γ ′ and 𝜂 ∈ R𝑛 \Γ. Since
𝜑 ∈ E ′ (R𝑛 ) we have š(𝜑𝑢) (𝜂) ≲ (1 + |𝜂|) 𝑘 for some 𝑘 > 0 and all 𝜂 ∈ R𝑛 (Theorem
2.1.2). If we select 𝑁 ≥ 𝑚 + 𝑘 + 𝑛 + 1 we get
∫
(1 + |𝜉 |) 𝑚 b (𝜉 − 𝜂) š
𝜓 (𝜑𝑢) (𝜂) d𝜂

≲ |𝜓| 𝑚+𝑛+1 |𝜑𝑢| 𝑚,Γ + 𝜀 −𝑚−𝑁 |𝜓| 𝑚+𝑁 +𝑛+1 . □

Definition 2.1.5 By the wave-front set of 𝑢 ∈ D ′ (Ω) we shall mean the set of points
(𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0}) that do not satisfy Condition (NWF). The wave-front set
of 𝑢 will be denoted by 𝑊 𝐹 (𝑢).

Later, we shall sometimes refer to 𝑊 𝐹 (𝑢) as the C ∞ wave-front set of 𝑢, to

distinguish it from the analytic wave-front set of 𝑢 (Definition 3.5.1).

Proposition 2.1.6 Let 𝑢 ∈ D ′ (Ω). If (𝑥 ◦ , 𝜉) ∉ 𝑊 𝐹 (𝑢) for all 𝜉 ∈ R𝑛 \ {0} then 𝑢 is

a C ∞ function in a neighborhood of 𝑥 ◦ .

Proof There are finitely many open cones Γ 𝑗 ( 𝑗 = 1, ..., 𝑟) such that Γ1 ∪ · · · ∪ Γ𝑟 =
R𝑛 \ {0} and such that, for each 𝑗 = 1, ..., 𝑟, there is a 𝜑 𝑗 ∈ Cc∞ (Ω), 𝜑 𝑗 = 1 in a
neighborhood 𝑈 𝑗 ⊂ Ω of 𝑥 ◦ , satisfying

∀𝑚 ∈ Z+ , sup (1 + |𝜉 |) 𝑚 𝜑 𝑗 𝑢 (𝜉) < +∞.
š
𝜉 ∈Γ 𝑗

𝑟
Ù
Set 𝑈 = 𝑈 𝑗 and let 𝜓 ∈ Cc∞ (𝑈) be equal to 1 in some open subset 𝑉 of 𝑈
𝑗=1
containing 𝑥 ◦ . We derive from Lemma 2.1.4:

∀ 𝑗 = 1, ..., 𝑟, 𝑚 ∈ Z+ , sup (1 + |𝜉 |) 𝑚 (𝜓𝑢)
š (𝜉) < +∞.
𝜉 ∈Γ 𝑗

This means that (𝜓𝑢)
š ∈ S R𝑛 , the Schwartz space, and therefore 𝜓𝑢 ∈ S R𝑛
𝜉 𝑥
implying that 𝑢 is smooth in 𝑉. □
24 2 Distributions in Euclidean Space

2.1.4 Action of differential operators on distributions

The action of a linear PDO on a distribution 𝑢 in Ω is defined by transposition:

⟨𝑃 (𝑥, D) 𝑢, 𝜑⟩ = ⟨𝑢, 𝑃 (𝑥, D) ⊤ 𝜑⟩, 𝜑 ∈ Cc∞ (Ω) . (2.1.6)

When 𝑢 ∈ C ∞ (Ω), (2.1.6) simply reflects integration by parts. Likewise,

⟨𝑃 (𝑥, D) 𝑢, 𝜑⟩ = ⟨𝑢, 𝑃 (𝑥, D) ∗ 𝜑⟩, 𝜑 ∈ Cc∞ (Ω) . (2.1.7)

It follows directly from (2.1.6) that the inclusion (1.3.2), supp 𝑃 (𝑥, D) 𝑓 ⊂
supp 𝑓 , remains valid when 𝑓 ∈ D ′ (Ω). It is also obvious that

singsupp 𝑃 (𝑥, D) 𝑓 ⊂ singsupp 𝑓 , (2.1.8)

and if the coefficients of 𝑃 (𝑥, D) are real-analytic, that

singsuppa 𝑃 (𝑥, D) 𝑓 ⊂ singsuppa 𝑓 . (2.1.9)

In other words, differential operators “decrease” the singular supports, just like they
decrease the supports.
Every linear PDO maps D ′ (Ω) linearly and continuously into itself, and E ′ (Ω)
into itself. In particular, 𝑃 (𝑥, D) acts in the distribution sense (often called “the weak
sense”) on a function 𝑓 ∈ 𝐿 loc 1 (Ω):

∫
⟨𝑃 (𝑥, D) 𝑓 , 𝜑⟩ = 𝑓 𝑃 (𝑥, D) ⊤ 𝜑d𝑥, 𝜑 ∈ Cc∞ (Ω) .

Actually [cf. (2.1.5)], every distribution 𝑢 ∈ D ′ (Ω) can be represented locally as a

finite sum of derivatives of continuous functions.

2.2 Sobolev Spaces

Possibly the most important scales of distribution spaces consist of the Sobolev
spaces. In this text we will solely make use of the Sobolev spaces based on 𝐿 2 ,
which we shall denote by 𝐻 𝑠 (R𝑛 ) with 𝑠 ∈ R: 𝐻 𝑠 (R𝑛 ) is the linear space of
tempered distributions 𝑢 whose Fourier transform 𝑢 is a square-integrable function
b
𝑠
in R𝑛 with respect to the density 1 + |𝜉 | 2 d𝜉. The Hermitian product
∫ 𝑠
−𝑛
(𝑢, 𝑣) 𝑠 = (2𝜋) 𝑣 (𝜉) 1 + |𝜉 | 2 d𝜉
𝑢 (𝜉) b
b (2.2.1)
R𝑛
2.2 Sobolev Spaces 25
√︁
defines a Hilbert space structure on 𝐻 𝑠 (R𝑛 ); we use the notation ∥𝑢∥ 𝑠 = (𝑢, 𝑢) 𝑠 .
′
We have 𝐻 0 (R𝑛 ) = 𝐿 2 (R𝑛 ); if 𝑠 ′ < 𝑠, 𝐻 𝑠 (R𝑛 ) ⊂ 𝐻 𝑠 (R𝑛 ); and 𝑢 ∈ 𝐻 𝑠 (R𝑛 ) =⇒
∥𝑢∥ 𝑠′ ≤ ∥𝑢∥ 𝑠 . All the Hilbert spaces 𝐻 𝑠 (R𝑛 ) are isomorphic: it is immediate to see
that the operators
∫ 𝑡/2
(1 − Δ 𝑥 ) 𝑡/2 𝜑 (𝑥) = (2𝜋) −𝑛 e−𝑖 𝑥· 𝜉 1 + |𝜉 | 2 b (𝜉) d𝜉, 𝑡 ∈ R,
𝜑 (2.2.2)
R𝑛

form a group of (continuous linear) automorphisms of S (R𝑛 ); (2.2.2) extends as an

isometry of 𝐻 𝑠 (R𝑛 ) onto 𝐻 𝑠−𝑡 (R𝑛 ), whatever the real numbers 𝑠, 𝑡.
We mention a useful inequality, valid for all 𝑠, 𝑡 ∈ R such that 𝑎 = 𝑠 − 𝑡 > 0, all
𝜀 > 0 and 𝑢 ∈ 𝐻 𝑠 (R𝑛 ):
1
∥𝑢∥ 2𝑡 ≤ 𝜀 ∥𝑢∥ 2𝑠 + ∥𝑢∥ 2𝑡−𝑎 , (2.2.3)
4𝜀

a direct consequence of the inequality 𝐴𝑡 ≤ 𝜀 𝐴𝑠 + 1

4𝜀 𝐴𝑡−𝑎 , 𝐴 = 1 + |𝜉 | 2 .
Proposition 2.2.1 For any 𝑠 ∈ R the Schwartz space S (R𝑛 ) is continuously embed-
ded and dense in 𝐻 𝑠 (R𝑛 ).
Proof The claim is well-known to be true when 𝑠 = 0; using the isometry
(1 − Δ 𝑥 ) −𝑠/2 : 𝐿 2 (R𝑛 ) −→ 𝐻 𝑠 (R𝑛 ) proves it for all 𝑠. □
Corollary 2.2.2 The space of test-functions Cc∞ (R𝑛 ) is dense in 𝐻 𝑠 (R𝑛 ).
For any 𝑠 ∈ R there is a natural duality between 𝐻 𝑠 (R𝑛 ) and 𝐻 −𝑠 (R𝑛 ). Here
“natural” means that transposing the continuous embedding S (R𝑛 ) ↩→ 𝐻 𝑠 (R𝑛 ) ↩→
S ′ (R𝑛 ) (which have dense range) yields the embedding S (R𝑛 ) ↩→ 𝐻 −𝑠 (R𝑛 ) ↩→
S ′ (R𝑛 ) (the Schwartz space S is reflexive: S is identical to the dual of its dual S ′).
The duality bracket between 𝐻 𝑠 (R𝑛 ) and its dual 𝐻 −𝑠 (R𝑛 ) is the extension of that
between S (R𝑛 ) and S ′ (R𝑛 ):
∫
⟨𝑢, 𝑣¯ ⟩ = (2𝜋) −𝑛 𝑢 (𝜉) b
b 𝑣 (𝜉)d𝜉 (2.2.4)

= (1 − Δ 𝑥 ) 𝑠/2 𝑢, (1 − Δ 𝑥 ) −𝑠/2 𝑣 2 .
𝐿

The spaces 𝐻 𝑠 (R𝑛 ) as we have defined them, by means of the Fourier transform,
are naturally related to the classical Sobolev spaces, defined for any open set Ω ⊂ R𝑛
and for any 𝑚 ∈ Z: if 𝑚 ≥ 0, 𝐻 𝑚 (Ω) is the space of functions whose partial
derivatives of order ≤ 𝑚 belong to 𝐿 2 (Ω); 𝐻 −𝑚 (Ω) is the space of distributions
that are finite sums of partial derivatives of order ≤ 𝑚 of functions belonging to
𝐿 2 (Ω). We shall adopt the commonly used notation, 𝐻◦𝑚 (Ω), for the closure of
Cc∞ (Ω) in 𝐻 𝑚 (Ω); 𝐻 −𝑚 (Ω) is naturally identified with the dual of 𝐻◦𝑚 (Ω). For
any 𝑚 ∈ Z, the multiplication Cc∞ (Ω) × 𝐻 𝑚 (Ω) ∋ (𝜑, 𝑢) ↦→ 𝜑𝑢 ∈ E ′ (Ω) is a
continuous linear map into 𝐻 𝑚 (Ω) [into 𝐻◦𝑚 (Ω) if 𝑚 ≥ 0].
We shall not deal with 𝐻 𝑠 (Ω) for 𝑠 ∉ Z; however the following spaces are easily
defined and useful: If 𝐾 ⊂ R𝑛 is a compact set,
26 2 Distributions in Euclidean Space

𝐻 𝑠 (𝐾) = {𝑢 ∈ 𝐻 𝑠 (R𝑛 ) ; supp 𝑢 ⊂ 𝐾 } ;

𝐻 𝑠 (𝐾) is a closed vector subspace of 𝐻 𝑠 (R𝑛 ) and the Hilbert norm and inner
product of 𝐻 𝑠 (R𝑛 ) make a Hilbert space of 𝐻 𝑠 (𝐾). We define
Ø
𝐻c𝑠 (Ω) = 𝐻 𝑠 (𝐾) .
𝐾 ⊂ ⊂Ω

Multiplication is a continuous bilinear map Cc∞ (Ω) × 𝐻 𝑚 (Ω) ∋ (𝜑, 𝑢) ↦→ 𝜑𝑢 ∈

𝐻c𝑠 (Ω). A sequence {𝑢 𝑘 } 𝑘=1,2,... converges in 𝐻c𝑠 (Ω) if there is a compact subset
𝐾 of Ω such that supp 𝑢 𝑘 ⊂ 𝐾 for all 𝑘, and the 𝑢 𝑘 converge in 𝐻 𝑠 (𝐾). Lastly we
𝑠 (Ω)
define 𝐻loc as the space of distributions 𝑢 ∈ D ′ (Ω) such that 𝜑𝑢 ∈ 𝐻c𝑠 (Ω)
∞
for all 𝜑 ∈ Cc (Ω). A sequence {𝑢 𝑘 } 𝑘=1,2,... converges in 𝐻loc
𝑠 (Ω)
if the sequence
𝑠 ∞
𝜑𝑢 𝑘 converges in 𝐻c (Ω) for each 𝜑 ∈ Cc (Ω); 𝐻loc (Ω) is a Fréchet space [not so
𝑠

𝐻c𝑠 (Ω): it is what is called an inductive limit of Fréchet spaces, actually of Hilbert
spaces].
Multiplication by a function 𝜑 ∈ C ∞ (Ω) defines continuous linear endomor-
phisms of each space 𝐻c𝑠 (Ω) and 𝐻loc 𝑠 (Ω).

We have
Ù Ù
Cc∞ (Ω) = 𝐻c𝑠 (Ω) , C ∞ (Ω) = 𝑠
𝐻loc (Ω) ,
𝑠 ∈R 𝑠 ∈R
Ø
′
E (Ω) = 𝐻c𝑠 (Ω) .
𝑠 ∈R
Ø
𝑠 (Ω)
The union 𝐻loc is the space of distributions of finite order, meaning that
𝑠 ∈R
its elements are (globally) equal to sums of derivatives of locally-𝐿 2 functions (cf.
Subsection 1.1.1).
′ 𝑠′ (Ω)
If 𝑠 ′ < 𝑠 the natural embeddings 𝐻c𝑠 (Ω) ↩→ 𝐻c𝑠 (Ω), 𝐻loc
𝑠 (Ω)
↩→ 𝐻loc are
continuous. The following variant of Rellich’s Lemma is worth recalling:
Proposition 2.2.3 Let 𝐾 ⊂ R𝑛 is a compact set. If 𝑠 ′ < 𝑠 the natural embedding
′
𝐻 𝑠 (𝐾) ↩→ 𝐻 𝑠 (𝐾) is compact [i.e., it transforms the unit ball in 𝐻 𝑠 (𝐾) into a
′
subset of 𝐻 𝑠 (𝐾) whose closure is compact].
The next proposition, of later use, is perhaps less well known; to prove it we need
the following
Lemma 2.2.4 Let 𝐾 be a compact subset of R𝑛 and 𝑠 ∈ R. For all 𝑢 ∈ 𝐻 𝑠 (𝐾),
∫
∥𝑢∥ 2𝑠 ≲ (2𝜋) −𝑛 𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉.
|b (2.2.5)
| 𝜉 | ≥1

Proof Suppose there were a sequence of distributions 𝑢 𝜈 ∈ 𝐻 𝑠 (𝐾) (𝜈 = 1, 2, ...)

such that ∥𝑢 𝜈 ∥ 𝑠 = 1 and
∫ 𝑠 1
(2𝜋) −𝑛 𝑢 𝜈 (𝜉)| 2 1 + |𝜉 | 2 d𝜉 ≤
|b .
| 𝜉 |>1 𝜈+1
2.2 Sobolev Spaces 27

Then there would be a subsequence 𝑢 𝜈𝑘 𝑘=1,2,... converging weakly in 𝐻 𝑠 (𝐾)
and therefore converging (strongly) in E ′ (R𝑛 ), necessarily to 0 since the Fourier
transform of their limit is an entire function of exponential type vanishing identically
for |𝜉 | > 1. As 𝜁 ranges over 𝔅1C = {𝜁 ∈ C𝑛 ; |𝜁 | < 1} the functions e−𝑖 𝑥·𝜁 remain
in a bounded subset of C ∞ (R𝑛 ), i.e., their partial derivatives of any given order
are bounded on each compact subset of R𝑛 independently of 𝜁. It follows that
the entire functions b𝑢 𝜈𝑘 (𝜁) = 𝑢 𝜈𝑘 (𝑥) , e−𝑖 𝑥·𝜁 converge uniformly to zero in 𝔅1C ,
contradicting the hypothesis that
∫ 𝑠 𝜈
(2𝜋) −𝑛 𝑢 𝜈 (𝜉)| 2 1 + |𝜉 | 2 d𝜉 ≥
|b .
| 𝜉 |<1 𝜈+1

We conclude that there is a 𝑐 𝐾 ,𝑠 > 0 such that

∫ 𝑠 ∫
2
𝑐 𝐾 ,𝑠 ≤ |b 2
𝑢 (𝜉)| 1 + |𝜉 | d𝜉 ≲ 𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉
|b
| 𝜉 |>1 | 𝜉 |>1

for all 𝑢 ∈ 𝐻 𝑠 (𝐾) such that ∥𝑢∥ 𝑠 = 1. □

Proposition 2.2.5 Let 𝐾 be an arbitrary compact subset of R𝑛 and 𝑠 ∈ [0, +∞). To

every 𝑠 ′ < 𝑠 there is a constant 𝐶𝐾 ,𝑠,𝑠′ > 0 such that
′
∥𝑢∥ 2𝑠′ ≤ 𝐶𝐾 ,𝑠,𝑠′ (diam 𝐾) 𝑠−𝑠 ∥𝑢∥ 2𝑠

for every 𝑢 ∈ 𝐻 𝑠 (𝐾).

Proof We use the notation 𝔅𝑟 = {𝑥 ∈ R𝑛 ; |𝑥| < 𝑟} (𝑟 > 0). Since 𝑠 ≥ 0, the
elements of 𝐻 𝑠 ⊂ 𝐿 2 are (locally integrable) functions. Now assume 𝐾 ⊂ 𝔅𝑟 ; if
𝑢 ∈ 𝐻 𝑠 (𝐾) we define 𝑢𝑟 (𝑥) = 𝑢 (𝑟𝑥). We have 𝑢𝑟 ∈ 𝐻 𝑠 (𝔅1 ) and
∫
𝑢b𝑟 (𝜉) = e−𝑖 𝑥· 𝜉 𝑢 (𝑟𝑥) d𝑥 = 𝑟 −𝑛 b
𝑢 (𝜉/𝑟)
R𝑛

and therefore
∫ ∫
2 2𝑠′
|b
𝑢 (𝜉/𝑟)| |𝜉 | d𝜉 ≤ 𝑢 (𝜉/𝑟)| 2 |𝜉 | 2𝑠 d𝜉,
|b
| 𝜉 | ≥1 | 𝜉 | ≥1

whence ∫ ∫
′ ′
𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉 ≤ 𝑟 2(𝑠−𝑠 )
|b 𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉.
|b
| 𝜉 | ≥𝑟 | 𝜉 | ≥𝑟

It suffices to deal with 𝑟 < 1; then

∫ ∫
2 2𝑠′ ′
|b
𝑢 (𝜉)| |𝜉 | d𝜉 ≤ 𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉
|b
| 𝜉 | ≥1 | 𝜉 | ≥𝑟

and, since 𝑠 ≥ 0,
28 2 Distributions in Euclidean Space
∫ ∫ 𝑠
𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉 ≤
|b 𝑢 (𝜉)| 2 1 + |𝜉 | 2 d𝜉.
|b
| 𝜉 | ≥𝑟 R𝑛

We reach the conclusion that

∫
′ ′
(2𝜋) −𝑛
𝑢 (𝜉)| 2 |𝜉 | 2𝑠 d𝜉 ≤ 𝑟 2(𝑠−𝑠 ) ∥𝑢∥ 2𝑠 .
|b
| 𝜉 | ≥1

The sought result follows by applying Lemma 2.2.4 to the left-hand side (with 𝔅1 in
the place of 𝐾 and 𝑠 ′ instead of 𝑠). □
The action of differential operators on Sobolev spaces is one of their most impor-
tant features:

Proposition 2.2.6 If 𝛼 ∈ Z+𝑛 , 𝑢 ↦→ D 𝛼 𝑢 is a continuous linear map 𝐻 𝑠 (R𝑛 ) −→

𝐻 𝑠− | 𝛼 | (R𝑛 ). If 𝑃 (𝑥, D) is a linear partial differential operator of order 𝑚 with C ∞
coefficients in the open set Ω ⊂ R𝑛 , 𝑢 ↦→ 𝑃 (𝑥, D) 𝑢 is a continuous linear map
𝑠 (Ω) 𝑠−𝑚 (Ω).
𝐻loc −→ 𝐻loc

A useful property is the following consequence of the Sobolev inequalities (see

e.g. [Friedman, 1969], pp. 22–32, [Treves, 1975], pp. 217–221).

Proposition 2.2.7 Let the open subset Ω of R𝑛 and the integer 𝑚 ≥ 0 be arbitrary.
If 𝑠 > 𝑚 + 21 𝑛 then 𝐻loc
𝑠 (Ω)
⊂ C 𝑚 (Ω) and the embedding 𝐻loc𝑠 (Ω)
↩→ C 𝑚 (Ω) is
∞
a continuous map. If moreover Ω is bounded and has a C boundary and if 𝑠 ∈ Z+
then 𝐻 𝑠 (Ω) ⊂ C 𝑚 (Ω) ∩ 𝐿 ∞ (Ω) and the embedding map is continuous.
𝑠 (Ω)
In this statement both 𝐻loc and C 𝑚 (Ω) are viewed as subspaces of D ′ (Ω);
the topology of C (Ω) is that of uniform convergence on each compact subset of
𝑚

Ω of the functions and of all their partial derivatives of order ≤ 𝑚.

2.3 Distribution Kernels

We must now introduce distributions 𝐹 (𝑥, 𝑦) on products Ω1 × Ω2 with Ω1 ⊂

R𝑛1 , Ω2 ⊂ R𝑛2 open sets. Distributions belonging to D ′ (Ω1 × Ω2 ) are often referred
to as kernels or distribution kernels. We can regard the product of two test-functions
𝜑 ∈ Cc∞ (Ω1 ) and 𝜓 ∈ Cc∞ (Ω2 ) as an element of Cc∞ (Ω1 × Ω2 ), denoted by 𝜑 ⊗ 𝜓,
and evaluate 𝐹 ∈ D ′ (Ω1 × Ω2 ) on it. Fixing 𝜓 defines a distribution in Ω1 :

Cc∞ (Ω1 ) ∋ 𝜑 ↦→ ⟨𝐹, 𝜑 ⊗ 𝜓⟩ ∈ C.

To emphasize
∫ this partial action it is convenient to adopt the “Volterra notation”: to
write
∫ 𝐹 (𝑥, 𝑦) 𝜓 (𝑦) d𝑦 rather than ⟨𝐹 (𝑥, 𝑦) , 𝜓 (𝑦)⟩. (Keep in mind, however, that
does not stand for a true integral!) In passing we point out that the Fubini formula
is always true in distribution theory:
2.3 Distribution Kernels 29
∫ ∫ ∫ ∫
𝐹 (𝑥, 𝑦) 𝜓 (𝑦) d𝑦 𝜑 (𝑥) d𝑥 = 𝐹 (𝑥, 𝑦) 𝜑 (𝑥) d𝑥 𝜓 (𝑦) d𝑦.

The map
∫
Cc∞ (Ω2 ) ∋ 𝜓 ↦→ 𝔗 𝐹 𝜓 (𝑥) = 𝐹 (𝑥, 𝑦) 𝜓 (𝑦) d𝑦 ∈ D ′ (Ω1 ) (2.3.1)

is linear and continuous. The Schwartz Kernel Theorem states that, actually, every
continuous linear map Cc∞ (Ω2 ) −→ D ′ (Ω1 ) is of the kind (2.3.1), and that the
correspondence between continuous linear maps and distribution kernels is one-to-
one. This is a very special property of D ′, obviously false for any infinite-dimensional
Banach space (but true for E ′, C ∞ , Cc∞ , if properly reformulated).
The composition 𝐴1,2 ◦ 𝐴2,3 of two linear operators 𝐴1,2 : Cc∞ (Ω2 ) −→ D ′ (Ω1 ),
𝐴2,3 : Cc∞ (Ω3 ) −→ D ′ (Ω2 ), puts requirements of regularity and support on the
factors. For instance, we might require that 𝐴2,3 maps Cc∞ (Ω3 ) into Cc∞ (Ω2 ), or
else that 𝐴1,2 extend as a continuous linear operator D ′ (Ω2 ) −→ D ′ (Ω1 ), which is
equivalent to requiring that the transpose 𝐴1,2 ⊤ maps C ∞ (Ω ) into C ∞ (Ω ). These
c 1 c 2
concerns are addressed in Definitions 2.3.1 and 2.3.6 below.

Definition 2.3.1 We shall say that the operator 𝔗 𝐹 [see (2.3.1)] is semiregular if
for
∫ every pair of test-functions
∫ 𝜑 ∈ Cc∞ (Ω1 ) and 𝜓 ∈ Cc∞ (Ω2 ) both distributions
𝐹 (𝑥, 𝑦) 𝜑 (𝑥) d𝑥 and 𝐹 (𝑥, 𝑦) 𝜓 (𝑦) d𝑦 are C ∞ functions, in Ω2 and Ω1 respec-
tively. In this case we shall also say that the distribution kernel 𝐹 (𝑥, 𝑦) is semiregular.
We shall say that the distribution kernel 𝐹 (𝑥, 𝑦) ∈ D ′ (Ω × Ω) is regular if
𝐹 is semiregular and if 𝐹 is a C ∞ function in the complement of the diagonal,
(Ω × Ω) \ diag (Ω × Ω).

When 𝐹 (𝑥, 𝑦) is semiregular we have 𝔗 𝐹 Cc∞ (Ω2 ) ⊂ C ∞ (Ω1 ). Moreover (and

very important in applications!), 𝔗 𝐹 can be extended as a continuous linear map
E ′ (Ω2 ) −→ D ′ (Ω1 ). Indeed, the transpose of 𝔗 𝐹 , the map

Cc∞ (Ω1 ) ∋ 𝜑 ↦→ 𝔗 𝐹⊤ 𝜑 (𝑦) = ⟨𝐹 (𝑥, 𝑦) , 𝜑 (𝑥)⟩ ∈ D ′ (Ω2 ) , (2.3.2)

is also semiregular and, therefore,

⊤ it is a continuous linear map Cc∞ (Ω1 ) −→
C (Ω2 ). Its transpose 𝔗 𝐹 is a continuous linear map E ′ (Ω2 ) −→ D ′ (Ω1 )
∞ ⊤

and coincides with 𝔗 𝐹 on Cc∞ (Ω2 ).

The next result, stated in the case Ω1 = Ω2 = Ω, is important in applications.

Theorem 2.3.2 If the distribution kernel 𝐹 (𝑥, 𝑦) ∈ D ′ (Ω × Ω) is regular then

∀𝑢 ∈ E ′ (Ω) , singsupp 𝔗 𝐹 𝑢 ⊂ singsupp 𝑢.

The property of decreasing the singular support (Definition 2.1.1) stated in The-
orem 2.3.2 is often referred to by saying that the operator 𝔗 𝐹 is pseudolocal, in
contrast to being local, i.e., decreasing the support, which would require 𝔗 𝐹 to be a
linear differential operator (see [Peetre, 1960]).
30 2 Distributions in Euclidean Space

Proof Let 𝑈 ⊂ Ω be an open set in which 𝑢 is C ∞ and let 𝜒 ∈ Cc∞ (𝑈) be equal to
1 in an open set 𝑉 ⊂⊂ 𝑈. We have:

𝔗 𝐹 𝑢 (𝑥) = ⟨𝐹 (𝑥, 𝑦) , 𝜒 (𝑦) 𝑢 (𝑦)⟩ (2.3.3)

+ ⟨𝐹 (𝑥, 𝑦) , (1 − 𝜒 (𝑦)) 𝑢 (𝑦)⟩ .

Since 𝜒𝑢 ∈ Cc∞ (𝑈) and 𝔗 𝐹 Cc∞ (Ω) ⊂ C ∞ (Ω) we see that the first bracket on the
right in (2.3.3) is C ∞ in Ω. Let 𝜒1 ∈ Cc∞ (𝑉) be equal to 1 in an open set 𝑊 ⊂ 𝑉.
We have

supp ( 𝜒1 (𝑥) 𝐹 (𝑥, 𝑦) (1 − 𝜒 (𝑦))) ⊂ (Ω × Ω) \ diag (Ω × Ω)

and therefore, by our hypothesis, 𝜒1 (𝑥) 𝐹 (𝑥, 𝑦) (1 − 𝜒 (𝑦)) ∈ C ∞ (Ω × Ω). This

implies that

⟨𝐹 (𝑥, 𝑦) , (1 − 𝜒 (𝑦)) 𝑢 (𝑦)⟩ = ⟨(1 − 𝜒 (𝑦)) 𝐹 (𝑥, 𝑦) , 𝑢 (𝑦)⟩

is C ∞ in 𝑊. We conclude that 𝔗 𝐹 𝑢 is also C ∞ in 𝑊. Of course, we can take 𝑊 to be

a neighborhood of an arbitrary point of Ω\ singsupp 𝑢, whence the result. □
The “most regular” operators are the regularizing ones.

Definition 2.3.3 A continuous linear operator 𝑅 : Cc∞ (Ω2 ) −→ D ′ (Ω1 ) is said to

be smoothing (or regularizing) if 𝑅 maps Cc∞ (Ω2 ) into C ∞ (Ω1 ) and extends as a
continuous linear map E ′ (Ω2 ) −→ C ∞ (Ω1 ).

Proposition 2.3.4 The transpose of a smoothing operator is smoothing.

Evident. If 𝑅 is smoothing then, whatever the multi-indices 𝛼, 𝛽 ∈ Z+ , the

composite D 𝛼 𝑅D𝛽 is also smoothing. The next statement readily ensues:

Proposition 2.3.5 For the continuous linear operator 𝑅 : Cc∞ (Ω2 ) −→ D ′ (Ω1 ) to
be smoothing it is necessary and sufficient that the corresponding distribution kernel
𝑅 (𝑥, 𝑦) belong to C ∞ (Ω1 × Ω2 ).

Next we deal with the concerns about the supports of the distribution kernels that
are factors in a composition.

Definition 2.3.6 We say that the kernel 𝐹 (𝑥, 𝑦) ∈ D ′ (Ω1 × Ω2 ) and the correspond-
ing operator (2.3.1) are properly supported if given any pair of compact subsets
𝐾𝑖 ⊂ Ω𝑖 (𝑖 = 1, 2) the sets

{(𝑥, 𝑦) ∈ supp 𝐹; 𝑥 ∈ 𝐾1 } ,
{(𝑥, 𝑦) ∈ supp 𝐹; 𝑦 ∈ 𝐾2 }

are compact.
2.4 Fundamental Solutions, Parametrix, Hypoelliptic PDOs 31

If 𝐹 (𝑥, 𝑦) is regular and properly supported then (2.3.1) extends as a continuous

linear map D ′ (Ω2 ) −→ D ′ (Ω1 ) mapping Cc∞ (Ω2 ) into Cc∞ (Ω1 ), C ∞ (Ω2 ) into
C ∞ (Ω1 ), and E ′ (Ω2 ) into E ′ (Ω1 ).
Suppose Ω1 = Ω2 = Ω. Select a properly supported function 𝜒 ∈ C ∞ (Ω × Ω)
identically equal to 1 in some neighborhood N of the diagonal of Ω × Ω,

diag Ω × Ω = {(𝑥, 𝑦) ∈ Ω × Ω; 𝑥 = 𝑦} .

Whatever 𝐹 ∈ D ′ (Ω × Ω), the kernel 𝜒𝐹 is properly supported and equal to 𝐹 in

N.

Example 2.3.7 If 𝑢 ∈ D ′ (R𝑛 ) we can define 𝑢 (𝑥 − 𝑦) ∈ D ′ (R𝑛 × R𝑛 ) by the

formula

⟨𝑢 (𝑥 − 𝑦) , 𝜑 (𝑥, 𝑦)⟩ = ⟨𝑢 (𝑦) , 𝜑 (𝑥, 𝑥 − 𝑦)⟩, 𝜑 ∈ Cc∞ (R𝑛 × R𝑛 ) .

The kernel 𝑢 (𝑥 − 𝑦) in R𝑛 × R𝑛 is regular; it is properly supported if and only if

supp 𝑢 is compact.

Example 2.3.8 The Dirac distribution 𝛿 (𝑥 − 𝑦) is a regular and properly supported

kernel in Ω × Ω:
∫
⟨𝛿 (𝑥 − 𝑦) , 𝜑 (𝑥, 𝑦)⟩ = 𝜑 (𝑥, 𝑥) d𝑥, 𝜑 ∈ C ∞ (Ω × Ω) .
Ω

We will often make use of the following natural concept.

Definition 2.3.9 By the restriction of an operator 𝐴 : Cc∞ (Ω2 ) −→ D ′ (Ω1 ) to

open sets 𝑈 𝑗 ⊂ Ω 𝑗 ( 𝑗 = 1, 2) we shall mean the composite map
injection 𝐴 restriction
Cc∞ (𝑈2 ) ↩→ Cc∞ (Ω2 ) −→ D ′ (Ω1 ) −→ D ′ (𝑈1 ) . (2.3.4)

The distribution kernel [belonging to D ′ (𝑈1 × 𝑈2 )] associated to the restriction

of 𝐴 to the open sets 𝑈 𝑗 ⊂ Ω 𝑗 ( 𝑗 = 1, 2) is the restriction to 𝑈1 ×𝑈2 of the distribution
kernel 𝐴 (𝑥, 𝑦) ∈ D ′ (Ω1 × Ω2 ) associated to 𝐴.

2.4 Fundamental Solutions, Parametrix, Hypoelliptic PDOs

Let 𝑃 (𝑥, D) be a differential operator with C ∞ coefficients in the open set Ω ⊂ R𝑛

[see (1.3.1)].

Definition 2.4.1 We say that 𝐺 (𝑥, 𝑦) ∈ D ′ (Ω × Ω) is a fundamental solution of

𝑃 (𝑥, D) if 𝐺 (𝑥, 𝑦) is semiregular and if 𝑃 (𝑥, D) 𝐺 (𝑥, 𝑦) = 𝛿 (𝑥 − 𝑦).
32 2 Distributions in Euclidean Space

The fundamental solution 𝐺 (𝑥, 𝑦) provides a right inverse for 𝑃 (𝑥, D): if 𝑓 ∈
E ′ (Ω) then ∫
𝑃 (𝑥, D) 𝐺 (𝑥, 𝑦) 𝑓 (𝑦) d𝑦 = 𝑓 (𝑥) . (2.4.1)
⊤
If one seeks a left inverse one must require that 𝑃 𝑦, D 𝑦 𝐺 (𝑥, 𝑦) = 𝛿 (𝑥 − 𝑦),
since then
∫ ∫
𝐺 (𝑥, 𝑦) 𝑓 (𝑥) 𝑃 𝑦, D 𝑦 𝜑 (𝑦) d𝑥d𝑦 = ⟨ 𝑓 , 𝜑⟩, 𝑓 ∈ E ′ (Ω) , 𝜑 ∈ Cc∞ (Ω) .

(2.4.2)
When the coefficients of 𝑃 are constant we write 𝑃 = 𝑃 (D) with 𝑃 (𝜉) ∈
C [𝜉1 , ..., 𝜉 𝑛 ]. Because of translation invariance we seek a fundamental solution
of 𝑃 (D) of the form 𝐸 (𝑥 − 𝑦), with 𝐸 a distribution solution of the equation
𝑃(D)𝐸 = 𝛿, the Dirac distribution (at the origin). A celebrated theorem of L.
Ehrenpreis and B. Malgrange states that every linear PDO with constant coefficients
in R𝑛 (not all of them equal to zero) has a fundamental solution (see [Ehrenpreis,
1954], [Malgrange, 1955]). On this topic, see Theorem 15.2.38.
In general, fundamental solutions of a linear PDO are difficult to construct. But in
the study of the regularity of the solutions of a linear PDE, approximate fundamental
solutions, sometimes easier to get, can be made to play the role of exact ones.

Definition 2.4.2 A semiregular kernel 𝐺 (𝑥, 𝑦) ∈ D ′ (Ω × Ω) is called a parametrix

of the differential operator 𝑃 (𝑥, D 𝑥 ) if 𝑃 (𝑥, D 𝑥 ) 𝐺 (𝑥, 𝑦) − 𝛿 (𝑥 − 𝑦) ∈ C ∞ (Ω × Ω)
and if 𝐺 (𝑥, 𝑦) is C ∞ in the complement of the diagonal in Ω × Ω.

The name “parametrix” is also used for the continuous linear operator defined by
the kernel 𝐺 (𝑥, 𝑦).

Proposition 2.4.3 If 𝑃 (𝑥, D 𝑥 ) has a parametrix 𝐺 (𝑥, 𝑦) it also has a properly

supported (Definition 2.3.6) parametrix.

Proof Let 𝜒 ∈ C ∞ (Ω × Ω) be properly supported and such that 𝜒 (𝑥, 𝑦) = 1 for all
(𝑥, 𝑦) in a neighborhood of the diagonal in Ω × Ω. Then

𝑃 (𝑥, D) ( 𝜒𝐺) (𝑥, 𝑦) = 𝑃 (𝑥, D) 𝐺 (𝑥, 𝑦)

in a neighborhood of the diagonal in Ω × Ω and it is C ∞ in the complement of the

diagonal in Ω × Ω. □
Suppose the parametrix 𝐺 of 𝑃 is properly supported. Then, whatever the distri-
bution 𝑓 in Ω,
∫ ∫
𝑃 (𝑥, D 𝑥 ) 𝐺 (𝑥, 𝑦) 𝑓 (𝑦)d𝑦 = 𝑓 (𝑥) + 𝑅 (𝑥, 𝑦) 𝑓 (𝑦) d𝑦, (2.4.3)

where 𝑅 ∈ C ∞ (Ω × Ω) [in other words, the operator defined by 𝑅 is smoothing

(Definition 2.3.3)]; (2.4.3) has the following consequence.
2.4 Fundamental Solutions, Parametrix, Hypoelliptic PDOs 33

Theorem 2.4.4 If the transpose 𝑃⊤ of 𝑃 = 𝑃 (𝑥, D 𝑥 ) has a parametrix then

∀𝑢 ∈ D ′ (Ω) , singsupp 𝑃𝑢 = singsupp 𝑢. (2.4.4)

Proof Let 𝑢 ∈ D ′ (Ω) be arbitrary; if we recall (2.1.8) it suffices to prove that

singsupp 𝑢 ⊂ singsupp 𝑃𝑢.

By Proposition 2.4.3 𝑃⊤ has a properly supported parametrix 𝐺 ′ (𝑥, 𝑦). Given an

arbitrary 𝜑 ∈ Cc∞ (Ω) we have
∫ ∫
𝐺 ′ (𝑥, 𝑦) 𝜑(𝑦)𝑃 (𝑥, D 𝑥 )𝑢 (𝑥) d𝑥d𝑦
∫ ∫
= 𝜑 (𝑥) 𝑢 (𝑥) d𝑥 + 𝑅 ′ (𝑥, 𝑦) 𝜑 (𝑦) 𝑢 (𝑥) d𝑥d𝑦,

where 𝑅 ′ ∈ C ∞ (Ω × Ω). In other words,

∫ ∫
′
𝑅 ′ (𝑦, 𝑥) 𝑢 (𝑦) d𝑦.

𝐺 (𝑦, 𝑥) 𝑃 𝑦, D 𝑦 𝑢 (𝑦) d𝑦 = 𝑢 (𝑥) + (2.4.5)

The second term on the right-hand side of (2.4.5) is smooth in Ω and therefore the
singular support of the right-hand side of (2.4.5) is equal to singsupp 𝑢. By Definition
2.4.5 𝐺 ′ (𝑦, 𝑥) is semiregular and smooth in (Ω × Ω) \ diag (Ω × Ω); by Theorem
2.3.2 the singular support of the left-hand side of (2.4.5) is contained in that of 𝑃𝑢,
which proves the sought inclusion. □
In connection with (2.4.4) we introduce the important

Definition 2.4.5 A linear partial differential operator 𝑃 in Ω is said to be hypoelliptic

at a point 𝑥 ◦ ∈ Ω if there is a neighborhood 𝑈 ⊂ Ω of 𝑥 ◦ such that singsupp 𝑃𝑢 =
singsupp 𝑢 whatever 𝑢 ∈ D ′ (𝑈); 𝑃 is said to be hypoelliptic in Ω if 𝑃 is hypoelliptic
at every point of Ω.

Functional Analysis easily provides a necessary condition for hypoellipticity:

Lemma 2.4.6 If the differential operator 𝑃 is hypoelliptic in Ω then to every integer

𝑁 ≥ 1 and every compact subset 𝐾 of Ω there is a compact subset 𝐾1 of Ω, an
integer 𝑁1 ≥ 1 and a constant 𝐶 > 0 such that

∑︁ ∑︁
∀𝜑 ∈ C ∞ (Ω) , max |D 𝛼 𝜑| ≤ 𝐶max ­ |𝜑| + |D 𝛼 𝑃𝜑| ® . (2.4.6)
© ª
𝐾 𝐾1
|𝛼|≤𝑁 « | 𝛼 | ≤ 𝑁1 ¬
Proof When 𝑃 is hypoelliptic the coarsest locally convex topology on C ∞ (Ω) such
that the linear mapping

C ∞ (Ω) ∋ 𝜑 ↦→ (𝜑, 𝑃𝜑) ∈ C (Ω) × C ∞ (Ω)

34 2 Distributions in Euclidean Space

is continuous [i.e., the topology defined by the kind of seminorms at the right-hand
side of (2.4.6)] turns C ∞ (Ω) into a complete metrizable space, i.e., a Fréchet space.
By the Open Mapping Theorem it must be equal to the standard topology [defined
by the kind of seminorms on the left-hand side of (2.4.6)]. □

Corollary 2.4.7 If the differential operator 𝑃 in Ω is hypoelliptic then every point

𝑥 ◦ ∈ Ω has a neighborhood 𝑈 in Ω such that the following is true:
(LCT) To every 𝑓 ∈ D ′ (𝑈) there is a 𝑢 ∈ D ′ (𝑈) satisfying the equation 𝑃⊤ 𝑢 = 𝑓 .

Proof In (2.4.6) we take 𝑈 = {𝑥 ∈ R𝑛 ; |𝑥 − 𝑥 ◦ | ≤ 𝑟 }, 𝐾 = 𝑈, with 𝑟 > 0 sufficiently

small that 𝐾 ⊂ Ω; obviously, 𝐾 ⊂ 𝐾1 . Radial integration starting from 𝑥 shows that
supp 𝜑 ⊂ 𝑈 entails ∑︁
max |𝜑| ≤ 𝑟 max |D 𝛼 𝜑| .
| 𝛼 |=1

Selecting 𝑁 ≥ 1 and 𝑟 < 1/2𝐶 in (2.4.6) yields

∑︁ ∑︁
max |D 𝛼 𝜑| ≤ 2𝐶 max |D 𝛼 𝑃𝜑| .
| 𝛼 | ≤𝑁 | 𝛼 | ≤ 𝑁1

This shows that, given an arbitrary 𝑓 ∈ D ′ (𝑈), the linear functional 𝑃𝜑 ↦→ ⟨ 𝑓 , 𝜑⟩ is

well defined in the linear subspace 𝑃Cc∞ (𝑈) of Cc∞ (𝑈) and continuous with respect
to the topology inherited from Cc∞ (𝑈). The Hahn–Banach Theorem implies that it
can be extended as a continuous linear functional on Cc∞ (𝑈), i.e. a distribution 𝑢 in
𝑈. We have ⟨𝑢, 𝑃𝜑⟩ = ⟨ 𝑓 , 𝜑⟩ for every 𝜑 ∈ Cc∞ (Ω), i.e. 𝑃⊤ 𝑢 = 𝑓 . □
The following stronger version of Corollary 2.4.7 is proved in [Treves, 1967]
(Theorem 52.2):

Theorem 2.4.8 If 𝑃 is hypoelliptic in Ω then every point of Ω has a neighborhood

in which 𝑃⊤ has a parametrix.
Chapter 3
Analytic Tools in Distribution Theory

This chapter is the true introduction, at a level as elementary as possible, to the

core of the book. In the first section the reader will find the needed definitions
and basic results that can be translated to the C 𝜔 category from the last section of
the preceding chapter. One of the handicaps in the analytic theory is the absence of
compactly supported cut-off functions (and therefore of partitions of unity submitted
to coverings by relatively compact open subsets). One way of circumventing this
hurdle is provided by the Ehrenpreis cutoffs, compactly supported C ∞ functions
that satisfy the standard analytic inequalities, but only up to a certain (finite) order
of differentiation. Their precise definition and how to use them is explained in the
second section. A totally different approach is to extend holomorphically, so to speak,
distributions in an open subset Ω of R𝑛 to an open wedge in C𝑛 , meaning a set of
points 𝑥+𝑖𝑦, 𝑥 ∈ Ω, 𝑦 ∈ Γ, |𝑦| < 𝛿, with Γ an open cone in R𝑛 with vertex at the origin
and 𝛿 > 0, or, more properly said, to regard a distribution as the boundary value
of a holomorphic function in the open wedge. Not all distributions can be viewed
in this manner but every distribution is, locally, a finite sum of distributions that
can. This is the subject matter of Section 3.3. Section 3.4 introduces, at a beginner’s
level, one of the main tools in the microlocal analysis of distributions (and, later
on, in Part II of the book, of hyperfunctions), the integral transform that replaces
(and resembles) the Fourier transform, the so-called FBI transform, named after
three French physicists, Joseph Fourier of course, and J. Brós and D. Iagolnitzer;
the latter were the first, in the early 1970s, to show its usefulness in analyzing
problems of causality in theoretical quantum electrodynamics. The Brós–Iagolnitzer
work concerned the so-called Wedge of the Edge Theorem, itself a result about the
holomorphic extendibility of distributions to wedges; this is a theme further explored
in Section 3.5 and recurs in different guises as we advance. Section 3.5 presents the
three equivalent definitions of the analytic wave-front set of a distribution, that is
to say, the C 𝜔 analogue of the C ∞ wave-front set (Definition 2.1.5). On these last
topics we must mention the similar treatment in Section 9.6 of [Hörmander, 1983,
I].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 35

F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_3
36 3 Analytic Tools in Distribution Theory

3.1 Analytic Parametrices, Analytic Hypoellipticity

Let Ω be an open subset of R𝑛 . The following somewhat improper terminology is

commonly used:
Definition 3.1.1 A semiregular kernel 𝐺 (𝑥, 𝑦) ∈ D ′ (Ω × Ω) (Definition 2.3.1) is
called an analytic parametrix of the differential operator 𝑃 (𝑥, D 𝑥 ) if

𝑃 (𝑥, D 𝑥 ) 𝐺 (𝑥, 𝑦) − 𝛿 (𝑥 − 𝑦) ∈ C 𝜔 (Ω × Ω)

and if 𝐺 (𝑥, 𝑦) is real-analytic in the complement of the diagonal in Ω × Ω.

The notion of the analytic singular support of 𝑢 ∈ D ′ (Ω) has been introduced in
Definition 2.1.1. There is an analytic analogue of Definition 2.4.5:
Definition 3.1.2 A real-analytic linear partial differential operator 𝑃 in Ω is said to
be analytic hypoelliptic at a point 𝑥 ◦ ∈ Ω if there is a neighborhood 𝑈 ⊂ Ω of 𝑥 ◦
such that singsuppa 𝑃 (𝑥, D) 𝑢 = singsuppa 𝑢 whatever 𝑢 ∈ D ′ (𝑈); 𝑃 is said to be
analytic hypoelliptic in Ω if 𝑃 is analytic hypoelliptic at every point of Ω.
Remark 3.1.3 There exist hypoelliptic linear PDOs that are not analytic hypoelliptic,
e.g. the heat equation [cf. (4.1.2) and Theorem 4.1.1].
We can prove the analytic analogue of Theorem 2.4.4, under an additional hy-
pothesis.
Theorem 3.1.4 Let 𝑃 be a real-analytic linear PDO in Ω and let 𝑃𝑚 (𝑥, 𝜉) be its
principal symbol (𝑚 ∈ Z+ ; see Definition 1.3.1). Suppose that for every 𝑥 ∈ Ω there
is an 𝜉 ∈ R𝑛 such that 𝑃𝑚 (𝑥, 𝜉) ≠ 0. If every point of Ω has a neighborhood in
which 𝑃⊤ has an analytic parametrix then 𝑃 is hypoelliptic and analytic hypoelliptic
in Ω.
Proof The claim that the hypotheses imply the hypoellipticity of 𝑃 in Ω is an
immediate consequence of Theorem 2.4.4. We must prove that they also imply the
analytic hypoellipticity of 𝑃 in Ω.
Let 𝑥 ◦ ∈ Ω be arbitrary. According to our hypotheses there is a hyperplane Π
defined by an equation ⟨𝜉 ◦ , 𝑥 − 𝑥 ◦ ⟩ = 0 such that 𝑃𝑚 (𝑥, 𝜉 ◦ ) ≠ 0 for all 𝑥 in a
suitably small neighborhood 𝑈 of 𝑥 ◦ : Π ∩ 𝑈 is noncharacteristic for 𝑃 at every one
of its points (Definition 1.3.3). Let 𝑢 ∈ D ′ (𝑈) be such that 𝑃𝑢 ∈ C 𝜔 (𝑈) whence
𝑢 ∈ C ∞ (𝑈). The Cauchy–Kovalevskaya Theorem (Theorem 5.2.4, below) implies
the existence of a solution 𝑣 ∈ C 𝜔 (𝑈) of the equation 𝑃𝑣 = 𝑃𝑢 in 𝑈 (possibly
contracted about 𝑥 ◦ ). It will suffice to prove that 𝑢 − 𝑣 ∈ C 𝜔 (𝑈) or, which amounts
to the same, hypothesize that 𝑃𝑢 ≡ 0 in 𝑈 and conclude that 𝑢 ∈ C 𝜔 (𝑈). Let
𝑉 ⊂⊂ 𝑈 be an open set and 𝜒 ∈ Cc∞ (𝑈) be identically equal to 1 in 𝑉; we have
𝑓 = 𝑃 ( 𝜒𝑢) ∈ Cc∞ (𝑈), 𝑓 ≡ 0 in 𝑉. Let 𝐺 ′ (𝑥, 𝑦) be an analytic parametrix 𝑃⊤ in 𝑈.
We now have [cf. (2.4.5)]
∫ ∫
𝑢 (𝑥) = 𝐺 ′ (𝑦, 𝑥) 𝑓 (𝑦) d𝑦 − 𝑅 ′ (𝑦, 𝑥) ( 𝜒𝑢) (𝑦) d𝑦
3.1 Analytic Parametrices, Analytic Hypoellipticity 37

in 𝑉. We know that the function (𝑥, 𝑦) ↦→ 𝐺 ′ (𝑦, 𝑥) is real-analytic in 𝑉 × (𝑈\𝑉): if

𝐾1 ⊂ 𝑉 is a compact set and 𝐾2 = supp 𝑓 , there is a constant 𝐶 > 0 such that

∀𝛼 ∈ Z+𝑛 , max D 𝑥𝛼 𝐺 ′ (𝑦, 𝑥) ≤ 𝐶 | 𝛼 |+1 𝛼!.

( 𝑥,𝑦) ∈𝐾1 ×𝐾2

Differentiation with respect to 𝑥 under the integral sign implies directly that
∫
𝐺 ′ (𝑦, 𝑥) 𝑓 (𝑦) d𝑦 ∈ C 𝜔 (𝑉) . □

Remark 3.1.5 There exist analytic hypoelliptic linear PDOs that are not hypoelliptic
and whose characteristic set intersects every cotangent space on a proper subvariety,
i.e., there are noncharacteristic hyperplanes through each point. This is the case for
the differential operator in R2

(D1 + 𝑖𝑥 12 𝑝 D2 ) 2 + 𝑐D2 (𝑝 ≥ 1, 𝑐 ≠ 0),

as shown in [Okaji, 1988]. The transpose of such a linear PDO cannot have a local
analytic parametrix.

Later we will need the analytic analogue of Lemma 2.4.6. For this we assume
that Ω is connected and we select a basis ΩC𝜈 𝜈=1,2,... of neighborhoods of Ω in

C𝑛 ; for each 𝜈, ΩC𝜈+1 ⊂ ΩC𝜈 , ΩC𝜈 a domain in C𝑛 . Let O ΩC𝜈 be the space of
holomorphic functions in ΩC𝜈 equipped with its natural Fréchet–Montel topology,
that of uniform convergence on compact subsets of Ω𝜈 . We have a continuous
C

embedding O Ω𝜈 ↩→ O Ω𝜈+1 . Since Ω𝜈 is connected, the restriction (i.e., trace)
C C C

map O ΩC𝜈 −→ C 𝜔 (Ω) is injective and allows us to identify O ΩC𝜈 with a linear
subspace of C 𝜔 (Ω) and to view C 𝜔 (Ω) as the union of the subspaces O ΩC𝜈 . We
can assume (see Theorem 6.2.14 below) that every ΩC𝜈 is a Runge domain, i.e., the
restrictions of entire functions in C 𝑛 to ΩC is dense in O ΩC (Definition 6.1.5,
𝜈 𝜈
below); thus every O ΩC𝜈 is dense (and not closed!) in C 𝜔 (Ω). We equip C (Ω)
𝜔

with the inductive limit topology defined by the subspaces O Ω𝜈 : a convex

C
subset
of C 𝜔 (Ω) is open if and only if its intersection with each and every O ΩC𝜈 is open
in O ΩC𝜈 . It is easily checked that this topology does not depend on the choice of
the basis ΩC𝜈 𝜈=1,2,... of neighborhoods of Ω in C𝑛 .
In what follows ΩC1 is selected so that the coefficients of 𝑃 = 𝑃 (𝑥, D 𝑥 ) extend
holomorphically to ΩC1 .

Lemma 3.1.6 If 𝑃 is analytic hypoelliptic in Ω then to every integer 𝜈 ≥ 1 and every

compact subset 𝐾 C of ΩC𝜈 there is a compact subset 𝐾1C of ΩC𝜈 and a compact subset
𝐾 ′ of Ω such that

∀ 𝑓 ∈ O ΩC𝜈 , max | 𝑓 | ≲ max
′
| 𝑓 | + max |𝑃 𝑓 | .
𝐾C 𝐾 𝐾1C
38 3 Analytic Tools in Distribution Theory

Proof For every 𝜈, the restriction of 𝑃 (𝑧, D𝑧 ) to ΩC𝜈 is a continuous linear operator
mapping O ΩC𝜈 into itself; 𝑓 ↦→ ( 𝑓 | Ω , 𝑃 𝑓 ) is a continuous linear map of O Ω𝜈
C

into C (Ω) × O ΩC𝜈 , per force injective since ΩC𝜈 is connected [the topology of the
space C (Ω) of continuous
functions in Ω is that of uniform convergence on compact
subsets of Ω]. Let 𝑓 𝑗 𝑗=1,2,... be a sequence in O ΩC𝜈 whose restrictions to Ω

converge to a function 𝑓◦ ∈ C (Ω) and such that 𝑃 𝑓 𝑗 converges in O ΩC𝜈1 for some
positive integer 𝜈1 ≥ 𝜈, necessarily to a holomorphic function whose restriction to
to 𝑃 (𝑥, D 𝑥 ) 𝑓◦ . If 𝑃 is analytic hypoelliptic in Ω then necessarily 𝑓◦ ∈
Ω is equal
O ΩC𝜈2 for some 𝜈2 ≥ 𝜈1 . We conclude that the graph G of the map

C 𝜔 (Ω) ∋ 𝑓 ↦→ ( 𝑓 , 𝑃 𝑓 ) ∈ C (Ω) × C 𝜔 (Ω) (3.1.1)

is closed. Since the Open Mapping Theorem applies to C 𝜔 (Ω) and to G (equipped
with the graph topology) we conclude that (3.1.1) is an isomorphism of C 𝜔 (Ω) onto
G. Lemma 3.1.6 expresses the fact that its inverse is continuous. □

3.2 Ehrenpreis’ Cutoffs and Analytic Regularity of Distributions

3.2.1 Ehrenpreis cutoffs

The fact that no analytic function can have compact support unless it vanishes
identically is a serious hindrance when one tries to exploit a priori estimates to
prove the analyticity of solutions of a linear PDE. To circumvent these difficulties
one frequently uses Ehrenpreis’ cutoff functions, whose distinguishing feature is that
they satisfy estimates of the kind (1.2.1) but only up to a given order of differentiation.
If X is a normed vector space, with norm ∥·∥, and 𝐴, 𝐵 two subsets of X, we use the
notation
dist ( 𝐴, 𝐵) = sup ∥𝑥 − 𝑦∥ . (3.2.1)
𝑥 ∈ 𝐴,𝑦 ∈𝐵

If 𝐴 consists of a single point we write dist (𝑥, 𝐵) rather than dist ({𝑥} , 𝐵).

Proposition 3.2.1 Let 𝑉 ⊂ 𝑈 be two open subsets of R𝑛 . If dist (𝑉, R𝑛 \𝑈) > 0
there is a sequence of functions 𝜑 𝑁 ∈ C ∞ (R𝑛 ) (𝑁 = 1, 2, ...) having the following
properties:
(1) for every 𝑁 and for every 𝑥 ∈ R𝑛 , 0 ≤ 𝜑 𝑁 (𝑥) ≤ 1;
(2) for every 𝑁, 𝑥 ∈ 𝑉 =⇒ 𝜑 𝑁 (𝑥) = 1 and 𝑥 ∉ 𝑈 =⇒ 𝜑 𝑁 (𝑥) = 0;
(3) there is a constant 𝐶◦ > 0 independent of 𝑁 such that

∀𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑁, max |𝜕 𝛼 𝜑 𝑁 | ≤ (𝐶◦ 𝑁) | 𝛼 | . (3.2.2)

Proof Let 𝜌 𝜀 (𝜀 > 0) denote a standard

∫ mollifier: 𝜌 ∈ C ∞ (R𝑛 ) is such that 𝜌 (𝑥) > 0
if |𝑥| < 1, 𝜌 (𝑥) = 0 if |𝑥| ≥ 1, R𝑛 𝜌 (𝑥) d𝑥 = +1; and define 𝜌 𝜀 (𝑥) = 𝜀 −𝑛 𝜌 (𝑥/𝜀).
Let 𝑓 ∈ 𝐿 𝑝 (R𝑛 ) (1 ≤ 𝑝 ≤ +∞) be arbitrary; the convolution
3.2 Ehrenpreis’ Cutoffs and Analytic Regularity of Distributions 39
∫
(𝜌 𝜀 ∗ 𝑓 ) (𝑥) = 𝜌 𝜀 −1 𝑥 − 𝑦 𝑓 (𝜀𝑦) d𝑦
R𝑛

is a C ∞ function, vanishing if dist (𝑥, supp 𝑓 ) > 𝜀. We have

∥ 𝜌 𝜀 ∗ 𝑓 ∥ 𝐿 𝑝 ≤ ∥ 𝜌 𝜀 ∥ 𝐿1 ∥ 𝑓 ∥ 𝐿 𝑝 = ∥ 𝑓 ∥ 𝐿 𝑝 . (3.2.3)
𝜕
If 𝜕𝑖 = 𝜕𝑥𝑖 , 1 ≤ 𝑖 ≤ 𝑛, we have
∫
𝜕𝑖 (𝜌 𝜀 ∗ 𝑓 ) (𝑥) = 𝜀 −1 𝜕𝑖 𝜌 𝜀 −1 𝑥 − 𝑦 𝑓 (𝜀𝑦) d𝑦,

whence
∥𝜕𝑖 (𝜌 𝜀 ∗ 𝑓 )∥ 𝐿 𝑝 ≤ 𝑀𝜀 −1 ∥ 𝑓 ∥ 𝐿 𝑝 , (3.2.4)
∫
where 𝑀 = max |𝜕𝑖 𝜌 (𝑥)| d𝑥.
1≤𝑖 ≤𝑛
Define 𝑈′ = 𝑥 ∈ 𝑈; dist (𝑥, 𝑉) < 21 dist (𝑉, R𝑛 \𝑈) , and let 𝜒𝑈′ denote the
1
characteristic function of 𝑈 ′. We can take 𝜀 = 2𝑁 dist (𝑉, R𝑛 \𝑈) and define

𝑁 factors
z }| {
𝜑𝑁 = 𝜌 𝜀 ∗ 𝜌 𝜀 ∗ · · · 𝜌 𝜀 ∗ 𝜒𝑈′ . (3.2.5)

We have 𝜑 𝑁 ∈ C ∞ (R𝑛 ); Property (1) in the statement is a direct consequence of

the fact that the convolution of nonnegative functions is nonnegative and of (3.2.3)
where 𝑝 = +∞. Property (2) is a direct consequence of the classical properties of
the support of a convolution. Lastly we note that if 𝜈 ≤ 𝑁 and 𝑖 1 , ..., 𝑖 𝜈 are integers
in the interval [1, ..., 𝑛],
𝑁 factors
z }| {

𝜕𝑖1 · · · 𝜕𝑖𝜈 𝜑 𝑁 = 𝜕𝑖1 𝜌 𝜀 ∗ · · · ∗ 𝜕𝑖𝜈 𝜌 𝜀 ∗ 𝜌 𝜀 ∗ · · · 𝜌 𝜀 ∗ 𝜒𝑈′ .

Applying (3.2.4) with 𝑝 = ∞ we get the estimates

𝜕𝑖1 · · · 𝜕𝑖𝜈 𝜑 𝑁 𝐿∞
≤ (𝑀/𝜀) 𝜈 , (3.2.6)

yielding immediately (3.2.2), with 𝐶◦ = 2𝑀/dist (𝑉, R𝑛 \𝑈). □

Definition 3.2.2 A sequence of functions 𝜑 𝑁 (𝑁 = 1, 2, ...) satisfying properties (1),

(2) and (3) in Proposition 3.2.1 will be called an Ehrenpreis sequence relative to
the pair (𝑈, 𝑉).

Remark 3.2.3 Let 𝜑 𝑁 (𝑁 = 1, 2, ...) be an Ehrenpreis sequence relative to some pair

(𝑈, 𝑉). Define 𝜑e𝑁 = 𝜑 𝑁◦ +𝑁 with 𝑁◦ a positive integer; then { 𝜑
e𝑁 } 𝑁 =1,2,... is also an
Ehrenpreis sequence relative to the pair (𝑈, 𝑉). Obviously it has Properties (1) and
(2) in Proposition 3.2.1. As for property (3) we note that, whatever 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑁,
40 3 Analytic Tools in Distribution Theory

e𝑁 | ≤ (𝐶◦ (𝑁 + 𝑁◦ )) | 𝛼 |
max |D 𝛼 𝜑
≤ (𝐶◦ (1 + 𝑁◦ ) 𝑁) | 𝛼 | .

3.2.2 Using Ehrenpreis cutoffs to prove local analyticity

Proposition 3.2.4 Let Ω be an open subset of R𝑛 . For 𝑓 ∈ C ∞ (Ω) to be real-

analytic at a point 𝑥 ◦ ∈ Ω it is necessary and sufficient that there be a pair of
relatively compact open subsets 𝑈, 𝑉 of Ω with 𝑥 ◦ ∈ 𝑉 ⊂⊂ 𝑈 and an Ehrenpreis
sequence 𝜑 𝑁 (𝑁 = 1, 2, ...) relative to the pair (𝑈, 𝑉) such that, for some constant
𝐶 > 0 independent of 𝑁 and for every 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑁,

max |D 𝛼 (𝜑 𝑁 𝑓 ) (𝑥)| ≲ 𝐶 𝑁 𝛼!. (3.2.7)

𝑥

Proof The condition is sufficient since 𝜑 𝑁 (𝑥) = 1 if 𝑥 ∈ 𝑉. To prove that it is

necessary we select a neighborhood 𝑈 of 𝑥 ◦ sufficiently small and a constant 𝐶1 > 0
sufficiently large that
|𝛽 |
∀𝛽 ∈ Z+𝑛 , sup D𝛽 𝑓 (𝑥) ≲ 𝐶1 𝛽!. (3.2.8)
𝑥 ∈𝑉

Taking |𝛼| ≤ 𝑁 we derive from the Leibniz rule (1.1.3) and from (3.2.2) and (3.2.8):
∑︁ 𝛼! |𝛽 |
|D 𝛼 (𝜑 𝑁 (𝑥) 𝑓 (𝑥))| ≲ 𝐶 (𝐶◦ 𝑁) | 𝛼−𝛽 |
𝛽⪯𝛼
(𝛼 − 𝛽)! 1

≲ 𝛼!𝐶1| 𝛼 | exp (𝐶◦ 𝑁) ,

whence (3.2.7) with 𝐶 = 𝐶1 exp 𝐶◦ . □

Proposition 3.2.5 For 𝑢 ∈ D ′ (R𝑛 ) to be real-analytic at a point 𝑥 ◦ ∈ R𝑛 it is
necessary and sufficient that there be a pair of bounded open subsets 𝑈, 𝑉 of R𝑛
with 𝑥 ◦ ∈ 𝑉 ⊂⊂ 𝑈 and an Ehrenpreis sequence 𝜑 𝑁 (𝑁 = 1, 2, ...) relative to the pair
(𝑈, 𝑉) such that
∀𝜉 ∈ R𝑛 , |𝜉 | 𝑘 | 𝜑d 𝑁
𝑁 𝑢 (𝜉)| ≲ 𝐶 𝑘! (3.2.9)
for some constant 𝐶 > 0 independent of 𝑁 and of every 𝑘 ∈ Z+ , 𝑘 ≤ 𝑁.
Proof I. Necessity of the condition. Select the open set 𝑈 so that 𝑢 ∈ C 𝜔 (Ω) with
Ω an open set containing the (compact) closure of 𝑈. We apply Proposition 3.2.4
with 𝑓 equal to 𝑢 in Ω. We derive from the Stirling’s Formula and from (3.2.8):
∑︁ ∑︁ ∫
𝛼
|𝜉 𝜑d 𝑁 𝑢 (𝜉)| = e−𝑖 𝑥· 𝜉 D 𝑥𝛼 (𝜑 𝑁 𝑢) (𝑥) d𝑥
| 𝛼 |=𝑘 | 𝛼 |=𝑘
∑︁
≤ (Vol 𝑈) max D 𝑥𝛼 (𝜑 𝑁 𝑢) (𝑥) ≲ 𝐶1𝑁 𝑘!
𝑥
| 𝛼 |=𝑘
3.2 Ehrenpreis’ Cutoffs and Analytic Regularity of Distributions 41

for a suitably large constant 𝐶1 > 0 independent of 𝑁, whence the claim.

II. Sufficiency of the condition. Select 𝑁 > 𝑛. The Fourier inversion formula
yields (2𝜋) 𝑛 𝜑 𝑁 𝑢 = 𝑣 𝑁 + 𝑤 𝑁 where
∫
𝑣 𝑁 (𝑥) = e𝑖 𝑥· 𝜉 𝜑d
𝑁 𝑢 (𝜉) d𝜉,
| 𝜉 | ≤1
∫
𝑤 𝑁 (𝑥) = e𝑖 𝑥· 𝜉 𝜑d
𝑁 𝑢 (𝜉) d𝜉.
| 𝜉 | ≥1

The first integral extends as an entire function in C𝑛 ; moreover we derive from (3.2.9)
∫
𝛼
|D 𝑣 𝑁 (𝑥)| ≤ |𝜉 𝛼 𝜑d
𝑁 𝑢 (𝜉)| d𝜉
| 𝜉 | ≤1
∫
𝑁 +1
≤ | 𝜑d
𝑁 𝑢 (𝜉)| d𝜉 ≲ 𝐶
| 𝜉 | ≤1

whatever 𝛼 ∈ Z+𝑛 . We also derive from (3.2.9)

∑︁ ∫
𝛼
|D 𝑤 𝑁 (𝑥)| ≤ |𝜉 | 𝑁 −𝑛−1 | 𝜑d
𝑁 𝑢 (𝜉)| d𝜉
| 𝛼 |=𝑁 −𝑛−1 | 𝜉 | ≥1
∫
d𝜉
𝑁 +1
≤ sup |𝜉 | 𝑁 | 𝜑d
𝑁 𝑢 (𝜉)| ≲ 𝐶 𝑁!
| 𝜉 | ≥1 |𝜉 | 𝑛+1 R𝑛

whence (assuming 𝑁 > 𝑛 + 1)

∑︁ 𝑁
𝑁 −𝑛−1
|D 𝛼 𝑤 𝑁 (𝑥)| ≲ 𝐶 (2𝐶) 𝑁 −𝑛−1 (𝑁 − 𝑛 − 1)!
| 𝛼 |=𝑁 −𝑛−1

and therefore, for a suitably large constant 𝐶2 > 0 independent of 𝑁,

∑︁
max
𝑛
|D 𝛼 (𝜑 𝑁 𝑢)| ≤ 𝐶2𝑁 −𝑛 (𝑁 − 𝑛 − 1)!.
R
| 𝛼 |=𝑁 −𝑛−1

Since 𝜑 𝑁 ≡ 1 in 𝑉 we conclude that 𝑢 ∈ C 𝜔 (𝑉). □

We are going to need partitions of unity consisting of Ehrenpreis’ cutoffs. Let
𝐾 ⊂ R𝑛 be a compact set, {𝑈 𝜄 } 𝜄=1,...,𝑟 and {𝑉 𝜄 } 𝜄=1,...,𝑟 be two families of 𝑟 ≥ 2 open
Ø 𝑟
sets such that 𝑉 𝜄 ⊂⊂ 𝑈 𝜄 for each 𝜄, and such that 𝐾 ⊂ 𝑉 𝜄 . For each 𝜄 we select an
𝜄=1
Ehrenpreis sequence 𝜑 𝑁 , 𝜄 𝑁 =1,2,... relative to the pair 𝑉 𝜄 , 𝑈 𝜄 : supp 𝜑 𝑁 , 𝜄 ⊂ 𝑈 𝜄 and

𝜑 𝑁 , 𝜄 ≡ 1 in 𝑉 𝜄 . Fixing 𝑁 we define a new set of cutoff functions 𝜓 𝑁 , 𝜄 𝑁 =1,2,... .

We take 𝜓 𝑁 ,1 = 𝜑 𝑁 ,1 , 𝜓 𝑁 ,2 = 1 − 𝜑 𝑁 ,1 𝜑 𝑁 ,2 ; we have supp 𝜓 𝑁 ,2 ⊂ 𝑉2 . We
have 𝜓 𝑁 ,1 = 1, 𝜓 𝑁 ,2 = 0 in 𝑉1 and 𝜓 𝑁 ,1 + 𝜓 𝑁 ,2 = 1 in 𝑉2 . If 𝑟 ≥ 3 we take
42 3 Analytic Tools in Distribution Theory

𝜓 𝑁 ,3 = 1 − 𝜑 𝑁 ,1 1 − 𝜑 𝑁 ,2 𝜑 𝑁 ,3 ; supp 𝜓 𝑁 , 𝜄 ⊂ 𝑈 𝜄 for 𝜄 = 1, 2, 3; 𝜓 𝑁 ,3 = 0 in
𝑉1 ∪ 𝑉2 and 𝜓 𝑁 ,1 + 𝜓 𝑁 ,2 + 𝜓 𝑁 ,3 = 1 in 𝑉3 . Etc. The Leibniz rule ensures that 𝜓 𝑁 , 𝜄
will satisfy inequalities of the kind (3.2.2).

3.3 Distribution Boundary Values of Holomorphic Functions

3.3.1 Function spaces in complex wedges

In this section we define the boundary value of a holomorphic function in a wedge

whose absolute value is allowed to have tempered growth at the edge (see Definition
3.3.2 below), and we show how every distribution can be represented as a sum of such
boundary values. This will be important for the lifting of the analysis to phase-space.
By a wedge we shall generally mean a subset of complex space C𝑛 of the kind

W𝛿 (Ω, Γ) = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; 𝑥 ∈ Ω, 𝑦 ∈ Γ, |𝑦| < 𝛿} . (3.3.1)

The edge Ω ⊂ R𝑛 of W𝛿 (Ω, Γ) is an open set and Γ ⊂ R𝑛 \ {0} is an open

and connected cone, never empty and (unless specified otherwise) acute, meaning
that the closure of Γ in R𝑛 \ {0} is contained in an open half-space 𝑦 · 𝜉 ◦ > 0
(0 ≠ 𝜉 ◦ ∈ R𝑛 ). In special instances it will be convenient to deal with a cone of
revolution, i.e., a cone (per force acute) that can be transformed by a rotation into a
cone {𝑦 ∈ R𝑛 ; |𝑦 ′ | < 𝜅𝑦 𝑛 } (0 < 𝜅 < +∞). Here and in the sequel, we frequently use
the notation 𝑦 ′ = (𝑦 1 , ..., 𝑦 𝑛−1 ). The positive number 𝛿 will sometimes be referred
to as the height of the wedge W𝛿 (Ω, Γ); most often, 𝛿 ≤ 1.
We recall that O (W𝛿 (Ω, Γ)) stands for the set of holomorphic functions in
W𝛿 (Ω, Γ); it is an algebra with respect to ordinary addition and multiplication.

Definition 3.3.1 Given any number 𝑚 ≥ 0 we define O (𝑚) (W𝛿 (Ω, Γ)) as the set
of functions ℎ ∈ O (W𝛿 (Ω, Γ)) such that |Im 𝑧| 𝑚 |ℎ (𝑧)| is bounded in W𝛿 (Ω, Γ).

The norm

ℎ ↦→ 𝑵 𝑚 (ℎ; Ω, Γ, 𝛿) = sup (|Im 𝑧| 𝑚 |ℎ (𝑧)|) (3.3.2)

𝑧 ∈W𝛿 (Ω,Γ)

turns O (𝑚) (W𝛿 (Ω, Γ)) into a Banach space. Obviously, for every 𝑚 ∈ Z+ ,

O (𝑚) (W𝛿 (Ω, Γ)) ⊂ O (𝑚+1) (W𝛿 (Ω, Γ))

and
∀ℎ ∈ O (𝑚) (W𝛿 (Ω, Γ)) , 𝑵 𝑚+1 (ℎ; Ω, Γ, 𝛿) ≤ 𝛿𝑵 𝑚 (ℎ; Ω, Γ, 𝛿) .
It will be convenient to use the notation
Ø
Otemp (W𝛿 (Ω, Γ)) = O (𝑚) (W𝛿 (Ω, Γ)) ; (3.3.3)
𝑚≥0
3.3 Distribution Boundary Values of Holomorphic Functions 43

Otemp (W𝛿 (Ω, Γ)) carries a natural locally convex topology: a convex set is open
in Otemp (W𝛿 (Ω, Γ)) if and only if its intersection with every Banach space
O (𝑚) (W𝛿 (Ω, Γ)) (𝑚 ≥ 0) is open in the latter.

Definition 3.3.2 We shall say that the function 𝐹 ∈ O (W𝛿 (Ω, Γ)) has tempered
growth at the edge if, given any open set 𝑈 ⊂⊂ Ω and any open cone Γ ′ in
R𝑛 \ {0} such that Γ ′ ∩ S𝑛−1 ⊂⊂ Γ, the restriction of 𝐹 to W𝛿 (𝑈, Γ ′) belongs to
Otemp (W𝛿 (𝑈, Γ ′)).

Note that the set of functions 𝐹 ∈ O (W𝛿 (Ω, Γ)) with tempered growth at the
edge is strictly larger than Otemp (W𝛿 (Ω, Γ)).

3.3.2 Derivatives and antiderivatives of holomorphic functions in

wedges

To define the boundary value of a function ℎ ∈ W𝛿 (𝑈, Γ) we shall need the

following lemma about antiderivatives of holomorphic functions in a wedge.

Í Let Γ be an open cone of revolution; there is a holomorphic vector

Lemma 3.3.3
field 𝑍 = 𝑛𝑗=1 𝑐 𝑗 𝜕𝑧𝜕 𝑗 with the following property:
Given an arbitrary open ball 𝑈 ⊂⊂ Ω there is a continuous linear map 𝔗 𝛿
of Otemp (W𝛿 (𝑈, Γ)) into itself such that ℎ = 𝑍𝔗 𝛿 ℎ for all ℎ ∈ Otemp (W𝛿 (𝑈, Γ));
𝔗 𝛿 maps O (0) (W𝛿 (𝑈, Γ)) into itself and, for any integer 𝑚 ≥ 2, maps
O (𝑚) (W𝛿 (𝑈, Γ)) into O (𝑚−1) (W𝛿 (𝑈, Γ)); and 𝔗 2𝛿 maps O (1) (W𝛿 (𝑈, Γ)) into
O (0) (W𝛿 (𝑈, Γ)).

Proof After a translation we may take 𝑈 to be a ball centered at 0 with ra-

dius 𝑟 > 0; and after a real rotation in 𝑧-space we may also assume that
Γ = {𝑦 ∈ R𝑛 ; |𝑦 ′ | < 𝜅𝑦 𝑛 }, 𝜅 > 0. With this choice of Γ it is convenient to use
the following norms in Otemp (W𝛿 (𝑈, Γ)):
′
𝑵𝑚 (ℎ; 𝑈, Γ, 𝛿) = sup 𝑦𝑚𝑛 |ℎ (𝑧)| .
𝑧 ∈W𝛿 (𝑈,Γ)

We have [see (3.3.2)]:

− 21 𝑚 ′
𝑵𝑚 (ℎ; 𝑈, Γ, 𝛿)
1 + 𝜅2 ≤ ≤1
𝑵 𝑚 (ℎ; 𝑈, Γ, 𝛿)

for every ℎ ∈ O (𝑚) (W𝛿 (𝑈, Γ)). For those same functions ℎ and using the notation
𝑧 ′ = (𝑧 1 , ..., 𝑧 𝑛−1 ) we define
∫ 𝑥𝑛 ∫ 𝜏
𝔗 𝛿 ℎ (𝑧) = ℎ (𝑧 ′, 𝑠 + 𝑖𝜏) d𝑠 − 𝑖 ℎ (𝑧 ′, 𝑥 𝑛 + 𝑖𝑡) d𝑡, (3.3.4)
0 𝑦𝑛
44 3 Analytic Tools in Distribution Theory
√
where 𝜏 = 𝛿/ 1 + 𝜅 2 . Note that, in the second integral, we have |𝑦 ′ | < 𝜅𝑦 𝑛 < 𝜅𝑡. On
the one hand,
∫ 𝜏
𝜕 𝜕
𝔗 𝛿 ℎ (𝑧) = ℎ (𝑧 ′, 𝑥 𝑛 + 𝑖𝜏) − ℎ (𝑧 ′, 𝑥 𝑛 + 𝑖𝑡) d𝑡 = ℎ (𝑧) .
𝜕𝑥 𝑛 𝑦𝑛 𝜕𝑡

𝜕
On the other hand, differentiation under the integral sign entails 𝜕 𝑧¯ 𝑗 𝔗 𝛿 ℎ ≡ 0 for
𝑗 = 1, ..., 𝑛 − 1: thus 𝔗 𝛿 ℎ is holomorphic and ℎ = 𝜕𝑧𝜕𝑛 𝔗 𝛿 ℎ in W𝛿 (𝑈, Γ).
We must now prove the bounds on 𝔗 𝛿 . We derive directly from (3.3.4):

|𝔗 𝛿 ℎ (𝑧)| ≤ (𝑟 + 𝜏 − 𝑦 𝑛 ) 𝑁0 (ℎ; 𝑈, Γ, 𝛿) .

Next we consider the cases 𝑚 ≥ 2. We have

𝑚−1 ∫ 𝑥𝑛
−1 𝑦 𝑛
𝑚−1
𝑦 𝑛 |𝔗 𝛿 ℎ (𝑧)| ≤ 𝜏 𝜏 𝑚 |ℎ (𝑧 ′, 𝑠 + 𝑖𝜏)| d𝑠
𝜏 0
∫ 𝜏
d𝑡
+ 𝑦 𝑚−1
𝑛 |𝑡 𝑚 ℎ (𝑧 ′, 𝑥 𝑛 + 𝑖𝑡)| 𝑚
𝑦𝑛 𝑡
𝑦 𝑚−1
−1 1 𝑛 ′
≤ 𝜏 𝑟+ 1− 𝑁𝑚 (ℎ; 𝑈, Γ, 𝛿) .
𝑚−1 𝜏

Lastly, we deal with the case 𝑚 = 1. We derive from (3.3.4), for 𝑧 ∈ W𝛿 (𝑈, Γ)
(and 𝑦 𝑛 ≤ 𝜏),
∫ 𝜏
d𝑡
|𝔗 𝛿 ℎ (𝑧)| ≤ 𝜏 −1 𝑟 + 𝑵 1′ (ℎ; 𝑈, Γ, 𝛿)
𝑦𝑛 𝑡

≤ 𝜏 𝑟 − log (𝑦 𝑛 /𝜏) 𝑵 1′ (ℎ; 𝑈, Γ, 𝛿)
−1

as well as
∫ 𝑥𝑛 ∫ 𝜏
𝔗 2𝛿 ℎ (𝑧) ≤ |𝔗 𝛿 ℎ (𝑧 ′, 𝑠 + 𝑖𝑡)| d𝑠 + |𝔗 𝛿 ℎ (𝑧 ′, 𝑥 𝑛 + 𝑖𝑡)| d𝑡.
0 𝑦𝑛

From these two estimates we get

𝔗 2𝛿 ℎ (𝑧)
∫ 𝑥𝑛 ∫ 𝜏
′ ≤ 𝜏 −1 𝑟 d𝑠 + 𝜏 −1 𝑟 − log 𝜏 −1 𝑡 d𝑡
𝑵 1 (ℎ; 𝑈, Γ, 𝛿) 0 𝑦𝑛
∫ 1
≤ 𝑟 + 𝜏 −1 𝑟 2 + 𝜏 |log 𝑡| d𝑡. □
0

Corollary 3.3.4 Let the open ball 𝑈 ⊂⊂ Ω, the cone Γ and the holomorphic vector
field 𝑍 be as in Lemma 3.3.3. Let 𝛿 > 0 and 𝑚 ∈ Z+ be arbitrary. To every
ℎ ∈ O (𝑚) (W𝛿 (𝑈, Γ)) there is an 𝑓 ∈ O (0) (W𝛿 (𝑈, Γ)) such that ℎ = 𝑍 𝑚+1 𝑓 in
W𝛿 (𝑈, Γ).
3.3 Distribution Boundary Values of Holomorphic Functions 45

Proof It suffices to take 𝑓 = 𝔗 𝑚+1

𝛿 ℎ. □
It is also useful to know that the partial derivatives act continuously between the
appropriate spaces of holomorphic functions in wedges.

Lemma 3.3.5 Let Γ∗ be an open cone of revolution in R𝑛 \ {0}, Γ∗ ∩ S𝑛−1 ⊂⊂ Γ. If

ℎ ∈ O (𝑚) (W𝛿 (Ω, Γ)) (𝑚 ∈ Z+ , 𝛿 > 0) then, given any positive number 𝛿 ′ < 𝛿 and
any multi-index 𝛼 ∈ Z+ , we have 𝜕𝑧𝛼 ℎ ∈ O (𝑚+| 𝛼 |) (W𝛿′ (Ω, Γ∗ )). Moreover, to every
open set 𝑈 ⊂⊂ Ω there is a positive constant 𝐶 independent of 𝑚, 𝛼 and ℎ, such that

𝑵 𝑚+| 𝛼 | 𝜕𝑧𝛼 ℎ; 𝑈, Γ∗ , 𝛿 ′ ≤ 𝐶 𝑚+| 𝛼 | 𝛼!𝑵 𝑚 (ℎ; 𝑈, Γ, 𝛿) .

(3.3.5)

Proof After a real rotation in C𝑛 we may take Γ∗ = {𝑦 ∈ R𝑛 ; |𝑦 ′ | < 𝜅𝑦 𝑛 }, 0 < 𝜅 <

+∞. Fix 𝑧◦ = 𝑥 ◦ + 𝑖𝑦 ◦ ∈ 𝑈 + 𝑖Γ∗ , |𝑦 ◦ | < 𝛿 ′. We select a number 𝜌 satisfying the
following conditions:

1 1 𝛿
0 < 𝜌 < min , ′ − 1 , (3.3.6)
2 2 𝛿
1 ′
𝜌𝛿 < |𝑥 − 𝑥 ◦ | whatever 𝑥 ∉ Ω, 𝑥 ◦ ∈ 𝑈, (3.3.7)
2
𝑦 𝑦◦
𝜌< − ◦ whatever 0 ≠ 𝑦 ∉ Γ, 𝑦 ◦ ∈ Γ∗ . (3.3.8)
|𝑦| |𝑦 |

They ensure that the polydisk

1
Δ = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 − 𝑧 ◦𝑗 ≤ √ 𝜌 |𝑦 ◦ | , 𝑗 = 1, ..., 𝑛
2 𝑛

is entirely contained in W𝛿 (Ω, Γ). Indeed, 𝑥 +𝑖𝑦 ∈ Δ entails |𝑦| ≤ |𝑦 ◦ | 1 + 21 𝜌 < 𝛿
by (3.3.6). By (3.3.7) we get |𝑥 − 𝑥 ◦ | ≤ 21 𝜌 |𝑦 ◦ | < 12 𝜌𝛿 ′. We also have

|𝑦| 𝑦 𝑦◦ 1
1− ≤ − ≤ 𝜌, (3.3.9)
|𝑦 ◦ | |𝑦 ◦ | |𝑦 ◦ | 2

whence, by (3.3.8),

𝑦 𝑦◦ |𝑦| 𝑦 𝑦◦
− ◦ ≤ 1 − ◦ + ◦ − ◦ ≤ 𝜌.
|𝑦| |𝑦 | |𝑦 | |𝑦 | |𝑦 |
•
Let Δ denote the distinguished boundary of Δ; according to (3.3.9),

• 1 1
𝑧 ∈ Δ =⇒ |𝑦| ≥ 1 − 𝜌 |𝑦 ◦ | ≥ |𝑦 ◦ | .
2 2

This allows us to derive from the Cauchy inequalities

46 3 Analytic Tools in Distribution Theory
√ ! |𝛼|
|𝑦 ◦ | | 𝛼 | 𝜕 𝛼 ℎ ◦ 2𝑛
(𝑧 ) ≤ max |ℎ|
𝛼! 𝜕𝑧 𝛼 𝜌 Δ

𝑚 √ ! |𝛼|
2 2𝑛
≤ max (|𝑦| 𝑚 |ℎ (𝑧)|) ,
|𝑦 ◦ | 𝜌 𝑧 ∈Δ

and therefore
√ ! |𝛼|
|𝑦 ◦ | 𝑚+| 𝛼 | 𝜕 𝛼 ℎ ◦ 2𝑛
(𝑧 ) ≤ 2𝑚 𝑵 𝑚 (ℎ; 𝑈, Γ, 𝛿). □
𝛼! 𝜕𝑧 𝛼 𝜌

3.3.3 Distribution boundary values of holomorphic functions in wedges

The wedge W𝛿 (Ω, Γ) is defined in (3.3.1).

Theorem 3.3.6 Let Ω ⊂ R𝑛 be an open set and the cone Γ ⊂ R𝑛 \ {0} be open and
connected. To each function ℎ ∈ Otemp (W𝛿 (Ω, Γ)) there is a distribution 𝑢 in Ω
such that, for any 𝑦 ∈ Γ and any 𝜑 ∈ Cc∞ (Ω),
∫
⟨𝑢, 𝜑⟩ = lim ℎ(𝑥 + 𝑖𝜆𝑦)𝜑 (𝑥) d𝑥. (3.3.10)
𝜆↘0

The assignment ℎ ↦→ 𝑢 is an injective linear map Otemp (W𝛿 (Ω, Γ)) → D ′ (Ω).
It is understood that 𝜆 |𝑦| < 𝛿 in (3.3.10).
Proof I. Existence of 𝑢. Let ℎ ∈ O (𝑚) (W𝛿 (Ω, Γ)), 𝑚 ∈ Z+ . Let 𝑦 ◦ ∈ Γ be
arbitrary. We select arbitrarily an open cone of revolution around the axis spanned
by 𝑦 ◦ , Γ∗ ⊂ Γ. Let 𝑍 = 𝑛𝑗=1 𝑐 𝑗 𝜕𝑧𝜕 𝑗 (𝑐 𝑗 ∈ C) be the vector field associated to the cone
Í

Γ∗ as per Lemma 3.3.3, i.e., the holomorphic derivative in the direction of 𝑦 ◦ . Let
{𝑈 𝜄 } 𝜄 ∈𝐼 be a locally finite covering of Ω by open balls whose closures are contained
in Ω and let {𝑔 𝜄 } 𝜄 ∈𝐼 be a C ∞ partition of unity in Ω subordinate to this covering.
By Corollary 3.3.4, whatever 𝜄 ∈ 𝐼 we have ℎ = 𝑍 𝑚+1 𝑓 𝜄 in W𝛿 (𝑈 𝜄 , Γ) for some
function 𝑓 𝜄 ∈ O (0) (W𝛿 (𝑈 𝜄 , Γ)). Thanks to this we can write, for any 𝜑 ∈ Cc∞ (Ω),
∫
lim ℎ(𝑥 + 𝑖𝜆𝑦 ◦ )𝑔 𝜄 (𝑥) 𝜑 (𝑥) d𝑥
𝜆↘0
𝑚+1
∫ 𝑛
© ∑︁ 𝜕 ª
= lim 𝑓 𝜄 (𝑥 + 𝑖𝜆𝑦 ◦ ) ­− 𝑐𝑗 ® (𝑔 𝜄 (𝑥) 𝜑 (𝑥)) d𝑥
𝜆↘0 𝜕𝑥 𝑗
« 𝑗=1 ¬
𝑚+1
∫ 𝑛
© ∑︁ 𝜕 ª
= 𝑓 𝜄 (𝑥) ­− 𝑐𝑗 ® (𝑔 𝜄 (𝑥) 𝜑 (𝑥)) d𝑥,
𝑗=1
𝜕𝑥 𝑗
« ¬
3.3 Distribution Boundary Values of Holomorphic Functions 47
∫
whence lim ℎ(𝑥 + 𝑖𝜆𝑦 ◦ )𝜑 (𝑥) d𝑥 = ⟨𝑢, 𝜑⟩ with
𝜆↘0

𝑚+1
𝑛
∑︁ ©∑︁ 𝜕 ª
𝑢= 𝑔𝜄 ­ 𝑐𝑗 ® 𝑓𝜄.
𝜄 ∈𝐼 𝑗=1
𝜕𝑥 𝑗
« ¬
Let 𝑦 ∗ ∈ Γ, 𝑦 ∗ ≠ 𝑦 ◦ , and let [0, 1] ∋ 𝑡 ↦→ 𝛾 (𝑡) ∈ Γ be a smooth curve such that
𝛾 (0) = 𝑦 ◦ and 𝛾 (1) = 𝑡 ∗ . We have, for any 𝜓 ∈ Cc∞ (Ω),
∫
lim ( 𝑓 𝜄 (𝑥 + 𝑖𝜆𝑦 ∗ ) − 𝑓 𝜄 (𝑥 + 𝑖𝜆𝑦 ◦ )) 𝜓 (𝑥) d𝑥
𝜆↘0
∫ ∫ 1
• 𝜕 𝑓𝜄
= lim 𝑖𝜆 𝛾 (𝑡) · (𝑥 + 𝑖𝜆𝛾 (𝑡)) d𝑡 𝜓 (𝑥) d𝑥
𝜆↘0 0 𝜕𝑧
∫ ∫ 1
• 𝜕𝜓
= − lim 𝑖𝜆 𝑓 𝜄 (𝑥 + 𝑖𝜆𝛾 (𝑡)) 𝛾 (𝑡) · (𝑥) d𝑡d𝑥 = 0.
𝜆↘0 0 𝜕𝑥

This proves that 𝑢 does not depend on the choice of 𝑦 ◦ .

II. Injectivity of the map ℎ ↦→ 𝑢. There is no loss of generality in assuming Ω
is connected. We are going to show that if ℎ ∈ O (Ω + 𝑖Γ) is such that, for some
𝑦 ◦ ∈ Γ, ∫
∀𝜑 ∈ Cc∞ (𝑈) , lim 𝜑 (𝑥) ℎ(𝑥 + 𝑖𝜆𝑦 ◦ )d𝑥 = 0,
𝜆↘0

then ℎ ≡ 0 in Ω Í
+ 𝑖Γ. Here it is convenient to carry out a linear change of variables
𝑛
𝑧 𝑗 1≤ 𝑗 ≤𝑛 ⇝ 𝑘=1 𝑎 𝑗,𝑘 𝑧 𝑘 1≤ 𝑗 ≤𝑛 with real entries 𝑎 𝑗,𝑘 , allowing us to assume
◦
that 𝑦 = (0, ..., 0, 1) ∈ Γ♭ = {𝑦 ∈ R𝑛 ; 0 < 𝑦 1 < 𝑦 2 · · · < 𝑦 𝑛 } ⊂ Γ; we also select
𝑚+1
𝑈 = (−𝑟, 𝑟) 𝑛 ⊂ Ω, 𝑟 > 0. Corollary 3.3.4 allows us to write ℎ = 𝜕𝜕𝑧 𝑚+1𝑓 in 𝑈 + 𝑖Γ ′ for
𝑛
a suitable choice of the function 𝑓 ∈ O (0) (W𝛿 (𝑈, Γ)) and of the integer 𝑚 ≥ 0. It
follows that ∫
𝜕 𝑚+1 𝜑
∀𝜑 ∈ Cc∞ (𝑈) , 𝑓 (𝑥) 𝑚+1 (𝑥) d𝑥 = 0.
𝜕𝑥 𝑛
𝑚+1
This means that 𝜕𝜕𝑥 𝑚+1𝑓 = 0 in 𝑈; in other words, 𝑓 (𝑥) is a polynomial 𝑝 𝑚 (𝑥 ′, 𝑥 𝑛 )
𝑛
of degree ≤ 𝑚 with respect to 𝑥 𝑛 whose coefficients are continuous functions of
𝑥 ′ = (𝑥 1 , ..., 𝑥 𝑛−1 ) ∈ [−𝑟, 𝑟] 𝑛−1 . We can write, for each 𝑥 ∈ (−𝑟, 𝑟) 𝑛 and 0 < 𝑦 𝑛 < 𝛿,

𝜕 𝑚+1 𝑓 ′
ℎ (𝑥 ′, 𝑥 𝑛 + 𝑖𝑦 𝑛 ) = (𝑥 , 𝑥 𝑛 + 𝑖𝑦 𝑛 ) ,
𝜕𝑧 𝑚+1
𝑛
𝑓 (𝑥 ′, 𝑥 𝑛 + 𝑖𝑦 𝑛 ) = 𝑝 𝑚 (𝑥 ′, 𝑥 𝑛 + 𝑖𝑦 𝑛 ) + 𝑔 (𝑥 ′, 𝑥 𝑛 + 𝑖𝑦 𝑛 )

with 𝑔 (𝑥 ′, 𝑧 𝑛 ) holomorphic with respect to 𝑧 𝑛 in the open set (−𝑟, 𝑟) + 𝑖 (0, 𝛿) ⊂ C,

continuous in the semiclosed set (−𝑟, 𝑟) + 𝑖[0, 𝛿) and such that 𝑔 (𝑥) = 0 for all
𝑥 ∈ (−𝑟, 𝑟) 𝑛 . It is a well-known property of holomorphic functions of a single
complex variable that, for fixed 𝑥 ′ ∈ (−𝑟, 𝑟) 𝑛−1 , the function 𝑧 𝑛 ↦→ 𝑔 (𝑥 ′, 𝑧 𝑛 ) must
48 3 Analytic Tools in Distribution Theory

vanish identically, implying ℎ (𝑥 ′, 𝑧 𝑛 ) ≡ 0 in (−𝑟, 𝑟) 𝑛−1 × ((−𝑟, 𝑟) + 𝑖 (0, 𝛿)). Next

we note that, for any (𝑥 1 , ..., 𝑥 𝑛−2 , 𝑥 𝑛 ) ∈ (−𝑟, 𝑟) 𝑛−1 and any 𝑦 𝑛 ∈ (0, 𝛿), the function
𝑧 𝑛−1 ↦→ ℎ (𝑥 1 , ..., 𝑥 𝑛−2 , 𝑧 𝑛−1 , 𝑥 𝑛 + 𝑖𝑦 𝑛 ) is holomorphic in the rectangle (−𝑟, 𝑟) ×
(0, 𝑦 𝑛 ) and continuous in (−𝑟, 𝑟) × [0, 𝑦 𝑛 ). Since its restriction to 𝑦 𝑛−1 = 0 vanishes
identically we conclude, by the same one-complex variable argument used above,
that ℎ (𝑥 1 , ..., 𝑥 𝑛−2 , 𝑧 𝑛−1 , 𝑧 𝑛 ) ≡ 0 in the region under consideration. We can repeat
the reasoning with 𝑧 𝑛−2 , 𝑧 𝑛−3 , etc., to conclude that ℎ ≡ 0 in 𝑈 + 𝑖Γ♭ and therefore
in the domain Ω + 𝑖Γ. □
Definition 3.3.7 The distribution 𝑢 in (3.3.10) is called the boundary value of
ℎ ∈ Otemp (W𝛿 (𝑈, Γ)) in Ω and shall be denoted by 𝑏 Ω ℎ.
The continuity of the operator 𝑏 Ω : Otemp (W𝛿 (Ω, Γ)) −→ D ′ (Ω) can easily
be deduced from Part I of the proof of Theorem 3.3.6. Obviously, 𝜕𝑥𝜕 𝑗 𝑏 Ω ℎ = 𝑏 Ω 𝜕𝑧
𝜕ℎ
𝑗
(which makes sense thanks to Lemma 3.3.5). Note also that the boundary values
of holomorphic functions belonging to Otemp (W𝛿 (Ω, Γ)) form an algebra for the
multiplication (𝑏 Ω ℎ1 ) (𝑏 Ω ℎ2 ) = 𝑏 Ω (ℎ1 ℎ2 ).
Remark 3.3.8 The argument at the end of Part I of the proof of Theorem 3.3.6
makes it clear that, in order to define 𝑏 Ω ℎ, any nonempty open and connected cone
contained in Γ could have been used in place of Γ.

3.3.4 Representation of distributions as sums of boundary values

In this subsection we show that, locally, any distribution can be equated to the sum
of an analytic function and of a finite number of boundary values of holomorphic
functions in wedges “of infinite height” Ω + 𝑖Γ. For greater precision in the selection
of the latter functions we need to introduce new linear subspaces of O (Ω + 𝑖Γ).
Definition 3.3.9 Given 𝑚 ∈ R+ and 𝜏 > 0 we shall denote by O 𝜏(𝑚) (Ω + 𝑖Γ) the set
of functions ℎ ∈ O (Ω + 𝑖Γ) with the property that, for some constant 𝐶 > 0 and for
all 𝑘 ∈ Z+ , 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑘,

𝜏𝑘
sup |Im 𝑧| 𝑚+𝑘 𝜕𝑧𝛼 ℎ (𝑧) ≤ 𝐶. (3.3.11)
𝑘! 𝑧 ∈Ω+𝑖Γ

The infimum 𝑵 𝑚, 𝜏 (ℎ; Ω, Γ) of the constants 𝐶 in (3.3.11) can be taken as the

norm of ℎ ∈ O 𝜏(𝑚) (Ω + 𝑖Γ); 𝑵 𝑚, 𝜏 (ℎ; Ω, Γ) turns O 𝜏(𝑚) (Ω + 𝑖Γ) into a Banach
space. If 𝜏 < 𝜏 ′ then O 𝜏(𝑚)
′ (Ω + 𝑖Γ) ⊂ O 𝜏(𝑚) (Ω + 𝑖Γ) and the injection has norm
≤ 1; thus the numbers 𝜏 must be thought of as small and approaching zero. A function
ℎ (𝑥 + 𝑖𝑦) that belongs to O 𝜏(𝑚) (Ω + 𝑖Γ) decays exponentially as Γ ∋ 𝑦 −→ ∞:
Lemma 3.3.10 If ℎ ∈ O 𝜏(𝑚) (Ω + 𝑖Γ) then, for every 𝛼 ∈ Z+𝑛 ,

(𝜏/2) | 𝛼 | 1

sup |Im 𝑧| 𝑚+| 𝛼 | e 2 𝜏 |Im 𝑧 | 𝜕𝑧𝛼 ℎ (𝑧) ≤ 2𝑵 𝑚, 𝜏 (ℎ; Ω, Γ) . (3.3.12)
|𝛼|! 𝑧 ∈Ω+𝑖Γ
3.3 Distribution Boundary Values of Holomorphic Functions 49

Proof We have, for every 𝛼 ∈ Z+𝑛 , 𝑧 ∈ Ω + 𝑖Γ,

(𝜏/2) | 𝛼 | © ∑︁ (𝜏/2)
𝑘− | 𝛼 |
|Im 𝑧| 𝑚+| 𝛼 | ­ |Im 𝑧| 𝑘− | 𝛼 | ® 𝜕𝑧𝛼 ℎ (𝑧)
ª
|𝛼|! (𝑘 − |𝛼|)!
«𝑘 ≥ | 𝛼 | ¬
∑︁ 𝑘! 1 𝑘 𝑚+𝑘 𝛼
= (𝜏/2) |Im 𝑧| 𝜕𝑧 ℎ (𝑧)
|𝛼|! (𝑘 − |𝛼|)! 𝑘!
𝑘≥|𝛼|
∑︁ 𝑘!
≤ 2−𝑘 𝑵 𝑚, 𝜏 (ℎ; Ω, Γ) = 2𝑵 𝑚, 𝜏 (ℎ; Ω, Γ) . □
|𝛼|! (𝑘 − |𝛼|)!
𝑘≥|𝛼|

Clearly, the restriction from Ω + 𝑖Γ to W𝛿 (Ω, Γ) [𝛿 > 0, see (3.3.1)] defines

a continuous linear map O 𝜏(𝑚) (Ω + 𝑖Γ) −→ O (𝑚) (W𝛿 (Ω, Γ)) which is injective.
Thus it is permitted to talk of the boundary value 𝑏 Ω ℎ of an arbitrary function
ℎ ∈ O 𝜏(𝑚) (Ω + 𝑖Γ).
Lemma 3.3.11 Let Ω ⊂ R𝑛 be open and the cone Γ ⊂ R𝑛 \ {0} be open and con-
nected. Given any 𝑚 ∈ Z+ and any 𝛼 ∈ Z+𝑛 , the partial derivative 𝜕𝑧𝛼 defines a con-
|𝛼|
tinuous linear map O 𝜏(𝑚) (Ω + 𝑖Γ) −→ O 𝜏/2
(𝑚+ | 𝛼 |)
(𝑈 + 𝑖Γ) with norm ≤ |𝛼|! 𝜏2 .

Proof Indeed, if ℎ ∈ O 𝜏(𝑚) (Ω + 𝑖Γ) and |𝛽| ≤ 𝑘 then, for 𝑧 ∈ Ω + 𝑖Γ,

1 𝜏 𝑘
|Im 𝑧| 𝑚+| 𝛼 |+𝑘 𝜕𝑧 𝜕𝑧𝛼 ℎ (𝑧)
𝛽
𝑘! 2
(|𝛼| + 𝑘)! − | 𝛼 | 1
= 2−𝑘 𝜏 | 𝛼 |+𝑘 |Im 𝑧| 𝑚+| 𝛼 |+𝑘 𝜕𝑧 ℎ (𝑧) ,
𝛼+𝛽
𝜏
𝑘! (|𝛼| + 𝑘)!

whence (𝜏/2) | 𝛼 | 𝑵 𝑚+| 𝛼 |, 𝜏/2 𝜕𝑧𝛼 ℎ; Ω, Γ ≤ |𝛼|!𝑵 𝑚, 𝜏 (ℎ; Ω, Γ).

□
Theorem 3.3.12 Let ℭ ( 𝑗)
( 𝑗 = 1, ..., 𝜈) be pairwise disjoint,
acute cones in R𝑛 \ {0}
(1)
such that the Lebesgue measure of R \ ℭ ∪ · · · ∪ ℭ
𝑛 (𝜈) is equal to zero. Select a
number 𝑐 ◦ > 0 such that, whatever 𝑗, 1 ≤ 𝑗 ≤ 𝜈, the convex cone
n o
Γ ( 𝑗) = 𝑦 ∈ R𝑛 ; ∀𝜉 ∈ ℭ ( 𝑗) , 𝑦 · 𝜉 > 𝑐 ◦ |𝑦| |𝜉 |

is not empty. Let Ω′ be an arbitrary open subset of R𝑛 whose closure is compact and
contained in Ω . To every distribution 𝑢 in Ω there
exist an integer
𝑁 ≥ 0, a number
𝜏 > 0 and functions 𝐹 ∈ C 𝜔 (R𝑛 ), 𝑓 𝑗 ∈ O 𝜏( 𝑁 ) Ω′ + 𝑖Γ ( 𝑗) (1 ≤ 𝑗 ≤ 𝜈) such that
𝑢 = 𝐹 + 𝜈𝑗=1 𝑏 Ω′ 𝑓 𝑗 in Ω′. Moreover, 𝐹 can be chosen to be the restriction of an
Í
entire function of exponential type in C𝑛 .
Proof We can assume supp 𝑢 to be compact. The Paley–Wiener–Schwartz Theorem
2.1.2 tells us that the Fourier transform b
𝑢 extends to C𝑛 as an entire
function
of expo- −𝑚
𝑣 (𝜉) = 1 + |𝜉 | 2
nential type in C𝑛 and that there is an 𝑚 ∈ Z+ such that b 𝑢 (𝜉) ∈
b
𝐿1 (R𝑛 ). The Fourier inversion formula implies 𝑢 = (1 − Δ 𝑥 ) 𝑣 with 𝑚
50 3 Analytic Tools in Distribution Theory
∫
𝑣 (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 b
𝑣 (𝜉) d𝜉.
R𝑛

By a classical theorem of Lebesgue (a direct consequence of the density of Cc∞ in

𝐿 1 ) 𝑣 is continuous in R𝑛 and tends to zero as |𝑥| → +∞. Suppose we have proved
that there exist functions 𝐺 ∈ O (C𝑛 ) of exponential type and 𝑔 𝑗 ∈ O 𝜏(0) Ω′ + 𝑖Γ ( 𝑗)
(𝜏 > 0, 1 ≤ 𝑗 ≤ 𝜈) such that 𝑣 = 𝐺 + 𝜈𝑗=1 𝑏 Ω′ ,Γ 𝑔 𝑗 in Ω′. Then we would derive
Í

𝜈 𝑛
!𝑚
∑︁ ∑︁
𝑢 = (1 − Δ 𝑥 ) 𝐺 + 𝑚
𝑏 Ω′ ,Γ 1 − 𝜕𝑧2ℓ 𝑔𝑗
𝑗=1 ℓ=1

in ′ ; (1 − Δ ) 𝑚 𝐺 ∈ O (C𝑛 ) is of exponential type and Lemma 3.3.11 tells us that

ΩÍ 𝑧
𝑚
𝜕2 (2𝑚)
𝑛
1 − ℓ=1 𝜕𝑧 2 𝑔 𝑗 ∈ O 𝜏/2 Ω′ + 𝑖Γ ( 𝑗) , thereby implying the general result.
ℓ
Now, our hypotheses entail
∫ 𝜈 ∫
∑︁
𝑛 𝑖 𝑥· 𝜉
(2𝜋) 𝑣 (𝑥) = e 𝑣 (𝜉) d𝜉 +
b e𝑖 𝑥· 𝜉 b
𝑣 (𝜉) d𝜉.
| 𝜉 | ≤1 𝑗=1 𝜉 ∈ℭ ( 𝑗) , | 𝜉 | ≥1

The first integral on the right can be directly extended as an entire function of
exponential type in C𝑛 (the “easy part” of the Paley–Wiener theorem). Define, for
𝑧 ∈ R𝑛 + 𝑖Γ ( 𝑗) , ∫
𝑔 𝑗 (𝑧) = (2𝜋) −𝑛 e𝑖𝑧· 𝜉 b
𝑣 (𝜉) d𝜉.
𝜉 ∈ℭ ( 𝑗) , | 𝜉 | ≥1

Since e𝑖𝑧· 𝜉 ≤ e−𝑐◦ |𝑦 | | 𝜉 | for all 𝑧 = 𝑥 + 𝑖𝑦 ∈ R𝑛 + 𝑖Γ ( 𝑗) , 𝜉 ∈ ℭ ( 𝑗) , it follows that, if

𝑘 ∈ Z+ and 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑘,
∫
|𝑦| 𝑘 D𝑧𝛼 𝑔 𝑗 (𝑧) = (2𝜋) −𝑛 |𝑦| 𝑘 𝜉 𝛼 e𝑖𝑧· 𝜉 b
𝑣 (𝜉) d𝜉
𝜉 ∈ℭ ( 𝑗) , | 𝜉 | ≥1

is a holomorphic function in R𝑛 + 𝑖Γ ( 𝑗) such that

∫
|𝑦| 𝑘 𝜕𝑧𝛼 𝑔 𝑗 (𝑧) ≤ (2𝜋) −𝑛 (|𝑦| |𝜉 |) 𝑘 e−𝑐◦ |𝑦 | | 𝜉 | |b
𝑣 (𝜉)| d𝜉
𝜉 ∈ℭ ( 𝑗) , | 𝜉 | ≥1
∫
≤ (2𝜋) −𝑛 𝑐−𝑘
◦ 𝑘! |b
𝑣 (𝜉)| d𝜉.
R𝑛

We conclude that 𝑔 𝑗 ∈ O 𝜏(0) R𝑛 + 𝑖Γ ( 𝑗) with 𝜏 = 𝑐 ◦ . □

𝑏 Ω′ 𝑓 𝑗 (in Ω′) is not unique.

Í𝜈
Needless to say, the representation 𝑢 = 𝐹 + 𝑗=1
3.4 The FBI Transform of Distributions. An Introduction 51

3.4 The FBI Transform of Distributions. An Introduction

3.4.1 The classical FBI transform and its inversion

Following [Brós-Iagolnitzer, 1973], [Brós-Iagolnitzer, 1975] (see also [Iagolnitzer-

Stapp, 1972]) we introduce the following transform of an arbitrary function 𝑓 ∈
𝐿 1 (R𝑛 ): ∫
2
𭟋 𝑓 (𝑥, 𝜉) = e𝑖 𝜉 · ( 𝑥−𝑦)− | 𝜉 | | 𝑥−𝑦 | 𝑓 (𝑦) d𝑦, (3.4.1)
R𝑛
where (𝑥, 𝜉) ∈ R𝑛 × R𝑛 ; this can be rewritten as the convolution
2
𭟋 𝑓 (𝑥, 𝜉) = 𝑓 (𝑥) ∗ e𝑖 𝜉 ·𝑥− | 𝜉 | | 𝑥 | .

The Lebesgue Dominated Convergence Theorem implies that 𭟋 𝑓 is a bounded and

continuous function of (𝑥, 𝜉) in R2𝑛 . The standard estimates for convolution further
give ∫
1 1
|𝜉 | 2 𝑛 |𭟋 𝑓 (𝑥, 𝜉)| d𝑥 ≤ 𝜋 2 𝑛 ∥ 𝑓 ∥ 𝐿 1 (R𝑛 ) . (3.4.2)
R𝑛
We shall exploit the formula
2
2
D 𝑥𝛼 D 𝜉 e𝑖 𝜉 · ( 𝑥−𝑦)−| 𝜉 | |𝑥−𝑦 | = 𝑃 𝛼,𝛽 (𝑥 − 𝑦, 𝜉) e𝑖 𝜉 · ( 𝑥−𝑦)− | 𝜉 | | 𝑥−𝑦 | ,
𝛽
(3.4.3)

where 𝑃 𝛼,𝛽 (𝑥, 𝜉) is a polynomial with respect to 𝑥 of degree |𝛼 + 𝛽| whose coeffi-

cients are smooth functions of 𝜉 in R𝑛 \ {0} satisfying estimates of the kind | 𝑓 (𝜉)| ≲
1 2
|𝜉 | | 𝛼+𝛽 | + |𝜉 | 1−|𝛽 | . Then, by making use of the estimate |𝑥| 𝑚 ≤ 𝐶𝑚 |𝜉 | − 2 𝑚 e− | 𝜉 | | 𝑥 |
we conclude immediately that 𭟋 𝑓 (𝑥, 𝜉) is a C ∞ function of (𝑥, 𝜉) ∈ R𝑛 × (R𝑛 \ {0}).
And from the fact that 𝑃 𝛼,0 (𝜉, 𝑥 − 𝑦) is a polynomial with respect to (𝜉, |𝜉 | (𝑥 − 𝑦))
we conclude that 𭟋 𝑓 (𝑥, 𝜉) is a C ∞ function of 𝑥 ∈ R𝑛 for all 𝜉 ∈ R𝑛 . The growth as
|𝜉 | → +∞ of each partial derivative D 𝑥𝛼 𭟋 𝑓 is at most polynomial.
The transformation 𭟋 can be extended to an arbitrary distribution 𝑢 ∈ D 𝐿′ 1 (R𝑛 ),
by which we mean a finite sum 𝑢 = 𝛼 D 𝛼 𝑓 𝛼 of derivatives of 𝐿 1 functions 𝑓 𝛼 :
Í

2
𭟋𝑢(𝑥, 𝜉) = 𝑢 (𝑥) ∗ e𝑖 𝜉 ·𝑥−| 𝜉 | |𝑥 | . (3.4.4)

Since partial derivatives commute with convolution,

∑︁
𭟋𝑢(𝑥, 𝜉) = D 𝑥𝛼 𭟋 𝑓 𝛼 (𝑥, 𝜉) (3.4.5)
𝛼

is a C ∞ function of (𝑥, 𝜉) in R𝑛 × (R𝑛 \ {0}); 𭟋𝑢 extends continuously to R𝑛 × R𝑛 . It

is by now traditional to refer to (3.4.4) (as well as to its variants, see Remark 3.4.4
below and the sequel) as the Fourier–Brós–Iagolnitzer (in short, FBI) transform
of the distribution 𝑢.
52 3 Analytic Tools in Distribution Theory

Lemma 3.4.1 Let 𝑧 = (𝑧 1 , ..., 𝑧 𝑛 ) ∈ C𝑛 be arbitrary. The Jacobian determinant of

the map R𝑛 \ {0} ∋ 𝜉 ↦→ 𝜁 = 𝜉 + |𝜉 | 𝑧 ∈ C𝑛 is equal to 1 + |𝜉 | −1 𝑧 · 𝜉.
𝜉
Proof We write 𝜉 = 𝜌𝜔, 𝜌 = |𝜉 |, 𝜔 𝑗 = | 𝜉𝑗 | ( 𝑗 = 1, ..., 𝑛). The fact that d𝜔1 ∧ · · · ∧
d𝜔 𝑛 = 0 implies
Û𝑛
d𝜁1 ∧ · · · ∧ d𝜁 𝑛 = 𝜌d𝜔 𝑗 + 𝜔 𝑗 + 𝑧 𝑗 d𝜌
𝑗=1
𝑛
∑︁
= 𝜌 𝑛−1 d𝜌 ∧ (−1) 𝑗−1 𝜔 𝑗 + 𝑧 𝑗 Λ 𝑗 ,
𝑗=1

where we have used the notation

Û
Λ𝑗 = d𝜔 𝑘 .
1≤𝑘 ≤𝑛, 𝑘≠ 𝑗

We reason near the point 𝜔◦ = (0, ..., 0, 1), where

𝑛−1
!
∑︁
d𝜔 𝑛 = −𝜔−1
𝑛 𝜔ℓ d𝜔ℓ .
ℓ=1

We get

𝑛−1 𝑛−1
∑︁ ©∑︁ 𝜔𝑗 ª
(−1) 𝑗−1 𝜔 𝑗 + 𝑧 𝑗 Λ 𝑗 = (−1) 𝑛−1 ­ 𝜔𝑗 + 𝑧𝑗 ® Λ𝑛
𝑗=1 𝑗=1
𝜔𝑛
« ¬
𝑛−1
∑︁
= (−1) 𝑛−1 𝜔−1 2
𝑛 ­1 − 𝜔 𝑛 + 𝑧 𝑗 𝜔 𝑗 ® Λ𝑛 ,
© ª

« 𝑗=1 ¬
whence
𝑛
∑︁ 𝑛
∑︁
𝜔−1
𝑗−1 𝑛−1
(−1) 𝜔 𝑗 + 𝑧 𝑗 Λ 𝑗 = (−1) 𝑛 ­1 + 𝑧 𝑗 𝜔 𝑗 ® Λ𝑛 .
© ª
𝑗=1 « 𝑗=1 ¬
Thus
∑︁ 𝑛
d𝜁1 ∧ · · · ∧ d𝜁 𝑛 = (−1) 𝑛−1 ­1 + 𝑧 𝑗 𝜔 𝑗 ® 𝜔−1 𝑛−1
d𝜌 ∧ Λ𝑛 .
© ª
𝑛 𝜌
« 𝑗=1 ¬
It suffices then to note that

d𝜉1 ∧ · · · ∧ d𝜉 𝑛 = d𝜁1 ∧ · · · ∧ d𝜁 𝑛 | 𝑧=0

= (−1) 𝑛−1 𝜔−1
𝑛 𝜌
𝑛−1
d𝜌 ∧ Λ𝑛 ,

which yields
3.4 The FBI Transform of Distributions. An Introduction 53

∑︁ 𝑛
d𝜁1 ∧ · · · ∧ d𝜁 𝑛 = ­1 + 𝑧 𝑗 𝜔 𝑗 ® d𝜉1 ∧ · · · ∧ d𝜉 𝑛 . □
© ª

« 𝑗=1 ¬
We now introduce the Jacobian determinant of the map 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜅 |𝜉 | 𝑧
(𝜅 > 0 an arbitrary constant) which is, by Lemma 3.4.1,

Δ 𝜅 (𝑧, 𝜉) = 1 + 𝑖𝜅 |𝜉 | −1 (𝑧 · 𝜉) = Δ1 (𝜅𝑧, 𝜉) . (3.4.6)

Lemma 3.4.2 We have, for every 𝑧 ∈ C𝑛 and every 𝑛 × 𝑛 matrix 𝑩 with complex
entries, ∫
′ 2 1
e− | 𝑥 | Δ 𝜅 (𝑧 − 𝑩𝑥 ′, 𝜉) d𝑥 ′ = 𝜋 2 𝑛 Δ 𝜅 (𝑧, 𝜉) .
R𝑛
Proof Indeed,
∫
1 ′ |2
𝜋− 2 𝑛 e−| 𝑥 1 + 𝑖𝜅 |𝜉 | −1 (𝑧 − 𝑩𝑥 ′) · 𝜉 d𝑥 ′
R𝑛
= 1 + 𝑖𝜅 |𝜉 | −1 𝑧 · 𝜉
∫ 2 1
since R𝑛
e−| 𝑥 | d𝑥 = 𝜋 2 𝑛 and
∫
2
e− | 𝑥 | 𝑃 (𝑥) d𝑥 = 0
R𝑛

whatever the polynomial 𝑃 such that 𝑃 (−𝑥) = −𝑃 (𝑥). □

The next result provides a left inverse for the transform 𭟋.
Theorem 3.4.3 If 𝑢 ∈ D ′ (R𝑛 ) is a finite sum of derivatives of 𝐿 1 functions then
3
2𝑛 𝜋 2 𝑛 𝑢 (𝑥) (3.4.7)
∫
′ ′ 2 1
= e𝑖 𝜉 · ( 𝑥−𝑥 )−| 𝜉 | |𝑥−𝑥 | (𭟋𝑢) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉.
R2𝑛

Proof Proving (3.4.7) for 𝑢 ∈ 𝐿 1 (R𝑛 ) will also prove it for a finite sum 𝑢 =
Í 1
𝛼 D 𝑓 𝛼 , 𝑓 𝛼 ∈ 𝐿 (R ). Indeed, (3.4.5) implies
𝛼 𝑛

∫ !
2 1
′ ′
∑︁
e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | | 𝑥−𝑥 | 𭟋 D 𝑥𝛼′ 𝑓 𝛼 (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉
𝛼
∫ !
𝑖 𝜉 · ( 𝑥−𝑥 ′ )− | 𝜉 | |𝑥−𝑥 ′ | 2 1
∑︁
= e D 𝑥𝛼′ 𭟋 𝑓 𝛼 (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉
𝛼
∫
1
𝑖 𝜉 · ( 𝑥−𝑥 ′ )−| 𝜉 | |𝑥−𝑥 ′ | 2
∑︁
= D 𝑥𝛼 e (𭟋 𝑓 𝛼 ) (𝑥 ′, 𝜉) |Δ1 (𝑥 − 𝑥 ′, 𝜉) 𝜉 | 2 𝑛 d𝑥 ′d𝜉
𝛼 R2𝑛
3
∑︁
= 2𝑛 𝜋 2 𝑛 D 𝑥𝛼 𝑓 𝛼 .
𝛼
54 3 Analytic Tools in Distribution Theory

Thus suppose 𝑢 ∈ 𝐿 1 (R𝑛 ); we have

∫
′ ′ 2 1
e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | | 𝑥−𝑥 | 𭟋𝑢(𝑥 ′, 𝜉) |2𝜉 | 2 𝑛 Δ1 (𝑥 − 𝑥 ′, 𝜉) d𝑥 ′d𝜉
R2𝑛
∫
′ 2 ′ 2 1
= e𝑖 𝜉 · ( 𝑥−𝑦)−| 𝜉 | |𝑥−𝑥 | − | 𝜉 | |𝑦−𝑥 | 𝑢(𝑦) |2𝜉 | 2 𝑛 Δ1 (𝑥 − 𝑥 ′, 𝜉) d𝑥 ′d𝑦d𝜉
R3𝑛
∫ ∫
2
𝑖 𝜉 · ( 𝑥−𝑦)− 21 | 𝜉 | |𝑥−𝑦 | 2 −2| 𝜉 | | 𝑥 ′ − 𝑥+𝑦 | 1
𝑛 ′ ′
= e e 2 |2𝜉 | Δ1 (𝑥 − 𝑥 , 𝜉) d𝑥 𝑢(𝑦)d𝑦d𝜉
2
R2𝑛 R𝑛
∫ ∫ √︄ ! !
1 2 ′ 2 2
= e𝑖 𝜉 · ( 𝑥−𝑦)− 2 | 𝜉 | |𝑥−𝑦 | e−| 𝑥 | Δ 1 𝑥 − 𝑦 − 𝑥 ′, 𝜉 d𝑥 ′ 𝑢(𝑦)d𝑦d𝜉.
R2𝑛 R𝑛 2 |𝜉 |

By Lemma 3.4.2,
∫ √︄ !
−| 𝑥 ′ | 2 2 ′ 1
e Δ1 𝑥 − 𝑦 − 𝑥 , 𝜉 d𝑥 ′ = 𝜋 2 𝑛 Δ 1 (𝑥 − 𝑦, 𝜉)
R𝑛 2 |𝜉 | 2

whence
∫
′ 1 ′ 2 1
e𝑖 𝜉 · ( 𝑥−𝑥 )− 2 | 𝜉 | | 𝑥−𝑥 | 𭟋𝑢(𝑥 ′, 𝜉)Δ 1 (𝑥 − 𝑦, 𝜉) |2𝜉 | 2 𝑛 d𝑥 ′d𝜉
2
R2𝑛
∫
1 1 2
= 𝜋2𝑛 e𝑖 𝜉 · ( 𝑥−𝑦)− 2 | 𝜉 | |𝑥−𝑦 | Δ 1 (𝑥 − 𝑦, 𝜉) 𝑢(𝑦)d𝑦d𝜉
2𝑛 2
∫R ∫
1
= 𝜋2𝑛 e𝑖𝜁 · ( 𝑥−𝑦) 𝑢(𝑦)d𝑦d𝜁
R𝑛 Λ

After pulling back the domain of 𝜁-integration from Λ to R𝑛 the Fourier inversion
formula produces the desired result. □

Remark 3.4.4 If we are willing to modify the definition of the FBI transform of a
compactly supported distribution from (3.4.4) to
∫
2
𭟋 𝜅 𝑓 (𝑥, 𝜉) = e𝑖 𝜉 · ( 𝑥−𝑦)−𝜅 | 𝜉 | |𝑥−𝑦 | 𝑓 (𝑦) Δ 𝜅 (𝑥 − 𝑦, 𝜉) d𝑦 (3.4.8)
R𝑛
2
= 𝑓 (𝑥) ∗ e𝑖 𝜉 ·𝑥−𝜅 | 𝜉 | |𝑥 | Δ 𝜅 (𝑥, 𝜉)

(𝜅 > 0 arbitrary) then 𭟋 𝜅 admits an inversion formula even simpler than (3.4.7).
Assume that 𝑓 ∈ 𝐿 1 (R𝑛 ) has compact support, i.e., 𝑓 ∈ 𝐿 c1 (R𝑛 ); the Fourier
inversion formula reads
∫
2
𝑓 (𝑥) = (2𝜋) −𝑛 lim e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜀 | 𝜉 | 𝑓 (𝑦) d𝑦d𝜉. (3.4.9)
𝜀↘0 R2𝑛

It suffices to reason under the hypothesis that 𝜅 diam supp 𝑓 < 21 since an arbitrary
function belonging to 𝐿 c1 (R𝑛 ) can be decomposed as a finite sum of functions
𝑓 ∈ 𝐿 c1 (R𝑛 ) with that property. If we apply Stokes’ Theorem and deform the
3.4 The FBI Transform of Distributions. An Introduction 55

domain of 𝜉-integration in (3.4.9) from R𝑛 to the image of R𝑛 under the map

𝜉 ↦→ 𝜉 + 𝑖𝜅 |𝜉 | (𝑥 − 𝑦) we get (cf. Remark 1.1.1)

𝑓 (𝑥) =
∫
2
−𝜀 ⟨ 𝜉 +𝑖𝜅 | 𝜉 | ( 𝑥−𝑦) ⟩ 2
(2𝜋) −𝑛 lim e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜅 | 𝜉 | | 𝑥−𝑦 | 𝑓 (𝑦) Δ 𝜅 (𝑥 − 𝑦, 𝜉) d𝑦d𝜉.
𝜀↘0 R2𝑛

If we agree to carry out the integration with respect to 𝑦 first then the Lebesgue
Dominated Convergence Theorem yields
∫
2
−𝑛
𝑓 (𝑥) = (2𝜋) e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜅 | 𝜉 | | 𝑥−𝑦 | 𝑓 (𝑦) Δ 𝜅 (𝑥 − 𝑦, 𝜉) d𝑦d𝜉,
R2𝑛

equivalent to ∫
𝑓 (𝑥) = (2𝜋) −𝑛 𭟋 𝜅 𝑓 (𝑥, 𝜉)d𝜉. (3.4.10)
R𝑛
Using the analogue of (3.4.5) for 𭟋 𝜅 makes it clear that (3.4.10) is valid for all
𝑓 ∈ E ′ (R𝑛 ).

3.4.2 The FBI transform and analyticity

In this section we show how to characterize the real-analyticity of a distribution by

the exponential decay of its FBI transform (3.4.4) as |𝜉 | ↗ +∞. This palliates the
unsatisfactory aspects of the Fourier transform as remarked following Proposition
2.1.3. We begin by proving a special case of (part of) the claim. As before Ω will be
an open subset of R𝑛 .

Lemma 3.4.5 If 𝑢 ∈ E ′ (Ω) and 𝑢 ≡ 0 in a neighborhood of 𝑥 ◦ then the following is

true:
(EXP DECAY 1) There is a neighborhood 𝑈 ⊂ Ω of 𝑥 ◦ such that

∃𝑐 > 0, ∀ (𝑥, 𝜉) ∈ 𝑈 × R𝑛 , |𭟋𝑢 (𝑥, 𝜉)| ≲ e−𝑐 | 𝜉 | . (3.4.11)

Proof If 𝑢 ≡ 0 in a neighborhood of 𝑥 ◦ then 𝑢 admits a representation (1.3.5) with

functions 𝑓 𝛼 ∈ 𝐿 c1 (R𝑛 ) which vanish identically in a neighborhood of 𝑥 ◦ : it suffices to
select 𝜀 < dist (𝑥 ◦ , supp 𝑢) in the construction leading to (1.3.5). As a consequence,
it suffices to prove the claim for 𝑢 = D 𝑥𝛼 𝑓 with 𝛼 ∈ Z+𝑛 and 𝑓 ∈ 𝐿 1 (R𝑛 ), 𝑥 ◦ ∉ supp 𝑓 .
In this case, going back to (3.4.1) we get
∫
2
|𝛼|
𭟋𝑢 (𝑥, 𝜉) = (−1) 𝑓 (𝑦) D 𝑦𝛼 e𝑖 𝜉 · ( 𝑥−𝑦)−𝜅 | 𝜉 | |𝑥−𝑦 | d𝑦
R𝑛
∫
2
= 𝑓 (𝑦) e𝑖 𝜉 · ( 𝑥−𝑦)−𝜅 | 𝜉 | | 𝑥−𝑦 | 𝑃 𝛼 (|𝜉 | (𝑥 − 𝑦) , 𝜉) d𝑦,
R𝑛
56 3 Analytic Tools in Distribution Theory

where 𝑃 𝛼 (𝑥, 𝜉) is a polynomial with respect to (𝑥, 𝜉) [cf. (3.4.3)]. We can use the
1
𝑚 1 1 2
inequalities |𝜉 | 2 |𝑥 − 𝑦| ≤ 𝐶𝑚 𝜅 − 2 𝑚 e− 2 𝜅 | 𝜉 | | 𝑥−𝑦 | (𝑚 ∈ Z+ , 𝜅 > 0) to obtain
∫
1 2
|𭟋𝑢 (𝑥, 𝜉)| ≲ (1 + |𝜉 |) | 𝛼 | | 𝑓 (𝑦)| e− 2 𝜅 | 𝜉 | | 𝑥−𝑦 | d𝑦.
R𝑛

By selecting 𝑐 > 0 such that 41 𝜅 dist2 (𝑥 ◦ , supp 𝑢) ≥ 3𝑐 we conclude that

1 ◦ 2
2 𝜅 |𝑥 − 𝑥 | ≤ 𝑐 together with 𝑦 ∈ supp 𝑢 implies

1 1 1
𝜅 |𝑥 − 𝑦| 2 ≥ 𝜅 |𝑥 ◦ − 𝑦| 2 − 𝜅 |𝑥 − 𝑥 ◦ | 2 ≥ 2𝑐,
2 4 2
whence ∫
|𭟋𝑢 (𝑥, 𝜉)| ≲ (1 + |𝜉 |) | 𝛼 | e−2𝑐 | 𝜉 | | 𝑓 (𝑦)| d𝑦
R𝑛
and thereby (3.4.11). □
Corollary 3.4.6 Suppose 𝑢 ∈ E ′ (Ω) has Property (EXP DECAY 1) and 𝑣 ∈
D ′ (R𝑛 ) is equal to 𝑢 in a neighborhood of 𝑥 ◦ in Ω. Then 𝜒𝑣 also has Property
(EXP DECAY 1) whatever 𝜒 ∈ Cc∞ (Ω) such that 𝜒 ≡ 1 in a neighborhood of 𝑥 ◦ .
We can now prove the announced characterization of (local) real-analyticity.
Theorem 3.4.7 Condition (EXP DECAY 1) is necessary and sufficient for 𝑢 ∈
D ′ (Ω) to be real-analytic in some neighborhood of 𝑥 ◦ ∈ Ω.
Proof I. Necessity of the condition. Suppose 𝑢 ∈ C 𝜔 (Ω′) with Ω′ ⊂ Ω open,
𝑥 ◦ ∈ Ω′. Let 𝜒 ∈ Cc∞ (Ω) be a function equal to 1 in a suitable neighborhood of
𝑥 ◦ . By Lemma 3.4.1 we know that (1 − 𝜒) 𝑢 satisfies Condition (EXP DECAY 1).
It suffices therefore to prove that the same is true of 𝜒𝑢. For the sake of simplicity
we take 𝑥 ◦ = 0. Thus we may as well assume that 𝑢 ∈ Cc∞ (Ω) and that there is
a number 𝑟 ◦ , 0 < 𝑟 ◦ < 1, such that 𝑢 extends as a holomorphic function in the
ball {𝑧 ∈ C𝑛 ; |𝑧| < 𝑟 ◦ }. Thanks to the Cauchy Integral Theorem we can deform
the domain of integration in (3.4.1) from R𝑛 to the contour Λ 𝜉 in C𝑛 described by
𝑧 = 𝑦 + 𝑖𝜑 (𝑦) | 𝜉𝜉 | , where 𝜑 ∈ C ∞ (R𝑛 ), 𝜑 (𝑦) = 0 if |𝑦| > 𝑟, 0 < 𝜑 (𝑦) ≤ 𝑟 if
|𝑦| < 𝑟, 0 < 𝑟 < 𝑟 ◦ . Provided 𝑟 > 0 is sufficiently small there is a 𝑐 > 0 such that, if
|𝑥| < 𝑟/2, 0 ≠ 𝜉 ∈ R𝑛 and 𝑧 ∈ Λ 𝜉 then

𝑛
© 𝜉 ∑︁ 2ª
Re ­𝑖 · (𝑥 − 𝑧) + 𝑥 𝑗 − 𝑧 𝑗 ® = 𝜑 (𝑦) + |𝑥 − 𝑦| 2 − 𝜑2 (𝑦) (3.4.12)
|𝜉 | 𝑗=1
« ¬
1 2
≥ 𝜑 (𝑦) + |𝑥 − 𝑦| ≥ 𝑐.
2
We conclude that if |𝑥| < 𝑟/2 then
∫
−𝑐 | 𝜉 |
|𭟋𝑢 (𝑥, 𝜉)| ≤ e |𝑢 (𝑧)| |d𝑧| (3.4.13)
Λ𝜉
3.5 The Analytic Wave-Front Set of a Distribution 57

for all 𝜉 ∈ R𝑛 .
II. Sufficiency of the condition. Let 𝑣 be as in (EXP DECAY 1). If 𝜓 ∈ Cc∞ (𝑈),
𝜓 ≡ 1 in a neighborhood 𝑉 of 𝑥 ◦ , then each of the two integrals
∫
′ ′ 2 1
e𝑖 𝜉 · ( 𝑥−𝑥 )−𝜅 | 𝜉 | | 𝑥−𝑥 | 𝜓 (𝑥 ′) (𭟋𝑣) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉,
2𝑛
∫R
′ ′ 2 1
e𝑖 𝜉 · ( 𝑥−𝑥 )−𝜅 | 𝜉 | | 𝑥−𝑥 | (1 − 𝜓 (𝑥 ′)) (𭟋𝑣) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉,
R2𝑛

can be extended holomorphically to a suitably small neighborhood of 𝑥 ◦ in C𝑛 : the

first one because 𝜓 (𝑥 ′) ≠ 0 =⇒ |(𭟋𝑣) (𝑥 ′, 𝜉)| ≲ e−𝑐 | 𝜉 | ; the second one because
𝜓 (𝑥 ′) ≠ 1 and |𝑥 − 𝑥 ◦ | sufficiently small ensure that
′ |2 ′|𝜉 |
∃𝑐 ′ > 0, e−𝜅 | 𝜉 | | 𝑥−𝑥 |(𭟋𝑣) (𝑥 ′, 𝜉)| ≲ e−𝑐 .

The inversion formula (3.4.7) implies directly that 𝑣 and therefore 𝑢 can be extended
holomorphically to a neighborhood of 𝑥 ◦ in C𝑛 . □

Remark 3.4.8 Inspection of the proofs of Lemma 3.4.5 and Theorem 3.4.3 shows
that if 𝑢 extends as a holomorphic function 𝑢˜ in the ball {𝑧 ∈ C𝑛 ; |𝑧| < 𝑟 ◦ } then
(3.4.11) will hold for 𝑢˜ with the same constant 𝑐 > 0.

3.5 The Analytic Wave-Front Set of a Distribution

3.5.1 The analytic wave-front set. Microlocal analytic hypoellipticity

In the real-analytic context the next definition is central to the whole idea of lifting
the analysis of distributions to phase-space, in other words, of microlocalization.
We introduce the concept of analytic wave-front set of a distribution (cf. Definition
2.1.5) through the FBI transform. Later (in Ch. 7), we shall introduce the analytic
wave-front set of a hyperfunction 𝑢 (called its essential singular support in [Sato-
Kawai-Kashiwara, 1973] and numerous other texts) using the representation of 𝑢 as
a sum of boundary values (Definition 7.4.7).
We use standard terminology: a set U ⊂ R𝑛 × (R𝑛 \ {0}) is said to be conic when
(𝑥, 𝜉) ∈ U =⇒ ∀𝜆 > 0, (𝑥, 𝜆𝜉) ∈ U. The transform 𭟋 𝜅 is defined in (3.4.8).

Definition 3.5.1 We say that 𝑣 ∈ E ′ (R𝑛 ) is microanalytic at a point (𝑥 ◦ , 𝜉 ◦ ) ∈

R𝑛 × (R𝑛 \ {0}) if there is a conic neighborhood U of (𝑥 ◦ , 𝜉 ◦ ) in R𝑛 × (R𝑛 \ {0})
such that |𭟋 𝜅 𝑣 (𝑥, 𝜉)| ≲ e−𝑐 | 𝜉 | for some positive numbers 𝜅, 𝑐, and all (𝑥, 𝜉) ∈ U.
Let Ω ⊂ R𝑛 be an open set. We say that a distribution 𝑢 ∈ D ′ (Ω) is microanalytic
at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0}) if there is a 𝑣 ∈ E ′ (R𝑛 ) equal to 𝑢 in some neigh-
borhood of 𝑥 ◦ and microanalytic at (𝑥 ◦ , 𝜉 ◦ ). The (closed) subset of Ω × (R𝑛 \ {0})
consisting of the points (𝑥 ◦ , 𝜉 ◦ ) at which 𝑢 is not microanalytic is called the analytic
wave-front set of 𝑢 and shall be denoted by 𝑊 𝐹a (𝑢).
58 3 Analytic Tools in Distribution Theory

Proposition 3.5.2 The analytic singular support (Definition 1.3.1) of a distribution

𝑢 ∈ D ′ (Ω) is the image of 𝑊 𝐹a (𝑢) under the projection Ω × (R𝑛 \ {0}) ∋ (𝑥, 𝜉) ↦→
𝑥 ∈ Ω.
Proof By Theorem 3.4.7 the real-analyticity of 𝑢 at 𝑥 ◦ ∈ Ω entails its microlocal
analyticity at every point (𝑥, 𝜉) ∈ 𝑈 × (R𝑛 \ {0}) for a suitable choice of the neighbor-
hood 𝑈 ⊂ Ω of 𝑥 ◦ . Conversely, suppose that 𝑢 is microanalytic at (𝑥 ◦ , 𝜉 ◦ ) whatever
𝜉 ◦ ∈ R𝑛 \ {0}. We can find positive constants 𝑐, 𝑀, a neighborhood 𝑉 ⊂ Ω of 𝑥 ◦ and
finitely many conic open sets U𝑖 ⊂ Ω × (R𝑛 \ {0}) (𝑖 = 1, ..., 𝑟) with the following
properties:
(1) to each 𝜉 ∈ R𝑛 \ {0} there is an index 𝑖 such that (𝑥 ◦ , 𝜉) ∈ U𝑖 ;
(2) for each 𝑖 there is a 𝑣 𝑖 ∈ E ′ (R𝑛 ) equal to 𝑢 in 𝑉 such that (3.4.11) holds for
𝑣 = 𝑣 𝑖 and all (𝑥, 𝜉) ∈ U𝑖 .
Define 𝑣 = 𝑟1 (𝑣 1 + · · · + 𝑣 𝑟 ); in 𝑉 we have 𝑣 = 𝑣 𝑖 for every 𝑖. After applying
Lemma 3.4.5 to 𝑣 − 𝑣 𝑖 and possibly contracting 𝑉 about 𝑥 ◦ we see that (3.4.11) will
hold for all (𝑥, 𝜉) ∈ U𝑖 such that 𝑥 ∈ 𝑉. We conclude that (3.4.11) will hold for all
(𝑥, 𝜉) ∈ 𝑉 × (R𝑛 \ {0}). □
The concept of analytic wave-front set enables us to microlocalize that of analytic
hypoellipticity (Definition 3.1.2). Let 𝑃 be a linear partial differential operator with
C 𝜔 coefficients in Ω.
Definition 3.5.3 The differential operator 𝑃 is said to be analytic hypoelliptic in
a conic subset U of Ω × (R𝑛 \ {0}) if U ∩ 𝑊 𝐹a (𝑢) ⊂ U ∩ 𝑊 𝐹a (𝑃𝑢) whatever
𝑢 ∈ D ′ (Ω).
The microlocal versions of Definitions 3.1.1 and 3.1.2 are obvious: 𝑃 is analytic
hypoelliptic at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0}) if there is a conic neighborhood
U ⊂ Ω× (R𝑛 \ {0}) of (𝑥 ◦ , 𝜉 ◦ ) in which 𝑃 is analytic hypoelliptic; 𝑃 is germ analytic
hypoelliptic at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0}) if to each conic neighborhood
U ⊂ Ω × (R𝑛 \ {0}) of (𝑥 ◦ , 𝜉 ◦ ) there is a conic neighborhood V ⊂ U of (𝑥 ◦ , 𝜉 ◦ )
such that
∀𝑢 ∈ D ′ (Ω) , V ∩ 𝑊 𝐹a (𝑢) ⊂ V ∩ 𝑊 𝐹a (𝑃𝑢) . (3.5.1)
The next statement is a straightforward consequence of Proposition 3.5.2:
Proposition 3.5.4 Let 𝑥 ◦ ∈ Ω. The following two properties are equivalent:
(1) 𝑃 is analytic hypoelliptic at 𝑥 ◦ ;
(2) 𝑃 is germ analytic hypoelliptic at (𝑥 ◦ , 𝜉) whatever 𝜉 ∈ R𝑛 \ {0}.

3.5.2 Analytic wave-front sets and holomorphic extension

In this subsection we relate the representation of a distribution 𝑢 as the sum of

boundary values of holomorphic functions in a wedge (Theorem 3.3.6) to the analytic
wave-front set of 𝑢. If Γ ⊂ R𝑛 \ {0} is a cone we denote by Γ◦ its polar (sometimes
3.5 The Analytic Wave-Front Set of a Distribution 59

called its dual), i.e., the set {𝜉 ∈ R𝑛 ; ∀𝑦 ∈ Γ, 𝑦 · 𝜉 ≥ 0}; Γ◦ is a closed and convex
cone in R𝑛 (with 0 ∈ Γ◦ , obviously); Γ◦ is also the polar of the convex hull of Γ,
i.e., the intersection of all the convex cones containing Γ. We also recall the notation
(3.3.1): W𝛿 (𝑈, Γ) = {𝑧 = 𝑥 + 𝑖𝑦 ∈ 𝑈 + 𝑖Γ; |𝑦| < 𝛿}.

Theorem 3.5.5 Let Ω ⊂ R𝑛 be an open set and Γ ⊂ R𝑛 \ {0} a convex open cone.
The following properties of a distribution 𝑢 ∈ D ′ (Ω) are equivalent:
(a) whatever the open set 𝑈 ⊂⊂ Ω and the open cone Γ∗ ⊂ Γ in R𝑛 \ {0} such that
Γ∗ ∩ S𝑛−1 ⊂⊂ Γ the restriction of 𝑢 to 𝑈 is the boundary value of a function
ℎ ∈ Otemp (W𝛿 (𝑈, Γ∗ ));
(b) 𝑊 𝐹a (𝑢) ⊂ Ω × Γ◦ .

Proof I. (a)=⇒(b). Let 𝜉 ◦ ∈ R𝑛 be such that 𝑦 ◦ · 𝜉 ◦ < 0 for some 𝑦 ◦ ∈ Γ. Let 𝑥 ◦ ∈ 𝑈

be arbitrary and select 𝜑 ∈ Cc∞ (𝑈), 𝜑 (𝑥) = 1 if |𝑥 − 𝑥 ◦ | < 𝜌, 𝜌 > 0. By Lemma
3.4.5 it suffices to show that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝜑𝑢). We have
∫
′ ′ 2
𭟋 𝜅 (𝜑𝑢) (𝑥, 𝜉) = e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | | 𝑥−𝑥 | 𝜑 (𝑥 ′) 𝑢 (𝑥 ′) d𝑥 ′
∫
′ ′ 2
= lim e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | | 𝑥−𝑥 | 𝜑 (𝑥 ′) ℎ (𝑥 ′ + 𝑖𝑡𝑦 ◦ ) d𝑥 ′.
0<𝑡→0 R𝑛

We introduce an additional cutoff function 𝜒 ∈ C ∞ (𝑈), 0 ≤ 𝜒 (𝑥) ≤ 1 for all 𝑥,

𝜒 (𝑥) = 0 if |𝑥 − 𝑥 ◦ | > 43 𝜌, 𝜒 (𝑥) = 1 if |𝑥 − 𝑥 ◦ | < 12 𝜌. We deform the domain of
integration
√ in the last integral from R𝑛 to the image Λ of R𝑛 under the map 𝑥 ′ ↦→ 𝑧 ′ =
𝑥 + −1𝑎 𝜒 (𝑥 ′) 𝑦 ◦ , 0 < 𝑎 |𝑦 ◦ | < 𝛿. We observe that if 𝑧 ′ ∈ Λ and | 𝜉𝜉 | · 𝑦 ◦ < −2𝑎 |𝑦 ◦ | 2
′

then

𝜉
Re 𝑖 · (𝑥 − 𝑧 ′) − (𝑥 − 𝑧 ′) 2
|𝜉 |

𝜉
≤ 𝑎 𝜒 (𝑥 ′) · 𝑦 ◦ + 𝑎 |𝑦 ◦ | 2 − |𝑥 − 𝑥 ′ | 2
|𝜉 |
≤ −𝑎 2 𝜒 (𝑥 ′) |𝑦 ◦ | 2 − |𝑥 − 𝑥 ′ | 2 ≤ −𝑐

for some 𝑐 > 0, provided 𝑥 belongs to a sufficiently small neighborhood 𝑉 of 𝑥 ◦ . This

◦
implies |𭟋 𝜅 (𝜑𝑢) (𝑥, 𝜉)| ≲ e−𝑐 | 𝜉 | for all 𝑥 ∈ 𝑉 and all 𝜉 such that | 𝜉𝜉 | − | 𝜉𝜉 ◦ | < 𝜀
provided 𝜀 > 0 is sufficiently small.
II. (b)=⇒(a). The proof of this entailment is based on the inversion formula
(3.4.7). Let 𝑣 ∈ E ′ (Ω) be equal to 𝑢 in the open set 𝑈 ⊂⊂ Ω; let 𝑉 ⊂ 𝑈 be an open
set with closure 𝐾 ⊂ 𝑈. Consider first the function 𝑤 ◦ defined by
21 𝑛
2𝜋 3 𝑤 ◦ (𝑥)
∫ ∫
′ ′ |2 1
= e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | |𝑥−𝑥 (𭟋 𝜅 𝑣) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉.
R𝑛 𝑥 ′ ∉𝐾
60 3 Analytic Tools in Distribution Theory

[Δ1 (𝑧, 𝜉) = 1 + |𝜉 | −1 (𝑧 · 𝜉)]. Since (𭟋 𝜅 𝑣) (𝑥 ′, 𝜉) has at most polynomial growth at

infinity in R2𝑛 (cf. Subsection 3.4.1) we see directly that 𝑤 ◦ extends as a holomorphic
function in an open subset 𝑉 C of C𝑛 such that 𝑉 = 𝑉 C ∩ R𝑛 . By (3.4.7) we have
21 𝑛
2𝜋 3 (𝑣 (𝑥) − 𝑤 ◦ (𝑥))
∫ ∫
′ ′ |2 1
= e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | | 𝑥−𝑥 (𭟋 𝜅 𝑣) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉.
R𝑛 𝑥 ′ ∈𝐾

Let the open cone Γ ′ in R𝑛 \ {0} be such that Γ ′ ∩ S𝑛−1 ⊂⊂ Γ. Each 𝜉 ∗ ∈ R𝑛 \Γ ′◦

is contained in an open cone ℭ such that ℭ ∩ S𝑛−1 ⊂⊂ R𝑛 \Γ◦ and

∃𝑐 > 0, ∀ (𝑥, 𝜉) ∈ 𝐾 × ℭ, |(𭟋 𝜅 𝑣) (𝑥, 𝜉)| ≲ e−𝑐 | 𝜉 | . (3.5.2)

Since the distance between Γ ′ ∩ S𝑛−1 and 𝜕Γ ∩ S𝑛−1 is not equal to zero, the Borel–
Lebesgue Lemma allows us to select a cone ℭ such that (3.5.2) holds and

(R𝑛 \Γ ′◦ ) ∩ S𝑛−1 ⊂⊂ ℭ ∩ S𝑛−1 ⊂⊂ (R𝑛 \Γ◦ ) ∩ S𝑛−1 , (3.5.3)

independently of 𝜉 ∗ . Let us then introduce the function 𝑤 1 defined by

21 𝑛
2𝜋 3 𝑤 1 (𝑥)
∫ ∫
′ ′ |2 1
= e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | | 𝑥−𝑥 (𭟋 𝜅 𝑣) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉.
𝜉 ∈ℭ 𝑥 ′ ∈𝐾

It follows immediately from (3.5.2) that 𝑤 1 extends as a holomorphic function in an

open subset of C𝑛 containing 𝑉 ⊂ R𝑛 .
Lastly we must deal with the function 𝑤 2 defined by
21 𝑛
2𝜋 3 𝑤 2 (𝑥)
∫ ∫
′ ′ |2 1
= e𝑖 𝜉 · ( 𝑥−𝑥 )−| 𝜉 | |𝑥−𝑥 (𭟋 𝜅 𝑣) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉
𝜉 ∉ℭ 𝑥 ′ ∈𝐾

since 𝑤 2 = 𝑣 −𝑤 ◦ −𝑤 1 according to (3.4.7). By (3.5.3) we have R𝑛 \ℭ ⊂Γ ′◦ , implying

𝜉 · 𝑦 > 0 for all 𝜉 ∉ ℭ and 𝑦 ∈ Γ ′. If Γ♭ is a cone in R𝑛 \ {0} such that Γ♭ ∩S𝑛−1 ⊂⊂ Γ ′,
there is a 𝑐 ◦ > 0 such that

∀ (𝑦, 𝜉) ∈ Γ♭ × (R𝑛 \ℭ) , 𝜉 · 𝑦 ≥ 𝑐 ◦ |𝜉 | |𝑦| .

It follows directly from this that, for 𝑦 ∈ Γ♭ ,

Re 𝑖𝜉 · (𝑥 + 𝑖𝑦 − 𝑥 ′) − |𝜉 | ⟨𝑥 + 𝑖𝑦 − 𝑥 ′⟩ 2 ≤ − (𝑐 ◦ − |𝑦|) |𝑦| |𝜉 | .
3.5 The Analytic Wave-Front Set of a Distribution 61

By availing ourselves once again of the tempered growth at infinity in 𝜉-space of

(𭟋 𝜅 𝑣) (𝑥 ′, 𝜉) we conclude easily that 𝑤 2 (𝑥 + 𝑖𝑦) ∈ Otemp (W𝛿 (𝑉, Γ)) provided
0 < 𝛿 < 𝑐 ◦ . This completes the proof of Theorem 3.5.5. □
The next statement is the distribution version of the Edge of the Wedge Theorem
(cf. Theorem 7.5.1).
Corollary 3.5.6 Let 𝑢 ∈ D ′ (𝑈) and let the cones Γ ( 𝑗) ( 𝑗 = 1, ..., 𝜈) be open, convex
and such that R𝑛 = Γ (1) + · · · + Γ (𝜈) (vector sum). Suppose
to each 𝑗 = 1, ..., 𝜈
that
there is a holomorphic function ℎ 𝑗 in the wedge W𝛿 𝑈, Γ ( 𝑗) such that 𝑢 = 𝑏𝑈 ℎ 𝑗
(𝛿 > 0). Then 𝑢 is real-analytic in 𝑈.
Proof On the one hand, according to Theorem 3.5.5, the hypotheses imply
𝑊 𝐹a (𝑢) ⊂ 𝑈 × Γ ( 𝑗)◦ for every 𝑗 = 1, ..., 𝜈. On the other hand they imply
Γ (1)◦ ∩ · · · ∩ Γ (𝜈)◦ = {0}, whence the claim by Proposition 3.5.2. □
Corollary 3.5.7 Let 𝑢 ∈ D ′ (𝑈) and let the cone Γ ⊂ R𝑛 \ {0} be open and convex.
Suppose there is a holomorphic function ℎ+ (resp., ℎ− ) in the wedge W𝛿 (𝑈, Γ)
[resp., W𝛿 (𝑈, −Γ)] such that, in 𝑈, 𝑢 = 𝑏𝑈 ℎ+ and 𝑢 = 𝑏𝑈 ℎ− . Then 𝑢 is real-
analytic in 𝑈.
Theorem 3.5.8 For a distribution 𝑢 ∈ D ′ (Ω) to be microanalytic at a point
(𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0}) (Definition 3.5.1) it is necessary and sufficient
that there be
◦
a neighborhood 𝑈 ⊂ Ω of 𝑥 , finitely many wedges W𝛿 𝑈, Γ ( 𝑗) ( 𝑗 = 1, ..., 𝜈) such

that 𝑦 · 𝜉 < 0 for all 𝑦 ∈ Γ ∪ · · · ∪ Γ and functions ℎ 𝑗 ∈ Otemp W𝛿 𝑈, Γ ( 𝑗)
◦ (1) (𝜈)

such that 𝑢 = 𝑏𝑈 ℎ1 + · · · + 𝑏𝑈 ℎ 𝜈 .
Proof That the condition is sufficient ensues directly from Theorem 3.5.5: for each
𝑗 = 1, ..., 𝜈, 𝑊 𝐹a 𝑏𝑈 ℎ 𝑗 ⊂ 𝑈 × Γ ( 𝑗)◦ and 𝜉 ◦ ∉ Γ ( 𝑗)◦ , implying that 𝑏𝑈 ℎ 𝑗 is
microanalytic at (𝑥 ◦ , 𝜉 ◦ ). The proof of the necessity is based on the inversion
formula (3.4.7). There is no loss of generality in dealing with a compactly sup-
ported distribution 𝑢. First of all we select a neighborhood 𝑈 of 𝑥 ◦ and an open
cone ℭ (0) ⊂ R𝑛 \ {0} such that 𝜉 ◦ ∈ ℭ (0) and |(𭟋 𝜅 𝑢) (𝑥, 𝜉)| ≲ e−𝑐◦ | 𝜉 | for some
𝑐 ◦ > 0 and all (𝑥, 𝜉) ∈ 𝑈 ×ℭ (0) . Next we select pairwise disjoint open cones ℭ ( 𝑗)
( 𝑗 = 1, ..., 𝜈) such that R𝑛 \ ℭ (0) ∪ ℭ (1) ∪ · · · ∪ ℭ (𝜈) have zero Lebesgue measure
and ℭ ( 𝑗) ∩ ℭ (𝑘) = ∅ if 0 ≤ 𝑗 < 𝑘 ≤ 𝜈. We also require that, for some 𝑐 > 0 and
each 𝑗 = 1, ..., 𝜈, the cone
n o
Γ ( 𝑗) = 𝑦 ∈ R𝑛 ; 𝑦 · 𝜉 ◦ < 0, ∀𝜉 ∈ ℭ ( 𝑗) , 𝑦 · 𝜉 > 𝑐 |𝑦| |𝜉 |
Í𝜈
be nonempty. Writing 𝑢 = 𝑗=0 𝑢 𝑗 with
21 𝑛
2𝜋 3 𝑢 𝑗 (𝑥)
∫ ∫
′ ′ |2 1
= e𝑖 𝜉 · ( 𝑥−𝑥 )− | 𝜉 | |𝑥−𝑥 (𭟋 𝜅 𝑢) (𝑥 ′, 𝜉)Δ1 (𝑥 − 𝑥 ′, 𝜉) |𝜉 | 2 𝑛 d𝑥 ′d𝜉
ℭ ( 𝑗) R𝑛
62 3 Analytic Tools in Distribution Theory

one can directly check that, for a suitably

√ small 𝛿 > 0, the following is true: 1) 𝑢 0
extends holomorphically to a set 𝑈 + −1𝔅 𝛿 where 𝔅 𝛿 = {𝑥 ∈ R𝑛 ; |𝑥| < 𝛿}; 2)
for each 𝑗 = 1, ..., 𝜈, 𝑢 𝑗 extends holomorphically to the wedge W𝛿 𝑈, Γ ( 𝑗) in the
statement. The latter follows from the fact that, if 𝑦 ∈ Γ ( 𝑗) and 𝜉 ∈ ℭ ( 𝑗) then

𝜉
Re 𝑖 · (𝑥 + 𝑖𝑦 − 𝑥 ′) − |𝑥 + 𝑖𝑦 − 𝑥 ′ | 2 ≤ −𝑐 |𝑦| − |𝑥 − 𝑥 ′ | 2 + |𝑦| 2 . □
|𝜉 |

A direct application of Theorem 3.5.8 is the following.

Proposition 3.5.9 Let 𝑃 = 𝑃 (𝑥, 𝜕𝑥 ) be a differential operator with analytic coeffi-

cients in the open set Ω ⊂ R𝑛 . For every 𝑢 ∈ D ′ (Ω) we have 𝑊 𝐹a (𝑃𝑢) ⊂ 𝑊 𝐹a (𝑢).

Proof Let (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢) ⊂ Ω × (R𝑛 \ {0}). The necessity of the condition in

Theorem 3.5.8 allows us to write 𝑢 = 𝑏𝑈 ℎ1 + · · · + 𝑏𝑈 ℎ 𝜈 in a neighborhood 𝑈 ⊂ Ω of
◦
𝑥 . For each 𝑗 = 1, ..., 𝜈, ℎ 𝑗 ∈ Otemp W𝛿 𝑈, Γ ( 𝑗) (see Subsection 8.2.2) where,
as before, 𝑦 · 𝜉 ◦ < 0 for all 𝑦 ∈ Γ ( 𝑗) . We can extend 𝑃 as a holomorphic differential
operator 𝑃 (𝑧, 𝜕𝑧 ) in an open set ΩC ⊂ C𝑛 , Ω ⊂ ΩC ∩ R𝑛 . It follows from Lemma
3.3.5 that 𝑃 (𝑧, 𝜕𝑧 ) ℎ 𝑗 ∈ Otemp W𝛿 𝑈, Γ ( 𝑗) for each 𝑗 (with 𝛿 > 0 possibly
decreased) and that

𝑃𝑢 = 𝑏𝑈 (𝑃 (𝑧, 𝜕𝑧 ) ℎ1 ) + · · · + 𝑏𝑈 (𝑃 (𝑧, 𝜕𝑧 ) ℎ 𝜈 ) in 𝑈,

which implies (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑃𝑢) by the sufficiency of the condition in Theorem

3.5.8. □

Corollary 3.5.10 The differential operator 𝑃 is analytic hypoelliptic in a conic subset

U of Ω × (R𝑛 \ {0}) (Definition 3.5.3) if U ∩ 𝑊 𝐹a (𝑃𝑢) = U ∩ 𝑊 𝐹a (𝑢) for every
𝑢 ∈ D ′ (Ω).

The following consequence of the preceding arguments is noteworthy.

Proposition 3.5.11 Given arbitrary (𝑥 ◦ , 𝜉 ◦ ) ∈ R𝑛 × S𝑛−1 there is a distribution in

R𝑛 whose analytic wave-front set is exactly the ray of points (𝑥 ◦ , 𝜌𝜉 ◦ ), 𝜌 > 0.

Proof There is no loss of generality in assuming 𝑥 ◦ = 0. We shall prove that the

stated property holds for the inverse Fourier transform 𝑢 of

𝜉
𝑢 (𝜉) = exp −
b − 𝜉◦ . (3.5.4)
|𝜉 |

We begin by proving that 𝑢 ∈ C 𝜔 (R𝑛 \ {0}). We use the fact that 𝑢 = lim 𝑢 𝜀 , where
𝜀↘0
∫
𝑖 𝑥· 𝜉 − 𝜉
|𝜉| −𝜉
◦ −𝜀 | 𝜉 | 2
𝑢 𝜀 (𝑥) = (2𝜋) −𝑛 e d𝜉.
R𝑛
3.5 The Analytic Wave-Front Set of a Distribution 63

It is evident that 𝑢 𝜀 extends to C𝑛 as an entire function. If 𝑥 stays in R𝑛 \ {0} we

can deform the domain of integration from R𝑛 to the image of R𝑛 under the map
𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜅 |𝜉 | | 𝑥𝑥 | with 0 < 𝜅 ≪ 1; we get
∫ D E
𝑖 𝑥· 𝜉 −𝜅 | 𝑥 | | 𝜉 |− 𝜁
⟨𝜁 ⟩ − 𝜉
◦ −𝜀 ⟨𝜁 ⟩ 2 𝑥
𝑢 𝜀 (𝑥) = (2𝜋) −𝑛 e Δ𝜅 , 𝜉 d𝜉,
R𝑛 |𝑥|

where ⟨𝜁⟩ is the principal branch of the square root of ⟨𝜁⟩ 2 = 𝑛𝑗=1 𝜁 2𝑗 and Δ 𝜅 | 𝑥𝑥 | , 𝜉
Í

is the Jacobian determinant of 𝜁 with respect to 𝜉. The absolute value of the integrand
is ≲ exp (−𝜅 |𝑥| |𝜉 |) and the analogue is true in the integral
∫ D E
−𝑛 𝑖𝑧· 𝜉 −𝜅 ⟨𝑧 ⟩ | 𝜉 |− ⟨𝜁𝜁 ⟩ − 𝜉 ◦ −𝜀 ⟨𝜁 ⟩ 2 𝑧
𝑢 𝜀 (𝑧) = (2𝜋) e Δ𝜅 , 𝜉 d𝜉,
R𝑛 ⟨𝑧⟩

if 𝑧 = 𝑥 + 𝑖𝑦 , |𝑦| < 𝜅 |𝑥| (with 𝜅 decreased as needed). In this union of two convex
cones 𝑢 𝜀 (𝑧) converges uniformly as 𝜀 ↘ 0 to the holomorphic function
∫ D E
𝑖𝑧· 𝜉 −𝜅 ⟨𝑧 ⟩ | 𝜉 |− ⟨𝜁𝜁 ⟩ − 𝜉 ◦ ⟨𝜁 ⟩ 𝑧
𝑢 (𝑧) = (2𝜋) −𝑛 e Δ𝜅 , 𝜉 d𝜉.
R𝑛 ⟨𝑧⟩

This proves the claim.

Next we prove that (0, 𝜉) ∉ 𝑊 𝐹a (𝑢) unless 𝜉 = 𝜌𝜉 ◦ for some 𝜌 > 0. Consider,
for arbitrary 𝑥 ∈ R𝑛 ,
∫
𝑖 𝑥· 𝜉 − | 𝑥 | 2 | 𝜉 |− | 𝜉𝜉 | − 𝜉 ◦ | 𝜉 |
𝑣 𝑑 (𝑥) = (2𝜋) −𝑛 e d𝜉.
𝜉 ◦
|𝜉| −𝜉 >𝑑

It is evident that, however small 𝑑 > 0, the function 𝑣 𝑑 extends holomorphically to

a slab {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; |𝑦| < 𝛿} provided 0 < 𝛿 ≪ 𝑑. This proves that 𝑊 𝐹a 𝑢 =
𝑊 𝐹a (𝑢 − 𝑣 𝑑 ). The argument in this subsection (cf. the proof of Theorem 3.5.8)
leads easily to the conclusion that if | 𝜉𝜉 | ≠ 𝜉 ◦ then 𝑑 ≪ | 𝜉𝜉 | − 𝜉 ◦ implies 𝜉 ∉
𝑊 𝐹a (𝑢 − 𝑣 𝑑 ). We reach the conclusion that

2𝑛 𝜉 ◦
𝑊 𝐹a 𝑢 ⊂ (𝑥, 𝜉) ∈ R ; 𝑥 = 0, =𝜉 .
|𝜉 |

Thus, by Proposition 3.5.2, in order to prove that this inclusion is an equality it suffices
to prove that 𝑢 (𝑥) is not analytic at the origin. As the reader can easily ascertain this
is a direct consequence of Proposition 3.2.5 and of the fact that b 𝑢 (𝜌𝜉 ◦ ) = 1 for all
𝜌 > 0. □
An application of the preceding results is the following:
Theorem 3.5.12 Let Ω be a domain in R𝑛 , 𝑢 ∈ D ′ (Ω) and 𝑥 ◦ ∈ supp 𝑢. Suppose
there is a neighborhood 𝑈 ⊂ Ω of 𝑥 ◦ and 𝑓 ∈ C ∞ (𝑈; R) such that 𝑓 (𝑥 ◦ ) = 0,
𝜉 ◦ · d𝑥 = d 𝑓 (𝑥 ◦ ) ≠ 0 and that supp 𝑢 ⊂ {𝑥 ∈ Ω; 𝑓 (𝑥) ≥ 0}. Then (𝑥 ◦ , ±𝜉 ◦ ) ∈
𝑊 𝐹a (𝑢).
64 3 Analytic Tools in Distribution Theory

Proof Obviously, the hypotheses in Theorem 3.5.12 remain valid after we contract 𝑈
about 𝑥 ◦ . Using a finite Taylor expansion of 𝑓 about 𝑥 ◦ we can form a real polynomial
𝑝 (𝑥) such that 𝑝 (𝑥 ◦ ) = 0, d𝑝 (𝑥 ◦ ) = d 𝑓 (𝑥 ◦ ) and 𝑝 (𝑥) < 0 =⇒ 𝑓 (𝑥) < 0 whatever
𝑥 ∈ 𝑈. We may as well assume that 𝑓 itself is a (real) polynomial. Then a C 𝜔 change
of variables allows us to assume that 𝑓 (𝑥) = 𝑥 𝑛 . In the new coordinates we take
𝑈 = 𝑈 ′ × (−𝑇, 𝑇), 𝑈 ′ ⊂ R𝑛−1 open, ∫0 ∈ 𝑈 ′, 𝑇 > 0. Let 𝜑 ∈ Cc∞ (𝑈 ′) be arbitrary and
consider the distribution 𝑢 𝜑 (𝑥 𝑛 ) = 𝜑 (𝑥 ′) 𝑢 (𝑥) d𝑥 ′ in (−𝑇, 𝑇) [𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛−1 );
the integral stands for a duality bracket]. Suppose (0, 1) ∉ 𝑊 𝐹a 𝑢 𝜑 ; the necessity
of the condition in Theorem 3.5.8 implies that 𝑢 𝜑 is the boundary value in (−𝜀, 𝜀)
of a function ℎ ∈ Otemp ((−𝜀, 𝜀) × (0, −𝛿)) (0 < 𝜀 < 𝑇, 𝛿 > 0 suitably small).
Since 𝑢 𝜑 ≡ 0 in (−𝜀, 0) Theorem 3.3.6 implies ℎ ≡ 0 in (−𝜀, 0) × (0, −𝛿) hence in
(−𝜀, 𝜀) × (0, −𝛿), which in turn implies 𝑢 𝜑 ≡ 0 in (−𝜀, 𝜀), meaning
∫
∀𝜓 ∈ Cc∞ (−𝜀, 𝜀) , 𝜑 (𝑥 ′) 𝜓 (𝑥 𝑛 ) 𝑢 (𝑥) d𝑥 ′d𝑥 𝑛 = 0.

Since the products 𝜑 (𝑥 ′) 𝜓 (𝑥 𝑛 ) span a dense subspace of Cc∞ (𝑈 ′ × (−𝜀, 𝜀)) we

reach the conclusion that 0 ∉ supp 𝑢, a contradiction. The same argument is valid
with −𝜉 ◦ substituted for 𝜉 ◦ and 𝛿 for −𝛿. □

3.5.3 Analytic wave-front sets and Ehrenpreis’ cutoffs

Given 𝜉 ◦ ∈ S𝑛−1 and 𝜌 > 0 we shall use the notation

𝜉
ℭ𝜌 (𝜉 ◦ ) = 𝜉 ∈ R𝑛 \ {0} ; − 𝜉◦ < 𝜌 . (3.5.5)
|𝜉 |

Theorem 3.5.13 For 𝑢 ∈ D ′ (Ω) to be microanalytic at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω ×

(R𝑛 \ {0}) it is necessary and sufficient that there be a pair of open sets 𝑈, 𝑉, with
𝑥 ◦ ∈ 𝑉 ⊂⊂ 𝑈 ⊂⊂ Ω, an Ehrenpreis sequence 𝜑 𝑁 (𝑁 = 1, 2, ...) relative to the pair
(𝑈, 𝑉) (Definition 3.2.2) and positive numbers 𝜌, 𝐶, such that, for all 𝑁 ∈ Z+ and
all 𝜉 ∈ ℭ𝜌 (𝜉 ◦ ),
𝑁 +1
∀𝑘 ∈ Z+ , 𝑘 ≤ 𝑁, |𝜉 | 𝑘 | 𝜑d
𝑁 𝑢 (𝜉)| ≤ 𝐶 𝑘!. (3.5.6)

The “hat” means Fourier transform.

Proof I. Necessity of the condition. Suppose (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢). Thanks to Theo-
rem 3.3.12, we may limit our attention to a distribution 𝑢 which is equal, in an open
set Ω′ ⊂ Ω, to the boundary value of a function ℎ ∈ O (𝑚) (W𝛿 (Ω′, Γ)) (𝑚 ∈ Z+ ,
𝛿 > 0;
see Definition 3.3.1). After a rotation and a contraction we may assume that
Í 21
𝑛−1 2
Γ = 𝑦 ∈ R𝑛 ; 𝑗=1 𝑦 𝑗 < 𝑦 𝑛 and that 𝑦 · 𝜉 ≤ −𝑐 ◦ 𝑦 𝑛 |𝜉 | for some 𝑐 ◦ > 0 and
all 𝑦 ∈ Γ, 𝜉 ∈ ℭ𝜌 (𝜉 ◦ ). We take 𝑉 ⊂⊂ 𝑈 ⊂⊂ Ω′ and select an arbitrary Ehrenpreis
sequence 𝜑 𝑁 (𝑁 = 1, 2, ...) relative to the pair (𝑉, 𝑈). Fixing 𝑁 we introduce the
3.5 The Analytic Wave-Front Set of a Distribution 65

function
𝑁 −1
∑︁ 1 𝜕𝑘 𝜑𝑁
e𝑁 (𝑥, 𝑦 𝑛 ) =
𝜑 (𝑖𝑦 𝑛 ) 𝑘 (𝑥) , (3.5.7)
𝑘=0
𝑘! 𝜕𝑥 𝑛𝑘
which is an element of Cc∞ (Ω′) depending on the parameter 𝑦 𝑛 ∈ R. Therefore,
taking →
−𝑣 = (0, ..., 0, 1) we can assert that

−𝑖 𝑥· 𝜉
𝑁 𝑢 (𝜉) = ⟨e
𝜑d 𝑢 (𝑥) , 𝜑 𝑁 (𝑥)⟩
∫
→
−

= lim e−𝑖 ( 𝑥+𝑖𝑡 𝑣 ) · 𝜉 ℎ 𝑥 + 𝑖𝑡→
−𝑣 𝜑e𝑁 (𝑥, 𝑡) d𝑥.
𝛿>𝑡↘0 R𝑛

The inequalities (3.2.2) show that, for a suitably large constant 𝐶 > 0 independent
of 𝑁, 𝑁
e𝑁 (𝑥, 𝑦 𝑛 )| ≤ 𝐶e |𝑦𝑛 |
|𝜑 (3.5.8)

as well as
𝜕𝜑
(𝑥, 𝑦 𝑛 ) ≤ 𝐶 𝑁 |𝑦 𝑛 | 𝑁 −1 ,
e𝑁
(3.5.9)
𝜕 𝑧¯𝑛
the latter since
𝜕𝜑 𝜕𝜑 𝜕𝜑 1 𝜕 𝑁 𝜑𝑁
(𝑖𝑦 𝑛 ) 𝑁 −1
e𝑁 e𝑁 e𝑁
2 = +𝑖 = .
𝜕 𝑧¯𝑛 𝜕𝑥 𝑛 𝜕𝑦 𝑛 (𝑁 − 1)! 𝜕𝑥 𝑛𝑁
Stokes’ Theorem implies
∫
→
−

e−𝑖 ( 𝑥+𝑖𝑡 𝑣 ) · 𝜉 ℎ 𝑥 + 𝑖𝑡→ −𝑣 𝜑 e𝑁 (𝑥, 𝑡) d𝑥
R𝑛
∫
→−

= e−𝑖 ( 𝑥+𝑖 𝛿 𝑣 ) · 𝜉 ℎ 𝑥 + 𝑖𝛿→ −𝑣 𝜑 e𝑁 (𝑥, 𝛿) d𝑥
R𝑛
∫ ∫ 𝑠= 𝛿
1 →
−
e
e−𝑖 ( 𝑥+𝑖𝑠 𝑣 ) · 𝜉 ℎ 𝑥 + 𝑖𝑠→−𝑣 𝜕 𝜑 𝑁
+ (𝑥, 𝑠) d𝑠d𝑥.
2𝑖 R𝑛 𝑠=𝑡 𝜕 𝑧¯𝑛

On the one hand, using (3.5.8) and the fact that → −𝑣 · 𝜉 ≤ −𝑐 |𝜉 | if 𝜉 ∈ ℭ (𝜉 ◦ ) we get
◦ 𝜌
∫
→−

e−𝑖 ( 𝑥+𝑖 𝛿 𝑣 ) · 𝜉 ℎ 𝑥 + 𝑖𝛿→
−𝑣 𝜑 e𝑁 (𝑥, 𝛿) d𝑥 (3.5.10)
R𝑛
∫
≤ e−𝑐◦ 𝛿 | 𝜉 | ℎ 𝑥 + 𝑖𝛿→ −𝑣 d𝑥
𝑈
𝑁
≤ 𝐶e 𝛿
(Vol 𝑈) e−𝑐◦ 𝛿 | 𝜉 | .

On the other hand, (3.5.9) yields, for 0 ≤ 𝑘 ≤ 𝑁 − 𝑚 − 1,

66 3 Analytic Tools in Distribution Theory
∫ ∫ 𝑠= 𝛿 →
−
e
e−𝑖 ( 𝑥+𝑖𝑠 𝑣 ) · 𝜉 ℎ 𝑥 + 𝑖𝑠→
−𝑣 𝜕 𝜑 𝑁
(𝑥, 𝑠) d𝑠d𝑥 (3.5.11)
R𝑛 𝑠=𝑡 𝜕 𝑧¯1
∫ ∫ 𝑠= 𝛿
≤ 𝐶 𝑁 max 𝑠 𝑘 e−𝑐◦ 𝑠 | 𝜉 | 𝑠 𝑚 ℎ 𝑥 + 𝑖𝑠→ −𝑣 d𝑠d𝑥
𝑠>0 R𝑛 𝑠=0
𝑘
𝑘
≤ 𝑀𝐶 𝑁 .
𝑐 ◦ 𝑒 |𝜉 |

The claim ensues directly from (3.5.10) and (3.5.11) by applying Stirling’s Formula.
II. Sufficiency of the condition. There is no loss of generality in assuming that
supp 𝑢 is compact. We select a C ∞ partition of unity 𝜓 𝑗 ( 𝑗 = 0, 1, ..., 𝜈) on S𝑛−1
having the following properties:
(1) for every 𝑗 = 0, 1, ..., 𝜈, 𝜓 𝑗 ≥ 0 (hence 𝜓 𝑗 ≤ 1);
(2) supp 𝜓0 ⊂ S𝑛−1 ∩ ℭ𝜌 (𝜉 ◦ );
(3) if 1 ≤ 𝑗 ≤ 𝜈, 𝜉 ◦ does not belong to the convex hull of supp 𝜓 𝑗 [and thus
◦
𝜓0 | 𝜉𝜉 ◦ | = 1].
Then, provided 𝑐 ◦ > 0 is sufficiently small, for each 𝑗 = 0, 1, ..., 𝜈 we can select
an open cone Γ ( 𝑗) ≠ ∅ such that (𝑦, 𝜉) ∈ Γ ( 𝑗) × supp 𝜓 𝑗 =⇒ 𝑦 · 𝜉 ≥ 𝑐 ◦ |𝑦|; and
furthermore, for 𝑗 ≥ 1, such that 𝑦 ∈ Γ ( 𝑗) =⇒ 𝑦 · 𝜉 ◦ < 0. We can define, for
𝑗 = 0, 1...,𝜈, ∫
𝜉
𝑢 𝑗 (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 b
𝑢 (𝜉) 𝜓 𝑗 d𝜉. (3.5.12)
R 𝑛 |𝜉 |
Remark 3.5.14 The integral in (3.5.12) is an oscillatory integral, in the following
sense. There is an integer 𝑚 ≥ 0 such that
−𝑚
1 + |𝜉 | 2 𝑢 (𝜉)| ∈ 𝐿 1 (R𝑛 ) .
|b (3.5.13)

Then 𝑢 𝑗 = (1 − Δ 𝑥 ) 𝑚 𝑣 𝑗 in the distribution sense, with

∫ −𝑚
−𝑛 𝑖 𝑥· 𝜉 2 𝜉
𝑣 𝑗 (𝑥) = (2𝜋) e 1 + |𝜉 | 𝑢 (𝜉) 𝜓 𝑗
b d𝜉 (3.5.14)
R𝑛 |𝜉 |

a continuous function in R𝑛 tending to zero at infinity. □

Í
The Fourier inversion formula implies 𝑢 = 𝑢 0 + 𝜈𝑗=1 𝑢 𝑗 . If 0 ≤ 𝑗 ≤ 𝜈 then
∫
𝜉
𝑢˜ 𝑗 (𝑧) = (2𝜋) −𝑛 e𝑖𝑧· 𝜉 b
𝑢 (𝜉) 𝜓 𝑗 d𝜉
R𝑛 |𝜉 |

defines a holomorphic function in R𝑛 + 𝑖Γ ( 𝑗) ; (3.5.13) implies easily that 𝑢˜ 𝑗 ∈

Otemp R𝑛 + 𝑖Γ ( 𝑗) and 𝑢 𝑗 = 𝑏𝑈 𝑢˜ 𝑗 . This being so, Proposition 3.5.11 implies that
(𝑥, 𝜉 ◦ ) ∉ 𝑊 𝐹a 𝑢 𝑗 if 𝑗 ≥ 1. It follows that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢) if and only if

(𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢 0 ). As a matter of fact we are going to prove that 𝑢 0 is real-analytic
in a neighborhood of 𝑥 ◦ .
3.5 The Analytic Wave-Front Set of a Distribution 67

Let 𝜑 𝑁 (𝑁 = 1, 2, ...) be an Ehrenpreis sequence relative to the pair (𝑈, 𝑉) with

𝑥 ◦ ∈ 𝑉 ⊂⊂ 𝑈 ⊂⊂ 𝑢, such that (3.5.6) holds true. In what follows we shall only
consider integers 𝑁 ≥ 𝑛 + 2. We have
∫
𝜂
𝜑š
𝑁 0𝑢 (𝜉) = 𝜑
b 𝑁 (𝜉 − 𝜂) 𝑢
b (𝜂) 𝜓 0 d𝜂
R 𝑛 |𝜂|

and

|𝜉 | 𝑘 | 𝜑š
𝑁 𝑢 0 (𝜉)|
𝑘 ∫
∑︁ 𝑘 𝜂
≤ |𝜉 − 𝜂| ℓ | 𝜑
b𝑁 (𝜉 − 𝜂)| |𝜂| 𝑘−ℓ b
𝑢 (𝜂) 𝜓0 d𝜂.
ℓ=0
ℓ R𝑛 |𝜂|

Now we assume 0 ≤ 𝑘 ≤ 𝑁 − 𝑛 − 1; thanks to our hypothesis (3.5.6) the preceding

inequality implies
𝑘
!
|𝜂| ℓ
∫ ∑︁
𝑘 𝑁 +1
|𝜉 | | 𝜑š
𝑁 𝑢 0 (𝜉)| ≤ 𝐶 𝑘! |𝜑
b𝑁 (𝜂)| d𝜂.
R𝑛 ℓ=0 ℓ!

We also have
𝑘
!
|𝜂| ℓ
∫ ∑︁
|𝜑
b𝑁 (𝜂)| d𝜂
R𝑛 ℓ=0
ℓ!
𝑘
|𝜂| ℓ |𝜂| 𝑛+1
∫
d𝜂 ∑︁
≤ 1+ b𝑁 (𝜂)
𝜑
R𝑛 1+ | 𝜂 | 𝑛+1 ℓ! (𝑛 + 1)!
(𝑛+1)! ℓ=0 𝐿∞
𝑘+𝑛+1
∑︁ 3ℓ
≤ 𝐶 (𝑛) |𝜂| ℓ 𝜑
b𝑁 (𝜂) 𝐿∞
,
ℓ=0
ℓ!

where 𝐶 (𝑛) > 0 depends only on the dimension 𝑛. Below 𝐶1 and 𝐶2 are positive
constants independent of 𝑁; 𝐶 is the constant in (3.5.6). From (3.2.6) where we take
𝜈 = ℓ we derive
sup |𝜂| ℓ | 𝜑
b𝑁 (𝜂)| 𝐿 ∞ ≤ 𝐶1ℓ+1 𝑁 ℓ
R𝑛

whence

∞
∑︁ 1
|𝜉 | 𝑘 | 𝜑š 𝑁
𝑁 𝑢 0 (𝜉)| ≤ 𝐶 (𝑛) 𝐶1 𝐶 𝑘! (3𝐶1 𝑁) ℓ
ℓ=0
ℓ!
≤ 𝐶 (𝑛) 𝐶1 (𝐶 exp 3𝐶1 ) 𝑁 𝑘! ≤ 𝐶2𝑁 +1 𝑘!.

Since here 𝜉 ∈ R𝑛 and 𝑘 ≤ 𝑁 − 𝑛 − 1 are arbitrary, Proposition 3.2.5 and Remark

3.2.3 imply the analyticity of 𝑢 0 in a neighborhood of 𝑥 ◦ . □
Chapter 4
Analyticity of Solutions of Linear PDEs. Basic
Results

In the late 1930s I.G. Petrowski proved that all classical solutions of a linear PDE
with constant coefficients are analytic if and only if the equation is elliptic (he
extended this result to a class of systems of linear PDEs with constant coefficients).
Petrowski’s theorem signals clearly that this is XXth century analysis: it characterizes
an entire class of PDEs. With the (notable!) exception of the Cauchy–Kovalevskaya
Theorem, this generalizing tendency is not apparent in the 1800s; it shows up right
at the start of the next century in works such as [Holmgren, 1901], [Levi, 1907],
[Hilbert, 1910], [Hadamard, 1923], perhaps impelled by what was happening in
other areas of mathematics. Generally speaking, earlier mathematicians would study
special equations in depth, for sure with some (limited) variations, but would not
seek grand generalizations of an important prototype.
We prove the if part in Petrowski’s theorem (or, rather, in its modern formulation
in terms of analytic hypoellipticity, see Definition 3.1.2) using a parametrix. This
approach also proves that all elliptic differential operators with constant coefficients
are C ∞ hypoelliptic (Definition 2.4.5); the latter statement for the Laplacian was first
proved in [Weyl, 1940] and is still known as Weyl’s Lemma. In Section 4.1 we prove
the analytic hypoellipticity of all elliptic PDEs with analytic coefficients; actually,
we give two distinct proofs of this fact, the first based on the construction of a local
analytic parametrix (Definition 3.1.1), the second based on the Gårding inequality.
The result is easily extended to all elliptic square systems of PDEs with analytic
coefficients.
Section 4.2 introduces examples of second-order linear PDOs which are not
elliptic yet are analytic hypoelliptic, the simplest among them being

𝐿 𝑝 = D2𝑥 + 𝑥 2 𝑝 D2𝑦 , 𝑝 = 1, 2, ..., (4.1)

in R2 . The operator 𝐿 1 is intimately related to the Heisenberg group: if we set

𝜕 𝜕 𝜕
𝑋 = 𝜕𝑥 = 𝑖D 𝑥 , 𝑌 = 𝑥 𝜕𝑦 = 𝑖𝑥D 𝑦 , the equation [𝑋, 𝑌 ] = 𝑋𝑌 − 𝑌 𝑋 = 𝜕𝑦 = 𝑍 is the
𝜕2 𝜕2
defining relation of the Heisenberg Lie algebra and −𝐿 1 = 𝜕𝑥 2
+ 𝑥 2 𝜕𝑦 2 is its Casimir
d2 2
operator. The differential operator in a single variable, − d𝑥 2 + 𝑥 , is commonly

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 69

F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_4
70 4 Analyticity of Solutions of Linear PDEs. Basic Results

d 2
referred to as the harmonic oscillator whereas, when 𝑝 ≥ 2, − d𝑥 2 𝑝 is referred
2 + 𝑥
to as the anharmonic oscillator. Sometimes and, admittedly, somewhat abusively,
we shall refer to 𝐿 1 as the harmonic oscillator operator, merely because Fourier
d2 2 2
transform with respect to 𝑦 transforms it into − d𝑥 2 + 𝜂 𝑥 .
The bulk of Section 4.2, however, is devoted to a fairly complete analysis of a
multi-dimensional model of the following type:
𝑛
∑︁ 𝜕
𝑰𝑑 𝑎 𝑗,𝑘 D 𝑥 𝑗 + 𝑖𝑥 𝑗 D 𝑦 D 𝑥𝑘 − 𝑖𝑥 𝑘 D 𝑦 + 𝑩 , (4.2)
𝑗,𝑘=1
𝜕𝑦

where 𝑨 = 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 is a real 𝑛 × 𝑛 matrix, symmetric positive definite, 𝑩 is a
complex 𝑑 × 𝑑 matrix and 𝑰 𝑑 is the 𝑑 × 𝑑 identity matrix. We construct a (global,
i.e., in R𝑛+1 × R𝑛+1 ) fundamental solution and prove its analyticity off the diagonal
– this under a condition on the behavior of the spectrum of 𝑩 relative to that of 𝑨
that is shown to be necessary and sufficient for the analytic hypoellipticity of (4.2).
The main motivation for this study is to contrast the features of second-order elliptic
degenerate PDEs to those of elliptic PDEs, and more precisely to stress the influence
of lower-order terms on the regularity of solutions of the former and to present a
method of dealing with them. Variable coefficients generalization of (4.2) in the
microlocal (i.e., phase-space) set-up are in part based on what is done in this chapter
(cf. [Treves, 1978]). Our approach makes use, at certain points, of the Hermite
functions whose role in connection with the Schwartz space S is summarized (with
proofs) in the Appendix to this chapter. Needless to say, the latter material is widely
available in other texts (as well as in Wikipedia).
We close Section 4.2 with the first example of a so-called sum of squares operator
which is C ∞ hypoelliptic but not analytic hypoelliptic, the Baouendi–Goulaouic
operator in R3 ,
𝜕2 2 𝜕
2 𝜕2
+ 𝑥 + .
𝜕𝑥 2 𝜕𝑦 2 𝜕𝑧 2

4.1 Analyticity of Solutions of Elliptic Linear PDEs

4.1.1 Elliptic linear PDOs with constant coefficients

We refer the reader to Definitions 1.3.2 (ellipticity), 2.4.5 (C ∞ hypoellipticity) and

3.1.2 (C 𝜔 hypoellipticity). In this subsection we consider a linear PDO with constant
coefficients, ∑︁
𝑃 = 𝑃 (D) = 𝑐 𝛼 D 𝛼 , 𝑐 𝛼 ∈ C. (4.1.1)
𝛼

Since (4.1.1) is translation invariant, for it to be hypoelliptic or analytic hypoelliptic

at a point is equivalent to it being so at every point. If 𝑃 (D) is hypoelliptic (resp.,
analytic hypoelliptic) obviously every fundamental solution (Definition 2.4.1) of
4.1 Analyticity of Solutions of Elliptic Linear PDEs 71

𝑃 (D) must be C ∞ (resp., C 𝜔 ) in R𝑛 \ {0}. The classical fundamental solutions of

the Laplacian Δ 𝑥 = 𝜕𝑥21 + · · · + 𝜕𝑥2𝑛 have this property. The classical fundamental
solution of the heat equation 𝜕𝑡 − Δ 𝑥 in R𝑛+1 ,

|𝑥| 2

𝐻 (𝑡)
𝐸 (𝑡, 𝑥) = 1
exp − , 𝐻: Heaviside’s function, (4.1.2)
(4𝜋𝑡) 2 𝑛 4𝑡

is C ∞ but not C 𝜔 in R𝑛 \ {0}. Actually, it is fairly easy to characterize the analytic

hypoelliptic PDEs with constant coefficients (in contrast to those that are C ∞ hy-
poelliptic; see [Hörmander, 1955]); this is essentially what is done in [Petrowsky,
1939].

Theorem 4.1.1 The following properties of 𝑃 (D) are equivalent:

(a) 𝑃 (D) is elliptic;
(b) 𝑃 (D) is analytic hypoelliptic;
(c) ∀𝜑 ∈ C ∞ (R𝑛 ), 𝑃 (D) 𝜑 ∈ C 𝜔 (R𝑛 ) implies 𝜑 ∈ C 𝜔 (R𝑛 ).

We sketch the fairly instructive proof. The order of 𝑃 will be 𝑚 ≥ 1.

Proof Obviously (b)=⇒(c). First we prove that (a)=⇒(b). To say that 𝑃 (D) is
elliptic is the same as saying that there are positive constants 𝑅 and 𝑐 such that
|𝜉 | ≥ 𝑅 =⇒ |𝑃 (𝜉)| ≥ 𝑐 |𝜉 | 𝑚 . Define then the distribution in R𝑛 ,
∫
𝐺 ◦ (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 𝑃 (𝜉) −1 d𝜉.
| 𝜉 |>𝑅

If 𝑚 ≤ 𝑛 the integral is not absolutely convergent: it is an “oscillatory integral”. We

can use differentiation under the integral sign:
∫
𝐺 ◦ (𝑥) = (2𝜋) −𝑛 Δ 𝑥𝑁 e𝑖 𝑥· 𝜉 𝑃 (𝜉) −1 |𝜉 | −2𝑁 d𝜉
| 𝜉 |>𝑅

with a suitably large integer 𝑁. Another approach is to define 𝐺 ◦ as the distribution

limit as 𝜀 ↘ 0 of the smooth (analytic, actually) functions
∫
2
−𝑛
𝐺 𝜀 (𝑥) = (2𝜋) e𝑖 𝑥· 𝜉 −𝜀 | 𝜉 | 𝑃 (𝜉) −1 d𝜉.
| 𝜉 |>𝑅

The property that 𝐺 ◦ ∈ C 𝜔 (R𝑛 \ {0}) follows from estimates of the integrals
∫
|𝑥| 2𝑘 D 𝛼 𝐺 ◦ (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 Δ 𝑘𝜉 𝜉 𝛼 𝑃 (𝜉) −1 d𝜉
| 𝜉 |>𝑅

for any 𝛼 ∈ and positive integers 𝑘 proportionally large. Lastly we note that
Z+𝑛
𝑃 (D) 𝐺 ◦ = 𝛿 − 𝑅 in R𝑛 ;
∫
𝑅 (𝑧) = (2𝜋) −𝑛 e𝑖𝑧· 𝜉 d𝜉
| 𝜉 | ≤𝑅
72 4 Analyticity of Solutions of Linear PDEs. Basic Results

is an entire function in C𝑛 . It suffices then to apply Theorem 3.1.4 with 𝐺 (𝑥, 𝑦) =

𝐺 ◦ (𝑥 − 𝑦).
(c)=⇒(a). If 𝑃 is not elliptic there exist vectors v, v′ ∈ R𝑛 \ {0} such that
𝑃𝑚 (v) ≠ 0, 𝑃𝑚 (v′) = 0. We select the coordinates in R𝑛 so that v = (1, 0, ..., 0),
v′ = (0, 1, 0, ..., 0). We seek a nonanalytic solution ℎ (𝑥1 , 𝑥2 ) ∈ C ∞ R2 of the
homogeneous equation 𝑃 (D) ℎ = 𝑃 (D1 , D2 , 0, ..., 0) ℎ = 0. We use the fact that
𝑚
∑︁
𝑃 (𝜉1 , 𝜉2 , 0, ..., 0) = 𝑐𝜉1𝑚 + 𝜉1𝑚−𝑘 𝑞 𝑘 (𝜉2 ) ,
𝑘=1

with 0 ≠ 𝑐 ∈ C and polynomials 𝑞 𝑘 in a single variable such that deg 𝑞 𝑘 ≤ 𝑘 and

deg 𝑞 𝑚 < 𝑚. Applying Puiseux’ Theorem it is not difficult to check that the equation
𝑃 (𝜉1 , 𝜉2 , 0, ..., 0) = 0 has a solution 𝜉1 = Φ (𝜉2 ) which is holomorphic and satisfies
an estimate |Φ (𝜉2 )| ≤ const. |𝜉2 | 𝑟 (𝑟 ∈ Q, 0 < 𝑟 < 1) in some truncated sector
|arg 𝜉2 | < 𝛿, |𝜉2 | > 𝑅 (𝛿, 𝑅 > 0). The sought function is then
∫ +∞
e𝑖 𝑥1 Φ( 𝜉2 )+𝑖 𝑥2 𝜉2 − 𝜉2 d𝜉2
𝜃
ℎ (𝑥1 , 𝑥2 ) =
𝑅

with 𝑟 < 𝜃 < 1: if 𝑥2 is replaced by 𝑥2 + 𝑖𝑦 2 with 𝑦 2 < 0 the integral diverges. □

4.1.2 Elliptic linear PDOs with variable coefficients. Constructing a

parametrix

How
Í much of Theorem 4.1.1 remains valid for a linear PDO 𝑃 = 𝑃 (𝑥, D) =
|𝛾 | ≤𝑚 𝑐 𝛾 (𝑥) D with variable C coefficients? The sufficiency part does:
𝛾 𝜔

Theorem 4.1.2 If the coefficients of 𝑃 are real-analytic in Ω and if 𝑃 is elliptic then

𝑃 is analytic hypoelliptic in Ω.

The proof of this statement is usually based on one of two approaches: construction
of a local analytic parametrix (Definition 3.1.1) and application of Theorem 3.1.4;
or the exploitation of 𝐿 2 estimates, starting from the classical Gårding inequality.
Proof One can construct a local parametrix of the elliptic linear PDO 𝑃 as an
oscillatory integral
∫
𝐺 (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑦, 𝜉) d𝜉.
R𝑛

Here we reason in some neighborhood 𝑈 of a generic point 𝑥 ◦ ∈ Ω. The “amplitude”

𝑎 (𝑦, 𝜉) is a smooth function of both variables 𝑦, 𝜉, determined by the requirement
that 𝐺 (𝑥, 𝑦) be real-analytic in (𝑈 × 𝑈) \ diag (𝑈 × 𝑈) and that

𝑃⊤ 𝑦, D 𝑦 𝐺 (𝑥, 𝑦) − 𝛿 (𝑥, 𝑦) ∈ C 𝜔 (𝑈 × 𝑈)

(4.1.3)
4.1 Analyticity of Solutions of Elliptic Linear PDEs 73

where 𝑃⊤ stands for the transpose of 𝑃. The smoothness of 𝑎 with respect to 𝑦

entails the regularity (Definition 2.3.1) of the kernel 𝐺. Differentiation under the
integral sign, commutation with the exponential followed by Taylor expansion show
immediately that
∫
𝑃⊤ 𝑦, D 𝑦 𝐺 (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑃⊤ 𝑦, D 𝑦 𝑎 (𝑦, 𝜉) d𝜉

R𝑛
∑︁ 1 ∫
e𝑖 ( 𝑥−𝑦) · 𝜉 D 𝑦𝛼 𝑎 (𝑦, 𝜉) 𝜕 𝜉𝛼 𝑃⊤ (𝑦, −𝜉) d𝜉.

=
𝛼
𝛼! R 𝑛

One begins by defining 𝑎 as a “formal symbol”:

∞
∑︁
𝑎 (𝑦, 𝜉) = 𝑎 𝑘 (𝑦, 𝜉) (4.1.4)
𝑘=0

with 𝑎 𝑘 (𝑦, 𝜉) decaying as |𝜉 | ↗ +∞ increasingly fast (as 𝑘 ↗ +∞). For |𝜉 | large

the 𝑎 𝑘 are determined successively, and unambiguously, by the recurrence equations

𝑎 0 (𝑦, 𝜉) 𝑃⊤ (𝑦, −𝜉) = 1 (4.1.5)

and for 𝑘 = 1, 2, ...,

𝑘−1 ∑︁
∑︁ 1 𝛼
𝑎 𝑘 (𝑦, 𝜉) 𝑃⊤ (𝑦, −𝜉) = − D 𝑦 𝑎 ℓ (𝑦, 𝜉) 𝜕 𝜉𝛼 𝑃⊤ (𝑦, −𝜉) .

(4.1.6)
ℓ=0 | 𝛼 |=𝑘−ℓ
𝛼!

Indeed, 𝑃⊤ (𝑥, −D 𝑥 ) is elliptic of the same order, 𝑚, as 𝑃. For 𝑈 suitably small and
𝑅 > 0 suitably large, 𝑦 ∈ 𝑈 and |𝜉 | ≥ 𝑅 =⇒ |𝑃⊤ (𝑦, −𝜉)| ≥ 𝑐 |𝜉 | 𝑚 . When this is true,
solving (4.1.5)–(4.1.6) is straightforward; each term 𝑎 𝑘 is a real-analytic function
of (𝑦, 𝜉), 𝑦 ∈ 𝑈, 𝜉 ∈ R𝑛 , |𝜉 | ≥ 𝑅. Reasoning by induction on 𝑘 it is possible to
prove that there is a constant 𝐶◦ > 0 such that, for all 𝑘 and all pairs of multi-indices
𝛼, 𝛽 ∈ Z+𝑛 ,
𝑘+|𝛽 |+1
𝛼!𝛽! |𝜉 | −𝑚−𝑘−|𝛽 |
𝛽
𝜕𝑦𝛼 𝜕 𝜉 𝑎 𝑘 (𝑦, 𝜉) ≤ 𝐶◦ (4.1.7)

for all 𝑦 ∈ 𝑈, |𝜉 | ≥ 𝑅 (1 + |𝛽|). To ensure that 𝐺 (𝑥, 𝑦) is analytic in

𝑈 × 𝑈\ diag (𝑈 × 𝑈) and that (4.1.3) holds we multiply each “symbol” 𝑎 𝑘 (𝑦, 𝜉)
by a cutoff function 𝜒 (𝜉/𝑅 𝑘 ) with 𝜒 ∈ C ∞ (R𝑛 ), 𝜒 (𝜉) = 0 if |𝜉 | ≤ 1 and 𝜒 (𝜉) = 1
if |𝜉 | > 2. The positive numbers 𝑅 𝑘 ↗ +∞ must, and can, be chosen so as to ensure
the required analyticity. In conclusion, we take [in the place of (4.1.4)]
∞
∑︁
𝑎 (𝑦, 𝜉) = 𝑎 𝑘 (𝑦, 𝜉) 𝜒 (𝜉/𝑅 𝑘 ) . (4.1.8)
𝑘=0

The details can be found, in a wider context, in Section 17.2 (also in Ch. V, [Treves,
1980]). As with the Green function of the Laplacian we have in 𝑈 × 𝑈 [at least for
𝑛 > max (𝑚, 2)]:
74 4 Analyticity of Solutions of Linear PDEs. Basic Results
∫
−𝑛 −𝑛+1 𝑖𝜃·𝜉 𝜉 𝐵 (𝑥, 𝑦)
𝐺 (𝑥, 𝑦) = (2𝜋) |𝑥 − 𝑦| e 𝑎 𝑦, d𝜉 = ,
R𝑛 |𝑥 − 𝑦| |𝑥 − 𝑦| 𝑛−𝑚
𝑥−𝑦
where 𝜃 = | 𝑥−𝑦 | and 𝐵 (𝑥, 𝑦) has “good” regularity. □

Remark 4.1.3 The preceding argument is easier in the C ∞ category, where the
estimates of the derivatives, e.g. (4.1.7), require less precision. It proves that every
elliptic operator in Ω with C ∞ coefficients is C ∞ hypoelliptic (Definition 2.4.5). We
shall avail ourselves of this knowledge in the sequel.

4.1.3 Exploiting the Gårding inequality

In this subsection we transcribe the proof of Theorem 4.1.2 as presented in pp. 178–
180 of [Hörmander, 1969]. We do this because the proof is the most basic illustration
of the techniques used to deduce analytic hypoellipticity from a priori 𝐿 2 estimates.
The starting point will be a special case of the classical Gårding inequality ([Gård-
ing, 1953]; for a proof see, e.g., [Treves, 1975], Ch. 36, p. 348 et seq.).

Theorem 4.1.4 If the coefficients of 𝑃 = 𝑃 (𝑥, D) = |𝛾 | ≤𝑚 𝑐 𝛾 (𝑥) D𝛾 are C ∞ in Ω

Í
and if 𝑃 is elliptic of order 𝑚 ≥ 1 then to every open set Ω′ ⊂⊂ Ω there is a constant
𝐶 > 0 such that, for all 𝜑 ∈ Cc∞ (Ω′),

∑︁ ∑︁
∥D 𝛼 𝜑∥ ≤ 𝐶 ­ ∥𝑃𝜑∥ + ∥D 𝛼 𝜑∥ ® . (4.1.9)
© ª
| 𝛼 | ≤𝑚 « | 𝛼 |<𝑚 ¬
∫ 21
We have used the notation ∥ 𝑓 ∥ = R𝑛 | 𝑓 (𝑥)| 2 d𝑥 .
Actually, the conclusion in Theorem 4.1.4 demands that 𝑃 be elliptic. This is
easily proved by reasoning by contradiction: suppose that 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) = 0 for some
𝑥 ◦ ∈ Ω′, 0 ≠ 𝜉 ◦ ∈ R𝑛 and take 𝜑 (𝑥) = 𝜑𝜆 (𝑥) =
𝑛
−𝑚 𝑖𝜆𝑥· 𝜉 ◦ ◦
∫ 𝜆 2 e 2 𝜓 (𝜆 (𝑥 − 𝑥 )), with
∞
𝜆 > 0 suitably large and 𝜓 ∈ Cc (R ) such that R𝑛 |𝜓 (𝑥)| d𝑥 = 1. The right-hand
𝑛

side in (4.1.9) tends to zero as 𝜆 → +∞ whereas the left-hand side does not.
Proof (of Theorem 4.1.2 starting from the Gårding inequality) We denote by 𝔅𝑟
the ball of radius 𝑟 centered at an arbitrary point of Ω′ which, for convenience,
we take to be the origin of R𝑛 . Let 𝑢 ∈ D ′ (Ω) be such that 𝑃𝑢 ∈ C 𝜔 (Ω); then
𝑢 ∈ C ∞ (Ω) (Remark 4.1.3). For 𝑟 ◦ > 0 sufficiently small we may apply the Cauchy–
Kovalevskaya Theorem (Theorem 5.2.4, or, for the linear version, [Treves, 1975],
pp. 148–150) to solve 𝑃𝑣 = 𝑃𝑢 in 𝔅𝑟◦ with 𝑣 ∈ C 𝜔 𝔅𝑟◦ . Substituting 𝑢 − 𝑣 for 𝑢
shows that there is no loss of generality in hypothesizing that 𝑃𝑢 ≡ 0 in 𝔅𝑟◦ .
In order to estimate the partial derivatives of 𝑢 we shall avail ourselves of special
Ehrenpreis cutoff functions (Section 4.2). In what follows we shall always assume
0 < 𝑟 < 𝑟 ◦ /2. We select a positive number 𝜀 < 𝑟/2 and we define [cf. (3.2.5);
𝑚 = order 𝑃]
4.1 Analyticity of Solutions of Elliptic Linear PDEs 75

𝑚 factors
z }| {
𝜒𝑟 , 𝜀,𝑚 = 𝜌 𝜀/2𝑚 ∗ 𝜌 𝜀/2𝑚 ∗ · · · ∗ 𝜌 𝜀/2𝑚 ∗ 𝜒𝑟 , 𝜀 (4.1.10)
where 𝑚 = order 𝑃 and 𝜒𝑟 , 𝜀 is the characteristic function of 𝔅𝑟−𝜀/2 . We have

𝜒𝑟 , 𝜀,𝑚 (𝑥) = 1 if |𝑥| < 𝑟 − 𝜀, (4.1.11)

𝜒𝑟 , 𝜀,𝑚 (𝑥) = 0 if |𝑥| > 𝑟, (4.1.12)

and [cf. (3.2.2)]

∀𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑚, max 𝜕 𝛼 𝜒𝑟 , 𝜀,𝑚 ≤ (𝐶◦ 𝑚) | 𝛼 | .

As shown in the proof of Proposition 3.2.1, we can take 𝐶◦ = 𝐾𝜀 −1 since

dist (𝔅𝑟−𝜀 , R𝑛 \𝔅𝑟 ) = 𝜀; 𝐾 is the constant in (3.2.4), ibid. Thus

∀𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑚, max 𝜕 𝛼 𝜒𝑟 , 𝜀,𝑚 ≤ (𝑚𝐾/𝜀) | 𝛼 | . (4.1.13)

In what follows we deal with two multi-indices 𝛼, 𝛽 ∈ Z+𝑛 such that |𝛼| = 𝑚,
|𝛽| = 𝑘 ≥ 0, otherwise arbitrary. Taking (4.1.11) into account we deduce from
(4.1.10):
∫ 21
𝛼+𝛽 2
𝑢 (𝑥) d𝑥 ≤ D 𝛼 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢
D (4.1.14)
| 𝑥 |<𝑟−𝜀
∑︁
≤ 𝐶 𝑃 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢 + 𝐶 D𝛾 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢 .
|𝛾 |<𝑚

The Leibniz rule (1.1.3) yields immediately

𝑃 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢 ≤ 𝜒𝑟 , 𝜀,𝑚 𝑃D𝛽 𝑢
∑︁ ∑︁ 𝛼 ′ ′ ′′ ′′
+ ′′
𝑐 𝛼′ D 𝛼 −𝛼 𝜒𝑟 , 𝜀,𝑚 D 𝛼 +𝛽 𝑢 .
′ 𝛼′′ ≺ 𝛼′
𝛼
| 𝛼 | ≤𝑚

By applying (4.1.12) and 4.1.13) we then deduce

∫ 21
2
𝑃 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢 ≤ 𝑃D𝛽 𝑢 d𝑥 (4.1.15)
| 𝑥 | ≤𝑟
∑︁ 𝛼 ′ ∫ 21
2
| 𝛼′ −𝛼′′ |
∑︁
𝛼′′ +𝛽
+𝐶 (𝐾1 /𝜀) D 𝑢 d𝑥 ,
𝛼′′ ≺ 𝛼′
𝛼 ′′ | 𝑥 | ≤𝑟
| 𝛼′ | ≤𝑚

where ∑︁
𝐶 = sup |𝑐 𝛼′ | .
𝔅𝑟◦ | 𝛼′ | ≤𝑚
76 4 Analyticity of Solutions of Linear PDEs. Basic Results

By the same token we get

∑︁
D𝛾 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢 (4.1.16)
|𝛾 |<𝑚

∑︁ ∑︁ 𝛾 ∫ 21
2
|𝛾−𝛾′ | 𝛽+𝛾′
≤ (𝑚𝐾/𝜀) D 𝑢 d𝑥 .
𝛾′ ⪯𝛾
𝛾′ | 𝑥 | ≤𝑟
|𝛾 |<𝑚

The expression (2.1.1), the Leibniz rule (1.1.3) and the fact that 𝑃𝑢 ≡ 0 in 𝔅𝑟◦ imply
∫ 21 ∫ 21
𝛽 2 𝛽 𝛽 2
𝑃D 𝑢 d𝑥 = 𝑃D 𝑢 − D 𝑃𝑢 d𝑥
| 𝑥 | ≤𝑟 | 𝑥 | ≤𝑟
∑︁ ∑︁ 𝛽
! ∫ 12
2
𝛽−𝛽′ 𝛽′ +𝛾
≤ sup D 𝑐𝛾 D 𝑢 d𝑥
𝛽′ ≺𝛽
𝛽 ′ 𝔅𝑟◦ | 𝑥 | ≤𝑟
|𝛾 | ≤𝑚

where 𝛾 ≺ 𝛽 means that 𝛾 𝑗 ≤ 𝛽 𝑗 for every 𝑗 but 𝛾 𝑗 < 𝛽 𝑗 for some 𝑗 (1 ≤ 𝑗 ≤ 𝑛).
The analyticity of the coefficients 𝑐 𝛼′ of 𝑃 means that, possibly after increasing the
constant 𝐶, we have, for every 𝛾 ∈ Z+𝑛 ,
∑︁
sup |D𝛾 𝑐 𝛼′ | ≤ 𝐶 |𝛾 |+1 𝛾!.
𝔅𝑟◦ | 𝛼′ | ≤𝑚

This implies
∫ 21 ∫ 12
2
∑︁ ∑︁ 𝛽! 2
𝛽 |𝛽−𝛽′ | 𝛽′ +𝛾
𝑃D 𝑢 d𝑥 ≤𝐶 𝐶 D 𝑢 d𝑥
| 𝑥 | ≤𝑟 𝛽′ ≺𝛽
𝛽 ′! | 𝑥 | ≤𝑟
|𝛾 | ≤𝑚

whence, by (4.1.15),

𝐶1−1 𝑃 𝜒𝑟 , 𝜀,𝑚 D𝛽 𝑢 (4.1.17)
∫ 12
∑︁ ∑︁ 𝛽! 2
|𝛽−𝛽′ | 𝛽′ +𝛾
≤ 𝐶 D 𝑢 d𝑥
𝛽′ ≺𝛽
𝛽 ′! |𝑥 | ≤𝑟
|𝛾 | ≤𝑚

∑︁ 𝛼 ′ ∫ 21
2
| 𝛼′ −𝛼′′ |
∑︁
𝛼′′ +𝛽
+ (𝑚𝐾/𝜀) D 𝑢 d𝑥 .
𝛼′′ ≺ 𝛼′
𝛼 ′′ | 𝑥 | ≤𝑟
| 𝛼′ | ≤𝑚

We combine (4.1.17) with (4.1.14) and 4.1.16):

4.1 Analyticity of Solutions of Elliptic Linear PDEs 77
∫ 21
2
𝐶1−1 D 𝛼+𝛽
𝑢 (𝑥) d𝑥
| 𝑥 |<𝑟−𝜀
∫ 12
∑︁ ∑︁ 𝛽! 2
|𝛽−𝛽′ | 𝛽′ +𝛾
≤ 𝐶 D 𝑢 d𝑥
𝛽′ ≺𝛽
𝛽 ′! | 𝑥 | ≤𝑟
|𝛾 | ≤𝑚

∑︁ 𝛼 ′ ∫ 21
2
| 𝛼′ −𝛼′′ |
∑︁
𝛼′′ +𝛽
+ (𝑚𝐾/𝜀) D 𝑢 d𝑥
𝛼′′ ≺ 𝛼′
𝛼 ′′ | 𝑥 | ≤𝑟
| 𝛼′ | ≤𝑚

∑︁ ∑︁ 𝛾 ∫ 21
2
|𝛾−𝛾′ | 𝛽+𝛾′
+ (𝑚𝐾/𝜀) D 𝑢 d𝑥
𝛾′ ⪯𝛾
𝛾′ |𝑥 | ≤𝑟
|𝛾 |<𝑚

for a sufficiently large positive constant 𝐶1 depending only on 𝐶.

Actually we shall apply the last inequality after replacing 𝑟 by 𝑟 − (𝑘 − 1) 𝜀 with
𝑘 ≥ 1, assuming now 𝜀 < 𝑘+1 𝑟
. The latter insures that 𝜀 < 21 (𝑟 − (𝑘 − 1) 𝜀), in
agreement with our initial requirement. We reach the following conclusion:
∫ 21
2
𝐶1−1 D 𝛼+𝛽
𝑢 (𝑥) d𝑥 (4.1.18)
| 𝑥 |<𝑟−𝑘 𝜀
∫ 12
2
|𝛽−𝛽′ |
∑︁ ∑︁ 𝛽!
𝛽′ +𝛾
≤ 𝐶 D 𝑢 d𝑥
𝛽′ ≺𝛽
𝛽 ′! ◦ |𝑥 | ≤𝑟−(𝑘−1) 𝜀
|𝛾 | ≤𝑚

∑︁ 𝛼 ′ ∫ 21
2
| 𝛼′ −𝛼′′ |
∑︁
𝛼′′ +𝛽
+ (𝑚𝐾/𝜀) D 𝑢 d𝑥
𝛼′′ ≺ 𝛼′
𝛼 ′′ | 𝑥 | ≤𝑟−(𝑘−1) 𝜀
| 𝛼′ | ≤𝑚

∑︁ ∑︁ 𝛾 ∫ 21
2
|𝛾−𝛾′ | 𝛽+𝛾′
+ (𝑚𝐾/𝜀) D 𝑢 d𝑥 .
𝛾′ ⪯𝛾
𝛾′ | 𝑥 | ≤𝑟−(𝑘−1) 𝜀
|𝛾 |<𝑚

This will allow us to reason by induction on 𝑘 to derive from (4.1.18), for a suitably
large constant 𝐵 > 0 independent of 𝑘,
∫ 21
2
D 𝛼+𝛽 𝑢 (𝑥) d𝑥 ≤ 𝐵 (𝐵/𝜀) 𝑘+𝑚 . (4.1.19)
| 𝑥 |<𝑟−𝑘 𝜀

Recall that |𝛼| = 𝑚, |𝛽| = 𝑘. Induction on 𝑘 in each term of the right-hand side of
(4.1.18) leads to the estimate
78 4 Analyticity of Solutions of Linear PDEs. Basic Results
∫ 12
2
𝐶1−1 𝐵−1 D 𝛼+𝛽
𝑢 (𝑥) d𝑥 (4.1.20)
| 𝑥 |<𝑟−𝑘 𝜀
|𝛽−𝛽′ |
∑︁ ∑︁ 𝛽! ′
≤ 𝐶
′! ◦
(𝐵/𝜀) |𝛽 +𝛾 |
𝛽
|𝛾 | ≤𝑚 𝛽′ ≺𝛽
∑︁ ∑︁ 𝛼 ′ ′ ′′ ′′
+ ′′
(𝑚𝐾/𝜀) | 𝛼 −𝛼 | 𝐵 | 𝛼 +𝛽 | 𝜀 − |𝛽 |
𝛼
| 𝛼′ | ≤𝑚 𝛼′′ ≺ 𝛼′
∑︁ ∑︁ 𝛾! ′ ′
+ ′!
(𝑚𝐾/𝜀) |𝛾−𝛾 | (𝐵/𝜀) |𝛽+𝛾 | .
𝛾 ′ ⪯𝛾
𝛾
|𝛾 |<𝑚

We get directly:
∑︁ ∑︁ 𝛼 ′ ′ ′′ ′′
𝜀 𝑘+𝑚
′′
(𝑚𝐾/𝜀) | 𝛼 −𝛼 | 𝐵 | 𝛼 +𝛽 | 𝜀 − |𝛽 |
𝛼
| 𝛼′ | ≤𝑚 𝛼′′ ≺ 𝛼′
∑︁ ∑︁ 𝛼 ′ ′ ′′ ′′
≤ 𝐵𝑘 ′′
𝜅 1| 𝛼 −𝛼 | 𝐵 | 𝛼 | ≤ 𝐶 (𝑚) 𝐵 𝑘+𝑚−1
′ 𝛼′′ ≺ 𝛼′
𝛼
| 𝛼 | ≤𝑚

if 𝐵 > 𝑚𝐾; here and below 𝐶 (𝑚) stands for a positive constant depending only on
𝑚. Likewise,
∑︁ ∑︁ 𝛾! ′ ′
𝜀 𝑘+𝑚 ′!
(𝑚𝐾/𝜀) |𝛾−𝛾 | (𝐵/𝜀) |𝛽+𝛾 |
𝛾
|𝛾 |<𝑚 𝛾′ ⪯𝛾
∑︁ ∑︁ 𝛾 ′
≤ (𝑚 − 1)!𝜀 ′
𝑚𝐾 𝐵 |𝛽+𝛾 | ≤ 𝐶 (𝑚) 𝜀𝐵 𝑘+𝑚−1 .
𝛾′ ⪯𝛾
𝛾
|𝛾 |<𝑚

𝑟
Finally, recalling that 𝜀 < 𝑘+𝑚 and assuming 𝐵 > 2𝐶 we get
∑︁ ∑︁ 𝛽! ′ ′
𝜀 𝑘+𝑚 ′!
𝐶 |𝛽−𝛽 | (𝐵/𝜀) |𝛽 +𝛾 |
𝛽
|𝛾 | ≤𝑚 𝛽′ ≺𝛽
!
𝛽! ′
∑︁ ′
≤ 𝐶 (𝑚) 𝐵 𝑘+𝑚 sup ′ 𝜀 |𝛽−𝛽 | (𝐶/𝐵) |𝛽−𝛽 |
′
𝛽 ≺𝛽 𝛽 ! 𝛽′ ≺𝛽
|𝛽−𝛽′ |
𝛽! 1
≤ 2𝐶𝐶 (𝑚) 𝐵 𝑘+𝑚−1 sup ′ ≤ 2𝐶𝐶 (𝑚) 𝐵 𝑘+𝑚−1
′
𝛽 ≺𝛽 𝛽 ! |𝛽| + 𝑚

since 𝑝−𝑞 Ö𝑝−𝑞

𝑝! 1 𝑞+ 𝑗
𝑝, 𝑞 ∈ Z+ , 𝑞 < 𝑝 =⇒ = ≤ 1.
𝑞! 𝑝 𝑗=1
𝑝
4.1 Analyticity of Solutions of Elliptic Linear PDEs 79

Putting all this into (4.1.19) yields

∫ 21
2
D 𝛼+𝛽 𝑢 (𝑥) d𝑥 ≤ 𝐶2 (𝐵/𝜀) 𝑘+𝑚
| 𝑥 |<𝑟−𝑘 𝜀

for a suitably large constant 𝐶2 > 0 independent of 𝑘 and 𝜀; (4.1.19) ensues provided
𝐵 ≥ 𝐶2 .
𝑟
The end of the proof is clear: by taking 𝜀 = 2(𝑘+𝑚) we derive from (4.1.19):

∫ ! 21 𝑘+𝑚
2 2𝐵
𝛾
|D 𝑢 (𝑥)| d𝑥 ≤𝐵 (𝑘 + 𝑚) 𝑘+𝑚
|𝑥 |< 21 𝑟 𝑟

𝑘+𝑚
|𝛾| = 𝑘 + 𝑚. Stirling’s Formula tells us that (𝑘 + 𝑚)
for ≤ (𝑚 + 𝑘)!e 𝑘+𝑚 ≤
(𝑘+𝑚)! 𝑘+𝑚 𝑚+𝑘
𝛾! e 𝛾! ≤ (2𝑒) , and therefore

∫ ! 21
2
𝛾
|D 𝑢 (𝑥)| d𝑥 ≤ 𝐵𝑀 |𝛾 | 𝛾!
| 𝑥 |< 21 𝑟

for 𝑀 = 4𝐵𝑒/𝑟 and all 𝛾 ∈ Z+𝑛 . The proof of Theorem 4.1.4 is then completed by
applying Proposition 1.2.1. □

4.1.4 A word about elliptic systems

We may have to deal with elliptic systems of linear PDEs instead of a single equation.
When the system is square (meaning that there are as many unknowns as there are
equations) the extension of the preceding result is relatively easy. We are talking of
a system of equations
𝑁
∑︁
𝑃 𝑗,𝑘 (𝑥, D) 𝑢 𝑘 = 𝑓 𝑗 , 𝑗 = 1, ..., 𝑁, (4.1.21)
𝑘=1

where the operators 𝑃 𝑗,𝑘 (𝑥, D) are all of the same order 𝑚 ≥ 1 and all have
coefficients in C 𝜔 (Ω), Ω a domain in R𝑛 . The system (4.1.21), which we now
denote by 𝑷, is said to be elliptic if the 𝑁 × 𝑁 (complex matrix) matrix

𝜎 (𝑷) (𝑥, 𝜉) = 𝜎 𝑃 𝑗,𝑘 (𝑥, 𝜉) 𝑗,𝑘=1,..., 𝑁 (4.1.22)

[𝜎 𝑃 𝑗,𝑘 : principal symbol of 𝑃 𝑗,𝑘 (𝑥, D)] is nonsingular for all (𝑥, 𝜉) ∈ Ω × R𝑛 ,
𝜉 ≠ 0. Let then 𝑨 (𝑥, 𝜉) = 𝑎 𝑗,𝑘 (𝑥, 𝜉) be the adjugate matrix of 𝜎 (𝑷) (𝑥, 𝜉); we
have
80 4 Analyticity of Solutions of Linear PDEs. Basic Results

𝑨 (𝑥, 𝜉) 𝜎 (𝑷) (𝑥, 𝜉) = 𝜎 (𝑷) (𝑥, 𝜉) 𝑨 (𝑥, 𝜉) = (det 𝜎 (𝑷)) (𝑥, 𝜉) 𝑰 𝑁 (4.1.23)

(𝑰 𝑁 : the 𝑁 × 𝑁 identity matrix). We see that 𝑷 is elliptic, meaning that the scalar
operator det 𝜎 (𝑷) (𝑥, D) is elliptic, if and only if the following composites are
elliptic:

𝑨 (𝑥, D) 𝑷 (𝑥, D) = 𝑰 𝑁 (det 𝜎 (𝑷)) (𝑥, D) + 𝑹 (𝑥, D) (4.1.24)

𝑷 (𝑥, D) 𝑨 (𝑥, 𝜉) = 𝑰 𝑁 (det 𝜎 (𝑷)) (𝑥, D) + 𝑺 (𝑥, D) (4.1.25)

where 𝑹 (𝑥, D) and 𝑺 (𝑥, D) have order ≤ 𝑚 − 1. We know that det 𝜎 (𝑷) (𝑥, D)
satisfies the Gårding inequality (4.1.9); it follows that the same is true of the matrix
operators 𝑨 (𝑥, D) 𝑷 (𝑥, D) and 𝑷 (𝑥, D) 𝑨 (𝑥, 𝜉). Inspection of the proof in the
preceding subsection readily shows that it extends to the latter operators and leads
to the conclusion that both are analytic hypoelliptic (and C ∞ hypoelliptic). Now let
′

𝒖 ∈ D Ω; C be such that 𝑷𝒖 ∈ C Ω; C ; it ensues that 𝑨𝑷𝒖 ∈ C 𝜔 Ω; C 𝑁
𝑁 𝜔 𝑁

and therefore that 𝒖 ∈ C 𝜔 Ω; C 𝑁 . Applying this reasoning to every subdomain

of Ω proves that 𝑷 is analytic hypoelliptic. Exploiting (4.1.25) instead of (4.1.24)
allows us to conclude that 𝑨 is also analytic hypoelliptic.

4.2 Degenerate Elliptic Equations. Influence of Lower Order

Terms

In the previous section we have seen that the entailment elliptic=⇒analytic hypoel-
liptic in Petrowsky’s Theorem 6.3.1 extends to partial differential operators with
variable C 𝜔 coefficients. The converse, however, does not: there exist analytic hy-
poelliptic PDOs which are not elliptic. This is true for a whole class of second-order
degenerate elliptic operators of which the simplest example is the harmonic oscillator
operator in R2 ,
𝐿 1 = D2𝑥 + 𝑥 2 D2𝑦 . (4.2.1)
In this section we prove this fact for the model of this class, namely the matrix
differential operator in R𝑛+1
𝑛
∑︁
𝑎 𝑗,𝑘 𝑍 𝑗 𝑍 𝑘∗ + 𝑩D 𝑦

𝑷 𝑥, D 𝑥 , D 𝑦 = −𝑰 𝑑 (4.2.2)
𝑗,𝑘=1

where 𝑨 = 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 is a symmetric real 𝑛 × 𝑛 matrix, 𝑩 is a complex 𝑑 × 𝑑
matrix (𝑑 ≥ 1) satisfying special conditions and

𝑍 𝑗 = D 𝑥 𝑗 + 𝑖𝑥 𝑗 D 𝑦 , 𝑍 ∗𝑗 = D 𝑥 𝑗 − 𝑖𝑥 𝑗 D 𝑦 . (4.2.3)
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 81

As usual, D 𝑦 = √1 𝜕𝑦 𝜕
and 𝑰 𝑑 is the 𝑑 × 𝑑 identity matrix. The system (4.2.2) shall
−1
be assumed, throughout, to be elliptic with respect to 𝑥: we shall assume that the
matrix 𝑨 is positive definite. Under this hypothesis the study of (4.2.2) will fully
reveal the effect of the first-order terms 𝑩D 𝑦 .
On the (more) general case, in which 𝑨 and 𝑩 are allowed to be C 𝜔 functions
of (𝑥, 𝑦), we refer the reader to [Treves, 1978], [Métivier, 1980], [Métivier, 1981],
[Bove-Mughetti, 2020], also [Boutet-Grigis-Helffer, 1976].

4.2.1 The associated elliptic system of PDEs and the special condition
on the matrix 𝑩

In this subsection we consider the elliptic system in R𝑛 (sometimes referred to as

a Grushin operator, see [Grushin, 1970]) derived from (4.2.2) simply by replacing
D 𝑦 by 1:
𝑛
© ∑︁
𝑷◦ (𝑥, D 𝑥 ) = 𝑰 𝑑 ­ 𝑎 𝑗,𝑘 𝑋 𝑗 𝑋 𝑘′ ® + 𝑩 (4.2.4)
ª

« 𝑗,𝑘=1 ¬
where [cf. (4.2.3)]
𝜕 𝜕
𝑋𝑗 = − 𝑥 𝑗 , 𝑋 ′𝑗 = + 𝑥 𝑗 , 𝑗 = 1, ..., 𝑛. (4.2.5)
𝜕𝑥 𝑗 𝜕𝑥 𝑗
h i
We have 𝑋 𝑗 , 𝑋 𝑘 = 𝑋 𝑗 , 𝑋 𝑘′ = 0 if 𝑗 ≠ 𝑘, 𝑋 𝑗 , 𝑋 ′𝑗 = 2; 𝑋 𝑗 and −𝑋 ′𝑗 are the

transpose of each other.

The motivation for this brief prologue is to show the relevance of the relation
between the spectra of the matrices 𝑨 and 𝑩. The ellipticity of 𝑷◦ is equivalent to
the property that the eigenvalues 𝜒 𝑗 ( 𝑗 = 1, ..., 𝑛) of 𝑨 are either all > 0 or all < 0.
We denote by S R𝑛 ; C𝑑 the Schwartz space of C𝑑 -valued C ∞ functions rapidly


decaying at infinity; S R𝑛 ; C𝑑 (S (R𝑛 )) 𝑑 . We now state and prove a basic global
property of the operator 𝑷◦ .

Theorem 4.2.1 If the matrix 𝑨 is symmetric positive definite the following properties
of the operator (4.2.4) are equivalent:

(a) 𝑷◦ defines a Fréchet space automorphism ofÍS R𝑛 ; C𝑑 ;
(b+ ) no eigenvalue of the matrix 𝑩 is of the form 2 𝑗=1 𝑚 𝑗 𝜒 𝑗 with (𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z+𝑛 .
𝑛

Í
Proof We begin by carrying out a linear change of variables 𝑥˜𝑖 = 𝑛𝑗=1 𝛾𝑖, 𝑗 𝑥 𝑗 with

𝚪 = 𝛾𝑖, 𝑗 1≤ 𝑗,𝑘≤𝑛 orthogonal; we have

𝑛
𝜕 ∑︁ 𝜕
± 𝑥𝑗 = 𝛾𝑖, 𝑗 ± 𝑥˜𝑖
𝜕𝑥 𝑗 𝑖=1
𝜕 𝑥˜𝑖
82 4 Analyticity of Solutions of Linear PDEs. Basic Results

whence, in obvious notation,

𝑛
∑︁ 𝑛 ∑︁
∑︁ 𝑛
𝑎 𝑗,𝑘 𝑋 𝑗 𝑋 𝑘′ = 𝛾ℓ,𝑘 𝑎 𝑗,𝑘 𝛾𝑖, 𝑗 𝑋 e′ .
e𝑖 𝑋
ℓ
𝑗,𝑘=1 𝑗,𝑘=1 𝑖,ℓ=1

The matrix 𝑨 is transformed into 𝚪 𝑨𝚪⊤ , which is diagonal if we select 𝚪 appropri-

ately. We can assume this to be the case from the start, with 𝜒1 , ..., 𝜒𝑛 the diagonal
entries of 𝑨:
𝑛
©∑︁
𝑷◦ (𝑥, D 𝑥 ) = 𝑰 𝑑 ­ 𝜒 𝑗 𝑋 𝑗 𝑋 ′𝑗 ® + 𝑩. (4.2.6)
ª

« 𝑗=1 ¬
We avail ourselves of the results about the Hermite functions in the Appendix
to this chapter. Suppose 𝑚 = (𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z+𝑛 . Just reasoning formally we derive
from (4.A.7) that if
∑︁
𝒇 = H𝑚1 (𝑥 1 ) · · · H𝑚𝑛 (𝑥 𝑛 ) 𝒄 𝑚 , 𝒄 𝑚 ∈ C𝑑 ,
𝑚∈Z+𝑛

then

∑︁ ∑︁ 𝑛
𝑷◦ 𝒇 = H𝑚1 (𝑥1 ) · · · H𝑚𝑛 (𝑥 𝑛 ) ­ 𝑩 − 2 𝑚 𝑗 𝜒 𝑗 𝑰 𝑑 ® H𝑚1 (𝑥1 ) · · · H𝑚𝑛 (𝑥 𝑛 ) 𝒄 𝑚 .
© ª
𝑚∈Z+𝑚 « 𝑗=1 ¬

We deduce from Theorem 4.A.6 that the operator 𝑷◦ on S R𝑛 ; C𝑑 is conjugate to
the operator on S Z+𝑛 ; C𝑑 S Z+𝑛 𝑑 (cf. Definition 4.A.5),


 𝑛
∑︁ 


{𝒄 𝑚 } 𝑚∈Z+𝑛 ↦→ 𝑩𝒄 𝑚 − 2 𝑚 𝑗 𝜒 𝑗 𝒄𝑚 , (4.2.7)
 
 𝑗=1  𝑚∈Z+𝑛

under the isomorphism

∑︁
S Z+𝑛 ; C𝑑 ∋ {𝒄 𝑚 } 𝑚∈Z+𝑛 ↦→ 𝒄 𝑚 H𝑚1 (𝑥1 ) · · · H𝑚𝑛 (𝑥 𝑛 ) ∈ S R𝑛 ; C𝑑 (4.2.8)
𝑚∈Z+𝑛

[cf. (4.A.9)]. The fact that the spectrum of 𝑩 is finite and independent of 𝑚 implies
that (4.2.7) is a Fréchet space automorphism of S Z+𝑛 ; C𝑑 if and only if (b+ ) holds.□

Remark 4.2.2 Theorem 4.2.1 extends to complex matrices 𝑨 that are self-adjoint
positive definite. The proof in the complex case is more complicated due to the
impossibility, in general, of diagonalizing 𝑨 by a linear change of the real variables
natural, and easy, extension to complex matrices 𝑨 and functions in
𝑥 𝑗 . The more
S R𝑛 ; C𝑑 is to operators in R2𝑛 of the kind
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 83

𝑛
©∑︁ 𝜕 𝜕
𝑷◦ (𝑥, D 𝑥 ) = 𝑰 𝑑 ­ − 𝑧¯ 𝑗 + 𝑧 𝑗 ® + 𝑩. (4.2.9)
ª
𝑎 𝑗,𝑘
𝜕𝑧 𝑗 𝜕 ¯
𝑧 𝑗
« 𝑗=1 ¬
To deal with (4.2.9) we can diagonalize the matrix 𝑨 by a unitarylineartransforma-
tion of C𝑛 and then rely on the complex analogues, based on exp − |𝑧| 2 (𝑧 ∈ C), of
the real Hermite functions. The properties of the “complex Hermite functions” that
we need are direct consequences of those of the real Hermite functions.

Remark 4.2.3 The properties in Theorem 4.2.1 have an important stability feature:
(b+ ) remains true if we replace the matrices 𝑨 and 𝑩 by matrices sufficiently close
to them. This stability is exploited in the analysis of related differential operators.

Remark 4.2.4 The fact that S R𝑛 ; C𝑑 and S ′ R𝑛 ; C𝑑 are nuclear spaces (see

[Grothendieck, 1952], [Treves, 1967]) implies that the space of (continuous linear)
endomorphisms of either one of them can be identifiedwith the space of matrix-
⊗S ′ R𝑛𝑦 ; C𝑑 ; 𝑲 (𝑥, 𝑦) defines

valued kernel distributions 𝑲 (𝑥, 𝑦) ∈ S R𝑛𝑥 ; C𝑑 b
the maps 𝝋 → ⟨𝑲 (𝑥, 𝑦) , 𝝋 (𝑦)⟩ and u ↦→ ⟨u (𝑥) , 𝑲 (𝑥, 𝑦)⟩. Theorem 4.2.1 can be
rephrased as follows: Property (b) is equivalent
to the fact that there is a kernel
⊗ S ′ R𝑛𝑦 ; C𝑑 such that 𝑷◦ (𝑥, D 𝑥 ) 𝑬 (𝑥, 𝑦) =

distribution 𝑬 (𝑥, 𝑦) ∈ S R𝑛𝑥 ; C𝑑 b
𝑰 𝑑 𝛿 (𝑥 − 𝑦). In connection with this result note that the transpose of 𝑷◦ , 𝑷⊤
◦ , is of the
same type as 𝑷◦ , the only difference being that 𝑩 must be replaced by its transpose.
This implies that 𝑷◦ (𝑥, D 𝑥 ) also has a left-inverse, namely the transpose of the
′ ′

operator defined by the kernel distribution 𝑬 (𝑥, 𝑦) ∈ S R 𝑥 ; C b 𝑛 𝑑 ⊗S R𝑛𝑦 ; C𝑑
such that 𝑷◦ (𝑥, D 𝑥 ) ⊤ 𝑬 (𝑥, 𝑦) = 𝑰 𝑑 𝛿 (𝑥 − 𝑦).

Remark 4.2.5 In connection with the preceding remark we point out that the kernel
distribution 𝑬 (𝑥, 𝑦) such that 𝑷◦ (𝑥, D 𝑥 ) 𝑬 (𝑥, 𝑦) = 𝑰 𝑑 𝛿 (𝑥 − 𝑦) can be seen as a
2
holomorphic function of 𝑩 (identified with a point in C𝑑 ) provided the spectrum of
Í𝑛
𝑩 stays away from the lattice of real points 2 𝑗=1 𝑚 𝑗 𝜒 𝑗 , 𝑚 = (𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z𝑛 .
This can be checked by looking at the resolventÍof the map (4.2.7). We do not discuss
the behavior of 𝑬 (𝑥, 𝑦) at the critical points 2 𝑛𝑗=1 𝑚 𝑗 𝜒 𝑗 ; obviously, the dimensions

of the kernel and cokernel of the operator 𝑷◦ (𝑥, D 𝑥 ) : S R𝑛𝑥 ; C𝑑 ←↪ cease to be
zero.

4.2.2 Statement of the main theorem. Necessity of condition

We go back to (4.2.2) and state the main of result of this section:

Theorem 4.2.6 Suppose that the matrix 𝑨 is symmetric positive definite. For the
differential operator 𝑷 in (4.2.2) to be analytic (or C ∞ ) hypoelliptic in any open
subset of R𝑛 intersecting the 𝑦-axis it is necessary and sufficient that the following
condition be satisfied:
84 4 Analyticity of Solutions of Linear PDEs. Basic Results
Í𝑛
(b) No eigenvalue of the matrix 𝑩 is of the form 2 𝑗=1 𝑚 𝑗 𝜒 𝑗 with (𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z𝑛 .
The necessity of (b) is a direct consequence of Theorem 4.2.1. We formalize it
in the following
Proposition 4.2.7 Suppose that the matrix 𝑨 is symmetric positive definite. If the
matrix-valued differential operator 𝑷 in (4.2.2) is analytic (or C ∞ ) hypoelliptic in
any open subset of R𝑛 intersecting the 𝑦-axis then Condition (b) holds.
Proof We carry out a Fourier transform with respect to 𝑦; (4.2.2) is transformed into
𝑛
∑︁ 𝜕 𝜕
𝑷 (𝑥, D 𝑥 , 𝜂) = 𝑰 𝑑 𝜒𝑗 − 𝑥 𝑗𝜂 + 𝑥 𝑗 𝜂 + 𝜂𝑩
𝑗=1
𝜕𝑥 𝑗 𝜕𝑥 𝑗
𝑛
!
∑︁ 𝜕2 2 2
= 𝑰𝑑 𝜒𝑗 − 𝑥 𝑗 𝜂 + 𝜂 (𝑩 + (tr 𝑨) 𝑰 𝑑 )
𝑗=1
𝜕𝑥 2𝑗
Í Í
where tr 𝑨 = 𝑛𝑗=1 𝜒 𝑗 . Suppose 𝑩 has an eigenvalue 𝜒 = 2 𝑛𝑗=1 𝑚 𝑗 𝜒 𝑗 with
(𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z+𝑛 and let v ∈ C𝑑 \ {0} be such that 𝑩v = 𝜒v. By Theorem 4.2.1 the
function b 𝝋 (𝑥) = H𝑚1 (𝑥1 ) · · · H𝑚𝑛 (𝑥 𝑛 ) 𝒗 satisfies the equation
𝑛
!
∑︁ 𝜕2b
𝝋 2
𝜒𝑗 𝝋 + 𝑩b
− 𝑥 𝑗b 𝝋 + (tr 𝑨) b𝝋 = 0.
𝑗=1
𝜕𝑥 2𝑗

The function
∫ ∞
1 1 d𝜂
𝒖 (𝑥, 𝑦) = e𝑖𝑦 𝜂 b
𝝋 𝜂2𝑥 (𝜅 > 0)
2𝜋 0 1 + 𝜂 𝑛+𝜅

satisfies 𝑷 𝑥, D 𝑥 , D 𝑦 𝒖 (𝑥, 𝑦) = 0. From Lemma 4.A.3 in the Appendix we know
that there is an 𝛼 = (𝛼1 , ..., 𝛼𝑛 ) with 𝛼 𝑗 = 0 or 1 for each 𝑗, such that 𝜕 𝛼 b
𝝋 (0) ≠ 0.
The function of 𝑦,
∞ 1
𝜂 2 | 𝛼 | d𝜂
∫
1
𝜕𝑥𝛼 𝒖 (0, 𝑦) = e𝑖𝑦 𝜂 𝜕 𝛼 b
𝝋 (0) ,
2𝜋 0 1 + 𝜂 𝑛+𝜅

is not C ∞ .
The Ísame reasoning but with 𝜂 ∈ (−∞, 0) proves that 𝑩 has no eigenvalue
𝜒 = −2 𝑛𝑗=1 𝑚 𝑗 𝜒 𝑗 with (𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z+𝑛 . □
To prove that (b) is sufficient for 𝑷 to be analytic hypoelliptic we shall construct
a fundamental solution of 𝑷 (Definition 2.4.1). This is made easier by the additional
assumption that 𝑩 is semisimple (i.e., diagonalizable): it enables us to select 𝚪 ∈
GL(𝑑, C) such that 𝚲 = 𝚪𝑩𝚪−1 is a diagonal matrix with entries 𝜆1 , ..., 𝜆 𝑑 . The
system
𝑛
∑︁
𝚪𝑷𝚪−1 = 𝑰 𝑑 𝜒 𝑗 𝑍 𝑗 𝑍 ′𝑗 + 𝚲D 𝑦
𝑗=1
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 85

is completely uncoupled:

𝜒 𝑗 𝑍 𝑗 𝑍 ′𝑗 𝑢 1 + 𝜆1 D 𝑦 𝑢 1
Í𝑛
𝑢1 𝑗=1
−1 ­ .. ® ..
© ª © ª
𝚪𝑷𝚪 ­ . ® = ­­ .
®.
®
Í𝑛 ′
« 𝑢𝑑 ¬ « 𝑗=1 𝜒 𝑗 𝑍 𝑗 𝑍 𝑗 𝑢 𝑛 + 𝜆 𝑑 D 𝑦 𝑢 𝑑 ¬

It suffices, then, to deal with each scalar operator 𝐿 𝑘 = 𝑛𝑗=1 𝜒 𝑗 𝑍 𝑗 𝑍 ′𝑗 + 𝜆 𝑘 D 𝑦 .

Í

If 𝑩 has a nilpotent part one can select 𝚪 ∈ GL(𝑑, C) such that 𝚲 = 𝚪𝑩𝚪−1 is
lower diagonal and then solve 𝚪𝑷𝚪−1 𝒖 = 𝒇 by solving a system of equations of the
type
𝑑−1
∑︁
𝐿 1 𝑢 1 = 𝑓1 , 𝐿 2 𝑢 2 = 𝑓2 + 𝑐 2,1 𝑢 1 , ..., 𝐿 𝑑 𝑢 𝑑 = 𝑓 𝑑 + 𝑐 𝑑, 𝑗 𝑢 𝑗
𝑗=1

with the appropriate coefficients 𝑐 𝑗,𝑘 , 𝑘 < 𝑗. We leave the details as an exercise;
from now on we limit ourselves to the case of B diagonal and therefore focus our
attention on a scalar operator.

4.2.3 Construction of a fundamental solution

We construct a fundamental solution for a scalar differential operator

𝑛
∑︁
𝑃 𝑥, D 𝑥 , D 𝑦 = 𝜒 𝑗 D2𝑥 𝑗 + 𝑥 2𝑗 D2𝑦 + 𝜆D 𝑦 . (4.2.10)
𝑗=1

To this end we seek a kernel distribution 𝐸 (𝑥, 𝑥 ′, 𝑦) satisfying the equation

𝑃 𝑥, D 𝑥 , D 𝑦 𝐸 = 𝛿 (𝑥 − 𝑥 ′) ⊗ 𝛿 (𝑦) ;

(4.2.11)

then 𝐸 (𝑥, 𝑥 ′, 𝑦 − 𝑦 ′) will be a fundamental solution of (4.2.10). For the time being
the notation 𝐸 ignores the dependence on 𝜆 but the latter will become essential
further on. We construct 𝐸 through the initial value problem for the associated “heat
equation”:

𝜕𝑡 𝐹 + 𝑃 𝑥, D 𝑥 , D 𝑦 𝐹 = 0, (4.2.12)
′
𝐹 | 𝑡=0 = 𝛿 (𝑥 − 𝑥 ) ⊗ 𝛿 (𝑦) .

Proposition 4.2.9 below allows us to differentiate the following distribution under

the integral sign with respect to (𝑥, 𝑦)
∫ ∞
′
𝐸 (𝑥, 𝑥 , 𝑦) = 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) d𝑡 (4.2.13)
0

and thus verify that 𝐸 (𝑥, 𝑥 ′, 𝑦) indeed satisfies (4.2.11):

86 4 Analyticity of Solutions of Linear PDEs. Basic Results
∫ ∞
𝜕𝐹
𝑃 𝑥, D 𝑥 , D 𝑦 𝐸 (𝑥, 𝑥 ′, 𝑦) = − (𝑡, 𝑥, 𝑥 ′, 𝑦) d𝑡,

(4.2.14)
0 𝜕𝑡
𝐹 (0, 𝑥, 𝑥 ′, 𝑦) = 𝛿 (𝑥 − 𝑥 ′) ⊗ 𝛿 (𝑦) .

Definition 4.2.8 We shall say that 𝜆 ∈ C is basal for 𝑨 if either |Re 𝜆| < tr 𝑨 or else
|Re 𝜆| = tr 𝑨, Im 𝜆 ≠ 0.

Proposition 4.2.9 If 𝜆 ∈ C is basal for 𝑨 then (4.2.12) admits a solution 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦)

which is a C ∞ function of 𝑡 ≥ 0 valued in D ′ R𝑛𝑥 × R𝑛𝑥′ × R 𝑦 .

Proof Fourier transformation with respect to 𝑦 transform (4.2.12) into

𝑛
!
∑︁ 𝜕 2𝐹
b
b− 2 2
𝜕𝑡 𝐹 𝜒𝑗 − 𝜂 𝑥𝑗𝐹b + 𝜆𝜂 𝐹
b = 0, (4.2.15)
𝑗=1
𝜕𝑥 2𝑗

𝐹
b = 𝛿 (𝑥 − 𝑥 ′) ;
𝑡=0

then ∫ +∞
′ 1 b (𝑡, 𝑥, 𝑥 ′, 𝜂) d𝜂.
𝐹 (𝑡, 𝑥, 𝑥 , 𝑦) = e𝑖𝑦 𝜂 𝐹
2𝜋 −∞
In (4.2.15) we make the substitution

b (𝑡, 𝑥, 𝑥 ′, 𝜂) = 𝑢 (𝑡, 𝑥, 𝑥 ′, 𝜂) exp 1 |𝜂| |𝑥| 2 − |𝑥 ′ | 2 − 𝜆𝜂𝑡 .

𝐹 (4.2.16)
2

Then 𝑢 must satisfy

𝑛
!
𝜕𝑢 ∑︁ 𝜕2𝑢 𝜕𝑢
− 𝜒𝑗 + 2 |𝜂| 𝑥 𝑗 + |𝜂| 𝑢 = 0,
𝜕𝑡 𝑗=1
𝜕𝑥 2𝑗 𝜕𝑥 𝑗
𝑢| 𝑡=0 = 𝛿 (𝑥 − 𝑥 ′) .

𝜕𝑢
We carry out a Fourier transformation with respect to 𝑥; since 𝑥š 𝑢 −𝜉 𝑗 𝜕𝜕b
𝑗 𝜕𝑥 𝑗 (𝜉) = −b
𝑢
𝜉𝑗
we get
𝑛
𝜕b
𝑢 ∑︁ 2 𝜕b
𝑢
+ 𝑢 + 2 |𝜂| 𝜉 𝑗
𝜒𝑗 𝜉 𝑗 b + (tr 𝑨) |𝜂| b
𝑢 = 0. (4.2.17)
𝜕𝑡 𝑗=1 𝜕𝜉 𝑗

𝑢 | 𝑡=0 =
Straightforward calculation shows that the solution of (4.2.17) satisfying b
′
e−𝑖 𝜉 ·𝑥 is given by

∑︁ 𝑛 ∑︁ 1 − e−4𝜒 𝑗 𝑡 | 𝜂 | ª
𝑛
𝑢 (𝑡, 𝜉, 𝑥 ′, 𝜂) = exp ­− (tr 𝑨) 𝑡 |𝜂| − 𝑖 e−2𝜒 𝑗 𝑡 | 𝜂 | 𝑥 ′𝑗 𝜉 𝑗 − 𝜉 2𝑗 ® .
©
b
𝑗=1 𝑗=1
4 |𝜂|
« ¬
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 87

We derive

𝑢 (𝑡, 𝑥, 𝑥 ′, 𝜂)
1 − e−4𝜒 𝑗 𝑡 | 𝜂 | 2 d𝜉
∫ 𝑛
∑︁
−(tr 𝑨)𝑡 | 𝜂 |
=e exp 𝑖 𝑥𝑗 − e−2𝜒 𝑗 𝑡 | 𝜂 | 𝑥 ′𝑗 𝜉𝑗 − 𝜉𝑗 ,
R𝑛 𝑗=1
4 |𝜂| (2𝜋) 𝑛

whence
∫
b (𝑡, 𝑥, 𝑥 ′, 𝜂) = e−(tr 𝑨+𝜃𝜆)𝑡 | 𝜂 |− 21 𝑄 (𝑡 | 𝜂 |, 𝑥, 𝑥′ ) | 𝜂 | ′ d𝜉
𝐹 e−Ψ(𝑡 | 𝜂 |, 𝑥, 𝑥 , 𝜉 ) , (4.2.18)
R𝑛 (2𝜋) 𝑛

where 𝜃 = 𝜂/|𝜂|(= 1 or −1) and

2
𝑛
∑︁ e 𝜒 𝑗 𝑠 𝑥 𝑗 − e−𝜒 𝑗 𝑠 𝑥 ′𝑗
𝑄 (𝑠, 𝑥, 𝑥 ′) = |𝑥 ′ | 2 − |𝑥| 2 + , (4.2.19)
𝑗=1
sinh 2𝜒 𝑗 𝑠
!2
𝑛
∑︁ 1 − e−4𝜒 𝑗 𝑠 𝑥 𝑗 − e−2𝜒 𝑗 𝑠 𝑥 ′𝑗
Ψ (𝑠, 𝑥, 𝑥 ′, 𝜉) = 𝜉 𝑗 − 2𝑖 |𝜂| .
𝑗=1
4 |𝜂| 1 − e−4𝜒 𝑗 𝑠

The Cauchy Integral Theorem implies

©∑︁ 1 − e−4𝜒 𝑗 𝑠 2 ª
∫ ∫ 𝑛
′
e−Ψ(𝑡 | 𝜂 |, 𝑥, 𝑥 , 𝜉 ) d𝜉 = exp ­ 𝜉 𝑗 ® d𝜉
R𝑛 R𝑛 𝑗=1
4 |𝜂|
« ¬
𝑛 −4𝜒 | |
− 12
Ö 1−e 𝑗 𝑡 𝜂
= 2𝑛 𝜋 𝑛/2
𝑗=1
|𝜂|

whence, by (4.2.18),
− 1
1 − e−4𝜒 𝑗 𝑡 | 𝜂 | 2
𝑛
b (𝑡, 𝑥, 𝑥 ′, 𝜂) = 𝜋 −𝑛/2 e−(tr 𝑨+𝜃𝜆)𝑡 | 𝜂 |− 21 𝑄 (𝑡 | 𝜂 |, 𝑥, 𝑥′ ) | 𝜂 |
Ö
𝐹 (4.2.20)
𝑗=1
|𝜂|

and
+∞
|𝜂| 𝑛/2 d𝜂
∫
1 1 ′
𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) = e𝑖𝑦 𝜂−(tr 𝑨+𝜃𝜆)𝑡 | 𝜂 |− 2 𝑄 (𝑡 | 𝜂 |, 𝑥, 𝑥 ) | 𝜂 | 𝑛 √ .
2𝜋 1+𝑛/2 −∞ Ö
1− e−4𝜒 𝑗 𝑡 | 𝜂 |
𝑗=1
∫ (4.2.21)
Now let 𝜑 ∈ Cc∞ R𝑛𝑥 × R 𝑦 be arbitrary and 𝜑
b (𝑥, 𝜂) = R e−𝑖 𝜂 𝑦 𝜑 (𝑥, 𝑦) d𝑦. By the

Plancherel identity we have
88 4 Analyticity of Solutions of Linear PDEs. Basic Results
∫
𝜑 (𝑥, 𝑦) 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦)d𝑥d𝑦 (4.2.22)
∫
1 b (𝑡, 𝑥, 𝑥 ′, 𝜂)d𝑥d𝜂
= b (𝑥, 𝜂) 𝐹
𝜑
2𝜋
+∞ ∫
|𝜂| 𝑛/2 d𝑥d𝜂
∫
1 1 ′
= e− ( tr 𝑨+𝜃𝜆) 𝑡 | 𝜂 |− 2 𝑄 (𝑡 | 𝜂 |, 𝑥, 𝑥 ) | 𝜂 | 𝜑
b (𝑥, 𝜂) 𝑛 √ .
(2𝜋) 1+𝑛/2 −∞ R𝑛 Ö
−4𝜒 | |
1−e 𝑗 𝑡 𝜂

𝑗=1

We use the fact that, for arbitrary real numbers 𝑎, 𝑏, 𝑠 ≠ 0,

2
(e𝑠 𝑎 − e−𝑠 𝑏) 2

1 𝑏
𝑏2 − 𝑎2 + = 𝑎− + 𝑏 2 tanh (2𝑠)
sinh (2𝑠) tanh (2𝑠) cosh (2𝑠)

to obtain
!2
𝑛
∑︁ 𝑥 ′𝑗
′
coth 2𝜒 𝑗 𝑡 |𝜂| + 𝑄 1 (𝑡 |𝜂| , 𝑥 ′) ,

𝑄 (𝑡 |𝜂| , 𝑥, 𝑥 ) = 𝑥𝑗 −
𝑗=1
cosh 2𝜒 𝑗 𝑡 |𝜂|
(4.2.23)
𝑛
∑︁
𝑄 1 (𝑡 |𝜂| , 𝑥 ′) = 𝑥 ′2

𝑗 tanh 2𝜒 𝑗 𝑡 |𝜂| .
𝑗=1

˜ 𝑥 ′), with
The change of variables 𝑥 = ℎ (𝑡 |𝜂| , 𝑥,

˜ 𝑥 ′) = ℎ 𝑗 2𝜒 𝑗 𝑡 |𝜂| , 𝑥˜ 𝑗 , 𝑥 ′𝑗
ℎ (𝑡 |𝜂| , 𝑥, , (4.2.24)
𝑗=1,...,𝑛
√︃ 𝑥 ′𝑗
ℎ 𝑗 𝑠 𝑗 , 𝑥˜ 𝑗 , 𝑥 ′𝑗 = tanh 𝑠 𝑗 𝑥˜ 𝑗 + , 𝑗 = 1, ..., 𝑛,
cosh 𝑠 𝑗

yields
˜ 2 + 𝑄 1 (𝑡 |𝜂| , 𝑥 ′) .
𝑄 (𝑡 |𝜂| , 𝑥, 𝑥 ′) = | 𝑥|
Also,
𝑛 1/2
d𝑥 Ö tanh 2𝜒 𝑗 𝑡 |𝜂|
= d𝑥˜
𝑛 √
Ö 1 − e−4𝜒 𝑗 𝑡 | 𝜂 |
1− e−4𝜒 𝑗 𝑡 | 𝜂 | 𝑗=1

𝑗=1
𝑛
Ö 1
= √ d𝑥.
˜
𝑗=1 1 + e−4𝜒 𝑗 𝑡 | 𝜂 |
√︁
After the change of variable |𝜂| 𝑥˜ ⇝ 𝑥˜ in (4.2.22) we obtain
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 89
∫
2𝜋 1+𝑛/2 𝜑 (𝑥, 𝑦) 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦)d𝑥d𝑦 (4.2.25)
∫ +∞ ∫ √︁ e− 21 | 𝑥˜ |2 − 12 𝑄1 (𝑡 | 𝜂 |, 𝑥′ ) | 𝜂 |
−(tr 𝑨)𝑡 | 𝜂 |−𝜆𝑡 𝜂 ′
= e b ℎ 𝑡 |𝜂| , 𝑥/
𝜑 ˜ |𝜂|, 𝑥 , 𝜂 𝑛 √ d𝑥d𝜂
˜
−∞ R𝑛 Ö
−4𝜒 | |
1+e 𝑗 𝑡 𝜂

𝑗=1

(recall that 𝜃 = 𝜂/|𝜂| = 1 or −1). Since 𝜆 is basal for 𝑨 we have (tr 𝑨) 𝑡 |𝜂| +
(Re 𝜆) 𝑡𝜂 ≥ 0 for all 𝑡 ≥ 0, 𝜂 ∈ R, and by (4.2.23)–(4.2.24),

0 ≤ 𝑄 1 (𝑡 |𝜂| , 𝑥 ′) ≤ |𝑥 ′ | 2 ,
lim𝑄 1 (𝑡 |𝜂| , 𝑥 ′) = 0,
𝑡↘0
√︁
˜ |𝜂|, 𝑥 ′ = 𝑥 ′, 𝑗 = 1, ..., 𝑛.
lim ℎ 𝑗 2𝜒 𝑗 𝑡 |𝜂| , 𝑥/
𝑡↘0

The Lebesgue Dominated Convergence Theorem implies that (4.2.25) converges, as

𝑡 ↘ 0, to ∫ ∫ +∞ 2
1
b (𝑥 ′, 𝜂) e− 2 | 𝑥˜ | d𝜂d𝑥˜ = 2𝜋 1+𝑛/2 𝜑 (𝑥 ′, 0) .
𝜑
R𝑛 −∞
′
that 𝐹 (𝑡, 𝑥, 𝑥 , 𝑦) is a′ continuous function of 𝑡 ≥ 0
It follows directly from (4.2.25)
′
valued in D R 𝑥 × R 𝑥′ × R 𝑦 , converging to 𝛿 (𝑥 − 𝑥 ) ⊗ 𝛿 (𝑦) as 𝑡 ↘ 0.
𝑛 𝑛

Since (4.2.12) implies, for every 𝑘 = 1, 2, ....,

𝜕𝑡 𝜕𝑡𝑘 𝐹 = −𝑃 𝑥, D 𝑥 , D 𝑦 𝜕𝑡𝑘 𝐹 ,
𝜕𝑡𝑘 𝐹 𝑡=0 = (−1) 𝑘 𝑃 𝑘 𝑥, D 𝑥 , D 𝑦 (𝛿 (𝑥 − 𝑥 ′) ⊗ 𝛿 (𝑦)) ,

we conclude that 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) is a C ∞ function of 𝑡 ≥ 0 valued in D ′ R𝑛𝑥 × R𝑛𝑥′ × R 𝑦 ,

satisfying (4.2.12). □

Proposition 4.2.10 Suppose |Re 𝜆| < tr 𝑨 and let 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) be the solution of
(4.2.12) in Proposition 4.2.9. If 𝑡 ↗ +∞ then 𝑡 1+𝑛/2 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) converges to the
constant
− 1
+∞
1 − e−4𝜒 𝑗 | 𝜂 | 2
∫ 𝑛
Ö
𝐶 ( 𝑨, 𝜆) = (2𝜋) −1−𝑛/2 e− ( tr 𝑨+𝜃𝜆) | 𝜂 | d𝜂 (4.2.26)
−∞ 𝑗=1
|𝜂|

while 𝑡 2+𝑛/2 𝜕𝐹 ′ 𝑛

𝜕𝑡 (𝑡, 𝑥, 𝑥 , 𝑦) converges to − 1 + 2 𝐶 ( 𝑨, 𝜆).

Proof We keep the notation of the proof of Proposition 4.2.9. The change of variable
𝑡𝜂 ⇝ 𝜂 yields
!2
𝑛
∑︁ 𝑥 ′𝑗
′
coth 2𝜒 𝑗 |𝜂| + 𝑥 ′2

𝑄 (|𝜂| , 𝑥, 𝑥 ) = 𝑥𝑗 − 𝑗 tanh 2𝜒 𝑗 |𝜂|
𝑗=1
cosh 2𝜒 𝑗 |𝜂|
90 4 Analyticity of Solutions of Linear PDEs. Basic Results

and transforms (4.2.22) into

∫
𝑡 1+𝑛/2 𝜑 (𝑥, 𝑦) 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦)d𝑥d𝑦 (4.2.27)
− 1
+∞ ∫
1 − e−4𝜒 𝑗 | 𝜂 | 2
∫ 𝑛
− ( tr 𝑨+𝜃𝜆) | 𝜂 |− 2𝑡1 𝑄 ( | 𝜂 |, 𝑥, 𝑥 ′ ) | 𝜂 |
Ö d𝑥d𝜂
= e b (𝑥, 𝜂/𝑡)
𝜑 .
−∞ R𝑛 𝑗=1
|𝜂| (2𝜋) 1+𝑛/2

As 𝑡 ↗ +∞ this integral converges to

∫
𝐶 ( 𝑨, 𝜆) 𝜑 (𝑥, 𝑦) d𝑥d𝑦,
R𝑛+1

where 𝐶 ( 𝑨, 𝜆) is the constant (4.2.26). Moreover, by differentiating (4.2.27) with

respect to 𝑡 we see that
∫
(2𝜋) 1+𝑛/2 𝑡 2+𝑛/2 𝜑 (𝑥, 𝑦) 𝜕𝑡 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦)d𝑥d𝑦 (4.2.28)

𝑛 +∞ |𝜂| 𝑛/2 d𝑥d𝜂

∫ ∫
1 ′
=− 1+ e− ( tr 𝑨+𝜃𝜆) | 𝜂 |− 2𝑡 𝑄 ( | 𝜂 |, 𝑥, 𝑥 ) | 𝜂 | 𝜑
b (𝑥, 𝜂/𝑡) 𝑛
2 −∞ R𝑛 Ö 1
1 − e−4𝜒 𝑗 | 𝜂 | 2
𝑗=1
∫ +∞ ∫
1 1 ′
+ e− ( tr 𝑨+𝜃𝜆) | 𝜂 |− 2𝑡 𝑄 ( | 𝜂 |, 𝑥, 𝑥 ) | 𝜂 |
2𝑡 −∞ R𝑛
|𝜂| 𝑛/2 d𝑥d𝜂
×𝑄 (|𝜂| , 𝑥, 𝑥 ′) |𝜂| 𝜑
b (𝑥, 𝜂/𝑡) 𝑛 .
Ö 12
1 − e−4𝜒 𝑗 | 𝜂 |
𝑗=1

Since
1 1
lim 𝑄 (|𝜂| , 𝑥, 𝑥 ′) |𝜂| exp 𝑄 (|𝜂| , 𝑥, 𝑥 ′) |𝜂| = 0
𝑡↗+∞ 𝑡 2𝑡
the Lebesgue Dominated Convergence Theorem implies that, if tr 𝑨 + 𝜃 Re 𝜆 > 0
and 𝑡 ↗ +∞, then (4.2.28) converges to
∫
𝑛
− 1+ 𝐶 ( 𝑨, 𝜆) 𝜑 (𝑥, 𝑦) d𝑥d𝑦. □
2 R𝑛+1

Proposition 4.2.11 Suppose |Re 𝜆| = tr 𝑨, Im 𝜆 ≠ 0, and let 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) be the so-

lution of (4.2.12) in Proposition 4.2.9. If 𝑡 ↗ +∞ then 𝑡 1+𝑛/2 𝐹 (𝑡, 𝑥, 𝑥 ′, 𝑦) converges
to a constant 𝐶 ♮ ( 𝑨, 𝜆 ′) while 𝑡 2+𝑛/2 𝜕𝐹
𝜕𝑡 (𝑡, 𝑥, 𝑥 ′ , 𝑦) converges to − 1 + 𝑛 𝐶 ♮ ( 𝑨, 𝜆 ′ ).
2

Proof Let us suppose Re 𝜆 = − tr 𝑨 and 𝜆 ′ = Im 𝜆 ≠ 0. The integral on the right-

hand side of (4.2.27) is the sum of two integrals:
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 91
− 1
+∞ ∫
1 − e−4𝜒 𝑗 𝜂 2
∫ 𝑛
−𝑖𝜆′ 𝜂− 2𝑡1 𝑄 ( 𝜂, 𝑥, 𝑥 ′ ) 𝜂
Ö d𝑥d𝜂
𝐽1 = e b (𝑥, 𝜂/𝑡)
𝜑
0 R𝑛 𝑗=1
𝜂 (2𝜋) 1+𝑛/2
(4.2.29)
and
− 1
+∞ ∫
1 − e−4𝜒 𝑗 𝜂 2
∫ 𝑛
−(tr 𝑨) 𝜂+𝑖𝜆′ 𝜂− 2𝑡1 𝑄 ( 𝜂, 𝑥, 𝑥 ′ ) 𝜂
Ö d𝑥d𝜂
𝐽2 = e b (𝑥, 𝜂/𝑡)
𝜑 .
0 R𝑛 𝑗=1
𝜂 (2𝜋) 1+𝑛/2

We deal with 𝐽2 in the same manner as the case |Re 𝜆| < tr 𝑨 was dealt with, in the
proof of Proposition 4.2.10. We conclude that 𝑡 1+𝑛/2 𝐽2 (𝑡, 𝑥, 𝑥 ′, 𝑦) converges to
∫
1
𝐶 ( 𝑨, − tr 𝑨 + 𝑖𝜆 ′) 𝜑 (𝑥, 𝑦) d𝑥d𝑦,
2 R𝑛+1

where 𝐶 ( 𝑨, 𝜆) is the constant (4.2.26), while 𝑡 2+𝑛/2 𝜕𝐽 2 ′

𝜕𝑡 (𝑡, 𝑥, 𝑥 , 𝑦) converges to
∫
1 𝑛 ′
− 1+ 𝐶 ( 𝑨, − tr 𝑨 + 𝑖𝜆 ) 𝜑 (𝑥, 𝑦) d𝑥d𝑦.
2 2 R𝑛+1

We shall therefore focus on (4.2.29). We apply the Cauchy Integral Theorem and
deform the domain of 𝜂 integration in (4.2.29) from R+ to the image Σ+ (𝜆 ′) ⊂ C of
R+ under the map 𝜂 ↦→ 𝜁 = (1 − 𝑖𝜆 ′/|𝜆 ′ |) 𝜂. The Paley–Wiener Theorem implies

|𝜑
b (𝑥, 𝜁/𝑡)| ≲ max |𝜑 (𝑥, 𝑦)| exp (𝐶◦ 𝜂/𝑡) ,

where 𝐶◦ only depends on the size of the (convex hull of) the support of the function
𝑦 ↦−→ 𝜑 (𝑥, 𝑦). We go back to (4.2.19):
2
𝑛
∑︁ e 𝜒 𝑗 𝜁 𝑥 𝑗 − e−𝜒 𝑗 𝜁 𝑥 ′𝑗
𝑄 (𝜁, 𝑥, 𝑥 ′) = |𝑥 ′ | 2 − |𝑥| 2 + .
𝑗=1
sinh 2𝜒 𝑗 𝜁

It is clear that for |Im 𝜁 | = 𝜂 large, |𝑄 (𝜁, 𝑥, 𝑥 ′)| ≲ |𝑥| 2 + |𝑥 ′ | 2 . It is also obvious that,
for 𝜂 large,
−1
1 − e−4𝜒 𝑗 𝜁 2
Ö 𝑛
≲ 𝜂 𝑛\2 .
𝑗=1
𝜁

We reach the conclusion that the absolute value of the integrand in (4.2.29) is
bounded by (a constant times)

𝜂 𝑛\2 exp − |𝜆 ′ | 𝜂 + 𝐶1 𝑡 −1 |𝜂| .

We are therefore allowed to apply the Lebesgue Dominated Convergence Theorem

and state that, as 𝑡 ↗ +∞, (4.2.29) converges to
92 4 Analyticity of Solutions of Linear PDEs. Basic Results
∫
𝐶 ♭ ( 𝑨, 𝜆 ′) 𝜑 (𝑥, 𝑦) d𝑥d𝑦,
R𝑛+1

where
− 1
1 − e−4𝜒 𝑗 𝜁 2
∫ 𝑛
Ö
♭ ′ −𝑖𝜆′ 𝜁
𝐶 ( 𝑨, 𝜆 ) = e d𝜁.
Σ+ (𝜆′ ) 𝑗=1
𝜁

The same reasoning allows us to state that 𝑡 2+𝑛/2 𝜕𝐽1 ′

𝜕𝑡 (𝑡, 𝑥, 𝑥 , 𝑦) converges to
∫
1 𝑛 ♭ ′
− 1+ 𝐶 ( 𝑨, 𝜆 ) 𝜑 (𝑥, 𝑦) d𝑥d𝑦.
2 2 R𝑛+1

The case Re 𝜆 = tr 𝑨 and 𝜆 ′ = Im 𝜆 ≠ 0 is handled in a similar fashion. □

Integration over [0, +∞) of a function valued in a locally convex topological vector
space [such as D ′ R𝑛𝑥 × R𝑛𝑥′ × R 𝑦 ] in which every Cauchy sequence converges is

defined by using Riemann sums.

Corollary 4.2.12 Under the hypotheses of Propositions 4.2.10 and 4.2.11 the inte-
gral (4.2.13) converges in D ′ R𝑛𝑥 × R𝑛𝑥′ × R 𝑦 and satisfies (4.2.14).

4.2.4 The result for basal 𝝀s

To emphasize the dependence on 𝜆 we write 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦) instead of 𝐸 (𝑥, 𝑥 ′, 𝑦);

we derive from (4.2.13) and (4.2.21):

(2𝜋) −1−𝑛/2 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦) (4.2.30)

∫ +∞ ∫ +∞ 𝑛/2
1 ′ |𝜂| d𝜂d𝑡
= e𝑖𝑦 𝜂−(tr 𝑨+𝜃𝜆)𝑡 | 𝜂 |− 2 𝑄 (𝑡 | 𝜂 |, 𝑥, 𝑥 ) | 𝜂 | 𝑛 √ .
𝑡=0 −∞ Ö
1− e−4𝜒 𝑗 𝑡 | 𝜂 |
𝑗=1

Remark 4.2.13 Expressions of the type (4.2.30) are sometimes referred to as

Mehler’s formulas; they are related to the so-called Feynman propagator for
the harmonic oscillator ([Feynman, Hibbs, 1965], [Barone, Boschi-Filho, Farina,
2002]).

Proposition 4.2.14 If 𝜆 is basal for 𝑨 (Definition 4.2.8) the distribution 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦)
is real-analytic in the domain Ω in R2𝑛+1 defined by |𝑦| + |𝑥 − 𝑥 ′ | ≠ 0.

Proof If we take into account, in (4.2.19), the following identity, for arbitrary real
numbers 𝑎, 𝑏, 𝑠 ≠ 0,

(e𝑠 𝑎 − e−𝑠 𝑏) 2 (𝑎 − 𝑏) 2 2
𝑏2 − 𝑎2 + = + 𝑎 + 𝑏 2 tanh 𝑠,
sinh (2𝑠) sinh (2𝑠)
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 93

we see that
2
𝑛
∑︁ 𝑥 𝑗 − 𝑥 ′𝑗
𝑄 (𝑠, 𝑥, 𝑥 ′) = + 𝑥 2𝑗 + 𝑥 ′2

𝑗 tanh 𝜒 𝑗 𝑠 . (4.2.31)
𝑗=1
sinh 2𝜒 𝑗 𝑠

It follows that 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦) is symmetric with respect to 𝑥 and 𝑥 ′ and that we have,
in Ω,
𝑛
∑︁
𝜒 𝑗 D2𝑥 𝑗 𝐸 (𝜆) + 𝑥 2𝑗 D2𝑦 𝐸 (𝜆) + 𝜆D 𝑦 𝐸 (𝜆) = 0, (4.2.32)
𝑗=1
𝑛
∑︁
𝜒 𝑗 D2𝑥′ 𝐸 (𝜆) + 𝑥 ′2 2 (𝜆)
𝑗 D𝑦 𝐸 + 𝜆D 𝑦 𝐸 (𝜆) = 0. (4.2.33)
𝑗
𝑗=1

By adding (4.2.32) and (4.2.33) we obtain an elliptic equation in the subset of Ω

defined by |𝑥| 2 + |𝑥 ′ | 2 ≠ 0; by Theorem 4.1.2 we deduce that 𝐸 (𝜆) is of class C 𝜔
with respect to (𝑥, 𝑥 ′, 𝑦) in said subset.
It remains to study 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦) in a suitably small open ball 𝔅 ⊂ Ω centered at
(0, 0, 𝑦 ◦ ), 𝑦 ◦ ≠ 0; we shall reason in the case 𝑦 ◦ > 0, the analogous reasoning being
valid when 𝑦 ◦ < 0. We point out that it suffices to prove that D 𝑦 𝐸 (𝜆) ∈ C 𝜔 (𝔅).
Indeed, the system (4.2.32)–(4.2.33) is elliptic with respect to (𝑥, 𝑥 ′) and therefore
D 𝑦 𝐸 (𝜆) ∈ C 𝜔 (𝔅) implies that 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦 ◦ ) is of class C 𝜔 with respect to (𝑥, 𝑥 ′);
in turn this implies
∫ 𝑦
𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦) = 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦 ◦ ) + 𝑖 D 𝑦 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦 ′) d𝑦 ′ ∈ C 𝜔 (𝔅) . (4.2.34)
𝑦◦

We shall first deal with the case |Re 𝜆| < tr 𝑨. We introduce the partial derivatives
of 𝐸 (𝜆) of order 𝑘 ≥ 1 with respect to (the complex variable) 𝜆; they satisfy the
equations

𝜕 𝑘 𝐸 (𝜆) 𝑘 (𝜆) 𝜕 𝑘 𝐸 (𝜆) 𝜕 𝑘−1 𝐸 (𝜆)

𝑛
2 2𝜕 𝐸
∑︁
𝜒𝑗 D2𝑥 𝑗 + 𝑥 𝑗 D 𝑦 + 𝜆D 𝑦 = −𝑘D 𝑦 , (4.2.35)
𝑗=1
𝜕𝜆 𝑘 𝜕𝜆 𝑘 𝜕𝜆 𝑘 𝜕𝜆 𝑘−1

𝜕 𝑘 𝐸 (𝜆) 𝑘 (𝜆) 𝜕 𝑘 𝐸 (𝜆) 𝜕 𝑘−1 𝐸 (𝜆)

𝑛
′2 2 𝜕 𝐸
∑︁
𝜒 𝑗 D2𝑥′ 𝑘
+ 𝑥 𝑗 D 𝑦 𝑘
+ 𝜆D 𝑦 𝑘
= −𝑘D 𝑦 . (4.2.36)
𝑗=1
𝑗 𝜕𝜆 𝜕𝜆 𝜕𝜆 𝜕𝜆 𝑘−1

𝜕 𝑘 𝐸 (𝜆) 𝑘−1 (𝜆)

It is clear that if 𝜕𝜆 𝑘
∈ C 𝜔 (𝔅) the same is true of D 𝑦 𝜕 𝜕𝜆𝑘−1
𝐸
and therefore of
𝜕 𝑘−1 𝐸 (𝜆)
𝜕𝜆 𝑘−1
, by (4.2.34); by descending induction on 𝑘 this proves that 𝐸 (𝜆) ∈ C 𝜔 (𝔅).
We derive from (4.2.30), after the change of variables 𝑠 = 𝑡 |𝜂|,
94 4 Analyticity of Solutions of Linear PDEs. Basic Results

𝜕 𝑘 𝐸 (𝜆)
D𝑦 (𝑥, 𝑥 ′, 𝑦)
𝜕𝜆 𝑘
+∞ ∫
𝑠 𝑘 |𝜂| 𝑛/2
∫
1 1 ′
= lim e𝑖𝑦 𝜂−(tr 𝑨+𝜃𝜆)𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) | 𝜂 | 𝜃d𝜂d𝑠
(2𝜋) 1+𝑛/2 𝜀↘0 −∞ 𝑠> 𝜀/ | 𝜂 | Δ (𝑛) (𝑠)

where
𝑛 √︁
Ö
Δ (𝑛) (𝑠) = 1 − e−4𝜒 𝑗 𝑠 .
𝑗=1

Since √
Δ (𝑛) (𝑠) = 2𝑛 det 𝑨𝑠 𝑛/2 (1 + 𝑂 (𝑠)) if 𝑠 ≈ 0,
when 𝑘 > 1 + 𝑛/2 we can let 𝜀 go to zero. We have

𝜕 𝑘 𝐸 (𝜆)
(2𝜋) 1+𝑛/2 D 𝑦 = 𝐽+(𝑘) − 𝐽−(𝑘) ,
𝜕𝜆 𝑘
∫ +∞ ∫ +∞
1 ′ 𝑠 𝑘 𝜂 𝑛/2
𝐽+(𝑘) ′
(𝑥, 𝑥 , 𝑦) = e𝑖𝑦 𝜂−(tr 𝑨+𝜆)𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 (𝑛) d𝜂d𝑠,
0 0 Δ (𝑠)
∫ +∞ ∫ +∞
1 ′ 𝑠 𝑘 𝜂 𝑛/2
𝐽−(𝑘) (𝑥, 𝑥 ′, 𝑦) = e−𝑖𝑦 𝜂−(tr 𝑨−𝜆)𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 (𝑛) d𝜂d𝑠;
0 0 Δ (𝑠)

𝐽±(𝑘) (𝑥, 𝑥 ′, 𝑦) are oscillatory integrals in 𝜂. Dealing with 𝐽+(𝑘) we use the fact that

𝐽+(𝑘) (𝑥, 𝑥 ′, 𝑦)
𝑛 ∫ +∞ ∫ +∞
𝜕2

1 ′ 𝑠 𝑘 𝜂 𝑛/2 d𝜂d𝑠
= 1− e𝑖𝑦 𝜂−(tr 𝑨+𝜆)𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 .
𝜕𝑦 2
𝑛
0 0 1 + 𝜂2 Δ (𝑛) (𝑠)

We deform the domain of 𝜂-integration from (0, +∞) to the image of (0, +∞) under
the map 𝜂 → 𝜂 (1 + 𝑖𝛿) with 0 < 𝛿 ≪ 𝑦 ◦ . We get, by the Cauchy Integral Theorem,
∫ +∞ ∫ +∞
1 ′ 𝑠 𝑘 𝜂 𝑛/2 d𝜂d𝑠
e𝑖𝑦 𝜂−(tr 𝑨+𝜆)𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 𝑛
0 0 1 + 𝜂2 Δ (𝑛) (𝑠)
+∞ ∫ +∞
𝑠 𝑘 (1 + 𝑖𝛿) 𝑛/2 𝜂 𝑛/2 d𝜂d𝑠
∫
1 ′
= e𝑖𝑦 𝜂− 𝛿 𝑦 𝜂−(tr 𝑨+𝜆)𝑠− 2 (1+𝑖 𝛿)𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 𝑛 .
0 0 1 + (1 + 𝑖𝛿) 2 𝜂2 Δ (𝑛) (𝑠)

If tr 𝑨 + Re 𝜆 > 0 it is evident that the integral on the right-hand side converges if we

let (𝑥, 𝑥 ′, 𝑦) vary in a neighborhood 𝑈 C of 𝔅 (contracted as needed) in C2𝑛+1 whose
diameter is small compared to 𝛿 and thus it defines a holomorphic function in 𝑈 C .
𝑘 𝐸 (𝜆)
The analogous argument applies to 𝐽−(𝑘) . This proves that D 𝑦 𝜕 𝜕𝜆 𝑘 ∈ C 𝜔 (𝔅).
′
Lastly, we deal with the cases Re 𝜆 = ± tr 𝑨, 𝜆 = Im 𝜆 ≷ 0. Assume that
Re 𝜆 = − tr 𝑨, 𝜆 ′ > 0; the other cases are dealt with in a similar fashion. The
previous reasoning still works for 𝐽−(𝑘) without modification but now
4.2 Degenerate Elliptic Equations. Influence of Lower Order Terms 95
∫ +∞ ∫ +∞
′ 1 ′ 𝑠 𝑘 𝜂 𝑛/2
𝐽+(𝑘) (𝑥, 𝑥 ′, 𝑦) = e𝑖𝑦 𝜂−𝑖𝜆 𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 d𝜂d𝑠
0 0 Δ (𝑛) (𝑠)

is an oscillatory integral in the (𝜂, 𝑠) variables. We use the fact that

𝐽+(𝑘) (𝑥, 𝑥 ′, 𝑦)
𝑘+1 ∫ +∞ ∫ +∞
𝜕2

′ 1 ′ 𝑠 𝑘 𝜂 𝑛/2 d𝜂d𝑠
= 1 − ′2 e𝑖𝑦 𝜂−𝑖𝜆 𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 𝑘+1 (𝑛)
𝜕𝜆 −∞ 0 1 + 𝑠2 Δ (𝑠)

and the argument in the preceding paragraph can be applied to the integral (oscillatory
solely in the 𝜂 variable)
∫ +∞ ∫ +∞
′ 1 ′ 𝑠 𝑘 𝜂 𝑛/2 d𝜂d𝑠
e𝑖𝑦 𝜂−𝑖𝜆 𝑠− 2 𝑄 (𝑠, 𝑥, 𝑥 ) 𝜂 𝑘+1 (𝑛) ,
−∞ 0 1 + 𝑠2 Δ (𝑠)

to prove that it, and therefore 𝐽+(𝑘) (𝑥, 𝑥 ′, 𝑦), is an analytic function of (𝑥, 𝑥 ′, 𝑦) in
𝑈 C . This completes the proof of Proposition 4.2.11. □

Corollary 4.2.15 If 𝜆 is basal for 𝑨 the differential operator (4.2.10) is analytic as

well as C ∞ hypoelliptic in every open subset of R𝑛+1 .

4.2.5 Completing the proof of Theorem 4.2.6

We are now able to prove the sufficiency of Condition (b) in Theorem 4.2.6. Recall
that the proof has been reduced to proving the analogous result for the scalar operator
(4.2.10), which we rewrite in the form
𝑛
∑︁
𝑃 (𝜆) 𝑥, D 𝑥 , D 𝑦 = − 𝜒 𝑗 𝑍 𝑗 𝑍 ∗𝑗 + (𝜆 − tr 𝑨) D 𝑦

(4.2.37)
𝑗=1

using the notation (4.2.3). By Corollary 4.2.15 we know that 𝑃 (𝜆) 𝑥, D 𝑥 , D 𝑦 is

analytic hypoelliptic if 𝜆 is basal for 𝑨. We must prove that this is also true for all
Í
𝜆 ∈ C, 𝜆 ≠ 2 𝑛𝑘=1 𝑚 𝑘 𝜒𝑘 , whatever (𝑚 1 , ..., 𝑚 𝑛 ) ∈ Z𝑛 .
We shall follow the concatenations approach (see [Treves, 1973]). This is made
easy by the fact that 𝑍 𝑗 , 𝑍 𝑘 = 𝑍 𝑗 , 𝑍 𝑘∗ = 0 if 𝑗 ≠ 𝑘 and both 𝑍 𝑗 and 𝑍 ∗𝑗 commute

h i
with 𝑍 𝑗 , 𝑍 ∗𝑗 = 2D 𝑦 . Formula (4.2.37) implies

𝑍 𝑛 𝑃 (𝜆−𝜒𝑛 ) = 𝑃 (𝜆+𝜒𝑛 ) 𝑍 𝑛 (4.2.38)

𝑍 𝑛∗ 𝑃 (𝜆+𝜒𝑛 ) = 𝑃 (𝜆−𝜒𝑛 ) 𝑍 𝑛∗ (4.2.39)
96 4 Analyticity of Solutions of Linear PDEs. Basic Results

Let Ω be an open subset of R𝑛+1 . We start from the assumption that 𝑃 (𝜆−𝜒𝑛 ) is
analytic hypoelliptic in Ω (which is true when 𝜆 − 𝜒𝑛 is basal for 𝑨). Suppose there
is a 𝑢 ∈ D ′ (Ω), 𝑢 ∉ C 𝜔 (Ω), such that 𝑃 (𝜆+𝜒𝑘 ) 𝑢 ∈ C 𝜔 (Ω). Then, by (4.2.39),
𝑃 (𝜆−𝜒𝑛 ) 𝑍 𝑘∗ 𝑢 ∈ C 𝜔 (Ω) implying 𝑍 𝑛∗ 𝑢 ∈ C 𝜔 (Ω) and, as a consequence,

𝑛−1
∑︁
− 𝜒 𝑗 𝑍 𝑗 𝑍 ∗𝑗 𝑢 + (𝜆 + 𝜒𝑛 − tr 𝑨) D 𝑦 𝑢 = 𝑃 (𝜆+𝜒𝑛 ) 𝑢 + 𝜒𝑛 𝑍 𝑛 𝑍 𝑛∗ 𝑢 ∈ C 𝜔 (Ω) .
𝑗=1

At this point we reason by induction on 𝑛Íand assume that the claim has been proved
up to dimension 𝑛 − 1. Since tr 𝑨 − 𝜒𝑛 = 𝑛−1 𝑗=1 𝜒 𝑗 and Condition (b) implies that 𝜆 is
Í
not of the form 2 𝑛−1 𝑗=1 𝑚 𝜒
𝑗 𝑗 , 𝑚 𝑗 ∈ Z + , we reach the conclusion that 𝑢 is an analytic
function of (𝑥1 , ..., 𝑥 𝑛−1 , 𝑦). Since we also have 𝑍 𝑛∗ 𝑢 ∈ C 𝜔 (Ω) it is not difficult to
deduce that 𝑢 ∈ C 𝜔 (Ω): the Cauchy–Kovalevskaya Theorem implies the existence
of a function 𝑣 ∈ C 𝜔 (𝑈), with 𝑈 ⊂⊂ Ω a sufficiently small domain (otherwise
arbitrary), such that 𝑍 𝑛∗ 𝑣 = 𝑍 𝑛∗ 𝑢. The distribution ℎ = 𝑢 − 𝑣 satisfies 𝑍 𝑛∗ ℎ = 0 and
is analytic with respect to 𝑦. It is an easy application of the FBI transform to prove
that ℎ ∈ C 𝜔 (𝑉), where 𝑉 is any domain such that 𝑉 ⊂⊂ 𝑈. This implies that
𝑢 ∈ C 𝜔 (Ω) and proves that 𝑃 (𝜆+𝜒𝑛 ) is also analytic hypoelliptic in Ω. To move in
the other direction on the 𝜆-axis we rely on (4.2.38). The preceding reasoning, after
substitution of 𝑍 𝑛 for 𝑍 𝑛∗ , proves that the analytic hypoellipticity of 𝑃 (𝜆+𝜒𝑛 ) implies
that of 𝑃 (𝜆−𝜒𝑛 ) .
It remains to prove the result when 𝑛 = 1, i.e, for

𝑃 (𝜆) 𝑥, D 𝑥 , D 𝑦 = −𝜒𝑍 𝑍 ∗ + (𝜆 − 𝜒) D 𝑦

where 𝑍 = D 𝑥 + 𝑖𝑥D 𝑦 with 𝑥 a real variable, 𝜒 > 0. Repeating the argument

for a general 𝑛 ≥ 1 we use the equation 𝑍 ∗ 𝑃 (𝜆+𝜒) = 𝑃 (𝜆−𝜒) 𝑍 ∗ and the analytic
hypoellipticity of 𝑃 (𝜆−𝜒) to derive that

𝑃 (𝜆+𝜒) 𝑢 ∈ C 𝜔 (Ω) =⇒ 𝑍 ∗ 𝑢 ∈ C 𝜔 (Ω) .

In turn this implies

𝜆D 𝑦 𝑢 = 𝑃 (𝜆+𝜒) 𝑢 + 𝑍 𝑍 ∗ 𝑢 ∈ C 𝜔 (Ω)

and therefore, assuming 𝜆 ≠ 0 (which is the case if 𝜆 − 𝜒 is basal for 𝑨 = 𝜒),

𝑃 (𝜆−𝜒) 𝑢 = 𝑃 (𝜆+𝜒) 𝑢 − 2𝜒D 𝑦 𝑢 ∈ C 𝜔 (Ω)

implying 𝑢 ∈ C 𝜔 (Ω). This proves that 𝑃 (𝜆−𝜒+2𝑚𝜒) is analytic hypoelliptic in Ω for

every 𝑚 ∈ Z+ if this is true of 𝑃 (𝜆−𝜒) . To prove the analogous result for the operators
𝑃 (𝜆−𝜒−2𝑚𝜒) we repeat the preceding reasoning with 𝑍 substituted for 𝑍 ∗ , starting
from the equation 𝑃 (𝜆−𝜒) 𝑍 = 𝑍 𝑃 (𝜆−3𝜒) [cf. (4.2.38)].
This completes the proof of the sufficiency of Condition (b) and thus the proof of
Theorem 4.2.6.
4.3 A Generalization of the Harmonic Oscillator 97

Remark 4.2.16 Throughout this subsection analytic hypoelliptic can be replaced by

C ∞ hypoelliptic.

Remark 4.2.17 The existence and semiregularity (Definition 2.3.1) of the funda-
mental solution 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦 − 𝑦 ′) [cf. (4.2.30)] prove the global solvability
of the
inhomogeneous equation, in the following sense: to every 𝑓 ∈ C ∞ R𝑛+1 [resp.,
c
E ′ R𝑛+1 ] there is a 𝑢 ∈ C ∞ R𝑛+1 [resp., D ′ R𝑛+1 ] such that 𝑃 𝑥, D 𝑥 , D 𝑦 𝑢 = 𝑓 .

Exercise 4.2.18 Can we use Hermite functions (see the Appendix to present
chapter) to derive from the existence and properties of the fundamental solu-
tion 𝐸 (𝜆) (𝑥, 𝑥 ′, 𝑦 − 𝑦 ′) that to 𝑛+1 [resp., S ′ R𝑛+1 ] there is a

every 𝑓 ∈ S R
𝑢 ∈ S R𝑛+1 [resp., S ′ R𝑛+1 ] such that 𝑃 (𝜆) 𝑥, D 𝑥 , D 𝑦 𝑢 = 𝑓 ?

4.3 A Generalization of the Harmonic Oscillator

4.3.1 Exploiting a priori estimates

In this subsection we show how a priori estimates can be used to prove the analytic
hypoellipticity of a class of second-order, degenerate elliptic differential operators
whose significant coefficients vanish to order > 2. We content ourselves with ap-
plying the method to a Grushin operator in 2 variables slightly more general than
(4.2.1), specifically the operator

𝐿 𝑝 = D2𝑥 + 𝑥 2 𝑝 D2𝑦 , (4.3.1)

where 𝑝 is an arbitrary positive integer.

Let 𝜑 (𝑥, 𝑦) be a C ∞ function in a slab (𝑥, 𝑦) ∈ R2 ; |𝑥| < 𝑟 ◦ which vanishes

identically if |𝑦| > 𝑟 ◦ (𝑟 ◦ > 0). If 0 < 𝑟 < 𝑟 ◦ integration by parts shows that
∫ 𝑟 ∫ +∞ ∫ +∞
𝜑𝐿 𝑝 𝜑d𝑥d𝑦 = 𝜑 (−𝑟, 𝑦) 𝜑 𝑥 (−𝑟, 𝑦) d𝑦
∫ +∞−𝑟 −∞ ∫−∞𝑟 ∫ +∞
2
− 𝜑 (𝑟, 𝑦) 𝜑 𝑥 (𝑟, 𝑦) d𝑦 + |𝜑 𝑥 | 2 + 𝑥 𝑝 𝜑 𝑦 d𝑥d𝑦.
−∞ −𝑟 −∞

Our proof of the analytic hypoellipticity of the operator (4.3.1) will be based on the
resulting estimate:
∫ 𝑟 ∫ +∞ ∫ 𝑟 ∫ +∞
2
|𝜑 𝑥 | 2 + 𝑥 𝑝 𝜑 𝑦 d𝑥d𝑦 ≤ 𝜑𝐿 𝑝 𝜑d𝑥d𝑦 (4.3.2)
−𝑟 −∞
∫ +∞ ∫ +∞−𝑟 −∞
+ |𝜑 (𝑟, 𝑦)| |𝜑 𝑥 (𝑟, 𝑦)| d𝑦 + |𝜑 (−𝑟, 𝑦)| |𝜑 𝑥 (−𝑟, 𝑦)| d𝑦.
−∞ −∞

Theorem 4.3.1 The operator (4.3.1) is analytic hypoelliptic in R2 .

98 4 Analyticity of Solutions of Linear PDEs. Basic Results

Proof Let Ω be an open subset of the (𝑥, 𝑦)-plane, intersecting the axis 𝑥 = 0
(𝐿 𝑝 is elliptic, hence analytic hypoelliptic, in the region 𝑥 ≠ 0). We deal with a
distribution 𝑢 in Ω such that 𝐿 𝑝 𝑢 ∈ C 𝜔 (Ω). We are going to assume straightaway
that 𝑢 ∈ C ∞ (Ω), i.e., that 𝐿 𝑝 is hypoelliptic (Definition 6.1.1), a result which is a
very special case of the classical theorem in ([Hörmander, 1967]).
We are going to show that 𝑢 is real-analytic in an open square

𝔔𝑟◦ = (𝑥, 𝑦) ∈ R2 ; |𝑥| < 𝑟 ◦ , |𝑦 − 𝑦 ◦ | < 𝑟 ◦ ⊂⊂ Ω.

For convenience we take 𝑦 ◦ = 0. We select 𝑟 ◦ > 0 small enough that the equation
𝐿 𝑝 𝑣 = 𝐿 𝑝 𝑢 has an analytic solution 𝑣 in a ball 𝔔𝑟◦′ ⊂⊂ Ω, 𝑟 ◦′ > 𝑟 ◦ . Since 𝑢 can be
replaced by 𝑢 − 𝑣 there is no loss of generality in assuming that 𝐿 𝑝 𝑢 ≡ 0 in 𝔔𝑟◦′ .
Then we also have 𝐿 𝑝 D 𝑘𝑦 𝑢 ≡ 0 in 𝔔𝑟◦′ whatever 𝑘 ∈ Z+ .
Now let 𝜒𝑟 , 𝜀,2 (𝑦) be an Ehrenpreis cutoff function of the single variable 𝑦 as
defined in (7.1.2) (here 𝑚 = 2); as before, we assume 0 < 2𝜀 < 𝑟 < 𝑟 ◦ . We apply
(4.3.2) to 𝜑 (𝑥, 𝑦) = 𝜒𝑟 , 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 (𝑥, 𝑦); taking into account (7.1.4) and (7.1.5) we
get
∫ 𝑟∫ 𝑟 ∫ ∫
2 1 𝑟 𝑟 2𝑝 2
𝜒𝑟 , 𝜀,2 (𝑦) D 𝑥 D 𝑘𝑦 𝑢 d𝑥d𝑦 + 𝑥 𝜒𝑟 , 𝜀,2 (𝑦) D 𝑘+1
𝑦 𝑢 d𝑥d𝑦
−𝑟 −𝑟 2 −𝑟 −𝑟
∫ 𝑟∫ 𝑟
≤ 𝜒𝑟 , 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 𝐿 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 d𝑥d𝑦 (4.3.3)
−𝑟 −𝑟
∫ 𝑟
+ D 𝑘𝑦 𝑢 (𝑟, 𝑦) D 𝑥 D 𝑘𝑦 𝑢 (𝑟, 𝑦) d𝑦
−𝑟
∫ 𝑟 ∫ 𝑟∫ 𝑟
2
+ D 𝑦 𝑢 (−𝑟, 𝑦) D 𝑥 D 𝑘𝑦 𝑢 (−𝑟, 𝑦) d𝑦 + (4𝐾/𝜀) 2
𝑘
𝑥 2 𝑝 D 𝑘𝑦 𝑢 d𝑥d𝑦.
−𝑟 −𝑟 −𝑟

The ellipticity of 𝐿 𝑝 in the region 𝑥 ≠ 0 entails

∫ 𝑟 ∫ 𝑟
𝑘 𝑘
D 𝑦 𝑢 (𝑟, 𝑦) D 𝑥 D 𝑦 𝑢 (𝑟, 𝑦) d𝑦 + D 𝑘𝑦 𝑢 (−𝑟, 𝑦) D 𝑥 D 𝑘𝑦 𝑢 (−𝑟, 𝑦) d𝑦
−𝑟 −𝑟
≤ 𝐶12𝑘+3 𝑘! (𝑘 + 1)!. (4.3.4)

A priori the positive constant 𝐶1 depends on 𝑟; but we can bound 𝐶1 independently

of 𝑟 if we prescribe a lower limit for 𝑟, say 𝑟 > 21 𝑟 ◦ .
Next we note that 𝐿 𝑝 D 𝑘𝑦 𝑢 ≡ 0 implies

𝐿 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 = −2𝑥 2 𝑝 𝜒𝑟′ , 𝜀,2 (𝑦) 𝜕𝑦 D 𝑘𝑦 𝑢 − 𝑥 2 𝑝 𝜒𝑟′′, 𝜀,2 (𝑦) D 𝑘𝑦 𝑢.

We derive from (4.2.21) and (4.2.23):

4.3 A Generalization of the Harmonic Oscillator 99
∫ 𝑟∫ 𝑟−𝜀 ∫ 𝑟∫ 𝑟−𝜀
2 2
D 𝑥 D 𝑘𝑦 𝑢 d𝑥d𝑦 + 𝑥 2 𝑝 D 𝑘+1
𝑦 𝑢 d𝑥d𝑦
−𝑟 −𝑟+𝜀 −𝑟 −𝑟+𝜀
∫ 𝑟∫ 𝑟
≤ −2 Re 𝑥 2 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) 𝜒𝑟′ , 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 𝜕𝑦 D 𝑘𝑦 𝑢d𝑥d𝑦
∫ 𝑟 ∫−𝑟 𝑟 −𝑟
2
−2 𝑥 2 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) 𝜒𝑟′′, 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 d𝑥d𝑦
−𝑟 −𝑟
∫ 𝑟∫ 𝑟
2
+ 𝐶1 𝜀 −2 𝑥 2 𝑝 D 𝑘𝑦 𝑢 d𝑥d𝑦 + 𝐶22𝑘+3 𝑘! (𝑘 + 1)!.
−𝑟 −𝑟

To the first integral on the right we apply the obvious integration by parts formula:
∫ +∞
1 +∞ ′
∫
− Re 𝑓 (𝑦) 𝜑 (𝑦)𝜑 ′ (𝑦) d𝑦 = 𝑓 (𝑦) |𝜑 (𝑦)| 2 d𝑦,
−∞ 2 −∞

valid for all 𝜑 ∈ Cc∞ (R) and all real-valued 𝑓 ∈ C 1 (R). We get
∫ 𝑟∫ 𝑟
− 2 Re 𝑥 2 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) 𝜒𝑟′ , 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 𝜕𝑦 D 𝑘𝑦 𝑢d𝑥d𝑦
−𝑟 −𝑟
∫ 𝑟∫ 𝑟
2
−2 𝑥 2 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) 𝜒𝑟′′, 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 d𝑥d𝑦
∫−𝑟𝑟 ∫−𝑟𝑟
= Re 𝑥 2 𝑝 𝜒𝑟′2, 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 𝜕𝑦 D 𝑘𝑦 𝑢d𝑥d𝑦
∫ 𝑟−𝑟∫ 𝑟−𝑟
2
− 𝑥 2 𝑝 𝜒𝑟 , 𝜀,2 (𝑦) 𝜒𝑟′′, 𝜀,2 (𝑦) D 𝑘𝑦 𝑢 d𝑥d𝑦,
−𝑟 −𝑟

and therefore, applying once again (4.3.4),

∫ 𝑟 ∫ 𝑟−𝜀 ∫ 𝑟 ∫ 𝑟−𝜀
2 2
D 𝑥 D 𝑘𝑦 𝑢 d𝑥d𝑦 + 𝑥 2 𝑝 D 𝑘+1
𝑦 𝑢 d𝑥d𝑦 (4.3.5)
−𝑟 −𝑟+𝜀 −𝑟 −𝑟+𝜀
∫ 𝑟∫ 𝑟
2
≤ 𝐶22𝑘+3 𝑘! (𝑘 + 1)! + 𝐶3 𝜀 −2 𝑥 2 𝑝 D 𝑘𝑦 𝑢 d𝑥d𝑦.
−𝑟 −𝑟

The positive constant 𝐶3 only depends on the constant 𝐾 in (3.2.4).

At this point we replace 𝑟 by 𝑟 − 𝑘𝜀 < 𝑟 ◦ − 𝜀𝑘. We must keep in mind, however,
that we have required 𝑟 > 21 𝑟 ◦ (to maintain control of the constant 𝐶2 ). In view of
this we shall henceforth require 𝑟 > 43 𝑟 ◦ and we take 𝜀 = 4𝑘1
𝑟 (with 𝑘 ≥ 1 given).
3 1
The latter ensures 𝑟 − 𝜀𝑘 > 4 𝑟 ◦ − 𝜀𝑘 > 2 𝑟 ◦ as well as

1 1 1
𝑟 − 𝜀 (𝑘 + 1) = 3− 𝑟 ≥ 𝑟.
4 𝑘 2

With this we deduce from (4.3.5):

100 4 Analyticity of Solutions of Linear PDEs. Basic Results
∫ 𝑟−𝜀𝑘 ∫ 𝑟−𝜀 (𝑘+1) ∫ 𝑟−𝜀𝑘 ∫ 𝑟−𝜀 (𝑘+1)
2 2
D 𝑥 D 𝑘𝑦 𝑢 d𝑥d𝑦 + 𝑥 2 𝑝 D 𝑘+1
𝑦 𝑢 d𝑥d𝑦
−𝑟−𝜀𝑘 −𝑟−𝜀 (𝑘+1) −𝑟−𝜀𝑘 −𝑟−𝜀 (𝑘+1)
∫ 𝑟−𝜀𝑘 ∫ 𝑟−𝜀𝑘
2
≤ 𝐶22𝑘+3 𝑘! (𝑘 + 1)! + 16𝐶3 𝑟 −2 𝑘 2 𝑥 2 𝑝 D 𝑘𝑦 𝑢 d𝑥d𝑦. (4.3.6)
−𝑟−𝜀𝑘 −𝑟−𝜀𝑘

Induction on 𝑘 = 1, 2, ... allows us to derive from (4.3.6):

∫ 𝑟−𝜀𝑘 ∫ 𝑟−𝜀 (𝑘+1) 21
2𝑝 2
𝑥 D 𝑘+1
𝑦 𝑢 d𝑥d𝑦
−𝑟−𝜀𝑘 −𝑟−𝜀 (𝑘+1)
12
≤ (𝐶2 /𝑀) 2𝑘+2 𝐶2 + 16𝐶3 𝑟 −2 𝑀 𝑘+1 (𝑘 + 1)!,

whence
∫ 𝑟−𝜀𝑘 ∫ 𝑟−𝜀 (𝑘+1) 21
2𝑝 2
𝑥 D 𝑘+1
𝑦 𝑢 d𝑥d𝑦 ≤ 𝑀 𝑘+2 (𝑘 + 1)!
−𝑟−𝜀𝑘 −𝑟−𝜀 (𝑘+1)

under the sole proviso that 𝑀 ≥ 𝐶2 and

21 12
𝑀 ≥ 𝐶2 + 2𝐶3 𝑟 ◦−2 ≥ (𝐶2 /𝑀) 2𝑘+2 𝐶2 + 16𝐶3 𝑟 −2 .

Then this also implies

∫ 1 ∫ 1
! 12
2𝑟 2𝑟 2
D 𝑥 D 𝑘𝑦 𝑢 d𝑥d𝑦 ≤ 𝑀 𝑘+2 (𝑘 + 1)!. (4.3.7)
− 12 𝑟 − 12 𝑟

We use the fact that

∫ 1𝑟 2
2
𝑘 2 𝑘 1
D 𝑦 𝑢 d𝑥 = 𝑟 D 𝑦 𝑢 𝑟, 𝑦
− 12 𝑟 2
∫ 1𝑟
2 1
− 2 Re 𝑥 + 𝑟 D 𝑘𝑦 𝑢𝜕𝑥 D 𝑘𝑦 𝑢 d𝑥.
− 12 𝑟 2

Exploiting once again the analyticity of 𝑢 in the region 𝑥 ≠ 0 we get

∫ 1 ∫ 1
2𝑟 2 2𝑟 2
D 𝑘𝑦 𝑢 d𝑥 ≤ 𝐶42𝑘+2 𝑘! +𝑟 2
D 𝑘𝑦 𝑢 d𝑥
− 12 𝑟 − 12 𝑟
∫ 1
2𝑟 2
+ D 𝑥 D 𝑘𝑦 𝑢 d𝑥.
− 12 𝑟

Taking 𝑟 ◦ << 1 and combining the last inequality with (4.3.7) yields
4.3 A Generalization of the Harmonic Oscillator 101

∫ 1 ∫ 1
! 12
2𝑟 2𝑟 2
D 𝑘𝑦 𝑢 d𝑥d𝑦 ≤ 𝑀1𝑘+1 𝑘! (4.3.8)
− 12 𝑟 − 12 𝑟

for a suitably large 𝑀1 > 0.

Let us use the notation
∫ 1 ∫ 1
! 12
2𝑟 2𝑟
2
N𝑟 (𝑣) = |𝑣| d𝑥d𝑦 .
− 12 𝑟 − 12 𝑟

Suppose we have proved, for some 𝑗 ∈ Z+ and all 𝛼 ≤ 𝑗, that

N𝑟 𝜕𝑥𝛼 D 𝑘𝑦 𝑢 ≤ 𝑀2𝛼+𝑘+1 𝛼!𝑘! (4.3.9)

with a constant 𝑀2 > 0 independent of 𝑗. In the square 𝔔𝑟◦ we have D2𝑥 𝑢 = 𝑥 2 𝑝 D2𝑦 𝑢,
whence
min( 𝑗,2 𝑝)
∑︁ 𝑗! (2𝑝)! 2 𝑝−ℓ 𝑗−ℓ 𝑘
𝜕𝑥 D2𝑥 D 𝑘𝑦 𝑢 =
𝑗
𝑥 𝜕𝑥 D 𝑦 𝑢.
ℓ=0
ℓ! ( 𝑗 − ℓ)! (2𝑝 − ℓ)!
We derive
min( 𝑗,2 𝑝)
𝑗+2
∑︁ 𝑗 𝑗−ℓ
N𝑟 𝜕𝑥 D 𝑘𝑦 𝑢 ≤ (2𝑝)! N𝑟 𝜕𝑥 D 𝑘𝑦 𝑢
ℓ=0
ℓ
(2𝑝)! exp 𝑀2−1 .
𝑘+ 𝑗+1
≤ ( 𝑗 + 2)!𝑘!𝑀2

Requiring 𝑀22 ≥ (2𝑝)! ensures the validity of (4.3.9) for 𝛼 = 𝑗 + 2. Since the
estimates (4.3.7) and (4.3.8) show that (4.3.9) holds true for 𝛼 = 1, induction proves
(4.3.9) for all 𝛼 ∈ Z+ . □

Exercise 4.3.2 By adapting the proof of Theorem 4.3.1 prove the analytic hypoel-
lipticity of the operator 𝐿 = D2𝑥 + 𝑎 (𝑥, 𝑦) 𝑥 2 D2𝑦 in R2 , where 𝑎 ∈ C 𝜔 R2 , 𝑎 > 0 at
every point.

Remark 4.3.3 The method in the preceding proof can be greatly extended to an
important class of so-called sums of square operators as shown in [Tartakoff, 1980].

4.3.2 Differential operators akin to the harmonic oscillator operator

that are not analytic hypoelliptic

There exist linear differential operators seemingly very close to (4.2.1) that are not
analytic hypoelliptic. The simplest and first discovered (see [Baouendi-Goulaouic,
1972]) is the object of the following
102 4 Analyticity of Solutions of Linear PDEs. Basic Results

Theorem 4.3.4 The equation

D2𝑥 ℎ + 𝑥 2 D2𝑦 ℎ + D2𝑧 ℎ = 0 (4.3.10)

has a solution in R3 that is not of class C 𝜔 at the origin.

Proof For any 𝜅, 1 < 𝜅 < 2, the C ∞ function in R3 ,

∫ +∞
2 1 2
e𝑖𝑡 ( 𝑦+ 2 𝑖 𝑥 ) +𝑡 𝑧−𝑡 d𝑡,
𝜅
ℎ (𝑥, 𝑦, 𝑧) =
0

satisfies (4.3.10) but cannot be extended to any neighborhood of 0 in C2 as a

holomorphic function of (𝑥, 𝑦). □
Equation (4.3.10) belongs to a large class of the same type that are not analytic
hypoelliptic, for example the Oleinik PDEs (see [Oleinik, 1973])
𝑛−1
∑︁
D2𝑥 ℎ + 𝑥 2 𝑝 𝑗 D2𝑦 𝑗 ℎ = 0 in R𝑛 , 𝑛 ≥ 3, (4.3.11)
𝑗=1

where 𝑝 𝑗 ≠ 𝑝 𝑘 for some ( 𝑗, 𝑘), 1 ≤ 𝑗 < 𝑘 ≤ 𝑛 − 1. (Notice that it suffices to

prove the claim when 𝑛 = 3.) There are even examples in R2 : D2𝑥 + 𝑥 2 + 𝑦 2 D2𝑦 (see

[Métivier, 1980]). On this class of problems for the sums of squares operators (cf.
Remark 4.3.3), generalizing the Métivier operator as well as (4.3.11), we refer the
reader to the survey in [Bove-Mughetti, 2020].
In Ch. 24 it will be proved that for linear differential operators of order 𝑚 ≥ 1 with
analytic coefficients and simple real characteristics [meaning that d 𝜉 𝑃𝑚 (𝑥, 𝜉) ≠ 0
if 0 ≠ 𝜉 ∈ R𝑛 ] hypoellipticity is equivalent to analytic hypoellipticity.

4.A Appendix: Hermite’s Functions and the Schwartz Space

d
We introduce the following two first-order real differential operators 𝑋 = d𝑥 − 𝑥,
′ d ′
𝑋 = d𝑥 + 𝑥; 𝑋 and −𝑋 are the adjoint (or transpose) of one another. We have
[𝑋, 𝑋 ′] = 2 and, for 𝑓 ∈ D ′ (R),

d2 𝑓
𝑋𝑋′ 𝑓 = − 𝑥2 𝑓 + 𝑓 , (4.A.1)
d𝑥 2
d2 𝑓
𝑋 ′ 𝑋 𝑓 = 2 − 𝑥2 𝑓 − 𝑓 . (4.A.2)
d𝑥
We define the 𝑚 th Hermite function as follows:

2−𝑚/2 𝑚

1 2
H𝑚 (𝑥) = √ 𝑋 exp − 𝑥 . (4.A.3)
𝑚! 2
4.A Appendix: Hermite’s Functions and the Schwartz Space 103

We have the following recurrence relations, for 𝑚 ≥ 0,

1
H𝑚+1 (𝑥) = √︁ 𝑋H𝑚 (𝑥) , (4.A.4)
2 (𝑚 + 1)

and the estimate

𝑚/2 1
|H𝑚 (𝑥)| ≤ 𝐶𝑚 1 + 𝑥 2 exp − 𝑥 2 (4.A.5)
2
with 𝐶𝑚 > 0 independent of 𝑥. In particular, H𝑚 ∈ S (R) for all 𝑚 ∈ Z+ .
√
Lemma 4.A.1 The Fourier transform H b𝑚 (𝜉) of H𝑚 (𝑥) is equal to 2𝜋𝑖 −𝑚 H𝑚 (𝜉).

Proof We have
∫ +∞
1 2
b0 (𝜉) =
H e−𝑖 𝑥 𝜉 − 2 𝑥 d𝑥
−∞
∫ +∞ √
2
− 12 𝜉 2 1 1 2
=e e− 2 ( 𝑥+𝑖 𝜉 ) d𝑥 = 2𝜋e− 2 𝜉 ,
−∞

the latter equality a consequence of the Cauchy Integral Theorem, as it allows us to

move the integration from the line Im 𝑧 = 𝜉 to the 𝑥-axis. We derive from (4.A.4)
√︁ ∫ +∞
2 (𝑚 + 1) H
b𝑚+1 (𝜉) = − H𝑚 (𝑥) 𝑋 ′e−𝑖 𝑥 𝜉 d𝑥
−∞
∫ +∞
= e−𝑖 𝑥 𝜉 H𝑚 (𝑥) (𝑖𝜉 − 𝑥) d𝑥
−∞

𝜕 b𝑚 (𝜉) .
= −𝑖 −𝜉 H
𝜕𝜉

Induction on 𝑚 yields
√
√

𝜋 −𝑚−1 𝜕
H𝑚+1 (𝜉) = √
b 𝑖 − 𝜉 H𝑚 (𝜉) = 2𝜋𝑖 −𝑚−1 H𝑚+1 (𝜉) . □
𝑚+1 𝜕𝜉

Lemma 4.A.2 We have

d2 H𝑚
− 𝑥 2 H𝑚 = − (2𝑚 + 1) H𝑚 (4.A.6)
d𝑥 2
for every 𝑚 ∈ Z+ .

By (4.A.1) Equation (4.A.6) can be rewritten as

𝑋 𝑋 ′H𝑚 = −2𝑚H𝑚 . (4.A.7)

Proof The claim is trivial for 𝑚 = 0. We derive from (4.A.4)

104 4 Analyticity of Solutions of Linear PDEs. Basic Results

1 1
𝑋 ′H𝑚 = √ 𝑋 ′ 𝑋H𝑚−1 = √ (𝑋 𝑋 ′ − 2) H𝑚−1 .
2𝑚 2𝑚
Induction on 𝑚 ≥ 1 yields
√
𝑋 ′H𝑚 = − 2𝑚H𝑚−1 (4.A.8)

whence √
𝑋 𝑋 ′H𝑚 = − 2𝑚𝑋H𝑚−1 .
Combining this with (4.A.4) proves (4.A.7). □
√
(2 𝑝)!
Lemma 4.A.3 If 𝑝 ∈ Z+ then H2 𝑝 (0) = (−1) 𝑝 and H2 𝑝+1 (0) = 0,
√ 𝑝!
dH2 𝑝+1 𝑝+1 (2 𝑝+1)!
d𝑥 (0) = (−1) 𝑝! .

dH1
√
Proof We have H0 (0) = 1, H1 (0) = 0, d𝑥 (0) = − 2. From (4.A.8) we derive
√ √︁
𝑋 ′2 H𝑚 = − 2𝑚𝑋 ′H𝑚−1 = 2 𝑚 (𝑚 − 1)H𝑚−2 .

By Lemma 4.A.2 we have

d2 H𝑚

d dH𝑚
𝑋 ′2 H𝑚 𝑥=0 = +𝑥 + 𝑥H𝑚 = (0) + H𝑚 (0) = −2𝑚H𝑚 (0)
d𝑥 d𝑥 d𝑥 2

whence, for 𝑚 ≥ 2, √︂
𝑚−1
H𝑚 (0) = − H𝑚−2 (0) .
𝑚
√
(2 𝑝)!
We derive H2 𝑝 (0) = (−1) 𝑝 𝑝! and H2 𝑝+1 (0) = 0 for all 𝑝. Applying again
(4.A.8) we get
dH2 𝑝+1 √︁
(0) = − 2𝑝 + 1H2 𝑝 ,
d𝑥
which implies the last formula in the statement. □

Lemma 4.A.4 The functions √1 H𝑚 (𝑚 ∈ Z+ ) form an orthonormal system in

𝜋
𝐿2 (R).
∫ +∞
Proof We have −∞
H02 (𝑥) d𝑥 = 𝜋. For 𝑚 ≥ 1 , (4.A.4) implies
∫ +∞ ∫ +∞
1
H𝑚 (𝑥) H0 (𝑥) d𝑥 = √ H𝑚−1 (𝑥) 𝑋 ′H0 (𝑥) d𝑥 = 0.
−∞ 2𝑚 −∞

If 𝑚, 𝑚 ′ ∈ Z+ , we have, again by (4.A.4),

4.A Appendix: Hermite’s Functions and the Schwartz Space 105
∫ +∞ +∞∫
1
H𝑚 (𝑥) H𝑚′ (𝑥) d𝑥 = √ 𝑋H𝑚−1 (𝑥) H𝑚′ (𝑥) d𝑥
−∞ 2𝑚 −∞
∫ +∞
1
= −√ H𝑚−1 (𝑥) 𝑋 ′H𝑚′ (𝑥) d𝑥
2𝑚 −∞
√︂ ∫ +∞
𝑚′
= H𝑚−1 (𝑥) H𝑚′ −1 (𝑥) d𝑥,
𝑚 −∞
the last equation a consequence of (4.A.8). Induction on 𝑚 yields
∫ +∞
H𝑚2 (𝑥) d𝑥 = 𝜋,
−∞
∫ +∞
H𝑚 (𝑥) H𝑚′ (𝑥) d𝑥 = 0 if 𝑚 ′ < 𝑚. □
−∞

Definition 4.A.5 We denote by S Z+𝑛 the set of sequences 𝒄 = {𝑐 𝛼 } 𝛼∈Z+𝑛 ⊂ C such

that sup (1 + |𝛼|) 𝑘 |𝑐 𝛼 | < +∞ for every 𝑘 ∈ Z+ , and by S ′ Z+𝑛 the space of

𝛼∈Z+𝑛
sequences 𝒄 = {𝑐 𝛼 } 𝛼∈Z+𝑛 ⊂ C with the property that there is a 𝑘 ≥ 0 such that

sup (1 + |𝛼|) −𝑘 |𝑐 𝛼 | < +∞.
𝛼∈Z+𝑛

We refer to the elements of S Z+𝑛 as the rapidly decaying sequences in Z+𝑛 ;
S Z+𝑛 is a Fréchet–Montel algebra when equipped with the norms

𝒄 ↦→ sup (1 + |𝛼|) 𝑘 |𝑐 𝛼 | .
𝛼∈Z+𝑛

We refer to S ′ Z+𝑛 as the space of tempered (i.e., of polynomial

growth at infinity)
𝑛 and S ′ Z𝑛 are the dual of

sequences in Z+𝑛 . The topological vector spaces S Z + +
each other, the duality bracket being ⟨𝒄, 𝒄 ′⟩ = 𝛼∈Z+𝑛 𝑐 𝛼 𝑐 ′𝛼 .
Í

Theorem 4.A.6 If {𝑐 𝑚 } ∈ S (Z+ ) then the series ∞

Í
𝑚=0 𝑐 𝑚 H𝑚 converges to 𝑓 ∈
S (R) in 𝐿 2 (R). The map
∞
∑︁
S (Z+ ) ∋ {𝑐 𝑚 } ↦→ 𝑐 𝑚 H𝑚 ∈ S (R) (4.A.9)
𝑚=0

is an isomorphism of Fréchet–Montel algebras.

Proof It is immediate that if {𝑐 𝑚 } ∈ S (Z+ ) then the series ∞

Í
𝑚=0 𝑐 𝑚 H𝑚 converges
to 𝑓 ∈ 𝐿 2 (R). By (4.A.4) and (4.A.8) we have (in the distribution sense)
106 4 Analyticity of Solutions of Linear PDEs. Basic Results
∞ √︁
∑︁
𝑋𝑓 = 2 (𝑚 + 1)𝑐 𝑚 H𝑚+1 .
𝑚=0
∞ √
∑︁
𝑋′ 𝑓 = − 2𝑚𝑐 𝑚 H𝑚−1 .
𝑚=1

d𝑓 Í∞
We derive from this that both d𝑥 and 𝑥 𝑓 are series of the same type as 𝑚=0 𝑐 𝑚 H𝑚 .
ℓ
Induction on 𝑘,ℓ ∈ Z+ yields to the conclusion that 𝑥 𝑘 dd𝑥 ℓ𝑓
∈ (R) 𝐿2 for all (𝑘, ℓ) ∈ Z2+
thereby proving that 𝑓 ∈ S (R). ∫ +∞
Now let 𝑓 ∈ S (R) be such that −∞ H𝑚 (𝑥) 𝑓 (𝑥) d𝑥 = 0 for all 𝑚 ∈ Z+ . We
apply once again (4.A.4) and (4.A.8) and derive, for all 𝑚 ∈ Z+ ,
∫ +∞ ∫ +∞
H𝑚 (𝑥) 𝑋 𝑓 (𝑥) d𝑥 = H𝑚 (𝑥) 𝑋 ′ 𝑓 (𝑥) d𝑥 = 0,
−∞ −∞
∫ +∞ ∫ +∞
implying −∞ H𝑚 (𝑥) 𝑓 (𝑥) 𝑥d𝑥 = 0 whence −∞ H𝑚 (𝑥) 𝑓 (𝑥) 𝑥 𝑘 d𝑥 = 0 for all
𝑘 ∈ Z+ . In particular, we get, whatever 𝑦 ∈ R𝑛 ,
∫ +∞
1 2
∀𝑘 ∈ Z+ , e− 2 𝑥 𝑓 (𝑥) (𝑥 − 𝑦) 𝑘 d𝑥 = 0,
−∞

which implies, for all 𝜈 > 0,

√︂ ∑︁ ∞ ∫ +∞
𝜈 (−𝜈) 𝑘 1 2
e− 2 𝑥 𝑓 (𝑥) (𝑥 − 𝑦) 2𝑘 d𝑥
𝜋 𝑘=0 𝑘! −∞
√︂ ∫ +∞
𝜈 2 1 2
= e−𝜈 ( 𝑥−𝑦) − 2 𝑥 𝑓 (𝑥) d𝑥 = 0.
𝜋 −∞
1 2
Letting 𝜈 go to +∞ yields e− 2 𝑦 𝑓 (𝑦) = 0 for all 𝑦 ∈ R.
Lastly, let 𝑓 ∈ S (R) be arbitrary and define, for each 𝑚 ∈ Z+ ,
∫ +∞
1
𝑐𝑚 = √ 𝑓 (𝑥) H𝑚 (𝑥) d𝑥. (4.A.10)
𝜋 −∞

We apply (4.A.7):
∫ +∞
1
2𝑚𝑐 𝑚 = √ 𝑋 𝑋 ′ 𝑓 (𝑥) H𝑚 (𝑥) d𝑥.
𝜋 −∞

By the Cauchy–Schwarz inequality we get

1
|𝑐 𝑚 | ≤ ∥ 𝑋 𝑋 ′ 𝑓 ∥ 𝐿2 .
2𝑚
4.A Appendix: Hermite’s Functions and the Schwartz Space 107

Since (𝑋 𝑋 ′) 𝑘 𝑓 ∈ S (R) we can repeat this procedure indefinitely. We conclude that

sup 𝑚 𝑘 |𝑐 𝑚 | < +∞ for all 𝑘 ∈ Z+ and thus ∞
Í
𝑚=0 𝑚 𝑚 ∈ S (R). It follows directly
𝑐 H
𝑚
from Lemma 4.A.4 that
∞
!
∫ +∞ ∑︁
∀ℓ ∈ Z+ , 𝑓 (𝑥) − 𝑐 𝑚 H𝑚 (𝑥) Hℓ (𝑥) d𝑥 = 0, (4.A.11)
−∞ 𝑚=0

whence ∞
Í
𝑚=0 𝑐 𝑚 H𝑚 = 𝑓 . To prove that the map (4.A.9) is continuous is easy;
as a continuous bijection of a Fréchet space onto another one it is necessarily a
homeomorphism. □
1
Corollary 4.A.7 For 𝑢 ∈ S ′ (R) define 𝑐 𝑚 = 𝜋 − 2 ⟨𝑢, H𝑚 ⟩; then the map
∞
∑︁
S ′ (Z+ ) ∋ {𝑐 𝑚 } ↦→ 𝑐 𝑚 H𝑚 ∈ S ′ (R) (4.A.12)
𝑚=0

is an isomorphism of duals of Fréchet–Montel spaces.

Proof The map (4.A.12) is the contragredient (i.e., the inverse of the transpose) of
the map (4.A.9). □
Chapter 5
The Cauchy–Kovalevskaya Theorem

The bulk of this chapter is devoted to the fundamental theorem in analytic PDE theory
and one of the most important mathematical discoveries of the XIXth century: that
the Cauchy problem for an analytic, fully nonlinear PDEs, with Cauchy data on
a noncharacteristic hypersurface Σ has a unique analytic solution in a sufficiently
small neighborhood of an arbitrary point of Σ. (The uniqueness of solutions allows
their analytic extension to a neighborhood of Σ whose shape determination is one
the deepest questions in PDE analysis.) Most of the time we interpret analytic as
complex-analytic, the real-analytic version then ensuing by restriction to real space.
We shall refrain from selecting a direction transversal to Σ in which to relinquish
analyticity and only require continuity, as done in [Nirenberg, 1972] using the Nash–
Moser Implicit Function Theorem. There are two reasons for this restrictive choice:
1) Preserving invariance under analytic transformations is an important imperative
from the standpoint of this book. 2) Assuming analyticity in all directions makes the
proof extremely simple and self-contained. This approach has its origins in the work
[Ovsyannikov, 1965]; it is essentially the proof in [Treves, 1970]. But we have chosen
to frame it in an overarching concept, that of Ovsyannikov1 analyticity, the defining
property of nonlinear maps 𝐹 between scales of Banach spaces {𝑬 𝑠 }0<𝑠<1 whose
Fréchet derivatives of every order satisfy locally the standard analytic inequalities –
up to the factor (𝑠 − 𝑠 ′) −1 . The precise definition is the content of the first subsection;
it is immediately followed by the solution of the Cauchy problem for a first-order
ODE in the Ovsyannikov analytic category. Section 5.2 presents the derivation of the
Cauchy–Kovalevskaya Theorem for a fully nonlinear analytic PDE and, by a simple
extension, for a class of square systems of such equations. The derivation consists
in a small number of changes of independent variables and of unknowns, rather
routine for at least a couple of centuries, to bring the initial equation to the form of a
quasilinear square system recognizable as a special case of the abstract ODEs solved
in Section 5.1. Section 5.3 discusses the (not unexpected) applications to a linear
analytic PDE and the derivation, using duality, of the Holmgren Theorem. Section

1 The misspelling of the last name of L.V. Ovsiannikov by the author in prior publications is
corrected in this book.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 109
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_5
110 5 The Cauchy–Kovalevskaya Theorem

5.4 shows how Ovsyannikov analyticity can be exploited to solve Cauchy problems
for certain nonlinear integrodifferential equations, with the Camassa–Holm equation
as an example.
A warning: the main theorem in this chapter is not the ultimate statement about
the Cauchy problem for fully nonlinear (or even linear) analytic PDEs. As a matter
of fact, in this very book, we shall use (sometimes without proofs, regarding the
claims as self-evident) versions of the Cauchy–Kovalevskaya Theorem not covered
in the present chapter, for instance the version applicable to systems of equations
𝐿 𝑗 𝑢 = 𝑓 𝑗 , 𝑗 = 1, ..., 𝑟, where 𝐿 1 , ..., 𝐿 𝑛 are analytic vector fields in an open subset
Ω of Euclidean space K𝑛 (as usual, K = R or C). It will alwaysbe assumed that
the 𝐿 𝑗 satisfy an involution condition: the commutation brackets 𝐿 𝑗 , 𝐿 𝑘 must be
linear combinations of 𝐿 1 , ..., 𝐿 𝑛 with coefficients in C 𝜔 (Ω) or in O (Ω). As a
consequence, the right-hand sides 𝑓 𝑗 must satisfy the corresponding compatibility
conditions. Such systems will be discussed in detail in Ch. 12; here we mention
them only as a pointer to the truly general Cauchy–Kovalevskaya Theorem, the one
dealing with the Cauchy problem for systems of nonlinear PDEs

𝑃 𝑗 𝑥, D 𝑥𝛼1 𝑢 1 .., D 𝑥𝛼𝜈 𝑢 𝜈 = 0, 𝑗 = 1, ..., 𝑟, (5.1)

where 𝑢 1 , ..., 𝑢 𝜈 are the unknowns, analytic functions in a neighborhood of the

origin in K𝑛 . Each 𝑃 𝑗 𝑥, D 𝑥𝛼1 𝑢 1 .., D 𝑥𝛼𝜈 𝑢 𝜈 is a polynomial with respect to the partial
derivatives of 𝑢 1 , ..., 𝑢 𝜈 , of order ≤ 𝑚, with coefficients in C 𝜔 (Ω) or in O (Ω).
The nonspecialist reader will be readily aware of the enormous number and depth of
conditions that must be imposed on the equations (5.1) and on the Cauchy conditions
on the putative solutions 𝑢 1 , ..., 𝑢 𝜈 . For instance, is the maximum order on each 𝑢 𝑗 to
be determined individually, or should it be the same for all 𝑢 𝑗 ? What will we mean
by requiring the submanifold that carries the Cauchy data to be noncharacteristic?
Most importantly, what relations between the 𝑃 𝑗 themselves (of the involution kind)
need to be imposed?
Part of the answers can be extracted from the linearization of the problem (as we
do in the first section of this chapter, in the case of a single equation with a single
unknown). Since we are here solely interested in analytic equations and analytic
solutions, a crucial part of the answers is provided by the formal approach: Replace
coefficients and solutions by their Taylor expansions about a central point in Ω, taken
to be the origin of K𝑛 . Check then that (5.1) and the attached “initial conditions”
make sense in the framework of formal series in the powers of the “independent
variables”, 𝑥1 , ..., 𝑥 𝑛 . There is an abundant literature on this and related topics,
going back to the XIXth century (for instance, on Pfaffian systems). Last, perhaps
the hardest part: prove that the series do converge. An introduction to the modern
treatment can be found in [Goldschmidt, 1986]; for the technical details, references
are [Goldschmidt-Spencer, 1976], [Goldschmidt-Spencer, 1978].
An open question is whether, after validation of the Cauchy problem for (5.1)
in terms of formal power series, Ovsyannikov analyticity applied to the appropriate
scales of Banach spaces of analytic functions could yield a relatively simple proof
of the existence and uniqueness of the solutions.
5.1 A Nonlinear Ovsyannikov Theorem 111

5.1 A Nonlinear Ovsyannikov Theorem

5.1.1 Ovsyannikov analyticity in a scale of Banach spaces

By a scale of complex Banach spaces we shall mean a one-parameter family of

complex Banach spaces 𝑬 𝑠 (0 < 𝑠 < 1) satisfying the following condition:
(Scale) If 𝑠 > 𝑠 ′ there is a continuous linear injection 𝑬 𝑠 ↩→ 𝑬 𝑠′ with norm ≤ 1.
We denote by 𝑵 𝑠 (·) the norm in 𝑬 𝑠 . If 0 < 𝑠 ′ < 𝑠 < 1 we identify 𝑬 𝑠 to
a linear subspace of 𝑬 𝑠′ ; with this identification the property (Scale) states that
𝑵 𝑠′ (𝑢) ≤ 𝑵 𝑠 (𝑢) for every 𝑢 ∈ 𝑬 𝑠 . We introduce the following vector spaces:
Ø
𝑬0 = 𝑬𝑠, (5.1.1)
0<𝑠<1
( )
Ù
𝑬1 = 𝑢 ∈ 𝑬 𝑠 ; sup 𝑵 𝑠 (𝑢) < +∞ . (5.1.2)
0<𝑠<1 0<𝑠<1

There is a natural locally convex topology on 𝑬 0 , that of a so-called (nonstrict)

inductive limit of Banach spaces (see [Treves, 1967]): a convex subset of 𝑬 0 is open
if and only if its intersection with 𝑬 𝑠 is open for every 𝑠 ∈ (0, 1). We regard 𝑬 1 as
a normed space, with the norm 𝑢 ↦→ 𝑵 1 (𝑢) = sup 𝑵 𝑠 (𝑢). The case 𝑬 1 = {0} is
0<𝑠<1
not precluded but if true the main applications in the present chapter are not very
interesting; that 𝑬 1 is dense in 𝑬 0 is not a requirement although it happens to be true
in the most significant applications.
Proposition 5.1.1 The normed space 𝑬 1 is a Banach space.
Proof A Cauchy sequence {𝑢 𝑘 } 𝑘=1,2,... in 𝑬 1 for the norm 𝑵 1 is a Cauchy sequence
(for the norm 𝑵 𝑠 ) whatever 𝑠 ∈ (0, 1) and therefore converges in 𝑬 𝑠 to some
in 𝑬 𝑠Ù
𝑢∈ 𝑬 𝑠 . Since sup 𝑵 1 (𝑢 𝑘 ) < +∞ we have 𝑵 1 (𝑢) < +∞. □
0<𝑠<1 𝑘=1,2,...
Ù Ø
In general it is not true that 𝑬 𝑠 = 𝑬 𝑠′ nor that 𝑬 𝑠 = 𝑬 𝑠′ , as seen in
0<𝑠′ <𝑠 𝑠<𝑠′ <1
the following
Example 5.1.2 For each 𝑠 ∈ (0, 1) set Δ𝑠 = {𝑧 ∈ C; |𝑧| < 𝑠} and let Δ𝑠 denote the
closure of Δ𝑠 . Take 𝑬 𝑠 to be the linear space of complex functions 𝑓 ∈ C Δ𝑠
holomorphic in Δ𝑠 ; 𝑬 𝑠 is a Banach space for the norm ℎ ↦→ max |ℎ(𝑧|. If 𝑠 > 𝑠 ′ the
|𝑧 | ≤𝑠
“natural” injection 𝑬 𝑠 ↩→ 𝑬 𝑠′ is defined
Ù by the restriction from Δ𝑠 to Δ𝑠′ . For every
𝑠 ∈ (0, 1), on the one hand we have 𝑬 𝑠′ = O (Δ𝑠 ) ≠ 𝑬 𝑠 [O (Δ𝑠 ): the space of
0<𝑠′ <𝑠 Ø
holomorphic functions in Δ𝑠 ]. On the other hand 𝑬 𝑠′ is the space of functions
𝑠<𝑠′ <1
defined and holomorphic in a neighborhood of Δ𝑠 . Note, however, that
112 5 The Cauchy–Kovalevskaya Theorem

∀𝑢 ∈ 𝑬 𝑠 , 𝑵 𝑠 (𝑢) = sup 𝑵 𝑠′ (𝑢) = inf′ 𝑵 𝑠′ (𝑢) . (5.1.3)

0<𝑠′ <𝑠 𝑠<𝑠 <1

The union 𝑬 0 is the space of holomorphic functions in some neighborhood of 0

and 𝑬 1 is the subspace of O (Δ1 ) consisting of the bounded holomorphic functions
in the open unit disk, Δ1 .

We introduce the following “balls” centered at an arbitrary element 𝑢 ◦ ∈ 𝑬 1 :

𝔅𝑠,𝑅 (𝑢 ◦ ) = {𝑢 ∈ 𝑬 𝑠 ; 𝑵 𝑠 (𝑢 − 𝑢 ◦ ) < 𝑅} , 𝑅 > 0, 0 < 𝑠 < 1; (5.1.4)

Ø
𝔅𝑅 (𝑢 ◦ ) = 𝔅𝑠,𝑅 (𝑢 ◦ ) .
0<𝑠<1

Lemma 5.1.3 The subset 𝔅𝑅 (𝑢 ◦ ) of 𝑬 0 is convex and open.

Proof As the union of nested convex sets 𝔅𝑅 (𝑢 ◦ ) is convex. If 0 < 𝑠 ′ < 𝑠 < 1 the
continuity of the injection 𝑬 𝑠 ↩→ 𝑬 𝑠′ implies that 𝔅𝑠′ ,𝑅 (𝑢 ◦ ) ∩ 𝑬 𝑠 is open (but not
necessarily bounded) in 𝑬 𝑠 ; if 𝑠 ≤ 𝑠 ′ < 1 then 𝔅𝑠′ ,𝑅 (𝑢 ◦ ) ⊂ 𝔅𝑠,𝑅 (𝑢 ◦ ). It follows
that Ø
𝔅𝑅 (𝑢 ◦ ) ∩ 𝑬 𝑠 = 𝔅𝑠′ ,𝑅 (𝑢 ◦ ) ∩ 𝑬 𝑠
0<𝑠′ ≤𝑠<1

is open in 𝑬 𝑠 . □
We shall deal with maps 𝐹 : 𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 that have special differential
properties. In the remainder of this section (𝑠, 𝑠 ′) will denote a pair of real numbers
such that 0 < 𝑠 ′ < 𝑠 < 1. For each such pair (𝑠, 𝑠 ′), 𝐹 shall be a bounded map
𝔅𝑠,𝑅 (𝑢 ◦ ) −→ 𝑬 𝑠′ . Its bound will depend on the difference 𝑠 − 𝑠 ′ as follows:

𝑵 𝑠′ (𝐹 (𝑢)) ≤ 𝐶◦ (𝑠 − 𝑠 ′) −1 , (5.1.5)

where the constant 𝐶◦ > 0 is independent of the pair (𝑠, 𝑠 ′).

Example 5.1.4 Consider the scale {𝑬 𝑠 }0<𝑠<1 in Example 5.1.2 and set 𝐹 (𝑢) = 𝜕𝑧 𝑢:
by the Cauchy inequality, if 0 < 𝑠 ′ < 𝑠 < 1 the map 𝐹 : 𝑬 𝑠 −→ 𝑬 𝑠′ is linear and
bounded as follows:

max |𝜕𝑧 𝑢 (𝑧)| ≤ (𝑠 − 𝑠 ′) −1 max |𝑢 (𝑧)| .

|𝑧 | ≤𝑠′ |𝑧 | ≤𝑠

Of course, 𝐹 does not map 𝑬 𝑠 into itself.

Throughout the sequel, unless stated otherwise, 𝐹 : 𝔅𝑠,𝑅 (𝑢 ◦ ) −→ 𝑬 𝑠′ will

be a Fréchet differentiable map (nonlinear, generally). This means that, given any
𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) and any 𝑣 ∈ 𝑬 𝑠 , as 𝜀 > 0 goes to zero the difference quotient

𝐹 (𝑢 + 𝜀𝑣) − 𝐹 (𝑢)
(5.1.6)
𝜀
converges in 𝑬 𝑠′ . In the situations of interest to us (5.1.6) will not converge in 𝑬 𝑠 .
5.1 A Nonlinear Ovsyannikov Theorem 113

We refer to the limit of (5.1.6) as the Fréchet derivative of 𝐹 at 𝑢 in the direction

𝑣 and denote it by D𝐹 (𝑢) 𝑣. It is evident that for fixed 𝑢, the map 𝑬 𝑠 ∋ 𝑣 ↦→
D𝐹 (𝑢) 𝑣 ∈ 𝑬 𝑠′ is linear. We shall hypothesize that it is bounded, in the following
precise fashion: there is a 𝐶1 > 0 independent of the pair (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1,
and of 𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) such that

∀𝑣 ∈ 𝑬 𝑠 , 𝑵 𝑠′ (D𝐹 (𝑢) 𝑣) ≤ 𝐶◦ 𝐶1 (𝑠 − 𝑠 ′) −1 𝑵 𝑠 (𝑣) , (5.1.7)

where 𝐶◦ is the constant in (5.1.5). This defines a map

𝔅𝑠,𝑅 (𝑢 ◦ ) ∋ 𝑢 ↦→ D𝐹 (𝑢) ∈ 𝔏 (𝑬 𝑠 ; 𝑬 𝑠′ ) (5.1.8)

with 𝔏 (𝑬 𝑠 ; 𝑬 𝑠′ ) the Banach space of bounded linear operators 𝑬 𝑠 −→ 𝑬 𝑠′ .

Our next step is to hypothesize that D𝐹 itself is differentiable in the same sense
as 𝐹: whatever 𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) and 𝑤 ∈ 𝑬 𝑠 , the difference quotient

D𝐹 (𝑢 + 𝜀𝑤) − D𝐹 (𝑢)
𝜀
converges in 𝔏 (𝑬 𝑠 ; 𝑬 𝑠′ ) as 𝜀 ↘ 0, equivalent to saying that

D𝐹 (𝑢 + 𝜀𝑤) 𝑣 − D𝐹 (𝑢) 𝑣
(5.1.9)
𝜀
converges in 𝑬 𝑠′ whatever (𝑣, 𝑤) ∈ 𝑬 2𝑠 . We denote by D2 𝐹 (𝑢 ◦ ) (𝑣, 𝑤) the limit of
(5.1.9); it is easily seen to be bilinear with respect to 𝑣, 𝑤. We get a map

𝔅𝑠,𝑅 (𝑢 ◦ ) ∋ 𝑢 ↦→ D2 𝐹 (𝑢) ∈ 𝔏2 (𝑬 𝑠 ; 𝑬 𝑠′ )

with 𝔏2 (𝑬 𝑠 ; 𝑬 𝑠′ ) the Banach space of bounded bilinear maps 𝑬 2𝑠 −→ 𝑬 𝑠′ . We shall

hypothesize that it is bounded, in the following precise fashion:

𝑵 𝑠′ D2 𝐹 (𝑢) (𝑣, 𝑤) ≤ 2𝐶◦ 𝐶12 (𝑠 − 𝑠 ′) −1 𝑵 𝑠 (𝑣) 𝑵 𝑠 (𝑤)

for all 𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) and (𝑣, 𝑤) ∈ 𝑬 2𝑠 . As before the positive constants 𝐶◦ and 𝐶1
are independent of (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1, here also of (𝑢, 𝑣, 𝑤) ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) × 𝑬 2𝑠 .
We can repeat our hypothesizing ad infinitum. Letting the positive integer 𝑘 be
arbitrary, we posit that the Fréchet derivative of 𝐹 of order 𝑘 at 𝑢 ◦ has been defined:
it is a bounded map

𝔅𝑠,𝑅 (𝑢 ◦ ) ∋ 𝑢 ↦→ D 𝑘 𝐹 (𝑢) ∈ 𝔏 𝑘 (𝑬 𝑠 ; 𝑬 𝑠′ ) , (5.1.10)

where 𝔏 𝑘 (𝑬 𝑠 ; 𝑬 𝑠′ ) is the Banach space of bounded 𝑘-linear maps 𝑬 𝑠𝑘 −→ 𝑬 𝑠′ . The

bound of (5.1.10) is given by

𝑵 𝑠′ D 𝑘 𝐹 (𝑢) (𝑣 1 , ..., 𝑣 𝑘 ) ≤ 𝐶◦ 𝐶1𝑘 𝑘! (𝑠 − 𝑠 ′) −1 𝑵 𝑠 (𝑣 1 ) · · · 𝑵 𝑠 (𝑣 𝑘 ) (5.1.11)
114 5 The Cauchy–Kovalevskaya Theorem

for all 𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) and all (𝑣 1 , ..., 𝑣 𝑘 ) ∈ 𝑬 𝑠𝑘 ; here the constants 𝐶◦ and 𝐶1 are
independent of (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1, and of (𝑣 1 , ..., 𝑣 𝑘 ) ∈ 𝑬 𝑠𝑘 . We assume that the
map (5.1.10) is itself Fréchet differentiable in the sense of (5.1.8) and we denote by
D 𝑘+1 𝐹 (𝑢) its Fréchet derivative. Etc.

Remark 5.1.5 If the map 𝐹 : 𝔅𝑠,𝑅 (𝑢 ◦ ) −→ 𝑬 𝑠′ is infinitely Fréchet differentiable

then, whatever 𝑘 ∈ Z+ , the 𝑘th Fréchet derivative (5.1.10) is a continuous map and
D 𝑘 𝐹 (𝑢) ∈ 𝔏 𝑘 (𝑬 𝑠 ; 𝑬 𝑠′ ) is symmetric, meaning that

D 𝑘 𝐹 (𝑢) (𝑣 1 , ..., 𝑣 𝑘 ) = D 𝑘 𝐹 (𝑢) 𝑣 𝑖1 , ..., 𝑣 𝑖𝑘 (5.1.12)

for all (𝑣 1 , ..., 𝑣 𝑘 ) ∈ 𝑬 𝑠𝑘 and all permutations (𝑖1 , ..., 𝑖 𝑘 ) of the set of integers
(1, ..., 𝑘). The proofs of these properties are the same as in the finite-dimensional
case.

Definition 5.1.6 Let 𝔄 be an open subset of 𝑬 0 and 𝐹 a map 𝔄 −→ 𝑬 0 . We

shall say that the map 𝐹 is Ovsyannikov analytic at 𝑢 ◦ ∈ 𝔄 ∩ 𝑬 1 if there exist a
ball 𝔅𝑅 (𝑢 ◦ ) ⊂ 𝔄 and positive constants 𝐶◦ , 𝐶1 , such that, for every pair (𝑠, 𝑠 ′),
0 < 𝑠 ′ < 𝑠 < 1, and every 𝑘 ∈ Z+ , the Fréchet derivative (5.1.10) is well-defined
and satisfies the inequality (5.1.11) for all 𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) and all (𝑣 1 , ..., 𝑣 𝑘 ) ∈ 𝑬 𝑠𝑘 .

Proposition 5.1.7 If the map 𝐹 : 𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 is Ovsyannikov analytic at 𝑢 ◦ ∈ 𝑬 1

then there is a constant 𝛿 > 0 such that, given any pair (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1, and
any 𝑣 ∈ 𝔅𝑠, 𝛿 (0), the Taylor expansion
𝑘
∞
∑︁ 1 𝑘 ©z }| {ª
D 𝐹 (𝑢 ◦ ) ­­𝑣, ..., 𝑣 ®® (5.1.13)
𝑘=0
𝑘!
« ¬
converges uniformly to 𝐹 (𝑢 ◦ + 𝑣) in 𝑬 𝑠′ .

Proof We note that, if 𝑵 𝑠 (𝑣) < 𝑅 − 𝑵 𝑠 (𝑢 ◦ ), then the Lagrange remainder formula
holds:
𝑘
𝑚
◦
∑︁ 1 𝑘 ©z }| {ª
𝐹 (𝑢 + 𝑣) = D 𝐹 (𝑢 ◦ ) ­­𝑣, ..., 𝑣 ®® (5.1.14)
𝑘=0
𝑘!
« ¬
∫ 1 𝑚+1
1 ◦
©z }| {ª
+ D 𝐹 (𝑢 + 𝑡𝑣) ­𝑣, ..., 𝑣 ®® (1 − 𝑡) 𝑚 d𝑡,
𝑚+1 ­
𝑚! 0
« ¬
whence, by (5.1.11),
𝑘
𝑚
∑︁ 1 𝑘 ©z }| {ªª
𝑵 𝑠′ ­­𝐹 (𝑢 ◦ + 𝑣) − D 𝐹 (𝑢 ◦ ) ­­𝑣, ..., 𝑣 ®®®® ≤ 𝐶◦ 𝐶1𝑚+1 (𝑠 − 𝑠 ′) −1 𝑵 𝑠 (𝑣) 𝑚+1 .
©

𝑘=0
𝑘!
« « ¬¬
5.1 A Nonlinear Ovsyannikov Theorem 115

As 𝑚 ↗ +∞ the right-hand side converges to zero if 𝑵 𝑠 (𝑣) < 𝐶1−1 . Thus we can
select 𝛿 ≤ min 𝐶1−1 , 𝑅 − 𝑵 𝑠 (𝑢 ◦ ) .

□
Naturally, we will refer to (5.1.13) as the Taylor expansion of 𝐹 about 𝑢 ◦ .

Corollary 5.1.8 Let the map 𝐹 : 𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 be Ovsyannikov analytic at 𝑢 ◦ ∈

𝑬 1 . If D 𝑘 𝐹 (𝑢 ◦ ) = 0 for every 𝑘 ∈ Z+ then 𝐹 vanishes identically in 𝔅 𝛿 (𝑢 ◦ ).

Now let 𝑠◦ ∈ (0, 1) be arbitrary. The family 𝑬 𝜃 𝑠◦ 0< 𝜃 <1 is also a scale of Banach
spaces: it satisfies (5.1.5) with 𝐶1 replaced by 𝐶1 /𝑠◦ . It is evident that the preceding
reasoning applies to the scale 𝑬 𝜃 𝑠◦ 0< 𝜃 <1 . In particular, Proposition 5.1.7 is valid
for 𝑢 ◦ ∈ 𝑬 𝜃 𝑠◦ with 𝛿 replaced by 𝛿𝑠◦ . We can state

Corollary 5.1.9 Let the map 𝐹 : 𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 be Ovsyannikov analytic at 𝑢 ◦ ∈

𝑬 𝑠◦ (0 < 𝑠◦ < 1). If D 𝑘 𝐹 (𝑢 ◦ ) = 0 for every 𝑘 ∈ Z+ then 𝐹 vanishes identically in
𝔅 𝛿𝑠◦ (𝑢 ◦ ).

We conclude that the unique continuation property holds for Ovsyannikov analytic
maps, in the following sense:

Proposition 5.1.10 Let 𝔄 be a connected open subset of 𝑬 0 and the map 𝐹 : 𝔄 −→

𝑬 0 be Ovsyannikov analytic in 𝔄. If 𝑢 ◦ ∈ 𝔄 and D 𝑘 𝐹 (𝑢 ◦ ) = 0 for every 𝑘 ∈ Z+ then
𝐹 vanishes identically in 𝔄.

Standard proof: the subset of 𝔄 in which 𝐹 ≡ 0 is obviously closed; by Lemma

5.1.3 and Corollary 5.1.9 it is open.

5.1.2 The Ovsyannikov Theorem for a first-order analytic ODE

We introduce a complex variable 𝑧 = 𝑥 + 𝑖𝑦 in C and a disk Δ𝜌 = {𝑧 ∈ C; |𝑧| < 𝜌}

(𝜌 > 0). In the sequel 𝑢 ◦ will be an arbitrary element of 𝑬 1 ; the balls 𝔅𝑅 (𝑢 ◦ ),
𝔅𝑠,𝑅 (𝑢 ◦ ) are defined in (5.1.4). The main ingredient in the differential equation that
we propose to solve will be a map 𝐹 : Δ𝜌 × 𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 . As in the preceding
subsection, for every pair (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1, 𝑢 ↦→ 𝐹 (𝑧, 𝑢) is a map of 𝔅𝑠,𝑅 (𝑢 ◦ )
into 𝑬 𝑠′ , now for each 𝑧 ∈ Δ𝜌 . We hypothesize that 𝐹 is holomorphic with respect
to 𝑧 ∈ Δ𝜌 and Ovsyannikov analytic (Definition 5.1.6) with respect to 𝑢 ∈ 𝔅𝑅 (𝑢 ◦ );
more precisely:
(Ov) We have 𝜕¯𝑧 𝐹 (𝑧, 𝑢) = 0 for all (𝑧, 𝑢) ∈ Δ𝜌 × 𝔅𝑅 (𝑢 ◦ ) and there is a constant
𝐶◦ > 0 such that

𝑵 𝑠′ D 𝑘 𝐹 (𝑧, 𝑢) (𝑣 1 , ..., 𝑣 𝑘 ) ≤ 𝐶◦𝑘+1 𝑘! (𝑠 − 𝑠 ′) −1 𝑵 𝑠 (𝑣 1 ) · · · 𝑵 𝑠 (𝑣 𝑘 )
(5.1.15)
for every pair (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1, all 𝑘 ∈ Z+ , all (𝑧, 𝑢) ∈ Δ𝜌 × 𝔅𝑠,𝑅 (𝑢 ◦ )
and all 𝑘-tuples (𝑣 1 , ..., 𝑣 𝑘 ) ∈ 𝑬 𝑠𝑘 .
116 5 The Cauchy–Kovalevskaya Theorem

The Cauchy inequalities in the 𝑧-plane imply directly

Lemma 5.1.11 Suppose that (Ov) holds and that 0 < 𝑠 ′ < 𝑠 < 1. We have, for all
( 𝑗, 𝑘) ∈ Z2+ , all (𝑧, 𝑢) ∈ Δ𝜌/2 × 𝔅𝑠,𝑅 (𝑢 ◦ ) and all (𝑣 1 , ..., 𝑣 𝑘 ) ∈ 𝑬 𝑠𝑘 ,

𝑵 𝑠′ 𝜕𝑧 D 𝑘 𝐹 (𝑧, 𝑢) (𝑣 1 , ..., 𝑣 𝑘 ) ≤ (2/𝜌) 𝑗 𝐶◦𝑘+1 𝑗!𝑘! (𝑠 − 𝑠 ′) −1 𝑵 𝑠 (𝑣 1 ) · · · 𝑵 𝑠 (𝑣 𝑘 ) .
𝑗

(5.1.16)

In the proof of Theorem 5.1.13 below we are going to need the following lemma.

Lemma 5.1.12 Let 𝑘, 𝑝 be two integers such that 1 ≤ 𝑘 ≤ 𝑝. Then

𝑘
∑︁ ∑︁ 1 1 2 2
··· ≤ 2 𝜋 . (5.1.17)
𝛼1 ≥1 | 𝛼𝑘
𝛼2 · · · 𝛼2𝑘
| ≥1 1
𝑝 3
𝛼1 +···+𝛼𝑘 = 𝑝

Proof We reason by induction on 𝑘. When 𝑘 = 1, (5.1.17) is the statement that

𝜋 2 > 23 . Let us call Ξ 𝑝,𝑘 the left-hand side in (5.1.17). When 𝑘 ≥ 2, the induction
hypothesis yields
𝑝−𝑘+1 𝑝−1
𝑘−1 ∑︁
∑︁ 1 2 2 1 1
Ξ 𝑝,𝑘 = Ξ 𝑝−ℓ,𝑘−1 ≤ 𝜋 .
ℓ=1
ℓ2 3 ℓ=1
ℓ 2
( 𝑝 − ℓ) 2

We have
𝑝−1 𝑝−1 2
2
∑︁ 1 1 ∑︁ 1 1
𝑝 = +
ℓ=1
ℓ 2 ( 𝑝 − ℓ) 2 ℓ=1 ℓ 𝑝 − ℓ
𝑝−1 ∞
∑︁ 1 1 ∑︁ 1 2
≤2 2
+ 2
≤ 4 2
= 𝜋2 . □
ℓ=1
ℓ ( 𝑝 − ℓ) ℓ=1
ℓ 3

Theorem 5.1.13 Let 𝐹 : Δ𝜌 ×𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 satisfy Hypothesis (Ov). For arbitrary

𝑢 ◦ ∈ 𝑬 1 the following is true:
(Existence) For 𝜀 > 0 sufficiently small there is a map

Δ 𝜀 ∋ 𝑧 ↦→ 𝑢 (𝑧) ∈ 𝔅𝑅 (𝑢 ◦ ) (5.1.18)

with the following properties:

(1) for each 𝑠 ∈ (0, 1), 𝑢 is a holomorphic map

Δ (1−𝑠) 𝜀 −→ 𝔅𝑠,𝑅 (𝑢 ◦ ) ; (5.1.19)

(2) in the open disk Δ 𝜀 ,

𝜕𝑧 𝑢 = 𝐹 (𝑧, 𝑢) ; (5.1.20)
5.1 A Nonlinear Ovsyannikov Theorem 117

(3) 𝑢 (0) = 𝑢 ◦ .
(Uniqueness) If for some 𝑠 ′, 0 < 𝑠 ′ < 1, 𝜀 ′ > 0, 𝑅 ′ > 0, a holomorphic map
𝑣 : Δ 𝜀′ −→ 𝔅𝑠′ ,𝑅′ (𝑢 ◦ ) satisfies (5.1.20) in Δ 𝜀′ and if 𝑣 (0) = 𝑢 ◦ then there is a
𝛿 > 0, 𝛿 ≤ min (𝜀, 𝜀 ′), such that 𝑣 (𝑧) = 𝑢 (𝑧) for all 𝑧 ∈ Δ 𝛿 .

Note that the ODE in (5.1.20) makes sense in 𝑬 𝑠′ , to which 𝐹 (𝑧, 𝑢) belongs if
0 < 𝑠 ′ < 𝑠 < 1 and if 𝑧 ∈ Δ (1−𝑠) 𝜀 .
The problem of solving (5.1.20) with the prescribed value 𝑢 (0) = 𝑢 ◦ is called a
Cauchy problem; the “initial” value 𝑢 ◦ is often referred to as the Cauchy data.
Proof In terms of the Taylor expansion (5.1.13) the ODE in (5.1.20) reads
𝑘
∞
1 𝑘 ©z
∑︁ }| {ª
𝜕𝑧 𝑢 = D 𝐹 (𝑧, 𝑢 ◦ ) ­𝑢 (𝑧) − 𝑢 ◦ , ..., 𝑢 (𝑧) − 𝑢 ◦ ® . (5.1.21)
­ ®
𝑘=0
𝑘! ­ ®
« ¬
We (𝑧) =
Í∞ begin 𝑚by reasoning formally:◦ let 𝑢 be given as a formal power series 𝑢
𝑚=0 𝑢 𝑚 𝑧 , 𝑢 𝑚 ∈ 𝑬 𝑠 and 𝑢 0 = 𝑢 ; (5.1.21) translates into

∞ ∞ ∑︁∞ ∞ ∞
!
∑︁
𝑚−1
∑︁ 1 𝑗 𝑗 𝑘 ◦
∑︁
ℓ1
∑︁
ℓ𝑘
𝑚𝑢 𝑚 𝑧 = 𝑧 𝜕𝑧 D 𝐹 (0, 𝑢 ) 𝑢 ℓ1 𝑧 , ..., 𝑢 ℓ𝑘 𝑧 .
𝑚=1 𝑗=0 𝑘=0
𝑗!𝑘! ℓ =1 ℓ =1
1 𝑘
(5.1.22)
Keep in mind that 𝜕𝑧 D 𝑘 𝐹 (0, 𝑢 ◦ ) is a 𝑘-linear map 𝑬 𝑠𝑘 −→ 𝑬 𝑠′ , which enables us to
𝑗

rewrite the right-hand side in (5.1.22) as a power series in 𝑧:

∞ ∞ ∑︁
𝑚
∑︁ ∑︁ ∑︁ 1 𝑗 𝑘
𝜕𝑧 D 𝐹 (0, 𝑢 ◦ ) 𝑢 ℓ1 , ..., 𝑢 ℓ𝑘 ,

(𝑚 + 1) 𝑢 𝑚+1 𝑧 𝑚 = 𝑧𝑚
𝑚=0 𝑚=0 𝑘=0 𝑗+ℓ1 +···+ℓ𝑘 =𝑚
𝑗!𝑘!
ℓ 𝛼 ≥1, 𝛼=1,...,𝑘
(5.1.23)
equivalent to the sequence of “transport equations”
1
𝑢 𝑚+1 = 𝜕 𝑚 𝐹 (0, 𝑢 ◦ ) (5.1.24)
(𝑚 + 1)! 𝑧
𝑚
1 ∑︁ ∑︁ 1 𝑗 𝑘
𝜕 D 𝐹 (0, 𝑢 ◦ ) 𝑢 ℓ1 , ..., 𝑢 ℓ𝑘 .

+
𝑚 + 1 𝑘=1 𝑗+ℓ +···+ℓ =𝑚 𝑗!𝑘! 𝑧
1 𝑘
ℓ 𝛼 ≥1, 𝛼=1,...,𝑘

On the right-hand side we have necessarily ℓ 𝛼 ≤ 𝑚 for every 𝛼 = 1, ..., 𝑘. Thus

𝑢 𝑚+1 is determined unambiguously by (5.1.24). By Corollary 5.1.8 this entails the
uniqueness part in Theorem 5.1.13.
In the remainder of the proof 𝑠 ∈ (0, 1) will be fixed but arbitrary. We must now
show that the power series ∞
Í
𝑚=0 𝑢 𝑚 𝑧 converges uniformly in 𝑬 𝑠 when 𝑧 ∈ Δ (1−𝑠) 𝜀
𝑚

for a suitable choice of 𝜀 ∈ (0, 𝜌). First of all we note that (5.1.24) implies 𝑢 1 =
𝐹 (0, 𝑢 ◦ ). For 𝑚 ≥ 1 we derive from (5.1.16) and (5.1.24):
118 5 The Cauchy–Kovalevskaya Theorem
𝑚
1 ∑︁ 𝑘+1 ∑︁
(2/𝜌) 𝑗 𝛿−1

𝑵 𝑠 (𝑢 𝑚+1 ) ≤ 𝐶 𝑗 𝑵 𝑠+ 𝛿 𝑗 𝑢 ℓ1 · · · 𝑵 𝑠+ 𝛿 𝑗 𝑢 ℓ𝑘 ,
𝑚 + 1 𝑘=0 ◦ 𝑗+ℓ1 +···+ℓ𝑘 =𝑚
ℓ 𝛼 ≥1, 𝛼=1,...,𝑘
(5.1.25)
1−𝑠
where 𝛿 𝑗 > 0, 𝛿 𝑗 = 𝑚− 𝑗+1 . By using induction on 𝑚 = 1, 2, ..., we are going to
deduce from (5.1.25): 𝑚
𝜅 𝐵
𝑵 𝑠 (𝑢 𝑚 ) ≤ 2 (5.1.26)
𝑚 1−𝑠
for suitable positive constants 𝜅, 𝐵. In this connection recall that, according to (5.1.5),
we have
(1 − 𝑠) 𝑵 𝑠 (𝐹 (0, 𝑢 ◦ )) ≤ 𝐶◦ .
In view of this we begin by requiring 𝐵𝜅 ≥ 𝐶◦ , thus ensuring that (5.1.26) is true
when 𝑚 = 1. Putting 𝑧 = 0, 𝑗 = 𝑚, 𝑘 = 0, in (5.1.16) we derive, for 𝑚 > 1:
1 𝐶◦
𝑵 𝑠′ 𝜕𝑧𝑚 𝐹 (0, 𝑢 ◦ ) ≤

(2/𝜌) 𝑚
𝑚! 𝑠 − 𝑠′
and for 𝑗 < 𝑚, 𝑘 ≥ 1,

1 𝐶 𝑘+1
𝑵 𝑠′ 𝜕𝑧 D 𝑘 𝐹 (0, 𝑢 ◦ ) (𝑣 1 , ..., 𝑣 𝑘 ) ≤ (2/𝜌) 𝑗 ◦ ′ 𝑵 𝑠 (𝑣 1 ) · · · 𝑵 𝑠 (𝑣 𝑘 ) .
𝑗
𝑗!𝑘! 𝑠−𝑠
(5.1.27)
In the last two estimates we take 𝑠 = 1 and replace 𝑠 ′ by 𝑠 + 𝛿 𝑗 < 1. When
𝑘 ≥ 1 the induction hypothesis yields
𝑚− 𝑗
𝜅𝑘 𝐵
𝑵 𝑠+ 𝛿 𝑗 𝑢 ℓ1 · · · 𝑵 𝑠+ 𝛿 𝑗 𝑢 ℓ𝑘 ≤
ℓ12 · · · ℓ𝑘2 1 − 𝑠 − 𝛿 𝑗
𝑚− 𝑗 𝑚− 𝑗
𝜅𝑘 𝑚− 𝑗 +1 𝐵
= 2 .
ℓ1 · · · ℓ𝑘2 𝑚− 𝑗 1−𝑠

(Notice that 𝑘 = 𝑚 =⇒ 𝑗 = 0 and that we always have 𝑗 + 𝑘 ≤ 𝑚.) We get

𝑚− 𝑗
𝜅𝑘 𝐵
𝑵 𝑠+ 𝛿 𝑗 𝑢 ℓ1 · · · 𝑵 𝑠+ 𝛿 𝑗 𝑢 ℓ𝑘 ≤ 𝑒 .
ℓ12 · · · ℓ𝑘2 1 − 𝑠

Combining this last inequality with (5.1.25) and (5.1.27) yields

𝐶◦ 1
𝑵 𝑠 (𝑢 𝑚+1 ) ≤ (2/𝜌) 𝑚 (5.1.28)
𝑚+1 1−𝑠
−1
𝑚
𝑚+1 ∑︁ 𝑗
𝐵 𝐶◦ e 𝐵 ∑︁ 2 1
+ (𝜅𝐶◦ ) 𝑘 .
𝑚+1 1−𝑠 𝑘=1 𝑗+ℓ1 +···+ℓ𝑘 =𝑚
𝐵𝜌 ℓ12 · · · ℓ𝑘2
ℓ 𝛼 ≥1, 𝛼=1,...,𝑘

Taking 𝐵−1 𝐶◦ ≤ 𝜅 into account yields

5.1 A Nonlinear Ovsyannikov Theorem 119

𝑵 𝑠 (𝑢 𝑚+1 )
𝑚+1 𝑚 𝑚 𝑚−𝑘 𝑗
𝜅 𝐵 © 2 ∑︁ ∑︁ 2
≤ +e (𝜅𝐶◦ ) 𝑘 (𝑚 − 𝑗 + 1) Ξ𝑚− 𝑗,𝑘 ® ,
ª
𝑚+1 1−𝑠
­
𝐵𝜌 𝑘=1 𝑗=0
𝐵𝜌
« ¬
where Ξ 𝑝,𝑘 is the quantity in the left-hand side in (5.1.17). By Lemma 5.1.12 (and
changing summation indices to 𝑚 − 𝑗 = ℓ) we reach the following conclusion:
𝑚+1 𝑚 𝑚 ∑︁𝑚 𝑘 𝑚−ℓ !
𝜅 𝐵 2 ∑︁ ℓ+1 2 2 2
𝑵 𝑠 (𝑢 𝑚+1 ) ≤ +e 𝜋 𝜅𝐶◦ .
𝑚+1 1−𝑠 𝐵𝜌 ℓ=1 𝑘=ℓ
ℓ2 3 𝐵𝜌

3
−1
Selecting 𝜅 ≤ 4 𝜋 2 𝐶◦ yields
𝑚+1
𝜅 𝐵
𝑵 𝑠 (𝑢 𝑚+1 ) ≤ 𝑐 𝑚 (5.1.29)
(𝑚 + 1) 2 1−𝑠

where
𝑚 ℓ !
∑︁ 1 2
𝑐 𝑚 = (𝑚 + 1) (2/𝐵𝜌) 𝑚 + 4e 𝜋 2 𝜅𝐶◦ (2/𝐵𝜌) 𝑚−ℓ .
ℓ=1
ℓ 3
√︁ 𝑞
We require 𝐵 to be large enough that 𝐵𝜌/2 ≥ 𝑞 + 1 for every 𝑞 ∈ Z+ . We get

𝑚 ℓ
√︁ 𝑚 ∑︁ 𝑚+1 2 2 √︁ 𝑚−ℓ
𝑐𝑚 ≤ 2/𝐵𝜌 + 4e 𝜋 𝜅𝐶◦ 2/𝐵𝜌
ℓ=1
ℓ(𝑚 + 1 − ℓ) 3
𝑚
√︁ 𝑚 2
≤ 2/𝐵𝜌 + 8e 𝜋 2 𝜅𝐶◦
3
𝑚−1 ℓ
∑︁ 𝑚+1 2 2 √︁ 𝑚−ℓ
+ 4e 𝜋 𝜅𝐶◦ 2/𝐵𝜌 .
ℓ=1
ℓ(𝑚 + 1 − ℓ) 3

If 1 ≤ ℓ ≤ 𝑚 − 1 then
𝑚+1 𝑚 𝑚!
≤2 ≤2 .
ℓ(𝑚 + 1 − ℓ) ℓ(𝑚 − ℓ) ℓ!(𝑚 − ℓ)!
We conclude that 𝑚
2 2 √︁
𝑐 𝑚 ≤ 8e 𝜋 𝜅𝐶◦ + 2/𝐵𝜌
3
and thus 𝑐 𝑚 ≤ 1 whatever the integer 𝑚 ≥ 1, provided 𝜅 and 𝐵−1 are sufficiently
small. This proves (5.1.26) with 𝑚 + 1 in place of 𝑚. The estimate (5.1.26) for all
𝑚 ∈ Z+ implies that the power series ∞
Í
𝑚=0 𝑢 𝑚 𝑧 converges uniformly in 𝑬 𝑠 in the
𝑚
−1
disk |𝑧| ≤ 𝐵 (1 − 𝑠). □
120 5 The Cauchy–Kovalevskaya Theorem

Remark 5.1.14 A more general version of Theorem 5.1.13 is proved in [Nirenberg,

1972] using the Nash–Moser Implicit Function Theorem: analyticity with respect to
the complex variable 𝑧 is replaced by continuity with respect to a real-variable 𝑡.

5.2 Application: the Nonlinear Cauchy–Kovalevskaya Theorem

5.2.1 Well-posed Cauchy problem for a holomorphic fully nonlinear

PDE

We shall limit our attention to the holomorphic version of the Cauchy–Kovalevskaya

Theorem. The real-analytic version ensues by restricting the variables to real space.
We consider an equation of order 𝑚 ≥ 1,

𝑃 (𝑧, 𝑢, 𝜕𝑢, ..., 𝜕 𝑚 𝑢) = 0, (5.2.1)

in a polydisk Δ𝑟𝑛 = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 ; 𝑟 = (𝑟 1 , ..., 𝑟 𝑛 ) is the polyra-
dius; we assume that 𝑟 𝑗 > 0 for every 𝑗. In (5.2.1) 𝑢 will be a holomorphic function
of 𝑧 in Δ𝑟𝑛 ; all partial derivatives are holomorphic derivatives, i.e., compositions of
the operators 𝜕𝑧𝜕 𝑗 ( 𝑗 = 1, ..., 𝑛); thus 𝜕 stands for the complex gradient, 𝜕 𝑘 for the set
of partial derivatives
𝛼1 𝛼𝑛
𝜕 𝜕
𝜕𝑧𝛼 = ··· , 𝛼 = (𝛼1 , ..., 𝛼𝑛 ) ∈ Z+ , |𝛼| = 𝛼1 +· · ·+𝛼𝑛 = 𝑘. (5.2.2)
𝜕𝑧1 𝜕𝑧 𝑛

It is convenient to introduce the symbol of 𝑃 (𝑧, 𝑢, 𝜕𝑢, ..., 𝜕 𝑚 𝑢),

𝑝 (𝑧, 𝑤) = 𝑃 𝑧, 𝑤 0 , 𝑤 𝑗 𝑗=1..,𝑛 , ..., (𝑤 𝛼 ) | 𝛼 |=𝑚 , (5.2.3)

a holomorphic function in a suitably large domain where 𝑧, (𝑤 𝛼 ) | 𝛼 | ≤𝑚 varies
(each 𝑤 𝛼 is scalar); and the following linear PDO
∑︁ 𝜕 𝑝
𝑝 lin (𝑧, 𝑤, 𝜕𝑧 ) = (𝑧, 𝑤) 𝜕𝑧𝛼 . (5.2.4)
𝛼∈Z 𝑛 𝜕𝑤 𝛼
+

The order of 𝑝 lin (𝑧, 𝑤, 𝜕𝑧 ) does not exceed 𝑚.

Given a function 𝑢 ◦ ∈ O Δ𝑟𝑛 we may introduce the linearization of (5.2.1) at 𝑢 ◦ :
∑︁ 𝜕𝑝
𝑃lin,𝑢◦ (𝑧, 𝜕𝑧 ) = (𝑧, 𝑢 ◦ (𝑧) , 𝜕𝑢 ◦ (𝑧) , ..., 𝜕 𝑚 𝑢 ◦ (𝑧)) 𝜕𝑧𝛼 (5.2.5)
𝛼∈Z+𝑚
𝜕𝑤 𝛼

and its principal symbol,

5.2 Application: the Nonlinear Cauchy–Kovalevskaya Theorem 121
∑︁ 𝜕𝑝
(𝑧, 𝑢 ◦ (𝑧) , 𝜕𝑢 ◦ (𝑧) , ..., 𝜕 𝑚 𝑢 ◦ (𝑧)) 𝜁 𝛼 . (5.2.6)
𝜕𝑤 𝛼
| 𝛼 |=𝑚

Our basic hypothesis will be that (5.2.6) does not vanish identically at 𝑧 = 0. Possibly
after a linear change of the variables 𝑧 𝑗 we may assume that

𝜕𝑝
(0, 𝑢 ◦ (0) , 𝜕𝑢 ◦ (0) , ..., 𝜕 𝑚 𝑢 ◦ (0)) ≠ 0. (5.2.7)
𝜕𝑤 (𝑚,0,...,0)

This enables us to apply the Implicit Function Theorem in the holomorphic category
and write
𝑝 (𝑧, 𝑤) = 𝑞 (𝑧, 𝑤) 𝑤 (𝑚,0,...,0) − 𝑓 (𝑧, 𝑤 ′)

(5.2.8)
in a suitably small neighborhood of 𝑧 = 0 and 𝑤 ◦ = (𝜕 𝛼 𝑢 ◦ (0)) | 𝛼 | ≤𝑚 in which the
factor 𝑞 (𝑧, 𝑤) does not vanish. We have also used the notation

𝑤 ′ = (𝑤 𝛼 ) | 𝛼 | ≤𝑚, 𝛼≠(𝑚,0,...,0) .

This means that if the radii 𝑟 𝑗 > 0 are sufficiently small and if the holomorphic
function 𝑢 is sufficiently close to 𝑢 ◦ in O Δ𝑟𝑛 , Eq. (5.2.1) is equivalent to the

equation

𝜕𝑧𝑚1 𝑢 = 𝑓 𝑧, 𝜕𝑧𝛼 𝑢 | 𝛼 | ≤𝑚, 𝛼≠(𝑚,0,...,0) . (5.2.9)

Remark 5.2.1 The meaning of (5.2.7) is that the covector 𝜉 ◦ = (1, 0, ..., 0) at the
origin is not characteristic for the linear differential operator (5.2.5) [i.e., (0, 𝜉 ◦ ) ∉
Char 𝑃lin,𝑢◦ ; cf. Definition 1.3.5]. It implies that the hyperplane 𝑧 1 = 0, as well
as any complex hypersurface Σ close to it and passing through the origin of C𝑛 , is
noncharacteristic at 0 for 𝑃lin,𝑢◦ (cf. loc. cit.). One then says that Σ is noncharacteristic
at 0 for the nonlinear equation (5.2.1).

With Eq. (5.2.9) or, equivalently, with (5.2.1) we associate the following “initial”
(or Cauchy) conditions:

𝜕𝑧𝑘1 𝑢 (0, 𝑧 ′) = 𝑢 (𝑘) (𝑧 ′) , 𝑘 = 0, ..., 𝑚 − 1, (5.2.10)

where the 𝑢 (𝑘) are holomorphic functions of the variable 𝑧 ′ = (𝑧 2 , ..., 𝑧 𝑛 ) in some
neighborhood of 0 in C𝑛−1 .

5.2.2 Reduction to a quasilinear first-order square system

Our objective, here, is to transform (5.2.9) into a system of first-order PDEs and
(5.2.10) into the corresponding Cauchy conditions. The equation (5.2.9) is scalar but
the argument that follows could also be applied to a system of 𝜈 × 𝜈 equations of the
same type, with the values of 𝑢, 𝑓 and 𝑢 ( 𝑗) belonging to C𝜈 . Such a vector equation
is usually called a square system of PDEs or, more precisely, a 𝜈 × 𝜈 system of PDEs.
122 5 The Cauchy–Kovalevskaya Theorem

We reason by induction on the order 𝑚. There is nothing to do when 𝑚 = 1;

𝜕𝑢
suppose 𝑚 ≥ 2. We introduce the unknowns 𝑢 𝑗 = 𝜕𝑧 𝑗
, 𝑗 = 1, ..., 𝑛. To simplify the
notation we set 𝑢 𝑛+1 = 𝑢. Let ⟨ 𝑗⟩ denote the 𝑛-tuple 𝛾 = (𝛾1 , ..., 𝛾𝑛 ) such that 𝛾𝑖 = 0
if 𝑖 ≠ 𝑗 and 𝛾 𝑗 = 1. Given any 𝛼 ∈ Z+𝑛 , |𝛼| ≥ 2, we set 𝑇 (𝛼) = 𝛼 − ⟨ 𝑗⟩ where 𝑗 is the
largest integer such that 𝛼 𝑗 ≠ 0. We have 𝜕𝑧𝛼 𝑢 = 𝜕𝑧𝑇 ( 𝛼) 𝑢 𝑗 . We can rewrite (5.2.9) as

𝑚−1 𝛽
𝜕𝑧1 𝑢 1 = 𝜑 𝑧, 𝜕𝑧 𝑢 𝑗 (5.2.11)
|𝛽 | ≤𝑚−1, 𝑗=1,...,𝑛+1

with 𝜑 related to 𝑓 in an obvious manner; in particular, the right-hand side does

not involve 𝜕𝑧𝑚−1
1
𝑢 1 . To the system of equations (5.2.11) we adjoin the following
equations

𝜕𝑧𝑚−1
1
𝑢 𝑗 = 𝜕𝑧 𝑗 𝜕𝑧𝑚−2
1
𝑢 1 , 𝑗 = 2, ..., 𝑛, (5.2.12)
𝜕𝑧𝑚−1
1
𝑢 𝑛+1 = 𝜕𝑧𝑚−2
1
𝑢1 .

The conditions (5.2.10) are equivalent to

𝑢 𝑛+1 (0, 𝑧 ′) = 𝑢 (0) (𝑧 ′) , (5.2.13)

′ (𝑘+1) ′
𝜕𝑧𝑘1 𝑢 1 (0, 𝑧 ) = 𝑢 (𝑧 ) , 𝑘 = 0, 1, ..., 𝑚 − 2;

they imply, for 𝑘 = 0, 1, ..., 𝑚 − 1,

𝜕𝑧𝑘1 𝑢 𝑗 (0, 𝑧 ′) = 𝜕𝑧 𝑗 𝑢 (𝑘) (𝑧 ′) , 𝑗 = 2, ..., 𝑛, (5.2.14)

′ (𝑘) ′
𝜕𝑧𝑘1 𝑢 𝑛+1 (0, 𝑧 ) = 𝑢 (𝑧 ) . (5.2.15)

Remark 5.2.2 When we put 𝑧1 = 0 in (5.2.12) we deduce the relations 𝜕𝑧𝑚−1 1

𝑢 𝑗 (0, 𝑧 ′)
= 𝜕𝑧 𝑗 𝜕𝑧𝑚−2
1
𝑢 1 (0, 𝑧 ′) = 𝜕𝑧 𝑗 𝑢 (𝑚−1) (𝑧 ′), 𝑗 = 2, ..., 𝑛. Similarly, we deduce from
(5.2.12) the relation 𝜕𝑧𝑚−11
𝑢 𝑛+1 (0, 𝑧 ′) = 𝑢 (𝑚−1) (𝑧 ′). This shows that nothing would
have been lost if we had required 𝑘 ≤ 𝑚 − 2 in (5.2.14)–(5.2.15).

Combining (5.2.12) with (5.2.11), in which we have made the substitutions

(5.2.12), we obtain a system of equations of order 𝑚 − 1,

𝛽
𝜕𝑧𝑚−1
1
𝑈 = Φ 𝑧, 𝜕𝑧 𝑈 , (5.2.16)
|𝛽 | ≤𝑚−1

where Φ and 𝑈 = (𝑢 1 , ..., 𝑢 𝑛+1 ) are valued in C𝜈 (𝑛+1) and 𝜕𝑧𝑚−1

1
𝑈 does not enter in the
right-hand side. The Cauchy conditions (5.2.13)–(5.2.14)–(5.2.15) can be equated
to the conditions:

𝜕𝑧𝑘1 𝑈 (0, 𝑧 ′) = 𝑈 (𝑘) (𝑧 ′) , 𝑘 = 0, 1, ..., 𝑚 − 2, (5.2.17)

where 𝑈 (𝑘) (𝑧 ′) = 𝑢 1(𝑘) (𝑧 ′) , ..., 𝑢 𝑛+1
(𝑘)
(𝑧 ′) (cf. Remark 5.2.2).
5.2 Application: the Nonlinear Cauchy–Kovalevskaya Theorem 123

Induction on 𝑚 allows us to conclude that (5.2.9)–(5.2.10) is equivalent to a

Cauchy problem for a first-order square system, of the following type:

𝜕𝑧1 𝑈 = Φ 𝑧, 𝑈, 𝜕𝑧2 𝑈, ..., 𝜕𝑧𝑛 𝑈 , (5.2.18)
′ ◦ ′
𝑈 (0, 𝑧 ) = 𝑈 (𝑧 ) . (5.2.19)

The precise dimensions of the ranges are not important; we shall simply say that Φ,
𝑈 and 𝑈 ◦ are valued in one and the same Euclidean space C𝑞 .
It is convenient to go two steps further. Firstly, by setting 𝑈 𝑗 = 𝜕𝑧 𝑗 𝑈 ( 𝑗 = 2, ..., 𝑛)
and also 𝑈1 = 𝑈, (5.2.18) can be rewritten as

𝜕𝑧1 𝑈1 = Φ (𝑧, 𝑈1 , 𝑈2 , ..., 𝑈𝑛 ) , (5.2.20)

𝜕𝑧1 𝑈 𝑗 = 𝜕𝑧 𝑗 Φ (𝑧, 𝑈1 , 𝑈2 , ..., 𝑈𝑛 ) , 𝑗 = 2, ..., 𝑛.

Secondly, we transform 𝑧 ′ = (𝑧2 , ..., 𝑧 𝑛 ) from variable to unknown (we prefer not to
regard it as a parameter). We adjoin to (5.2.20) the anodine equations:

𝜕𝑧1 𝑧 𝑗 = 0, 𝑗 = 2, ..., 𝑛. (5.2.21)

Thus the values of all the unknowns combined will belong to C𝑛𝑞+𝑛−1 . We formulate
a Cauchy problem for the system (5.2.20)–(5.2.21) by assigning a set of values:

𝑧𝑗 𝑧1 =0
= 𝑧 𝑗 , 𝑗 = 2, ..., 𝑛; (5.2.22)
𝑈𝑗 𝑧1 =0
= 𝑈 ◦𝑗 (𝑧 ′) , 𝑗 = 1, ..., 𝑛. (5.2.23)

All the Cauchy data 𝑈 ◦𝑗 (𝑧 ′) will be C𝑞 -valued holomorphic functions of 𝑧 ′ in some

polydisk centered at the origin in C𝑛−1 .

5.2.3 End of the proof of the Cauchy–Kovalevskaya Theorem

We must reformulate the system of equations (5.2.20)–(5.2.21) with the “initial”

conditions (5.2.22)–(5.2.23) as a Cauchy problem (5.1.20). This demands that we be
more precise about all our data. We recall that (𝑈1 , ..., 𝑈𝑛 ) varies in C𝑛𝑞 . We return
to the notation 𝑢, this time meaning the vector-valued unknown

𝑢 = (𝑧 ′, 𝑈1 , ..., 𝑈𝑛 ) ∈ C𝑛−1+𝑛𝑞 . (5.2.24)

In the application of Theorem 5.1.13 the Banach space 𝑬 𝑠 will be the space of
C𝑛−1+𝑛𝑞 -valued holomorphic functions of 𝑧 ′ in the polydisk
(𝑛−1)
= 𝑧 ′ ∈ C𝑛−1 ; 𝑧 𝑗 < 𝑠𝑟, 𝑗 = 2, ..., 𝑛

Δ𝑠𝑟
124 5 The Cauchy–Kovalevskaya Theorem

(𝑛−1)
that can be extended continuously to the closure Δ𝑠𝑟 , with 𝑟 > 0 fixed and
0 < 𝑠 < 1. Recall that 𝑬 0 is the union of the spaces 𝑬 𝑠 . We denote by 𝐹 (𝑧1 , 𝑢) the
map
𝑛−1
©z }| { ª
𝑬 0 ∋ 𝑢 ↦→ ­0, ..., 0, Φ (𝑧, 𝑈1 , ..., 𝑈𝑛 ) , 𝜕𝑧 𝑗 Φ (𝑧, 𝑈1 , ..., 𝑈𝑛 ) ® ∈ C𝑛−1+𝑛𝑞 .
­ ®
­ 𝑗=2,...,𝑛 ®
« ¬
(5.2.25)
We assume that 𝐹 (𝑧1 , 𝑢) depends holomorphically on 𝑧 1 ∈ Δ𝑟1 (𝑟 1 > 0); 𝑧1 plays
the role of the single complex variable 𝑧 in (5.1.20).
We select the Cauchy data in (5.2.23) to be continuous functions of 𝑧 ′ in Δ𝑟𝑛−1 ,
holomorphic in Δ𝑟𝑛−1 . Combining the initial values in (5.2.22)–(5.2.23) yields an
element 𝑢 ◦ ∈ 𝑬 1 . The range of the map Δ𝑟𝑛−1 ∋ 𝑧 ′ ↦→ 𝑢 ◦ (𝑧 ′) is a compact subset 𝐾
of C𝑛−1+𝑛𝑞 . Obviously, by (5.2.22), the projection of 𝐾 into C𝑛−1+𝑛𝑞 = C𝑛−1 is equal
to Δ𝑟𝑛−1
We must require that the function 𝐹 (𝑧1 , 𝑢), or equivalently Φ (𝑧, 𝑈1 , ..., 𝑈𝑛 ), be
defined and holomorphic in Δ𝑟 × Ω, where Ω ⊂ C𝑛−1+𝑛𝑞 is an open set that contains
𝐾. Actually we select 𝑅◦ > 0 such that

𝐾 𝑅◦ = 𝑤 ∈ C𝑛−1+𝑛𝑞 ; dist (𝑤, 𝐾) ≤ 𝑅◦ ⊂⊂ Ω. (5.2.26)

Since
sup |𝑢 (𝑧 ′) − 𝑢 ◦ (𝑧 ′)| ≤ 𝑁 𝑠 (𝑢 − 𝑢 ◦ ) (5.2.27)
𝑧 ′ ∈Δ𝑛−1
𝑠𝑟

(5.2.26) ensures that the following condition is satisfied, whatever 𝑅 ∈ (0, 𝑅◦ ],

(★) If 𝑠 ∈ (0, 1) and 𝑢 ∈ 𝔅𝑠,𝑅 (𝑢 ◦ ) then the range of the map Δ𝑛−1 ′ ′
𝑠𝑟 ∋ 𝑧 ↦→ 𝑢 (𝑧 ) is
contained in 𝐾 𝑅 .

Lemma 5.2.3 Under the above hypotheses the map (5.2.25) is Ovsyannikov analytic
(Definition 5.1.6) at 𝑢 ◦ ∈ 𝑬 1 in the scale {𝑬 𝑠 }0<𝑠<1 .

Proof It is convenient to use both notations 𝐹 (𝑧 1 , 𝑢) and 𝐹 (𝑧, 𝑈1 , ..., 𝑈𝑛 ). The

Fréchet derivative (with respect to 𝑢) of the map (5.2.25) is given by
𝑛 𝑛
∑︁ 𝜕𝐹 ∑︁ 𝜕𝐹
D𝑢 𝐹 (𝑧1 , 𝑢) (ℎ, 𝑣) = ℎ𝑗 + 𝑣𝑘, (5.2.28)
𝑗=2
𝜕𝑧 𝑗 𝑘=1
𝜕𝑈 𝑘

where ℎ = (ℎ2 , ..., ℎ 𝑛 ) ∈ C𝑛−1 , 𝑣 = (𝑣 1 , ..., 𝑣 𝑛 ) ∈ C𝑛𝑞 and

𝑛−1
2
𝜕𝐹 ©­z }| { 𝜕Φ 𝜕 Φ ª
= ­0, ..., 0, (𝑧, 𝑈1 , ..., 𝑈𝑛 ) , (𝑧, 𝑈1 , ..., 𝑈𝑛 ) ® , (5.2.29)
®
𝜕𝑧 𝑗 ­ 𝜕𝑧 𝑗 𝜕𝑧 𝑗 𝜕𝑧ℓ ℓ=2,...,𝑛
®
« ¬
5.2 Application: the Nonlinear Cauchy–Kovalevskaya Theorem 125

𝑛−1
2
𝜕𝐹 ©z }| { 𝜕Φ 𝜕 Φ ª
= ­0, ..., 0, (𝑧, 𝑈1 , ..., 𝑈𝑛 ) , (𝑧, 𝑈1 , ..., 𝑈𝑛 ) ® . (5.2.30)
­ ®
𝜕𝑈 𝑘 ­ 𝜕𝑈 𝑘 𝜕𝑈 𝑘 𝜕𝑧ℓ ℓ=2,...,𝑛 ®
« ¬
More generally, whatever 𝑀 ≥ 1,
∑︁
𝛽
D𝑢𝑀 𝐹 (𝑧1 , 𝑢) (ℎ, 𝑣) = 𝜕𝑧𝛼′ 𝜕𝑈 𝐹 (𝑧, 𝑈1 , ..., 𝑈𝑛 ) ℎ 𝛼 𝑣 𝛽 (5.2.31)
𝛼+𝛽=𝑀

𝛽 𝛽 𝛽
in the multi-index notation 𝜕𝑧𝛼 = 𝜕𝑧𝛼22 · · · 𝜕𝑧𝛼𝑛𝑛 , 𝜕𝑈 = 𝜕𝑈11 · · · 𝜕𝑈𝑛𝑛 , ℎ 𝛼 = ℎ 𝛼2 · · · ℎ 𝛼𝑛 ,
𝑣 𝛽 = 𝑣 𝛽1 · · · 𝑣 𝛽𝑛 , etc. We have, by (5.2.25),

2 2 𝑛
∑︁ 2
𝛽 𝛽 𝛽
𝜕𝑧𝛼′ 𝜕𝑈 𝐹 = 𝜕𝑧𝛼′ 𝜕𝑈 Φ + 𝜕𝑧𝛼′ 𝜕𝑈 𝜕𝑧𝑘 Φ .
𝑘=2

In what follows 𝑅 ∈ (0, 𝑅◦ ) will be fixed. Let 𝑠, 𝑠 ′ ∈ (0, 1), 𝑠 ′ < 𝑠, and 𝑢 ∈
𝔅𝑠,𝑅 (𝑢 ◦ ) be arbitrary. We assume that (★) holds. For 𝑧 1 ∈ Δ𝑟 the Cauchy inequalities
imply

max 𝜕𝑧𝛼′ 𝜕𝑈 Φ (𝑧 1 , 𝑢 (𝑧 ′)) ≤ max 𝜕𝑧𝛼′ 𝜕𝑈 Φ (𝑧 1 , ·)

𝛽 𝛽
|𝑧 ′ | ≤𝑠𝑟 𝐾𝑅

≤ 𝛼!𝛽! (𝑅◦ − 𝑅) −𝑀 max |Φ (𝑧1 , ·)| .

𝐾 𝑅◦

They also imply, for each 𝑗 = 2, ..., 𝑛,

1
max 𝜕𝑧𝛼′ 𝜕𝑈 𝜕𝑧 𝑗 Φ (𝑧 1 , 𝑢 (𝑧 ′)) ≤ max 𝜕 𝛼′ 𝜕 Φ (𝑧1 , 𝑢 (𝑧 ′)) .
𝛽 𝛽
|𝑧 ′ | ≤𝑠′ 𝑟 𝑟 (𝑠 − 𝑠 ′) |𝑧′ | ≤𝑠𝑟 𝑧 𝑈

From all of this we conclude easily that, for some 𝐶◦ > 0, all 𝑠, 𝑠 ′ ∈ (0, 1), 𝑠 ′ < 𝑠,
and all 𝑧 ′ ∈ Δ𝑛−1
𝑠𝑟 ,

𝑁 𝑠′ D𝑢𝑀 𝐹 (𝑧1 , 𝑢 (𝑧 ′)) ≤ 𝐶◦𝑀+1 𝑀! (𝑠 − 𝑠 ′) −1 . (5.2.32)

□
By decreasing the radius 𝑟 1 of the disk in which 𝑧1 varies we can render the
constant 𝐶◦ in (5.2.32) independent of 𝑧1 , thus validating Cauchy inequalities of the
type (5.1.16).
Lemma 5.2.3 enables us to apply Theorem 5.1.13 in the scale of Banach spaces
𝑬 𝑠 selected here. The selection of the positive constants 𝑟 and 𝑟 1 is determined by
Φ (𝑧, 𝑈1 , ..., 𝑈𝑛 ) and the “initial” value 𝑈 ◦ . In any event we have proved the classical
Cauchy–Kovalevskaya Theorem for a fully nonlinear analytic PDE, which we state
without any pretense of relating the domains of definition of the solution to that of
the data.
126 5 The Cauchy–Kovalevskaya Theorem

Theorem 5.2.4 Let Ω be an open subset of C𝑛 containing the origin and let
𝑢 ◦ ∈ O (Ω). Let Σ ⊂ Ω be an analytic hypersurface passing through 0 and non-
characteristic for the linearization of (5.2.1) at 𝑢 ◦ [see (5.2.5)]. There is a function
𝑢 ∈ O (Ω′), with Ω′ open and 0 ∈ Ω′ ⊂ Ω, such that (5.2.1) and

|𝑢(𝑧) − 𝑢 ◦ (𝑧)| ≲ dist (𝑧, Σ) 𝑚 (5.2.33)

hold for all 𝑧 ∈ Ω′. Any other function with the same properties is equal to 𝑢 in a
neighborhood of 0.

Remark 5.2.5 Let 𝑁 ≥ 2 be an integer. The analogue of Theorem 5.2.4 holds for an
𝑁 × 𝑁 system of PDEs provided it is reducible (by analytic changes of coordinates
and unknowns) to an equation

𝜕𝑧𝑚1 u = F 𝑧, 𝜕𝑧𝛼 u | 𝛼 | ≤𝑚, 𝛼≠(𝑚,0,...,0) . (5.2.34)

In (5.2.34) both u and F are holomorphic functions (in the appropriate domains)
valued in C 𝑁 .

Remark 5.2.6 The real version (in the C 𝜔 class in R𝑛 ) of Theorem 5.2.4 is an im-
mediate corollary of Theorem 5.2.4, derived by holomorphic extension to a complex
neighborhood of the origin.

5.3 Applications to Linear PDE

5.3.1 Linear version of Theorem 5.2.4

The linear version of Eq. (5.2.1) reads

∑︁
𝑐 𝛼 (𝑧) 𝜕𝑧𝛼 𝑢 = 𝑓 (𝑧) . (5.3.1)
| 𝛼 | ≤𝑚

The coefficients 𝑐 𝛼 and 𝑓 are holomorphic functions in some neighborhood Ω of 0

in
Í C . We assume
𝑛 that the principal symbol of the differential operator 𝑃 (𝑧, 𝜕𝑧 ) =
| 𝛼 | ≤𝑚 𝑐 𝛼 (𝑧) 𝜕𝑧
𝛼 𝑢,
∑︁
𝑃𝑚 (𝑧, 𝜁) = 𝑐 𝛼 (𝑧) 𝜁 𝛼 , (5.3.2)
| 𝛼 |=𝑚

does not vanish identically at 𝑧 = 0. After a linear change of variables we may assume
that 𝑐 (𝑚,0,...,0) (0) ≠ 0. After division by 𝑐 (𝑚,0,...,0) (𝑧) (and a change of notation)
we can assume that, in some neighborhood of the origin, (5.3.1) is equivalent to
∑︁
𝜕𝑧𝑚1 𝑢 + 𝑐 𝛼 (𝑧) 𝜕𝑧𝛼 𝑢 = 𝑓 . (5.3.3)
| 𝛼 |=𝑚, 𝛼1 <𝑚
5.3 Applications to Linear PDE 127

This is a special case of (5.2.9) and we can associate to (5.3.1) the same “initial”
conditions (5.2.10) associated to (5.2.9). The procedure described in Subsection
5.2.2 transforms (5.3.1) into a square system of first-order linear PDEs:
𝑛−1
∑︁
𝜕𝑧1 𝑈 = 𝐴 𝑗 (𝑧)𝜕𝑧 𝑗 𝑈 + 𝐴0 (𝑧)𝑈 + 𝐹 (𝑧) . (5.3.4)
𝑗=1

In (5.3.4) 𝐹 is a holomorphic function of 𝑧 ∈ Ω valued in C 𝑁 and the coefficients

𝐴 𝑗 are also holomorphic functions of 𝑧 ∈ Ω valued in the space of 𝑁 × 𝑁 complex
matrices. Ovsyannikov analyticity is a direct consequence of the fact that if 0 < 𝑠 ′ <
𝑠 < 1 and if 𝑧 1 ∈ Δ𝑟 we have
1
sup 𝜕𝑧 𝑗 𝑈 (𝑧1 , 𝑧 ′) ≤ sup |𝑈 (𝑧 1 , 𝑧 ′)| . (5.3.5)
𝑧 ′ ∈Δ𝑛−1 𝑠 − 𝑠 ′ 𝑧′ ∈Δ𝑛−1
𝑠′ 𝑟 𝑠𝑟

The linear version of Theorem 5.2.4 is evident and needs no restating here. However,
different applications suggest taking a closer look at the linear version of Theorem
5.1.13.

5.3.2 The Ovsyannikov approach to linear PDEs

The direct proof of the linear version of Theorem 5.1.13 (see [Treves, 1975], pp.
148–150) is simpler than that of the general nonlinear case. It shows that there is no
need of holomorphy with respect to the independent variable 𝑧. As a matter of fact
𝑧 can be replaced by a real variable 𝑡 and continuity with respect to 𝑡 suffices (cf.
Remark 5.1.14). Still within the scale of Banach spaces 𝑬 𝑠 (0 < 𝑠 < 1), in place of
the nonlinear map 𝐹 : 𝔅𝑅 (𝑢 ◦ ) −→ 𝑬 0 we consider a linear map 𝐴 (𝑡) : 𝑬 0 −→ 𝑬 0
satisfying the following hypothesis:
(Ovlin) For every pair of numbers 𝑠, 𝑠 ′, 0 < 𝑠 ′ < 𝑠 < 1, 𝐴 (𝑡) is a continuous
function of 𝑡 ∈ (−𝑇, 𝑇) (𝑇 > 0) valued in the space of bounded linear
operators 𝔏 (𝑬 𝑠 , 𝑬 𝑠′ ) with the operator norm bound
𝐶
sup ∥ 𝐴 (𝑡)∥ 𝔏(𝑬 𝑠 ,𝑬 𝑠′ ) ≤ , (5.3.6)
|𝑡 |<𝑇 𝑠 − 𝑠′

where the constant 𝐶 > 0 is independent of 𝑠, 𝑠 ′.

We can apply Theorem 5.1.13 and state (recall that 𝑬 1 is a normed space):
Theorem 5.3.1 Let 𝐴 (𝑡) satisfy (Ovlin); let the continuous map 𝑓 : (−𝑇, 𝑇) → 𝑬 1
and the element 𝑢 ◦ ∈ 𝑬 1 be arbitrary. Then the following is true:
(Existence) For 𝜀 > 0 sufficiently small there is a map (−𝜀, 𝜀) ∋ 𝑡 ↦→ 𝑢 (𝑡) ∈ 𝑬 0
having the following properties:
128 5 The Cauchy–Kovalevskaya Theorem

(1) for each 𝑠 ∈ (0, 1), 𝑢 is a C 1 map (− (1 − 𝑠) 𝜀, (1 − 𝑠) 𝜀) −→ 𝑬 𝑠 ;

(2) in the interval (−𝜀, 𝜀), 𝑢 satisfies the equation

𝜕𝑡 𝑢 = 𝐴𝑢 + 𝑓 ; (5.3.7)

(3) 𝑢 satisfies the initial condition

𝑢 (0) = 𝑢 ◦ . (5.3.8)

(Uniqueness) If for some 𝑠, 0 < 𝑠 < 1, and some 𝜀 ′ > 0, a C 1 map 𝑣 :

(−𝜀 ′, 𝜀 ′) −→ 𝑬 𝑠 satisfies (5.3.7) in the interval (−𝜀 ′, 𝜀 ′) and if 𝑣 (0) = 𝑢 ◦ then
𝑣 = 𝑢 in an interval (−𝛿, 𝛿), 0 < 𝛿 < min ((1 − 𝑠) 𝜀, 𝜀 ′).

Below we need a smooth version of Theorem 5.3.1; this requires that we strengthen
the hypotheses on 𝐴 (𝑡). It is convenient to introduce the following

Definition 5.3.2 LetØ𝐴 (𝑡) be a function of 𝑡 ∈ (0, 1) valued in the space of linear
operators of 𝑬 0 = 𝑬 𝑠 into itself. Let 𝑘 ∈ Z+ or 𝑘 = +∞. We shall say that 𝐴 (𝑡)
0<𝑠<1
is of Ovsyannikov class C 𝑘 in the scale {𝑬 𝑠 }0<𝑠<1 in an interval (𝑎, 𝑏) ⊂ R if, for
every pair (𝑠, 𝑠 ′), 0 < 𝑠 ′ < 𝑠 < 1, the map 𝐴 (𝑡) : (𝑎, 𝑏) −→ 𝔏 (𝑬 𝑠 , 𝑬 𝑠′ ) is of class
C 𝑘 and if to every integer ℓ < 𝑘 + 1 there is a 𝐶ℓ > 0 such that
𝐶ℓ
sup 𝜕𝑡ℓ 𝐴 (𝑡) 𝔏(𝑬 𝑠 ,𝑬 𝑠′ )
≤ . (5.3.9)
𝑎<𝑡 <𝑏 𝑠 − 𝑠′

We shall say that 𝐴 (𝑡) is of Ovsyannikov class C 𝜔 in the scale {𝑬 𝑠 }0<𝑠<1 if 𝐴 (𝑡)
is of Ovsyannikov class C ∞ in said scale and if, moreover, we can take 𝐶ℓ = 𝐶◦ℓ+1 ℓ!
in (5.3.9) for every ℓ ∈ Z+ , with 𝐶◦ > 0 independent of 𝑠, 𝑠 ′, ℓ.

Proposition 5.3.3 Suppose that 𝐴 is of Ovsyannikov class C 𝑘 in the scale {𝑬 𝑠 }0<𝑠<1

with 𝑘 ∈ Z+ ∪ {+∞} (resp., C 𝜔 ) in the interval (−𝑇, 𝑇) and that 𝑓 : (−𝑇, 𝑇) −→ 𝑬 1
is of class C 𝑘 (resp., C 𝜔 ). If 0 < 𝑠 < 1 every C 1 solution 𝑢 : (−𝑇, 𝑇) −→ 𝑬 1 of
(5.3.7)–(5.3.8) is of class C 𝑘+1 (resp., C 𝜔 ) when viewed as a map (−𝑇, 𝑇) −→ 𝑬 𝑠 .

Proof We focus on C 𝑘 regularity for 𝑘 ∈ Z+ . Let 𝑢 ∈ C 1 (−𝑇, 𝑇; 𝑬 1 ) be a

solution of (5.3.7). Let 𝑠 ∈ (0, 1) be arbitrary. Suppose we have proved that
𝑢 ∈ C 𝑘 (−𝑇, 𝑇; 𝑬 𝑠 ). Our hypotheses imply, then, for 0 < 𝑠 ′ < 𝑠 arbitrary, that
𝜕𝑡𝑘 ( 𝐴𝑢 + 𝑓 ) ∈ C (−𝑇, 𝑇; 𝑬 𝑠′ ). By (5.3.7) we have, for |𝑡| < 𝑇,
∫ 𝑡
𝜕𝑡𝑘 𝑢 (𝑡) = 𝜕𝑡𝑘 𝑢 (0) + 𝜕𝑡𝑘 ( 𝐴𝑢 + 𝑓 ) (𝜏) d𝜏. (5.3.10)
0

The right-hand side is a C 1 map (−𝑇, 𝑇) −→ 𝑬 𝑠′ , proving that 𝑢 ∈ C 𝑘+1 (−𝑇, 𝑇; 𝑬 𝑠′ ).

To complete the proof it suffices to note that 𝑠 ′ is arbitrarily close to 1.
Now suppose that 𝐴 is of Ovsyannikov class C 𝜔 in the interval (−𝑇, 𝑇) and that
𝑓 : (−𝑇, 𝑇) −→ 𝑬 1 is of class C 𝜔 . Let 𝑢 ∈ C ∞ (−𝑇, 𝑇; 𝑬 𝑠 ) satisfy (5.3.7). Let
𝑘 ∈ Z+ and 𝑠 ∈ (0, 1) be arbitrary. We derive from (5.3.9):
5.3 Applications to Linear PDE 129

𝑘
1 ∑︁ 𝑘!
𝑵 𝑠 𝜕𝑡𝑘+1 𝑢 ≤ 𝐶◦ℓ+1 𝑵 𝑠+ 𝛿 𝜕𝑡𝑘−ℓ 𝑢 + 𝑁1 𝜕𝑡𝑘 𝑓 (5.3.11)
𝛿 ℓ=0 (𝑘 − ℓ)!

1−𝑠
for some 𝐶◦ > 0 independent of 𝑠 and all 𝛿 ∈ (0, 1 − 𝑠). Actually we take 𝛿 = 𝑘+1 .
Suppose we have proved, for every 𝑗 ∈ Z+ , 𝑗 ≤ 𝑘, and every 𝑠 ∈ (0, 1),
𝑗
𝑗 𝐴
sup 𝑵 𝑠 𝜕𝑡 𝑢 ≤𝐵 𝑗! (5.3.12)
|𝑡 |<𝑇 1−𝑠

where 𝐴 and 𝐵 are suitably large positive constants. As a matter of fact, we start by
requiring 𝐵 ≥ sup 𝑁1 (𝑢), thus ensuring that (5.3.12) holds when 𝑘 = 0. Combining
|𝑡 |<𝑇
(5.3.11) and (5.3.12) yields, for some constant 𝐶1 > 0 independent of 𝑠 and 𝑘,
𝑘 𝑘−ℓ
1
𝑘+1
1
𝑘
∑︁
ℓ 1
𝑵 𝑠 𝜕𝑡 𝑢 ≤ 𝐵𝐶◦ 𝐴 (𝐶◦ /𝐴) + 𝐶1𝑘+1 .
𝑘! 𝛿 ℓ=0
1 − 𝑠 − 𝛿

We require 𝐴 > 𝐶◦ , whence

𝑘−ℓ
1 + 𝑘 −1
𝑘
1 1 ∑︁ 1
𝑵 𝑠 𝜕𝑡𝑘+1 𝑢 ≤ 𝐵𝐶◦ 𝐴 𝑘+1 (𝐶◦ /𝐴) ℓ + 𝐶1𝑘+1
(𝑘 + 1)! 1−𝑠 ℓ=0
1 − 𝑠 𝑘 + 1
𝑘+1
𝐵𝐶◦ 𝐴 1
≤ + 𝐶 𝑘+1
𝐴 − 𝐶◦ 1 − 𝑠 𝑘 +1 1
𝑘+1 ! 𝑘+1
𝐶◦ 𝐶1 𝐴
≤𝐵 + .
𝐴 − 𝐶◦ 𝐴𝐵 1−𝑠

𝐶◦ 𝐶1
We take 𝐴 sufficiently large that 𝐴−𝐶◦ + 𝐴𝐵 ≤ 1, which proves (5.3.12) for 𝑗 = 𝑘 +1.□

5.3.3 Distribution solutions of the abstract linear ODE

We wish to make sense of Eq. (5.3.7) when 𝑢 ∈ D ′ (−𝑇, 𝑇; 𝑬 𝑠 ), i.e., 𝑢 (𝑡) is a

distribution with respect to 𝑡 ∈ (−𝑇, 𝑇) valued in 𝑬 𝑠 . This means that the linear
functional
Cc∞ (−𝑇, 𝑇) ∋ 𝜑 ↦→ ⟨𝑢, 𝜑⟩ ∈ 𝑬 𝑠
is bounded, in the usual sense: whatever 𝜀 ∈ (0, 𝑇), if {𝜑 𝑘 } 𝑘=1,2,... is a sequence in
Cc∞ (−𝑇 + 𝜀, 𝑇 − 𝜀) whose derivatives of all orders converge uniformly to zero then
⟨𝑢, 𝜑 𝑘 ⟩ → 0 in 𝑬 𝑠 . It is a basic fact of distribution theory that, in (−𝑇 + 𝜀, 𝑇 − 𝜀),
we have 𝑢 = 𝜕𝑡𝑝 𝑣 where 𝑣 ∈ C (−𝑇 + 𝜀, 𝑇 − 𝜀; 𝑬 𝑠 ) and 𝑝 ∈ Z+ depend on 𝜀. As a
matter of fact, the reader may take this as an operative definition of an 𝑬 𝑠 -valued
distribution in the interval (−𝑇, 𝑇).
130 5 The Cauchy–Kovalevskaya Theorem

Returning to (5.3.7) we now require that 𝐴 be of Ovsyannikov class C ∞ in the

scale {𝑬 𝑠 }0<𝑠<1 . Under this hypothesis it is permitted to take 𝑢 ∈ D ′ (−𝑇, 𝑇; 𝑬 1 ).
If 𝑢 = 𝜕𝑡𝑝 𝑣 in (−𝑇 + 𝜀, 𝑇 − 𝜀), with 𝑣 ∈ C (−𝑇 + 𝜀, 𝑇 − 𝜀; 𝑬 1 ) and 𝑝 ∈ Z+ , we may
write [cf. (1.1.5)]
𝑝
∑︁ 𝑝!
𝐴𝜕𝑡𝑝 𝑣 = 𝜕𝑞 𝜕𝑡𝑝−𝑞 𝐴 𝑣 .

(−1) 𝑝−𝑞 (5.3.13)
𝑞=0
𝑞! ( 𝑝 − 𝑞)! 𝑡

Since 𝜕𝑡𝑝−𝑞 𝐴 ∈ C ∞ (−𝑇, 𝑇; 𝔏 (𝑬 1 , 𝑬 𝑠 )) we have 𝜕𝑡𝑝−𝑞 𝐴 𝑣 ∈ C (−𝑇 + 𝜀, 𝑇 − 𝜀; 𝑬 𝑠 )

whatever 𝑠 ∈ (0, 1) and 𝑞 = 0, ..., 𝑝: the right-hand side in (5.3.7) is indeed an 𝑬 𝑠 -
valued distribution in (−𝑇 + 𝜀, 𝑇 − 𝜀). In the proof of the next theorem we need the
following
Lemma 5.3.4 Let 𝑝 ∈ Z+ and let 𝑩 be a Banach space. If a distribution 𝒖 ∈
D ′ (−𝑇, 𝑇; 𝑩) is such that 𝜕𝑡𝑝+1 𝒖 = 0 then 𝒖 is a polynomial of degree 𝑝 in 𝑡 with
coefficients in 𝑩.
Proof We first prove the claim when 𝑝 = 0. If 𝒖 is not a constant then (say, by the
Hahn–Banach Theorem) there is a continuous linear functional 𝒃 ∗ on 𝑩 such that the
scalar distribution ⟨𝒃 ∗ , 𝒖⟩ in (−𝑇, 𝑇) is not a constant. If we had 𝜕𝑡 𝒖 = 0 we would
have ⟨𝒖, 𝒃 ∗ 𝜕𝑡 𝜑⟩ = 0 for every 𝜑 ∈ Cc∞ (−𝑇, 𝑇). For 𝜓 ∈ Cc∞ (−𝑇, 𝑇) to be of the
∫𝑇
form 𝜓 = 𝜕𝑡 𝜑, 𝜑 ∈ Cc∞ (−𝑇, 𝑇), it is necessary and sufficient that −𝑇 𝜓 (𝜏) d𝜏 = 0.
This implies that such functions 𝜓 form a hyperplane in Cc∞ (−𝑇, 𝑇); its orthogonal
in D ′ (−𝑇, 𝑇) is one-dimensional and therefore must consist of the constants. We
have reached a contradiction.
Induction on 𝑝 ≥ 1 allows us to deduce from 𝜕𝑡𝑝+1 𝒖 = 0 that 𝜕𝑡 𝒖 = 𝑘=0
Í 𝑝−1
𝒃𝑘𝑡 𝑘,
𝒃 𝑘 ∈ 𝑩. It follows that
𝑝−1
!
∑︁ 1
𝜕𝑡 𝒖 − 𝒃 𝑘 𝑡 𝑘+1 = 0,
𝑘=0
𝑘 + 1
Í 𝑝−1 1
whence 𝒖 − 𝑘=0 𝑘+1 𝒃 𝑘 𝑡
𝑘+1 = 𝒄 ∈ 𝑩, again from the result when 𝑝 = 0. □
Theorem 5.3.5 Suppose that 𝐴 is of Ovsyannikov class C ∞ in the scale {𝑬 𝑠 }0<𝑠<1 .
Let 𝜃 ∈ (0, 1) and 𝑓 ∈ C ∞ (−𝑇, 𝑇; 𝑬 𝜃 ) be arbitrary. Every distribution solution
𝑢 ∈ D ′ (−𝑇, 𝑇; 𝑬 𝜃 ) of (5.3.7) belongs to C ∞ (−𝑇, 𝑇; 𝑬 𝑠 ) for every 𝑠 ∈ (0, 𝜃). The
same is true with C 𝜔 substituted for C ∞ .
Proof Let 𝑢 ∈ D ′ (−𝑇, 𝑇; 𝑬 𝜃 ) be a solution of (5.3.7). We exploit what was just
said above: possibly after decreasing 𝑇 > 0 we can assume that 𝑢 = 𝜕𝑡𝑝 𝑣 in (−𝑇, 𝑇),
with 𝑣 ∈ C (−𝑇, 𝑇; 𝑬 𝜃 ). By putting (5.3.13) in (5.3.7) we can write
𝑝
∑︁ 𝑝!
𝜕𝑡𝑝+1 𝑣 = 𝜕𝑞 𝜕𝑡𝑝−𝑞 𝐴 𝑣 + 𝑓 .

(−1) 𝑝−𝑞 (5.3.14)
𝑞=0
𝑞! ( 𝑝 − 𝑞)! 𝑡

Let us define the following operator C (−𝑇, 𝑇; 𝑬 𝑠 ) −→ C 1 (−𝑇, 𝑇; 𝑬 𝑠 ):

5.3 Applications to Linear PDE 131
∫ 𝑡
𝜕𝑡−1 𝜑 (𝑡) = 𝜑 (𝜏) d𝜏,
0
𝑟
and its powers 𝜕𝑡−𝑟 = 𝜕𝑡−1 : C (−𝑇, 𝑇; 𝑬 𝑠 ) −→ C 𝑟 (−𝑇, 𝑇; 𝑬 𝑠 ), 𝑟 = 2, 3....
Still for 𝜑 ∈ C (−𝑇, 𝑇; 𝑬 𝑠 ) we have 𝜕𝑡𝑞 𝜑 = 𝜕𝑡𝑞+1 𝜕𝑡−1 𝜑 (𝑡); more generally, 𝜕𝑡𝑞 𝜑 =
𝜕𝑡𝑞+𝑟 𝜕𝑡−𝑟 𝜑 (𝑡) for every 𝑟 = 1, 2, .... From (5.3.14) we get 𝜕𝑡𝑝+1 (𝑣 − 𝑤) = 0, where
𝑝
∑︁ 𝑝!
𝜕 𝑞− 𝑝−1 𝜕𝑡𝑝−𝑞 𝐴 𝑣 + 𝜕𝑡− 𝑝−1 𝑓 .

𝑤= (−1) 𝑝−𝑞
𝑞=0
𝑞! ( 𝑝 − 𝑞)! 𝑡

Since 𝐴 is of Ovsyannikov class C ∞ and 𝑞 − 𝑝 ≤ 0 in the sum we obtain that

𝑤 ∈ C 1 −𝑇, 𝑇; 𝑬 𝜂 whatever 𝜂 ∈ (0, 𝜃). Since 𝜕𝑡𝑝+1 (𝑣 − 𝑤) = 0 it follows from

Lemma 5.3.4 that 𝑣 − 𝑤 is a polynomial of degree 𝑝 in 𝑡 with coefficients in
1
𝑝−1
𝑣 ∈ C −𝑇, 𝑇; 𝑬 𝜂 and that 𝑢 = 𝜕𝑡 𝑣 1 in (−𝑇,
𝑬 𝜂 . This proves that 𝑇), with
𝑣 1 ∈ C −𝑇, 𝑇; 𝑬 𝜂 . Induction on 𝑝 leads to the conclusion that 𝑢 ∈ C 1 −𝑇, 𝑇; 𝑬 𝜂 .

(Each step in the induction requires that 𝜂 be decreased, but by an arbitrarily small
amount and 𝜂 was chosen arbitrarily close to 𝜃.)
We can now apply Proposition 5.3.3 to the scale 𝑬 𝑠 𝜂 0<𝑠<1 . This is permitted
because 𝐴 is obviously of Ovsyannikov class C ∞ in the scale 𝑬 𝑠 𝜂 0<𝑠<1 and

( )
Ù
𝑬𝜂 ⊂ 𝑤 ∈ 𝑬 𝑠 𝜂 ; sup 𝑵 𝑠 𝜂 (𝑤) < +∞ . (5.3.15)
0<𝑠<1 0<𝑠<1

The Banach
space (Proposition 5.1.1) at the right in (5.3.15) plays the role for
the scale 𝑬 𝑠 𝜂 0<𝑠<1 that 𝐸 1 plays for the scale {𝑬 𝑠 }0<𝑠<1 . We conclude that
𝑢 ∈ C ∞ −𝑇, 𝑇; 𝑬 𝑠 𝜂 whatever 𝑠 ∈ (0, 1).

□
One can view Theorem 5.3.5 as a hypoellipticity result (cf. Definition 3.1.1).
In the next subsection we shall apply the following consequence of Theorems
5.3.1 and 5.3.5:

Theorem 5.3.6 Let 𝐴 (𝑡) be an operator ∞

′
of Ovsyannikov class C in the scale
{𝑬 𝑠 }0<𝑠<1 and let 𝑢 ∈ D −𝑇, 𝑇; 𝑬 𝑠◦ (0 < 𝑠◦ < 1) be a distribution solution of
the homogeneous equation
𝜕𝑡 𝑢 = 𝐴 (𝑡) 𝑢. (5.3.16)
If supp 𝑢 ⊂ [𝑡◦ , 𝑇) with −𝑇 < 𝑡◦ < 𝑇 then 𝑢 ≡ 0.

Proof From Theorem 5.3.5 we derive that 𝑢 ∈ C ∞ (−𝑇, 𝑇; 𝑬 𝑠 ) for every 𝑠 ∈ (0, 𝑠◦ ).
The uniqueness part of Theorem 5.3.1 (with the Cauchy data at 𝑡◦ ) implies that
supp 𝑢 ⊂ [𝑡◦ + 𝜀, 𝑇) for some 𝜀 > 0. This proves that the set (−𝑇, 𝑇) \ supp 𝑢 is open
and closed and therefore must be equal to (−𝑇, 𝑇). □
132 5 The Cauchy–Kovalevskaya Theorem

5.3.4 The Holmgren Theorem

We shall provisionally make use of the following notation: for 0 < 𝑠 < 1, 𝑩 𝑠 will
denote the Banach space (for the max modulus norm) of continuous functions in the
closed polydisk

(𝑛−1)
= 𝑧 ′ = (𝑧 1 , ..., 𝑧 𝑛−1 ) ∈ C𝑛−1 ; 𝑧 𝑗 ≤ 𝑠𝑅, 𝑗 = 1, ..., 𝑛 − 1

Δ𝑠𝑅
(𝑛−1)
that are holomorphic in the interior Δ𝑠𝑅 (𝑅 > 0 is kept fixed).
We shall avail ourselves of the following version of Runge’s Theorem:

Lemma 5.3.7 The space O C𝑛−1 of entire holomorphic functions in C𝑛−1 is dense
in 𝑩 𝑠 .

Proof An arbitrary function ℎ (𝑧 ′) ∈ 𝑩 𝑠 is the limit in 𝑩 𝑠 as 𝜀 ↘ 0 of the function

ℎ ((1 − 𝜀) 𝑧 ′) ∈ 𝑩 𝑠′ (𝑠 ′ = 1−𝜀
𝑠
). By Taylor expansion we find a sequence of poly-
nomials 𝑃 𝜀,𝑘 (𝑧 ) (𝑘 = 1, 2, ...) that converges to ℎ ((1 − 𝜀) 𝑧 ′) uniformly on Δ𝑠𝑟 . A
′

diagonal process produces a sequence of polynomials 𝑃 𝜀𝜈 ,𝑘𝜈 (𝑧 ′) (𝜈 = 1, 2, ...) that

converges to ℎ uniformly on Δ𝑠𝑟 . □

Corollary 5.3.8 If 0 < 𝑠 ′ < 𝑠 < 1, 𝑩 𝑠 is dense in 𝑩 𝑠′ .

Let 𝑩∗𝑠 denote the Banach space dual of 𝑩 𝑠 . As a consequence of Lemma 5.3.7
the transpose of the restriction map O C𝑛−1 ∋ ℎ ↦→ ℎ| Δ𝑠𝑅 ∈ 𝑩 𝑠 is an injective map
of 𝑩∗𝑠 into the dual of O C𝑛−1 , the space O ′ C𝑛−1 of analytic functionals in C𝑛−1

(see Section 6.1). In the sequel we identify 𝑩∗𝑠 with a linear subspace of O ′ C𝑛−1 .

It is obvious that the analytic functionals in C𝑛−1 that belong to 𝑩∗𝑠 are carried by
(𝑛−1)
Δ𝑠𝑅 (Definition 6.1.1). If 0 < 𝑠 ′ < 𝑠 < 1, 𝑩∗𝑠′ ⊂ 𝑩∗𝑠 .
We “tensor” these various vector spaces with C𝑞 , 1 ≤ 𝑞 ∈ Z+ : we will be
dealing with C𝑞 -valued functions, analytic functionals and distributions; 𝑩 𝑠 ⊗ C𝑞
(𝑛−1)
is the Banach space of continuous maps Δ𝑠𝑅 −→ C𝑞 holomorphic in the interior
(𝑛−1)
Δ𝑠𝑅 . The norm in 𝑩 𝑠 ⊗ C𝑞 is ∥h∥ 𝑠 = max |h (𝑧)|. We will apply Theorem 5.3.1
|𝑧 | ≤𝑠𝑅
putting
𝑬 𝑠 = 𝑩∗1−𝑠 ⊗ C𝑞 = (𝑩1−𝑠 ⊗ C𝑞 ) ∗ (5.3.17)
equipped with the dual norm

𝑬 𝑠 ∋ 𝜇 → 𝑵 𝑠 (𝜇) = sup |⟨𝜇, h⟩| .

∥h ∥ 1−𝑠 ≤1

The elements of 𝑬 𝑠 are C𝑞 -valued analytic functionals carried by Δ (1−𝑠) 𝑅 .

Proposition 5.3.9 The Banach spaces 𝑬 𝑠 (0 < 𝑠 < 1) satisfy Condition (Scale).
5.3 Applications to Linear PDE 133

Proof Suppose 0 < 𝑠 ′ < 𝑠 < 1; then Δ (𝑛−1) (1−𝑠) 𝑅

⊂ Δ (𝑛−1)
(1−𝑠′ ) 𝑅
and restriction to the
smaller polydisk induces a continuous linear injection 𝑩1−𝑠′ ⊗ C𝑞 ↩→ 𝑩1−𝑠 ⊗ C𝑞
whose range is dense, by Corollary 5.3.8. It follows that the transpose of this injection
is a continuous linear injection 𝑩∗1−𝑠 ⊗ C𝑞 ↩→ 𝑩∗1−𝑠′ ⊗ C𝑞 . The norm of each one of
these linear maps is ≤ 1. □
We take 𝐴 to be a first-order linear differential operator
𝑛
∑︁ 𝜕
𝐴 (𝑧 ′, 𝑡, 𝜕𝑧′ ) = 𝐴 𝑗 (𝑧 ′, 𝑡) + 𝐴0 (𝑧 ′, 𝑡) (5.3.18)
𝑗=2
𝜕𝑧 𝑗

with coefficients 𝐴 𝑗 (𝑧 ′, 𝑡) valued in the space of 𝑞 × 𝑞 complex matrices; the entries

of 𝐴 𝑗 (𝑧 ′, 𝑡) are C ∞ with respect to 𝑡 in an interval (−𝑟, 𝑟) and holomorphic with
respect to 𝑧 ′ in some open set containing the closed polydisk Δ𝑅(𝑛−1) . It is directly
checked that (5.3.18) defines an operator 𝐴 (𝑡) of Ovsyannikov class C ∞ in the scale
{𝑬 𝑠 }0<𝑠<1 .
Below we use the notation

Ω = 𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛−1 ) ∈ R𝑛−1 ; 𝑥 𝑗 < 𝑅, 𝑗 = 1, ..., 𝑛 − 1 × (−𝑇, 𝑇) .

We now restrict our attention to a problem associated to (5.3.18) in real space, namely
the equation in Ω,
𝜕u
= 𝐴 (𝑥 ′, 𝑡, 𝜕𝑥′ ) u (5.3.19)
𝜕𝑡
where u ∈ D ′ (Ω; C𝑞 ).
Lemma
′ 5.3.10 If the support of the solution u ∈ D ′ (Ω; C𝑞 ) of (5.3.19) is contained
in 𝑥 ∈ R ; 𝑥 𝑗 < 𝜃𝑅, 𝑗 = 1, ..., 𝑛 − 1 × [0, 𝑇) for some 𝜃 ∈ (0, 1) then u ≡ 0.
𝑛−1

Proof The hypothesis means that, for every 𝜑 ∈ Cc∞ (−𝑇, 𝑇), the duality bracket
with respect to 𝑡, ⟨u, 𝜑⟩, is a C𝑞 -valued distribution in R𝑛−1 , with compact support
contained in the hypercube 𝑥 ′ ∈ R𝑛−1 ; 𝑥 𝑗 < 𝜃𝑅, 𝑗 = 1, ..., 𝑛 − 1 . By Proposi-
tion 6.2.6 ⟨u, 𝜑⟩ can be equated to an unambiguously defined analytic functional
belonging to 𝑬 𝜃 and u to an element of D ′ (−𝑇, 𝑇; 𝑬 𝜃 ). Then the claim ensues from
Theorem 5.3.6. □
Next we revisit (5.3.19) after we relax the hypothesis on supp u.
Theorem 5.3.11 If the support of the solution u ∈ D ′ (Ω; C𝑞 ) of (5.3.19) is con-
tained in [0, 𝑇) × 𝑥 ′ ∈ R𝑛−1 ; 𝑥 𝑗 < 𝑅, 𝑗 = 1, ..., 𝑛 − 1 then u ≡ 0 in a neighbor-
hood of the origin.
Proof We make the change of variable 𝑡 ↦→ 𝑥 𝑛 = 𝑡 + 21 𝜅 |𝑥 ′ | 2 , with 𝜅 > 0 and 𝑥 ′
unchanged. This transforms (5.3.19) into
𝑛
𝜕u 1 ∑︁ 1 𝜕u
= 𝐴 𝑥 𝑛 − 𝜅 |𝑥 ′ | 2 , 𝑥 ′, 𝜕𝑥′ u + 𝜅 𝐴 𝑗 𝑥 𝑛 − 𝜅 |𝑥 ′ | 2 , 𝑥 ′ 𝑥 𝑗 .
𝜕𝑥 𝑛 2 𝑗=2
2 𝜕𝑥 𝑛
134 5 The Cauchy–Kovalevskaya Theorem

For sufficiently small 𝜅 and

𝑥 ∈ (−𝑇, 𝑇) × 𝑥 ′ = (𝑥 2 , ..., 𝑥 𝑛 ) ∈ R𝑛−1 ; 𝑥 𝑗 < 𝑅, 𝑗 = 1, ..., 𝑛 − 1

the matrix
𝑛
∑︁ 1
𝑀 (𝑥) = 𝐼𝑞 − 𝜅 𝐴𝑗 𝑥 𝑛 − 𝜅 |𝑥 ′ | 2 , 𝑥 ′
𝑗=2
2

is nonsingular and we can replace (5.3.19) with the equation

𝑛
𝜕u ∑︁ e 𝜕
= 𝐴 𝑗 (𝑥) e0 (𝑥) ,
+𝐴 (5.3.20)
𝜕𝑥 𝑛 𝑗=2
𝜕𝑥 𝑗

e𝑗 (𝑥) = 𝑀 (𝑥) −1 𝐴 𝑗 𝑥 𝑛 − 1 𝜅 |𝑥 ′ | 2 , 𝑥 ′ , 𝑗 = 0, 1, ..., 𝑛 − 1. In the coordinates
where 𝐴 2
𝑥1 , 𝑥2 , ..., 𝑥 𝑛 the support of u is contained in the set 𝑥 𝑛 ≥ 21 𝜅 |𝑥 ′ | 2 . If 𝑥 𝑛 < 𝑇 ′ =
∞ ′ ′
√︁ 𝜅 2 2
2 𝜃 𝑅 then, for every 𝜑 ∈ Cc (−𝑇 , 𝑇 ), the duality bracket with respect to 𝑥 𝑛 ,
⟨u, 𝜑⟩, is a C -valued distribution in R 𝑥′ , with compact support contained in the
𝑞 𝑛−1

hypercube 𝑥 ′ ∈ R𝑛−1 ; 𝑥 𝑗 < 𝜃𝑅, 𝑗 = 1, ..., 𝑛 − 1 . We may apply Lemma 5.3.7:

u ≡ 0 in the open set 𝑥 ∈ R𝑛 ; 𝑥 𝑗 < 𝑅, 𝑗 = 1, ..., 𝑛 − 1, 𝑥 𝑛 < 𝑇 ′ .

□
Now we look at a linear PDE
∑︁
𝑃 (𝑥, 𝜕𝑥 ) 𝑢 = 𝑐 𝛼 (𝑥) 𝜕𝑥𝛼 𝑢 = 0 (5.3.21)
| 𝛼 | ≤𝑚

in a domain 𝔄 in R𝑛 , under the hypothesis that 𝑐 𝛼 ∈ C 𝜔 (𝔄) and that the principal
symbol 𝑃𝑚 (𝑥, 𝜉) of 𝑃 (see Definition 2.1.1) does not vanish for all 𝜉 ∈ R𝑛 when
𝑥 = 𝑥 ◦ ∈ 𝔄. We can now state and prove the Holmgren Theorem ([Holmgren, 1901]):
Theorem 5.3.12 Let 𝑆 ⊂ 𝔄 be a C ∞ hypersurface containing 𝑥 ◦ and noncharacter-
istic at 𝑥 ◦ for the analytic differential operator 𝑃 (𝑥, 𝜕𝑥 ) (Definition 1.3.3). Suppose
that, for some 𝑟 > 0, 𝑆 subdivides the open ball

B𝑟 (𝑥 ◦ ) = {𝑥 ∈ R𝑛 ; |𝑥 − 𝑥 ◦ | < 𝑟} ⊂ 𝔄

into two parts, i.e., B𝑟 (𝑥 ◦ ) \ (B𝑟 (𝑥 ◦ ) ∩ 𝑆) has two connected components. If a

solution 𝑢 ∈ D ′ (𝔄) of (5.3.21) vanishes identically in one of these two components
then necessarily 𝑢 ≡ 0 in a ball B 𝜀 (𝑥 ◦ ), 0 < 𝜀 ≤ 𝑟.
Proof Suppose that B𝑟 (𝑥 ◦ ) ∩ supp 𝑢 lies on one side of 𝑆. It is easy to find a C 𝜔
hypersurface Σ in B𝑟 (𝑥 ◦ ) tangent to 𝑆 at 𝑥 ◦ and such that (Σ\ {𝑥 ◦ }) ∩ supp 𝑢 = ∅.
After decreasing 𝑟 if need be, we can make a C 𝜔 change of variables transforming
Σ ∩ B𝑟 (𝑥 ◦ ) into a neighborhood of 𝑥 ◦ in the hyperplane 𝑥 𝑛 = 𝑥 𝑛◦ and, possibly after
division by a nowhere vanishing C 𝜔 function, (5.3.21) into
∑︁
𝜕𝑥𝑚𝑛 𝑢 + 𝑐 𝛼 (𝑥) 𝜕𝑥𝛼 𝑢 = 0. (5.3.22)
| 𝛼 | ≤𝑚, 𝛼𝑛 <𝑚
5.3 Applications to Linear PDE 135

We refer the reader to Subsection 5.2.2 on how to reduce a higher order PDE such
as (5.3.22) to a first-order system (5.3.19). Such a reduction enables us to derive that
𝑢 ≡ 0 in B𝑟 (𝑥 ◦ ) from Theorem 5.3.11. □
Remark 5.3.13 In what precedes the equation (5.3.21) is assumed to be scalar. But
it is evident that Theorem 5.3.12 is also valid for the class of systems of linear PDEs
that are equivalent, mod analytic changes of variables and linear substitutions of the
unknowns, to a system (5.3.19) [cf. Remark 5.2.5].

5.3.5 An application: global solvability of elliptic linear PDEs with

analytic coefficients

Let 𝑃 (𝑥, D) be a linear PDO of order 𝑚 ≥ 1 with analytic coefficients in an open

subset Ω of R𝑛 . Recall that E ′ ( Ω) is the space of compactly supported distributions
in Ω. We derive directly from Theorem 4.1.2:
Lemma 5.3.14 If 𝑃 (𝑥, D) is elliptic and 𝜇 ∈ E ′ (Ω), 𝑃 (𝑥, D) 𝜇 = 0 implies 𝜇 = 0.
We exploit the Holmgren Theorem 5.3.12 to prove a particular case of a result in
[Malgrange, 1955], Ch. III.
Theorem 5.3.15 If 𝑃 (𝑥, D) is elliptic then 𝑃 (𝑥, D) C ∞ (Ω) = C ∞ (Ω).
Proof As we do several times in this book we apply a classical characterization of
linear surjections of Fréchet–Montel spaces (see Lemma 7.1.17): Theorem 5.3.15 is
a consequence of the fact that the transpose 𝑃 (𝑥, D) ⊤ of 𝑃 (𝑥, D), as a linear map
of E ′ ( Ω) into itself, is injective and its range is sequentially closed. The injectivity
follows at once from Lemma 5.3.14 since 𝑃 (𝑥, D) ⊤ is elliptic (and its coefficients
are analytic). Let 𝜇 𝑗 ( 𝑗 = 1, 2, ...) be a sequence in E ′ ( Ω) such that 𝑃 (𝑥, D) ⊤ 𝜇 𝑗
converges to 𝜆 in E ′ ( Ω). The nature of the seminorms defining the topology of
C ∞ (Ω) requires that there be a compact subset 𝐾 of Ω and 𝑁 ∈ Z such that
supp 𝑃 (𝑥, D) ⊤ 𝜇 𝑗 ⊂ 𝐾 for all 𝑗 and 𝑃 (𝑥, D) ⊤ 𝜇 𝑗 𝑗=1,2,... is a bounded sequence in

the Sobolev space 𝐻 𝑁 (𝐾) (cf. Section 2.2). Let Ω′ be an arbitrary open subset of R𝑛
such that 𝐾 ⊂ Ω′ ⊂⊂ Ω and whose boundary 𝜕Ω′ is a C ∞ hypersurface. Since 𝜕Ω′
is noncharacteristic for 𝑃 (𝑥, D) ⊤ at every one of its points it follows from Theorem
5.3.12 that if 𝑥 ◦ ∈ 𝜕Ω′ but 𝑥 ◦ ∉ 𝐾 then 𝑥 ◦ ∉ supp 𝜇 𝑗 . By contracting Ω′ about 𝐾 we
supp 𝜇 𝑗 ⊂ 𝐾. It follows readily from the Gårding inequality (Theorem
conclude that
4.1.4) that 𝜇 𝑗 𝑗=1,2,... is a bounded sequence in 𝐻 𝑁 +𝑚 (𝐾). The latter implies (by
Proposition 2.2.3) that a subsequence converges in 𝐻 𝑁 (𝐾), hence in E ′ ( Ω), to 𝜇;
necessarily 𝑃 (𝑥, D) ⊤ 𝜇 = 𝜆, which proves the claim. □
Corollary 5.3.16 If 𝑃 (𝑥, D) is elliptic then 𝑃 (𝑥, D) C 𝜔 (Ω) = C 𝜔 (Ω).
Remark 5.3.17 Theorem 5.3.15 is stated for scalar linear PDEs with analytic coef-
ficients; it is also valid for square elliptic systems of linear PDEs. It is not valid for
very general elliptic systems, such as those introduced in Ch. 9, (9.4.2), (9.4.4).
136 5 The Cauchy–Kovalevskaya Theorem

Exercise 5.3.18 Prove that if 𝑃 (𝑥, D) is a linear PDO with analytic coefficients in
an open subset Ω of R𝑛 and 𝑠 ∈ R is arbitrary then 𝑃 (𝑥, D) 𝐻loc
𝑠 (Ω) 𝑠−𝑚 (Ω).
= 𝐻loc
Does this prove that 𝑃 (𝑥, D) D ′ (Ω) = D ′ (Ω)?

5.4 Application to Integrodifferential Cauchy Problems

5.4.1 Cauchy Problem for the Camassa–Holm Equation

The Camassa–Holm equation is the following nonlinear PDE in two real variables
𝑡, 𝑥:
(1 − 𝜕𝑥2 )𝜕𝑡 𝑢 = 𝑢𝜕𝑥3 𝑢 + 2 (𝜕𝑥 𝑢) 𝜕𝑥2 𝑢 − 3𝑢𝜕𝑥 𝑢. (5.4.1)
We rewrite it as an integrodifferential equation:

𝜕𝑡 𝑢 = 𝜕𝑥 (1 − 𝜕𝑥2 ) −1 𝑄 [𝑢] (5.4.2)

where
1 3
𝑄 [𝑢] = 𝑢𝜕𝑥2 𝑢 + (𝜕𝑥 𝑢) 2 − 𝑢 2 (5.4.3)
2 2
and for 𝑓 ∈ 𝐿 1 (R),
∫
1 d𝜉
(1 − 𝜕𝑥2 ) −1 𝑓 (𝑥) = e𝑖 𝑥 𝜉 b
𝑓 (𝜉) . (5.4.4)
2𝜋 R 1 + 𝜉2
The Ovsyannikov approach to a Cauchy problem for (5.4.1) requires that we select
a scale of Banach algebras 𝑬 𝑠 (0 < 𝑠 < 1) on which (1 − 𝜕𝑥2 ) −1 acts in a convenient
manner. There are many possible choices; we select one of the simplest.
Definition 5.4.1 For 0 < 𝑠 < 1 and 𝑚 ∈ R we denote by 𝑬 𝑚 𝑠 be the linear space of
tempered distributions 𝑢 in R such that
∫
1
∥𝑢∥ 𝑠(𝑚) = 𝑢 (𝜉)| (1 + |𝜉 |) 𝑚 e𝑠𝜅 | 𝜉 | d𝜉 < +∞,
|b (5.4.5)
2𝜋 R
where 𝜅 > 0.
If 0 < 𝑠 ′ < 𝑠 ≤ 1 the natural injection 𝑬 𝑚
𝑠 ↩→ 𝑬 𝑠′ has norm 1; note also that
𝑚
′
𝑬𝑚 ′
𝑠 ⊂ 𝑬 𝑠′ whatever 𝑚, 𝑚 ∈ R.
𝑚

Proposition 5.4.2 For 0 < 𝑠 ≤ 1 and 𝑚 ∈ R the elements of 𝑬 𝑚

𝑠 are restrictions to R
of holomorphic functions in the open slab {𝑧 ∈ C; |Im 𝑧| < 𝜅𝑠} that tend uniformly
to zero as |Re 𝑧| → +∞, |Im 𝑧| ≤ 𝜅 ′ 𝑠 (0 ≤ 𝜅 ′ < 𝜅).

𝑠 be arbitrary. If |𝑦| < 𝜅𝑠 the Fourier inversion formula

Proof Let 𝑢 ∈ 𝑬 𝑚
∫
1
𝑢 (𝑥 + 𝑖𝑦) = e𝑖 𝜉 𝑥−𝑦 𝜉 b
𝑢 (𝜉) d𝜉
2𝜋
5.4 Application to Integrodifferential Cauchy Problems 137

implies
|𝑢 (𝑥 + 𝑖𝑦)| ≤ sup (1 + 𝑡) −𝑚 e ( |𝑦 |−𝜅 𝑠)𝑡 ∥𝑢∥ 𝑠(𝑚) .
𝑡 >0

The claim that |𝑢 (𝑥 + 𝑖𝑦)| → 0 as |𝑥| → +∞ is a direct consequence of the fact that
the Schwartz space S (R) is dense in 𝐿 1 (R) and therefore also in 𝑬 𝑚
𝑠 . □
Proposition 5.4.3 Equipped with the norm (5.4.5) 𝑬 𝑚𝑠 is a Banach space. For 𝑚 ≥ 0
it is a Banach algebra with respect to ordinary multiplication.
Proof The first claim is self-evident. The second claim is equivalent to the assertion
that the space of Fourier transforms of elements of 𝑬 𝑚 𝑠 is a Banach algebra with
respect to convolution. This is a direct consequence of the fact that 𝐿 1 (R) is such
an algebra. Indeed,
∫
(𝑚)
2
(2𝜋) ∥𝑢𝑣∥ 𝑠 = |b 𝑣 (𝜂)| (1 + |𝜉 |) 𝑚 e𝑠𝜅 | 𝜉 | d𝜉d𝜂
𝑢 (𝜉 − 𝜂) b
R2
∫
≤ |b 𝑣 (𝜂)| (1 + |𝜉 − 𝜂|) 𝑚 (1 + |𝜂|) 𝑚 e𝑠𝜅 | 𝜉 −𝜂 |+𝑠𝜅 | 𝜂 | d𝜉d𝜂
𝑢 (𝜉 − 𝜂) b
R2
≤ (2𝜋) 2 ∥𝑢∥ 𝑠(𝑚) ∥𝑣∥ 𝑠(𝑚) . □

Proposition 5.4.4 The partial derivative 𝜕𝑥 defines a bounded linear operator

𝑬 𝑚+1 −→ 𝑬 𝑚 2 −1 defines a bounded linear
𝑠 𝑠 . The convolution operator (1 − 𝜕𝑥 )
operator 𝑬 𝑚
𝑠 −→ 𝑬 𝑚+2
𝑠 .
Self-evident.
Proposition 5.4.5 If 0 < 𝑠 ′ < 𝑠 ≤ 1 the derivation 𝜕𝑥 defines an Ovsyannikov
bounded (linear) operator 𝑬 𝑚
𝑠 −→ 𝑬 𝑠′ .
𝑚

Proof We have
∫
1 ′
∥𝜕𝑥 𝑢∥ 𝑠(𝑚)
′ = 𝑢 (𝜉)| (1 + |𝜉 |) 𝑚 e𝑠 𝜅 | 𝜉 | d𝜉
|𝜉b
2𝜋 R
∫
1
−(𝑠−𝑠′ ) 𝜅𝑡
≤ sup 𝑡e 𝑢 (𝜉)| (1 + |𝜉 |) 𝑚 e𝑠𝜅 | 𝜉 | d𝜉
|b
2𝜋 𝑡 >0 R
e𝜅 −1
= ∥𝑢∥ 𝑠(𝑚) . □
𝑠 − 𝑠′
The map 𝑢 ↦→ 𝑄 [𝑢] defines a quadratic polynomial map 𝑬 𝑚+2 𝑠 −→ 𝑬 𝑚
𝑠 and
2 −1
therefore 𝑢 ↦→ (1 − 𝜕𝑥 ) 𝑄 [𝑢] defines a quadratic operator 𝑬 𝑠 −→ 𝑬 𝑚+2
𝑚+2
. The
𝑠
Fréchet derivatives of 𝑄 [𝑢] at an arbitrary 𝑢 ◦ ∈ 𝑬 𝑚+2
𝑠 are

𝐷𝑄 [𝑢 ◦ ] 𝑣 = 𝑢 ◦ 𝜕𝑥2 𝑣 + 𝑣𝜕𝑥2 𝑢 ◦ + 2 (𝜕𝑥 𝑢 ◦ ) 𝜕𝑥 𝑣 − 3𝑢 ◦ 𝑣,

D2 𝑄 [𝑢 ◦ ] (𝑣, 𝑤) = 𝜕𝑥2 (𝑣𝑤) − 3𝑣𝑤,
D 𝑘 𝑄 [𝑢 ◦ ] = 0 if 𝑘 ≥ 3.
138 5 The Cauchy–Kovalevskaya Theorem

Combining this with Proposition 5.4.5 yields

Proposition 5.4.6 Let 𝑚 ≥ 0 be an integer. The function 𝑢 ◦ ↦→ 𝜕𝑥 (1 − 𝜕𝑥2 ) −1 𝑄 [𝑢 ◦ ]

is Ovsyannikov analytic in the whole space 𝑬 1𝑚+2 .

Recall that Δ 𝜀 is the open disk of radius 𝜀 and center 0 in C. Theorem 5.1.13
implies

Theorem 5.4.7 Suppose 2 ≤ 𝑚 ∈ R and let 𝑢 ◦ ∈ 𝑬 1(𝑚) be arbitrary. For 𝜀 > 0

sufficiently small there is a map
Ø
Δ 𝜀 ∋ 𝑡 ↦→ 𝑢 (𝑡) ∈ 𝑬 𝑠(𝑚)
0<𝑠 ≤1

with the following properties: 𝑢 (0) = 𝑢 ◦ and

(1) for each 𝑠 ∈ (0, 1), 𝑢 is a holomorphic map Δ (1−𝑠) 𝜀 −→ 𝑬 𝑠(𝑚) ;

(2) in the open disk Δ 𝜀 ,
𝜕𝑡 𝑢 = 𝜕𝑥 (1 − 𝜕𝑥2 ) −1 𝑄 [𝑢] . (5.4.6)

If for some 𝑠 ′, 0 < 𝑠 ′ ≤ 1, 𝜀 ′ > 0, a holomorphic map 𝑣 : Δ 𝜀′ −→ 𝑬 𝑠(𝑚) satisfies

(5.4.6) in Δ 𝜀′ and if 𝑣 (0) = 𝑢 ◦ then there is a 𝛿 > 0, 𝛿 ≤ min (𝜀, 𝜀 ′), such that
𝑣 (𝑡) = 𝑢 (𝑡) for all 𝑡 ∈ Δ 𝛿 .
Part II
Hyperfunctions in Euclidean Space
Chapter 6
Analytic Functionals in Euclidean Space

The aim, in Part II of this book, is to describe, in as simple terms as the author was
able to, the theory of hyperfunctions in R𝑛 . Hyperfunctions, introduced by M. Sato
in the late 1950s, are the “generalized functions” naturally suited to the study of
analytic PDEs. Distributions, the natural generalized functions in the C ∞ class, are
special hyperfunctions. Our approach is that proposed by A. Martineau in the 1960s,
through analytic functionals, whose introductory study is the content of the present
chapter. Analytic functionals in an open subset Ω of C𝑛 are the continuous linear
functionals on the Fréchet space O (Ω) of holomorphic functions in Ω, just as the
compactly supported distributions in an open subset 𝔄 of R𝑛 are the continuous
linear functionals on the Fréchet space of C ∞ functions in 𝔄. Section 6.1 is devoted
to the definitions and basic properties of analytic functionals with particular attention
to the Runge open (alternatively, compact) subsets of Ω.
An open subset 𝑈 of Ω is Runge (or has the Runge property) in Ω if the restriction
to 𝑈 maps O (Ω) onto a dense subspace of the Fréchet space O (𝑈), in which case
its dual O ′ (𝑈), the set of analytic functionals in 𝑈, can be identified with a linear
subspace of O ′ (Ω). When Ω = C𝑛 this is the same as saying that the polynomials are
dense in O (𝑈), in which case we simply say that 𝑈 is Runge; the Taylor expansion
of entire functions shows that every open polydisk in C𝑛 is Runge.
A compact set 𝐾 ⊂ Ω is Runge in Ω if every holomorphic function in a neighbor-
hood of 𝐾 in Ω is the limit, in a possibly smaller neighborhood of 𝐾, of functions
belonging to O (Ω). Note that if 𝐾 ⊂ Ω is Runge in C𝑛 it is automatically Runge in
Ω (in this case we simply say that 𝐾 is Runge); in particular, every finite set of points
in Ω is Runge. If a compact set 𝐾 ⊂ Ω is Runge in Ω we can define O ′ (𝐾) as a
linear subspace of O ′ (Ω), whose elements are said to be carried by 𝐾; furthermore,
O ′ (𝐾) is dense in O ′ (Ω) (Proposition 6.1.7 below). In particular, if 𝑧◦ ∈ Ω then
O ′ ({𝑧 ◦ }) is dense in O ′ (Ω). This stands in striking contrast with the properties of
compactly supported distributions in a real domain 𝔄: the distributions with support
concentrated at a single point 𝑥 ◦ are the linear combinations of finitely many deriva-
tives of the Dirac distribution 𝛿 (𝑥 ◦ ); they form a “tiny” closed subspace of E ′ (𝔄).
This is at the root of the crucial difference between compactly supported distribu-

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 141
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_6
142 6 Analytic Functionals in Euclidean Space

tions in a real domain and analytic functionals in a complex domain Ω: the latter are
not localizable; they do not have a support; they have compact carriers, as indicated
above (see Definition 6.1.3 below), always infinitely many different carriers.
One way to see this is to regard O (Ω) as a closed linear subspace of C (Ω), the
space of complex-valued continuous functions in Ω, equipped with the topology of
uniform convergence on the compact subsets of Ω. By the Hahn–Banach Theorem,
an arbitrary 𝜇 ∈ O ′ (Ω) extends as a continuous linear functional on C (Ω), which is
to say, a Radon measure 𝜇♭ . The support of 𝜇♭ is well-defined but 𝜇♭ is not unique:
for instance, 𝜇♭ + 𝜕𝜕𝑧𝑓¯ is also an extension of 𝜇 whatever 𝑓 ∈ Cc∞ (Ω) and can have a
very different support. Actually, carriers of 𝜇 may not intersect (see Example 6.1.4
below).
Let us agree to say that a class of generalized functions, defined in a Hausdorff
topological space X, are localizable if we are allowed to state whether two of them are
equal or not in a neighborhood of an arbitrary point of X (in a neighborhood of, not
at, the point, because of a possibly lurking uncertainty principle or, to put it simply,
because any measuring apparatus has an extension). The prerequisite that a class of
generalized functions be localizable has generally been regarded as essential to their
use in theoretical physics, at least by theoretical physicists1 who insist on localness,
on the axiom that “things” are somewhere at a given time and do not communicate
instantaneously. Analytic functionals in C𝑛 are not localizable, as indicated above,
but analytic functionals carried by compact subsets of R𝑛 are (in R𝑛 ). The basis
for this all important property (which implies that hyperfunctions are localizable,
see Ch. 7) lies in the fact that every compact subset of R𝑛 is Runge in C𝑛 (and
polynomially convex, see Theorems 6.2.13, 6.2.14, in Subsection 6.2.3). Denote by
O ′ (R𝑛 ) the set of analytic functionals carried by some compact subset of R𝑛 . The
intersection of all the carriers of 𝜇 ∈ O ′ (R𝑛 ) contained in R𝑛 is a carrier of 𝜇
which, naturally, we call the support of 𝜇.
Earlier, in Subsection 6.2.1, we introduce the Laplace–Borel transform of ana-
lytic functionals in C𝑛 , a generalization in complex space of the Fourier transform
of compactly supported distributions in real space. The Laplace–Borel transform
establishes a ring isomorphism between O ′ (C𝑛 ), regarded as a ring with respect
to convolution, and the set of entire functions of exponential type in C𝑛 , a ring
with respect to ordinary multiplication. This is the Paley–Wiener theorem extended
to analytic functionals. The convolution of analytic functionals in C𝑛 is defined in
Subsection 6.2.2; it is the inverse Laplace–Borel transform of the (multiplicative)
product of their Laplace–Borel transforms.
In Section 6.3 we depart from the rule generally adhered to in this book: we
apply results to be found later, in order to interpret the analytic functionals in
O ′ (R𝑛 ) as cohomology classes and derive from this interpretation that they can be
decomposed into smaller pieces and “reconstituted” from such pieces. Introducing
the cohomology groups (vector spaces) is not really a problem if we use differential
forms of bidegree (𝑛, 𝑞) with smooth coefficients and the exterior derivative (which
coincide with 𝜕¯ on such forms). The problem is to explain when these groups

1 such as Albert Einstein, see [Einstein-Podolsky-Rosen, 1935].

6.1 Analytic Functionals in Complex Domains 143

vanish. To do this we exploit the Bochner–Martinelli formula (11.1.24) (Section

11.2) pertaining to the theory of the 𝜕¯ differential complex introduced in Section
9.4.

6.1 Analytic Functionals in Complex Domains

In this section Ω shall denote an arbitrary open subset of C𝑛 ; O (Ω) shall stand for
the space of holomorphic functions in Ω, equipped
with the topology of uniform
convergence on compact subsets of Ω. Let 𝐾 𝑗 𝑗=1,2,... be an exhaustive sequence
Ø∞
of compact subsets of Ω, meaning that 𝐾 𝑗 ⊂ 𝐾 𝑗+1 and Ω = 𝐾 𝑗 . A basis of
𝑗=1
neighborhoods of the origin in O (Ω) consists of the (open) “semiballs” (“balls” for
seminorms)
ℎ ∈ O (Ω) ; max |ℎ| < 𝑅 .
𝐾𝑗

If 𝐾 is a compact subset of Ω, to say that the seminorm ℎ ↦→ max |ℎ| is a norm

𝐾
is the same as saying that every function ℎ ∈ O (Ω) that vanishes identically on 𝐾
vanishes identically in Ω. This is the case when Ω is connected and the interior of
𝐾 ⊂ Ω is not empty, also in much more general circumstances (think of an infinite
sequence of distinct points converging in Ω).
We denote by O ′ (Ω) the dual of O (Ω), which is to say, the space of continuous
linear functionals 𝜇 : O (Ω) −→ C. The evaluation of 𝜇 at a function ℎ ∈ O (Ω) will
be denoted, interchangeably,
by 𝜇 (ℎ) or by ⟨𝜇, ℎ⟩. That 𝜇 is continuous means that
lim 𝜇 (ℎ 𝑛 ) = 𝜇 lim ℎ 𝑛 whatever the convergent sequence {ℎ 𝑛 } 𝑛=1,2,... in O (Ω).
𝑛→∞ 𝑛→∞
Thus, to every 𝜇 ∈ O ′ (Ω) there is a compact set 𝐾 ⊂ Ω and a constant 𝐶 > 0 such
that
∀ℎ ∈ O (Ω) , |𝜇 (ℎ)| ≤ 𝐶max |ℎ| . (6.1.1)
𝐾

Definition 6.1.1 The elements of O′ (Ω) are called the analytic functionals in Ω.

Multiplication of an analytic functional 𝜇 in Ω by a function ℎ ∈ O (Ω) is the

transpose of the ordinary multiplication 𝑔 ↦→ ℎ𝑔 in O (Ω): ⟨ℎ𝜇, 𝑔⟩ = ⟨𝜇, ℎ𝑔⟩. Thus
O ′ (Ω) can be viewed as an O (Ω)-module.
We shall make use of the weak topology in O ′ (Ω): a sequence {𝜇 𝑛 } 𝑛=1,2,...
converges in O ′ (Ω) if the numbers 𝜇 𝑛 (ℎ) converge in C whatever ℎ ∈ O (Ω).

Remark 6.1.2 The topology of O (Ω) is the same as that inherited from C (Ω), the
space of continuous functions in Ω, whose dual is the space of compactly supported
Radon measures. It follows from the Hahn–Banach Theorem that to each analytic
functional
∫ 𝜇 there is a compactly supported Radon measure d𝑚 in Ω such that
⟨𝜇, ℎ⟩ = ℎd𝑚 for every ℎ ∈ O (Ω).
144 6 Analytic Functionals in Euclidean Space

Definition 6.1.3 The analytic functional 𝜇 is said to be carried by the compact subset
𝐾 of Ω and 𝐾 is called a carrier of 𝜇 if to every open subset 𝑈 of Ω containing 𝐾
there is a constant 𝐶𝑈 > 0 such that

∀ℎ ∈ O (Ω) , |𝜇 (ℎ)| ≤ 𝐶𝑈 sup |ℎ| . (6.1.2)

𝑈

Generally, when the open set 𝑈 contracts about 𝐾 one should expect the constant
𝐶𝑈 to increase to +∞.
The notion of carrier is not to be confused with that of support (of a function, of
a distribution). In general, an analytic functional has many different, even disjoint,
carriers. That is why one says that analytic functionals are not localizable.
Example 6.1.4 Let 𝛿 (𝑧 ◦ ) denote the Dirac distribution at 𝑧 ◦ ∈ C𝑛 ; it defines the
analytic functional ℎ ↦→ ℎ (𝑧 ◦ ). If 𝑈 and Ω are bounded domains in C𝑛 such that
𝑧◦ ∈ 𝑈 ⊂⊂ Ω then |ℎ (𝑧 ◦ )| ≤ max |ℎ| whatever ℎ ∈ O (Ω), which proves that the
𝜕𝑈
boundary of 𝑈 is a carrier of 𝛿 (𝑧 ◦ ) regarded as an analytic functional in Ω.
We shall denote by O (C𝑛 ) the sheaf of germs of holomorphic functions at points
of C𝑛 ; the stalk of this sheaf at a point 𝑧◦ ∈ C𝑛 , O𝑧 ◦ , is a ring, isomorphic (via Taylor
expansion) to the ring of convergent power series in 𝑧1 − 𝑧◦1 , ..., 𝑧 𝑛 − 𝑧◦𝑛 .
Let 𝐾 be a compact subset of C𝑛 (𝐾 ≠ ∅ always). Definition 1.2.14 allows us to
introduce the set of germs of holomorphic functions at 𝐾. We shall denote it by
O (𝐾); it has a natural ring structure. If 𝑈 is an open subset of C𝑛 , 𝑈 ⊃ 𝐾, we denote
by𝜌𝑈 𝐾 : O (𝑈) −→ O (𝐾) the map that assigns to a function 𝑓 ∈ O (𝑈) its germ
at 𝐾. If 𝑓1 , 𝑓2 ∈ O (𝑈) then, by definition, 𝜌𝑈 𝐾 ( 𝑓1 ) + 𝜌 𝐾 ( 𝑓2 ) = 𝜌 𝐾 ( 𝑓1 + 𝑓2 ) and
𝑈 𝑈

𝜌 𝐾 𝑓1 𝜌 𝐾 𝑓2 = 𝜌 𝐾 𝑓1 𝑓2 ). There is a natural topological vector space (actually,

𝑈 ( ) 𝑈 ( ) 𝑈 (

topological ring) structure on O (𝐾) of the type called a nonstrict LF structure (cf.
[Treves, 1967]).
Define 𝑈𝜈 = 𝑧 ∈ C𝑛 ; dist (𝑧, 𝐾) < 𝜈1 (𝜈 = 1, 2, ...); O (𝐾) can be defined as
the union of the subspaces O (𝑈𝜈 ) (𝜈 = 1, 2, ...) after we equate any two functions
ℎ 𝑗 ∈ O (𝑈𝜈 𝑗 ) ( 𝑗 = 1, 2) which coincide in a neighborhood 𝑈 ⊂ 𝑈𝜈1 ∩ 𝑈𝜈2 of 𝐾.
Formally, the topology of O (𝐾) is defined by specifying a basis of open sets: a
convex subset ℭ of O (𝐾) is open if and only if ℭ ∩ 𝜌𝑈 𝐾 O (𝑈 𝜈 ) is open for every 𝜈.
𝜈

In practice the following two properties are useful:

(1) A sequence {ℎ 𝜈 } 𝜈=1,2,... converges in O (𝐾) if and only if all the ℎ 𝜈 extend
holomorphically to one and the same open subset 𝑈 of C𝑛 , 𝐾 ⊂ 𝑈, and these
extensions converge in O (𝑈).
(2) Let 𝑬 be a Fréchet space (possibly finite-dimensional or even 𝑬 = C). For a
linear map 𝜙 : O (𝐾) −→ 𝑬 to be continuous it is necessary and sufficient that
the composite 𝜙 ◦ 𝜌𝑈𝐾 : O (𝑈) −→ 𝑬 be continuous, whatever the open subset
𝑈 ⊃ 𝐾 of C𝑛 .
Property (2) facilitates the handling of the continuous linear functionals on O (𝐾);
they form the dual of O (𝐾), which we shall denote by O ′ (𝐾). There is a natural
Fréchet space structure, hence a complete metrizable topology, on O ′ (𝐾). The
topology of O ′ (𝐾) can be defined by the norms
6.1 Analytic Functionals in Complex Domains 145

𝒑 𝜈 (𝜇) = sup |⟨𝜇, ℎ⟩| ; ℎ ∈ O (𝑈𝜈 ) , sup |ℎ| ≤ 1 . (6.1.3)
𝑈𝜈

In general, if Ω is an open subset of C𝑛 and 𝐾 a compact subset of Ω, it is not

permitted to identify O ′ (𝐾) with a linear subspace of O ′ (Ω). Although there is
a natural “restriction” map O (Ω) −→ O (𝐾) its transpose, to which we refer as
the natural map O ′ (𝐾) → O ′ (Ω), might not be injective. For later reference we
formalize the definitions already given in the Introduction.

Definition 6.1.5 An open subset 𝑈 ≠ ∅ of Ω is said to have the Runge property in

Ω or to be a Runge open subset of Ω if the restriction map O (Ω) −→ O (𝑈) has
a dense image. A compact subset 𝐾 ≠ ∅ of Ω is said to have the Runge property in
Ω or to be a Runge compact subset of Ω if to every open subset 𝑈 of Ω such that
𝐾 ⊂ 𝑈 there is another open set 𝑉 such that 𝐾 ⊂ 𝑉 ⊂ 𝑈 and such that, given an
arbitrary ℎ ∈ O (𝑈), the restriction of ℎ to 𝑉 is the limit in O (𝑉) of a sequence of
functions belonging to O (Ω).

If 𝐾 (resp., 𝑈) is a Runge compact (resp., open) subset of C𝑛 , most of the time

we will simply say that 𝐾 (resp., 𝑈) is Runge.

Proposition 6.1.6 Let Ω be an open subset of C𝑛 . If 𝐾 is a Runge compact subset

of Ω then the transpose of the natural mapping O (Ω) −→ O (𝐾) is an injective
continuous linear map O ′ (𝐾) → O ′ (Ω).

Actually, the converse is also true: If 𝐾 ⊂ Ω is compact and if the natural map
O ′ (𝐾) → O ′ (Ω) is injective then 𝐾 is a Runge subset of Ω. If this is so we identify
O ′ (𝐾) with a linear subspace of O ′ (Ω). If 𝐾1 ⊂ 𝐾2 are two Runge compact subsets
of Ω it follows from the maximum principle that O ′ (𝐾1 ) ⊂ O ′ (𝐾2 ) (cf. Example
6.1.4).

Proposition 6.1.7 Suppose the open set Ω is connected and let 𝐾 be an arbitrary
Runge compact subset of Ω. The subspace O ′ (𝐾) is dense in O ′ (Ω).

Proof It suffices to prove the claim when 𝐾 consists of a single point ℘. The dual
of O ′ (Ω) (equipped with either the weak or the strong dual topology) is the space
O (Ω) (the latter and its strong dual are both reflexive, see [Treves, 1967], Ch.
36). Let (𝑈, 𝑧1 , ..., 𝑧 𝑛 ) be a holomorphic local chart in Ω centered at ℘ (i.e., the
holomorphic coordinates 𝑧 𝑗 all vanish at ℘). Let 𝛿 be the Dirac distribution at ℘ and
let 𝛿 ( 𝛼) = 𝜕𝑧𝛼 𝛿 if 𝛼 ∈ Z+𝑛 . It follows directly from the Cauchy inequalities that every
analytic functional 𝛿 ( 𝛼) is carried by {℘}. On the other hand, if ℎ ∈ O (Ω) is such
that 𝛿 ( 𝛼) , ℎ = (−1) | 𝛼 | ℎ ( 𝛼) (0) = 0 for all 𝛼 ∈ Z+𝑛 then ℎ must vanish identically
in Ω. By the Hahn–Banach Theorem, for a linear subspace 𝑽 of a locally convex
space E not to be dense in E it is necessary (and sufficient) that there be a continuous
linear functional on E that vanishes identically on 𝑽 but not on E. □
The space O ′ (Ω) is a module over the commutative ring O (Ω): multiplication of
any analytic functional by a holomorphic function yields an analytic functional. More
generally, if 𝑃 : O (Ω) ←↪ is a continuous linear map its transpose is a continuous
146 6 Analytic Functionals in Euclidean Space

endomorphism 𝑃⊤ : O ′ (Ω) ←↪. This applies, in particular, to complex-analytic

linear partial differential operators, i.e., differential operators
∑︁
𝑃 (𝑧, 𝜕𝑧 ) = 𝑐 𝛼 (𝑧) 𝜕𝑧𝛼 , 𝑐 𝛼 ∈ O (Ω) . (6.1.4)
| 𝛼 | ≤𝑚

There is a natural map of the space of compactly supported distributions in

Ω ⊂ C𝑛 , E ′ (Ω), into O ′ (Ω), the transpose of the natural injection O (Ω) −→
C ∞ (Ω). The natural map E ′ (Ω) −→ O ′ (Ω) is not injective since O (Ω) is a closed
subspace of C ∞ (Ω): O (Ω) is the null space of the Cauchy–Riemann operator 𝜕. ¯
𝜕𝜇 ′
Any distribution 𝜕𝑧¯ 𝑗 with 𝜇 ∈ E (Ω) is orthogonal to O (Ω).

6.2 Analytic Functionals in C 𝒏

6.2.1 The Laplace–Borel transform

Definition 6.2.1 By the Laplace–Borel transform of an analytic functional 𝜇 ∈

O ′ (C𝑛 ) we mean the function of 𝜁 ∈ C𝑛 , L𝜇 (𝜁) = 𝜇 e−𝑧·𝜁 .

We are using the notation 𝑧 · 𝜁 = 𝑧 1 𝜁1 + · · · + 𝑧 𝑛 𝜁 𝑛 .

We recall that an entire function ℎ in C𝑛 is said to be of exponential type if there
are two positive constants 𝐴, 𝐵 such that

∀𝑧 ∈ C𝑛 , |ℎ (𝑧)| ≤ 𝐴 exp (𝐵 |𝑧|) . (6.2.1)

We shall denote by Exp (C𝑛 ) the space of entire functions of exponential type in C𝑛 .

Theorem 6.2.2 The Laplace–Borel transform is a linear isomorphism of O ′ (C𝑛 )

onto Exp (C𝑛 ). If 𝐾 is a compact subset of C𝑛 and 𝜇 is carried by 𝐾 then to every
𝜀 > 0 there is a 𝐶 𝜀 > 0 such that

∀𝜁 ∈ C𝑛 , |L𝜇 (𝜁)| ≤ 𝐶 𝜀 e 𝐻𝐾 (𝜁 )+𝜀 |𝜁 | , (6.2.2)

where 𝐻𝐾 (𝜁) = max (− Re (𝑧 · 𝜁)).

𝑧 ∈𝐾

Proof Let 𝐾 ⊂ C𝑛 be a compact set and let 𝜇 ∈ O ′ (𝐾) be arbitrary. We see that
𝜕𝜁¯ L𝜇 (𝜁) = 𝜇, 𝜕𝜁¯ e−𝑧·𝜁 vanishes identically and therefore LO ′ (C𝑛 ) ⊂ O (C𝑛 ).
To each 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that

∀𝜁 ∈ C𝑛 , |L𝜇 (𝜁)| ≤ 𝐶 𝜀 max e− Re(𝑧·𝜁 ) . (6.2.3)

dist(𝑧,𝐾) ≤ 𝜀

Since
max (− Re (𝑧 · 𝜁)) ≤ 𝐻𝐾 (𝜁) + 𝜀 |𝜁 |
dist(𝑧,𝐾) ≤ 𝜀
6.2 Analytic Functionals in C𝑛 147

(6.2.3) entails (6.2.2). Since |𝐻𝐾 (𝜁)| ≤ 𝐵 𝐾 |𝜁 | we see that (6.2.2) entails (6.2.1)
and thus LO ′ (C𝑛 ) ⊂ Exp (C𝑛 ).
Let ℎ ∈ Exp (C𝑛 ) satisfy (6.2.1); then the coefficients in the Taylor expansion
Í
ℎ (𝜁) = 𝛼∈Z+𝑛 𝑐 𝛼 𝜁 𝛼 satisfy the estimate

1 𝛼 1 1
|𝑐 𝛼 | = 𝜕 ℎ (0) ≤ | 𝛼 | max |ℎ(𝜁)| ≤ | 𝛼 | 𝐴 exp (𝐵𝑅) (6.2.4)
𝛼! 𝜁 𝑅 |𝜁 |=𝑅 𝑅

whatever 𝑅 > 0. We choose 𝑅 = | 𝐵𝛼 | so as to minimize the far right side in (6.2.4);

Stirling’s Formula implies, for a suitably large 𝐶 > 0,
|𝛼|
𝐵|𝛼|

𝐵
|𝑐 𝛼 | ≤ 𝐴 e ≤𝐶 . (6.2.5)
|𝛼| |𝛼|!
Í
It follows from (6.2.5) that whatever 𝑓 ∈ O (C𝑛 ) the series 𝛼∈Z+𝑛 𝑐 𝛼 𝜕𝑧𝛼 𝑓 (0)
converges absolutely and thus defines an analytic functional 𝜇 whose Laplace–Borel
transform is ℎ. This proves that L is surjective. If L𝜇 vanishes identically then

∀𝛼 ∈ Z+𝑛 , 𝜕𝜁𝛼 L𝜇 (𝜁) = 𝜇 (𝑧 𝛼 ) = 0.

𝜁 =0

This means that 𝜇 = 0 since the polynomials in 𝑧 are dense in O (C𝑛 ), proving that
L is injective. □

Remark 6.2.3 Note that if 𝐾 b stands for the closed convex hull of the compact set
𝐾 ⊂ C𝑛 (i.e., the intersection of all the convex compact sets that contain 𝐾) then
𝐻𝐾 = 𝐻𝐾b . The last part of Theorem 6.2.2 has a converse: if to every 𝜀 > 0 there
is a 𝐶 𝜀 > 0 such that (6.2.2) holds then the analytic functional 𝜇 is carried by 𝐾. b
The proof of this fact, however, is not elementary in dimension 𝑛 ≥ 2; the hurdle
is to prove that if 𝜇 is carried by two distinct convex sets then it is carried by their
intersection (see [Martineau, 1964], [Martineau, 1967/68], also [Hörmander, 1966],
Theorem 4.5.3). This “converse” together with Theorem 6.2.2 can be viewed as the
analytic functionals version of the Paley–Wiener Theorem (cf. Theorem 2.1.2).

Let ℎ = L𝜇 have the Taylor expansion (6.2.3) with coefficients 𝑐 𝛼 satisfy-

ing (6.2.5). It is seen immediately that the series 𝛼∈Z+𝑛 (−1) | 𝛼 | 𝑐 𝛼 𝛿 ( 𝛼) converges
Í

(weakly) in O ′ (C𝑛 ) (𝛿 ( 𝛼) : derivative of order 𝛼 of the Dirac distribution at 0). Since

L𝛿 ( 𝛼) = 𝜁 𝛼 the injectivity of L implies
∑︁
𝜇= 𝑐 𝛼 𝛿 ( 𝛼) . (6.2.6)
𝛼∈Z+𝑛

We point out the contrast between the representation (6.2.6) and that in Remark
6.1.2.
148 6 Analytic Functionals in Euclidean Space

Proposition 6.2.4 If ℎ = L𝜇 (𝜁) satisfies (6.2.1) then the closed ball

{𝑧 ∈ C𝑛 ; |𝑧| ≤ 𝐵}

is a carrier of 𝜇.

Proof We have seen that (6.2.1) entails (6.2.5). By (6.2.5) and (6.2.6) the Cauchy
inequalities imply, for every ℎ ∈ O (C𝑛 ) and 𝜀 > 0,
∑︁
|𝜇 (ℎ)| ≤ 𝑐 𝛼 ℎ ( 𝛼) (0)
𝛼∈Z+𝑛
∑︁ 𝐵 | 𝛼 | 𝛼!
≲ max |ℎ (𝑧)|
𝛼∈Z+𝑛
|𝛼|! (𝐵 + 𝜀) | 𝛼 | |𝑧 |=𝐵+𝜀

≲ 1 + 𝐵𝜀 −1 max |ℎ (𝑧)| ,
|𝑧 |=𝐵+𝜀

whence the claim. □

Remark 6.2.5 There is a natural locally convex space structure on Exp (C𝑛 ): call
Exp (C𝑛 , 𝐵) the linear subspace consisting of the functions ℎ that satisfy (6.2.1) for
some 𝐴 > 0; Exp (C𝑛 , 𝐵) is a Banach space with respect to the norm

Exp (C𝑛 , 𝐵) ∋ ℎ ↦→ sup ℎ (𝑧) e−𝐵 |𝑧 | .

A basis of open sets in the topology in Exp (C𝑛 ) consists of the convex sets whose
intersection with Exp (C𝑛 , 𝐵) is open in Exp (C𝑛 , 𝐵) whatever 𝐵 > 0. On the
O ′ (C𝑛 ) as the union of the linear subspaces O ′ 𝐾 𝑗 ,

other hand we can regard

𝑗 = 1, 2, ..., where 𝐾 𝑗 𝑗=1,2,... is an exhausting sequence (meaning 𝐾 𝑗 ⊂ 𝐾 𝑗+1
Ø∞
for all 𝑗 and C𝑛 = 𝐾 𝑗 ) of Runge compact subsets of C𝑛 . We can take 𝐾 𝑗 to
𝑗=1
be the ball {𝑧 ∈ C𝑛 ; |𝑧| ≤ 𝑗 }. Then the strong dual topology on O ′ (C𝑛 ) can be
described asfollows: a convex subset of O ′ (C𝑛 ) is open if and only if its intersection
with O ′ 𝐾 𝑗 is an open subset of O ′ 𝐾 𝑗 whatever 𝑗 = 1, 2, .... Inspection of the
proofs of Theorem 6.2.2 and Proposition 6.2.4 shows directly that the Laplace–Borel
transform is a homeomorphism of O ′ (C𝑛 ) onto Exp (C𝑛 ). It is worthwhile to point
out that a subset 𝔅 of O ′ (C𝑛 ) [resp., Exp (C𝑛 )] is bounded if and only if 𝔅 is a
bounded subset of O ′ 𝐾 𝑗 for some 𝑗 [resp., Exp (C𝑛 , 𝐵) for some 𝐵 > 0].
6.2 Analytic Functionals in C𝑛 149

6.2.2 Convolution of analytic functionals

Let 𝜇 𝑗 ∈ O ′ (C𝑛 ), 𝑗 = 1, 2. The convolution 𝜇1 ★ 𝜇2 of two analytic functionals can

be defined as follows. First we define the tensor products 𝜇1,𝑧 ⊗ 𝜇2,𝑤 by the rule

𝜇1,𝑧 ⊗ 𝜇2,𝑤 , ℎ1 (𝑧) ℎ2 (𝑤) = 𝜇1,𝑧 , ℎ1 (𝑧) 𝜇2,𝑤 , ℎ2 (𝑤) (6.2.7)

for ℎ 𝑗 ∈ O (C𝑛 ), 𝑗 = 1, 2; (6.2.7) determines 𝜇1,𝑧 ⊗𝜇2,𝑤 ∈ O ′ C2𝑛

𝑧,𝑤 unambiguously
since the polynomials 𝑃 (𝑧, 𝑤) ∈ C [𝑧 1 , ..., 𝑧 𝑛 , 𝑤 1 , ..., 𝑤 𝑛 ] are dense in O C2𝑛

𝑧,𝑤 . We
then define, for arbitrary ℎ ∈ O (C𝑛 ),

⟨𝜇1 ★ 𝜇2 , ℎ⟩ = 𝜇1,𝑧 ⊗ 𝜇2,𝑤 , ℎ (𝑧 + 𝑤) (6.2.8)

with the subscripts indicating in which variable the analytic functional is acting;
(6.2.8) is equivalent (as seen by taking ℎ to be a polynomial) to

⟨𝜇1 ★ 𝜇2 , ℎ⟩ = 𝜇1,𝑧 , 𝜇2,𝑤 ℎ (𝑧 + 𝑤) . (6.2.9)

The convolution of analytic functionals is a continuous bilinear map O ′ (C𝑛 ) ×

O ′ (C𝑛 ) −→ O ′ (C𝑛 ). Obviously it is associative and commutative. The unit element
in the convolution algebra O ′ (C𝑛 ) is the Dirac distribution 𝛿. If 𝑃 (𝜕𝑧 ) is a constant
coefficients “holomorphic” differential operator in C𝑛 , i.e., a polynomial with respect
to 𝜕𝑧𝜕1 , ..., 𝜕𝑧𝜕𝑛 , we have

𝑃 (𝜕𝑧 ) (𝜇1 ★ 𝜇2 ) = (𝑃 (𝜕𝑧 ) 𝜇1 ) ★ 𝜇2 = 𝜇1 ★ 𝑃 (𝜕𝑧 ) 𝜇2 (6.2.10)

whence, for every 𝜇 ∈ O ′ (C𝑛 ),

𝑃 (𝜕𝑧 ) 𝜇 = 𝜇 ★ 𝑃 (𝜕𝑧 ) 𝛿. (6.2.11)

Proposition 6.2.6 If 𝜇 𝑗 ∈ O ′ (C𝑛 ) is carried by a compact set 𝐾 𝑗 ⊂ C𝑛 ( 𝑗 = 1, 2)

then 𝜇1 ★ 𝜇2 is carried by 𝐾1 + 𝐾2 .

Proof Follows easily from Definition 6.1.3 and (6.2.8). The proof is left as an
exercise. □

Proposition 6.2.7 For all pairs 𝜇1 , 𝜇2 of analytic functionals in O (C𝑛 ), we have

L (𝜇1 ★ 𝜇2 ) = (L𝜇1 ) L𝜇2 . (6.2.12)

Proof Put ℎ (𝑧) = exp (−𝜁 · 𝑧) in (6.2.8). □

Eq. (6.2.12) shows that we could have defined the convolution of analytic func-
tionals by making use of the inverse of the Laplace–Borel transform, which is
well-defined by Theorem 6.2.2:

𝜇1 ★ 𝜇2 = L −1 (L𝜇1 L𝜇2 ) . (6.2.13)

150 6 Analytic Functionals in Euclidean Space

On the one hand, Exp (C𝑛 ) is an algebra (obviously commutative) with respect to
ordinary multiplication of functions; on the other hand, O ′ (C𝑛 ) is an algebra with
respect to convolution. The Paley–Wiener Theorem 6.2.2 and Formula (6.2.12) tell
us that the Laplace–Borel transform is an algebra isomorphism of O ′ (C𝑛 ) onto
Exp (C𝑛 ).
The compactly supported distributions in R𝑛 i.e., the elements of E ′ (R𝑛 ), can be
identified with analytic functionals in C𝑛 . Indeed, the restriction map O (C𝑛 ) −→
C ∞ (R𝑛 ) has a dense image since polynomials are dense in C ∞ (R𝑛 ) (Weierstrass
Approximation Theorem). Its transpose is a linear injection E ′ (R𝑛 ) −→ O ′ (C𝑛 ).
We can state:

Proposition 6.2.8 Each compactly supported distribution 𝑢 in R𝑛 can be identified

to the analytic functional 𝜇𝑢 in C𝑛 defined as follows:

∀ℎ ∈ O (C𝑛 ) , 𝜇𝑢 (ℎ) = ⟨𝑢, ℎ| R𝑛 ⟩ ;

𝜇𝑢 is carried by supp 𝑢.

Remark 6.2.9 Proposition 6.2.6 may be contrasted with the natural map E ′ R2𝑛 −→

O ′ (C𝑛 ), definitely not injective: its kernel consists of the distributions of the form
Í𝑛 𝜕𝜇 𝑗 ′ 2𝑛 .

𝑗=1 𝜕 𝑧¯ 𝑗 , 𝜇 𝑗 ∈ E R

6.2.3 Analytic functionals carried by R𝒏

The main point of this subsection is to prove that every compact subset of R𝑛 has
the Runge property (Definition 6.1.5). But prior to this we introduce an important
concept in the theory of Several Complex Variables (see, e.g., [Gunning and Rossi,
1965], Ch. I, Section D; [Hörmander, 1966], pp. 52–59).

Definition 6.2.10 A compact subset 𝐾 of C𝑛 is said to be polynomially convex

if given any point 𝑧◦ ∈ C𝑛 \𝐾 there is a polynomial 𝑃 ∈ C [𝑧 1 , ..., 𝑧 𝑛 ] such that
max |𝑃| < |𝑃 (𝑧 ◦ )|. If 𝐾 is an arbitrary compact subset of C𝑛 the polynomially
𝐾
convex hull of 𝐾 is the intersection of all the polynomially convex compact subsets
of C𝑛 that contain 𝐾.

The existence of a 𝑃 ∈ C [𝑧1 , ..., 𝑧 𝑛 ] such that max |𝑃| < |𝑃 (𝑧◦ )| is equivalent
𝐾
to the existence of an ℎ ∈ O (C𝑛 ) such that max |ℎ| < |ℎ (𝑧◦ )| since to every 𝜀 > 0
𝐾
there is a 𝑃 ∈ C [𝑧 1 , ..., 𝑧 𝑛 ] such that max◦ |ℎ − 𝑃| < 𝜀.
𝐾∪{𝑧 }
Later we will need the following consequence of Theorem 2.7.3 in [Hörmander,
1966]:

Theorem 6.2.11 Every polynomially convex compact subset 𝐾 of C𝑛 has the Runge
property (Definition 6.1.5).
6.2 Analytic Functionals in C𝑛 151

Proposition 6.2.12 Every convex compact subset of C𝑛 is polynomially convex.

Proof Let 𝐾 be a convex compact subset of C𝑛 , 𝑧◦ ∈ C𝑛 \𝐾. There is a linear
functional 𝜆 : R2𝑛 −→ R such that 𝜆 (𝑧◦ ) > 𝑎 and max𝜆 < 𝑎 (say, by the Hahn–
𝐾
Banach Theorem). Let 𝐿 (𝑧) be the holomorphic linear function such that 𝜆 (𝑥, 𝑦) =
Re 𝐿 (𝑥 + 𝑖𝑦). If ℎ = exp 𝐿 we have |ℎ (𝑧◦ )| = e𝑎 > max |ℎ|. □
𝐾

Theorem 6.2.13 An arbitrary compact subset 𝐾 of R𝑛 is polynomially convex.

Proof Let 𝑧◦ = 𝑥 ◦ + 𝑖𝑦 ◦ ∉ 𝐾. Suppose 𝑥 ◦ ∉ 𝐾. If ℎ (𝑧) = exp − ⟨𝑥 + 𝑖𝑦 − 𝑥 ◦ ⟩ 2 we

have ℎ (𝑧 ◦ ) = exp |𝑦 ◦ | 2 ≥ 1 while

max |ℎ (𝑥)| = max exp − |𝑥 − 𝑥 ◦ | 2 ≤ exp − dist (𝑥 ◦ , 𝐾) 2 < 1.
𝑥 ∈𝐾 𝑥 ∈𝐾

Suppose 𝑥 ◦ ∈ 𝐾; then 𝑦 ◦ ≠ 0. If ℎ (𝑧) = exp (−𝑖𝑧 · 𝑦 ◦ ) then |ℎ (𝑧◦ )| = exp |𝑦 ◦ | 2 > 1
while |ℎ (𝑥)| = 1 for all 𝑥 ∈ R𝑛 . □
It can be proved directly that every polynomially convex compact subset of C𝑛 is
Runge and has the Cousin property. The latter means that the 𝜕¯ differential complex
(see Section 9.4) is exact in the class of germs (of differential forms) at 𝐾. Here we
prove directly the announced result about compact subsets of R𝑛 :
Theorem 6.2.14 Every compact subset of R𝑛 is Runge.
Proof Let 𝐾 ⊂ R𝑛 be compact and 𝑈 ⊂ C𝑛 be an open set such that 𝐾 ⊂ 𝑈 R =
𝑈 ∩ R𝑛 . Let 𝑉 ⊂⊂ 𝑈 be an open subset of C𝑛 such that 𝐾 ⊂ 𝑉 R = 𝑉 ∩ R𝑛 and which
has the following property:

∀𝑧 ∈ 𝑉, |Im 𝑧| ≤ 𝜀𝑑, (6.2.14)

and 𝜀 > 0 is as small as needed. We select 𝜓 ∈ Cc∞ 𝑈 R

where 𝑑 = dist 𝑉 R , R𝑛 \𝑈 R

such that
1
dist 𝑥, 𝑉 R < 𝑑 =⇒ 𝜓 (𝑥) = 1,
3
2
dist 𝑥, 𝑉 R > 𝑑 =⇒ 𝜓 (𝑥) = 0,
3
∀𝑥 ∈ 𝑈 R , 0 ≤ 𝜓 (𝑥) ≤ 1.

Given ℎ ∈ O (𝑈) and 𝑧 ∈ 𝑉 we form

12 𝑛 ∫ 𝑛
𝑘 © ∑︁ 2ª
ℎ 𝑘 (𝑧) = 𝜓 (𝜉) ℎ (𝜉) exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 ® d𝜉, 𝑘 = 1, 2, .... (6.2.15)
𝜋 R𝑛 𝑗=1
« ¬
By letting 𝜕𝑧¯ act under the integral sign we see immediately that ℎ 𝑘 ∈ O (C𝑛 ). We
are going to exploit the fact that, whatever 𝑧 ∈ C𝑛 ,
152 6 Analytic Functionals in Euclidean Space

12 𝑛 ∫ 𝑛
𝑘 © ∑︁ 2ª
exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 ® d𝜉 = 1. (6.2.16)
𝜋 R𝑛 𝑗=1
« ¬
[Proof: (6.2.16) is true when 𝑧 ∈ R𝑛 ; since both sides are entire functions of 𝑧 it
must be true in the whole of C𝑛 .] We derive from (6.2.16):
12 𝑛 ∫ 𝑛
𝑘 © ∑︁ 2ª
ℎ 𝑘 (𝑧) − ℎ (𝑧) = 𝜓 (𝜉) (ℎ (𝜉) − ℎ (𝑧)) exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 ® d𝜉
𝜋 R𝑛 𝑗=1
« ¬
(6.2.17)
12 𝑛 ∫ 𝑛
𝑘 © ∑︁ 2ª
− ℎ (𝑧) (1 − 𝜓 (𝜉)) exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 ® d𝜉.
𝜋 R𝑛 𝑗=1
« ¬
Let us call 𝐼1 (𝑧) and 𝐼2 (𝑧) respectively the first and second integral in the right-
hand side of (6.2.17). Let 𝑊 be an open subset of C𝑛 such that 𝐾 ⊂ 𝑊 ⊂⊂ 𝑉; we
take 𝑧 = 𝑥 + 𝑖𝑦 ∈ 𝑊. In handling 𝐼2 we exploit (6.2.14), where we require 0 < 𝜀 < 31 :
1 1 √
1 − 𝜓 (𝜉) ≠ 0 =⇒ dist 𝜉, 𝑉 R > 𝑑 =⇒ |𝑥 − 𝜉 | > 𝑑 > 2 |𝑦| ,
3 3
whence
12 𝑛 ∫
𝑘 1 2
|𝐼2 (𝑧)| ≤ |ℎ (𝑧)| e− 2 𝑘 | 𝑥− 𝜉 | d𝜉
𝜋 1
| 𝑥− 𝜉 |> 3 𝑑
12 𝑛 ∫
1 2 𝑘 1 2
≤ e− 12 𝑘𝑑 |ℎ (𝑧)| e− 4 𝑘 | 𝜉 | d𝜉
𝜋 R 𝑛
1 2
≤ 2𝑛 e− 12 𝑘𝑑 |ℎ (𝑧)| .

We conclude that 𝐼2 → 0 uniformly on 𝑊 as 𝑘 → +∞.

We regard the integrand in 𝐼1 as the restriction of the function
12 𝑛 𝑛
𝑘 © ∑︁ 2ª
𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑡𝑦) = 𝜓 (𝜉) (ℎ (𝜉 + 𝑖𝑡𝑦) − ℎ (𝑧)) exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 − 𝑖𝑡𝑦 ®
𝜋 𝑗=1
« ¬
to the “floor” 𝑡 = 0 of the (𝑛 + 1)-dimensional “cylinder” 𝑉 R + 𝑖 [0, 1] 𝑦. We select
𝜀 in (6.2.14) to be sufficiently small that 𝜉 + 𝑖𝑡𝑦 stays in a compact subset of 𝑈 if
|𝜉 | ≤ 32 𝑑, |𝑦| ≤ 𝜀𝑑, 0 ≤ 𝑡 ≤ 1. The restriction to the “roof” 𝑡 = 1 is
12 𝑛
𝑘
𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑦) = 𝜓 (𝜉) (ℎ (𝜉 + 𝑖𝑦) − ℎ (𝑧)) exp −𝑘 |𝑥 − 𝜉 | 2 ;
𝜋

we have
6.2 Analytic Functionals in C𝑛 153
∫ ∫ ∫ 1∫
𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑦) d𝜉 − 𝐹𝑘 (𝑧, 𝜉) d𝜉 = 𝜕𝑡 𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑡𝑦) d𝜉d𝑡,
R𝑛 R𝑛 0 R𝑛

and

𝜕𝑡 𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑡𝑦)
12 𝑛 𝑛
𝑘 © ∑︁ 2 ªª
=𝑖 𝜓 (𝜉) 𝑦 · 𝜕 𝜉 ­ (ℎ (𝜉 + 𝑖𝑡𝑦) − ℎ (𝑧)) exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 − 𝑖𝑡𝑦 ®® .
©
𝜋 𝑗=1
« « ¬¬
Integration by parts yields
−𝑛/2 ∫ 1 ∫
𝑘
𝑖 𝜕𝑡 𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑡𝑦) d𝜉d𝑡 (6.2.18)
𝜋 0 R𝑛
∫ 1∫ 𝑛
© ∑︁ 2ª
= 𝑦 · 𝜕 𝜉 𝜓 (𝜉) (ℎ (𝜉 + 𝑖𝑡𝑦) − ℎ (𝑧)) exp ­−𝑘 𝑧 𝑗 − 𝜉 𝑗 − 𝑖𝑡𝑦 ® d𝜉d𝑡.
0 R𝑛 𝑗=1
« ¬
We have 𝐾1 = supp 𝜕 𝜉 𝜓 ⊂ supp 𝜓 ∩ supp (1 − 𝜓). We can repeat the argument used
in estimating |𝐼2 | to derive an estimate
∫ 1∫
1 2 1
𝜕𝑡 𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑡𝑦) d𝜉d𝑡 ≤ 𝐶e− 12 𝑘𝑑 sup ℎ 𝑧 + 𝜉 − ℎ (𝑧) ,
0 R𝑛 𝜉 ∈𝐾1 𝑘

showing that (6.2.18) converges to zero, again uniformly on 𝑊, as 𝑘 → +∞.

It remains to prove the same for
∫ 12 𝑛 ∫
𝑘 2
𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑦) d𝜉 = 𝜓 (𝜉) (ℎ (𝜉 + 𝑖𝑦) − ℎ (𝑧)) e−𝑘 | 𝑥− 𝜉 | d𝜉
R𝑛 𝜋 𝑛
∫ R 2
− 21 𝑛
=𝜋 𝜓 𝑥 + 𝑘 −1 𝜉 ℎ 𝑧 + 𝑘 −1 𝜉 − ℎ (𝑧) e− | 𝜉 | d𝜉.
R𝑛

The desired property is a direct consequence of the Lebesgue Dominated Conver-

gence Theorem, since
∫
lim 𝐹𝑘 (𝑧, 𝜉 + 𝑖𝑦) d𝜉
𝑘→+∞ R𝑛
∫
1 2
= 𝜋− 2 𝑛 lim 𝜓 𝑥 + 𝑘 −1 𝜉 ℎ 𝑧 + 𝑘 −1 𝜉 − ℎ (𝑧) e− | 𝜉 | d𝜉.
R𝑛 𝑘→+∞

[The last integrand makes sense because 𝜓 𝑥 + 𝑘 −1 𝜉 ≠ 0 and 𝑧 ∈ 𝑊 entail 𝑧+𝑘 −1 𝜉 ∈

𝑉 for 𝑘 suitably large.] □
Corollary 6.2.15 If 𝐾 is a compact subset of R𝑛 the natural map O ′ (𝐾) −→ O ′ (C𝑛 )
is injective.
Proof Combine Proposition 6.1.6 and Theorem 6.2.14. □
154 6 Analytic Functionals in Euclidean Space

When 𝐾 ⊂⊂ R𝑛 we may regard O ′ (𝐾) as a vector subspace of O ′ (C𝑛 ). Recall

that O ′ (𝐾) has the natural structure of a Fréchet space defined by the norms (6.1.3);
O ′ (R𝑛 ) can be equipped with the inductive limit of the spaces O ′ (𝐾) as 𝐾 ranges
over the family of compact subsets of R𝑛 : a convex subset of O ′ (R𝑛 ) is open if
its intersection with each O ′ (𝐾) is open in O ′ (𝐾); a linear map of O ′ (R𝑛 ) into
a locally convex topological vector space E is continuous if its restriction to each
O ′ (𝐾) is continuous.
Recall how the topology of C 𝜔 (R𝑛 ) is defined (cf. Subsection 1.2.1). There is
a basis of neighborhoods of R𝑛 in C𝑛 , {U 𝜄 } 𝜄 ∈𝐼 , each one of which is Runge (this
follows easily from the results about compact subsets of R𝑛 stated above); C 𝜔 (R𝑛 )
can be regarded as the projective limit of the topological vector spaces O (U 𝜄 ) (this
is essentially the same as saying that a C 𝜔 function in R𝑛 extends holomorphically
to a neighborhood of R𝑛 in C𝑛 ). The dual of C 𝜔 (R𝑛 ) can be identified with the
intersection of all the spaces O ′ (U 𝜄 ). Restriction to R𝑛 maps O (C𝑛 ) onto a dense
linear subspace of C 𝜔 (R𝑛 ) (Weierstrass Approximation Theorem); it follows that
there is a natural (continuous) linear injection of the dual of C 𝜔 (R𝑛 ) into O ′ (C𝑛 );
the range of this map is obviously equal to O ′ (R𝑛 ).
Proposition 6.2.16 The topological vector spaces C 𝜔 (R𝑛 ) and O ′ (R𝑛 ) can be
identified with the (strong) dual of each other.
Proof Both C 𝜔 (R𝑛 ) and O ′ (R𝑛 ) are reflexive (i.e., each is the dual of its own dual,
both being regarded as locally convex topological vector spaces). □
If we combine Theorem 6.2.2 and Remark 6.2.3 we can state
Proposition 6.2.17 For 𝜇 ∈ O ′ (C𝑛 ) to be carried by a convex compact subset 𝐾 of
R𝑛 it is necessary and sufficient that the Laplace–Borel transform of 𝜇 (Definition
6.2.1) have the following property: there are constants 𝑀 > 0 and, to every 𝜀 > 0,
𝐶 𝜀 > 0, such that

∀𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 , |L𝜇 (𝜁)| ≤ 𝐶 𝜀 e 𝐻𝐾 ( 𝜉 )+𝜀 |𝜁 | , (6.2.19)

where 𝐻𝐾 (𝜉) = max (−𝑥 · 𝜉).

𝑥 ∈𝐾

6.3 Analytic Functionals in R 𝒏 as Cohomology Classes

6.3.1 Analytic functionals in R𝒏 defined by integrals over closed

hypersurfaces

Let 𝐾 ⊂ R𝑛 be compact; we consider a differential form in C𝑛 \𝐾 of the following

kind:
𝑛
∑︁
𝑓 (𝑧, d𝑧, d𝑧¯) = (−1) 𝑛− 𝑗 𝑓 𝑗 (𝑥, 𝑦) d𝑧 ∧ d𝑧¯1 ∧ · · · ∧ dc
𝑧¯ 𝑗 ∧ · · · ∧ d𝑧¯𝑛 (6.3.1)
𝑗=1
6.3 Analytic Functionals in R𝑛 as Cohomology Classes 155

where d𝑧 = d𝑧1 ∧ · · · ∧ d𝑧 𝑛 ; the “hat” covers a factor to omit; the coefficients

𝑓 𝑗 belong to C ∞ (C𝑛 \𝐾). In real terms 𝑓 (𝑧, d𝑧, d𝑧¯) is a special kind of (2𝑛 − 1)-
differential form in R2𝑛 \𝐾 to which we refer as a differential form of bidegree
(𝑛, 𝑛 − 1) or an (𝑛, 𝑛 − 1)-form.
Such forms make up a linear space which we denote
∞
by C C \𝐾; 𝑇𝑛 (𝑛,𝑛−1) in this section. The action of the exterior derivative d on
(6.3.1) coincides with that of the operator 𝜕: ¯

𝑛
©∑︁ 𝜕 𝑓 𝑗 ª

d𝑓 = ­ ® d𝑧 ∧ d𝑧¯ ∈ C ∞ C𝑛 \𝐾; 𝑇 (𝑛,𝑛) . (6.3.2)
𝜕 𝑧¯ 𝑗
« 𝑗=1 ¬

The differential operator 𝑓 ↦→ d 𝑓 is the tail end of the 𝜕 differential complex in

bidegree (𝑛, 𝑞) (see Section 9.4).
Let 𝑈 be a bounded domain in C𝑛 , 𝑈 ⊃ 𝐾, whose boundary 𝜕𝑈 is a C ∞ hyper-
surface in R2𝑛 . A form (6.3.1) defines a linear functional
∫
𝑛
O (C ) ∋ ℎ ↦→ ℎ (𝑧) 𝑓 (𝑧, d𝑧, d𝑧¯) , (6.3.3)
𝜕𝑈

obviously continuous with respect to the topology on O (C𝑛 ) of uniform convergence

on compact subsets of C𝑛 ; in other words, (6.3.3) defines an analytic functional 𝜇 𝑓
carried by the closure of 𝑈. If 𝑓 is closed, meaning d 𝑓 ≡ 0, it follows from Stokes’
Theorem that 𝜇 𝑓 is independent of the choice of 𝑈 and thus defines an analytic
functional
carried by𝐾. Now suppose 𝑓 is exact, meaning 𝑓 = 𝜕𝑢 = d𝑢 for some 𝑢 ∈
C C \𝐾; 𝑇 (𝑛,𝑛−2) ; let 𝜒 ∈ C ∞ R2𝑛 , supp 𝜒 ⊂ C𝑛 \𝐾, 𝜒 ≡ 1 in a neighborhood
∞ 𝑛

of 𝜕𝑈. We have, again by Stokes’ Theorem,

∫ ∫
ℎ𝑓 = ℎ𝜕 ( 𝜒𝑢)
𝜕𝑈
∫𝜕𝑈
= 𝜕 ℎ𝜕 ( 𝜒𝑢)
∫𝑈
= d (ℎd ( 𝜒𝑢)) = 0.
𝑈

Thus, if we denote by ker 𝜕 (resp., im 𝜕) the linear space of the closed (resp., exact)
(𝑛, 𝑛 − 1)-forms we see that 𝜇 𝑓 does not depend on the individual 𝑓 ∈ ker 𝜕 but
solely on its coset [ 𝑓 ] ∈ ker 𝜕/im 𝜕. This quotient linear space is usually denoted
by 𝐻 𝑛,𝑛−1 (C𝑛 \𝐾) and referred to as the cohomology space on C𝑛 \𝐾 of bidegree
(𝑛, 𝑛 − 1); its elements are the 𝜕 cohomology classes of said bidegree. We have
defined a natural linear map

𝐻 (𝑛,𝑛−1) (C𝑛 \𝐾) −→ O ′ (𝐾) . (6.3.4)

156 6 Analytic Functionals in Euclidean Space

Our purpose, in this section, is to establish and exploit the properties of this map.
There is an important difference between the cases 𝑛 = 1 and 𝑛 ≥ 2, beyond
the fact that 𝑇 (𝑛,𝑛−2) makes no obvious sense when 𝑛 = 1. On the one hand, in one
dimension the tools of analysis are simpler; on the other hand, when 𝑛 ≥ 2 there is the
consequential Hartog’s Theorem (see [Hörmander, 1966], p. 30, [Stein-Shakarchi,
2011], Vol. IV, p. 286) according to which every ℎ ∈ O (C𝑛 \𝐾) extends as an entire
function in C𝑛 .

6.3.2 The one-dimensional case

In this subsection 𝐾 is a compact subset of C; a coset [ 𝑓 ] ∈ 𝐻 1,0 (C\𝐾) has a single

representative 𝑓 (𝑧) d𝑧 with 𝑓 ∈ O (C\𝐾). The map (6.3.4) reduces to the map
∫
O (C\𝐾) ∋ 𝑓 ↦→ 𝜇 𝑓 = O (C) ∋ ℎ ↦→ ℎ (𝑧) 𝑓 (𝑧) d𝑧 ∈ O ′ (C)
𝜕𝑈

whose null space obviously contains the subspace O (C) of O (C\𝐾). To get the
correct replacement of the map (6.3.4) we introduce the space O◦ (C\𝐾) of holo-
morphic functions in C\𝐾 that tend to zero as 𝑧 −→ ∞. We may identify O◦ (C\𝐾)
with the space of holomorphic functions in S2 \𝐾 (S2 : the Riemann sphere) that
vanish at ∞; as such O◦ (C\𝐾) is a closed linear subspace of O S2 \𝐾 and inherits
from the latter a Fréchet space structure. The main result of this subsection can now
be stated.
Theorem 6.3.1 If 𝐾 is a Runge compact subset of C the map O◦ (C\𝐾) ∋ 𝑓 ↦→ 𝜇 𝑓
is a bijection onto O ′ (𝐾).
Actually, it is easy to construct an inverse of the map O◦ (C\𝐾) ∋ 𝑓 ↦→ 𝜇 𝑓 and
show that the inverse is an isomorphism. Let 𝜇 ∈ O ′ (𝐾). If 𝑧 ∉ 𝐾 the function
(𝑧 − 𝑤) −1 is holomorphic with respect to 𝑤 in every neighborhood 𝑈 ⊂ C\ {𝑧} of 𝐾
and we can therefore form

1 1
ℭ𝜇 (𝑧) = 𝜇𝑤 , , (6.3.5)
2𝜋𝑖 𝑧−𝑤

where the subscript indicates that the functional 𝜇 𝑤 acts in 𝑤-space.

Definition 6.3.2 The function (6.3.5) is called the Cauchy transform of the analytic
functional 𝜇.
In what follows we are going to make use of two simple results of Functional
Analysis. We denote by E a topological vector space (over C): this means that the
topology of E is such that vector addition E × E ∋ (u, v) ↦→ u + v ∈ E and scalar
multiplication C × E ∋ (𝑐, u) ↦→ 𝑐u ∈ E are continuous operations. All the vector
space topologies encountered in this book will be locally convex, meaning that every
neighborhood of a point contains a convex neighborhood of the point.
6.3 Analytic Functionals in R𝑛 as Cohomology Classes 157

Lemma 6.3.3 Let R ⊃ (𝑎, 𝑏) ∋ 𝑡 → u (𝑡) ∈ E be a C 1 mapping. Then, whatever the

continuous linear functional 𝜆 on E we have

d du
⟨𝜆, u (𝑡)⟩ = 𝜆, (𝑡) .
d𝑡 d𝑡

Proof Indeed, by continuity of 𝜆,

1 u (𝑡 + ℎ) − u (𝑡)
lim (⟨𝜆, u (𝑡 + ℎ) − u (𝑡)⟩) = 𝜆, lim . □
0≠ℎ→0 ℎ 0≠ℎ→0 ℎ

The obvious generalization of Lemma 6.3.3 to several variables is immediate.

Lemma 6.3.4 Let E be a Fréchet space. Let [𝑎, 𝑏] ⊂⊂ R and [𝑎, 𝑏] ∋ 𝑡 → u (𝑡) ∈ E
be a continuous mapping. Then, whatever the continuous linear functional 𝜆 on E
we have ∫ 𝑏 ∫ 𝑏
⟨𝜆, u (𝑡)⟩ d𝑡 = 𝜆, u (𝑡) d𝑡 .
𝑎 𝑎
Í 𝑁 𝑏−𝑎 𝑘

Proof It is not difficult to see that the Riemann sums 𝑘=0 𝑁 u 𝑎 + 𝑁 (𝑏 − 𝑎) ,
∫𝑏
𝑁 = 1, 2, ..., form a Cauchy sequence in E; 𝑎 ⟨𝜆, u (𝑡)⟩ d𝑡 is the limit of

𝑁
∑︁ 𝑏−𝑎 𝑘
𝜆, u 𝑎 + (𝑏 − 𝑎) ,
𝑘=0
𝑁 𝑁

whence the claim. □

Recall that O ′ (𝐾) carries a natural Fréchet space structure, defined by the semi-
norms (6.1.3). Theorem 6.3.1 is a consequence of the following.
Theorem 6.3.5 If 𝐾 is a Runge compact subset of C then the Cauchy transform
defines a Fréchet space isomorphism of O ′ (𝐾) onto O◦ (C\𝐾).
Proof Suppose 𝐾 is Runge. Thanks to Lemma 6.3.3 we may let 𝜕𝑧¯ act under the
duality bracket in (6.3.5); we see directly that ℭ𝜇 (𝑧) is a holomorphic function in
C\𝐾. Let 𝑈 be as in (6.3.3), a bounded domain containing 𝐾 and having a smooth
boundary 𝜕𝑈 which inherits an orientation from the plane, and Ω be a domain in
C such that 𝑈 ⊂⊂ Ω. Since O (Ω) is a Fréchet space we can apply Lemma 6.3.4:
whatever ℎ ∈ O (Ω) we have
∮ * ∮ +
1 1
ℎ (𝑧) ℭ𝜇 (𝑧) d𝑧 = 𝜇 𝑤 , ℎ (𝑧) d𝑧 = ⟨𝜇, ℎ⟩ . (6.3.6)
2𝜋𝑖 𝑧−𝑤
𝜕𝑈 𝜕𝑈

In the language of the previous subsection, (6.3.6) can be rewritten as 𝜇 𝑓 = 𝜇 if

𝑓 = ℭ𝜇 as well as ℭ𝜇 𝑓 = 𝑓 , proving that the Cauchy transform is bijective. By the
Closed Graph Theorem, for ℭ to be a homeomorphism it suffices that the linear map
ℭ𝜇 ↦→ 𝜇 be continuous. We derive from (6.3.6):
158 6 Analytic Functionals in Euclidean Space

∀ℎ ∈ O (Ω) , sup |ℎ| ≤ 1 =⇒ |⟨𝜇, ℎ⟩| ≤ 𝐶𝑈 max |ℭ𝜇| . □

𝑈 𝜕𝑈

It is a classical result of one complex variable theory that a compact set 𝐾 ⊂ C

is Runge (Definition 6.1.5) if and only if C\𝐾 is connected (see, e.g., [Hörmander,
1966], Theorem 1.3.1, or [Conway, 1973], Ch. VIII).

Theorem 6.3.6 Let 𝐾1 and 𝐾2 be two Runge compact subsets of C such that 𝐾1 ∪ 𝐾2
is Runge. The following properties hold, for every 𝜇 ∈ O ′ (C),

(1) If 𝜇 ∈ O ′ 𝐾 𝑗 , 𝑗 = 1, 2, then 𝜇 ∈ O ′ (𝐾1 ∩ 𝐾2 ).

(2) If 𝜇 ∈ O ′ (𝐾1 ∪ 𝐾2 ) then there exist 𝜇 𝑗 ∈ O ′ 𝐾 𝑗 , 𝑗 = 1, 2, such that 𝜇 = 𝜇1 +𝜇2 .

Proof By Proposition 6.2.8, (1) is equivalent to the following claim:

O◦ (C\𝐾1 ) ∩ O◦ (C\𝐾2 ) = O◦ (C\ (𝐾1 ∩ 𝐾2 )) , (6.3.7)

whereas (2) is equivalent to

O◦ (C\𝐾1 ) + O◦ (C\𝐾2 ) = O◦ (C\ (𝐾1 ∪ 𝐾2 )) . (6.3.8)

The proof of (6.3.7) is a simple exercise in one complex variable. We limit ourselves
to the proof of (6.3.8). Let Ω 𝑗 = C\𝐾 𝑗 , 𝑗 = 1, 2. We can find a function 𝜒 ∈
C ∞ (Ω1 ∪ Ω2 ) such that supp 𝜒 ⊂ Ω1 , supp (1 − 𝜒) ⊂ Ω2 and therefore supp 𝜕𝑧¯ 𝜒 ⊂
Ω1 ∩ Ω2 . Now let ℎ ∈ O◦ (C\ (Ω1 ∩ Ω2 )). We can regard (1 − 𝜒) ℎ as a C ∞ function
in Ω1 , 𝜒ℎ as a C ∞ function in Ω2 and ℎ𝜕𝑧¯ 𝜒 as a C ∞ function in Ω1 ∪ Ω2 , by setting
these three functions to be equal to zero in (Ω1 ∪ Ω2 ) \ (Ω1 ∩ Ω2 ). After selecting a
solution 𝑢 ∈ C ∞ (Ω1 ∪ Ω2 ) of 𝜕𝑧¯ 𝑢 = ℎ𝜕𝑧¯ 𝜒 we define

ℎ1 = (1 − 𝜒) ℎ + 𝑢 ∈ O (Ω1 ) , ℎ2 = 𝜒ℎ − 𝑢 ∈ O (Ω2 ) .

It is obvious that ℎ = ℎ1 + ℎ2 in Ω1 ∩ Ω2 . By hypothesis Ω1 ∩ Ω2 is connected and

contains the region {𝑧 ∈ C; |𝑧| > 𝑅} provided 𝑅 > 0 is sufficiently large. Consider
then the Laurent expansions of ℎ1 and ℎ2 :
∑︁
𝑐 𝑗,𝑛 𝑧 𝑛 , 𝑗 = 1, 2.
𝑛∈Z

Since ℎ (𝑧) → 0 as 𝑧 −→ ∞ necessarily 𝑐 1,𝑛 + 𝑐 2,𝑛 = 0for all 𝑛 ∈ Z+ and we can

simply replace ℎ 𝑗 by ℎ˜ 𝑗 = ℎ 𝑗 − +∞ ˜ 𝑗 ∈ O◦ Ω 𝑗 is unambiguously defined
Í 𝑛; ℎ
𝑛=0 𝑐 𝑗,𝑛 𝑧
because Ω 𝑗 is connected. □

Remark 6.3.7 Unless one assumes that 𝐾1 ∪ 𝐾2 is Runge, statement (2) in Theorem
6.3.6 does not make sense, since O ′ (𝐾1 ∪ 𝐾2 ) cannot be regarded as a subspace of
O ′ (C) like O ′ 𝐾 𝑗 , 𝑗 = 1, 2. Moreover (6.3.8) might not be true, as shown in the
following.
6.3 Analytic Functionals in R𝑛 as Cohomology Classes 159

Example 6.3.8 Take 𝐾1 (resp., 𝐾2 ) to be the closed upper (resp., lower) semicircle
𝑧 = e𝑖 𝜃 , 0 ≤ 𝜃 ≤ 𝜋 (resp., −𝜋 ≤ 𝜃 ≤ 0); set Ω 𝑗 = C\𝐾 𝑗 . We see that

Ω1 ∩ Ω2 = {𝑧 ∈ C; |𝑧| ≠ 1} , Ω1 ∪ Ω2 = {𝑧 ∈ C; 𝑧 ≠ ±1} ;

Ω1 ∩ Ω2 is not connected.
Take ℎ (𝑧) = 1 if |𝑧| < 1, ℎ (𝑧) = 0 if |𝑧| > 1. If there were
functions ℎ 𝑗 ∈ O Ω 𝑗 such that ℎ = ℎ1 + ℎ2 we would have ℎ1 + ℎ2 = 0 in the region
{𝑧 ∈ C; |𝑧| > 1} and therefore ℎ1 and −ℎ2 would have holomorphic extensions to
the whole of Ω1 ∪ Ω2 necessarily identical, implying ℎ ≡ 0 if |𝑧| < 1 and thereby
contradicting the definition of ℎ.

Corollary 6.3.9 If an analytic functional is carried by two compact subsets of R then

𝜇 is carried by their intersection.

Proof Follows from Theorem 6.3.6 and the fact that every compact set 𝐾 ⊂ R is
Runge since C\𝐾 is obviously connected. □
Corollary 6.3.9 has the implication that, for an analytic functional 𝜇 in C carried
by the real line, the notion of support of 𝜇 makes sense (cf. Corollary 6.3.12 below):
supp 𝜇 is the intersection of all the compact subsets of R that carry 𝜇.

6.3.3 The higher-dimensional cases

In this subsection we prove the following

Theorem 6.3.10 If 𝑛 ≥ 2 the map (6.3.4), 𝐻 (𝑛,𝑛−1) (C𝑛 \𝐾) −→ O ′ (𝐾), is a linear
bijection.

Exceptionally, we are going to invoke and use concepts and results which will be
described only later. We are referring to the 𝜕¯ complex (Subsection 9.4.5) and the
Bochner–Martinelli formula (11.1.24).
Proof I. The map (6.3.4) is surjective. Let 𝜇 ∈ O ′ (𝐾) be arbitrary and let Ω be a
bounded domain in C𝑛 such that 𝐾 ⊂ Ω. The topology of O (Ω) is the same as that
induced on O (Ω) by D ′ (Ω); it follows from the Hahn–Banach Theorem that there
is a function 𝜑 ∈ Cc∞ (Ω) such that
∫
∀ℎ ∈ O (C𝑛 ) , ⟨𝜇, ℎ⟩ = ℎ (𝑧) 𝜑 (𝑥, 𝑦) d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯1 ∧ · · · ∧ d𝑧¯𝑛 .
Ω

We apply (11.1.24), with Φ = 𝜑 (𝑥, 𝑦) d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯1 ∧ · · · ∧ d𝑧¯𝑛 hence 𝑞 = 𝑛

and Φ = 𝜕K¯ (𝑛) Φ. We define 𝑓 = K (𝑛) Φ ∈ C ∞ R𝑛 ; 𝑇 (𝑛,𝑛−1) and select a domain
Ω Ω
𝑈 such that 𝐾 ⊂ 𝑈 ⊂⊂ Ω and 𝜕𝑈 is a C ∞ hypersurface. Applying Stokes’ Theorem
yields 𝜇 = 𝜇 𝑓 [see (6.3.3)] thereby proving the surjectivity of (6.3.4).
160 6 Analytic Functionals in Euclidean Space

II. The map (6.3.4) is injective. Let 𝑓 ∈ C ∞ C𝑛 \𝐾; 𝑇 (𝑛,𝑛−1) be closed and
∫
satisfy 𝜕𝑈 ℎ 𝑓 = 0 for all ℎ ∈ O (C𝑛 ); the latter remains true if we multiply 𝑓 by
𝜒 ∈ C ∞ R2𝑛 , supp 𝜒 ⊂ C𝑛 \𝐾, 𝜒 ≡ 1 in C𝑛 \𝐾 ′ with 𝐾 ′ a compact subset of 𝑈

whose interior contains 𝐾. We define 𝜓 by the equation

𝜓d𝑧 ∧ d𝑧¯ = 𝜕¯ ( 𝜒 𝑓 ) = −𝜕¯ ((1 − 𝜒) 𝑓 ) ;

since (1 − 𝜒) 𝑓 ∈ Cc∞ C𝑛 \𝐾; 𝑇 (𝑛,𝑛−1) Stokes’ Theorem implies
∫
𝑛
∀ℎ ∈ O (C ) , ℎ𝜓d𝑥d𝑦 = 0. (6.3.9)
R2𝑛

We have 𝜓 ∈ C ∞ (C𝑛 ), supp 𝜓 ⊂ 𝐾 ′\𝐾.

Since 𝑛 ≥ 2 we can then use the fact that the range of 𝜕¯ : Cc∞ (C𝑛 ) −→
Cc C𝑛 ; 𝑇 0,1 is exactly equal to the kernel of the map 𝜕¯ : Cc∞ C𝑛 ; 𝑇 (0,1) −→
∞


Cc∞ C𝑛 ; 𝑇 (0,2) . This is easy to prove: let 𝑔 = 𝑛𝑗=1 𝑔 𝑗 d𝑧¯ 𝑗 , 𝑔 𝑗 ∈ Cc∞ (C𝑛 ) be such
Í
𝜕𝑔
that 𝜕𝑧¯ 𝑘𝑗 = 𝜕𝑔
𝜕 𝑧¯ 𝑗 for all 𝑗, 𝑘 = 1, ..., 𝑛. We apply Formula (11.1.24), with Ω = C :
𝑘 𝑛
∞
there is a 𝑢 ∈ C (C ) such that 𝜕𝑧¯ 𝑢 = 𝑔. We have 𝑢 ∈ O (Ω1 ), where Ω1 is the
𝑛
Ø𝑛
unbounded connected component of C𝑛 \ supp 𝑔 𝑗 . We apply Hartog’s theorem
𝑗=1
(see [Hörmander, 1966], p. 30, Th. 2.3.2, [Stein-Shakarchi, 2011], Vol. IV, p. 286,
Th. 3.3): there is a 𝑢˜ ∈ O (C𝑛 ) such that 𝑢˜ ≡ 𝑢 in Ω1 ; it follows that 𝑢 − 𝑢˜ ∈ Cc∞ (C𝑛 )
and 𝜕¯ (𝑢 − 𝑢)
˜ = 𝑔 in C𝑛 , which proves the claim.
We exploit this result in the framework of the duality between scalar distributions
in R2𝑛 and compactly supported C ∞ differential 2𝑛-forms in R2𝑛 . In this sense the
¯ ′ 2𝑛 ′ 2𝑛 (0,1)

orthogonal of O (C ), the kernel of 𝜕 : D R
𝑛 −→ D R ; 𝑇 , is equal to
the range of
𝜕¯ : Cc∞ R2𝑛 ; 𝑇 (𝑛,𝑛−1) −→ Cc∞ R2𝑛 ; 𝑇 (𝑛,𝑛) .

In view of this fact (6.3.9) entails 𝜓 = 𝜕𝜑 ¯ for some 𝜑 ∈ Cc∞ C𝑛 ; 𝑇 (𝑛,𝑛−1) . It
′ ¯
follows directly that 𝜑 ≡ 0 in C \𝐾 ⊃ 𝜕𝑈. We have 𝜕 ( 𝜒 𝑓 − 𝜑) = 0 whence
𝑛

¯
𝜒 𝑓 = 𝜑 + 𝜕𝜆 for some 𝜆 ∈ C C ; 𝑇 ∞ 𝑛 (𝑛,𝑛−2) . We conclude that 𝜒 𝑓 = 𝑓 = 𝜕𝜆 ¯ in
C𝑛 \𝐾 ′. By contracting 𝐾 ′ about 𝐾 one can easily conclude that the cohomology
class [ 𝑓 ] ∈ 𝐻 (𝑛,𝑛−1) (C𝑛 \𝐾) vanishes identically. □
Thus, if 𝐾 ⊂ R𝑛 (𝑛 ≥ 2) we have the isomorphism

O ′ (𝐾) 𝐻 (𝑛,𝑛−1) (C𝑛 \𝐾) . (6.3.10)

We are now in a position to extend Theorem 6.3.6 to higher dimensions and

compact subsets of real space:
6.3 Analytic Functionals in R𝑛 as Cohomology Classes 161

Theorem 6.3.11 Let 𝐾1 and 𝐾2 be two compact subsets of R𝑛 (𝑛 ≥ 2). The following
properties hold, for every 𝜇 ∈ O ′ (C):
(1) If 𝜇 ∈ O ′ 𝐾 𝑗 , 𝑗 = 1, 2, then 𝜇 ∈ O ′ (𝐾1 ∩ 𝐾2 ).

(2) If 𝜇 ∈ O ′ (𝐾1 ∪ 𝐾2 ) then there exist 𝜇 𝑗 ∈ O ′ 𝐾 𝑗 , 𝑗 = 1, 2, such that 𝜇 = 𝜇1 +𝜇2 .
Proof We shall avail ourselves repeatedly of (6.3.10). Let Ω 𝑗 = C𝑛 \𝐾 𝑗 and 𝜑 𝑗 ∈
C ∞ (Ω1 ∪ Ω2 ) be such that supp 𝜑 𝑗 ⊂ Ω 𝑗 ( 𝑗 = 1, 2) and 𝜑1 + 𝜑2 = 1 everywhere in
Ω1 ∪ Ω2 . We equate 𝜕¯ to d when acting on C ∞ Ω 𝑗 ; 𝑇 (𝑛,𝑛−1) .
Proof of (1). Suppose 𝜇 ∈ O ′ 𝐾 𝑗 , 𝑗 = 1, 2; there are differential forms 𝑓 𝑗 ∈
∞ (𝑛,𝑛−1)

C Ω𝑗;𝑇 such that d 𝑓 𝑗 = 0 and 𝜇 = 𝜇 𝑓 𝑗 ( 𝑗 = 1, 2). By (6.3.10) there is a
𝑢 ∈ C ∞ Ω1 ∩ Ω2 ; 𝑇 (𝑛,𝑛−2) such that 𝑓1 − 𝑓2 = d𝑢 in Ω1 ∩ Ω2 . Define

𝑢 1 = 𝜑2 𝑢 in Ω1 ∩ Ω2 , 𝑢 1 = 0 in Ω1 \ (Ω1 ∩ Ω2 ) ;
𝑢 2 = 𝜑1 𝑢 in Ω1 ∩ Ω2 , 𝑢 2 = 0 in Ω2 \ (Ω1 ∩ Ω2 ) ;

we have 𝑢 1 +𝑢 2 = 𝑢 in Ω1 ∩Ω2 . Next define 𝐹1 = 𝑓1 −d𝑢 1 in Ω1 and 𝐹2 = 𝑓2 +d𝑢 2 in Ω2.

We have 𝐹1 −𝐹2 = 0 in Ω1 ∩Ω2 , meaning that there is an 𝐹 ∈ C ∞ Ω1 ∪ Ω2 ; 𝑇 (𝑛,𝑛−1)
whose restriction to Ω 𝑗 is equal to 𝐹 𝑗 for 𝑗 = 1, 2. Now let the number 𝑅 be sufficiently
large that 𝑧 ∈ 𝐾1 ∪ 𝐾2 =⇒ |𝑧| < 𝑅; note that, then, the sphere |𝑧| = 𝑅 is contained
in Ω1 ∩ Ω2 . If ℎ ∈ O (C𝑛 ) we have
∫ ∫
ℎ (𝑧) 𝑓1 (𝑧, d𝑧, d𝑧¯) = ℎ (𝑧) 𝐹1 (𝑧, d𝑧, d𝑧¯) (6.3.11)
|𝑧 |=𝑅 |𝑧 |=𝑅
∫
= ℎ (𝑧) 𝐹 (𝑧, d𝑧, d𝑧¯) ,
|𝑧 |=𝑅

proving that 𝜇 = 𝜇 𝐹 and therefore 𝜇 ∈ O ′ (𝐾1 ∩ 𝐾2 ).

Proof of (2). Now suppose there is an 𝐹 ∈ C ∞ Ω1 ∩ Ω2 ; 𝑇 (𝑛,𝑛−1) such that

d𝐹 = 0 and 𝜇 = 𝜇 𝐹 . We define 𝐹 𝑗 ∈ C ∞ Ω 𝑗 ; 𝑇 (𝑛,𝑛−1) as follows

𝐹1 = 𝜑2 𝐹 in Ω1 ∩ Ω2 , 𝐹1 = 0 in Ω1 \ (Ω1 ∩ Ω2 ) ;
𝐹2 = 𝜑1 𝐹 in Ω1 ∩ Ω2 , 𝐹2 = 0 in Ω2 \ (Ω1 ∩ Ω2 ) ;

we have d𝐹1 = −d𝐹2 in Ω1 ∩ Ω2 . We call 𝛼 the (𝑛, 𝑛)-form in Ω1 ∪ Ω2 equal to

d𝐹1 in Ω1 and to −d𝐹2 in Ω2 . We refer the reader to a result stated and proved later,
in Ch. 9, Theorem 9.4.15: there is a 𝑣 ∈ C ∞ Ω1 ∪ Ω2 ; 𝑇 (𝑛,𝑛−1) such that d𝑣 = 𝛼.
Define 𝑓1 = 𝐹1 − 𝑣 ∈ C ∞ Ω1 ; 𝑇 (𝑛,𝑛−1) , 𝑓2 = 𝐹2 + 𝑣 ∈ C ∞ Ω2 ; 𝑇 (𝑛,𝑛−1) . We have
d 𝑓 𝑗 = 0 in Ω 𝑗 , 𝑗 = 1, 2. With 𝑅 > 0 and ℎ ∈ O (C𝑛 ) we have
∫
𝜇 𝑓1 , ℎ + 𝜇 𝑓2 , ℎ = ℎ (𝑧) ( 𝑓1 + 𝑓2 ) (𝑧, d𝑧, d𝑧¯)
|𝑧 |=𝑅
∫
= ℎ (𝑧) (𝐹1 + 𝐹2 ) (𝑧, d𝑧, d𝑧¯)
|𝑧 |=𝑅
∫
= ℎ (𝑧) 𝐹 (𝑧, d𝑧, d𝑧¯) = ⟨𝜇, ℎ⟩ . □
|𝑧 |=𝑅
162 6 Analytic Functionals in Euclidean Space

It is convenient to denote by O ′ (R𝑛 ) the subspace of O ′ (C𝑛 ) consisting of the

analytic functionals carried by a compact subset of R𝑛 , 𝑛 = 1, 2, ....

Corollary 6.3.12 The intersection of all the compact subsets of R𝑛 that carry 𝜇 ∈
O ′ (R𝑛 ) is a compact subset of R𝑛 that carries 𝜇 (called the support of 𝜇 and denoted
by supp 𝜇 in the sequel).

Proof Denote by 𝔎 𝜇 the family of all compact subsets of R𝑛 that carry 𝜇 and let 𝔉 be
a totally ordered (with respectÙ
to inclusion) subfamily of 𝔎 𝜇 ; an arbitrary open subset
𝑈 of R𝑛 containing 𝐾 (𝔉) = 𝐾 contains some 𝐾 ∈ 𝔉 and therefore 𝜇 ∈ O ′ (𝑈);
𝐾 ∈𝔉
this proves that 𝜇 is carried by 𝐾 (𝔉). It follows from Zorn’s lemma that 𝔎 𝜇 admits
at least one minimal element 𝐾 ∗ ; if there were 𝐾 ∈ 𝔎 𝜇 , 𝐾 ∗ ⊄ 𝐾, then Theorem 6.3.6
would imply that 𝐾 ∗ ∩ 𝐾 ⫋ 𝐾 ∗ belongs to 𝔎 𝜇 , contradicting the minimality of 𝐾 ∗ .□
Restriction to R𝑛 of entire functions in C𝑛 defines a continuous, linear injective
map O (C𝑛 ) −→ C ∞ (R𝑛 ) whose image is dense, by the Weierstrass Approximation
Theorem. Its transpose is therefore an injective map of the dual of C ∞ (R𝑛 ), the
space of compactly supported distributions in R𝑛 , E ′ (R𝑛 ), into O ′ (R𝑛 ).

Proposition 6.3.13 The support of a compactly supported distribution 𝑢 in R𝑛 is

equal to the support of the analytic functional 𝜇𝑢 ∈ O ′ (R𝑛 ) defined by 𝑢.

Proof Let 𝜀 > 0 be arbitrary and 𝜒 ∈ Cc∞ (R𝑛 ), 𝜒 (𝑥) = 1 if dist (𝑥, supp 𝑢) < 𝜀
and 𝜒 (𝑥) = 0 if dist (𝑥, supp 𝑢) > 2𝜀. For all ℎ ∈ O (C𝑛 ) we have ⟨𝜇𝑢 , ℎ⟩ = ⟨𝑢, 𝜒ℎ⟩
where the brackets express the two different dualities. The topology of C ∞ (R𝑛 ) is
such that there is a constant 𝐶 > 0 and 𝑚 ∈ Z+ (independent of ℎ) such that
∑︁
|⟨𝑢, 𝜒ℎ⟩| ≤ 𝐶 max |𝜕 𝛼 ℎ| .
supp 𝜒
| 𝛼 | ≤𝑚

Cauchy’s inequalities then imply, if 𝑥 ∈ supp 𝜒,

|𝜕 𝛼 ℎ (𝑥)| ≤ 𝐶 𝜀′ max {|ℎ (𝑧)| ; 𝑧 ∈ C𝑛 , dist (𝑧, supp 𝜒) ≤ 𝜀}

≤ 𝐶 𝜀′ max {|ℎ (𝑧)| ; 𝑧 ∈ C𝑛 , dist (𝑧, supp 𝑢) ≤ 3𝜀}

(where supp 𝜒 ⊂ R𝑛 and supp 𝑢 ⊂ R𝑛 are viewed as subsets of C𝑛 ). This proves that
supp 𝑢 carries 𝜇𝑢 and therefore supp 𝜇𝑢 ⊂ supp 𝑢. Let now 𝐾 be a compact subset of
R𝑛 that carries 𝜇𝑢 and 𝜑 ∈ Cc∞ (R𝑛 ) such that 𝐾 ∩ supp 𝜑 = ∅. Consider the entire
functions

𝜈 𝑛2 ∫ 𝑛
© ∑︁ 2ª
𝜑 𝜈 (𝑧) = 𝜑 (𝜉) exp ­−𝜈 𝑧 𝑗 − 𝜉 𝑗 ® d𝜉, 𝜈 = 1, 2, .... (6.3.12)
𝜋 R𝑛
« 𝑗=1 ¬
6.3 Analytic Functionals in R𝑛 as Cohomology Classes 163

Whatever 𝛼 ∈ Z+𝑛 , the restriction to R𝑛 of 𝜕𝑧𝛼 𝜑 𝜈 converges uniformly to 𝜕𝑥𝛼 𝜑 (whence

the Weierstrass Approximation Theorem) and thus ⟨𝑢, 𝜑 𝜈 ⟩ −→ ⟨𝑢, 𝜑⟩ as 𝜈 → +∞.
It is readily seen that the 𝜑 𝜈 converge uniformly to zero in a complex neighborhood
of 𝐾, implying lim ⟨𝜇𝑢 , 𝜑 𝜈 ⟩ = 0 and therefore ⟨𝑢, 𝜑⟩ = 0. This proves that
𝜈→+∞
supp 𝑢 ⊂ 𝐾; in particular, supp 𝑢 ⊂ supp 𝜇𝑢 (by Corollary 6.3.12). □
Chapter 7
Hyperfunctions in Euclidean Space

Hyperfunctions are defined in real space R𝑛 , where analytic functionals are local-
izable, i.e., have a uniquely defined support (Corollary 6.3.12). The first equation
in the present chapter defines the space of hyperfunctions in a bounded open set
𝑈 ⊂ R𝑛 , B (𝑈); this simple definition was proposed by A. Martineau in the 1960s.
The right-hand side in Equation (7.1.1) is the quotient of two Fréchet spaces with
the denominator everywhere dense in the numerator (Proposition 6.1.7), a fact that
precludes the possibility of defining a reasonable Hausdorff topological vector space
structure on B (𝑈), thereby limiting the relevance of Functional Analysis (without
totally eliminating it) and strongly shifting the development of Hyperfunction The-
ory towards (Homological) Algebra and Sheaf Theory (see Subsection 1.2.2 and Ch.
10).
Hyperfunctions in 𝑈 have restrictions to an arbitrary open set 𝑉 ⊂ 𝑈; this
property leads directly to the definition of the presheaf of hyperfunctions in an
arbitrary domain Ω in R𝑛 . In Subsection 7.1.1 it is shown that this presheaf is a
sheaf in the sense that one can patch up, unambiguously, local continuous sections
that agree on overlaps, a consequence of the properties of analytic functionals in real
space proved in Ch. 6.
In Subsection 12.1.2, after the formal definition of the sheaf B (Ω) of germs of
hyperfunctions in Ω we prove that the restriction mapping from 𝑈 ⊂ Ω to 𝑉 ⊂ 𝑈
(𝑈, 𝑉: open sets) of continuous sections of B (𝑈) is surjective. If by a hyperfunction
in Ω we mean a continuous section of the sheaf B (Ω) (as Proposition 7.1.7 allows
us to do) the surjectivity of the restriction mappings means that every hyperfunction
in 𝑈 extends as a hyperfunction in Ω and, for that matter, as a hyperfunction in
R𝑛 : the sheaf B (R𝑛 ) is flabby. This stands in total contrast to the state of affairs
with distributions: there is a great number of distributions in Ω ≠ R𝑛 that are not
extendible to R𝑛 . A simple example is the locally integrable function exp (1/|𝑥|) in
R𝑛 \ {0}. Nonextendible distributions will be looked at closely in Section 14.2.
A noteworthy property is the bijective correspondence between compactly sup-
ported hyperfunctions in R𝑛 and analytic functionals in R𝑛 (Theorem 7.1.14).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 165
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_7
166 7 Hyperfunctions in Euclidean Space

Linear differential operators with C 𝜔 coefficients in Ω act naturally on B (Ω)

since they do so on analytic functionals. But we cannot say that they act continuously
on B (Ω), given that there is no reasonable topological vector space structure on
B (Ω). Given a bounded domain 𝑈 ⊂ R𝑛 , B (𝑈) is the quotient of two Fréchet
spaces; but the denominator is dense in the numerator, precluding that the quotient
topology be Hausdorff.
In Subsection 7.1.3 a relatively simple argument shows that every hyperfunction
in a domain Ω ⊂ R𝑛 can be divided by an arbitrary C 𝜔 function Ω that does not
vanish identically, the division yielding another hyperfunction. The proof of the
analogous statement for distributions is quite difficult; it will be fully presented in
Section 15.2. The penultimate subsection of Section 7.1 is devoted to establishing
an elegant sufficient condition for the solvability in hyperfunctions of a linear PDE
with analytic coefficients. I am indebted to P.D. Cordaro for calling my attention to
these two properties of hyperfunctions.
In the last subsection of Section 7.1 we introduce singularity hyperfunctions, the
result of modding off analytic functions from hyperfunctions, the concept at the core
of the investigation of the analytic singularities of solutions of PDEs. In this book
the distinction between hyperfunctions and singularity hyperfunctions is formalized,
to dispel some confusion in the terminology introduced in other texts. For instance,
what is defined in Formula 1.1, p. 1 of [Komatsu, 1992], as Sato’s hyperfunctions,
is precisely what we call singularity hyperfunctions (on the real line). This is per-
fectly acceptable and can be done throughout, in all dimensions; but then one cannot
say that hyperfunctions generalize distributions. What (singularity) hyperfunctions
generalize are equivalence classes of distributions modulo analytic functions. Con-
sidering the elegance of Martineau’s approach to hyperfunctions through analytic
functionals there is no need to complicate it from the start by a double modding off.
As we have seen in Section 3.2, if a holomorphic function ℎ in a wedge 𝑈 + 𝑖Γ in
C𝑛 (𝑈, Γ ⊂ R𝑛 both open, Γ a cone with vertex 0) has tempered growth at the edge
𝑈 (meaning slower growth than some power of |𝑦| −1 , 𝑦 ∈ Γ) its boundary value can
be defined as a distribution in 𝑈. The beautiful thing about hyperfunctions is that
one can remove all restrictions on the growth of |ℎ (𝑥 + 𝑖𝑦)| as 𝑦 ∈ Γ tends to 0 and
yet the boundary value of ℎ can be defined as a hyperfunction in 𝑈, 𝑏𝑈 ℎ. This is
explained in Section 7.2 using the boundary limit of appropriate analytic functionals.
This same approach [combined with the isomorphism (6.3.10)] will allow us (in Ch.
10) to interpret 𝑏𝑈 ℎ as a relative cohomology class. In Section 7.3 the FBI transform
of analytic functionals is exploited to locally represent every hyperfunction as a
finite sum of boundary values of holomorphic functions in wedges (all with the same
edge).
As the understanding of hyperfunctions in this book expands and deepens it is
difficult to avoid the impression that, in nonglobal analysis, much can be achieved by
using solely analytic functionals (and thereby, as a side benefit, Functional Analysis).
We can derive the same impression from other texts, such as [Hörmander, 1994] –
quite a remarkable development, considering the somewhat modest beginnings of
the theory of analytic functionals in the first half of the XXth century (they were
introduced by L. Fantappiè c. 1920/1930; see [Fantappiè, 1943]). This impression is
7.1 The Sheaf of Hyperfunctions in Euclidean Space 167

reinforced when dealing with the FBI transform in Section 7.3, where its inversion
is also described. Analytic functionals are the natural objects to transform, extend-
ing the FBI transform of compactly supported distributions. The FBI transform of
analytic functionals is used to define the essential support of a hyperfunction 𝑓 in
R𝑛 – to which we refer as the analytic wave-front set of 𝑓 (to preserve the link with
the C ∞ wave-front set of distributions) and which we denote by 𝑊 𝐹a 𝑓 ; 𝑊 𝐹a 𝑓 is a
conic subset of phase-space R𝑛 × (R𝑛 \ {0}). Analytic differential operators decrease
the analytic wave-front set of hyperfunctions.
As already mentioned, the FBI transform of analytic functionals also enables us
to represent every hyperfunction as a finite sum of boundary values of holomorphic
functions in wedges. Such representations lead to a different definition of the analytic
wave-front set of a hyperfunction 𝑓 and of hyperfunctions themselves. The problem
is that these representations are not unique and this complicates matters, especially
in dimensions higher than one. The celebrated Edge of the Wedge Theorem 7.5.1
allows us to mod off the finite families of holomorphic functions in wedges whose
boundary values on the common edge add up to zero or to an analytic function. The
latter point makes clear that they are really representations of the singularity hyper-
function defined by 𝑓 . The latter is a strong argument for dealing with singularity
hyperfunctions (and, perhaps, calling them hyperfunctions).
The chapter closes with the definition of the sheaf of microfunctions in Euclidean
space, preparatory to the extension of microlocal analysis to hyperfunctions, explored
and used in later parts of the book.

7.1 The Sheaf of Hyperfunctions in Euclidean Space

7.1.1 Definition based on analytic functionals

Let 𝑈 ⊂ R𝑛 be a bounded open set, 𝑈 its closure and 𝜕𝑈 = 𝑈 ∩ (R𝑛 \𝑈) its boundary.
We introduce the quotient vector space

B (𝑈) = O ′ 𝑈 /O ′ (𝜕𝑈) . (7.1.1)

Definition 7.1.1 The elementsofB (𝑈) will be called the hyperfunctions in 𝑈. If

an analytic functional 𝜇 ∈ O ′ 𝑈 belongs to the coset 𝑓 ∈ B (𝑈) we will say that
𝜇 represents the hyperfunction 𝑓 in 𝑈.
Example 7.1.2 Let the series +∞
Í
𝑘=0 𝑐 𝑘 𝑧 (𝑧 ∈ C) have the following property: to
𝑘

every 𝜀 > 0 there is aÍ𝐶 𝜀 > 0 such that |𝑐 𝑘 | ≤ 𝐶 𝜀 𝜀 𝑘 /𝑘! for every 𝑘 = 0, 1... (cf.
Section 8.1). Let 𝜇 = ∞ 𝑘=0 𝑐 𝑘 𝛿
(𝑘) be the associated analytic functional [𝛿 (𝑘) = 𝑘 th

derivative of the Dirac measure]. If ℎ ∈ O (C𝑛 ) the Cauchy inequalities imply

+∞
∑︁ 1 𝑘 (𝑘)
|⟨𝜇, ℎ⟩| ≤ 𝐶 𝜀 𝜀 ℎ (0) ≤ 2𝐶 𝜀 max |ℎ (𝑧)| .
𝑘! |𝑧 |<2𝜀
𝑘=0
168 7 Hyperfunctions in Euclidean Space

Since 𝜀 > 0 is arbitrary we see that 𝜇 is carried by the origin and therefore defines
a hyperfunction 𝑓 on the real line; 𝑓 is not a distribution unless 𝑐 𝑘 = 0 for all but
finitely many 𝑘.

The topologies of O ′ 𝑈 and O ′ (𝜕𝑈) do not allow us to make a Hausdorff

topological vector space of B (𝑈) because O ′ (𝜕𝑈) is dense in O ′ 𝑈 (Proposition
6.1.7).
If 𝐾 ⊂ R𝑛 is compact, an arbitrary (linear) differential operator 𝑃 (𝑥, D 𝑥 ) with C 𝜔
coefficients in an open subset 𝑈 of R𝑛 , 𝐾 ⊂ 𝑈, acts on O ′ (𝐾) through a holomorphic
extension (6.1.4) to an open subset 𝑈 C of C𝑛 such that 𝑈 ⊂ 𝑈 C ∩R𝑛 . It is an immediate
consequence of the Cauchy inequalities that 𝑃 (𝑥, D 𝑥 ) O ′ (𝐾) ⊂ O ′ (𝐾). In turn this
implies that 𝑃 (𝑥, D 𝑥 ) defines a linear operator B (𝑈) ←↪ (𝑈 bounded).

Proposition 7.1.3 Let 𝑉 ⊂ 𝑈 be two bounded open subsets of R𝑛 . There is a natural

restriction map 𝜌𝑈 𝑉 : B (𝑈) −→ B (𝑉). The restriction map is surjective.

Proof Let 𝜇 ∈ O ′ 𝑈 be arbitrary. Set 𝐾1 = 𝑉 and 𝐾2 = 𝑈\𝑉; we have 𝑈 = 𝐾1 ∪ 𝐾2
and 𝜕𝑉 = 𝐾1 ∩ 𝐾 2 . By Theorems 6.3.6, 6.3.11, we can decompose 𝜇 = 𝜇1 + 𝜇2
with 𝜇 𝑗 ∈ O′ 𝐾 𝑗 , 𝑗 = 1, 2. Consider a different decomposition 𝜇 = 𝜇1′ + 𝜇2′ with
𝜇 ′𝑗 ∈ O ′ 𝐾 𝑗 , 𝑗 = 1, 2; we have

𝜇1 − 𝜇1′ = 𝜇2′ − 𝜇2 ∈ O ′ (𝐾1 ∩ 𝐾2 ) = O ′ (𝜕𝑉) .

It follows that the coset [𝜇1 ] of 𝜇1 ∈ O ′ 𝑉 mod O ′ (𝜕𝑉) is unambiguously defined.
′
to 𝑉 of the coset [𝜇] ∈ B (𝑈) /O (𝜕𝑈).
We call it the restriction
′
Now let 𝜇1 ∈ O 𝑉 be arbitrary. Thanks to the fact that all compact subsets of

R𝑛 are Runge (Theorem 6.2.14) we have a natural injection O ′ 𝑉 −→ O ′ 𝑈 . We

can therefore view 𝜇1 as an element of O ′ 𝑈 ; the restriction to 𝑉 of 𝜇1 ’s coset
mod O ′ (𝜕𝑈) is evidently equal to the coset of 𝜇1 mod O ′ (𝜕𝑉). □

Proposition 7.1.4 Let 𝑉 ⊂ 𝑈 be two bounded open subsets of R𝑛 and 𝑓 ∈ B (𝑈).

′
𝑉 𝑓 = 0 then supp 𝜇 ⊂ 𝑈\𝑉 whatever the representative 𝜇 ∈ O 𝑈 of 𝑓 .
If 𝜌𝑈

Conversely, if supp 𝜇 ⊂ 𝑈\𝑉 for some representative 𝜇 ∈ O ′ 𝑈 of 𝑓 then 𝜌𝑈
𝑉 𝑓 = 0.

Proof I. Suppose 𝜌𝑈 𝑓 = 0. Let 𝜇 ∈ O ′ 𝑈 be an arbitrary representative of 𝑓 .
𝑉

Applying Theorems 6.3.6, 6.3.11, we decompose 𝜇 = 𝜇˜ + 𝜈 with 𝜇˜ ∈ O ′ 𝑈\𝑉

and 𝜈 ∈ O ′ 𝑉 . The coset [𝜈] mod O ′ (𝜕𝑉) is the restriction to 𝑉 of the coset

[𝜇] ∈ O ′ 𝑈 /O ′ (𝜕𝑈). Since the said restriction is equal to zero we must have

𝜈 ∈ O ′ (𝜕𝑉); since 𝜕𝑉 ⊂ 𝑈\𝑉 we get 𝜈 ∈ O ′ 𝑈\𝑉 whence 𝜇 ∈ O ′ 𝑈\𝑉 .
7.1 The Sheaf of Hyperfunctions in Euclidean Space 169

II. Suppose 𝜇 ∈ O ′ 𝑈\𝑉 represents 𝑓 in 𝑈. Consider an arbitrary decomposition

𝜇 = 𝜇˜ + 𝜈 with 𝜇˜ ∈ O ′ 𝑈\𝑉 and 𝜈 ∈ O ′ 𝑉 . We have 𝜈 = 𝜇 − 𝜇˜ ∈ O ′ 𝑈\𝑉 ; since

𝑈\𝑉 ∩ 𝑉 ⊂ 𝜕𝑉 we conclude that the coset [𝜈] of 𝜈 mod O ′ (𝜕𝑉) vanishes. Since
𝑉 𝑓 (see the proof of Proposition 7.1.3) we must have 𝜌𝑉 𝑓 = 0.□
this coset defines 𝜌𝑈 𝑈

Proposition 7.1.5 Let {𝑈 𝜄 } 𝜄 ∈𝐼 be an open covering of a bounded open subset 𝑈 of

R𝑛 (𝑈 𝜄 ⊂ 𝑈 for every 𝜄 ∈ 𝐼). If 𝑓 ∈ B (𝑈) is such that 𝜌𝑈
𝑈
𝜄
𝑓 = 0 for every 𝜄 ∈ 𝐼 then
𝑓 = 0.

Proof Suppose we had 𝑓 ≠ 0; there would be a 𝜇 ∈ O ′ 𝑈 representing 𝑓 with
support 𝐾 such that 𝐾 ∩ 𝑈 ≠ ∅. Let 𝜄 ∈ 𝐼 be such that 𝐾 ∩ 𝑈 𝜄 ≠ ∅. Since 𝜌𝑈𝑈
𝜄
𝑓 =0

Proposition 7.1.4 demands that there be a representative 𝜇 ′ ∈ O ′ 𝑈 of 𝑓 such that
supp 𝜇 ′ = 𝐾 ′ ⊂ 𝑈\𝑈 𝜄 implying 𝐾 ∩ 𝑈 ≠ 𝐾 ′ ∩ 𝑈; yet 𝜇 − 𝜇 ′ ∈ O ′ (𝜕𝑈), whence a
contradiction. □

Proposition 7.1.6 Let the open covering {𝑈 𝜄 } 𝜄 ∈𝐼 of the bounded open set 𝑈 be as
in Proposition 7.1.5. Let 𝐼 ∋ 𝜄 ↦→ 𝑓 𝜄 ∈ B (𝑈 𝜄 ) be an indexing that satisfies the
following condition:
𝑈𝜄 𝑈𝜅
(•) For every pair of indices 𝜄, 𝜅 ∈ 𝐼 such that 𝑈 𝜄 ∩ 𝑈 𝜅 ≠ ∅, 𝜌𝑈 𝜄 ∩𝑈𝜅
𝑓 = 𝜌𝑈 𝜄 ∩𝑈𝜅
𝑓.

Under this hypothesis there is an 𝑓 ∈ B (𝑈) such that 𝜌𝑈

𝑈
𝜄
𝑓 = 𝑓 𝜄 for every 𝜄 ∈ 𝐼.

Proof Case 𝐼 = {1, 2}. We are dealing with a pair of bounded open subsets of R𝑛 ,
𝑈1 , 𝑈2 , such that 𝑈 = 𝑈1∪𝑈2 . We assume that 𝑉 = 𝑈1 ∩𝑈2 ≠ ∅ otherwise the claim
is trivial. Let 𝑓 𝑗 ∈ B 𝑈 𝑗 , 𝑗 = 1, 2, be such that 𝜌𝑈 1 𝑈2
𝑉 𝑓1 = 𝜌𝑉 𝑓2 . Since the restriction
maps are surjective (Proposition 7.1.3) there exist elements 𝐹 𝑗 ∈ B (𝑈) such that
𝜌𝑈𝑈
𝑗
𝐹 𝑗 = 𝑓 𝑗 , 𝑗 = 1, 2, whence 𝜌𝑈
𝑉 𝐹1 = 𝜌𝑉 𝐹2 . Proposition 7.1.4 implies that there is
𝑈

a representative 𝜆 of 𝐹1 − 𝐹2 in 𝑈 supported by 𝑈\𝑉 = 𝑈 1 \𝑉 ∪ 𝑈 2 \𝑉 . We use a

decomposition 𝜆 = 𝜆1 + 𝜆2 with 𝜆 𝑗 ∈ O ′ 𝑈 𝑗 \𝑉 . Given an arbitrary representative

𝜇˜ 𝑗 ∈ O ′ 𝑈 of 𝐹 𝑗 we define 𝜇1 = 𝜇˜ 1 − 𝜆1 , 𝜇2 = 𝜇˜ 2 + 𝜆2 . Since

𝜇1 − 𝜇2 = 𝜇˜ 1 − 𝜇˜ 2 − 𝜆

we have 𝜇1 𝜇2 mod O ′ (𝜕𝑈); their coset 𝑓 ∈ B (𝑈) is such that 𝜌𝑈 𝑈

𝑗
𝑓 = 𝑓 𝑗 for
𝑗 = 1, 2.
Case 𝐼 = {1, ..., 𝑁 }, 𝑁 ≥ 3.
Induction on 𝑁 reduces this case to that of a pair of open sets.
General case. We can find a locally finite sequence of open sets 𝑈 𝑘′ (𝑘 = 1, 2, ...)
whose union is equal to 𝑈 and such that to every 𝑘 = 1, 2, ..., there is an 𝜄 ∈ 𝐼 such
that 𝑈 𝑘′ ⊂ 𝑈 𝜄 for some 𝜄 ∈ 𝐼, We set 𝑓 𝑘′ = 𝜌𝑈
𝑈𝜄
′ 𝑓 𝜄 ; it follows directly from Hypothesis
𝑘
(•) that this definition does not depend on 𝜄. From the result about finite sets of
170 7 Hyperfunctions in Euclidean Space

indices 𝐼 we know that to every 𝑁 ∈ Z+ there is an 𝐹𝑁 ∈ B 𝑈1′ ∪ · · · ∪ 𝑈 𝑁 ′ whose

restriction to 𝑈 𝑘′ is equal to 𝑓 𝑘′ whatever 𝑘 = 1, ..., 𝑁. It follows that the restriction of
𝐹𝑁 +1 to 𝑈1′ ∪ · · · ∪𝑈 𝑁
′ is equal to 𝐹 . Letting 𝑁 ↗ +∞ yields the sought 𝑓 ∈ B (𝑈).
𝑁
Its uniqueness follows directly from Proposition 7.1.5. □

7.1.2 The sheaf of hyperfunctions in R𝒏

At this stage it makes sense to switch to Sheaf Theory terminology (the reader is
referred to Subsection 1.2.2). We form the presheaf B (𝑈) , 𝜌𝑈 𝑉 : 𝑈 ranges over
the family of all bounded open subsets of R𝑛 , 𝑉 ranges over the family of all open
subsets of 𝑈, 𝜌𝑈𝑉 : B (𝑈) −→ B (𝑉) is
the restriction map introduced in the previous
subsection. The presheaf B (𝑈) , 𝜌𝑈𝑉 defines a sheaf B (R𝑛 ) to which we shall refer
as the sheaf of hyperfunctions or, more accurately, of germs of hyperfunctions in
R𝑛 . Given any bounded open set 𝑈 ⊂ R𝑛 we can regard B (𝑈) defined in (7.1.1) as
a vector subspace of Γ (𝑈, B (R𝑛 )), the vector space of continuous sections of the
sheaf B (R𝑛 ) over 𝑈.

Proposition 7.1.7 If the open subset 𝑈 of R𝑛 is bounded then Γ (𝑈, B (R𝑛 )) = B (𝑈).

Proof Let 𝒔 ∈ Γ (𝑈, B (R𝑛 )) be arbitrary. There is a covering {𝑉 𝜄 } 𝜄 ∈𝐼 of 𝑈 such that,

for each 𝜄 ∈ 𝐼, the restriction of 𝑠 to 𝑉 𝜄 is equal to an element 𝑠 𝜄 ∈ B (𝑉 𝜄 ). Let 𝜅 ∈ 𝐼
be such that 𝑉 𝜄 ∩ 𝑉𝜅 ≠ ∅. By the transitivity of restrictions 𝑠 𝜄 = 𝑠 𝜅 in 𝑉 𝜄 ∩ 𝑉𝜅 . Then,
by Proposition 7.1.6, we conclude that there is an 𝑠˜ ∈ B (𝑈) such that 𝜌𝑈 𝑈 ( )
𝜄
𝑠˜ = 𝑠 𝜄 ;
necessarily 𝑠 = 𝑠˜. □
Proposition 7.1.7 shows that the following is compatible with Definition 7.1.1.

Definition 7.1.8 By a hyperfunction in an open subset Ω of R𝑛 we shall mean a

continuous section of the sheaf B (R𝑛 ) over Ω. The vector space of hyperfunctions
in Ω will be denoted by B (Ω).

If 𝑓 ∈ B (Ω) we define its support, supp 𝑓 , to be the set of points 𝑥 ∈ Ω where

𝑓 (𝑥) ≠ 0. Recall that the set of points in Ω where a continuous section of a sheaf
(of vector spaces, say) over Ω vanishes is an open subset of Ω; thus supp 𝑓 is closed
in Ω.
An arbitrary differential operator 𝑃 (𝑥, D 𝑥 ) with C 𝜔 coefficients in Ω defines a
linear operator B (Ω) ←↪; we have supp 𝑃 (𝑥, D 𝑥 ) 𝑓 ⊂ supp 𝑓 whatever 𝑓 ∈ B (Ω).

Proposition 7.1.9 Let 𝑉 ⊂ 𝑈 be a pair of open subsets of R𝑛 (not necessarily

bounded). Every hyperfunction in 𝑉 is equal to the restriction of a hyperfunction in
𝑈.

Proof When 𝑈 is bounded the result is stated in Proposition 7.1.6. When 𝑈 is

unbounded the proof in the general case of Proposition 7.1.6 applies if we specify
that the sets 𝑈 𝑘′ are bounded. □
7.1 The Sheaf of Hyperfunctions in Euclidean Space 171

Corollary 7.1.10 Every hyperfunction in an open subset 𝑉 of R𝑛 is equal to the

restriction of a hyperfunction in R𝑛 .
In sheaves parlance Proposition 7.1.9 states that the sheaf B (R𝑛 ) is flabby.
Let D′ (R𝑛 ) be the sheaf of germs of distributions in R𝑛 . It can be defined
′ , 𝜌𝑉 where D ′ (𝑈) is the space of distributions in the

through the presheaf D (𝑈) 𝑈

(arbitrary) open set 𝑈 and 𝜌𝑉 is the standard restriction to 𝑉 ⊂ 𝑈 of distributions in

𝑈

𝑈. The sheaf D′ (R𝑛 ) is definitely not flabby.

Example 7.1.11 The function exp |𝑥| −1 in R𝑛 \ {0} cannot be extended as a distri-
bution in the whole of R𝑛 . More generally, see Section 15.2.
Proposition 7.1.12 There is a natural sheaf homomorphism D′ (R𝑛 ) −→ B (R𝑛 ).
This map is injective.
Proof It is convenient to reason in terms of the stalks: let 𝒇 be a germ of distribution
at an arbitrary point 𝑥 ◦ ∈ R𝑛 . There is a bounded open set 𝑈 ∋ 𝑥 ◦ in which 𝒇 has a
representative 𝑓 ∈ D ′ (𝑈). Select a function 𝜒 ∈ Cc∞ (𝑈), 𝜒 (𝑥) = 1 for all 𝑥 such
that |𝑥 − 𝑥 ◦ | < 𝜀, with 𝜀 > 0 suitably small. The germs at𝑥 ◦ defined by 𝑓 and 𝜒 𝑓
are the same. We can identify 𝜒 𝑓 with an element of O ′ 𝑈 (Proposition 6.3.13)
which in turn defines a hyperfunction in 𝑈 and a germ of hyperfunction 𝒇˜ at 𝑥 ◦ .
It is a simple exercise to check that the germ 𝒇˜ does not depend on the choice of
the representative 𝑓 or on that of the cutoff 𝜒. If we had 𝒇˜ = 0 it would mean that
𝑥 ◦ ∉ supp ( 𝜒 𝑓 ) (again by Proposition 6.3.13) requiring that 𝑓 vanish identically in
a full neighborhood of 𝑥 ◦ and therefore, that 𝒇 = 0. □
Corollary 7.1.13 Let Ω be an open subset of R𝑛 . There is a natural linear map
D ′ (Ω) −→ B (Ω). This map is injective.
In the sequel we may identify D ′ (Ω) with a linear subspace of B (Ω). Keep in
mind that we have not defined a topology on B (Ω).
Theorem 7.1.14 To every hyperfunction 𝑓 in R𝑛 whose support is compact there
is a unique analytic functional 𝜇 𝑓 in R𝑛 representing 𝑓 [in any bounded open set
containing supp 𝑓 ] such that supp 𝜇 𝑓 = supp 𝑓 .
Proof Let us write 𝐾 = supp 𝑓 ⊂⊂ R𝑛 . Select a bounded open subset 𝑈 of R𝑛 such
that 𝐾 ⊂ 𝑈. According to Proposition 7.1.4 there is a representative 𝜆 ∈ O ′ 𝑈 of

𝑓 such that supp 𝜆 ⊂ 𝑈\ 𝑈\𝐾 ⊂ 𝐾 ∪ 𝜕𝑈. By Theorems 6.3.6, 6.3.11, there is a
decomposition 𝜆 = 𝜇 + 𝜈 with 𝜇 ∈ O ′ (𝐾) and 𝜈 ∈ O ′ (𝜕𝑈). Since 𝐾 ∩ 𝜕𝑈 = ∅ we
see that 𝜇 alone represents 𝑓 in 𝑈 and it is the unique analytic functional in R𝑛 to
do so. We apply again Proposition 7.1.4: there cannot exist an open set 𝑉 ⊂ 𝐾 such
that 𝑉 ∩ supp 𝜇 = ∅, hence supp 𝜇 = 𝐾. Obviously, 𝜇 does not depend on the choice
of 𝑈 provided 𝐾 ⊂ 𝑈; and 𝑓 is the unique hyperfunction defined by 𝜇 in some open
set containing 𝐾 and equal to zero in R𝑛 \𝐾: we can take 𝜇 𝑓 = 𝜇. □
Thus we can identify O ′ (R𝑛 ) with the space Bc (R𝑛 ) of compactly supported
hyperfunctions in R𝑛 . This remains true if we replace R𝑛 by any one of its open
subsets, thanks to the flabbiness of the sheaf B (R𝑛 ).
172 7 Hyperfunctions in Euclidean Space

7.1.3 Division of hyperfunctions by analytic functions

In this subsection1 we are going to show that arbitrary hyperfunctions can be divided
by arbitrary analytic functions that do not vanish identically. For this we need parti-
tions of unity in the hyperfunctions sense; these are constructed following a general
argument in sheaf theory (see, e.g., [Godement, 1964], p. 156). As before, Ω will be
an open subset of R𝑛 .

Lemma 7.1.15 Let {𝑈 𝑗 } 𝑗 ∈Z+ be a countable, locally finite open covering of Ω such
that 𝑈 𝑗 ⊂⊂ ΩÍfor each 𝑗. Given 𝑢 ∈ B (Ω) there are 𝑢 𝑗 ∈ B (Ω), supp 𝑢 𝑗 ⊂ 𝑈 𝑗 ,
such that 𝑢 = 𝑗 ∈Z+ 𝑢 𝑗 .

Proof Select open sets 𝑉 𝑗 ⊂⊂ 𝑈 𝑗 such that ∪ 𝑗 ∈Z+ 𝑉 𝑗 = Ω and let E denote the
families (𝑢 𝑗 ) 𝑗 ∈𝐽 with 𝐽 ⊂ Z+ and 𝑢 𝑗 ∈ B 𝑈 𝑗 such that supp 𝑢 𝑗 ⊂ 𝑈 𝑗
set of all Í
and 𝑢 = 𝑗 ∈𝐽 𝑢 𝑗 in 𝑉𝐽 = ∪ 𝑗 ∈𝐽 𝑉 𝑗 (we extend 𝑢 𝑗 to 𝑉𝐽 by setting 𝑢 𝑗 = 0 in the
complement of supp 𝑢 𝑗 ). We have E ≠ ∅ since taking 𝐽 = { 𝑗 ◦ }, 𝑗◦ ∈ Z+ , and 𝑢 𝑗◦ = 𝑢
in 𝑉 𝑗◦ yields (𝑢 𝑗 ) 𝑗 ∈𝐽 ∈ E. In E we introduce a partial order: (𝑢 𝑗 ) 𝑗 ∈𝐽 ⪯ (𝑢 𝑗 ) 𝑗 ∈𝐽 ′
if 𝐽 ⊂ 𝐽 ′. It is easily seen that every totally ordered subset of E has an upper
bound and therefore, by Zorn’s lemma, E has a maximal element (𝑢 𝑗 ) 𝑗 ∈𝐽★ . We are
going to show that 𝐽★ = Z+ . Suppose there is a 𝑗◦ ∈Z+ \𝐽★ and consider the open
set Ω′ = 𝑉𝐽★ ∪ 𝑉 𝑗◦ ∪ (Ω \ 𝑉¯ 𝑗◦ ). Since (𝑉𝐽★ ∪ 𝑉 𝑗◦ ) ∩ (Ω \ 𝑉¯ 𝑗◦ ) ⊂ 𝑉𝐽★ we obtain a
well-defined hyperfunction 𝑣 ∈ 𝐵(Ω′) simply by setting
Í
𝑢 − 𝑗 ∈𝐽★ 𝑢 𝑗 in 𝑉𝐽★ ∪ 𝑉 𝑗◦ ,
𝑣=
0 in Ω \ 𝑉¯ 𝑗◦ .

If 𝑢 𝑗◦ ∈ B (Ω) is an extension of 𝑣 then supp 𝑢 𝑗◦ ⊂ 𝑉¯ 𝑗◦ ⊂ 𝑈 𝑗◦ and (𝑢 𝑗 ) 𝑗 ∈𝐽★∪{ 𝑗◦ } ∈ E,

contradicting the maximality of (𝑢 𝑗 ) 𝑗 ∈𝐽★ . □
We recall that if 𝐾 ⊂ C𝑛 is compact then O (𝐾) denotes the ring of germs of
holomorphic functions at 𝐾 (see Section 6.1).

Lemma 7.1.16 Let 𝐾 be a compact subset of C𝑛 and let 𝑓 ∈ O (𝐾). Then the ideal
of O (𝐾), ℑ 𝑓 = {𝑔 𝑓 ; 𝑔 ∈ O (𝐾)}, is sequentially closed in O (𝐾).

Proof Let {𝑔 𝑗 𝑓 } 𝑗 ∈Z+ be a sequence in ℑ 𝑓 converging to a germ ℎ in O (𝐾); since

O (𝐾) is the (nonstrict) inductive limit of the Fréchet spaces O (𝑈), 𝑈 open subsets
of C𝑛 containing 𝐾, there exists an open set 𝑈 ⊃ 𝐾 such that 𝑔 𝑗 𝑓 ∈ O (𝑈) for every
𝑗 ∈ Z+ and 𝑔 𝑗 𝑓 → ℎ in O (𝑈) (see Subsection 7.1.1; here we are identifying germs
and their representatives). Fix 𝑧◦ ∈ 𝐾 and assume that the germ of 𝑓 at 𝑧 ◦ , f𝑧 ◦ ∈ O𝑧 ◦ ,
is not zero. We claim that
(•) the germ of ℎ at 𝑧◦ belongs to the ideal f𝑧 ◦ O𝑧 ◦ .

1 This subsection and the next one are contributions of P.D. Cordaro.
7.1 The Sheaf of Hyperfunctions in Euclidean Space 173

We are going to apply the Weierstrass Division Theorem (Theorem 14.3.5).

After a translation and a linear change of variables we can assume that 𝑧◦ = 0 and
𝑓 (0, . . . , 0, 𝑧 𝑛 )/𝑧 𝑛𝑘 is holomorphic and not zero at the origin for some 𝑘 ∈ Z+ . We
can then write ℎ = 𝑄 𝑓 + 𝑅 in a polydisk Δ𝑟(𝑛) in C𝑛 centered at 0 with all radii equal
to 𝑟 > 0, 𝑄, 𝑅 ∈ O (Δ𝑟(𝑛) ) and 𝑅 a polynomial in 𝑧 𝑛 of degree < 𝑘 whose coefficients
are holomorphic functions of (𝑧1 , . . . , 𝑧 𝑛−1 ). In Δ𝑟(𝑛) we have

𝑔 𝑗 𝑓 − ℎ = (𝑔 𝑗 − 𝑄) 𝑓 − 𝑅 . (7.1.2)

From the uniqueness part in Theorem 14.3.5 and the standard estimate for the quotient
in the division of holomorphic functions, we obtain

sup |𝑔 𝑗 − 𝑄| ≤ 𝐶 sup |𝑔 𝑗 𝑓 − ℎ|,

(𝑛) (𝑛)
Δ𝑟 Δ𝑟

where 𝐶 > 0 depends solely on 𝑓 . Thus 𝑔 𝑗 → 𝑄 uniformly in Δ𝑟(𝑛) which, by (7.1.2),

implies 𝑅 ≡ 0, thereby proving (•).
Property (•) enables us to cover 𝐾 with a finite number of polydisks Δ𝑟(𝑛)
1
, ..., Δ𝑟(𝑛)
𝑁
,
such that
(1) 𝑓 does not vanish identically in Δ𝑟(𝑛)
𝑗
for any 𝑗 = 1 . . . , 𝑚 (0 ≤ 𝑚 ≤ 𝑁);
(2) 𝑓 ≡ 0 in Δ𝑟(𝑛)
𝑗
if 𝑗 = 𝑚 + 1, . . . , 𝑁;
(3) for each 𝑗 = 1, . . . , 𝑚, there is a 𝑄 𝑗 ∈ O (Δ𝑟(𝑛)
𝑗
) such that ℎ = 𝑄 𝑗 𝑓 in Δ𝑟(𝑛)
𝑗
.

Set 𝑈1 = Δ𝑟(𝑛)1
∪ · · · ∪ Δ𝑟(𝑛)
𝑚
(𝑛)
and 𝑈2 = Δ𝑚+1 (𝑛)
∪ . . . ∪ Δ𝑁 . Properties (1) and (2)
require 𝑈1 ∩ 𝑈2 = ∅. Property (3) entails the existence of 𝑔∗ ∈ O (𝑈1 ) such that
ℎ = 𝑓 𝑔∗ in 𝑈1 . We can define 𝑔 ∈ O (𝑈1 ∪ 𝑈2 ) by setting 𝑔 = 𝑔∗ in 𝑈1 and 𝑔 ≡ 0 in
𝑈2 ; then ℎ = 𝑓 𝑔 in a neighborhood of 𝐾 . Thus ℎ ∈ ℑ 𝑓 as claimed. □
We recall that the dual O ′ (𝐾) of O (𝐾) carries the natural topology of a reflexive
Fréchet space: O (𝐾) can be regarded as the dual of the Fréchet space O ′ (𝐾). Let
𝑓 ∈ O (𝐾); the transpose of the linear map

O ′ (𝐾) ∋ 𝜇 ↦→ 𝑓 𝜇 ∈ O ′ (𝐾) (7.1.3)

is the multiplication
O (𝐾) ∋ 𝑔 ↦→ 𝑓 𝑔 ∈ O (𝐾) . (7.1.4)
We are going to deduce the surjectivity of (7.1.3) from Lemma 7.1.16. For this we
need an important theorem of Functional Analysis due to S. Banach; below we state
it as a lemma. We recall that a Fréchet space 𝑬 is a locally convex topological vector
space which is metrizable (i.e., the topology of 𝑬 can be defined by a metric) and
complete (every Cauchy sequence converges). A Fréchet–Montel space is a Fréchet
space 𝑬 that has the following property:
(F-M) Every bounded and closed subset of 𝑬 is compact.
174 7 Hyperfunctions in Euclidean Space

Let 𝑬 ∗ denote the (topological) dual of 𝑬; its elements are the continuous linear
functionals 𝜆 : 𝑬 −→ C. The strong topology on 𝑬 ∗ is the topology of uniform
convergence on every compact subset of 𝑬. Property (F-M) holds for 𝑬 if and only
if it holds for 𝑬 ∗ (in general 𝑬 ∗ is not a Fréchet space). Furthermore 𝑬 is reflexive: it
is canonically isomorphic to the dual of its dual, 𝑬 (𝑬 ∗ ) ∗ ; the isomorphism is a
homeomorphism: if we identify 𝑬 with (𝑬 ∗ ) ∗ the strong dual topology is the same
as the original topology of 𝑬. We must also point out that a set 𝐵 in either 𝑬 or 𝑬 ∗ is
bounded in both the strong topology and the weak dual topology (Mackey’s theorem,
see [Treves, 1967], p. 371). The same applies to the convergence of sequences.
Examples: The spaces C ∞ (Ω) (Ω ⊂ R 𝑁 open) and O ΩC (ΩC ⊂ C 𝑁 open) are

Fréchet–Montel spaces when equipped with their natural topologies. In the present
∗
subsection the important case is 𝑬 = O ′ (𝐾); that it holds for 𝑬 = O (𝐾) is a direct
consequence of the fact that it holds for O Ω (the original Montel Theorem).
C

Lemma 7.1.17 Let 𝑬, 𝑭 be Fréchet–Montel spaces and let 𝜑 : 𝑬 → 𝑭 be a contin-

uous linear map with 𝜑(𝑬) dense in 𝑭. The following properties are equivalent:
(1) 𝜑(𝑬) = 𝑭;
(2) 𝜑 is an open mapping;
(3) the transpose 𝜑⊤ : 𝑭 ∗ → 𝑬 ∗ is an open mapping, in the sense of the strong
topologies on both 𝑬 ∗ and 𝑭 ∗ ;
(4) ∀𝐵 ⊂ 𝑭 ∗ , if 𝜑⊤ (𝐵) is bounded in 𝑬 ∗ then 𝐵 bounded;
(5) whatever the sequence {𝜆 𝑗 } 𝑗=1,2,... ⊂ 𝑭 ∗ , if { 𝜑⊤ (𝜆 𝑗 )} 𝑗=1,2,... is bounded in 𝑬 ∗
then {𝜆 𝑗 } 𝑗=1,2,... is bounded in 𝑭 ∗ ;
(6) 𝜑⊤ (𝑭 ∗ ) is sequentially closed in 𝑬 ∗ .

For a proof, see e.g. [Köthe, 1979], pp. 18–22 (also [Treves, 1967], Chapters 37,
38).

Corollary 7.1.18 If 𝐾 is a connected compact subset of R𝑛 and if 𝑓 ∈ O (𝐾) does

not vanish identically then the map (7.1.3) is surjective.

Proof Since 𝐾 is connected and 𝑓 . 0 the map (7.1.4) is injective, implying that
the range of (7.1.3) is dense (by the Hahn–Banach Theorem). Lemma 7.1.16 implies
that its range, ℑ 𝑓 , is sequentially closed. The corollary is then a consequence of
Lemma 7.1.17. □
The division theorem for hyperfunctions (first proved in [Kantor-Schapira, 1978])
is a direct consequence of the previous results:

Theorem 7.1.19 If Ω is a domain in R𝑛 and if 𝑓 ∈ C 𝜔 (Ω) does not vanish identically

then the map B (Ω) ∋ 𝑢 → 𝑓 𝑢 ∈ B (Ω) is surjective.

Proof Let {𝑈 𝑗 } 𝑗 ∈Z+ be a countable, locally finite open covering of Ω such that
𝑈 𝑗 ⊂⊂ (Ω). By Lemma 7.1.15 we can write
Í Ω is an open ball for each 𝑗. Let 𝑢 ∈ B
𝑢 = 𝑗 ∈Z+ 𝑢 𝑗 , where 𝑢 𝑗 ∈ B (Ω) is supported in 𝑈 𝑗 . By Corollary 7.1.18, for each
7.1 The Sheaf of Hyperfunctions in Euclidean Space 175

𝑗 there is a 𝑣 𝑗 ∈ B (Ω), also

Í supported in 𝑈 𝑗 , such that 𝑓 𝑣 𝑗 = 𝑢 𝑗 . Since {𝑈 𝑗 } is
locally finite the series 𝑣 = 𝑗 ∈Z+ 𝑣 𝑗 defines an element in B (Ω) and
∑︁ ∑︁
𝑓𝑣 = 𝑓𝑣𝑗 = 𝑢 𝑗 = 𝑢. □
𝑗 ∈Z+ 𝑗 ∈Z+

7.1.4 Solvability of analytic linear PDEs in hyperfunctions

Let Ω be a domain in R𝑛 . We have pointed out (right after Definition 7.1.1) that
an arbitrary differential operator 𝑃 = 𝑃 (𝑥, D 𝑥 ) with coefficients in C 𝜔 (Ω) acts on
the hyperfunctions in Ω; it thereby defines a sheaf homomorphism P of B (Ω) [the
restriction to Ω of the sheaf B (R𝑛 ), see Subsection 7.1.2] into itself. We shall denote
by P 𝑥 the linear endomorphism B 𝑥 ←↪ induced by P.

Proposition 7.1.20 Let 𝑥 ◦ ∈ Ω. The following properties are equivalent:

(i) P 𝑥 ◦ B 𝑥 ◦ = B 𝑥 ◦ ;
(ii) there is a neighborhood 𝑈 of 𝑥 ◦ in Ω such that 𝑃B (𝑈) = B (𝑈).

Proof We use repeatedly the fact that the sheaf B (R𝑛 ) is flabby (Proposition 7.1.9).
It entails that every f 𝑥 ◦ ∈ B 𝑥 ◦ is the germ at 𝑥 ◦ of some 𝑓 ∈ B (Ω). If 𝑢 ∈ B (𝑈)
satisfies 𝑃𝑢 = 𝑓 in 𝑈 then P 𝑥 ◦ u 𝑥 ◦ = f 𝑥 ◦ . This proves (ii)=⇒(i).
Let 𝑈 𝑗 𝑗=1,2,... be a basis of neighborhoods of 𝑥 ◦ in Ω such that 𝑈 𝑗+1 ⊂ 𝑈 𝑗 ⊂
𝑈1 ⊂⊂ Ω for every 𝑗. To say that to a given f 𝑥 ◦ ∈ B 𝑥 ◦ there is u 𝑥 ◦ ∈ B 𝑥 ◦ satisfying
that if f 𝑥 ◦ ∈ B 𝑥 ◦ is the germ at
P 𝑥 ◦ u 𝑥 ◦ = f 𝑥 ◦ is the same as saying 𝑥 ◦ of 𝑓 ∈ B (𝑈1 )
◦
there is a 𝑗 ≥ 1 and 𝑢 ∈ B 𝑈 𝑗 such that u 𝑥 ◦ is the germ at 𝑥 of 𝑢 and 𝑃𝑢 = 𝑓
in 𝑈 𝑗 . By (7.1.1) this last equation can be translated interms of analytic functionals
representing the hyperfunctions 𝑓 and 𝑢: if 𝜇 ∈ O ′ 𝑈1 represents 𝑓 in 𝑈1 and

𝜈 ∈ O ′ 𝑈 𝑗 represents 𝑢 in 𝑈 𝑗 then 𝑃𝜇 − 𝜈 ∈ O ′ 𝑈1 \𝑈 𝑗 . Define
n o
𝑬 𝑗 = (𝜇, 𝜈, 𝜌) ∈ O ′ 𝑈1 × O ′ 𝑈 𝑗 × O ′ 𝑈1 \𝑈 𝑗 ; 𝑃𝜇 − 𝜈 = 𝜌 ,

𝑗 = 1, 2, .... Since 𝑬 𝑗 is a closed linear subspace of the Fréchet space O ′ (𝐾) it is itself

a Fréchet space and its image 𝜋𝑬 𝑗 under the projection 𝜋 : (𝜇, 𝜈, 𝜌) ↦→ 𝜇 ∈ O ′ 𝑈1
is closed. Property (i) implies
Ø ∞
O ′ 𝑈1 = 𝜋𝑬 𝑗 ,
𝑗=1

whence O ′ 𝑈1 = 𝜋𝑬 𝑗◦ for some integer 𝑗 ◦ ≥ 1, by the Baire category theorem.
This implies (ii) with 𝑈 = 𝑈 𝑗◦ . □
176 7 Hyperfunctions in Euclidean Space

In the next statement we use the notation 𝑆 𝜀 = {𝑧

∈ C ; dist (𝑧, 𝑆) < 𝜀}, 𝑆 ⊂ C ,
𝑛 𝑛

𝜀 > 0. Notice that 𝑆 𝜀 is open and also that 𝑆 𝜀 = 𝑆 . We need a preliminary remark.
𝜀
Let 𝑈 ⊂⊂ Ω be open and define Ob (𝑈 𝜀 ) = O (𝑈 𝜀 ) ∩ 𝐿 ∞ (𝑈 𝜀 ); Ob (𝑈 𝜀 ) equipped
with the 𝐿 ∞ norm is a Banach space. We observe that the restriction mapping
Ob (𝑈 𝜀 ) −→ O 𝑈 is injective; we can identify Ob (𝑈 𝜀 ) with its image and thus
write
Ø ∞

O 𝑈 = Ob 𝑈1/𝑘 .
𝑘=1

The locally convex topological vector space O 𝑈 is the (nonstrict) inductive limit
of the Banach spaces Ob (𝑈 𝜀 ). This has the following consequences:

(1) For a subset 𝑬 of O 𝑈 to be bounded it is necessary and sufficient that there
𝑬
exists an 𝜀 > 0 such that ⊂ Ob (𝑈 𝜀 ) and 𝑬 is bounded in Ob (𝑈 𝜀 ).
(2) For a linear map 𝑇 : O 𝑈 ×O (𝜕𝑈) −→ O 𝑈 to be bounded (i.e., continuous)
it is necessary and sufficient that to every 𝜀 > 0 there exists a 𝛿 > 0 such that

𝑇 (Ob (𝑈 𝜀 ) × Ob ((𝜕𝑈) 𝜀 )) ⊂ Ob (𝑈 𝛿 )

and the restriction of 𝑇 to Ob (𝑈 𝜀 ) × Ob ((𝜕𝑈) 𝜀 ) is a bounded map into Ob (𝑈 𝛿 ).

The latter means that there is a 𝐶 > 0 such that

∥𝑇 ( 𝑓1 , 𝑓2 )∥ 𝐿 ∞ (𝑈 𝛿 ) ≤ 𝐶 ∥ 𝑓1 ∥ 𝐿 ∞ (𝑈 𝜀 ) + ∥ 𝑓2 ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 )

for all ( 𝑓1 , 𝑓2 ) ∈ Ob (𝑈 𝜀 ) × Ob ((𝜕𝑈) 𝜀 ).

Lemma 7.1.21 Let 𝑈 ⊂⊂ Ω be an open set. The following conditions are equivalent:
(a) 𝑃B (𝑈) = B (𝑈);

(b) an arbitrary sequence ℎ 𝑗 𝑗=1,2,...
⊂ O 𝑈 is bounded if the sequence

𝑃 (𝑧, D𝑧 ) ⊤ ℎ 𝑗 , ℎ 𝑗

𝑗=1,2,...

is bounded in O 𝑈 × O (𝜕𝑈);
(c) to every 𝜀 > 0 there are 𝛿 > 0, 𝐶 > 0 such that, for every 𝑓 ∈ O (C𝑛 ),

∥ 𝑓 ∥ 𝐿 ∞ (𝑈 𝛿 ) ≤ 𝐶 𝑃 (𝑧, D𝑧 ) ⊤ 𝑓 𝐿 ∞ (𝑈 𝜀 ) + ∥ 𝑓 ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 ) .

As usual, 𝑃 (𝑧, D𝑧 ) ⊤ is the transpose of 𝑃 (𝑧, D𝑧 ).

shows that 𝑃B (𝑈)

Proof A look at the proof of Proposition 7.1.20 = B (𝑈) is
equivalent to the surjectivity of the map O ′ 𝑈 × O ′ (𝜕𝑈) −→ O ′ 𝑈 defined as

(𝜇, 𝜌) ↦→ 𝑃 (𝑥, D 𝑥 ) 𝜇 − 𝜌. (7.1.5)

7.1 The Sheaf of Hyperfunctions in Euclidean Space 177

The transpose of (7.1.5) is the (obviously injective) map

O 𝑈 ∋ ℎ ↦→ 𝑃 (𝑧, D𝑧 ) ⊤ ℎ, ℎ ∈ O 𝑈 × O (𝜕𝑈) .

Since O ′ 𝑈 and O ′ (𝜕𝑈) are Fréchet–Montel spaces the equivalence of (1) and
(5) in Lemma 7.1.17 implies
the equivalence of (a) and (b). Since, after restriction,
O (C ) is dense in O 𝑈 (Theorem 6.2.14) (c) implies (b) directly. We prove that
𝑛

the negation of (c) implies that of

(b). Suppose there exist 𝜀 > 0, a sequence of
numbers 𝛿 𝑗 ↘ 0 and a sequence 𝑓 𝑗 𝑗=1,2,... ⊂ O (C𝑛 ) such that

∥ 𝑓 ∥ 𝐿 ∞ 𝑈 > 𝑗 𝑃 (𝑧, D𝑧 ) ⊤ 𝑓 𝐿 ∞ (𝑈 𝜀)
+ ∥ 𝑓 ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 ) . (7.1.6)
𝛿𝑗

−1
If ℎ 𝑗 = 𝑗 𝑓 𝑗 𝐿 ∞ 𝑈 𝑓 𝑗 then ℎ 𝑗 𝐿 ∞ 𝑈 = 𝑗; given 𝛿 > 0 arbitrarily small, the
𝛿 𝛿𝑗
n o 𝑗
sequence ℎ 𝑗 𝑈 𝛿 is not bounded in O (𝑈 𝛿 ). But (7.1.6) implies
𝑗=1,2,...

𝑃 (𝑧, D𝑧 ) ⊤ ℎ 𝑗 𝐿 ∞ (𝑈 𝜀 )
+ ℎ𝑗 𝐿 ∞ ( (𝜕𝑈) 𝜀 )
< 1,

negating (b). □
We are now ready to prove the main result of this subsection (first proved in
[Cordaro-Trépreau, 1998]).

Theorem 7.1.22 Let 𝑃 (𝑥, D 𝑥 ) be a differential operator in Ω with C 𝜔 coefficients

and 𝑈 ⊂⊂ Ω be an open set. Suppose the following is true, whatever 𝑢 ∈ C ∞ (Ω):
(A) If 𝑢 ∈ C 𝜔 (𝑉), 𝑉 ⊂ Ω open, 𝜕𝑈 ⊂ 𝑉, then 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 ∈ C 𝜔 (𝑈) implies
𝑢 ∈ C 𝜔 (𝑈).

Under this hypothesis, 𝑃 (𝑥, D 𝑥 ) B (𝑈) = B (𝑈).

Proof Let 𝜀 > 0 be sufficiently small that 𝑈 𝜀 ∩ R𝑛 ⊂⊂ Ω. Consider the Fréchet

space
𝑬 = C ∞ (𝑈 𝜀 ∩ R𝑛 ) × Ob (𝑈 𝜀 ) × Ob ((𝜕𝑈) 𝜀 )
and its closed linear subspace 𝑭 consisting of the triples (𝑢, 𝑓 , 𝑔) such that
𝑃 (𝑥, D) ⊤ 𝑢 = 𝑓 in 𝑈 and 𝑢 = 𝑔 in (𝜕𝑈) 𝜀 ∩ R𝑛 ; Hypothesis (A) entails
𝑢 ∈ C 𝜔 (𝑈 𝜀 ∩ R𝑛 ). For each integer 𝑘 ≥ 𝑘 ◦ ≥ 𝜀 −1 let 𝑽 𝑘 denote the subset of
𝑭 consisting of the triples (𝑢, 𝑓 , 𝑔) ∈ 𝑭 such that there is a 𝑢♭ ∈ Ob 𝑈1/𝑘 with
the following properties: 𝑢♭ = 𝑢 in 𝑈 and 𝑢♭ 𝐿 ∞ (𝑈 ) ≤ 𝑘. The set 𝑽 𝑘 is convex
1/𝑘
and balanced, meaning that (𝑢, 𝑓 , 𝑔) ∈ 𝑽 𝑘 implies (𝜁𝑢, 𝜁 𝑓 , 𝜁 𝑔) ∈ 𝑽 𝑘 for all 𝜁 ∈ C,
Ø∞
|𝜁 | ≤ 1. Since 𝑭 = 𝑽 𝑘 the Baire category theorem entails that there is a neigh-
𝑘=𝑘◦
borhood V of 0 in 𝑬 such that V ∩ 𝑭 ⊂ 𝑽 𝑘 for some 𝑘 ≥ 𝑘 ◦ ; from now on 𝑘 shall
178 7 Hyperfunctions in Euclidean Space

be such an integer. This has the following consequence: there exist a compact set
𝐾 ⊂ 𝑈 𝜀 ∩ R𝑛 and numbers 𝑐 > 0, 𝑚 ∈ Z+ , such that, if ℎ ∈ O (C𝑛 ) satisfies
∑︁
𝑁 𝑃,𝐾 ,𝑚, 𝜀 (ℎ) = max |D 𝛼 ℎ| + 𝑃 (𝑧, D𝑧 ) ⊤ ℎ 𝐿 ∞ (𝑈 𝜀 ) + ∥ℎ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 ) ≤ 𝑐,
𝐾
| 𝛼 | ≤𝑚

then ∥ℎ∥ 𝐿 ∞ (𝑈1/𝑘 ) ≤ 𝑘. For each ℎ ∈ O (C𝑛 ), ℎ . 0, there is a 𝜆 > 0 such that
𝑁 𝑃,𝐾 ,𝑚, 𝜀 (ℎ) = 𝑐/𝜆, implying ∥ℎ∥ 𝐿 ∞ (𝑈1/𝑘 ) ≤ 𝑘/𝜆. Consequently,

𝑘
∀ℎ ∈ O (C𝑛 ) , ∥ℎ∥ 𝐿 ∞ (𝑈1/𝑘 ) ≤ 𝑁 𝑃,𝐾 ,𝑚, 𝜀 (ℎ) . (7.1.7)
𝑐
To complete the proof we show that (7.1.7) implies Property (c)in Lemma 7.1.21.
Suppose
(c) does not hold, then we can select 𝛿 ∈ 0, 𝑘 −1 and a sequence
ℎ 𝑗 𝑗=1,2,... ⊂ O (C𝑛 ) such that ℎ 𝑗 𝐿 ∞ (𝑈 𝛿 ) = 1 and

𝑃 (𝑧, D𝑧 ) ⊤ ℎ 𝑗 𝐿 ∞ (𝑈 𝜀 )
+ ℎ𝑗 𝐿 ∞ ( (𝜕𝑈) 𝜀 )
≤ 1/ 𝑗 (7.1.8)

for every 𝑗. By the Cauchy inequalities there is a 𝐶𝑚, 𝛿 > 1, depending solely on 𝑚
and 𝛿, such that
∑︁
max D 𝛼 ℎ 𝑗 ≤ 𝐶𝑚, 𝛿 ℎ 𝑗 𝐿 ∞ (𝑈 𝛿 ) = 𝐶𝑚, 𝛿 .
𝐾
| 𝛼 | ≤𝑚

If we put this into (7.1.7) we derive that

sup ℎ𝑗 𝐿 ∞ (𝑈1/𝑘 )
< +∞.
𝑗=1,2,...

By the Montel Theorem there is a subsequence ℎ 𝑗ℓ (1 ≤ 𝑗ℓ < 𝑗ℓ+1 , ℓ = 1, 2, ...)

converging in O 𝑈1/𝑘 , necessarily to zero in 𝑈1/𝑘 ∩ (𝜕𝑈) 𝜀 by (7.1.8), and there-
fore to zero in 𝑈1/𝑘 by the unique continuation principle, thereby contradicting
ℎ 𝑗ℓ 𝐿 ∞ (𝑈 𝛿 ) = 1 for all ℓ. □

Corollary 7.1.23 If 𝑃 (𝑥, D 𝑥 ) ⊤ is analytic-hypoelliptic in 𝑈 (Definition 3.1.2) then

𝑃 (𝑥, D 𝑥 ) B (𝑈) = B (𝑈).

7.1.5 Singularity hyperfunctions in Euclidean space

We denote by C 𝜔 (R𝑛 ) the sheaf of germs of real analytic, complex-valued functions

at points of R𝑛 . The stalk of C 𝜔 (R𝑛 ) at an arbitrary point 𝑥 ◦ ∈ R𝑛 , C 𝑥𝜔◦ , can be
identified with the ring O 𝑥 ◦ of germs at 𝑥 ◦ of holomorphic functions in C𝑛 , also,
through Taylor expansion about 𝑥 ◦ , with the ring of convergent power series (with
complex coefficients).
7.1 The Sheaf of Hyperfunctions in Euclidean Space 179

Definition 7.1.24 We shall refer to the quotient sheaf

Bsing (R𝑛 ) = B (R𝑛 ) /C 𝜔 (R𝑛 )

as the sheaf of germs of singularity hyperfunctions in R𝑛 . By a singularity hyper-

function in an open subset Ω of R𝑛 we shall mean a continuous section of Bsing (R𝑛 )
over Ω; singularity hyperfunctions in Ω form a linear space B sing (Ω).

We shall say that 𝑓 ∈ B (Ω) represents 𝑓 ♮ ∈ B sing (Ω) in a subset 𝐸 of Ω if the

germ of 𝑓 ♮ at an arbitrary point 𝑥 ◦ ∈ 𝐸 is the image of the germ of 𝑓 at 𝑥 ◦ under
sing
the quotient map B 𝑥 ◦ −→ B 𝑥 ◦ . If 𝑓 represents 𝑓 ♮ in Ω then supp 𝑓 ♮ = singsuppa 𝑓
(Definition 7.4.7). The stalk at 𝑥 ◦ ∈ R𝑛 of Bsing (R𝑛 ) is the quotient vector space
B 𝑥 ◦ /C 𝑥𝜔◦ .
In studying the singularities of solutions of differential equations we will be
mainly interested in singularity hyperfunctions rather than hyperfunctions proper.

Remark 7.1.25 It should be pointed out that singularity hyperfunctions are not a
generalization of distributions; they generalize equivalence classes of distributions
modulo analytic functions.

We are going to make use of an explicit formulation of Theorem 10.2.8.

Lemma 7.1.26 Let {𝑈 𝜄 } 𝜄 ∈𝐼 be an open covering of Ω and, for each pair of indices
(𝜄, 𝜅) ∈ 𝐼 2 such that 𝑈 𝜄 ∩ 𝑈 𝜅 ≠ ∅, a function 𝑔 𝜄,𝜅 ∈ C 𝜔 (𝑈 𝜄 ∩ 𝑈 𝜅 ) such that
𝑔 𝜄,𝜅 +𝑔 𝜅 ,𝜆 +𝑔𝜆, 𝜄 = 0 in 𝑈 𝜄 ∩𝑈 𝜅 ∩𝑈𝜆 (𝜆 ∈ 𝐼, 𝑈 𝜄 ∩𝑈 𝜅 ∩𝑈𝜆 ≠ ∅). Under this hypothesis,
to each 𝜄 ∈ 𝐼 there is a function ℎ 𝜄 ∈ C 𝜔 (𝑈 𝜄 ) such that 𝑔 𝜄,𝜅 = (ℎ 𝜄 − ℎ 𝜅 )|𝑈 𝜄 ∩𝑈𝜅 if
𝑈 𝜄 ∩ 𝑈 𝜅 ≠ ∅.

Theorem 7.1.27 Let the domain Ω in R𝑛 and 𝑓 ♭ ∈ B sing (Ω) be arbitrary. There is
an 𝑓 ∈ B (Ω) representing 𝑓 ♭ in Ω.

Proof There exist an open covering {𝑈 𝜄 } 𝜄 ∈𝐼 of Ω and, for each 𝜄 ∈ 𝐼, a hyperfunction

𝑓 𝜄 ∈ B (𝑈 𝜄 ) whose equivalence class mod C 𝜔 (𝑈 𝜄 ) is 𝑓 ♭ 𝑈 𝜄 . If 𝑈 𝜄 ∩𝑈 𝜅 ≠ ∅ (𝜄, 𝜅 ∈ 𝐼)
then 𝑔 𝜄,𝜅 = 𝑓 𝜄 |𝑈 𝜄 ∩𝑈𝜅 − 𝑓 𝜅 |𝑈 𝜄 ∩𝑈𝜅 ∈ C 𝜔 (𝑈 𝜄 ∩ 𝑈 𝜅 ). Since 𝑔 𝜄,𝜅 + 𝑔 𝜅 ,𝜆 + 𝑔𝜆, 𝜄 = 0 in
𝑈 𝜄 ∩ 𝑈 𝜅 ∩ 𝑈𝜆 (≠ ∅) Lemma 7.1.26 states that to every 𝜄 ∈ 𝐼 there is a function
ℎ 𝜄 ∈ C 𝜔 (𝑈 𝜄 ) such that 𝑔 𝜄,𝜅 = (ℎ 𝜄 − ℎ 𝜅 )|𝑈 𝜄 ∩𝑈𝜅 . Then 𝑓 𝜄 + ℎ 𝜄 ∈ B (𝑈 𝜄 ) also
represent 𝑓 ♭ 𝑈 𝜄 in 𝑈 𝜄 ; if 𝑈 𝜄 ∩ 𝑈 𝜅 ≠ ∅ we have

( 𝑓 𝜄 − ℎ 𝜄 )|𝑈 𝜄 ∩𝑈𝜅 = ( 𝑓 𝜅 − ℎ 𝜅 )|𝑈 𝜄 ∩𝑈𝜅 .

This proves that there is an 𝑓 ∈ B (Ω) such that 𝑓 |𝑈 𝜄 = 𝑓 𝜄 − ℎ 𝜄 for every 𝜄 ∈ 𝐼,

thereby representing 𝑓 ♭ in Ω. □

Corollary 7.1.28 The quotient map B (Ω) −→ B (Ω) /C 𝜔 (Ω) induces a natural
isomorphism B (Ω) /C 𝜔 (Ω) B sing (Ω).

Proof If 𝑓 ∈ B (Ω) represents the zero singularity hyperfunction in Ω then 𝑓 ∈

C 𝜔 (Ω). □
180 7 Hyperfunctions in Euclidean Space

Corollary 7.1.29 The sheaf Bsing (R𝑛 ) is flabby.

Proof We have just seen that there is an 𝑓 ∈ B (Ω) that represents 𝑓 ♭ in Ω and
so does an arbitrary extension 𝐹 ∈ B (R𝑛 ) of 𝑓 (Corollary 7.1.10). The singularity
hyperfunction 𝐹 ♭ represented by 𝐹 in R𝑛 extends 𝑓 ♭ . □

7.2 Boundary values of holomorphic functions in wedges

We are going to revisit the concept of boundary value of a holomorphic function in a

wedge W𝛿 (Ω, Γ) = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; 𝑥 ∈ Ω, 𝑦 ∈ Γ, |𝑦| < 𝛿} [with height 𝛿 > 0,
see (3.3.1)], this time without the requirement of temperedness on the growth of
the functions at the edge Ω (an arbitrary open subset of R𝑛 ). The forthcoming
Lemmas 7.2.1, 7.2.2, and the ensuing Definition 7.2.3 are of crucial importance in
the presentation of Hyperfunction Theory in this book.
In this section Γ shall always be a convex open cone in R𝑛 \ {0}, ℎ an arbi-
trarily given element of O (W𝛿 (Ω, Γ)) [the algebra of holomorphic functions in
W𝛿 (Ω, Γ)], 𝑈 ⊂⊂ Ω an open subset of R𝑛 . Given arbitrarily v ∈ Γ, |v| < 𝛿,
𝜑 ∈ O (C𝑛 ), we define
D E ∫
𝜇𝑈
ℎ,v , 𝜑 = ℎ (𝑧) 𝜑 (𝑧) d𝑧 (7.2.1)
∫𝑈+𝑖v
= ℎ (𝑥 + 𝑖v) 𝜑 (𝑥 + 𝑖v) d𝑥.
𝑈

As usual d𝑧 = d𝑧1 · · · d𝑧 𝑛 and likewise for d𝑥; (7.2.1) defines an analytic functional
carried by the compact subset of C𝑛 , 𝑈 +𝑖v. In passing note that 𝑈 +𝑖v has the Runge
property, being a translate of 𝑈 ⊂ R𝑛 .
Lemma 7.2.1 Let ℎ ∈ O (W𝛿 (Ω, Γ)) and 𝑈 ⊂⊂ Ω be an open subset of R𝑛 .
Suppose that the boundary 𝜕𝑈 of 𝑈 is smooth and let v 𝑗 ∈ Γ, v 𝑗 < 𝛿, 𝑗 = 1, 2. The
′ (C𝑛 ) is carried by ℭ = 𝜕𝑈 + 𝑖 [v , v ].
ℎ,v1 − 𝜇 ℎ,v2 ∈ O
analytic functional 𝜇𝑈 𝑈
1 2

We have denoted by [v1 , v2 ] the straight-line segment joining v1 and v2 ; we have

[v1 , v2 ] ⊂ Γ since Γ is convex; and |𝑡v1 + (1 − 𝑡) v2 | < 𝛿 for every 𝑡 ∈ [0, 1].
Proof The closure of the set

𝔄 = {𝑧 ∈ C𝑛 ; Re 𝑧 ∈ 𝑈, Im 𝑧 = 𝑡v1 + (1 − 𝑡) v2 , 0 < 𝑡 < 1 }

is a compact subset of W𝛿 (Ω, Γ). Its boundary 𝔄\𝔄 consists of the “horizontal”
leaves 𝑈 + 𝑖v1 and 𝑈 + 𝑖v2 and the “collar” ℭ. Let 𝜑 ∈ O (C𝑛 ) be arbitrary. Since the
(𝑛, 0)-form ℎ (𝑧) 𝜑 (𝑧) d𝑧 is closed in W𝛿 (Ω, Γ) Stokes’ Theorem entails
D E ∫
𝑈 𝑈
𝜇 ℎ,v1 − 𝜇 ℎ,v2 , 𝜑 = ℎ (𝑧) 𝜑 (𝑧) d𝑧, (7.2.2)
ℭ

whence the claim. □

7.2 Boundary values of holomorphic functions in wedges 181

Lemma 7.2.2 Let ℎ ∈ O (W𝛿 (Ω, Γ)) and 𝑈 ⊂⊂ Ω be an open subset of R𝑛 .

Suppose that the boundary 𝜕𝑈 of 𝑈 is smooth and let v ∈ Γ, |v| < 𝛿. There is an
′
analytic functional 𝜇 ℎ ∈ O 𝑈 having the following property:
𝑈

♣ To every compact neighborhood 𝐾 of 𝜕𝑈 in C𝑛 there is an integer 𝑝 𝐾 ≥ 1 such

that if Z+ ∋ 𝑝 ≥ 𝑝 𝐾 then the analytic functional 𝜇𝑈 1 − 𝜇𝑈
ℎ is carried by 𝐾. ℎ, 𝑝 v

The coset mod O ′ (𝜕𝑈) of the analytic functional 𝜇𝑈

𝑈
ℎ , 𝜇 ℎ , is independent of
the choice
of the vector
𝑈 v. If 𝑉 is an open subset of 𝑈 whose boundary 𝜕𝑉 is smooth
then 𝜇𝑉ℎ = 𝜌𝑈 𝑉 𝜇ℎ .

Proof As before let 𝜑 ∈h O (Ci𝑛 ) be arbitrary. Let 𝑞 > 𝑝 ≥ 1 be integers; we apply

(7.2.2) with ℭ = 𝜕𝑈 + 𝑖 𝑞1 , 𝑝1 v:
∫ ∫
𝜇𝑈
ℎ, 1 v
− 𝜇𝑈
ℎ, 1 v
, 𝜑 = h i ℎ (𝑧 + 𝑖𝑡v) 𝜑 (𝑧 + 𝑖𝑡v) d (𝑧 + 𝑖𝑡v) . (7.2.3)
1 1
𝑝 𝑞 𝑧 ∈𝜕𝑈 𝑡∈ 𝑞, 𝑝

We use a finite Taylor expansion of 𝜓 (𝑧, 𝑡) = 𝜑 (𝑧 + 𝑖𝑡v) with respect to 𝑡:

𝑁 ∫ 𝑡
∑︁ 𝑡𝑗 𝑗 1
𝜑 (𝑧 + 𝑖𝑡v) = 𝜕𝑡 𝜓 (𝑧, 0) + (𝑡 − 𝑠) 𝑁 𝜕𝑠𝑁 +1 𝜓 (𝑧, 𝑠) d𝑠.
𝑗=0
𝑗! 𝑁! 0

We define new analytic functionals:

𝑁 ∫ ∫
∑︁ 𝑡𝑗 𝑗
𝜆 𝑝+1, 𝑁 , 𝜑 = 𝜕 𝜓 (𝑧, 0) ℎ (𝑧 + 𝑖𝑡v) d (𝑧 + 𝑖𝑡v) ,
𝑗! 𝑡
h i
1 1
𝑗=0 𝑧 ∈𝜕𝑈 𝑡∈ 𝑝+1 , 𝑝
∫ ∫ 𝑡= 𝑝1 ∫
1
𝑅 𝑝, 𝑁 , 𝜑 = (𝑡 − 𝑠) 𝑁 ℎ (𝑧 + 𝑖𝑡v) 𝜕𝑠𝑁 +1 𝜓 (𝑧, 𝑠) d𝑠d (𝑧 + 𝑖𝑡v) .
𝑁! 1
𝑧 ∈𝜕𝑈 𝑡= 𝑝+1 𝑠 ∈ [0,𝑡 ]

Taking 𝑞 = 𝑝 + 1 in (7.2.3) yields

𝑈 𝑈
𝜇 ℎ, 1 v − 𝜇 ℎ, 1 v , 𝜑 = 𝜆 𝑝+1, 𝑁 + 𝑅 𝑝, 𝑁 , 𝜑 . (7.2.4)
𝑝 𝑝+1

It is evident that the analytic functionals 𝜆 𝑝, 𝑁 are carried by 𝜕𝑈. Concerning 𝑅 𝑝, 𝑁

we make use of the Cauchy inequalities:

1 𝑝 𝑁 +1
sup 𝜕𝑠𝑁 +1 𝜓 (𝑧, 𝑠) ≤ sup |𝜑 (𝑧 + 𝑖𝑧 ◦ v)| .
(𝑁 + 1)! 0≤𝑠 ≤ 1 2 𝑧◦ ∈C, |𝑧◦ |= 3
𝑝 𝑝

If we set ∫ ∫ 𝑡= 𝑝1
𝑀𝑝 = |ℎ (𝑧 + 𝑖𝑡v)| |d (𝑧 + 𝑖𝑡v)|
1
𝑧 ∈𝜕𝑈 𝑡= 𝑝+1
182 7 Hyperfunctions in Euclidean Space

we obtain
𝑝 𝑁 +1 ∫ 1
𝑝
𝑅 𝑝, 𝑁 , 𝜑 ≤ 𝑀 𝑝 (𝑁 + 1) 𝑠 𝑁 d𝑠 sup |𝜑 (𝑧 + 𝑖𝑧 ◦ v)|
2 0 𝑧◦ ∈C, |𝑧◦ |= 𝑝3

≤ 𝑀 𝑝 2−𝑁 −1 sup |𝜑 (𝑧 + 𝑖𝑧 ◦ v)| .

𝑧◦ ∈C, |𝑧◦ |= 𝑝3

For each 𝑝 = 1, 2, ..., we select an integer 𝑁 sufficiently large that 𝑀 𝑝 2−𝑁 −1 ≤ 2− 𝑝 ;

with such a choice of 𝑁 we define 𝜆 𝑝 = 𝜆 𝑝, 𝑁 . We also define

3𝛿
𝐾 𝑝 = 𝑧 ∈ C𝑛 ; dist (𝑧, 𝜕𝑈) ≤ , 𝑝 = 1, 2, ....
𝑝

We have shown that

𝜇𝑈
ℎ, 1 v
− 𝜇𝑈 1
v
− 𝜆 𝑝+1 , 𝜑 ≤ 2− 𝑝 max |𝜑 (𝑧)| . (7.2.5)
𝑝 ℎ, 𝑝+1 𝑧 ∈𝐾 𝑝

Now define
𝑝
∑︁
𝜇˜ ℎ, 𝑝 = 𝜇𝑈
ℎ, 𝑝1 v
− 𝜆𝑘 ;
𝑘=1

(7.2.5) reads
𝜇˜ ℎ, 𝑝 − 𝜇˜ ℎ, 𝑝+1 , 𝜑 ≤ 2− 𝑝 max |𝜑 (𝑧)| ,
𝑧 ∈𝐾 𝑝

telling us that
the analytic functionals 𝜇˜ ℎ, 𝑝+𝑘 (𝑘 = 1, 2, ...) form a Cauchy sequence
in O ′ 𝐾 𝑝 . It is readily seen that its limit 𝜇𝑈 ℎ is independent of 𝑝 and that it satisfies
the requirements of the statement.
′ 𝑈 /O ′ (𝜕𝑈) is independent of v is an easy

The fact that the coset 𝜇𝑈 ℎ ∈ O
consequence of Lemma 7.2.1 and of Property ♣.
The last part of the statement is a direct consequence of the formula
∫
1 1
𝜇𝑈
ℎ, 𝑝1 v
− 𝜇 𝑉
ℎ, 𝑝1 v
, 𝜑 = ℎ 𝑥 + 𝑖 v 𝜑 𝑥 + 𝑖 v d𝑥.
𝑈\𝑉 𝑝 𝑝

Letting 𝑝 ↗ +∞ shows that 𝜇𝑈

ℎ − 𝜇 ℎ is carried by 𝑈\𝑉, which is the claim.
𝑉
□

If ℎ ∈ O (W𝛿 (Ω, Γ)) and 𝑈 ⊂⊂ Ω is an open subset of R𝑛 with a C ∞ boundary

𝜕𝑈 we denote by 𝑏𝑈 ℎ the hyperfunction in 𝑈 defined by 𝜇𝑈 ℎ ; 𝑏𝑈 ℎ is another notation

for the coset 𝜇𝑈ℎ ∈ B (𝑈). If 𝑉 ⊂ 𝑈 and if 𝜕𝑉 is smooth the restriction of 𝑏𝑈 ℎ to
𝑉 is equal to 𝑏 𝑉 ℎ. If {𝑈𝜈 } 𝜈=1,2,... is an increasing sequence of open subsets of R𝑛
whose closures 𝑈 𝜈 are compact subsets of Ω and whose boundaries are smooth, and
7.2 Boundary values of holomorphic functions in wedges 183
∞
Ø
if Ω = 𝑈𝜈 , then there is a hyperfunction in Ω whose restriction to 𝑈𝜈 is equal to
𝜈=1
𝑏𝑈𝜈 ℎ; we
shall
denote it by 𝑏 Ω ℎ. If Ω is bounded it follows from Lemma 7.2.2 that
𝑏 Ω ℎ = 𝜇Ωℎ .

Definition 7.2.3 The hyperfunction 𝑏 Ω ℎ is called the hyperfunction boundary

value in Ω of the function ℎ ∈ O (W𝛿 (Ω, Γ)).

Cf. Definition 3.3.7. We shall often omit the qualificative hyperfunction.

Remark 7.2.4 (cf. Remark 3.3.8) The independence from the choice of the vector
v ∈ Γ of the definitions of 𝜇𝑈 ℎ (Lemma 7.2.2) and, as a consequence, of those of 𝑏𝑈 ℎ
and 𝑏 Ω ℎ, shows that if Γ ′ is a nonempty convex open cone in R𝑛 \ {0}, Γ ′ ⊂ Γ, and
′ : O (W (Ω, Γ ′ )) −→ B (Ω) is the corresponding boundary value map, then
if 𝑏 Ω 𝛿

′
𝑏Ω ℎ| W𝛿 (Ω,Γ′ ) = 𝑏 Ω ℎ

for every ℎ ∈ O (W𝛿 (Ω, Γ)).

Proposition 7.2.5 The boundary value map 𝑏 Ω : O (W𝛿 (Ω, Γ)) −→ B (Ω) is
injective.

Proof Let the open subset 𝑈 of R𝑛 have compact closure contained in Ω, be

connected and have a smooth boundary 𝜕𝑈, and let v ∈ Γ, |v| < 𝛿. Let
ℎ ∈ O (W𝛿 (Ω, Γ)) be such that 𝑏𝑈 ℎ = 0 everywhere in 𝑈. By Lemma 7.2.2,
given any compact neighborhood 𝐾 of 𝜕𝑈 in C𝑛 we have supp 𝜇𝑈 1 ⊂ 𝐾 provided
ℎ, 𝑝 v
𝑝 ∈ Z+ is sufficiently large. Consider then the integral
𝜏 12 𝑛 ∫ 2
𝑔 𝑝, 𝜏 (𝑤) = ℎ (𝑧) e−𝜏 ⟨𝑧−𝑤 ⟩ d𝑧,
𝜋 𝑈+𝑖 𝑝1 v

where 𝜏 > 0 and ⟨𝑧⟩ 2 = 𝑧 2𝑗 . Thus our hypothesis implies

Í𝑛
𝑗=1

𝜏 12 𝑛
𝑔 𝑝, 𝜏 (𝑤) ≤ 𝐶 𝑝 max exp −𝜏 |𝑥 − Re 𝑤| 2 − |𝑦 − Im 𝑤| 2 . (7.2.6)
𝑥+𝑖𝑦 ∈𝐾 𝜋
We derive from (7.2.6) that, provided 𝐾 is a sufficiently “thin” tubular neighborhood
of 𝜕𝑈 and 𝑝 is correspondingly large, there is an open subset V of C𝑛 such that
𝑉 = 𝑈 ∩ V ≠ ∅ and for all 𝑧 ∈ 𝐾 and 𝑤 ∈ V,

|𝑥 − Re 𝑤| 2 − |𝑦 − Im 𝑤| 2 ≥ 𝑐 > 0. (7.2.7)

By keeping 𝑝 fixed and letting 𝜏 ↗ +∞, we derive from (7.2.6) and (7.2.7) that
𝑔 𝑝, 𝜏 (𝑤) → 0 for all 𝑤 ∈ V. But if 𝑝 is large enough we can find an open subset
𝑊 ≠ ∅ of 𝑉 such that 𝑤 = 𝜉 + 𝑖 𝑝1 v ∈ V whatever 𝜉 ∈ 𝑊. We have
184 7 Hyperfunctions in Euclidean Space

1
𝜏 12 𝑛 ∫
1

2
𝑔 𝑝, 𝜏 𝜉 +𝑖 v = ℎ 𝑥 + 𝑖 v e−𝜏 ⟨𝑥− 𝜉 ⟩ d𝑥
𝑝 𝜋 𝑈 𝑝

and it is well known (cf. the proof of Theorem 6.2.14) that

1 1
∀𝜉 ∈ 𝑊, lim 𝑔 𝑝, 𝜏 𝜉 + 𝑖 v = ℎ 𝜉 + 𝑖 v .
𝜏→+∞ 𝑝 𝑝

We conclude that ℎ 𝜉 + 𝑖 𝑝1 v = 0 for all 𝜉 ∈ 𝑊. This demands that ℎ ≡ 0 in
W𝛿 (Ω, Γ) (because, e.g., the submanifold 𝑊 + 𝑖 𝑝 −1 v is “parallel” to the open subset
𝑊 of R𝑛 ). □

Now suppose that we have a finite family of wedges W𝛿 𝑗 Ω, Γ 𝑗 ( 𝑗 = 1, ..., 𝑁)
Í
and a corresponding set of functions ℎ 𝑗 ∈ O W𝛿 𝑗 Ω, Γ 𝑗 . The sum 𝑁𝑗=1 𝑏 Ω ℎ 𝑗 is
a hyperfunction in Ω. This raises two questions:
(1) Can every hyperfunction in Ω be represented as a sum of boundary values of
holomorphic functions in wedges with edge Ω?
Í
(2) How do we recognize that two different sums of the type 𝑁𝑗=1 𝑏 Ω ℎ 𝑗 represent
the same hyperfunction in Ω?
To answer these questions we shall rely (in the next section) on the extension to
analytic functionals of the FBI transform (see Definition 7.3.2 below). But before
doing this we avail ourselves of Corollary 7.1.28 to adapt the concept of boundary
value of holomorphic functions in wedges to singularity hyperfunctions
Definition 7.2.6 Let Γ be an open cone in R𝑛 \ {0} and W𝛿 (Ω, Γ) be the wedge
(3.3.1). By the singularity hyperfunction boundary value map in W𝛿 (Ω, Γ) we
shall mean the composite map
𝑏Ω
O (W𝛿 (Ω, Γ)) −→ B (Ω) −→ B (Ω) /C 𝜔 (Ω) B sing (Ω)

where 𝑏 Ω is the hyperfunction boundary map (Definition 7.2.3) and the right-hand
arrow is the quotient map.
We may use the notation 𝑏 Ω also for the singularity hyperfunction boundary map.

7.3 The FBI Transform of Analytic Functionals

7.3.1 The FBI transform of an analytic functional. Definition and basic

properties

Unless specified otherwise, henceforth all analytic functionals under consideration

will be carried by compact subsets of R𝑛 , i.e., will belong to O ′ (R𝑛 ). We use the
notation ⟨𝑧⟩ 2 = |𝑥| 2 − |𝑦| 2 + 2𝑖𝑥 · 𝑦 if 𝑧 = 𝑥 + 𝑖𝑦. We introduce a kind of “dual
7.3 The FBI Transform of Analytic Functionals 185

variable” 𝜁 that varies in a “thin” conic neighborhood of R𝑛 \ {0}; its thinness will
be measured by a small parameter 𝛾 > 0:

𝚪𝛾(𝑛) = {𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 ; |𝜂| < 𝛾 |𝜉 |} . (7.3.1)

For 𝜁 ∈ 𝚪 𝛾(𝑛) the notation ⟨𝜁⟩ makes sense

√ (cf. Remark 1.1.1); we have ⟨𝜁⟩ = |𝜉 |
when 𝜂 = 0; ⟨𝜁⟩ is the main branch of 𝜁 · 𝜁. Let us state right-away

Lemma 7.3.1 If 𝜁 ∈ 𝚪 𝛾(𝑛) then

𝛾
|Im ⟨𝜁⟩| ≤ 𝛾 |𝜉 | ≤ √︁ Re ⟨𝜁⟩ , (7.3.2)
1 − 𝛾2
√︄
1 + 𝛾2
|⟨𝜁⟩| ≤ |𝜁 | ≤ Re ⟨𝜁⟩ . (7.3.3)
1 − 𝛾2

Proof Suppose 𝜁 ∈ 𝚪 𝛾(𝑛) ; we have

(Re ⟨𝜁⟩) 2 − (Im ⟨𝜁⟩) 2 = Re ⟨𝜁⟩ 2 = |𝜉 | 2 − |𝜂| 2 ≥ 1 − 𝛾 2 |𝜉 | 2 ,

1 + 𝛾 2 |𝜉 | 2 ≥ |𝜉 | 2 + |𝜂| 2 = |𝜁 | 2 ≥ ⟨𝜁⟩ 2 = (Re ⟨𝜁⟩) 2 + (Im ⟨𝜁⟩) 2 .

By adding these two inequalities we get

1 + 𝛾 2 |𝜉 | 2 + (Re ⟨𝜁⟩) 2 − (Im ⟨𝜁⟩) 2 ≥ 1 − 𝛾 2 |𝜉 | 2 + (Re ⟨𝜁⟩) 2 + (Im ⟨𝜁⟩) 2 ,
√︁
whence |Im ⟨𝜁⟩| ≤ 𝛾 |𝜉 |. From
√︁ the first one we also derive Re ⟨𝜁⟩ ≥ 1 − 𝛾 2 |𝜉 |,
whence (7.3.2). Since |𝜁 | ≤ 1 + 𝛾 2 |𝜉 | we also get (7.3.3). □

We introduce a second parameter 𝜅 > 0 and the Jacobian determinant of the map
𝜁 ↦→ 𝜁 + 𝑖𝜅 ⟨𝜁⟩ 𝑧, Δ 𝜅 (𝑧, 𝜁) = 1 + 𝑖𝜅 ⟨𝜁⟩ −1 (𝑧 · 𝜁) (see Remark 3.4.4).

Definition 7.3.2 By the Fourier–Brós–Iagolnitzer (in short, FBI) transform of

an analytic functional 𝜇 ∈ O ′ (R𝑛 ) we shall mean the holomorphic function of
(𝑧, 𝜁) ∈ C𝑛 × 𝚪𝛾(𝑛) , depending on 𝜅 > 0,
D E
F𝜅 𝜇 (𝑧, 𝜁) = 𝜇 𝑤 , Δ 𝜅 (𝑧 − 𝑤, 𝜁) exp 𝑖𝜁 · (𝑧 − 𝑤) − 𝜅 ⟨𝜁⟩ ⟨𝑧 − 𝑤⟩ 2 . (7.3.4)

Remark 7.3.3 Both parameters 𝛾, 𝜅 have been introduced for technical reasons, in
particular to facilitate the inversion of the FBI transform acting on analytic function-
als with large supports in R𝑛 . The 𝑧 → 𝑧/𝜅, 𝜁 →
change of variables 𝜅𝜁 transforms
Δ 𝜅 (𝑧, 𝜁) exp 𝑖𝜁 · 𝑧 − 𝜅 ⟨𝜁⟩ ⟨𝑧⟩ 2 into Δ1 (𝑧, 𝜁) exp 𝑖𝜁 · 𝑧 − ⟨𝜁⟩ ⟨𝑧⟩ 2 .
186 7 Hyperfunctions in Euclidean Space

The essential feature of the FBI transform resides, obviously, in the nature of the
exponent
𝜑 𝜅 (𝑧 − 𝑤, 𝜁) = 𝑖𝜁 · (𝑧 − 𝑤) − 𝜅 ⟨𝜁⟩ ⟨𝑧 − 𝑤⟩ 2 . (7.3.5)
Let 𝐾 a compact subset of R𝑛 . Given 𝜇 ∈ O ′ (𝐾) arbitrarily, to every 𝜀 > 0 there
are positive constants 𝑀 𝜀 and 𝑀 𝜀′ such that

|⟨𝜇 𝑤 , Δ 𝜅 (𝑧 − 𝑤) exp 𝜑 𝜅 (𝑧 − 𝑤, 𝜁)⟩| (7.3.6)

≤ 𝑀 𝜀 max (|Δ 𝜅 (𝑧 − 𝑤, 𝜁)| exp Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁))
dist(𝑤,𝐾) ≤ 𝜀

≤ 𝑀 𝜀′ max (1 + |𝑧 − 𝑤|) 𝑛 exp Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁) .

dist(𝑤,𝐾) ≤ 𝜀

We will exploit the fact that Re 𝜑 𝜅 (𝑧, 𝜁) = −𝜅 |𝜁 | |𝑧| 2 for 𝑧, 𝜁 ∈ R𝑛 ; and that
Re 𝜑 𝜅 (𝑧, 𝜁) has good bounds provided |Im 𝑧| and |Im 𝜁 | /|Re 𝜁 | are sufficiently small.
More precisely, let 𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 , 𝜁 = 𝜉 + 𝑖𝜂 ∈ 𝚪𝛾(𝑛) with 0 < 𝛾 < 1; we have

Re 𝜑 𝜅 (𝑧, 𝜁) = −𝑥 · 𝜂 − 𝑦 · 𝜉 − 𝜅 Re ⟨𝜁⟩ |𝑥| 2 − |𝑦| 2 − 2𝜅 Im ⟨𝜁⟩ (𝑥 · 𝑦) .

We deduce from (7.3.2)–(7.3.3), now replacing 𝑧 by 𝑧 − 𝑤, 𝑤 = 𝑢 + 𝑖𝑣,

Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁) ≤ (𝛾 |𝑥 − 𝑢| + |𝑦 − 𝑣|) |𝜉 |
!
2 2 2𝛾
− 𝜅 |𝑥 − 𝑢| − |𝑦 − 𝑣| − √︁ |𝑥 − 𝑢| |𝑦 − 𝑣| Re ⟨𝜁⟩ .
1 − 𝛾2

From (7.3.3) we deduce

√︄
1 + 𝛾2
Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁) ≤ (𝛾 |𝑥 − 𝑢| + |𝑦 − 𝑣|) Re ⟨𝜁⟩
1 − 𝛾2
!
2 2 2𝛾
− 𝜅 |𝑥 − 𝑢| − (|𝑦 − 𝑣|) − √︁ |𝑥 − 𝑢| |𝑦 − 𝑣| Re ⟨𝜁⟩ .
1 − 𝛾2

If 𝛾 ≤ 1/2 we get √︁
Re ⟨𝜁⟩ ≤ |𝜉 | ≤ 5/3 Re ⟨𝜁⟩ (7.3.7)
and

Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁) /Re ⟨𝜁⟩ ≤ 2 (𝛾 |𝑥 − 𝑢| + |𝑦 − 𝑣|) (7.3.8)

− 𝜅 |𝑥 − 𝑢| 2 − |𝑦 − 𝑣| 2 − 3𝛾 |𝑥 − 𝑢| |𝑦 − 𝑣| .
7.3 The FBI Transform of Analytic Functionals 187

Proposition 7.3.4 Let 𝜅 > 0 be arbitrary. If 𝐾 ⊂ R𝑛 is compact and 𝜇 ∈ O ′ (𝐾) the

following properties hold.
(i) To every 𝜀 > 0 there are positive constants 𝛾 ≤ 1/2, 𝑟, 𝐶 𝜀 , such that

|F𝜅 𝜇 (𝑧, 𝜁)| ≤ 𝐶 𝜀 exp (𝜀 |𝜉 |) (7.3.9)

for every 𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 , |𝑦| ≤ 𝑟, and every 𝜁 = 𝜉 + 𝑖𝜂 ∈ 𝚪𝛾(𝑛) .

(ii) To every 𝑑 > 0 there are positive constants 𝛾 ≤ 1/2, 𝑟, 𝐶, such that

|F𝜅 𝜇 (𝑧, 𝜁)| ≤ 𝐶 exp −2−4 𝜅𝑑 2 |𝜉 | (7.3.10)

for every 𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 such that dist (𝑥, 𝐾) ≥ 𝑑, |𝑦| < 𝑟, and every 𝜁 =
𝜉 + 𝑖𝜂 ∈ 𝚪𝛾(𝑛) .

Proof We use the notation 𝑤 = 𝑢 + 𝑖𝑣; if we put |𝑣| ≤ 𝜀 ′, |𝑦| < 𝑟, in (7.3.8) we get

Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁) /Re ⟨𝜁⟩ ≤ −𝜅 |𝑥 − 𝑢| 2 + 2𝛾 |𝑥 − 𝑢| + 3𝛾𝜅 |𝑥 − 𝑢| (𝑟 + 𝜀 ′)

+ 2 (𝑟 + 𝜀 ′) + 𝜅 (𝑟 + 𝜀 ′) 2 .

Given 𝜀 > 0 arbitrarily we can select 𝜀 ′ , 𝛾 and 𝑟 so small that

1 1
Re 𝜑 𝜅 (𝑧 − 𝑤, 𝜁) /Re ⟨𝜁⟩ ≤ − 𝜅 |𝑥 − 𝑢| 2 + 𝜀.
2 2
Applying (7.3.6) with 𝜀 ′ substituted for 𝜀 and taking (7.3.7) into account yields

|⟨𝜇 𝑤 , Δ 𝜅 (𝑧 − 𝑤) exp 𝜑 𝜅 (𝑧 − 𝑤, 𝜁)⟩| (7.3.11)

1 1
≤ 𝑀 𝜀′′′ sup (1 + |𝑥 − 𝑢| |𝜉 |) 𝑛 exp 𝜀 − 𝜅 |𝑥 − 𝑢| 2 |𝜉 |
dist(𝑢,𝐾) ≤ 𝜀′ 2 2

for a sufficiently large constant 𝑀 𝜀′′′ > 0 and all 𝑧 = 𝑥 + 𝑖𝑦, |𝑦| < 𝑟, 𝜁 = 𝜉 + 𝑖𝜂 ∈ 𝚪𝛾(𝑛) .
Since
∀𝑎 > 0, ∀𝜁 ∈ Γ𝛾(𝑛) , (1 + 𝑎 |𝜉 |) 𝑛 exp −𝑎 2 |𝜉 | ≲ (1 + |𝜉 |) 𝑛/2

we see that (7.3.11) implies (i).

If |𝑥 − 𝑢| ≥ 𝑑 then (7.3.7) and (7.3.11) imply

|⟨𝜇 𝑤 , Δ 𝜅 (𝑧 − 𝑤) exp 𝜑 𝜅 (𝑧 − 𝑤, 𝜁)⟩|

1 2 1 1
≤ 𝑀 𝜀′′′ e ( 𝜀− 8 𝜅 𝑑 ) | 𝜉 | sup 𝑛 2
(1 + |𝑥 − 𝑢| |𝜉 |) exp − 𝜀 + 𝜅 |𝑥 − 𝑢| |𝜉 | .
dist(𝑢,𝐾) ≤ 𝜀′ 2 4

1 2
We select 𝜀 < 16 𝑐𝜅𝑑 (and then 𝛾 and 𝑟 correspondingly) to deduce directly (ii). □
188 7 Hyperfunctions in Euclidean Space

Corollary 7.3.5 Let 𝑈 be a bounded open subset of R𝑛 and let 𝜇 𝑗 ∈ O ′ 𝑈 ,
𝑗 = 1, 2, be two analytic functionals representing the same hyperfunction in 𝑈,
i.e., 𝜇1 − 𝜇2 ∈ O ′ (𝜕𝑈). Let 𝑥 ◦ ∈ 𝑈 be arbitrary. There exist an open subset 𝑉 C of
C𝑛 , 𝑥 ◦ ∈ 𝑉 = 𝑉 C ∩ R𝑛 ⊂ 𝑈, and positive constants 𝛾, 𝐶, 𝑅, such that

∀ (𝑧, 𝜁) ∈ 𝑉 C × Γ𝛾(𝑛) , |F𝜅 (𝜇1 − 𝜇2 ) (𝑧, 𝜁)| ≤ 𝐶e− |𝜁 |/𝑅 . (7.3.12)

7.3.2 Inversion of the FBI transform

What information can one draw from the FBI transform of an analytic functional
𝜇 and, first of all, can one recover 𝜇 from F𝜅 𝜇 (𝑧, 𝜁)? We need the following
complement to the Fourier transform inversion formula.

Lemma 7.3.6 Let Ω be a bounded open subset of R𝑛 with a smooth boundary

𝜕Ω and set ΩC = {𝑤 = 𝑢 + 𝑖𝑣 ∈ C𝑛 ; 𝑢 ∈ Ω, |𝑣| < dist (𝑢, 𝜕Ω)}. Then, given any
ℎ ∈ O (C𝑛 ) and any 𝑤 ∈ ΩC , we have
∫ ∫
2
−𝑛
ℎ (𝑤) = lim (2𝜋) e𝑖 𝜉 · ( 𝑥−𝑤)−𝜀 | 𝜉 | ℎ (𝑥) d𝑥d𝜉. (7.3.13)
𝜀↘0 𝜉 ∈R𝑛 𝑥 ∈Ω

(Limit in the sense of uniform convergence on every compact subset of ΩC .)

1
Proof We have, for 𝑧 ∈ C𝑛 and 𝜏 = 4𝜀 ,
∫ ∫
2 2 1 2
e𝑖 𝜉 ·𝑧−𝜀 | 𝜉 | d𝜉 = e−𝜏 ⟨𝑧 ⟩ e− 4𝜏 ⟨ 𝜉 −2𝑖 𝜏𝑧 ⟩ d𝜉
R 𝑛 R 𝑛

√ 𝑛 −𝜏 ⟨𝑧 ⟩2
∫
2 1 2
= 2 𝜏 e e−| 𝜉 | d𝜉 = (4𝜋𝜏) 2 𝑛 e−𝜏 ⟨𝑧 ⟩ .
R𝑛

This shows that (7.3.13) is equivalent to

𝜏 12 𝑛 ∫ 2
ℎ (𝑤) = lim e−𝜏 ⟨𝑥−𝑤 ⟩ ℎ (𝑥) d𝑥.
𝜏→+∞ 𝜋 Ω

We apply Stokes’ Theorem:

∫ ∫
2 2
e−𝜏 ⟨𝑥−𝑤 ⟩ ℎ (𝑥) d𝑥 = e−𝜏 | 𝑥−𝑢 | ℎ (𝑥 + 𝑖𝑣) d𝑥
Ω Ω
∫ ∫ 𝑡=1
2
+ e−𝜏 ⟨𝑥−𝑢−𝑖 (1−𝑡) 𝑣 ⟩ ℎ (𝑥 + 𝑖𝑡𝑣) d𝜎 (𝑥) d𝑡
𝑥 ∈𝜕Ω 𝑡=0

where d𝜎 (𝑥) is the appropriate measure on 𝜕Ω. If 𝑥 ∈ 𝜕Ω and 𝑢 + 𝑖𝑣 ∈ ΩC we have

Re ⟨𝑢 − 𝑥 + 𝑖 (1 − 𝑡) 𝑣⟩ 2 = |𝑢 − 𝑥| 2 − (1 − 𝑡) |𝑣| 2 > 0
7.3 The FBI Transform of Analytic Functionals 189

and 𝑛 𝑛/2 −𝑛/2

2
1
−(1−𝑡) |𝑣 | 2 )
𝜏 2 𝑛 e−𝜏 ( |𝑢−𝑥 | |𝑢 − 𝑥| 2 − (1 − 𝑡) |𝑣| 2
≤ .
2e
The Lebesgue Dominated Convergence Theorem implies directly
∫ ∫ 𝑡=1
1 2
lim 𝜏 2𝑛 e−𝜏 (𝑢−𝑥+𝑖 (1−𝑡) 𝑣) ℎ (𝑥 + 𝑖𝑡𝑣) d𝜎 (𝑥) d𝑡 = 0.
𝜏→+∞ 𝑥 ∈𝜕Ω 𝑡=0

On the other hand,

𝜏 12 𝑛 ∫ 2
lim e−𝜏 |𝑢−𝑥 | ℎ (𝑥 + 𝑖𝑣) d𝑥 = ℎ (𝑢 + 𝑖𝑣) . □
𝜏→+∞ 𝜋 Ω

Theorem 7.3.7 We have, for every analytic functional 𝜇 ∈ O ′ (R𝑛 ) and every func-
tion ℎ ∈ O (C𝑛 ),
∫ ∫
2
−𝑛
⟨𝜇, ℎ⟩ = lim (2𝜋) ℎ (𝑥) F𝜅 𝜇(𝑥, 𝜉)e−𝜀 | 𝜉 | d𝑥d𝜉, (7.3.14)
𝜀↘0 𝜉 ∈R𝑛 𝑥 ∈Ω

where Ω is a bounded open subset of R𝑛 containing supp 𝜇.

Proof Let 𝐾 = supp 𝜇. We begin by proving the claim under the hypothesis that
diam 𝐾 is suitably small. Let Ω be a bounded open subset of R𝑛 such that 𝐾 ⊂ Ω and
ΩC an open subset of C𝑛 such that ΩC ∩ R𝑛 = Ω and |Im 𝑧| < 𝑐 for every 𝑧 ∈ ΩC ,
with 𝑐 > 0 as small as need be. In (7.3.13) we deform the domain of 𝜉-integration
from R𝑛 to the image of R𝑛 under the map 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜅 (𝑥 − 𝑤) |𝜉 |:
∫ ∫
2
𝑛
(2𝜋) ℎ(𝑤) = lim e𝑖 𝜉 · ( 𝑥−𝑤)−𝜀 | 𝜉 | ℎ (𝑥) d𝑥d𝜉 (7.3.15)
𝜀↘0 𝜉 ∈R𝑛 𝑥 ∈Ω
∫ ∫
2 2
= lim e𝑖 𝜉 · ( 𝑥−𝑤)−𝜅 | 𝜉 | ⟨𝑥−𝑤 ⟩ −𝜀 ⟨𝜁 ⟩ ℎ (𝑥) Δ 𝜅 (𝑥 − 𝑤, 𝜉)d𝑥d𝜉.
𝜀↘0 𝜉 ∈R𝑛 𝑥 ∈Ω

where Δ 𝜅 (𝑥 − 𝑤, 𝜉) = 1 + 𝑖𝜅 | 𝜉𝜉 | · (𝑥 − 𝑤) (cf. Lemma 3.4.1) and 𝑤 ∈ ΩC . We

deform the domain of 𝑥 integration from Ω to the image Ω𝑡◦ (𝜉) of Ω under the map
𝑥 ↦→ 𝑧 = 𝑥 + 𝑖𝑡◦ | 𝜉𝜉 | , 𝑡◦ > 0; Stokes’ Theorem implies
∫
2
−𝜀 ⟨𝜁 ⟩ 2
e𝑖 𝜉 · ( 𝑥−𝑤)−𝜅 | 𝜉 | ⟨𝑥−𝑤 ⟩ ℎ (𝑥) Δ 𝜅 (𝑥 − 𝑤, 𝜉)d𝑥 (7.3.16)
𝑥 ∈Ω
∫
2
−𝜀 ⟨𝜁 ⟩ 2
= e𝑖 𝜉 · (𝑧−𝑤)−𝜅 | 𝜉 | ⟨𝑧−𝑤 ⟩ ℎ (𝑧) Δ 𝜅 (𝑧 − 𝑤, 𝜉)d𝑧
𝑧 ∈Ω𝑡◦ ( 𝜉 )
∫ 𝑡◦ ∫
2
−𝜀 ⟨𝜁 ⟩ 2
+ e𝑖 𝜉 · (𝑧−𝑤)−𝜅 | 𝜉 | ⟨𝑧−𝑤 ⟩ ℎ (𝑥) Δ 𝜅 (𝑧 − 𝑤, 𝜉)d𝑡d𝜎 (𝑥) ,
0 𝑧 ∈𝜕Ω𝑡 ( 𝜉 )

where 𝜕Ω𝑡 (𝜉) = 𝜕Ω + 𝑖𝑡 | 𝜉𝜉 | and d𝜎 (𝑥) is the appropriate measure on the boundary
𝜕Ω. Keep in mind that, in the right-hand side of (7.3.16),
190 7 Hyperfunctions in Euclidean Space

𝜁 = (1 − 𝜅𝑡)𝜉 + 𝑖𝜅 (𝑥 − 𝑤) |𝜉 | .

Let 𝑉 C be a bounded open subset of C𝑛 such that 𝐾 ⊂ 𝑉 = 𝑉 C ∩ R𝑛 ⊂⊂ Ω and let

𝑤 = 𝑢 + 𝑖𝑣 ∈ 𝑉 C be arbitrary. We focus on the exponent

𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡) = 𝑖𝜉 · (𝑥 − 𝑤) − 𝑡 |𝜉 |
2
𝜉
− 𝜅 𝑥 − 𝑤 + 𝑖𝑡 |𝜉 | − 𝜀 ⟨(1 − 𝜅𝑡) 𝜉 + 𝑖𝜅 |𝜉 | (𝑥 − 𝑤)⟩ 2
|𝜉 |

and its real part,

2
𝜉
Re 𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡) = −𝜅 |𝑥 − 𝑢| 2 |𝜉 | + 𝜉 · 𝑣 − 𝑡 |𝜉 | + 𝜅 𝑡 − 𝑣 |𝜉 |
|𝜉 |
!
2
𝜉
− 𝜀 (1 − 𝜅𝑡) + 𝜅𝑣 − 𝜅 2 |𝑥 − 𝑢| 2 |𝜉 | 2 .
|𝜉 |

First of all, we select 𝑡 ◦ and 𝛿 = sup |𝑣| sufficiently small to ensure that 𝑡 ≤ 𝑡◦
𝑤 ∈𝑉 C
implies
2
𝜉 1
𝜅 𝑡 − 𝑣 ≤ 𝑡 + 2𝜅𝛿2 .
|𝜉 | 2
Next, we require 𝜅 diam 𝐾 to be sufficiently small that, for all 𝑥 ∈ Ω, 𝑤 ∈ 𝑉 C and
𝑡 ∈ [0, 𝑡◦ ], we have
2
𝜉
(1 − 𝜅𝑡) + 𝜅𝑣 − 𝜅 2 |𝑥 − 𝑢| 2 > 0.
|𝜉 |

We derive
1
|𝜉 | −1 Re 𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡) ≤ −𝜅 |𝑥 − 𝑢| 2 − 𝑡 + (2𝜅𝛿 + 1) 𝛿. (7.3.17)
2

We require 𝛿 < 1
4 (2𝜅 + 1) −1 𝑡◦ ; in the first integral on the right-hand side of (7.3.16)
we get
1
Re 𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡 ◦ ) ≤ − 𝑡◦ |𝜉 | .
4
In the second integral in the right-hand side of (7.3.16) 𝑥 ∈ 𝜕Ω implies |𝑥 − 𝑢| ≥
𝑑 = dist (𝑉, 𝜕Ω) > 0; we derive from (7.3.17):

|𝜉 | −1 Re 𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡) ≤ (2𝜅 + 1) 𝛿 − 𝜅𝑑 2 .
𝜅 𝑑2
Here we require 𝛿 < 2(2𝜅+1) , which implies

1
Re 𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡) ≤ − 𝜅𝑑 2 |𝜉 | .
2
7.3 The FBI Transform of Analytic Functionals 191

In all cases we have

Re 𝜑 𝜀 (𝑥, 𝑤, 𝜉, 𝑡) ≤ −𝑐 ◦ |𝜉 |
for some 𝑐 ◦ > 0. In both integrands in the right-hand side of (7.3.16) we obtain
2
| 𝜉 |−𝜀 ⟨𝜁 ⟩ 2
e𝑖 𝜉 · (𝑧−𝑤)−𝜅 ⟨𝑧−𝑤 ⟩ ℎ (𝑧) Δ 𝜅 (𝑧 − 𝑤, 𝜉) ≤ 𝐶◦ e−𝑐◦ | 𝜉 |

for some constant 𝐶◦ > 0 depending only on Ω and ℎ. This allows us to apply
the Lebesgue Dominated Convergence Theorem to (7.3.15) and let 𝜀 go to zero, to
derive

(2𝜋) 𝑛 ℎ(𝑤) (7.3.18)

∫ ∫
2
= e𝑖 𝜉 · ( 𝑥−𝑤)−𝜅 | 𝜉 | ⟨𝑥−𝑤 ⟩ ℎ (𝑥) Δ 𝜅 (𝑥 − 𝑤, 𝜉)d𝑥d𝜉.
𝜉 ∈R𝑛 𝑥 ∈Ω

Since supp 𝜇 ⊂ 𝑉 we can let 𝜇 act (in 𝑤-space) under the integral sign, thus proving
(7.3.14) under the hypothesis that 𝜅 diam supp 𝜇 is sufficiently small.
To prove the claim in the general case we apply Theorems 6.3.6, 6.3.11: if 𝜅 > 0
and 𝜇 ∈ O ′ (R𝑛 ) are arbitrary we decompose 𝜇 as a finite sum 𝜇 = 𝑟𝑗=1 𝜇 𝑗 with
Í
𝜇 𝑗 ∈ O ′ (R𝑛 ) such that 𝜅 diam supp 𝜇 𝑗 is sufficiently small. Formula (7.3.14) is
obviously additive in the argument 𝜇. □

Corollary 7.3.8 If 𝜇 ∈ E ′ (R𝑛 ) then, for every 𝜑 ∈ Cc∞ (R𝑛 ),

∫
−𝑛
⟨𝜇, 𝜑⟩ = (2𝜋) 𝜑 (𝑥) F𝜅 𝜇(𝑥, 𝜉)d𝑥d𝜉. (7.3.19)
R2𝑛

Proof Assume that 𝐾 = supp 𝜇 ⊂ Ω, a bounded open subset of R𝑛 ; it suffices to

prove (7.3.19) for every 𝜑 ∈ Cc∞ (Ω); such a 𝜑 is the limit in C ∞ (Ω) of a sequence of
entire functions in C𝑛 , by the Weierstrass Approximation Theorem, and thus (7.3.19)
is a direct consequence of (7.3.14). □

Corollary 7.3.9 Let 𝑓 be a compactly supported 𝐿 1 function in R𝑛 ; we have

∫
𝑓 (𝑥) = (2𝜋) −𝑛 F𝜅 𝑓 (𝑥, 𝜉)d𝜉 a.e.
R𝑛

Proof Follows directly from Corollary 7.3.8, keeping in mind that

∫
′ ′ 2
F𝜅 𝑓 (𝑥, 𝜉) = e𝑖 𝜉 · ( 𝑥−𝑥 )−𝜅 | 𝜉 | | 𝑥−𝑥 | 𝑓 (𝑥 ′) Δ 𝜅 (𝑥 − 𝑥 ′, 𝜉)d𝑥 ′. (7.3.20)
R𝑛

□
192 7 Hyperfunctions in Euclidean Space

7.4 Analytic Wave-front Set of a Hyperfunction

7.4.1 FBI transform and local analyticity

The next result is the analogue for hyperfunctions of Theorem 3.4.7.

Theorem 7.4.1 Let 𝑓 be a hyperfunction in R𝑛 . For 𝑓 to be an analytic function

in a neighborhood of a point 𝑥 ◦ ∈ R𝑛 it is necessary and sufficient that, given any
bounded open subset 𝑈 of R𝑛 , 𝑥 ◦ ∈ 𝑈, and any analytic functional 𝜇 ∈ O ′ 𝑈
representing 𝑓 in 𝑈 the following condition be satisfied (cf. Lemma 3.4.5):
(EXP DECAY 2) There exist an open subset 𝑉 C of C𝑛 , 𝑉 = 𝑉 C ∩ R𝑛 ⊂ 𝑈, 𝑥 ◦ ∈ 𝑉,
and positive constants 𝛾, 𝑅, such that

∀ (𝑧, 𝜁) ∈ 𝑉 C × Γ𝛾(𝑛) , |F𝜅 𝜇(𝑧, 𝜁)| ≲ e− |𝜁 |/𝑅 . (7.4.1)

The cone Γ𝛾(𝑛) is defined in (7.3.1).

Proof It follows directly from Corollary 7.3.5 that Condition (EXP DECAY 2) is
′
independent of the choice of analytic functional 𝜇 ∈ O 𝑈 representing 𝑓 in 𝑈.
I. Necessity of the condition. There is no loss in assuming that 𝑓 |𝑈 ∈ C 𝜔 (𝑈) ∩
∞
𝐿 (𝑈) and, as a matter of fact, that 𝑓 vanishes identically in R𝑛 \𝑈. Let then 𝑉 ⊂⊂ 𝑈
be an open set (in R𝑛 ) such that 𝑥 ◦ ∈ 𝑉; and let 𝜓 ∈ Cc∞ (𝑈), 0 ≤ 𝜓 ≤ 1, 𝜓 (𝑥) = 1
for all 𝑥 ∈ 𝑉. We select 𝛿 > 0 sufficiently small that 𝑓 extends as a holomorphic
function in the set 𝑉 C = {𝑧 ∈ C𝑛 ; dist(𝑧, supp 𝜓) < 𝛿}. In the integral on the right
in (7.3.20) we can deform the domain of integration from R𝑛 to the image of R𝑛
𝜉
under the map 𝑥 ′ → 𝑧 ′ = 𝑥 ′ − 𝑖𝛿𝜓 (𝑥 ′) . We look at the real part of the exponent
|𝜉 |
in the integrand, after this deformation. Provided 𝛿 is sufficiently small and 𝑧 = 𝑥 +𝑖𝑦
stays in a sufficiently small neighborhood of 𝑥 ◦ we get, for every 𝑥 ′ ∈ 𝑈,

|𝜉 | −1 Re 𝑖𝜉 · (𝑧 − 𝑧 ′) − 𝜅 |𝜉 | ⟨𝑧 − 𝑧 ′⟩ 2

= −𝛿𝜓 (𝑥 ′) + 𝜅 |𝑦 − 𝛿𝜓 (𝑥 ′)| 2 − 𝜅 |𝑥 − 𝑥 ′ | 2
1
≤ − 𝛿𝜓 (𝑥 ′) − 𝜅 |𝑥 − 𝑥 ′ | 2 + 𝜅 |𝑦| 2 ≤ −𝑐 ◦ < 0.
2
Property (EXP DECAY 2) ensues directly.
II. Sufficiency of the condition. Suppose (EXP DECAY 2) holds; it suffices to
prove that the restriction of 𝑓 to the neighborhood 𝑉 of 𝑥 ◦ is analytic in some, possibly
smaller, neighborhood of 𝑥 ◦ . From Theorem 7.3.7 we deduce that the function equal
to ∫
(2𝜋) −𝑛 F𝜅 𝜇(𝑥, 𝜉)d𝜉 (7.4.2)
𝜉 ∈R𝑛
7.4 Analytic Wave-front Set of a Hyperfunction 193

in 𝑉 and to zero in R𝑛 \𝑉 defines an analytic functional that represents the hyper-

function 𝑓 in 𝑉. It follows directly from (7.4.1) that (7.4.2) extends as a holomorphic
function in the open subset 𝑉 C of C𝑛 . □

7.4.2 Hyperfunctions as sums of boundary values of holomorphic

functions

In this subsection we answer the question as to whether every hyperfunction in a

bounded open set can be represented as a sum of boundary values of holomorphic
functions in wedges with edge that open set. The wedges W𝛿 (𝑈, Γ) used below are
defined according to (3.3.1). We shall make use of the following two lemmas.

Lemma 7.4.2 Let ℭ be an acute open cone and Γ a convex open cone in R𝑛 \ {0}
such that 𝑦 · 𝜉 > 𝑐 ◦ |𝑦| |𝜉 | for some 𝑐 ◦ > 0, all 𝜉 ∈ ℭ, 𝑦 ∈ Γ. Let 𝜇 ∈ O ′ (R𝑛 ), a
bounded open subset Ω of R𝑛 and 𝛿 > 0 be arbitrary. If 𝜅𝛿 is sufficiently small then,
as 𝜀 ↘ 0, the entire function of 𝑧 ∈ C𝑛 ,
𝜅 𝑛2 ∫ ∫
𝑛 ′ ′ 2 2

3
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉, (7.4.3)
2𝜋 𝑥 ′ ∈Ω 𝜉 ∈ℭ

converges (uniformly on compact subsets) to a holomorphic function in W𝛿 (R𝑛 , Γ).

Proposition 7.3.4 implies that the integral in (7.4.3) converges. That the cone ℭ
is acute means that its closure in R𝑛 \ {0} is contained in an open half-space.
Proof As in the proofs of Lemma 7.3.6 and Theorem 7.3.7 we focus on the exponent
or rather, here, on the quantity

𝜉 ′ 2 𝜉
𝑄 = Re 𝑖 ′
· (𝑧 − 𝑥 ) − 𝜅 ⟨𝑧 − 𝑥 ⟩ = − · 𝑦 − 𝜅 |𝑥 − 𝑥 ′ | 2 − |𝑦| 2 . (7.4.4)
|𝜉 | |𝜉 |

For arbitrary 𝑥 ∈ R𝑛 and 𝑦 ∈ Γ we have

𝑄 ≤ − |𝑦| (𝑐 ◦ − 𝜅 |𝑦|) − 𝜅 |𝑥 − 𝑥 ′ | 2 . (7.4.5)

We require 2𝜅 ≤ 𝑐 ◦ 𝛿−1 whence |𝑦| < 𝛿 ≤ 21 𝑐 ◦ 𝜅 −1 , whence 𝑄 ≤ − 21 𝑐 ◦ |𝑦|−𝜅 |𝑥 − 𝑥 ′ | 2 .

We avail ourselves of (7.3.9) where we take 𝑧 = 𝑥 ′ ∈ R𝑛 , 𝜁 = 𝜉 ∈ R𝑛 and we replace
𝜀 by 𝜀 ′ > 0, and of (7.3.10):
′ ′ 2 2
e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉)

1
≤ 𝐶 𝜀′ exp 𝜀 ′ − 𝑐 ◦ |𝑦| |𝜉 | − 𝜅 |𝑥 − 𝑥 ′ | 2 |𝜉 | − 𝜀 |𝜉 | 2 .
2

It is permitted to select 𝜀 ′ as small as needed; here we take 𝜀 ′ < 41 𝑐 ◦ |𝑦|, whence

194 7 Hyperfunctions in Euclidean Space
′ ′ 2 2
e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉)

1
≤ 𝐶 𝜀′ exp − 𝑐 ◦ |𝑦| |𝜉 | − 𝜅 |𝑥 − 𝑥 ′ | 2 |𝜉 | − 𝜀 |𝜉 | 2 .
4

It follows (by the Lebesgue Dominated Convergence Theorem) that we can let 𝜀 go
to zero in (7.4.3), proving the claim. □

Lemma 7.4.3 Let 𝐾 ⊂ R𝑛 be a compact set and 𝜇 ∈ O ′ (𝐾). Consider the entire
function of 𝑧,
𝜅 𝑛2 ∫ 𝑛 ′ ′ 2 2
𝜀
ℎ (𝑧) = 3
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉.
2𝜋 R2𝑛

Let Ω be a bounded open subset of R𝑛 , 𝐾 ⊂ Ω. For arbitrary 𝜑 ∈ O (C𝑛 ) we have

∫
⟨𝜇, 𝜑⟩ = lim ℎ 𝜀 (𝑥) 𝜑 (𝑥) d𝑥.
𝜀↘0 Ω

Proof According to Definition 7.3.2 we have

𝑛2
2𝜋 3

ℎ 𝜀 (𝑥)
𝜅
∫
𝑛
𝑖 𝜉 · ( 𝑥−𝑤)−𝜅 | 𝜉 | ( ⟨𝑥 ′ −𝑤 ⟩ 2 +|𝑥−𝑥 ′ | 2 ) −𝜀 | 𝜉 | 2 ′ ′
= 𝜇𝑤 , |𝜉 | e
2 Δ 𝜅 (𝑥 − 𝑤, 𝜉)d𝑥 d𝜉 .
R2𝑛

We carry out the change of variables 𝑥 ′ = 𝑥 ′′ + 𝑥+𝑤

2 :

𝑛2
2𝜋 3

ℎ 𝜀 (𝑥)
𝜅
∫
𝑛
𝑖 𝜉 · ( 𝑥−𝑤)−𝜅 | 𝜉 | ( 21 ⟨𝑥−𝑤 ⟩ 2 +2| 𝑥 ′′ | 2 ) −𝜀 | 𝜉 | 2 ′′ 𝑥−𝑤 ′′
= 𝜇𝑤 , |𝜉 | e
2 Δ 𝜅 (𝑥 + , 𝜉)d𝑥 d𝜉 .
R2𝑛 2

We observe that

𝑥−𝑤 ′′ 𝑥−𝑤 ′′ 𝜉
Δ 𝜅 (𝑥 + , 𝜉) = Δ 𝜅 ( , 𝜉) + 𝑃 𝑥 − 𝑤, 𝑥 ,
2 2 |𝜉 |

where 𝑃 𝑥 − 𝑤, 𝑥 ′′, | 𝜉𝜉 | is a polynomial in which each monomial is linear with
respect to some 𝑥 ′′𝑗 . Since
∫ +∞ 2
−𝜅 | 𝜉 | 𝑥 ′′𝑗
e 𝑥 ′′𝑗 d𝑥 ′′𝑗 =0
−∞

1
and Δ 𝜅 2 𝑧, 𝜉 = Δ 𝜅/2 (𝑧, 𝜉) we get
7.4 Analytic Wave-front Set of a Hyperfunction 195
∫ 𝑛2
′′ | 2 𝑥−𝑤 𝜋
e−2𝜅 | 𝜉 | | 𝑥 Δ 𝜅 (𝑥 ′′ + , 𝜉)d𝑥 ′′ = Δ 𝜅/2 (𝑥 − 𝑤, 𝜉)
R𝑛 2 2𝜅 |𝜉 |

and, as a consequence,
∫ D E
1 2 2
−𝑛
𝜀
ℎ (𝑥) = (2𝜋) 𝜇 𝑤 , e𝑖 𝜉 · ( 𝑥−𝑤)− 2 𝜅 | 𝜉 | ⟨𝑥−𝑤 ⟩ Δ 𝜅/2 (𝑥 − 𝑤, 𝜉) e−𝜀 | 𝜉 | d𝜉
𝑛
∫R
2
−𝑛
= (2𝜋) F𝜅/2 𝜇 (𝑥, 𝜉) e−𝜀 | 𝜉 | d𝜉.
R𝑛

For fixed 𝜀 > 0 we have ℎ 𝜀 ∈ O (C𝑛 ) and we can form, for arbitrary 𝜑 ∈ O (C𝑛 ),
∫ ∫ ∫
2
−𝑛
𝜀
ℎ (𝑥) 𝜑 (𝑥) d𝑥 = (2𝜋) 𝜑 (𝑥) F𝜅/2 𝜇 (𝑥, 𝜉) e−𝜀 | 𝜉 | d𝑥d𝜉.
Ω 𝑥 ∈Ω 𝜉 ∈R𝑛

The claim follows directly from this and from (7.3.14). □

We can now prove the hyperfunction version of Theorem 3.3.12; its precise
statement is of crucial importance to hyperfunction theory as approached in this
text, by way of analytic functionals and the FBI transform.

Theorem 7.4.4 Let 𝑈 be a bounded open subset of R𝑛 and ℭ 𝑗 ( 𝑗 = 1, ..., 𝜈 < +∞)
be pairwise disjoint acute open cones in R𝑛 \ {0} such that the Lebesgue measure of
R𝑛 \ (ℭ1 ∪ · · · ∪ ℭ 𝜈 ) is equal to zero. Let 𝑐 ◦ > 0 be sufficiently small that the convex
open cone
Γ 𝑗 = 𝑦 ∈ R𝑛 ; ∀𝜉 ∈ ℭ 𝑗 , 𝑦 · 𝜉 > 𝑐 ◦ |𝑦| |𝜉 | (7.4.6)

𝜈. Given arbitrary 𝑓 ∈ B (R ) and 𝛿 > 0, there are

is not empty whatever 𝑗 = 1, ..., 𝑛

functions ℎ 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 such that (cf. Definition 7.2.3)

𝜈
∑︁
𝑓 |𝑈 = 𝑏𝑈 ℎ 𝑗 . (7.4.7)
𝑗=1

Proof Let 𝜇 ∈ O ′ 𝑈 represent the hyperfunction 𝑓 in 𝑈 and Ω′ be a bounded
domain in R𝑛 , Ω′ ⊃ 𝑈. We introduce the entire function of 𝑧,
𝜅 𝑛2 ∫ ∫
𝑛 ′ ′ 2 2
𝜀
ℎ∞ (𝑧) = 3
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉.
2𝜋 𝑥 ′ ∈R𝑛 \Ω′ 𝜉 ∈R𝑛

From (ii), Proposition 7.3.4, we derive

′ 𝑛 ′ 1 ′ ′ 2
∀𝑥 ∈ R \Ω , |F𝜅 𝜇 (𝑥 , 𝜉)| ≲ exp − 𝜅 dist (𝑥 , 𝑈) |𝜉 | .
4
196 7 Hyperfunctions in Euclidean Space

This implies
𝜀
ℎ∞ (𝑧) ≲
∫ ∫
1
|𝜉 | 2 exp |𝜉 | |𝑦| + 𝜅 |𝑦| 2 − 𝜅 dist (𝑥 ′, 𝑈) 2 − 𝜀 |𝜉 | 2 d𝑥 ′d𝜉.
𝑛
𝜅 𝑛/2
𝑥 ′ ∈R𝑛 \Ω′ 𝜉 ∈R𝑛 4

We select Ω′ such that 𝜅𝛿2 + 𝛿 ≤ 81 𝜅 dist (𝑈, R𝑛 \Ω′) 2 ; as 𝜀 ↘ 0, ℎ∞ 𝜀 converges

uniformly in the “horizontal” tube {𝑧 ∈ C ; |Im 𝑧| < 𝛿} to the holomorphic function

𝑛

𝜅 𝑛2 ∫ ∫
𝑛 ′ ′ 2
ℎ∞ (𝑧) = 3
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉.
2𝜋 𝑥 ′ ∈R𝑛 \Ω′ 𝜉 ∈R𝑛

If |Im 𝑧| < 𝛿 we have

𝜅 𝑛2 ∫ ∫
1 ′ 2
2 e− 8 𝜅 dist ( 𝑥 ,𝑈 ) | 𝜉 | d𝑥 ′ d𝜉.
𝑛
|ℎ∞ (𝑧)| ≲ |𝜉 |
2𝜋 3 𝑥 ′ ∈R𝑛 \Ω′ 𝜉 ∈R𝑛

The entire function in Lemma 7.4.3 can be decomposed as follows:

𝜈
∑︁
ℎ 𝜀 = ℎ∞
𝜀
+ ℎ 𝜀𝑗 (7.4.8)
𝑗=1

where
𝜅 𝑛2 ∫ ∫
𝑛 ′ ′ 2 2
ℎ 𝜀𝑗 (𝑧) = |𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉.
2𝜋 3 𝑥 ′ ∈Ω′ 𝜉 ∈ℭ 𝑗

By Lemma 7.4.2, if 𝜅𝛿 is sufficiently small then, as 𝜀 ↘ 0, each entire function ℎ 𝜀𝑗

converges normally to a holomorphic function ℎ 𝑗 in W𝛿 R𝑛 , Γ 𝑗 .
Let Ω be a bounded open subset of R𝑛 with a C ∞ boundary and such that 𝑈 ⊂ Ω;
Lemma 7.4.3 shows that, for arbitrary 𝜑 ∈ O (C𝑛 ),
∫
⟨𝜇, 𝜑⟩ = lim ℎ 𝜀 (𝑥) 𝜑 (𝑥) d𝑥. (7.4.9)
𝜀↘0 Ω

For each 𝑗 = 1, ..., 𝜈, let 𝑦 ( 𝑗) ∈ Γ 𝑗 , 𝑦 ( 𝑗) < 𝛿. Stokes’ theorem yields

∫ ∫
lim ℎ 𝜀𝑗 (𝑥) 𝜑 (𝑥) d𝑥 = ℎ 𝑗 𝑥 + 𝑖𝑦 ( 𝑗) 𝜑 𝑥 + 𝑖𝑦 ( 𝑗) d𝑥 (7.4.10)
𝜀↘0 Ω Ω
∫ ∫ 𝑡=1
− ℎ 𝑗 𝑥 + 𝑖𝑡𝑦 ( 𝑗) 𝜑 𝑥 + 𝑖𝑡𝑦 ( 𝑗) d 𝑥 + 𝑖𝑡𝑦 ( 𝑗) .
𝑥 ∈𝜕Ω 𝑡=0

As Γ ′𝑗 ∋ 𝑦 ( 𝑗) → 0 (Γ ′𝑗 : a cone such that Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 ) the first integral in

D E
the right-hand side of (7.4.10) converges to 𝜇Ω ℎ𝑗 , 𝜑 (cf. Lemma 7.2.2) where the
analytic functional 𝜇Ω
ℎ 𝑗 is carried by Ω and represents the hyperfunction 𝑏 Ω ℎ 𝑗 in Ω.
7.4 Analytic Wave-front Set of a Hyperfunction 197
D E
The second integral in the right-hand side of (7.4.10) converges to 𝜇 𝜕Ω
ℎ 𝑗 , 𝜑 , where
the analytic functional 𝜇 𝜕Ω
ℎ 𝑗 is carried by 𝜕Ω. Combining this with (7.4.8) and (7.4.9)
yields
∫ 𝜈 D
∑︁ E
⟨𝜇, 𝜑⟩ = ℎ∞ (𝑥) 𝜑 (𝑥) d𝑥 + 𝜇Ω
ℎ𝑗 + 𝜇 𝜕Ω
ℎ𝑗 , 𝜑 ,
Ω 𝑗=1

which implies
𝜈
∑︁
𝑓 |𝑈 = 𝑏𝑈 ℎ 𝑗 + ℎ ∞ | 𝑈 ;
𝑗=1

(7.4.7) ensues if we replace ℎ1 by ℎ1 − ℎ∞ . □

Remark 7.4.5 The case 𝑛 = 1 is especially simple: given arbitrary 𝑓 ∈ B (R) and
𝛿 > 0, if −∞ < 𝑎 < 𝑏 < +∞ there are functions ℎ± ∈ O ((𝑎, 𝑏) × (0, ±𝛿)) such that
𝑓 | 𝑈 = 𝑏𝑈 ℎ + + 𝑏𝑈 ℎ − .

Remark 7.4.6 We must stress the following two facts:

(1) the choice of cones Γ 𝑗 is independent of the hyperfunction 𝑓 ;
(2) 𝛿 can be taken arbitrarily large. Naturally, if we modify 𝛿 the ℎ 𝑗 may also need
to be modified.

Let 𝑈 be a bounded open subset of R𝑛 andÍ𝛿 > 0, both arbitrary. We shall denote
by E 𝛿 (𝑈) the vector space of finite sums 𝜍 = 𝜈𝑗=1 ℎ 𝑗 (𝜈 can vary) such that, for each

𝑗 = 1.., 𝜈, ℎ 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 with Γ 𝑗 a convex open cone in R𝑛 \ {0}. If 𝑉 ⊂ 𝑈
′
is open we have a natural restriction map 𝜌𝑈 𝑉 : E 𝛿 (𝑈) → E 𝛿 (𝑉); if 0 < 𝛿 < 𝛿
the restriction maps O (W𝛿 (𝑈, Γ)) −→ O (W𝛿′ (𝑈, Γ)) lead to a restriction map
𝒓 𝛿𝛿′ : E 𝛿 (𝑈) → E 𝛿′ (𝑈). Note that each one of these restriction mappings is
injective; we can view E 𝛿 (𝑈) as a vector subspace of E 𝛿′ (𝑈). We also remind the
reader that each boundary value map O W𝛿′ 𝑈, Γ 𝑗 −→ B (𝑈) (Definition 7.2.3)
is injective (Proposition 7.2.5). We can form the direct limit E (𝑈) of the vector
spaces E 𝛿 (𝑈), 0 < 𝛿 < +∞, namely the quotient of the disjoint union
•

+
𝛿>0
E 𝛿 (𝑈)

for the relation 𝒓 𝛿𝛿′ 𝜍 = 𝜍 ′ ∈ E 𝛿′ (𝑈), 0 < 𝛿 ′ < 𝛿, 𝜍 ∈ E 𝛿 (𝑈). We have a natural
boundary value map 𝒃 E(𝑈) : E (𝑈) −→ B (𝑈). A restatement of Theorem 7.4.4 is
that the map 𝒃 E(𝑈) is surjective.
Before studying the nullspace ker 𝒃 E(𝑈) we relate the exponential decay, in certain
frequency directions, of the FBI transform of an analytic functional 𝜇 ∈ O ′ (R𝑛 ) to
the representation of the hyperfunction defined by 𝜇 as a sum of boundary values of
holomorphic functions in special wedges.
198 7 Hyperfunctions in Euclidean Space

7.4.3 Analytic wave-front set of a hyperfunction

Definition 7.4.7 We shall say that a hyperfunction 𝑓 in an open subset Ω of R𝑛 is

microanalytic at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0}) if there exist a neighborhood 𝑈
of 𝑥 ◦ in Ω and finitely many convex open cones Γ 𝑗 ⊂ R𝑛 \ {0} ( 𝑗 = 1, ..., 𝜈) with the
following properties:
𝜈
Ø
(1) 𝜉 ◦ · 𝑦 < 0 for all 𝑦 ∈ Γ𝑗;
𝑗=1
Í𝜈
(2) there are functions ℎ 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 (𝛿 > 0) such that 𝑓 = 𝑗=1 𝑏𝑈 ℎ 𝑗 in 𝑈.
The set of points (𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}) at which 𝑓 is not microanalytic shall be
called the analytic wave-front set of 𝑓 and denoted by 𝑊 𝐹a ( 𝑓 ); the base projection
of 𝑊 𝐹a ( 𝑓 ) shall be called the analytic singular support of 𝑓 and denoted by
singsuppa 𝑓 .

We shall say that 𝑓 ∈ B (Ω) is microanalytic in a subset U of Ω × (R𝑛 \ {0})

if it is microanalytic at every point of U; then this will also be true in some conic
open set containing U. Indeed, we can assume (cf. Remark 7.2.4) in Property (1),
Definition 7.4.7, that the cones Γ 𝑗 are such that 𝜉 ◦ · 𝑦 < −𝑐 |𝑦| for some 𝑐 > 0 and
Ø𝜈
all 𝑦 ∈ Γ 𝑗 . Thus 𝑊 𝐹a ( 𝑓 ) is a closed conic subset of Ω × (R𝑛 \ {0}).
𝑗=1

Remark 7.4.8 Suppose 𝑛 = 1. In the notation of Remark 7.4.5 (𝑥 ◦ , −1) ∉ 𝑊 𝐹a ( 𝑓 )

if 𝑓 = 𝑏𝑈 𝜑+ whereas 𝑓 is analytic at 𝑥 ◦ if and only if either 𝜑+ or 𝜑− extend
holomorphically to a neighborhood of 𝑥 ◦ in C.

There are a variety of other names for analytic wave-front set: essential support,
analytic singular spectrum, microsupport (in this book, the latter will be used in
a slightly different sense, see Definition 7.6.1 below). If 𝑓 ∈ B (Ω) and if 𝑥 ◦ ∈ R𝑛
we denote by 𝑊 𝐹a ( 𝑓 )| 𝑥 ◦ the set of points (𝑥 ◦ , 𝜉) ∈ Ω × (R𝑛 \ {0}) that belong to
𝑊 𝐹a ( 𝑓 ).
We shall denote by Γ◦ the polar of a cone Γ:

Γ◦ = {𝜉 ∈ R𝑛 \ {0} ; ∀𝑥 ∈ Γ, 𝜉 · 𝑥 ≥ 0} ; (7.4.11)

Γ◦ is always a convex cone, closed in R𝑛 \ {0}. If Γ1 ⊂ Γ2 then Γ2◦ ⊂ Γ1◦ . If ch Γ

denotes the convex hull of Γ in R𝑛 \ {0} then Γ◦ = (ch Γ) ◦ ; and (Γ◦ ) ◦ is equal to the
closure of ch Γ. Below, 𝜅 > 0 is arbitrary.

Theorem 7.4.9 For a hyperfunction 𝑓 in R𝑛 to be microanalytic at (𝑥 ◦ , 𝜉 ◦ ) ∈ R𝑛 ×

S𝑛−1 it is necessary and sufficient that the following property holds:
Let 𝑈 be a suitably small neighborhood of 𝑥◦ in R𝑛 whose boundary 𝜕𝑈 is
smooth. To each analytic functional 𝜇 ∈ O ′ 𝑈 representing 𝑓 in 𝑈 there are
positive constants 𝑐, 𝑟, such that
7.4 Analytic Wave-front Set of a Hyperfunction 199

|F𝜅 𝜇 (𝑧, 𝜁)| ≲ e−𝑐 |𝜁 | (7.4.12)

for all (𝑧, 𝜁) ∈ C2𝑛 , |𝑧 − 𝑥 ◦ | < 𝑟, 𝜉 ◦ − 𝜁

|𝜁 | < 𝑟.

The FBI transform F𝜅 𝜇 is defined in (7.3.4).

Proof I. Necessity of the condition. By Definition 7.4.7 it suffices to prove the claim
when 𝑓 = 𝑏𝑈 ℎ, ℎ ∈ O (W𝛿 (𝑈, Γ)), 𝛿 > 0, and Γ is a convex open cone contained
in the half-space {𝑥 ∈ R𝑛 ; 𝜉 ◦ · 𝑥 < 0}. We select a vector 𝑣 ∈ Γ such that |𝑣| < 𝛿;
according to Definition 7.2.3 the fact that 𝑓 = 𝑏𝑈 ℎ means that if 1 > 𝑡 → +0, the
analytic functionals
∫
𝑛 𝜇𝑡
O (C ) ∋ 𝜑 ↦→ 𝜑 (𝑤) ℎ (𝑤) d𝑤
𝑈+𝑖𝑡 𝑣

converge in O ′ (C𝑛 ) to some analytic functional 𝜇 ∈ O ′ 𝑈 representing 𝑓 in 𝑈. In
particular,
∫
2
F𝜅 𝜇 (𝑧, 𝜁) = lim e𝑖𝜁 · (𝑧−𝑤)−𝜅 ⟨𝜁 ⟩ ⟨𝑧−𝑤 ⟩ ℎ (𝑤) Δ 𝜅 (𝑧 − 𝑤, 𝜁)d𝑤.
𝑡→+0 𝑈+𝑖𝑡 𝑣

As we have done before, we apply Stokes’ Theorem:

∫
2
e𝑖𝜁 · (𝑧−𝑤)−𝜅 ⟨𝜁 ⟩ ⟨𝑧−𝑤 ⟩ ℎ (𝑤) Δ 𝜅 (𝑧 − 𝑤, 𝜁)d𝑤 (7.4.13)
𝑈+𝑖𝑡 𝑣
∫
2
= e𝑖𝜁 · (𝑧−𝑤)−𝜅 ⟨𝜁 ⟩ ⟨𝑧−𝑤 ⟩ ℎ (𝑤) Δ 𝜅 (𝑧 − 𝑤, 𝜁)d𝑤
𝑈+𝑖𝑣
∫ 𝑠=1 ∫ 2
− e𝑖𝜁 · (𝑧−𝑤)−𝜅 ⟨𝜁 ⟩ ⟨𝑧−𝑤 ⟩ ℎ (𝑤) Δ 𝜅 (𝑧 − 𝑤, 𝜁)d𝑤.
𝑠=𝑡 𝑤 ∈𝜕𝑈+𝑖𝑠𝑣

It is again a matter of analyzing the exponents 𝐸 (𝑧, 𝑤, 𝜁). We provisionally fix

𝑧 = 𝑥◦, 𝜁 = 𝜉 ◦:

𝐸 (𝑥 ◦ , 𝑢 + 𝑖𝑠𝑣, 𝜉 ◦ ) = 𝑖𝜉 ◦ · (𝑥 ◦ − 𝑢 − 𝑖𝑠𝑣) − 𝜅 ⟨𝑥 ◦ − 𝑢 − 𝑖𝑠𝑣⟩ 2

and require 2𝜅 |𝑣| 2 < −𝜉 ◦ · 𝑣; we get

Re 𝐸 (𝑥 ◦ , 𝑢 + 𝑖𝑠𝑣, 𝜉 ◦ ) = 𝑠𝜉 ◦ · 𝑣 − 𝜅 |𝑥 ◦ − 𝑢| 2 + 𝑠𝜅 |𝑣| 2 (7.4.14)

1
< −𝜅 |𝑥 ◦ − 𝑢| 2 − 𝑠𝜅 |𝑣| 2 .
2
When 𝑢 ∈ 𝑈, 𝑠 = 1 as in the integral over 𝑈 + 𝑖𝑣 in the right-hand side of (7.4.13),
we get
1
𝐸 (𝑧, 𝑢 + 𝑖𝑣, 𝜁) ≤ − 𝜅 |𝜁 | |𝑣| 2
2
provided |𝑧 − 𝑥 ◦ | + 𝜉 ◦ − 𝜁
|𝜁 | is sufficiently small, in which case
200 7 Hyperfunctions in Euclidean Space
∫
2 1
e𝑖𝜁 · (𝑧−𝑤)−𝜅 ⟨𝜁 ⟩ ⟨𝑧−𝑤 ⟩ ℎ (𝑤) Δ 𝜅 (𝑧 − 𝑤, 𝜁)d𝑤 ≲ e− 4 𝜅 |𝜁 | . (7.4.15)
𝑈+𝑖𝑣

Next we look at (7.4.14) when 𝑢 ∈ 𝜕𝑈 and 𝑡 ≤ 𝑠 ≤ 1; in this case |𝑥 ◦ − 𝑢| ≥

dist (𝑥 ◦ , 𝜕𝑈) = 𝑑 and we require |𝑣| < 21 𝑑, implying

1
𝐸 (𝑧, 𝑢 + 𝑖𝑠𝑣, 𝜁) ≤ − 𝜅𝑑 2 |𝜁 | ,
2

again provided |𝑧 − 𝑥 ◦ | + 𝜉 ◦ − 𝜁
|𝜁 | is sufficiently small. We obtain
∫ 𝑠=1 ∫ 2 1 2 |𝜁
e𝑖𝜁 · (𝑧−𝑤)−𝜅 ⟨𝜁 ⟩ ⟨𝑧−𝑤 ⟩ ℎ (𝑤) Δ 𝜅 (𝑧 − 𝑤, 𝜁)d𝑤 ≲ e− 4 𝜅 𝑑 |
.
𝑠=𝑡 𝑤 ∈𝜕𝑈+𝑖𝑠𝑣

Combining this with (7.4.15) proves (7.4.12). Suppose that 𝜇 ′ ∈ O ′ 𝑈 is another
analytic functional representing 𝑓 in 𝑈. To see that an estimate of type (7.4.12) is
also valid for 𝜇 ′ it suffices to exploit Property (ii), Proposition 7.3.4, where 𝐾 = {𝑥 ◦ }
and 𝜇 is replaced by 𝜇 − 𝜇 ′ ∈ O ′ (𝜕𝑈).
II. Sufficiency of the condition. The proof of this
claim is a variant of that of
′
Theorem 7.4.4. Let the analytic functional 𝜇 ∈ O 𝑈 represent 𝑓 in 𝑈 and let Ω
be a bounded open subset of R𝑛 with smooth boundary and such that 𝑈 ⊂ Ω. We
′
apply Theorem 7.4.1: for every analytic functional 𝜇 ∈ O 𝑈 and every function
𝜑 ∈ O (C𝑛 ),
∫ ∫
−𝑛
⟨𝜇, 𝜑⟩ = (2𝜋) 𝜑 (𝑥) F𝜅 𝜇(𝑥, 𝜉)d𝑥d𝜉. (7.4.16)
𝑥 ∈Ω 𝜉 ∈R𝑛

We introduce finitely many pairwise disjoint, acute open cones ℭ 𝑗 ( 𝑗 = 0, 1, ..., 𝜈)

in R𝑛 \ {0} such that the Lebesgue measure of R𝑛 \ (ℭ0 ∪ ℭ1 ∪ · · · ∪ ℭ 𝜈 ) is equal to
zero. The cone ℭ0 is convex and such that

𝜉
𝜉 ◦ ∈ ℭ0 ∩ S𝑛−1 and 𝜉 ∈ ℭ0 =⇒ 𝜉 ◦ − < 𝑟. (7.4.17)
|𝜉 |

For 1 ≤ 𝑗 ≤ 𝜈 the cone ℭ 𝑗 will not necessarily be convex but we require 𝜉 ◦ ∉

◦
ℭ ◦𝑗 , the closed convex hull of ℭ 𝑗 . Let ℭ ′𝑗 be a convex open cone in 𝜉-space such
◦
that ℭ ◦𝑗 ⊂ ℭ ′𝑗 and 𝜉 ◦ ∉ ℭ ′𝑗 . Call Γ 𝑗 the interior of the polar of ℭ ′𝑗 ; we have
Γ◦𝑗 = ℭ ′𝑗 \ {0}; it follows that 𝜉 ◦ ∉ Γ◦𝑗 . This is equivalent to saying that there is a
𝑦 ( 𝑗) ∈ R𝑛 such that 𝑦 ( 𝑗) · 𝜉 ◦ < 0 whereas 𝑦 ( 𝑗) · 𝜉 > 0 for all 𝜉 ∈ ℭ 𝑗 . (The existence of
𝑦 ( 𝑗) could also be deduced from the Hahn–Banach Theorem – in finite dimension!)
For 𝑗 ≥ 1 let the function ℎ 𝜀𝑗 (𝑧) be defined as in (7.4.3), where ℭ = ℭ 𝑗 and Ω is
replaced by Ω′ ⊃ 𝑈; Lemma 7.4.2 implies that the entire function ℎ 𝜀𝑗 (𝑧) converges
7.4 Analytic Wave-front Set of a Hyperfunction 201

normally to a holomorphic function ℎ 𝑗 in W𝛿 Ω, Γ 𝑗 , 𝛿 > 0. As shown in the proof
of Theorem 7.4.4 the analytic functional
∫
O (C𝑛 ) ∋ 𝜑 ↦→ lim ℎ 𝜀𝑗 (𝑥) 𝜑 (𝑥) d𝑥
𝜀↘0 𝑥 ∈Ω

represents 𝑏 Ω ℎ 𝑗 .
We are left to deal with
𝜅 𝑛2 ∫ ∫
𝑛 ′ ′ 2 2
ℎ0𝜀 (𝑧) = 3
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉
2𝜋 𝑥 ′ ∈Ω′ 𝜉 ∈ℭ0
(7.4.18)
where now we assume that 𝜕Ω′ is smooth. Let 𝑦 ∈ R𝑛 ; Stokes’ Theorem entails
𝑛2
2𝜋 3

ℎ0𝜀 (𝑧) (7.4.19)
𝜅
∫ ∫
2
𝑛 ′ ′ −𝜀 | 𝜉 | 2
= |𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 −𝑖𝑦)−𝜅 | 𝜉 | ⟨𝑧−𝑥 −𝑖𝑦 ⟩ F𝜅 𝜇 (𝑥 ′ + 𝑖𝑦, 𝜉) d𝑥 ′d𝜉−
𝑥 ′ ∈Ω′ 𝜉 ∈ℭ0
∫ ∫ 𝑡=1 ∫
2
𝑛 ′ ′ −𝜀 | 𝜉 | 2
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 −𝑖𝑡 𝑦)−𝜅 | 𝜉 | ⟨𝑧−𝑥 −𝑖𝑡 𝑦 ⟩ F𝜅 𝜇 (𝑥 ′ + 𝑖𝑡𝑦, 𝜉) d𝑥 ′d𝜉.
𝑥 ′ ∈𝜕Ω′ 𝑡=0 𝜉 ∈ℭ0

The number 𝑟 in the statement of our condition will be sufficiently small that
|𝑥 ′ − 𝑥 ◦ | ≤ 𝑟/2 =⇒ 𝑥 ′ ∈ Ω′; we shall assume |𝑦| < 𝑟/2; then
𝑛2
2𝜋 3

ℎ0𝜀 (𝑥 + 𝑖𝑦) (7.4.20)
𝜅
∫ ∫
𝑛 ′ ′ | 2 −𝜀 | 𝜉 | 2
= |𝜉 | 2 e𝑖 𝜉 · ( 𝑥−𝑥 )−𝜅 | 𝜉 | | 𝑥−𝑥 F𝜅 𝜇 (𝑥 ′ + 𝑖𝑦, 𝜉) d𝑥 ′d𝜉
| 𝑥 ′ −𝑥 ◦ |<𝑟/2 𝜉 ∈ℭ0
∫ ∫
𝑛 ′ ′ | 2 −𝜀 | 𝜉 | 2
+ |𝜉 | 2 e𝑖 𝜉 · ( 𝑥−𝑥 )−𝜅 | 𝜉 | |𝑥−𝑥 F𝜅 𝜇 (𝑥 ′ + 𝑖𝑦, 𝜉) d𝑥 ′d𝜉
𝑥 ′ ∈Ω′ , |𝑥 ′ −𝑥 ◦ |>𝑟/2 𝜉 ∈ℭ0
∫ ∫ 𝑡=1 ∫
2
𝑛 ′ ′ −𝜀 | 𝜉 | 2
− |𝜉 | 2 e𝑖 𝜉 · ( 𝑥−𝑥 +𝑖 (1−𝑡) 𝑦)−𝜅 | 𝜉 | ⟨𝑥−𝑥 +𝑖 (1−𝑡) 𝑦 ⟩
𝑥 ′ ∈𝜕Ω′ 𝑡=0 𝜉 ∈ℭ0

× F𝜅 𝜇 (𝑥 ′ + 𝑖𝑡𝑦, 𝜉) d𝑥 ′d𝜉.

In the first integral in the right-hand side of (7.4.20) we take advantage of (7.4.12),
more precisely, of the fact that |𝑥 ′ − 𝑥 ◦ | < 𝑟, |𝑦| < 𝑟 and 𝜉 ∈ ℭ0 imply

|F𝜅 𝜇 (𝑥 ′ + 𝑖𝑦, 𝜉)| ≲ e−𝑐 | 𝜉 | .

As 𝜀 ↘ 0 that first integral converges to a holomorphic function in 𝔅 1 𝑟 (𝑥 ◦ ) +

2
𝑖𝔅 1 𝑟 (0); we are using the notation 𝔅𝑟 (𝑥 ◦ ) = {𝑥 ∈ R𝑛 ; |𝑥 − 𝑥 ◦ | < 𝑟}. In the second
2
and third integrals in the right-hand side of (7.4.20) we exploit once again Property
(i), Proposition 7.3.4 (where 𝜂 > 0 replaces 𝜀):
202 7 Hyperfunctions in Euclidean Space
′ ′ 2 ′ | 2 +𝜅 |𝑦 | 2
e𝑖 𝜉 · ( 𝑥−𝑥 +(1−𝑖𝑡) 𝑦)−𝜅 | 𝜉 | ⟨𝑥−𝑥 +𝑖 (1−𝑡) 𝑦 ⟩ F𝜅 𝜇 (𝑥 ′ + 𝑖𝑡𝑦, 𝜉) ≤ 𝐶 𝜂 e ( 𝜂−𝜅 | 𝑥−𝑥 ) | 𝜉 |.

1 4 1 1
We take |𝑥 − 𝑥 ◦ | < 10 𝑟 (whence |𝑥 − 𝑥 ′ | > 10 𝑟), |𝑦| < 10 𝑟 and 𝜂 < 10 𝜅𝑟 2 . We
can then let 𝜀 go to zero; these two integrals converge to holomorphic functions
in 𝔅 1 𝑟 (𝑥 ◦ ) + 𝑖𝔅 1 𝑟 (0). It follows that ℎ0𝜀 converges, as 𝜀 ↘ 0, to a holomorphic
10 10
function in a full neighborhood of 𝑥 ◦ . The proof of Theorem 7.4.9 is complete. □

Corollary 7.4.10 Let Ω be an open subset of R𝑛 and 𝑢 ∈ D ′ (Ω). The analytic

wave-front set of 𝑢 in the distribution sense is equal to the analytic wave-front set of
𝑢 in the hyperfunction sense (Definition 7.4.7).

Proof Combine Theorem 7.4.9 and Definition 3.5.1. □

Proposition 7.4.11 Let 𝑓 be a hyperfunction in R𝑛 , Ω a bounded open subset of

R𝑛 and Γ1 ,...,Γ𝜈 convex open cones in R𝑛 \ {0}. Suppose that, given arbitrarily
an open subset 𝑈 ⊂⊂ Ω of R𝑛 and convex open cones Γ1′ ,...,Γ𝜈′ such that ∅ ≠
Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 for each 𝑗, there exist a neighborhood 𝑉 C of 𝑈 in C𝑛 and functions

ℎ 𝑗 ∈ O 𝑉 C ∩ R𝑛 + 𝑖Γ ′𝑗 ( 𝑗 = 1, ..., 𝜈) such that 𝑓 |𝑈 = 𝜈𝑗=1 𝑏𝑈 ℎ 𝑗 . Under this
Í

hypothesis,
𝑊 𝐹a ( 𝑓 | Ω ) ⊂ Ω × Γ1◦ ∪ · · · ∪ Γ𝜈◦ .

(7.4.21)

We have used the notation 𝑏𝑈 for the hyperfunction boundary value and Γ◦ for
the polar of Γ.
Proof Suppose 𝜉 ◦ ∉ Γ1◦ ∪ · · · ∪ Γ𝜈◦ . For each 𝑗 we can select Γ ′𝑗 such that 𝜉 ◦ · 𝑦 < 0 for
′
every 𝑦 ∈ Γ 𝑗 . By Definition 7.4.7 𝑓 is microanalytic at every point (𝑥, 𝜉 ◦ ), 𝑥 ∈ 𝑈.□
We can prove a sort of converse to Proposition 7.4.11.

Proposition 7.4.12 Let 𝑓 , Ω, Γ1 ,...,Γ𝜈 , be as in Proposition 7.4.11 and let Γ ′𝑗 ( 𝑗 =

1, ..., 𝜈) be convex open cones in R𝑛 \ {0} such that ∅ ≠ Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 . If (7.4.21)
holds then every 𝑥 ◦ ∈ Ω has a neighborhood 𝑈 C in C𝑛 such that 𝑈 = 𝑈 C ∩ R𝑛 ⊂⊂ Ω
and the following is true:
There are functions ℎ♭ ∈ O 𝑈 C , ℎ 𝑗 ∈ O W𝛿 𝑈, Γ ′𝑗 ( 𝑗 = 1, ..., 𝜈, 𝛿 > 0)

such that
∑︁𝜈
𝑓 |𝑈 = ℎ ♭ + 𝑏𝑈 ℎ 𝑗 . (7.4.22)
𝑈
𝑗=1

Note that there are no (lower or upper) bounds on 𝛿.

Proof For each 𝑗 = 1, ..., 𝜈 we select convex open cones Γ ′′𝑗 such that

Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ ′′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 (7.4.23)

and we define the following open cones in R𝑛 \ {0}:

7.4 Analytic Wave-front Set of a Hyperfunction 203

ℭ1 = interior of Γ1′′◦ , (7.4.24)

ℭ2 = interior of Γ2′′◦ \ Γ2′′◦ ∩ ℭ1 ...,

ℭ 𝜈 = interior of Γ𝜈′′◦ \ Γ𝜈′′◦ ∩ (ℭ1 ∪ · · · ∪ ℭ 𝜈−1 ) ;

lastly,

ℭ0 = interior of R𝑛 \ (ℭ1 ∪ · · · ∪ ℭ 𝜈 )
Ù𝜈
= (interior of R𝑛 \ℭ 𝑗 ).
𝑗=1

By construction ℭ0 , ℭ1 , ..., ℭ 𝜈 are pairwise disjoint; and R𝑛 \ (ℭ0 ∪ ℭ1 ∪ · · · ∪ ℭ 𝜈 )

is a set of Lebesgue measure zero.
Assume 1 ≤ 𝑗 ≤ 𝜈. The polar of Γ ′′𝑗 is acute and so is therefore ℭ 𝑗 ; moreover
◦
Γ ′′𝑗 is equal to the interior of Γ ′′◦
𝑗 ⊂ ℭ ◦𝑗 . It ensues that 𝑦 · 𝜉 > 𝑐 ◦ |𝑦| |𝜉 | for some
𝑐 ◦ > 0 and all 𝑦 ∈ Γ ′𝑗 , 𝜉 ∈ ℭ 𝑗 .
Let 𝑈 ⊂⊂ Ω be a neighborhood of 𝑥 ◦ in R𝑛 , soon to be chosen more precisely.
Let 𝜇 ∈ O ′ 𝑈 represent 𝑓 in 𝑈 and let Ω′ be a bounded open subset of R𝑛 such
that 𝑈 ⊂ Ω′; we assume that both 𝑈 and Ω′ have smooth boundaries. We have
𝜈
∑︁
ℎ 𝜀 = ℎ0𝜀 + ℎ 𝜀𝑗 + ℎ∞
𝜀
, (7.4.25)
𝑗=1

where ℎ 𝜀𝑗 (1 ≤ 𝑗 ≤ 𝜈) and ℎ∞ 𝜀 are defined exactly as in (7.4.8) but with the choice

(7.4.24) of the cones ℭ 𝑗 , whereas

𝜅 𝑛2 ∫ ∫
𝑛 ′ ′ 2 2
ℎ0𝜀 (𝑧) = 3
|𝜉 | 2 e𝑖 𝜉 · (𝑧−𝑥 )−𝜅 | 𝜉 | ⟨𝑧−𝑥 ⟩ −𝜀 | 𝜉 | F𝜅 𝜇 (𝑥 ′, 𝜉) d𝑥 ′d𝜉.
2𝜋 𝑥 ′ ∈Ω′ 𝜉 ∈ℭ0
(7.4.26)
By Lemma 7.4.2, if 𝜅𝛿 is sufficiently small and 1 ≤ 𝑗 ≤ 𝜈, ℎ 𝜀𝑗 converges as

𝜀 ↘ 0 to a holomorphic function ℎ 𝑗 in the wedge W𝛿 R𝑛 , Γ ′𝑗 . In the proof of
Theorem 7.4.4 it is shown that ℎ∞ 𝜀 converges to a holomorphic function ℎ in the tube

{𝑧 ∈ C ; |Im 𝑧| < 𝛿}.

𝑛
Ø 𝜈
We must deal with ℎ0𝜀 . Since ℭ1 ∪ · · · ∪ ℭ 𝜈 = (interior of Γ ′′◦
𝑗 ) we have
𝑗=1

𝜈
Ù 𝜈
Ù
ℭ0 ∩ S𝑛−1 ⊂ R𝑛 \Γ ′′◦
𝑗 ∩ S𝑛−1 ⊂⊂ R𝑛 \Γ◦𝑗 ∩ S𝑛−1 .
𝑗=1 𝑗=1

The hypothesis (7.4.21) implies that if 𝑥 ◦ ∈ 𝑈 and 0 ≠ 𝜉 ∈ ℭ0 ∩ S𝑛−1 then 𝑓 is

microanalytic at (𝑥 ◦ , 𝜉). Let 𝜇 ∈ O ′ 𝑈 represent 𝑓 in 𝑈. Using a finite covering
of ℭ0 ∩ S𝑛−1 by suitably small open subsets of S𝑛−1 we derive from Theorem 7.4.9
204 7 Hyperfunctions in Euclidean Space

that if diam Ω′ (and therefore diam 𝑈) and 𝑐 > 0 are sufficiently small then

∀ (𝑥 ′, 𝜉) ∈ Ω′ × ℭ0 , |F𝜅 𝜇 (𝑥 ′, 𝜉)| ≲ e−𝑐 | 𝜉 | .

Taking this into account in (7.4.26) shows that ℎ0𝜀 converges uniformly to a function

ℎ0 ∈ O 𝑈 C for a suitable choice of 𝑈 C . Using (7.4.9) we conclude the proof like
that of Theorem 7.4.4: (7.4.25) entails (7.4.22) where ℎ♭ = ℎ0 + ℎ∞ . □
Actually, it is natural to reason mod C 𝜔 functions, i.e., to deal with singularity
hyperfunctions (Definition 7.1.24). Propositions 7.4.11, 7.4.12 yield the following,
where 𝑓 𝑔 means 𝑊 𝐹a ( 𝑓 − 𝑔) = ∅.

Theorem 7.4.13 Let Ω be a bounded open subset of R𝑛 and Γ1 ,...,Γ𝜈 convex open
cones in R𝑛 \ {0}. The following properties of a hyperfunction 𝑓 in R𝑛 are equivalent:
(i) 𝑊 𝐹a ( 𝑓 | Ω ) ⊂ Ω × Γ1◦ ∪ · · · ∪ Γ𝜈◦ .

(ii) Given arbitrarily 𝛿 > 0 and convex open cones Γ ′𝑗 such that ∅ ≠ Γ ′𝑗 ∩S𝑛−1 ⊂⊂ Γ 𝑗
( 𝑗 = 1, ..., 𝜈), every point of Ω has aneighborhood
in R𝑛 , 𝑈 ⊂⊂ Ω, such that
𝑓 |𝑈 𝑗=1 𝑏𝑈 ℎ 𝑗 with ℎ 𝑗 ∈ O W𝛿 𝑈, Γ ′𝑗 .
Í𝜈

We have the analogue of Theorem 3.5.5:

Corollary 7.4.14 Let 𝑈 ⊂ R𝑛 be an open set and Γ ⊂ R𝑛 \ {0} an open and convex
cone. The following properties of a hyperfunction 𝑢 ∈ B (𝑈) are equivalent:
(a) 𝑢 is the boundary value of a function ℎ ∈ O (W𝛿 (𝑈, Γ)), 𝛿 > 0;
(b) 𝑊 𝐹a (𝑢) ⊂ 𝑈 × Γ◦ .

Distributions in an open subset Ω of R𝑛 are special hyperfunctions in Ω. The

separate definitions, for distributions and for hyperfunctions, of the boundary value
of a holomorphic function in a wedge W𝛿 (Ω, Γ) call for harmonization:

Theorem 7.4.15 If the hyperfunction boundary value of ℎ ∈ O (W𝛿 (Ω, Γ)) is a

distribution in Ω then ℎ has tempered growth at the edge (Definition 3.3.2) and
the distribution boundary value of ℎ (Definition 3.3.7) is equal to its hyperfunction
boundary value (Definition 7.2.3).

Proof Let 𝑢 ∈ D ′ (Ω) be the hyperfunction boundary value of ℎ ∈ O (W𝛿 (Ω, Γ)).
Corollary 7.4.10 tells us that the analytic wave-front sets in the distribution and
hyperfunction sense are equal. Proposition 7.4.11 implies that 𝑊 𝐹a (𝑢) ⊂ 𝑈 × Γ◦ .
By Theorem 3.5.5, 𝑊 𝐹a (𝑢) ⊂ 𝑈 × Γ◦ implies that 𝑢 is the distribution boundary
value of a function ℎ♭ ∈ Otemp (W𝛿 (Ω, Γ)). The injectivity of the boundary value
map (Proposition 7.2.5) demands ℎ = ℎ♭ . □
The analytic singular support of 𝑓 ∈ B (Ω), singsuppa 𝑓 , is the support of the
singularity hyperfunction 𝑓 ♮ represented by 𝑓 in Ω. If 𝑓1 ∈ B (Ω) is another repre-
sentative of 𝑓 ♮ in Ω then, obviously, 𝑊 𝐹a ( 𝑓1 − 𝑓 ) = ∅.
7.4 Analytic Wave-front Set of a Hyperfunction 205

Definition 7.4.16 By the analytic wave-front set of a singularity hyperfunction 𝑓 ♮

in Ω we shall mean
that of any one of its hyperfunction
representatives; it shall be
◦ ◦
denoted by 𝑊 𝐹a 𝑓 . If (𝑥 , 𝜉 ) ∉ 𝑊 𝐹a 𝑓 we shall say that 𝑓 ♮ is microanalytic
♮ ♮

at (𝑥 ◦ , 𝜉 ◦ ).

7.4.4 Differential operators and the analytic wave-front set

Let Ω be an open subset of R𝑛 and 𝑃 (𝑥, D) be a linear PDO of order 𝑚 with complex
coefficients belonging to C 𝜔 (Ω):
∑︁
𝑃 (𝑥, D 𝑥 ) = 𝑐 𝛼 (𝑥) D 𝑥𝛼 .
| 𝛼 | ≤𝑚

Let ΩC be an open subset of C𝑛 such that Ω ⊂ ΩC ∩R𝑛 , to which each function 𝑐 𝛼 has
a holomorphic extension 𝑐 𝛼 (𝑧). Let 𝑈 be an open subset of R𝑛 whose closure 𝑈 is
a compact subset of Ω and let 𝛿 > 0 be such that 𝑈 C𝛿 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝑈) < 𝛿} ⊂
ΩC . It follows that the extension 𝑃 (𝑧, D𝑧 ) is defined and holomorphic in 𝑈 C𝛿 and
therefore also in any wedge W𝛿 (𝑈, Γ) [Γ ⊂ R𝑛 \ {0}: a convex cone, cf. (3.3.1)].
We have
𝑃 (𝑧, D𝑧 ) O (W𝛿 (𝑈, Γ)) ⊂ O (W𝛿 (𝑈, Γ)) .
The next two statements are easy consequences of the definition of hyperfunctions
and boundary values. Actually, they are special cases of Proposition 8.1.35 and
Corollary 8.1.36.

Proposition 7.4.17 For every ℎ ∈ O (W𝛿 (Ω, Γ)) we have

𝑏 Ω 𝑃 (𝑧, 𝜕𝑧 ) ℎ = 𝑃 (𝑥, 𝜕𝑥 ) 𝑏 Ω ℎ.

Proposition 7.4.18 For every hyperfunction 𝑢 in Ω we have

𝑊 𝐹a ( 𝑃 (𝑥, 𝜕𝑥 ) 𝑢) ⊂ 𝑊 𝐹a (𝑢). (7.4.27)

In connection with (7.4.27) we state, without proof, the PDE version of the
celebrated fundamental principle of M. Sato. Sato’s theorem will be proved in much
greater generality in Section 17.5.

Theorem 7.4.19 Let 𝑃 = 𝑃 (𝑥, D 𝑥 ) be a linear PDO of order 𝑚 with C 𝜔 coefficients

in Ω. Whatever the hyperfunction 𝑢 in Ω, if 𝑃𝑢 is microanalytic at (𝑥 ◦ , 𝜉 ◦ ) ∈
Ω × (R𝑛 \ {0}) and if 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) ≠ 0 then 𝑢 is microanalytic at (𝑥 ◦ , 𝜉 ◦ ).

Proposition 7.4.18 and Theorem 7.4.19 are obviously valid with “singularity
hyperfunction” substituted for “hyperfunction”.
206 7 Hyperfunctions in Euclidean Space

7.5 Edge of the Wedge

7.5.1 The theorem of the Edge of the Wedge

Let 𝑈 be a bounded open subset of R𝑛 . The results of the preceding section enable
us to describe the kernel of the boundary value operator 𝒃 E(𝑈) : E (𝑈) ↦→ B (𝑈)
(see the end of Subsection 7.4.2). We point out that, if W𝛿 (𝑈, Γ) is a wedge with
edge 𝑈 and ℎ ∈ O (W𝛿 (𝑈, Γ)), to say that 𝑏𝑈 ℎ ∈ C 𝜔 (𝑈) is to say that ℎ extends
holomorphically to a neighborhood of 𝑈 in C𝑛 . We also recall that the vector sum
of two open convex cones in R𝑛 \ {0}, Γ + Γ ′, is a convex open cone in R𝑛 that
might be equal to the whole of R𝑛 . In what follows this case is not precluded: when
Γ + Γ ′ = R𝑛 we have

W𝛿 (𝑈, Γ + Γ ′) = W𝛿 (𝑈, R𝑛 )
= {𝑧 ∈ C𝑛 ; Re 𝑧 ∈ 𝑈, |Im 𝑧| < 𝛿} .

The following statement is the celebrated Theorem of the Edge of the Wedge:

convex open cones in R \ {0}.

Theorem 7.5.1 Let Γ1 , ..., Γ𝜈 (𝜈 ≥ 2) be nonempty 𝑛

Assume that the functions ℎ 𝑗 ∈ O (W𝛿 𝑈, Γ 𝑗 ) ( 𝑗 = 1, ..., 𝜈, 𝛿 > 0) have the

following property:
Í
(Z) 𝜈𝑗=1 𝑏𝑈 ℎ 𝑗 ∈ C 𝜔 (𝑈).
Then the following is true, for any set of convex open cones Γ ′𝑗 such that ∅ ≠
Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 ( 𝑗 = 1, ..., 𝜈):
(Z’) To every 𝑥 ◦ ∈ 𝑈 there exist a neighborhood 𝑈 ′ ⊂ 𝑈 of 𝑥 ◦ in R𝑛 , a num-
ber 𝛿 ′ ∈ (0, 𝛿) and, to each pair
( 𝑗, 𝑘), 1 ≤ 𝑗 < 𝑘 ≤ 𝜈, func-
of indices
′ ′ ′
tions 𝑔 𝑗,𝑘 = −𝑔 𝑘, 𝑗 ∈ O W𝛿 𝑈 , Γ 𝑗 + Γ𝑘 such that, for each 𝑗 = 1, ..., 𝜈,
′

𝑏𝑈′ ℎ 𝑗 − 𝜈𝑘=1 𝑔 𝑗,𝑘 ∈ C 𝜔 (𝑈 ′).

Í

Remark 7.5.2 Once 𝑥 ◦ selected, we are allowed to contract 𝑈 about 𝑥 ◦ and decrease
𝛿 > 0 as needed. In particular, this allows us Í
Í to replace (Z) by the property that
𝜈 𝜈
𝑗=1 𝑏 𝑈 ℎ 𝑗 ≡ 0. Indeed, if we assume that 𝑓 = 𝑗=1 ℎ 𝑗 extends holomorphically to
W𝛿 (𝑈, Γ1 ) we can replace ℎ1 by ℎ1 − 𝑓 .
(𝑞)
Proof For each pair of positive integers 𝑗 ≤ 𝜈, 𝑞 ≤ 𝜈, let Γ 𝑗 be a convex open
cone such that
(𝑞+1) (𝑞)
Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 .

The recursive argument in the proof will require us to select appropriately neighbor-
hoods of 𝑥 ◦ , 𝑈 (𝑞+1) ⊂ 𝑈 (𝑞) ⊂ 𝑈.
7.5 Edge of the Wedge 207

(
Let 𝔖𝜈 𝑝) denote the set of multi-indices 𝐼 = 𝑖 1 , ..., 𝑖 𝑝 such that 1 ≤ 𝑖1 <

· · · < 𝑖 𝑝 ≤ 𝜈; it is convenient to introduce also the set 𝔖𝜈(0) consisting of the single
multi-index ∅, the empty multi-index (with 𝜈 arbitrary). Assuming that the cones
(𝑞) (
Γ 𝑗 for some 𝑞 ∈ Z+ have been selected, then to each multi-index 𝐼 ∈ 𝔖𝜈 𝑝) we
(𝑞) (𝑞) (𝑞)
shall associate the cone Γ𝐼 = Γ𝑖1 + · · · + Γ𝑖 𝑝 (cf. Remark 7.5.2); if 𝑝 = 0 we set
(𝑞)
Γ∅ = ∅.
We shall reason by induction on 𝑞 ≥ 1 and start from the hypothesis that there
(𝑞−1)
is a neighborhood 𝑈 (𝑞) ⊂ 𝑈 of 𝑥 ◦ in R𝑛 and, for each 𝐼 ∈ 𝔖𝜈 , each integer
(𝑞) (𝑞)
𝑗 ∈ (1, ..., 𝜈), 𝑗 ∉ 𝐼, a function ℎ 𝐼, 𝑗 ∈ O W𝛿 𝑈 (𝑞) , Γ𝐼 + Γ 𝑗 , such that
∑︁ ∑︁
𝑏𝑈 (𝑞) ℎ 𝐼, 𝑗 0 [i.e., ∈ C 𝜔 𝑈 (𝑞) ]. (7.5.1)
𝐼 ∈𝔖𝜈
(𝑞−1) 1≤ 𝑗 ≤𝜈
𝑗∉𝐼

Hypothesis (Z) states that (7.5.1) is true when 𝑞 = 1, 𝑈 (1) = 𝑈, Γ (1)

𝑗 = Γ 𝑗 , ℎ ∅, 𝑗 = ℎ 𝑗 .
(𝑞)
For arbitrary 𝐽 ∈ 𝔖𝜈 we introduce the symmetrization
∑︁
ℎ 𝐽𝜎 = ℎ 𝐼, 𝑗 . (7.5.2)
(𝑞−1)
𝐼 ∈𝔖𝜈
𝐼 ⊂𝐽 , 𝑗 ∈𝐽\𝐼

(𝑞)
Obviously ℎ 𝐽𝜎 ∈ O W𝛿 (𝑞) 𝑈 (𝑞) , Γ𝐽 and, as a consequence of (7.5.1),
∑︁
𝑏𝑈 (𝑞) ℎ 𝐽𝜎 0. (7.5.3)
(𝑞)
𝐽 ∈𝔖𝜈

(𝑞) Í
Given any 𝐽 ∈ 𝔖𝜈 , we define 𝑓 𝐽 = 𝑏𝑈 (𝑞) ℎ 𝐽𝜎 ; (7.5.3) reads (𝑞)
𝐽 ∈𝔖𝜈
𝑓 𝐽 0. We
apply Proposition 7.4.11 (or Theorem 7.4.13) twice:
◦
(𝑞)
𝑊 𝐹a ( 𝑓 𝐽 ) ⊂ 𝑈 (𝑞) × Γ𝐽

and Ø ◦
(𝑞)
𝑊 𝐹a ( 𝑓 𝐽 ) ⊂ 𝑈 (𝑞) × Γ𝐽 ′
(𝑞)
𝐽 ′ ∈𝔖𝜈
𝐽 ′ ≠𝐽

since ∑︁
𝑓𝐽 = − 𝑓𝐽′ .
(𝑞)
𝐽 ′ ∈𝔖𝜈
𝐽 ′ ≠𝐽
208 7 Hyperfunctions in Euclidean Space

It follows that
Ø ◦ ◦
(𝑞) (𝑞)
𝑊 𝐹a ( 𝑓 𝐽 ) ⊂ 𝑈 (𝑞) × Γ𝐽 ∩ Γ𝐽 ′
(𝑞)
𝐽 ′ ∈𝔖𝜈
𝐽 ′ ≠𝐽
Ø ◦
(𝑞) (𝑞)
= 𝑈 (𝑞) × Γ𝐽 + Γ𝐽 ′ .
(𝑞)
𝐽 ′ ∈𝔖𝜈
𝐽 ′ ≠𝐽

We apply Proposition 7.4.12) [or Theorem 7.4.13, (ii)]: if the neighborhood 𝑈 (𝑞+1)
(𝑞+1) (𝑞) (𝑞)
of 𝑥 ◦ in R𝑛 is sufficiently small and if ∅ ≠ Γ𝐽 ∩ S𝑛−1 ⊂⊂ Γ𝐽 for all 𝐽 ∈ 𝔖𝜈 ,
′ (𝑞) ′
then,
to each pair of multi-indices 𝐽, 𝐽 ∈ 𝔖𝜈 , 𝐽 ≠ 𝐽 , there is a function 𝜓 𝐽 , 𝐽 ′ ∈
(𝑞+1) (𝑞+1)
O W𝛿 𝑈 (𝑞+1) , Γ𝐽 + Γ𝐽 ′ such that
∑︁
𝑓 𝐽 |𝑈 (𝑞+1) = 𝑏𝑈 (𝑞+1) ℎ 𝐽𝜎 𝑏𝑈 (𝑞+1) 𝜓 𝐽 , 𝐽 ′ . (7.5.4)
(𝑞)
𝐽 ′ ∈𝔖𝜈
𝐽 ′ ≠𝐽

(𝑞)
For 𝐽 ∈ 𝔖𝜈 , 𝑘 ∈ (1, ..., 𝜈) \𝐽, we define
1 ∑︁
ℎ 𝐽 ,𝑘 = 𝜓𝐽 , 𝐽′ ,
𝑁 (𝑞, 𝑘) (𝑞)
𝐽 ′ ∈𝔖𝜈 , 𝐽 ′ ∋𝑘

(𝑞)
where 𝑁 (𝑞, 𝑘) is the number of multi-indices 𝐽 ′ ∈ 𝔖𝜈 that contain 𝑘; (7.5.4) can
be rewritten as ∑︁
𝑓 𝐽 |𝑈 (𝑞+1) = 𝑏𝑈 (𝑞+1) ℎ 𝐽𝜎 𝑏𝑈 (𝑞+1) ℎ 𝐽 ,𝑘 . (7.5.5)
1≤𝑘 ≤𝜈
𝑘∉𝐽

A consequence of (7.5.5) and of the injectivity of the boundary value map (Propo-
sition 7.2.5) is that
∑︁
(𝑞) e(𝑞+1) ,
∀𝐽 ∈ 𝔖𝜈 , ℎ 𝐽𝜎 − ℎ 𝐽 ,𝑘 ∈ O 𝑈 (7.5.6)
1≤𝑘 ≤𝜈
𝑘∉𝐽

where 𝑈e(𝑞+1) is an open subset of C𝑛 such that 𝑈e(𝑞+1) ∩R𝑛 = 𝑈 (𝑞+1) . A consequence
of (7.5.3) and (7.5.5) is that
∑︁ ∑︁
𝑏𝑈 (𝑞+1) ℎ 𝐼, 𝑗 0.
𝐼 ∈𝔖𝜈
(𝑞) 1≤ 𝑗 ≤𝜈
𝑗∉𝐼

This is the analogue of (7.5.1) with 𝑞 + 1 substituted for 𝑞, thereby proving (7.5.1)
for all 𝑞 ≤ 𝜈; (7.5.1) is vacuous for 𝑞 = 𝜈 + 1.
7.5 Edge of the Wedge 209

Since (7.5.1)=⇒(7.5.3) the latter holds for all 𝑞 ≥ 0, 𝑞 ≤ 𝜈. From (7.5.3), using
the injectivity of the boundary value map we deduce, for an appropriate choice of
the neighborhoods 𝑈 e(𝑞) of 𝑥 ◦ in C𝑛 [in particular, 𝑈
e(𝑞+1) ⊂ 𝑈
e(𝑞) for all 𝑞 ≤ 𝜈 − 1],
∑︁
ℎ 𝐼𝜎 ∈ O 𝑈 e(𝑞) . (7.5.7)
(𝑞)
𝐼 ∈𝔖𝜈

Note that the last case is ℎ (1,...,𝜈)
𝜎 ∈O 𝑈 e(𝜈) .
Next we proceed with the definition of the functions 𝑔 𝑗,𝑘 in (Z’). If 𝑘 > 1 we
define 𝑔1,𝑘 = ℎ1,𝑘 in W𝛿 𝑉, Γ1′ + Γ𝑘′ . Now suppose 2 ≤ 𝑗 < 𝑘 ≤ 𝜈; we define
∑︁ ∑︁
𝑔 𝑗,𝑘 = ℎ ( 𝑗),𝑘 + ℎ (𝑖1 , 𝑗),𝑘 + ℎ (𝑖1 ,𝑖2 , 𝑗),𝑘 + · · · + ℎ (1,2,..., 𝑗),𝑘 . (7.5.8)
𝑖1 < 𝑗 𝑖1 <𝑖2 < 𝑗

By restriction, we regard 𝑔 𝑗,𝑘 as a holomorphic function in

(𝑞+1) (𝑞+1)
W𝛿 𝑈 (𝑞+1) , Γ 𝑗 + Γ𝑘

and, as a matter of fact, in

O W𝛿 𝑈 (𝜈) , Γ (𝜈) (𝜈)
𝑗 + Γ𝑘 .

We derive from (7.5.8):

𝜈
∑︁ ∑︁ ∑︁
𝑔 𝑗,𝑘 = 𝑔 𝑗,𝑘 − 𝑔 𝑘, 𝑗 (7.5.9)
𝑘=1 𝑘 >𝑗 𝑘 <𝑗
!
∑︁ ∑︁ ∑︁
= ℎ ( 𝑗),𝑘 + ℎ (𝑖1 , 𝑗),𝑘 + ℎ (𝑖1 ,𝑖2 , 𝑗),𝑘 + · · · + ℎ (1,2,..., 𝑗),𝑘
𝑘 >𝑗 𝑖1 < 𝑗 𝑖1 <𝑖2 < 𝑗
!
∑︁ ∑︁ ∑︁
− ℎ (𝑘), 𝑗 + ℎ (𝑖1 ,𝑘), 𝑗 + ℎ (𝑖1 ,𝑖2 ,𝑘), 𝑗 + · · · + ℎ (1,2,...,𝑘), 𝑗 .
𝑘 <𝑗 𝑖1 <𝑘 𝑖1 <𝑖2 <𝑘

Since ℎ (𝜎𝑗) = ℎ 𝑗 [cf. (7.5.2)] putting 𝑞 = 1 in (7.5.6) yields

∑︁ ∑︁
ℎ𝑗 − ℎ ( 𝑗),𝑘 − ℎ ( 𝑗),𝑘 ∈ O 𝑈e(2) .
𝑘< 𝑗 𝑗<𝑘 ≤𝜈
210 7 Hyperfunctions in Euclidean Space

This and (7.5.9) imply

𝜈
∑︁ ∑︁
𝑔 𝑗,𝑘 − ℎ 𝑗 ≈ − ℎ ( 𝑗),𝑘 + ℎ (𝑘), 𝑗 (7.5.10)
𝑘=1 𝑘< 𝑗
!
∑︁ ∑︁ ∑︁
+ ℎ (𝑖1 , 𝑗),𝑘 + ℎ (𝑖1 ,𝑖2 , 𝑗),𝑘 + · · · + ℎ (1,2,..., 𝑗),𝑘
𝑘 >𝑗 𝑖1 < 𝑗 𝑖1 <𝑖2 < 𝑗
!
∑︁ ∑︁ ∑︁
− ℎ (𝑖1 ,𝑘), 𝑗 + ℎ (𝑖1 ,𝑖2 ,𝑘), 𝑗 + · · · + ℎ (1,2,...,𝑘), 𝑗
𝑘 <𝑗 𝑖1 <𝑘 𝑖1 <𝑖2 <𝑘

where ≈ means that the left-hand and right-hand side differ by some function 𝐻
holomorphically extendible to 𝑈 e(𝜈) . We propose to prove (Z’) with 𝑈 ′ = 𝑈 e(𝜈) ,
′ (𝜈) Í𝜈
Γ 𝑗 = Γ 𝑗 , in other words 𝑘=1 𝑔 𝑗,𝑘 ≈ ℎ 𝑗 . By (7.5.2) with 𝑞 = 2 we have ℎ (𝑘, 𝑗) = 𝜎

ℎ ( 𝑗),𝑘 + ℎ (𝑘), 𝑗 . We must therefore prove

!
∑︁ ∑︁ ∑︁ ∑︁
𝜎
ℎ (𝑖1 , 𝑗) ≈ ℎ (𝑖1 , 𝑗),𝑘 + ℎ (𝑖1 ,𝑖2 , 𝑗),𝑘 + · · · + ℎ (1,2,..., 𝑗),𝑘 (7.5.11)
𝑖1 < 𝑗 𝑘 >𝑗 𝑖1 < 𝑗 𝑖1 <𝑖2 < 𝑗
!
∑︁ ∑︁ ∑︁
− ℎ (𝑖1 ,𝑘), 𝑗 + ℎ (𝑖1 ,𝑖2 ,𝑘) , 𝑗 + · · · + ℎ (1,2,...,𝑘), 𝑗 .
𝑘 <𝑗 𝑖1 <𝑘 𝑖1 <𝑖2 <𝑘

Proof of (7.5.11): Í When 𝑗 = 1, (7.5.11) states that 0 = 0; when 𝑗 = 2, (7.5.11)

states that ℎ (1,2)
𝜎 ≈ 𝑘>2 ℎ (1,2),𝑘 , which is true by (7.5.6). Now suppose 𝑗 ≥ 3; for
arbitrary 𝑝 ≥ 1, 𝑝 ≤ 𝑗 − 2, (7.5.6) implies
∑︁ ∑︁ ∑︁ ∑︁
ℎ 𝜎𝑖 ,...,𝑖 , 𝑗 ≈ ℎ ( 𝑖1 ,...,𝑖 𝑝 , 𝑗 ) ,𝑘 + ℎ ( 𝑖1 ,...,𝑖 𝑝 , 𝑗 ) ,𝑘
(1 𝑝 )
𝑖1 <···<𝑖 𝑝 < 𝑗 𝑘< 𝑗 𝑖1 <···<𝑖 𝑝 < 𝑗 𝑖1 <···<𝑖 𝑝 < 𝑗<𝑘
𝑖 𝛼 ≠𝑘, 𝛼=1,..., 𝑝
∑︁ ∑︁
= ℎ ( 𝑖1 ,...,𝑖b𝛼 ,...,𝑖 𝑝+1 , 𝑗 ) ,𝑖 𝛼 + ℎ ( 𝑖1 ,...,𝑖 𝑝 , 𝑗 ) ,𝑘 ,
𝑖1 <···<𝑖 𝑝+1 < 𝑗 𝑖1 <···<𝑖 𝑝 < 𝑗<𝑘
𝛼=1,..., 𝑝+1

where we sum over all indicated indices (varying between 1 and 𝜈) and the hatted
index is deleted; this can be rewritten as
∑︁ ∑︁
ℎ 𝜎𝑖 ,...,𝑖 , 𝑗 − ℎ 𝜎𝑖 ,...,𝑖 , 𝑗
(1 𝑝 ) (1 𝑝+1 )
𝑖1 <···<𝑖 𝑝 < 𝑗 𝑖1 <···<𝑖 𝑝+1 < 𝑗
∑︁ ∑︁ ∑︁
≈ ℎ ( 𝑖1 ,...,𝑖 𝑝 , 𝑗 ) ,𝑘 − ℎ ( 𝑖1 ,...,𝑖 𝑝 ,𝑘 ) , 𝑗 .
𝑖1 <···<𝑖 𝑝 < 𝑗<𝑘 𝑘< 𝑗 𝑖1 <···<𝑖 𝑝 <𝑘

We add all these equivalences over 𝑝 = 1, ..., 𝑗 − 2:

7.5 Edge of the Wedge 211
∑︁ ∑︁
𝜎 𝜎
ℎ (𝑖1 , 𝑗)
− ℎ (1,..., 𝑗)
𝑖1 < 𝑗 𝑖1 <···<𝑖 𝑝+1 < 𝑗
𝑗−2 ∑︁
∑︁ ∑︁ 𝑗−2 ∑︁
∑︁ ∑︁
≈ ℎ ( 𝑖1 ,...,𝑖 𝑝 , 𝑗 ) ,𝑘 − ℎ ( 𝑖1 ,...,𝑖 𝑝 ,𝑘 ) , 𝑗 .
𝑝=1 𝑘> 𝑗 𝑖1 <···<𝑖 𝑝 < 𝑗 𝑝=1 𝑘< 𝑗 𝑖1 <···<𝑖 𝑝 <𝑘

This shows that the right-hand side in (7.5.11) is equal to

∑︁ ∑︁
𝜎 𝜎
ℎ (𝑖1 , 𝑗)
− ℎ (1,..., 𝑗) + ℎ (1,2,..., 𝑗),𝑘
𝑖1 < 𝑗 𝑘> 𝑗
Í
and thus (7.5.11) reduces to ℎ (1,...,
𝜎
𝑗)
≈ 𝑘> 𝑗 ℎ (1,2,..., 𝑗),𝑘 , true by (7.5.6). The proof
of Theorem 7.5.1 is complete. □
The case 𝜈 = 2 is particularly illuminating.

Corollary 7.5.3 Let 𝑈 be a bounded open subset of R𝑛 , 𝑥 ◦ a point of 𝑈, and Γ1 , Γ2

two nonempty convex open cones in R𝑛 \ {0}. Let 𝑓 𝑗 ∈ O (W𝛿 𝑈, Γ 𝑗 ) ( 𝑗 = 1, 2,
𝛿 > 0) be such that 𝑏𝑈 𝑓1 + 𝑏𝑈 𝑓2 ∈ C 𝜔 (𝑈). Let Γ ′𝑗 be a convex open cone such that
∅ ≠ Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 . Then there is a function 𝑔 ∈ O (W𝛿′ 𝑈 ′, Γ1′ + Γ2′ ) (𝑈 ′ ⊂ 𝑈

a neighborhood 𝑥 ◦ in R𝑛 , 0 < 𝛿 ′ < 𝛿) such that 𝑓1 = 𝑔 in W𝛿′ 𝑈 ′, Γ1′ and 𝑓2 = −𝑔

in W𝛿′ 𝑈 ′, Γ2′ .

Proof We can assume that 𝑏𝑈 𝑓1 + 𝑏𝑈 𝑓2 ≡ 0 in 𝑈 (cf. Remark 7.5.2). Theorem

7.5.1 implies that there is a 𝑔1,2 ∈ O (W𝛿′ 𝑈 ′, Γ1′ + Γ2′ such that ℎ1 = 𝑓1 − 𝑟 1 𝑔1,2 ,

ℎ2 = 𝑓2 + 𝑟 2 𝑔1,2 [𝑟 𝑗 : restriction to W𝛿′ 𝑈 ′, Γ ′𝑗 ] extend holomorphically to a
neighborhood of 𝑈 ′ in C𝑛 . After contracting
𝑈 ′ ◦ ′
about 𝑥 and decreasing 𝛿 we can
assume that ℎ 𝑗 is holomorphic in W𝛿′ 𝑈 ′, Γ ′𝑗 . We have 𝑏𝑈′ ℎ1 = −𝑏𝑈′ ℎ2 , implying
ℎ1 = −ℎ2 in a neighborhood of 𝑈 ′ in C𝑛 . The function 𝑔 = 𝑔1,2 + ℎ1 fulfills the
requirements after contracting 𝑈 ′ about 𝑥 ◦ and decreasing 𝛿 ′. □
In the context of Corollary 7.5.3 there are two cases to ponder:

(1) Γ1 + Γ2 is contained in an open half-space, in which case the convex cone Γ1′ + Γ2′
is acute and 𝑔 extends holomorphically 𝑓1 and − 𝑓2 to W𝛿′ 𝑈 ′, Γ1′ + Γ2′ .
(2) Γ1 + Γ2 = R𝑛 , in which case Γ ′𝑗 ( 𝑗 = 1, 2) can be chosen so that Γ1′ + Γ2′ = R𝑛 .
Incidentally, this simply means that there is a vector v ∈ Γ1′ ∩ −Γ2′ . Then the

conclusion in Corollary 7.5.3 is that 𝑓1 extends as a holomorphic function, 𝑔, in a
full neighborhood of 𝑥 ◦ in C𝑛 , while 𝑓2 extends as −𝑔 in the same neighborhood.

Remark 7.5.4 Theorem 7.5.1 provides a new definition of hyperfunctions in 𝑈, by

allowing us to identify B (𝑈) with E (𝑈) /ker 𝒃 E(𝑈) (see the end of Subsection
7.3.2).
212 7 Hyperfunctions in Euclidean Space

7.5.2 One more characterization of the analytic wave-front set

Given 𝜃 ◦ ∈ S𝑛−1 we use the notation Γ 𝜀 (𝜃 ◦ ) = {𝑦 ∈ R𝑛 \ {0} ; 𝑦 · 𝜃 ◦ > 𝜀 |𝑦|},

0 < 𝜀 < 1. We denote by Γ◦𝜀 (𝜃 ◦ ) the polar of Γ 𝜀 (𝜃 ◦ ), i.e., the set of 𝜃 ∈ R𝑛 such
that 𝑦 · 𝜃 ≥ 0 for all 𝑦 ∈ Γ 𝜀 (𝜃 ◦ ). When 𝑛 = 1, 𝜃 ◦ = ±1,

Γ 𝜀 (𝜃 ◦ ) = {𝑦 ∈ R\ {0} ; 𝑦 · 𝜃 ◦ > 0} ,
Γ◦𝜀 (𝜃 ◦ ) = Γ 𝜀 (𝜃 ◦ ).

Lemma 7.5.5 Suppose 𝑛 ≥ 2; then

𝑛−1 1
√︁
◦ ◦ 𝑛−1 ◦ 2 2
Γ 𝜀 (𝜃 ) ∩ S = 𝜃 ∈ S ; |𝜃 − 𝜃 | ≤ 1 − 1 − 𝜀 . (7.5.12)
2

Proof Formula (7.5.12) is equivalent to

n √︁ o
Γ◦𝜀 (𝜃 ◦ ) ∩ S𝑛−1 = 𝜃 ∈ S𝑛−1 ; 𝜃 · 𝜃 ◦ ≥ 1 − 𝜀 2 . (7.5.13)

After a rotation we can assume 𝜃 ◦ = (0, ..., 0, 1); then 𝑦 · 𝜃 ◦ = 𝑦 𝑛 and 𝑦 ∈ Γ 𝜀 (𝜃 ◦ ) ∩

S𝑛−1 ⇐⇒ 𝑦 𝑛 > 𝜀; we have 𝜃 ∈ Γ◦𝜀 (𝜃 ◦ ) if and only if

∀𝑦 ∈ S𝑛−1 , 𝑦 𝑛 ≥ 𝜀 =⇒ 𝜃 · 𝑦 ≥ 0.

In particular, if 𝑦 ′ = (𝑦 1 , ..., 𝑦 𝑛−1 ) √

and 𝑦 𝑛 = 𝜀, we see that 𝜃 √∈ Γ◦𝜀 (𝜃 ◦ ) requires
′
𝜀𝜃 𝑛 ≥ −𝜃 · 𝑦 . We select 𝑦 = − 1 − 𝜀 2 | 𝜃𝜃 ′ | to get 𝜀𝜃 𝑛 ≥ 1 − 𝜀 2 |𝜃 ′ |, which
′ ′ ′
√
demands 𝜃 𝑛 ≥ 1 − 𝜀 2 since |𝜃 ′ | 2 = 1 − 𝜃 2𝑛 . This proves that the left-hand side in
(7.5.13) is contained in the√right-hand side. To prove the converse suppose |𝑦| = 1,
𝑦 𝑛 > 𝜀, and |𝜃| = 1, 𝜃 𝑛 ≥ 1 − 𝜀 2 , therefore |𝜃 ′ | ≤ 𝜀; we have

𝜃 · 𝑦 ≥ 𝜃 𝑛 𝑦 𝑛 − 𝜀 |𝑦 ′ | > 𝜀 (𝜃 𝑛 − |𝑦 ′ |) ≥ 0. □

Lemma 7.5.6 Let ℭ be an acute, convex, open cone in R𝑛 \ {0}. Define

Γ = {𝑦 ∈ R𝑛 ; ∀𝜃 ∈ ℭ, 𝑦 · 𝜃 > 𝑐 ◦ |𝑦| |𝜃|} , 𝑐 ◦ > 0. (7.5.14)

Assume that 𝜃 ◦ ∉ ℭ. Then, provided 𝑐 ◦ and 𝜀 > 0 are sufficiently small we have
Γ 𝜀 (𝜃 ◦ ) + Γ = R𝑛 .
Proof We can select 𝑦 ∗ ∈ S𝑛−1 such that 𝑦 ∗ · 𝜉 > 0 for every 𝜃 ∈ ℭ whereas
𝑦 ∗ · 𝜃 ◦ = −𝜀, i.e., −𝑦 ∗ ∈ Γ 𝜀 (𝜃 ◦ ) (𝜀 > 0 suitably small). We take

𝑦∗ · 𝜃
𝑐 ◦ < inf ,
𝜃 ∈ℭ |𝜃|

ensuring that 𝑦 ∗ ∈ Γ. Since Γ 𝜀 (𝜃 ◦ )+Γ is open and convex we must have Γ 𝜀 (𝜃 ◦ )+Γ =
R𝑛 . □
7.5 Edge of the Wedge 213

In the sequel we use the notation [cf. (3.3.1)]

W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) = {𝑥 + 𝑖𝑦 ∈ C𝑛 ; 𝑥 ∈ 𝑈, 𝑦 ∈ Γ 𝜀 (𝜃 ◦ ) , |𝑦| < 𝛿} (7.5.15)

with 𝑈 a domain in R𝑛 , 𝛿 > 0. As 𝜀 ↗ 1 the wedge W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) contracts

to the set 𝑈 + 𝑖(0, 𝛿)𝜃 ◦ whereas, if 𝜀 ↘ 0, the cone Γ 𝜀 (𝜃 ◦ ) expands to the open
half-space {𝑦 ∈ R𝑛 ; 𝑦 · 𝜃 ◦ > 0}.

Proposition 7.5.7 Let (𝑥 ◦ , 𝜃 ◦ ) ∈ R𝑛 × S𝑛−1 , the neighborhood 𝑈 of 𝑥 ◦ in R𝑛 and

𝛿 > 0 be arbitrary. The following properties of ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) are
equivalent:
(1) (𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a (ℎ);
(2) ℎ extends holomorphically to a neighborhood of 𝑥 ◦ in C𝑛 .

Proof That (1)=⇒(2) follows from Definition 7.4.7, Lemma 7.5.6 and the Edge of
the Wedge Theorem 7.5.1; (2)=⇒(1) is trivial. □
Let now 𝑈 be a connected neighborhood of 𝑥 ◦ in R𝑛 . We shall denote by
O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) ∩ O 𝑥 ◦ the subalgebra (with respect to ordinary multiplication)
of O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) consisting of the functions that extend holomorphically to
a neighborhood of 𝑥 ◦ in C𝑛 . To shorten the notation we define

O 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) = O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) /(O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) ∩ O 𝑥 ◦ ) .
(7.5.16)
Let 𝑈 ′ be another connected neighborhood of 𝑥 ◦ , 𝑈 ′ ⊂ 𝑈, 0 < 𝛿 ′ ≤ 𝛿, 𝜀 ′ ≥ 𝜀; then
restriction from W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) to W𝛿′ (𝑈 ′, Γ 𝜀′ (𝜃 ◦ )) induces a linear injection

O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) ∋ ℎ ↦→ ℎ| W𝛿′ (𝑈′ ,Γ 𝜀′ ( 𝜃 ◦ )) ∈ O (W𝛿′ (𝑈 ′, Γ 𝜀′ (𝜃 ◦ ))) .

Obviously, if ℎ| W𝛿′ (𝑈′ ,Γ 𝜀′ ( 𝜃 ◦ )) extends holomorphically to a neighborhood of 𝑥 ◦ in

C𝑛 then the same is true of ℎ; the preceding map induces a linear injection

O 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) −→ O 𝑥+◦ (W𝛿′ (𝑈 ′, Γ 𝜀′ (𝜃 ◦ ))) . (7.5.17)

We can regard the left-hand side as a vector subspace of the right-hand side and form
the direct limit of the vector spaces O 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) as 𝑈 ranges over the set
of connected neighborhoods of 𝑥 ◦ , 𝛿 ↘ 0, 0 < 𝜀 ↗ 1,

O+( 𝑥 ◦ , 𝜃 ◦ ) = limO 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) . (7.5.18)

−→

We remind the reader that (7.5.18) is the quotient of the disjoint sum
•

+
𝑈∋𝑥 ◦ , 𝛿>0,0< 𝜀<1
O 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )))
214 7 Hyperfunctions in Euclidean Space

for the following equivalence relation: ℎ 𝑗 ∈ O 𝑥+◦ W𝛿 𝑗 𝑈 𝑗 , Γ 𝜀 𝑗 (𝜃 ◦ ) , 𝑗 = 1, 2,
are equivalent if there exist a connected neighborhood of 𝑥 ◦ , 𝑈 ⊂ 𝑈1 ∩ 𝑈2 , and
positive numbers 𝛿 ≤ min (𝛿1 , 𝛿2 ), 𝜀 ≥ max (𝜀1 , 𝜀2 ), 𝜀 < 1, such that ℎ1 = ℎ2 in
W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) [in the sense of the injections (7.5.17)].
As (𝑥 ◦ , 𝜃 ◦ ) ranges over R𝑛 × S𝑛−1 , 𝑈 ranges over the set of connected neigh-
borhoods of 𝑥 ◦ and 𝜀 < 1 gets arbitrarily close to 1, the sets 𝑈 × Γ 𝜀 (𝜃 ◦ ) ∩ S𝑛−1
form a basis of the topology of R𝑛 × S𝑛−1 . Let 𝑏𝑈 : O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) −→
B (𝑈) /C 𝜔 (𝑈) be the singularity hyperfunction boundary value map (Definition
7.2.6). If 0 < 𝛿 ′ < 𝛿 the restriction mapping induces a linear injection

O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) /ker 𝑏𝑈 −→ O (W𝛿′ (𝑈, Γ 𝜀 (𝜃 ◦ ))) /ker 𝑏𝑈 .

We can therefore form the direct limit

e+ (𝑈 × Γ 𝜀 (𝜃 ◦ )) = limO (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) /ker 𝑏𝑈 .
O (7.5.19)
−→

e+ (𝑈 × Γ 𝜀 (𝜃 ◦ )) and the natural restriction mappings

The vector spaces O
e+ (𝑈 × Γ 𝜀 (𝜃 ◦ )) −→ O
O e+ (𝑈 ′ × Γ 𝜀′ (𝜃 ◦ )) (7.5.20)

(𝑈 ′ ⊂ 𝑈 a connected neighborhood of 𝑥 ◦ , 𝜀 ≤ 𝜀 ′ < 1) define a presheaf O e + over

R ×S .
𝑛 𝑛−1

Proposition 7.5.8 The stalk at (𝑥 ◦ , 𝜃 ◦ ) of the sheaf O+ R𝑛 × S𝑛−1 defined by the

presheaf O e + is O+ ◦ ◦ .
(𝑥 ,𝜃 )

Proof Let ℎ 𝑗 ∈ O 𝑥+◦ W𝛿 𝑗 𝑈 𝑗 , Γ 𝜀 𝑗 (𝜃 ◦ ) , 𝑗 = 1, 2, represent the same h ( 𝑥 ◦ , 𝜃 ◦ ) ∈
O+( 𝑥 ◦ , 𝜃 ◦ ) (𝑈 𝑗 : connected neighborhoods of 𝑥 ◦ , 𝛿 𝑗 > 0, 0 < 𝜀 𝑗 < 1). Then there exist a
connected neighborhood of 𝑥 ◦ , 𝑈 ⊂ 𝑈1 ∩𝑈2 , and positive numbers 𝛿 ≤ min (𝛿1 , 𝛿2 ),
𝜀 ≥ max (𝜀1 , 𝜀2 ), 𝜀 < 1, such that ℎ1 − ℎ2 ∈ ker 𝑏𝑈 in W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) [in
the sense of the injections (7.5.17)]. As before, 𝑏𝑈 : O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) −→
B (𝑈) /C 𝜔 (𝑈) is the singularity hyperfunction boundary value map; note that there
are natural injections

O 𝑥+◦ W𝛿 𝑗 𝑈 𝑗 , Γ 𝜀 𝑗 (𝜃 ◦ ) −→ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) /ker 𝑏𝑈 .

This proves that O+( 𝑥 ◦ , 𝜃 ◦ ) is equal to the direct limit of the spaces (7.5.19) as 𝑈 ranges
over the set of connected neighborhoods of 𝑥 ◦ and 𝜀 < 1 gets arbitrarily close to 1.□
We denote by O + (U) the vector space of continuous sections of O+ R𝑛 × S𝑛−1

over an open subset U of R𝑛 × S𝑛−1 . We leave it as an exercise to check whether
e+ (𝑈 × Γ 𝜀 (𝜃 ◦ )) = O + (𝑈 × Γ 𝜀 (𝜃 ◦ )) or not (𝑈: an arbitrary open subset of R𝑛 ).
O
We denote by ℭ 𝜀 the interior of Γ◦𝜀 (𝜃 ◦ ); we have
Ù
ℭ 𝜀 = {𝜃 ∈ R𝑛 ; ∃𝜆 > 0, 𝜃 = 𝜆𝜃 ◦ } . (7.5.21)
𝜀>0
7.5 Edge of the Wedge 215

Proposition 7.5.9 Let (𝑥 ◦ , 𝜃 ◦ ) ∈ R𝑛 × S𝑛−1 , the neighborhood 𝑈 of 𝑥 ◦ in R𝑛 and

𝛿 > 0 be arbitrary. If 𝜀 > 0 is sufficiently small then to every 𝑓 ∈ B (R𝑛 ) there is a
function ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) such that

𝑊 𝐹a ( 𝑓 − 𝑏𝑈 ℎ) ∩ (𝑈 × ℭ 𝜀 ) = ∅. (7.5.22)

If 𝑈 ′, 𝛿 ′, 𝜀 ′, ℎ♭ ∈ O (W𝛿′ (𝑈 ′, Γ 𝜀′ (𝜃 ◦ ))) have the analogous properties, then ℎ − ℎ♭

extends holomorphically to a neighborhood of 𝑥 ◦ in C𝑛 .
For (𝑥 ◦ , 𝜃 ◦ ) to belong to 𝑊 𝐹a ( 𝑓 ) it is necessary and sufficient that ℎ not be
holomorphically extendible to a neighborhood of 𝑥 ◦ in C𝑛 .

Proof Recall that ℭ 𝜀 is the interior of Γ◦𝜀 (0 < 𝜀 < 1) and that ℭ 𝜀 ∩ S𝑛−1 ↘ {𝜃 ◦ }
as 𝜀 ↘ 0. We select a special family of pairwise disjoint acute (but, in general, not
convex) open cones ℭ1 , ..., ℭ 𝑁 in R𝑛 \ {0} such that ℭ2𝜀 ∩ (ℭ1 ∪ · · · ∪ ℭ 𝑁 ) = ∅ and
R𝑛 \ (ℭ2𝜀 ∪ ℭ1 ∪ · · · ∪ ℭ 𝑁 ) has measure zero. The cones ℭ1 , ..., ℭ 𝑁 are required to
satisfy the following condition: the closed convex hull of each ℭ 𝑗 does not intersect
the closure of ℭ 𝜀 in R𝑛 \ {0}. (This might require 𝑁 to be very large.) Let then
Γ 𝑗 be the cones defined in (7.4.6) [cf. (7.5.14)]; the polar Γ◦𝑗 of Γ 𝑗 contains the
closed convex hull of ℭ 𝑗 and converges to it as 𝑐 ◦ → 0; this allows us to select 𝑐 ◦
sufficiently small that Γ◦𝑗 ∩ ℭ 𝜀 = {0} for every 𝑗 = 1, ..., 𝑁. By Theorem 7.4.4, given
arbitrarily 𝑓 ∈ B (R𝑛 ), a neighborhood 𝑈 of ◦
𝑥 in R and 𝛿 > 0, there are functions
𝑛

ℎ ∈ O (W𝛿 (𝑈, Γ2𝜀 )), ℎ 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 , such that

𝑁
∑︁
𝑓 | 𝑈 = 𝑏𝑈 ℎ + 𝑏𝑈 ℎ 𝑗 .
𝑗=1

By Corollary 7.4.14 we have 𝑊 𝐹a 𝑏𝑈 ℎ 𝑗 ⊂ 𝑈 × Γ◦𝑗 , 𝑗 = 1, ..., 𝑁, which implies

𝑊 𝐹a ( 𝑓 − 𝑏𝑈 ℎ) ∩ (𝑈 × ℭ 𝜀 ) = ∅.

Let ℎ♭ ∈ W𝛿′ (𝑈 ′, Γ 𝜀′ (𝜃 ◦ )) be as in the statement; then

(𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a 𝑏𝑈 ♭ ℎ − ℎ♭ .

It follows from Proposition 7.5.7 that ℎ − ℎ♭ extends holomorphically to a neighbor-

hood of 𝑥 ◦ in C𝑛 . Applying the latter result to ℎ♭ ≡ 0 proves that if (𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a ( 𝑓 )
then ℎ extends holomorphically to a neighborhood of 𝑥 ◦ in C𝑛 . Conversely, if this
last property holds then (7.5.22) and Proposition 7.5.7 entail (𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a ( 𝑓 ). □

Corollary 7.5.10 Let (𝑥 ◦ , 𝜃 ◦ ) ∈ R𝑛 × S𝑛−1 , the neighborhood 𝑈 of 𝑥 ◦ in R𝑛 and

𝛿 > 0 be arbitrary. To every 𝑓 ∈ B (R𝑛 ) there is a unique h ( 𝑥 ◦ , 𝜃 ◦ ) ∈ O+( 𝑥 ◦ , 𝜃 ◦ ) with
the following property: if ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) is a representative of h ( 𝑥 ◦ , 𝜃 ◦ )
then (𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a ( 𝑓 − 𝑏𝑈 ℎ).
216 7 Hyperfunctions in Euclidean Space

Proposition 7.5.9 and Corollary 7.5.10 imply the existence of a boundary value
map
+
𝑏𝑈 : O + (𝑈 × Γ 𝜀 (𝜃 ◦ )) −→ B (𝑈) /C 𝜔 (𝑈) (7.5.23)
[cf. (7.5.19)] and of the corresponding sheaf homomorphism

b+ : O+ R𝑛 × S𝑛−1 −→ Bsing (R𝑛 ) ; (7.5.24)

Corollary 7.5.10 states that b+ is surjective, injective when restricted to a given

continuous section over 𝑈 × {𝜃 ◦ }, 𝜃 ◦ ∈ S𝑛−1 fixed. The sheaf O+ R𝑛 × S𝑛−1 will
be given another interpretation in the next section.

7.5.3 The Holmgren Theorem for hyperfunctions

Theorem 3.5.12 can be extended to singularity hyperfunctions in R𝑛 .

Theorem 7.5.11 Let 𝜇 ∈ O ′ (R𝑛 ) represent 𝑢 ∈ B sing (R𝑛 ) in a bounded open subset
Ω of R𝑛 . Suppose that 𝑥 ◦ ∈ Ω ∩ supp 𝜇 and that there is a neighborhood 𝑈 ⊂ Ω
of 𝑥 ◦ and 𝑓 ∈ C ∞ (𝑈; R) such that 𝑓 (𝑥 ◦ ) = 0, 𝜉 ◦ · d𝑥 = d 𝑓 (𝑥 ◦ ) ≠ 0 and that
supp 𝜇 ⊂ {𝑥 ∈ Ω; 𝑓 (𝑥) ≥ 0}. Then (𝑥 ◦ , ±𝜉 ◦ ) ∈ 𝑊 𝐹a (𝑢).

Proof If 𝑉 ⊂ 𝑈 is a neighborhood of 𝑥 ◦ we can apply Theorem 6.3.11 and use a

decomposition 𝜇 = 𝜇1 + 𝜇2 such that supp 𝜇1 ⊂ 𝑉 and 𝑥 ◦ ∉ supp 𝜇2 . In other words,
the hypotheses in Theorem 7.5.11 remain valid after we contract 𝑈 about 𝑥 ◦ . The
same argument as in the proof of Theorem 3.5.12 allows us to assume that 𝑥 ◦ = 0
and 𝑓 (𝑥) = 𝑥 𝑛 . In the new coordinates we take 𝑈 = 𝑈 ′ × (−𝑇, 𝑇), 𝑈 ′ ⊂ R𝑛 open,
′

0 ∈ 𝑈 , 𝑇 > 0. Let 𝜑 ∈ O C 𝑛−1 be arbitrary; the linear functional

O (C) ∋ ℎ ↦→ ⟨𝜇, 𝜑 (𝑧 ′) ℎ (𝑧 𝑛 )⟩

[𝑧 ′ = (𝑧1 , ..., 𝑧 𝑛−1 )] defines an analytic functional 𝜇 𝜑 in C which, in turn, defines a

hyperfunction 𝑢 𝜑 in (−𝑇, 𝑇). It follows from the hypotheses that supp 𝜇 𝜑 ⊂ [0, +∞).
Suppose (0, 1) ∉ 𝑊 𝐹a 𝑢 𝜑 ; the one-dimensional version of Theorem 7.4.13 implies
that 𝑢 𝜑 is the boundary value in (−𝜀, 𝜀) of a function ℎ ∈ O ((−𝜀, 𝜀) × (0, −𝛿))
(0 < 𝜀 < 𝑇, 𝛿 > 0, suitably small). Since 𝑢 𝜑 ≡ 0 in (−𝜀, 0) the injectivity of the
boundary map (Proposition 7.2.5) implies ℎ ≡ 0 in (−𝜀, 0) × (0, −𝛿) hence in C𝑛 .
We reach the conclusion that 𝑢 𝜑 ≡ 0 in (−𝜀, 𝜀), implying supp 𝜇 𝜑 ⊂ [𝜀, +∞). Since
the products 𝜑 (𝑥 ′) ℎ (𝑥 𝑛 ) span a dense subspace of O (C𝑛 ) we reach the conclusion
that supp 𝜇 ⊂ R𝑛−1 × [𝜀, +∞), a contradiction. The same argument is valid with −𝜉 ◦
substituted for 𝜉 ◦ and 𝛿 for −𝛿. □

Theorem 7.5.11 enables us to derive the hyperfunction version of the Holmgren

Theorem 5.3.12:
7.6 Microfunctions in Euclidean space 217

Corollary 7.5.12 Let 𝜇, 𝑢, Ω, 𝑓 , be as in Theorem 7.5.11. Let 𝑃 (𝑥, D 𝑥 ) be a dif-

ferential operator of order 𝑚 ≥ 1 with C 𝜔 coefficients in Ω, 𝑃𝑚 (𝑥, 𝜉) its principal
symbol. If 𝑃 (𝑥, D 𝑥 ) 𝑢 ∈ C 𝜔 (𝑈) then necessarily 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) = 0.
Proof If 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) ≠ 0, Sato’s Theorem 7.4.19 states that 𝑢 is microanalytic at
(𝑥 ◦ , 𝜉 ◦ ), contradicting Theorem 7.5.11. □

7.6 Microfunctions in Euclidean space

In this section we take advantage of the flabbiness of the sheaves B (R𝑛 ), Bsing (R𝑛 )
(Proposition 7.1.9, Corollary 7.1.29), and mainly deal with hyperfunctions and sin-
gularity hyperfunctions in R𝑛 .
All concepts related to microlocal analyticity and the analytic wave-front set are
invariant under the dilations R𝑛 × (R𝑛 \ {0}) ∋ (𝑥, 𝜉) ↦→ (𝑥, 𝜆𝜉), 𝜆 > 0. It is natural
to mod off the conical feature of our set-up and attach the properties of interest to the
traces on the unit sphere, i.e., replace R𝑛 × (R𝑛 \ {0}) by R𝑛 × S𝑛−1 . This simplifies
the formulation of the theory when we put it in the language of sheaves, as we will
do now. By the natural projection of R × (R \ {0}) onto R × S
𝑛 𝑛 𝑛 𝑛−1 we mean the

map (𝑥, 𝜉) ↦→ 𝑥, | 𝜉𝜉 | .

Definition 7.6.1 By the microsupport of 𝑓 ∈ B (R𝑛 ) [resp., 𝑓 ∈ B sing (R𝑛 )],

henceforth denoted by microsupp 𝑓 , we shall mean the image of 𝑊 𝐹a ( 𝑓 ) under the
natural projection R𝑛 × (R𝑛 \ {0}) −→ R𝑛 × S𝑛−1 .
Since 𝑊 𝐹a ( 𝑓 ) is closed in R𝑛 × (R𝑛 \ {0}), microsupp 𝑓 is a closed subset of
R𝑛 × S𝑛−1 .
Proposition 7.6.2 Let 𝑃 (𝑥, D) be a linear differential operator with analytic coeffi-
cients in R𝑛 ; then microsupp 𝑃 (𝑥, D) 𝑓 ⊂ microsupp 𝑓 for every 𝑓 ∈ B (R𝑛 ).
This is a rephrasing of Proposition 7.4.18.
For arbitrary (𝑥 ◦ , 𝜃 ◦ ) ∈ R𝑛 × S𝑛−1 let Amicro( 𝑥◦ , 𝜃 ◦)
denote the linear space of germs of
hyperfunctions 𝑓 in R𝑛 microanalytic (Definition 7.4.7) at (𝑥 ◦ , 𝜃 ◦ ) [and consequently
at all points (𝑥 ◦ , 𝜆𝜃 ◦ ), 𝜆 > 0], in other words, such that (𝑥 ◦ , 𝜃 ◦ ) ∉ microsupp 𝑓 . The
latter is an open property: if true there is a neighborhood 𝑈 (resp., Θ) of 𝑥 ◦ in
R𝑛 (resp., 𝜃 ◦ in S𝑛−1 ) such that (𝑈 × Θ) ∩ microsupp 𝑓 = ∅. This shows that the
assignment (𝑥 ◦ , 𝜃 ◦ ) ↦→ Amicro
( 𝑥◦ , 𝜃 ◦)
defines a subsheaf (of complex vector spaces) of
B (R ) [resp., B (R )] which we denote by Amicro (R𝑛 ).
𝑛 sing 𝑛

Definition 7.6.3 We refer to Amicro (R𝑛 ) as the sheaf of germs of hyperfunctions

(or of singularity hyperfunctions) in R𝑛 microanalytic at points of R𝑛 × S𝑛−1 .
Let 𝑈 ⊂ R𝑛 , Θ ⊂ S𝑛−1 be open sets as above. The continuous sections of
Amicro(R𝑛 ) over 𝑈 × Θ are the hyperfunctions in R𝑛 microanalytic
in 𝑈 × Θ; they
form a linear space A (𝑈 × Θ). Note that A 𝑈 × S𝑛−1 C 𝜔 (𝑈).
218 7 Hyperfunctions in Euclidean Space

The natural injection Amicro (R𝑛 ) ↩→ B (R𝑛 ) followed up with the quotient map
B (R𝑛 ) −→ Bsing (R𝑛 ) is a sheaf homomorphism and we can form the quotient sheaf.

Definition 7.6.4 We shall refer to the quotient

Bmicro (R𝑛 ) = Bsing (R𝑛 ) /Amicro (R𝑛 )

B (R𝑛 ) /Amicro (R𝑛 )

as the sheaf of germs of microfunctions in R𝑛 × S𝑛−1 .

The continuous sections of Bmicro (R𝑛 ) over 𝑈 × Θ are the microfunctions in

𝑈 × Θ; they form a linear space B micro (𝑈 × Θ). We have

B micro (𝑈 × Θ) = B sing (R𝑛 ) /A (𝑈 × Θ) (7.6.1)

B (R𝑛 ) /A (𝑈 × Θ) .

If 𝑈 ′ ⊂ 𝑈, Θ′ ⊂ Θ are also open there are natural restriction mappings

A (𝑈 × Θ) −→ A (𝑈 ′ × Θ′), B micro (𝑈 ′ × Θ′) −→ B micro (𝑈 ′ × Θ′), yielding
presheaves that (re)define the sheaves Amicro (R𝑛 ) and Bmicro (R𝑛 ).

Remark 7.6.5 We are diverging from common practice in hyperfunction theory: in

the latter we should use the notation C and C instead of B micro and Bmicro . But in a
text like this one, with all the connotations with continuity (and complex structures),
we have thought wiser not to use cees.

All the sequences of sheaf homomorphisms in the following diagram are exact
and the diagram is commutative:

0
↘
C 𝜔 (R𝑛 ) −→ B (R𝑛 ) (7.6.2)
↓ ↗ ↓ ↘
0 −→ Amicro (R𝑛 ) −→ Bsing (R𝑛 ) −→ Bmicro (R𝑛 ) −→ 0.

The support of 𝑓 ♮ ∈ B micro (𝑈 × Θ), denoted by supp 𝑓 ♮ , is equal to

(𝑈 sing (R𝑛 ) is an arbitrary representative of

♮× Θ) ∩ microsupp 𝑓 , where 𝑓 ∈♮ B
♮ ♮

𝑓 . The base projection of supp 𝑓 is equal to supp 𝑓 ♮ = singsuppa 𝑓 (Defini-

tion 7.4.7) if 𝑓 ∈ B (R𝑛 ) is a representative of 𝑓 ♮ . A linear differential operator
𝑃 (𝑥, D) with analytic
♮ coefficients in an open subset Ω of R𝑛 can be made to act on a
microfunction 𝑓 in Ω × S 𝑛−1 by acting on a hyperfunction in Ω representing it;
♮ ♮
supp 𝑃 (𝑥, D) 𝑓 ⊂ supp 𝑓 . This is a direct consequence of Proposition 7.6.2.
The surjection Bsing (R𝑛 ) −→ Bmicro (R𝑛 ) implies the following result in [Kashi-
wara, 1971].

Theorem 7.6.6 The sheaf Bmicro (R𝑛 ) is flabby.

7.6 Microfunctions in Euclidean space 219

At the end of Subsection 7.4.2 we introduced the boundary value map (7.5.23),
leading to the sheaf homomorphism (7.5.24) and thence to the composite
b+
O+ R𝑛 × S𝑛−1 −→ Bsing (R𝑛 ) −→ Bmicro (R𝑛 ) (7.6.3)

where the second arrow is the same surjection as in (7.6.2). This leads to an alternative
definition of microfunctions.

Theorem 7.6.7 The composite map (7.6.3) is a sheaf isomorphism.

Proof Corollary 7.5.10 states that b+ is surjective and therefore the same is true
of (7.6.3). The map (7.6.3) is injective. Indeed, to say that h ( 𝑥 ◦ , 𝜃 ◦ ) ∈ O+(𝑥 ◦ , 𝜃 ◦ )
belongs to the kernel of (7.6.3) is the same as saying that (𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a (𝑏𝑈 ℎ),
where ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) is an arbitrary representative of h ( 𝑥 ◦ , 𝜃 ◦ ) (with 𝑈 a
neighborhood of 𝑥 ◦ in R𝑛 , 𝛿 > 0, 𝜀 > 0). But then, by Proposition 7.5.7, ℎ extends
holomorphically to a neighborhood of 𝑥 ◦ in C𝑛 , whence h ( 𝑥 ◦ , 𝜃 ◦ ) = 0. □
We shall also use the following terminology.

Definition 7.6.8 Let Ω (resp., Θ) be an open subset of R𝑛 (resp., S𝑛−1 ). A micro-

function [ 𝑓 ] in Ω × Θ will be called a microdistribution if [ 𝑓 ] has a representative
𝑓 ∈ B (R𝑛 ) whose restriction to Ω is a distribution.
Chapter 8
Hyperdifferential Operators

This chapter is devoted to a class of differential operators 𝐽 (𝑧, 𝜕𝑧 ) of infinite order

that act on holomorphic functions in a domain ΩC of C𝑛 . They are defined by the
property that
𝐽 (𝑧, 𝜁) = e−𝑧·𝜁 𝐽 (𝑧, 𝜕𝑧 ) e𝑧·𝜁 (8.1)

is an entire function of infra-exponential type with respect to 𝜁 valued in O ΩC ,
meaning that |𝐽 (𝑧, 𝜁)| exp (−𝜀 |𝜁 |) is bounded at ∞ in C𝑛𝜁 uniformly on compact
subsets of ΩC𝑧 for every 𝜀 > 0 (Definition 8.1.1 below). The function 𝐽 (𝑧, 𝜁) can be
referred to as the symbol (or total symbol) of the hyperdifferential operator 𝐽 (𝑧, 𝜕𝑧 ).
The microlocal version of this class of symbols, in which the variation of 𝜁 is limited
to an open cone with vertex at the origin in C𝑛 , will play a central role in Part VI of
this book, to a somewhat different purpose.
Hyperdifferential operators act on analytic functionals in C𝑛 and decrease their
carriers; as a consequence, they act on hyperfunctions in domains in R𝑛 and decrease
their supports.
The motivation for introducing and studying hyperdifferential operators lies, in
part, in two important results in Sections 8.1 and 8.2:
1) There is a one-to-one correspondence between hyperdifferential operators in
ΩC and endomorphisms of the sheaf of germs of holomorphic functions in ΩC
(Theorem 8.1.17). Were it not that this is a book about PDEs the title of this chapter
might have read something like Sheaf Morphisms in the Complex-Analytic Category.
2) Locally, every hyperfunction can be represented by the action of a hyperdiffer-
ential operator with constant coefficients on a C ∞ function (Theorem 8.2.3).
Section 8.3 is devoted to elliptic hyperdifferential operators; they are the obvious
generalization of elliptic (analytic) differential operators. Elliptic hyperdifferential
operators with constant coefficients 𝑃 (D) are analytic hypoelliptic in hyperfunctions,
meaning that, if 𝑢 ∈ B (𝑈) and 𝑃 (D) 𝑢 ∈ C 𝜔 (𝑈) (𝑈 ⊂ R𝑛 open) then 𝑢 ∈ C 𝜔 (𝑈).
The microlocal version of this result is stated and proved at the end of the chapter.
Using entire functions of infra-exponential type allows one to prove an extension
(due to P. Schapira) of the classical Petrowski Theorem 4.1.1: a differential operator
with constant coefficients, 𝑃 (D), is elliptic if and only if every solution 𝑢 ∈ B (𝑈) of

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 221
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_8
222 8 Hyperdifferential Operators

the homogeneous equation 𝑃 (D) 𝑢 = 0 is a distribution in 𝑈 (and then, of course, 𝑢 is

C 𝜔 in 𝑈). A somewhat curious consequence of this is that the heat equation, or more
generally C ∞ hypoelliptic partial differential equations with constant coefficients that
are not elliptic, have solutions that are not even distributions while the right-hand
side is analytic.
Section 8.4 presents the proofs of the following two results about the solvability
of a hyperdifferential operator with constant coefficients, 𝐽 (𝜕𝑧 ) ≠ 0 (generalizing
earlier results about PDEs with constant coefficients):

If ΩC ⊂ C𝑛 is convex then 𝐽 (𝜕𝑧 ) O ΩC = O ΩC (proved in [Martineau, 1967])
and its consequence:
If Ω ⊂ R𝑛 is bounded then 𝐽 (𝜕𝑥 ) B (Ω) = B (Ω).
Much of the material in this chapter has been suggested or provided by P.D. Cor-
daro.

8.1 Action on Holomorphic Functions and on Hyperfunctions

8.1.1 Hyperdifferential operators with constant coefficients

Definition 8.1.1 An entire function 𝐽 ∈ O (C𝑛 ) is said to be of infra-exponential

type if to every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that

∀𝜁 ∈ C𝑛 , |𝐽 (𝜁)| ≤ 𝐶 𝜀 e 𝜀 |𝜁 | . (8.1.1)

Classically, entire functions of infra-exponential type are called entire functions

of exponential order 1 and type 0. For this reason we shall denote by Exp1,0 (C𝑛 )
the set of entire functions of infra-exponential type in C𝑛 . These functions can be
characterized by a growth condition on the coefficients of their Taylor series:

Proposition 8.1.2 For 𝐽 ∈ O (C𝑛 ) to be of infra-exponential type it is necessary and

sufficient that to every 𝜀 > 0 there is a 𝐵 𝜀 > 0 such that

∀𝛼 ∈ Z+𝑛 , |𝐽 ( 𝛼) (0) | ≤ 𝐵 𝜀 𝜀 | 𝛼 | . (8.1.2)

Í
If we write 𝐽 (𝜁) = 𝛼∈Z+𝑛 𝑐 𝛼 𝜁 𝛼 , (8.1.2) is equivalent to

𝜀|𝛼|
∀𝛼 ∈ Z+𝑛 , |𝑐 𝛼 | ≤ 𝐵 𝜀 . (8.1.3)
𝛼!
Proof Apply the Cauchy inequalities and Stirling’s formula (cf. the proof of Theorem
6.2.2). □

Corollary 8.1.3 If 𝐽 ∈ Exp1,0 (C𝑛 ) then the same is true of all its partial derivatives
𝐽 (𝛽) , 𝛽 ∈ Z+𝑛 .
8.1 Action on Holomorphic Functions and on Hyperfunctions 223

(𝛽) (𝛽)
Proof Indeed, (8.1.2) implies |𝐽 (𝛽+𝛼) (0)| ≤ 𝐵 𝜀 𝜀 | 𝛼 | for all 𝛼, 𝛽 ∈ Z+𝑛 with 𝐵 𝜀 =
𝜀 |𝛽 | 𝐵 𝜀 . □
The relevance of Definition 8.1.1 to the topics of this chapter lies in the following
Proposition 8.1.4 For the analytic functional 𝜇 = 𝛼∈Z+𝑛 𝑐 𝛼 𝛿 ( 𝛼) to be carried by
Í
the origin it is necessary and sufficient that its Laplace–Borel transform L𝜇 (𝜁) be
of infra-exponential type.
Proof The condition is necessary: if 𝜇 ∈ O ′ ({0}) then, given any neighborhood
𝑈 of 0 in C𝑛 , 𝜇, e−𝑧·𝜁 ≤ 𝐶𝑈 sup e |𝑧 | |𝜁 | . Its sufficiency follows from Proposition
𝑧 ∈𝑈
6.2.4. □
We note that, on the one hand, Exp1,0 (C𝑛 ) is a ring with respect to ordinary
addition and multiplication, stable under translations and linear changes of variables.
On the other hand O ′ ({0}) is a convolution ring, by Proposition 6.2.6.
Proposition 8.1.5 The Laplace–Borel transform is a ring isomorphism of O ′ ({0})
(with convolution as the composition law) onto Exp1,0 (C𝑛 ).
We also derive from Proposition 6.2.6:
Proposition 8.1.6 Let 𝐾 be a Runge compact subset of C𝑛 . If 𝜈 ∈ O ′ ({0}) the
convolution 𝜇 ↦→ 𝜇 ∗ 𝜈 is a continuous linear map of O ′ (𝐾) into itself.
Actually we can interpret the operator 𝜈∗ as an analytic differential operator of
infinite order with constant coefficients. The coefficients in the series
∑︁
𝜈= 𝑐 𝛼 𝛿 ( 𝛼) (8.1.4)
𝛼∈Z+𝑛

satisfy (8.1.3). Thanks to the convergence of the series, using the Laplace–Borel
transform L𝜈 (𝜁) we get
∑︁
𝜇∗𝜈= 𝑐 𝛼 𝜕 𝛼 𝜇 = L𝜈 (𝜕𝑧 ) 𝜇. (8.1.5)
𝛼∈Z+𝑛

By (6.2.12), we have
L𝜈 (𝜕𝑧 ) 𝜇 = L −1 (L𝜈L𝜇) . (8.1.6)
In view of this it is natural to introduce the following terminology.
Í
Definition 8.1.7 If 𝐽 (𝜁) = 𝛼∈Z+𝑛 𝑐 𝛼 𝜁 𝛼 is an entire function of infra-exponential
type in C𝑛 then the operator
∑︁
𝐽 (𝜕𝑧 ) = 𝑐 𝛼 𝜕𝑧𝛼 , (8.1.7)
𝛼∈Z+𝑛

shall be called a hyperdifferential operator with constant coefficients in C𝑛 with

symbol 𝐽 (𝜁).
224 8 Hyperdifferential Operators

Proposition 8.1.6 can be rephrased:

Proposition 8.1.8 If 𝐾 is a Runge compact subset of C𝑛 then every hyperdifferential

operator 𝐽 (𝜕𝑧 ) defines a continuous linear map of O ′ (𝐾) into itself.

By combining Proposition 8.1.8 with Corollary 6.3.12 we obtain

Corollary 8.1.9 If 𝐽 (𝜕𝑥 ) is a hyperdifferential operator in C𝑛 and if 𝜇 ∈ O ′ (R𝑛 )

then 𝐽 (𝜕𝑥 ) 𝜇 ∈ O ′ (R𝑛 ) and supp 𝐽 (𝜕𝑧 ) 𝜇 ⊂ supp 𝜇.

By transposition Proposition 8.1.8 yields

Proposition 8.1.10 Let 𝐾 be a Runge compact subset of C𝑛 . If 𝐽 ∈ Exp1,0 (C𝑛 ) then

𝑓 ↦→ 𝐽 (𝜕𝑧 ) 𝑓 is a continuous endomorphism of O (𝐾).

A direct proof of Proposition 8.1.10, based on estimates, is also possible. Actually,

we have
Í
Proposition 8.1.11 Let 𝐽 (𝜁) = 𝛼∈Z_+𝑛 𝑐 𝛼 𝜁 𝛼 ∈ Exp1,0 (C𝑛 ) and let ΩC be an open
subset of C𝑛 . The linear operator
∑︁
𝑓 ↦→ 𝐽 (𝜕𝑧 ) 𝑓 = 𝑐 𝛼 𝜕𝑧𝛼 𝑓 (8.1.8)
𝛼∈Z+𝑛

maps O (ΩC ) continuously into itself.

Proof If 𝐾 and 𝐾 ′ are compact subsets of ΩC with 𝐾 ⊂ Int 𝐾 ′ then, by the Cauchy
estimates, there is a constant 𝑀 = 𝑀 (𝐾, 𝐾 ′) > 0 such that

sup |𝜕 𝛼 𝑓 | ≤ 𝑀 | 𝛼 | 𝛼! sup | 𝑓 | (8.1.9)
𝐾 𝐾′

for all 𝑓 ∈ O (ΩC ), 𝛼 ∈ Z+ . Let 𝜀 > 0 be such that 𝜀𝑀 < 1 and 𝐵 𝜀 > 0 such that
(8.1.3) holds; we derive from (8.1.9)

∑︁ © ∑︁
sup 𝑐 𝛼 𝜕𝑧𝛼 𝑓 ≤ 𝐵 𝜀 ­ (𝜀𝑀) | 𝛼 | ® sup | 𝑓 | ,
ª
𝐾 𝛼∈Z+𝑛 𝑛 𝐾′
« 𝛼∈Z+ ¬

which shows that the series in (8.1.8) converges in O (ΩC ) and that (8.1.8) is a
continuous map. □
If 𝑓 ∈ O (ΩC ) and 𝜇 ∈ O ′ (ΩC ) then

⟨𝐽 (𝜕𝑧 )𝜇, 𝑓 ⟩ = ⟨𝜇, 𝐽 (−𝜕𝑧 ) 𝑓 ⟩ . (8.1.10)

Recall that O ′ (𝐾) is an O (𝐾)-module. The next statement is the Leibniz rule for
hyperdifferential operators.
8.1 Action on Holomorphic Functions and on Hyperfunctions 225

Proposition 8.1.12 Let 𝐾 be a Runge compact set and 𝐽 (𝜕𝑧 ) a hyperdifferential

operator in C𝑛 . If 𝜇 ∈ O ′ (𝐾) and 𝑓 ∈ O (𝐾) then
∑︁ 1
𝐽 (𝜕𝑧 )( 𝑓 𝜇) = 𝐽 ( 𝛼) (𝜕𝑧 ) 𝑓 𝜕𝑧𝛼 𝜇 . (8.1.11)
𝛼∈Z𝑛
𝛼!
+

Proof Since O ′ (𝐾) can be identified with a linear subspace of O ′ (C𝑛 ) we may as
well prove the statement for all 𝜇 ∈ O ′ (C𝑛 ) and all entire functions 𝑓 . It is readily
derived from (8.1.2) and the Cauchy inequalities that the series
∑︁ (−1) | 𝛼 |
𝜕𝑧𝛼 ℎ𝐽 ( 𝛼) (𝜕𝑧 ) 𝑓
𝛼∈Z𝑛
𝛼!
+

converges in O (C𝑛 ) for every ℎ ∈ O (C𝑛 ); (8.1.11) ensues directly by duality. □

8.1.2 Hyperdifferential operators with variable coefficients. Definition

and localness

Presently we are going to need the natural topology on the ring Exp1,0 (C𝑛 ) of
entire functions of infra-exponential type (Definition 8.1.1). This topology is locally
convex; it can be defined by the norms:

Exp1,0 (C𝑛 ) ∋ ℎ ↦→ ∥ℎ∥ 𝜀 = sup e−𝜀 |𝜁 | ℎ (𝜁) (𝜀 > 0). (8.1.12)

𝜁 ∈C𝑛

Letting 𝜀 range over the sequence 𝑁 −1 , 𝑁 = 1, 2, ..., shows that Exp1,0 (C𝑛 ) is
metrizable; it is immediately checked that Exp1,0 (C𝑛 ) is complete. The topology is
compatible with the ring structure: Exp1,0 (C𝑛 ) is a Fréchet algebra.
Let ΩC be an open subset of C𝑛 and E be a complex topological vector space
(TVS). A map 𝒉 : ΩC −→ E is said to be holomorphic if 𝑤 −1 (𝒉 (𝑧 + 𝑤) − 𝒉 (𝑧))
converges in E as |𝑤| → 0, uniformly with respect to 𝑧 on compact subsets of ΩC .
Most of the properties (standard when dim E < ∞) of holomorphic functions, such
as the Cauchy–Riemann equations, convergent Taylor expansions, etc., hold when
dim E = +∞ (generalization of integration on curves and of the Cauchy Integral
Theorem requires that every Cauchy sequence in E converges, i.e., that E is complete).
We equip O (ΩC ; E) with the topology of uniform convergence on compact subsets
of ΩC . If E is a locally convex TVS, a sequence {𝒉 𝜈 } 𝜈=1,2,... ⊂ O (ΩC ; E) converges
to 𝒉 ∈ O (ΩC ; E) if and only if, given any continuous seminorm ℘ on E and any
compact set 𝐾 ⊂ ΩC , lim sup𝑧 ∈𝐾 ℘ (𝒉 (𝑧) − 𝒉 𝜈 (𝑧)) = 0.
𝜈→+∞
We shall denote by O ΩC ; Exp1,0 (C𝑛 ) the set of holomorphic maps 𝐽 : ΩC −→

Exp1,0 (C𝑛 ). [For the initiated let us mention that both O ΩC and Exp1,0 (C𝑛 ) are

nuclear spaces and that O ΩC ; Exp1,0 (C𝑛 ) can be identified with the completion

tensor product O ΩC b ⊗Exp1,0 (C𝑛 ).] The map 𝐽 can be identified with a holomor-
226 8 Hyperdifferential Operators

phic function 𝐽 (𝑧, 𝜁) in ΩC × C𝑛 , of infra-exponential type with respect to 𝜁. The

latter means that to every compact set 𝐾 ⊂ ΩC and to every 𝜀 > 0 there is a constant
𝐶𝐾 , 𝜀 such that
∀𝜁 ∈ C𝑛 , max |𝐽 (𝑧, 𝜁)| ≤ 𝐶𝐾 , 𝜀 e 𝜀 |𝜁 | . (8.1.13)
𝑧 ∈𝐾

Of course, this is equivalent to estimates of the kind

∀𝛼 ∈ Z+𝑛 , max |𝜕𝜁𝛼 𝐽 (𝑧, 0) | ≤ 𝐵 𝐾 , 𝜀 𝜀 | 𝛼 | , (8.1.14)

𝑧 ∈𝐾

or
𝜀|𝛼|
∀𝛼 ∈ Z+𝑛 , max |𝑐 𝛼 (𝑧) | ≤ 𝐵 𝐾 , 𝜀 (8.1.15)
𝑧 ∈𝐾 𝛼!
for the coefficients of the Taylor expansion
∑︁
𝐽 (𝑧, 𝜁) = 𝑐 𝛼 (𝑧) 𝜁 𝛼 . (8.1.16)
𝛼∈Z+𝑛

Proposition 8.1.13 As 𝑁 → +∞ the differential operators

∑︁
O ΩC ∋ ℎ ↦→ 𝐽 𝑁 (𝑧, 𝜕𝑧 ) ℎ = 𝑐 𝛼 𝜕𝑧𝛼 ℎ (8.1.17)
|𝛼|≤𝑁

converge to a continuous linear endomorphism O ΩC ←↪.

Proof Let 𝐾, 𝐾 ′ ⊂ ΩC be two compact sets such that 𝐾 ⊂ Interior (𝐾 ′). By

combining the Cauchy inequalities for ℎ with (8.1.15) we get, for a suitably large
constant 𝐶 > 0 and all 𝛼 ∈ Z+𝑛 ,

sup 𝑐 𝛼 (𝑧) 𝜕𝑧𝛼 ℎ (𝑧) ≤ 𝐵 𝐾 , 𝜀 𝐶 | 𝛼 |+1 𝜀 | 𝛼 | sup |ℎ (𝑧)| .

𝑧 ∈𝐾 𝑧 ∈𝐾 ′

The claim is a direct consequence of this inequality where we take 𝜀 < 𝐶 −1 . □

Definition 8.1.14 The limit of the differential operators 𝐽 𝑁 (𝑧, 𝜕𝑧 ) as 𝑁 → +∞ shall

be denoted by 𝐽 (𝑧, 𝜕𝑧 ) and called a hyperdifferential operator in ΩC .

We shall denote by Hyperdiff ΩC the set of hyperdifferential operators in ΩC ;
by Proposition 8.1.13 Hyperdiff ΩC is a linear subspace of the algebra of endo-
morphisms of O ΩC .

Proposition 8.1.15 Let 𝑈 ⊂ 𝑉 be two open subsets of C𝑛 and let 𝜌𝑈 𝑉

: O (𝑉) −→
O (𝑈) denote the restriction mapping. If 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff (𝑉) then
𝑉 𝑉
𝐽 (𝑧, 𝜕𝑧 ) 𝜌𝑈 = 𝜌𝑈 𝐽 (𝑧, 𝜕𝑧 ) . (8.1.18)

The claim is self-evident.

The Leibniz rule (8.1.11) extends to hyperdifferential operators 𝐽 (𝑧, 𝜕𝑧 ) if we
define 𝐽 ( 𝛼) (𝑧, 𝜁) = 𝜕𝜁𝛼 𝐽 (𝑧, 𝜁).
8.1 Action on Holomorphic Functions and on Hyperfunctions 227

Following [Ishimura, 1978] we now show that the concept of hyperdifferential

operator is identical to that of local operator in O Ω C , by which we mean a

continuous endomorphism of the presheaf O (𝑈) , 𝜌𝑈 𝑉 over Ω and, thereby, an

C

endomorphism of the sheaf of germs of holomorphic functions in ΩC (Subsection

1.2.2). This statement can be regarded as the complex-analytic analogue of the
theorem in [Peetre, 1960] according to which every local endomorphism of C ∞ (Ω),
Ω a domain in R𝑛 (in the C ∞ class local means that the map decreases supports), is
a differential operator. We give an explicit definition:

Definition 8.1.16 Let ΩC be an open subset
of C𝑛 . By a local operator in O ΩC
we shall mean a linear map 𝐿 : O ΩC ←↪ such that, given an arbitrary open subset
𝑈 of ΩC , there is a continuous linear map 𝐿𝑈 : O (𝑈) ←↪ having the properties that

(1) 𝐿𝑈 ( ℎ|𝑈 ) = 𝐿ℎ|𝑈 for every ℎ ∈ O ΩC ;
(2) if 𝑉 ⊂ 𝑈 is open then 𝐿𝑈 ℎ| 𝑉 = 𝐿 𝑉 ( ℎ| 𝑉 ) for every ℎ ∈ O (𝑈).

When 𝐿 is a local operator in O ΩC we shall say that the associated maps 𝐿𝑈
are induced by 𝐿 on O (𝑈).

8.1.17 Let Ω be an open subset of C . For a linear operator 𝐿 :

Theorem C 𝑛

O Ω C ←↪ to be local (Definition 8.1.16) it is necessary and sufficient that

𝐿 ∈ Hyperdiff ΩC .

Proof It suffices to prove the necessity of the condition since the sufficiency is the
content of Proposition 8.1.13. Thus suppose 𝐿 is local. We have 𝐿 (𝑧 𝛼 ) ∈ O ΩC
whatever 𝛼 ∈ Z+𝑛 . Define
∑︁ ∑︁ (−1) | 𝛼−𝛽 |
𝐽 (𝑧, 𝜕𝑧 ) = 𝑐 𝛼 (𝑧) 𝜕𝑧𝛼 , 𝑐 𝛼 (𝑧) = 𝑧 𝛼−𝛽 𝐿 𝑧 𝛽 . (8.1.19)
𝛼∈Z+𝑛 𝛽⪯𝛼
(𝛼 − 𝛽)!

Then, for arbitrary 𝛾 ∈ Z+𝑛 ,

∑︁ ∑︁ (−1) | 𝛼−𝛽 | 𝛾!
𝐽 (𝑧, 𝜕𝑧 ) 𝑧 𝛾 = 𝐿 𝑧 𝛽 𝑧 𝛾−𝛽
𝛼 ⪯𝛾 𝛽 ⪯ 𝛼
(𝛼 − 𝛽)! (𝛾 − 𝛼)!
!
∑︁ ∑︁
| 𝛼−𝛽 | (𝛾 − 𝛽)! 𝛾!
= (−1) 𝐿 𝑧 𝛽 𝑧 𝛾−𝛽 ,
𝛽 ⪯𝛾 𝛽 ⪯ 𝛼 ⪯𝛾
(𝛼 − 𝛽)! (𝛾 − 𝛼)! (𝛾 − 𝛽)!

implying 𝐿 (𝑧 𝛾 ) = 𝐽 (𝑧, 𝜕𝑧 ) 𝑧 𝛾 and therefore 𝛾!𝑐 𝛾 (𝑧 ∗ ) = 𝐽 (𝑧, 𝜕𝑧 ) (𝑧 − 𝑧∗ ) 𝛾 | 𝑧=𝑧 ∗ =

𝐿 (𝑧 − 𝑧∗ ) 𝛾 | 𝑧=𝑧 ∗ whatever 𝑧 ∗ ∈ ΩC . Let then 𝑧 ◦ ∈ ΩC be arbitrary and 𝑟 > 0 be such
that Δ𝑟(𝑛) (𝑧◦ ) = {𝑧 ∈ C𝑛 ; |𝑧 − 𝑧◦ | < 𝑟} ⊂⊂ ΩC . We avail ourselves of the continuity
(𝑛)
of the map 𝐿𝑈 induced by 𝐿 on O (𝑈) when 𝑈 = Δ𝑟/2 (𝑧 ◦ ): to each 𝜀 ∈ (0, 𝑟/2)
◦
there is a 𝐶 𝜀 (𝑧 ) > 0 independent of 𝛾 and such that

𝛾! sup 𝑐 𝛾 (𝑧) ≤ 𝐶 𝜀 (𝑧 ◦ ) sup (𝑧 − 𝑧◦ ) 𝛾 = 𝐶 𝜀 (𝑧◦ ) 𝜀 |𝛾 | .

|𝑧−𝑧 ◦ |< 𝜀/2 |𝑧−𝑧 ◦ |< 𝜀
228 8 Hyperdifferential Operators

By the Borel–Lebesgue Lemma it follows that if 𝐾 is an arbitrary compact subset of

ΩC and if 𝜀 < dist 𝐾, C𝑛 \ΩC then

− |𝛾 |
sup 𝛾!𝜀 sup 𝑐 𝛾 (𝑧) < +∞. □
𝛾 ∈Z+𝑛 𝑧 ∈𝐾

Now let 𝐾 ⊂ C𝑛 be a compact set and E be a complex topological vector

space. We introduce the germs at 𝐾 (Definition 1.2.14) of E-valued holomorphic
functions; they form a vector space O (𝐾; E). The topology of O (𝐾; E) is defined
in the same manner as in the finite-dimensional case, at least when E is locally
convex: it is the weakest locally convex topology that renders the “restriction” map
𝜌𝑈𝐾 : O (𝑈; E) −→ O (𝐾; E) continuous, whatever the open set 𝑈 ⊂ C such that
𝑛

𝐾 ⊂ 𝑈.
We now limit our attention to E = Exp1,0 (C𝑛 ), which is a Fréchet space. A
sequence {𝐹𝜈 } 𝜈=1,2,... ⊂ O (𝐾; Exp1,0 (C𝑛 )) converges in O (𝐾; Exp1,0 (C𝑛 )) if and
only if there is an open set 𝑈 ⊂ C𝑛, 𝐾 ⊂ 𝑈, such that {𝐹𝜈 } 𝜈=1,2,... is contained
and converges in O 𝑈; Exp1,0 (C𝑛 ) . We exploit Propositions 8.1.13 and 8.1.15:
if 𝐽 (𝑧, 𝜁) ∈ O (𝑈; Exp1,0 (C𝑛 )) and if 𝑉 ranges over the family of open sets such
that 𝐾 ⊂ 𝑉 ⊂ 𝑈, the linear operators 𝐽 (𝑧, 𝜕𝑧 ) : O (𝑉) ←↪ are continuous (which
is the same as bounded). Since every germ of holomorphic function at 𝐾 has a
representative belonging to O (𝑉) for some such 𝑉 those linear operators define a
continuous linear operator O (𝐾) ←↪, which we also denote by 𝐽 (𝑧, 𝜕𝑧 ).

Definition 8.1.18 Let 𝐾 ⊂ C𝑛 be a compact set. The linear operator 𝐽 (𝑧, 𝜕𝑧 ) :

O (𝐾) ←↪ will be called the germ of a hyperdifferential operator at 𝐾. The set of
germs of hyperdifferential operators at 𝐾 shall be denoted by Hyperdiff (𝐾).

In dealing with elements of Hyperdiff (𝐾) most of the time we omit the terms
“germ of” and simply speak of hyperdifferential operators at 𝐾. For later reference
we restate the observation that led to Definition 8.1.18:

Proposition 8.1.19 If 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff (𝐾) then the linear operator 𝐽 (𝑧, 𝜕𝑧 ) :
O (𝐾) ←↪ is continuous.

The definition of hyperdifferential operators in real space is practically self-

evident. Let Ω be an arbitrary open subset of R𝑛 and let E be a locally convex
TVS. Then C 𝜔 (Ω; E) can be defined as the space of “germs at Ω” of E-valued
holomorphic functions
C in C , which is to say: the quotient of the disjoint
𝑛 union of
the spaces O Ω , Ω ranging over the family of open subsets of C𝑛 that contain
C

Ω, modulo the equivalence

relation of being equal when restricted to Ω. If 𝐽 (𝑥, 𝜁) ∈
C 𝜔 Ω; Exp1,0 (C𝑛 ) we define an operator acting on a function 𝑓 ∈ C 𝜔 (Ω) in
the obvious way: there is an open subset ΩC of C𝑛 that contains Ω to which both
𝐽 (𝑥, 𝜁) and 𝑓 (𝑥) extend holomorphically, as 𝐽˜ (𝑧, 𝜁) ∈ O ΩC ; Exp1,0 (C𝑛 ) and
𝑓˜ (𝑧) ∈ O ΩC . We define, then,

𝐽 (𝑥, 𝜕𝑥 ) 𝑓 = 𝐽˜ (𝑧, 𝜕𝑧 ) 𝑓˜ Ω . (8.1.20)

8.1 Action on Holomorphic Functions and on Hyperfunctions 229

Proposition 8.1.20 The map 𝑓 ↦→ 𝐽 (𝑥, 𝜕𝑥 ) 𝑓 is a continuous endomorphism of

C 𝜔 (Ω).

Proof Follows directly from the definition of the topology of C 𝜔 (Ω) and from
Proposition 8.1.13. □

Definition 8.1.21 Let Ω be an open subset of R𝑛 . If 𝐽 (𝑥, 𝜁) ∈ C 𝜔 Ω; Exp1,0 (C𝑛 )
we shall refer to 𝐽 (𝑥, 𝜕𝑥 ) as a hyperdifferential operator in Ω. The set of hyperdif-
ferential operators in Ω shall be denoted by Hyperdiff (Ω).

8.1.3 Composition and transposition

As before ΩC denotes an open subset of C𝑛 .

Proposition 8.1.22 The set
Hyperdiff ΩC is a subring of the ring of continuous
endomorphisms of O ΩC .

Proof Let 𝐽 𝜄 (𝑧, 𝜁) ∈ O ΩC ; Exp1,0 (C𝑛 ) , 𝜄 = 1, 2. We have, for 𝑓 ∈ O ΩC ,

𝐽1 (𝑧, 𝜕𝑧 ) (𝐽2 (𝑧, 𝜕𝑧 ) 𝑓 ) = (𝐽1 #𝐽2 ) (𝑧, 𝜕𝑧 ) 𝑓 ,

where ∑︁ 1
(𝐽1 #𝐽2 ) (𝑧, 𝜁) = 𝜕𝜁𝛼 𝐽1 (𝑧, 𝜁) 𝜕𝑧𝛼 𝐽2 (𝑧, 𝜁) . (8.1.21)
𝛼∈Z𝑛
𝛼!
+

Indeed this is true for differential operators such as (8.1.17) [cf. (1.3.9)] and therefore
remains so when 𝑁 → +∞. [Since we have used the Laplace–Borel transform to
define the symbols, Formula (8.1.21) does not make use of D𝑧 , contrary to (1.3.8).]
Taylor expansions with respect to 𝜁 allows us to rewrite (8.1.21) as
∑︁ 1
𝛽

𝛾

(𝐽1 #𝐽2 ) (𝑧, 𝜁) = 𝜕𝜁𝛼 𝜕𝜁 𝐽1 (𝑧, 0) 𝜕𝑧𝛼 𝜕𝜁 𝐽2 (𝑧, 0) 𝜁 𝛽+𝛾 (8.1.22)
𝛼∈Z+𝑛
𝛼!𝛽!𝛾!
∑︁ ∑︁ 𝑖 𝛽−𝛾
𝛼+𝛽−𝛾

𝛾

= 𝜕𝜁 𝐽1 (𝑧, 0) 𝜕𝑧𝛼 𝜕𝜁 𝐽2 (𝑧, 0) 𝜁 𝛽 .
𝛼,𝛽 ∈Z+𝑛 𝛾 ⪯𝛽
𝛼! (𝛽 − 𝛾)!𝛾!

Let 𝐾, 𝐾 ′ ⊂ ΩC be two compact sets such that 𝐾 ⊂ Interior (𝐾 ′). The Cauchy
inequalities imply, for some constant 𝐶2 > 0 and all 𝛼 ∈ Z+𝑛 ,

1
sup 𝜕𝑧𝛼 𝜕𝜁 𝐽2 (𝑧, 0) ≤ 𝐶2| 𝛼 |+1 sup 𝜕𝜁 𝐽2 (𝑧, 0) .
𝛾 𝛾
𝛼! 𝑧 ∈𝐾 𝑧 ∈𝐾 ′

We also know by (8.1.14) that there is a constant 𝐵 𝐾 ′ , 𝜀 > 0 such that

∀𝛼 ∈ Z+𝑛 , sup |𝜕𝜁𝛼 𝐽 𝜄 (𝑧, 0) | ≤ 𝐵 𝐾 ′ , 𝜀 𝜀 | 𝛼 | , 𝜄 = 1, 2.

𝑧 ∈𝐾 ′
230 8 Hyperdifferential Operators

These two estimates imply, for every 𝑧 ∈ 𝐾,

1 𝛼+𝛽−𝛾
𝐽1 (𝑧, 0) 𝜕𝑧𝛼 𝜕𝜁 𝐽2 (𝑧, 0) ≤ 𝐵 𝐾 , 𝜀 𝐶2| 𝛼 |+1 𝜀 | 𝛼+𝛽−𝛾 | sup 𝜕𝜁 𝐽2 (𝑧, 0)
𝛾 𝛾
𝜕𝜁
𝛼! 𝑧 ∈𝐾 ′

≤ 𝐵 𝐾 , 𝜀 𝐵 𝐾 ′ , 𝜀 𝐶2| 𝛼 |+1 𝜀 | 𝛼+𝛽 | .

Assuming 𝐶1 𝜀 ≪ 1 and taking (8.1.22) into account we get, for a suitably large
constant 𝐵 𝐾 , 𝜀 ,

∑︁ ∑︁ 𝛽! 𝜀 |𝛽 | 𝛽
sup |(𝐽1 #𝐽2 ) (𝑧, 𝜁)| ≤ 𝐵 𝐾 , 𝜀 𝜁
𝑧 ∈𝐾 𝛽 ∈Z+𝑛 𝛾 ⪯𝛽
(𝛽 − 𝛾)!𝛾! 𝛽!

≤ 𝐵 𝐾 , 𝜀 exp (2𝑛𝜀 |𝜁 |) .

This proves that 𝐽1 #𝐽2 ∈ O ΩC ; Exp1,0 (C𝑛 ) . □

Corollary 8.1.23 Let 𝐾 be a compact subset of C𝑛 and 𝐽 (𝑧, 𝜕𝑧 ) a hyperdifferential

operator at 𝐾. If 𝜇 ∈ O ′ (𝐾) and 𝑔 ∈ O (𝐾) then
∑︁ 1
𝐽 (𝑧, 𝜕𝑧 )(𝑔𝜇) = 𝜕𝜁𝛼 𝐽 (𝑧, 𝜕𝑧 )𝑔 𝜕𝑧𝛼 𝜇 . (8.1.23)
𝛼∈Z𝑛
𝛼!
+

Proposition 8.1.24 Let ΩC𝑗 ( 𝑗 = 1, 2) be two open subsets of C𝑛 and let the map
Φ : ΩC1 −→ ΩC2 be a biholomorphism (i.e., a diffeomorphism which is holomorphic

and whose inverse is holomorphic). For arbitrary 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC1 and

ℎ ∈ O ΩC2 define
(Φ∗ 𝐽) (𝑧, 𝜕𝑧 ) ℎ = 𝐽 (𝑧, 𝜕𝑧 ) (ℎ ◦ Φ) . (8.1.24)

Then 𝐽 (𝑧, 𝜕𝑧 ) −→ (Φ∗ 𝐽) (𝑧, 𝜕𝑧 ) is a ring isomorphism of Hyperdiff ΩC1 onto

Hyperdiff ΩC2 .

Proof That (Φ∗ 𝐽) (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC2 follows immediately from Theorem
8.1.17 and the fact that (Φ∗ 𝐽) (𝑧, 𝜕𝑧 ) is local in ΩC2 . The map Φ∗ is a ring isomor-
phism because it is invertible and obviously transforms addition into addition and
composition into composition (of operators). □

Next we introduce the transpose of 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC . Assuming that
𝐽 (𝑧, 𝜁) is given by (8.1.16) and that ℎ ∈ O ΩC we define
∑︁
𝐽 (𝑧, 𝜕𝑧 ) ⊤ ℎ = (−1) | 𝛼 | 𝜕𝑧𝛼 (𝑐 𝛼 ℎ) . (8.1.25)
𝛼∈Z+𝑛

Proposition 8.1.25 If 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC the same is true of 𝐽 (𝑧, 𝜕𝑧 ) ⊤ .

Proof According to (8.1.25) the symbol of 𝐽 (𝑧, 𝜕𝑧 ) ⊤ is

8.1 Action on Holomorphic Functions and on Hyperfunctions 231
∑︁
𝐽 (𝑧, 𝜁) ⊤ = 𝑏 𝛼 (𝑧) 𝜁 𝛼 , (8.1.26)
𝛼∈Z+𝑛

where ∑︁ (𝛽 + 𝛼)! 𝛼
𝑏𝛼 = (−1) |𝛽+𝛼 | 𝜕 𝑎 𝛼+𝛽 .
𝛽 ∈Z+𝑛
𝛽!𝛼! 𝑧

Once again consider two compact sets 𝐾, 𝐾 ′ ⊂ ΩC such that 𝐾 ⊂ Interior (𝐾 ′). The
Cauchy inequalities imply, for some constant 𝐶 > 0 and all 𝛼,𝛽 ∈ Z+𝑛 ,

(𝛽 + 𝛼)!
sup 𝜕𝑧 𝑎 𝛼+𝛽 ≤ 𝐶 |𝛽 |+1 (𝛽 + 𝛼)!sup 𝑎 𝛼+𝛽
𝛽
𝛽! 𝐾 𝐾′
≤ 𝐶 |𝛽 |+1 𝐵 𝐾 ′ , 𝜀 𝜀 | 𝛼+𝛽 | ,

the latter inequality ensuing from (8.1.15). The sought conclusion is a consequence
of the estimates
∑︁
𝛼!𝜀 − | 𝛼 | sup |𝑏 𝛼 | ≤ 𝐶𝐵 𝐾 ′ , 𝜀 (𝐶𝜀) |𝛽 |
𝐾 𝛽 ∈Z+𝑛

< +∞ if 𝐶𝜀 < 1. □

It is a routine exercise to show that if 𝐾 ⊂ C𝑛 is a compact set then Hyperdiff (𝐾)

(Definition 8.1.18) is a subring of the ring of continuous endomorphisms of O (𝐾),
stable under the involution 𝐽 (𝑧, 𝜕𝑧 ) ↦→ 𝐽 (𝑧, 𝜕𝑧 ) ⊤ .
The next statements in this subsection pertain to C 𝜔 functions in open subsets
of R𝑛 ; they are direct consequences of Definition 8.1.21 and Propositions 8.1.22,
8.1.24, 8.1.25.

Proposition 8.1.26 If Ω is an open subset of R𝑛 then Hyperdiff (Ω) is a subring of

the ring of continuous endomorphisms of C 𝜔 (Ω).

Proposition 8.1.27 Let Ω 𝑗 ( 𝑗 = 1, 2) be two open subsets of R𝑛 and let the map
Φ : Ω1 −→ Ω2 be a C 𝜔 diffeomorphism. For arbitrary 𝐽 (𝑥, 𝜕𝑥 ) ∈ Hyperdiff (Ω1 )
and ℎ ∈ O (Ω2 ) define

(Φ∗ 𝐽) (𝑥, 𝜕𝑥 ) ℎ = 𝐽 (𝑥, 𝜕𝑥 ) (ℎ ◦ Φ) . (8.1.27)

Then 𝐽 (𝑥, 𝜕𝑥 ) ↦→ (Φ∗ 𝐽) (𝑥, 𝜕𝑥 ) is a ring isomorphism of Hyperdiff (Ω1 ) onto

Hyperdiff (Ω2 ).
232 8 Hyperdifferential Operators

8.1.4 Action of hyperdifferential operators on analytic functionals

As before, let ΩC ⊂ C𝑛 be an open set. Proposition 8.1.24 enables us to define the

action of 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC on the space of analytic functionals in ΩC :

⟨𝐽 (𝑧, 𝜕𝑧 ) 𝜇, ℎ⟩ = 𝜇, 𝐽 (𝑧, 𝜕𝑧 ) ⊤ ℎ , (8.1.28)

for 𝜇 ∈ O ′ ΩC , ℎ ∈ O ΩC . In view of (8.1.28) Propositions 8.1.13 and 8.1.19

have the following immediate consequences:

Proposition 8.1.28 If 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC then, whatever the open subset 𝑈
of ΩC , 𝜇 ↦→ 𝐽 (𝑧, 𝜕𝑧 ) 𝜇 is a continuous linear map of O ′ (𝑈) into itself. If 𝐾 ⊂ C𝑛
is a compact set and 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff (𝐾) then 𝜇 ↦→ 𝐽 (𝑧, 𝜕𝑧 ) 𝜇 is a continuous
linear map of O ′ (𝐾) into itself.

Below we are going to need the following

Proposition 8.1.29 Let 𝐾 be a compact of C𝑛 and 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff (𝐾). If

𝜇 ∈ O ′ (𝐾) and 𝑔 ∈ O (𝐾) then
∑︁ 1
𝐽 (𝑧, 𝜕𝑧 )(𝑔𝜇) = 𝜕𝜁𝛼 𝐽 (𝑧, 𝜕𝑧 )𝑔 𝜕𝑧𝛼 𝜇 . (8.1.29)
𝛼∈Z𝑛
𝛼!
+

The same statement holds with 𝐾 replaced by an arbitrary open subset ΩC of C𝑛 .

Proof It suffices to apply (8.1.11) taking into account the series expansion (8.1.16).□

A special case of Proposition 8.1.29 that will be useful in the next subsection is
that of 𝐾 = closure of 𝑉, 𝑉 being an open and bounded subset∫ of R , and 𝜇 = 𝜒𝐾 ,
𝑛

the characteristic function of 𝐾. In other words, ⟨𝜇, ℎ⟩ = 𝐾 ℎ(𝑥)d𝑥, ℎ ∈ O (C𝑛 ).

We recall that every compact subset of R𝑛 is Runge (Theorem 6.2.14). Denoting by
𝜕𝑉 the boundary of 𝑉 in R𝑛 we set (𝜕𝑉) 𝛿 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝜕𝑉) < 𝛿}, 𝛿 > 0. We
need the following

Lemma 8.1.30 There is a 𝐶 > 0 such that, for every 𝛿 > 0 sufficiently small and
every multi-index 𝛼, |𝛼| > 0,

| 𝜕𝑧𝛼 𝜒𝐾 , ℎ | ≤ 𝐶 | 𝛼 |+1 𝛿− | 𝛼 | 𝛼! sup |ℎ|, ℎ ∈ O (C𝑛 ). (8.1.30)

(𝜕𝑉) 𝛿

Proof We use Ehrenpreis’ cutoffs (Section 3.2). According to Proposition 3.2.1,

there is a constant 𝐶◦ > 0 with the following property: given arbitrary 𝑁 ∈ Z+ and
small 𝛿 > 0, there is a 𝜙 𝑁 , 𝛿 ∈ 𝐶c∞ (R𝑛 ) satisfying 𝜙(𝑥) = 1 if dist (𝑥, 𝜕𝑉) ≤ 𝛿/4,
|𝛽 |+1 |𝛽 |
𝑁 /𝛿 |𝛽 | if |𝛽| ≤ 𝑁. We have
𝛽
𝜙(𝑥) = 0 if dist (𝑥, 𝜕𝑉) ≥ 𝛿/2 and |𝜕𝑥 𝜙| ≤ 𝐶◦
∫ ∫
| 𝜕𝑧𝛼 𝜒𝐾 , ℎ | = 𝐾 𝜕𝑥𝛼 (𝜙ℎ)d𝑥 since 𝐾 𝜕𝑥𝛼 ((1 − 𝜙)ℎ) d𝑥 = 0. Consequently, taking
𝑁 = |𝛼| leads to the estimates
8.1 Action on Holomorphic Functions and on Hyperfunctions 233

| 𝜕𝑧𝛼 𝜒𝐾 , ℎ | ≤ 𝐶1 sup |𝜕𝑧𝛼 (𝜙 𝑁 , 𝛿 ℎ)|

(𝜕𝑉) 𝛿/2

𝛼! | 𝛼−𝛽 |
| 𝛼−𝛽 |+1 𝑁
∑︁
≤ 𝐶1 𝐶◦ sup |ℎ (𝛽) | .
𝛽⪯𝛼
(𝛼 − 𝛽)!𝛽! 𝛿 | 𝛼−𝛽 | (𝜕𝑉) 𝛿/2

The Cauchy inequalities yield

|𝛽 |
1 (𝛽) 2
sup |ℎ | ≤ sup |ℎ|,
𝛽! (𝜕𝑉) 𝛿/2 𝛿 (𝜕𝑉) 𝛿

whence
∑︁ 𝛼! | 𝛼−𝛽 | | 𝛼−𝛽 |
| 𝜕𝑧𝛼 𝜒𝐾 , ℎ | ≤ 𝐶◦ 𝐶1 𝛿− | 𝛼 | 𝐶 𝑁 sup |ℎ|.
𝛽⪯𝛼
(𝛼 − 𝛽)! ◦ (𝜕𝑉) 𝛿

| 𝛼|
If 𝛽 ⪯ 𝛼 we have 𝑁 | 𝛼−𝛽 | = |𝛼| | 𝛼−𝛽 | ≤ |𝛼| | 𝛼 | |𝛽| − |𝛽 | ≤ 𝑀 |𝛽 |+1 | 𝛼𝛽!| by Stirling’s
Formula. We obtain
|𝛼|
𝛼 | 𝛼 | |𝛼|
| 𝜕𝑧 𝜒𝐾 , ℎ | ≤ 𝐶◦ 𝐶1 𝑀 (𝐶◦ + 𝑀) sup |ℎ| .
𝛿 (𝜕𝑉) 𝛿

Applying Stirling’s Formula once more yields (8.1.30). □

Proposition 8.1.31 Let 𝑉 be an open and bounded subset of R𝑛 and 𝐾 = 𝑉 its
closure. If 𝐽 (𝑧, 𝜕𝑧 ) is a hyperdifferential operator at 𝐾 and if 𝑔 ∈ O (𝐾) then
𝐽 (𝑧, 𝜕𝑧 ) (𝑔 𝜒𝐾 ) − (𝐽 (𝑧, 𝜕𝑧 ) 𝑔) 𝜒𝐾 ∈ O ′ (𝜕𝑉). More precisely, to each sufficiently
small 𝛿 > 0 there is a 𝐶 𝛿 > 0 such that, for all ℎ ∈ O (C𝑛 ),
!
|⟨ 𝐽 (𝑧, 𝜕𝑧 ) (𝑔 𝜒𝐾 ) − (𝐽 (𝑧, 𝜕𝑧 ) 𝑔) 𝜒𝐾 , ℎ⟩| ≤ 𝐶 𝛿 sup |𝑔| sup |ℎ|. (8.1.31)
(𝜕𝑉)2 𝛿 (𝜕𝑉) 𝛿

Proof By (8.1.23) we have, for an arbitrary ℎ ∈ O (C𝑛 ),

𝑇 (ℎ) = |⟨ 𝐽 (𝑧, 𝜕𝑧 ) (𝑔 𝜒𝐾 ) − ( 𝐽 (𝑧, 𝜕𝑧 ) 𝑔) 𝜒𝐾 , ℎ⟩|

∑︁ 1 D E
= 𝜕𝑧𝛼 𝜒𝐾 𝜕𝜁𝛼 𝐽 (𝑧, 𝜕𝑧 )𝑔, ℎ
𝛼!
| 𝛼 |>0

and thus, by Lemma 8.1.30,

∑︁ 𝐶 | 𝛼 |+1
𝛼
𝑇 (ℎ) ≤ sup ℎ 𝜕 𝐽 (𝑧, 𝜕𝑧 )𝑔 .
| 𝛼 |>0
𝛿 | 𝛼 | (𝜕𝑉) 𝛿 𝜁

Í
Writing 𝜕𝜁𝛼 𝐽 (𝑧, 𝜁) = 𝛽 ∈Z+𝑛 𝑐 𝛼,𝛽 (𝑧) 𝜁 𝛽 we deduce from (8.1.14) that to each 𝜀 > 0
there is a 𝐶 𝜀 > 0 such that |𝑐 𝛼,𝛽 | ≤ 𝐶 𝜀 𝜀 | 𝛼+𝛽 | /𝛽!. This implies
234 8 Hyperdifferential Operators
∑︁
𝛽
sup |𝜕𝜁𝛼 𝐽 (𝑧, 𝜕𝑧 )𝑔| ≤ |𝑐 𝛼,𝛽 | sup |𝜕𝑧 𝑔|
(𝜕𝑉) 𝛿 𝛽 ∈Z+𝑛 (𝜕𝑉) 𝛿

© ∑︁ 𝛽!
≤­ |𝑐 | ® sup |𝑔|,
ª
𝛿 |𝛽 | 𝛼,𝛽
(𝜕𝑉)2 𝛿
«𝛽 ∈𝑍+
𝑛
¬
the last inequality by the Cauchy inequalities. As a consequence,
!
| 𝛼 |+1 |𝑐
© ∑︁ ∑︁ 𝐶 𝛼,𝛽 |𝛽! ª
𝑇 (ℎ) ≤ ­ ® sup |𝑔|
𝛿 | 𝛼 |+ |𝛽 | (𝜕𝑉)2 𝛿
« | 𝛼 |>0 𝛽 ∈Z+
𝑛
¬
!
© ∑︁ ∑︁ 𝐶 | 𝛼 |+1 𝜀 | 𝛼 |+ |𝛽 | ª
≤ 𝐶𝜀 ­ ® sup |𝑔| sup |ℎ|.
| 𝛼 |>0 𝛽 ∈Z 𝑛 𝛿 | 𝛼 |+ |𝛽 | (𝜕𝑉)2 𝛿 (𝜕𝑉) 𝛿
« + ¬
Taking 𝜀 > 0 appropriately small gives the desired conclusion. □

8.1.5 Action of hyperdifferential operators on hyperfunctions

Let 𝐽 (𝑥, 𝜕𝑥 ) be a hyperdifferential operator in the open set Ω ⊂ R𝑛 . Let an open

subset ΩC of C𝑛 contain
Ω and be such that 𝐽 (𝑥, 𝜁) has an extension to 𝐽 (𝑧, 𝜁) ∈
O ΩC ; Exp1,0 (C𝑛 ) . Applying Proposition 8.1.28 to an arbitrary compact subset 𝐾
of Ω we can state (see Corollary 6.3.12):

Proposition 8.1.32 If 𝜇 ∈ O ′ (R𝑛 ) then supp 𝐽 (𝑧, 𝜕𝑧 ) 𝜇 ⊂ supp 𝜇.

Let then 𝑈 be an open subset of R𝑛 whose closure 𝑈 is a compact subset of Ω.

Since 𝐽 (𝑧, 𝜕𝑧 ) O ′ (𝑈) ⊂ O ′ (𝑈) and 𝐽 (𝑧, 𝜕𝑧 ) O ′ (𝜕𝑈) ⊂ O ′ (𝜕𝑈) we get canonically
an endomorphism 𝐽 (𝑥, 𝜕𝑥 ) of B (𝑈) = O ′ (𝑈)/O ′ (𝜕𝑈). Considering (8.1.20) this
notation is justified as follows. Through our identification of D ′ (Ω) with a linear
subspace of B (Ω) (Corollary 7.1.13) we may identify each subspace of D ′ (Ω), in
particular 𝐶 𝜔 (Ω), with a linear subspace of B (Ω). We know that 𝐽 (𝑥, 𝜕𝑥 ) defines
an endomorphism of 𝐶 𝜔 (Ω) (cf. Proposition 8.1.26).

Proposition 8.1.33 The diagram

𝐶 𝜔 (Ω) ↩→ B (Ω)
𝐽 (𝑥, 𝜕𝑥 ) ↓ ↓ 𝐽 (𝑥, 𝜕𝑥 ) (8.1.32)
𝐶 𝜔 (Ω) ↩→ B (Ω)

is commutative.
8.1 Action on Holomorphic Functions and on Hyperfunctions 235

Proof Indeed, each 𝑔 ∈ 𝐶 𝜔 (Ω) defines a hyperfunction on Ω which, on an arbitrary

open set 𝑉 ⊂⊂ Ω, is represented by the analytic functional 𝑔 𝜒𝑉¯ . Thus, the real-
analytic function 𝐽 (𝑥, 𝜕𝑥 ) 𝑔 is identified with the hyperfunction which is represented
on 𝑉 by the analytic functional (𝐽 (𝑥, 𝜕𝑥 ) 𝑔) 𝜒𝑉¯ . By Proposition 8.1.31 we have
𝐽 (𝑥, 𝜕𝑥 ) (𝑔 𝜒𝑉¯ ) − (𝐽 (𝑥, 𝜕𝑥 ) 𝑔) 𝜒𝑉¯ ∈ O ′ (𝜕𝑉), whence the claim. □
We continue to use the notation 𝜌𝑈
𝑉 for the restriction mapping, even when applied
to hyperfunctions: if 𝑉 ⊂ 𝑈 are open subsets of R𝑛 (possibly unbounded) then 𝜌𝑈 𝑉
maps B (𝑈) onto B (𝑉) (Proposition 7.1.9). If there is no risk of confusion we shall

𝑉 𝑢 when 𝑢 ∈B (𝑈).
also use the notation 𝑢| 𝑉 instead of 𝜌𝑈

Proposition 8.1.34 Let Ω be an open subset of R𝑛 and 𝐽 (𝑥, 𝜕𝑥 ) ∈ Hyperdiff (Ω). If

𝑉 ⊂ 𝑈 are bounded open subsets of Ω then 𝐽 (𝑥, 𝜕𝑥 ) 𝜌𝑈
𝑉 𝑓 = 𝜌𝑉 𝐽 (𝑥, 𝜕𝑥 ) 𝑓 whatever
𝑈

𝑓 ∈B (𝑈).

Proof Let 𝜇 ∈ O ′ (𝑈) represent 𝑢 ∈ B (𝑈), i.e., 𝑢 = [𝜇]. If we decompose 𝜇 = 𝜈 + 𝛼

(cf. the proof of Proposition 7.1.3) with 𝜈 ∈ O ′ (𝑉) and 𝛼 ∈ O ′ (𝑈 \ 𝑉), then [𝜈] =
′
𝜌𝑈𝑉 𝑢 and therefore 𝐽 (𝑥, 𝜕𝑥 ) 𝜌𝑉 𝑢 ∈ B (𝑉) is the coset of 𝐽 (𝑧, 𝜕𝑧 ) 𝜈 mod O (𝜕𝑉). On
𝑈

the other hand we have 𝐽 (𝑧, 𝜕𝑧 ) 𝜇 = 𝐽 (𝑧, 𝜕𝑧 ) 𝜈+ 𝐽 (𝑧, 𝜕𝑧 ) 𝛼, with 𝐽 (𝑧, 𝜕𝑧 ) 𝜈 ∈ O ′ (𝑉)
and 𝐽 (𝑧, 𝜕𝑧 ) 𝛼 ∈ O ′ (𝑈 \𝑉), which shows that 𝜌𝑈𝑉 𝐽 (𝑥, 𝜕𝑥 ) 𝑢 is the coset of 𝐽 (𝑥, 𝜕𝑥 ) 𝜈
mod O ′ (𝜕𝑉). □
Whether Ω ⊂ R𝑛 is bounded or not, given any open set 𝑈 ⊂⊂ Ω we have
Ω Ω
∀𝑢 ∈ B (Ω), 𝜌𝑈 𝐽 (𝑥, 𝜕𝑥 ) 𝑢 = 𝐽 (𝑥, 𝜕𝑥 ) 𝜌𝑈 𝑢. (8.1.33)

Next we turn our attention to boundary values of holomorphic functions in wedges

W𝛿 (Ω, Γ) [see (3.3.1); in this section we always take Ω to be an open subset of
R𝑛 , Γ ⊂ R𝑛 \ {0} an open cone and 𝛿 > 0]. Concerning the boundary value map
𝑏 Ω : O (W𝛿 (Ω, Γ)) −→ B (Ω) we refer to Definition 7.2.3 and Proposition 7.2.5.

Proposition 8.1.35 Let Ω be an open set in R𝑛 and let 𝐽 (𝑧, 𝜕𝑧 ) ∈ Hyperdiff ΩC𝛿 ,
where ΩC𝛿 = {𝑧 ∈ C𝑛 ; Re 𝜁 ∈ Ω, |Im 𝜁 | < 𝛿}. For every 𝑓 ∈ O (W𝛿 (Ω, Γ)) we
have
𝑏 Ω 𝐽 (𝑧, 𝜕𝑧 ) 𝑓 = 𝐽 (𝑥, 𝜕𝑥 ) 𝑏 Ω 𝑓 . (8.1.34)

Proof Let 𝑈 ⊂⊂ Ω be open and have a smooth boundary 𝜕𝑈 and let v ∈ Γ, |v| < 𝛿.
We shall use the notation (7.2.1): if 𝜙 ∈ O (W𝛿 (Ω, Γ)) and ℎ ∈ O (C𝑛 ),
D E ∫
𝜇𝑈
𝜙,v , ℎ = 𝜙(𝑧)ℎ(𝑧)d𝑧. (8.1.35)
𝑈+𝑖v

To simplify the notation we also set 𝑔 = 𝐽 (𝑧, 𝜕𝑧 ) 𝑓 ∈ O (W𝛿 (Ω, Γ)). In the proof
of Lemma 7.2.2 it is shown that there exist representatives 𝛼, 𝛽 ∈ O ′ (𝑈) of 𝑏 Ω 𝑓 |𝑈
and 𝑏 Ω 𝑔|𝑈 respectively, having the following properties: to each open subset of C𝑛 ,
𝑉 C ⊃ 𝜕𝑈, there is an 𝜀 > 0 such that |v| < 𝜀 implies that 𝛼 − 𝜇𝑈𝑓,v and 𝛽 − 𝜇𝑈𝑔,v are
236 8 Hyperdifferential Operators

carried by compact subsets of 𝑉 C . Since 𝐽 (𝑥, 𝜕𝑥 ) 𝑏 Ω 𝑓 |𝑈 is represented by 𝐽 (𝑧, 𝜕𝑧 ) 𝛼,

we must show that 𝐽 (𝑧, 𝜕𝑧 ) 𝛼 − 𝛽 ∈ O ′ (𝜕𝑈). We can write

𝐽 (𝑧, 𝜕𝑧 ) 𝛼−𝛽 = 𝐽 (𝑧, 𝜕𝑧 ) 𝛼 − 𝐽 (𝑧, 𝜕𝑧 ) 𝜇𝑈𝑓,v + 𝐽 (𝑧, 𝜕𝑧 ) 𝜇𝑈𝑓,v − 𝜇𝑈 𝑈
𝑔,v − 𝛽 − 𝜇 𝑔,v ,

which shows that it suffices to prove the following claim:

(•) given any open subset of C𝑛 , 𝑉 C ⊃ 𝜕𝑈, there is an 𝜀◦ > 0 such that
𝐽 (𝑧, 𝜕𝑧 ) 𝜇𝑈𝑓,v − 𝜇𝑈 ′
𝑔,v ∈ O (𝑉 ) if |v| < 𝜀 ◦ .
C

If we denote by 𝜆𝑈+𝑖v the analytic functional

∫
𝑛
O (C ) ∋ ℎ ↦→ ⟨𝜆𝑈+𝑖v , ℎ⟩ = ℎ(𝑧) d𝑧
𝑈+𝑖v

then, by (7.2.1), we can write 𝜇𝑈 𝜙,v = 𝜙𝜆𝑈+𝑖v . Thus 𝐽 (𝑧, 𝜕𝑧 ) 𝜇 𝑓 ,v = 𝐽 (𝑧, 𝜕𝑧 ) [ 𝑓 𝜆𝑈+𝑖v ]
𝑈

and 𝜇𝑔,v = (𝐽 (𝑧, 𝜕𝑧 ) 𝑓 )𝜆𝑈+𝑖v . We apply the Leibniz rule (8.1.29):

𝑈

∑︁ 1
𝐽 (𝑧, 𝜕𝑧 ) 𝜇𝑈𝑓,v − 𝜇𝑈
𝑔,v = 𝜕𝜁𝛼 𝐽 (𝑧, 𝜕𝑧 ) 𝑓 𝜕𝑧𝛼 𝜆𝑈+𝑖v .
𝛼!
| 𝛼 |>0

It is readily verified (cf. Proposition 8.1.6) that 𝐽 (𝑧, 𝜕𝑧 ) 𝜇𝑈𝑓,v − 𝜇𝑈

𝑔,v is carried by
𝜕𝑈 + 𝑖v, proving (•). □

Corollary 8.1.36 If Ω ⊂ R𝑛 is an open set and 𝐽 (𝑥, 𝜕𝑥 ) ∈ Hyperdiff (Ω) then

∀𝑢 ∈ B (Ω), 𝑊 𝐹a ( 𝐽 (𝑥, 𝜕𝑥 ) 𝑢) ⊂ 𝑊 𝐹a (𝑢). (8.1.36)

Proof Let (𝑥0 , 𝜉0 ) ∉ 𝑊 𝐹a (𝑢); according to Definition 7.4.7 there exist an open
subset 𝑈 of Ω and finitely many convex open cones Γ 𝑗 ⊂ R𝑛 \ {0} ( 𝑗 = 1, ..., 𝜈) with
the following properties:
𝜈
Ø
(1) 𝑥 ◦ ∈ 𝑈 and 𝜉 ◦ · 𝑦 < 0 for all 𝑦 ∈ Γ𝑗;
𝑗=1
Í
(2) there are functions 𝑓 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 (𝛿 > 0) such that 𝑢 = 𝜈𝑗=1 𝑏𝑈 𝑓 𝑗 .

Proposition 8.1.35 implies 𝜌𝑈 Ω 𝐽 (𝑥, 𝜕 ) 𝑢 = Í 𝑁 𝑏 ( 𝐽 (𝑧, 𝜕 ) 𝑓 ) in 𝑈 and,

𝑥 𝑗=1 𝑈 𝑧 𝑗
consequently, (𝑥0 , 𝜉0 ) ∉ 𝑊 𝐹a (𝐽 (𝑥, 𝜕𝑥 )𝑢). □
8.2 Local Representation of Hyperfunctions 237

8.2 Local Representation of Hyperfunctions

8.2.1 A class of entire functions of infra-exponential type

We shall use once again the notation ⟨𝜁⟩ 2 = 𝑛𝑗=1 𝜁 2𝑗 , 𝜁 = (𝜁1 , . . . , 𝜁 𝑛 ) ∈ C𝑛 . Given
Í
a monotone increasing map 𝜒 of (1, +∞) onto (0, +∞) we define
∞
⟨𝜁⟩ 2
Ö
𝑄 𝜒 (𝜁) = 1+ , 𝜁 ∈ C𝑛 . (8.2.1)
𝑝=1
𝑝 2 𝜒( 𝑝) 2

For 𝜌 > 0 (in practice, small) we introduce the cone

ℭ𝜌 = {𝜁 ∈ C𝑛 ; | Im 𝜁 | ≤ 𝜌| Re 𝜁 |}. (8.2.2)

Below we use the following functions of 𝜌 ∈ (0, 1):

√︄
1 − 𝜌2 1
𝜌˜ = 2
, 𝑐 𝜌 = log(1 + 𝜌),
˜ 𝐴𝜌 = .
1+𝜌 1 + 𝜌˜

Proposition 8.2.1 The function (8.2.1) is an entire function of infra-exponential type.

Moreover, there is an 𝑅 > 0 (depending only on 𝜒) such that

|𝑄 𝜒 (𝜁)| ≥ 𝐴𝜌 e𝑐𝜌 |𝜁 |/𝜒 ( |𝜁 |+1) (8.2.3)

for all 𝜁 ∈ ℭ𝜌 , |𝜁 | ≥ 𝑅.

Proof Let 𝜀 > 0 be arbitrary and 𝑞 ∈ Z+ be such that 𝜒( 𝑝) ≥ 1/𝜀 if 𝑝 > 𝑞; we have
∞ ∞
|𝜁 | 2 𝜀 2 |𝜁 | 2
Ö Ö
|𝑄 𝜒 (𝜁)| ≤ 1+ ≤ 𝐶 𝜀 1 + (8.2.4)
𝑝=1
𝑝 2 𝜒( 𝑝) 2 𝑝=1
𝑝2

where −1
𝑞
|𝜁 | 2 𝜀 2 |𝜁 | 2
Ö
𝐶 𝜀 = sup 1+ 2 1+ .
𝜁 ∈C𝑛 𝑝=1 𝑝 𝜒( 𝑝) 2 𝑝2

The Hadamard factorization ([Conway, 1973], I, p. 291)

∞
𝑧2

sin(𝜋𝑧) Ö
= 1 − 2 , 𝑧 ∈ C,
𝜋𝑧 𝑝=1
𝑝

implies, if 0 ≠ 𝜁 ∈ C𝑛 ,
∞
𝜀 2 |𝜁 | 2 e2 𝜋 𝜀 |𝜁 | − 1

Ö sin(𝜋𝑖𝜀|𝜁 |)
1+ = ≤ ≤ e2 𝜋 𝜀 |𝜁 | .
𝑝=1
𝑝2 𝜋𝑖𝜀|𝜁 | 2𝜋𝜀|𝜁 |
238 8 Hyperdifferential Operators

Putting this into (8.2.4) shows that 𝑄 𝜒 ∈ Exp1,0 (C𝑛 ).

Next we prove (8.2.3). Let 𝜁 ∈ ℭ𝜌 be arbitrary; then

Re⟨𝜁⟩ 2 = | Re 𝜁 | 2 − | Im 𝜁 | 2
≥ (1 − 𝜌 2 )| Re 𝜁 | 2 = 𝜌˜ 2 (1 + 𝜌 2 )| Re 𝜁 | 2
≥ 𝜌˜ 2 |𝜁 | 2

and thus, for every 𝑞 = 1, 2, . . .,

∞ ∞
⟨𝜁⟩ 2 Re⟨𝜁⟩ 2
Ö Ö
|𝑄 𝜒 (𝜁)| = 1+ 2 ≥ 1+ 2 ,
𝑝=1
𝑝 𝜒( 𝑝) 2 𝑝=1
𝑝 𝜒( 𝑝) 2

whence
𝑞
𝜌˜ 2 |𝜁 | 2
Ö
|𝑄 𝜒 (𝜁)| ≥ 1+ . (8.2.5)
𝑝=1
𝑝 2 𝜒( 𝑝) 2

Now assume |𝜁 | ≥ 𝜒(1) (with 𝜁 ∈ ℭ𝜌 still). We select 𝑞 ∈ Z+ such that

0 < 𝑞 𝜒(𝑞) ≤ |𝜁 | < (𝑞 + 1) 𝜒(𝑞 + 1). (8.2.6)

We deduce from (8.2.6) that 𝑝 𝜒( 𝑝) ≤ |𝜁 | if 1 ≤ 𝑝 ≤ 𝑞; it then follows from (8.2.5)

that
˜ 𝑞 = exp {𝑞 log(1 + 𝜌)}
|𝑄 𝜒 (𝜁)| ≥ (1 + 𝜌) ˜ . (8.2.7)
We also deduce from (8.2.6) that |𝜁 | ↗ +∞ demands 𝜒(𝑞) ↗ +∞; in particular
there is an 𝑅 ≥ 1 such that |𝜁 | ≥ 𝑅 =⇒ 𝜒(𝑞) ≥ 1. Since 𝜒 is increasing and
𝑞 ≤ 𝑞 𝜒(𝑞) ≤ |𝜁 | we get

|𝜁 | |𝜁 |
𝑞> −1 ≥ −1.
𝜒(𝑞 + 1) 𝜒(|𝜁 | + 1)

Putting this into (8.2.7) implies, for |𝜁 | ≥ 𝑅,

|𝜁 |
|𝑄 𝜒 (𝜁)| ≥ exp − 1 log(1 + 𝜌)
˜ ,
𝜒(|𝜁 | + 1)

which is precisely (8.2.3). □

8.2.2 Local representation of hyperfunctions in term of smooth

functions

Let Ω be an open subset of R𝑛 ; we view C ∞ (Ω) as a vector subspace of the

distribution space D ′ (Ω) and D ′ (Ω) as a vector subspace of B (Ω), the vector space
of hyperfunctions in Ω. A hyperdifferential operator 𝐽 (𝜕𝑥 ) maps C ∞ (Ω) into B (Ω),
8.2 Local Representation of Hyperfunctions 239

generally not into D ′ (Ω). In the present section we are going to show that, locally,
every hyperfunction can be represented as the action of a hyperdifferential operator
with constant coefficients on a smooth function. This is the local representation
theorem for hyperfunctions, first proved in [Kaneko, 1972].
If 𝐾 is a compact subset of R𝑛 we define

𝐻𝐾 (𝜁) = max (𝑥 · 𝜉) , 𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 . (8.2.8)

𝑥 ∈𝐾

Lemma 8.2.2 Let 𝐾 be a compact subset of R𝑛 and L𝜇(𝜁) denote the Laplace–Borel
transform of 𝜇 ∈ O ′ (𝐾) (Definition 6.2.1). To every positive number 𝜃 < 1 there
exist a monotone increasing function 𝜓 : [0, +∞) → [1, +∞) satisfying 𝜓(0) = 1,
𝜓(𝑟) ↗ +∞ as 𝑟 ↗ +∞, sup 𝑟 −𝜃 𝜓(𝑟) < +∞, and a constant 𝐶 > 0, such that
𝑟 >0

|𝜁 |
∀𝜁 ∈ C𝑛 , |L𝜇(𝜁)| ≤ 𝐶 exp 𝐻𝐾 (𝜁) + . (8.2.9)
𝜓(|𝜁 |)

Proof If 𝐾 𝛿 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝐾) ≤ 𝛿} (𝛿 > 0) then there is a 𝐶 𝛿 > 0 such that

|L𝜇(𝜁)| ≤ 𝐶 𝛿 exp 𝐻𝐾 𝛿 (𝜁) for all 𝜁 ∈ C𝑛 . Since 𝐻𝐾 𝛿 (𝜁) ≤ 𝐻𝐾 (𝜁) + 𝛿|𝜁 | we have

∀𝜁 ∈ C𝑛 , e−𝐻𝐾 ( 𝜁 ) |L𝜇(𝜁)| ≤ 𝐶 𝛿 e 𝛿 |𝜁 | . (8.2.10)

If we define, for 𝑟 > 0,

𝑟
𝜓1 (𝑟) = sup e−𝐻𝐾 (𝜁 ) L𝜇 (𝜁) , 𝜓2 (𝑟) = ,
|𝜁 |=𝑟 log(e + 𝜓1 (𝑟))

we get
h i
|L𝜇(𝜁)| ≤ 𝜓1 (|𝜁 |)e 𝐻𝐾 (𝜁 ) = e |𝜁 |/𝜓2 ( |𝜁 |) − e e 𝐻𝐾 ( 𝜁 ) ≤ e 𝐻𝐾 (𝜁 )+ |𝜁 |/𝜓2 ( |𝜁 |) .

By (8.2.10) we know that 𝜓1 (𝑟) ≤ 𝐶 𝛿 e 𝛿𝑟 , implying 𝜓2 (𝑟) ≥ 𝑟/log(𝑒 + 𝐶 𝛿 e 𝛿𝑟 )

whatever 𝛿 > 0. Since lim 𝑟 −1 log(𝑒 + 𝐶 𝛿 e 𝛿𝑟 ) = 𝛿 we deduce that 𝜓2 (𝑟) → +∞ as
𝑟↗+∞
𝑟 → +∞. Let us then define 𝜓3 (𝑟) = min{𝜓2 (𝑟), 𝑟 𝜃 }; clearly 𝜓3 (0) = 0, 𝜓3 (𝑟) ≤ 𝑟 𝜃 ,
𝜓3 (𝑟) → +∞ as 𝑟 → +∞, and

|L𝜇(𝜁)| ≤ e 𝐻𝐾 (𝜁 )+ |𝜁 |/𝜓3 ( |𝜁 |) .

The function 𝜓(𝑟) = max{1, inf 𝑠 ≥𝑟 𝜓3 (𝑠)} has all the properties required if we take,
in (8.2.9),
𝐶 = sup e−𝐻𝐾 (𝜁 )−|𝜁 | L𝜇 (𝜁) . □
𝜓3 ( |𝜁 |) <1

We can now prove the announced local representation theorem.

Theorem 8.2.3 Let 𝑢 ∈ B (R𝑛 ) be arbitrary. If 𝑈 ⊂ R𝑛 is open and bounded there is
a hyperdifferential operator 𝑄 𝜒 (𝐷) with symbol (8.2.1) and a function 𝜑 ∈ C ∞ (R𝑛 )
satisfying 𝑢 = 𝑄 𝜒 (D 𝑥 )𝜑 in 𝑈.
240 8 Hyperdifferential Operators
√
Recall that 𝐷 = − −1𝜕.

Proof Let 𝜇 ∈ O ′ (𝑈) be arbitrary. Let 𝜓 be the function in Lemma 8.2.2 where
𝐾 = 𝑈 and 𝜃 ∈ (0, 1) and 𝑄 𝜒 (D𝑧 ) be the hyperdifferential operator with symbol
(8.2.1) and 𝜒(𝑟) = 𝜓(𝑟 − 1)/2 if 𝑟 ≥ 2. Thanks to Proposition 8.2.1 we get the
estimate
(2 log 2)|𝜉 |
𝐴◦ exp ≤ |𝑄 𝜒 (𝜉)|,
𝜓(|𝜉 |)
for some 𝐴◦ > 0 and all 𝜉 ∈ R𝑛 , |𝜉 | ≥ 𝑅, and thus

𝐶 |𝜉 |
𝑄 𝜒 (𝜉) −1 L𝜇(𝑖𝜉) ≤ exp −(2 log 2 − 1) .
𝐴◦ 𝜓(|𝜉 |)

Since 2 log 2 > 1 we can find constants 𝐶1 > 0, 𝑐 > 0 such that

𝑄 𝜒 (𝜉) −1 L𝜇(𝑖𝜉) ≤ 𝐶1 exp −𝑐|𝜉 | 1−𝜃

(8.2.11)

for all 𝜉 ∈ R𝑛 , |𝜉 | ≥ 𝑅. It follows from (8.2.11) that

∫
1 L𝜇(𝑖𝜉) 𝑖 𝑥· 𝜉
𝜑(𝑥) = e d𝜉 (8.2.12)
(2𝜋) 𝑛 R𝑛 𝑄 𝜒 (𝜉)

defines a smooth function in R𝑛 .

Let 𝑉 ⊂ R𝑛 be a bounded open subset of R𝑛 such that 𝑈 ⊂⊂ 𝑉. The theorem will
be proved if we show that

𝑄 𝜒 (D𝑧 )(𝜑𝜒𝑉 ) − 𝜇 ∈ O ′ (𝑉 \ 𝑈). (8.2.13)

Let us define
∫
1 L𝜇(𝑖𝜉) 𝑖 𝑥· 𝜉 − 1 | 𝜉 |2
𝜑 𝑘 (𝑥) = e 𝑘 d𝜉 (8.2.14)
(2𝜋) R𝑛 𝑄 𝜒 (𝜉)
𝑛
∫
1 1 𝑖 (𝑧−𝑤) · 𝜉 −𝜀 | 𝜉 | 2
= 𝜇𝑤 , e d𝜉
(2𝜋) 𝑛 R𝑛 𝑄 𝜒 (𝜉)

(𝑘 ∈ Z+ ); then 𝜑 𝑘 ∈ 𝐶 𝜔 (R𝑛 ) and

∫
1 𝑖 (𝑧−𝑤) · 𝜉 − 𝑘1 | 𝜉 | 2
𝑄 𝜒 (D𝑧 )𝜑 𝑘 (𝑧) = 𝜇 𝑤 , e d𝜉
(2𝜋) 𝑛 R𝑛
* √︂ 𝑛 ! +
1 𝑘 𝑘 2
= 𝜇𝑤 , e− 4 ⟨𝑧−𝑤 ⟩
2 𝜋

is an entire function in C𝑛 . We can therefore consider the analytic functionals carried

by 𝑉,
8.2 Local Representation of Hyperfunctions 241
√︂ ! 𝑛 ∫
1 𝑘 𝑘 2
𝑛
O (C ) ∋ ℎ ↦→ 𝜒𝑉 𝑄 𝜒 (D𝑧 )𝜑 𝑘 , ℎ = 𝜇𝑤 , e− 4 ⟨𝑥−𝑤 ⟩ ℎ (𝑥) d𝑥 .
2 𝜋 𝑉

On the one hand, since the entire functions

√︂ ! 𝑛 ∫
1 𝑘 𝑘 2
ℎ 𝑘 (𝑤) = e− 4 ⟨𝑥−𝑤 ⟩ ℎ (𝑥) d𝑥
2 𝜋 𝑉

converge uniformly to ℎ in a neighborhood of 𝑈 in C𝑛 as 𝑘 → +∞, it follows that

lim (𝑄 𝜒 (D𝑧 )𝜑 𝑘 ) 𝜒𝑉 = 𝜇 in O ′ (C𝑛 ). (8.2.15)

𝑘→+∞

On the other hand, 𝜑 𝑘 𝜒𝑉 −→ 𝜑𝜒𝑉 in O ′ (C𝑛 ) as 𝑘 → +∞ (by the Lebesgue

Dominated Convergence Theorem) implying

lim 𝑄 𝜒 (D𝑧 )(𝜑 𝑘 𝜒𝑉 ) = 𝑄 𝜒 (D𝑧 ) (𝜑𝜒𝑉 ) in O ′ (C𝑛 ). (8.2.16)

𝑘→+∞

We are going to show that the sequence of analytic functionals

𝜆 𝑘 = 𝑄 𝜒 (D𝑧 )(𝜑 𝑘 𝜒𝑉 ) − (𝑄 𝜒 (D𝑧 )𝜑 𝑘 ) 𝜒𝑉

is bounded in O ′ (𝑉\𝑈). By (8.2.15) and (8.2.16) that would mean that lim 𝑘→+∞ 𝜆 𝑘 =
𝑄 𝜒 (D𝑧 ) 𝜑 𝜒𝑉 − 𝜇 satisfies (8.2.13).
Thanks to (3.5.3) it suffices to prove that to each sufficiently small 𝛿 ∈ (0, 1) there
is a 𝐶 𝛿 > 0 such that
∀𝑘 ∈ Z+ , sup |𝜑 𝑘 | ≤ 𝐶 𝛿 . (8.2.17)
(𝜕𝑈) 𝛿

We have ∫
1 2 d𝜉
𝜑 𝑘 (𝑧) = 𝜇 𝑤 , e𝑖 (𝑧−𝑤) · 𝜉 −𝜀 | 𝜉 | .
(2𝜋) 𝑛 R𝑛 𝑄 𝜒 (𝜉)
Let 𝛿◦ > 0 be such that the distance between the compact sets (𝜕𝑉) 𝛿◦ ∩ R𝑛 and
𝑈 𝛿◦ ∩ R𝑛 is at least equal to some 𝑟 > 0. Since 𝜇 ∈ O ′ (𝑈), to each 𝛿 ∈ (0, 𝛿◦ ) there
is a 𝐶1 > 0 such that
∫
2 d𝜉
|𝜑 𝑘 (𝑧)| ≤ 𝐶1 sup e𝑖 (𝑧−𝑤) · 𝜉 −𝜀 | 𝜉 | .
𝑤 ∈𝑈 𝛿 R𝑛 𝑄 𝜒 (𝜉)

We restrict 𝑧 to the set (𝜕𝑉) 𝛿 and write 𝑧 = 𝑥 + 𝑖𝑦, 𝑤 = 𝑠 + 𝑖𝑡. In the integral we
perform the change of variables 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜆(𝑥 − 𝑠)|𝜉 |. Here, 𝜆 > 0 is chosen
sufficiently small that 𝜁 stays in the cone ℭ1/2 [cf. (8.2.2)] for all 𝑥 ∈ (𝜕𝑉) 𝛿◦ ∩ R𝑛
and 𝑠 ∈ 𝑈 𝛿◦ ∩ R𝑛 . Since |𝑄 𝜒 (𝜉)| ≥ 1 for all 𝜉 ∈ R𝑛 [see (8.2.1)] and since

Re(𝑖(𝑧 − 𝑤) · 𝜁) = − (𝑦 − 𝑡) |𝜉 | ≤ (−𝜆𝑟 2 + 2𝛿)|𝜉 |,

242 8 Hyperdifferential Operators

for all 𝑧 ∈ (𝜕𝑈) 𝛿 , 𝑤 ∈ 𝑈 𝛿 , (8.2.17) follows at once. The proof of Theorem 8.2.3 is
complete. □

8.3 Elliptic Hyperdifferential Operators

8.3.1 Elliptic hyperdifferential operators. Definition

Let Ω be an open subset of R𝑛 and let 𝐽 (𝑥, 𝜁) ∈ C 𝜔 Ω; Exp1,0 (C𝑛 ) ; ℭ𝜌 is the
cone (8.2.2).
Definition 8.3.1 The hyperdifferential operator 𝐽 (𝑥, 𝜕𝑥 ) in Ω (Definition 8.1.21)
will be called elliptic if there is an open subset ΩC of C𝑛 containing Ω, to which
the function 𝐽 (𝑥, 𝜁) has an extension 𝐽 (𝑧, 𝜁) ∈ O ΩC ; Exp1,0 (C𝑛 ) that has the
following property:
(ELL) given any compact set 𝐾 C ⊂ ΩC there are positive numbers 𝜌 < 1, 𝑅 and 𝑐
such that
∀ (𝑧, 𝜁) ∈ 𝐾 C × ℭ𝜌 , |𝜁 | > 𝑅 =⇒ |𝐽 (𝑧, 𝜁)| ≥ 𝑐. (8.3.1)

If (ELL) holds we shall also say that the function 𝐽 (𝑥, 𝜁) ∈ C 𝜔 Ω; Exp1,0 (C𝑛 )
is elliptic.
The terminology of Definition 8.3.1 is justified by the following
Proposition 8.3.2 Let Ω be an open and connected subset of R𝑛 . Given
∑︁
𝑃 (𝑥, 𝜉) = 𝑐 𝛼 (𝑥) 𝜉 𝛼 , 𝑐 𝛼 ∈ C 𝜔 (Ω) , (8.3.2)
| 𝛼 | ≤𝑚
Í
assume that the principal symbol 𝑃𝑚 (𝑥, 𝜉) = | 𝛼 |=𝑚 𝑐 𝛼 (𝑥) 𝜉 𝛼 does not vanish
identically in Ω × R𝑛 . Under this hypothesis the following properties are equivalent:
(1) The differential operator 𝑃 (𝑥, 𝜕𝑥 ) is elliptic (Definition 1.3.2).
(2) The hyperdifferential operator 𝑃 (𝑥, 𝜕𝑥 ) in Ω is elliptic (Definition 8.3.1).

Proof When 𝑚 = 0 both properties (1) and (2) mean that 𝑃 (𝑥, 𝜉) = 𝑎 (𝑥) ∈ C 𝜔 (Ω),
𝑎 (𝑥) ≠ 0 for every 𝑥 ∈ Ω. We shall therefore assume 𝑚 ≥ 1. Suppose (1) holds,
meaning that there is a continuous function Ω ∋ 𝑥 ↦→ 𝛾 (𝑥) > 0 such that

∀ (𝑥, 𝜉) ∈ Ω × R𝑛 , |𝑃𝑚 (𝑥, 𝜉)| ≥ 𝛾 (𝑥) |𝜉 | 𝑚 . (8.3.3)

Possibly after decreasing 𝛾 we may also suppose that the coefficients in (8.3.2) extend
as holomorphic functions 𝑐 𝛼 (𝑧) to the open subset

ΩC = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; 𝑥 ∈ Ω, |𝑦| < 2𝛾 (𝑥)} .

8.3 Elliptic Hyperdifferential Operators 243

Using (8.3.3) and Taylor expansion shows that there is a continuous function Ω ∋
𝑥 ↦→ 𝛿 (𝑥) ∈ (0, 𝛾 (𝑥)] such that
1
∀ (𝑥, 𝜉) ∈ Ω × R𝑛 , |𝑦| < 𝛿 (𝑥) =⇒ |𝑃𝑚 (𝑥 + 𝑖𝑦, 𝜉)| ≥ 𝛾 (𝑥) |𝜉 | 𝑚 .
2
Let us then define

ΩC = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; 𝑥 ∈ Ω, |𝑦| < 𝛿 (𝑥)} ;

let 𝐾 C be an arbitrary compact subset of ΩC and let us denote by 𝐾 its coordinate

projection in R𝑛 . We have, for some 𝐶𝐾 > 0 and all 𝑧 ∈ 𝐾 C ,
1
|𝑃 (𝑧, 𝜉)| ≥ min𝛾 |𝜉 | 𝑚 − 𝐶𝐾 (1 + |𝜉 |) 𝑚−1 .
2 𝐾

It follows that there is a 𝐶𝐾′ > 0 such that, whatever (𝑧, 𝜉 + 𝑖𝜂) ∈ 𝐾 C × C𝑛 ,
1
|𝑃 (𝑧, 𝜉 + 𝑖𝜂)| ≤ 1 implies |𝜉 | 𝑚 ≤ 𝐶𝐾′ (1 + |𝜂|) 𝑚 and therefore |𝜉 | ≤ 2𝐶𝐾′ 𝑚 |𝜂|
if |𝜉 | > 𝑅 provided 𝑅 > 0 is large enough. This proves (2).
Now suppose (1) is not true; this means that there is an (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0})
such that 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) = 0. We distinguish two cases. Case I: there is an N ∈ S𝑛−1
such that 𝑃𝑚 (𝑥 ◦ , N) ≠ 0. In this case we look at the equation
𝑚
∑︁
𝑃 (𝑥 ◦ , 𝜏𝜉 ◦ + 𝑧N) = 𝑞 𝑗 (𝜏) 𝑧 𝑚− 𝑗 = 0, (8.3.4)
𝑗=0

where 𝜏 ∈ R and the 𝑞 𝑗 are complex polynomials, deg 𝑞 𝑗 ≤ 𝑗; we have

𝑞 0 (𝜏) = 𝑃𝑚 (𝑥 ◦ , N) ≠ 0,

𝑞 𝑚 (𝜏) = 𝑃 (𝑥 ◦ , 𝜏𝜉 ◦ ) = 𝑂 𝜏 𝑚−1 .

We derive immediately that, given any 𝜏 ∈ R+ , there is a root 𝑧 (𝜏) of Eq. (8.3.4) such
𝑚−1
that |𝑧 (𝜏)| ≤ 𝐶 (1 + 𝜏) 𝑚 which implies that (ELL) does not hold when 𝐽 = 𝑃.
Case II: 𝑃𝑚 (𝑥 ◦ , 𝜉) = 0 for all 𝜉 ∈ R𝑛 . Let 𝑟 > 0 be sufficiently small that the
closed ball in C𝑛 , with center 𝑥 ◦ and radius 𝑟, is contained in ΩC . The hypothesis
that 𝑃𝑚 . 0 in Ω × R𝑛 ensures that there are unit vectors v, N ∈ R𝑛 such that
𝑧 ∈ C, 0 < |𝑧| < 𝑟, implies 𝑃𝑚 (𝑥 ◦ + 𝑧v, N) ≠ 0. Consider
𝑚
∑︁
𝐹 (𝑤, 𝑧) = 𝑤 𝑚 𝑃 𝑥 ◦ + 𝑧v, 𝑤 −1 N = 𝑎 𝑗 (𝑧) 𝑤 𝑗 .
𝑗=0

Observe that 𝐹 (0, 𝑧) = 𝑃𝑚 (𝑥 ◦ + 𝑧v, N) = 𝑧 𝑘 𝜑 (𝑧) for some 𝑘 > 0 and some
𝜑 ∈ O ({𝑧 ∈ C; |𝑧| < 𝜀}) (𝜀 < 𝑟), 𝜑 (0) ≠ 0. By the Weierstrass Preparation
Theorem 14.3.4 we have
244 8 Hyperdifferential Operators

∑︁ 𝑘
𝐹 (𝑤, 𝑧) = 𝐸 (𝑤, 𝑧) ­𝑧 𝑘 + 𝑏 𝑗 (𝑤) 𝑧 𝑘− 𝑗 ® ,
© ª

« 𝑗=1 ¬
where 𝐸 (𝑤, 𝑧) ∈ O (𝑧, 𝑤) ∈ C2 ; |𝑧| + |𝑤| < 𝜀 , 𝐸 (0, 0) ≠ 0 and where 𝑏 𝑗 ∈

O ({𝑧 ∈ C; |𝑧| < 𝜀}) with 𝑏 𝑗 (0) = 0 for every 𝑗 = 1, ..., 𝑘. It follows that 𝐹 (𝑤, 𝑧) =
0 is equivalent to the equation
𝑘
∑︁
𝑧𝑘 + 𝑏 𝑗 (𝑤) 𝑧 𝑘− 𝑗 = 0
𝑗=1

and entails |𝑧| ≤ 𝐶 |𝑤| 𝜃 for some positive constants 𝐶, 𝜃. Taking 𝑤 = 𝜏 −1 ∈ R+ we

see that there are real numbers 𝜏 → +∞ and points 𝑥 ◦ + 𝑧 (𝜏) v ∈ ΩC converging to
𝑥 ◦ such that 𝑃 (𝑥 ◦ + 𝑧 (𝜏) v, 𝜏N) = 0, which implies that (ELL) does not hold when
𝐽 = 𝑃. □
Example 8.3.3 The hyperdifferential operator 𝑄 𝜒 (𝜕𝑥 ) with symbol (8.2.1) is elliptic.

8.3.2 Analytic hypoellipticity in hyperfunctions of elliptic

hyperdifferential operators with constant coefficients

Theorem 8.3.4 If 𝐽 (𝜁) ∈ Exp1,0 (C𝑛 ) is elliptic then, for any 𝑢 ∈ B (R𝑛 ) and
Ω ⊂ R𝑛 open, 𝐽 (D 𝑥 )𝑢 ∈ C 𝜔 (Ω) implies 𝑢 ∈ C 𝜔 (Ω).
√
Keep in mind that 𝐷 = − −1𝜕.
Proof It suffices to prove the claim when 𝑢 ∈ C ∞ (R𝑛 ). Indeed, assume the latter has
been proved and now let 𝑢 ∈ B (R𝑛 ), 𝐽 (D 𝑥 )𝑢 ∈ C 𝜔 (Ω). Let 𝑈 ⊂⊂ Ω be open and
apply Theorem 8.2.3: there are 𝑄 𝜒 (D 𝑥 ) and 𝑣 ∈ C ∞ (R𝑛 ) such that 𝑢 = 𝑄 𝜒 (D 𝑥 )𝑣 in
𝑈; then 𝐽 (D 𝑥 )𝑄 𝜒 (D 𝑥 )𝑣 ∈ C 𝜔 (𝑈). The composite 𝐽 (D 𝑥 )𝑄 𝜒 (D 𝑥 ) = 𝑄 𝜒 (D 𝑥 )𝐽 (D 𝑥 )
is a hyperdifferential operator with symbol 𝐽 (𝜁)𝑄 𝜒 (𝜁); it is also elliptic. Hence
𝑣 ∈ C 𝜔 (𝑈) and therefore, by Proposition 8.1.20, 𝑢 ∈ C 𝜔 (𝑈).
Thus, let 𝑢 ∈ C ∞ (R𝑛 ) be such that 𝐽 (D 𝑥 )𝑢 = 𝑓 ∈ C 𝜔 (Ω). We select arbitrarily
an open set 𝑉 ⊂⊂ Ω; the analytic functional
∫
𝑛
O (C ) ∋ ℎ ↦→ ⟨𝜇, ℎ⟩ = (𝑢 (𝑠) 𝐽 (−D𝑠 )ℎ (𝑠) − 𝑓 (𝑠) ℎ (𝑠)) d𝑠 (8.3.5)
𝑉

is carried by the boundary 𝜕𝑉. Assuming that (8.3.1) holds we introduce the entire
functions ∫
−𝑛 2
𝐺 𝜀 (𝑧) = (2𝜋) 𝐽 (𝜉) −1 e𝑖𝑧· 𝜉 −𝜀 | 𝜉 | d𝜉 (8.3.6)
| 𝜉 |>𝑅

(𝜀 > 0). We are going to look at the convolution

(𝜇 ∗ 𝐺 𝜀 ) (𝑧) = ⟨𝜇 𝑤 , 𝐺 𝜀 (𝑧 − 𝑤)⟩ .
8.3 Elliptic Hyperdifferential Operators 245

First of all,
∫
2
𝐽 (−D𝑠 )𝐺 𝜀 (𝑧 − 𝑠) = (2𝜋) −𝑛 e𝑖 (𝑧−𝑠) · 𝜉 −𝜀 | 𝜉 | d𝜉
| 𝜉 |>𝑅
∫ ∫
2 2
= (2𝜋) −𝑛 e𝑖 (𝑧−𝑠) · 𝜉 −𝜀 | 𝜉 | d𝜉 − (2𝜋) −𝑛 e𝑖 (𝑧−𝑠) · 𝜉 −𝜀 | 𝜉 | d𝜉
R𝑛 | 𝜉 |<𝑅
∫
1 1 2 2
= (4𝜋𝜀) − 2 𝑛 e− 4𝜀 ⟨𝑧−𝑠⟩ − (2𝜋) −𝑛 e𝑖 (𝑧−𝑠) · 𝜉 −𝜀 | 𝜉 | d𝜉,
| 𝜉 |<𝑅

whence
∫ ∫
1 1 2
𝑢 (𝑠) 𝐽 (−D𝑠 )𝐺 𝜀 (𝑧 − 𝑠)d𝑠 = (4𝜋𝜀) − 2 𝑛 e− 4𝜀 ⟨𝑧−𝑠⟩ 𝑢 (𝑠) d𝑠
𝑉 𝑉
∫ ∫
−𝑛 2
− (2𝜋) e𝑖 (𝑧−𝑠) · 𝜉 −𝜀 | 𝜉 | 𝑢 (𝑠) d𝑠d𝜉.
𝑠 ∈𝑉 | 𝜉 |<𝑅

We have ∫
1 1 2
lim (4𝜋𝜀) − 2 𝑛 e− 4𝜀 ⟨𝑥−𝑠⟩ 𝑢 (𝑠) d𝑠 = 𝑢 (𝑥) if 𝑥 ∈ 𝑉,
𝜀↘0 𝑉
while
∫ ∫
2
𝜓 (𝑧) = lim e𝑖 (𝑧−𝑠) · 𝜉 −𝜀 | 𝜉 | 𝑢 (𝑠) d𝑠d𝜉
𝜀↘0 𝑠 ∈𝑉 | 𝜉 |<𝑅
∫
𝑖𝑧· 𝜉
= ( 𝜒d
𝑉 𝑢) (𝜉) e d𝜉
| 𝜉 |<𝑅

is an entire function of exponential type.

By (8.3.1) we know that
∫
( 𝜀)
𝐹𝑉 (𝑥) = 𝑓 (𝑠) 𝐺 𝜀 (𝑥 − 𝑠)d𝑠
𝑉
∫ ∫
2
= (2𝜋) −𝑛 𝐽 (𝜉) −1 𝑓 (𝑠)e𝑖 ( 𝑥−𝑠) · 𝜉 −𝜀 | 𝜉 | d𝜉d𝑠
𝑠 ∈𝑉 | 𝜉 |>𝑅

converges, uniformly on compact subsets of 𝑉, to a function 𝐹𝑉 ∈ C 𝜔 (𝑉) as 𝜀 ↘ 0.

It follows that 𝜇 ∗ 𝐺 𝜀 converges (in the same sense) to the function 𝑢 − 𝜓 − 𝐹𝑉 . Since
𝜇 is carried by 𝜕𝑉, to every 𝛿 > 0 there is a 𝐶 𝛿 > 0 such that

|(𝜇 ∗ 𝐺 𝜀 ) (𝑧)| ≤ 𝐶 𝛿 sup |𝐺 𝜀 (𝑧 − 𝑤)| . (8.3.7)

𝑤 ∈(𝜕𝑉) 𝛿

We carry out the deformation of the domain of 𝜉-integration in (8.3.6) from the
region |𝜉 | > 𝑅 to its image under the map 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜅(𝑧 − 𝑤) (|𝜉 | − 𝑅) , now
requiring that 𝜅 > 0 be small enough to ensure that 𝜁 ∈ ℭ𝜌 if 𝑤 ∈ (𝜕𝑉) 𝛿 and
|𝑧 − 𝑥 ◦ | ≤ 𝑟 ◦ :
246 8 Hyperdifferential Operators

𝐺 𝜀 (𝑧 − 𝑤)
∫
2
( | 𝜉 |−𝑅)−𝜀 | 𝜉 | 2
= (2𝜋) −𝑛 𝐽 (𝜁) −1 𝑓 (𝑤)e𝑖 (𝑧−𝑤) · 𝜉 −𝜅 ⟨𝑧−𝑤 ⟩ 𝐷 (𝑧 − 𝑤, 𝜉) d𝜉.
| 𝜉 |>𝑅

We conclude that there is a 𝐶 > 0 such that

∫
2
|𝐺 𝜀 (𝑧 − 𝑤)| ≤ 𝐶 e−𝜅 ( | 𝜉 |−𝑅) Re ⟨𝑧−𝑤 ⟩ d𝜉
| 𝜉 |>𝑅

if 𝑤 ∈ (𝜕𝑉) 𝛿 and |𝑧 − 𝑥 ◦ | ≤ 𝑟 ◦ . We can select the positive numbers 𝛿 and 𝑟 ◦

so small that Re ⟨𝑧 − 𝑤⟩ 2 > 𝛿, as a consequence of which sup |𝐺 𝜀 (𝑧 − 𝑤)|
𝑤 ∈ (𝜕𝑉) 𝛿
remains bounded in the ball {𝑧 ∈ C; |𝑧 − 𝑥 ◦ | < 𝑟 ◦ }; and therefore, by (8.3.7), so do
the holomorphic functions (𝜇 ∗ 𝐺 𝜀 ) (𝑧). Thanks to Montel’s Theorem there is a
sequence of functions 𝜇 ∗ 𝐺 𝜀𝑘 (𝜀 𝑘 ↘ 0, 𝑘 = 1, 2, ...) converging normally, in that
same ball, to a holomorphic function 𝐺. If 𝑥 ∈ 𝑉, |𝑥 − 𝑥 ◦ | ≤ 𝑟 ◦ , necessarily

𝐺 (𝑥) = 𝑢 (𝑥) − 𝐹𝑉 (𝑥) − 𝜓 (𝑥) ,

proving that 𝑢 is analytic in an open subset of 𝑉 that contains 𝑥 ◦ . Since 𝑥 ◦ can be

any point of 𝑉 we reach the conclusion that 𝑢 ∈ C 𝜔 (𝑉). □
The following variant of (Petrowski’s) Theorem 4.1.1 attests to the difference
between hyperfunction and distribution solutions of a linear PDE with constant
coefficients ([Schapira, 1970], p. 103).

Theorem 8.3.5 The following properties of a linear PDO, 𝑃(D 𝑥 ), are equivalent:
(a) 𝑃(D 𝑥 ) is elliptic;
(b) ∀𝑢 ∈ B (R𝑛 ), 𝑃(D 𝑥 )𝑢 = 0 =⇒ 𝑢 ∈D ′ (R𝑛 ).

Proof In view of Theorem 8.3.4 it suffices to prove that (b)=⇒(a). We return to the
proof of the entailment (c)=⇒(a) in the proof of Theorem 4.1.1. Under the hypothesis
that 𝑃(D 𝑥 ) is not elliptic we have shown, after a suitable linear change of coordinates,
that the symbol of the operator 𝑃(D 𝑥1 , D 𝑥2 , 0, . . . , 0) can be expressed as
𝑚
∑︁
𝑃(𝜉1 , 𝜉2 , 0, . . . , 0) = 𝑐𝜉1𝑚 + 𝑞 𝑘 (𝜉2 )𝜉1𝑚−𝑘 , (8.3.8)
𝑘=1

where 0 ≠ 𝑐 ∈ C, the degree of 𝑞 𝑘 is ≤ 𝑘 and, furthermore, the degree of 𝑞 𝑚

is ≤ 𝑚 − 1. Such conditions imply the existence of a holomorphic function Φ(𝑠),
defined in some truncated sector in the complex plane |arg 𝑠| < 𝛿, |𝑠| > 𝜌, satisfying
Φ(𝑠) = 𝑂 (|𝑠| 𝑟 ) (𝑟 ∈ Q, 0 < 𝑟 < 1) and such that 𝑃(Φ(𝑠), 𝑠, 0, . . . , 0) = 0. It follows
that if 𝐽 (𝜁) ∈ Exp1,0 (C) then
∫ ∞
ℎ(𝑧) = 𝐽 (𝜆) e𝑖𝑧2 𝜆+𝑖𝑧1 Φ(𝜆) d𝜆
𝜌
8.4 Solvability of Constant Coefficients Hyperdifferential Equations 247

defines a holomorphic function in the half-space {𝑧 ∈ C𝑛 ; Im 𝑧 2 > 0} satisfy-

ing 𝑃(D𝑧 )ℎ = 0. If 𝑢 ∈ B (R𝑛 ) is the hyperfunction boundary value of ℎ then
𝑃(D 𝑥 )𝑢 = 0. If 𝑢 were a distribution in R𝑛 it would have a well-defined distribution
trace 𝑢 ◦ on the hyperplane 𝑥1 = 0 [since, according to (8.3.8), this hyperplane is
noncharacteristic with respect to 𝑃(D)]. Such a trace would be the boundary value
of the holomorphic function of (𝑧2 , ..., 𝑧 𝑛 ),
∫ ∞ ∫ 𝜌
1
𝐽 (𝜆) e𝑖𝑧2 𝜆 d𝜆 = 𝐽 (D𝑧2 ) − 𝐽 (𝜆)e𝑖𝑧2 𝜆 d𝜆 ( Im 𝑧2 > 0).
𝜌 𝑖𝑧 2 0

We leave as an exercise to find 𝐽 (𝜁) ∈ Exp1,0 (C) such that 𝐽 (D 𝑥2 ) 𝑥21+𝑖0 ∉ D ′ (R),
contradicting (b). □
Corollary 8.3.6 If 𝑃(D) is hypoelliptic but not elliptic there are nonsmooth hyper-
function solutions of the equation 𝑃(D)𝑢 = 0.
Needless to say, the nonsmooth solutions in Corollary 8.3.6 are not distributions.

8.4 Solvability of Constant Coefficients Hyperdifferential

Equations

8.4.1 Two minimum principles for functions of infra-exponential type

The minimum principles for entire functions of infra-exponential type proved in this
subsection will be used in the next one to establish the solvability in holomorphic
functions of hyperdifferential equations with constant coefficients. We shall make
use of classical theorems about polynomials and entire functions of a single complex
variable, stated without proofs but with references. The first result is due to G. Valiron
(see [Valiron, 1949], p. 80).
Lemma 8.4.1 There is a number 𝛾 ∈ (0, 1) such that the following property holds,
for every complex polynomial 𝑝 in a single variable such that 𝑝(0) = 1:
(V) If |𝑧| > 𝑅 > 0 implies 𝑝(𝑧) ≠ 0 then there is an 𝑟 ∈ [ 81 𝑅, 14 𝑅] such that

min | 𝑝(𝑧)| ≥ 𝛾 deg 𝑝 . (8.4.1)

|𝑧 |=𝑟

The next result is due to C. Carathéodory; for a proof see [Boas, 1954], p. 3.
Throughout, log stands for the branch of the logarithm function that is real in
(0, +∞).
Lemma 8.4.2 Suppose that 𝑔 ∈ O (C) has no zeros in the disk |𝑧| < 𝑅. If 𝑔(0) = 1
and 0 < 𝑟 < 𝑅 then
2𝑟
max | log 𝑔(𝑧)| ≤ log max |𝑔(𝑧)| . (8.4.2)
|𝑧 |=𝑟 𝑅−𝑟 |𝑧 |=𝑟
248 8 Hyperdifferential Operators

The next result is the Jensen identity (loc. cit., p. 2).

Lemma 8.4.3 Let 𝑓 be a continuous function in the closed disk |𝑧| ≤ 𝑅, holomorphic
in the interior |𝑧| < 𝑅 and such that 𝑓 (0) = 1. For 0 < 𝑟 ∫< 𝑅 denote by 𝑛 𝑓 (𝑟) the
𝑟
number of zeros of 𝑓 in the disk |𝑧| ≤ 𝑟 and set 𝑁 𝑓 (𝑟) = 0 𝑠−1 𝑛 𝑓 (𝑠)d𝑠. Then, for
0 < 𝑟 < 𝑅, ∫ 2𝜋
1
𝑁 𝑓 (𝑟) = log | 𝑓 (𝑟e𝑖 𝜃 )|d𝜃 . (8.4.3)
2𝜋 0
In the statement of Lemma 8.4.3 zeros are counted with their multiplicity: 𝑧 𝑚 has
𝑚 zeros; this rule applies through the remainder of this section.
The preceding three lemmas will now be applied to prove the first minimum
principle needed in the proof of the surjectivity theorems. This result is a special
case of Theorem 2, [Ehrenpreis, 1955]. We shall use the notation

𝑀 𝑓 (𝑟) = max | 𝑓 (𝑧)|, 𝑓 ∈ O (C) . (8.4.4)

|𝑧 |=𝑟

Theorem 8.4.4 There is an integer 𝜅 > 0 such that the following is true:
• To every entire function 𝑓 in C such that 𝑓 (0) = 1 and every 𝑅 > 0 there is an
𝑟 ∈ [ 81 𝑅, 41 𝑅] such that

min log | 𝑓 (𝑧)| ≥ −𝜅 log 𝑀 𝑓 (2𝑅) . (8.4.5)

|𝑧 |=𝑟

Proof Let 𝑅 > 0 be arbitrary and 𝑓 ∈ O (C) be such that 𝑓 (0) = 1. Let 𝛼1 , ..., 𝛼𝑚
be the zeros (with multiplicities) of 𝑓 in the disk |𝑧| ≤ 𝑅 [thus 𝑚 = 𝑛 𝑓 (𝑅)]; we
define
𝑚
Ö 𝑧
𝑝(𝑧) = 1− .
𝑗=1
𝛼𝑗

Then 𝑓 (𝑧) = 𝑝(𝑧)𝑔(𝑧), where 𝑔 ∈ O (C), 𝑔(0) = 1 and |𝑧| ≤ 𝑅 =⇒ 𝑔 (𝑧) ≠ 0. By

Lemma 8.4.1 there is an 𝑟 ∈ [ 81 𝑅, 41 𝑅] such that

min log | 𝑝(𝑧)| ≥ −𝑚 |log 𝛾| ;

|𝑧 |=𝑟

we derive

log | 𝑓 (𝑧)| = log | 𝑝(𝑧)| + log |𝑔(𝑧)| ≥ −𝑚 |log 𝛾| − |log 𝑔(𝑧)| . (8.4.6)

We apply Lemma 8.4.2 with 2𝑅 in place of 𝑅; we get

2𝑟
max | log 𝑔(𝑧)| ≤ log 𝑀𝑔 (𝑟) . (8.4.7)
|𝑧 |=𝑟 2𝑅 − 𝑟

By (8.4.6) and (8.4.7) we get

8.4 Solvability of Constant Coefficients Hyperdifferential Equations 249

min log | 𝑓 (𝑧)| ≥ −𝑚 |log 𝛾| − max |log 𝑔(𝑧)| (8.4.8)

|𝑧 |=𝑟 |𝑧 |=𝑟
2𝑟
≥ −𝑚 |log 𝛾| − log 𝑀𝑔 (𝑟).
2𝑅 − 𝑟
We have
𝑀 𝑓 (𝑟)
𝑀𝑔 (𝑟) ≤ ≤ 𝛾 −𝑚 𝑀 𝑓 (𝑟),
min |𝑧 |=𝑟 | 𝑝(𝑧)|
that is,
log 𝑀𝑔 (𝑟) ≤ log 𝑀 𝑓 (𝑟) + 𝑚 |log 𝛾| .
Putting this into (8.4.8) yields

2𝑟 2𝑟
min log | 𝑓 (𝑧)| ≥ − 1 + 𝑚 |log 𝛾| − log 𝑀 𝑓 (𝑟).
|𝑧 |=𝑟 2𝑅 − 𝑟 2𝑅 − 𝑟

Since 18 𝑅 ≤ 𝑟 ≤ 14 𝑅 we have 2𝑟
2𝑅−𝑟 < 13 , whence

4 1
min log | 𝑓 (𝑧)| ≥ − 𝑚 |log 𝛾| − log 𝑀 𝑓 (𝑟) . (8.4.9)
|𝑧 |=𝑟 3 3

We can estimate 𝑚 = 𝑛 𝑓 (𝑅):

∫ 2𝑅
𝑛 𝑓 (𝑅)
𝑚 log 2 = d𝑠
𝑅 𝑠
∫ 2𝑅
𝑛 𝑓 (𝑠)
≤ d𝑠 ≤ 𝑁 𝑓 (2𝑅).
𝑅 𝑠
Then Lemma 8.4.3 implies
log 𝑀 𝑓 (2𝑅)
𝑚≤ . (8.4.10)
log 2
Putting this into (8.4.9) yields

1 |log 𝛾|
min log | 𝑓 (𝑧)| ≥ − 1 + 4 𝑀 𝑓 (2𝑅),
|𝑧 |=𝑟 3 log 2

|log 𝛾 |
entailing (8.4.5) with 𝜅 ≥ 31 1 + 4 log 2 . □

The following corollary is the first minimum principle for entire functions of
infra-exponential type announced earlier.
Corollary 8.4.5 Let 𝐽 ∈ Exp1,0 (C) be arbitrary. To every 𝑅 > 0 there is an 𝑟 ∈
[ 81 𝑅, 41 𝑅] such that

min |𝐽 (𝑧)| ≥ |𝐽 (0)| 𝜅+1 |𝐽 | −𝜅

𝜀 e
−2𝜅 𝜀𝑅
, (8.4.11)
|𝑧 |=𝑟

where 𝜅 is the positive integer in (8.4.5) and

250 8 Hyperdifferential Operators

|𝐽 | 𝜀 = max sup |𝐽 (𝑧)|e−𝜀 |𝑧 | , 1 . (8.4.12)
𝑧 ∈C

Proof We suppose 𝐽 (0) ≠ 0 otherwise the statement is trivial. We have max |𝐽 (𝑧)| ≤
|𝑧 | ≤𝑟
|𝐽 | 𝜀e 𝜀𝑟 |
[cf. (8.1.1)] and therefore log 𝑀 𝐽 (2𝑅) ≤ 2𝜀𝑅 + log |𝐽 | 𝜀 . Taking 𝑓 (𝑧) =
𝐽 (𝑧) /𝐽 (0) in (8.4.4) we get

min log | 𝑓 (𝑧)| ≥ −𝜅 log 𝑀 𝑓 (2𝑅)

|𝑧 |=𝑟

≥ −𝜅 (2𝜀𝑅 + log | 𝑓 | 𝜀 ) .

Exponentiating yields (8.4.11). □

Our second minimum principle is based on a result of P. Boutroux and H. Cartan
that can be viewed as a complement to Valiron’s Lemma 8.4.1. For a proof we refer
to [Boas, 1954], p. 46.

Lemma 8.4.6 Let 𝑃 ∈ C [𝑧] be a monic polynomial (meaning, its highest power has
coefficient 1) of degree 𝑚 ≥ 1. To every 𝜏 > 0 there are 𝑚 disks
n o
Δ𝑟 𝑗 𝑧 ( 𝑗) = 𝑧 ∈ C; 𝑧 − 𝑧 ( 𝑗) < 𝑟 𝑗 (8.4.13)

with 𝑟 1 + · · · + 𝑟 𝑚 ≤ 2𝜏, such that |𝑃(𝑧)| ≥ (𝜏/𝑒) 𝑚 in the set

𝑚
Ø
𝑺 (𝑃, 𝜏) = C\ Δ𝑟 𝑗 𝑧 ( 𝑗) . (8.4.14)
𝑗=1

The disks (8.4.13) need not be distinct: think of the special case 𝑃 (𝑧) = 𝑧 𝑚 .
We introduce the following function of 𝜏 ∈ (0, e):

1 − log 𝜏
𝑏𝜏 = 2 + 3 . (8.4.15)
log 2
The next statement can be found in [Levin, 1964] (Theorem 11).

Theorem 8.4.7 Let 𝑓 ∈ O (C), 𝑓 (0) = 1, and the numbers 𝜏 ∈ (0, e) and 𝑅 > 0 be
arbitrary. Let 𝑚 be the number of zeros of 𝑓 in the disk |𝑧| ≤ 𝑅 and let 𝑃 be the
monic polynomial of degree 𝑚 having those same zeros. There are, then, 𝑚 disks
(8.4.13) with 𝑟 1 + · · · + 𝑟 𝑚 ≤ 2𝜏𝑅, such that

log | 𝑓 (𝑧)| ≥ −𝑏 𝜏 log 𝑀 𝑓 (2𝑅) (8.4.16)

for all 𝑧 ∈ 𝑺 (𝑃, 𝑅𝜏) [defined in (8.4.14)], |𝑧| ≤ 21 𝑅.

Proof We have the factorization 𝑓 (𝑧) = 𝑝(𝑧)𝑔(𝑧) with 𝑚 = deg 𝑝 and no zero of 𝑔
lying in the disk |𝑧| ≤ 𝑅; we may, and shall, assume that 𝑝 (0) = 𝑔 (0) = 1. Consider
the monic polynomial 𝑃(𝑧) = 𝐴𝑝(𝑧), where 𝐴 (≠ 0) is the product of all the roots 𝛼 𝑗
8.4 Solvability of Constant Coefficients Hyperdifferential Equations 251

( 𝑗 = 1, ..., 𝑚) of 𝑝; since sup 𝑗=1,...,𝑚 |𝛼 𝑗 | ≤ 𝑅 we have | 𝑝(𝑧)| ≥ 𝑅 −𝑚 |𝑃(𝑧)|. We apply

Lemma 8.4.6 with 𝑅𝜏 substituted for 𝜏: if 𝑧 ∈ 𝑺 (𝑃, 𝑅𝜏) we have | 𝑝(𝑧)| ≥ (𝜏/e) 𝑚
and therefore
log | 𝑓 (𝑧)| ≥ − log |𝑔(𝑧)| − 𝑚 log(e/𝜏) . (8.4.17)
If 𝑟 ≤ 21 𝑅 we can apply (8.4.7) with 𝑅 in place of 2𝑅. Combining the result with
(8.4.17) yields, for every 𝑧 ∈ 𝑺 (𝑃, 𝑅𝜏) such that |𝑧| = 𝑟,
2𝑟
log | 𝑓 (𝑧)| ≥ − log max |𝑔(𝑧)| − 𝑚 log(e/𝜏)
𝑅−𝑟 |𝑧 |=𝑟
2𝑟 𝑅+𝑟
≥− log max | 𝑓 (𝑧)| − 𝑚 log(e/𝜏)
𝑅−𝑟 |𝑧 |=𝑟 𝑅−𝑟

whence, for those same 𝑧’s,

1
log | 𝑓 (𝑧)| ≥ −3𝑚 log(e/𝜏) − 2 log 𝑀 𝑓 ( 𝑅).
2
We combine this with (8.4.10) to get

log(e/𝜏) 1
log | 𝑓 (𝑧)| ≥ −3 log 𝑀 𝑓 (2𝑅) − 2 log 𝑀 𝑓 ( 𝑅),
log 2 2
which implies directly (8.4.16). □
The following corollary is the sought second minimum principle.
Corollary 8.4.8 Let 𝑓 ∈ O (C) such that 𝑓 (0) ≠ 0 and the numbers 𝜏 ∈ (0, e),
𝑅 > 0, be arbitrary. Let 𝑚 be the number of zeros of 𝑓 in the disk |𝑧| ≤ 𝑅 and let 𝑃
be the monic polynomial of degree 𝑚 having those same zeros. There are, then, 𝑚
disks (8.4.13) with 𝑟 1 + · · · + 𝑟 𝑚 ≤ 2𝜏𝑅, such that

| 𝑓 (𝑧)| ≥ | 𝑓 (0)| 1+𝑏 𝜏 | 𝑓 | −𝑏

𝜀 e
𝜏 −2𝑏 𝜏 𝜀𝑅
(8.4.18)

for every 𝑧 ∈ 𝑺 (𝑃, 𝑅𝜏), |𝑧| ≤ 𝑅/2.

Proof It suffices to apply (8.4.16) combined with the fact that log 𝑀 𝑓 (𝑅) ≤
log | 𝑓 | 𝜀 + 𝜀𝑅 when 𝑓 ∈ Exp1,0 (C𝑛 ). □

8.4.2 Solvability in holomorphic functions of hyperdifferential

equations with constant coefficients

This subsection presents the precise statement and the proof of the result announced
at the beginning of the preceding subsection. We need a couple of additional lemmas
based on the minimum principles proved above (Corollaries 8.4.5, 8.4.8). We extend
the functional (8.4.12) to functions in higher dimensions: if 𝐽 ∈ Exp1,0 (C𝑛 ), 𝑛 ≥ 1,
we define
252 8 Hyperdifferential Operators

|𝐽 | 𝜀 = max sup |𝐽 (𝑧)|e−𝜀 |𝑧 | , 1 . (8.4.19)
𝑧 ∈C𝑛

Lemma 8.4.9 Let 𝐽 ∈ Exp1,0 (C𝑛 ) be such that 𝐽 (0) = 1. Let the numbers 𝜀 > 0
and 𝜏 ∈ (0, 2−5 ] be arbitrary. Then we have, for every Φ ∈ O (C𝑛 ),
𝜏 𝜀𝑐 𝜏 |𝜁 |
∀𝜁 ∈ C𝑛 , |Φ(𝜁)| ≤ |𝐽 | 𝜅+𝑏
𝜀 e sup |Φ(𝑧)𝐽 (𝑧)|, (8.4.20)
|𝑧−𝜁 | ≤26 𝜏 |𝜁 |

where 𝜅 is the positive integer in (8.4.5), 𝑏 𝜏 is the number (8.4.15) and 𝑐 𝜏 =

23 𝑏 𝜏 + 27 𝜅𝜏.

Proof Fix 𝜁 ≠ 0 and consider 𝑓 (𝑤) = 𝐽 (𝑤𝜁/|𝜁 |), 𝑤 ∈ C. Then 𝑓 ∈ Exp1,0 (C) and
| 𝑓 | 𝜀 ≤ |𝐽 | 𝜀 ; also 𝑓 (0) = 1. Fix 0 < 𝜏 ≤ 2−5 and let 𝑚 ∈ Z+ be arbitrary. The disk
{𝑤 ∈ C : |𝑤 − |𝜁 || ≤ 24 𝜏|𝜁 |}, whose area is equal to 28 𝜋𝜏 2 |𝜁 | 2 , cannot be covered
by the union of 𝑚 disks (8.4.13) with 𝑟 1 + · · · + 𝑟 𝑚 ≤ 23 𝜏|𝜁 |, whose total area does
not exceed 𝜋 𝑟 1 + · · · + 𝑟 𝑚 ≤ 𝜋 (𝑟 1 + · · · + 𝑟 𝑚 ) 2 ≤ 26 𝜋𝜏 2 |𝜁 | 2 . Thus we are allowed
2 2

Ø 𝑚
to apply Corollary 8.4.8 with 𝑅 = |𝜁 |: there is a 𝑤 ◦ ∈ C\ Δ𝑟 𝑗 𝑧 ( 𝑗) such that
𝑗=1
1
|𝑤 ◦ − |𝜁 || ≤ 24 𝜏|𝜁 | ≤ 2 |𝜁 | and
3𝑏
| 𝑓 (𝑤 ◦ )| ≥ |𝐽 | −𝑏
𝜀 e
𝜏 −2 𝜏 𝜀 |𝜁 |
. (8.4.21)

Next we apply Corollary 8.4.5 with 𝑅 = 27 𝜏|𝜁 |. We derive that there is an 𝑟 ∈

24 𝜏|𝜁 |, 25 𝜏 |𝜁 | such that
8 𝜅 𝜏 𝜀 |𝜁
min | 𝑓 (𝑤)| ≥ |𝐽 | −𝜅
𝜀 | 𝑓 (𝑤 ◦ )|
𝜅+1 −2
e |
,
|𝑤−𝑤◦ |=𝑟

and therefore, by (8.4.21),

|𝑤 − |𝜁 || ≥ min | 𝑓 (𝑤)| ≥ |𝐽 | −(
𝜀
𝜅+𝑏 𝜏 ) −𝑐 𝜏 𝜀 |𝜁 |
e . (8.4.22)
|𝑤−𝑤◦ |=𝑟

If now 𝑔 ∈ O (C) then, by the maximum principle, there is a 𝑤★ ∈ C with |𝑤★ −𝑤 ◦ | =

𝑟 ≥ 24 𝜏|𝜁 | such that |𝑔(|𝜁 |)| ≤ |𝑔(𝑤★)|. Since

|𝑤 ∗ − |𝜁 || ≤ |𝑤★ − 𝑤 ◦ | + |𝑤 ◦ − |𝜁 || ≤ 26 𝜏|𝜁 |,

(8.4.22) implies

|𝑔(|𝜁 |)| ≤ |𝐽 | (𝜅+𝑏

𝜀
𝜏 ) 𝑐 𝜏 𝜀 |𝜁 |
e |𝑔(𝑤★) 𝑓 (𝑤★)| (8.4.23)
≤ |𝐽 | (𝜅+𝑏
𝜀
𝜏 ) 𝑐 𝜏 𝜀 |𝜁 |
e sup |𝑔(𝑤) 𝑓 (𝑤)| .
|𝑤− |𝜁 | | ≤26 𝜏 |𝜁 |

Now let Φ ∈ O (C𝑛 ) be arbitrary and set 𝑔(𝑤) = Φ(𝑤𝜁/|𝜁 |). Since
8.4 Solvability of Constant Coefficients Hyperdifferential Equations 253

sup |Φ (𝑤𝜁/|𝜁 |) 𝐽 (𝑤𝜁/|𝜁 |)| ≤ sup |Φ(𝑧)𝐽 (𝑧)|,

|𝑤− |𝜁 | | ≤26 𝜏 |𝜁 | |𝑧−𝜁 | ≤26 𝜏 |𝜁 |

(8.4.23) implies (8.4.20). □

We are now ready to prove the crucial lemma needed in the proof of the holomor-
phic solvability theorem:

Lemma 8.4.10 Let 𝜓 : C𝑛 → R satisfy

∀𝑧, 𝜁 ∈ C𝑛 , |𝜓(𝑧) − 𝜓(𝜁)| ≤ 𝐵|𝑧 − 𝜁 | (8.4.24)

for some 𝐵 > 0. Let 𝐽 ∈ O (C𝑛 ) be of infra-exponential type, satisfying 𝐽 (0) = 1.

Under these hypotheses, if ℎ ∈ O (C𝑛 ) satisfies

∀𝜁 ∈ C𝑛 , |ℎ(𝜁)𝐽 (𝜁)| ≤ 𝐶e 𝜓 (𝜁 ) , (8.4.25)

for some 𝐶 > 0 then, for every 𝛿 ∈ (0, 1), ℎ also satisfies

∀𝜁 ∈ C𝑛 , |ℎ(𝜁)| ≤ 𝐶 𝛿 e 𝜓 (𝜁 )+ 𝛿 |𝜁 | , (8.4.26)

where 𝐶 𝛿 > 0 is a constant independent of ℎ (but depending on 𝐽, 𝐵, 𝐶, 𝛿).

Proof Let 𝜁 ∈ C𝑛 be arbitrary and 𝜏 ∈ (0, 1/32]. Since

6 𝑀 𝜏 |𝜁
e−𝜓 (𝜁 ) sup e−𝜓 (𝑧) ≤ e2 |
,
|𝑧−𝜁 | ≤26 𝜏 |𝜁 |

we derive from (8.4.20):

6𝑀𝜏
𝜏 ) ( 𝜀𝑐 𝜏 +2 ) |𝜁 |
e−𝜓 (𝜁 ) |ℎ(𝜁)| ≤ |𝐽 | (𝜅+𝑏
𝜀 e sup |e−𝜓 (𝑧) ℎ(𝑧)𝐽 (𝑧)|
|𝑧−𝜁 | ≤26 𝜏 |𝜁 |
6𝑀𝜏
𝜏 ) ( 𝜀𝑐 𝜏 +2 ) |𝜁 | .
≤ 𝐶 |𝐽 | (𝜅+𝑏
𝜀 e

If we recall the definitions of 𝑏 𝜏 , (8.4.15), and 𝑐 𝜏 (Lemma 8.4.9) it is easy to select

𝜀 > 0 and 𝜏 ∈ (0, 2−5 ] such that 𝜀𝑐 𝜏 + 26 𝑀𝜏 ≤ 𝛿 ∈ (0, 1). □

We are now in a position to prove the announced solvability theorem for hyper-
differential operators with constant coefficients. According to Proposition 8.1.11 an
arbitrary hyperdifferential operator 𝐽 (𝜕𝑧 ) maps O (ΩC ) into itself.

Theorem 8.4.11 Let ΩC ⊂ C𝑛 be a convex domain and let the entire function of
infra-exponential type 𝐽 (𝜁) not vanish identically. Then the linear map 𝐽 (𝜕𝑧 ) :
O (ΩC ) −→ O (ΩC ) is surjective.

Proof Possibly after a translation in C𝑛 we may assume that 𝐽 (0) ≠ 0; then we may
replace 𝐽 by 𝐽 (0) −1 𝐽 or, equivalently, assume that 𝐽 (0) = 1. In order to prove that
𝐽 (𝜕𝑧 ) O (ΩC ) is dense in O (ΩC ) we must show that 𝐽 (𝜕𝑧 ) : O ′ (ΩC ) −→ O ′ (ΩC ) is
254 8 Hyperdifferential Operators

injective. Let 𝜇 ∈ O ′ (ΩC ) be such that 𝐽 (𝜕𝑧 ) 𝜇 = 0, equivalent to 𝐽 (𝜁)L𝜇(𝜁) = 0 for

all 𝜁 ∈ C𝑛 (L: Laplace–Borel transform). This implies that L𝜇 vanishes identically
and therefore, by Theorem 6.2.2, that 𝜇 = 0.
This established, we are going to apply the general epimorphism theorem for
Fréchet–Montel spaces (Lemma 7.1.17): the desired surjectivity will ensue from the
boundedness of every sequence {𝜇 𝑗 } 𝑗=1,2,... ⊂ O ′ (ΩC ) such that 𝐽 (−𝜕𝑧 )𝜇 𝑗 𝑗=1,2,...
is bounded. The latter means that there are a compact set 𝐾 ⊂ ΩC and a constant
𝐶 (𝐾) > 0 such that

∀ 𝑓 ∈ O (C𝑛 ), sup | 𝐽 (−𝜕𝑧 )𝜇 𝑗 , 𝑓 | ≤ 𝐶 (𝐾) sup | 𝑓 |.

𝑗=1,2,... 𝐾

We have, for all 𝜁 ∈ C𝑛 and all 𝑗 = 1, 2, . . .,

|𝐽 (−𝜁)L𝜇 𝑗 (𝜁)| = | 𝐽 (−𝜕𝑧 )𝜇 𝑗 , e−𝑧·𝜁 | ≤ 𝐶 (𝐾)e 𝐻𝐾 (𝜁 ) (8.4.27)

where 𝐻𝐾 (𝜁) = max (− Re (𝑧 · 𝜁)). Thus we can apply Lemma 8.4.10, with the
𝑧 ∈𝐾
choices ℎ(𝜁) = L𝜇 𝑗 (𝜁), 𝐺 (𝜁) = 𝐽 (−𝜁) and 𝜓(𝜁) = 𝐻𝐾 (𝜁). We conclude that to
each 𝜀 > 0 there is a constant 𝐶 𝜀 > 0, depending only on 𝐶 (𝐾), 𝜀 and |𝐽 | 𝜀 , such
that
∀𝜁 ∈ C𝑛 , ∀ 𝑗 = 1, 2, . . . , |L𝜇 𝑗 (𝜁)| ≤ 𝐶 𝜀 e 𝐻𝐾 (𝜁 )+𝜀 |𝜁 | . (8.4.28)
Since ΩC is convex we can select 𝐾 convex; then (8.4.28) implies that every 𝜇 𝑗 is
carried by 𝐾 (see Remark 6.2.3) and since L : O ′ (C𝑛 ) −→ Exp (C𝑛 ) is a topological
vector spaces isomorphism we see {𝜇 𝑗 } 𝑗=1,2,... is bounded in O ′ (C𝑛 ) and therefore
also in O ′ (ΩC ). □

8.4.3 Solvability in hyperfunctions of hyperdifferential equations with

constant coefficients

The following statement is a kind of “microlocal structure theorem” (cf. Theorem

8.2.3).

Proposition 8.4.12 Let Ω ⊂ R𝑛 be open, Γ ⊂ R𝑛 \ {0} an acute convex open cone

and 𝐹 ∈ O (W𝛿 (Ω, Γ) (𝛿 > 0). If 𝑈 ⊂⊂ Ω is open and convex there is a hyperdif-
ferential operator 𝑄 𝜒 (𝜕) with symbol (8.2.1) and a function

𝐺 ∈ O (W𝛿′ (𝑈, Γ) ∩ C ∞ W𝛿′ (𝑈, Γ)

(0 < 𝛿 ′ < 𝛿) such that 𝑏𝑈 𝐹 = 𝑄 𝜒 (𝜕𝑥 ) ( 𝐺 |𝑈 ).

Wedges are defined in (3.3.1); 𝐺 ∈ C ∞ W𝛿′ (𝑈, Γ) means that all the partial

derivatives of 𝐺 extend continuously to the closure W𝛿′ (𝑈, Γ).

8.4 Solvability of Constant Coefficients Hyperdifferential Equations 255

Proof Let 𝑉 ⊂⊂ Ω be open and convex, 𝑈 ⊂⊂ 𝑉. By Theorem 8.2.3 there are

𝜑 ∈ C ∞ (R𝑛 ) and 𝑄 𝜒 (𝜕𝑥 ) such that 𝑏 Ω 𝐹 = 𝑄 𝜒 (𝜕𝑥 )𝜑 in 𝑉. Since W𝛿 (𝑉, Γ) is
a convex open subset of C𝑛 we can apply to it Theorem 8.4.11: there is an 𝐻 ∈
O (W𝛿 (𝑉; Γ) such that 𝑄 𝜒 (𝜕𝑧 )𝐻 = 𝐹 in W𝛿 (𝑉; Γ). Proposition 8.1.15 implies
𝑄 𝜒 (𝜕𝑥 ) (𝑏 Ω 𝐻 − 𝜑) = 0 in 𝑉; then Theorem 8.3.4 implies that 𝑏 Ω 𝐻 − 𝜑 ∈ C 𝜔 (𝑉).
In particular the restriction of 𝑏 Ω 𝐻 − 𝜑 to 𝑈 extends as a holomorphic function
𝐻• defined in some tube 𝑈 + 𝑖{𝑦 ∈ R𝑛 ; |𝑦| < 𝛿 ′ } (0 < 𝛿 ′ < 𝛿). It follows that
𝐺 = 𝐻 − 𝐻• ∈ O (W𝛿′ (𝑈, Γ)) and 𝜑 = 𝑏𝑈 𝐺 in 𝑈, thereby proving our claim. □
The same kind of reasoning enables us to prove the surjectivity of hyperdifferential
operators with constant coefficients in the space of hyperfunctions in R𝑛 .
Theorem 8.4.13 Let 𝐽 (𝜕𝑥 ) be a hyperdifferential operator with constant coefficients,
not identically zero. If Ω is an open bounded subset of R𝑛 then 𝐽 (𝜕𝑥 )B (Ω) = B (Ω).
Proof By the flabbiness of the sheaf of hyperfunctions we can assume that Ω is
convex. We must show that to every 𝑓 ∈ B (Ω) there is a 𝑢 ∈ B (Ω) such that
𝐽 (𝐷)𝑢 = 𝑓 in Ω. Since every hyperfunction in Ω is the sum of a finite number of
boundary values of holomorphic functions defined in wedges with edge Ω, we are
reduced to the case when 𝑓 = 𝑏 Ω 𝐹, 𝐹 ∈ O (W𝛿 (Ω, Γ) with Γ ⊂ R𝑛 \ {0} a convex
open cone and 𝛿 > 0. Since W𝛿 (Ω, Γ) is a convex subset of C𝑛 , Theorem 8.4.11
implies that there is a 𝐺 ∈ O (W𝛿 (Ω, Γ) satisfying 𝐽 (𝑧, 𝜕𝑧 ) 𝐺 = 𝐹 in W𝛿 (Ω, Γ).
Proposition 8.1.15 yields 𝐽 (𝜕𝑥 )𝑏 Ω 𝐺 = 𝑏 Ω 𝐹 , which completes the proof. □
Essentially the same argument yields what can be viewed as a microlocal version
of Theorem 8.4.13:
Proposition 8.4.14 Let 𝐽 ∈ Exp1,0 (C𝑛 ) not vanish identically. Let Ω be an open
subset of R𝑛 and 𝑓 ∈ B (Ω). If (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω×(R𝑛 \ {0}) does not belong to the analytic
wave-front set of 𝑓 (Definition 7.4.7) then there is an open set 𝑈 ⊂ Ω, 𝑥 ◦ ∈ 𝑈, and
a hyperfunction 𝑢 ∈ B (𝑈) such that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢) and 𝐽 (𝜕𝑥 )𝑢 = 𝑓 in 𝑈.
Proof The hypothesis means that there exist a convex open set 𝑈 ⊂ Ω, 𝑥 ◦ ∈ 𝑈,
finitely many convex open cones Γ1 , . . . , Γ𝑁 ⊂ R𝑛 \ 0 such that 𝜉 ◦ · Γ 𝑗 < 0 for every
𝑗 = 1, . . . , 𝑁, andÍfor some 𝛿 > 0, functions ℎ 𝑗 ∈ O (W𝛿 (𝑈; Γ 𝑗 ), 𝑗 = 1, . . . , 𝑁
such that 𝑓 |𝑈 = 𝑁𝑗=1 𝑏𝑈 ℎ 𝑗 . As before we apply Theorem 8.4.11: for each 𝑗 we
can find a function 𝑔 𝑗 ∈ O (W Í 𝛿 (𝑈; Γ 𝑗 ), 𝑗 = 1, . . . , 𝑁, such that 𝐽 (𝜕𝑥 )𝑔 𝑗 = ℎ 𝑗 in
W𝛿 (𝑈; Γ 𝑗 ). If we define 𝑢 = 𝑁𝑗=1 𝑏 Γ 𝑗 (𝑔 𝑗 ) then (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢) and Proposition
8.1.15 implies that 𝐽 (𝜕𝑥 )𝑢 = 𝑓 in 𝑈.

Corollary 8.4.15 Let 𝐽 (𝜕𝑥 ) be an elliptic hyperdifferential operator (Definition

8.1.1). Let Ω be an open subset of R𝑛 and 𝑢 ∈ B (Ω). If (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × (R𝑛 \ {0})
does not belong to the analytic wave-front set of 𝐽 (𝜕𝑥 )𝑢 then (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢).

Proof By Proposition 8.4.14 we can find 𝑣 ∈ B (R𝑛 ) such that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑣)

and 𝐽 (𝜕𝑥 )𝑣 = 𝐽 (𝜕𝑥 )𝑢 in an open set 𝑈 ⊂ Ω, 𝑥 ◦ ∈ 𝑈. Theorem 8.3.4 implies
𝑣 − 𝑢 ∈ C 𝜔 (𝑈) and therefore (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢). □
Part III
Geometric Background
Chapter 9
Elements of Differential Geometry

The purpose of this chapter is to recall a number of basic definitions in differential

geometry needed for the transition from analysis in Euclidean phase-space Ω × R𝑛
to analysis in the cotangent bundle of a C 𝜔 manifold M. Since much of the content
is valid not only for C 𝜔 manifolds but also for C ∞ as well as complex-analytic
manifolds we introduce the notion of regular manifolds to cover the three classes;
the base field will be R or C, definitely C when dealing with complex-analytic
objects. The headings of the sections (and of the subsections) say all there is to be
said. The last section is devoted to differential complexes, introducing (in the last two
subsections) the two most important of these in geometric analysis, the De Rham
and the Dolbeault (or 𝜕, also referred to as d-bar) complexes. The same material
(and much more) can be found in numerous other texts. The reason for replicating it
here is twofold: to make the presentation in this book as self-contained as possible;
and to ensure that the terminology and notation in the book is sufficiently coherent.

9.1 Regular Manifolds

9.1.1 Regular functions and maps in Euclidean space

So far, the concepts introduced and the results proved relate to objects defined in open
subsets of Euclidean space. In other words we have assumed that the coordinates
in our analysis are kept unchanged throughout. But the study of a PDE frequently
requires changing coordinates. The natural framework is then that of manifolds and
fiber bundles. Here we are faced with an exposition problem: almost every concept
introduced in this chapter and used in the sequel can be formulated within one of
three categories: smooth, real-analytic, complex-analytic. It is convenient to follow
a unified approach: throughout this chapter we shall use the adjective “regular” and
the notation R for either C ∞ , C 𝜔 or O (the latter meaning holomorphic). When

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 259
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_9
260 9 Elements of Differential Geometry

R = O the base field is K = C; in the other two cases K = R. Depending on the

meaning of “regular” one often says that we are reasoning within the smooth (i.e.,
𝐶 ∞ ), or the real-analytic (i.e., 𝐶 𝜔 ) or the complex-analytic category.
In this section we recall the definition of a regular manifold and some of the
terminology associated with such a structure.

Remark 9.1.1 It should be made clear at the outset that practically everything that
will be said in this chapter about our “regular” structures, i.e., either C ∞ , real- or
complex-analytic structures, applies as well to many other structures, for instance
Gevrey structures or quasi-analytic structures, intermediate between C ∞ and C 𝜔 .
These generalizations are self-evident and will not be discussed here.

In K𝑛 we make use of the Cartesian coordinates which, until specified otherwise,

shall be denoted by 𝑥1 , ..., 𝑥 𝑛 even when K = C.

Notation 9.1.2 Let Ω be an open subset of K𝑛 . We shall denote by R (Ω) the ring
of K-valued regular (i.e., C ∞ , C 𝜔 or holomorphic) functions in Ω.

With our choice of the meaning of “regular” the following can be asserted:
(1) The polynomial functions in K𝑛 are regular.
(2) If Ω ⊂ K𝑛 is an open set then R (Ω) can be identified with a subring of the ring
of K-valued smooth functions, C ∞ (Ω; K) .
(3) If Ω1 ⊂ Ω2 ⊂ K𝑛 the restriction of functions from Ω2 to Ω1 induces a ring
homomorphism of R (Ω2 ) into R (Ω1 ).
If Ω1 ⊂ K𝑛1 and Ω2 ⊂ K𝑛2 are two open sets we shall say that a map 𝑓 : Ω1 −→ Ω2
is regular if every component 𝑓𝑖 (𝑖 = 1, ..., 𝑛2 ) of 𝑓 is regular in Ω1 . We shall say
that 𝑓 is a regular isomorphism (in which case, necessarily, 𝑛1 = 𝑛2 = 𝑛) if 𝑓 is a
−1
regular homeomorphism and if its inverse 𝑓 is regular. This means that the Jacobian
determinant 𝐷𝐷((𝑓𝑥11,..., 𝑓𝑛 )
...𝑥𝑛 ) does not vanish at any point of Ω1 . We recall that when
K = R what we call here a regular isomorphism is usually called a diffeomorphism
(of class C ∞ or C 𝜔 ). When K = C and R = O a regular isomorphism is often called
a biholomorphism.

9.1.2 Regular manifolds

For us a regular manifold will be a topological space M (always countable at

infinity, i.e., equal to the union of a sequence of compact subsets) equipped with
an assignment U ↦→ R (U), U being an arbitrary open subset of M and R (U)
a special subring of the ring C (U; K) of the K-valued continuous functions in U.
The elements of R (U) are the regular functions (i.e., C ∞ , C 𝜔 or O) in U (U itself
regarded as a regular manifold). The correspondence U ↦→ R (U) must satisfy the
following three “axioms”:
9.1 Regular Manifolds 261

(M1) Restriction to any open set V ⊂ U defines a ring homomorphism R (U) −→

R (V). Ø
(M2) Let U = U 𝜄 be the union of a family of open subsets of M. If the restriction
𝜄 ∈𝐼
to U 𝜄 of a continuous function 𝑓 in U belongs to R (U 𝜄 ) for every 𝜄 ∈ 𝐼 then
𝑓 ∈ R (U).
(M3) Every point ℘ ∈ M is contained in an open set U having the following
property: there is a regular isomorphism 𝜑 of U onto an open subset of K𝑛
such that the pullback map R (𝜑 (U)) ∋ 𝑓 ↦→ 𝑓 ◦𝜑 is an algebra isomorphism
onto R (U). The integer 𝑛 is independent of the point ℘.
Simply phrased, this definition states that we know what the regular (i.e., C ∞ , C 𝜔
or O) functions in our manifold are. According to (M2)–(M3) the regularity of a
function is completely determined by its regularity in a neighborhood of each point
of its domain of definition.
The number 𝑛 is the dimension of the manifold M and is denoted by dimK M or
simply by dim M if there is no danger of confusion.
Property (M3) states that the topological space M is locally Euclidean. If U and
𝜑 are as in (M3) the pair (U, 𝜑) is called a local chart in M. We can make use of
the map 𝜑 to pullback from K𝑛 to U the Cartesian coordinates 𝑥 𝑗 : if ℘ ∈ U we write
𝑥 𝑗 (℘) = 𝑥 𝑗 (𝜑 (℘)). This defines a system (𝑥 1 , ..., 𝑥 𝑛 ) of local coordinates in U; it
is then customary to write (U, 𝑥1 , ..., 𝑥 𝑛 ) rather than (U, 𝜑); we shall often refer to
(U, 𝑥1 , ..., 𝑥 𝑛 ) as a coordinate chart (or patch). Note that this convention allows us
to use the same notation 𝑓 (𝑥1 , ..., 𝑥 𝑛 ) for a function in 𝜑 (U) and for its pullback
𝑓 ◦ 𝜑 in the local coordinates (𝑥 1 , ..., 𝑥 𝑛 ). A coordinate change in U is equivalent to
a modification of the map 𝜑.
Ø
Proposition 9.1.3 Let U = U 𝜄 be the union of a family of open subsets of M.
𝜄 ∈𝐼
Suppose that to each 𝜄 ∈ 𝐼 there is an 𝑓 𝜄 ∈ R (U 𝜄 ) such that, whatever 𝜄1 , 𝜄2 ∈ 𝐼, the
restrictions of 𝑓 𝜄1 and 𝑓 𝜄2 to U 𝜄1 ∩ U 𝜄2 are equal. Under this hypothesis there is a
unique function 𝑓 ∈ R (U) whose restriction to U 𝜄 is equal to 𝑓 𝜄 for each 𝜄 ∈ 𝐼.

Proof Indeed, there is a unique continuous function in U whose restriction to U 𝜄 is

equal to 𝑓 𝜄 for each 𝜄 ∈ 𝐼. The claim is therefore a consequence of (M2). □

Proposition 9.1.4 The constant functions belong to R (M).

Proof By (M3) the constant functions belong to R (U) if U is the domain of a local
chart. Domains of local charts form a covering of M, whence the claim, by (M2).□

Remark 9.1.5 If U is an arbitrary open subset of M the constants might very well
be the sole elements of R (U). This is always the case when R = O and U is a
compact connected component of the complex manifold M.

Proposition 9.1.6 Let (U 𝜄 , 𝜑 𝜄 ), 𝜄 = 1, 2, be two local charts in M such that U1 ∩

U2 ≠ ∅. Then the map 𝜑1 ◦ 𝜑−1 2 : 𝜑2 (U1 ∩ U2 ) −→ 𝜑1 (U1 ∩ U2 ) is a regular
isomorphism.
262 9 Elements of Differential Geometry

Proof The map 𝜑1 ◦ 𝜑−1 2 is a homeomorphism; its components are regular by (M3)
−1
and its inverse, 𝜑2 ◦ 𝜑1 , has the same properties. □
By a regular map 𝜒 : M1 −→ M2 between manifolds of the same regularity
set U2 ⊂ M2
class we shall mean a continuous map such that, given any open
−1
and any function 𝑓 ∈ R (U2 ), the pullback 𝑓 ◦ 𝜒 belongs to R 𝜒 (U2 ) . This is
seen easily to agree with the definition in Subsection 9.1.1 when M 𝑗 is an open set
Ω 𝑗 ⊂ K𝑛 𝑗 . The definition of a regular isomorphism 𝜒 (i.e., “diffeomorphism” in the
C ∞ or C 𝜔 cases, “biholomorphism” in the complex-analytic case) is obvious: the
−1
regular map 𝜒 must be a homeomorphism and its inverse 𝜒 : M2 −→ M1 must be
regular. A regular isomorphism of M onto itself is a regular automorphism of M.
The definition of a product manifold M1 × M2 is self-evident: its structure is
defined by the local charts of the kind (U1 × U2 , (𝜑1 , 𝜑2 )).

Definition 9.1.7 By a submanifold of M of dimension 𝑑, 0 ≤ 𝑑 < 𝑛, we shall mean

a subset L of M every point of which lies in the domain U of a system of regular
local coordinates 𝑥 1 , ..., 𝑥 𝑛 so that

L ∩ U = {℘ ∈ U; 𝑥 𝑑+1 (℘) = · · · = 𝑥 𝑛 (℘) = 0} .

A zero-dimensional submanifold of M is a discrete (possibly infinite) set of

points. An 𝑛-dimensional submanifold of M is an open subset of M.
A submanifold L of M is itself a manifold, as we see by positing that a function 𝑓
in an open subset V of L is regular in V if its restriction to any domain L ∩ U ⊂ V,
with U as in Definition 9.1.7, is a regular function of 𝑥 1 , ..., 𝑥 𝑑 . It is readily checked
that the conditions (M1)–(M2)–(M3) are satisfied. The number codim L = dim M −
dim L is the codimension of L. The natural injection L ↩→ M is a regular map. It
is not difficult to check that a submanifold L of M is countable at infinity (if this is
true of M).
If “regular” means C ∞ (resp., C 𝜔 , resp., complex-analytic) we shall say that L is
a C ∞ , or smooth (resp., C 𝜔 , or real-analytic, resp., complex-analytic) submanifold
of M.
A regular submanifold of M need not be a closed subset of M. Examples: any
convergent sequence of distinct points, not including their limit; or

Example 9.1.8 The spiral in R2 defined by 𝑟 = exp (−𝜃) in polar coordinates (0 <
𝜃 < +∞) is a real-analytic submanifold of R2 .

If L and L ′ are two analytic submanifolds of M and if L ∩ L ′ is an open subset

of each, L ∪L ′ might not be an analytic submanifold of M.

Example 9.1.9 Let Λ be the Bernoulli lemniscate, defined by the equation

2
𝑥2 + 𝑦2 = 𝑥2 − 𝑦2
9.1 Regular Manifolds 263

in the plane. Both L = {(𝑥, 𝑦) ∈ Λ; 𝑥 > 0} and L ′ = {(𝑥, 𝑦) ∈ Λ; 𝑥𝑦 ≥ 0} are

connected C 𝜔 submanifolds of R2 ; it is evident for L. It is true for L ′ since it is
obvious in the complement of the origin and since L ′ is defined by the equation
√︃
𝑦 = 𝑥 1 − 4𝑥 2 + 16𝑥 4 + 𝑂 𝑥 6

in a sufficiently small neighborhood of (0, 0). The origin is a singularity of L ∪L ′.

Remark 9.1.10 Later (Definition 9.3.7) we shall encounter a “weaker” notion of
submanifold, that of an immersed submanifold. When the need arises to distinguish
between the two notions we shall refer to the submanifolds just defined as embedded
submanifolds.
Definition 9.1.11 By a regular embedding of a regular manifold M1 into another
one (of the same category), M2 , we shall mean a regular isomorphism of M1 onto
a regular submanifold of M2 .
It is evident that a complex-analytic manifold of complex dimension 𝑛 can be
viewed as a C 𝜔 manifold of real dimension 2𝑛.
The following result, stated here without proof (see [Grauert, 1958]), is often
useful.
Theorem 9.1.12 Let M be a C 𝜔 manifold, dim M = 𝑛. There is a C 𝜔 embedding
of M into Euclidean space R2𝑛+1 .
We can however prove the following by applying Theorem 9.1.12.
Corollary 9.1.13 Let M be a C 𝜔 manifold, dim M = 𝑛. There is a C 𝜔 embedding
of M in a complex-analytic manifold M
f of complex dimension 𝑛.

Proof By Theorem 9.1.12 there is no loss of generality in assuming that M is an

analytic submanifold of R 𝑁 (𝑁 ≥ 2𝑛 + 1). There is a covering of M by open balls
𝑈 𝜄 (𝜄 ∈ 𝐼) of R 𝑁 such that

M ∩ 𝑈 𝜄 = 𝑥 ∈ 𝑈 𝜄 ; 𝑓 𝜄,1 (𝑥) = · · · = 𝑓 𝜄, 𝑁 −𝑛 (𝑥) = 0 ,

where the functions 𝑓 𝜄, 𝑗 ∈ C 𝜔 (𝑈 𝜄 ) are real-valued and the differentials

d 𝑓 𝜄,1 , ..., d 𝑓 𝜄, 𝑁 −𝑛

are linearly independent at every point of 𝑈 𝜄 . Regarding R 𝑁 as the real part of

C 𝑁 there is a unique holomorphic extension 𝑓˜𝜄, 𝑗 of each 𝑓 𝜄, 𝑗 to an open subset 𝑈 f𝜄
(homeomorphic to a 2𝑁-dimensional open ball) of C such that 𝑈 𝜄 = 𝑈 𝜄 ∩ R and𝑁 e 𝑁

𝜕𝑧 𝑓˜𝜄,1 ∧ · · · ∧ 𝜕𝑧 𝑓˜𝜄, 𝑁 −𝑛 ≠ 0 at every point of 𝑈 f𝜄 . On nonempty overlaps 𝑈 𝜄 ∩ 𝑈 𝜄′ we

have 𝑓 𝜄′ , 𝑗 = Φ 𝜄, 𝜄′ , 𝑗 ( 𝑓 𝜄,1 , ..., 𝑓 𝜄, 𝑁 −𝑛 ); these relations are preserved in the holomorphic
extensions. The union (as 𝜄 ranges over the index set 𝐼) of the sets
n o
M f𝜄 = 𝑧 ∈ 𝑈 e𝜄 ; 𝑓˜𝜄,1 (𝑧) = · · · = 𝑓˜𝜄, 𝑁 −𝑛 (𝑧) = 0
264 9 Elements of Differential Geometry
Ø
is a complex-analytic submanifold M
f of the open subset 𝑈e𝜄 of C 𝑁 in which M
𝜄 ∈𝐼
is C 𝜔 embedded; obviously, dimC M
f = 𝑛. □

Definition 9.1.14 Let M be a C 𝜔 manifold, dimR M = 𝑛. A complex-analytic

manifold Mf of complex dimension 𝑛 in which M is C 𝜔 embedded as a totally real
submanifold is called a complexification of M.

Totally real means that every point ℘ ∈ M belongs to a domain U e ⊂ M

f of
complex-analytic coordinates 𝑧1 , ..., 𝑧 𝑛 such that
n o
M∩U e = ℘ ∈ U; e Im 𝑧1 = · · · = Im 𝑧 𝑛 = 0 .

Remark 9.1.15 It is not difficult to construct a complexification of M directly, which

is to say, without applying Theorem 9.1.12. One uses a locally finite covering of M
by domains of real-analytic local coordinates, carries out the local embeddings and
patches together the local constructions.

Remark 9.1.16 The complex-analytic analogue of Theorem 9.1.12 is not true, unless
an additional condition is imposed upon M: M must be a Stein manifold. On this
topic we refer the reader to standard texts in Several Complex Variables theory, e.g.
[Gunning and Rossi, 1965] and [Hörmander, 1966]. In [Grauert, 1958] it is proved
that every C 𝜔 manifold has a Stein complexification.

Another result that may be of interest is the following

Proposition 9.1.17 Let M be a complex-analytic manifold and L a connected C 𝜔

submanifold of M. If a nonempty open subset of L is a complex-analytic submanifold
of M then L is a complex-analytic submanifold of M.

Needless to say, both the hypothesis and the conclusion require dimR L to be
even.
Proof Let (U, 𝑧1 , ..., 𝑧 𝑛 ), U =Δ𝑟(𝑛) = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 < 𝑟, 𝑗 = 1...𝑛 , be a complex-

analytic coordinate patch in M such that U ∩ L is defined in U by the equations

𝑧 𝑘 = 𝑓 𝑘 (𝑧1 , ..., 𝑧 𝑚 , 𝑧¯1 , ..., 𝑧¯𝑚 ), 𝑘 = 𝑚 + 1, ..., 𝑛, 𝑓 𝑘 ∈ C 𝜔 Δ𝑟𝑚 (𝑚 = 21 dimR L). If
𝜕 𝑓𝑘 𝜕 𝑓𝑘
𝜕 𝑧¯ 𝑗 ≡ 0 (1 ≤ 𝑗 ≤ 𝑚, 1 ≤ 𝑘 ≤ 𝑛) in a nonempty open subset of Δ𝑟 we have 𝜕 𝑧¯ 𝑗 ≡ 0
𝑚

in the whole of Δ𝑟𝑚 . Since L is connected this proves the claim. □

An important concept is that of orientation of a manifold. Here we give the most
direct definition. A linear change of coordinates in R𝑛 is orientation preserving if its
determinant is positive. We can partition the general linear group GL (𝑛, R) into two
connected components, the sign of the determinant of their elements being constant
in each but opposite to that in the other. What the sign actually is depends on the
choice of the canonical coordinates in R𝑛 , meaning a starting linear basis, which
allows us to represent each element of GL (𝑛, R) by a nonsingular 𝑛 × 𝑛 real matrix.
9.1 Regular Manifolds 265

Definition 9.1.18 A real C ∞ manifold M is said to be orientable if there is a

covering of M by local charts (U, 𝑥1 , ..., 𝑥 𝑛 ) such that the Jacobian determinants of
the change of coordinates in overlaps are all positive.

A familiar example of a nonorientable submanifold is the Möbius band in R3 .

Equipping R2𝑛 with a complex structure, in other words, specifying canonical
complex coordinates 𝑧1 , ..., 𝑧 𝑛 , directly fixes a preferred orientation on the real
vector space R2𝑛 , the one determined by the coordinates (𝑥1 , 𝑦 1 , ..., 𝑥 𝑛 , 𝑦 𝑛 ) or the
one determined by (𝑥1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 ) (these two orientations are opposite unless
𝑛 = 4𝑚, 𝑚 ∈ Z+ ). It follows that every complex-analytic manifold is orientable when
regarded as a real C 𝜔 manifold.

9.1.3 Function algebras on regular manifolds

With the atlas (the set of all local charts) of the (regular) manifold M at our
disposal we can define a great variety of function algebras on M, at least those that
can be defined by their local properties and provided these properties are invariant
under (regular) coordinate changes. First of all, complex-valued C 𝑘 (0 ≤ 𝑘 ≤ +∞)
functions are well-defined in both real-analytic and complex-analytic manifolds,
and real-analytic functions in the latter. This can be rephrased by saying that a
real-analytic manifold structure on M determines a smooth manifold structure on
M of the same dimension; and that a complex-analytic manifold structure on M
determines a real-analytic manifold structure on M and, as a consequence, also one
of a C ∞ manifold. In the latter case dimR M = 2 dimC M. Of course the notion of a
C 𝜔 function does not make sense in an arbitrary C ∞ manifold.
We shall denote by C 𝑘 (M) (0 ≤ 𝑘 ≤ +∞) the algebra of complex-valued C 𝑘
functions in M. By definition, the support of a function 𝑓 ∈ C 0 (M) is the closure
of the set of points where 𝑓 ≠ 0 and will be denoted by supp 𝑓 . We shall denote by
Cc𝑘 (M) the subalgebra of C 𝑘 (M) consisting of the compactly supported functions
𝑓 ∈ C 𝑘 (M).
Since M is countable at infinity there exist C ∞ partitions of unity subordinate to
any open covering of M.
It is permissible to talk of linear partial differential operators with smooth coef-
ficients in M. They are the linear endomorphisms 𝑃 : C ∞ (M) ←↪ that have the
following property: given any local chart (U, 𝜑) in M there is a differential operator
𝑃 (𝑥, 𝜕𝑥 ) : C ∞ (𝜑 (U)) ←↪ such that, for any function 𝑓 ∈ Cc∞ (U),
−1

𝑃 𝑓 = 𝑃 (𝑥, 𝜕𝑥 ) 𝑓 ◦ 𝜑 ◦ 𝜑. (9.1.1)

Just as in the Euclidean situation we have supp 𝑃 𝑓 ⊂ supp 𝑓 and, as in the Euclidean
situation, this property characterizes differential operators amongst linear endomor-
phisms of C ∞ (M). We say that the differential operator 𝑃 has regular coefficients
if given any open subset U of M, 𝑃 maps the set R (U) of regular functions in
266 9 Elements of Differential Geometry

U into itself. This is the same as saying that in any local chart (U, 𝑥1 , ..., 𝑥 𝑛 ) the
coefficients of the differential operator 𝑃 (𝑥, 𝜕𝑥 ) representing 𝑃 are regular (i.e., C ∞ ,
C 𝜔 or holomorphic).

9.2 Fibre Bundles, Vector Bundles

9.2.1 Fibre bundles on a manifold

We start by introducing the terminology of fiber bundles. For the moment, by a fiber
bundle over a regular (i.e., C ∞ , C 𝜔 or O) manifold M we simply mean a regular (of
the same class) manifold B together with a regular surjective map 𝜋 : B −→ M.
The map 𝜋 is the base projection; actually, M itself is often referred to as the base
−1
manifold or, simply, the base. The set 𝜋 (℘) is the fiber at the point ℘; it will often
be denoted by B℘ . The fibers B℘ will be regular submanifolds of B; note that they
−1
are closed sets. If 𝐸 is a subset of M we shall often use the notation B| 𝐸 = 𝜋 (𝐸).
A section of B over an open set U ⊂ M is a map 𝑠 : U −→ B such that 𝜋 ◦ 𝑠 is
the identity map of U. The section 𝑠 can be continuous, C 𝑘 , etc., as well as regular.
We shall denote by Γ (U; B) the set of all sections of B over the open set U.
In practice the fiber bundle B is always assumed to admit a regular manifold F
as typical fiber and a structure group G. We explain what this means. First of all the
following condition must be satisfied:
(FB 1) Every point of M is contained in an open set U such that there is a regular
−1
isomorphism 𝜓 of 𝜋 (U) onto U × F mapping the fiber B℘ onto {℘} × F
for each ℘ ∈ U.
The pair (U, 𝜓) in (FB 1) is often called a trivialization of the fiber bundle B
over U.
Our second condition constrains the admissible changes of local trivializations.
This is where the structure group G comes into play. We shall require that

• G is a group of regular automorphisms of F , with G itself equipped with the

structure of a regular manifold.
In this text, most often, G will be a finite-dimensional Lie group, either real
or complex (and therefore always a C 𝜔 or a complex-analytic manifold). In some
instances G might be a finite group. But, mainly, it will be one of the “simpler” linear
groups: the full 𝑟 th linear group GL (𝑟, K), whose elements are the nonsingular 𝑟 × 𝑟
matrices with entries in the field K = R or C, or one of its usual subgroups such as
the group of rotations, the orthogonal group, the unitary group, etc. The manifold
structure underlying the Lie group structure of GL (𝑟, K) is that of a C 𝜔 or of a
complex-analytic manifold.
9.2 Fibre Bundles, Vector Bundles 267

With (U, 𝜓) as in (FB 1) we introduce the composite map

𝜓 (℘) : B℘ −→ {℘} × F −→ F , (9.2.1)

in which the first arrow stands for the restriction of 𝜓 to B℘ and the second arrow
for the second coordinate projection
(℘, v) ↦→ v. Obviously 𝜓 (℘) is a regular
isomorphism. Now let U, 𝜓 𝑗 , 𝑗 = 1, 2, be two trivializations of B over U and for
−1
z }| {
each ℘ ∈ U, let 𝜓 𝑗 (℘) be the analogue of (9.2.1). The composite 𝜓1 (℘) ◦ 𝜓2 (℘) is
a regular automorphism of F . We require that the following condition be satisfied:

(FB 2) If U, 𝜓 𝑗 , 𝑗 = 1, 2, are two trivializations of B over U then U ∋℘ ↦→
−1
z }| {
𝜓1 (℘) ◦ 𝜓2 (℘) is a regular map into G.
This completes, as far as we are concerned, the definition of a regular fiber
bundle. Once again we underline the fact that “regular” means either “smooth”, i.e.,
C ∞ , “real-analytic”, i.e., C 𝜔 , or “complex-analytic”, i.e., O: holomorphic.
Let (B, 𝜋) be a fiber bundle over M. A regular section 𝑠 of B over an open
subset U of M is a regular map U −→ M which is a section, i.e., such that 𝜋 ◦ 𝑠 =
identity map of U. We shall denote by R (U; B) the set of regular sections of B
over U.
By a regular bundle map of B into a fiber bundle B ′ over another (possibly the
same) regular manifold M ′, we shall mean a regular map 𝑓 of the manifold B onto
the manifold B ′ mapping each fiber B℘ into some fiber B℘′ ′ . Using trivializations
one sees readily that the map M ∋ ℘ ↦→ ℘′ ∈ M ′ must be a regular map. The
definitions of regular bundle isomorphisms and automorphisms are obvious.
The regular product manifold M × F (see Subsection 9.1.2) with the first coordi-
nate projection (℘, v) ↦→ ℘ as the base projection can be regarded as a regular fiber
bundle over M, with typical fiber F and structure group consisting solely of the
identity map of F . Any regular fiber bundle B over M isomorphic to the product
M × F is called a trivial bundle.
One can restrict to a regular submanifold L of M (see Subsection 9.1.2) a regular
fiber bundle B over M to form a regular fiber bundle over L which we shall denote
by B| L : the fiber of B| L at ℘ ∈ L is simply the fiber of B at ℘, B℘ .
A simple example of a fiber bundle over M is that of the orientation bundle
of M: the fiber at each point has two elements, corresponding to the two possible
orientations of the regular local charts (U, 𝑥1 , ..., 𝑥 𝑛 ), ℘ ∈ U. Over U this bundle
is isomorphic to U× ({−1} × {+1}). Thus the typical fiber is the set {−1} × {+1},
the structure group G is the group of permutations of two objects, isomorphic to
Z2 = Z/2Z. A regular section over an open subset of M is a locally constant section;
to say that M is orientable (Definition 9.1.18) is the same as saying that there is a
regular section of the orientation bundle over M.
268 9 Elements of Differential Geometry

9.2.2 Vector bundles

Definition 9.2.1 A regular vector bundle over a regular manifold M is a regular

fiber bundle B with typical fiber K𝑟 (𝑟 ∈ Z+ ) and structure group GL (𝑟, K).

The integer 𝑟 is called the rank (or the fiber-dimension) of the vector bundle B
over M and shall be denoted by rankK B or simply rank B if there is no danger of
confusion.
The zero-section of the vector bundle B is invariantly defined, thanks to the
property (FB 2). The complement of the zero section in B will be denoted by B\0.
If (U, 𝜑) is a local trivialization (also called affinization) of the bundle B then,
for each ℘ ∈ U, the bijection (9.2.1) defines on the fiber B℘ the structure of an 𝑟-
dimensional vector space over the field K, precisely that structure which makes (9.2.1)
into a linear isomorphism. The canonical basis e1 , ..., e𝑟 of K𝑟 can be identified with
a system of regular sections of U × K𝑟 ; their pullbacks under the map 𝜑, 𝑠1 , ..., 𝑠𝑟 ,
form a system of regular sections of B over U, often referred to as a frame (here
as a regular, meaning C ∞ , C 𝜔 or holomorphic, frame); their values at an arbitrary
point ℘ ∈ U form a basis of the vector space B℘ . A change of trivialization over U
results in a linear automorphism of B℘ . The following statement is obvious.

Proposition 9.2.2 For the regular vector bundle B of rank 𝑟 over M to be trivial it
is necessary and sufficient that there be a regular frame of B over M.

The dual bundle B ∗ of B can be defined “fiberwise”: its fiber at an arbitrary point
℘ ∈ M is the dual B℘∗ of B℘ , that is to say, the vector space of linear functionals
B℘ −→ K. Let (U, 𝜓) be a trivialization of B over an open set U ⊂ M. The
restriction of 𝜓 to B℘ , ℘ ∈ U, induces the isomorphism 𝜓 (℘) : B℘ −→ K𝑟 ; its
transpose 𝜓 (℘) ⊤ : K𝑟 −→ B℘∗ is also an isomorphism, and so is its contragredient
−1
𝜓 (℘) ⊤ : B℘∗ −→ K𝑟 . We have identified K𝑟 with its own dual by means of the
dot product 𝑥 · 𝑦 = 𝑥1 𝑦 1 + · · · + 𝑥 𝑛 𝑦 𝑛 . Denote by 𝜋 ∗ the base projection B ∗ −→ M;
it is not difficult to check that the maps
−1 −1
𝜋 ∗ (U) ∋ (℘, 𝜃) ↦→ ℘, 𝜓 (℘) ⊤ 𝜃 ∈ U × K𝑟

define on B ∗ the structure of a regular vector bundle. The key point is that if (U, 𝜒)
is any other trivialization of B over U the map
−1 −1
⊤
U ∋ ℘ ↦→ 𝜒 (℘) ⊤ ◦ 𝜓 (℘) ⊤ = 𝜓 (℘) ◦ 𝜒 (℘) ∈ GL (𝑟, K)

is regular.
The standard operations on vector spaces, such as direct sum, product, tensor
product, exterior product, can also be extended “fiberwise” to vector bundles on
one and the same manifold, in the obvious fashion. If B (1) and B (2) are two regular
vector bundles over the regular manifold M (which presumes the base field K fixed)
the direct sum B (1) ⊕ B (2) is the vector bundle whose fiber at an arbitrary point
9.2 Fibre Bundles, Vector Bundles 269

℘ ∈ M is the direct sum B℘(1) ⊕ B℘(2) of the fibers of the factors (sometimes this is
referred to as the Whitney sum of B (1) and B (2) ). Likewise for the product ×, the
tensor product ⊗ (over the field K), and the exterior product ∧. Checking that the
resulting vector bundles are regular is straightforward.
When K = R a very important operation is complexification, understood in the
sense that the fibers (not the base!) are complexified. The complexification of the
vector bundle B, which we shall denote by CB, is simply the tensor product B⊗
(M × C), with the understanding that the tensor products of the fibers at an arbitrary
point ℘ ∈ M are taken over R, i.e., B℘ ⊗ C = B℘ ⊗R C. If the fibers B℘ ⊗ C
are regarded as real vector spaces then CB is a regular vector bundle over M;
rankR CB = 2 rankR B.
The concept of a regular vector subbundle B ′ of the regular vector bundle B
is self-evident and then so is the concept of the quotient vector bundle B/B ′,
automatically regular, and of the associated maps: the injection B ′ ↩→ B and the
quotient map B −→ B/B ′. These two bundles are also defined fiberwise.

9.2.3 Section spaces and differential operators

Henceforth we shall always assume that M is a regular manifold (occasionally

abbreviated to “manifold”).
Let B be a regular vector bundle over M. It is very convenient to denote by
R (U; B) the set of regular sections of B over an open subset U of M, hoping
that the context will prevent R (U; B) from being confused with the set of regular
maps U −→ B (based on the tacit agreement that B is not just a manifold; it is
a vector bundle). The set R (U; B) carries a natural vector space structure over K,
actually of a module over the ring R (U) of regular K-valued functions in U. It also
carries a natural locally convex space structure. Likewise, it is convenient to denote
by C 𝑘 (U; B) (0 ≤ 𝑘 ≤ ∞, or 𝑘 = 𝜔 when regular means analytic) the set of C 𝑘
sections of B over U; C 𝑘 (U; B) is also a module over R (U), meaning that C 𝑘
sections of B over U can be added, or multiplied by functions belonging to R (U),
to yield C 𝑘 sections of B over U.
Actually, a frame {𝑠1 , ..., 𝑠𝑟 } of B over U consisting of regular sections allows
us to define on each one of the spaces R (U; B) and C 𝑘 (U; B) the structure of a
module over the ring M𝑟 (R (U)) of 𝑟 × 𝑟 matrices with entries in R (U).
The support of a continuous section 𝑠 of B is the closure supp 𝑠 of the set of
points ℘ ∈ M such that 𝑠 (℘) ≠ 0. We shall denote by Cc𝑘 (U; B) (0 ≤ 𝑘 ≤ ∞) the
space of compactly supported C 𝑘 sections of B over U. Needless to say, there is no
question of considering the compactly supported C 𝜔 sections of B over U unless
U is a compact connected component of M.
A differential operator 𝑃 with smooth coefficients transforms linearly the C ∞
sections of a C ∞ vector bundle B (1) over a C ∞ manifold M into C ∞ sections of a
C ∞ vector bundle B (2) over the same manifold M:
270 9 Elements of Differential Geometry

𝑃 : C ∞ M; B (1) −→ C ∞ M; B (2) , (9.2.2)

with the additional property that supp 𝑃𝑠 ⊂ supp 𝑠 for every 𝑠 ∈ C ∞ M; B (1) . An
easy way of defining 𝑃 is by reasoning within affinizations of the vector bundles.
Let 𝑠1 , ..., 𝑠𝑟1 be a C ∞ frame of B (1) over an open set U ⊂ M and 𝑡1 , ..., 𝑡𝑟2 be

a C ∞ frame of B (2) , also over U. An arbitrary C ∞ section 𝑓 of B (1) over U can
be written as a linear combination 𝑓 = 𝑟𝑗=1 𝑓 𝑗 𝑠 𝑗 with coefficients 𝑓 𝑗 that are C ∞
Í1

functions in U. The action of 𝑃 on 𝑓 will result in a C ∞ section of B (2) over U

which is a linear combination
𝑟2 ∑︁
∑︁ 𝑟1
𝑃𝑓 = (9.2.3)
© ª
­ 𝑃 𝑗,𝑘 𝑓 𝑗 ® 𝑡 𝑘
𝑘=1 « 𝑗=1 ¬
where the 𝑃 𝑗,𝑘 are differential operators with C ∞ coefficients in U acting on func-
tions. It is important to keep in mind that 𝑃 is globally defined, and therefore the
right-hand side in (9.2.3) must be independent not only of local coordinates (as the
differential operators 𝑃 𝑗,𝑘 are meant to be) but also of the local frames: any changes
of these frames by smooth linear substitutions must be reflected in the conjugacy of
the 𝑟 1 × 𝑟 2 matrix 𝑃 𝑗,𝑘 with the matrices representing these substitutions.
When “regular” means “real-analytic” or “complex-analytic” we shall say that
𝑃 is an analytic differential
operator, or that 𝑃 has analytic coefficients, if
𝑃R U; B (1) ⊂ R U; B (2) whatever the open set U ⊂ M. Of course, to preserve
the analyticity of 𝑃 the changes of coordinates and of frames alluded to above must
also be analytic.
When “regular” means “smooth” or “real-analytic”, i.e., when the base field is
R, we shall often deal with differential operators with complex coefficients, meaning
linear differential operators

𝑃 : R M; CB (1) −→ R M; CB (2) , (9.2.4)

CB ( 𝑗) being the complexification of B ( 𝑗) .

9.2.4 Associated sphere and projective bundles

Definition 9.2.3 Let B be a vector bundle over a manifold M.

(1) By the sphere bundle SB associated with B we shall mean the quotient of B\0
modulo the equivalence relation

(℘, v) (℘′, v′) meaning: ℘ = ℘′ and ∃𝜆 > 0, v = 𝜆v′ .

9.3 Tangent and Cotangent Bundles of a Manifold 271

(2) By the projective bundle PB associated with B we shall mean the quotient of
B\0 modulo the equivalence relation

(℘, v) (℘′, v′) meaning: ℘ = ℘′ and ∃𝜆 ∈ K\ {0} , v = 𝜆v′ .

The projective bundle PB associated with a regular vector bundle B can be

equipped with the structure of a regular fiber bundle, with GL (𝑟, K) as structure
group (𝑟 = rank B). The same is true of the sphere bundle SB provided K = R, i.e.,
provided B is of class C ∞ or C 𝜔 . But when K = C and the bundle B is complex-
analytic the sphere bundle SB is merely real-analytic (it does carry another structure,
that of a CR bundle – which we shall not discuss here).
A point of SB ℘ (℘ ∈ M) can be identified with a ray of B℘ , namely the set of
points 𝜆v, v ∈ B℘ \ {0}, 𝜆 > 0. A point of PB℘ can be identified with a “line” in B℘ ,
the set of points 𝜆v, v ∈ B℘ \ {0}, 𝜆 ∈ K.
The “conic terminology” so often used in Euclidean microlocal analysis (starting
with Definition 3.5.1) is very convenient when extended to vector bundles:

Definition 9.2.4 A subset U of a vector bundle B is said to be conic if (℘, v) ∈ U

implies (℘, 𝜆v) ∈ U for all 𝜆 > 0.

In practice we shall often deal with conic subsets U ⊂ B\0. Every subset 𝑬 of
the sphere bundle SB determines a conic subset E of B\0, namely the preimage of
𝑬 under the quotient map B\0 −→ SB: E is the set of rays corresponding to the
points of 𝑬.
For both fiber bundles SB and PB the structure group GL (𝑟, K) can be replaced
by a “smaller” group. Indeed, the multiples 𝜆𝐼𝑟 , 𝜆 > 0, of the 𝑟 × 𝑟 identity matrix 𝐼𝑟
and the multiples 𝜆𝐼𝑟 , 0 ≠ 𝜆 ∈ K, make up normal subgroups of GL (𝑟, K) which we
can denote, respectively, by R+ and K∗ . It is clear that the fiber bundle SB (resp., PB)
admits the quotient group GL (𝑟, K) /R+ [resp., GL (𝑟, K) /K∗ ] as structure group.
The first quotient is naturally isomorphic to the subgroup of GL (𝑟, K) consisting
of the 𝑟 × 𝑟 matrices 𝑀 such that |det 𝑀 | = 1 while GL (𝑟, K) /K∗ is naturally
isomorphic to the subgroup SL (𝑟, K) consisting of the 𝑟 × 𝑟 matrices 𝑀 such that
det 𝑀 = 1.

9.3 Tangent and Cotangent Bundles of a Manifold

From our viewpoint the most important vector bundles on a manifold are the tangent
bundle and the cotangent bundle. We recall their definition, mainly with the aim of
clarifying our terminology and notation.
272 9 Elements of Differential Geometry

9.3.1 Tangent bundle

Let (U, 𝜑) be a local chart in a manifold M. Consider a vector field

𝑛
∑︁ 𝜕
𝑋= 𝑎 𝑗 (𝑥) (9.3.1)
𝑗=1
𝜕𝑥 𝑗

in the open set 𝜑 (U) ⊂ K𝑛 . The coefficients of 𝑋 can be selected in any function
class of our choice; for the time being we shall take them to be smooth, although
this requirement may be very much weakened (or strengthened) if need be. What is
more important is that they should be valued in K, i.e., they should be real in the
C ∞ and C 𝜔 cases. We point out, however, that when K = C, the partial derivative
𝜕 𝜕 1 𝜕
√ 𝜕

𝜕𝑥 𝑗 means the holomorphic derivative: 𝜕𝑥 𝑗 = 2 𝜕 Re 𝑥 𝑗 − −1 𝜕 Im 𝑥 𝑗 . In Euclidean
space K𝑛 the vector field 𝑋 is often identified with the vector a = (𝑎 1 (𝑥) , ..., 𝑎 𝑛 (𝑥))
visualized as originating at the point 𝑥. This attaches to 𝑥 a vector space over K,
the tangent space 𝑇𝑥 K𝑛 at 𝑥. We also have the option of regarding 𝑋 as a first-order
differential operator in 𝜑 (U) without zero-order term, which is what we shall do
most of the time.
Consider an arbitrary C 1 map 𝜓 : 𝜑 (U) −→ U. The pushforward 𝜓∗ 𝑋 of the
vector field 𝑋 is a differential operator acting on a function 𝑓 ∈ C 1 (U) according
to the rule
−1
𝜓∗ 𝑋 𝑓 = (𝑋 ( 𝑓 ◦ 𝜓)) ◦ 𝜓 . (9.3.2)
−1
Now suppose 𝜓 = 𝜑 . In the local coordinates 𝑥 𝑗 of the chart (U, 𝜑) we can write
Í 𝜕𝑓
𝑋 𝑓 (℘) = 𝑛𝑗=1 𝑎 𝑗 (𝑥 (℘)) 𝜕𝑥 𝑗
(℘). In practice we shall use the same notation (9.3.1)
−1
for 𝜑 ∗ 𝑋 as for 𝑋; in other words, we think of (9.3.1) as a vector field in U, with
the understanding that this is only permitted as long as the local coordinates are not
changed.
Obviously 𝑋 defines a linear map C 1 (U) −→ C 0 (U); if we follow it up with the
valuation at an arbitrary point ℘ ∈ U we get the linear functional C 1 (U) ∋ 𝑓 −→
𝑋 𝑓 ( 𝑝) ∈ K which we shall denote here by 𝑋℘ and interpret as a tangent vector
to M at ℘. One often refers to 𝑋℘ as the freezing of the vector field 𝑋 at the point
℘. The tangent vectors at ℘ form an 𝑛-dimensional vector space (with the scalars
in K) denoted by 𝑇℘ M. This definition is coordinate-free: the choice of coordinates
𝑥1 , ..., 𝑥 𝑛 in U merely provides us with a basis of the vector space 𝑇℘ M, the partial
derivatives 𝜕𝑥𝜕 𝑗 evaluated at ℘; a change of coordinates produces a linear change of
basis in conformity with the chain-rule of differentiation, i.e., a linear automorphism
of 𝑇℘ M. It ensues Ø directly from this and from (M1)–(M2)–(M3) that the disjoint
union 𝑇M = 𝑇℘ M carries a natural regular vector bundle structure over M;
℘∈M
this is the tangent bundle of M. Of course, rank 𝑇M = 𝑛.
9.3 Tangent and Cotangent Bundles of a Manifold 273

Remark 9.3.1 A warning about the “regular” terminology is in order: when regular
means complex-analytic the tangent space at an arbitrary point ℘ ∈ M, 𝑇℘ M, must
be understood in the complex sense: its regular sections are the holomorphic vector
Í
fields; in local holomorphic coordinates 𝑧1 , ..., 𝑧 𝑛 these are of the type 𝑛𝑗=1 𝑎 𝑗 (𝑧) 𝜕𝑧𝜕 𝑗
with holomorphic coefficients 𝑎 𝑗 .

In agreement with our notation for general vector bundles we denote by

C 𝑘 (U; 𝑇M) (𝑘 ∈ Z+ or 𝑘 = +∞ or 𝑘 = 𝜔) the vector space of C 𝑘 sections of
𝑇M over the open subset U of M. An element of C 𝑘 (U; 𝑇M) is a C 𝑘 vector field
in U.
Let N be a regular submanifold of M (see Subsection 9.1.1). Recalling that the
regular functions (in the sense of the manifold structure of N ) in an arbitrary open
subset V of N are those functions in V that can be extended as regular functions
in neighborhoods in M of the points of V we see that 𝑇N can be identified with
a vector subbundle of 𝑇M| N : 𝜗 ∈ 𝑇℘ M (℘ ∈ N ) belongs to 𝑇℘ N if 𝜗 𝑓 (℘) = 0
whenever the regular function 𝑓 in a neighborhood U of ℘ in M is constant on
U ∩ N.
An orientation of U (Definition 9.1.18) induces an orientation of 𝑇℘ M, ℘ ∈ U. If
U is the domain of regular coordinates 𝑥1 , ..., 𝑥 𝑛 the orientation defined by 𝑥1 , ..., 𝑥 𝑛
induces on 𝑇℘ M the orientation defined by the basis 𝜕𝑥𝜕 1 , ..., 𝜕𝑥𝜕𝑛 . If M is orientable
so is 𝑇M.

9.3.2 Cotangent and cosphere bundle

The “dual” 𝑇 ∗ M of 𝑇M is the cotangent bundle of M: given any point ℘ ∈ M

the vector spaces 𝑇℘ M and 𝑇℘∗ M are the dual of each other. The fiber 𝑇℘∗ M is the
cotangent space to M at ℘ ∈ M and an element of 𝑇℘∗ M is a cotangent vector to M
(or, often, a covector) at ℘. We use any of the two notations ⟨𝜉, v⟩ or ⟨v, 𝜉⟩ to mean
the evaluation at v ∈ 𝑇℘ M of the linear functional 𝜉 ∈ 𝑇℘∗ M, i.e., the duality bracket
of 𝜉 and v. Here we shall denote by 𝜋 the base projection 𝑇 ∗ M ∋ (℘, 𝜉) −→ ℘ ∈ M.

Remark 9.3.2 In the complex-analytic category the cotangent space 𝑇℘∗ M is spanned
by the differentials dℎ at ℘ of holomorphic functions in a neighborhood of ℘ (cf.
Remark 9.3.1); what we denote here by 𝑇℘∗ M will be denoted by 𝑇℘(1,0) M in the
notation of the forthcoming Subsection 9.4.5.

We denote by C 𝑚 (U; 𝑇 ∗ M) (𝑚 ∈ Z+ or 𝑚 = +∞ or 𝑚 = 𝜔) the vector space of

C𝑚 sections of 𝑇 ∗ M over the open subset U of M. An element of C 𝑚 (U; 𝑇M)
is a C 𝑚 differential form (or a C 𝑚 one-form) in U. Assuming that (U, 𝑥1 , ..., 𝑥 𝑛 )
is
a coordinate patch in the regular manifold M we can make use of the frame
𝜕 𝜕 ∗
𝜕𝑥1 , ..., 𝜕𝑥𝑛 in 𝑇M| U and of the dual frame (d𝑥 1 , ..., d𝑥 𝑛 ) in 𝑇 M| U , meaning
that ⟨d𝑥 𝑗 , 𝜕𝑥𝜕𝑚 ⟩ = 𝛿 𝑗,𝑚 , the Kronecker index. This provides us with dual coordinates
(or cocoordinates) 𝜉1 , ..., 𝜉 𝑛 in each cotangent space 𝑇℘∗ M, ℘ ∈ U; if the functions 𝑥 𝑗
274 9 Elements of Differential Geometry
Í
are regular so are the 𝜉 𝑚 . Given a vector field 𝑋 = 𝑛𝑗=1 𝑎 𝑗 (𝑥) 𝜕𝑥𝜕 𝑗 with coefficients
Í
𝑎 𝑗 ∈ C 𝑚 (U) and a differential form 𝑓 (𝑥, d𝑥) = 𝑛𝑗=1 𝑏 𝑗 (𝑥) d𝑥 𝑗 with coefficients
𝑏 𝑗 ∈ C 𝑚 (U), we have the natural bilinear map
𝑛
∑︁
𝑚 𝑚 ∗
C (U; 𝑇M) × C (U; 𝑇 M) ∋ (𝑋, 𝑓 (𝑥, d𝑥)) −→ 𝑎 𝑗 𝑏 𝑗 ∈ C 𝑚 (U) . (9.3.3)
𝑗=1

The noteworthy property of (9.3.3) is that it is invariant under regular coordinate

changes, as we see by applying the chain-rule (see the proof of Proposition 9.3.6
below); it follows that U can be taken to be any open subset, even one that cannot
be coordinatized.
When K = R we shall often deal with the complexification C𝑇 ∗ M (see Subsection
9.2.2) of the cotangent bundle 𝑇 ∗ M (when K = C, 𝑇 ∗ M is already a complex
vector
Í𝑛 bundle). In other words, we shall deal with differential forms 𝑓 (𝑥, d𝑥) =
𝑗=1 𝑗 (𝑥) d𝑥 𝑗 with complex coefficients 𝑏 𝑗 .
𝑏
Definition 9.3.3 By the cosphere bundle of M we shall mean the sphere bundle
(Definition 9.2.3) of the vector bundle 𝑇 ∗ M.
Let ℘ be an arbitrary point of a regular submanifold
⊥ L of M. The vector subspace
𝑇℘ L of 𝑇℘ M admits an orthogonal 𝑇℘ L in 𝑇℘∗ M: the linear space of all linear
functionals 𝑇℘ M −→ K that vanish identically on 𝑇℘ L. As ℘ ranges over L the
⊥
linear spaces 𝑇℘ L make up a regular vector subbundle 𝑁 ∗ L of 𝑇 ∗ M| L .
Definition 9.3.4 The vector bundle 𝑁 ∗ L is called the conormal bundle of the
submanifold L.
Remark 9.3.5 When M is equipped with a regular Riemannian metric (now assum-
ing K = R), meaning that each fiber 𝑇℘ M is equipped with an inner product varying
either smoothly or analytically with ℘, then 𝑇℘ M and its dual 𝑇℘∗ M can be identified
(the finite-dimensional
⊥ Riesz Representation Theorem!). The Riesz map transforms
𝑇℘ L into the orthogonal of 𝑇℘ L for the Riemannian scalar product, namely the
subspace of 𝑇℘ M consisting of the normal vectors to L at ℘. This identifies the
conormal bundle of L with its normal bundle in M.

9.3.3 The tautological one-form

Unless specified otherwise U continues to be the domain of the local coordinates

𝑥1 , ..., 𝑥 𝑛 . The invariance of (9.3.3) can be reinterpreted as follows. Keeping in mind
that 𝑇 ∗ M is itself a manifold we can define the following differential form in the
−1 −1 Í
open set 𝜋 (U): at a point (𝑥, 𝜉) ∈ 𝜋 (U) its value is 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗 .
Proposition 9.3.6 Let 𝜉 𝑗 , 𝑗 = 1, ..., 𝑛, denote the coordinates in the cotangent
Í spaces
at points of U with respect to the frame d𝑥 1 , ..., d𝑥 𝑛 . The one-form 𝜎 = 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗
is invariant under a change of the regular coordinates 𝑥 𝑗 .
9.3 Tangent and Cotangent Bundles of a Manifold 275

It follows that the domain of definition of 𝜎 can be taken to be the whole of

𝑇 ∗ M. One often refers to 𝜎 as the tautological one-form on 𝑇 ∗ M. The reason
is evident: in the linear basis d𝑥1 , ..., d𝑥 𝑛 of the cotangent space at a point of U
Í
the covector 𝜉 is precisely equal to 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗 . (In Ch. 13 we shall encounter a
symplectic interpretation of 𝜎.)
Proof Let 𝑥 𝑗 = 𝑥 𝑗 (𝑦) be a regular coordinate change in U: 𝑦 1 , ..., 𝑦 𝑛 are new
𝜕𝑥 𝜕𝑥 𝑗
regular coordinates in U. The Jacobian matrix 𝜕𝑦 = 𝜕𝑦𝑘 is invertible
1≤ 𝑗,𝑘 ≤𝑛
𝜕𝑦
at every point of U; its inverse is the Jacobian matrix 𝜕𝑥 = 𝜕𝑦 𝜕𝑥 𝑗 1≤ 𝑗,𝑘 ≤𝑛 . Let
𝑘

Í
𝑋 = 𝑛𝑗=1 𝑎 𝑗 (𝑥) 𝜕𝑥𝜕 𝑗 be a regular vector field in U. The invariance of ⟨𝜎, 𝑋⟩ follows
immediately from the chain-rule:
𝑛 𝑛
∑︁ 𝜕 ∑︁ 𝜕𝑦 𝑘 𝜕
𝑋= 𝑎𝑗 = 𝑎𝑗 ,
𝑗=1
𝜕𝑥 𝑗 𝑗=1
𝜕𝑥 𝑗 𝜕𝑦 𝑘
𝑛 𝑛
∑︁ ∑︁ 𝜕𝑥 𝑗
𝜎= 𝜉 𝑗 d𝑥 𝑗 = 𝜉 𝑗 d𝑦 𝑘 .
𝑗=1 𝑗,𝑘=1
𝜕𝑦 𝑘

The coordinates in the fibers 𝑇℘∗ M with respect to the basis (d𝑦 1 , ..., d𝑦 𝑛 ) being
𝑛
∑︁ 𝜕𝑥 𝑗
𝜂𝑘 = 𝜉 𝑗 , 𝑘 = 1, ..., 𝑛;
𝑗=1
𝜕𝑦 𝑘

Í𝑛
we conclude that 𝜎 = 𝑗=1 𝜂 𝑗 d𝑦 𝑗 . □

Let us view the frames as vectors:

in 𝑇M| U the gradient operators ∇ 𝑥 =
𝜕 𝜕 𝜕 𝜕 ∗
𝜕𝑥1 , ..., 𝜕𝑥𝑛 , ∇ 𝑦 = 𝜕𝑦1 , ..., 𝜕𝑦𝑛 ; in 𝑇 M| U the vector-differentials d𝑥 =
(d𝑥1 , ..., d𝑥 𝑛 ), d𝑦 = (d𝑦 1 , ..., d𝑦 𝑛 ); then we can write
⊤
𝜕𝑦 𝜕𝑥
∇𝑥 = ∇ 𝑦 , d𝑥 = d𝑦, (9.3.4)
𝜕𝑥 𝜕𝑦

where the upper sign ⊤ indicates matrix transposition. This can also be stated by
saying that the change of variables 𝑥 ⇝ 𝑦 determines the change of coordinates in
the fibers of 𝑇 ∗ M given by
⊤ −1 ⊤
𝜕𝑥 𝜕𝑦
𝜉= 𝜂= 𝜂. (9.3.5)
𝜕𝑦 𝜕𝑥

Keep in mind that the matrix in the right-hand sides of (9.3.5) varies with the base
point ℘ ∈ U.
276 9 Elements of Differential Geometry

9.3.4 Differential of functions, of maps

Í
The invariance of 𝜎 = 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗 means that it is possible to define the differential,
or exterior derivative, of a function 𝑓 ∈ C 1 (M) as a section of 𝑇 ∗ M over M: in
Í 𝜕𝑓
an arbitrary coordinate patch (U, 𝑥1 , ..., 𝑥 𝑛 ) it is the one-form d 𝑓 = 𝑛𝑗=1 𝜕𝑥 𝑗
d𝑥 𝑗 in
U. The differential defines a linear map

d : C 𝑘+1 (M) −→ C 𝑘 (M; 𝑇 ∗ M) , (9.3.6)

where 0 ≤ 𝑘 ≤ +∞ or, when M is a C 𝜔 manifold, 𝑘 = 𝑘 + 1 = 𝜔. When K = C and

we use the notation 𝑧 𝑗 = 𝑥 𝑗 + 𝑖𝑦 𝑗 for the holomorphic local coordinates in U, we
have d𝑧 𝑗 = d𝑥 𝑗 + 𝑖d𝑦 𝑗 and the holomorphic (or complex) differential

𝜕 : O (U) ∋ ℎ −→ 𝜕𝑧 ℎ = 𝜕𝑧1 ℎ, ..., 𝜕𝑧𝑛 ℎ ∈ O (U; 𝑇 ∗ M)

(9.3.7)

𝜕ℎ 1 𝜕 𝜕
where 𝜕𝑧 𝑗 ℎ = 𝜕𝑧 𝑗
= 2 𝜕𝑥 𝑗
− 𝑖 𝜕𝑦 𝑗
. In all cases d is a differential operator, in the
sense that supp d 𝑓 ⊂ supp 𝑓 .
It is traditional to interpret the d𝑥 𝑗 as infinitesimal displacements and the freezing
Í 𝜕𝑓
of d 𝑓 at a point ℘ ∈ M, d 𝑓 (℘) = 𝑛𝑗=1 𝜕𝑥 𝑗
(𝑥 (℘)) d𝑥 𝑗 , as an infinitesimal variation.
Selecting a tangent vector 𝑋℘ ∈ 𝑇℘ M leads to the interpretation of the duality bracket
⟨d 𝑓 (℘) , 𝑋℘ ⟩ as the infinitesimal variation in the direction of 𝑋℘ (note however that
implicit to this is not only a direction but also an “intensity”, i.e., some “length” of
the tangent vector 𝑋℘ ). If we let ℘ vary and consider a vector field 𝑋 in M, then

∀ 𝑓 ∈ C 1 (M) , ⟨d 𝑓 , 𝑋⟩ = 𝑋 𝑓 . (9.3.8)

The differential satisfies the Leibniz rule: if 𝑓 , 𝑔 ∈ C 1 (M),

d ( 𝑓 𝑔) = 𝑓 d𝑔 + 𝑔d 𝑓 . (9.3.9)

The differential of a regular map 𝐹 : M −→ N between two manifolds (with,

possibly, dimK M = 𝑛 ≠ dimK N = 𝑚) defines a regular vector bundle map 𝐹∗ :
𝑇M −→ 𝑇N , mapping 𝑇℘ M into 𝑇𝐹 (℘) N whatever ℘ ∈ M. To see this it is
convenient to make use of a coordinate patch (U, 𝑥1 , ..., 𝑥 𝑛 ) in M and of one in
N , (V, 𝑦 1 , ..., 𝑦 𝑚 ), such that 𝐹 (U) ⊂ V. We can write 𝐹 = (𝐹1 , ..., 𝐹𝑚 ) with
Í𝑛 𝜕
𝑦 𝑗 (𝐹 (℘)) = 𝐹 𝑗 (𝑥 (℘)) for ℘ ∈ U. If 𝑋 = 𝑖=1 𝑎 𝑖 (𝑥) 𝜕𝑥 𝑖
is a smooth vector field
in U we can form the tangent vector to N at 𝐹 (℘),
𝑚
∑︁ 𝜕
𝐹∗ 𝑋℘ = 𝑋 𝐹 𝑗 (℘) . (9.3.10)
𝑗=1
𝜕𝑦 𝑗

Obviously 𝐹∗ 𝑋℘ depends solely on the value 𝑋℘ of 𝑋 at ℘. In order to

see that the
definition of 𝐹∗ 𝑋℘ is coordinate-free we observe that in the frames 𝜕𝑥𝜕 1 , ..., 𝜕
𝜕𝑥𝑛

and 𝜕𝑦𝜕 1 , ..., 𝜕𝑦𝜕𝑚 the transformation 𝐹∗ is expressed by the equations
9.3 Tangent and Cotangent Bundles of a Manifold 277
𝑚
∑︁ 𝜕𝐹 𝑗
𝑏𝑗 = 𝑎 𝑖 , 𝑗 = 1, ..., 𝑚,
𝑗=1
𝜕𝑥𝑖

which transform precisely as prescribed by the changes of frames if we carry out a

change of coordinates 𝑥 𝑖 and 𝑦 𝑗 . These same equations show directly that the bundle
map 𝐹∗ is regular; 𝐹∗ 𝑋℘ is the pushforward of the tangent vector 𝑋℘ under the map
𝐹.
By the rank of the map 𝐹 at the point ℘ we mean the rank of the linear map
𝑇 ℘ M ∋𝑋℘ −→ 𝐹∗ 𝑋℘ . In the above coordinate patches it is the rank of the matrix
𝜕𝐹 𝑗
𝜕𝑥𝑖 𝑖=1,...,𝑛 .
𝑗=1,...,𝑚

9.3.5 Immersions and submersions

The next definition introduces a subtle but important concept. Let N be a regular
manifold countable at infinity (like M).
Definition 9.3.7 A regular map 𝐹 : N −→ M is called an immersion if 𝐹∗ is a linear
injection of 𝑇℘ N into 𝑇𝐹 (℘) M for every ℘ ∈ N . By an immersed submanifold of
M we shall mean the image of an immersion.
By the dimension of the immersed submanifold 𝐹 (N ) we mean dim N . If L is
a regular submanifold of M as per Definition 9.1.7 the natural injection L ↩→ M is
an immersion. More generally we have
Proposition 9.3.8 If every point of a subset L of M has a neighborhood in L which
is a regular submanifold of M then L is an immersed submanifold.
Proof Let U 𝑗 , 𝑗 = 1, 2, be open subsets of L that are regular submanifolds of M
and such that U1 ∩ U2 ≠ ∅; then U1 ∩ U2 is also a regular submanifold of M,
implying that dim U1 = dim U2 (= ℓ = dim L). After contracting U1 and U2 (while
keeping U1 ∩ U2 ≠ ∅) we can assume that, for each 𝑗 = 1, 2, there are regular
isomorphisms 𝑓 𝑗 (i.e., regular bijections whose inverse is regular) of the open unit
−1
ball 𝔅 (ℓ) in Kℓ onto U 𝑗 . Then 𝑓2 ◦ 𝑓1 is a regular isomorphism of 𝔅 (ℓ) onto itself.
This proves that L is a regular manifold and that the natural injection L ↩→ M is
an immersion. □
A modification of the proof of Proposition 9.3.8 shows that, for a subset L of M,
to be an immersed submanifold of M is a local property: If every point of a subset
L of M has a neighborhood in L which is an immersed submanifold then L is an
immersed submanifold.
Remark 9.3.9 In our terminology an immersion 𝐹 : N −→ M need not be injective,
even if N is connected. Example: take N to be the helix in M = R3 , defined by the
equation 𝑥1 = cos 𝑥3 , 𝑥2 = sin 𝑥3 , and 𝐹 to be the projection 𝑥 ↦→ (𝑥1 , 𝑥2 ).
278 9 Elements of Differential Geometry

The condition in Proposition 9.3.8 is not necessary: it excludes a number of curves

that self-intersect at isolated points, as in the following

Example 9.3.10 Let N be the analytic curve in R3 defined by the equations 𝑥 1 =

sin 2𝑥3 cos 𝑥3 , 𝑥2 = sin 2𝑥3 sin 𝑥 3 . The projection (𝑥 1 , 𝑥2 , 𝑥3 ) −→ (𝑥1 , 𝑥2 ) induces an
immersion of N into R2 since
𝜕 𝜕 𝜕
+ (2 cos 2𝑥 3 cos 𝑥 3 − sin 2𝑥3 sin 𝑥3 ) + (2 cos 2𝑥 3 sin 𝑥3 + sin 2𝑥3 cos 𝑥 3 )
𝜕𝑥 3 𝜕𝑥1 𝜕𝑥2
is tangent to N at every point, and we cannot have

2 cos 2𝑥 3 cos 𝑥3 − sin 2𝑥3 sin 𝑥3 = 0,

2 cos 2𝑥 3 sin 𝑥3 + sin 2𝑥3 cos 𝑥3 = 0.

In polar coordinates 𝐹 (N ) is defined by 𝑟 = sin 2𝜃 where 𝑟 is allowed to take

negative values; 𝐹 (N ) is the classical four-clover curve or quadrifolium.

Let 𝐹 : N −→ M be an immersion. The pushforward map 𝐹∗ is an immersion

of the tangent bundle 𝑇N into 𝑇M| 𝐹 ( N) ; if 𝐹 is injective then the range of 𝐹∗ is
an immersed regular vector subbundle of 𝑇M| 𝐹 ( N) and the orthogonal of 𝐹∗𝑇N in
𝑇 ∗ M| 𝐹 ( N) is the conormal bundle of N in M (cf. Subsection 9.3.2).

Proposition 9.3.11 For a regular map 𝐹 : N −→ M to be an immersion it is

necessary and sufficient that every point ℘ ∈ N have a neighborhood V such that
𝐹 | V is a regular isomorphism of V onto a regular submanifold of M (cf. Definition
9.1.7).

Proof Use local frames of 𝑇M and 𝑇N and apply the Implicit Function Theorem.□
If dim M = dim N an immersion 𝐹 : N −→ M is a local regular isomorphism:
every point ℘ ∈ N has a neighborhood V such that 𝐹 | V is a regular isomorphism
of V onto an open subset of M.

Corollary 9.3.12 If 𝐹 : N −→ M is a local regular isomorphism then the preimage

−1
𝐹 (℘) of a point ℘ ∈ 𝐹 (N ) is a discrete subset of N .

Proposition 9.3.13 Let 𝐹 : N −→ M be an immersion. An arbitrary point ℘◦ ∈

𝐹 (N ) is contained in an open subset U of M such that the connected component
of U ∩ 𝐹 (N ) containing ℘◦ is a set

℘ ∈ U; 𝜑 𝑗 (℘) = 0, 𝑗 = 1, ..., 𝜈

where the 𝜑 𝑗 are regular functions in U.

−1
Proof Let 𝑛 ′ = dimK N and℘∗ ∈ 𝐹 (℘◦ ) ⊂ N be arbitrary. We may assume that ℘◦
and ℘∗ are the centers of (regular) local charts (U, 𝑥1 , ..., 𝑥 𝑛 ) and (V, 𝑦 1 , ..., 𝑦 𝑛′ )
respectively, and transfer the analysis to a neighborhood of the origin, Ω [resp.,
9.3 Tangent and Cotangent Bundles of a Manifold 279
′
Θ], in R𝑛 (resp., R𝑛 ). After suitable contractions of U and V about ℘◦ and ℘∗
respectively, we may equate 𝐹 | V to a regular immersion 𝑓 : Θ −→ Ω expressed, in
Θ, by equations 𝑥 𝑗 = 𝑓 𝑗 (𝑦), 𝑗 = 1, ..., 𝑛, with 𝑦 = (𝑦 1 , ..., 𝑦 𝑛′ ). We can assume that
𝜕( 𝑓1 ,..., 𝑓 𝑛 ) ′
the rank of the Jacobian matrix 𝜕( 𝑦1 ,...,𝑦𝑛′ ) is equal to 𝑛 (≤ 𝑛, necessarily) at every
point of Θ. A linear change
of the coordinates 𝑥𝑖 in Ω (possibly contracted about 0)
𝜕 𝑓𝑖
can ensure that det 𝜕𝑦 𝑗 ′
does not vanish in Θ. Let us write 𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛′ ),
1≤𝑖, 𝑗 ≤𝑛
𝑥 ′′ = (𝑥 𝑛′ +1 , ..., 𝑥 𝑛 ) and select Ω = Ω′ × Ω′′ where Ω′ (resp., Ω′′) is a connected
′ ′
neighborhood of the origin in 𝑥 ′-space R𝑛 (resp., 𝑥 ′′-space R𝑛−𝑛 ). The Implicit
Function Theorem allows us to solve the equations 𝑥𝑖 = 𝑓𝑖 (𝑦), 𝑖 = 1, ..., 𝑛 ′, with
respect to the 𝑦 𝑗 : there are regular functions 𝑔 𝑗 in Ω′ (possibly further contracted
about 0) such that 𝑦 𝑗 = 𝑔 𝑗 (𝑥 ′) and 𝑔 𝑗 (0) = 0, 𝑗 = 1, ..., 𝑛 ′. We select Ω′ such
that ( 𝑓𝑛′ +1 (𝑔 (𝑥 ′)) , ..., 𝑓𝑛 (𝑔 (𝑥 ′))) ∈ Ω′′ whatever 𝑥 ′ ∈ Ω′. The equations 𝑥𝑖 −
𝑓𝑖 (𝑔 (𝑥 ′)) = 0, 𝑖 = 𝑛 ′ + 1, ..., 𝑛 − 𝑛 ′, define the connected component of the image
of 𝑓 in Ω that contains the origin. □

Remark 9.3.14 The open set U in Proposition 9.3.13 might contain several distinct
connected components of U ∩ 𝐹 (N ); these might even be dense in U as in the
archetypical example of an everywhere dense geodesic of the torus [see Subsection
12.2.2].

In the (real- or complex-)analytic category Proposition 9.3.13 has the following

useful consequence.

Proposition 9.3.15 Let L1 and L2 be two immersed analytic submanifolds of an

analytic manifold M. If Λ is an open subset of L1 contained in L2 and if ℘◦ is
a boundary point of Λ with respect to L1 then either ℘◦ ∉ L2 or else there is a
neighborhood of ℘◦ in L1 contained in L2 .

Proof Suppose ℘◦ ∈ L2 . By Proposition 9.3.13 there is a neighborhood U of ℘◦ in

M such that the connected component of U ∩ L2 containing ℘◦ is a set

℘ ∈ U; 𝜑 𝑗 (℘) = 0, 𝑗 = 1, ..., 𝜈

with the 𝜑 𝑗 analytic in U. We have 𝜑 𝑗 ≡ 0 in Λ ∩ U ; Λ ∩ U is an open subset of

L1 whose closure contains ℘◦ ; it follows that 𝜑 𝑗 ≡ 0 in the connected component of
L1 ∩ U that contains ℘◦ . □
It is instructive to compare the next statement to Example 9.1.9.

Corollary 9.3.16 Let L1 and L2 be two immersed analytic submanifolds of an

analytic manifold M. Suppose that L1 is connected and that the two following
properties hold:
(1) there is an open subset Λ◦ of L1 contained in L2 ;
(2) if an open subset Λ of L1 is contained in L2 then Λ ⊂ L2 .
Under these hypotheses L1 ⊂ L2 .
280 9 Elements of Differential Geometry

Proof Property (1) implies that L1 ∩ L2 contains an open subset of L1 ; if L∗ is

a maximal open subset of L1 contained in L2 Property (2) implies that its closure
in L1 is contained in L2 and therefore must be open in L1 by Proposition 9.3.15,
implying L∗ = L1 . □

Corollary 9.3.17 Let L1 and L2 be two immersed analytic submanifolds of an

analytic manifold M. Suppose that L1 ∩ L2 is an open subset of both L1 and L2
(thus dim L1 = dim L2 ). A boundary point of L1 ∩ L2 cannot belong to both L1
and L2 .

In Corollary 9.3.17 by the boundary of a set 𝐸 we mean 𝐸\𝐸.

A terminology also used is the following.

Definition 9.3.18 A regular map 𝐹 : N −→ M is called a submersion if 𝐹∗ is a

linear surjection of 𝑇℘ N onto 𝑇𝐹 (℘) M whatever ℘ ∈ N .

The next statement is self-evident.

Proposition 9.3.19 For a regular map 𝐹 : N −→ M to be an immersion (resp., a

submersion) it is necessary and sufficient that 𝐹 have constant rank equal to dim N
(resp., to dim M ≤ dim N ).

If dim N = dim M a submersion 𝐹 : N −→ M is also an immersion and a local

regular isomorphism.

9.4 Differential Complexes and Grassman Algebras

9.4.1 Differential operators between vector bundles and their symbols.

Differential complexes

First consider a scalar differential operator 𝑃 of order 𝑚, with complex-valued C ∞

coefficients, in a C ∞ manifold M; “scalar” means that 𝑃 acts on C ∞ functions and
transforms them into C ∞ functions. In a local chart (U, 𝑥1 , ..., 𝑥 𝑛 ) we may represent
𝑃 as a differential operator
∑︁
𝑃 (𝑥, D) = 𝑐 𝛼 (𝑥) D 𝛼 .
| 𝛼 | ≤𝑚

In the local chart (U, 𝑥1 , ..., 𝑥 𝑛 ) its symbol is 𝑃 (𝑥, 𝜉) , its principal
symbol 𝑃𝑚 (𝑥, 𝜉)
⊤
𝜕𝑦
[see (2.1.8)]. Suppose we change coordinates: 𝑥 = 𝑥 (𝑦), 𝜉 = 𝜕𝑥 𝜂 [see (9.3.5)].
The chain rule implies
that the principal
symbol of 𝑃 in the new coordinates is
⊤
𝜕𝑦
transformed into 𝑃𝑚 𝑥 (𝑦) , 𝜕𝑥 𝜂 , which shows that 𝑃𝑚 can be regarded as a
smooth function in the whole of the cotangent bundle 𝑇 ∗ Ω. In other words, 𝑃𝑚 (𝑥, 𝜉)
9.4 Differential Complexes and Grassman Algebras 281

is coordinate free if the only permissible changes of coordinates in 𝑇 ∗ Ω are those

that leave the tautological one-form 𝜎 unchanged. This is also true, trivially, of the
zero-order term in 𝑃, equal to the action of 𝑃 on the function identically equal to 1
in M, 𝑃1. In general, however, this is not true of other parts of the symbol of 𝑃.
d2
Example 9.4.1 Let M = K, 𝑃 = .
Let 𝑥 ⇝ 𝑦 be a regular change of variable
d𝑥 2
2 2
d𝑦 d d2 𝑦 d
in an open subset of K. In the coordinate 𝑦 we have 𝑃 = d𝑥 d𝑦 2
+ d𝑥 2 d𝑦 . The

homogeneous part of degree 1 of 𝑃 is modified.

Now let B 𝑗 ( 𝑗 = 1, 2) be two regular vector bundles over the same manifold M
and let CB 𝑗 denote their complexifications (of course, CB 𝑗 = B 𝑗 when K = C). We
introduce [cf. (9.2.3)] differential operators acting from sections over an open subset
U of M of CB1 into those of CB2 :

𝑃 : C ∞ (U; CB1 ) −→ C ∞ (U; CB2 ) . (9.4.1)

Suppose U is the domain of C ∞ local coordinates 𝑥1 , ..., 𝑥 𝑛 and let 𝜉1 , ..., 𝜉 𝑛

be the dual coordinates in the cotangent spaces to M at points of U; we
can write 𝑃 𝑗,𝑘 = 𝑃 𝑗,𝑘 (𝑥, D 𝑥 ). This assigns to 𝑃 a symbol in U, the matrix
𝑃 (𝑥, 𝜉) = 𝑃 𝑗,𝑘 (𝑥, 𝜉) 𝑗=1,...,𝑟1 . In general the definition of an invariant princi-
𝑘=1,...,𝑟2
pal symbol for such a matrix raises delicate questions. This definition, however, is
clear-cut if we assume that the order of every entry 𝑃 𝑗,𝑘 is at most equal to the
same number 𝑚 independently of ( 𝑗, 𝑘) and that it is equal to 𝑚 for some ( 𝑗, 𝑘)
at some points of M. In this case the principal symbol of 𝑃 shall be the matrix
𝑃𝑚 (𝑥, 𝜉) = 𝑃 𝑗,𝑘,𝑚 (𝑥, 𝜉) 𝑗=1,...,𝑟1 , where 𝑃 𝑗,𝑘,𝑚 (𝑥, 𝜉) is the homogeneous part of
𝑘=1,...,𝑟2
degree 𝑚 of 𝑃 𝑗,𝑘 (𝑥, 𝜉), a function globally defined in C𝑇 ∗ M (possibly vanishing).
In order to give an invariant meaning to 𝑃𝑚 (𝑥, 𝜉) one may introduce the vector
bundle 𝑳 (CB1 ; CB2 ) over M: the fiber of this bundle at ℘ ∈ M is the vector
space 𝑳 (CB1 ) ℘ ; (CB2 ) ℘ of linear maps (CB1 ) ℘ −→ (CB2 ) ℘ . We can also in-
troduce the pullback of 𝑳 (CB1 ; CB2 ) to 𝑇 ∗ M under the base projection 𝜋, which
we denote by 𝜋 ∗ 𝑳 (CB ∗ ∗
1 ; CB2 ): the fiber of 𝜋 𝑳 (CB1 ; CB2 ) at (℘, 𝜉) ∈ 𝑇 M is
𝑳 (CB1 ) ℘ ; (CB2 ) ℘ independent of 𝜉. We can regard 𝑃𝑚 (𝑥, 𝜉) as a smooth sec-
tion of 𝜋 ∗ 𝑳 (CB1 ; CB2 ), regarding (𝑥, 𝜉) as the variable point in 𝑇 ∗ M and 𝜉 as the
variable point in 𝑇𝑥∗ M. Clearly the section 𝑃𝑚 (𝑥, 𝜉) is homogeneous of degree 𝑚
with respect to 𝜉 and therefore vanishes for 𝜉 = 0 if 𝑚 ≥ 1. This interpretation incor-
porates well the effect of changes of local frames in the bundles B 𝑗 : such changes
translate into conjugations of the matrix representing 𝑃𝑚 (𝑥, 𝜉). It agrees with the
initial definition of the principal symbol in the scalar case, when B 𝑗 = M × K,
𝑗 = 1, 2.
An important topic in geometry and analysis is that of differential complexes. The
following definition is not the most general but will amply suffice for our needs in
the remainder of the book.
282 9 Elements of Differential Geometry

Definition 9.4.2 A differential complex of order 𝑚 over the manifold M, associated

to a sequence of vector bundles B ( 𝑗) ( 𝑗 ∈ Z+ ) over M, shall be a sequence of
differential operators of order 𝑚, 𝑃 ( 𝑗) : C ∞ M; CB ( 𝑗−1) −→ C ∞ M; CB ( 𝑗)
( 𝑗 = 1, 2, ...) such that 𝑃 ( 𝑗+1) ◦ 𝑃 ( 𝑗) = 0 for every 𝑗.

The property in Definition 9.4.2 is often expressed by saying that the image
(i.e., the range) of 𝑃 ( 𝑗) is contained in the kernel (i.e., the null-space) of 𝑃 ( 𝑗+1) :
Im 𝑃 ( 𝑗) ⊂ ker 𝑃 ( 𝑗+1) . In practice, our differential complex will end at a finite stage,
𝑁 ≥ 1, with the null operator:

𝜄
𝑃 (1)
0 −→ ker 𝑃 (1) −→ C ∞ M; CB (0) −→ · · · (9.4.2)
𝑃 ( 𝑗)
· · · −→ C ∞ M; CB ( 𝑗−1) −→ C ∞ M; CB ( 𝑗+1) −→ · · ·
𝑃(𝑁)
· · · −→ C ∞ M; CB ( 𝑁 −1) −→ C ∞ M; CB ( 𝑁 ) −→ 0
𝑚

where the first map, 𝜄, is the natural injection.

Remark 9.4.3 In the applications we may need variants or generalizations of (9.4.2).

For instance, we may need to replace C ∞ sections by generalized classes of sections,
such as distribution sections of the vector bundles CB ( 𝑗) , or hyperfunction sections
when the base manifold M and the coefficients of the differential operators 𝑃 ( 𝑗) are
analytic. Such “generalized sections” will be defined in later chapters.
What can already be done is to replace M by arbitrarily small neighborhoods of
a point ℘ ∈ M and define the analogue of (9.4.2) for the space of germs at ℘ of C ∞
sections of the CB ( 𝑗) , and from there for the corresponding sheaves. The same will
be doable when C ∞ is replaced by more general classes.

A sequence of linear maps between vector spaces

𝜑 ( 𝑗) 𝜑 ( 𝑗+1)
· · · −→ E ( 𝑗) −→ E ( 𝑗+1) −→ E ( 𝑗+2) −→ · · ·

is said to be exact if Im 𝜑 ( 𝑗) = ker 𝜑 ( 𝑗+1) for every 𝑗. In accordance with this we

say that the differential complex (9.4.2) is exact if Im 𝑃 ( 𝑗) = ker 𝑃 ( 𝑗+1) for every
𝑗 = 1, 2, .... A less rare occurrence is for (9.4.2) to be locally exact, meaning that
every point of M has a neighborhood such that (9.4.2) is exact if we replace M by
said neighborhood. Often we may be interested in whether the short sequence
𝑃 ( 𝑗) 𝑃 ( 𝑗+1)
C ∞ M; CB ( 𝑗−1) −→ C ∞ M; CB ( 𝑗+1) −→ C ∞ M; CB ( 𝑗+1)

is exact for a given value of 𝑗 ≥ 1. If this is true we may say that (9.4.2) is exact in
degree 𝑗.
9.4 Differential Complexes and Grassman Algebras 283

Proposition 9.4.4 Let B ( 𝑗) , 𝑃 ( 𝑗) be a differential complex of order 𝑚 over
𝑗=1,2,...
( 𝑗)
M. If for each 𝑗 = 1, 2, ..., 𝑃𝑚 (𝑥, 𝜉) is the principal symbol of 𝑃 ( 𝑗) then 𝑇 ∗M ∋
( 𝑗+1) ( 𝑗)
(𝑥, 𝜉) ↦→ 𝑃𝑚 (𝑥, 𝜉) ◦ 𝑃𝑚 (𝑥, 𝜉) is the zero section of 𝜋 ∗ 𝑳 CB ( 𝑗) ; CB ( 𝑗+2) .

Proof It suffices to prove the statement in frames over the domains of local coordi-
nates in M. In other words, it suffices to prove the statement for systems of linear
partial differential operators in an open subset of Euclidean space K𝑛 . For this we
let the systems act on vector-valued functions v exp (𝑖𝑥 · 𝜉), with v ∈ K𝑛 arbitrary.
We leave the details as an exercise. □
Definition 9.4.5 The differential complex (9.4.2) of order 𝑚 over M is said to
( 𝑗) ( 𝑗) ( 𝑗+1)
be elliptic if the principal symbols 𝑃𝑚 (𝑥, 𝜉) ∈ 𝑳 CB 𝑥 ; CB 𝑥 satisfy the
(1) ( 𝑗) ( 𝑗+1)
conditions ker 𝑃𝑚 (𝑥, 𝜉) = 0, Im 𝑃𝑚 (𝑥, 𝜉) = ker 𝑃𝑚 (𝑥, 𝜉) for every 𝑗 = 1, 2, ...,
i.e., the following sequence is exact:
(1) (2)
𝑃𝑚 ( 𝑥, 𝜉 ) 𝑃𝑚 ( 𝑥, 𝜉 )
0 −→ CB 𝑥(0) −→ CB 𝑥(1) −→ ··· (9.4.3)
( 𝑗)
( 𝑗−1) 𝑃𝑚 ( 𝑥, 𝜉 ) ( 𝑗)
· · · −→ CB 𝑥 −→ CB 𝑥 −→ · · ·
𝑃 ( 𝑁 ) ( 𝑥, 𝜉 )
· · · −→ CB 𝑥( 𝑁 −1) −→ CB 𝑥( 𝑁 ) −→ 0.

Thus, when the differential complex (9.4.2) is elliptic the first linear map in the
(1)
principal symbols sequence (9.4.3), 𝑃𝑚 (𝑥, 𝜉), must be injective and the last one,
( 𝑁 −1)
𝑃𝑚 (𝑥, 𝜉), must be surjective. In the case of the “differential complex” consisting
of a single operator (9.4.1) ellipticity means that, whatever (𝑥, 𝜉) ∈ 𝑇 ∗ M\0, the
principal symbol sequence
𝑃𝑚 ( 𝑥, 𝜉 )
0 −→ CB 𝑥(1) −→ CB 𝑥(2) −→ 0 (9.4.4)

is exact, which is equivalent to saying that 𝑃𝑚 (𝑥, 𝜉) is a linear bijection. In particular

we must have rank B (1) = rank B (2) . In this case the operator (9.4.1) is most often
referred to as an elliptic system of differential operators, sometimes also as a square
elliptic system, to emphasize the equality of the ranks of the vector bundles B (1)
and B (2) .

9.4.2 Exterior algebra of a vector space, of a vector bundle

Let E be a vector space over K with dimK E = 𝑁 < +∞. We recall the definition of
the exterior algebra of E. For 𝑝 ≥ 1 we denote by E⊗ 𝑝 the 𝑝 th tensor power of E,
𝑝 factors
z }| {
E⊗ 𝑝 = E ⊗ · · · ⊗ E. (9.4.5)
284 9 Elements of Differential Geometry

If 𝑝 ≥ 2 we denote by N ( 𝑝) the linear subspace of E⊗ 𝑝 spanned by the tensor

products 𝜃 1 ⊗ x ⊗ x ⊗ 𝜃 2 where x ∈ E and 𝜃 𝑗 ∈ E⊗ 𝑝 𝑗 , 𝑝 1 + 𝑝 2 = 𝑝 − 2. For instance,
in E⊗2 the two-tensor

x ⊗ y + y ⊗ x = (x + y) ⊗ (x + y) −x ⊗ x − y ⊗ y

belongs to N (2) . We denote by Λ 𝑝 E the quotient vector space E⊗ 𝑝 /N ( 𝑝) ; we also

define Λ0 E = K and Λ1 E = E.
We denote by x1 ∧ · · · ∧ x 𝑝 the coset in the quotient E⊗ 𝑝 /N ( 𝑝) represented by
x1 ⊗ · · · ⊗ x 𝑝 ∈ E⊗ 𝑝 ; x1 ∧ · · · ∧ x 𝑝 is often called a 𝑝-vector. We have x1 ∧ · · · ∧ x 𝑝 =
±x𝑖1 ∧ · · · ∧ x𝑖 𝑝 if 𝑖 1 , ..., 𝑖 𝑝 is a permutation of (1, ..., 𝑝), with the sign + or −
depending on whether the permutation is even or odd. By decomposing elements 𝜃 ∈
Λ 𝑝 E and 𝜔 ∈ Λ𝑞 E into sums of 𝑝-vectors x1 ∧ · · · ∧ x 𝑝 and 𝑞-vectors y1 ∧ · · · ∧ y𝑞
we can define their exterior (or wedge) product 𝜃 ∧ 𝜔 = (−1) 𝑝𝑞 𝜔 ∧ 𝜃 ∈ Λ 𝑝+𝑞 E;
obviously (𝜃, 𝜔) −→ 𝜃 ∧ 𝜔 is a bilinear map. When 𝑝 = 0, 𝜃 ∧ 𝜔 is simply the
product of the 𝑞-vector 𝜔 by the scalar 𝜃. It is checked directly that the exterior
product is associative.
Select at random a linear basis of E, e1 , ..., e 𝑁 . It is readily seen that a basis of
Λ 𝑝 E consists of the exterior products e𝐼 = e𝑖1 ∧ · · · ∧ e𝑖 𝑝 where the multi-index
𝐼 = 𝑖 1 , ..., 𝑖 𝑝 is ordered, in the sense that 1 ≤ 𝑖 1 < · · · < 𝑖 𝑝 ≤ 𝑁. It follows directly

that Λ 𝑝 E = {0} if 𝑝 > 𝑁 and that dim Λ 𝑝 E = 𝑁𝑝 if 2 ≤ 𝑝 ≤ 𝑁.
Ê𝑁
This said, the exterior algebra of E is the direct sum ΛE = Λ 𝑝 E; ΛE is an
𝑝=0
algebra with respect to vector addition and the exterior multiplication.
Let us now replace the vector space E by its dual E∗ . The following interpretation
of the exterior power Λ 𝑝 E∗ is common: an element 𝜃 ∗ ∈ E∗ defines an alternating
𝑝 factors
z }| {
𝑝-linear functional on E 𝑝 = E × · · · × E. Indeed, let 𝜃 ∗ = x∗1 ∧ · · · ∧ x∗𝑝 with x∗𝑖 ∈ E∗ .

If x1 , ..., x 𝑝 ∈ E 𝑝 we can define
∑︁
𝜃 ∗ x1 , ..., x 𝑝 = 𝜀 (𝐼) ⟨x∗1 , x𝑖1 ⟩ · · · ⟨x∗𝑝 , x𝑖 𝑝 ⟩,

(9.4.6)
𝐼= ( 𝑖1 ,...,𝑖 𝑝 )

where the sum is taken over all permutations 𝐼 = 𝑖1 , ..., 𝑖 𝑝 of the set of integers
(1, ..., 𝑝) and 𝜀 (𝐼) = +1 if the permutation is even, −1 if it is odd. Take for instance
𝑝 = 2; (9.4.6) reads

x∗1 ∧x∗2 (x1 , x2 ) = ⟨x∗1 , x1 ⟩⟨x∗2 , x2 ⟩ − ⟨x∗1 , x2 ⟩⟨x∗2 , x1 ⟩. (9.4.7)

9.4 Differential Complexes and Grassman Algebras 285

If B is a regular vector bundle of rank 𝑁 over a regular manifold M we can define

its 𝑝 th exterior power Λ 𝑝 B fiberwise: the fiber of Λ 𝑝 B at a point ℘ ∈ M is simply
the 𝑝 th exterior power Λ 𝑝 B℘ . Using local trivializations of B it is readily checked
that Λ 𝑝 B is a regular vector bundle over M and therefore, that the same is true of
Ê 𝑁
the Whitney sum ΛB = Λ 𝑝 B.
𝑝=0

9.4.3 The Grassman algebra of a manifold

For us the Grassman algebra of a manifold M shall be the regular vector bundle
𝑛
Ê
ΛC𝑇 ∗ M = Λ𝑞 C𝑇 ∗ M, 𝑛 = dimK M. (9.4.8)
𝑞=0

We remind the reader that Λ0 C𝑇 ∗ M is the trivial bundle M × C and that Λ1 C𝑇 ∗ M =

C𝑇 ∗ M, the complexification of the cotangent bundle 𝑇 ∗ M. The rank (over C) of the
complex vector bundle Λ C𝑇 M is equal to 𝑞 and that of ΛC𝑇 ∗ M is equal to 2𝑛 .
𝑞 ∗ 𝑛

For each ℘ ∈ M the elements of Λ𝑞 C𝑇℘∗ M can be regarded as alternating 𝑞-linear

functionals C𝑇℘ M −→ C.
Let (U, 𝑥1 , ..., 𝑥 𝑛 ) be a regular local chart in M; the differentials d𝑥 1 , ..., d𝑥 𝑛
form a frame (with the complex numbers as scalars) of C𝑇 ∗ M over U. It ensues that
the exterior products d𝑥 𝐼 = d𝑥𝑖1 ∧ · · · ∧ d𝑥 𝑖𝑞 , where 𝐼 = 𝑖 1 , ..., 𝑖 𝑞 ranges over the set
of all multi-indices such that 1 ≤ 𝑖1 < · · · < 𝑖 𝑞 ≤ 𝑛, form a frame of Λ𝑞 C𝑇 ∗ M over
U. We shall refer to 𝑞 as the length of the multi-index 𝐼 and write |𝐼 | = 𝑞. A section
of Λ𝑞 C𝑇 ∗ M (frequently called a differential 𝑞-form or simply a 𝑞-form) over U can
be represented as a linear combination
∑︁
𝑓 (𝑥, d𝑥) = 𝑓 𝐼 (𝑥) d𝑥 𝐼 . (9.4.9)
|𝐼 |=𝑞

At each point ℘ ∈ U (9.4.9) defines a linear form on Λ𝑞 𝑇℘ M or, equivalently, a

skew-symmetric (or alternating) 𝑞-linear form on C𝑇℘ M. Here “skew-symmetric”
means that the form is invariant up to sign under permutations and the change of
sign occurs when the permutation is odd. Let
𝑛
∑︁ 𝜕
𝑋𝑗 = 𝑎 𝑗,𝑘 (𝑥) , 𝑗 = 1, ..., 𝑞,
𝑘=1
𝜕𝑥 𝑘

be vector fields, defined and regular in U (i.e., the coefficients 𝑎 𝑗,𝑘 are regular in
U). The 𝑞-linear form defined by 𝑓 is
286 9 Elements of Differential Geometry
∑︁
𝑓 , 𝑋1 , ..., 𝑋𝑞 = 𝑓 𝐼 (𝑥) d𝑥 𝐼 , 𝑋1 , ..., 𝑋𝑞 (9.4.10)
|𝐼 |=𝑞
∑︁ ∑︁
= 𝜀 (𝐼, 𝐾) 𝑓 𝐼 (𝑥) 𝑎 1,𝑘1 (𝑥) · · · 𝑎 𝑞,𝑘𝑞 (𝑥) ,
|𝐼 |=𝑞 𝐾= ( 𝑘1 ,...,𝑘𝑞 )

where 𝜀 (𝐼, 𝐾) = 0 unless 𝐾 is a permutation of 𝐼 and, in this case, 𝜀 (𝐼, 𝐾) = ±1

depending on whether the permutation is even or odd.
Let 𝜑 : N −→ M be a regular map between regular manifolds. In (9.3.2) we
have defined the pushforward map 𝜑∗ of tangent vectors and of vector fields. By
transposition we may define the pullback 𝜑∗ 𝑓 of a 𝑞-form (9.4.9). Let U be an
′ −1 −1
open subset of M and 𝑋1 , ..., 𝑋𝑞 be regular vector fields in 𝜑 (U).′If ℘ ∈ 𝜑 (U)
we denote by (𝜑∗ 𝑋1 ) ℘ , ..., 𝜑∗ 𝑋𝑞 ℘ the pushforwards at ℘ = 𝜑 (℘ ); we can then
define D E
(𝜑∗ 𝑓 ) ℘′ , 𝑋1 , ..., 𝑋𝑞 = 𝑓℘ , (𝜑∗ 𝑋1 ) ℘ , ..., 𝜑∗ 𝑋𝑞 ℘ .

The 𝑞-linear functional (𝜑∗ 𝑓 ) ℘′ : C𝑇 ∗ M −→ C is alternating. As ℘′ varies this

−1
defines a 𝑞-form in 𝜑 (U).
Now suppose U is the domain of regular local coordinates 𝑥 1 , ..., 𝑥 𝑛 and let
−1
𝑦 1 , ..., 𝑦 𝑛′ be regular local coordinates in a neighborhood V ⊂ 𝜑 (U) of ℘′ (𝑛 ′ =
dim N ). The map 𝜑 restricted to V can be expressed by equations 𝑥𝑖 = 𝑥 𝑖 (𝑦 1 , ..., 𝑦 𝑛′ ),
𝑖 = 1, ..., 𝑛. We see directly that, in V,
∑︁ D𝑥 𝐼
(𝜑∗ 𝑓 ) (𝑦, d𝑦) = 𝑓 𝐼 (𝑥 (𝑦)) d𝑦 𝐽 , (9.4.11)
D𝑦 𝐽
|𝐼 |= | 𝐽 |=𝑞

where
D𝑥 𝐼 D 𝑥 𝑖1 , ..., 𝑥 𝑖𝑞

𝜕𝑥 𝑖

= = det . (9.4.12)
D𝑦 𝐽 D 𝑦 , ..., 𝑦 𝜕𝑦 𝑗 𝑖 ∈𝐼, 𝑗 ∈𝐽
𝑗1 𝑗𝑞

It is obvious that 𝜑∗ 𝑓 is regular if this is true of 𝑓 . Two additional observations:

(1) What has just been said remains valid when N = M and the map 𝜑 stands
simply for a regular coordinate change.
−1
(2) The pullback to 𝜑 (U) of functions in U may be regarded as the special case
of the pullback of 𝑞-forms when 𝑞 = 0.

Remark 9.4.6 Suppose M is a complex manifold and (U, 𝑧1 , ..., 𝑧 𝑛 ) a holomorphic

local coordinates chart in M. Suppose 𝑛 ≥ 2; writing 𝑧 𝑗 = 𝑥 𝑗 + 𝑖𝑦 𝑗 we get
(𝑛−1) 𝑛
d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯1 ∧ · · · ∧ d𝑧¯𝑛 = (−1) 2 d𝑧1 ∧ d𝑧¯1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯𝑛
(9.4.13)
(𝑛−1) 𝑛
= (−1) (−2𝑖) 𝑛 d𝑥1 ∧ d𝑦 1 ∧ · · · ∧ d𝑥 𝑛 ∧ d𝑦 𝑛
2

= (−2𝑖) 𝑛 d𝑥 1 ∧ · · · ∧ d𝑥 𝑛 ∧ d𝑦 1 ∧ · · · ∧ d𝑦 𝑛 .
9.4 Differential Complexes and Grassman Algebras 287

We see that the orientations on U defined by d𝑧1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯1 ∧ · · · ∧ d𝑧¯𝑛 (resp.,

d𝑥1 ∧ d𝑦 1 ∧ · · · ∧ d𝑥 𝑛 ∧ d𝑦 𝑛 ) and d𝑧 1 ∧ d𝑧¯1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯𝑛 (resp., d𝑥 1 ∧ · · · ∧ d𝑥 𝑛 ∧
d𝑦 1 ∧ · · · ∧ d𝑦 𝑛 ) are the same if and only if either 𝑛 − 1 or 𝑛 is a multiple of 4.

9.4.4 Exterior derivative and De Rham complex

In this subsection we take K = R: M shall be a C ∞ manifold, dimR M = 𝑛 ≥ 1. We

deal with the smooth sections (9.4.9) of Λ𝑞 C𝑇 ∗ M, meaning that the coefficients 𝑓 𝐼
will be complex-valued C ∞ functions in U. By the exterior derivative of (9.4.9)
we mean the (𝑞 + 1)-form
𝑛
∑︁ ∑︁ ∑︁ 𝜕 𝑓𝐼
d 𝑓 (𝑥, d𝑥) = d 𝑓 𝐼 (𝑥) ∧ d𝑥 𝐼 = (𝑥) d𝑥 𝑗 ∧ d𝑥 𝐼 . (9.4.14)
𝜕𝑥 𝑗
|𝐼 |=𝑞 |𝐼 |=𝑞
𝑗=1

Í 𝜕 𝑓𝐼
When 𝑞 = 0, d 𝑓 = 𝑛𝑗=1 𝜕𝑥 𝑗
(𝑥) d𝑥 𝑗 . When 𝑞 ≥ 1 we have d𝑥 𝑗 ∧ d𝑥 𝐼 = 0 if 𝑗 ∈ 𝐼
(implying d 𝑓 = 0 if 𝑞 = 𝑛); and d𝑥 𝑗 ∧ d𝑥 𝐼 = ±d𝑥 𝐽 if 𝑗 ∉ 𝐼, where 𝐽 is the ordering
of the set of integers { 𝑗 } ∪ 𝐼 and where the sign is + if the permutation transforming
{ 𝑗 } ∪ 𝐼 into 𝐽 is even, − if it is odd. The 𝑞-form 𝑓 (𝑥, d𝑥) is said to be closed in U if
d 𝑓 ≡ 0. Assuming 𝑞 ≥ 1, 𝑓 (𝑥, d𝑥) is said to be exact in U if there is a (𝑞 − 1)-form
𝑢 (𝑥, d𝑥) = |𝐼 |=𝑞−1 𝑢 𝐼 (𝑥) ∧ d𝑥 𝐼 , 𝑢 𝐼 ∈ C ∞ (U) for all multi-indices 𝐼, such that
Í
d𝑢 = 𝑓 .
Let 𝜑 : N −→ M be a regular map between regular manifolds. It follows directly
from (9.4.11) that
d𝜑∗ 𝑓 = 𝜑∗ d 𝑓 (9.4.15)
where d stands for the exterior derivative in M at the right, in N at the left.
In particular Formula (9.4.14) is coordinate free. Consequently, the exterior
derivative defines a sequence of linear maps

d𝑞 : C ∞ (M; Λ𝑞 C𝑇 ∗ M) −→ C ∞ M; Λ𝑞+1 C𝑇 ∗ M , 𝑞 = 0, 1.... (9.4.16)

When 𝑞 ≥ 𝑛 we have Λ𝑞+1 C𝑇 ∗ M = 0 and so d𝑛 = 0. Obviously d𝑞 is a first-order

differential operator with regular coefficients (Subsection 9.2.3); supp d𝑞 𝑓 ⊂ supp 𝑓 .
It is immediately checked that d𝑞+1 ◦ d𝑞 = 0, i.e.,

Im d𝑞 ⊂ ker d𝑞+1 . (9.4.17)

In other words, every exact form is closed. This is the lemma originally stated by H.
Poincaré; its converse (when valid) is now called the Poincaré Lemma (cf. Theorem
9.4.10 below).
In the sequel, when there is no danger of confusion, we shall omit the subscripts
and simply write d for d𝑞 whatever the integer 𝑞 ≥ 0; (9.4.17) will simply read d2 =
d ◦ d = 0. It is traditional to “augment” the sequence of maps (9.4.16) by taking d−1
288 9 Elements of Differential Geometry

to be the injection (here denoted by 𝜄) of the space of locally constant functions in M,

here denoted by Constloc (M; C), into the space C ∞ M; Λ0 C𝑇 ∗ M = C ∞ (M; C).
Condition (9.4.17) is trivially valid for 𝑞 = −1. Thus the sequence of maps (where
the first arrow simply indicates that 𝜄 is injective)
𝜄 d d
0 −→ Constloc (M; C) −→ C ∞ (M; C) −→ C ∞ (M; C𝑇 ∗ M) −→ · · · (9.4.18)
d d d
−→ C ∞ (M; Λ𝑞 C𝑇 ∗ M) −→ C ∞ M; Λ𝑞+1 C𝑇 ∗ M −→ · · ·
d d
−→ C ∞ (M; Λ𝑛 C𝑇 ∗ M) −→ 0 ,

constitutes a differential complex (Definition 9.4.2).

Definition 9.4.7 The differential complex (9.4.18) is called the De Rham complex
of the manifold M (of class C ∞ ).
In the terminology of homological algebra, closed forms, the elements of ker d, are
cocycles and exact forms, the elements of Im d, are coboundaries (see next chapter);
every coboundary is a cocycle. The quotient spaces 𝐻 (𝑞) (M) = ker d𝑞 /Im d𝑞−1 1
are the De Rham cohomology spaces (or groups) of the manifold M. Note that
𝐻 (0) (M) = Constloc (M; C).
Proposition 9.4.8 The De Rham complex is elliptic.
Proof The principal symbol complex associated to the De Rham complex is the
sequence
·𝜎 𝜎∧
0 −→ M × C −→ C𝑇𝑥∗ M −→ · · · (9.4.19)
𝜎∧ 𝜎∧ 𝜎∧
· · · −→ Λ𝑞 C𝑇𝑥∗ M −→ Λ𝑞+1 C𝑇𝑥∗ M −→ · · ·
𝜎∧ 𝜎∧
· · · −→ Λ𝑛 C𝑇 ∗ M −→ 0 ,
Í
where 𝜎 = 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗 is the tautological one-form and ·𝜎 (second arrow) stands
for scalar multiplication of 𝜎. Of course, if 𝜃 ∈ Λ𝑞 C𝑇𝑥∗ M (𝑞 = 1, ..., 𝑛 − 1) then
𝜎 ∧ 𝜎 ∧ 𝜃 = 0. Conversely, assuming 𝑞 ≥ 2, it is a simple exercise in exterior algebra
to show that 𝜎 ∧ 𝜃 = 0 =⇒ 𝜃 = 𝜎 ∧ 𝜔 for some 𝜔 ∈ Λ𝑞−1 C𝑇𝑥∗ M. When 𝑞 = 1 the
latter means that 𝜃 is proportional to 𝜎. When 𝑞 = 𝑛 obviously 𝜎 ∧ 𝜃 = 0. Thus the
sequence (9.4.19) is exact. □
In order to prove that the De Rham complex (9.4.18) is locally exact (cf. Subsection
9.4.1) we shall apply a homotopy formula, (9.4.21) below. To each (ordered) multi-
index 𝐼 = 𝑖 1 , ..., 𝑖 𝑞 (𝑞 ≥ 1) we associate the 𝑞-form
𝑞
∑︁
𝜛𝐼 = (−1) 𝛼−1 𝑥𝑖 𝛼 d𝑥𝑖1 ∧ · · · ∧ d𝑥
d 𝑖 𝛼 ∧ · · · ∧ d𝑥 𝑖𝑞 , (9.4.20)
𝛼=1

1 In the remainder of this section Im stands for “image of”, meaning “range of”, and not “imaginary
part of”.
9.4 Differential Complexes and Grassman Algebras 289

where the hatted factor must be omitted. It is also convenient to write

𝑞
∑︁
𝜛𝐼 = (−1) 𝛼−1 𝑥𝑖 𝛼 d𝑥 𝐼\𝑖 𝛼 .
𝛼=1

We have d𝜛𝐼 = 𝑞d𝑥 𝐼 . We define the homotopy operator 𝐾 (𝑞) transforming 𝑞-forms
into (𝑞 − 1)-forms
∑︁ ∫ 1
(𝑞) 𝑞−1
𝐾 𝑓 (𝑥, d𝑥) = 𝑓 𝐼 (𝜆𝑥) 𝜆 d𝜆 𝜛𝐼 .
|𝐼 |=𝑞 0

Lemma 9.4.9 Let Ω ⊂ R𝑛 be a star-shaped open set centered at the origin and
𝑞 ∈ Z+ , 𝑞 ≥ 1. If 𝑓 ∈ C ∞ (Ω; Λ𝑞 C𝑇 ∗ R𝑛 ) then

𝑓 = d𝐾 (𝑞) 𝑓 + 𝐾 (𝑞+1) d 𝑓 . (9.4.21)

When 𝑞 = 0, 𝑓 ∈ C ∞ (Ω) and

𝑓 = 𝑓 (0) + 𝐾 (1) d 𝑓 . (9.4.22)

Proof When 𝑞 = 0 the claim is immediate:

∫ 1
𝜕
𝑓 (𝑥) = 𝑓 (0) + 𝑓 (𝜆𝑥) d𝜆
0 𝜕𝜆
∑︁ ∫ 1 𝜕 𝑓
= 𝑓 (0) + 𝑥𝑗 (𝜆𝑥) d𝜆.
𝑗 ∈𝐼 0 𝜕𝑥 𝑗

Now suppose 𝑞 ≥ 1. Given 𝐼 = 𝑖 1 , ..., 𝑖 𝑞 we have
𝑞
!
∑︁ 𝜕 𝑓𝐼 ∑︁ 𝜕 𝑓 𝐼
d𝑥 𝑖 ∧ 𝜛𝐼 = 𝑥𝑗 d𝑥 𝐼
𝛼=1
𝜕𝑥𝑖 𝛼 𝛼 𝑗 ∈𝐼
𝜕𝑥 𝑗

whence
∑︁ ∑︁ ∫ 1
(𝑞) 𝜕 𝑓𝐼 𝑞
d 𝐾 𝑓 (𝑥, d𝑥) = 𝑥𝑗 (𝜆𝑥) 𝜆 d𝜆 d𝑥 𝐼
0 𝜕𝑥 𝑗
|𝐼 |=𝑞 𝑗 ∈𝐼
∑︁ ∑︁ ∫ 1
𝜕 𝑓𝐼
+ (𝜆𝑥) 𝜆 𝑞 d𝜆 d𝑥 𝑗 ∧ 𝜛𝐼
0 𝜕𝑥 𝑗
|𝐼 |=𝑞 𝑗∉𝐼
∫ 1
+𝑞 𝑓 (𝜆𝑥, d𝑥) 𝜆 𝑞−1 d𝜆.
0
290 9 Elements of Differential Geometry

An integration by parts,
𝑛 ∫ 1 ∑︁ ∫ 1
∑︁ ∑︁ 𝜕 𝑓𝐼 𝑞 𝑞 𝜕
𝑥𝑗 (𝜆𝑥) 𝜆 d𝜆 d𝑥 𝐼 = 𝜆 𝑓 𝐼 (𝜆𝑥) d𝜆 d𝑥 𝐼
0 𝜕𝑥 𝑗 0 𝜕𝜆
|𝐼 |=𝑞 𝑗=1 |𝐼 |=𝑞
∫ 1
= 𝑓 (𝑥, d𝑥) − 𝑞 𝑓 (𝜆𝑥, d𝑥) 𝜆 𝑞−1 d𝜆,
0

yields
∑︁ ∑︁ ∫ 1
𝜕 𝑓𝐼
d 𝐾 (𝑞) 𝑓 (𝑥, d𝑥) = 𝑓 (𝑥, d𝑥) −

(𝜆𝑥) 𝜆 𝑞 d𝜆 𝑥 𝑗 d𝑥 𝐼 − d𝑥 𝑗 ∧ 𝜛𝐼 .
0 𝜕𝑥 𝑗
|𝐼 |=𝑞 𝑗∉𝐼

We have ∑︁ ∑︁ 𝜕 𝑓 𝐼
d 𝑓 (𝑥, d𝑥) = d𝑥 𝑗 ∧ d𝑥 𝐼 ,
𝜕𝑥 𝑗
|𝐼 |=𝑞
𝑗∉𝐼

whence
∫ 1
(𝑞+1)
∑︁ ∑︁
𝜀 (𝐼, 𝑗) 𝜕 𝑓𝐼 𝑞
𝐾 d𝑓 = (−1) (𝜆𝑥) 𝜆 d𝜆 𝜛 [𝐼, 𝑗 ],
0 𝜕𝑥 𝑗
|𝐼 |=𝑞 𝑗∉𝐼

where [𝐼, 𝑗] is the ordered multi-index such that d𝑥 𝑗 ∧ d𝑥 𝐼 = (−1) 𝜀 (𝐼, 𝑗) d𝑥 [𝐼, 𝑗 ] .
Applying (9.4.20) with 𝑞 + 1 substituted for 𝑞 and counting the commutations shows
readily that
(−1) 𝜀 (𝐼, 𝑗) 𝜛 [𝐼, 𝑗 ] = 𝑥 𝑗 d𝑥 𝐼 − d𝑥 𝑗 ∧ 𝜛𝐼 ,
whence the claim. □
What is known as the local Poincaré Lemma is the following corollary of Lemma
9.4.9.

Theorem 9.4.10 The De Rham complex (9.4.18) is locally exact.

In terms of the De Rham cohomology spaces Theorem 9.4.10 states that there is
a covering of the manifold M by open sets U such that 𝐻 (𝑞) (U) = {0} for every
𝑞 = 1, 2, ....

Remark 9.4.11 One could also consider the “holomorphic De Rham complex” over
an open subset U of a complex-analytic manifold M,
𝜄 𝜕 𝜕
0 −→ Constloc (U; C) −→ O (U; C) −→ O (U; 𝑇 ∗ U) −→ · · · (9.4.23)
𝜕 𝜕 𝜕
−→ O (U; Λ𝑞 𝑇 ∗ U) −→ O U; Λ𝑞+1𝑇 ∗ U −→ · · ·
𝜕 𝜕
−→ O (U; Λ𝑛𝑇 ∗ U) −→ 0 ,
9.4 Differential Complexes and Grassman Algebras 291

where O (U; ·) are the spaces of holomorphic sections over U and 𝜕 stands for the
“holomorphic” derivative: for a holomorphic function 𝑓 defined in local coordinates
Í 𝜕𝑓
𝑧1 , ..., 𝑧 𝑛 , 𝜕 𝑓 = 𝑛𝑗=1 𝜕𝑧 𝑗
d𝑧 𝑗 (which would be the meaning of d so far, when K = C).
Also keep in mind that in (9.4.23) 𝑇 ∗ M stands for the complex cotangent space.
With these meanings Proposition 9.4.8 and Theorem 9.4.10 and their proofs remain
valid.

9.4.5 The 𝝏 complex

In this subsection we take K = C and M to be a complex-analytic manifold. In most

aspects of Function Theory on such a manifold the differential complex of interest
is not likely to be the complex (9.4.18) [nor (9.4.23)], but rather one of the factors in
a special “direct sum” decomposition of its complexification, the 𝜕 complex which
we shall now define. Let 𝑛 = dimC M. Let us agree on the following notation:
𝑇 ∗ M means the real cotangent bundle of the real 2𝑛-dimensional C 𝜔 manifold M;
C𝑇 ∗ M = 𝑇 ∗ M ⊗ C is its complexification (here C stands for the vector bundle
M × C over M and the tensor product is taken over R; the rank of C𝑇 ∗ M over
R is equal to 4𝑛). Let d stand for the exterior derivative in M. Given an arbitrary
point ℘ ∈ M we denote by 𝑇℘(1,0) M the subspace of C𝑇℘∗ M whose elements are the
covectors dℎ (℘) with ℎ a holomorphic function in some neighborhood of ℘; and
by 𝑇℘(0,1) M its complex conjugate, spanned by the covectors d ℎ¯ (℘) (the complex
conjugate ℎ¯ is an antiholomorphic function). Select an open set U containing √ ℘, the
domain of complex coordinates 𝑧1 , ..., 𝑧 𝑛 ; if we write 𝑧 𝑗 = 𝑥 𝑗 + 𝑖𝑦 𝑗 (𝑖 = −1) we can
regard (𝑥1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 ) as a system of local real coordinates. We see that the
differentials d𝑧 𝑗 at ℘ span 𝑇℘(1,0) M over C while the d𝑧¯ 𝑘 = d𝑥 𝑘 − 𝑖d𝑦 𝑘 span 𝑇℘(0,1) M.
This shows that ℘ ↦→ 𝑇℘(1,0) M defines a complex vector bundle of rank 𝑛 over M,
𝑇 (1,0) M; ℘ ↦→ 𝑇℘(0,1) M defines the complex vector bundle 𝑇 (0,1) M. The latter is
the complex conjugate of the former; we have the direct sum decomposition

C𝑇 ∗ M = 𝑇 (1,0) M ⊕ 𝑇 (0,1) M. (9.4.24)

Now let 𝑓 ∈ C ∞ (U); we can form

𝑛
∑︁ 𝜕𝑓 𝜕𝑓 1 𝜕𝑓 𝜕𝑓
𝜕𝑓 = d𝑧¯ 𝑗 , = +𝑖 . (9.4.25)
𝑗=1
𝜕 𝑧¯ 𝑗 𝜕 𝑧¯ 𝑗 2 𝜕𝑥 𝑗 𝜕𝑥 𝑗

It is immediately checked that 𝜕 𝑓 is invariant under holomorphic changes of coor-

dinates and thus the local definition (9.4.25) leads to a globally defined differential
operator
𝜕 : C ∞ (M; C) −→ C ∞ M; 𝑇 (0,1) M . (9.4.26)
292 9 Elements of Differential Geometry

The right-hand side in (9.4.26) is the space of smooth sections of 𝑇 (0,1) M. The next
step is to form the exterior powers of 𝑇 (0,1) M, Λ𝑞 𝑇 (0,1) M; we shall use the notation
𝑇 (0,𝑞) M rather than Λ𝑞 𝑇 (0,1) M; of course, 𝑇 (0,𝑛+1) M = 0. Over the holomorphic
local chart (U, 𝑧1 , ..., 𝑧 𝑛 ) the smooth sections of 𝑇 (0,𝑞) M are the 𝑞-forms
∑︁
𝑓 (𝑧, d𝑧¯) = 𝑓 𝐼 (𝑥, 𝑦) d𝑧¯ 𝐼 , (9.4.27)
|𝐼 |=𝑞

where 𝑓 𝐼 ∈ C ∞ (U) and 𝐼 = 𝑖 1 , ..., 𝑖 𝑞 is an ordered multi-index of integers 1 ≤

𝑖 1 < · · · < 𝑖 𝑞 ≤ 𝑛. Over U we can define
∑︁
𝜕𝑓 = 𝜕 𝑓 𝐼 ∧ d𝑧¯ 𝐼 , (9.4.28)
|𝐼 |=𝑞

which is a smooth section of 𝑇 (0,𝑞+1) M over U. The invariance under holomorphic

changes of coordinates is readily checked; it allows us to define globally a differential
operator

𝜕 𝑞 : C ∞ M; 𝑇 (0,𝑞) M → C ∞ M; 𝑇 (0,𝑞+1) M , 𝑞 = 0, 1.... (9.4.29)

In local coordinates we have

𝑛 ∑︁
𝑛 ∑︁
∑︁ 𝜕 2 𝑓𝐼
𝜕 𝑞+1 𝜕 𝑞 𝐹 (𝑧, d𝑧¯) = d𝑧¯ 𝑗 ∧ d𝑧¯ 𝑘 ∧ d𝑧¯ 𝐼
𝑗=1 𝑘=1 |𝐼 |=𝑞
𝜕 𝑧¯ 𝑗 𝜕 𝑧¯ 𝑘
∑︁ ∑︁ 𝜕2 𝑓𝐼
= d𝑧¯ 𝑗 ∧ d𝑧¯ 𝑘 + d𝑧¯ 𝑘 ∧ d𝑧¯ 𝑗 ∧ d𝑧¯ 𝐼 = 0.
1≤ 𝑗<𝑘 ≤𝑛 |𝐼 |=𝑞
𝜕 𝑧¯ 𝑗 𝜕 𝑧¯ 𝑘

The sequence (9.4.29) is augmented by 𝜕 −1 = 𝜄, the natural injection of the space

of holomorphic functions in M, O (M), into C ∞ (M; C); thus Im 𝜄 = ker 𝜕 0 . This
proves that the sequence

𝜄 𝜕
𝜕
0 −→ O (M) −→ C ∞ (M; C) −→ C ∞ M; 𝑇 (0,1) M −→ · · · (9.4.30)

𝜕
𝜕 𝜕
−→ C ∞ M; 𝑇 (0,𝑞) M −→ C ∞ M; 𝑇 (0,𝑞+1) M −→ · · ·
𝜕
𝜕
−→ C ∞ M; 𝑇 (0,𝑛) M −→ 0

forms a differential complex (for the sake of simplicity we write 𝜕 rather than 𝜕 𝑞
and we shall do so in the future, when there is no risk of confusion). We define the
𝜕-cohomology spaces (or groups)

𝐻 (0,𝑞) (M) = ker 𝜕 𝑞 /Im 𝜕 𝑞−1 , 𝑞 = 1, 2, ...; (9.4.31)

9.4 Differential Complexes and Grassman Algebras 293

𝐻 (0,0) (M) is identified with the space of entire holomorphic functions in M. Of

course, 𝐻 (0,𝑞) (M) = {0} if 𝑞 > 𝑛.

Definition 9.4.12 The differential complex (9.4.30) is called the 𝜕-complex (or Dol-
beault) complex of the manifold M.

Proposition 9.4.13 The 𝜕-complex is elliptic.

Cf. Proposition 9.4.8 and its proof.

Proof We return to the expression (9.4.25) in the local chart (U, 𝑧1 , ..., 𝑧 𝑛 ). We call
𝜉 𝑗 and 𝜂 𝑗 the coordinates in the basis d𝑥 1 , ..., d𝑥 𝑛 , d𝑦 1 , ..., d𝑦 𝑛 of the real cotangent
spaces; we define 𝜁 𝑗 = 𝜉 𝑗 +𝑖𝜂 𝑗 . We see right away that the symbol of 𝜕 is the operator

1 ∑︁𝑛 𝑛
1 © ∑︁ 1
𝜎 𝜕 = 𝜁 𝑗 d𝑧¯ 𝑗 ∧ = ­𝑑 𝜁 𝑗 𝑧¯ 𝑗 ® ∧ = d (𝜁 · 𝑧¯) ∧ . (9.4.32)
ª
2 𝑗=1 2 2
« 𝑗=1 ¬

Let 𝜁 ∈ C𝑛 \ {0} be arbitrary. We can identify 𝑇℘(0,0) M with C and the action of

𝜎 𝜕 on 𝑇℘(0,0) M with the map 𝑧 ↦→ 21 𝑧d (𝜁 · 𝑧¯) ∈ 𝑇℘(0,1) M. It is evident that this
Í
map is injective. Suppose 𝑞 ≥ 2; if 𝜃 = |𝐼 |=𝑞 𝜃 𝐼 d𝑧¯ 𝐼 satisfies d (𝜁 · 𝑧¯) ∧ 𝜃 = 0 then
𝜃 is “divisible” by d (𝜁 · 𝑧¯), meaning that there is a (𝑞 − 1)-covector 𝜔 such that
𝜃 = 𝜔 ∧ d (𝜁 · 𝑧¯). □
The next statement is the Poincaré Lemma for the 𝜕-complex (cf. Theorem 9.4.10);
it is often referred to as the Dolbeault (or Dolbeault–Grothendieck) lemma.

Theorem 9.4.14 There is a covering of the manifold M by open sets U such that
𝐻 (0,𝑞) (U) = {0} for 𝑞 = 1, 2, ....

In other words, the 𝜕-complex (9.4.30) is locally exact.

Using a local chart it suffices to prove that 𝐻 (0,𝑞) Δ𝑛𝑅 = {0}, where Δ𝑛𝑅 =

Proof

𝑧 ∈ C𝑛 ; 𝑧 𝑗 < 𝑅, 𝑗 = 1, ..., 𝑛 . We are going to reason by induction on 𝑛.
The case 𝑛 = 1 is settled by the inhomogeneous Cauchy formula: given an arbitrary
form 𝑓 (𝑥, 𝑦) d𝑧¯, 𝑓 ∈ Cc∞ (Δ𝑅 ), we introduce the convolution,
∫ ∫
1 𝑓 (𝜉, 𝜂) 1 𝑓 (𝜉, 𝜂) ¯
(𝐸 ★ 𝑓 ) (𝑥, 𝑦) = d𝜉d𝜂 = d𝜁 ∧ d𝜁 (9.4.33)
𝜋 R2 𝑧 − 𝜁 2𝜋𝑖 R2 𝑧 − 𝜁
(𝑧 = 𝑥 + 𝑖𝑦, 𝜉 = 𝜉 + 𝑖𝜂). It is an elementary fact (see e.g. [Treves, 1975]) that

𝜕
𝐸 ★ 𝑓 = 𝑓 in R2 . (9.4.34)
𝜕 𝑧¯
Select a sequence 𝜒𝜈 ∈ Cc∞ (Δ𝑅 ), 𝜈 = 2, 3..., such that 𝜒𝜈 (𝑥, 𝑦) = 1 if 𝑧 ∈ Δ ( 1−𝜈 −1 ) 𝑅 .
Let 𝐹 ∈ C ∞ (Δ𝑅 ) be arbitrary and define 𝑢 𝜈 = 𝐸 ★ ( 𝜒𝜈 𝐹); (9.4.34) implies 𝜕𝑢 𝜕 𝑧¯ = 𝐹
𝜈
Í∞
in Δ ( 1−𝜈 −1 ) 𝑅 . In general the series 𝑢 3 + 𝜈=3 (𝑢 𝜈+1 − 𝑢 𝜈 ) does not converge in
294 9 Elements of Differential Geometry

C ∞ (Δ𝑅 ). But we can use the Mittag-Leffler procedure: we have 𝜕𝜕𝑧¯ (𝑢 𝜈+1 − 𝑢 𝜈 ) = 0
in Δ ( 1−𝜈 −1 ) 𝑅 ; using the Taylor expansion of 𝑢 𝜈+1 − 𝑢 𝜈 we can find a polynomial
𝑃 𝜈 ∈ C [𝑧] such that

sup |𝑢 𝜈+1 (𝑧) − 𝑢 𝜈 (𝑧) − 𝑃 𝜈 (𝑧)| ≤ 𝜀 𝜈

1
|𝑧 |< ( 1− 𝜈−1 )𝑅

where 𝜀 𝜈 ↘ 0 can be chosen in such a manner that the series

∞
∑︁
𝑢3 + (𝑢 𝜈+1 − 𝑢 𝜈 − 𝑃 𝜈 ) (9.4.35)
𝜈=3

converges in C ∞ (Δ𝑅 ), meaning that its derivatives of all orders converge uniformly
on disks Δ ( 1−𝜈 −1 ) 𝑅 , 𝜈 ↗ +∞. The choice of the sequence {𝜀 𝜈 } is left as an exercise.
We point out that, when 𝑓 happens to depend, say in C ∞ fashion, on some parameters
(𝑡1 , ..., 𝑡 𝑚 ) (that vary in R𝑚 ), Formula (9.4.33) ensures that the same is true of 𝑢 𝜈
and, by way of consequence, also of the truncated Taylor expansions 𝑃 𝜈 . By refining
the choice of the 𝜀 𝜈 one can prove the same property for (9.4.35).
Now suppose 𝑛 ≥ 1 and let a form (9.4.27) in Δ𝑛𝑅 belong to ker 𝜕. We exploit a
decomposition
𝐹 = 𝜑 ∧ d𝑧 𝑛 + 𝜓

where 𝜑 ∈ C ∞ Δ𝑛𝑅 ; 𝑇 (0,𝑞−1) and 𝜓 ∈ C ∞ Δ𝑛𝑅 ; 𝑇 (0,𝑞) do not involve any exterior
product factor d𝑧 𝑛 . Since 𝜕𝐹 = 0 we must have

′ 𝑛 𝜕𝜓 ′
𝜕 𝜑 − (−1) ∧ d𝑧 𝑛 + 𝜕 𝜓 = 0,
𝜕 𝑧¯𝑛
′ 𝜕
where 𝜕 stands for the 𝜕 operator in C𝑛−1 and acts on the 𝑞-form 𝜓 coefficient-
𝜕𝑧𝑛
wise. We must therefore have, in Δ𝑛𝑅 ,

′ 𝜕𝜓 ′
𝜕 𝜑 = (−1) 𝑛 , 𝜕 𝜓 = 0.
𝜕 𝑧¯𝑛

By the induction hypothesis there is a 𝜃 ∈ C ∞ Δ𝑛𝑅 ; 𝑇 (0,𝑞−1) C𝑛 , depending smoothly
′
on 𝑧 𝑛 ∈ Δ𝑅 and such that 𝜓 = 𝜕 𝜃. We derive

′ 𝜕𝜃
𝜕 𝜑 − (−1) 𝑛 = 0.
𝜕 𝑧¯𝑛

Again by the induction hypothesis we can find 𝜔 ∈ C ∞ Δ𝑛𝑅 ; 𝑇 (0,𝑞−2) , depending
smoothly on 𝑧 𝑛 ∈ Δ𝑅 and such that
9.4 Differential Complexes and Grassman Algebras 295

′ 𝜕𝜃
𝜕 𝜔 = 𝜑 − (−1) 𝑛 .
𝜕 𝑧¯𝑛
If we define 𝛼 = 𝜔 ∧ d𝑧 𝑛 + 𝜃 we get
′ 𝜕𝜃
𝜕𝛼 = 𝜕 𝜔 ∧ d𝑧 𝑛 + 𝜓 − (−1) 𝑛 ∧ d𝑧 𝑛 = 𝐹. □
𝜕 𝑧¯𝑛
The following global result is noteworthy.

Theorem 9.4.15 We have 𝐻 (0,𝑞) (C𝑛 ) = {0} for every 𝑞 = 1, 2, ..., 𝑛 − 1. We also
have 𝐻 (0,𝑛) (Ω) = {0} whatever the open subset Ω of C𝑛 .

that 𝐻 (0,𝑞) Δ𝑛𝑅 = {0} for every 𝑞 ≥ 1

Proof The proof of Theorem
9.4.14 shows

and 𝑅 > 0. Let 𝑓 ∈ C ∞ C𝑛 ; 𝑇 (0,𝑞) C𝑛 be such that 𝜕 𝑓 = 0.
Case 𝑞 = 1. We have 𝑓 (𝑧, d𝑧¯) = 𝑛𝑗=1 𝑓 𝑗 (𝑧) d𝑧¯ 𝑗 , 𝑓 𝑗 ∈ C ∞ (C𝑛 ). For each
Í

𝜈 = 2, 3..., let 𝑢 𝜈 ∈ Cc∞ Δ𝑛𝜈 be such that 𝜕𝑢 𝜈 = 𝑓 in Δ𝑛 1 . We have 𝑢 𝜈+1 −

𝜈− 2

𝑢 𝜈 ∈ O Δ𝑛 1 ; we can therefore find a polynomial 𝑃 𝜈 ∈ C [𝑧 1 , ..., 𝑧 𝑛 ] such that
𝜈− 2
|𝑢 𝜈+1 − 𝑢 𝜈 − 𝑃 𝜈 | ≤ 2−𝜈 at every point of Δ𝑛𝜈−1 . Define
∞
∑︁
𝑢 = 𝑢2 + (𝑢 𝜈+1 − 𝑢 𝜈 − 𝑃 𝜈 ) . (9.4.36)
𝜈=2

If 𝐾 is an arbitrary compact subset of C𝑛 and if the integer 𝜈◦ is so large that

𝐾 ⊂ Δ𝑛𝜈◦ −1 then we have, in 𝐾,

𝜈◦
∑︁
𝑢 − 𝑢2 + (𝑢 𝜈+1 − 𝑢 𝜈 − 𝑃 𝜈 ) ≤ 21−𝜈◦ .
𝜈=2

Thus (9.4.36) is a continuous function in C𝑛 that satisfies 𝜕𝑢 = 𝑓 in C𝑛 in the

distribution sense. This implies 𝑢 ∈ C ∞ R2𝑛 . Indeed, it implies
∑︁
Δ𝑢 = 4 𝜕 𝑧 𝑗 𝑓 𝑗 ∈ C ∞ R2𝑛
𝑗=1

and the Laplacian Δ is hypoelliptic (cf. Definition 3.1.1, Remark 4.1.3). [Actually,
one can select the 𝑃 𝜈 with more care, to achieve that the series (9.4.36) converges in
C ∞ R2𝑛 .]


Case 2 ≤ 𝑞 < 𝑛. For each 𝜈 = 2, 3..., we select 𝑢 𝜈 ∈ C ∞ Δ𝑛𝜈 ; 𝑇 (0,𝑞−1) C𝑛
such that 𝜕𝑢 𝜈 = 𝑓 in Δ𝑛 . We have 𝜕 (𝑢 𝜈+1 − 𝑢 𝜈 ) = 0 in Δ𝑛 1 . Starting with
𝜈− 21
𝜈− 2
𝜙2 ≡ 0 and using induction on 𝜈 we can find 𝜙 𝜈+1 ∈ C Δ𝑛𝜈+1 ; 𝑇 (0,𝑞−2) C𝑛
∞

satisfying 𝜕𝜙 𝜈+1 = 𝑢 𝜈+1 − 𝑢 𝜈 + 𝜕𝜙 𝜈 in Δ𝑛𝜈−1 . In other words, 𝑢 𝜈+1 − 𝜕𝜙 𝜈+1 ∈

296 9 Elements of Differential Geometry

C ∞ Δ𝑛𝜈+1 ; 𝑇 (0,𝑞−1) C𝑛 is an extension of the pullback (i.e., the restriction) to Δ𝑛𝜈−1 of

𝑢 𝜈 − 𝜕𝜙 𝜈 ∈ C ∞ Δ𝑛𝜈 ; 𝑇 (0,𝑞−1) C𝑛 . It follows that there is a 𝑢 ∈ C ∞ C𝑛 ; 𝑇 (0,𝑞−1) C𝑛
whose pullback to Δ𝑛𝜈−1 is equal to that of 𝑢 𝜈 − 𝜕𝜙 𝜈 ; obviously 𝜕𝑢 = 𝑓 in C𝑛 .
Case 𝑞 = 𝑛. Let us write

𝑓 (𝑧, d𝑧¯) = 𝑓◦ (𝑧) d𝑧¯1 ∧ · · · ∧ d𝑧¯𝑛 , 𝑓◦ ∈ C ∞ (Ω) .

As a linear operator C ∞ (Ω) ←↪ the Laplacian Δ is surjective (Theorem 5.3.15). Let

𝑢 ◦ ∈ C ∞ (Ω) satisfy
𝑛
∑︁
Δ𝑢 ◦ = 4 𝜕𝑧 𝑗 𝜕 𝑧 𝑗 𝑢 ◦ = 𝑓◦ .
𝑗=1

It follows that the (0, 𝑛 − 1)-form

𝑛
∑︁
𝑢= (−1) 𝑗 𝜕 𝑧 𝑗 𝑢 ◦ d𝑧¯1 ∧ · · · ∧ dc
𝑧¯ 𝑗 ∧ · · · ∧ d𝑧¯𝑛
𝑗=1

(hatted factor to be omitted) satisfies 𝜕𝑢 = 𝑓 . □

Back to a generic complex manifold M. Instead of the vector bundles 𝑇 (0,𝑞) M
we could have used the vector bundles 𝑇 ( 𝑝,𝑞) M with 1 ≤ 𝑝 ≤ 𝑛. In a holomorphic
local chart (U, 𝑧1 , ..., 𝑧 𝑛 ) the sections of 𝑇 ( 𝑝,𝑞) M have expressions
∑︁ ∑︁
𝐹 (𝑧, d𝑧, d𝑧¯) = 𝑓 𝐼, 𝐽 (𝑧) d𝑧 𝐼 ∧ d𝑧¯ 𝐽 (9.4.37)
|𝐼 |= 𝑝 | 𝐽 |=𝑞

in which 𝐼 and 𝐽 are (ordered) multi-indices and 𝑓 𝐼, 𝐽 ∈ C ∞ (U). We shall say that
𝐹 (𝑧, d𝑧, d𝑧¯) is a differential form of bidegree ( 𝑝, 𝑞) in the complex manifold M.
We have
𝑟 ∑︁𝑛
∑︁ 𝜕 𝑓 𝐼, 𝐽
𝜕𝐹 (𝑧, d𝑧, d𝑧¯) = (𝑧) d𝑧¯ 𝑘 ∧ d𝑧 𝐼 ∧ d𝑧¯ 𝐽 . (9.4.38)
𝑗=1 𝑘=1
𝜕 𝑧¯ 𝑘

We get the differential complexes

𝜄 𝜕 𝜕
0 −→ O M; 𝑇 ( 𝑝,0) M −→ C ∞ M; 𝑇 ( 𝑝,0) M −→ C ∞ M; 𝑇 ( 𝑝,1) M −→
(9.4.39)
𝜕
𝜕
𝜕
· · · −→ C ∞ M; 𝑇 ( 𝑝,𝑞) M −→ C ∞ M; 𝑇 ( 𝑝,𝑞+1) M −→ · · ·
𝜕
𝜕
−→ C ∞ M; 𝑇 ( 𝑝,𝑛) M −→ 0

where O M; 𝑇 ( 𝑝,0) M consists of the sections of 𝑇 ( 𝑝,0) M that are holomorphic.
Í
In a holomorphic local chart (U, 𝑧1 , ..., 𝑧 𝑛 ) these have the form | 𝐽 |= 𝑝 ℎ 𝐽 (𝑧) d𝑧 𝐽
with ℎ 𝐽 ∈ O (U). The standard notation for the corresponding cohomology groups
9.4 Differential Complexes and Grassman Algebras 297

is
𝐻 ( 𝑝,𝑞) (M) = ker 𝜕 𝑞 /Im 𝜕 𝑞−1 (9.4.40)

where 𝜕 𝑞 is the map 𝜕 : C ∞ M; 𝑇 ( 𝑝,𝑞) M −→ C ∞ M; 𝑇 ( 𝑝,𝑞+1) M , 𝑞 = 1, 2, ....
The isomorphism
∑︁
C ∞ M; 𝑇 (0,𝑞) M ∋ 𝑓 𝐽 d𝑧¯ 𝐽 (9.4.41)
| 𝐽 |=𝑞
∑︁
−→ 𝑓 𝐽 d𝑧1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯ 𝐽 ∈ C ∞ M; 𝑇 (𝑛,𝑞) M
|𝐽 |=𝑞

commutes with 𝜕 and thus induces an isomorphism between the differential complex
(9.4.39) where 𝑝 = 𝑛 and (9.4.30); it follows that 𝐻 (0,𝑞) (M) 𝐻 (𝑛,𝑞) (M). In
particular, we could have stated Theorems 9.4.14 and 9.4.15 with 𝐻 (𝑛,𝑞) substituted
for 𝐻 (0,𝑞) .

Proposition 9.4.16 If 𝑝 = 𝑛 in the differential complex (9.4.39) the operator 𝜕 is

equal to the exterior derivative 𝑑.

Proof Indeed, in local coordinates, for any multi-index 𝐽 and any C 1 function 𝑓 , we
have
𝑛
∑︁ 𝜕𝑓
d ( 𝑓 d𝑧1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯ 𝐽 ) = (−1) 𝑛 d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯ 𝑘 ∧ d𝑧¯ 𝐽 . □
𝑘=1
𝜕 𝑧¯ 𝑘

About the global solvability of the differential complex (9.4.39), the following
result is important.

Theorem 9.4.17 If Ω is a Runge domain in C𝑛 (Definition 6.1.5) then 𝐻 𝑝,𝑞 (Ω) = 0

for all 𝑝 ≥ 0, 𝑞 ≥ 1.

The proof of this statement (Theorem 2.7.8 in [Hörmander, 1966]) is fairly el-
ementary: it is based on the Oka lemma (Lemma 2.7.5, loc. cit.) and does not use
the full machinery needed in the proof of the analogous result for general pseudo-
convex domains (see Definition 11.2.15). Traditionally, the fact that 𝐻 0,1 (Ω) = 0 is
expressed by saying that the first Cousin problem has a solution in Ω.
Chapter 10
A Primer on Sheaf Cohomology

This chapter is a first modest step toward a link with the algebraic (i.e., cohomo-
logical) approach to hyperfunctions and microfunctions favored by many mathe-
maticians, to start with, by the initiator of the theory, M. Sato and his school (see
[Sato-Kawai-Kashiwara, 1973], [Morimoto, 1973]). We have strived to keep the
exposition as elementary as feasible.
The first section presents a very elementary introduction to Sheaf Cohomology
based on sheaf-valued cochains. The reader who finds our presentation too skimpy
is referred to [Godement, 1964], Ch. II.
Section 10.2 introduces fine sheaves, leading to the cohomology groups of a sheaf
E of Abelian groups on a paracompact Hausdorff topological space 𝑋 through its
fine resolutions. This is applied to the De Rham and 𝜕 differential complexes (over
domains of R𝑛 and C𝑛 respectively), introduced in the last section of the preceding
chapter.
The final section defines relative cohomology and explains the redefinitions, of
analytic functionals as cohomological classes, of hyperfunctions and microfunctions
as relative cohomology classes. We have refrained, perhaps unwisely, from going
deeper into the matter, such as the important tools of Ext, Tor, spectral sequences (see
[Godement, 1964], Ch. I) and from using more modern language such as Verdier’s
derived categories. Here again there are many texts devoted to such topics where the
interested student can find needed information.

10.1 Basics on Sheaf Cohomology

10.1.1 Cohomology valued in a sheaf

Let 𝑋 be a topological space and F a sheaf of Abelian groups F 𝑥 over 𝑋 (see

Subsection 1.2.2). All groups considered in this section will be Abelian. We denote
by Γ (𝑈; F) the group of continuous sections of F over an open subset 𝑈 of 𝑋.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 299
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_10
300 10 A Primer on Sheaf Cohomology

Let U = {𝑈 𝛼 } 𝛼∈ℑ be an open covering of 𝑋 with ℑ a set of indices. For each

multi-index 𝐼 = 𝛼0 , ..., 𝛼𝑞 ∈ ℑ𝑞+1 we set 𝑈𝐼 = 𝑈 𝛼0 ∩ · · · ∩ 𝑈 𝛼𝑞 .

Definition 10.1.1 Let 𝑞 ∈ Z+ . By a 𝑞-cochain of the open covering U of 𝑋 valued

in the sheaf F we mean a map

𝒔 : ℑ𝑞+1 ∋ 𝐼 ↦→ 𝒔 𝐼 ∈ Γ (𝑈𝐼 ; F) (10.1.1)

such that 𝒔 𝐼 = −𝒔 𝐼 ′ if the multi-indices 𝐼 = 𝛼0 , ..., 𝛼𝑞 and 𝐼 ′ = 𝛼0′ , ..., 𝛼𝑞′ differ

by an odd permutation. We denote by ℭ 𝑞 (U; F) the group of 𝑞-cochains.

Every even permutation of 𝐼 is the composite of two odd permutations of 𝐼 and

therefore 𝒔 𝐼 = 𝒔 𝐼 ′ if 𝐼 and 𝐼 ′ differ by an even permutation.
We define the coboundary operator 𝛿 (𝑞) : ℭ 𝑞 (U; F) −→ ℭ 𝑞+1 (U; F) (we
write 𝛿 when there is no risk of confusion): if 𝒔 ∈ ℭ 𝑞 (U; F) and 𝐼 = 𝛼0 , ..., 𝛼𝑞+1 ∈
ℑ𝑞+2 we have

𝑞+1
∑︁
𝛿 (𝑞) 𝒔 = (−1) 𝑗 𝒔 𝛼0 ,..., 𝛼 𝑗−1 , 𝛼 𝑗+1 ..., 𝛼𝑞+1 in 𝑈𝐼 , (10.1.2)
𝐼
𝑗=0

with the understanding that when 𝑗 = 0 (resp, 𝑗 = 𝑞 + 1) 𝒔 𝛼0 ,..., 𝛼 𝑗−1 , 𝛼 𝑗+1 ..., 𝛼𝑞+1 =
𝒔 𝛼1 ,..., 𝛼𝑞+1 (resp., 𝒔 𝛼0 ,..., 𝛼 𝑗−1 , 𝛼 𝑗+1 ..., 𝛼𝑞+1 = 𝒔 𝛼0 ,..., 𝛼𝑞 ). It is checked directly that 𝛿 (𝑞+1) ◦
𝛿 (𝑞) = 0 for all 𝑞. The standard notation is1

𝑍 𝑞 (U; F) = ker 𝛿 (𝑞) , 𝐵𝑞 (U; F) = Im 𝛿 (𝑞−1) , 𝑞 ≥ 1; (10.1.3)

𝐵𝑞 (U; F) is a subgroup of 𝑍 𝑞 (U; F); the elements of 𝑍 𝑞 (U; F) are called 𝑞-

cocycles, those of 𝐵𝑞 (U; F) 𝑞-coboundaries; the quotient

𝐻 𝑞 (U; F) = 𝑍 𝑞 (U; F) /𝐵𝑞 (U; F)

is the 𝑞 th cohomology group associated to the covering U.

To say that 𝒔 ∈ ℭ 0 (U; F) is a cocycle is the same as saying that, for arbitrary pairs
(𝛼0 , 𝛼1 ) ∈ ℑ2 , if 𝑈 𝛼0 ∩𝑈 𝛼1 ≠ ∅ then 𝒔 𝛼0 = 𝒔 𝛼1 in 𝑈 𝛼0 ∩𝑈 𝛼1 . Since 𝐵0 (U; F) = {0}
we have
𝐻 0 (U; F) Γ (U; F) . (10.1.4)
Here and throughout the sequel of this section denotes an isomorphism (of whatever
type – category
– of objects is in play).
Let V = 𝑉𝛽 𝛽 ∈𝔎 be another open covering of 𝑋. Assume that V is a refinement
of U, meaning that there is a map 𝜙 : 𝔎 −→ ℑ such that 𝑉𝛽 ⊂ 𝑈 𝜙 (𝛽) for every
𝛽 ∈ 𝔎. Introducing the map

𝔎𝑞+1 ∋ 𝐾 = 𝛽0 , ..., 𝛽𝑞 −→ 𝜙 (𝐾) = 𝜙 (𝛽0 ) , ..., 𝜙 𝛽𝑞 ∈ ℑ𝑞+1 (10.1.5)

1 In this section Im stands for “image of”, meaning “range of”, and not “imaginary part of”.
10.1 Basics on Sheaf Cohomology 301

allows us to define a homomorphism ℭ 𝑞 (U; F) −→ ℭ 𝑞 (V; F) by assigning to an

arbitrary cochain (10.1.1) the cochain

𝔎𝑞+1 ∋ 𝐾 −→ 𝒔 𝜙 (𝐾) 𝑉𝐾
∈ Γ (𝑉𝐾 ; F) .

This homomorphism commutes with the coboundary operator and thus induces a
homomorphism
(𝑞)
𝜏F, 𝜙,U, V : 𝐻 𝑞 (U; F) −→ 𝐻 𝑞 (V; F) . (10.1.6)
Proposition 10.1.2 Let V be a refinement of U. The homomorphism (10.1.6) does
not depend on the choice of the map 𝜙.
Proof In view of (10.1.4) the claim is trivial if 𝑞 = 0; suppose 𝑞 ≥ 1. Let 𝜓 : 𝔎 −→ ℑ
be another map such that 𝑉𝛽 ⊂ 𝑈 𝜓 (𝛽) . We define a map 𝜒 : ℭ 𝑞+1 (U; F) −→
ℭ 𝑞 (V; F) by the formula
𝑞
∑︁
∀𝒔 ∈ ℭ 𝑞+1 (U; F) , ( 𝜒𝒔) 𝐾 = (−1) ℓ 𝒔 𝐼ℓ (10.1.7)
ℓ=0

with 𝐾 = 𝛽0 , ..., 𝛽𝑞 ∈ 𝔎𝑞+1 , 𝐼ℓ = 𝜙 (𝛽0 ) , . . . , 𝜙 (𝛽ℓ ) , 𝜓 (𝛽ℓ ) , . . . , 𝜓 𝛽𝑞 ; note
that 𝐼0 = (𝜙 (𝛽0 ) , 𝜓 (𝐾)) [cf. (10.1.5)] and 𝐼𝑞 = 𝜙 (𝐾) , 𝜓𝑞 𝛽𝑞 . Now let us take
𝒔 ∈ ℭ 𝑞 (U; F), 𝑞 ≥ 1; we have, by (10.1.2),
𝑞−1
∑︁
( 𝜒 (𝛿𝒔)) 𝐾 = (𝛿𝒔) 𝐼0 + (−1) ℓ (𝛿𝒔) 𝐼ℓ + (−1) 𝑞 (𝛿𝒔) 𝐼𝑞 ,
ℓ=1

whence
𝑞−1
∑︁
( 𝜒 (𝛿𝒔)) 𝐾 − 𝒔 𝜓 (𝐾) + 𝒔 𝜙 (𝐾) = (−1) ℓ (𝛿𝒔) 𝐼ℓ (10.1.8)
ℓ=1
𝑞
∑︁
+ (−1) 𝑗 𝒔 𝜙 (𝛽0 ) ,..., 𝜙 ( 𝛽 𝑗−1 ) , 𝜙 ( 𝛽 𝑗+1 ) ,..., 𝜙 ( 𝛽𝑞 ) , 𝜓 ( 𝛽𝑞 )
𝑗=0
𝑞
∑︁
+ (−1) 𝑞 (−1) 𝑗 𝒔 𝜙 (𝛽0 ), 𝜓 (𝛽0 ),..., 𝜓 ( 𝛽 𝑗−1 ) , 𝜓 ( 𝛽 𝑗+1 ) ,..., 𝜓 ( 𝛽𝑞 )
𝑗=0

with the understanding that

𝑗 = 0 =⇒ 𝒔 𝜙 (𝛽0 ), 𝜓 (𝛽0 ),..., 𝜓 ( 𝛽 𝑗−1 ) , 𝜓 ( 𝛽 𝑗+1 ) ,..., 𝜓 ( 𝛽𝑞 ) = 𝒔 𝜙 (𝛽0 ), 𝜓 (𝛽1 ),..., 𝜓 ( 𝛽𝑞 ) ,

𝑗 = 𝑞 =⇒ 𝒔 𝜙 (𝛽0 ),..., 𝜙 ( 𝛽 𝑗−1 ) , 𝜙 ( 𝛽 𝑗+1 ) ,..., 𝜙 ( 𝛽𝑞 ) , 𝜓 ( 𝛽𝑞 ) = 𝒔 𝜙 (𝛽0 ),..., 𝜙 ( 𝛽𝑞−1 ) , 𝜓 ( 𝛽𝑞 ) .

With the same multi-index 𝐾 we have

𝑞
∑︁
(𝛿 𝜒𝒔) 𝐾 = (−1) 𝑗 ( 𝜒𝒔) 𝛽0 ,...,𝛽 𝑗−1 ,𝛽 𝑗+1 ...,𝛽𝑞 .
𝑗=0
302 10 A Primer on Sheaf Cohomology

Replacing 𝑞 by 𝑞 − 1 and 𝐾 by 𝛽0 , ..., 𝛽 𝑗−1 , 𝛽 𝑗+1 . . . , 𝛽𝑞 in (10.1.7) we get
𝑗−1
∑︁ 𝑞
∑︁
( 𝜒𝒔) 𝛽0 ,...,𝛽 𝑗−1 ,𝛽 𝑗+1 ...,𝛽𝑞 = 𝒔 𝐼 𝑗,ℓ + 𝒔 𝐼 𝑗,ℓ (10.1.9)
ℓ=0 ℓ= 𝑗+1

where

𝐼 𝑗,ℓ = 𝜙 (𝛽0 ) , . . . , 𝜙 (𝛽ℓ ) , 𝜓 (𝛽ℓ ) , . . . , 𝜓 𝛽 𝑗−1 , 𝜓 𝛽 𝑗+1 , . . . , 𝜓 𝛽𝑞

if 0 ≤ ℓ ≤ 𝑗 − 1 whereas

𝐼 𝑗,ℓ = 𝜙 (𝛽0 ) , . . . , 𝜑 𝛽 𝑗−1 , 𝜑 𝛽 𝑗+1 , . . . , 𝜙 (𝛽ℓ ) , 𝜓 (𝛽ℓ ) , . . . , 𝜓 𝛽𝑞

if 𝑗 + 1 ≤ ℓ ≤ 𝑞. It is not difficult (but tedious) to show that the right-hand sides in

(10.1.8) and (10.1.9) are equal, implying

( 𝜒 (𝛿𝒔)) 𝐾 − (𝛿 𝜒𝒔) 𝐾 = 𝒔 𝜓 (𝐾) − 𝒔 𝜙 (𝐾) .

This proves that

(𝑞) (𝑞)
∀𝒔 ∈ ℭ 𝑞 (U; F) , 𝜒𝛿𝒔 − 𝛿 𝜒𝒔 = 𝜏F, 𝜓, U, V 𝒔− 𝜏F, 𝜙,U, V 𝒔.

(𝑞) (𝑞)
If 𝛿𝒔 = 0 we have 𝜏F, 𝜓, U, V 𝒔− 𝜏F, 𝜙,U, V 𝒔 = 𝛿 𝜒𝒔 ∈ 𝐵𝑞 (U; F) [recall that 𝜒𝒔 ∈
ℭ 𝑞−1 (U; F) ]. □
(𝑞)
The homomorphism (10.1.6) shall henceforth be denoted by 𝜏F, U, V .
(𝑞)
Corollary 10.1.3 For every covering U of 𝑋 and every 𝑞 ∈ Z+ , 𝜏F, U, U is the identity
map of 𝐻 𝑞 (U; F).
Proof Indeed, we can take 𝜙 in (10.1.5) to be the identity map of the index set ℑ. □
Two coverings U, V are said to be equivalent if V is a refinement of U and U
is a refinement of V.
Let W be another open covering of 𝑋, this time a refinement of V; we have
(𝑞) (𝑞) (𝑞)
𝜏F, V, W ◦ 𝜏F, U, V = 𝜏F, U, W . This has the following direct consequence.
(𝑞)
Corollary 10.1.4 If the coverings of 𝑋, U, V, are equivalent then 𝜏F, U, V :
𝐻 𝑞 (U; F) −→ 𝐻 𝑞 (V; F) is a group isomorphism for every 𝑞 ∈ Z+ .
(𝑞) (𝑞) (𝑞)
Proof We have 𝜏F, V, U ◦ 𝜏F, U, V = 𝜏F, U, U , the identity map of 𝐻 𝑞 (U; F) by
Corollary 10.1.3. □
Proposition 10.1.5 Let V be a refinement of U. The homomorphism
(1)
𝜏F,U,V : 𝐻 1 (U; F) −→ 𝐻 1 (V; F)

is injective.
10.1 Basics on Sheaf Cohomology 303

Proof Let 𝜙 be the map (10.1.5) and 𝒔 ∈ 𝑍 1 (U, F) such that 𝜙𝒔 = 𝛿𝒄,
𝒄 ∈ ℭ 0 (V, F), meaning that, given an arbitrary pair 2
(𝛽1 , 𝛽2 ) ∈ 𝔎 , we have
0
𝒔 𝜙 (𝛽1 ), 𝜙 (𝛽2 ) = 𝒄 𝛽1 − 𝒄 𝛽2 in 𝑉𝛽1 ∩ 𝑉𝛽2 for some 𝒄 𝛽 𝑗 ∈ ℭ V𝛽 𝑗 , F , 𝑗 = 1, 2. Since the

coboundary of 𝒔 = 𝒔 𝛼1 , 𝛼2 ( 𝛼1 , 𝛼2 ) ∈𭟋2 vanishes identically, given any index 𝛼 ∈ ℑ we
have

𝒔 𝛼, 𝜙 (𝛽1 ) + 𝒔 𝜙 (𝛽1 ), 𝜙 (𝛽2 ) + 𝒔 𝜙 (𝛽2 ), 𝛼 = 𝒔 𝛼, 𝜙 (𝛽1 ) + 𝒄 𝛽1 − 𝒄 𝛽2 − 𝒔 𝛼, 𝜙 (𝛽2 ) = 0

identically in 𝑈 𝛼 ∩ 𝑉𝛽1 ∩ 𝑉𝛽2 . This allows us to define 𝒔 ′ ∈ ℭ 0 (U, F) by the formula

𝒔 ′𝛼 = 𝒔 𝛼, 𝜙 (𝛽) + 𝒄 𝛽 in 𝑈 𝛼 ∩ 𝑉𝛽 for every 𝛼 ∈ ℑ and 𝛽 ∈ 𝔎. If (𝛼1 , 𝛼2 ) ∈ ℑ2 and
𝛽 ∈ 𝔎 we obtain, in 𝑈 𝛼1 ∩ 𝑈 𝛼2 ∩ 𝑉𝛽 ,

𝒔 ′𝛼1 − 𝒔 ′𝛼2 = 𝒔 𝛼1 , 𝜙 (𝛽) − 𝒔 𝛼2 , 𝜙 (𝛽) = 𝒔 𝛼1 , 𝛼2

which proves that 𝒔 = 𝛿𝒔 ′. □

Let 𝔓 (𝑋) denote the power set of 𝑋; an open covering of 𝑋 can be regarded as
a subset of 𝔓 (𝑋); the equivalence class of a given open covering U of 𝑋, i.e., the
family [U] of all the open coverings of 𝑋 equivalent to U, can be regarded as a
subset of 𝔓 (𝔓 (𝑋)). Corollary 10.1.4 allows us to associate to [U] the cohomology
groups 𝐻 𝑞 (U; F) and to denote them by 𝐻 𝑞 ([U] ; F).
By the natural ordering of these equivalence classes we mean the following
partial order relation: [U] ≺ [V] if [U] ≠ [V] and if some representative of [V]
is a refinement of some representative of [U] (in which case every representative of
[V] is a refinement of every representative of [U]).

Definition 10.1.6 For each 𝑞 = 1, 2, ..., we define 𝐻 𝑞 (𝑋; F) to be the direct limit
of the groups 𝐻 𝑞 ([U] ; F) with respect to the natural ordering of the equivalence
classes [U].

Concretely, the direct limit of the groups 𝐻 𝑞 ([U] ; F) is obtained by equating

two elements 𝑢 ∈ 𝐻 𝑞 ([U] ; F) and 𝑣 ∈ 𝐻 𝑞 ([V] ; F) when there is [W] ≺ [U],
(𝑞) (𝑞)
[W] ≺ [V] such that 𝜏F, U, W
𝑢 = 𝜏F, V, W 𝑣.

Proposition 10.1.7 If there is an open covering U of 𝑋 such that 𝐻 1 (U; F) = 0

then 𝐻 1 (𝑋; F) = 0.

This follows immediately from Proposition 10.1.5 and Definition 10.1.6.

We will need one functorial feature of the cohomology groups 𝐻 𝑞 (𝑋; F).

Proposition 10.1.8 Given an arbitrary pair of topological spaces 𝑋, 𝑌 , to each

continuous map 𝑓 : 𝑌 −→ 𝑋 there correspond group homomorphisms

−1
∗ 𝑞 𝑞
𝑓𝑞 : 𝐻 (𝑋; F) −→ 𝐻 𝑌 ; 𝑓 F , 𝑞 ∈ Z+ , (10.1.10)

−1
where 𝑓 F is the sheaf on 𝑌 whose stalk at each 𝑦 ∈ 𝑌 is F 𝑓 ( 𝑦) .
304 10 A Primer on Sheaf Cohomology

−1 −1
Proof Given an open covering U = {𝑈 𝛼 } 𝛼∈𝐼 of 𝑋 then 𝑓 (U) = 𝑓 (𝑈 𝛼 )
𝛼∈ℑ
is an open covering of 𝑌 . By pullback 𝑓 every cochain 𝒔 ∈ ℭ (U; F)
under 𝑞

−1 −1
defines a cochain 𝒔 ◦ 𝑓 ∈ ℭ 𝑞 𝑓 (U) ; 𝑓 F . The map 𝑠 ↦→ 𝒔 ◦ 𝑓 commutes
with the coboundary operator
𝛿; it therefore defines a natural homomorphism
−1 −1
∗

𝑓 : 𝐻 (U; F) −→ 𝐻 𝑓 (U) ; 𝑓 F . If V = 𝑉𝛽 𝛽 ∈𝔎 is an arbitrary open cov-
𝑞 𝑞

ering of 𝑌 then W = 𝑈 𝛼 ∩ 𝑉𝛽 𝛼∈ℑ,𝛽 ∈𝔎 is a refinement of V; by restriction 𝒔 ◦ 𝑓
defines an element of ℭ 𝑞 (W; F). We get thus a map 𝐻 𝑞 (U; F) −→ 𝐻 𝑞 (W; F)
and thereby maps 𝐻 𝑞 ([U] ; F) −→ 𝐻 𝑞 ([W] ; F). By going to the direct limits we
end up with the “natural” maps (10.1.10). □
When 𝑞 = 0 (10.1.10) is the standard pullback map

−1
𝑓 ∗ : Γ (𝑋; F) −→ Γ 𝑌 ; 𝑓 F . (10.1.11)

10.1.2 Connecting homomorphism and long exact sequences

We recall the standard terminology and notation: If 𝑮 1 , 𝑮 2 , 𝑮 3 are Abelian groups,

a pair of group homomorphisms 𝜗1 : 𝑮 1 → 𝑮 2 , 𝜗2 : 𝑮 2 → 𝑮 3 , define an exact
sequence if 𝜗1 is injective, 𝜗2 surjective and Im 𝜗1 (the range of 𝜗1 ) is equal to
ker 𝜗2 (the nullspace of 𝜗2 ); this is expressed by the formula
𝜗1 𝜗2
0 → 𝑮 ! → 𝑮 2 → 𝑮 3 → 0.

Likewise, a sequence of homomorphisms of sheaves of Abelian groups on 𝑋 (an

arbitrary topological space)

𝜄 𝜗0 𝜗1 𝜗𝑝 𝜗 𝑝+1
0 → E → F0 → F1 → · · · → F 𝑝 → F 𝑝+1 → · · · (10.1.12)

is said to be exact if 𝜄 is injective, Im 𝜄 = ker 𝜗0 and Im 𝜗 𝑝 = ker 𝜗 𝑝+1 for every

𝑝 ∈ Z+ .
Consider a “short” exact sequence of sheaves of group homomorphisms:

𝜄 𝜗
0→E→F→G→0 (10.1.13)

(thus 𝜗 is surjective).
10.1 Basics on Sheaf Cohomology 305

Proposition 10.1.9 Let Ω be an open subset of 𝑋. The sequence of Abelian group

homomorphisms
𝜄 𝜗
0 −→ Γ (Ω, E) −→ Γ (Ω; F) −→ Γ (Ω, G) (10.1.14)

is exact.

Proof It is clear that Im 𝜄 ⊂ ker 𝜗 in (10.1.14). Let 𝒇 ∈ ker 𝜗 be arbitrary. Given an

arbitrary point 𝑥 ◦ ∈ Ω there exist a neighborhood 𝑉𝑥 ◦ ⊂ Ω of 𝑥 ◦ and 𝒆 ∈ Γ (𝑉𝑥 ◦ , E)
such that 𝜄𝒆 (𝑥) = 𝒇 (𝑥) for every 𝑥 ∈ 𝑉𝑥 ◦ . Let 𝑥 ∗ ∈ Ω, 𝑉𝑥 ∗ ⊂ Ω a neighborhood of 𝑥 ∗
and 𝒆 ∗ ∈ Γ (𝑉𝑥 ∗ , E) be such that 𝜄𝒆 ∗ (𝑥) = 𝒇 (𝑥) for every 𝑥 ∈ 𝑉𝑥 ∗ . If 𝑉𝑥 ◦ ∩ 𝑉𝑥 ∗ ≠ ∅
then 𝜄𝒆 (𝑥) = 𝜄𝒆 ∗ (𝑥) for every 𝑥 ∈ 𝑉𝑥 ◦ ∩ 𝑉𝑥 ∗ . Since 𝜄 is injective we conclude that
there is 𝒆 ∈ Γ (Ω,E) such that 𝜄𝒆 = 𝒇 in the whole of Ω. □

Corollary 10.1.10 Let U = {𝑈 𝛼 } 𝛼∈ℑ be an open covering of 𝑋. For each 𝑞 ∈ Z+

the sequence
𝜄 𝜗
0 −→ ℭ 𝑞 (U, E) −→ ℭ 𝑞 (U; F) −→ ℭ 𝑞 (U, G) (10.1.15)

is exact. The maps 𝜄 and 𝜗 commute with the coboundary operator 𝛿.

Proof Follows directly from (10.1.1), (10.1.2) and Proposition 10.1.9. □

Corollary 10.1.11 For every 𝑞 ∈ Z+ ,

𝜄Z 𝑞 (U, E)⊂Z 𝑞 (U; F) , 𝜄B 𝑞 (U, E) ⊂B 𝑞 (U; F) ,

𝜗Z 𝑞 (U; F)⊂Z 𝑞 (U, G) , 𝜗B 𝑞 (U; F) ⊂B 𝑞 (U, G) .
𝑞
The range of 𝜗 in (10.1.15) shall be denoted by ℭlift (U, G) and its elements
referred to as the liftable 𝑞-cochains. The exactness of (10.1.15) entails that of the
sequence
𝜄 𝑞 𝜗
0 −→ ℭ 𝑞 (U, E) −→ ℭ 𝑞 (U; F) −→ ℭlift (U, G) → 0. (10.1.16)

We define
𝑞 𝑞
Zlift (U, G) = ℭlift (U, G) ∩ Z 𝑞 (U, G) , (10.1.17)
𝑞 𝑞−1
Blift (U, G) = 𝛿ℭlift (U, G) ,
𝑞 𝑞 𝑞
𝐻lift (U, G) = Zlift (U, G) /Blift (U, G) .

Corollary 10.1.11 shows that the maps 𝜄 and 𝜗 define the sequence of homomor-
phisms
𝜄∗ 𝑞 𝜗∗
𝐻 𝑞 (U, E) −→ 𝐻 𝑞 (U; F) −→ 𝐻lift (U, G) . (10.1.18)
In the sequel we denote by [ 𝑓 ] ∈ (U, E) the coset (or cohomology class) of 𝑓 ∈
𝐻𝑞
𝑞 𝑞
Z 𝑞 (U, E); likewise, we denote by [𝑔] ∈ 𝐻lift (U, G) the coset of 𝑔 ∈ Zlift (U, G).
306 10 A Primer on Sheaf Cohomology

Proposition 10.1.12 The range of 𝜄∗ is equal to ker 𝜗∗ .

Proof Let 𝑓 ∈ Z 𝑞 (U; F) be such that 𝜗∗ [ 𝑓 ] = 0; by (10.1.17) this is equivalent to

𝑞−1
the property that 𝜗 𝑓 = 𝛿𝑔 for some 𝑔 ∈ ℭlift (U, G). By the exactness of (10.1.16)
there is an 𝑓1 ∈ ℭ 𝑞−1 (U; F) such that 𝜗 𝑓1 = 𝑔 whence 𝜗 ( 𝑓 − 𝛿 𝑓1 ) = 0 and therefore
there is an 𝑒 ∈ ℭ 𝑞 (U, G) such that 𝜄𝑒 = 𝑓 − 𝛿 𝑓1 whence 𝜄𝛿𝑒 = 0. Since 𝜄 is injective
we see that 𝑒 ∈ Z 𝑞 (U, E) and 𝜄∗ [𝑒] = [ 𝑓 ]. □
Definition of the connecting homomorphism 𝛿∗ .
We now define a group homomorphism

𝛿∗ : 𝐻lift
𝑞
(U, G) −→ 𝐻 𝑞+1 (U, E) (10.1.19)

that will allow us to insert the short exact sequence (10.1.18) into a long exact
sequence of the cohomology groups.
𝑞
Let 𝒈 ∈ Zlift (U, G) and 𝒇 ∈ ℭ 𝑞 (U; F) be such that 𝒈 = 𝜗 𝒇 . Corollary 10.1.10
implies both 𝜗𝛿 𝒇 = 𝛿𝒈 = 0 and the existence of 𝒆 ∈ Z 𝑞+1 (U, E) such that 𝜄𝒆 = 𝛿 𝒇 .
If 𝒇 1 ∈ ℭ 𝑞 (U; F) and 𝒆 1 ∈ Z 𝑞+1 (U, E) are such that 𝒈 = 𝜗 𝒇 1 and 𝜄𝒆 1 = 𝛿 𝒇 1
then 𝒇 − 𝒇 1 = 𝜄𝒆 2 for some 𝒆 2 ∈ ℭ 𝑞 (U, E) whence 𝜄 (𝒆 − 𝒆 1 ) = 𝜄𝛿𝒆 2 and therefore
𝒆 − 𝒆 1 = 𝛿𝒆 2 . This allows us to define the map

𝛿 ′ : Zlift
𝑞
(U, G) ∋ 𝒈 −→ [𝒆] ∈ 𝐻 𝑞+1 (U, E) .

When 𝑞 = 0 we take 𝛿∗ = 𝛿 ′. Suppose 𝑞 ≥ 1, 𝒈 = 𝛿𝒈 ◦ , 𝒈 ◦ ∈ ℭlift 𝑞−1

(U, G);
let 𝒇 ◦ ∈ ℭ 𝑞−1 (U; F) be such that 𝒈 ◦ = 𝜗 𝒇 ◦ whence 𝒈 = 𝜗 𝒇 = 𝜗𝛿 𝒇 ◦ . By the
exactness of (10.1.16) there is an 𝒆 1 ∈ ℭ 𝑞 (U, E) such that 𝜄𝒆 1 = 𝒇 − 𝛿 𝒇 ◦ . Back
to 𝒆 ∈ Z 𝑞+1 (U, E) such that 𝜄𝒆 = 𝛿 𝒇 we get 𝜄 (𝒆 − 𝛿𝒆 1 ) = 0 whence 𝒆 = 𝛿𝒆 1 .
This proves that [𝒆] = 0; thus 𝛿 ′ Blift
𝑞
(U, G) = {0}. This shows that 𝛿 ′ induces a
homomorphism
𝑞 𝑞 𝑞
𝐻lift (U, G) = Zlift (U, G) /Blift (U, G) −→ 𝐻 𝑞+1 (U, E) ,

the sought operator 𝛿∗ .

Theorem 10.1.13 Let the sheaves E, F, G be as in (10.1.13). The sequence of Abelian

groups
𝜄∗ 𝜗∗ 𝛿∗ 𝜄∗
0 −→ 𝐻 0 (U, E) −→ 𝐻 0 (U; F) −→ 𝐻lift
0
(U, G) −→ 𝐻 1 (U, E) −→ · · ·
(10.1.20)
𝜄∗ 𝑞 𝜗∗ 𝛿∗ 𝜄∗
𝐻 𝑞 (U, E) −→ 𝐻 𝑞 (U; F) −→ 𝐻lift (U, G) −→ 𝐻 𝑞+1 (U, E) −→ · · ·

is exact.
𝑞
Proof Let 𝒈 ∈ Zlift (U, G), 𝒇 ∈ ℭ 𝑞 (U; F), 𝒆 ∈ Z 𝑞+1 (U, E) be related as in the
above definition of the connecting homomorphism. Proposition 10.1.5 states that
Im 𝜄∗ = ker 𝜗∗ .
10.1 Basics on Sheaf Cohomology 307

Now suppose 𝛿∗ [𝒈] = [𝒆] = 0: there is an 𝒆 1 ∈ ℭ 𝑞 (U, E) such that 𝒆 = 𝛿𝒆 1 . We

have 𝛿 ( 𝒇 − 𝜄𝒆 1 ) = 0 and 𝜗 ( 𝒇 − 𝜄𝒆 1 ) = 𝒈, whence [𝒈] = 𝜗∗ [ 𝒇 − 𝜄𝒆 1 ]. This proves
that ker 𝛿∗ = Im 𝜗∗ .
Lastly, 𝜄∗ [𝒆] = 0 means that there is an 𝒇 ∈ ℭ 𝑞 (U; F) such that 𝜄𝒆 = 𝛿 𝒇 . If
𝑞
𝒈 = 𝜗 𝒇 ∈ ℭlift (U, G) then 𝛿𝒈 = 𝜗𝜄𝒆 = 0; the latter is equivalent to saying that
𝒈 ∈ Zlift (U, G) and [𝒆] = 𝛿∗ [𝒈]. This proves that ker 𝜄∗ = Im 𝛿∗ .
𝑞
□

Let the open covering V = 𝑉𝛽 𝛽 ∈𝔎 of 𝑋 be a refinement of U and 𝜙 : 𝔎 −→
ℑ a map such that 𝑉𝛽 ⊂ 𝑈 𝜙 (𝛽) for every 𝛽 ∈ 𝔎. The analogue of Proposition
10.1.2 is valid for lifted cochains, with exactly the same proof: the homomorphism
𝑞 𝑞
𝐻lift (U, G) −→ 𝐻lift (V, G) based on 𝜙 does not depend on the choice of the map
(𝑞)
𝜙; denote it by 𝜏G, U, V,lift . If W is another open covering of 𝑋 which is a refinement
(𝑞) (𝑞) (𝑞)
of V, then 𝜏G, U, V,lift ◦ 𝜏G, V, W,lift = 𝜏G, U, W,lift .
The following statement shows the importance (for our purposes) of restricting
the choice of the base space 𝑋.
Proposition 10.1.14 Let 𝑋 be Hausdorff and paracompact. The natural map of the
𝑞
direct limit of the groups 𝐻lift ([U] , G) into 𝐻 𝑞 (𝑋, G) is a group isomorphism.
Proof The paracompactness of 𝑋 ensures that every open covering of 𝑋 admits a
locally finite refinement. Let U = {𝑈 𝜄 } 𝜄 ∈𝐼 be a locally finite open covering of 𝑋.
We return to (10.1.15); let 𝒈 ∈ ℭ 𝑞 (U, G) be arbitrary. By the exactness of (10.1.13)
each 𝑥 ∈ 𝑋 has a neighborhood 𝑉𝑥 intersecting finitely many open sets 𝑈 𝜄 and such,
moreover, that 𝜗 𝒇 𝑥 = 𝒈 in 𝑉𝑥 for some 𝒇 𝑥 ∈ ℭ 𝑞 (U; F). We can replace 𝑉𝑥 by
𝑉𝑥 ∩ 𝑈 𝜄1 ∩ · · · ∩ 𝑈 𝜄𝜈 if 𝑈 𝜄1 , . . . , 𝑈 𝜄𝜈 are the members of U that contain 𝑥. This
has the effect that 𝑥 ∈ 𝑈 𝜄 =⇒ 𝑉𝑥 ⊂ 𝑈 𝜄 whatever 𝜄 ∈ 𝐼. Let 𝜙 : 𝑋 −→ ℑ be a
map such that 𝑉𝑥 ⊂ 𝑈 𝜙 ( 𝑥) for every 𝑥 ∈ 𝑋; thus V = {𝑉𝑥 } 𝑥 ∈𝑋 is a refinement of
U. If 𝑥 0 , ..., 𝑥 𝑞 ∈ 𝑋 𝑞+1 we have 𝒈 = 𝜗 𝒇 𝑥0 over 𝑉𝑥0 ∩ · · · ∩ 𝑉𝑥𝑞 ; in other words,
𝑞
𝒈 ∈ ℭlift (V, G). This directly implies the claim. □
It is readily seen that combining Theorem 10.1.13 and Proposition 10.1.14 yields
the following results, under the hypothesis that 𝑋 is Hausdorff and paracompact and
the sheaves of Abelian groups, E, F, G, are as in (10.1.13).
Theorem 10.1.15 The sequence of Abelian groups
𝜄∗ 𝜗∗ 𝛿∗ 𝜄∗
0 −→ 𝐻 0 (𝑋, E) −→ 𝐻 0 (𝑋; F) −→ 𝐻 0 (𝑋, G) −→ 𝐻 1 (𝑋, E) −→ · · · (10.1.21)
𝜄∗ 𝜗∗ 𝛿∗ 𝜄∗
𝐻 𝑞 (𝑋, E) −→ 𝐻 𝑞 (𝑋; F) −→ 𝐻 𝑞 (𝑋, G) −→ 𝐻 𝑞+1 (𝑋, E) −→ · · ·

is exact.
Corollary 10.1.16 If there is an open covering U of 𝑋 such that 𝐻 1 (U; E) = 0 then
the short sequence of Abelian groups
𝜄∗ 𝜗∗ 𝛿∗
0 −→ Γ (𝑋, E) −→ Γ (𝑋; F) −→ Γ (𝑋, G) −→ 0

is exact.
308 10 A Primer on Sheaf Cohomology

Proof Follows from Proposition 10.1.7 and Theorem 10.1.15. □

In other words, under the hypotheses of Corollary 10.1.16 we have the isomor-
phism
Γ (𝑋, G) Γ (𝑋; F) /Γ (𝑋, E) . (10.1.22)

10.2 Fine Sheaves and Fine Resolutions

10.2.1 Fine sheaves on paracompact Hausdorff spaces

From now on the topological space 𝑋 shall be paracompact (i.e., every open covering
of 𝑋 has a locally finite refinement) and Hausdorff. In this framework the following
(classical) definition is important.

Definition 10.2.1 Let the topological space 𝑋 be Hausdorff and paracompact. A

sheaf F of Abelian groups over 𝑋 is said to be fine if, given an arbitrary locally
finite open covering U = {𝑈 𝛼 } 𝛼∈ℑ of 𝑋, to every 𝛼 ∈ ℑ there is a (continuous)
Í
sheaf endomorphism 𝒇 𝛼 of F such that 𝒇 𝛼 (𝑥) = 0 unless 𝑥 ∈ 𝑈 𝛼 and 𝛼∈ℑ 𝒇 𝛼
is identically equal to the identity map of F. We refer to the family 𝒇 𝛼 𝛼∈ℑ as a
partition of the identity subordinate to the covering U.

Each 𝒇 𝛼 defines an endomorphism of ℭ 𝑞 (𝑌 , F) (𝑌 ⊂ 𝑋, 𝑞 ∈ Z+ ); these endo-

morphisms commute with the coboundary operator 𝛿 as seen directly on (10.1.2).
The importance of Definition 10.2.1 is illustrated by the following result.

Theorem 10.2.2 Let 𝑋 be Hausdorff and paracompact. If the sheaf F of Abelian

groups over 𝑋 is fine then 𝐻 𝑞 (𝑋, F) = 0 for every 𝑞 = 1, 2, ....

Proof Let V = {𝑉 𝛼 } 𝛼∈ℑ be a locally finite open covering of 𝑋 and 𝒇 𝛼 𝛼∈ℑ a
partition of the identity subordinate to V. Let 𝒔 = {𝒔 𝐼 } 𝐼 ∈ℑ𝑞+1 ∈ Z 𝑞 (V, F) be
arbitrary. For each 𝐽 = 𝛽0 , ..., 𝛽𝑞−1 ∈ ℑ we define
𝑞

∑︁
𝒔 ′𝐽 = 𝒇 𝛼 𝒔 𝛼, 𝐽 .
𝛼∈ℑ

It is clear that 𝒔 ′ = 𝒔 ′𝐽 𝐽 ∈ℑ𝑞 ∈ ℭ 𝑞−1 (V, F) since 𝒔 𝛼, 𝐽 changes sign under an odd

permutation of the indices 𝛽 𝑗 ; (10.1.2) and the fact that 𝛿𝒔 = 0 imply, for arbitrary
𝐼 = 𝛼0 , ..., 𝛼𝑞 ∈ ℑ𝑞+1 ,
𝑞+1
∑︁
(𝛿𝒔 ′) 𝐼 = (−1) 𝑗 𝒔 ′𝛼0 ,..., 𝛼 𝑗−1 , 𝛼 𝑗+1 ..., 𝛼𝑞+1
𝑗=0
𝑞+1
∑︁ ∑︁
= (−1) 𝑗 𝒇 𝛼 𝒔 𝛼, 𝛼0 ,..., 𝛼 𝑗−1 , 𝛼 𝑗+1 ..., 𝛼𝑞+1
𝛼∈ℑ 𝑗=0
10.2 Fine Sheaves and Fine Resolutions 309
∑︁ ∑︁
=− 𝒇 𝛼 (𝛿𝒔) { 𝛼}∪𝐼 + 𝒇 𝛼 𝒔𝐼 = 𝒔𝐼
𝛼∈ℑ 𝛼∈ℑ

at points of 𝑉𝐼 . We conclude that 𝐻 𝑞 ([V] , F) = 0; since every open covering of 𝑋

has a locally finite refinement it ensues that the direct limit of the groups 𝐻 𝑞 ( [U] ; F)
is equal to 0. □

10.2.2 Fine resolutions and sheaf cohomology

From now on, unless specified otherwise, the topological space 𝑋 shall be Hausdorff
and paracompact.

Definition 10.2.3 Let E be a sheaf of Abelian groups on 𝑋. A sequence of homo-

morphisms of sheaves of Abelian groups on 𝑋,

𝛼 𝜗0 𝜗 𝑝−1 𝜗𝑝 𝜗 𝑝+1
0 → E → F0 → · · · → F 𝑝 → F 𝑝+1 → · · ·, (10.2.1)

is called a fine resolution of E if (10.2.1) is exact and every sheaf F 𝑝 is fine.

Theorem 10.2.2 directly implies

Proposition 10.2.4 The fine resolution (10.2.1) is acyclic, meaning 𝐻 𝑞 𝑋, F 𝑝 = 0
for all ( 𝑝, 𝑞) ∈ Z2+ , 𝑞 ≥ 1.

We decompose the exact sequence (10.2.1) into a system of short exact sequences
(where G 𝑝+1 = Im 𝜗 𝑝 = ker 𝜗 𝑝+1 )

𝛼 𝜗0
0 −→ E −→ F0 −→ G1 −→ 0, (10.2.2)
𝛼 𝜗1
0 −→ G1 −→ F1 −→G2 −→ 0, . . . ,
𝛼 𝜗𝑝
0 −→ G 𝑝 −→ F 𝑝 −→ G 𝑝+1 −→ 0, . . . .

We apply Theorem 10.1.15 to each short sequence (10.2.2):

𝛼∗ 𝜗0∗
0 −→ 𝐻 0 (𝑋, E) −→ 𝐻 0 (𝑋; F0 ) −→ 𝐻 0 (𝑋, G1 )
𝛿0∗ 𝛼∗ 𝜗0∗
−→ 𝐻 1 (𝑋, E) −→ 𝐻 1 (𝑋; F0 ) −→ (10.2.3)
𝛼∗ 𝜗0∗ 𝛿0∗ 𝛼∗ 𝜗0∗
· · · −→ 𝐻 𝑞 (𝑋; F0 ) −→ 𝐻 𝑞 (𝑋, G1 ) −→ 𝐻 𝑞+1 (𝑋, E) −→ 𝐻 𝑞+1 (𝑋; F0 ) −→ · · ·

and for 𝑝 = 1, 2, ...,

310 10 A Primer on Sheaf Cohomology

𝛼∗ 𝜗 ∗𝑝 𝛿 ∗𝑝 𝛼∗
0 −→ 𝐻 0 𝑋, G 𝑝 −→ 𝐻 0 𝑋; F 𝑝 −→ 𝐻 0 𝑋, G 𝑝+1 −→ 𝐻 1 𝑋, G 𝑝 −→ · · ·
(10.2.4)
∗ 𝜗 ∗ 𝛿 ∗ ∗
𝛼 𝑝 𝑝 𝛼
𝐻 𝑞 𝑋, G 𝑝 −→ 𝐻 𝑞 𝑋; F 𝑝 −→ 𝐻 𝑞 𝑋, G 𝑝+1 −→ 𝐻 𝑞+1 𝑋, G 𝑝 −→ · · ·.

Using the fact that (10.2.1) is acyclic (Proposition 10.2.4) we derive from (10.2.3):
𝜗0∗ 𝛿0∗
Γ (𝑋; F0 ) −→ Γ (𝑋, G1 ) −→ 𝐻 1 (𝑋, E) → 0, (10.2.5)

and that 𝛿0∗ defines isomorphisms

𝐻 𝑞+1 (𝑋, E) 𝐻 𝑞 (𝑋, G1 ) , 𝑞 = 1, 2, ..., (10.2.6)

and from (10.2.4) that the following sequences, for 𝑝 = 1, 2, ..., are exact:

𝜗 ∗𝑝 𝛿 ∗𝑝
Γ 𝑋; F 𝑝 −→ Γ 𝑋, G 𝑝+1 −→ 𝐻 1 𝑋, G 𝑝 → 0,

𝛿 ∗𝑝 𝛼∗
0 −→ 𝐻 𝑞 𝑋, G 𝑝+1 −→ 𝐻 𝑞+1 𝑋, G 𝑝 −→ 0, 𝑞 = 1, 2, ...,
implying that 𝛿∗𝑝 defines isomorphisms

𝐻 1 𝑋, G 𝑝 Γ 𝑋, G 𝑝+1 /𝜗∗𝑝 Γ 𝑋; F 𝑝 ,

(10.2.7)

𝐻 𝑞 𝑋, G 𝑝+1 𝐻 𝑞+1 𝑋, G 𝑝 , 𝑞 = 1, 2, .... (10.2.8)

Theorem 10.2.5 Given the fine resolution (10.2.1) on the (paracompact and Haus-
dorff) topological space 𝑋 we have, for each integer 𝑞 ≥ 1,
∗
𝐻 𝑞 (𝑋, E) Γ 𝑋; ker 𝜗𝑞 /𝜗𝑞−1 Γ 𝑋; F𝑞−1 . (10.2.9)

Proof Recall that ker 𝜗𝑞 = G𝑞 . For 𝑞 = 1 (10.2.9) follows directly from (10.2.5).
From (10.2.6), (10.2.7) and (10.2.8) we deduce, for 𝑞 ≥ 2,
∗
𝐻 𝑞 (𝑋, E) 𝐻 𝑞−1 (𝑋, G1 ) · · · 𝐻 1 𝑋, G𝑞−1 Γ 𝑋, G𝑞 /𝜗𝑞−1

Γ 𝑋; F𝑞−1 .
□

10.2.3 Examples

10.2.3.1 De Rham sheaf cohomology

Let M be a C ∞ manifold countable at infinity (thus paracompact); dimR M = 𝑛. We

refer the reader to the De Rham differential complex (9.4.18), which is locally exact
(Theorem 9.4.10). It follows that the sequence of sheaf homomorphisms
10.2 Fine Sheaves and Fine Resolutions 311

𝜄 𝑑 𝑑
0 −→ C −→ C∞ (M) −→ C∞ (𝑇 ∗ M) −→ · · · (10.2.10)
𝑑 𝑑 𝑑 𝑑 𝑑
−→ C∞ (Λ𝑞 𝑇 ∗ M) −→ C∞ Λ𝑞+1𝑇 ∗ M −→ · · · −→ C∞ (Λ𝑛𝑇 ∗ M) −→ 0

is exact. In (10.2.10) C∞ (Λ𝑞 𝑇 ∗ M) stands for the sheaf on M of germs of C ∞

sections of the (complex) vector bundle Λ𝑞 C𝑇 ∗ M (𝑞 = 1, 2, ...), C∞ (M) stands for
the sheaf on M of germs of C ∞ functions in M, and C for the constant sheaf on M.
Below the stalk at ℘ ∈ M of Λ𝑞 C𝑇 ∗ M is denoted by C∞ ℘ (Λ C𝑇 ).
𝑞 ∗

Proposition 10.2.6 The sheaves C∞ (Λ𝑞 𝑇 ∗ M) are fine.

Proof Let U = 𝑈 𝑗 𝑗 ∈Z+ be a locally finite open covering of M and {𝜑} 𝑗 ∈Z+
a C ∞ partition of the unity subordinate to the covering U, meaning that 𝜑 𝑗 ∈
Cc∞ 𝑈 𝑗 , ∞ ∞ (Λ𝑞 C𝑇 ∗ ) ∋ f −→ 𝜑 f
Í
𝜑
𝑗=0 𝑗 ≡ 1. Then the maps C ℘ ℘ 𝑗 ℘ (℘ ∈ M) are
endomorphisms that satisfy the requirements in Definition 10.2.1. □
Theorem 10.2.5 applies directly: we obtain the classical description

𝐻 𝑞 (M, C) (10.2.11)
.n o
∞ 𝑞 ∗ ∞ 𝑞−1 ∗
= { 𝑓 ∈ C (M; Λ C𝑇 M) ; d 𝑓 = 0} d𝑓; 𝑓 ∈ C M; Λ C𝑇 M .

10.2.3.2 Sheaf cohomology and the 𝝏 complex

In this subsection we take M to be a complex-analytic manifold, countable at infinity,

dimC M = 𝑛. The differential complex (9.4.39) is locally exact (Theorem 9.4.14). It
follows that the sequence of sheaf homomorphisms

𝜄 𝜕 𝜕
0 −→ O𝑇 ( 𝑝,0) (M) −→ C∞𝑇 ( 𝑝,0) (M) −→ C∞𝑇 ( 𝑝,1) (M) −→ · · · (10.2.12)
𝜕 𝜕 𝜕 𝜕
−→ C∞𝑇 ( 𝑝,𝑞) (M) −→ C∞𝑇 ( 𝑝,𝑞+1) (M) −→ · · · −→ C∞𝑇 ( 𝑝,𝑛) (M) → 0

is exact. In (10.2.12) O𝑇 ( 𝑝,0) (M) stands for the sheaf of germs of holomorphic
sections of the complex vector bundle 𝑇 ( 𝑝,0) M and C∞𝑇 ( 𝑝,𝑞) (M) for the sheaf of
germs of C ∞ sections of the vector bundle 𝑇 ( 𝑝,𝑞) M. Thus O (M) [resp., C∞ (M)]
stands for the sheaf of germs of holomorphic (resp., smooth) functions in M.

Proposition 10.2.7 The sheaves C∞

( M)
𝑇 ( 𝑝,𝑞) are fine.

The proof is the same as that of Proposition 10.2.6. Here Theorem 10.2.5 also
applies directly:

𝐻 𝑞 M, O𝑇 ( 𝑝,0) (M)
n o.n o
𝑓 ∈ C ∞ M; 𝑇 ( 𝑝,𝑞) M ; 𝜕 𝑓 = 0 𝜕 𝑓 ; 𝑓 ∈ C ∞ M; 𝑇 ( 𝑝,𝑞−1) M ,
312 10 A Primer on Sheaf Cohomology

which is to say, by (9.4.31),

𝐻 𝑞 M, O𝑇 ( 𝑝,0) (M) 𝐻 𝑝,𝑞 (M) . (10.2.13)

10.2.3.3 Application: analytic cohomology in real space

We close this section with an application of (10.2.13) leading to a result already used
in this book (in Subsection 7.1.5). We denote by C 𝜔 = C 𝜔 (R𝑛 ) the sheaf of germs
of (complex-valued) analytic functions in R𝑛 .

Theorem 10.2.8 Let Ω be an open subset of R𝑛 ; we have 𝐻 1 (Ω; C 𝜔 ) = 0.

Proof We have proved that an arbitrary compact subset 𝐾 of Ω is Runge (Theorem

6.2.14); in fact we have proved that there is a basis of neighborhoods of 𝐾 in C𝑛
consisting of Runge open subsets 𝑈 C𝑗 of C𝑛 ( 𝑗 = 1, 2, ...). Let then 𝑓 be a C 𝜔 -valued
1-cocycle in Ω . There is a 𝑗 such that 𝑓 extends as an O (𝑛) -valued 1-cocycle 𝑓˜
in 𝑈 C𝑗 (O (𝑛) : the sheaf of germs of holomorphic functions in C𝑛 ). Theorem 9.4.17

and (10.2.13) imply 𝐻 1 𝑈 C𝑗 , O (𝑛) = 0. Therefore there is a 𝑢 𝑗 ∈ O 𝑈 C𝑗 such
that 𝛿C 𝑢 𝑗 = 𝑓˜ in 𝑈 C𝑗 (𝛿C : the coboundary operator on O (𝑛) -valued cochains). By

restriction to 𝑈 𝑗 = 𝑈 C𝑗 ∩ R𝑛 we get 𝛿 𝑢 𝑗 𝑈 𝑗 = 𝑓˜ (𝛿: the coboundary operator
on C 𝜔 -valued cochains); 𝑢 𝑗 is unique by (10.1.4). If 𝐾 expands to Ω and, as a
consequence, 𝑈 C𝑗 (which depends on 𝑓 ) expands to a neighborhood ΩC of Ω in C𝑛 ,

then the 𝑢 𝑗 converge to 𝑢 ∈ O ΩC such that 𝛿 ( 𝑢| Ω ) = 𝑓 . □
Lemma 7.1.26 is a more explicit formulation of Theorem 10.2.8.
It is an easy exercise to prove, by combining Theorem 6.2.14, and (10.2.13), that
𝐻 𝑞 (Ω; C 𝜔 ) = 0 for every 𝑞 ≥ 1 (see also Theorem 9.4.17).

10.3 Relative Sheaf Cohomology

10.3.1 Relative sheaf cohomology. Definition

We return to an arbitrary sheaf F of Abelian groups over a paracompact Hausdorff

space 𝑋. We apply Proposition 10.1.8 with 𝑌 an open subset of 𝑋 and 𝑓 the natural
−1
injection 𝜌𝑌 : 𝑌 ↩→ 𝑋. In this situation, 𝑓 F is the sheaf on 𝑌 whose stalk at each
𝑦 ∈ 𝑌 is F 𝑦 ; we denote it by F|𝑌 : we get the homomorphism of Abelian groups
𝜌𝑌∗
(10.1.10) with 𝑞 = 0: 𝐻 0 (𝑋; F) −→ 𝐻 0 (𝑌 ; F|𝑌 ). If there is no risk of confusion we
write 𝐻 0 (𝑌 ; F) rather than 𝐻 0 (𝑌 ; F|𝑌 ).
10.3 Relative Sheaf Cohomology 313

Let 𝑍 = 𝑋\𝑌 ; the kernel F 𝑍 of the restriction map F −→ F|𝑌 is the sheaf on 𝑋
with stalk F 𝑥 at 𝑥 ∈ 𝑍 and {0} at 𝑥 ∈ 𝑌 . Its continuous sections have their support
in 𝑍: they vanish identically in 𝑌 . The short exact sequence
𝜄
0 −→ F 𝑍 −→ F −→ F|𝑌 → 0 (10.3.1)

is a particular case of (10.1.13). By Theorem 10.1.15, we get the long exact sequence
of Abelian groups

𝜄∗ 𝜌𝑌∗ 𝛿∗ 𝜄∗
0 −→ 𝐻 0 (𝑋; F 𝑍 ) −→ 𝐻 0 (𝑋; F) −→ 𝐻 0 (𝑌 ; F) −→ 𝐻 1 (𝑋; F 𝑍 ) −→ · · · (10.3.2)
𝜄∗ 𝜌𝑌∗ 𝛿∗ 𝜄∗
𝐻 𝑞 (𝑋; F 𝑍 ) −→ 𝐻 𝑞 (𝑋; F) −→ 𝐻 𝑞 (𝑌 ; F) −→ 𝐻 𝑞+1 (𝑋; F 𝑍 ) −→ · · ·.

In conformity with the terminology when F is the constant sheaf [meaning that every
𝑓 ∈ Γ (𝑆; F) is locally constant whatever 𝑆 ⊂ 𝑋] we shall refer to 𝐻 𝑞 (𝑋; F 𝑍 ) as the
𝑞 th relative cohomology group of 𝑋 mod 𝑌 valued in the sheaf F of Abelian groups
and we denote it by 𝐻 𝑞 (𝑋, 𝑌 ; F); it is often denoted by 𝐻 𝑍𝑞 (𝑋; F) and referred to
as the 𝑞 th cohomology group of 𝑋 with support in 𝑍; generally, its elements are
referred to as relative cohomology classes. We rewrite (10.3.2) as

𝜄∗ 𝜌𝑌∗ 𝛿∗ 𝜄∗
0 −→ 𝐻 0 (𝑋, 𝑌 ; F) −→ 𝐻 0 (𝑋; F) −→ 𝐻 0 (𝑌 ; F) −→ 𝐻 1 (𝑋, 𝑌 ; F) −→ · · ·
(10.3.3)
𝜄∗ 𝜌𝑌∗ 𝛿∗ 𝜄∗
𝐻 𝑞 (𝑋, 𝑌 ; F) −→ 𝐻 𝑞 (𝑋; F) −→ 𝐻 𝑞 (𝑌 ; F) −→ 𝐻 𝑞+1 (𝑋, 𝑌 ; F) −→ · · ·.

Suppose 𝑌 ′ is another open subset of 𝑋, 𝑌 ′ ⊂ 𝑌 , and 𝑍 ′ = 𝑋\𝑌 ′ ⊃ 𝑍; we have

the commutative diagram of sheaf homomorphisms

0 −→ F 𝑍 ′ −→ F −→ F|𝑌 ′ → 0
↓ ↕ ↑
0 −→ F 𝑍 −→ F −→ F|𝑌 → 0

where the left and right vertical arrows correspond to the natural restriction mappings
(the middle arrow is the identity map). If we compare (10.3.2) with the analogous
exact sequence for 𝑌 ′ substituted for 𝑌 , 𝑍 ′ for 𝑍, we get commutative diagrams of
group homomorphisms
𝜄𝑞′∗ 𝜌𝑌∗ ′ 𝛿∗
· · · −→ 𝐻 𝑞 (𝑋; F 𝑍 ′ ) −→ 𝐻 𝑞 (𝑋; F) −→ 𝐻 𝑞 (𝑌 ′; F) −→ · · ·
↓ ↕ ↑
𝜄𝑞 𝜌𝑌∗ 𝛿∗
· · · −→ 𝐻 𝑞 (𝑋; F 𝑍 ) −→ 𝐻 𝑞 (𝑋; F) −→ 𝐻 𝑞 (𝑌 ; F) −→ · · ·

whence the “natural” restriction mapping

𝐻 𝑍𝑞 ′ (𝑋; F) = 𝐻 𝑞 (𝑋; F 𝑍 ′ ) −→ 𝐻 𝑞 (𝑋; F 𝑍 ) = 𝐻 𝑍𝑞 (𝑋; F) ,

314 10 A Primer on Sheaf Cohomology

i.e.,
𝐻 𝑞 (𝑋, 𝑌 ′; F) −→ 𝐻 𝑞 (𝑋, 𝑌 ; F) . (10.3.4)
In standard Euclidean situations nothing interesting comes from relative coho-
mology. This is a consequence of the following observation:
Lemma 10.3.1 Suppose 𝐻 𝑞 (𝑋, F) = 0 for every 𝑞 = 1, ..., 𝑛, and 𝐻 𝑛 (𝑌 , F) = 0.
Then 𝐻 0 (𝑋, 𝑌 ; F) is the group of continuous sections of F over 𝑋 that vanish
identically in 𝑌 and 𝐻 𝑞 (𝑋, 𝑌 ; F) = 𝐻 𝑞−1 (𝑌 ; F) for all 𝑞 = 1, ..., 𝑛.
This follows directly by taking the hypotheses into account in (10.3.3).
When 𝑋 = R𝑛 , F = R or C (the sheaf of germs of constant real or complex
functions), or when 𝑋 = C𝑛 , F = O (𝑛) , and 𝑌 is an arbitrary (nonempty) open
subset of 𝑋, we can interpret the cohomology groups according to (10.2.11) and
(10.2.13). This directly implies 𝐻 𝑞 (𝑋, F) = 0 for every 𝑞 = 1, ..., 𝑛. The fact that
𝐻 𝑛 (𝑌 , F) = 0 is a direct consequence of Theorem 5.3.15, applied to
the Laplacian
Δ
(cf. the proof of Theorem 9.4.15). In the complex case we have 𝐻 0 C𝑛 , 𝑌 ; O (𝑛) = 0

since no 𝑓 ∈ O (C𝑛 ) = 𝐻 0 C𝑛 ; O (𝑛) can vanish identically in 𝑌 unless 𝑓 ≡ 0. In
the next subsection we discuss relative cohomologies that will be shown, later in this
book, to have important significance.

10.3.2 Relative 𝝏 -cohomology classes carried by a real domain

We apply the contents of the preceding subsection as preparation for the cohomo-
logical definition of hyperfunctions. In the remainder of this section 𝑈 shall be
a bounded domain in R𝑛 whose boundary 𝜕𝑈 is smooth. Let the open polydisk
Δ𝑅(𝑛) = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 < 𝑅, 𝑗 = 1, ..., 𝑛 contain the closure 𝑈. We propose to ap-

ply the results of the preceding subsection when 𝑋 = Δ𝑅(𝑛) \𝜕𝑈, 𝑌 = Δ𝑅(𝑛) \𝑈 (thus
𝑍 = 𝑈 ∩ 𝑋 = 𝑈 is closed in 𝑋) and F = O𝑇 (𝑛,0) Δ (𝑛) \𝜕𝑈 [cf.(10.2.12)]. To simplify
𝑅
the notation we write O𝑇 (𝑛,0) for either O𝑇 (𝑛,0) Δ (𝑛) \𝜕𝑈 or O𝑇 (𝑛,0) (𝑛)
Δ 𝑅 \𝑈
; then
𝑅
(10.2.13) and (10.3.2) yields the long exact sequence
𝜄∗ 𝜌∗
0 −→𝐻 0 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ 𝐻 0 Δ𝑅(𝑛) \𝜕𝑈; O𝑇 (𝑛,0) −→ (10.3.5)
𝛿∗ 𝜄∗
𝐻 0 Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ 𝐻 1 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ · · ·
𝜄∗ 𝜌∗
𝐻 𝑞 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ 𝐻 𝑞 Δ𝑅(𝑛) \𝜕𝑈; O𝑇 (𝑛,0) −→
𝛿∗ 𝜄∗
𝐻 𝑞 Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ 𝐻 𝑞+1 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ · · ·,

where 𝜌 ∗ = 𝜌 ∗ (𝑛) ; 𝐻 𝑞 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) are the relative cohomology
Δ 𝑅 \𝑈
groups.
10.3 Relative Sheaf Cohomology 315

We exploit the fine resolution (10.2.1) for M = 𝑋 = Δ𝑅(𝑛) \𝜕𝑈 and 𝑝 = 𝑛:

𝜄 𝜕
0 −→ O𝑇 (𝑛,0) −→ C∞
𝑋𝑇
(𝑛,0)
−→ · · · (10.3.6)
𝑋
𝜕 (𝑛,𝑞) 𝜕 (𝑛,𝑞+1) 𝜕 𝜕(𝑛,𝑛)
−→ C∞
𝑋𝑇 −→ C∞
𝑋𝑇 −→ · · · −→ C∞
𝑋𝑇 → 0.

Here Formula (10.2.13) is equivalent to

𝐻 𝑞 Δ𝑅(𝑛) \𝜕𝑈; O𝑇 (𝑛,0) 𝐻 𝑛,𝑞 Δ𝑅(𝑛) \𝜕𝑈 , (10.3.7)

keeping in mind that 𝐻 𝑛,𝑞 Δ𝑅(𝑛) \𝜕𝑈 𝐻 0,𝑞 Δ𝑅(𝑛) \𝜕𝑈 (see the end of the pre-
ceding section). Likewise, we have

𝐻 𝑞 Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) 𝐻 𝑛,𝑞 Δ𝑅(𝑛) \𝑈 . (10.3.8)

Note, in passing, that (10.3.7) is trivially true when 𝑞 = 0: both sides are equal to

O Δ𝑅(𝑛) \𝜕𝑈; Λ𝑛,0 C𝑛 O Δ𝑅(𝑛) \𝜕𝑈 ;

likewise with 𝑈 in the place of 𝜕𝑈. When 𝑛 ≥ 2 a classical result of Hartogs

([Hörmander, 1966], Theorem 2.3.2, pp. 30–31) implies

O Δ𝑅(𝑛) \𝜕𝑈 = O Δ𝑅(𝑛) \𝑈 = O Δ𝑅(𝑛) . (10.3.9)

We look at the tail end of the exact sequence (10.3.5), taking (10.3.7) and (10.3.8)
into account:
𝜌∗ 𝛿∗
𝐻 𝑛,𝑛−1 Δ𝑅(𝑛) \𝜕𝑈 −→ 𝐻 𝑛,𝑛−1 Δ𝑅(𝑛) \𝑈 −→ (10.3.10)
𝜄∗
𝐻 𝑛 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) −→ 𝐻 𝑛,𝑛 Δ𝑅(𝑛) \𝜕𝑈 → 0.

By Theorem 9.4.15, 𝐻 𝑛,𝑛 Δ𝑅(𝑛) \𝜕𝑈 = 0 whence

𝐻 𝑛 Δ𝑅(𝑛) \𝜕𝑈, Δ𝑅(𝑛) \𝑈; O𝑇 (𝑛,0) 𝐻 𝑛,𝑛−1 Δ𝑅(𝑛) \𝑈 /𝐻 𝑛,𝑛−1 Δ𝑅(𝑛) \𝜕𝑈 .
(10.3.11)
When 𝑛 ≥ 2 we can let 𝑅 go to +∞ and thus write

𝐻 𝑛 C𝑛 \𝜕𝑈, C𝑛 \𝑈; O𝑇 (𝑛,0) 𝐻 𝑛,𝑛−1 C𝑛 \𝑈 /𝐻 𝑛,𝑛−1 (C𝑛 \𝜕𝑈) . (10.3.12)

When 𝑛 = 1 the following can be said. A (1, 0)-cocycle in an open subset Ω of the
complex plane is a one-form 𝑓 (𝑧) d𝑧 with 𝑓 ∈ O (Ω) and there are no nonvanishing
(1, 0)-coboundaries in Ω. Thus 𝐻 1,0 (Ω) 𝐻 0,0 (Ω) = O (Ω) and (10.3.11) can be
316 10 A Primer on Sheaf Cohomology

rewritten as

𝐻 1 Δ𝑅(1) \𝜕𝑈, Δ𝑅(1) \𝑈; O𝑇 (1,0) O Δ𝑅(1) \𝑈 /O Δ𝑅(1) \𝜕𝑈 . (10.3.13)

For our purpose it suffices

to observe that (10.3.13) remains when 𝑅 = +∞
valid
provided we replace O C\𝑈 by its linear subspace O◦ C\𝑈 consisting of the
holomorphic functions in C\𝑈 vanishing at infinity; likewise with 𝜕𝑈 substituted
for 𝑈:

𝐻 1 C\𝜕𝑈, C\𝑈; O𝑇 (1,0) O◦ C\𝑈 /O◦ (C\𝜕𝑈) . (10.3.14)

Since 𝑈 is a closed subset of C𝑛 \𝜕𝑈 (now 𝑛 ≥ 2), the complement of the open
subset C𝑛 \𝑈, the standard notation in algebraic topology for the right-hand sides
in (10.3.12) and (10.3.14) would be 𝐻𝑈𝑛 (C𝑛 \𝜕𝑈). But since there is no ambiguity

about what the ambient space and the coefficient sheaf are we may as well follow
[Morimoto, 1973] and adopt the further abbreviated notation for the relative 𝜕
cohomology supported in 𝑈,

𝐻 𝑛 [𝑈] = 𝐻 𝑛 C𝑛 \𝜕𝑈, C𝑛 \𝑈; O𝑇 (𝑛,0) . (10.3.15)

10.3.3 Hyperfunctions in Euclidean space as relative cohomology

classes

We are going to exploit Theorem 6.3.10 when 𝑛 ≥ 2 (Theorem 6.3.1 when 𝑛 = 1).
Let 𝑈 ⊂ R𝑛 be as in the preceding subsection and suppose 𝑛 ≥ 2. We apply (6.3.10)
twice: to 𝐾 = 𝑈 and to 𝐾 = 𝜕𝑈. We get the two isomorphisms (the left-hand sides
being spaces of analytic functionals)

O ′ (𝜕𝑈) 𝐻 𝑛,𝑛−1 (C𝑛 \𝜕𝑈) , (10.3.16)

O ′ 𝑈 𝐻 𝑛,𝑛−1 C𝑛 \𝑈 , (10.3.17)

allowing us [cf. (7.1.1)] to identify each hyperfunction in 𝑈 with a relative cohomol-

ogy class in C𝑛 \𝜕𝑈 mod C𝑛 \𝑈 (i.e., a cohomology class in C𝑛 \𝜕𝑈 with support in
𝑈): .
B (𝑈) 𝐻 𝑛,𝑛−1 C𝑛 \𝑈 𝐻 𝑛,𝑛−1 (C𝑛 \𝜕𝑈) 𝐻 𝑛 [𝑈] (10.3.18)

[cf. (10.3.15)].
In the case 𝑛 = 1, in accordance with Theorem 6.3.1 we replace (10.3.16)–
(10.3.17) by
10.3 Relative Sheaf Cohomology 317

O ′ (𝜕𝑈) O◦ (C\𝜕𝑈) , (10.3.19)

O ′ 𝑈 O◦ C\𝑈 , (10.3.20)

where the right-hand sides are the spaces of holomorphic functions in C\𝜕𝑈 and
1
at infinity. As for (10.3.18) it remains valid if we equate 𝐻 [𝑈]
C\𝑈 vanishing

to O◦ C\𝑈 /O◦ (C𝑛 \𝜕𝑈). Keep in mind that 𝑈 is a closed subset of C𝑛 \𝜕𝑈, its
complement being C\𝑈.
Let 𝑈1 ⊂ 𝑈 be another open subset of R𝑛 with a C ∞ boundary 𝜕𝑈1 in R𝑛 , 𝑛 ≥ 2.
On the one hand we have the natural “restriction” map

O ′ 𝑈1 𝐻 𝑛,𝑛−1 C𝑛 \𝑈1 −→ 𝐻 𝑛,𝑛−1 C𝑛 \𝑈 O ′ 𝑈

which is injective since both 𝑈 and 𝑈1 are Runge. On the other hand Theorem 6.3.11
yields (see the proof of Proposition 7.1.3) a surjective map

𝐻 𝑛,𝑛−1 C𝑛 \𝑈 −→ 𝐻 𝑛,𝑛−1 C𝑛 \𝑈1 /𝐻 𝑛,𝑛−1 (C𝑛 \𝜕𝑈1 ) (10.3.21)

whence a surjective map

𝐻 𝑛,𝑛−1 C𝑛 \𝑈 /𝐻 𝑛,𝑛−1 (C𝑛 \𝜕𝑈) −→ 𝐻 𝑛,𝑛−1 C𝑛 \𝑈1 /𝐻 𝑛,𝑛−1 (C𝑛 \𝜕𝑈1 ) .

This remains true when 𝑛 = 1 if we replace 𝐻 1,0 by O◦ (cf. Theorem 6.3.1). In all
dimensions it is a restatement of the fact that the sheaf of hyperfunctions in R𝑛 is
flabby (Proposition 7.1.9).

10.3.4 Microfunctions in Euclidean space as relative cohomology

classes

In this subsection we follow an approach to the sheaf Bmicro (R𝑛 ) slightly different
from that of Ch. 7. We will rely on important concepts and results of Several
Complex Variables (SCV) theory, for which we refer to [Gunning and Rossi, 1965],
[Hörmander, 1966]. The first concept is that of a class of domains in C𝑛 , crucial in
SCV theory.

Definition 10.3.2 A domain Ω in C𝑛 is called a domain of holomorphy if, to every

point 𝑧◦ ∈ 𝜕Ω and neighborhood 𝑈 of 𝑧◦ , there are a neighborhood of 𝑧◦ , 𝑉 ⊂ 𝑈,
and a function ℎ ∈ O (Ω ∩ 𝑉) that cannot be extended as a holomorphic function in
a neighborhood of 𝑧 ◦ in C𝑛 .

To every domain Ω in C𝑛 there is a smallest domain of holomorphy Ω

b containing
Ω, called the envelope of holomorphy of Ω (see [Hörmander, 1966], pp. 137 et
seq.).
318 10 A Primer on Sheaf Cohomology

Theorem 10.3.3 Every Runge domain (Definition 6.1.5) is a domain of holomorphy.

When 𝑛 = 1 every open subset Ω of the plane is a domain of holomorphy since

the function (𝑧 − 𝑧 ◦ ) −1 ∈ O (Ω), 𝑧◦ ∈ 𝜕Ω, cannot be extended to a neighborhood of
𝑧◦ . Note that only the simply connected ones have the Runge property.
All convex domains in C𝑛 are domains of holomorphy. They are also Runge
domains, by Theorem 6.2.11 and Proposition 6.2.12.
Section 11.2 of the next chapter provides another interpretation of domains of
holomorphy: they are shown to be identical to pseudoconvex domains (Definition
11.2.15). We state here a corollary of one of the fundamental theorems of SCV
theory (cf. Theorem 11.2.21).

Theorem 10.3.4
If adomain Ω in C𝑛 is a domain of holomorphy then the cohomology
groups 𝐻 Ω; O (𝑛) vanish for all 𝑝 ≥ 1.
𝑝

Throughout this subsection 𝑧 = 𝑥 + 𝑖𝑦. Recall that

W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) = {𝑧 ∈ C𝑛 ; 𝑥 ∈ 𝑈, 𝑦 · 𝜃 ◦ > 𝜀 |𝑦| , |𝑦| < 𝛿} ,

now with 𝑈 a convex neighborhood of 𝑥 ◦ in R𝑛 , 𝜃 ◦ ∈ S𝑛−1 , 𝛿 > 0, 0 < 𝜀 < 1; thus

W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) itself is convex. We denote by W
c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) the envelope of
holomorphy of

W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) ∪ {𝑧 ∈ C𝑛 ; |𝑧 − 𝑥 ◦ | < 𝑟, 𝑦 · 𝜃 ◦ ≤ 0} (10.3.22)

with 𝑟 > 0 such that 𝔅𝑟 (𝑥 ◦ ) = {𝑥 ∈ R𝑛 ; |𝑥 − 𝑥 ◦ | < 𝑟} ⊂ 𝑈; (10.3.22) is an open

subset of C𝑛 .

Proposition 10.3.5 We have

n o
W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) = 𝑧 ∈ W
c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; 𝑦 · 𝜃 ◦ > 0 . (10.3.23)

Moreover, 𝑈𝑟 = W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ∩ R𝑛 is a closed subset of W

c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ ))
◦
and 𝔅𝑟 (𝑥 ) ⊂ 𝑈𝑟 .

Proof The convex hull ℭ 𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) of (10.3.22) contains W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ ))
◦ ◦
since ℭ 𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 )) is a domain of holomorphy. Since 𝔅𝑟 (𝑥 ) ⊂ 𝑈 we have

W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) = 𝑧 ∈ ℭ 𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; 𝑦 · 𝜃 ◦ > 0

implying
n o
c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; 𝑦 · 𝜃 ◦ > 0 ⊂ W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) .
𝑧∈W

Since the converse inclusion is trivial this proves (10.3.23). We have

𝑈𝑟 ⊂ ℭ 𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ∩ R𝑛 = 𝑈;
10.3 Relative Sheaf Cohomology 319

c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) since 𝑈 is closed in ℭ 𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )).

𝑈𝑟 is closed in W □

Proposition 10.3.6 If 0 < 𝛿 ′ < 𝛿 we have

n o n o
𝑧∈Wc𝛿′ ,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; 𝑦 · 𝜃 ◦ < 0 = 𝑧 ∈ W
c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; 𝑦 · 𝜃 ◦ < 0 .
(10.3.24)

Proof The left-hand side is obviously contained in the right-hand side. The union
n o
W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) ∪ 𝑧 ∈ Wc𝛿′ ,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; 𝑦 · 𝜃 ◦ ≤ 0 (10.3.25)

is a domain of holomorphy since the boundary points of (10.3.25) that do not

belong to the boundary of W c𝛿′ ,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) are points 𝑥 + 𝑖𝑦 ∈ W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))
′
such that |𝑦| ≥ 𝛿 . Since (10.3.22) is contained in (10.3.25) the latter must contain
Wc𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )). □

Example 10.3.7 When 𝑛 = 1 we have 𝑈 = (𝑥 ◦ − 𝑎, 𝑥 ◦ + 𝑏), 𝑎 > 0, 𝑏 > 0, 𝜃 ◦ = 1 or

−1. Suppose 𝑥 ◦ = 0, 𝜃 ◦ = 1, 𝑟 < min (𝑎, 𝑏); then

W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) = (−𝑎, 𝑏) × (0, 𝛿) ,

c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) = (𝑈 × (0, 𝛿)) ∪ {𝑧 ∈ C; |𝑧| < 𝑟, 𝑦 ≤ 0} ,
W

and therefore
𝑈𝑟 = {𝑥 ∈ R; |𝑥| < 𝑟} ≠ 𝑈.

Proposition 10.3.8 There is an 𝑟 > 0 such that the following properties of an

arbitrary function ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) are equivalent:
(1) (𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a (𝑏𝑈 ℎ);
(2) ℎ ∈ O W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) .

Proof By Proposition 7.5.7, (1) implies that to each ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )))
there is an 𝑟 > 0 such that ℎ is holomorphic in (10.3.22), which entails (2). Since
O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) is a Fréchet space the standard Baire category argument shows
that 𝑟 can be selected independently of ℎ. Property (2) implies that ℎ|𝑈𝑟 is the
boundary value of a holomorphic function in the set {𝑧 ∈ C𝑛 ; |𝑧| < 𝑟, 𝑦 · 𝜃 ◦ < 0},
whence (1) by Definition 7.4.7. □

Corollary 10.3.9 The existence of 𝑟 > 0 in Proposition 10.3.8 is equivalent to the

holomorphic extendibility of ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) to a neighborhood of 𝑥 ◦ in
C𝑛 .

Proof Combine Propositions 7.5.7, 10.3.8. □

320 10 A Primer on Sheaf Cohomology

Given the definition (7.5.16), Corollary 10.3.9 directly implies

Corollary 10.3.10 With 𝑟 as in Proposition 10.3.8 there is a natural isomorphism

.
O 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) O Wc𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) .
(10.3.26)

Remark 10.3.11 Since W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) is convex the restrictions of entire functions

form a dense subspace of O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))). This implies that the denominator
in the right-hand side of (10.3.26) is dense in the numerator, which in turn implies
that there is no reasonable Hausdorff topology (compatible with the vector space
structure) on this quotient.

We now look at the long sequence (10.3.3) taking 𝑋 = W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )), 𝑌 =
W𝛿(𝑈, Γ 𝜀 (𝜃 ◦ (𝑛)
)), F =
O . Both
𝑋 and 𝑌 are domains of holomorphy in C , whence
𝑛

𝐻 𝑞 𝑋; O (𝑛) = 𝐻 𝑞 𝑋; O (𝑛) = 0 for all 𝑞 ≥ 1 (Theorem 10.3.4); (10.3.3) reduces

to
𝜄∗ 𝜌∗
0 −→𝐻 0 𝑋, 𝑌 ; O (𝑛) −→ 𝐻 0 𝑋; O (𝑛) −→
𝑌

𝛿∗
𝐻 0 𝑌 ; O (𝑛) −→ 𝐻 1 𝑋, 𝑌 ; O (𝑛) → 0.

Also recall that 𝐻 0 𝑋, 𝑌 ; O (𝑛) is the space of continuous sections of O (𝑛) over
𝑋 vanishing identically in 𝑌 ; with our choice of 𝑋 and 𝑌 these sections vanish
identically in 𝑋. It follows that the short sequence reduces to
𝜌∗ 𝛿∗
𝐻 0 𝑋; O (𝑛) −→ 𝐻 0 𝑌 ; O (𝑛) −→ 𝐻 1 𝑋, 𝑌 ; O (𝑛) → 0
𝑌
(10.3.27)

We derive directly

𝐻1 W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) , W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛) (10.3.28)
.
O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) O W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) .

Taking 𝑟 such that the isomorphism (10.3.26) holds yields the vector space iso-
morphism

𝐻1 W c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) , W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛) O +◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) .
𝑥
(10.3.29)
The left-hand side in (10.3.29) is the first relative cohomology group of
Wc𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) with values in the sheaf O (𝑛) and support in the closed sub-
set 𝑈𝑟 (Proposition 10.3.5), also denoted by

𝐻𝑈1 c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛) .
W
𝑟
10.3 Relative Sheaf Cohomology 321

If 𝛿 ′ < 𝛿, 𝑟 ′ < 𝑟, 𝜀 ′ ∈ (𝜀, 1), and if 𝑈 ′ is such that 𝔅𝑟(𝑛) ◦ ′

′ (𝑥 ) ∩ R ⊂ 𝑈 ⊂ 𝑈, then
𝑛

combining (10.3.29) with (7.5.17) yields the linear injection

1
𝐻𝑈 Wc𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛) −→ 𝐻 1 ′ W c𝛿′ ,𝑟 ′ (𝑈 ′, Γ 𝜀′ (𝜃 ◦ )) ; O (𝑛) ,
𝑟 𝑈 𝑟′
(10.3.30)
allowing us to regard the left-hand side as a vector subspace of the right-hand side.
Combining (10.3.29) with the definition (7.5.18) yields the isomorphism

O+(𝑥 ◦ , 𝜃 ◦ ) lim𝐻𝑈
1 c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛)
W (10.3.31)
−→ 𝑟

where the right-hand side is the direct limit of the groups

𝐻𝑈 1 c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛)
W
𝑟

as 𝑈 ranges over the set of convex neighborhoods of 𝑥 ◦ , 𝑟 ↘ 0 accordingly [such

that the isomorphism (10.3.26) holds], 𝛿 ↘ 0, 0 < 𝜀 ↗ 1.
Now let us keep 𝑈 and 𝜀 fixed. In Proposition 10.3.8 the upper bound of the
numbers 𝑟 such that the isomorphism (10.3.26) holds might depend on (𝑥 ◦ , 𝜃 ◦ ); but
◦
valid whatever (𝑥,𝜃) ∈ 𝑈 × Γ 𝜀 (𝜃 ). It follows
if we let 𝑟 ↘ 0 the conclusion will be
that the direct limit of the spaces O W ◦
c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 )) as 𝑟 ↘ 0 can be identified
with the space of functions ℎ ∈ O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) such that ℎ ∈ ker 𝑏𝑈 [𝑏𝑈 :
the singularity hyperfunctions boundary value, acting on O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) ].
Combining this with (10.3.28) yields

lim 𝐻𝑈1 c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛) O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) /ker 𝑏𝑈 . (10.3.32)
W
𝑟→0 𝑟

If then we take the direct limit as 𝛿 ↘ 0 we obtain [cf. Proposition 10.3.6 and
(7.5.19)]

1
lim 𝐻𝑈 c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )) ; O (𝑛) O
W e+ (𝑈 × Γ 𝜀 (𝜃 ◦ )) . (10.3.33)
𝛿,𝑟→0 𝑟

The sets 𝑈 × Γ 𝜀 (𝜃 ◦ ) ∩ S𝑛−1 form a basis of the topology of R𝑛 × S𝑛−1 . We reach

the following conclusion: with the mappings (10.3.30) the direct limits in the left-
hand side of (10.3.33) define a presheaf on R𝑛 × S𝑛−1 . The stalks of the associated
sheaf H1rel R𝑛 × S𝑛−1 are the right-hand sides in (10.3.31); in other words, there is
a natural isomorphism

H1rel R𝑛 × S𝑛−1 O+ R𝑛 × S𝑛−1 (10.3.34)

(cf. Proposition 7.5.8). The connection with the introduction of microfunctions in

Section 7.6 is clear:
322 10 A Primer on Sheaf Cohomology

H1rel R𝑛 × S −→ Bmicro R𝑛 × S𝑛−1

𝑛−1

↕ ↑ (10.3.35)
b+
O+ R𝑛 × S𝑛−1 −→ Bsing (R𝑛 )

where the left ↕ is the isomorphism (10.3.34), b+ is the sheaf homomorphism (7.5.24),
and the right ↑ is the sheaf homomorphism Bsing −→ Bmicro [see (7.6.2)]. Theorem
7.6.7 implies

Theorem 10.3.12 The upper horizontal arrow in (10.3.35) is a sheaf of vector spaces
isomorphism.

10.4 Edge of the Wedge in (Co)homological Terms

It is clear that the theorem of the Edge of the Wedge (Theorem 7.5.1) must have
a homological meaning: certain kinds of cycles must be boundaries if they lie in
the kernel of the operator 𝑏𝑈 . Rather than introducing sheaf-valued chains and
simplicial homology we avail ourselves of a simple transformation of a (special)
simplicial complex into a (very) simple subcomplex of the De Rham differential
complex (with coefficients in the appropriate sheaf) over a real Euclidean space.
This artifice allows us to reformulate the Edge of the Wedge in cohomological
terms, sort of. We refer the reader to Condition (Z), Theorem 7.5.1, and to the 𝜈
functions ℎ 𝑗 ∈ O (W𝛿 Ω, Γ 𝑗 ) ( 𝑗 = 1, ..., 𝜈, 𝜈 ≥ 2, 𝛿 > 0); 𝑈 ⊂⊂ Ω is a domain in
R𝑛 and Γ1 , ..., Γ𝜈 are convex open cones in R𝑛 \ {0}.
We introduce a set 𝜉 = (𝜉1 , 𝜉2 , ..., 𝜉 𝜈 ) of 𝜈 variables, say real. For 1 ≤ 𝑞 ≤
𝜈 we denote by Λ𝑞 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) the complex vector space of 𝑞-forms 𝑔 =
Í c
1 < · · · < 𝑖 𝑞 ≤ 𝜈, 𝐼 = (1, ..., 𝜈) \𝐼
|𝐼 |=𝑞 𝑔 𝐼 c 𝜉 𝐼 c d𝜉 𝐼 where 𝐼 = 𝑖 1 , ..., 𝑖 𝑞 , 1 ≤ 𝑖Í
ordered, 𝜉 𝐼 = 𝜉𝑖1 · · · 𝜉𝑖𝑞 , 𝑔 𝐼 (𝑧) ∈ O W𝛿 𝑈, 𝑖 𝛼 ∈𝐼 Γ𝑖 𝛼 (very important: 𝑔 𝐼 does
not depend on 𝜉). We define

𝜈
© ∑︁ ªª
Λ0 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) = 𝜉1 · · · 𝜉 𝜈 O ­W𝛿 ­𝑈, Γ 𝑗 ®® .
©

« « 𝑗=1 ¬¬
We restrict the exterior derivative d 𝜉 in 𝜉-space R𝜈 to the forms

𝑔 ∈ Λ𝑞 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) ;

since d 𝜉 Λ𝑞 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) ⊂ Λ𝑞+1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) we get a differential com-

(∗)
plex Λ A 𝛿 ,
d𝜉 d𝜉
0 −→ Λ0 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ Λ1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ (10.4.1)
d𝜉 d𝜉 d𝜉
· · · −→ Λ𝜈−1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ Λ𝜈 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ 0.
10.4 Edge of the Wedge in (Co)homological Terms 323

Let us provisionally use the notation

𝐼Í ..., 𝜈}. If 𝑔 ∈ Λ0 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 )
◦ = {1,
𝜈
then 𝑔 = 𝜉 𝐼◦ 𝑔 𝐼◦ with 𝑔 𝐼◦ ∈ O W𝛿 𝑈, 𝑗=1 Γ 𝑗 and

𝜈
∑︁
d 𝜉 𝑔 = 𝑔 𝐼◦ (𝑧) 𝜉 𝐼◦ /𝜉 𝑗 d𝜉 𝑗 .
𝑗=1

If, instead,
𝜈
∑︁
𝑔 𝐼◦ \{ 𝑗 } (𝑧) 𝜉 𝐼◦ /𝜉 𝑗 d𝜉 𝑗 ∈ Λ1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) ,

𝑔 (𝑧, 𝜉, d𝜉) =
𝑗=1
Í
with 𝑔 𝐼◦ \{ 𝑗 } ∈ O W𝛿 𝑈, 𝑘=1,...,𝜈,𝑘≠ 𝑗 Γ𝑘 , then

𝜈
∑︁ 𝜉 𝐼◦
d 𝜉 𝑔 (𝑧, 𝜉, d𝜉) = 𝑔 𝐼◦ \{ 𝑗 } (𝑧) d𝜉 𝑘 ∧ d𝜉 𝑗 = 0
𝑗,𝑘=1
𝜉 𝑗 𝜉𝑘
Í
requires 𝑔 𝐼◦ \{ 𝑗 } = 𝑔 𝐼◦ \{𝑘 } in W𝛿 𝑈, ℓ=1,...,𝜈,ℓ≠ 𝑗,ℓ≠𝑘 Γ𝑘 for all 𝑗, 𝑘, meaning that
Í Í
there is a 𝑔 𝐼◦ ∈ O W𝛿 𝑈, 𝜈𝑗=1 Γ 𝑗 whose restriction to W𝛿 𝑈, ℓ=1,...,𝜈,ℓ≠ 𝑗 Γℓ
is equal to 𝑔 𝐼◦ \{ 𝑗 } for each 𝑗 = 1, ..., 𝜈. This shows that 𝐻 1 (Λ∗ A 𝛿 ) ≠ 0.
We also introduce the “parallel” differential complex Λ∗ B sing (𝑈),
d𝜉
0 −→ Λ0 B sing (𝑈) −→ (10.4.2)
d𝜉 d𝜉 d𝜉 d𝜉
Λ1 B sing (𝑈) −→ · · · −→ Λ𝜈−1 B sing (𝑈) −→ Λ𝜈 B sing (𝑈) −→ 0,

where Λ𝑞 B sing (𝑈) is the complex vector space of 𝑞-forms 𝑓 = |𝐼 |=𝑞 𝑓 𝐼 c 𝜉 𝐼 c d𝜉 𝐼 ,

Í

𝑓 𝐼 c ∈ B sing (𝑈). We have the homomorphism from (10.4.1) to (10.4.2) defined by

the maps
∑︁ 𝑏𝑈 ∑︁
Λ𝑞 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) ∋ 𝑔 𝐼 c 𝜉 𝐼 c d𝜉 𝐼 ↦→ (𝑏𝑈 𝑔 𝐼 c ) 𝜉 𝐼 c d𝜉 𝐼 ∈ Λ𝑞 B sing (𝑈) .
|𝐼 |=𝑞 |𝐼 |=𝑞

We pay particular attention to the tail-end of the sequences (10.4.1), (10.4.2) and to
the following commutative diagram:
d𝜉 d𝜉
Λ𝜈−2 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ Λ𝜈−1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ Λ𝜈 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ 0
↓ 𝑏𝑈 ↓ 𝑏𝑈 ↘ ↓ 𝑏𝑈
d𝜉 d𝜉
Λ𝜈−2 B sing (𝑈) −→ Λ𝜈−1 B sing (𝑈) −→ Λ𝜈 B sing (𝑈) −→ 0.
Í (10.4.3)
Note that if ℎ 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 then ℎ = 𝜈𝑗=1 ℎ 𝑗 𝜉 𝑗 d𝜉1 ∧ · · · ∧ d𝜉 c𝑗 ∧ · · · ∧ d𝜉 𝜈
(with the hatted factor omitted) belongs to Λ𝜈−1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) and
324 10 A Primer on Sheaf Cohomology

𝜈
©∑︁ ª
d𝜉 ℎ = ­ ℎ 𝑗 ® d𝜉1 ∧ · · · ∧ d𝜉 𝜈 .
« 𝑗=1 ¬

If 𝑢 𝑗,𝑘 ∈ O W𝛿 𝑈, Γ 𝑗 + Γ𝑘 then
∑︁
𝑢= 𝑢 𝑗,𝑘 𝜉 𝑗 𝜉 𝑘 d𝜉1 ∧ · · · ∧ d𝜉 𝑗−1 ∧ d𝜉 𝑗+1 ∧ · · · ∧ d𝜉 𝑘−1 ∧ d𝜉 𝑘+1 ∧ · · · ∧ d𝜉 𝜈
1≤ 𝑗<𝑘 ≤𝜈
∑︁
+ 𝑢 𝑗,𝑘 𝜉 𝑗 𝜉 𝑘 d𝜉1 ∧ · · · ∧ d𝜉 𝑘−1 ∧ d𝜉 𝑘+1 ∧ · · · ∧ d𝜉 𝑗−1 ∧ d𝜉 𝑗+1 ∧ · · · ∧ d𝜉 𝜈
1≤𝑘< 𝑗 ≤𝜈

belongs to Λ𝜈−2 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ). We have

∑︁
d𝜉 𝑢 = (−1) 𝑗−1 𝑢 𝑗,𝑘 + 𝑢 𝑘, 𝑗 𝜉 𝑘 d𝜉1 ∧ · · · ∧ d𝜉 𝑘−1 ∧ d𝜉 𝑘+1 ∧ · · · ∧ d𝜉 𝜈
1≤ 𝑗<𝑘 ≤𝜈
∑︁
− (−1) 𝑗−1 𝑢 𝑗,𝑘 + 𝑢 𝑘, 𝑗 𝜉 𝑘 d𝜉1 ∧ · · · ∧ d𝜉 𝑘−1 ∧ d𝜉 𝑘+1 ∧ · · · ∧ d𝜉 𝜈 .
1≤𝑘< 𝑗 ≤𝜈

Í𝜈
We see that d 𝜉 𝑢 = 𝑗=1 ℎ 𝑗 𝜉 𝑗 d𝜉1 ∧ · · · ∧ d𝜉
c𝑗 ∧ · · · ∧ d𝜉 𝜈 means

© ∑︁ ∑︁ ª
ℎ 𝑗 = (−1) 𝑗 ­ 𝑢 𝑗,𝑘 + 𝑢 𝑘, 𝑗 − 𝑢 𝑗,𝑘 + 𝑢 𝑘, 𝑗 ® , 𝑗 = 1, ..., 𝜈. (10.4.4)
«1≤𝑘< 𝑗 𝑗<𝑘 ≤𝜈 ¬
If we define

𝑔 𝑗,𝑘 = (−1) 𝑗 𝑢 𝑗,𝑘 + 𝑢 𝑘, 𝑗 if 1 ≤ 𝑘 < 𝑗,

𝑔 𝑗,𝑘 = (−1) 𝑗−1 𝑢 𝑗,𝑘 + 𝑢 𝑘, 𝑗 if 𝑗 < 𝑘 ≤ 𝜈,
Í
(10.4.4) can be rewritten as ℎ 𝑗 = 𝜈𝑘=1 𝑔 𝑗,𝑘 with 𝑔 𝑗,𝑘 = −𝑔 𝑘, 𝑗 ∈ O W𝛿 𝑈, Γ 𝑗 + Γ𝑘 .
This leads to the following reformulation of Theorem 7.5.1.
Theorem 10.4.1 Let Γ ′𝑗 ( 𝑗 = 1, ..., 𝜈, 𝜈 ≥ 2) be convex open cones in R𝑛 \ {0} such
that ∅ ≠ Γ ′𝑗 ∩ S𝑛−1 ⊂⊂ Γ 𝑗 for every 𝑗. Given 𝑥 ◦ ∈ 𝑈, the neighborhood 𝑈 ′ ⊂ 𝑈
of 𝑥 ◦ in R𝑛 and 𝛿 ′ ∈ (0, 𝛿) can be selected so that, in the following commutative
diagram,
Λ𝜈 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 )
d𝜉 ↗ 𝑏𝑈 ↓
𝑏𝑈 ◦d 𝜉
Λ𝜈−1 A 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) −→ Λ𝜈 B sing (𝑈)
↓𝑟
d𝜉
𝑈 , Γ1 , ..., Γ𝜈 −→ Λ A 𝛿′ 𝑈 ′ , Γ1′ , ..., Γ𝜈′
′ ′ ′

Λ𝜈−2 A 𝛿′ 𝜈−1

where 𝑟 is the map induced by restriction from

W𝛿 (𝑈, Γ1 , ..., Γ𝜈 ) to W𝛿′ 𝑈 ′, Γ1′ , ..., Γ𝜈′ ,

10.4 Edge of the Wedge in (Co)homological Terms 325

we have
𝑟 ker 𝑏𝑈 ◦ d 𝜉 = d 𝜉 Λ𝜈−2 A 𝛿′ 𝑈 ′, Γ1′ , ..., Γ𝜈′ ∩ Im 𝑟.

To rid ourselves of the restriction mappings we take advantage of the commutation

between exact sequences and direct limits (see [Godement, 1964], pp. 9–11). We
let 𝑈 ′ range over a basis of neighborhoods of 𝑥 ◦ and 𝛿 ′ ↘ 0; for each 𝑗n = 1,o..., 𝜈,
we select a fixed vector 𝜃 ◦( 𝑗) ∈ Γ ′𝑗 ∩ S𝑛−1 and let Γ ′𝑗 ∩ S𝑛−1 contract to 𝜃 ◦( 𝑗) . We

denote by 𝜃 ◦ the set of unit vectors 𝜃 ◦(1) , ..., 𝜃 ◦(𝜈) and by Γ (𝜃 ◦ ) the set of vectors
Í𝜈 ◦ ◦ )∩S𝑛−1 is closed. We define Λ𝑞 A (𝑥 ◦ , 𝜃 ◦ ) as the
𝑗=1 𝜆 𝑗 𝜃 ( 𝑗) , 𝜆 𝑗 > 0 for every 𝑗; Γ (𝜃 Í
complex vector space of 𝑞-forms g = |𝐼 |=𝑞 g𝐼 c 𝜉 𝐼 c d𝜉𝐼 with germ coefficients g𝐼 c and
𝐼 ranging over the set of multi-indices 𝐼 = 𝑖1 , ..., 𝑖 𝑞 , 1 ≤ 𝑖1 < · · · <𝑖 𝑞 ≤ 𝜈. More
precisely, g𝐼 c is the germ at {𝑥 ◦ } + 𝑖Γ (𝜃 ◦ ) of a function ℎ 𝐼 c ∈ O W𝛿 𝑈, 𝑗 ∈𝐼 Γ 𝑗
Í

where 𝑈 is a neighborhood of 𝑥 ◦ and the Γ 𝑗 are convex open cones in R𝑛 \ {0},

Γ 𝑗 ∋ 𝜃 ◦( 𝑗) . The interesting cases are those where Γ (𝜃 ◦ ) = R𝑛 . Theorem 10.4.1 leads
to the following

Theorem 10.4.2 Given 𝑥 ◦ ∈ R𝑛 , 𝜃 ◦( 𝑗) ∈ S𝑛−1 , 𝑗 = 1, ..., 𝜈, the lower row in the

commutative diagram

Λ𝜈 A (𝑥 ◦ , 𝜃 ◦ )
d𝜉 ↗ b↓
d𝜉 sing
Λ𝜈−2 A (𝑥 ◦ , 𝜃 ◦ ) −→ Λ𝜈−1 A (𝑥 ◦ , 𝜃 ◦ ) −→ Λ𝜈 B 𝑥 ◦

is an exact sequence.

The operators d 𝜉 and b reflect (on germs) the action of d 𝜉 and the boundary
value maps 𝑏𝑈 on their representatives.

Corollary 10.4.3 We have

𝐻 𝜈−1 (A (𝑥 ◦ , 𝜃 ◦ )) ker d 𝜉 /ker b ◦ d 𝜉 .

Chapter 11
Distributions and Hyperfunctions on a Manifold

This chapter aims at providing the conceptual basis of distribution (resp., hyperfunc-
tion) theory on a smooth (resp., real-analytic) manifold. An important aspect of how
we define distributions, often called currents after the pioneering text [De Rham,
1955], on a C ∞ manifold is the distinction made necessary by the orientation, or
lack thereof, of the manifold. This reflects the basic fact of a change of variables
in an integral: when carried out, does the Jacobian determinant of the change of
variables or its absolute value enter? This choice determines how we want to in-
terpret a locally integrable function as a distribution in the domain of local (C ∞ )
coordinates. When the absolute value of the Jacobian determinant enters we shall
talk of density-distributions; when it is the Jacobian determinant itself, we talk of
an 𝑛-current, or of a distribution 𝑛-form. The latter raises the question of defining
distributions, or density-distributions, sections of a C ∞ vector bundle: the natural
approach is to use local coordinates and this, of course, raises the question of the
representation of the object in the overlap of local charts and the manner in which
the Jacobian determinant in the change of variables enters the picture. All this is
dealt with in Section 11.1.
Section 11.2 is intended as an introduction to the (very important) role of plurisub-
harmonic functions in Several Complex Variables theory, in particular their relation
to pseudoconvex domains and the global exactness of the 𝜕¯ differential complex
(introduced in the last section of Ch. 9; see also Section 10.2). The properties of
plurisubharmonic functions and of pseudoconvex domains are listed without proofs;
the latter can all be found in [Hörmander, 1966]; the precise references to Hörman-
der’s textbook are provided in each instance. The motivation for such a listing is
that the concepts and results introduced here will be exploited (and referenced) at
various points in later parts of the book, and the reader might want to be apprised of
the statements (of definitions or results) within the confines of the book. In the last
subsection of Section 11.2 we take a look at real quadratic forms from the standpoint
of plurisubharmonicity and provide the (elementary) proofs of the claims.
Section 11.3 defines hyperfunctions in a C 𝜔 manifold M and their microlo-
cal version, microfunctions, starting from their local definitions, transferred from
Euclidean space, in C 𝜔 coordinates charts. Microfunctions are best defined in the

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 327
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_11
328 11 Distributions and Hyperfunctions on a Manifold

cosphere bundle of M. Here we may be talking of any of the diverse definitions

of hyperfunctions (and, afterward, of microfunctions) introduced in the preceding
chapters, be they the definitions through analytic functionals or as sums of boundary
values of holomorphic functions (both in some, local or global, complexification of
M) or those based on the concepts and results of Sheaf Cohomology, as sketched
in the preceding chapter. In this context we limit our attention to orientable C 𝜔
manifolds. Extension to nonorientable manifolds would require tensoring the vari-
ous intervening sheaves with the orientation sheaf of M, thereby complicating the
notation, in a text that does not pretend to be a treatise, simply an introduction.

11.1 Distributions and Currents on a Manifold

11.1.1 Densities and distribution-densities

Unless specified otherwise M shall denote a C ∞ manifold, countable at infinity. The

main problem with selecting the right concept of a distribution on a manifold is that
of deciding how it will be affected by coordinate changes or, if one prefers, what is
the effect of a diffeomorphism.
First consider an open subset Ω of R𝑛 and let d𝑥 = d𝑥1 · · · d𝑥 𝑛 be the Lebesgue
measure in R𝑛 . Let 𝐹 be a diffeomorphism of Ω onto another open subset Ω′ of R𝑛 .
The pullback under 𝐹 defines an isomorphism 𝐹 ∗ : 𝜑 −→ 𝜑 ◦ 𝐹 of Cc∞ (Ω′) onto
Cc∞ (Ω) and yields by transposition an isomorphism 𝐹∗ : D ′ (Ω) −→ D ′ (Ω′), a
pushforward map: ⟨𝐹∗ 𝑢, 𝜑⟩ = ⟨𝑢, 𝜑 ◦ 𝐹⟩. If 𝑢 is a locally integrable function in Ω
and 𝜑 ∈ Cc∞ (Ω′) then
∫ ∫
−1
−1
𝜕𝐹 −1
⟨𝐹∗ 𝑢, 𝜑⟩ = 𝜑 (𝐹 (𝑥)) 𝑢 (𝑥) d𝑥 = 𝑢 𝐹 (𝑦) 𝜑 (𝑦) det 𝐹 (𝑦) d𝑦,
𝜕𝑥
𝜕𝐹
where 𝜕𝑥 is the differential of the map 𝐹, in coordinates the Jacobian matrix. Thus,

−1
−1
𝜕𝐹 −1
𝐹∗ 𝑢 (𝑦) = 𝑢 𝐹 (𝑦) det 𝐹 (𝑦) . (11.1.1)
𝜕𝑥

We may, however, slightly modify the meaning of a distribution in Ω. Sup-

pose we regard a distribution in Ω as a continuous linear functional not on the
space of test-functions but on the space of test-densities 𝜑 (𝑥) d𝑥, 𝜑 ∈ Cc∞ (Ω). A
change of coordinates 𝑦 = 𝐹 (𝑥) transforms the density 𝜓 (𝑦) d𝑦, 𝜓 ∈ Cc∞ (Ω′), into
𝐹 ∗ 𝜓 (𝑦) d𝑦 = 𝜓 (𝐹 (𝑥)) det 𝜕𝐹
𝜕𝑥 (𝑥) d𝑥. In this case,

𝜕𝐹
⟨𝐹∗ 𝑢, 𝜓 (𝑦) d𝑦⟩ = ⟨𝑢, 𝜓 (𝐹 (𝑥)) det (𝑥) d𝑥⟩. (11.1.2)
𝜕𝑥
11.1 Distributions and Currents on a Manifold 329

1 (Ω) we see that the pushforward has now the effect that 𝐹 𝑢 (𝑦) =
If 𝑢 ∈ 𝐿 loc ∗

−1
𝑢 𝐹 (𝑦) .
To summarize: in the first instance we were dealing with density distributions,
defined as continuous linear functionals on the space of test-functions; in the second
instance we are dealing with distributions defined as continuous linear functionals on
the space of test-densities. Depending on our choice the effect of change of variables
is different.
We must now make our notation more precise: we shall temporarily write |d𝑥| for
the Lebesgue measure to emphasize the fact that the Lebesgue measure is a positive
(Radon) measure and that the meaning of d𝑥1 · · · d𝑥 𝑛 does not depend on the order of
its factors (whereas it does in the 𝑛-form d𝑥 1 ∧· · ·∧d𝑥
𝑛 ). We can then define densities
of arbitrary (but real) degree 𝜆. We denote by C 𝑘 Ω, |d𝑥| 𝜆 the space of 𝜆-densities
𝜑 (𝑥) |d𝑥| 𝜆 , 𝜑 ∈ C 𝑘 (Ω) (0 ≤ 𝑘 ≤ +∞; obviously other function spaces
than C can
𝑘

be used). The pullback under the diffeomorphism 𝐹 is defined in C 𝑘 Ω′, |d𝑦| 𝜆 by

𝜆
𝜕𝐹
𝐹 ∗ 𝜓 (𝑦) |d𝑦| 𝜆 = 𝜓 (𝐹 (𝑥)) det (𝑥) |d𝑥| 𝜆 . (11.1.3)
𝜕𝑥

We denote by Cc𝑘 Ω, |d𝑥| 𝜆 the space of test 𝜆-densities 𝜑 (𝑥) |d𝑥| 𝜆 , 𝜑 ∈ Cc𝑘 (Ω).

The continuous linear functionals on Cc∞ Ω, |d𝑥| 𝜆 are interpreted as distribu-

tion (1 − 𝜆)-densities. The space they form will be denoted by D ′ Ω, |d𝑥| 1−𝜆 and
we shall use the notation 𝑢 (𝑥) |d𝑥| 1−𝜆 for the distribution (1 − 𝜆)-density corre-
sponding to the ordinary distribution 𝑢. The duality bracket between distributions
(1 − 𝜆)-densities and test 𝜆-densities is simply given by

⟨𝑢 (𝑥) |d𝑥| 1−𝜆 , 𝜑 (𝑥) |d𝑥| 𝜆 ⟩ = ⟨𝑢, 𝜑 |d𝑥|⟩.

1 (Ω) the pushforward of the density 𝑢 (𝑥) |d𝑥| 1−𝜆 is given by
If 𝑢 ∈ 𝐿 loc

⟨𝐹∗ 𝑢 (𝑥) |d𝑥| 1−𝜆 , 𝜓 (𝑦) |d𝑦| 𝜆 ⟩ = ⟨𝑢 (𝑥) |d𝑥| 1−𝜆 , 𝐹 ∗ 𝜓 (𝑦) |d𝑦| 𝜆 ⟩
∫ 𝜆
𝜕𝐹
= 𝑢 (𝑥) 𝜓 (𝐹 (𝑥)) det (𝑥) |d𝑥|
𝜕𝑥
∫ 𝜆−1

−1
𝜕𝐹 −1
= 𝑢 𝐹 (𝑦) 𝜓 (𝑦) det 𝐹 (𝑦) |d𝑦| ,
𝜕𝑥

i.e.,
𝜆−1

1−𝜆

−1
𝜕𝐹 −1
𝐹∗ 𝑢 (𝑥) |d𝑥| =𝑢 𝐹 (𝑦) det 𝐹 (𝑦) |d𝑦| 1−𝜆 . (11.1.4)
𝜕𝑥
330 11 Distributions and Hyperfunctions on a Manifold

As should be expected (11.1.4) can also be interpreted as the pullback of the (1 − 𝜆)-
density 𝑢 (𝑥) |d𝑥| 1−𝜆 under the map 𝐹 −1 : Ω′ −→ Ω.
At the beginning of this subsection we were dealing with zero-density distribu-
tions, after which we were dealing with one-density distributions. Symmetry between
distribution-densities and test-densities obtains when 𝜆 = 21 .
It is clear that distribution-densities can be defined in a C ∞ manifold M whether it
is orientable or not (Definition 9.1.18). As a matter of fact, it suffices to define smooth
densities 𝜑 (𝑥) |d𝑥| 1−𝜆 , 𝜑 ∈ C ∞ (M). The definition is obvious in any local chart
(U, 𝑥1 , ..., 𝑥 𝑛 ) and changes of coordinates have the effect prescribed by (11.1.3).
Patching together the definitions in overlapping coordinates charts regardless of
orientation presents no difficulty. The notion of support
of the density 𝜑 (𝑥) |d𝑥| 1−𝜆
is self-evident: it is supp 𝜑. Then D ′ M, |d𝑥| 𝜆 can be defined as the dual of

Cc∞ M, |d𝑥| 1−𝜆 . Any element of D ′ M, |d𝑥| 𝜆 can be written in the form 𝑢 |d𝑥| 𝜆

with 𝑢 a distribution (zero-density) in M. We assume that D ′ M, |d𝑥| 𝜆 is equipped

with the weak convergence topology. Every space D ′ M, |d𝑥| 𝜆 is isomorphic to

D ′ M, |d𝑥| 0 = D ′ (M).

The support of a distribution-density 𝑢 ∈ D ′ M, |d𝑥| 𝜆 is defined in the usual
manner, by using C ∞ partitions of unity in M, which are available since M is
countable at infinity (hence paracompact). Multiplication of a density distribution
𝑢 ∈ D ′ M, |d𝑥| 𝜆 by 𝜒 ∈ C ∞ (M, |d𝑥| 𝜇 ) yields a density distribution 𝜒𝑢 ∈

D ′ M, |d𝑥| 𝜆+𝜇 . If 𝜆 + 𝜇 = 1 and if either 𝜒 or 𝑢 has compact support we have
⟨𝑢, 𝜒⟩ = ⟨𝜒𝑢, 1⟩; if moreover 𝜒𝑢 is a locally ∫ integrable function 𝑓 (which makes
sense in a manifold) we have ⟨𝜒𝑢, 1⟩ = 𝑓 (𝑥) d𝑥.

11.1.2 Currents on a manifold

In contrast to the definitions in the preceding subsection the notion introduced next,
of a current on a C ∞ manifold M (see [De Rham, 1955]), is sensitive to orientation.
A 𝑝-current in M (1 ≤ 𝑝 ≤ 𝑛) is a differential 𝑝-form with distribution coefficients.
In a local chart (U, 𝑥1 , ..., 𝑥 𝑛 ) a 𝑝-current will be represented by a sum
∑︁
𝑢= 𝑢 𝐽 d𝑥 𝑗1 ∧ · · · ∧ d𝑥 𝑗 𝑝 , (11.1.5)
| 𝐽 |= 𝑝

where the 𝑢 𝐽 are “ordinary” distributions in U, with 𝐽 = 𝑗 1 , ..., 𝑗 𝑝 an ordered
multi-index: 1 ≤ 𝑗 1 < · · · < 𝑗 𝑝 ≤ 𝑛. Changes of coordinates have the same effect
on 𝑢 as if the coefficients 𝑢 𝐽 were, say, continuous functions. If 𝑦 1 , ..., 𝑦 𝑛 are new
(smooth) coordinates in U we have
11.1 Distributions and Currents on a Manifold 331

∑︁ D 𝑥 𝑗1 , ..., 𝑥 𝑗 𝑝
𝑢= 𝑢𝐽 d𝑦 𝑘1 ∧ · · · ∧ d𝑦 𝑗 𝑝 . (11.1.6)
| 𝐽 |= 𝑝 D 𝑦 𝑘1 , ..., 𝑦 𝑘 𝑝

Notation 11.1.1 We denote by D ′ (M; CΛ 𝑝 𝑇 ∗ M) the linear space of 𝑝-currents on

the manifold M.

We equip the linear space D ′ (M; CΛ𝑞 𝑇 ∗ M) with the topology of weak con-
vergence of their coefficients: a sequence of 𝑞-currents converges to zero if their
coefficients tend to zero. The notion of support of a current is obvious: it is the union
of the supports of its coefficients.

Remark 11.1.2 The transformation from (11.1.5) to (11.1.6) means that we are
viewing the 𝑢 𝐽 as distribution zero-densities. In other words,

D ′ (M; CΛ 𝑝 𝑇 ∗ M) = D ′ M, |d𝑥| 0 ⊗ C ∞ (M; CΛ 𝑝 𝑇 ∗ M) ,

where the right-hand side is the ordinary tensor product (with the obvious topology).

There is a natural exterior multiplication operation

D ′ (M; CΛ 𝑝 𝑇 ∗ M) ×C ∞ (M; CΛ𝑞 𝑇 ∗ M) ∋ (𝑢, 𝜑)

−→ 𝑢 ∧ 𝜑 ∈ D ′ M; CΛ 𝑝+𝑞 𝑇 ∗ M .

Í
With 𝑢 as in (11.1.5) and 𝜑 = |𝐾 |=𝑞 𝜑 𝐾 d𝑥 𝑖1 ∧ · · · ∧ d𝑥𝑖𝑞 we have
∑︁
𝑢∧𝜑= 𝜑 𝐾 𝑢 𝐽 d𝑥 𝐽 ∧ d𝑥 𝐾 .
|𝐾 |=𝑞, | 𝐽 |= 𝑝

The exterior multiplication (𝜑, 𝑢) −→ 𝜑 ∧ 𝑢 is similarly defined, with d𝑥 𝐾 ∧ d𝑥 𝐽

substituted for d𝑥 𝐽 ∧ d𝑥 𝐾 . That these exterior products are coordinate invariant
follows at once from their invariance when both factors are (smooth, continuous,
locally-𝐿 1 ) functions. In particular, we have the natural bilinear map

D ′ (M; CΛ 𝑝 𝑇 ∗ M) ×C ∞ (M; CΛ𝑛− 𝑝 𝑇 ∗ M) ∋ (𝑢, 𝜑)

−→ 𝑢 ∧ 𝜑 ∈ D ′ (M; CΛ𝑛𝑇 ∗ M) .

Now CΛ𝑛𝑇 ∗ M is a complex line bundle: in a local coordinate chart (U, 𝑥1 , ..., 𝑥 𝑛 )
we have
𝑢 ∧ 𝜑 = 𝑣d𝑥 1 ∧ · · · ∧ d𝑥 𝑛 , 𝑣 ∈ D ′ (U) . (11.1.7)
It is important to note that we cannot equate 𝑢 ∧ 𝜑 to the distribution one-density
𝑣 |d𝑥|: coordinate changes do not have the same effect on 𝑣d𝑥 1 ∧ · · · ∧ d𝑥 𝑛 as on
𝑣 |d𝑥| (an odd permutation of the coordinates 𝑥1 , ..., 𝑥 𝑛 change the sign of the former
and has no effect on the latter). If the manifold M is not orientable it might be
possible to find a closed path 𝔠 covered by domains of local charts which reverts the
orientation of the frame of the tangent space at one of its points after an odd number
332 11 Distributions and Hyperfunctions on a Manifold

of “round-trips” along 𝔠. If we wish to interpret D ′ (M; CΛ 𝑝 𝑇 ∗ M) as the dual of a

space of forms the latter must be twisted forms (meaning that they change sign under
reversal of the orientation). We shall not pursue the discussion in this direction and
refer the reader to [Schwartz, 1966], pp. 315–318.
If the manifold M is orientable the situation becomes simpler. In this case it is
possible and convenient to use only local frames that agree with a chosen orientation:
all pairs of admissible coordinate systems 𝑥 1 , ..., 𝑥 𝑛 and 𝑦 1 , ..., 𝑦 𝑛 defined in one and
( 𝑥1 ,..., 𝑥𝑛 )
the same domain U must satisfy 𝐷 𝐷 ( 𝑦1 ,...,𝑦𝑛 ) > 0. With such a choice (11.1.7)
can be interpreted as a one-density 𝑣. If 𝜑 is∫compactly supported we can define
⟨𝑢, 𝜑⟩ = ⟨𝑣, 1⟩. If 𝑣 ∈ 𝐿 c1 (U) we get ⟨𝑢, 𝜑⟩ = R𝑛 𝑣 (𝑥) d𝑥.
In the general case of a possibly nonorientable manifold M we forego using
duality. We define the standard operations on currents simply by extending those
defined on smooth differential forms (which form a dense linear subspace of the
space of currents). Thus the exterior derivative extends as a continuous linear map

d : D ′ (M; CΛ 𝑝 𝑇 ∗ M) −→ D ′ M; CΛ 𝑝+1𝑇 ∗ M , (11.1.8)

where 𝑝 = 1, ..., 𝑛. Of course D ′ M; CΛ𝑛+1𝑇 ∗ M = {0}. We can also allow 𝑝 = 0

by introducing the zero-currents; these are the elements of D ′ (M) or, if one prefers,
the distribution zero-densities. In local coordinates with 𝑢 given by (11.1.5) we have
𝑛 ∑︁
∑︁ 𝜕𝑢 𝐽
d𝑢 = d𝑥 𝑘 ∧ d𝑥 𝑗1 ∧ · · · ∧ d𝑥 𝑗 𝑝 . (11.1.9)
𝜕𝑥 𝑘
𝑘=1 | 𝐽 |= 𝑝

The sequence of maps (11.1.8), augmented at the first step by the injection of
the locally constant functions into D ′ (M), constitutes the De Rham complex for
currents [cf. (9.4.18)]. Here also the Poincaré Lemma is valid: the De Rham complex
for currents is locally exact. The proof is easily derived from Theorem 9.4.10:
obviously, it suffices to reason in an open subset of R𝑛 and use the fact that, there,
the exterior derivative commutes with smoothing operators, i.e., convolution with
C ∞ functions. An important theorem is that the cohomology spaces of the De Rham
complex for currents are naturally isomorphic to those of the De Rham complex for
smooth forms (see Subsection 9.4.4).

11.1.3 The Bochner–Martinelli formulas

In this subsection we describe a kernel current in C𝑛 × C𝑛 R2𝑛 × R2𝑛 which

plays an important role in several complex variables theory, generalizing the Cauchy
kernel
1 1
d𝑤 (11.1.10)
2𝜋𝑖 𝑧 − 𝑤
in one complex variable theory. The relevant differential complex here is
11.1 Distributions and Currents on a Manifold 333

𝜕¯ : D ′ Ω; 𝑇 ( 𝑝,𝑞) C𝑛 −→ D ′ Ω; 𝑇 ( 𝑝,𝑞+1) C𝑛 (11.1.11)

where Ω ⊂ C𝑛 is an open set. The vector bundles 𝑇 ( 𝑝,𝑞) M (M: a complex manifold)
have been introduced in Subsection 9.4.5. The distribution sections of 𝑇 ( 𝑝,𝑞) C𝑛 have
expressions ∑︁ ∑︁
𝑓 = 𝑓 𝐼, 𝐽 d𝑧 𝐼 ∧ d𝑧¯ 𝐽 (11.1.12)
|𝐼 |= 𝑝 | 𝐽 |=𝑞

with coefficients 𝑓 𝐼, 𝐽 ∈ D′ (Ω). We shall

also say that 𝑓 is a current of bidegree
( 𝑝, 𝑞). The multi-indices 𝐼 = 𝑖 1 , ..., 𝑖 𝑝 and 𝐽 = 𝑗1 , ..., 𝑗 𝑞 are ordered: 1 ≤ 𝑖 1 <
· · · < 𝑖 𝑝 ≤ 𝑛, 1 ≤ 𝑗 1 < · · · < 𝑗 𝑞 ≤ 𝑛. As with smooth coefficients,
𝑛
∑︁ ∑︁ ∑︁ 𝜕 𝑓 𝐼, 𝐽
𝜕¯ 𝑓 = d𝑧¯ 𝑘 ∧ d𝑧 𝐼 ∧ d𝑧¯ 𝐽 . (11.1.13)
𝜕 𝑧¯ 𝑘
|𝐼 |= 𝑝 | 𝐽 |=𝑞
𝑘=1

We are now going to introduce a kernel current of bidegree (𝑛, 𝑛 − 1) in the

whole of C𝑛 × C𝑛 . First consider the locally-𝐿 1 functions in R2𝑛 :

(𝑛 − 1)! 𝑧¯ 𝑗
𝐸 𝑗 (𝑧) = , 𝑗 = 1, ..., 𝑛. (11.1.14)
(2𝜋𝑖) 𝑛 |𝑧| 2𝑛

Lemma 11.1.3 We have

𝑛
∑︁ 𝜕𝐸 𝑗
= (2𝑖) −𝑛 𝛿0 (11.1.15)
𝑗=1
𝜕 ¯
𝑧 𝑗

where 𝛿0 is the Dirac distribution in R2𝑛 .

Proof When 𝑛 = 1 we have 𝐸 1 = 2𝑖1𝜋𝑧 and (11.1.15) is a restatement of the fact that
1
𝜋𝑧 is a fundamental solution of 𝜕𝑧¯ (see e.g. [Treves, 1975], p. 34 et seq.). For 𝑛 > 1
we note that
(𝑛 − 2)! 1 (𝑛 − 1)! 1
𝐸 𝑗 (𝑧) = − 𝜕𝑧 = 𝜕𝑧 |𝑧| 2
(2𝜋𝑖) 𝑛 𝑗 |𝑧| 2𝑛−2 (2𝜋𝑖) 𝑛 |𝑧| 2𝑛 𝑗

and therefore
𝑛
∑︁ (𝑛 − 2)! 1
𝜕𝑧¯ 𝑗 𝐸 𝑗 = − (2𝑖) −𝑛 Δ ,
𝑗=1
4𝜋 𝑛 |𝑧| 2𝑛−2

where Δ is the Laplacian in R2𝑛 . The locally-𝐿 1 function − (𝑛−2)! 1

is a funda-
4 𝜋𝑛 |𝑧 | 2𝑛−2
mental solution of Δ in R2𝑛 (see [Hörmander, 1983, I], Th. 3.3.2, [Treves, 1975],
p. 68 et seq.). □
334 11 Distributions and Hyperfunctions on a Manifold

We introduce the following currents of bidegree (𝑛, 𝑛 − 1) in C𝑛 :

E 𝑗 (𝑧, d𝑧, d𝑧¯) = 𝐸 𝑗 (𝑧) d𝑧¯1 ∧ · · · ∧ dc

𝑧¯ 𝑗 ∧ · · · ∧ d𝑧¯𝑛 ∧ d𝑧, (11.1.16)
𝑛
∑︁
E (𝑧, d𝑧, d𝑧¯) = (−1) 𝑗−1 E 𝑗 (𝑧, d𝑧, d𝑧¯) , (11.1.17)
𝑗=1

where the hatted factor must be omitted and d𝑧 = d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 . We derive directly

from (11.1.15) and (11.1.16)–(11.1.17):
¯ (𝑧, d𝑧, d𝑧¯) = (2𝑖) −𝑛 𝛿0 d𝑧¯ ∧ d𝑧.
𝜕E (11.1.18)

We recall Remark 9.4.6:

d𝑧¯ ∧ d𝑧 = (2𝑖) 𝑛 d𝑥 ∧ d𝑦. (11.1.19)

Lemma 11.1.4 If 𝜑 (𝑧, d𝑧, d𝑧¯) ∈ C ∞ C𝑛 ; 𝑇 (𝑛,𝑞) , 0 ≤ 𝑞 ≤ 𝑛, then
∫
−𝑛
(2𝑖) 𝛿(𝑧 − 𝑤)d ( 𝑧¯ − 𝑤)
¯ ∧ d (𝑧 − 𝑤) ∧ 𝜑 (𝑤, d𝑤, d𝑤)
¯ = 𝜑 (𝑧, d𝑧, d𝑧¯) .
R2𝑛
(11.1.20)
∞
Proof It suffices to prove the claim for 𝜑 (𝑧, d𝑧, d𝑧¯) = 𝜑 𝐼 (𝑧) d𝑧¯ 𝐼 ∧d𝑧, 𝜑 𝐼 ∈ C (C ),
𝑛

𝐼 = 𝑖 1 , ..., 𝑖 𝑞 . Actually, a permutation of the variables allows us to take 𝐼 =

(𝑛 − 𝑞 + 1, ..., 𝑛). In this case,
∫
𝛿(𝑧 − 𝑤)d ( 𝑧¯ − 𝑤) ¯ ∧ d (𝑧 − 𝑤) ∧ 𝜑 (𝑤, d𝑤, d𝑤) ¯
R2𝑛
∫
= 𝜑 𝐼 (𝑤) 𝛿 (𝑧 − 𝑤) (d𝑧¯ − d𝑤) ¯ ∧ d𝑧 ∧ d𝑤¯ 𝐼 ∧ d𝑤
R2𝑛
∫
= (−1) 𝑛−𝑞 𝜑 𝐼 (𝑤) 𝛿 (𝑧 − 𝑤) d𝑤¯ 1 ∧ · · · ∧ d𝑤¯ 𝑛−𝑞 ∧ d𝑧¯ 𝐼 ∧ d𝑧 ∧ d𝑤¯ 𝐼 ∧ d𝑤
R2𝑛
∫
(𝑛−𝑞) (𝑞+1)+𝑛(𝑛+𝑞)
= (−1) d𝑧¯ 𝐼 ∧ 𝜑 𝐼 (𝑤) 𝛿 (𝑧 − 𝑤) d𝑤¯ ∧ d𝑤 ∧ d𝑧
R2𝑛
= (2𝑖) 𝑛 𝜑 𝐼 (𝑧) d𝑧¯ 𝐼 ∧ d𝑧

since (𝑛 − 𝑞) (𝑞 + 1) + 𝑛 (𝑛 + 𝑞) = 𝑛 (𝑛 + 1) − 𝑞 (𝑞 + 1) + 2𝑛𝑞 is an even integer. □

Now let Ω be a bounded open subset of C𝑛 with a C 1 boundary 𝜕Ω; Ω will stand
for the closure of Ω. We use the current kernel E (𝑧, d𝑧, d𝑧¯) to define operators

(𝑞)
KΩ : C ∞ Ω; 𝑇 (𝑛,𝑞) −→ C ∞ Ω; 𝑇 (𝑛,𝑞−1) ,

(𝑞)
K𝜕Ω : C 0 𝜕Ω; 𝑇 (𝑛,𝑞) −→ C ∞ Ω; 𝑇 (𝑛,𝑞) ,

by the integrals
11.1 Distributions and Currents on a Manifold 335
∫
(𝑞)
KΩ 𝑓 (𝑧, d𝑧, d𝑧¯) = (−1) 𝑛−𝑞 𝑓 (𝑤, d𝑤, d𝑤)
¯ ∧ E (𝑧 − 𝑤, d (𝑧 − 𝑤) , d ( 𝑧¯ − 𝑤))
¯ ,
Ω
(11.1.21)
∫
(𝑞)
K𝜕Ω 𝑔 (𝑧, d𝑧, d𝑧¯) = (−1) 𝑛−𝑞 𝑔 (𝑤, d𝑤, d𝑤)
¯ ∧ E (𝑧 − 𝑤, d (𝑧 − 𝑤) , d ( 𝑧¯ − 𝑤))
¯ .
𝜕Ω
(11.1.22)

Since 𝑓 and 𝑔 are of bidegree (𝑛, 𝑞) we have, for 𝑘 = 1, ..., 𝑛,

∫
𝑓 (𝑤, d𝑤, d𝑤)
¯ ∧ E 𝑘 (𝑧 − 𝑤, d (𝑧 − 𝑤) , d ( 𝑧¯ − 𝑤))
¯ =
Ω
∑︁ ∑︁ ∫
𝑎 𝐼, 𝐽 ,𝑘 𝑓 (𝑤, d𝑤, d𝑤)
¯ ∧ 𝐸 𝑘 (𝑧 − 𝑤) d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯ 𝐼 ∧ d𝑤¯ 𝐽 ;
|𝐼 |=𝑞−1 |𝐽 |=𝑛−𝑞 Ω
∫
𝑔 (𝑤, d𝑤, d𝑤)
¯ ∧ E 𝑘 (𝑧 − 𝑤, d (𝑧 − 𝑤) , d ( 𝑧¯ − 𝑤))
¯ =
𝜕Ω
∑︁ ∑︁ ∫
𝑏 𝐼, 𝐽 ,𝑘 𝑔 (𝑤, d𝑤, d𝑤)
¯ ∧ 𝐸 𝑘 (𝑧 − 𝑤) d𝑧 1 ∧ · · · ∧ d𝑧 𝑛 ∧ d𝑧¯ 𝐼 ∧ d𝑤¯ 𝐽 ,
|𝐼 |=𝑞 | 𝐽 |=𝑛−𝑞−1 𝜕Ω

where the coefficients 𝑎 𝐼, 𝐽 ,𝑘 and 𝑏 𝐼, 𝐽 ,𝑘 are real numbers independent of 𝑓 and 𝑔.

¯
We can now state the Bochner–Martinelli formulas for the 𝜕-complex.

Theorem 11.1.5 Given any 𝑓 ∈ C ∞ Ω; 𝑇 (𝑛,𝑞) we have

¯ (𝑞) 𝑓 + K (𝑞+1) 𝜕¯ 𝑓 + K (𝑞) 𝑓 .

𝑓 = 𝜕K (11.1.23)
Ω Ω 𝜕Ω

The proof is a straightforward exploitation of the identities (11.1.20), (11.1.21),

(11.1.22), facilitated by the fact that for (𝑛, 𝑞)-currents the 𝜕¯ operator coincides
with the exterior derivative (Proposition 9.4.16). For the details see [Treves, 1992],
Theorem VI.1.1.
When supp 𝑓 ⊂ Ω, Formula (11.1.23) reduces to the homotopy formula

¯ (𝑞) 𝑓 + K (𝑞+1) 𝜕¯ 𝑓 .
𝑓 = 𝜕K (11.1.24)
Ω Ω

Remark 11.1.6 Formula (11.1.24) remains valid when Ω = R2𝑛 and 𝑓 is compactly
(𝑞)
supported (or decays sufficiently fast at infinity). In this case KΩ is a convolution
operator.
¯
We know by Theorem 9.4.14 that the 𝜕-complex on an arbitrary complex manifold
is locally exact. Formula (11.1.23) implies that it is globally exact in special (to be
determined) open subsets of C𝑛 . In the next section we introduce domains in C𝑛 in
which the latter is true.
336 11 Distributions and Hyperfunctions on a Manifold

11.2 Plurisubharmonic functions and pseudoconvex domains

11.2.1 Subharmonic functions in an open subset 𝛀 of the plane

We recall some terminology:

(1) A function 𝑓 defined in a topological space 𝑬 and valued in [−∞, +∞) is said
to be upper semicontinuous if the “sublevel set” {℘ ∈ 𝑬; 𝑓 (℘) < 𝑡} is open,
whatever 𝑡 ∈ R.
(2) A harmonic function ℎ in an open subset Ω of C is a solution of the Laplace
equation Δℎ = 0; of course, ℎ ∈ C 𝜔 (Ω).

Definition 11.2.1 Let Ω ⊂ C be open. A function 𝑓 : Ω −→ [−∞, +∞) is said to

be subharmonic if 𝑓 is upper semicontinuous and if 𝑓 has the following equivalent
properties:

(1) Given any open set Ω′ ⊂⊂ Ω and any function ℎ ∈ C Ω′ harmonic in Ω′,
𝑓 ≤ ℎ on 𝜕Ω′ entails 𝑓 ≤ ℎ in Ω′.
(2) Given any closed disk (𝑥, 𝑦) ∈ R2 ; |𝑥 − 𝑥 ◦ | 2 + |𝑦 − 𝑦 ◦ | 2 ≤ 𝑟 2 ⊂ Ω,
∫ 2𝜋
◦ ◦ 1
𝑓 (𝑥 , 𝑦 ) ≤ 𝑓 (𝑥 ◦ + 𝑟 cos 𝜃, 𝑦 ◦ + 𝑟 sin 𝜃) d𝜃. (11.2.1)
2𝜋 0

Properties (1) and (2) imply that subharmonicity is a local property (see also
Proposition 11.2.3 below). Subharmonic functions form a convex cone in the vector
space of all functions Ω −→ [−∞, +∞). If 𝜑 is a convex monotone increasing
function in [−∞, +∞) and if 𝑓 is subharmonic in Ω the same is true of 𝜑 ◦ 𝑓 .
The limit of a decreasing sequence of subharmonic functions is subharmonic.
The supremum of a finite family of subharmonic functions is subharmonic. This is
also true of an infinite family of subharmonic functions provided the supremum of
these functions is finite and upper semicontinuous ([Hörmander, 1966], Theorem
1.6.2).
The following statement is not difficult to deduce from Definition 11.2.1:

Proposition 11.2.2 Let the open set Ω ⊂ R2 be connected and suppose that the
1 (Ω),
subharmonic function 𝑓 in Ω is not identically equal to −∞. Then 𝑓 ∈ 𝐿 loc
implying 𝑓 (𝑥, 𝑦) > −∞ a.e.

It follows that we can regard every subharmonic function 𝑓 in Ω, 𝑓 . −∞, as a

distribution in Ω. The next statement combines Theorems 1.6.10 and 1.6.11, loc. cit.
1 (Ω) is subharmonic if and only
Proposition 11.2.3 A real-valued function 𝑓 ∈ 𝐿 loc
if the distribution Δ 𝑓 is a positive Radon measure in Ω.

Corollary 11.2.4 A real-valued function 𝑓 ∈ C 2 (Ω) is subharmonic if and only if

Δ 𝑓 ≥ 0 at every point of Ω.
11.2 Plurisubharmonic functions and pseudoconvex domains 337

A function 𝑓 ∈ C 2 (Ω) is said to be strictly subharmonic if Δ 𝑓 > 0 at every

point of Ω.
Property (2), Definition 11.2.1, implies directly the following

Proposition 11.2.5 If ℎ ∈ O (Ω) then log |ℎ| is subharmonic.

If Ω is connected and ℎ ∈ O (Ω), ℎ . 0, then the set in which log |ℎ| = −∞ is

discrete.

11.2.2 Plurisubharmonic functions in an open subset of C𝒏

In the sequel Ω will be an open subset of C𝑛 , 𝑛 ≥ 2, where the variable point is

𝑧 = (𝑧 1 , ..., 𝑧 𝑛 ) (except for the terminology all results will be valid when 𝑛 = 1); we
write 𝑓 (𝑧) even when the function 𝑓 in Ω is not holomorphic.

Definition 11.2.6 A function 𝑓 : Ω −→ [−∞, +∞) is said to be plurisubharmonic if

𝑓 is upper semicontinuous and if, given an arbitrary affine complex line 𝐿 intersecting
Ω, the restriction of 𝑓 to 𝐿 ∩ Ω is subharmonic.

An affine complex line 𝐿 is the image of a map C ∋ 𝜁 ↦→ 𝑧◦ + 𝜁 𝑧 ∗ , where

𝑧◦ ,
𝑧 ∗ ∈ C𝑛 ; harmonic and subharmonic functions in 𝐿 ∩ Ω are defined in the 𝜁
coordinate. We shall denote by Psh (Ω) the set of plurisubharmonic functions in Ω.

Remark 11.2.7 In some instances we will need to handle functions 𝑓 defined and
plurisubharmonic in an open subset Ω of a complex affine subspace 𝑬 of C𝑛 . Their
definition is obvious: 𝑓 is upper semicontinuous and given an arbitrary affine complex
line 𝐿 ⊂ 𝑬 intersecting Ω, the restriction of 𝑓 to 𝐿 ∩ Ω is subharmonic.

We have an immediate generalization of the analogous properties for 𝑛 = 1 (see

the preceding subsection):

Proposition 11.2.8 The limit of a decreasing sequence of plurisubharmonic func-

tions is plurisubharmonic. The supremum of a finite family of plurisubharmonic
functions is plurisubharmonic. This is also true of an infinite family of plurisubhar-
monic functions provided the supremum of these functions is upper semicontinuous
and finite at every point.

Every pluriharmonic function is plurisubharmonic; log |ℎ| ∈ Psh (Ω) whatever

ℎ ∈ O (Ω) (Proposition 11.2.5); the function log |𝑧| is plurisubharmonic in C𝑛 .
The following statement is a direct consequence of Definition 11.2.6 and Corollary
11.2.4.

Proposition 11.2.9 A real-valued function 𝑓 ∈ C 2 (Ω) is plurisubharmonic if and

only if
𝑛
∑︁ 𝜕2 𝑓
∀𝑧 ∈ Ω, 𝑤 ∈ C𝑛 , (𝑧) 𝑤 𝑗 𝑤¯ 𝑘 ≥ 0. (11.2.2)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘
338 11 Distributions and Hyperfunctions on a Manifold

The quadratic form

𝑛
∑︁ 𝜕2 𝑓
𝑤 ↦→ (𝑧) 𝑤 𝑗 𝑤¯ 𝑘 (11.2.3)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

is commonly referred to as the Levi form of 𝑓 at 𝑧.

′
Corollary 11.2.10 Let Ω and Ω′ be open subsets of C𝑛 and C𝑛 respectively, and let
ℎ : Ω′ −→ Ω be a holomorphic map. If 𝑓 ∈ C 2 (Ω; R) is plurisubharmonic then
𝑓 ◦ ℎ ∈ C 2 (Ω′; R) is also plurisubharmonic.
′
Proof Indeed, we have, for all 𝑧 ′ ∈ Ω′, 𝑤 ∈ C𝑛 ,
𝑛′
∑︁ 𝜕2
𝑤 𝛼𝑤𝛽 ′ 𝑓 (ℎ (𝑧 ′))
𝛼,𝛽=1 𝜕𝑧 ′𝛼 𝜕 𝑧¯ 𝛽
𝑛′ 𝑛
∑︁ ∑︁ 𝜕2 𝑓 𝜕ℎ 𝑗 𝜕ℎ 𝑘
= 𝑤 𝛼𝑤𝛽 (ℎ (𝑧 ′)) ′ (𝑧 ′) ′ (𝑧 ′)
𝛼,𝛽=1 𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 ¯
𝑧 𝑘 𝜕𝑧 𝛼 𝜕 𝑧¯ 𝛽
𝑛′
𝑛
! 𝑛′
∑︁ 𝜕2 𝑓 ′
∑︁ 𝜕ℎ 𝑗 ′ ∑︁ 𝜕ℎ 𝑘
= (ℎ (𝑧 )) 𝑤 𝛼 ′ (𝑧 ) 𝑤 𝛽 ′ (𝑧 ′) ≥ 0. □
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 𝛼=1
𝜕𝑧 𝛼 𝛽=1
𝜕𝑧 𝛽

Definition 11.2.11 A real-valued function 𝑓 ∈ C 2 (Ω) is said to be strictly plurisub-

harmonic if to every compact subset 𝐾 of Ω there is a 𝑐 𝐾 > 0 such that
𝑛
∑︁ 𝜕2 𝑓
∀𝑧 ∈ 𝐾, 𝑤 ∈ C𝑛 , (𝑧) 𝑤 𝑗 𝑤¯ 𝑘 ≥ 𝑐 𝐾 |𝑤| 2 . (11.2.4)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

Let 𝜒 ∈ Cc∞ (R𝑛 ), 𝜒 ≥ 0, such that |𝑥| > 1 =⇒ 𝜒 (𝑥) = 0, and

∫
𝜒 (|𝑧 1 | , ..., |𝑧 𝑛 |) d𝜆 (𝑧) = 1,
C𝑛

where d𝜆 (𝑧) is the Lebesgue measure in R2𝑛 . If 𝑓 ∈ Psh (Ω) we define, for 𝑧 ∈
Ω 𝜀 = {𝑧 ∈ Ω; dist (𝑧, C𝑛 \Ω) > 𝜀}, with 𝜀 > 0 sufficiently small that Ω 𝜀 ≠ ∅,
∫
𝑓 𝜀 (𝑧) = 𝑓 (𝑧 − 𝜀𝑤) 𝜒 (|𝑤 1 | , ..., |𝑤 𝑛 |) d𝜆 (𝑤) . (11.2.5)
C𝑛

We have 𝑓 𝜀 ∈ C ∞ (Ω 𝜀 ).

Proposition 11.2.12 Assume that Ω is connected and that 𝑓 ∈ Psh (Ω), 𝑓 . −∞;
then 𝑓 𝜀 ∈ Psh (Ω 𝜀 ) and 𝑓 𝜀 ↘ 𝑓 as 𝜀 ↘ 0.

The proof is essentially that of the same result when 𝑛 = 1 ([Hörmander,

1966], Theorems 1.6.11 and 2.6.3). By combining Corollary 11.2.10 and Propo-
sition 11.2.12 one readily obtains
11.2 Plurisubharmonic functions and pseudoconvex domains 339
′
Proposition 11.2.13 Let Ω and Ω′ be domains in C𝑛 and C𝑛 respectively, and let
ℎ : Ω′ −→ Ω be a holomorphic map. If 𝑓 ∈ Psh (Ω), 𝑓 . −∞, then 𝑓 ◦ℎ ∈ Psh (Ω′).
In particular, plurisubharmonicity is invariant under biholomorphisms and there-
fore can be defined in complex-analytic manifolds.

11.2.3 Pseudoconvex domains. Global exactness of the 𝝏¯ -complex

We relate to plurisubharmonic functions a class of domains in C𝑛 that play a central

role in Several Complex Variables theory (see [Hörmander, 1966] to which loc. cit.
refers in this subsection).
Definition 11.2.14 Given an open subset Ω of C𝑛 we shall say that a compact subset
𝐾 of Ω is plurisubharmonic convex in Ω if to every point 𝑧 ∈ Ω\𝐾 there is a
continuous plurisubharmonic function 𝑓 in Ω such that 𝑓 (𝑧) > sup 𝑓 .
𝐾

A convex compact subset 𝐾 of Ω is plurisubharmonic convex in Ω.

Definition 11.2.15 An open subset Ω of C𝑛 is said to be pseudoconvex if the fol-
lowing equivalent conditions are satisfied:
(1) The function Ω ∋ 𝑧 ↦→ − log dist (𝑧, C𝑛 \Ω) is plurisubharmonic.
(2) There is a plurisubharmonic function 𝑓 in Ω such that the closure of the sublevel
set {𝑧 ∈ Ω; 𝑓 (𝑧) < 𝑡} is a compact subset of Ω whatever 𝑡 ∈ R.
(3) Every compact subset of Ω is contained in a plurisubharmonic convex compact
subset of Ω.
In (1), dist (𝑧, C𝑛 \Ω) is the Euclidean distance from 𝑧 to the complement of Ω.
On the equivalence of Conditions (1), (2), (3), in Definition 11.2.15 see Theorem
2.6.7, loc. cit.
When Ω = C𝑛 Condition (1) still makes sense if we replace C𝑛 \Ω by {∞}, in
which case − log dist (𝑧, C𝑛 \Ω) ≡ −∞; in Condition (2) we may take 𝑓 (𝑧) = log |𝑧|.
Perhaps not immediately apparent is the fact that the pseudoconvexity of Ω is a
local property of the boundary 𝜕Ω:
Proposition 11.2.16 If every point 𝑧◦ ∈ Ω has a neighborhood 𝑈 in C𝑛 such that
𝑈 ∩ Ω is pseudoconvex then Ω is pseudoconvex.
The proof is an exercise in exploiting both Properties (1) and (2), Definition
11.2.15 (cf. loc. cit., p. 48).
Proposition 11.2.17 If the open subset Ω of C𝑛 is pseudoconvex then the intersection
of Ω with a complex affine subspace 𝑬 of C𝑛 is a pseudoconvex open subset of 𝑬.
Proof Follows from Property (2), Definition 11.2.15, and the fact that restriction
to Ω ∩ 𝑬 of a plurisubharmonic function 𝑓 in Ω is a plurisubharmonic function in
Ω ∩ 𝑬 (cf. Remark 11.2.7). □
340 11 Distributions and Hyperfunctions on a Manifold

To say that Ω has a C 𝑘 (𝑘 ∈ Z+ or 𝑘 = ∞, or 𝑘 = 𝜔) boundary is the same as

saying that every point 𝑧◦ ∈ 𝜕Ω has a neighborhood 𝑈 in C𝑛 in which is defined
a real-valued C 𝑘 function 𝜌 (𝑧) such that 𝑈 ∩ 𝜕Ω = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0}. Excluding
the C 𝜔 case, one can patch the local defining functions together by means of a C ∞
partition of unity and assume that 𝑈 ⊃ 𝜕Ω, in which case we say that 𝜌 is a defining
function of 𝜕Ω. We recall the following

Definition 11.2.18 Let the domain Ω have a C 2 boundary 𝜕Ω and 𝜌 be a C 2 defining

function of 𝜕Ω. By the Levi form at 𝑧 ∈ 𝜕Ω we mean the Hermitian form
𝑛
∑︁ 𝜕2 𝜌
𝑳 𝜌(𝑧) (𝑤) = (𝑧) 𝑤 𝑗 𝑤¯ 𝑘 (11.2.6)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

Í 𝜕𝜌
defined on the complex tangent space to 𝜕Ω, meaning that 𝑛𝑗=1 𝜕𝑧 𝑗 (𝑧) 𝑤 𝑗 = 0. We
2
also refer to the self-adjoint 𝑛 × 𝑛 matrix 𝑳 𝜌(𝑧) = 𝜕𝑧𝜕𝑗 𝜕𝜌𝑧¯ 𝑘 (𝑧) as the Levi
1≤ 𝑗,𝑘 ≤𝑛
matrix of 𝜕Ω at 𝑧.

If 𝑔 ∈ C 2 (𝑈), 𝑔 > 0 in 𝑈 [so that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝑔 (𝑧) 𝜌 (𝑧) < 0}], we see that
𝑳 (𝑔𝜌) (𝑧) = 𝑔 (𝑧) 𝑳 𝜌(𝑧) . A local biholomorphism 𝜑 preserving 𝑧 transforms 𝑳 𝜌(𝑧)
into Γ∗ (𝑧) 𝑳 𝜌(𝑧) Γ (𝑧), Γ (𝑧) ∈ GL (C, 𝑛). It follows that the properties of being
positive (resp., negative) definite or semidefinite are holomorphic coordinates-free
and independent of the defining function 𝜌. We may therefore speak of the Levi form
of 𝜕Ω at 𝑧 without specifying what the defining function or the complex coordinates
are.

Proposition 11.2.19 (loc. cit., Theorem 2.6.12) A domain Ω in C𝑛 with a C 2 bound-

ary 𝜕Ω is pseudoconvex (Definition 11.2.15) if and only if the Levi form at every
point of 𝜕Ω is positive semidefinite.

We can also introduce the following important (and invariant) concept.

Definition 11.2.20 A domain Ω in C𝑛 with a C 2 boundary 𝜕Ω is said to be strictly

(or strongly) pseudoconvex at 𝑧◦ ∈ 𝜕Ω if the Levi form of 𝜕Ω is positive definite
at 𝑧◦ . The domain Ω is said to be strictly (or strongly) pseudoconvex if it is strictly
pseudoconvex at every point of 𝜕Ω.

Pseudoconvex domains are the domains in which the 𝜕¯ differential complex [cf.
Definition 9.4.12, also (9.4.30), (9.4.31)] is globally exact (loc. cit., Corollary 4.2.6).
It is important also to relate pseudoconvexity to another important concept, that of a
domain of holomorphy (Definition 10.3.2).
We combine in one simple statement one of the fundamental results in SCV
theory (cf. Theorem 10.3.4).
11.2 Plurisubharmonic functions and pseudoconvex domains 341

Theorem 11.2.21 The following properties of a domain Ω in C𝑛 are equivalent:

(1) Ω is pseudoconvex;
(2) Ω is a domain of holomorphy;
(3) 𝐻 (0,𝑞) (Ω) = 0, 𝑞 = 1, ..., 𝑛 − 1.
Proofs can be found in loc. cit., Ch. IV. More precisely, the entailment (1)=⇒(3) is
Corollary 4.2.6, loc. cit.; (3)=⇒(2) is Theorem 4.2.9, loc. cit.; (2)=⇒(1) is Theorem
4.2.8, loc. cit. For the case 𝑞 = 𝑛 in (3) recall Theorem 9.4.15.
Corollary 11.2.22 A Runge domain in C𝑛 is pseudoconvex.
Proof Combine Theorems 9.4.17, and 11.2.21. □
Recall that not every pseudoconvex domain is Runge: every domain in the complex
plane is pseudoconvex but only the simply connected domains are Runge.
¯
When Ω is pseudoconvex there are exactness theorems for the 𝜕-complex where
∞ ′ 2
C is replaced by D or by suitably weighted 𝐿 -spaces (loc. cit., Ch. IV). Theorem
11.2.21 can be extended to Stein manifolds (loc. cit., Ch. V).
Plurisubharmonic functions will be of further use later; we will need explicit
results about plurisubharmonic quadratic forms.

11.2.4 Plurisubharmonic quadratic forms in C𝒏

An arbitrary real quadratic form 𝑄 in C𝑛 is of the form

𝑄 (𝑧) = Re ℎ𝑄 (𝑧) + 𝐿 (𝑧) (11.2.7)

Í Í
with ℎ𝑄 (𝑧) = 𝑛𝑗,𝑘=1 𝑎 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 , 𝐿 (𝑧) = 𝑛𝑗,𝑘=1 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧¯ 𝑘 , 𝑎 𝑗,𝑘 , 𝑏 𝑗,𝑘 ∈ C, 𝑎 𝑗,𝑘 = 𝑎 𝑘, 𝑗 ,

𝑏 𝑗,𝑘 = 𝑏 𝑘, 𝑗 . Thus the matrix 𝐴 = 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 is complex symmetric and 𝐵 =

𝑏 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 is self-adjoint. Note that ℎ𝑄 ∈ O (C𝑛 ) and Re ℎ𝑄 is pluriharmonic,
meaning its restriction to any complex plane is harmonic.
Recall that the signature of a quadratic form 𝔮 defined in a real vector space E is
the difference sign 𝔮 = 𝜈+ (𝔮) − 𝜈− (𝔮) where 𝜈+ (𝔮) [resp., 𝜈− (𝔮)] is the maximum
dimension of the vector subspaces 𝑀 of E such that 𝔮 (𝑧) > 0 [resp., 𝔮 (𝑧) < 0] if
0 ≠ 𝑧 ∈ 𝑀. If we equip E with a scalar product in which we can write 𝔮 (v) = v · 𝑆v
with 𝑆 a symmetric (therefore semisimple) linear endomorphism of E then 𝜈+ (𝔮) =
#eigenvalues 𝜒 > 0 of 𝑆, 𝜈− (𝔮) = #eigenvalues 𝜒 < 0 of 𝑆. The form 𝔮 is said to be
nondegenerate if det 𝑆 ≠ 0 or, equivalently, if 𝜈+ (𝔮) + 𝜈− (𝔮) = dim E.
We recall that the Hessian of a C 2 function 𝑓 in an open subset of R2𝑛 is the
symmetric 2𝑛 × 2𝑛 matrix
2
𝜕 𝑥 𝑓 𝜕 𝑥 𝜕𝑦 𝑓
Hess 𝑓 = .
𝜕𝑥 𝜕𝑦 𝑓 𝜕𝑦2 𝑓

Taking 𝑓 = 𝑄 given by (11.2.7) we get

342 11 Distributions and Hyperfunctions on a Manifold

1 𝑥 𝑥
𝑄= · Hess 𝑄 .
2 𝑦 𝑦

To say that 𝑄 is nondegenerate is to say that its Hessian matrix, Hess 𝑄, is nonsingu-
lar; Hess 𝑄 is a real symmetric 2𝑛 × 2𝑛 matrix, obviously related to the self-adjoint
(complex) 2𝑛 × 2𝑛 matrix
2
𝜕 𝑄 𝜕𝑧2¯ 𝑄

HessC (𝑄) = 𝑧,2𝑧¯ 2 𝑄
𝜕𝑧 𝑄 𝜕𝑧, 𝑧¯

2 𝑄 = 𝜕 𝜕 𝑄). This enables us to represent 𝑄 as a Hermitian form:

(𝜕𝑧, 𝑧¯ 𝑧 𝑧¯

1 𝑧¯ 𝑧
𝑄 (𝑧) = · HessC (𝑄) .
2 𝑧 𝑧¯

Since
𝑧 𝐼 𝑖𝐼 𝑥
= 𝑛 𝑛
𝑧¯ 𝐼𝑛 −𝑖𝐼𝑛 𝑦
(𝐼𝑛 : 𝑛 × 𝑛 identity matrix) we have

𝐼𝑛 𝑖𝐼𝑛 C 𝐼𝑛 𝑖𝐼𝑛
Hess 𝑄 = Hess (𝑄) . (11.2.8)
𝐼𝑛 −𝑖𝐼𝑛 𝐼𝑛 −𝑖𝐼𝑛

𝐼𝑛 𝑖𝐼𝑛
The matrix √1 is unitary.
2 𝐼𝑛 −𝑖𝐼𝑛

Proposition 11.2.23 The spectra of Hess 𝑄 and HessC (𝑄) are the same and, as
a consequence, for the quadratic form 𝑄 to be nondegenerate it is necessary and
sufficient that the matrix HessC (𝑄) be nonsingular.

Proof Let v ∈ C2𝑛 be an eigenvector

of Hess 𝑄 corresponding to the eigenvalue 𝜒;
𝐼𝑛 𝑖𝐼𝑛
(11.2.8) implies that v is an eigenvector of HessC (𝑄) corresponding to
𝐼𝑛 −𝑖𝐼𝑛
the same eigenvalue 𝜒. □
We now look at plurisubharmonic (Definition 11.2.6) and strictly plurisubhar-
monic (Definition 11.2.11) quadratic forms in C𝑛 . The following statement is a
direct consequence of Proposition 11.2.9 and Definition 11.2.11:

Proposition 11.2.24 For 𝑄 to be plurisubharmonic (resp., strictly plurisubhar-

2 𝑄 be
monic) it is necessary and sufficient that the 𝑛 × 𝑛 self-adjoint matrix 𝜕𝑧, 𝑧¯
positive semidefinite (resp., definite).

A quadratic form can be plurisubharmonic, nondegenerate and yet not be strictly

plurisubharmonic.
11.2 Plurisubharmonic functions and pseudoconvex domains 343

Example 11.2.25 The quadratic form 𝑄 (𝑧) = Im 𝑛𝑗=1 𝑧2𝑗 is pluriharmonic and non-
Í

0 −𝑖𝐼𝑛
degenerate; the eigenvalues of Hess (𝑄) =
C
are −1, 1, both with multi-
𝑖𝐼𝑛 0
plicity 𝑛.

Proposition 11.2.26 If the quadratic form 𝑄 in C𝑛 is plurisubharmonic then

sign 𝑄 ≥ 0. If 𝑬 is a real vector subspace of C𝑛 2𝑛
√ R such that 𝑄 (𝑧) < 0
for all 𝑧 ∈ 𝑬\ {0} then 𝑬 is totally real, i.e., 𝑬 ∩ −1𝑬 = {0}, and therefore
dimR 𝑬 ≤ 𝑛.

Proof Let 𝑬 be as in the statement. From the decomposition (11.2.7) we deduce

that 𝑧 ∈ 𝑬\ {0} =⇒ Re ℎ𝑄 (𝑧) ≤ 𝑄 (𝑧) < 0 and therefore

𝑄 (𝑖𝑧) = − Re ℎ𝑄 (𝑧) + 𝐿 (𝑧) > 𝐿 (𝑧) ≥ 0,

implying 𝑬 ∩ 𝑖𝑬 = {0} and 𝜈+ (𝑄) ≥ 𝜈− (𝑄). □

Thus, if 𝑄 is plurisubharmonic and nondegenerate then

sign 𝑄 = 𝜈+ (𝑄) − 𝜈− (𝑄) = 0 ⇐⇒ 𝜈+ (𝑄) = 𝜈− (𝑄) = 𝑛. (11.2.9)

If (11.2.9) holds there is an 𝑛-dimensional totally real vector subspace 𝑬 of C𝑛 such

that 𝑄 (𝑧) < 0 if 0 ≠ 𝑧 ∈ 𝑬.

Corollary 11.2.27 If the quadratic form 𝑄 is plurisubharmonic, nondegenerate and

sign 𝑄 = 0 then the same is true of every plurisubharmonic quadratic form 𝑄 ♭ in C𝑛
such that 𝑄 ♭ ≤ 𝑄.

Proof Since 𝑄 ♭ ≤ 𝑄 we have necessarily 𝜈− 𝑄 ♭ ≥ 𝜈− (𝑄) = 𝑛. But 𝜈− 𝑄 ♭ ≤

𝜈+ 𝑄 ♭ ≤ 𝑛 by Proposition 11.2.26. □

There is a C-linear change of variables that transforms 𝐿 (𝑧) into a sum

𝑘
∑︁ 𝑚
∑︁
2 2
𝑧𝑗 − 𝑧 𝑗 , 𝑚 ≤ 𝑛.
𝑗=1 𝑗=𝑘+1

Í 2
If 𝑄 is plurisubharmonic, this form is 𝑘𝑗=1 𝑧 𝑗 ; 𝑄 is strictly plurisubharmonic if
and only if 𝑘 = 𝑛. There is also a complex orthogonal change of coordinates that
transforms ℎ𝑄 into a sum ℓ𝑗=1 𝑐 𝑗 𝑧 2𝑗 , 𝑐 𝑗 ∈ C. The following precise “diagonalization”
Í
of the whole form 𝑄 will be helpful later.

Lemma 11.2.28 If the real quadratic form 𝑄 is nondegenerate and sign 𝑄 = 0 then
there is a 𝑇 ∈ GL (𝑛, C) such that
𝑛
∑︁ 𝑛
∑︁
𝑄 (𝑇 𝑧) = − 𝜒 𝑗 𝑥 2𝑗 + 𝜒 ′𝑗 𝑦 2𝑗 (11.2.10)
𝑗=1 𝑗=1
344 11 Distributions and Hyperfunctions on a Manifold

with
𝜒 𝑗 > 0, 𝜒 ′𝑗 > 0, 𝑗 = 1, ..., 𝑛. (11.2.11)
For 𝑄 to be plurisubharmonic (Definition 11.2.6) it is necessary and sufficient
that 0 < 𝜒 𝑗 ≤ 𝜒 ′𝑗 , 𝑗 = 1, ..., 𝑛. For 𝑄 to be strictly plurisubharmonic (Definition
11.2.11) it is necessary and sufficient that 0 < 𝜒 𝑗 < 𝜒 ′𝑗 , 𝑗 = 1, ..., 𝑛. For 𝑄 to be
pluriharmonic it is necessary and sufficient that 𝜒 𝑗 = 𝜒 ′𝑗 for every 𝑗 = 1, ..., 𝑛.

𝑥 𝑥
Proof We can write 𝑄 (𝑧) = ·𝑆 , 𝑆 ∈ M2𝑛 (R) symmetric. Let 𝑬 + (resp.,
𝑦 𝑦
𝑬 − ) be the real vector subspace of R2𝑛 spanned by the eigenvectors of 𝑆 correspond-
ing to the positive (resp., negative) eigenvalues of 𝑆; we have dim 𝑬 + = dim 𝑬 − = 𝑛
and 𝑬 + ⊥𝑬 − . By Proposition 11.2.26 both 𝑬 + and 𝑬 − are totally real. For each
𝑗 = 1, ..., 𝑛 let u 𝑗 ∈ 𝑬 − be a unit eigenvector of 𝑆 corresponding to the eigenvalue
−𝜒 𝑗 < 0. There is a unitary transformation of C𝑛 which transforms u1 , ..., u𝑛 into
the canonical basis e1 , ..., e𝑛 ; it transforms 𝑬 + into 𝑖R𝑛 . We may as well assume that
u 𝑗 = e 𝑗 for every 𝑗 = 1, ..., 𝑛, in which case

−Δ◦ 0
𝑆=
0 𝑆◦

where Δ◦ is the diagonal matrix whose diagonal entries are the positive numbers
𝜒1 , ..., 𝜒𝑛 and 𝑆◦ is a real symmetric 𝑛 × 𝑛 matrix; 𝑆◦ has positive eigenvalues
1
𝜒1′ , ..., 𝜒𝑛′ . Let Δ◦2 denote the positive square root of Δ◦ ; it is a real diagonal 𝑛 ×
−1
𝑛 matrix. We may regard the C-linear transformation 𝑧 ↦→ Δ◦ 2 𝑧 as an R-linear
transformation of R2𝑛 , yielding

− 21 𝑥 ♭ 𝑥
𝑄 Δ◦ 𝑧 = ·𝑆
𝑦 𝑦

with !
♭
−𝐼𝑛 0
𝑆 = −1 −1 .
0 Δ◦ 2 𝑆◦ Δ◦ 2
−1 −1
There is a real orthogonal 𝑛 × 𝑛 matrix Γ such that Γ−1 Δ◦ 2 𝑆◦ Δ◦ 2 Γ is a diagonal
𝑛 × 𝑛 matrix (with diagonal entries 𝜒 ′𝑗 /𝜒 𝑗 – possibly after a relabeling of the 𝜒 ′𝑗 ).
We have 1 ! 1 !
Δ◦2 Γ−1 0 2

♭ ΓΔ◦ 0 −Δ◦ 0
𝑆 =
1
0 Δ◦2 Γ−1 0 ΓΔ◦2
1
0 Δ◦′

where Δ◦′ is the diagonal matrix with diagonal entries 𝜒 ′𝑗 . We conclude that (11.2.10)
−1 1
holds with 𝑇 = Δ◦ 2 ΓΔ◦2 . Since 𝑄 (𝑧) is plurisubharmonic the same is true of 𝑄 (𝑇 𝑧)
by Corollary 11.2.10, whence, for every 𝑤 ∈ C𝑛 ,
11.3 Hyperfunctions and Microfunctions in an Analytic Manifold 345
𝑛 𝑛
∑︁ 𝜕2 ∑︁ 2 𝜕2
𝑤 𝑗 𝑤¯ 𝑘 𝑄 (𝑇 𝑧) = 𝑤𝑗 𝑄 (𝑇 𝑧)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 𝑗=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑗
𝑛
∑︁
2
=2 𝜒 ′𝑗 − 𝜒 𝑗 𝑤 𝑗 ,
𝑗=1

which completes the proof of the claims. □

11.3 Hyperfunctions and Microfunctions in an Analytic Manifold

11.3.1 Hyperfunctions in an orientable C 𝝎 manifold

Throughout this section M shall be a (paracompact) orientable C 𝜔 manifold. There

are different ways of approaching the concept of hyperfunctions in M. The most
direct is to use analytic local charts to transfer the sheaves we are interested in from
open subsets of R𝑛 , 𝑛 = dim M. We use a covering of M by domains U 𝜄 (𝜄 in a set of
indices 𝐼) of local analytic coordinates 𝑥 𝜄,1 , ..., 𝑥 𝜄,𝑛 all having the same orientation,
meaning that the Jacobian determinants in the overlaps U 𝜄 ∩ U 𝜄′ (𝜄, 𝜄′ ∈ 𝐼) are all
positive. We denote by 𝜙 𝜄 the C 𝜔 diffeomorphism ℘ ↦→ 𝑥 𝜄,1 (℘) , ..., 𝑥 𝜄,𝑛 (℘)
of U 𝜄 onto 𝑈 𝜄 = 𝜑 𝜄 (U 𝜄 ) and by B (U 𝜄 ) the pullback of B (R𝑛 )|𝑈 𝜄 to U 𝜄 . The
only question is whether the local sheaves B (U 𝜄 ) can be patched together over the
overlaps U 𝜄 ∩ U 𝜄′ . That they can is a direct consequence of the following

Proposition 11.3.1 Let 𝑈 be an open subset of R𝑛 . The stalk B 𝑥 at a point 𝑥 ∈ 𝑈 is

independent of the choice of the analytic coordinates in 𝑈.

Proof A C 𝜔 diffeomorphism 𝜒 of 𝑈 onto another neighborhood of 𝑥 ◦ , 𝑉, preserving

𝑥 ◦ and whose Jacobian determinant in positive in 𝑈, extends as a biholomorphism,
which we also denote by 𝜒, of an open subset 𝑈 C of C𝑛 such that 𝑈 C ∩ R𝑛 = 𝑈 onto
an open subset 𝑉 C of C𝑛 such that 𝑉 C ∩ R𝑛 = 𝑉. Let then 𝑈 ′ ⊂ 𝑈 be a neighborhood
of 𝑥 ◦ , Γ an acute open cone in R𝑛 \ {0} and 𝛿 > 0 such that the closure of the wedge
W𝛿 (𝑈 ′, Γ) [cf. (3.3.1)] is a compact subset of 𝑈 C . Since the Jacobian determinant
of 𝜒 does not vanish there is a neighborhood of 𝑥 ◦ , 𝑉 ′ ⊂ 𝜒 (𝑈 ′), an acute open cone
Γ ′ and a number 𝛿 ′ > 0 such that

W𝛿′ (𝑉 ′, Γ ′) ⊂ 𝜒 (W𝛿 (𝑈 ′, Γ)) .

−1
The pullback under 𝜒 induces an isomorphism of O (W𝛿 (𝑈 ′, Γ)) onto a linear
subspace of O (W𝛿′ (𝑉 ′, Γ ′)) and therefore, by the boundary value map (see Propo-
sition 7.2.5), onto a linear subspace of B (𝑈 ′), the space of hyperfunctions in 𝑈 ′.
By Theorems 7.4.4 and 7.5.1(cf. Remark 7.5.4), the pullback map defines a linear
−1
injection of the stalk B 𝑥 into itself. Exchanging 𝜒 and 𝜒 shows that this injection is
a bijection. □
346 11 Distributions and Hyperfunctions on a Manifold

Remark 11.3.2 Proposition 11.3.1 is also a direct consequence of Formula (10.3.18).

We shall denote by B (M) the sheaf of hyperfunctions in M. If M ′ is an open

subset of M we denote by B (M ′) the vector space of continuous sections of B (M)
over M ′ and we refer to any element of B (M ′) as a hyperfunction in M ′. By the
support of a hyperfunction 𝑢 in M we shall mean the set of points of M at which
𝑢 ≠ 0, denoted by supp 𝑢 in the sequel; supp 𝑢 is a closed subset of M (cf. Definition
7.2.3).

Proposition 11.3.3 The sheaf B (M) is flabby.

Proof Let M ′ be an open subset of M U and let 𝒔 ∈ B (M ′) be arbitrary. By the

argument in the proof of Proposition 7.1.9, we see that there is a maximal open
subset M◦′ of M, M◦′ ⊃ M ′, to which 𝒔 extends as a continuous section of the
sheaf B (M). If M◦′ is a connected component of M then we can extend 𝒔 to the
whole of M by setting 𝒔 = 0 in the other connected components of M. If M◦′ is not
a connected component of M there is an 𝑥 ◦ ∈ 𝜕M◦′ ; in this case there is an open
subset U of M containing 𝑥 ◦ as well as points of M\M◦′ . By assuming that U is the
domain of local analytic coordinates we can transfer the analysis to R𝑛 ; when this is
done we see (by Proposition 7.1.9) that the restriction to U ∩ M ◦′ of the section 𝒔
can be extended to the whole of U and, as a consequence, that 𝒔 can be extended to
M◦′ ∪ U, contradicting the maximality of M◦′ . □
Using local analytic coordinates 𝑥 1 , ..., 𝑥 𝑛 in an open subset U of M and the dual
coordinates, 𝜉1 , ..., 𝜉 𝑛 in the cotangent bundle over U, 𝑇 ∗ M| U , i.e., the coordinates
with respect to the basis d𝑥 1 , ..., d𝑥 𝑛 in the cotangent spaces at points of U, we can
identify 𝑇 ∗ M| U with U × R𝑛 and apply Definition 7.4.7, to define the analytic wave
front set over U of an arbitrary hyperfunction 𝑓 ∈ B (M). Inspection of the proof of
Proposition 11.3.1 shows easily that this concept is independent of the choice of local
analytic coordinates. Thus the local affinizations can be patched up together, making
it possible to define globally the analytic wave front set of 𝑓 , 𝑊 𝐹a ( 𝑓 ), a closed
conic subset of 𝑇 ∗ M\0. If (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a ( 𝑓 ) we shall say that 𝑓 is microanalytic
at (𝑥 ◦ , 𝜉 ◦ ).

11.3.2 Singularity hyperfunctions in a manifold

We regard the sheaf of germs of analytic functions in the analytic manifold M,

C 𝜔 (M), as a subsheaf of B (M). We form the quotient sheaf Bsing (M) =
B (M) /C 𝜔 (M); in other words, the following sequence of sheaf homomorphisms
is exact:
0 −→ C 𝜔 (M) −→ B (M) −→ Bsing (M) → 0. (11.3.1)

Definition 11.3.4 The continuous sections of Bsing (M) over an open subset M ′ of
M will be called singularity hyperfunctions in M ′. We shall denote by B sing (M ′)
the vector space of singularity hyperfunctions in M ′.
11.3 Hyperfunctions and Microfunctions in an Analytic Manifold 347

The quotient sheaf map B (M) −→ B (M) /C 𝜔 (M) induces a natural map of
the continuous sections, B (M ′) → B sing (M ′); the image of 𝑢 ∈ B (M ′) under this
map shall be denoted by sing 𝑢. Theorem 7.1.27 can be generalized to the manifold
M. This is a consequence of Theorem 10.2.8, combined with the Grauert result that
M can be embedded as an analytic submanifold of a Euclidean space of sufficiently
high dimension (see [Grauert, 1958]).

Proposition 11.3.5 The linear map B (M ′) ∋ 𝑢 ↦→ sing 𝑢 ∈ B sing (M ′) induces an

isomorphism of the quotient vector space B (M ′) /C 𝜔 (M ′) onto B sing (M ′).

Proof The map B (M ′) ∋ 𝑢 ↦→ sing 𝑢 ∈ B sing (M ′) is obviously injective (a

hyperfunction C 𝜔 at every point of M ′ is a C 𝜔 function in M ′). To prove that it
is surjective consider an arbitrary 𝑓 ♮ ∈ B sing (M ′). There exist an open covering
{U 𝜄 } 𝜄 ∈𝑰 of M ′ and for each 𝜄 ∈ 𝐼, a hyperfunction 𝑓 𝜄 representing 𝑓 ♮ in U 𝜄 , such
that, for every pair of indices 𝜄, 𝜄′ ∈ 𝑰, the restriction of ℎ 𝜄, 𝜄′ = 𝑓 𝜄 − 𝑓 𝜄′ to U 𝜄 ∩U 𝜄′ ≠ ∅
belongs to C 𝜔 (U 𝜄 ∩ U 𝜄′ ). The ℎ 𝜄, 𝜄′ define a 2-cocycle in the cohomology of M ′
with values in the sheaf C 𝜔 . The cohomology with C 𝜔 coefficients of an open subset
of M is trivial, according to the results of Grauert alluded to above: it follows that
to each 𝜄 ∈ 𝑰 there is an ℎ 𝜄 ∈ C 𝜔 (U 𝜄 ) such that ℎ 𝜄 − ℎ 𝜄′ = ℎ 𝜄, 𝜄′ in U 𝜄 ∩ U 𝜄′ for all
pairs (𝜄, 𝜄′) ∈ 𝑰 2 . We introduce the hyperfunctions 𝑔 𝜄 = 𝑓 𝜄 − 21 ℎ𝑖 𝑓 𝜄 mod C 𝜔 (U 𝜄 ).
We have 𝑔 𝜄 = 𝑔 𝜄′ in U 𝜄 ∩ U 𝜄′ , implying the existence of a hyperfunction 𝑔 whose
restriction to U 𝜄 is equal to 𝑔 𝜄 for every 𝜄. It is clear that sing 𝑔 = 𝑓 ♮ . □

Corollary 11.3.6 The sheaf Bsing (M) is flabby.

Proof Let N be an arbitrary open subset of M. By Proposition 11.3.5 we know that

each element of B sing (N ) is of the form sing 𝑓 , 𝑓 ∈ B (N ). If 𝑓˜ ∈ B (M) extends
𝑓 then sing 𝑓˜ extends sing 𝑓 . □
We refer to supp (sing 𝑢) as the analytic singular support of 𝑢 ∈ B (M) and
denote it by singsuppa 𝑓 (cf. Definition 7.4.7).
Now let us reason in an analytic local chart (M ′, 𝑥1 , ..., 𝑥 𝑛 ) in M and, in fact, use
the coordinates 𝑥 𝑗 to transfer the analysis to an open subset Ω of R𝑛 . We use the (lo-
cal) representation of hyperfunctions as boundary values of holomorphic functions
(Definition 7.2.3) in a wedge W𝛿 (𝑈, Γ) = {𝑧 ∈ C𝑛 ; 𝑥 ∈ 𝑈, 𝑦 ∈ Γ, |𝑦| < 𝛿} (𝑈 an
open subset of Ω, Γ ⊂ R𝑛 \ {0} an open cone, 𝛿 > 0). The following statement is
self-evident.

Proposition 11.3.7 If 𝑓 𝑗 ∈ B (Ω), 𝑗 = 1, 2, satisfy sing 𝑓1 = sing 𝑓2 and if 𝑓1 is

equal in the open set 𝑈 ⊂⊂ Ω to the boundary value of a holomorphic function in
W𝛿 (𝑈, Γ) for some 𝛿 > 0 then the same is true of 𝑓2 .

Proposition 11.3.7 allows us to define the analytic wave-front set of 𝑓 ♮ = sing 𝑓

as that of 𝑓 (one of the hyperfunction
representatives of 𝑓 ♮ , selected arbitrarily); it
shall be denoted by 𝑊 𝐹a 𝑓 .♮

By a subcone of Γ we shall mean a cone in R𝑛 \ {0} contained in Γ.

348 11 Distributions and Hyperfunctions on a Manifold

Proposition 11.3.8 Let 𝔉Ω be a basis of the topology of Ω consisting of open subsets

of Ω; let 𝔉Γ be a family of open subcones of Γ such that every open subcone of Γ
is a union of cones Γ ′ ∈ 𝔉Γ . If to every Ω′ ∈ 𝔉Ω and every Γ ′ ∈ 𝔉Γ there exist
a wedge W𝛿 (Ω′, Γ ′) and a holomorphic function in W𝛿 (Ω′, Γ ′) whose boundary
value represents 𝑓 ♮ ∈ B sing (Ω) in Ω′ then
there
C 𝑛
subset Ω of C such
exist√an open
that Ω ⊂ Ω ∩R and a function ℎ ∈ O Ω ∩ Ω + −1Γ such that the boundary
C 𝑛 C

value of ℎ in every open set 𝑈 ⊂⊂ Ω represents 𝑓 ♮ in 𝑈.

Proof Let 𝑓 ∈ B (Ω) be a representative of 𝑓 ♮ ; by Proposition 11.3.7 our hypothesis
is that to every pair (𝑈, Γ ′), 𝑈 ∈ 𝔉Ω , Γ ′ ∈ 𝔉Γ there exist a wedge W𝛿 (𝑈, Γ ′)
and ℎ𝑈,Γ′ ∈ O (W𝛿 (𝑈, Γ)) such that 𝑓 |𝑈 = 𝑏𝑈 ℎ𝑈,Γ′ . Suppose 𝑈1 and 𝑈2 are two
elements of 𝔉Ω and Γ1 and Γ2 two elements of 𝔉Γ such that 𝑈1 ∩𝑈2 ≠ ∅, Γ1 ∩ Γ2 ≠ ∅
and ℎ 𝑗 ∈ O W𝛿 𝑈 𝑗 , Γ 𝑗 is such that 𝑓 |𝑈 = 𝑏𝑈 𝑗 ℎ 𝑗 in 𝑈 𝑗 ( 𝑗 = 1, 2). It follows from
Proposition 7.2.5 that the restrictions of ℎ1 and ℎ2 to W𝛿 (𝑈1 ∩ 𝑈2 , Γ1 ∩ Γ2 ) are
equal and, consequently, that there is an ℎ1,2 ∈ O (W𝛿 (𝑈1 ∪ 𝑈2 , Γ1 ∪ Γ2 )) whose
restriction to W𝛿 𝑈 𝑗 , Γ is equal to ℎ 𝑗 for 𝑗 = 1, 2. The claim follows readily from
this fact. □
If the conclusion in Proposition 11.3.8 holds we shall say that 𝑓 ♮ ∈ B sing (Ω) is
the boundary value of ℎ ∈ O (W (Ω, Γ)) and write 𝑓 ♮ = 𝑏 Ω ℎ.

11.3.3 Microfunctions in a manifold

Just as in Euclidean space it is convenient, here, to carry out the microlocal analysis
in the cosphere bundle 𝑆 ∗ M of M (Definition 9.3.3) rather than in 𝑇 ∗ M\0; recall
that 𝑆 ∗ M is the quotient of 𝑇 ∗ M\0 for the equivalence relation (𝑥, 𝜉) (𝑦, 𝜂) if
𝑥 = 𝑦 and 𝜂 = 𝑐𝜉 for some 𝑐 > 0. Over the domain U of local analytic coordinates
we can identify 𝑆 ∗ M with U × S𝑛−1 .
Definition 11.3.9 By the microsupport of 𝑓 ∈ B (M), in the sequel denoted
by microsupp 𝑓 , we shall mean the image of 𝑊 𝐹a ( 𝑓 ) under the quotient map
𝑇 ∗ M\0 −→ 𝑆 ∗ M.
We also define microsupp (sing 𝑓 ) = microsupp 𝑓 (see the preceding subsection);
microsupp 𝑓 is a closed subset of 𝑆 ∗ M.
Proposition 11.3.10 Let 𝑃 (𝑥, D) be a linear differential operator with analytic co-
efficients in an open subset M ′ of M; then microsupp 𝑃 (𝑥, D) 𝑓 ⊂ microsupp 𝑓 for
every 𝑓 ∈ B (M ′).
Proof It suffices to prove the claim locally. Using local coordinates we transfer
the analysis to an open subset Ω of R𝑛 ; if 𝑃 (𝑧, D𝑧 ) denotes the holomorphic ex-
tension of 𝑃 (𝑥, D 𝑥 ) to an open subset ΩC of C𝑛 such that Ω ⊂ ΩC ∩ R𝑛 then
𝑃 (𝑧, D𝑧 ) O (W𝛿 (𝑈, Γ)) ⊂ O (W𝛿 (𝑈, Γ)) for every wedge W𝛿 (𝑈, Γ) ⊂ ΩC .
From this fact and Definition 7.4.7, the claim follows easily. □
11.3 Hyperfunctions and Microfunctions in an Analytic Manifold 349

We may use the affinization (U, 𝑥1 , ..., 𝑥 𝑛 ) to pullback to U the sheaf of micro-
functions in the image 𝑈 of U under the coordinate map. Or we may define directly
the presheaf of microfunctions in 𝑆 ∗ M and thence consider the sheaf it defines.
Thus let U † be an open subset of 𝑆 ∗ M; we shall denote by A U † the linear space
of hyperfunctions 𝑓 in M such that U † ∩ microsupp 𝑓 = ∅. One can refer to the
elements of A U † as the hyperfunctions in M microanalytic in U † . We define

B micro U † = B (M) /A U † . (11.3.2)

Definition 11.3.11 The elements of the quotient vector space B micro U † are called

microfunctions in U † .

The support of [ 𝑓 ] ∈ B micro U † , denoted by supp [ 𝑓 ], is equal to U † ∩

microsupp 𝑓 , where 𝑓 ∈ B (M) is an arbitrary representative of [ 𝑓 ]. A singularity
hyperfunction 𝑓 ♭ in M defines a microfunction 𝑓 ♭ in 𝑆 ∗ M; the base projection


of supp 𝑓 ♭ is equal to supp 𝑓 ♭ .
We shall also use the following terminology.

Definition 11.3.12 A microfunction [ 𝑓 ] in U † will be called a microdistribution

if [ 𝑓 ] has a representative 𝑓 ∈ B (M) whose restriction to the base projection of
U † is a distribution.

If U † ⊂ V † are two open subsets of 𝑆 ∗ M there is a natural restriction map

V†
A V † −→ A U † and therefore natural restriction maps 𝑟 U † −→

† : A V
V†
A U† , 𝜌U micro V † −→ B micro U † . The two families of pairs

† : B

V† V†
A U† , 𝑟 U † , B micro U † , 𝜌 U †

form two presheaves on 𝑆 ∗ M: respectively, the sheaf of germs of microanalytic

hyperfunctions in M which we denote by Amicro (𝑆 ∗ M); the sheaf of germs of
∗ M which we denote by Bmicro (M); then B micro U † [resp.,

microfunctions in 𝑆
A U † ] can be identified with the linear space of continuous sections of Bmicro (M)

[resp., Amicro (𝑆 ∗ M)] over U † .
The stalk of Bmicro (M) at (𝑥 ◦ , 𝜃 ◦ ) ∈ 𝑆 ∗ M will be denoted by B 𝑥 ◦ , 𝜃 ◦ ; its elements
are the germs of microfunctions at (𝑥 ◦ , 𝜃 ◦ ). These, obviously, can be defined directly:
they are the equivalence classes of pairs [ 𝑓 ] , U † , [ 𝑓 ] ∈ B micro U † , U † an open

subset of 𝑆 ∗ M, for the relation [ 𝑓1 ] , U1† [ 𝑓2 ] , U2† , meaning that there is

V † ⊂ U †1 ∩ U2† such that 𝑓1 − 𝑓2 ∈ A V † ( 𝑓𝑖 : a representative of the coset [ 𝑓𝑖 ],

𝑖 = 1, 2). Likewise for the stalk Amicro micro (𝑆 ∗ M) at (𝑥 ◦ , 𝜃 ◦ ). Theorem 7.6.6,

𝑥 ◦ , 𝜃 ◦ of A
implies directly (cf. the proof of Proposition 11.3.3)

Proposition 11.3.13 The sheaf Bmicro (M) is flabby: every microfunction in an open
subset of 𝑆 ∗ M can be extended as a microfunction in the whole of 𝑆 ∗ M.
350 11 Distributions and Hyperfunctions on a Manifold

A linear differential operator 𝑃 (𝑥, D) with analytic coefficients in an open subset

M ′ of M can be made to act on a microfunction in 𝑆 ∗ M| M ; by acting on any
hyperfunction in M ′ representing it. This is a direct consequence of Proposition
11.3.10.
The following sequence of sheaf homomorphisms is exact:

0 −→ Amicro (𝑆 ∗ M) −→ B (M) −→ Bmicro (M) → 0. (11.3.3)

Remark 11.3.14 We may regard (11.3.3) as an exact sequence of sheaves of vector

spaces on one and the same topological space, 𝑆 ∗ M, by identifying B (M) with the
pullback of the sheaf B (M) on M under the base projection 𝑆 ∗ M −→ M; in other
words, by setting B 𝑥 ◦ ,𝜃 = B 𝑥 ◦ for all 𝜃 ∈ S𝑛−1 .

The interpretation of hyperfunctions and microfunctions as relative cohomology

classes (see Section 10.3) can be extended from R𝑛 to an arbitrary C 𝜔 manifold M
by using local analytic coordinates. Thus the vector spaces

𝐻 𝑛 [U] = 𝐻 𝑛,𝑛−1 M\U /𝐻 𝑛,𝑛−1 (M\𝜕U) ,

where U is an arbitrary open subset of M with compact closure, together with the
related restriction mappings, define a sheaf on M which we denote by H𝑛 [M]
(𝑛 = dimR M). We have
B (M) H𝑛 [M] . (11.3.4)
Likewise, we can use the isomorphisms (10.3.34) in an arbitrary local C 𝜔 coordinate
system (U, 𝑥1 , ..., 𝑥 𝑛 ) to define a sheaf H1rel (𝑆 ∗ M) (𝑆 ∗ M: cosphere bundle of M)
and use the local analogues of the isomorphisms in Theorem 10.1.1, to get a sheaf
isomorphism
Bmicro (M) H1rel (𝑆 ∗ M) . (11.3.5)
Chapter 12
Lie Algebras of Vector Fields

First-order linear PDEs are among the simplest PDEs (contenders for the simplicity
title are linear PDEs with constant coefficients). If a first-order PDO does not have a
zero-order term it can be regarded as a vector field, with an expression 𝑎 1 (𝑥) 𝜕𝑥𝜕 1 +
· · · + 𝑎 𝑛 (𝑥) 𝜕𝑥𝜕𝑛 in local coordinates. The vector fields of greatest interest to us
will have analytic coefficients 𝑎 𝑗 (𝑥) although the statements about them will, not
infrequently, be valid for C ∞ or much less “regular” vector fields. In the two most
basic results of this chapter, Frobenius’ Theorem (Theorem 12.2.5) and the Nagano
Theorem about analytic foliations (Theorem 12.4.2) these coefficients are required
to be real if the base manifold M is real. In the overall perspective of the book this
will not be the rule. Unlike in other areas of geometric analysis, the coefficients 𝑎 𝑗
will be allowed to be complex-valued even in domains in R𝑛 . The importance of
first-order linear PDEs is that many of the conjectures about linear PDEs of higher
order (with complex coefficients) can be tested on the first-order equations. And
some of the deepest discoveries about linear PDEs in the second half of the XXth
century were made on first-order models. To appreciate this fact just think of the
nonsolvability of the Lewy equation (see [Lewy, 1956]; much more on this in Part
VII). The usefulness as testing grounds is even more true of systems of PDEs and
we shall pay considerable attention to systems of analytic vector fields, both in the
C 𝜔 class and in the holomorphic class. To jump ahead (to Parts V and VII) this is
reinforced by the availability of the tools of microlocal analysis, which enable us to
transform into systems of complex vector fields, locally in phase-space, an important
class of systems of linear PDOs of higher order.
One demand on the vector fields generating the “systems” that is quasi-absolute
is that they must not have critical points, meaning that they not vanish at any point of
their domain of definition. The presence of critical points upends our entire approach
and the truth is that almost nothing is known about PDEs (of order ≥ 1, in dimension
≥ 2) that strongly degenerate at a point, concerning questions of solvability and
regularity of generalized solutions, essentially the topic of this book.
Among the topics touched upon in the last section of this chapter are elliptic
systems and systems of vector fields of finite type. Our definition of the latter is the
simplest, namely that the leaves of the associate foliation be open subsets of the base

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 351
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_12
352 12 Lie Algebras of Vector Fields

manifold. The concept of “type” is significant in the analysis of subelliptic PDEs (see
Ch. 24) and in Cauchy–Riemann geometry (see, e.g., [Kohn, 1972], [Bloom-Graham,
1977], [D’Angelo, 1979]).

12.1 The Lie Algebra of Smooth Vector Fields

Let M be a regular manifold over the field K (i.e., R or C); in this chapter the capitals
𝑋, 𝑌 , etc., shall denote smooth vector fields in some open subset 𝑈 of M, i.e., C ∞
sections over 𝑈 of the tangent bundle 𝑇M. We shall also consider complex vector
fields, i.e. vector fields with complex coefficients, even when M is a real manifold;
they are smooth sections of the complexification C𝑇M. We can identify vector fields
to first-order C ∞ differential operators without zero-order terms and then form their
commutator:
[𝑋, 𝑌 ] 𝑓 = 𝑋 (𝑌 𝑓 ) − 𝑌 (𝑋 𝑓 ) , 𝑓 ∈ C ∞ (𝑈) . (12.1.1)
We shall often refer to [𝑋, 𝑌 ] as the Lie bracket of 𝑋 and 𝑌 ; it is also a smooth
vector field in 𝑈. Indeed, suppose that 𝑈 is the domain of local coordinates 𝑥1 , ..., 𝑥 𝑛
and that, in those coordinates,
𝑛 𝑛
∑︁ 𝜕 ∑︁ 𝜕
𝑋= 𝑎 𝑗 (𝑥) ,𝑌 = 𝑏 𝑗 (𝑥) . (12.1.2)
𝑗=1
𝜕𝑥 𝑗 𝑗=1
𝜕𝑥 𝑗

It is seen at once that

𝑛
∑︁ 𝜕𝑏 𝑘 𝜕𝑎 𝑘 𝜕
[𝑋, 𝑌 ] = 𝑎 𝑗 (𝑥) − 𝑏 𝑗 (𝑥) . (12.1.3)
𝑗,𝑘=1
𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑘

We shall sometimes use the notation (Ad 𝑋) 𝑌 = [𝑋, 𝑌 ]. With the Lie bracket the
vector space C ∞ (𝑈; 𝑇M) becomes an infinite-dimensional Lie algebra.
By the symbol of a vector √ field 𝑋 we shall always mean its principal symbol
(Definition 1.3.1), namely −1⟨𝜎, 𝑋⟩, a well-defined function in the cotangent
bundle (𝜎 is the tautological one-form; see Subsection 9.3.3). In a coordinate patch
(𝑈, 𝑥1 , ..., 𝑥 𝑛 ) where 𝑋 and 𝑌 have the expressions (12.1.2) we have

√ 𝑛
∑︁
−1⟨𝜎, 𝑋⟩ = 𝑎 𝑗 (𝑥) 𝜉 𝑗 ; (12.1.4)
𝑗=1
√ 𝑛
∑︁ 𝜕𝑏 𝑘 𝜕𝑎 𝑘
−1⟨𝜎, [𝑋, 𝑌 ]⟩ = 𝑎 𝑗 (𝑥) − 𝑏 𝑗 (𝑥) 𝜉𝑘 (12.1.5)
𝑗,𝑘=1
𝜕𝑥 𝑗 𝜕𝑥 𝑗
= − {⟨𝜎, 𝑋⟩, ⟨𝜎, 𝑌 ⟩} ,

where {·, ·} is the Poisson bracket (see below, Section 13.3.2).

12.1 The Lie Algebra of Smooth Vector Fields 353

The two basic algebraic properties of the Lie bracket are: 1) its skew-symmetry:
[𝑋, 𝑌 ] = − [𝑌 , 𝑋]; 2) the Jacobi identity: if 𝑍 is a third smooth vector field in 𝑈 then

[𝑋, [𝑌 , 𝑍]] + [𝑌 , [𝑍, 𝑋]] + [𝑍, [𝑋, 𝑌 ]] = 0, (12.1.6)

which can be rewritten as (Ad 𝑋) [𝑌 , 𝑍] = [(Ad 𝑋) 𝑌 , 𝑍]+[𝑌 , (Ad 𝑋) 𝑍], the Leibniz
rule for the derivation Ad 𝑋 of the Lie algebra C ∞ (𝑈; 𝑇M).
Let 𝑋1 , ..., 𝑋𝑟 be C ∞ vector fields in the manifold M.

Notation 12.1.1 Given any nonempty set 𝐼 = (𝑖 1 , ..., 𝑖ℓ ) of integers 𝑖 𝛼 , 1 ≤ 𝑖 𝛼 ≤ 𝑟,

we write
𝑋𝐼 = Ad 𝑋𝑖1 · · · Ad 𝑋𝑖ℓ−1 𝑋𝑖ℓ . (12.1.7)

We refer to ℓ = |𝐼 | as the length of the multibracket (12.1.7); 𝑋𝐼 = 𝑋𝑖1 if ℓ = 1.

Notation 12.1.2 We shall denote by 𝔤 (𝑋1 , ..., 𝑋𝑟 ) the span over K of the multi-
brackets (12.1.7) of any length. If ℘ ∈ M we shall denote by 𝔤℘ (𝑋1 , ..., 𝑋𝑟 ) the
linear subspace of 𝑇℘ M consisting of the “freezing” at ℘ of the vector fields 𝑌 ∈
𝔤 (𝑋1 , ..., 𝑋𝑟 ).

Proposition 12.1.3 The Lie bracket makes a Lie algebra of 𝔤 (𝑋1 , ..., 𝑋𝑟 ).

Proof Let the multi-index 𝐼 be as in (12.1.7) and let 𝐽 = ( 𝑗1 , ..., 𝑗 𝑚 ), 1 ≤ 𝑗 𝛽 ≤ 𝑟.

We must show that [𝑋𝐼 , 𝑋 𝐽 ] is a linear combination (with coefficients in K) of
multibrackets. The claim is trivial when ℓ = 1. If ℓ ≥ 2 the Jacobi identity entails

[𝑋𝐼 , 𝑋 𝐽 ] = Ad 𝑋𝑖1 [𝑋𝐼 ′ , 𝑋 𝐽 ] − 𝑋𝐼 ′ , Ad 𝑋𝑖1 𝑋 𝐽 ,

where 𝐼 ′ = (𝑖 2 , ..., 𝑖ℓ ). The general result ensues by induction on ℓ. □

Keep in mind that dim 𝔤℘ (𝑋1 , ..., 𝑋𝑟 ) can vary with ℘: in general, the subspaces
𝔤℘ (𝑋1 , ..., 𝑋𝑟 ) do not make up a vector bundle over M.
The Ad operation and the exterior derivative of a 𝑞-form 𝑓 are related. The
evaluation of d 𝑓 on the 𝑞 + 1 regular vector fields 𝑋0 , 𝑋1 , ..., 𝑋𝑞 is given by
𝑞
∑︁
d 𝑓 𝑋0 , 𝑋1 , ..., 𝑋𝑞 = (−1) 𝑗 𝑋 𝑗 𝑓 𝑋0 , ..., c
𝑋 𝑗 , ..., 𝑋𝑞 (12.1.8)
𝑗=0
∑︁
+ (−1) 𝑗+𝑘 𝑓 𝑋 𝑗 , 𝑋 𝑘 , 𝑋0 , ..., c
𝑋 𝑗 , ..., 𝑋
c𝑘 , ..., 𝑋𝑞
0≤ 𝑗<𝑘 ≤𝑞

(the hatted factors must be omitted). The proof of (12.1.8) is straightforward (see
[Spivak, 1979], p. 289–291).
354 12 Lie Algebras of Vector Fields

12.2 Integral Manifolds. Frobenius’ Theorem

12.2.1 Frobenius’ Theorem

We continue to deal with a regular manifold M; we assume that M is countable

at infinity. Keep in mind that “regular” means either real of class C ∞ or C 𝜔 when
the base field K = R, or complex-analytic when K = C. In the latter case 𝑇℘ M and
𝑇℘∗ M are the complex tangent and cotangent spaces at the point ℘ ∈ M.

Definition 12.2.1 Let A be a set of regular vector fields in M. For each point ℘ ∈ M
let A ℘ ⊂ 𝑇℘ M be the set of values 𝑋℘ , 𝑋 ∈ A. By an integral manifold of A
in an open set 𝑈 ⊂ M we shall mean an immersed (Definition 9.3.7), connected,
regular submanifold L of 𝑈, without self-intersections, that satisfies the following
two conditions:
(1) 𝑇℘ L = A ℘ for every ℘ ∈ L;
(2) if L ′ is also an immersed, regular submanifold of 𝑈 having Property (1) then
either L ∩ L ′ = ∅ or else L ′ is an open subset of L.

If A is the linear space of all regular sections of a regular vector subbundle B of

𝑇M any integral manifold of A will also be called an integral manifold of B.

Property (1) implies that A ℘ is a linear subspace of constant dimension of 𝑇℘ M

for every ℘ ∈ L. If A ℘ = {0} then the single point {℘} is an integral manifold of
A. Property (2) states that L is maximal for Property (1).
By an integral manifold of a regular vector subbundle B of 𝑇M we mean an
integral manifold L of the set of all regular sections of B; Property (2) implies
dim L = rank B.

Remark 12.2.2 The integral manifolds of a set of regular vector fields need not be
locally closed. For instance, every integral manifold of the vector field 𝜕𝜃𝜕 1 − 𝜆 𝜕𝜃𝜕 2 ,
𝜆 ∉ Q, on the 2-torus T2 = R2 /Z2 is dense in T2 (see Subsection 12.2.2).

Definition 12.2.3 By a regular involutive vector subbundle of the tangent bundle

𝑇M we shall mean a regular vector subbundle B of 𝑇M such that the Lie bracket
[𝑋, 𝑌 ] of two regular sections 𝑋 and 𝑌 of B over an arbitrary open set 𝑈 ⊂ M
is also a section (per force regular) of B over 𝑈 (a property expressed by writing
[B, B] ⊂ B).

Every line subbundle B of 𝑇M (meaning rank B = 1) is involutive: [B, B] = 0.

The space R (𝑈; B) of regular sections over an open set 𝑈 ⊂ M of a regular
involutive vector subbundle B of 𝑇M is a Lie algebra (over the field K = R or C)
with respect to the Lie bracket. It has a special property in that this Lie algebra has
constant rank (equal to 𝑟 = rank B) in the following sense: if 𝑈 is the domain of a
regular frame of B, 𝑠1 , ..., 𝑠𝑟 , then every 𝑠 ∈ R (𝑈; B) is a linear combination of
𝑠1 , ..., 𝑠𝑟 , with coefficients in R (𝑈).
12.2 Integral Manifolds. Frobenius’ Theorem 355

Let 𝑋1 , ..., 𝑋𝑟 be regular vector fields in M. Obviously, if dim 𝔤℘ (𝑋1 , ..., 𝑋𝑟 ) is

the same for all ℘ ∈ M then the linear subspaces 𝔤℘ (𝑋1 , ..., 𝑋𝑟 ) ⊂ 𝑇℘ M make up
a regular involutive subbundle B of 𝑇M. This does not mean that the Lie algebra
𝔤 (𝑋1 , ..., 𝑋𝑟 ) contains a frame of B over M, although it contains frames of B over
sufficiently small neighborhoods of an arbitrary point of M.
Remark 12.2.4 Often, Íof regular vector fields in M, 𝑋1 , ..., 𝑋𝑟 , is said
a system
to be involutive if 𝑋 𝑗 , 𝑋 𝑘 = 𝑟ℓ=1 𝑐 𝑗,𝑘,ℓ 𝑋ℓ with 𝑐 𝑗,𝑘,ℓ ∈ R (M), for all indices
𝑗, 𝑘, ℓ = 1, ..., 𝑟. This terminology is used even when all the 𝑋 𝑗 vanish at some (but
not at every) point, in which case the 𝑋 𝑗 do not, of course, span a vector subbundle
of 𝑇M of positive rank.
We can now state and prove the all-important Frobenius Theorem.
Theorem 12.2.5 Let B be a regular vector subbundle of 𝑇M. The following two
conditions are equivalent:
(a) through every point ℘ ∈ M there passes a unique integral manifold L ℘ of B in
M; L ℘ is regular and countable at infinity;
(b) B is involutive.
We must emphasize the fact that, when K = R, the sections of the vector bundle
B in Theorem 12.2.5 are real vector fields.
Proof That (a)=⇒(b) is easy to prove. Indeed, if (a) holds then the elements of B℘
are the values at ℘ of the vectors tangent to the integral manifold L ℘ of B passing
through ℘. If 𝑋 and 𝑌 are two smooth sections of B in some neighborhood 𝑈 of ℘
then the restrictions to L ℘ of 𝑋 and 𝑌 are tangent to L ℘ and so is [𝑋, 𝑌 ], whose
value at ℘ must therefore belong to B℘ .
We shall now prove that (b)=⇒(a). The assertion is trivial if rank B = 0, each point
being then an integral manifold of B. We shall therefore assume rank B = 𝑟 ≥ 1. The
proof is in two steps. In Step 1 we prove that there is a covering of M by domains of
local charts (𝑈, 𝑥1 , ..., 𝑥 𝑛 ) such that
𝜕 𝜕
(•) the vector fields 𝜕𝑥1 , ..., 𝜕𝑥𝑟 form a regular frame of B over 𝑈.
In Step 2 we show how this implies (a).
Let (𝑈, 𝑥1 , ..., 𝑥 𝑛 ) be an arbitrary regular local chart centered at ℘, meaning ℘
is the point of 𝑈 at which 𝑥 𝑗 = 0 for every 𝑗 = 1, ..., 𝑛. Let 𝑋1 , ..., 𝑋𝑟 be a regular
frame of B over 𝑈; after a K-linear change of coordinates we can assume that
𝑋 𝑗 ℘ = 𝜕𝑥𝜕 𝑗 , 𝑗 = 1, ..., 𝑟, whence
𝑥=0

𝑛
∑︁ 𝜕
𝑋𝑗 = 𝑎 𝑗,𝑘 (𝑥) , 𝑗 = 1, ..., 𝑟,
𝑘=1
𝜕𝑥 𝑘

with 𝑎 𝑗,𝑘 (0) = 𝛿 𝑗,𝑘 (𝛿 𝑗,𝑘 : the Kronecker index). Provided 𝑈 is sufficiently small the
matrix 𝑎 𝑗,𝑘 (𝑥) 1≤ 𝑗,𝑘 ≤𝑟 is nonsingular; if we denote by 𝑏 𝑗,𝑘 (𝑥) the generic entry
of its inverse we see that
356 12 Lie Algebras of Vector Fields
𝑟 𝑛
∑︁ 𝜕 ∑︁ 𝜕
𝑋 ♭𝑗 = 𝑏 𝑗,𝑘 (𝑥) 𝑋 𝑘 = + 𝑎♭𝑗,𝑘 (𝑥) , 𝑗 = 1, ..., 𝑟, (12.2.1)
𝑘=1
𝜕𝑥 𝑗 𝑘=𝑟+1 𝜕𝑥 𝑘

with the coefficients 𝑎♭𝑗,𝑘 all vanishing at ℘; 𝑋1♭ , ..., 𝑋𝑟♭ is a frame of B over 𝑈 and
therefore we must have
h 𝑛
i ∑︁
𝑋 ♭𝑗 , 𝑋 ♭𝑗′ = 𝑐 𝑗 𝑗 ′ ,𝑘 (𝑥) 𝑋 ♭𝑗 .
𝑘=1

But the left-hand side is a linear combination of the partial derivatives 𝜕𝑥𝜕ℓ (ℓ =
𝑟 + 1, ..., 𝑛) whereas the right-hand side involves perforce the 𝜕𝑥𝜕ℓ (1 ≤ ℓ ≤ 𝑟) unless
h i
all the coefficients 𝑐 𝑗 𝑗 ′ ,𝑘 vanish identically in 𝑈. We conclude that 𝑋 ♭𝑗 , 𝑋 ♭𝑗′ ≡ 0 in
𝑈 for all 𝑗, 𝑗 ′, 1 ≤ 𝑗 < 𝑗 ′ ≤ 𝑟. Thus there is no loss of generality in assuming that the
original frame 𝑋1 , ..., 𝑋𝑟 consists of commuting vector fields having the expression
(12.2.1); in other words we can omit the superscripts ♭.
In order to prove the claim (•) we solve the following initial value-problem for a
system of 𝑛 − 𝑟 ODEs (in which 𝑥2 , ..., 𝑥𝑟 are parameters – if 𝑟 ≥ 2):

𝜕 𝑓𝑘
= 𝑎♭𝑗,𝑘 (𝑥1 , 𝑥2 , ..., 𝑥𝑟 , 𝑓𝑟+1 , ..., 𝑓𝑛 ) , 𝑓 𝑘 | 𝑥1 =0 = 𝑦 𝑘 , 𝑘 = 𝑟 + 1, ..., 𝑛.
𝜕𝑥1
By the fundamental theorem of ODE theory this problem admits a unique regular
so-
𝜕 𝑓𝑘
lution ( 𝑓 𝑘 (𝑥1 , 𝑥2 , ..., 𝑥𝑟 , 𝑦 𝑟+1 , ..., 𝑦 𝑛 )) 𝑟 <𝑘 ≤𝑛 in 𝑈. The Jacobian matrix 𝜕𝑦ℓ
𝑟 <𝑘,ℓ ≤𝑛
is the 𝑟 ×𝑟 identity matrix on the hypersurface 𝑥1 = 0. Thanks to the Implicit Function
Theorem we can solve in 𝑈 (sufficiently small) the system of equations

𝑥 𝑘 = 𝑓 𝑘 (𝑥1 , 𝑥2 , ..., 𝑥𝑟 , 𝑦 𝑟+1 , ..., 𝑦 𝑛 ) , 𝑘 = 𝑟 + 1, ..., 𝑛,

with respect to
𝑦 𝑘 = 𝑔 𝑘 (𝑥) , 𝑘 = 𝑟 + 1, ..., 𝑛. (12.2.2)
By adjoining to the equations (12.2.2) the equations 𝑦 𝑗 = 𝑥 𝑗 , 𝑗 = 1, ..., 𝑟, we obtain
a new coordinate patch (𝑈, 𝑦 1 , ..., 𝑦 𝑛 ) centered at ℘ in which
𝑛
𝜕 𝜕 ∑︁ 𝜕
𝑋1 = , 𝑋𝑗 = + 𝑏 𝑗,𝑘 (𝑦) if 𝑗 = 2, ..., 𝑟. (12.2.3)
𝜕𝑦 1 𝜕𝑦 𝑗 𝑘=𝑟+1 𝜕𝑦 𝑘

Since 𝑋1 , ..., 𝑋𝑟 commute, the coefficients 𝑏 𝑗,𝑘 (1 < 𝑗 ≤ 𝑘) must be independent of

𝑦 1 . We are thus reduced to dealing with a commuting system of 𝑟 − 1 vectors in 𝑛 − 1
variables. Induction proves (•) and concludes Step 1.
We now proceed with Step 2. We select a countable covering of M by the
domains of regular local charts 𝑈 (𝜈) , 𝑥1 , ..., 𝑥 𝑛 , 𝜈 ∈ Z+ , each one having Property
(•). We can assume that the coordinate map 𝑈 (𝜈) ∋ ℘ ↦→ (𝑥1 (℘) , ..., 𝑥 𝑛 (℘)) is
a regular isomorphism onto a domain in K𝑛 of the form L (𝜈) × pr 𝑈 (𝜈) where
12.2 Integral Manifolds. Frobenius’ Theorem 357

pr : ℘ ↦→ 𝑥 ′′ (℘) = (𝑥𝑟+1 , (℘) , ..., 𝑥 𝑛 (℘)), with L (𝜈) and pr 𝑈 (𝜈) geometrically
simple domains in K𝑟 and K𝑛−𝑟 respectively. Property (•) enables us to select the
coordinates 𝑥1 , ..., 𝑥𝑟 , so that, for each 𝑥 ′′ = (𝑥𝑟+1 , ..., 𝑥 𝑛 ) ∈ pr 𝑈 (𝜈) , L (𝜈) × {𝑥 ′′ } is
an integral manifold of 𝑋1 , ..., 𝑋𝑟 in 𝑈 (𝜈) .
Let 𝜈, 𝜇 ∈ Z+ , 𝜈 < 𝜇, and 𝑥 1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 , be the local coordinates in
𝑈 (𝜈)
,𝑈
( 𝜇) respectively, such that (•) is satisfied [𝑈 (𝜈) = 𝑈 ( 𝜇) is not precluded].

If L × {𝑥 ′′ } ∩ L ( 𝜇) × {𝑦 ′′ } ≠ ∅ then L (𝜈) × {𝑥 ′′ } ∪ L ( 𝜇) × {𝑦 ′′ } is a
(𝜈)

regular submanifold of the connected open set 𝑈 (𝜈) ∪ 𝑈 ( 𝜇) , actually an integral

manifold of B in 𝑈 (𝜈) ∪ 𝑈 ( 𝜇) . [If 𝑈 (𝜈) = 𝑈 ( 𝜇) this demands L (𝜈) × {𝑥 ′′ } =
L ( 𝜇) × {𝑦 ′′ }.] It has the following direct consequence: For any given L (𝜈◦ ) × {𝑥 ′′◦ }
let 𝚽 (𝜈◦ ) (𝑥 ′′◦ ) denote the set of connected regular submanifolds of M containing
L (𝜈◦ ) × {𝑥 ′′◦ }, each the union of finitely many submanifolds L (𝜈) × {𝑥 ′′ }. If L 𝑗 ∈
𝚽 (𝜈◦ ) (𝑥 ′′◦ ), 𝑗 = 1, 2, then L1 ∪ L2 ∈ 𝚽 (𝜈◦ ) (𝑥 ′′◦ ). From this we easily reach
the conclusion that the union of all the submanifolds containing L (𝜈◦ ) (𝑥 ′′◦ ) is a
pathwise connected, countable at infinity, immersed, regular submanifold L (𝑥 ′′◦ );
dim L (𝑥 ′′◦ ) = 𝑟. The manifolds L (𝑥 ′′◦ ) are pairwise disjoint and every point of M
belongs to one of them; each one is an integral manifold of B. □

Remark 12.2.6 In the preceding proof we have shown that aregularinvolutive vector
subbundle B of 𝑇M has local frames 𝑋1 , ..., 𝑋𝑟 such that 𝑋 𝑗 , 𝑋 𝑘 = 0 for all 𝑗, 𝑘.
If 𝑟 ≥ 2 this has an interesting side-effect. Let the 𝑋 𝑗 have the expressions (12.2.1)
(superscripts ♭ omitted). If we regard 𝑋 𝑗 as a differential operator (say, acting on
regular functions) it has a transpose, 𝑋 ⊤𝑗 = −𝑋 𝑗 − div 𝑋 𝑗 , where
𝑛
∑︁ 𝜕𝑎 𝑗,ℓ
div 𝑋 𝑗 = .
ℓ=𝑟+1
𝜕𝑥ℓ

The differential operators 𝑋 ⊤𝑗 also commute:

h i
𝑋 ⊤𝑗 , 𝑋 𝑘⊤ = div 𝑋 𝑗 𝑋 𝑘 − (div 𝑋 𝑘 ) 𝑋 𝑗 + 𝑋 𝑗 (div 𝑋 𝑘 ) − 𝑋 𝑘 div 𝑋 𝑗 = 0.

Because of (12.2.1) this demands div 𝑋 𝑗 ≡ div 𝑋 𝑘 ≡ 0. Thus, if rank B ≥ 2 and the
𝑋 𝑗 commute they are divergence free.

12.2.2 A classical example

The integral manifolds of the vector field in the plane, 𝑋 = 𝜕𝑥𝜕 1 + 𝑎 𝜕𝑥𝜕 2 (𝑎 ≥ 0), are
the straight lines

Λ (𝑎,𝑏) = 𝑥 ∈ R2 ; 𝑥2 = 𝑎𝑥1 + 𝑏 , 𝑏 ∈ R.

358 12 Lie Algebras of Vector Fields

The quotient map 𝜋 : R2 −→ T2 = R2 /Z2 (T2 : the two-dimensional torus) induces

a homeomorphism of the semiclosed square 𝔔 = 𝑥 ∈ R2 ; 0 ≤ 𝑥 𝑗 < 1, 𝑗 = 1, 2

onto T2 . Given arbitrarily 𝑛 𝑗 ∈ Z, 𝑗 = 1, 2, 𝜋Λ (𝑎,𝑏) = 𝜋Λ (𝑎,𝑏+𝑎𝑛1 −𝑛2 ) . We will

investigate how many parallel but distinct straight lines Λ (𝑎,𝑏+𝑎𝑛1 −𝑛2 ) intersect 𝔔.
If 𝑎 = 0, 0 ≤ 𝑥2 < 1 implies 𝑛2 ≤ 𝑏 < 𝑛2 + 1, i.e., 𝑛2 = [𝑏], the integer part of 𝑏;
this determines the horizontal line Λ (0,𝑏−𝑛2 ) uniquely.
In the remainder of this subsection we take 𝑎 > 0 fixed.

Remark 12.2.7 For 𝑎 < 0 the forthcoming argument would be essentially the same
after replacing 𝔔 by 𝑥 ∈ R2 ; − 1 < 𝑥1 ≤ 0 ≤ 𝑥2 ≤ 1 .

What distinguishes the straight lines Λ (𝑎,𝑏+𝑎𝑛1 −𝑛2 ) is the quantity 𝜏 = 𝑎𝑛1 + 𝑏 −𝑛2 .
If (0, 0) ∈ Λ (𝑎, 𝜏) ∩ 𝔔 then 𝜏 = 0 and

Λ (𝑎,0) = 𝑥 ∈ R2 ; 𝑥2 = 𝑎𝑥1 ,

which intersects 𝔔\ {(0, 0)}. Thus, either Λ (𝑎, 𝜏) ∩ 𝔔 = ∅ or else the following
holds:
(H) ∃𝑥 ∈ Λ (𝑎, 𝜏) ∩ 𝔔, 𝑥 ≠ (0, 0),
in which case Λ (𝑎, 𝜏) is completely determined by Λ (𝑎, 𝜏) ∩ 𝔔.
Claim: Property (H) is equivalent to 𝜏 ∈ (−𝑎, 1).
Proof Suppose there is an 𝑥 ≠ (0, 0) satisfying

0 ≤ 𝑥2 = 𝑎𝑥1 + 𝜏 < 1. (12.2.4)

Since 𝑎𝑥 1 ≥ 0 we have 𝜏 < 1; since 𝑥1 < 1 we have 𝜏 ≥ −𝑎𝑥 1 > −𝑎. Conversely,
suppose max (𝜏, −𝜏/𝑎) < 1. If 𝜏 ≤ 0 there is an 𝑥1 ∈ (−𝜏/𝑎, 1) such that 𝑥2 =
𝑎𝑥 1 + 𝜏 ∈ (0, 1). If 0 < 𝜏 < 1 there is an 𝑥1 ∈ (0, 1) such that 𝑥 2 = 𝑎𝑥 1 + 𝜏 ∈ (0, 1).□

Thus our question is: How many distinct values of 𝜏 are there in the interval
(−𝑎, 1) as 𝑛1 and 𝑛2 range over Z?
It is convenient to write 𝑛2 = [𝑏] + 𝑚, 𝑚 ∈ Z; thus 𝑏 − 𝑛2 = 𝑚 + 𝜃, 0 ≤ 𝜃 < 1,
and 𝜏 = 𝑎𝑛1 + 𝑚 + 𝜃; (H) is equivalent to

− 𝜃 < 𝑚 + 𝑎 (𝑛1 + 1) < 1 − 𝜃. (12.2.5)

Proposition 12.2.8 The range of the map Z2 ∋ (𝑚, 𝑛) −→ 𝑚 − 𝑎𝑛 is locally finite if

and only if 𝑎 ∈ Q.

Proof Suppose 𝑎 = 𝑝/𝑞, 𝑝, 𝑞 ∈ Z+ \ {0} and |𝑞𝑚 − 𝑝𝑛| < 𝑁𝑞, 1 ≤ 𝑁 ∈ Z. It is

obvious that there are only finitely many (𝑚, 𝑛), 𝑚𝑛 ≤ 0, that satisfy this condition.
Let us suppose 𝑚 ≥ 1, 𝑛 ≥ 1; there are 𝑘, ℓ ∈ Z+ such that 0 ≤ 𝑚 − 𝑘 𝑝 = 𝑚 ′ ≤ 𝑝 − 1,
0 ≤ 𝑛 − ℓ𝑞 = 𝑛 ′ ≤ 𝑞 − 1, and 𝑞𝑚 − 𝑝𝑛 = 𝑚 ′ 𝑞 − 𝑛 ′ 𝑝 − ℎ𝑝𝑞, ℎ = 𝑘 − ℓ. We have

|𝑚 ′ − 𝑎𝑛 ′ − ℎ𝑝| ≤ 𝑁. (12.2.6)
12.2 Integral Manifolds. Frobenius’ Theorem 359

There are at most 𝑝 × 𝑞 choices of the pair (𝑚 ′, 𝑛 ′) and therefore a finite number of
choices of ℎ ∈ Z satisfying (12.2.6).
If 𝑎 ∉ Q the classical Lagrange Theorem states that there are infinitely many pairs
of coprime positive integers 𝑚, 𝑛 such that
1 1
− √ < 𝑛𝑎 − 𝑚 < √ . □
5𝑛 5𝑛

Corollary 12.2.9 Finitely many straight lines Λ (𝑎,𝑏+𝑎𝑛1 −𝑛2 ) intersect 𝔔 if and only
if 𝑎 ∈ Q.

If 𝑎 ∉ Q the number of distinct straight lines Λ (𝑎,𝑏+𝑎𝑛1 −𝑛2 ) intersecting 𝔔 is

infinite but evidently countable. It follows easily from the Lagrange Theorem that
the intersections Λ (𝑎,𝑏+𝑎𝑛1 −𝑛2 ) ∩𝔔 form a dense subset of 𝔔. It ensues that the curve
𝜋Λ (𝑎,𝑏) in T2 is a dense subset of T2 .

12.2.3 Application: Local integrability of C 𝝎 complex involutive

structures

We are now going to complexify the concept of involutive structure introduced in

Definition 12.2.3. The complexification concerns the case K = R (thus regular here
means either C ∞ or C 𝜔 ) and requires the complexifications of 𝑇℘ M and 𝑇℘∗ M,
usually denoted by C ⊗R 𝑇℘ M and C ⊗R 𝑇℘∗ M, for every ℘ ∈ M. (We recall that
if 𝑬 is a real vector
Í space then C ⊗R 𝑬 is the complex vector space of finite linear
combinations 𝑖 𝑐 𝑖 𝒗 𝑖 where 𝒗 𝑖 ∈ 𝑬 and 𝑐 𝑖 ∈ C.) Here we shall denote these spaces
by C𝑇℘ M and C𝑇℘∗ M, and the bundles they form, by C𝑇M and C𝑇 ∗ M respectively.
The sections of C𝑇M are the complex vector fields in M; those of C𝑇 ∗ M are the
complex differential one-forms in M. Of course, when K = C we have 𝑇M = C𝑇M
and 𝑇 ∗ M = C𝑇 ∗ M. We have the complex version of Definition 12.2.3:

Definition 12.2.10 We shall say that a regular complex vector subbundle B of C𝑇M
is involutive if [B, B] ⊂ B and, if this is true, that B defines a complex involutive
structure on M.

Every complex line subbundle B of C𝑇M is involutive.

Let B be a regular complex vector subbundle of C𝑇M (not necessarily involutive).
We denote by B ⊤ the complex vector subbundle of C𝑇 ∗ M orthogonal to B, meaning
that 𝛾 ∈ C𝑇℘∗ M belongs to B℘⊤ if and only if ⟨𝛾, 𝒗⟩ = 0 whatever 𝒗 ∈ B℘ . If
⊤
𝑟 = rank B then rank B ⊤ = 𝑛 − 𝑟. Note also that B℘⊤ = B℘ .

Definition 12.2.11 We shall say that a regular complex vector subbundle B of C𝑇M
of rank 𝑟 ≥ 1 is locally integrable if every point ℘ ∈ M is contained in an open set
𝑈 in which there are 𝑛 − 𝑟 functions 𝑍 𝑘 ∈ ℜ (𝑈), 𝑘 = 1, ..., 𝑛 − 𝑟, whose differentials
360 12 Lie Algebras of Vector Fields

span B ⊤ at every point of 𝑈. We shall refer to the functions 𝑍 𝑘 as first integrals in

𝑈 of the system of homogeneous differential equations defined by the sections of B
or, in short, first integrals of B in 𝑈.
Generally, the first integrals 𝑍 𝑘 are complex-valued; their differentials are linearly
independent at every point of 𝑈. We have 𝐿𝑍 𝑘 ≡ 0 in 𝑈 whatever 𝐿 ∈ C 1 (𝑈; B).
Actually, a regular vector field 𝐿 in 𝑈 is a section of B if and only if ⟨d𝑍 𝑘 , 𝐿⟩ = 𝐿𝑍 𝑘
vanishes identically in 𝑈 for every 𝑘 = 1, ..., 𝑛 − 𝑟. If 𝐿 𝑖 ∈ ℜ (𝑈; B), 𝑖 = 1, 2, then

⟨d𝑍 𝑘 , [𝐿 1 , 𝐿 2 ]⟩ = 𝐿 1 (𝐿 2 𝑍 𝑘 ) − 𝐿 2 (𝐿 1 𝑍 𝑘 ) (12.2.7)

vanishes identically in U for each 𝑘 = 1, ..., 𝑟; (12.2.7) has the following direct
consequence:
Proposition 12.2.12 If a regular complex vector subbundle B of C𝑇M is locally
integrable then B is involutive.
Remark 12.2.13 In view of the Poincaré Lemma (Theorem 9.4.10) the defining
property of local integrability could be stated by saying that, locally, the complex
vector bundle B ⊤ has regular frames consisting of closed one-forms (since these are
locally exact).
Theorem 12.2.14 If a C 𝜔 complex vector subbundle B of C𝑇M is involutive it is
locally integrable.
Proof Let (𝑈, 𝑥1 , ..., 𝑥 𝑛 ) be a C 𝜔 local chart in M in which the vector fields
𝑛
∑︁ 𝜕
𝐿𝑗 = 𝑎 𝑖, 𝑗 (𝑥) , 𝑗 = 1, ..., 𝑟 = rank B, (12.2.8)
𝑖=1
𝜕𝑥 𝑗

span B at every point of 𝑈; 𝑎 𝑖, 𝑗 ∈ C 𝜔 (𝑈). Letting Ω ⊂ R𝑛 denote the image of

𝑈 under the C 𝜔 diffeomorphism ℘ → (𝑥1 (℘) , ..., 𝑥 𝑛 (℘)) we can find an open
subset ΩC of C𝑛 such that Ω ⊂ ΩC ∩ R𝑛 and to which the coefficients 𝑎 𝑗,𝑘 extend
as holomorphic functions. The holomorphic vector fields 𝐿˜ 𝑗 = 𝑖=1
Í𝑛
𝑎 𝑖, 𝑗 (𝑧) 𝜕𝑧𝜕 𝑗 ,
𝑗 = 1, ..., 𝑟, span an involutive vector subbundle BΩC of 𝑇ΩC to which we can apply
Frobenius’ Theorem (Theorem 12.2.5) with K = C. We directly reach the following
conclusion: every point 𝑥 ◦ ∈ Ω is contained in an open set 𝑉 C ⊂ ΩC in which there
𝜕
exist holomorphic coordinates 𝑤 1 , ..., 𝑤 𝑛 such that the vector fields 𝜕𝑤 𝑗
( 𝑗 = 1, ..., 𝑟)
span BΩC at every point of 𝑉 C . It follows immediately that the functions 𝑍 𝑘 = 𝑤 𝑟+𝑘 |𝑈
(𝑘 = 1, ..., 𝑛 − 𝑟) are first integrals. □
Remark 12.2.15 There are C ∞ complex involutive structures that are not locally
integrable, as shown in [Nirenberg, 1974]. If dim M = 3 most (in the sense of Baire
category) C ∞ complex involutive structures defined by line bundles are not locally
integrable (see [Jacobowitz-Treves, 1983]).
To each complex involutive structure on M, whether locally integrable or not,
one can associate a differential complex of order 1 (Definition 9.4.2). On this topic
we refer to [Treves, 1992] and [Berhanu, Cordaro, Hounie, 2008].
12.3 Local Flow of a Regular Vector Field 361

12.3 Local Flow of a Regular Vector Field

In this section the base manifold M shall be regular, as before meaning C ∞ , C 𝜔 or

complex-analytic; local coordinates will always be regular. But we shall often deal
with a vector field 𝑋 in M that is only C ∞ , unless we specify that it is “regular”. The
points of M where 𝑋 vanishes are the critical points of 𝑋; they form the critical
set Crit 𝑋 of 𝑋. By Frobenius’ Theorem its complement M\Crit 𝑋 is foliated by
the integral curves of 𝑋 which are smooth (regular if 𝑋 is regular), immersed,
self-intersection free, one-dimensional submanifolds of M.
Let (U, 𝑥1 , ..., 𝑥 𝑛 ) be a coordinate patch centered at an arbitraryÍpoint ℘◦ ∈ M
(meaning that 𝑥1 = · · · = 𝑥 𝑛 = 0 at ℘) where the expression 𝑋 = 𝑛𝑗=1 𝑎 𝑗 (𝑥) 𝜕𝑥𝜕 𝑗
holds, the coefficients 𝑎 𝑗 being real-valued when M is real. The Fundamental
Theorem on ODEs tells us that the initial value problem for a system of ODEs,

d𝑥 𝑗
= 𝑎 𝑗 (𝑥 (𝑡)) , 𝑥 𝑗 (0) = 0, 𝑗 = 1, ..., 𝑛, (12.3.1)
d𝑡
has a unique C ∞ solution 𝑥 (𝑡) in an interval |𝑡| < 𝑇 for some 𝑇 > 0 depending on
℘◦ ; this solution is regular if 𝑋 is regular. If 𝑎 𝑗 (0) = 0 for every 𝑗 = 1, ..., 𝑛, the
solution is 𝑥 (𝑡) = 0 for all 𝑡 ∈ R: if ℘◦ ∈ Crit 𝑋 the solution does not move away
from ℘◦ .
Suppose we carry out a change of regular coordinates in U: 𝑥 = 𝜙 (𝑦), or
equivalently 𝑦 = 𝜓 (𝑥), with 𝜙 = (𝜙1 , ..., 𝜙 𝑛 ), 𝜓 = (𝜓1 , ..., 𝜓 𝑛 ). We get the new
Í
expression 𝑋 = 𝑛𝑘=1 𝑏 𝑘 (𝑦) 𝜕𝑦𝜕𝑘 with

𝑛
∑︁ 𝜕𝜓 𝑘
𝑏 𝑘 (𝑦) = 𝑎 𝑗 (𝜙 (𝑦)) (𝜙 (𝑦)) .
𝑗=1
𝜕𝑥 𝑗

If we set 𝑦 (𝑡) = 𝜓 (𝑥 (𝑡)) then

𝑛
d𝑦 𝑘 ∑︁ 𝜕𝜓 𝑘 d𝑥 𝑗
= (𝑥 (𝑡)) = 𝑏 𝑘 (𝑦 (𝑡)) .
d𝑡 𝑗=1
𝜕𝑥 𝑗 d𝑡

We conclude that the equations (12.3.1) are invariant under (regular) coordinate
change. Solving them means finding a C ∞ (regular if 𝑋 is regular) map (−𝑇, 𝑇) ∋
𝑡 → 𝛾 (𝑋, ℘◦ , 𝑡) ∈ U such that 𝛾 (𝑋, ℘◦ , 0) = ℘◦ . It is customary to identify this
map with its image and call it a curve; below we denote it by 𝛾 (𝑋, ℘◦ , ·). The
pushforward under this map of the vector d𝑡d tangent to the open interval (−𝑇, 𝑇) is
called the velocity vector of the curve 𝛾 (𝑋, ℘◦ , ·). The equations (12.3.1) state that
the velocity vector is equal to 𝑋 at every point 𝛾 (𝑋, ℘◦ , 𝑡). In particular the curve
𝛾 (𝑋, ℘◦ , ·) is an arc of the integral curve of 𝑋 through ℘◦ .
If the closure of an open set Ω ⊂ M is compact there is a number 𝜏Ω , 0 < 𝜏Ω ≤
+∞, such that the map 𝛾 (𝑋, ℘, 𝑡) is defined and C ∞ for every ℘ ∈ Ω and every
𝑡 ∈ (−𝜏Ω , 𝜏Ω ). (Generally speaking, the number 𝜏Ω tends to zero as Ω expands.)
362 12 Lie Algebras of Vector Fields

Moreover, for each 𝑡 ∈ (−𝜏Ω , 𝜏Ω ) the map ℘ ↦→ 𝛾 (𝑋, ℘, 𝑡) is a diffeomorphism (a

regular isomorphism if 𝑋 is regular) of Ω onto an open subset of M, which we
denote by Φ𝑋 (𝑡) and call the local flow of 𝑋 in Ω (see [Dieudonné, 1960], Ch. X).
A fundamental property of the local flow is that to each point ℘ ∈ M there is an
𝜀 > 0 such that |𝑡| + |𝑡 ′ | < 𝜀 entails Φ𝑋 (𝑡) Φ𝑋 (𝑡 ′) ℘ = Φ𝑋 (𝑡 + 𝑡 ′) ℘. This property
justifies the commonly used notation exp 𝑡 𝑋 for Φ𝑋 (𝑡).

Proposition 12.3.1 Let 𝑋 be a C ∞ vector field in M, ℘ ∈ M and 𝑓 ∈ C 1 (M). Then

d
(𝑋 𝑓 ) (℘) = 𝑓 ((exp 𝑡 𝑋) ℘) . (12.3.2)
d𝑡 𝑡=0

Proof It suffices to prove (12.3.2) in a local coordinate patch where it is a straight-

forward matter. □
In connection with flows of vector fields it is worth recalling the classical
Campbell–Hausdorff (also Baker–Campbell–Hausdorff) formula. It is best stated
in purely algebraic terms, for the most general Lie algebra 𝔤 (𝑋, 𝑌 ) with two gen-
erators 𝑋 and 𝑌 , “most general” meaning that the brackets [ 𝐴, 𝐵] in 𝔤 (𝑋, 𝑌 ) are
not submitted to any relation except for skew-symmetry and the Jacobi identity. The
elements of 𝔤 (𝑋, 𝑌 ) are the linear combinations with coefficients in the field K of the
multibrackets 𝑍 𝐼 = (Ad 𝑍𝑟 ◦ · · · ◦ Ad 𝑍1 ) (𝑍0 ), where for each 𝑖 = 0, 1, ..., 𝑟 = |𝐼 |,
either 𝑍𝑖 = 𝑋 or 𝑍𝑖 = 𝑌 .

Theorem 12.3.2 There is a formal power series 𝐹 (𝑋, 𝑌 , 𝑠, 𝑡) in the powers of two
indeterminates 𝑠 and 𝑡 with coefficients in 𝔤 (𝑋, 𝑌 ) such that

exp (𝑠𝑋) exp (𝑡𝑌 ) = exp 𝐹 (𝑋, 𝑌 , 𝑠, 𝑡). (12.3.3)

The coefficient of 𝑠 𝑝 𝑡 𝑞 in 𝐹 (𝑋, 𝑌 , 𝑠, 𝑡) is a linear combination of multibrackets 𝑍 𝐼

where exactly 𝑝 elements 𝑍𝑖 are equal to 𝑋 and exactly 𝑞 are equal to 𝑌 .

For a proof see, e.g., [Hochschild, 1965], [Bonfiglioli-Fulci, 2012], and for an
𝜕 𝜕
elementary presentation, [Hall, 2015]. Letting 𝜕𝑠 , 𝜕𝑡 act repeatedly on both sides of
(12.3.3) and putting 𝑠 = 𝑡 = 0 shows directly that
1 1 2
𝐹 (𝑋, 𝑌 , 𝑠, 𝑡) = 𝑠𝑋 + 𝑡𝑌 + 𝑠𝑡 [𝑋, 𝑌 ] + 𝑠 𝑡 [𝑋, [𝑋, 𝑌 ]] (12.3.4)
2 12
1 1
+ 𝑠𝑡 2 [𝑌 , [𝑌 , 𝑋]] − 𝑠2 𝑡 2 [𝑌 , [𝑋, [𝑋, 𝑌 ]]]
12 24
+ terms of degree ≥ 5 with respect to (𝑠, 𝑡) .

As before let 𝑋 be a regular vector field in M. We go back to the open set

Ω ⊂⊂ M and the regular isomorphism ℘ ↦→ (exp 𝑡 𝑋) ℘ of Ω onto open sets Ω𝑡 ,
|𝑡| < 𝜏Ω . The pullback map 𝑓 −→ (exp 𝑡 𝑋) ∗ 𝑓 from C ∞ (Ω𝑡 ) to C ∞ (Ω) is defined
in the usual fashion:

∀℘ ∈ Ω, (exp 𝑡 𝑋) ∗ 𝑓 (℘) = 𝑓 ((exp 𝑡 𝑋) ℘) .

(12.3.5)
12.3 Local Flow of a Regular Vector Field 363

Proposition 12.3.3 We have, for all 𝑓 ∈ C ∞ (Ω𝑡 ),

d
(exp 𝑡 𝑋) ∗ 𝑓 = (exp 𝑡 𝑋) ∗ 𝑋 𝑓 . (12.3.6)
d𝑡
Proof We assume that all the open sets Ω𝑡 , |𝑡| < 𝜏Ω , are contained in the domain of
local regular coordinates 𝑥 𝑗 ( 𝑗 = 1, ..., 𝑛). Let us use the notation (exp 𝑡 𝑋) ∗ 𝑓 (𝑦) =

𝑓 (𝑥 (𝑡, 𝑦)), 𝑦 ∈ Ω; then
𝑛
∑︁ d𝑥 𝑗
d 𝜕𝑓
𝑓 (𝑥 (𝑡, 𝑦)) = (𝑡, 𝑦) (𝑥 (𝑡, 𝑦))
d𝑡 𝑗=1
d𝑡 𝜕𝑥 𝑗
𝑛
∑︁ 𝜕𝑓
= 𝑎 𝑗 (𝑥 (𝑡, 𝑦)) (𝑥 (𝑡, 𝑦)) = (𝑋 𝑓 ) (𝑥 (𝑡, 𝑦)) . □
𝑗=1
𝜕𝑥 𝑗

In the remainder of this section we assume that 𝑓 ∈ C ∞ (Ω𝑡 ). If 𝑌 is a C ∞ vector

field in Ω we can look at its pushforward under the map exp 𝑡 𝑋:

(exp 𝑡 𝑋) ∗ 𝑌 𝑓 = 𝑌 (exp 𝑡 𝑋) ∗ 𝑓 .

(12.3.7)

From (12.3.6) and (12.3.7) we derive

d
(exp 𝑡 𝑋) ∗ 𝑌 𝑓 = (exp 𝑡 𝑋) ∗ 𝑌 𝑋 𝑓 . (12.3.8)
d𝑡
For fixed 𝑡 ∈ (−𝜏Ω , 𝜏Ω ), (exp 𝑡 𝑋) ∗ defines a C ∞ vector bundle isomorphism of 𝑇Ω
onto 𝑇Ω𝑡 and therefore, by fiberwise transposition, a C ∞ vector bundle isomorphism
of 𝑇 ∗ Ω onto 𝑇 ∗ Ω𝑡 , which we also denote by (exp 𝑡 𝑋) ∗ . These isomorphisms are
regular if 𝑋 is regular.
Proposition 12.3.4 For every 𝑓 ∈ C ∞ (Ω𝑡 ),

(exp 𝑡 𝑋) ∗ d 𝑓 = d (exp 𝑡 𝑋) ∗ 𝑓 , (12.3.9)

d
(exp 𝑡 𝑋) ∗ d 𝑓 = d (exp 𝑡 𝑋) ∗ 𝑋 𝑓 .

(12.3.10)
d𝑡
Proof If 𝑌 is a C ∞ vector field in Ω then, by (12.3.7),

(exp 𝑡 𝑋) ∗ d 𝑓 , 𝑌 = d 𝑓 , (exp 𝑡 𝑋) ∗ 𝑌 = (exp 𝑡 𝑋) ∗ 𝑌 𝑓

= 𝑌 (exp 𝑡 𝑋) ∗ 𝑓 = d (exp 𝑡 𝑋) ∗ 𝑓 , 𝑌 ,

whence (12.3.9); combining (12.3.6) and (12.3.9) yields (12.3.10). □

If we use local regular coordinates 𝑥 𝑗 as above we derive from (12.3.10)

d
(exp 𝑡 𝑋) ∗ d𝑥 𝑘 = d (exp 𝑡 𝑋) ∗ 𝑎 𝑘 .

d𝑡
(𝑥) d𝑥 𝑘 is a C ∞ one-form in Ω𝑡 we get
Í𝑛
If then 𝐹 = 𝑘=1 𝑓 𝑘
364 12 Lie Algebras of Vector Fields

d
(exp 𝑡 𝑋) ∗ 𝐹 = d (exp 𝑡 𝑋) ∗ ⟨𝐹, 𝑋⟩ (12.3.11)
d𝑡
= (exp 𝑡 𝑋) ∗ d ⟨𝐹, 𝑋⟩ ,

the last equality a consequence of (12.3.9).

From (12.3.7) and (12.3.9) we can extend the pushforward and pullback maps as
isomorphisms of the exterior algebras of 𝑇Ω and 𝑇 ∗ Ω:

(exp 𝑡 𝑋) ∗ : Λ 𝑝 𝑇Ω −→ Λ 𝑝 𝑇Ω𝑡 , (12.3.12)

(exp 𝑡 𝑋) ∗ : Λ 𝑝 𝑇 ∗ Ω −→ Λ 𝑝 𝑇 ∗ Ω𝑡 . (12.3.13)

Obviously, they can also be naturally extended as isomorphisms of the tensor or of

the symmetric algebras; the inverse of (exp 𝑡 𝑋) ∗ is (exp (−𝑡 𝑋)) ∗ , that of (exp 𝑡 𝑋) ∗
is (exp (−𝑡 𝑋)) ∗ ; (exp 𝑡 𝑋) ∗ commutes with the exterior derivative d [cf. (12.3.9)].
Let 𝐹1 , ..., 𝐹 𝑝 be C ∞ 1-forms in Ω𝑡 (1 ≤ 𝑝 ≤ 𝑛); we derive directly from (12.3.11):

d
(exp (−𝑡 𝑋)) ∗ (exp 𝑡 𝑋) ∗ 𝐹1 ∧ · · · ∧ 𝐹 𝑝

(12.3.14)
d𝑡
𝑝
∑︁
= (−1) 𝑗 𝐹1 ∧ · · · ∧ 𝐹 𝑗−1 ∧ d 𝐹 𝑗 , 𝑋 ∧ 𝐹 𝑗+1 ∧ · · · ∧ 𝐹 𝑝 .
𝑗=1

This allows us to prove the following lemma, useful later on.

Lemma 12.3.5 If 𝜔 is a linear combination with constant coefficients of exterior

products 𝐹1 ∧ · · · ∧ 𝐹 𝑝 where each 𝐹 𝑗 is a closed C ∞ one-form in Ω𝑡 then

d
(exp 𝑡 𝑋) ∗ 𝜔 = d (exp 𝑡 𝑋) ∗ (𝑋 ⌟𝜔) (12.3.15)
d𝑡
= (exp 𝑡 𝑋) ∗ d (𝑋 ⌟𝜔) .

We have denoted by 𝑋 ⌟𝜔 the contraction of 𝜔 by 𝑋.

Proof If every 1-form 𝐹 𝑗 is closed then the right-hand side in (12.3.14) is equal to
d 𝑋 ⌟ 𝐹1 ∧ · · · ∧ 𝐹 𝑝 . □

12.4 Foliations Defined By Analytic Vector Fields

A particular case of the results in the preceding section is that a regular vector field 𝑋
in a regular manifold M defines a partition of M into the (regular) integral manifolds
of 𝑋 (in this case, of dimension 0 or 1). In the present section we generalize this
result to an arbitrary Lie algebra 𝔤 of regular vector fields in M but only when
regular means real-analytic or complex-analytic (using the adjective “analytic” in
both cases). Thus M will be an analytic manifold over K = R or C. For a C ∞
analogue, see [Sussmann, 1973] (in the C ∞ case the notion of “integral manifold”
12.4 Foliations Defined By Analytic Vector Fields 365

must be replaced by the considerably subtler notion of “orbit”). Generally speaking

and in contrast with the conclusion in Frobenius’ Theorem, the dimensions 𝑑 of the
integral manifolds of 𝔤 will be allowed to vary in the segment 0 ≤ 𝑑 ≤ 𝑛.
Definition 12.4.1 By an analytic foliation of the manifold M we shall mean a family
𝚽 of immersed (Definition 9.3.7) analytic submanifolds without self-intersections
satisfying the following conditions:
(1) Every submanifold L ∈ 𝚽 is connected.
(2) Every point of M lies in a unique submanifold L ∈ 𝚽.
We point out that dim L might vary with L ∈ 𝚽. We state and prove the Nagano
Theorem (see [Nagano, 1966]).
Theorem 12.4.2 Let M be an analytic manifold and 𝔤 be a Lie algebra (with respect
to the Lie bracket, with scalar field K = R if M is real, C if M is complex) of analytic
sections of 𝑇M. The integral manifolds of 𝔤 (Definition 12.2.1) form a foliation of
M.
The elements of 𝔤 are analytic vector fields in the whole of M, with real coeffi-
cients if K = R. It is tacitly assumed that dimK 𝔤 < +∞.
Proof Let 𝔤℘ denote the “freezing” of 𝔤 at an arbitrary point ℘ ∈ M; 𝔤℘ is the linear
subspace of 𝑇℘∗ M spanned (over K) by the tangent vectors 𝑋 | ℘ , 𝑋 ∈ 𝔤. Define, for
each integer 𝑟, 0 ≤ 𝑟 ≤ dimK 𝔤,

M𝑟 = ℘ ∈ M; dim 𝔤℘ = 𝑟 .

Each connected component of M 𝑛 is a connected component of M and an integral

manifold of 𝔤. Each point of M0 = ℘ ∈ M; 𝔤℘ = 0 is an integral manifold of 𝔤.
We shall therefore assume 1 ≤ 𝑟 ≤ 𝑛 − 1. We are going to show that M𝑟 (when
nonvoid) is an immersed analytic submanifold of M, in general, of course, not
connected.
It suffices to prove the claim in a neighborhood U of an arbitrary point ℘◦ ∈ M𝑟 .
To do this we can replace 𝔤 by the R (U)-module b 𝔤 (U) spanned by 𝔤, since the
freezing of b 𝔤 (U) at an arbitrary ℘ ∈ U is equal to 𝔤℘ . This allows us to use
nonsingular linear substitutions of an arbitrary system of vector fields 𝑋1 , ..., 𝑋𝑟 ∈
𝔤 (U),
b
𝑟
∑︁
𝑌𝑗 = 𝑐 𝑗,𝑘 𝑋 𝑘 , 𝑗 = 1, ..., 𝑟, (12.4.1)
𝑘=1

with coefficients 𝑐 𝑗,𝑘 defined and analytic in U, det 𝑐 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑟 nowhere vanishing.
We take 𝑋1 , ..., 𝑋𝑟 to be linearly independent; if U is sufficiently small this is the
same as saying that the tangent vectors 𝑋 𝑗 ℘◦ span 𝔤℘◦ (but the 𝑋 𝑗 ℘ might not span
𝔤℘ at ℘ ∈ U, ℘ ≠ ℘◦ ). We can find analytic coordinates 𝑥1 , ..., 𝑥 𝑛 in U, all vanishing
at ℘◦ , in which
𝑛
𝜕 ∑︁ 𝜕
𝑋𝑗 = + 𝑎 𝑗,𝑘 (𝑥) , 𝑗 = 1, ..., 𝑟,
𝜕𝑥 𝑗 𝑘=1 𝜕𝑥 𝑘
366 12 Lie Algebras of Vector Fields

with 𝑎 𝑗,𝑘 (0) = 0 for all 𝑗,𝑘. The reasoning in the proof of Theorem 12.2.5 leading
to (12.2.3) applies here and we can start from the premise that 𝑎 1,𝑘 (𝑥) ≡ 0 for all 𝑘
and, after substituting 𝑋 𝑗 for 𝑋 𝑗 − 𝑎 𝑗,1 (𝑥) 𝑋1 if 𝑗 ≥ 2, that
𝑛
𝜕 𝜕 ∑︁ 𝜕
𝑋1 = , 𝑋𝑗 = + 𝑎 𝑗,𝑘 (𝑥) if 𝑗 = 2, ..., 𝑟.
𝜕𝑥 1 𝜕𝑥 𝑗 𝑘=𝑟+1 𝜕𝑥 𝑘
𝑚 𝑚
We exploit the property that the multibrackets Ad 𝑋𝑖1 1 · · · Ad 𝑋𝑖ℓ ℓ 𝑋 𝑗 [see
(12.1.7)] must vanish at ℘◦ , whatever the integers 𝑖 𝛼 , 𝑚 𝛼 (𝛼 = 1, ..., ℓ), 𝑗, such that
1 ≤ 𝑖 𝛼 ≤ 𝑟, 𝑚 𝛼 ∈ Z+ , 𝑚 1 + · · · + 𝑚 ℓ ≥ 1, 2 ≤ 𝑗 ≤ 𝑟. Indeed, on the one hand they
are linear combinations of 𝑋1 , ..., 𝑋𝑟 at ℘◦ ; on the other hand they do not involve
any 𝜕𝑥𝜕 𝑗 , 1 ≤ 𝑗 ≤ 𝑟.
If 𝑗 ≥ 2 we derive (Ad 𝑋1 ) 𝑚 𝑋 𝑗 ℘◦ = 0 for all 𝑚 ∈ Z+ implying 𝑎 𝑗,𝑘 (𝑥1 , 0, ..., 0) ≡
0, whence
𝑛
∑︁
𝑎 𝑗,𝑘 (𝑥) = 𝑥ℓ 𝑏 𝑗,𝑘,ℓ (𝑥) , 𝑘 = 1, ..., 𝑛.
ℓ=2

Next we exploit the property that (Ad 𝑋1 ) 𝑚 𝑋ℓ , 𝑋 𝑗 ℘◦ = 0 for each ℓ = 2, ..., 𝑛,
and all 𝑚 ∈ Z+ . This implies 𝑏 𝑗,𝑘,ℓ (𝑥1 , 0, ..., 0) ≡ 0. Repeating this reasoning
ad infinitum proves that the coefficients 𝑎 𝑗,𝑘 (𝑥) are independent of 𝑥 1 ; the system
of vector fields 𝑋2 , ..., 𝑋𝑟 in U is independent of 𝑥 1 . Induction on 𝑟 leads to the
conclusion that, after a number of substitutions (12.4.1) and analytic changes of
variables we end up with vector fields
𝑛
𝜕 ∑︁ 𝜕
𝑋𝑗 = + 𝑥ℓ 𝑎 𝑗,𝑘,ℓ (𝑥) , 𝑗 = 1, ..., 𝑟. (12.4.2)
𝜕𝑥 𝑗 𝑘,ℓ=𝑟+1 𝜕𝑥 𝑘

This implies that the vector fields 𝑋1 , ..., 𝑋𝑟 form an analytic frame of the tangent
bundle of the analytic submanifold

L (℘◦ ) = {℘ ∈ U; 𝑥ℓ (℘) = 0, ℓ = 𝑟 + 1, ..., 𝑛} .

𝔤 (U); there are analytic functions

Now consider an arbitrary vector field 𝑋 ∈ b
𝑎 𝑗 , 𝑏 𝑘 (1 ≤ 𝑗 ≤ 𝑟 < 𝑘 ≤ 𝑛), in U such that
𝑟 𝑛
∑︁ ∑︁ 𝜕
𝑋− 𝑎 𝑗 (𝑥) 𝑋 𝑗 = 𝑏 𝑘 (𝑥) 𝔤 (U) .
=𝑌 ∈b
𝑗=1 𝑘=𝑟+1
𝜕𝑥 𝑘
𝑚 𝑚
The fact that 𝑋1 | ℘◦ , ..., 𝑋𝑟 | ℘◦ span 𝔤℘◦ implies Ad 𝑋𝑖1 1 · · · Ad 𝑋𝑖 𝛼 ℓ 𝑌 = 0 at
℘◦ , whatever the integers 𝑖 𝛼 ∈ [1, 𝑟], 𝑚 𝛼 ∈ Z+ (𝛼 = 1, ..., ℓ). In turn, this implies
𝑛
∑︁
𝑏 𝑘 (𝑥) = 𝑥ℓ 𝑏 𝑘,ℓ (𝑥) , 𝑘 = 𝑟 + 1, ..., 𝑛,
ℓ=𝑟+1
12.4 Foliations Defined By Analytic Vector Fields 367

proving that 𝑌 ≡ 0 on L (℘◦ ) and therefore that Ø

𝔤| L (℘◦ ) = 𝑇 L (℘◦ ).
Each connected component L of M𝑟 = L (℘◦ ) can be equipped with
℘◦ ∈M𝑟
the structure of an analytic manifold by means of the local coordinate patches
(L (℘◦ ) , 𝑥1 , ..., 𝑥𝑟 ), keeping in mind that the coordinates 𝑥1 , ..., 𝑥 𝑛 depend on ℘◦ .
It is obvious that the injection L ↩→ M is an immersion and that 𝔤| L = 𝑇 L. The
connected components of M𝑟 are the 𝑟-dimensional integral manifolds of 𝔤. □
We shall refer to any integral manifold of the Lie algebra 𝔤 as a Nagano leaf of
𝔤 in M. These leaves form the Nagano foliation of M defined by 𝔤.

Remark 12.4.3 An extension of Theorem 12.4.2 (with a completely different proof)

to Lie algebras of C ∞ vector fields can be found in [Sussmann, 1973].

Proposition 12.4.4 Each Nagano leaf of a Lie algebra 𝔤 of analytic vector fields in
the (countable at infinity) analytic manifold M is countable at infinity.

This is a direct consequence of the next result.

Proposition 12.4.5 An immersed, analytic submanifold of an analytic manifold

countable at infinity is countable at infinity.

Here we omit the (rather technical) proof of Proposition 12.4.5; it can be found
in various texts, for instance in [Chevalley, 1946], pp. 96–98.
A property of leaves that must be mentioned is a special case of a classical theorem
of C. Carathéodory (see [Carathéodory, 1909]) more generally valid for C ∞ vector
fields. We state it without proof.

Theorem 12.4.6 Let 𝑋1 , ..., 𝑋𝑟 be real, real-analytic vector fields in the real-analytic
manifold M and let L be a Nagano leaf of the Lie algebra 𝔤 (𝑋1 , ..., 𝑋𝑟 ). Given any
pair of points ℘1 , ℘2 of L there is a Lipschitz continuous curve joining ℘1 to ℘2
equal to the union of finitely many closed arcs of integral curves of vector fields 𝑋 𝑗
(1 ≤ 𝑗 ≤ 𝑟).

To the nesting of Lie subalgebras of R (M; 𝑇M) there corresponds the nesting
of their foliations:

Proposition 12.4.7 Let 𝔤1 and 𝔤2 be two Lie algebras of analytic sections of 𝑇M. If
𝔤1 is a subalgebra of 𝔤2 then every integral manifold of 𝔤1 is contained in an integral
manifold of 𝔤2 .

Proof Indeed, every integral curve of every vector field belonging to 𝔤1 is contained
in an integral manifold of 𝔤2 . □
368 12 Lie Algebras of Vector Fields

12.5 Systems of Vector Fields Generating Special Lie Algebras

12.5.1 Systems of finite type. Elliptic systems

Throughout this section we assume the base manifold M to be real-analytic. We shall

deal with real C 𝜔 vector fields in M, unless specified otherwise (as in Definition
12.5.1 below). It will be convenient to assume also that M is connected. We continue
to use Notations 12.1.1 and 12.1.2. We denote by rank℘ 𝔤 (𝑋1 , ..., 𝑋𝑟 ) the dimension
of the linear subspace 𝔤℘ (𝑋1 , ..., 𝑋𝑟 ) of 𝑇℘∗ M. The set of points ℘ ∈ M such that

rank 𝔤 (𝑋1 , ..., 𝑋𝑟 ) = max rank 𝔤 (𝑋1 , ..., 𝑋𝑟 )
℘ ℘∗ ∈M ℘∗

is an open and dense subset of M.

Definition 12.5.1 The system of real C ∞ vector fields 𝑋1 , ..., 𝑋𝑟 in M is said to be

of finite type at a point ℘ ∈ M if rank℘ 𝔤 (𝑋1 , ..., 𝑋𝑟 ) = dim M. It is said to be of
finite type if it is of finite type at every point of M. If the system (𝑋1 , ..., 𝑋𝑟 ) is not
of finite type at ℘ it is said to be of infinite type at ℘.
A system of complex C ∞ vector fields 𝑋1 , ..., 𝑋𝑟 in M is said to be of finite type (at
a point, in an open set) if this is true of the system Re 𝑋1 , ..., Re 𝑋𝑟 , Im 𝑋1 , ..., Im 𝑋𝑟 .

Proposition 12.5.2 Let 𝑋1 , ..., 𝑋𝑟 be real C 𝜔 vector fields in M and let ℘ ∈ M. The
following two properties are equivalent:
(1) The system of vector fields 𝑋1 , ..., 𝑋𝑟 is of finite type at ℘.
(2) The Nagano leaf of the algebra 𝔤 (𝑋1 , ..., 𝑋𝑟 ) through ℘ is an open subset of M.

Corollary 12.5.3 The system of vector fields 𝑋1 , ..., 𝑋𝑟 is of finite type if and only if
it has a single Nagano leaf (cf. Theorem 12.4.2), M itself.

Self-evident. Keep in mind that M is connected.

Example 12.5.4 Consider the commuting vector fields 𝑋1 = 𝜕𝑥𝜕 1 , 𝑋2 = 𝑥2 𝜕𝑥𝜕 2 in R2 .

There are three Nagano leaves of 𝔤 (𝑋1 , 𝑋2 ): the two open half-planes 𝑥2 > 0 and
𝑥2 < 0, and the 𝑥1 -axis.

Remark 12.5.5 Systems of vector fields 𝑋1 , ..., 𝑋𝑟 of finite type with 𝑟 < 𝑛 do exist.
Í 𝑗−1
Example: 𝑋1 = 𝜕𝑥𝜕 1 , 𝑋2 = 𝑛𝑗=2 𝑥 1 𝜕𝑥𝜕 𝑗 (𝑟 = 2, 𝑛 ≥ 3).

Proposition 12.5.6 Let the system of real C ∞ vector fields 𝑋1 , ..., 𝑋𝑟 be of finite type
at every point of a connected open subset U of M. If ℎ ∈ D ′ (U) satisfies the
equations 𝑋 𝑗 ℎ = 0, 𝑗 = 1, ..., 𝑟, in U then ℎ is a constant function in U.

Proof Under the hypotheses ℎ ∈ D ′ (U) satisfies the equation 𝑋 ℎ = 0 in U

whatever 𝑋 ∈ 𝔤 (𝑋1 , ..., 𝑋𝑟 ) and therefore whatever the smooth section 𝑋 of 𝑇M
over U. □
12.5 Systems of Vector Fields Generating Special Lie Algebras 369

Theorem 12.5.7 Let 𝑋1 , ..., 𝑋𝑟 be a system of real C 𝜔 vector fields in M. The

following properties are equivalent:
(a) the system 𝑋1 , ..., 𝑋𝑟 is of finite type;
(b) there is a covering of M by open sets U such that if 𝑢 ∈ D ′ (U) and if
𝑋 𝑗 𝑢 ∈ C 𝜔 (U) for every 𝑗 = 1, ..., 𝑟, then 𝑢 ∈ C 𝜔 (U);
(c) there is a covering of M by open sets U such that if 𝑢 ∈ D ′ (U) and if
𝑋 𝑗 𝑢 ∈ C ∞ (U) for every 𝑗 = 1, ..., 𝑟, then 𝑢 ∈ C ∞ (U).

Proof Proof
that (a)=⇒(b).
Let U be an arbitrary open subset of M. Let 𝑋 𝐽 =
Ad 𝑋 𝑗1 · · · Ad 𝑋 𝑗ℓ−1 𝑋 𝑗ℓ (𝐽 = { 𝑗1 , ..., 𝑗ℓ}, 1 ≤ 𝑗 𝛼 ≤ 𝑟, 𝛼 = 1, ..., ℓ). Suppose
𝑓 𝐽 = 𝑋 𝐽 𝑢 ∈ C 𝜔 (U) if |𝐽 | ≤ ℓ. We have 𝑋 𝑗◦ , 𝑋 𝐽 𝑢 = 𝑋 𝑗◦ 𝑓 𝐽 − 𝑋 𝐽 𝑓 𝑗◦ ∈ C 𝜔 (U)
whence the result that 𝑓 𝐽 = 𝑋 𝐽 𝑢 ∈ C 𝜔 (U) if |𝐽 | = ℓ + 1, and thus for any
value of ℓ. We apply Proposition 12.1.3 and conclude that 𝑋𝑢 ∈ C 𝜔 (U) for any
𝑋 ∈ 𝔤 (𝑋1 , ..., 𝑋𝑟 ). Since (a) states that 𝔤 (𝑋1 , ..., 𝑋𝑟 ) contains a C 𝜔 frame of 𝑇 ∗ M
over a neighborhood of any point of U the conclusion ensues. The same argument
with C ∞ in the place of C 𝜔 proves that (a)=⇒(c).
Proof that (c)=⇒(a). If (a) does not hold there is a Nagano leaf L of 𝔤 (𝑋1 , ..., 𝑋𝑟 )
in M of codimension 𝑚 ≥ 1. Let (U, 𝑥1 , ..., 𝑥 𝑛 ) be an analytic coordinates chart in
M centered at a point ℘◦ of L and such that

L ∩ U = {𝑥 ∈ U; 𝑥1 = · · · = 𝑥 𝑚 = 0} .

The following defines a distribution in U:

∫
Cc∞ (U) ∋ 𝜑 −→ ⟨𝛿 𝐿 , 𝜑⟩ = 𝜑 (0, ..., 0, 𝑥 𝑚+1 , ..., 𝑥 𝑛 ) d𝑥 𝑚+1 · · · d𝑥 𝑛 .
L∩U
Í𝑛
In L ∩ U we have 𝑋 𝑗 = 𝑘=𝑚+1 𝑎 𝑗,𝑘 𝜕𝑥𝜕𝑘 , 𝑎 𝑗,𝑘 ∈ C 𝜔 (L ∩ U), whence
𝑛 ∫
∑︁ 𝜕
𝑋 𝑗 𝛿𝐿 , 𝜑 = − 𝑎 𝑗,𝑘 𝜑 (0, ..., 0, 𝑥 𝑚+1 , ..., 𝑥 𝑛 ) d𝑥 𝑚+1 · · · d𝑥 𝑛 = 0.
𝑘=𝑚+1 L∩U 𝜕𝑥 𝑘

Thus 𝑋 𝑗 𝛿 𝐿 = 0 in U, 𝑗 = 1, ..., 𝑟, negating (c). The preceding argument also proves

that (b)=⇒(a). □

Corollary 12.5.8 If the system 𝑋1 , ..., 𝑋𝑟 is of finite type then, whatever 𝑢 ∈ D ′ (M),
the property that 𝑋 𝑗 𝑢 ∈ C 𝜔 (M) [resp., C ∞ (M)] for every 𝑗 = 1, ..., 𝑟, implies
𝑢 ∈ C 𝜔 (M) [resp., C ∞ (M)].

Definition 12.5.9 The system of real C 𝜔 vector fields 𝑋1 , ..., 𝑋𝑟 is said to be elliptic
in an open subset U ⊂ M if 𝑋1 , ..., 𝑋𝑟 span 𝑇℘ M at every point ℘ ∈ U.
A system of complex C 𝜔 vector fields 𝑋1 , ..., 𝑋𝑟 is said to be elliptic in U if this
is true of the system Re 𝑋1 , ..., Re 𝑋𝑟 , Im 𝑋1 , ..., Im 𝑋𝑟 .

Every elliptic system of vector fields is of finite type.

370 12 Lie Algebras of Vector Fields

Proposition 12.5.10 Let 𝑋1 , ..., 𝑋𝑟 be real C 𝜔 vector fields in an open set U ⊂ M.

The following properties are equivalent:
(1) The system of vector fields 𝑋1 , ..., 𝑋𝑟 is elliptic in U.
(2) The set

Σ = (𝑥, 𝜉) ∈ 𝑇 ∗ M; 𝑥 ∈ U, 0 ≠ 𝜉 ∈ 𝑇𝑥 M, ⟨𝜎, 𝑋 𝑗 ⟩ (𝑥, 𝜉) = 0, 𝑗 = 1, ..., 𝑟

(12.5.1)
is empty.
(3) The second-order differential operator 𝐿 = 𝑋12 + · · · + 𝑋𝑟2 is elliptic in U.
In (12.5.1) 𝜎 stands for the tautological form (Subsection 9.3.3).
Proof There is no loss of generality in assuming that U is the domain of local
analytic coordinates 𝑥1 , ..., 𝑥 𝑛 . If the expressions of 𝑋 𝑗 in those coordinates is 𝑋 𝑗 =
Í𝑛 𝜕 Í𝑛
𝑘=1 𝑎 𝑗,𝑘 (𝑥) 𝜕𝑥 𝑘 we have ⟨𝜎, 𝑋 𝑗 ⟩ (𝑥, 𝜉) = 𝑘=1 𝑎 𝑗,𝑘 (𝑥) 𝜉 𝑘 . It is then clear that
Properties (2) and (3) are equivalent, since Σ is the characteristic set of the differential
operator 𝐿. Property (1) states that the system of vector fields 𝑋1 , ..., 𝑋𝑟 contains a
basis of 𝑇𝑥 M whatever 𝑥 ∈ U, precluding that the one-form 𝜎 be orthogonal to all
of them at any point 𝜉 ∈ 𝑇𝑥∗ M, 𝜉 ≠ 0. □
For complex vector fields the analogous statement is true taking 𝐿 = 𝑋 1 𝑋1 + · · · +
𝑋 𝑟 𝑋𝑟 with 𝑋 𝑗 the complex conjugate of 𝑋 𝑗 .
If the system of (real, real-analytic) vector fields 𝑋1 , ..., 𝑋𝑟 is of finite type but not
elliptic the second-order differential operator 𝐿 = 𝑋12 + · · · + 𝑋𝑟2 need not be analytic
hypoelliptic (Definition 3.1.2). Example: the operator (4.3.10).

12.5.2 Vector fields generating finite-dimensional Lie algebras

12.5.2.1 Commuting vector fields

We continue to assume the vector fields 𝑋1 , ..., 𝑋𝑟 to be real and real-analytic. The
Lie algebra 𝔤 (𝑋1 , ..., 𝑋𝑟 ) over R need not be finite-dimensional.

Example 12.5.11 Consider the vector fields 𝑋1 = 𝜕/𝜕𝑥1 , 𝑋2 = exp − 21 𝑥 12 𝜕/𝑥1 in

R. We have dimR 𝔤 (𝑋1 , 𝑋2 ) = +∞ since the polynomials exp 21 𝑥 12 𝑋1𝑚 exp − 21 𝑥12
form an infinite-dimensional linear space (cf. the Appendix to Ch. 4).
When the vector fields 𝑋1 , ..., 𝑋𝑟 commute 𝔤 (𝑋1 , ..., 𝑋𝑟 ) is the real vector space
spanned by 𝑋1 , ..., 𝑋𝑟 and dimR 𝔤 (𝑋1 , ..., 𝑋𝑟 ) ≤ 𝑟. In the commuting case the Nagano
leaves need not all have the same dimension (see Example 12.5.4). The following
can be said:
Proposition 12.5.12 Let the real C 𝜔 vector fields 𝑋1 , ..., 𝑋𝑟 in M commute and let
L be a Nagano leaf of 𝑋1 , ..., 𝑋𝑟 . There is a frame of the tangent bundle 𝑇 L over the
whole of L consisting of vector fields 𝑋𝑖1 , ..., 𝑋𝑖𝜈 (1 ≤ 𝑖1 < · · · < 𝑖 𝜈 ≤ 𝑟, 𝜈 = dim L).
12.5 Systems of Vector Fields Generating Special Lie Algebras 371

Proof If 𝜈 = 𝑟 there is nothing to prove: the vector fields 𝑋1 , ..., 𝑋𝑟 must be linearly
independent at every point of L. Suppose dim L < 𝑟 and let ℘ ∈ L be arbitrary.
Since 𝑇 L is spanned by 𝑋1 , ..., 𝑋𝑟 at every point we can select the indices 𝑖 𝛼
(𝛼 = 1, ..., 𝜈) so that 𝑋𝑖1 , ..., 𝑋𝑖𝜈 span 𝑇 L in a neighborhood U of ℘ in the manifold
L. Let 𝑖 ∈ {1, ..., 𝑟 }, 𝑖 ≠ 𝑖 𝛼 for every 𝛼 = 1, ..., 𝜈. We have, in U,
𝜈
∑︁
𝑋𝑖 = 𝑐 𝛼 𝑋𝑖 𝛼 , 𝑐 𝛼 ∈ C 𝜔 (U) .
𝛼=1

Í
But 𝑋𝑖𝛽 , 𝑋𝑖 = 𝜈𝛼=1 𝑋𝑖𝛽 𝑐 𝛼 𝑋𝑖 𝛼 = 0 for each 𝛽 = 1, ..., 𝜈, implying 𝑋𝑖𝛽 𝑐 𝛼 ≡ 0
in U Ífor all 𝛼, 𝛽, and therefore 𝑐 𝛼 = const. for all 𝛼. The analytic vector field
𝑋𝑖 − 𝜈𝛼=1 𝑐 𝛼 𝑋𝑖 𝛼 vanishes identically in U and therefore also in the connected
analytic manifold L. □
After relabeling we may suppose that 𝑇 L is spanned over L by 𝑋1 , ..., 𝑋𝜈 . Let
𝜔 𝑗 ( 𝑗 = 1, ..., 𝜈) be the dual frame of 𝑇 ∗ L: 𝜔 𝑗 is the unique (analytic) one-form in
L such that ⟨𝜔 𝑗 , 𝑋 𝑘 ⟩ = 1 if 𝑘 = 𝑗 and 0 otherwise. By (12.1.8) we have

d𝜔 𝑗 (𝑋ℎ , 𝑋𝑖 ) = 𝑋ℎ ⟨𝜔 𝑗 , 𝑋𝑖 ⟩ − 𝑋𝑖 ⟨𝜔 𝑗 , 𝑋ℎ ⟩ − ⟨𝜔 𝑗 , [𝑋ℎ , 𝑋𝑖 ]⟩ = 0

implying that 𝜔 𝑗 is closed. According to Theorem 9.4.10, an arbitrary point ℘◦ ∈ L

has a neighborhood U (℘◦ ) in L in which 𝜔 𝑗 = d 𝑓 𝑗 for some analytic function 𝑓 𝑗 .
We have 𝑋 𝑘 𝑓 𝑗 = 1 if 𝑘 = 𝑗 and 0 otherwise. Thus (U (℘◦ ) , 𝑓1 , ..., 𝑓 𝜈 ) is an analytic
coordinate patch in L and we can write 𝑋 𝑗 = 𝜕𝜕𝑓 𝑗 in U (℘◦ ) ( 𝑗 = 1, ..., 𝜈). In general
the functions 𝑓 𝑗 cannot be extended to the entire leaf L. The leaf L is foliated by
the integral curves of 𝑋 𝑗 , 1 ≤ 𝑗 ≤ 𝜈. For any ℘ ∈ L let 𝛾 𝑗 (℘) denote the integral
curve of 𝑋 𝑗 passing through ℘. Since there exist no critical points of 𝑋 𝑗 in L the
only boundary points of 𝛾 𝑗 (℘) in L are the points of 𝛾 𝑗 (℘) itself. The functions 𝑓 𝑘
(1 ≤ 𝑘 ≤ 𝜈, 𝑘 ≠ 𝑗) are constant on each connected component of the intersection
U (℘◦ ) ∩ 𝛾 𝑗 (℘). Keep in mind that there might be several, even infinitely many such
connected components (see Subsection 12.2.2).

12.5.2.2 Nilpotent Lie algebras

The most basic noncommutative case is the following

Example 12.5.13 Consider the two vector fields in R2 ,

𝜕 𝜕
𝑋1 = , 𝑋2 = 𝑥 1 .
𝜕𝑥1 𝜕𝑥2
𝜕
We have [𝑋1 , 𝑋2 ] = 𝜕𝑥2 and thus the system 𝑋1 , 𝑋2 has finite type.
372 12 Lie Algebras of Vector Fields

A closely related case is a standard representation of the Heisenberg Lie algebra

𝔥1 :

Example 12.5.14 Consider the two vector fields in R3 ,

𝜕 1 𝜕 𝜕 1 𝜕
𝑋= − 𝑥2 ,𝑌 = + 𝑥1 .
𝜕𝑥1 2 𝜕𝑥3 𝜕𝑥2 2 𝜕𝑥3

We have [𝑋, 𝑌 ] = 𝜕𝑥𝜕 3 and thus 𝔤 (𝑋, 𝑌 ) has rank 3 at every point. The whole space
R3 is the Nagano leaf of 𝔤 (𝑋, 𝑌 ).

Examples 12.5.13 and 12.5.14 are prototypes of the class of (non-Abelian) Lie
algebras 𝔤 (𝑋1 , ..., 𝑋𝑟 ) called nilpotent, defined by the following property:
(Nil) There is a positive integer ℓ such that |𝐼 | > ℓ implies 𝑋𝐼 = 0 [see (12.1.7)].

If (Nil) holds, dim 𝔤 (𝑋1 , ..., 𝑋𝑟 ) < +∞. The converse is not true: 𝔤 (𝑋1 , ..., 𝑋𝑟 )
can be finite-dimensional (as well as of finite type) without being nilpotent, as shown
𝜕 𝜕
by the trivial example of the two vector fields 𝑋1 = 𝜕𝑥 and 𝑋2 = 𝑥 𝜕𝑥 on the real
line, or the more interesting

Example 12.5.15 Consider the three vector fields in R2 ,

𝜕 𝜕 𝜕
𝑋1 = , 𝑋2 = 𝑥 1 , 𝑋3 = −𝑥 2 .
𝜕𝑥1 𝜕𝑥2 𝜕𝑥2

We have [𝑋1 , 𝑋2 ] = 𝜕𝑥𝜕 2 implying that 𝔤 (𝑋1 , 𝑋2 , 𝑋3 ) is of finite type. But

(Ad 𝑋3 ) 𝑘 𝑋2 = 𝑋2 for every 𝑘 ∈ Z+ ; 𝑋1 , 𝑋2 , 𝑋3 , [𝑋1 , 𝑋2 ] form a linear basis of
𝔤 (𝑋1 , 𝑋2 , 𝑋3 ) over R. In passing note that the system of vector fields 𝑋1 , 𝑋2 , 𝑋3 is
not elliptic: the characteristic set of 𝑋12 + 𝑋22 + 𝑋32 is the union of the two rays in 𝑇 ∗ R2
defined by 𝑥1 = 𝑥2 = 𝜉1 = 0, 𝜉2 ≷ 0.

We know that the flow exp (𝑡 𝑋) of a regular vector field 𝑋 in M (see Section
12.3) defines local regular isomorphisms between relatively compact open subsets
of M for suitably small values of |𝑡|. The next statement is a direct consequence of
the Campbell–Hausdorff formula (Theorem 12.3.2).

Proposition 12.5.16 Let 𝑋, 𝑌 be two C ∞ vector fields in a C ∞ manifold M. For the

Lie algebra 𝔤 (𝑋, 𝑌 ) to be nilpotent it is necessary and sufficient that the formal power
series 𝐹 (𝑋, 𝑌 , 𝑠, 𝑡) in the Campbell–Hausdorff formula (12.3.3) be a polynomial.

The polynomial 𝐹 (𝑋, 𝑌 , 𝑠, 𝑡) is related to a group composition law between the

“operators” exp (𝑡 𝑋), exp (𝑡𝑌 ).

Example 12.5.17 Consider the vector fields 𝑋 and 𝑌 in R3 in Example 12.5.14.

Solving the differential equation (12.3.1) for the local flows of the vectors in the base
{𝑋, 𝑌 , [𝑋, 𝑌 ]} of 𝔤 (𝑋, 𝑌 ) yields directly
12.5 Systems of Vector Fields Generating Special Lie Algebras 373

1
exp (𝑡1 𝑋) (𝑥) = 𝑥1 + 𝑡 1 , 𝑥2 , 𝑥3 − 𝑡 1 𝑥 2 ,
2

1
exp (𝑡2𝑌 ) (𝑥) = 𝑥1 , 𝑥2 + 𝑡2 , 𝑥3 + 𝑡 2 𝑥1 ,
2
exp (𝑡 3 [𝑋, 𝑌 ]) (𝑥) = (𝑥 1 , 𝑥2 , 𝑥3 + 𝑡3 ) .

for any 𝑥 = (𝑥1 , 𝑥2 , 𝑥3 ) ∈ R3 . We observe that each of the above three vector flows
defines a one-parameter subgroup of the group of affine transformations of R3 ,
A (3, R), and combined, they define an injection R3 ∋ 𝑡 ↦→ 𝒈 (𝑡) ∈ A (3, R) 3 . A
simple calculation shows that

1
exp (𝑦 1 𝑋) exp (𝑦 2𝑌 ) exp 𝑦 3 + 𝑦 1 𝑦 2 [𝑋, 𝑌 ] (𝑥)
2

1
= 𝑥1 + 𝑦 1 , 𝑥2 + 𝑦 2 , 𝑥3 + 𝑦 3 + (𝑥 1 𝑦 2 − 𝑥2 𝑦 1 ) .
2

When equipped with the group structure defined by the right-hand side (taken as a
composition law 𝑥 ∗ 𝑦) R3 becomes the Heisenberg group H1 whose Lie algebra is
𝔥1 = 𝔤 (𝑋, 𝑌 ) (identifiable to the tangent space of H1 at the identity).

Example 12.5.17 raises the following question (which we content ourselves with
mentioning). Assuming that the Lie algebra 𝔤 (𝑋1 , ..., 𝑋𝑟 ) is nilpotent, is there a
nilpotent Lie group structure on each Nagano leaf L of 𝑋1 , ..., 𝑋𝑟 whose Lie algebra
is the restriction of 𝔤 (𝑋1 , ..., 𝑋𝑟 ) to L?
Many systems of vector fields 𝑋1 , ..., 𝑋𝑟 that serve as “models” in PDE theory
generate nilpotent Lie algebras 𝔤 (𝑋1 , ..., 𝑋𝑟 ).
Chapter 13
Elements of Symplectic Geometry

The first section of this chapter is a primer on symplectic algebra and geometry in
Euclidean spaces, real or complex. Concepts and results are extended to manifolds
in Section 13.3 and revolve around the Darboux Theorem 13.3.20, which asserts the
existence of local Darboux coordinates in an arbitrary symplectic manifold. Poisson
brackets, Hamiltonian vector fields and their integral manifolds play a crucial role
in later parts of the book; in the last subsection of Section 13.3 they are used
to introduce systems of functions (defined in a symplectic manifold) of finite type
(comprising elliptic systems). Section 13.2 presents the metaplectic group, important
in pseudodifferential calculus. The contents of the first three sections of this chapter
are classical, and are covered in numerous other texts, often in greater detail. The
last three sections of the chapter contain material that is somewhat less standard.
In Section 13.4 the reader can find the statement and proof of the local existence
and uniqueness of the regular solutions, satisfying the natural Cauchy conditions,
to the systems of eikonal equations that we call involutive and of principal type.
This result (not unexpected and not particularly difficult) enables us to construct the
phase-functions, real or complex, in the Fourier Integral Operators needed to simplify
involutive systems of PDEs of principal type discussed in the last Part of the book.1 A
step in this simplification √
is to transfer the analysis from the Euclidean phase-space,
now identified with R𝑛𝑥 × −1R𝑛𝜉 , to the outer conormal bundle, 𝑁out ∗ Ω, of a strictly

pseudoconvex domain Ω in C . As real submanifolds, of the same dimension

𝑛 2𝑛, of
2𝑛
√ 𝑛
C equipped with its standard complex symplectic structure, R 𝑥 × −1 R 𝜉 \ {0}
𝑛
∗ Ω are both I-symplectic and R-Lagrangian, important properties introduced
and 𝑁out
in [Schapira, 1981]. The symplectic aspect of all this is explained in the last two
sections of the chapter.

1 This material is also related to the existence and approximation of solutions of first-order nonlinear
PDEs (see [Métivier, 1985], [Treves, 1992], Ch. 10).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 375
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_13
376 13 Elements of Symplectic Geometry

13.1 Elements of Symplectic Algebra

13.1.1 Symplectic vector spaces

As in Chapter 9 the scalar field will be denoted by K which, for us, means R or C.
In either case we denote by 𝑥1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 the coordinates in K2𝑛 K𝑛 × K𝑛 .
It will be convenient, sometimes, to use the notation 𝑧 = (𝑥, 𝑦) for the point in K2𝑛 .
First we consider an arbitrary finite-dimensional vector space E over the field K;
we denote by E∗ its dual.
Definition 13.1.1 A bilinear form 𝛽 on E is called a symplectic form if 𝛽 is skew-
symmetric and nondegenerate. The pair (E, 𝛽) consisting of a vector space and a
symplectic form on it will be called a symplectic vector space.
That 𝛽 is nondegenerate means that to every 𝒗 ∈ E, 𝒗 ≠ 0, there is a 𝒗 ′ ∈ E such
that 𝛽 (𝒗, 𝒗 ′) ≠ 0. If 𝛽 is a symplectic form on E the map
𝛽∗
E ∋𝒗 ↦→ (𝒗 ′ ↦→ 𝛽 (𝒗, 𝒗 ′)) ∈ E∗ (13.1.1)

(𝒗, 𝒗 ′ ∈ E) is injective and therefore it is an isomorphism. If 𝒗 ∈ E then

∀𝒗 ′ ∈ E, 𝛽 (𝒗, 𝒗 ′) = ⟨𝛽∗ 𝒗, 𝒗 ′⟩. (13.1.2)

Example 13.1.2 Let E∗ be the dual of the finite-dimensional vector space E. The
skew symmetric bilinear form

𝜛E 𝒙 1 , 𝒙 ∗1 , 𝒙 2 , 𝒙 ∗2 = ⟨𝒙 ∗1 , 𝒙 2 ⟩ − ⟨𝒙 ∗2 , 𝒙 1 ⟩

(13.1.3)

is a symplectic form on E × E∗ .
Example 13.1.3 In Example 13.1.2 we take E = K𝑛 ; we identify K𝑛 with its own
dual through the scalar product 𝑥 · 𝑦 = 𝑥1 𝑦 1 + · · · + 𝑥 𝑛 𝑦 𝑛 and thus we identify
E × E∗ with K2𝑛 K𝑛 × K𝑛 . On K2𝑛 the form (13.1.3) reads, for vectors 𝑧 = (𝑥, 𝑦),
𝑧 ′ = (𝑥 ′, 𝑦 ′) in K2𝑛 ,
𝑛
∑︁
𝜛𝑛 (𝑧, 𝑧 ′) = 𝑥 ′𝑗 𝑦 𝑗 − 𝑥 𝑗 𝑦 ′𝑗 . (13.1.4)
𝑗=1

The isomorphism K2𝑛 −→ K 2𝑛 ∗ = K2𝑛 defined by (13.1.4) is the map (𝑥, 𝑦) ↦→

(𝑦, −𝑥).
Proposition 13.1.4 If (E, 𝛽) is a symplectic vector space then dim E is even.
Proof Let 𝒆 1 , ..., 𝒆 𝑁 be a basis of the vector space E and let 𝒆 ∗1 , ..., 𝒆 ∗𝑁 be the dual
basis in E∗ , meaning that ⟨𝒆 ∗𝑗 , 𝒆 𝑘 ⟩ = 𝛿 𝑗,𝑘 for all 𝑗, 𝑘 = 1, ..., 𝑁. If we write

𝑁
−1 ∑︁
𝛽∗ 𝒆 ∗𝑗 = 𝑏 𝑗,𝑘 𝒆 𝑘 , 𝑗 = 1, ..., 𝑁,
𝑘=1
13.1 Elements of Symplectic Algebra 377

then 𝑁
−1 −1 −1 ∑︁
𝛽 𝛽∗ 𝒆 ∗𝑖 , 𝛽∗ 𝒆 ∗𝑗 = ⟨𝒆 ∗𝑖 , 𝛽∗ 𝒆 ∗𝑗 ⟩ = 𝑏 𝑗,𝑘 ⟨𝒆 ∗𝑖 , 𝒆 𝑘 ⟩ = 𝑏 𝑗𝑖 ,
𝑘=1

which shows that the matrix 𝑏 𝑗,𝑘 𝑗=1,..., 𝑁 is skew-symmetric. Since 𝛽∗ is a linear
𝑘=1,..., 𝑁
bijection this matrix must also be nonsingular, which is only possible if 𝑁 is even.□
The next statement is self-evident.
Proposition 13.1.5 Let 𝑆 be a subset of E; then

𝑆 ⊥𝛽 = {𝒗 ∈ E; ∀𝒗 ′ ∈ 𝑆, 𝛽 (𝒗, 𝒗 ′) = 0}

is a K-linear subspace of E.
Let 𝑾 be a vector subspace of E; then 𝑾 ⊥𝛽 is called the orthogonal of 𝑾 for
the bilinear form 𝛽. The map (13.1.1) transforms 𝑾 ⊥𝛽 into 𝑾 ⊥ , the orthogonal of
𝑾 for the duality between E and E∗ . It follows that dim 𝑾 ⊥𝛽 = codim 𝑾 and that
⊥𝛽
𝑾 ⊥𝛽 = 𝑾.
Whereas 𝑾 ∩ 𝑾 ⊥ = {0}, generally speaking 𝑾 ∩ 𝑾 ⊥𝛽 ≠ {0}. The dimension of
the intersection 𝑾 ∩ 𝑾 ⊥𝛽 characterizes special classes of vector subspaces of E:
Definition 13.1.6 The vector subspace 𝑾 of E is said to be
(1) symplectic if 𝑾 ∩ 𝑾 ⊥𝛽 = {0};
(2) isotropic if 𝑾 ⊂ 𝑾 ⊥𝛽 ;
(3) coisotropic if 𝑾 ⊥𝛽 ⊂ 𝑾;
(4) Lagrangian if 𝑾 ⊥𝛽 = 𝑾.
Proposition 13.1.7 If the vector subspace 𝑾 of E is symplectic then necessarily
dim 𝑾 is even, 𝑾 ⊥𝛽 is also symplectic and E = 𝑾⊕𝑾 ⊥𝛽 . If 𝑾 is isotropic, dim 𝑾 ≤
1 1
2 dim E; if 𝑾 is coisotropic, dim 𝑾 ≥ 2 dim E. If 𝑾 is Lagrangian, dim 𝑾 =
1
2 dim E.

Proof The restriction of 𝛽 to a symplectic vector subspace 𝑾 of E is nondegenerate

and therefore dim 𝑾 must be even, by Proposition 13.1.4; 𝑾 ⊥𝛽 is symplectic because
the relation 𝑾∩𝑾 ⊥𝛽 = {0} is symmetric; dimensionality entails E = 𝑾⊕𝑾 ⊥𝛽 . If
𝑾 is isotropic dim 𝑾 ≤ dim 𝑾 ⊥𝛽 = dim E− dim 𝑾. If 𝑾 is coisotropic then 𝑾 ⊥𝛽
is isotropic. A Lagrangian subspace is both isotropic and coisotropic. □
Example 13.1.8 Suppose 1 ≤ 𝑚 < 𝑛 and let (13.1.4) define the symplectic structure
of K2𝑛 . We identify K𝑚 with the vector subspace of K𝑛 defined by 𝑥 𝑚+1 = · · · =
𝑥 𝑛 = 0, and K2𝑚 with the vector subspace of K2𝑛 defined by 𝑥 𝑚+1 = · · · = 𝑥 𝑛 =
𝑦 𝑚+1 = · · · = 𝑦 𝑛 = 0. Then K2𝑚 is a symplectic vector subspace of K2𝑛 ; we have
⊥𝛽 ⊥
K2𝑚 = K2𝑚 , the orthogonal of K2𝑛 for the dot product. If we were to identify
K2𝑚 with the vector subspace of K2𝑛 defined by 𝑥 2𝑚+1 = · · · = 𝑥2𝑛 = 0 (writing
𝑥 𝑛+ 𝑗 in lieu of 𝑦 𝑗 ) then K2𝑚 would be isotropic (resp. coisotropic) if 2𝑚 ≤ 𝑛 (resp.,
2𝑚 ≥ 𝑛) and Lagrangian if 2𝑚 = 𝑛.
378 13 Elements of Symplectic Geometry

Example 13.1.9 If K = R and if we write 𝑧 = 𝑥 + 𝑖𝑦 and 𝑧 ′ = 𝑥 ′ + 𝑖𝑦 ′ then (13.1.4)

can be viewed as a real symplectic form on C𝑛 . Note that (13.1.4) is equal to the
imaginary part of the hermitian inner product on C𝑛 ,

𝑧 · 𝑧 ′ = 𝑥 · 𝑥 ′ + 𝑦 · 𝑦 ′ + 𝑖 (𝑥 ′ · 𝑦 − 𝑥 · 𝑦 ′) , (13.1.5)

the real part of 𝑧 · 𝑧 ′ being the real scalar product.

Proposition 13.1.10 Let (E, 𝛽) be a symplectic vector space and let dim E = 2𝑛.
There is a linear basis 𝜺 1 , ..., 𝜺 2𝑛 of E such that

(•) 𝛽 𝜺 𝑗 , 𝜺 𝑘 = 0 for all 𝑗, 𝑘, 1 ≤ 𝑗 < 𝑘 ≤ 2𝑛, unless 𝑗 is odd and 𝑘 = 𝑗 + 1, in
which case 𝛽 𝜺 𝑗 , 𝜺 𝑘 = 1.

Proof We reason by induction on 𝑛 ≥ 1. Suppose 𝜺 1 , ..., 𝜺 2𝑛−2 satisfy (•) with 𝑛 − 1

instead of 𝑛. Let F be the linear span (over K) of 𝜺 1 , ..., 𝜺 2𝑛−2 . We select at random
two linearly independent linear functionals 𝒙 ∗𝑖 ∈ E∗ , 𝑖 = 1, 2, vanishing identically
−1
on F. If we define 𝒙 𝑖 = 𝛽∗ 𝜺 ∗𝑖 we get

𝛽 𝒙 𝑖 , 𝜺 𝑗 = ⟨𝒙 ∗𝑖 , 𝜺 𝑗 ⟩ = 0, 𝑖 = 𝑗 = 1, ..., 2𝑛 − 1.

We cannot have 𝛽 (𝒙 1 , 𝒙 2 ) = 0 otherwise 𝒙 1 and 𝒙 2 would be orthogonal with respect

to 𝛽 to the whole space E. As a consequence we can take 𝜺 2𝑛−1 = 𝛽 (𝒙 1 , 𝒙 2 ) −1 𝒙 1 ,
𝜺 2𝑛 = 𝒙 2 . □

Definition 13.1.11 Any basis 𝜺 1 , ..., 𝜺 2𝑛 of the symplectic vector space (E, 𝛽) that
has a permutation with Property (•) in Proposition 13.1.10 will be called a symplectic
basis of E.

Let now (E 𝜄 , 𝛽 𝜄 ) (𝜄 = 1, 2) be two symplectic vector spaces with scalar field K.

A linear map 𝑓 : E1 −→ E2 is called a symplectic map if for every pair of vectors
𝒗, 𝒗 ′ ∈ E1 we have
𝛽2 ( 𝑓 (𝒗) , 𝑓 (𝒗 ′)) = 𝛽1 (𝒗, 𝒗 ′) . (13.1.6)
A symplectic isomorphism is a bijective symplectic map; a symplectic automorphism
𝑓 of (E, 𝛽) is a symplectic isomorphism of E onto itself. In other words, 𝑓 is a linear
automorphism of E which preserves the form 𝛽:

∀𝒗, 𝒗 ′ ∈ E, 𝛽 ( 𝑓 (𝒗) , 𝑓 (𝒗 ′)) = 𝛽 (𝒗, 𝒗 ′) . (13.1.7)

A restatement of Proposition 13.1.10 is that any 2𝑛-dimensional symplectic vector

space over K is symplectically isomorphic to the symplectic space K2𝑛 , 𝜛𝑛 .
13.1 Elements of Symplectic Algebra 379

13.1.2 Lagrangian subspaces

In connection with Proposition 13.1.10 it is noteworthy that the choice of the sym-
plectic isomorphism (E, 𝛽) K2𝑛 , 𝜛𝑛 can be adjusted to give a special role to a
given pair of transverse Lagrangian vector subspaces of E:

Proposition 13.1.12 If 𝑳 1 and 𝑳 2 are two Lagrangian linear subspaces of E such

that 𝑳 1 ∩ 𝑳 2 = {0} then to every linear basis 𝜺 1,1 , ..., 𝜺 1,𝑛 of 𝑳 1 there is a linear

basis 𝜺 2,1 , ..., 𝜺 2,𝑛 of 𝑳 2 such that 𝛽 𝜺 1,𝑘 , 𝜺 2,ℓ = 0 if 𝑘 ≠ ℓ and 𝛽 𝜺 1,𝑘 , 𝜺 2,𝑘 = 1
for every 𝑘.

Proof Given an arbitrary v1 ∈ 𝑳 1 there is a v2 ∈ 𝑳 2 such that 𝛽 (v1 , v2 ) ≠ 0; if

this were not so the linear subspace 𝑳 2 ⊕ Kv1 of E would be isotropic and have
dimension > 𝑛. We can therefore identify 𝑳 2 with the dual of 𝑳 1 with 𝛽 as the
duality bracket (cf. Example 13.1.2). We can take 𝜺 2,1 , ..., 𝜺 2,𝑛 to be the basis in 𝑳 2
dual to 𝜺 1,1 , ..., 𝜺 1,𝑛 . □

The system of vectors 𝜺 1,1 , ..., 𝜺 1,𝑛 , 𝜺 2,1 , ..., 𝜺 2,𝑛 is then a symplectic basis of
(E, 𝛽) (Definition 13.1.11). Let 𝒆 1 , ..., 𝒆 2𝑛 be the canonical basis in K2𝑛 ; 𝒆 𝑗 is the
unit vector along the 𝑗 th axis (or coordinate plane when K = C). The linear map that
transforms 𝜺 1, 𝑗 into 𝒆 𝑗 and 𝜺 2, 𝑗 into 𝒆 𝑛+ 𝑗 (1 ≤ 𝑗 ≤ 𝑛) is a symplectic isomorphism
which transforms 𝑳 1 into K𝑛 × {0}, 𝑳 2 into {0} × K𝑛 and the bilinear form 𝛽 into
the standard symplectic form 𝜛𝑛 on K2𝑛 [see (13.1.4)].
The following result will be useful later:

Proposition 13.1.13 If 𝑳 1 and 𝑳 2 are two Lagrangian vector subspaces of E such

that 𝑳 1 ∩ 𝑳 2 = {0} then there is a third Lagrangian vector subspace 𝑳 3 of E such
that 𝑳 𝑗 ∩ 𝑳 3 = {0} , 𝑗 = 1, 2.

Proof As we have just seen it suffices to consider the case E = K2𝑛 , 𝛽 = 𝜛𝑛 ,

𝑳 1 = K𝑛 × {0}, 𝑳 2 = {0} × K𝑛 . Let then 𝜃 ∈ (0, 1) be arbitrary. The vector subspace
𝑉 𝜃 spanned by the vectors 𝜃𝒆 𝑗 + (1 − 𝜃) 𝒆 𝑛+ 𝑗 , 𝑗 = 1, ..., 𝑛, can be taken as the
Lagrangian subspace 𝑳 3 . □

We are also going to need the next statement:

Proposition 13.1.14 If 𝑴 1 and 𝑴 2 are two isotropic vector subspaces of E such

that 𝑴 1 ∩ 𝑴 2 = {0} then there are Lagrangian vector subspaces 𝑳 𝑗 of E such that
𝑴 𝑗 ⊂ 𝑳 𝑗 ( 𝑗 = 1, 2) and 𝑳 1 ∩ 𝑳 2 = {0}.

Proof Unless we have dim 𝑴 𝑗 < 𝑛 for 𝑗 = 1 or 𝑗 = 2 there is nothing to prove.

Suppose dim 𝑴 2 < 𝑛 and let 𝔉 denote the family of all isotropic vector subspaces
𝑽 of E such that 𝑴 1 ∩ 𝑽 = {0} and 𝑴 2 ⊂ 𝑽. Let 𝑾 be an element of 𝔉 that has
maximum dimension 𝑑 and denote by 𝑾 ⊥ the orthogonal of 𝑾 for the form 𝛽. We
have dim 𝑾 ⊥ = 2𝑛−𝑑 and 𝑾 ⊂ 𝑾 ⊥ (since 𝑾 is isotropic). Suppose there is 𝒗 ◦ ∈ 𝑾 ⊥ ,
𝒗 ◦ ∉ 𝑾 ⊕ 𝑴 1 . The vector subspace 𝑾e = 𝑾 ⊕ K𝒗 ◦ is isotropic, contains 𝑴 2 and
𝑴1 ∩ 𝑾e = {0}, contradicting the maximality of dim 𝑾. This proves 𝑾 ⊥ ⊂ 𝑴 1 ⊕ 𝑾.
380 13 Elements of Symplectic Geometry

If we had 𝑑 < 𝑛 there would be a subspace 𝑴 0 of 𝑾 ⊥ ∩ 𝑴 1 such that 𝑾 ⊥ = 𝑴 0 ⊕ 𝑾;

𝑴 0 ⊂ 𝑴 1 implies 𝑴 0 isotropic, which in turn implies 𝑴 0 ⊕ 𝑾 isotropic, hence
dim 𝑾 ⊥ ≤ 𝑛, contradicting 𝑑 < 𝑛. We conclude that 𝑑 = 𝑛 and 𝑾 is Lagrangian. If
dim 𝑴 1 < 𝑛 we apply the preceding reasoning to the pair 𝑴 1 , 𝑾. □

Combining Propositions 13.1.13 and 13.1.14 yields a strengthening of the former:

Proposition 13.1.15 Given arbitrarily two Lagrangian vector subspaces 𝑳 1 and 𝑳 2

of E there is a third Lagrangian vector subspace 𝑳 3 of E such that 𝑳 𝑗 ∩ 𝑳 3 =
{0} , 𝑗 = 1, 2.

Proof Let 𝑴 1 be a vector subspace of 𝑳 1 such that 𝑳 1 = 𝑴 1 ⊕ ( 𝑳 1 ∩ 𝑳 2 ). Since

𝑴 1 ∩ 𝑳 2 = {0} we can apply Proposition 13.1.14 taking 𝑴 2 = 𝑳 2 ; we see that then
𝑳 1 of E such that 𝑴 1 ⊂ e
there is a Lagrangian vector subspace e 𝑳 1 ∩ 𝑳 2 = {0}.
𝑳 1 and e
The claim then ensues from Proposition 13.1.13 where we substitute e 𝑳 1 for 𝑳 1 . □
Here is a useful application of the preceding results:

Proposition 13.1.16 Given arbitrarily two Lagrangian vector subspaces 𝑳 1 and 𝑳 2

of E there is a symplectic automorphism 𝐴 of (E, 𝛽) such that 𝐴𝑳 1 = 𝑳 2 .

Proof Let 𝑳 3 be a Lagrangian vector subspace of E such that 𝑳 𝑗 ∩ 𝑳 3 = {0}, 𝑗 = 1, 2.

Let 𝜺 𝑗,1 , ..., 𝜺 𝑗,𝑛 be a linear basis of 𝑳 𝑗 ; by Proposition
13.1.12 there are two
linear bases 𝜺 ∗𝑗,1 , ..., 𝜺 ∗𝑗,𝑛 of 𝐿 3 such that 𝛽 𝜺 𝑗,𝑘 , 𝜺 ∗𝑗,ℓ = 𝛿 𝑘,ℓ ( 𝑗 = 1, 2, 1 ≤
𝑘, ℓ ≤ 𝑛). The automorphism 𝐴 is the linear transformation of E which transforms
𝜺 1,1 , ..., 𝜺 1,𝑛 , 𝜺 ∗1,1 , ..., 𝜺 ∗1,𝑛 into 𝜺 2,1 , ..., 𝜺 2,𝑛 , 𝜺 ∗2,1 , ..., 𝜺 ∗2,𝑛 . □

13.1.3 The fundamental matrix

As above let 𝒆 1 , ..., 𝒆 2𝑛 be the canonical basis in K2𝑛 and 𝜛𝑛 the symplectic form
(13.1.4). With this choice we have

𝜛𝑛 𝒆 𝑗+𝑛 , 𝒆 𝑗 = 1 if 1 ≤ 𝑗 ≤ 𝑛; (13.1.8)

𝜛𝑛 𝒆 𝑗 , 𝒆 𝑘 = 0 if | 𝑗 − 𝑘 | ≠ 𝑛.

Definition 13.1.17 The (2𝑛) × (2𝑛) matrix 𭟋𝑛 such that 𝜛𝑛 (𝑧, 𝑧 ′) = 𭟋𝑛 𝑧 · 𝑧 ′ is called
the fundamental symplectic matrix, or simply the fundamental matrix, of the
symplectic space K2𝑛 , 𝜛𝑛 .

The matrix 𭟋𝑛 represents the linear bijection (13.1.1) in the canonical basis of
K2𝑛 when we identify the latter with its dual through the dot product. Thus 𭟋𝑛 is
nonsingular and since 𝜛𝑛 is skew-symmetric so is 𭟋𝑛 . We have, by (13.1.8),

𭟋𝑛 𝒆 𝑗+𝑛 · 𝒆 𝑗 = 1 if 1 ≤ 𝑗 ≤ 𝑛; 𭟋𝑛 𝒆 𝑗 · 𝒆 𝑘 = 0 if | 𝑗 − 𝑘 | ≠ 𝑛.
13.1 Elements of Symplectic Algebra 381

It follows that

𭟋𝑛 𝒆 𝑗 = −𝒆 𝑗+𝑛 if 1 ≤ 𝑗 ≤ 𝑛, 𭟋𝑛 𝒆 𝑗 = 𝒆 𝑗−𝑛 if 𝑛 + 1 ≤ 𝑗 ≤ 2𝑛. (13.1.9)

Using column notation we see that, if 𝑥, 𝑦 ∈ K𝑛 ,

𝑥 −𝑦
𭟋𝑛 = . (13.1.10)
𝑦 𝑥

We have
𭟋2𝑛 = −𝐼2𝑛 , det 𭟋𝑛 = 1, (13.1.11)
where 𝐼2𝑛 is the (2𝑛) × (2𝑛) identity matrix.
Suppose K = R; (13.1.1) shows that 𭟋𝑛 defines a complex structure on R2𝑛 ; this
structure identifies R2𝑛 with C𝑛 : if (𝑥, 𝑦) ∈ R2𝑛 then 𝑧 = 𝑥 + 𝑖𝑦 is identified with the
vector
𝑥 𝑥 𝑦
= + 𭟋𝑛 . (13.1.12)
𝑦 0 0
√
The transformation 𭟋𝑛 : R2𝑛 ←↪ is thus identified with multiplication by −1 in C𝑛 .

13.1.4 Linear groups

Let 𝑚 be a positive integer. We shall denote by

(1) M𝑚 (K) the ring (with respect to matrix addition and multiplication, also the
Lie algebra with respect to the commutation bracket 𝐴𝐵 − 𝐵𝐴) of the 𝑚 × 𝑚
matrices with entries in K, identified with the ring of linear automorphisms of
K𝑚 (with the understanding that we use the canonical basis 𝒆 1 , ..., 𝒆 𝑚 in K𝑚 );
(2) GL (𝑚, K) = {g ∈ M𝑚 (K) ; det g ≠ 0} the general linear group with coeffi-
cients in K, i.e., the group (with respect to matrix multiplication) of nonsingular
matrices g ∈ M𝑚 (K), identified with the group (with respect to composition)
of linear automorphisms of K𝑚 .
The group GL (𝑚, K) is an open (and dense) subset of the vector space M𝑚 (K)
2
K𝑚 from which it inherits the structure of an analytic manifold. Matrix inversion and
matrix multiplication are analytic maps, of GL (𝑚, K) and GL (𝑚, K) × GL (𝑚, K)
respectively, into GL (𝑚, K), making a Lie group of GL (𝑚, K), with the 𝑚 × 𝑚
identity matrix 𝐼𝑚 as the neutral element.
The tangent space to GL (𝑚, K) at the point 𝐼𝑚 can be identified with M𝑚 (K).
It can also be identified with the Lie algebra (with respect to the Lie bracket) of
the sections of the tangent bundle 𝑇GL (𝑚, K) that are invariant under every group
translation h ↦→ hg−1 . If 𝑋 ∈ M𝑚 (K) then exp 𝑡 𝑋 ∈ GL (𝑚, K) for all 𝑡 ∈ K.
Conversely, if g ∈ GL (𝑚, K) is sufficiently close to 𝐼𝑚 there is an 𝑋 ∈ M𝑚 (K) such
that g = exp 𝑋. Of course,
382 13 Elements of Symplectic Geometry

d
𝑋 = lim 𝑡 −1 ((exp 𝑡 𝑋) − 𝐼𝑚 ) = (exp 𝑡 𝑋) . (13.1.13)
0≠𝑡→0 d𝑡 𝑡=0

The Campbell–Hausdorff formula (12.3.3), applies: the series 𝐹 (𝑠, 𝑡, 𝑋, 𝑌 ) is

actually convergent in a domain (𝑠, 𝑡) ∈ K2 ; |𝑠| < 𝜀, |𝑡| < 𝛿 with the positive
numbers 𝜀 and 𝛿 depending on the matrix norms of 𝑋 and 𝑌 respectively.
By a linear Lie group, or simply a linear group, we shall mean an analytic
subgroup G of GL (𝑚, K) (for some value of 𝑚): 𝐼𝑚 is the neutral element in G;
matrix inversion and matrix multiplication are analytic maps, of G and G × G
respectively, into G. The tangent space to G at 𝐼𝑚 is the linear subspace 𝔤 of M𝑚 (K)
consisting of the matrices 𝑋 such that exp (𝑡 𝑋) ∈ G for all 𝑡 ∈ K. If 𝑋, 𝑌 ∈ 𝔤
then K ∋𝑡 ↦→ (exp 𝑋) (exp 𝑡𝑌 ) exp (−𝑋) is an analytic curve in G whose tangent at
the origin is (exp 𝑋) 𝑌 exp (−𝑋). It follows that K ∋𝑡 ↦→ (exp 𝑡 𝑋) 𝑌 exp (−𝑡 𝑋) is an
analytic curve in 𝔤 whose tangent vector at the origin is [𝑋, 𝑌 ]. Since 𝔤 is a linear
subspace of M𝑚 (K) we conclude that [𝑋, 𝑌 ] ∈ 𝔤: 𝔤 is a Lie subalgebra of M𝑚 (K).
We shall refer to 𝔤 as the Lie algebra of the linear group G.
Standard linear groups are the orthogonal groups O (𝑚, K) of matrices g ∈
GL (𝑚, K) such that g⊤ = g−1 , and therefore det g = ±1. The connected component
O+ (𝑚, K) of the identity 𝐼𝑚 in O (𝑚, K) is also a linear Lie group. Other important
examples are the unitary groups U (𝑚) of matrices g ∈ GL (𝑚, C) such that g∗ = g−1 ,
where g∗ is the adjoint of g, i.e., the complex conjugate of the transpose g⊤ . An all
important group is the special linear group SL (2, K) consisting of the 2 × 2 matrices
with entries in K whose determinant is equal to 1.
A matrix 𝐴 ∈ M𝑚 (K) belongs to the Lie algebra 𝔬 (𝑚, K) of O (𝑚, K) if and only
if 𝐴 + 𝐴⊤ = 0, i.e., if 𝐴 is skew-symmetric. This follows directly from differentiating
with respect to 𝑡 at 𝑡 = 0 the equation (exp 𝑡 𝐴) (exp 𝑡 𝐴⊤ ) = 𝐼𝑚 . Likewise, the matrix
𝐴 ∈ M𝑚 (K) belongs to the Lie algebra 𝔲 (𝑚) of U (𝑚) if and only if 𝐴 + 𝐴∗ = 0,
i.e., if 𝐴 is antiself-adjoint.

Proposition 13.1.18 A matrix 𝐴 ∈ M2 (K) belongs to the Lie algebra 𝔰𝔩 (2, K) of

SL (2, K) if and only if Tr 𝐴 = 0.

We have denoted by Tr 𝐴 the trace of the matrix 𝐴, i.e., the sum of its (two)
eigenvalues.
Proof Indeed we have

d d
Tr 𝐴 = det (𝐼2 + 𝑡 𝐴) = det (exp 𝑡 𝐴) (13.1.14)
d𝑡 𝑡=0 d𝑡 𝑡=0

for any 𝐴 ∈ M𝑚 (K). If 𝐴 ∈ 𝔰𝔩 (2, K) then det (exp 𝑡 𝐴) = 1 for all 𝑡 ∈ K, whence
Tr 𝐴 = 0. Now suppose Tr 𝐴 = 0 and det 𝐴 ≠ 0. In this case 𝐴 has two distinct √ eigen-
√
values, the complex numbers ± det 𝐴; the eigenvalues of exp 𝐴 are exp ± det 𝐴
and therefore det (exp 𝐴) = +1. The intersection 𝔰𝔩 (2, K) ∩ GL (2, K) is nonempty
and therefore it is an open and dense subset of the vector subspace 𝔰𝔩 (2, K). It follows
that the continuous function 𝐴 ↦→ det (exp 𝐴) is identically equal to 1 in 𝔰𝔩 (2, K).□
13.1 Elements of Symplectic Algebra 383

13.1.5 The symplectic group

Definition 13.1.19 The group of symplectic automorphisms of R2𝑛 , 𝜛𝑛 will be

called the 𝑛th linear symplectic group (or simply symplectic group) and will be
denoted by Sp (𝑛, R). The group of symplectic automorphisms of C2𝑛 , 𝜛𝑛 will be
called the 𝑛th complex symplectic group and will be denoted by Sp (𝑛, C).
For us the elements of Sp (𝑛, K) will be K-linear transformations of K2𝑛 or
equivalently, through use of the canonical basis of K2𝑛 , (2𝑛) × (2𝑛) matrices with
entries in K. If 𝒈 ∈ Sp (𝑛, K) we have

𝜛𝑛 ( 𝒈𝑧, 𝒈𝑧 ′) = (𭟋𝑛 𝒈𝑧) · 𝒈𝑧 ′ = 𭟋𝑛 𝑧 · 𝑧 ′. (13.1.15)

The equation (13.1.15) is equivalent to the relation

𝒈 ⊤ 𭟋𝑛 𝒈 = 𭟋𝑛 ,

(13.1.16)

which evidently requires 𝒈 to have an inverse 𝒈 −1 that also satisfies (13.1.16).

Proposition 13.1.20 If 𝒈 belongs to the group Sp (𝑛, K) so does its transpose 𝒈 ⊤ .
Proof Indeed 𭟋𝑛 = ( 𝒈 ⊤ ) −1 𭟋𝑛 𝒈 −1 =⇒ 𝒈𭟋𝑛 𝒈 ⊤ = 𭟋𝑛 by inversion. □
The equation (13.1.16) defines Sp (𝑛, K) as an analytic submanifold of GL (2𝑛, K);
matrix inversion and matrix multiplication are analytic. In other words Sp (𝑛, K) is
a linear group (either real or complex). In the proof of the next proposition we shall
apply the following
Lemma 13.1.21 The set 𝑆 𝑛 of pairs of vectors (𝒖, 𝒗) ∈ K2𝑛 ×K2𝑛 such that 𭟋𝑛 𝒖·𝒗 = 1
is connected.
Proof It suffices to deal with the case K = R. If we identify R2𝑛 with C𝑛 we see that
we must prove that the set 𝑆 C𝑛 of pairs of vectors (𝒖, 𝒗) ∈ C𝑛 ×C𝑛 such that 𭟋𝑛 𝒖 · 𝒗 = 1
(with the appropriate reinterpretation of 𭟋𝑛 ) is connected (when K = C we replace
𝑛 by 2𝑛). The automorphism (𝒖, 𝒗) ↦→ (−𝒖, 𭟋𝑛 𝒗) of C𝑛 × C𝑛 (the upper bar means
complex conjugation) induces a homeomorphism of 𝑆 𝑛 onto the set 𝑆 ◦𝑛 of pairs of
vectors (𝒖, 𝒗) ∈ C𝑛 × C𝑛 such that 𝒖 · 𝒗 = 1. To deal with 𝑆 ◦𝑛 rather than
𝑆 𝑛 lightens

the notation and helps to visualize what is being done. Let (𝒖 ◦ , 𝒗 ◦ ) , 𝒖♭ , 𝒗 ♭ ∈ 𝑆 ◦𝑛 ;
since 𝑆 ◦𝑛 is invariant under the transformation (𝒖, 𝒗) ↦→ 𝜆𝒖, 𝜆−1 𝒗 (0 ≠ 𝜆 ∈ C) we

can assume that |𝒖 ◦ | = 𝒖♭ = 1. The unitary group U𝑛 being connected there is a
smooth curve [0, 1] ∋ 𝑠 ↦→ 𝑇 (𝑠) ∈ U𝑛 such that 𝑇 (0) = 𝐼𝑛 and 𝑇 (1) 𝒖 ◦ = 𝒖♭ . We
have 𝑇 (𝑠) 𝒖 ◦ · 𝑇 (𝑠) 𝒗 ◦ = 1 for all 𝑠 ∈ [0, 1]: the curve

[0, 1] ∋ 𝑠 ↦→ (𝑇 (𝑠) 𝒖 ◦ , 𝑇 (𝑠) 𝒗 ◦ ) ∈ 𝑆 ◦𝑛

joins (𝒖 ◦ , 𝒗 ◦ ) to 𝒖♭ , 𝑇 (1) 𝒗 ◦ . We follow it up with the curve [0, 1] ∋ 𝑠 ↦→

𝒖♭ , 𝑠𝒗 ♭ + (1 − 𝑠) 𝒖♭ · 𝑇 (1) 𝒗 ◦ which joins 𝒖♭ , 𝑇 (1) 𝒗 ◦ to 𝒖♭ , 𝒗 ♭ . Since
384 13 Elements of Symplectic Geometry

∀𝑠 ∈ [0, 1] , 𝒖♭ · 𝑠𝒗 ♭ + (1 − 𝑠) 𝑇 (1) 𝒗 ◦ = 1,

we have thus connected (𝒖 ◦ , 𝒗 ◦ ) to 𝒖♭ , 𝒗 ♭ within 𝑆 ◦𝑛 . □

Proposition 13.1.22 The group Sp (𝑛, K) is connected.

Proof Let Σ𝑛 denote the set of all symplectic bases of K2𝑛 . There is a natural
bijection Σ𝑛 ∋ 𝐵 ↦→ g 𝐵 ∈ Sp (𝑛, K): g 𝐵 is the unique linear automorphism of K2𝑛
which transforms 𝐵 into the canonical basis {𝒆 1 ., ..., 𝒆 2𝑛 }. For every pair (𝒖, 𝒗) ∈
𝑆 𝑛 (see Lemma 13.1.21) let E𝒖,𝒗 be the vector subspace of K2𝑛 perpendicular
to both 𝒖 and 𝒗 in the sense of the dot product. The restriction of 𝜛𝑛 to E𝒖,𝒗 is a
symplectic form: indeed, it is skew-symmetric; it is nondegenerate since all elements
of E𝒖,𝒗 are 𝜛𝑛 -perpendicular to both 𭟋𝑛 𝒖 and 𭟋𝑛 𝒗. If we identify an arbitrary basis
{𝜺 1 ., ..., 𝜺 2𝑛−2 } ∈ Σ𝑛−1 with a symplectic basis of E𝒖,𝒗 , as we have the right to do,
the following map is a bijection

Σ𝑛−1 × 𝑆 𝑛 ∋ ({𝜺 1 ., ..., 𝜺 2𝑛−2 } , 𝒖, 𝒗) ↦→ {𝜺 1 ., ..., 𝜺 2𝑛−2 , 𭟋𝑛 𝒖, 𭟋𝑛 𝒗} ∈ Σ𝑛 .

This map can be equated to a bijection 𝜑 : Sp (𝑛 − 1, K) × 𝑆 𝑛 −→ Sp (𝑛, K). It is

immediately seen that it is a homeomorphism. (Actually it is not difficult to see that
𝜑 is an analytic diffeomorphism.) The induction hypothesis that Σ𝑛−1 is connected
combined with Lemma 13.1.21 entails the sought result. □

Corollary 13.1.23 If 𝒈 ∈ Sp (𝑛, K) then det 𝒈 = 1.

Proof By (13.1.15) we have (det 𝒈) 2 = det ( 𝒈 ⊤ 𭟋𝑛 𝒈) = 1, i.e., det 𝒈 = ±1, whence

the claim by Proposition 13.1.22. □

Remark 13.1.24 Let 𝒖 = (𝑢 1 , 𝑢 2 ), 𝒗 = (𝑣 1 , 𝑣 2 ) ∈ K2 be such that 𭟋𝑛 𝒖 · 𝒗 = 1,

i.e., 𝑢 1 𝑣 2 − 𝑢 2 𝑣 1 = 1. This shows that the special linear group SL (2, K) is equal to
Sp (1, K); note also that dim SL (2, K) = 3.

We shall denote by 𝔰𝔭 (𝑛, K) the Lie algebra of Sp (𝑛, K).

Proposition 13.1.25 For a matrix 𝐴 ∈ M2𝑛 (K) to belong to 𝔰𝔭 (𝑛, K) it is necessary

and sufficient that 𝐴⊤ 𭟋𝑛 + 𭟋𝑛 𝐴 = 0.

Proof If 𝒈 = exp 𝑡 𝐴 satisfies (13.1.15) for all 𝑡 ∈ K then

d
exp 𝑡 𝐴⊤ 𭟋𝑛 exp 𝑡 𝐴 = 𝐴⊤ 𭟋𝑛 + 𭟋𝑛 𝐴 = 0.

□
d𝑡 𝑡=0

Needless to say, when 𝑛 = 1 Propositions 13.1.18 and 13.1.25 are in agreement,

since
𝑎 𝑐 0 −1 0 −1 𝑎 𝑏 0 −𝑎 − 𝑑
+ = .
𝑏𝑑 1 0 1 0 𝑐𝑑 𝑎+𝑑 0
We are going to use the block decomposition of (2𝑛) × (2𝑛) matrices:
13.1 Elements of Symplectic Algebra 385

𝐴 𝐵
, 𝐴, 𝐵, 𝐶, 𝐷 ∈ M𝑛 (K) .
𝐶𝐷

Thus, by (13.1.10),
0 −𝐼𝑛
𭟋𝑛 = . (13.1.17)
𝐼𝑛 0

Lemma 13.1.26 Let M ⊂ GL (2𝑛, K) consist of the matrices

𝐴 0
, 𝐴 ∈ GL (𝑛, K) .
0 ( 𝐴⊤ ) −1

Let N ⊂ GL (2𝑛, K) consist of the matrices

𝐼𝑛 𝑋
0 𝐼𝑛

with 𝑋 ∈ M𝑛 (K) symmetric.

Then M and N are subgroups of Sp (𝑛, K) and every matrix 𝒈 ∈ Sp (𝑛, K) is a
product of finitely many elements of M, N and finitely many times 𭟋𝑛 .

Proof That M is a group follows directly from the fact that ( 𝐴⊤ ) −1 (𝐵⊤ ) −1 =
−1
( 𝐴𝐵) ⊤ . Moreover M ⊂ Sp (𝑛, K) since

𝐴⊤ 0

0 −𝐼𝑛 𝐴 0 0 −𝐼𝑛
−1 =
0 𝐴−1 𝐼𝑛 0 0 ( 𝐴⊤ ) 𝐼𝑛 0

[cf. (13.1.17)]. We have

𝐼𝑛 𝑋 𝐼𝑛 𝑌 𝐼𝑛 𝑋 + 𝑌
= ,
0 𝐼𝑛 0 𝐼𝑛 0 𝐼𝑛

𝐼𝑛 0 0 −𝐼𝑛 𝐼𝑛 𝑋 0 −𝐼𝑛
= ,
𝑋 𝐼𝑛 𝐼𝑛 0 0 𝐼𝑛 𝐼𝑛 0

which proves that N is an Abelian subgroup of Sp (𝑛, K).

Define N⊤ to be the subgroup of Sp (𝑛, K) obtained by transposing the matrices
that belong to N. Note that whatever 𝑋 ∈ M𝑛 (K),

𝐼𝑛 0 0 −𝐼𝑛 0 −𝐼𝑛 𝐼𝑛 −𝑋
= .
𝑋 𝐼𝑛 𝐼𝑛 0 𝐼𝑛 0 0 𝐼𝑛

It will suffice to show that every matrix

𝐴 𝐵
𝒈= ∈ Sp (𝑛, K)
𝐶𝐷

is equal to a product of finitely many matrices belonging to M, N or N⊤ . We have

386 13 Elements of Symplectic Geometry

𝐴⊤ 𝐶 ⊤

⊤ 0 −𝐼𝑛 𝐴 𝐵
𝒈 𭟋𝑛 𝒈 =
𝐵⊤ 𝐷 ⊤ 𝐼 𝑛 0 𝐶𝐷
⊤
𝐶 𝐴 − 𝐴 𝐶 𝐶 𝐵 − 𝐴⊤ 𝐷
⊤ ⊤

= = 𭟋𝑛 ,
𝐷 ⊤ 𝐴 − 𝐵⊤ 𝐶 𝐷 ⊤ 𝐵 − 𝐵⊤ 𝐷

whence, assuming that 𝐴 is nonsingular,

−1
𝐴⊤ 𝐶 ⊤ = 𝐶 𝐴−1 ,
𝐷 ⊤ 𝐵 = 𝐵⊤ 𝐷,
𝐴⊤ 𝐷 − 𝐶 ⊤ 𝐵 = 𝐷 ⊤ 𝐴 − 𝐵 ⊤ 𝐶 = 𝐼 𝑛 .

If 𝐴 is nonsingular then
−1 −1 −1
𝐷 = 𝐴⊤ − 𝐴⊤ 𝐶 ⊤ 𝐵 = 𝐴⊤ − 𝐶 𝐴−1 𝐵

and therefore

1 𝐴−1 𝐵

1 0 𝐴 0
𝐶 𝐴−1 1 0 ( 𝐴 ) −1
⊤ 0 1
1 𝐴−1 𝐵

1 0 𝐴 0
= = 𝒈.
𝐶 𝐴−1 1 0 ( 𝐴⊤ ) −1 0 1

Thus we have shown that a neighborhood of 𝐼2𝑛 in Sp (𝑛, K) consists of matrices

𝛽1⊤ 𝛼𝛽2 with 𝛼 ∈ M, 𝛽 𝑗 ∈ N, 𝑗 = 1, 2. We conclude that the subgroup H generated
by such products is open in Sp (𝑛, K). But any open subgroup of a topological group
is also closed. Since Sp (𝑛, K) is connected (Proposition 13.1.22) we must have
H = Sp (𝑛, K). □
Going now to the Lie algebra 𝔰𝔩 (𝑛, K) we isolate three Lie subalgebras:

0𝑋
(1) the (Abelian) Lie algebra 𝔫 of the subgroup N, consisting of the matrices
0 0
with 𝑋 ∈ 𝑀𝑛 (K) symmetric;
(2) the (Abelian) Lie algebra 𝔫 ⊤ of the subgroup N⊤ consisting of the transposes of
the matrices belonging to 𝔫;
𝐴 0
(3) the Lie algebra 𝔪 of the subgroup M, consisting of the matrices ,
0 −𝐴⊤
𝐴 ∈ 𝑀𝑛 (K).
The Lie algebra 𝔪 is non-Abelian if 𝑛 > 1. We denote by [𝔫, 𝔫 ⊤ ] the linear span
(over the field K) of the brackets [𝑁, 𝑁 ′] = 𝑁 𝑁 ′ − 𝑁 ′ 𝑁, 𝑁 ∈ 𝔫, 𝑁 ′ ∈ 𝔫 ⊤ .

Proposition 13.1.27 As a vector space 𝔰𝔩 (𝑛, K) is the direct sum 𝔫 ⊕ 𝔪 ⊕ 𝔫 ⊤ . As a

Lie algebra 𝔰𝔩 (𝑛, K) is generated by [𝔫, 𝔫 ⊤ ]. Moreover, 𝔪 ⊂ [𝔫, 𝔫 ⊤ ].

Proof The direct sum decomposition is an immediate consequence of Lemma

13.1.26. In order to prove the last two assertions we note first that
13.1 Elements of Symplectic Algebra 387

0𝑋 00 𝑋𝑌 0
, = .
0 0 𝑌 0 0 −𝑌 𝑋
Í
Every 𝑛 × 𝑛 matrix can be written as a sum 𝐴 = 𝑟𝑗=1 𝑋 𝑗 𝑌 𝑗 with 𝑋 𝑗 and 𝑌 𝑗 symmetric
matrices. This last claim ensues from the fact that every elementary matrix (i.e., a
matrix with one entry equal to 1 and all others equal to 0) is the product of two
symmetric matrices with entries equal to 0 or to 1. As a consequence,
𝑟
∑︁
𝐴 0 0 𝑋𝑗 0 0
= , ,
0 −𝐴⊤ 0 0 𝑌𝑗 0
𝑗=1

which proves that 𝔪 ⊂ [𝔫, 𝔫 ⊤ ]. A special case is

𝐼𝑛 0 0 𝐼𝑛 0 0
= , .
0 −𝐼𝑛 0 0 𝐼𝑛 0

With this established, the claim that the Lie algebra 𝔰𝔩 (𝑛, K) is generated by [𝔫, 𝔫 ⊤ ]
follows from the identities
1
0𝑋 𝐼𝑛 0 0 2𝑋
= , ;
0 0 0 −𝐼𝑛 0 0

0 0 0 0 𝐼𝑛 0
= 1 , . □
𝑋0 2𝑋 0
0 −𝐼𝑛

Corollary 13.1.28 The Lie algebra 𝔰𝔭 (𝑛, K) is equal to its commutator subalgebra
[𝔰𝔭 (𝑛, K) , 𝔰𝔭 (𝑛, K)].
Corollary 13.1.29 The group Sp (𝑛, K) is equal to the closure of its commutator
subgroup [the subgroup generated by the products 𝒈𝒈 ′ 𝒈 −1 𝒈 ′−1 , 𝒈, 𝒈 ′ ∈ Sp (𝑛, K) ]
as well as to the closure of the subgroup generated by the matrices n1 and 𭟋𝑛 n2 𭟋𝑛 =
−n⊤2 , n 𝑗 ∈ N, 𝑗 = 1, 2.
Proof By Proposition 13.1.27 and Corollary 13.1.28 the Lie algebras of the two
subgroups in the statement are equal to 𝔰𝔭 (𝑛, K). □
Corollary 13.1.30 We have dim Sp (𝑛, K) = 𝑛 (2𝑛 + 1).
Proof We have

dim 𝔰𝔭 (𝑛, K) = dim 𝔪 + 2 dim 𝔫 = dim GL (𝑛, K) + 2 dim Sym (𝑛, K) ,

where Sym (𝑛, K) is the linear space of symmetric 𝑛 × 𝑛 matrices with entries in the
field K, whose dimension is 21 𝑛 (𝑛 + 1), whence the result. □
Remark 13.1.31 Corollary 13.1.30 could also have been deduced from the proof of
Proposition 13.1.22. The result is true for 𝑛 = 1 as noted in Remark 13.1.24. The
dimension of the set 𝑆 𝑛 in the proof of Proposition 13.1.22 is equal to 4𝑛 − 1. It
follows that dim Sp (𝑛, K) = dim Sp (𝑛 − 1, K) + 4𝑛 − 1. Then induction on 𝑛 yields
the result.
388 13 Elements of Symplectic Geometry

Proposition 13.1.32 The center of Sp (𝑛, K) consists of the two elements 𝐼2𝑛 ,−𝐼2𝑛 .
The center of 𝔰𝔭 (𝑛, K) consists of the single element 0.

Proof The matrices in the center of Sp (𝑛, K) must commute with every matrix in the
subgroups M and N(Lemma 13.1.26). The only (2𝑛) × (2𝑛) matrices
that commute
𝐼𝑛 𝑋 𝐸 𝐹
with every matrix ∈ N are those of the form , 𝐸, 𝐹 ∈ M𝑛 (K).
0 𝐼𝑛 0 𝐸
For such a matrix to commute with every matrix belonging to the subgroup M it is
necessary that 𝐸 𝐴 = 𝐴𝐸 for every 𝐴 ∈ GL (𝑛, K), obviously implying 𝐸 = 𝜆𝐼𝑛 , and
𝐹 = 𝐴𝐹 𝐴⊤ , obviously
implying 𝐹 = 0 (just take 𝐴 = 2𝐼𝑛 ). Lastly we observe that
𝜆𝐼𝑛 0
a matrix belongs to Sp (𝑛, K) if and only if 𝜆2 = 1. Since the center of
0 𝜆𝐼𝑛
Sp (𝑛, K) is a finite set, that of its Lie algebra is equal to {0}. □
The properties above reflect the fact that 𝔰𝔭 (𝑛, K) is a semisimple Lie algebra;
Sp (𝑛, K) is a semisimple Lie group.

13.2 The Metaplectic Group

13.2.1 Some “elementary” unitary transformations in R𝒏

In this section we give a brief description of the metaplectic group Mp (𝑛, R). It is a
two-fold covering of the symplectic group Sp (𝑛, R) such that the “base projection”
𝜋 : Mp (𝑛, R) −→ Sp (𝑛, R) is a local analytic isomorphism, as well as a Lie
group homomorphism. The metaplectic group is not a linear group (see Subsection
13.1.4). It can be defined abstractly by a universal property among covering groups
of Sp (𝑛, R). Here, following most authors we identify Mp (𝑛, R) with its original
representation, as a finite-dimensional subgroup of U 𝐿 2 (R𝑛 ) , the group of unitary
transformations of 𝐿 2 (R𝑛 ), i.e., of linear isometries of 𝐿 2 (R𝑛 ) onto itself. Keep
in mind that the elements of 𝐿 2 (R𝑛 ) are (equivalence classes of) complex-valued
square-integrable functions (almost equal everywhere).
∗ 2
We denote∫ by 𝑇 the adjoint of a bounded linear operator 𝑇 of 𝐿 (R ) into itself: if
𝑛
2 ∗
( 𝑓 , 𝑔) 𝐿 2 = 𝑓 𝑔d𝑥 is the Hermitian product in 𝐿 (R ) then (𝑇 𝑓 , 𝑔) 𝐿 2 = ( 𝑓 , 𝑇 𝑔) 𝐿 2 .
𝑛

The operator 𝑇 is unitary if it is invertible and 𝑇 −1 = 𝑇 ∗ .

One of the most remarkable members of U 𝐿 2 (R𝑛 ) is the Fourier transfor-
mation. One must however modify our initial definition (1.3.3) to achieve that the
adjoint and the inverse of the Fourier transformation coincide. Thus we define, for
arbitrary 𝑢 ∈ 𝐿 2 (R𝑛 ),
∫
1 1
𝔉𝑢 (𝜉) = (2𝜋) − 2 𝑛 b
𝑢 (𝜉) = (2𝜋) − 2 𝑛 e−𝑖 𝑥· 𝜉 𝑢 (𝑥) d𝑥, (13.2.1)
R𝑛

whose inverse is given by

13.2 The Metaplectic Group 389
∫
1
𝔉−1 𝑢 (𝑥) = (2𝜋) − 2 𝑛 e𝑖 𝑥· 𝜉 𝔉𝑢 (𝜉) d𝜉 = 𝔉∗ 𝑢 (𝑥) . (13.2.2)
R𝑛

That 𝔉 ∈ U 𝐿 2 (R𝑛 ) is the content of the Plancherel Theorem.

There is a “natural” injection GL (𝑛, R) ∋ 𝐴 ↦→ 𝜒 ( 𝐴) ∈ U 𝐿 2 (R𝑛 ) , reflecting

1
the effect on 2 -densities of a linear change of variables in R𝑛 :
1
( 𝜒 ( 𝐴) 𝑓 ) (𝑥) = |det 𝐴| 2 𝑓 ( 𝐴𝑥) . (13.2.3)

Other basic elements of U 𝐿 2 (R𝑛 ) are the translations 𝑓 (𝑦) ↦→ 𝜏𝑎 𝑓 (𝑦) =

𝑓 (𝑦 − 𝑎), 𝑎 ∈ R𝑛 , and multiplication by any exp (𝑖𝜑 (𝑥)) with a (measurable) phase-
function 𝜑 : R𝑛 −→ R. In particular, the transformation

1
𝜇 ( 𝐴) : 𝑓 (𝑥) ↦→ 𝑓 (𝑥) exp 𝑖 𝐴𝑥 · 𝑥 (13.2.4)
2

is unitary provided the matrix 𝐴 ∈ M𝑛 (R) is symmetric.

13.2.2 Irreducible representations of the Heisenberg group

Consider the map 𝜌 : R2𝑛 −→ U 𝐿 2 (R𝑛 ) defined by

1
∀ 𝑓 ∈ 𝐿 2 (R𝑛 ) , (𝜌 (𝑥, 𝜉) 𝑓 ) (𝑦) = e−𝑖 𝜉 · ( 𝑦+ 2 𝑥 ) 𝑓 (𝑥 + 𝑦) . (13.2.5)

convenient to use vector notation, writing 𝜌 (u)

In the present context it is often
𝑥
rather than 𝜌 (𝑥, 𝜉) if u = . It is directly checked that the adjoint (which is also
𝜉
the inverse) of 𝜌 (u) is 𝜌 (−u).

Proposition 13.2.1 If a bounded linear operator 𝐴 in 𝐿 2 (R𝑛 ) commutes with the

unitary operator 𝜌 (u) for every u ∈ R2𝑛 then 𝐴 𝑓 = 𝜆 𝑓 for some 𝜆 ∈ C and all
𝑓 ∈ 𝐿 2 (R𝑛 ).

Proof By hypothesis 𝐴 commutes with multiplication of the functions belonging to

𝐿 2 (R𝑛 ) by exp (𝑖𝜉 · 𝑦), i.e., with 𝜌 (0, −𝜉), whatever 𝜉 ∈ R𝑛 . Let 𝑓 , 𝑔 ∈ 𝐿 2 (R𝑛 ) be
arbitrary; we have
∫
− 12 𝑛
𝑓 (𝑦) 𝑔 (𝑦) = (2𝜋) e𝑖 𝜉 ·𝑦 𝑓 (𝑦) 𝔉𝑔 (𝜉) d𝜉,

whence
𝐴 ( 𝑓 𝑔) = 𝑔 𝐴 𝑓
and therefore, for arbitrary 𝑓 ∈ 𝐿 2 (R𝑛 ),
2
2
2

𝐴 e−|𝑦 | 𝑓 (𝑦) = e− |𝑦 | 𝐴 𝑓 (𝑦) = 𝑓 (𝑦) 𝐴 e− |𝑦 | .
390 13 Elements of Symplectic Geometry
2
2

Thus 𝐴 𝑓 = ℎ 𝑓 with ℎ = e |𝑦 | 𝐴 e− |𝑦 | [in passing note that since 𝐴 is a bounded
linear operator we must have ℎ ∈ 𝐿 ∞ (R𝑛 )]. Our hypothesis also implies that 𝐴
commutes with all the translations 𝜏𝑥 = 𝜌 (𝑥, 0): 𝜏𝑥 𝐴 𝑓 = 𝐴𝜏𝑥 𝑓 = ℎ𝜏𝑥 𝑓 . But
𝜏𝑥 𝐴 𝑓 = (𝜏𝑥 ℎ) 𝜏𝑥 𝑓 , implying 𝜏𝑥 ℎ = ℎ for all 𝑥 ∈ R𝑛 , i.e., ℎ ≡ 𝜆, a constant, i.e.,
𝐴𝑓 = 𝜆𝑓. □

Corollary 13.2.2 Let E be a linear subspace of 𝐿 2 (R𝑛 ) such that 𝜌 (u) E ⊂ E for
all u ∈ R𝑛 . Then either E = {0} or E = 𝐿 2 (R𝑛 ).

Proof We have ( 𝑓 , 𝜌 (u) 𝑔) 𝐿 2 = (𝜌 (−u) 𝑓 , 𝑔) 𝐿 2 which shows that, if E⊥ denotes

the subspace of 𝐿 2 (R𝑛 ) orthogonal to E, 𝑔 ∈ E⊥ implies 𝜌 (u) 𝑔 ∈ E⊥ . It follows
that 𝜌 (u) induces an isometry of E (resp., E⊤ ) onto itself. Let then 𝑃E denote the
orthogonal projection of 𝐿 2 (R𝑛 ) onto E. For each 𝑓 ∈ 𝐿 2 (R𝑛 ) we have 𝜌 (u) 𝑓 =
𝜌 (u) 𝑃E 𝑓 + 𝜌 (u) (𝐼 𝐿 2 − 𝑃E ) 𝑓 , whence 𝑃E 𝜌 (u) 𝑓 = 𝜌 (u) 𝑃E 𝑓 . As u ∈ R𝑛 is
arbitrary, Proposition 13.2.1 implies 𝑃E = 𝜆𝐼 𝐿 2 hence 𝜆 = 0 or 1 since 𝑃E is an
orthogonal projection. □
We relate the transformations 𝜌 (𝑥, 𝜉) ∈ U 𝐿 2 (R𝑛 ) to the unitary transforma-

tions 𝔉, 𝜒 ( 𝐴), 𝜇 ( 𝐴) [𝐴 ∈ M𝑛 (R); see (13.2.3), (13.2.4)].

Proposition 13.2.3 For all (𝑥, 𝜉) ∈ R2𝑛 the following hold:

(1) 𝔉−1 𝜌 (𝑥, 𝜉) 𝔉 = 𝜌 (𝜉, −𝑥);
(2) if 𝐴 ∈ GL (𝑛, R) then 𝜌 (𝑥, 𝜉) 𝜒 ( 𝐴) 𝑓 = 𝜒 ( 𝐴) 𝜌 𝐴−1 𝑥, 𝐴⊤ 𝜉 𝑓 ;

(3) if 𝐴 ∈ M𝑛 (R) is a symmetric matrix then

𝜌 (𝑥, 𝜉) 𝜇 ( 𝐴) 𝑓 = 𝜇 ( 𝐴) 𝜌 (𝑥, 𝜉 − 𝐴𝑥) 𝑓 .

Proof Proof of (1): By (13.2.2) and (13.2.5) we have, for 𝑔 ∈ 𝐿 2 (R𝑛 ),

∫
− 21 𝑛 −𝑖 𝜉 · ( 𝑦+ 21 𝑥 )
−1
𝜌 (𝑥, 𝜉) 𝔉 𝑔 = (2𝜋) e e𝑖 𝜂· ( 𝑥+𝑦) 𝑔 (𝜂) d𝜂

whence, by (13.2.1),
∫ ∫

−𝑖𝑦· 𝜂 ′ −𝑖 𝜉 · ( 𝑦+ 21 𝑥 ) d𝜂d𝑦
−1
𝔉 𝜌 (𝑥, 𝜉) 𝔉 𝑔 (𝜂 ) = ′
e e𝑖 𝜂· ( 𝑥+𝑦) 𝑔 (𝜂)
(2𝜋) 𝑛
∫ ∫
′ 1 d𝜂d𝑦
= e𝑖 ( 𝑥−𝑦) · 𝜂 −𝑖 𝜉 · ( 𝑦− 2 𝑥 ) e𝑖 𝜂·𝑦 𝑔 (𝜂)
(2𝜋) 𝑛
∫ ∫
′ 1 ′ d𝜂d𝑦
= e𝑖 𝑥· ( 𝜂 + 2 𝜉 ) e−𝑖 ( 𝜂 + 𝜉 ) ·𝑦 e𝑖𝑦· 𝜂 𝑔 (𝜂)
(2𝜋) 𝑛
′ 1
= e𝑖 𝑥· ( 𝜂 + 2 𝜉 ) 𝑔 (𝜂 ′ + 𝜉) .

Proof of (2): We have, by (13.2.3),

13.2 The Metaplectic Group 391
1

𝜒 𝐴−1 𝜌 𝐴−1 𝑥, 𝐴⊤ 𝜉 𝑓 (𝑦) = 𝜒 𝐴−1 e−𝑖 𝜉 · ( 𝐴𝑦+ 2 𝑥 ) 𝑓 𝐴−1 𝑥 + 𝑦
1 1

= |det 𝐴| − 2 e−𝑖 𝜉 · ( 𝑦+ 2 𝑥 ) 𝑓 𝐴−1 (𝑥 + 𝑦)

= 𝜌 (𝑥, 𝜉) 𝜒 𝐴−1 𝑓 (𝑦) .

Proof of (3): We have, by (13.2.4),

1

𝜇 ( 𝐴) −1 𝜌 (𝑥, 𝜉) 𝑓 (𝑦) = 𝜇 (−𝐴) e−𝑖 𝜉 · ( 𝑦+ 2 𝑥 ) 𝑓 (𝑥 + 𝑦)
1 1
= e−𝑖 2 𝐴𝑦·𝑦 e−𝑖 𝜉 · ( 𝑦+ 2 𝑥 ) 𝑓 (𝑥 + 𝑦)
1 1
= e−𝑖 ( 𝜉 −𝐴𝑥) · ( 𝑦+ 2 𝑥 ) e−𝑖 2 𝐴( 𝑥+𝑦) · ( 𝑥+𝑦) 𝑓 (𝑥 + 𝑦)
= 𝜌 (𝑥, 𝜉 − 𝐴𝑥) 𝜇 (−𝐴) 𝑓 (𝑦) . □
′
𝑥 ′ 𝑥
If u = and u = ′ belong to R2𝑛 we have
𝜉 𝜉
1
′ 1 ′

𝜌 (u) 𝜌 (u′) 𝑓 = e−𝑖 𝜉 · ( 𝑦+ 2 𝑥 ) e−𝑖 𝜉 · ( 𝑦+𝑥+ 2 𝑥 ) 𝑓 (𝑦 + 𝑥 + 𝑥 ′)
1 ′ −𝑥 ′ · 𝜉 ) ′ 1 ′
= e− 2 𝑖 ( 𝑥· 𝜉 e−𝑖 ( 𝜉 + 𝜉 ) · ( 𝑦+ 2 ( 𝑥+𝑥 ) ) 𝑓 (𝑦 + 𝑥 + 𝑥 ′) ,

whence
1
𝜌 (u) 𝜌 (u ) = exp − 𝑖𝜛𝑛 (u, u ) 𝜌 (u + u′) .
′ ′
(13.2.6)
2
Formula (13.2.6) intimates that 𝜌 is “part of” a representation of the Heisenberg
group H𝑛 . The group law in H𝑛 is defined by the formula

1
(u, 𝑡) · (u′, 𝑡 ′) = u + u′, 𝑡 + 𝑡 ′ + 𝜛𝑛 (u, u′) , (13.2.7)
2

(u, 𝑡), (u′, 𝑡 ′) denoting elements of H𝑛 ( R2𝑛 × R; cf. Example 12.5.17). If then we
define 𝜌 ♮ (u, 𝑡) = e−𝑖𝑡 𝜌 (u) we can rewrite (13.2.6) as

𝜌 ♮ (u, 𝑡) 𝜌 ♮ (u′, 𝑡 ′) = 𝜌 ♮ ((u, 𝑡) · (u′, 𝑡 ′)) . (13.2.8)

In other words, 𝜌 ♮ is a unitary representation of H𝑛 . Then Corollary 13.2.2 can be

rephrased by saying that the representation 𝜌 ♮ of H𝑛 is irreducible.
Using 𝜌 ♮ we can form a simple picture of the set 𝔖 of all the equivalence classes
(or cosets) of irreducible unitary representations of H𝑛 . Two unitary representations
𝜌1 and 𝜌2 respectively acting on (complex) Hilbert spaces E1 and E2 (the repre-
sentation spaces) are equivalent if there is an isometry g : E1 −→ E2 such that
g𝜌1 = 𝜌2 g. We can label each class in 𝔖 by an element of R2𝑛 ∪ (R\ {0}) whose
selection is determined by the representation of the elements (0, 𝑡) ∈ H𝑛 , necessarily
multiplication by a number e𝑖𝜆𝑡 , 𝜆 ∈ R.
392 13 Elements of Symplectic Geometry

(1) For 𝜆 = 0 we get the one-dimensional representations: the representation space

is C with its standard Hermitian product. Each one-dimensional representation
is equivalent to a representation H𝑛 ∋ (u, 𝑡) ↦→ 𝜌◦ (u, 𝑡) ∈ U (1) defined by
∗
𝜁 ↦→ 𝜌◦ (u, 𝑡) 𝜁 = e𝑖u ·u 𝜁 for some u∗ ∈ R2𝑛 , the “label” of the representation
coset.
(2) The number 𝜆 ∈ R\ {0} labels the equivalence class of the unitary representa-

2 −𝑖𝜆𝑡 𝑥
tions acting on 𝐿 (R ), 𝜌 (𝜆 · u, 𝜆𝑡) = e
𝑛 ♮ 𝜌 (𝜆 · u) where 𝜆 · u = if
𝜆𝜉

𝑥
u= . In other words,
𝜉
1
𝜌 ♮ (𝜆 · u, 𝜆𝑡) 𝑓 (𝑦) = e−𝑖𝜆 ( 𝑡+ 𝜉 · ( 𝑦+ 2 𝑥) ) 𝑓 (𝑥 + 𝑦) , 𝑓 ∈ 𝐿 2 (R𝑛 ) .

For a proof see [Wallach, 1977], pp. 179–183.

13.2.3 The metaplectic group

We continue to make use of the map 𝜌 : R2𝑛 −→ U 𝐿 2 (R𝑛 ) defined in (13.2.5).

Proposition 13.2.4 Suppose that the operator 𝑇 ∈ U 𝐿 2 (R𝑛 ) has the following

property: to each u ∈ R2𝑛 there is a v ∈ R2𝑛 such that 𝑇 𝜌 (u) 𝑇 −1 = 𝜌 (v). Then the
assignment u ↦→ v is a symplectic transformation of R2𝑛 .

1
𝜌 (u) 𝜌 (u′) = exp − 𝑖𝜛𝑛 (u, u′) 𝜌 (u + u′) .
2

Proof The map 𝜌 obviously being injective the vector v corresponding to a given u ∈
R2𝑛 is unique. If u, u′ ∈ R2𝑛 the relation (13.2.6) holds. Suppose that 𝑇 𝜌 (u) 𝑇 −1 =
𝜌 (v) and 𝑇 𝜌 (u′) 𝑇 −1 = 𝜌 (v′). Then

1
exp − 𝑖𝜛𝑛 (v, v ) 𝜌 (v + v′) = 𝜌 (v) 𝜌 (v′)
′
2
= 𝑇 𝜌 (u) 𝜌 (u′) 𝑇 −1

1
= exp − 𝑖𝜛𝑛 (u, u′) 𝑇 𝜌 (u + u′) 𝑇 −1 .
2

But there is a unique w ∈ R2𝑛 such that 𝑇 𝜌 (u + u′) 𝑇 −1 = 𝜌 (w), which demands
1
𝜌 (w) = exp 𝑖 (𝜛𝑛 (u, u′) − 𝜛𝑛 (v, v′)) 𝜌 (v + v′) .
2
13.2 The Metaplectic Group 393

The definition of 𝜌 makes it evident that we must have w = v + v′ and 𝜛𝑛 (u, u′) =
𝜛𝑛 (v, v′). From the first one of these two equations one derives in standard fashion
that the map u ↦→ v is linear, and from the second one that it is a symplectic
automorphism of R2𝑛 . □
Notation 13.2.5 We shall denote by G◦ the group of transformations 𝑇 ∈ U 𝐿 2 (R𝑛 )

such that there is a transformation g ∈ Sp (𝑛, R) satisfying

∀u ∈ R2𝑛 , 𝑇 −1 𝜌 (u) 𝑇 = 𝜌 (gu) . (13.2.9)

That the set G◦ is a group with respect to operator composition is self-evident.

The (unique) element g ∈ Sp (𝑛, R) such that (13.2.9) holds for a given 𝑇 ∈ G◦ will
be denoted by 𝜋◦ (𝑇); 𝜋◦ is a group homomorphism G◦ −→ Sp (𝑛, R). We shall
denote by 𝐼 𝐿 2 the identity transformation of 𝐿 2 (R𝑛 ).
−1
Proposition 13.2.6 The kernel 𝜋◦ (𝐼2𝑛 ) of the group homomorphism 𝜋◦ : G◦ −→
Sp (𝑛, R) is the set of transformations e𝑖 𝜃 𝐼 𝐿 2 , 𝜃 ∈ R.
Proof Proposition 13.2.1 implies that ker 𝜋◦ consists of transformations 𝐿 2 (R𝑛 ) ∋
𝑓 ↦→ 𝜆 𝑓 , 0 ≠ 𝜆 ∈ C; since G◦ ⊂ U 𝐿 2 (R𝑛 ) we must have |𝜆| = 1. □
We assume that U 𝐿 2 (R𝑛 ) and G◦ are equipped with the strong convergence

topology: a sequence 𝑇 𝑗 𝑗=1,2,... of unitary transformations converges to 𝑇 if

lim 𝑇 𝑗 𝑓 − 𝑇 𝑓 𝐿2
=0
𝑗→+∞

for each 𝑓 ∈ 𝐿 2 (R𝑛 ).

Proposition 13.2.7 The group homomorphism 𝜋◦ : G◦ −→ Sp (𝑛, R) is continuous.
Proof It suffices to prove that if 𝑇 𝑗 𝑓 − 𝑓 𝐿 2 → 0 whatever 𝑓 ∈ 𝐿 2 (R𝑛 ) (𝑇 𝑗 ∈ G◦ )
then the matrix norms 𝜋◦ 𝑇 𝑗 − 𝐼2𝑛 tend to zero. We have, whatever u ∈ 𝐿 2 (R𝑛 ),

𝜌 𝜋◦ 𝑇 𝑗 u 𝑓 − 𝜌 (u) 𝑓 𝐿 2 ≤ 𝜌 𝜋◦ 𝑇 𝑗 u 𝑓 − 𝑇 𝑗 𝜌 (u) 𝑓 𝐿 2

+ 𝑇 𝑗 − 𝐼 𝐿 2 𝜌 (u) 𝑓 𝐿 2

≤ 𝜌 (u) 𝑇 𝑗−1 𝑓 − 𝑓

2
+ 𝑇 𝑗 − 𝐼 𝐿 2 𝜌 (u) 𝑓 𝐿 2
𝐿

= 𝑇 𝑗 − 𝐼 𝐿 2 𝑓 𝐿 2 + 𝑇 𝑗 − 𝐼 𝐿 2 𝜌 (u) 𝑓 𝐿 2

whence

𝜌 𝜋◦ 𝑇 𝑗 − 𝐼2𝑛 u 𝑓 − 𝑓 𝐿2
= 𝜌 𝜋◦ 𝑇 𝑗 u 𝑓 − 𝜌 (u) 𝑓 𝐿2
→ 0.

𝑥
Let us write 𝜋◦ 𝑇 𝑗 − 𝐼2𝑛 u = 𝑗 . We conclude that, for each 𝑓 ∈ 𝐿 2 (R𝑛 ),

𝜉𝑗
∫ 2
1
e−𝑖 𝜉 𝑗 · ( 𝑦+ 2 𝑥 𝑗 ) 𝑓 𝑥 𝑗 + 𝑦 − 𝑓 (𝑦) d𝑦 → 0.

(13.2.10)
394 13 Elements of Symplectic Geometry

The proof that this implies lim 𝑥 𝑗 , 𝜉 𝑗 = 0 is left as an exercise [Hint: apply
𝑗→+∞
1 2
(13.2.10) with 𝑓 (𝑦) = e − 2 |𝑦 | ]. □
Proposition 13.2.3 implies that the Fourier transformation 𝔉 and the operators
𝜒 ( 𝐴), 𝐴 ∈ GL (𝑛, R) [see (13.2.3)], and 𝜇 (𝑋), 𝑋 ∈ M𝑛 (R) symmetric [see
(13.2.4)], belong to G◦ . We have

𝜋◦ (𝔉) = 𭟋𝑛 , (13.2.11)
𝐴 0

𝜋◦ 𝜒 𝐴−1 = , (13.2.12)
0 ( 𝐴⊤ ) −1

𝐼𝑛 0
𝜋◦ (𝜇 (𝑋)) = . (13.2.13)
−𝑋 𝐼𝑛

Definition 13.2.8 The closure of the commutator subgroup of G◦ (the subgroup

generated by the products 𝑆𝑇 𝑆 −1𝑇 −1 , 𝑆, 𝑇 ∈ G◦ ) is called the metaplectic group
and shall be denoted by Mp (𝑛, R).
The commutator subgroup of a group is a normal subgroup; as a consequence
Mp (𝑛, R) is a normal subgroup of G◦ .
Proposition 13.2.9 The restriction 𝜋 : Mp (𝑛, R) −→ Sp (𝑛, R) of the group homo-
morphism 𝜋◦ : G◦ −→ Sp (𝑛, R) is surjective.
Proof Let 𝑋 ∈ M𝑛 (R) be symmetric and 𝜇 (𝑋) be defined in (13.2.4). Direct
computation shows that
√ √ −1
𝜇 (𝑋) = 𝜒 2𝐼𝑛 𝜇 (𝑋) 𝜒 2𝐼𝑛 𝜇 (𝑋) −1 ∈ Mp (𝑛, R) .

Since Mp (𝑛, R) is a normal subgroup of G◦ it also contains 𝔉𝜇 (𝑋) 𝔉−1 . It suffices

then to apply Corollary 13.1.29, taking into account (13.2.11) and (13.2.13). □
Corollary 13.2.10 The group homomorphism 𝜋◦ : G◦ −→ Sp (𝑛, R) is surjective.
Proposition 13.2.11 The map 𝜋◦ defines a Lie group structure on G◦ . The homo-
morphism 𝜋◦ is analytic (and therefore it is a “morphism” of Lie groups). The Lie
group G◦ is connected.
Proof Propositions 13.2.6 and 13.2.7 imply the exactness of the sequence
𝜄 𝜋◦
𝐼2𝑛 −→ S1 −→ G◦ −→ Sp (𝑛, R) −→ 𝐼2𝑛 , (13.2.14)

where the (injective) map 𝜄 transforms the number e𝑖 𝜃 ∈ S1 (unit circle) into the
unitary operator e𝑖 𝜃 𝐼 𝐿 2 . Since both S1 and Sp (𝑛, R) are Lie groups the exact sequence
above endows G◦ with the structure of a Lie group, by the general theory. The general
theory also states that any continuous homomorphism of a Lie group onto another is
analytic. The exact sequence (13.2.14) implies that G◦ is an (analytic) circle bundle
over the connected Lie group Sp (𝑛, R). □
13.2 The Metaplectic Group 395

The commutator subgroup of a connected Lie group is a connected subgroup (as

easily verified; cf. [Chevalley, 1946], pp. 125–127): it follows that Mp (𝑛, R) is a
closed, connected normal subgroup of G◦ . We equip Mp (𝑛, R) with the Lie group
structure inherited from G◦ . Its Lie algebra 𝔪𝔭 (𝑛, R) is equal to its commutator
subalgebra [𝔪𝔭 (𝑛, R) , 𝔪𝔭 (𝑛, R)] (the linear span of all commutators [𝑋, 𝑌 ]).
Proposition 13.2.12 The homomorphism 𝜋 : Mp (𝑛, R) −→ Sp (𝑛, R) is a local
analytic diffeomorphism.
Proof The Lie algebra 𝔪𝔭 (𝑛, R) of Mp (𝑛, R) is a proper vector subspace of the Lie
algebra 𝔤◦ of G◦ . Indeed, G◦ has a center isomorphic to S1 and therefore 𝔤◦ has a one-
dimensional center 𝔷 transverse to [𝔤◦ , 𝔤◦ ] [= 𝔪𝔭 (𝑛, R)]. It follows from (13.2.14)
that dim 𝔤◦ = 1 + dim 𝔰𝔭 (𝑛, R) > dim 𝔪𝔭 (𝑛, R). Since 𝜋∗ : 𝔪𝔭 (𝑛, R) −→ 𝔰𝔭 (𝑛, R)
is surjective we conclude that it is bijective, whence the claim. □
A rephrasing of Proposition 13.2.12 is that 𝜋 is a covering map. Proposition
−1 −1
13.2.12 implies that the kernel 𝜋 (𝐼2𝑛 ) is discrete; 𝜋 (𝐼2𝑛 ) is a closed subgroup
of the kernel of the map 𝜋◦ : G◦ −→ Sp (𝑛, R); the latter kernel is isomorphic
−1
to S1 according to Proposition 13.2.11, therefore 𝜋 (𝐼2𝑛 ) must be finite. Since
−1
𝜋◦ (−𝑇) = 𝜋◦ (𝑇) for any 𝑇 ∈ G◦ the number of elements of 𝜋 (𝐼2𝑛 ) must be
even. Actually, the kernel of the group homomorphism 𝜋 : Mp (𝑛, R) −→ Sp (𝑛, R)
consists of two elements, 𝐼 𝐿 2 and −𝐼 𝐿 2 . In other words, Mp (𝑛, R) is a two-fold
covering group of Sp (𝑛, R). The proof of this claim is rather involved and will not
be given here (see [Wallach, 1977], pp. 218–225).

13.2.4 The differential of the metaplectic representation and Grushin’s

operators

Let Ω◦ denote a suitably small neighborhood of 𝐼2𝑛 in Sp (𝑛, R) and let Ω ⊂

−1
Mp (𝑛, R) ⊂ U 𝐿 2 (R) be the component of the identity in 𝜋 (Ω◦ ) (𝜋 as in Propo-

sition 13.2.7). The restriction of 𝜋 to Ω is an analytic diffeomorphism; its inverse
−1
𝜋 can be viewed as a “local representation”. Let us view it as a map of Ω◦ into
the space 𝔏 S (R𝑛 ) ; 𝐿 2 (R𝑛 ) of continuous linear operators of the Schwartz space
S (R𝑛 ) into 𝐿 2 (R𝑛 ). This operator space can be topologized in a natural manner;
we shall not insist on this aspect here. The topology of 𝔏 S (R𝑛 ) ; 𝐿 2 (R𝑛 ) renders
−1 −1
the map 𝜋 : Ω◦ −→ 𝔏 S (R𝑛 ) ; 𝐿 2 (R𝑛 ) continuous, which is evident since 𝜋

is a continuous map of Ω◦ into the space of bounded linear operators on 𝐿 2 (R𝑛 )
equipped with the strong convergence topology. But actually, if viewed as valued
−1
in 𝔏 S (R𝑛 ) ; 𝐿 2 (R𝑛 ) the map 𝜋 is differentiable, in the following sense: given a

differentiable map (−𝜏, 𝜏) ∋ 𝑡 ↦→ g (𝑡) ∈ Ω◦ (𝜏 > 0) and any function 𝜑 ∈ S (R )
𝑛
−1
the composite (−𝛿, 𝛿) ∋ 𝑡 ↦→ 𝜋 g (𝑡) 𝜑 ∈ 𝐿 2 (R𝑛 ) is differentiable. If then we take
g (𝑡) = exp 𝑡 𝑋 with 𝑋 ∈ 𝔰𝔭 (𝑛, R) we obtain
396 13 Elements of Symplectic Geometry
−1
d −1
D 𝜋 𝑋𝜑 = 𝜋 exp 𝑡 𝑋 𝜑 ; (13.2.15)
d𝑡 𝑡=0
−1
𝑋 ↦→ D 𝜋 𝑋 is a linear map 𝔰𝔭 (𝑛, R) −→ 𝔏 S (R𝑛 ) ; 𝐿 2 (R𝑛 ) , to which it is

natural to refer as the differential at the origin

of the metaplectic representation. As
−1
we shall now see, in general the operators D 𝜋 𝑋 cannot be extended as bounded
linear operators on 𝐿 2 (R𝑛 ). It can be shown, however, that they are continuous linear
operators of S (R𝑛 ) into itself.
According to (13.2.13), for 𝐴 = 𝑎 𝑖 𝑗 1≤𝑖, 𝑗 ≤𝑛 ∈ M𝑛 (R) symmetric and for |𝑡|
small we have
−1 0 0 −1 𝐼𝑛 0
𝜋 exp 𝑡 ≈ 𝜋 = 𝜇 (𝑡 𝐴)
−𝐴 0 −𝑡 𝐴 𝐼𝑛
√
recalling (13.2.4), i.e., 𝜇 (𝑡 𝐴) : 𝑓 (𝑥) ↦→ 𝑓 (𝑥) exp 21 −1𝑡 𝐴𝑥 · 𝑥 , 𝑓 ∈ 𝐿 2 (R𝑛 ). We
have, for any 𝜑 ∈ S (R𝑛 ),

1√
0 0 𝑛
−1
∑︁
D𝜋 𝜑 (𝑥) = −1 𝑎 𝑗,𝑘 𝑥 𝑗 𝑥 𝑘 𝜑 (𝑥) . (13.2.16)
−𝐴 0 2 𝑗,𝑘=1

Next we recall that

0𝐴 0 0
= 𭟋−1 𭟋 ,
00 𝑛 −𝐴 0 𝑛
whence
0𝐴 0 0
exp 𝑡 = 𭟋−1 exp 𭟋
00 𝑛 −𝐴 0 𝑛
and therefore 0 𝐴 −1 0 0
−1 −1
D𝜋 =𝔉 D𝜋 𝔉.
00 −𝐴 0
For any 𝜑 ∈ S (R𝑛 ),
0 𝐴
−1
−1 0 0
−1
D𝜋 𝜑 (𝑥) = 𝔉 D𝜋 𝔉𝜑 (𝜉)
00 −𝐴 0
1√
= −1𝔉−1 (( 𝐴𝜉 · 𝜉) 𝔉𝜑 (𝜉)) ,
2
whence
𝑛
−1
0 𝐴 1 ∑︁ 𝜕2 𝜑
D𝜋 𝜑 (𝑥) = − 𝑎 𝑗,𝑘 (𝑥) . (13.2.17)
00 2𝑖 𝜕𝑥 𝑗 𝜕𝑥 𝑘
𝑗,𝑘=1

Last, according to (13.2.12) for 𝑀 = 𝑐 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 ∈ M𝑛 (R) and |𝑡| small we
have
13.2 The Metaplectic Group 397

−1 𝐼𝑛 + 𝑡 𝑀 0
𝜋 = 𝜒 (𝐼𝑛 + 𝑡 𝑀) ,
0 (𝐼𝑛 + 𝑡 𝑀 ⊤ ) −1
1
recalling (13.2.3), i.e., ( 𝜒 ( 𝒈) 𝑓 ) (𝑥) = |det 𝒈| 2 𝑓 ( 𝒈𝑥), 𝒈 ∈ GL (𝑛, R), 𝑓 ∈ 𝐿 2 (R𝑛 ).
We have, for any 𝜑 ∈ S (R𝑛 ),
1 1

lim |det (𝐼𝑛 + 𝑡 𝑀)| 2 𝜑 ((𝐼𝑛 + 𝑡 𝑀) 𝑥) − 𝜑 (𝑥)
0≠𝑡→0 𝑡
d 1
= |det (𝐼𝑛 + 𝑡 𝑀)| 2 𝜑 ((𝐼𝑛 + 𝑡 𝑀) 𝑥)
d𝑡 𝑡=0
𝑛 𝑛
∑︁ 𝜕𝜑 1 ©∑︁
= (𝑥) + ­ 𝑐 𝑗, 𝑗 ® 𝜑 (𝑥)
ª
𝑐 𝑗,𝑘 𝑥 𝑘
𝑗=1
𝜕𝑥 𝑗 2 𝑗=1
« ¬
1
d 1
since d𝑡 |det (𝐼𝑛 + 𝑡 𝑀)| 2 = 2 Tr 𝑀. We obtain
𝑡=0

𝑛 𝑛
−1
∑︁ 𝜕𝜑 1 ©∑︁
D 𝜋 𝜒 (𝑀) 𝜑 (𝑥) = + ­ 𝑐 𝑗, 𝑗 ® 𝜑 (𝑥) . (13.2.18)
ª
𝑐 𝑗,𝑘 𝑥 𝑘
𝑗=1
𝜕𝑥 𝑗 2
« 𝑗=1 ¬
We can now apply Proposition 13.1.27:
Proposition
−1 13.2.13 The differential at the origin of the metaplectic representation,
D 𝜋 , maps 𝔰𝔭 (𝑛, R) onto the Lie algebra 𝔐𝔭 (𝑛) (with respect to the commutation
√
bracket) of the second-order differential operators in R𝑛 of the form −1𝑃 (𝑥, D)
with

1 √ ∑︁
𝑛
∑︁ 𝑛
∑︁ 𝑛
∑︁ 𝑛
𝑃(𝑥, D) = 𝑎 𝑗,𝑘 D 𝑗 D 𝑘 + 𝑐 𝑗,𝑘 𝑥 𝑘 D 𝑗 + 𝑏 𝑗,𝑘 𝑥 𝑗 𝑥 𝑘 − −1 𝑐 𝑗, 𝑗 ,
𝑗,𝑘=1 𝑗=1 𝑗,𝑘=1
2 𝑗=1
(13.2.19)
where 𝑎 𝑗,𝑘 = 𝑎 𝑘 𝑗 , 𝑏 𝑗,𝑘 = 𝑏 𝑘 𝑗 and 𝑐 𝑗,𝑘 are real numbers and D 𝑗 = √1 𝜕
.
−1 𝜕𝑥 𝑗

∫ (13.2.19) maps S (R ) into

The differential operator ∫ itself and it is “formally”
𝑛

selfadjoint, meaning that R𝑛 𝜑 (𝑥) 𝑃 (𝑥, D) 𝜓 (𝑥)d𝑥 = R𝑛 𝜓 (𝑥)𝑃 (𝑥, D) 𝜑 (𝑥) d𝑥 if

√
𝜑,𝜓 ∈ S (R𝑛 ). This allows us to define exp −1𝑃 (𝑥, D) as a unitary operator on
𝐿 2 (R𝑛 ). One can say that 𝔐𝔭 (𝑛) is the Lie algebra of Mp (𝑛, R) when the latter is
viewed as a subgroup of U 𝐿 2 (R𝑛 ) .
1
Let us now carry out the dilations 𝑥 𝑗 ↦→ 𝜂− 2 𝑥 𝑗 , 𝑗 = 1, ..., 𝑛, 𝜂 ≠ 0. The transform
of 𝜂𝑃 (𝑥, D) is the differential operator

1 √ ∑︁
𝑛
∑︁ 𝑛
∑︁ 𝑛
∑︁ 𝑛
𝑎 𝑗,𝑘 D 𝑗 D 𝑘 + 𝑐 𝑗,𝑘 𝑥 𝑘 𝜂D 𝑗 + 𝑏 𝑗,𝑘 𝑥 𝑗 𝑥 𝑘 𝜂2 − −1 𝑐 𝑗, 𝑗 𝜂.
𝑗,𝑘=1 𝑗=1 𝑗,𝑘=1
2 𝑗=1

The latter differential operator can be viewed as the partial Fourier transform with
respect to a new variable 𝑦 of the differential operator
398 13 Elements of Symplectic Geometry
𝑛
∑︁ √ ∑︁ 𝑛
𝜕
𝑃(𝑥, D 𝑥 , D 𝑦 ) = 𝑎 𝑗,𝑘 D 𝑗 D 𝑘 + −1 𝑐 𝑗,𝑘 𝑥 𝑘 D 𝑗 (13.2.20)
𝑗,𝑘=1 𝑗=1
𝜕𝑦
𝑛 𝑛
∑︁ 𝜕2 1 ∑︁ 𝜕
− 𝑏 𝑗,𝑘 𝑥 𝑗 𝑥 𝑘 − 𝑐 𝑗, 𝑗 ,
𝑗,𝑘=1
𝜕𝑦 2 2 𝑗=1 𝜕𝑦

which is of Grushin type (Section 2.4).

Example 13.2.14 When 𝑛 = 1 the operators (13.2.19) are of the form

1√
𝑃(𝑥, D) = 𝑎D2 + 𝑐𝑥D + 𝑏𝑥 2 − −1𝑐. (13.2.21)
2
𝜕 2
Among these operators one finds the classical harmonic oscillator operator, − 𝜕𝑥 2 +
2
𝑥 . The operator (13.2.20) associated to the harmonic oscillator is the Grushin
𝜕2 2 𝜕2
operator 𝜕𝑥 2 + 𝑥 𝜕𝑦 2 .

13.3 Symplectic Manifolds

13.3.1 Symplectic manifolds

In this section M will denote a regular manifold countable at infinity; the meaning
of “regular” is the same as before: C ∞ , C 𝜔 or complex-analytic. The base field K is
R or C.

Definition 13.3.1 A symplectic structure on the manifold M is the datum of a

K-valued, closed, regular two-form 𝜛 on M nondegenerate at every point. The pair
(M, 𝜛) will be referred to as a symplectic manifold; 𝜛 will be referred to as the
fundamental symplectic two-form on M.

We spell out the properties of the fundamental symplectic two-form 𝜛:

(1) The value 𝜛℘ of 𝜛 at an arbitrary point ℘ ∈ M is a nondegenerate skew-
symmetric bilinear functional on 𝑇℘ M ×𝑇℘ M. In the case K = R, 𝜛℘ extends as
a skew-symmetric bilinear functional (also denoted by 𝜛℘ ) on C𝑇℘ M × C𝑇℘ M.
(2) Given any open set U ⊂ M and any pair of regular vector fields 𝑋, 𝑌 in U the
function 𝜛 (𝑋, 𝑌 ) : ℘ ↦→ 𝜛℘ 𝑋℘ , 𝑌℘ is regular in U.
(3) The form 𝜛 is closed.
Property (1) and Proposition 13.1.4 imply directly that dim M is even; we
write 𝑛 = 21 dim M. When K = R the skew-symmetric bilinear functional 𝜛℘
on C𝑇℘ M × C𝑇℘ M is nondegenerate. Indeed, if 𝑬 is a real (finite-dimensional)
vector space an arbitrary nondegenerate real-valued bilinear functional 𝛽 on 𝑬 × 𝑬
13.3 Symplectic Manifolds 399

has a unique complex-valued

bilinear extension to (𝑬⊗C) × (𝑬⊗C); this extension
√
is nondegenerate. For if 𝛽 u + −1v, w = 0, i.e., 𝛽 (u, w) = 𝛽 (v, w) = 0, for all
w ∈ 𝑬 then necessarily u = v = 0.

Remark 13.3.2 For the meaning of 𝑇M and 𝑇 ∗ M in the complex-analytic case

we refer the reader to Remark 9.3.2. When K = C we have 𝑇℘∗ M =𝑇℘(1,0) M (cf.
Subsection 9.4.5), the derivation d must be understood as the “holomorphic” exterior
derivative 𝜕𝑧 and by a C 𝑚 vector field (0 ≤ 𝑚 ≤ +∞ or 𝑚 = 𝜔) we mean a linear
combination with C 𝑚 complex coefficients of the partial derivations 𝜕𝑧𝜕 𝑗 if 𝑧1 , ..., 𝑧 2𝑛
are holomorphic local coordinates.

Let (U, 𝑥1 , ..., 𝑥 2𝑛 ) be a (regular) local chart in M; we can write

2𝑛
1 ∑︁
𝜛= 𝑎 𝑖 𝑗 (𝑥) d𝑥 𝑖 ∧ d𝑥 𝑗
2 𝑖, 𝑗=1

with 𝑎 𝑖 𝑗 ∈ R (U) (i.e., the coefficients 𝑎 𝑖 𝑗 are regular in U). We have 𝑎 𝑖 𝑗 + 𝑎 𝑗𝑖 = 0

for all 𝑖, 𝑗 = 1, ..., 2𝑛; and thus
∑︁
𝜛= 𝑎 𝑖 𝑗 (𝑥) d𝑥𝑖 ∧ d𝑥 𝑗 . (13.3.1)
1≤𝑖< 𝑗 ≤2𝑛

The skew-symmetric matrix 𝑎 𝑖 𝑗 1≤𝑖, 𝑗 ≤2𝑛 is nowhere singular. Moreover,

2𝑛
∑︁ ∑︁ 𝜕𝑎 𝑖 𝑗
d𝜛 = d𝑥𝑖 ∧ d𝑥 𝑗 ∧ d𝑥 𝑘 = 0. (13.3.2)
1≤𝑖< 𝑗 ≤2𝑛 𝑘=1
𝜕𝑥 𝑘

𝜕𝑎 𝜕𝑎𝑖𝑘 𝜕𝑎 𝑗,𝑘
The skew-symmetry of the exterior product implies 𝜕𝑥𝑖𝑘𝑗 − 𝜕𝑥 𝑗 + 𝜕𝑥𝑖 = 0 for any
triplet of integers 𝑖, 𝑗,𝑘 such that 1 ≤ 𝑖 < 𝑗 < 𝑘 ≤ 2𝑛.

Example 13.3.3 Suppose M = K𝑛𝑥 × K𝑛𝜉 is equipped with the symplectic structure
Í
defined by the 2-form 𝜛 = d𝜉 ∧ d𝑥 = 𝑛𝑗=1 d𝜉 𝑗 ∧ d𝑥 𝑗 ; we identify 𝑇( 𝑥, 𝜉 ) M with
K2𝑛 . Let
𝑛 𝑛
∑︁ 𝜕 𝜕 ∑︁ 𝜕 𝜕
𝑋= 𝑎𝑗 + 𝑏𝑗 , 𝑋′ = 𝑎 ′𝑗 + 𝑏 ′𝑗
𝑗=1
𝜕𝑥 𝑗 𝜕𝜉 𝑗 𝑗=1
𝜕𝑥 𝑗 𝜕𝜉 𝑗

be two smooth vector fields in an open subset U of M. We have

𝑛
∑︁
𝜛 (𝑋, 𝑋 ′) = d𝜉 𝑗 , 𝑋 d𝑥 𝑗 , 𝑋 ′ − d𝜉 𝑗 , 𝑋 ′ d𝑥 𝑗 , 𝑋
𝑗=1
𝑛
∑︁
= 𝑎 ′𝑗 𝑏 𝑗 − 𝑎 𝑗 𝑏 ′𝑗 .
𝑗=1
400 13 Elements of Symplectic Geometry

Let (M, 𝜛) and (N , 𝜛 ′) be two regular symplectic manifolds. A regular map

𝑓 : M −→ N will be called a symplectic map if the pullback 𝑓 ∗ 𝜛 ′ is equal to
𝜛. If moreover 𝑓 is a regular isomorphism (i.e., a C ∞ or C 𝜔 diffeomorphism when
K = R, a biholomorphism when K = C) then 𝑓 is called a symplectomorphism and
we shall say that M and N are symplectomorphic.

13.3.2 Hamiltonian vector fields, Poisson bracket

Let (M, 𝜛) be a regular symplectic manifold. For each ℘ ∈ M the nondegenerate

bilinear functional 𝜛℘ on 𝑇℘ M defines a bijection

e℘ : 𝑇℘ M −→ 𝑇℘∗ M
𝜛 (13.3.3)

by the formula
e℘ 𝑋, 𝑌 ⟩ = 𝜛℘ (𝑋, 𝑌 )
⟨𝜛 (13.3.4)
for any pair of regular vector fields 𝑋 and 𝑌 in a neighborhood of ℘ [cf. (13.1.1)–
(13.1.2)]; ⟨·, ·⟩ is the duality bracket between tangent and cotangent vectors. As ℘
ranges over M (13.3.3) defines a vector bundle isomorphism 𝜛 e : 𝑇M −→ 𝑇 ∗ M.
When K = R, 𝜛 e extends as a vector bundle isomorphism 𝜛 e C : C𝑇M −→ C𝑇 ∗ M.
One can also use the isomorphism 𝜛 e to transform the bilinear functional 𝜛 from
𝑇M × 𝑇M into a bilinear functional 𝜛 ∗ on 𝑇 ∗ M × 𝑇 ∗ M.

Definition 13.3.4 Let (M, 𝜛) be a symplectic manifold and 𝑓 a complex-valued C ∞

function in an open set U ⊂ M. The unique smooth vector field 𝑋 (with complex
coefficients) in U such that 𝜛𝑋
e = −d 𝑓 is called the Hamiltonian vector field of 𝑓
and shall be denoted by 𝐻 𝑓 .

It is evident that the action of 𝐻 𝑓 can be extended to much wider classes than C ∞
functions 𝑓 : definitely to C 1 and even to Lipschitz functions.

Example 13.3.5 Consider the symplectic space K𝑛𝑥 × K𝑛𝜉 with the symplectic 2-
form 𝜛 = d𝜉 ∧ d𝑥 = Í𝑛𝑗=1 d𝜉 𝑗 ∧ d𝑥 𝑗 . With 𝑋 and 𝑋 ′ as inÍExample 13.3.3, (13.3.4)
Í
e 𝑋 ′⟩ = 𝑛𝑗=1 𝑎 ′𝑗 𝑏 𝑗 − 𝑎 𝑗 𝑏 ′𝑗 whence 𝜛𝑋
implies ⟨𝜛𝑋, e = 𝑛𝑗=1 𝑏 𝑗 d𝑥 𝑗 − 𝑎 𝑗 d𝜉 𝑗 . It is
easily checked that 𝐻 𝑓 = 𝑛𝑗=1 𝜕𝜕𝜉𝑓𝑗 𝜕𝑥𝜕 𝑗 − 𝜕𝑥
𝜕𝑓 𝜕
Í
𝑗 𝜕 𝜉𝑗
for 𝑓 regular.

In the next statements U ⊂ M is an open set.

Proposition 13.3.6 If 𝑓 ∈ C ∞ (U) and 𝑋 is a smooth vector field in U then

𝜛 𝑋, 𝐻 𝑓 = 𝑋 𝑓 . (13.3.5)

Proof By (13.3.4) and Definition 13.3.1 we have, at every point of U,

𝜛 𝑋, 𝐻 𝑓 = −⟨𝜛𝐻 e 𝑓 , 𝑋⟩ = ⟨d 𝑓 , 𝑋⟩. □
13.3 Symplectic Manifolds 401

A restatement of (13.3.5) is

d 𝑓 = 𝜛⌟𝐻 𝑓 . (13.3.6)

Corollary 13.3.7 Let 𝑓 , 𝑔 ∈ C ∞ (U); we have

𝐻 𝑓 𝑔 = −𝐻𝑔 𝑓 = 𝜛 𝐻 𝑓 , 𝐻𝑔 . (13.3.7)

Definition 13.3.8 Let 𝑓 , 𝑔 ∈ C ∞ (U); 𝐻 𝑓 𝑔 is called the Poisson bracket of 𝑓 and

𝑔 and denoted by { 𝑓 , 𝑔}.

The Poisson bracket defines a skew-symmetric bilinear map

C ∞ (U) × C ∞ (U) ∋ ( 𝑓 , 𝑔) ↦→ { 𝑓 , 𝑔} ∈ C ∞ (U) .

The Poisson bracket satisfies the Jacobi identity [cf. (12.1.6)]:

Proposition 13.3.9 Let 𝑓 , 𝑔, ℎ ∈ C ∞ (U); we have

{ 𝑓 , {𝑔, ℎ}} + {𝑔, {ℎ, 𝑓 }} + {ℎ, { 𝑓 , 𝑔}} = 0. (13.3.8)

Proof We have, by (12.1.8),

0 = d𝜛 𝐻 𝑓 , 𝐻𝑔 , 𝐻 ℎ

= 𝐻 𝑓 𝜛 𝐻𝑔 , 𝐻 ℎ − 𝐻𝑔 𝜛 𝐻 𝑓 , 𝐻 ℎ + 𝐻 ℎ 𝜛 𝐻 𝑓 , 𝐻𝑔

− 𝜛 𝐻 𝑓 , 𝐻𝑔 , 𝐻 ℎ + 𝜛 𝐻 𝑓 , 𝐻 ℎ , 𝐻𝑔 − 𝜛 𝐻𝑔 , 𝐻 ℎ , 𝐻 𝑓 .

By applying (13.3.5) and (13.3.7) we obtain

𝐻 𝑓 {𝑔, ℎ} − 𝐻𝑔 { 𝑓 , ℎ} + 𝐻 ℎ { 𝑓 , 𝑔}
−𝐻 𝑓 𝐻𝑔 ℎ + 𝐻𝑔 𝐻 𝑓 𝑓 + 𝐻 𝑓 𝐻 ℎ 𝑔 − 𝐻 ℎ 𝐻 𝑓 𝑔
−𝐻𝑔 𝐻 ℎ 𝑓 + 𝐻 ℎ 𝐻𝑔 𝑓 = 0,

whence (13.3.8). □

Corollary 13.3.10 Let 𝑓 , 𝑔, ℎ ∈ C ∞ (U); then

𝐻 𝑓 {𝑔, ℎ} = 𝐻 𝑓 𝑔, ℎ + 𝑔, 𝐻 𝑓 ℎ , (13.3.9)

𝐻 𝑓 , 𝐻𝑔 ℎ = 𝐻 { 𝑓 ,𝑔 } ℎ. (13.3.10)

Proof Both (13.3.9) and (13.3.10) are rewritings of (13.3.8). □

Let 𝑓 ∈ C ∞ (U); obviously 𝐻 𝑓 𝑓 ≡ 0 in U: 𝑓 is constant on the integral curves
of 𝐻 𝑓 .
Now consider the pullback and pushforward isomorphisms defined by the flow
of the Hamiltonian vector field of 𝑓 , exp 𝑡𝐻 𝑓 (see the end of Section 12.3); here 𝑓
is a regular function in M. Let Ω be a relatively compact open subset of M and
402 13 Elements of Symplectic Geometry

𝜏Ω > 0 be such that exp 𝑡𝐻 𝑓 defines a regular isomorphism of Ω onto an open subset
Ω𝑡 of M whatever 𝑡 ∈ (−𝜏Ω , 𝜏Ω ). If |𝑡| < 𝜏Ω we have the regular isomorphism [cf.
(12.3.11)] ∗
exp 𝑡𝐻 𝑓 : Λ2𝑇 ∗ M Ω𝑡 −→ Λ2𝑇 ∗ M Ω . (13.3.11)

Proposition 13.3.11 If the function 𝑓 is regular the map (13.3.11) preserves the
fundamental 2-form 𝜛 on the symplectic manifold M.

Proof On the one hand Lemma 12.3.5, implies

d ∗ ∗
exp 𝑡𝐻 𝑓 𝜛 = exp 𝑡𝐻 𝑓 d 𝐻 𝑓 ⌟𝜛 ;
d𝑡

on the other hand (13.3.6) implies d 𝐻 𝑓 ⌟𝜛 = 0. □

Example 13.3.12 In symplectic space K𝑛𝑥 × K𝑛𝜉 with the symplectic 2-form 𝜛 =
d𝜉 ∧ d𝑥 let 𝑓 (𝑥, 𝜉) = 𝐴𝑥 · 𝜉 with 𝐴 an 𝑛 × 𝑛 matrix with (constant) entries in K; then
𝐻 𝑓 = 𝐴𝑥 · 𝜕𝑥 − 𝐴⊤ 𝜉 · 𝜕 𝜉 . The flow of 𝐻 𝑓 is defined by the equations

d𝑥 d𝜉
= 𝐴𝑥, = −𝐴⊤ 𝜉;
d𝑡 d𝑡
⊤

it is the map exp 𝑡𝐻 𝑓 : (𝑥, 𝜉) ↦→ e𝑡 𝐴𝑥, e−𝑡 𝐴 𝜉 . If ℎ ∈ C ∞ K𝑛𝑥 × K𝑛𝜉 we have
∗ ⊤

exp 𝑡𝐻 𝑓 ℎ (𝑥, 𝜉) = ℎ e𝑡 𝐴𝑥, e−𝑡 𝐴 𝜉 . The pullback of the differential dℎ under the
∗
map exp 𝑡𝐻 𝑓 is given by [cf. (12.3.9)]
∗ ∗
exp 𝑡𝐻 𝑓 dℎ = d exp 𝑡𝐻 𝑓 ℎ
⊤

= e𝑡 𝐴 (𝜕𝑥 ℎ) e𝑡 𝐴𝑥, e−𝑡 𝐴 𝜉 · d𝑥
⊤ ⊤

+ e−𝑡 𝐴 𝜕 𝜉 ℎ e𝑡 𝐴𝑥, e−𝑡 𝐴 𝜉 · d𝜉.
∗ ∗ ⊤
In particular,
∗ exp 𝑡𝐻 𝑓 d𝑥 = e𝑡 𝐴d𝑥, exp 𝑡𝐻 𝑓 d𝜉 = e−𝑡 𝐴 d𝜉. It is clear that
exp 𝑡𝐻 𝑓 (d𝜉 ∧ d𝑥) = d𝜉 ∧ d𝑥.

13.3.3 Submanifolds of a symplectic manifold

Let 𝜄 : N −→ M be a regular immersion (Definition 9.3.7). We can use the map

𝜄 to pullback the fundamental form 𝜛 to N as a regular, closed two-form 𝜄∗ 𝜛 [cf.
(9.4.11)]. Generally speaking, it is not true that 𝜄∗ 𝜛 is nondegenerate, not even that
it has constant rank (we recall that the rank of a two-form is equal to the rank of the
matrix representing it in an arbitrary coordinate patch). Actually, we shall always
assume that 𝜄 is injective, i.e., that N does not have any self-intersection, and we
13.3 Symplectic Manifolds 403

identify N with its image 𝜄 (N ); we regard N as a regular, immersed submanifold

of M and 𝑇N as a subbundle of 𝑇M. We denote by 𝜛| N the restriction of the
fundamental symplectic two-form to N .
We shall denote by 𝑇℘ N ⊤𝜛 (℘ ∈ N ) the vector subspace of 𝑇℘ M perpendicular
to 𝑇℘ N for the symplectic form 𝜛 and by 𝑇N ⊤𝜛 the subset of 𝑇M made up
by the linear subspaces 𝑇℘ N ⊤𝜛 as ℘ ranges over N . We shall sometimes refer
to 𝑇℘ N ⊤𝜛 as the symplectic orthogonal of 𝑇℘ N . Let 𝜛e : 𝑇M −→ 𝑇 ∗ M be the
vector bundle isomorphism defined by the fundamental symplectic form 𝜛. Recall
−1
that −𝜛d
e 𝑓 = 𝐻 𝑓 , the Hamiltonian vector field of 𝑓 (Definition 13.3.4). It follows
from (13.3.4) that 𝑇N ⊤𝜛 is the preimage of the conormal bundle 𝑇M ∗ N of N in

M (Subsection 9.3.3) under the map 𝜛 e and, as a consequence, that 𝑇N ⊤𝜛 is a

subbundle of 𝑇M of rank equal to codim N .

Proposition 13.3.13 Let N be a regular, immersed submanifold of M without self-

intersections and let ℘ ∈ N . For a vector 𝜗 ∈ 𝑇℘ N to belong to 𝑇℘ N ⊤𝜛 it is
necessary and sufficient that there be a regular function 𝑓 in a neighborhood of ℘ in
M which vanishes identically in a neighborhood of ℘ in N and is such that 𝜗 = 𝐻 𝑓
at ℘.

Proof Indeed, for 𝜛𝜗e ∈ 𝑇℘∗ M to belong to the conormal space 𝑇M,℘
∗ N it is necessary
and sufficient that there be a regular function 𝑓 in a neighborhood of ℘ in M which
vanishes identically in a neighborhood of ℘ in N and is such that d 𝑓 = 𝜛𝜗.
e □
Every point of N has a neighborhood V in N which is an embedded submanifold
of M and in fact, which can be defined by the vanishing of a number of functions
𝜑1 , ..., 𝜑𝑟 defined and regular in a neighborhood U of ℘ in M. If U is sufficiently
small we can select the functions 𝜑 𝑗 so that d𝜑1 ∧ · · · ∧ d𝜑𝑟 ≠ 0 at every point of
U, in which case 𝑟 = codim N . Keep in mind that V might not be equal to U ∩ N .
In the next statement, however, it is simpler to assume that V = U ∩ N .

Proposition 13.3.14 Let N be a regular submanifold of M and U ⊂ M an open

set intersecting N . Suppose there are regular functions 𝜑1 , ..., 𝜑𝑟 in U such that

U ∩ N = ℘ ∈ U; 𝜑 𝑗 (℘) = 0, 𝑗 = 1, ..., 𝑟

and, moreover, that d𝜑1 ∧ · · · ∧ d𝜑𝑟 ≠ 0 at every point of U. Under these conditions
the following is true.
(1) The symplectic orthogonal 𝑇N ⊤𝜛 is spanned over U ∩ N by the Hamiltonian
vector fields 𝐻 𝜑 𝑗 , 𝑗 = 1, ..., 𝑟.
(2) At every point of U ∩ N we have

rank 𝜄∗N 𝜛 + codim N = rank 𝜑𝑖 , 𝜑 𝑗 1≤𝑖, 𝑗 ≤𝑟 + dim N

(13.3.12)

where 𝜄∗N 𝜛 stands for the pullback of the 2-form 𝜛 to N .

404 13 Elements of Symplectic Geometry

Proof The differentials d𝜑1 , ..., d𝜑𝑟 span the conormal bundle of N over U ∩ N and
−1
e 𝑗 = −𝐻 𝜑 𝑗 , 𝑗 = 1, ..., 𝑟, whence Assertion (1). The bilinear form 𝜄∗N 𝜛 induces
𝜛d𝜑
a linear map 𝑇℘ N −→ 𝑇℘∗ N with nullspace 𝑇℘ N ∩ 𝑇℘ N ⊤𝜛 (℘ ∈ U ∩ N ) thus
rank℘ 𝜄∗N 𝜛 = dim N − dim 𝑇℘ N ∩ 𝑇℘ N ⊤𝜛 . On the other hand,

rank 𝜑𝑖 , 𝜑 𝑗 1≤𝑖, 𝑗 ≤𝑟
= rank 𝐻 𝜑 𝑗 𝜑𝑖
1≤𝑖, 𝑗 ≤𝑟

is equal to 𝑟 minus the dimension of the span of the 𝐻 𝜑 𝑗 that are tangent to N . By
⊤𝜛
the latter dimension is equal to dim 𝑇℘ N ∩ 𝑇℘ N ⊤𝜛 . We con-
Proposition 13.3.13
clude that rank 𝜑𝑖 , 𝜑 𝑗 (℘) 1≤𝑖, 𝑗 ≤𝑟 = codim N − dim 𝑇℘ N ∩ 𝑇℘ N , whence
Assertion (2). □
We underline the fact that both ranks in (13.3.12) are even numbers (as this is
always true of a skew-symmetric bilinear form).
Constant rank allows us to distinguish various classes of immersed submanifolds
(without self-intersections). Reflecting the classification of vector subspaces of a
symplectic vector space (Definition 13.1.6) we have

Definition 13.3.15 An immersed submanifold N of M without self-intersections is

said to be
(1) symplectic if 𝑇N ∩ 𝑇N ⊥𝜛 = 0, the zero section of 𝑇M| N ;
(2) isotropic if 𝑇N ⊂ 𝑇N ⊥𝜛 ;
(3) coisotropic if 𝑇N ⊥𝜛 ⊂ 𝑇N ;
(4) Lagrangian if 𝑇N ⊥𝜛 = 𝑇N .

Proposition 13.3.16 To say that the immersed submanifold N of M is symplectic

is the same as saying that 𝜄∗N 𝜛 is nondegenerate; in this case dim N is even and
𝑇M| N = 𝑇N ⊕𝑇N ⊥𝜛 . To say that N is isotropic (resp., Lagrangian) is the same as
saying that 𝜄∗N 𝜛 ≡ 0, in which case dim N ≤ 21 dim M (resp., dim N = 21 dim M).
If N is coisotropic dim N ≥ 21 dim M. In all these cases the rank of 𝜄∗N 𝜛 is constant
in N .

Proof The statements concerning dim N are direct consequences of Proposition

13.1.7. The rank of 𝜄∗N 𝜛 is equal to dim N when N is symplectic, to zero when N
is isotropic and to dim N − codim N when N is coisotropic. □

Proposition 13.3.17 Let N be a regular, immersed submanifold of M. To say that

the rank of 𝜄∗N 𝜛 is constant is equivalent to saying that 𝑇N ∩ 𝑇N ⊤𝜛 is a (regular)
vector subbundle of 𝑇N , which is then necessarily involutive (Definition 12.2.3).

Proof The first part of the statement follows from the fact that

dim 𝑇℘ N ∩ 𝑇℘ N ⊤𝜛 = dim N − rank℘ 𝜄∗N 𝜛℘ , ℘ ∈ N .

13.3 Symplectic Manifolds 405

Let 𝑓 and 𝑔 be regular functions in a neighborhood U of ℘ in M which vanish

identically in a neighborhood V of ℘ in N and are such that 𝐻 𝑓 and 𝐻𝑔 are tangent
to N in V. Then { 𝑓 , 𝑔} = 𝐻 𝑓 𝑔 vanishes in V and therefore 𝐻 𝑓 , 𝐻𝑔 = 𝐻 { 𝑓 ,𝑔 } is
a regular section of 𝑇℘ N ⊤𝜛 . Since 𝐻 𝑓 , 𝐻𝑔 is tangent to N it is a regular section

of 𝑇℘ N ∩ 𝑇℘ N ⊤𝜛 . The proof is completed by applying Proposition 13.3.13. □

We can apply Frobenius’ Theorem (Theorem 12.2.5): if 𝑇N ∩ 𝑇N ⊤𝜛 is a vector
subbundle of 𝑇N the integral manifolds of 𝑇N ∩ 𝑇N ⊤𝜛 foliate N .

Definition 13.3.18 Let N be a regular, immersed submanifold of M. Suppose there

is a 𝜌 > 0, 𝜌 < dim N , such that rank℘ 𝜄∗N 𝜛 = 𝜌 for every ℘ ∈ N . We shall refer
to the integral manifolds of 𝑇N ∩ 𝑇N ⊤𝜛 as the bicharacteristic leaves and to the
foliation they form as the bicharacteristic foliation of N .

The bicharacteristic leaves are regular immersed submanifolds of N whose di-

mension is equal to rank 𝑇N ∩𝑇N ⊤𝜛 = dim N −rank 𝜄∗N 𝜛. If rank 𝑇N ∩𝑇N ⊤𝜛 = 1
we shall say bicharacteristic curve rather than bicharacteristic leaf.

13.3.4 The Darboux Theorem

Let (M, 𝜛) be a regular symplectic manifold with dim M = 2𝑛, 𝑛 a positive integer.

Definition 13.3.19 We shall call Darboux coordinates any system of local coordi-
nate 𝑥1 , ..., 𝑥 2𝑛 satisfying,
possibly after relabeling, the conditions {𝑥𝑖+𝑛 , 𝑥𝑖 } = 1 for
𝑖 = 1, ..., 𝑛, and 𝑥 𝑗 , 𝑥 𝑘 = 0 if | 𝑗 − 𝑘 | ≠ 𝑛.

Equivalent (modulo relabeling) to the conditions in Definition 13.3.19 are the
following: { 𝑓𝑖+1 , 𝑓𝑖 } = 1 if 𝑖 is odd and 1 ≤ 𝑖 ≤ 2𝑛 − 1; 𝑓 𝑗 , 𝑓𝑖 = 0 if |𝑖 − 𝑗 | ≠ 1.

Theorem 13.3.20 Let N be a regular submanifold of M and let ℘◦ ∈ N . Suppose

there are regular functions 𝑓1 , ..., 𝑓𝑞 (1 ≤ 𝑞 < dim N ) in a neighborhood V of ℘◦
in N and an integer 𝑟, 0 ≤ 2𝑟 ≤ 𝑞, such that the following hold at every point of V:
(1) d 𝑓1 ∧ · · · ∧ d 𝑓𝑞 ≠ 0;
(2) if 1 ≤ 𝑗 < 𝑘 ≤ 𝑞, 𝑓 𝑘 , 𝑓 𝑗 = 1 if 𝑗 is odd and 𝑘 = 𝑗 + 1 ≤ 2𝑟, 𝑓 𝑘 , 𝑓 𝑗 = 0
otherwise.
Under these hypotheses, for each 𝑖 = 1, ..., 𝑞 there is a regular extension 𝐹𝑖 of
𝑓𝑖 | N to a neighborhood U of ℘◦ in M such that 𝐹1 , ..., 𝐹𝑞 are part of a system of
Darboux coordinates in U.

It is understood that if 𝑟 = 0 we have 𝑓 𝑗 , 𝑓 𝑘 = 0 in N ∩U for all 𝑗, 𝑘 ∈ {1, ..., 𝑞}.
Proof Substituting 𝑓 𝑗 − 𝑓 𝑗 (℘◦ ) for 𝑓 𝑗 allows us to assume that 𝑓 𝑗 (℘◦ ) = 0 for all
𝑗 = 1, ..., 𝑞. It is agreed that the neighborhood U shall be contracted about ℘ as
many times and to any extent as needed in the course of the argument.
406 13 Elements of Symplectic Geometry

I. Case dim N = dim M. If 𝑞 = 2𝑛 there is nothing to prove: 𝑓1 , ..., 𝑓𝑞 is a system

of Darboux coordinates in U. We shall therefore suppose 𝑞 < 2𝑛. We observe
that the vector fields 𝐻 𝑓 𝑗 in U are linearly independent and commute, according to
(13.3.10) and Property (2) in the statement. By Frobenius’ Theorem they define a
regular foliation of U; and the same is true of any subsystem of the system of vector
fields 𝐻 𝑓1 , ..., 𝐻 𝑓𝑞 .

First suppose 𝑟 = 0, i.e., 𝑓 𝑗 , 𝑓 𝑘 = 0 for all 𝑗, 𝑘 ∈ {1, ..., 𝑞} (which entails
𝑞 ≤ 𝑛). Let Σ◦ be a regular submanifold of U, containing ℘◦ and transverse to the
span H◦ of 𝐻 𝑓1 , ..., 𝐻 𝑓𝑞 , meaning 𝑇M = H◦ ⊕ 𝑇 Σ◦ along Σ◦ (thus codim Σ◦ = 𝑞).
We solve the system of differential equations

𝐻 𝑓1 𝑓𝑞+1 = 1, 𝐻 𝑓 𝑗 𝑓𝑞+1 = 0, 𝑗 = 2, ..., 𝑞,

with “initial datum” 𝑓𝑞+1 Σ◦ = 0. The vector field 𝐻 𝑓𝑞+1 is transverse to the hypersur-
face 𝑓1 = 0 and therefore linearly independent of 𝐻 𝑓1 , ..., 𝐻 𝑓𝑞 , vector fields that are
all tangent to said hypersurface. We end up with a set of 𝑞 + 1 functions 𝑓 𝑗 having
Properties (1) and (2) in the statement, except that now 𝑟 = 1.
Now suppose 𝑟 ≥ 1. Let Σ be the regular submanifold of U defined by the
equations 𝑓1 (℘) = · · · = 𝑓2𝑟 (℘) = 0 (and therefore containing ℘◦ ); Σ is symplectic
and the span H of 𝐻 𝑓1 , ..., 𝐻 𝑓2𝑟 is transversal to Σ: 𝑇M = H ⊕ 𝑇 Σ along Σ. Consider
first the case 𝑞 = 2𝑟; let 𝑔 be any regular function on Σ vanishing at ℘◦ but whose
derivative d𝑔 does not vanish at any point of Σ. Let 𝑓𝑞+1 be the (regular) solution of
the following Cauchy problem for a system of first-order linear PDEs in U,

𝐻 𝑓 𝑗 𝑓𝑞+1 = 0, 𝑗 = 1, ..., 2𝑟, 𝑓𝑞+1 Σ = 𝑔. (13.3.13)

If 𝑞 = 2𝑟 we stop here (keep in mind that dim N = dim M).

If 𝑞 > 2𝑟 we exploit the fact that the vector fields 𝐻 𝑓 𝑘 (2𝑟 + 1 ≤ 𝑘 ≤ 𝑞) are
tangent to Σ. It implies the existence of a regular function 𝑔 on Σ such that

{𝑔, 𝑓 𝑘 } = 0 if 2𝑟 + 1 ≤ 𝑘 < 𝑞, 𝑔, 𝑓𝑞 = 1, (13.3.14)

at every point of Σ. Let 𝑓𝑞+1 be the solution of (13.3.13) with this new choice of 𝑔.
Suppose 1 ≤ 𝑗 ≤ 2𝑟 < 𝑘 ≤ 𝑞. On the one hand

𝐻 𝑓 𝑗 𝑓𝑞+1 , 𝑓 𝑘 = 𝐻 𝑓 𝑗 𝑓𝑞+1 , 𝑓 𝑘 = 0,

implying that 𝑓𝑞+1 , 𝑓 𝑘 is constant along the flow of 𝐻 𝑓 𝑘 . On the other hand,
𝐻 𝑓 𝑘 𝑓𝑞+1 = { 𝑓 𝑘 , 𝑔} in Σ. We derive from (13.3.14) that

𝑓𝑞+1 , 𝑓 𝑘 = 0 if 1 ≤ 𝑘 < 𝑞, 𝑓𝑞+1 , 𝑓𝑞 = 1,

in the whole of U. We end up with a set of 𝑞 + 1 functions 𝑓 𝑗 having Properties (1)

and (2) in the statement, with 𝑟 + 1 substituting 𝑟. Induction on 𝑟 concludes the proof
in the case N = M.
13.3 Symplectic Manifolds 407

II. Case dim N < dim M. Provided U is sufficiently small there is a regular
function 𝑔 in U such that 𝐻𝑔 ∉ 𝑇N at every point of U ∩ N . Let Σ1 ⊂ U be
a regular hypersurface containing U ∩ N and transverse to 𝐻𝑔 at each one of its
points. We can assume that U is foliated by the integral curves of 𝐻𝑔 that intersect
Σ1 . The union of all the integral curves of 𝐻𝑔 that intersect U ∩ N form a regular
submanifold N1 of U; we have dim N1 = 1 + dim N . For each 𝑗 = 1, ..., 𝑞, we solve
the differential equation 𝐻𝑔 𝐹 𝑗 = 0 in U with initial value 𝐹 𝑗 Σ1 = 𝑓 𝑗 . The functions
𝐹 𝑗 are regular in U and constant along each integral curve of 𝐻𝑔 ; their Poisson
brackets are annihilated by 𝐻𝑔 and thus are constant in N1 . Induction on dim N
allows us to conclude that 𝐹1 | N1 , ..., 𝐹𝑞 N1 have extensions to U that are part of a
system of Darboux coordinates. □

Corollary 13.3.21 (Darboux’ Theorem) Every point ℘ ∈ M has a neighborhood

U symplectomorphic to an open subset of the symplectic vector space K2𝑛 , 𝜛𝑛
(𝜛𝑛 : the canonical symplectic two-form on K2𝑛 ).

Proof Select at random a regular function 𝑓1 in a sufficiently small neighborhood U

of ℘ whose differential does not vanish at any point of U. Apply Theorem 13.3.20
with N = M and 𝑞 = 1. □
In the sequel, given Darboux coordinates 𝑥 1 , ..., 𝑥 2𝑛 , we shall often use the notation
𝜉 𝑗 or 𝑦 𝑗 in the place of 𝑥 𝑛+ 𝑗 ( 𝑗 = 1, ..., 𝑛). Thus we have

𝜉 𝑗 , 𝑥 𝑘 = 𝛿 𝑗,𝑘 , the Kronecker index, 𝑥 𝑗 , 𝑥 𝑘 = 𝜉 𝑗 , 𝜉 𝑘 = 0 (1 ≤ 𝑗, 𝑘 ≤ 𝑛).
(13.3.15)
The following consequences of Theorem 13.3.20 are worth mentioning. In the next
two statements the meaning of the noun manifold accords with Definition 9.1.7 (thus
we are not dealing with arbitrary immersed submanifolds).

Proposition 13.3.22 The following properties of a submanifold N of M are equiv-

alent:
(1) N is symplectic.
(2) Every point ℘ ∈ N has a neighborhood U in M in which there are Darboux
coordinates 𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 such that N ∩ U is defined in U by the equation
𝑥𝑖 = 𝜉𝑖 = 0, 𝑖 = 1, ..., 𝑚 = 21 codim N .

Proof That (2) =⇒ (1) follows directly from (13.3.12). If (1) holds, by Corollary
13.3.21 there are Darboux coordinates 𝑓1 , ..., 𝑓𝑛−𝑚 , 𝑔1 , ..., 𝑔𝑛−𝑚 in a neighborhood
V of ℘ in N in the sense of the symplectic structure of N inherited from M. There
is no loss of generality in assuming that 𝑓1 , ..., 𝑓𝑛−𝑚 , 𝑔1 , ..., 𝑔𝑛−𝑚 vanish at ℘. By
Theorem 13.3.20 𝑓1 , ..., 𝑓𝑛−𝑚 , 𝑔1 , ..., 𝑔𝑛−𝑚 have regular extensions to a neighborhood
U of ℘ in M that are part of a system of Darboux coordinates in U; there is no loss
of generality in assuming that the latter coordinates also vanish at ℘. Let ℎ be any
member of this system that is not one of the 𝑓 𝑗 , 𝑔 𝑘 . Since 𝐻 𝑓 𝑗 ℎ = 𝐻𝑔𝑘 ℎ = 0 at every
point of N ∩ U we derive that ℎ is constant, and therefore vanishes identically in
N ∩ U. □
408 13 Elements of Symplectic Geometry

Proposition 13.3.23 The following properties of a submanifold N of M are equiv-

alent:
(1) N is Lagrangian;
(2) every point ℘ ∈ N has a neighborhood U in M in which there are Darboux
coordinates 𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 , such that N ∩ U is defined by the equation
𝑥𝑖 = 0, 𝑖 = 1, ..., 𝑛.

The proof of Proposition 13.3.23 is quite similar to that of Proposition 13.3.22

and is left as an exercise.
Let 𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 be Darboux coordinates in an open set U ⊂ M. For each
𝑗, 𝑘 = 1, ..., 𝑛 we derive from (13.3.15):

𝐻 𝑥 𝑗 𝑥 𝑘 = 𝐻 𝜉 𝑗 𝜉 𝑘 = 0; 𝐻 𝑥 𝑗 𝜉 𝑘 = 0 if 𝑗 ≠ 𝑘; 𝐻 𝜉 𝑗 𝑥 𝑗 = 1.
𝜕
We conclude that 𝐻 𝜉 𝑗 = 𝜕𝑥 𝑗 , 𝐻 𝑥 𝑗 = − 𝜕𝜕𝜉 𝑗 . We apply (13.3.7) and (13.3.15):

𝜕 𝜕
𝜛 , = 𝜛 𝐻 𝜉 𝑗 , 𝐻 𝑥𝑘 = 𝜉 𝑗 , 𝑥 𝑘 = 𝛿 𝑗,𝑘 , (13.3.16)
𝜕𝜉 𝑘 𝜕𝑥 𝑗

𝜕 𝜕 𝜕 𝜕
𝜛 , =𝜛 , = 0.
𝜕𝜉 𝑗 𝜕𝜉 𝑘 𝜕𝑥 𝑗 𝜕𝑥 𝑘

We reach the conclusion that, in the local chart (U, 𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 ),
𝑛
∑︁
𝜛= d𝜉 𝑗 ∧ d𝑥 𝑗 , (13.3.17)
𝑗=1

which we shall often abbreviate to 𝜛 = d𝜉 ∧ d𝑥. If 𝑓 ∈ C ∞ (U) then

𝑛
∑︁ 𝜕 𝜕
𝐻𝑓 = 𝐻𝑓𝑥𝑗 + 𝐻𝑓 𝜉𝑗
𝑗=1
𝜕𝑥 𝑗 𝜕𝜉 𝑗
𝑛
∑︁ 𝜕 𝜕
=− 𝐻𝑥 𝑗 𝑓 + 𝐻 𝜉𝑗 𝑓 ,
𝑗=1
𝜕𝑥 𝑗 𝜕𝜉 𝑗

whence
𝑛
∑︁ 𝜕𝑓 𝜕 𝜕𝑓 𝜕
𝐻𝑓 = − . (13.3.18)
𝑗=1
𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝜉 𝑗

The expression of the Poisson bracket (Definition 13.3.8) of 𝑓 , 𝑔 ∈ C ∞ (U) is

𝑛
∑︁ 𝜕 𝑓 𝜕𝑔 𝜕 𝑓 𝜕𝑔
{ 𝑓 , 𝑔} = − . (13.3.19)
𝑗=1
𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝜉 𝑗
13.3 Symplectic Manifolds 409

13.3.5 The cotangent bundle as a symplectic manifold

Let now Ω be a regular manifold (as always, C ∞ , C 𝜔 or complex-analytic, K = R

or C); here 𝑛 = dimK Ω. We denote the variable point in 𝑇 ∗ Ω by (𝑥, 𝜉); in local
coordinates in some open set 𝑈 ⊂ Ω this means 𝑥 = (𝑥1 , ..., 𝑥 𝑛 ) and 𝜉 = (𝜉1 , ..., 𝜉 𝑛 )
with 𝜉 𝑗 the coordinates with respect to the frame d𝑥1 , ..., d𝑥 𝑛 of 𝑇 ∗ Ω|𝑈 (often called
the dual coordinates of the 𝑥 𝑗 ). When K = C it is often preferable to use the notation
𝑧 𝑗 for the complex coordinates and 𝜁 𝑗 for the complex coordinates with respect
to the basis d𝑧1 , ..., d𝑧 𝑛 in the cotangent spaces at points of 𝑈. Most of the time,
however, we shall continue to use 𝑥 and 𝜉. In all cases the coordinates 𝑥 𝑗 and 𝜉 𝑘 are
regular functions in their domain of definition. We denote by 𝜋 : 𝑇 ∗ Ω −→ Ω the
base projection (𝑥, 𝜉) ↦→ 𝜉.
We apply what precedes in this section to the manifold M = 𝑇Ω. We take as
symplectic form 𝜛 = d𝜎 with 𝜎 the tautological form (Subsection 9.3.3); 𝜛 is often
Í
referred to as the fundamental two-form. In local coordinates 𝜎 = 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗 =
𝜉 · d𝑥 and 𝜛 has the expression (13.3.17). The Hamiltonian vector field 𝐻 𝑓 of a C ∞
function 𝑓 is given by (13.3.18). The local flow of 𝐻 𝑓 through a point (𝑥 ◦ , 𝜉 ◦ ) is
expressed by the solutions of the Hamilton–Jacobi equations:

d𝑥 𝑗 𝜕𝑓 d𝜉 𝑗 𝜕𝑓
= (𝑥, 𝜉) , =− (𝑥, 𝜉) , 𝑗 = 1, ..., 𝑛, (13.3.20)
d𝑡 𝜕𝜉 𝑗 d𝑡 𝜕𝑥 𝑗

satisfying the initial conditions 𝑥| 𝑡=0 = 𝑥 ◦ , 𝜉 | 𝑡=0 = 𝜉 ◦ . The integral curves of 𝐻 𝑓

are called the bicharacteristic curves (or bicharacteristics) of 𝑓 (cf. Definition
13.3.18). The function 𝑓 is constant on any one of its bicharacteristic curves (by
definition connected). Any level set of 𝑓 which is a regular submanifold of 𝑇Ω is
foliated by the bicharacteristics of 𝑓 .
The symplectic manifold 𝑇 ∗ Ω has one important property not shared by arbitrary
symplectic submanifolds: the radial dilations (𝑥, 𝜉) ↦→ (𝑥, 𝜆𝜉), 0 ≠ 𝜆 ∈ K, form a
group of (nonsymplectic!) automorphisms of the vector bundle 𝑇Ω. The pullback of
𝜎 under the dilation (𝑥, 𝜉) ↦→ (𝑥, 𝜆𝜉) is 𝜆𝜎 and that of 𝜛 is 𝜆𝜛.
Let 𝜛e : 𝑇 (𝑇 ∗ Ω) −→ 𝑇 ∗ (𝑇 ∗ Ω) be the vector bundle isomorphism defined by 𝜛
[cf. (13.3.3)]; if 𝑓 is a smooth function we have 𝜛𝐻 e 𝑓 = −d 𝑓 and 𝜛 𝐻 𝑓 , 𝑋 = 𝑋 𝑓
[Definition 13.3.1 and (13.3.5)]. Now let 𝑥1 , ..., 𝑥 𝑛 be local coordinates in some open
set 𝑈 ⊂ Ω and let 𝜉1 , ..., 𝜉 𝑛 denote the dual coordinates. According to (13.3.16) we
have
𝜕 𝜕
𝜛
e e 𝜉 𝑗 = −d𝜉 𝑗 , 𝜛
= 𝜛𝐻 e = −𝜛𝐻 e 𝑥 𝑗 = d𝑥 𝑗 . (13.3.21)
𝜕𝑥 𝑗 𝜕𝜉 𝑗
We derive at once
Í −1
Proposition 13.3.24 We have 𝜌 = 𝑛𝑗=1 𝜉 𝑗 𝜕𝜕𝜉 𝑗 = 𝜛 e (𝜎). In particular the vector 𝜌
is independent of the choice of local coordinates 𝑥 𝑗 .
We shall refer to the vector field 𝜌 as the radial vector field along the fibers of the
cotangent bundle 𝑇 ∗ M. It can be regarded as the generator of the “Lie algebra” of the
“Lie group” of radial dilations, under which it is obviously invariant. We can regard
410 13 Elements of Symplectic Geometry

its restriction to the complement of the zero section 𝑇 ∗ Ω\0 as a differential operator.
The kernel of the differential operator 𝜌 in C ∞ (𝑇 ∗ Ω\0) consists of the functions
𝑓 (𝑥, 𝜉) that are homogeneous of degree zero with respect to 𝜉. When K = R the
spectrum of 𝜌 is the whole real line; if 0 ≠ 𝑘 ∈ R the eigenfunctions corresponding
to 𝑘 are the functions 𝑓 (𝑥, 𝜉) that are homogeneous of degree 𝑘 with respect to 𝜉.
An important class of Lagrangian submanifolds of 𝑇 ∗ Ω are the conormal bundles
of submanifolds of Ω (Definition 9.3.4).

Proposition 13.3.25 The conormal bundle 𝑁 ∗ Σ of a regular, immersed submanifold

Σ of Ω is a regular Lagrangian submanifold of 𝑇 ∗ Ω.

Proof Let 𝑉 be a neighborhood in Σ of a point ℘ ∈ Σ which is an embedded

submanifold of Ω. We can choose a regular local chart (𝑈, 𝑥1 , ..., 𝑥 𝑛 ) with dual
coordinates 𝜉1 , ..., 𝜉 𝑛 , such that 𝑉 = {𝑥 ∈ 𝑈; 𝑥1 = · · · = 𝑥 𝜈 = 0} (𝜈 = codim Σ).
Then 𝑁 ∗ Σ is defined in 𝑇 ∗ Ω| 𝑉 by the equations 𝑥1 = · · · = 𝑥 𝜈 = 𝜉 𝜈+1 = · · · = 𝜉 𝑛 =
0. □
The following statement will be useful later on.

Proposition 13.3.26 If a regular submanifold L of 𝑇 ∗ Ω\0 is Lagrangian and conic

(Definition 9.2.4) then the tautological one-form 𝜎 vanishes identically on L.

Proof It suffices to prove the claim locally and therefore we can assume that Ω is
an open subset of K𝑛 and identify 𝑇 ∗ Ω\0 with Ω × (K𝑛 \ {0}). Since dim L = 𝑛 an
arbitrary point (𝑥 ◦ , 𝜉 ◦ ) ∈ L has a neighborhood U in 𝑇 ∗ Ω\0 such that

L ∩ U = {(𝑥, 𝜉) ∈ L; 𝑓1 (𝑥, 𝜉) = · · · = 𝑓𝑛 (𝑥, 𝜉) = 0} ,

where 𝑓1 , . . . , 𝑓𝑛 are regular functions in U such that d 𝑓1 ∧ · · · ∧ d 𝑓𝑛 ≠ 0 at

every point of L. Since L is conic we can take U to be conic and assume that
𝑓 𝑗 (𝑥, 𝜆𝜉) = 𝑓 (𝑥, 𝜉) for all (𝑥, 𝜉) ∈ U, 𝜆 > 0, 𝑗 = 1, ..., 𝑛. The Euler identities entail
* 𝑛 + 𝑛
∑︁ ∑︁ 𝜕 𝑓𝑗
𝜎, 𝐻 𝑓 𝑗 = 𝜉 𝑘 d𝑥 𝑘 , 𝐻 𝑓 𝑗 = 𝜉𝑘 = 0.
𝑘=1 𝑘=1
𝜕𝜉 𝑘

We apply Proposition 13.3.14: whatever (𝑥, 𝜉) ∈ L ∩ U the symplectic orthogonal

of the tangent space 𝑇( 𝑥, 𝜉 ) L is spanned by the Hamiltonian vector fields 𝐻 𝑓 𝑗 . Since
L is Lagrangian 𝐻 𝑓1 , ..., 𝐻 𝑓𝑛 span 𝑇 L over L ∩ U. □

Let Ω′ be a second regular manifold and let 𝑓 : Ω −→ Ω′ be a local regular

isomorphism. For 𝑥 ∈ Ω let D 𝑓 (𝑥) : 𝑇𝑥 Ω −→ 𝑇 𝑓 ( 𝑥) Ω′ be the differential of the map
𝑓 ; if 𝜗 is a vector tangent to Ω at the point 𝑥 then D 𝑓 (𝑥) 𝜗 is the pushforward of 𝜗
(also denoted by 𝑓∗ 𝜗). The transpose D 𝑓 (𝑥) ⊤ :𝑇 𝑓∗ ( 𝑥) Ω′ −→ 𝑇𝑥∗ Ω is a linear bijection
−1
and therefore it has an inverse D 𝑓 (𝑥) ⊤ : 𝑇𝑥∗ Ω −→ 𝑇 𝑓∗ ( 𝑥) Ω′, the contragredient
of D 𝑓 (𝑥).
13.3 Symplectic Manifolds 411

Proposition 13.3.27 Let 𝑓 : Ω −→ Ω′ be a local regular isomorphism. Every point

𝑥 ∈ Ω has a neighborhood 𝑈 such that the map
−1
(𝑥, 𝜉) ↦→ 𝐹 (𝑥, 𝜉) = 𝑓 (𝑥) , D 𝑓 (𝑥) ⊤ 𝜉

is a symplectomorphism of 𝑇 ∗ Ω|𝑈 onto 𝑇 ∗ Ω| 𝑓 (𝑈) .

Proof Since 𝐹 is obviously a regular isomorphism provided 𝑈 is sufficiently small
−1
one must prove that it is symplectic. Setting 𝑦 = 𝑓 (𝑥), 𝜂 = D 𝑓 (𝑥) ⊤ 𝜉 it is
immediately checked that 𝑥 · d𝜉 = 𝐹 ∗ (𝑦 · d𝜂); and d𝐹 ∗ = 𝐹 ∗ d. □
As a special case a regular change of local coordinates in the base Ω defines a
symplectomorphism of the cotangent bundle over the domains that are involved.

13.3.6 Systems of functions of finite type

Let M be a symplectic manifold. We assume that M is regular, meaning, as always,

C ∞ , C 𝜔 or complex-analytic (with the base field K = R or C respectively) and that
its symplectic structure, i.e., the fundamental two-form in M, is regular.
Let 𝔉 be a family of regular functions in M (real-valued when K = R; see Section
9.1). We define
𝑽 (𝔉) = {℘ ∈ M; ∀ 𝑓 ∈ 𝔉, 𝑓 (℘) = 0} . (13.3.22)
We are going to consider Poisson multibrackets of elements 𝑓 𝜄0 , 𝑓 𝜄1 , ..., 𝑓 𝜄 𝑁 of 𝔉
(𝑁 ≥ 1): we call 𝐼 the multi-index (𝜄0 , 𝜄1 , ..., 𝜄 𝑁 ) and we define

𝑓 𝐼 = 𝑓 𝜄 𝑁 , 𝑓 𝜄 𝑁 −1 , · · · 𝑓 𝜄1 , 𝑓 𝜄0 · · · . (13.3.23)

We extend the definition to the case 𝑁 = 0 by setting 𝑓 𝐼 = 𝑓 𝜄0 when 𝐼 = (𝜄0 ). We

denote by 𝔤 (𝔉) the Lie algebra over K for the Poisson bracket generated by the
elements of 𝔉, i.e., the K-linear span of all the multibrackets (13.3.23).
Definition 13.3.28 We shall say that the set of regular functions 𝔉 is of finite type
at a point ℘ ∈ 𝑽 (𝔉) if ℘ ∉ 𝑽 (𝔤 (𝔉)). If 𝔉 is not of finite type at ℘ it is said to be of
infinite type at ℘.
We shall say that 𝔉 is of type 𝑁 at ℘ if there are elements 𝑓 𝜄0 , ..., 𝑓 𝜄 𝑁 of 𝔉 (𝑁 ≥ 1)
such that 𝑓 𝐼 (℘) ≠ 0 while {𝑔 𝑘 , {𝑔 𝑘−1 , · · · {𝑔1 , 𝑔0 } · · · }} (℘) = 0 for every subset
(𝑔0 , ..., 𝑔 𝑘 ) of 𝔉 such that 𝑘 < 𝑁.
We shall say that 𝔉 is of finite (resp., infinite) type in a subset 𝑬 of 𝑽 (𝔉) if 𝔉 is
of finite (resp., infinite) type at every point of 𝑬 and that 𝔉 is of finite type (resp.,
infinite) if 𝔉 is of finite type (resp., infinite) at every point of 𝑽 (𝔉).
Example 13.3.29 Let M = R2𝑥 × R2𝜉 equipped with the symplectic structure defined
by the two-form 𝜛2 = d𝜉1 ∧ d𝑥 1 + d𝜉2 ∧ d𝑥 2 . The pair of functions 𝜉1 , 𝑥1𝑁 𝜉2
(1 ≤ 𝑁 ∈ Z) is of type 𝑁 at every point (𝑥, 𝜉) such that 𝑥1 = 𝜉1 = 0, 𝜉2 ≠ 0.
412 13 Elements of Symplectic Geometry

Example 13.3.30 Let M be as in Example 13.3.29. The pair of functions 𝜉1 , 𝑥 2 𝜉2

is of infinite type at every point (𝑥, 𝜉) such that 𝑥2 = 𝜉1 = 0.

The properties of the Poisson bracket make it evident that the finite type property
is invariant under symplectomorphisms. It is also invariant under nonsingular (and
regular) substitutions:

Proposition 13.3.31 Suppose 𝔉 = { 𝑓1 , ..., 𝑓 𝜈 } is a finite set of regular functions in

M and let the matrix 𝐴 = 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝜈 have entries 𝑎 𝑗,𝑘 that are regular functions
in M. If 𝔉 is of finite type 𝑁 at ℘ ∈ 𝑽 (𝔉) and 𝐴 is nonsingular at ℘ then the set of
regular functions
𝜈
∑︁
𝑓 𝑗♭ = 𝑎 𝑗,𝑘 𝑓 𝑘 , 𝑗 = 1, ..., 𝜈,
𝑘=1

is also of type 𝑁 at ℘.

The proof is an exercise in exterior linear algebra.

A natural question is how to relate Definition 13.3.28 to the concept of finite type
for a system of vector fields (Definition 12.5.1). Let 𝔉 be a set of regular functions in
M. We look at the Hamiltonian vector fields 𝐻 𝑓 (Definition 13.3.4) of the functions
𝑓 ∈ 𝔉, or more generally 𝑓 ∈ 𝔤 (𝔉). In the notation of (13.3.23) we have
h h h i ii
𝐻 𝑓 𝐼 = 𝐻 𝑓 𝜄 𝑁 , 𝐻 𝑓 𝜄 𝑁 −1 , · · · 𝐻 𝑓 𝜄1 , 𝐻 𝑓 𝜄0 · · · . (13.3.24)

That the Hamiltonian vector fields 𝐻 𝑓 𝐼 are of finite type at ℘ ∈ 𝑽 (𝔉) is a much
stronger property than 𝔉 being of finite type at ℘. As seen in Example 13.3.29, the
Hamiltonian vector fields of 𝜉1 , 𝑥1 𝜉2 , respectively 𝜕𝑥𝜕 1 , 𝑥1 𝜕𝑥𝜕 2 − 𝜉2 𝜕𝜕𝜉1 , do not form
a system of finite type at any point of 𝑇𝑥 R2 as their linear span does not capture 𝜕𝜕𝜉2 .
The following statement is a direct consequence of Definition 13.3.4.

Proposition 13.3.32 Let 𝔉 be a set of regular functions in M. For the system of

Hamiltonian vector fields 𝐻 𝑓 𝐼 to be of finite type at ℘ ∈ 𝑽 (𝔉) it is necessary and
sufficient that there be functions 𝑔 𝑘 ∈ 𝔤 (𝔉), 𝑘 = 1, ..., 2𝑛 = dim M, such that
d𝑔1 ∧ · · · ∧ d𝑔2𝑛 ≠ 0 at ℘.

More generally speaking, the rank of the set of differentials d 𝑓 , 𝑓 ∈ 𝔤 (𝔉), is

equal to that of the Hamiltonian vector fields 𝐻 𝑓 .

Corollary 13.3.33 Let 𝔉 be a set of regular functions in M. If the Hamiltonian

vector fields 𝐻 𝑓 , 𝑓 ∈ 𝔉, form a system of finite type at ℘ ∈ 𝑽 (𝔉) then 𝔉 is of finite
type at ℘.

Proof If 𝔉 were of infinite type at ℘ ∈ 𝑽 (𝔉) we would have { 𝑓 𝐼 , 𝑓 } (℘) =

𝐻 𝑓 𝐼 𝑓 (℘) = 0 for every multibracket (13.3.23) and every 𝑓 ∈ 𝔤 (𝔉). The hypothesis
that the tangent vectors 𝐻 𝑓 𝐼 ℘ span 𝑇℘ M would entail d 𝑓 (℘) = 0, contradicting
Proposition 13.3.32. □
13.3 Symplectic Manifolds 413

We now take M = 𝑇 ∗ Ω\0, Ω a regular manifold, and 𝑋1 , ..., 𝑋𝜈 vector fields in Ω

with regular coefficients. In the domain 𝑈 ⊂ Ω of local regular coordinates 𝑥1 , ..., 𝑥 𝑛
we can write
𝑛
∑︁ 𝜕
𝑋𝑗 = 𝑎 𝑗,𝑘 (𝑥) , 𝑗 = 1, ..., 𝜈,
𝑘=1
𝜕𝑥 𝑘
with 𝑎 𝑗,𝑘 regular in 𝑈. Let 𝜎 = 𝜉 · d𝑥 denote the tautological 1-form on M (see pre-
ceding subsection, also Proposition 9.3.6). If we denote by 𝜉1 , ..., 𝜉 𝑛 the coordinates
with respect to the basis d𝑥 1 , ..., d𝑥 𝑛 in the cotangent spaces 𝑇𝑥∗ Ω, 𝑥 ∈ 𝑈, we have
𝑛
∑︁
𝜎, 𝑋 𝑗 (𝑥, 𝜉) = 𝑎 𝑗,𝑘 (𝑥) 𝜉 𝑘 (13.3.25)
𝑘=1

where ⟨·, ·⟩ is the bracket for the duality between tangent and cotangent vectors and
(𝑥, 𝜉) ∈ 𝑇 ∗ Ω|𝑈 . Given a multi-index 𝐼 = (𝜄0 , ..., 𝜄 𝑁 ), 1 ≤ 𝜄 𝑘 ≤ 𝜈 (𝑘 = 0, ..., 𝑁), we
write
𝑋𝐼 = 𝑋 𝜄 𝑁 , 𝑋 𝜄 𝑁 −1 , · · · 𝑋 𝜄1 , 𝑋 𝜄0 · · · . (13.3.26)

Proposition 13.3.34 We have

⟨𝜎, 𝑋𝐼 ⟩ = 𝜎, 𝑋 𝜄 𝑁 , 𝜎, 𝑋 𝜄 𝑁 −1 , · · · 𝜎, 𝑋 𝜄1 , 𝜎, 𝑋 𝜄0 ··· . (13.3.27)

Proof Use induction on 𝑁 and the readily verified identity

𝜎, 𝑋 𝜄1 , 𝜎, 𝑋 𝜄0 = 𝜎, 𝑋 𝜄1 , 𝑋 𝜄0 . (13.3.28)

Proposition 13.3.35 If the system of vector fields 𝑋1 , ..., 𝑋𝜈 is of finite type at a point
𝑥 ∈ Ω (Definition 12.5.1) then the set of functions 𝜎, 𝑋 𝑗 , 𝑗 = 1, ..., 𝜈, is of finite
type in 𝑇𝑥 Ω\ {0} (Definition 13.3.28).

Proof If the set of functions 𝜎, 𝑋 𝑗 , 𝑗 = 1, ..., 𝜈, were of infinite type at (𝑥, 𝜉) ∈

𝑇𝑥 Ω\ {0} then, by Proposition 13.3.34, we would have ⟨𝜎, 𝑋𝐼 ⟩ (𝑥, 𝜉) = 0 whatever
the multi-index 𝐼 = (𝜄0 , ..., 𝜄 𝑁 ), 1 ≤ 𝜄 𝑘 ≤ 𝜈 (𝑘 = 0, ..., 𝑁). The hypothesis that the
Í
tangent vectors 𝑋𝐼 | 𝑥 span 𝑇𝑥 Ω would imply that 𝜎 = 𝑛𝑗=1 𝜉 𝑗 d𝑥 𝑗 vanishes identically
in 𝑇𝑥 Ω, contradicting the fact that 𝜉 ≠ 0. □
414 13 Elements of Symplectic Geometry

13.4 Involutive Systems of Functions of Principal Type

13.4.1 Involutive systems of functions of principal type

In this section we look at an important class of systems of regular functions of

infinite type (Definition 13.3.28) at points of a regular submanifold of the symplectic
manifold M. When K = R is the base field the functions 𝑓1 , ..., 𝑓 𝜈 considered below
shall be real-valued unless specified otherwise.
We will use the notation 𝐹 = ( 𝑓1 , ..., 𝑓 𝜈 ) and

𝑽 (𝐹) = {℘ ∈ M; 𝑓1 (℘) = · · · = 𝑓 𝜈 (℘) = 0} .

Definition 13.4.1 The system of regular functions 𝐹 = ( 𝑓1 , ..., 𝑓 𝜈 ) shall be said to be

of principal type at a point ℘◦ ∈ M if there is a neighborhood U of ℘◦ in M which
is the domain of a Darboux system of coordinates (Definition 13.3.19), 𝑥1 , ..., 𝑥 𝑛 ,
𝜉1 , ..., 𝜉 𝑛 , such that d 𝜉 𝑓1 ∧ · · · ∧ d 𝜉 𝑓 𝜈 ≠ 0 at ℘◦ . We shall say that 𝐹 is of principal
type in M if 𝐹 is of principal type at every point of M.

Note that if 𝐹 is of principal type at some point of M then necessarily 𝜈 ≤ 𝑛 =

1
2 dim M. If 𝐹 is of principal type at every point of M then d 𝑓1 ∧ · · · ∧ d 𝑓 𝜈 ≠ 0 at
every point of 𝑽 (𝐹) and 𝑽 (𝐹) is a closed regular submanifold of M of codimension
𝜈.

Definition 13.4.2 The system 𝐹 shall be said to be involutive in M if there is an

open covering {U 𝜄 } 𝜄 ∈𝑰 of M such that to every 𝜄 ∈ 𝑰 and to every pair of positive
integers 𝑗 < 𝑘 ≤ 𝜈 there are regular functions 𝑎 𝑗,𝑘,ℓ (ℓ = 1, ..., 𝜈) in U 𝜄 such that
𝜈
∑︁

𝑓 𝑗 , 𝑓𝑘 = 𝑎 𝑗,𝑘,ℓ 𝑓ℓ (13.4.1)
ℓ=1

at every point of U 𝜄 .

We point out that Condition (13.4.1) is meaningful only at points of U 𝜄 ∩ 𝑽 (𝐹).

Indeed, every point in M\𝑽 (𝐹) has a neighborhood in which some 𝑓ℓ does not
vanish and (13.4.1) is trivially valid. Property (13.4.1) implies that the system 𝐹 is
of infinite type at every point of 𝑽 (𝐹) (Definition 13.3.28).
We introduce the Hamiltonian vector fields of the functions 𝑓 𝑗 , 𝐻 𝑓 𝑗 (Definition
13.3.4). The hypothesis that 𝐹 is of principal type implies that the vector fields 𝐻 𝑓 𝑗
are linearly independent at every point of the regular submanifold 𝑽 (𝐹). It follows
from (13.3.10) and (13.4.1) that, at every point of 𝑽 (𝐹) and for all 𝑗, 𝑘 = 1, ..., 𝜈,
𝐻 𝑓 𝑗 𝑓 𝑘 = 0, implying that the vector fields 𝐻 𝑓 𝑗 are tangent to 𝑽 (𝐹) at every point of
𝑽 (𝐹). Then (13.4.1) implies that
𝜈
∑︁
𝐻 𝑓 𝑗 , 𝐻 𝑓𝑘 = 𝑎 𝑗,𝑘,ℓ 𝐻 𝑓ℓ (13.4.2)
ℓ=1
13.4 Involutive Systems of Functions of Principal Type 415

at every point of 𝑽 (𝐹). By Frobenius’ Theorem 12.4.2, (13.4.2) implies that the
vector fields 𝐻 𝑓 𝑗 define a foliation of 𝑽 (𝐹) of rank 𝜈 (every leaf is 𝜈-dimensional).
We shall refer to the leaves of this foliation as the bicharacteristics of the system 𝐹.
There is a natural symplectic interpretation of the bicharacteristic foliation. Let
𝑇𝑽 (𝐹) denote the (complex, when K = C) tangent bundle of 𝑽 (𝐹) and for each ℘ ∈
𝑽 (𝐹) let 𝑇℘𝑽 (𝐹) ⊥𝜛 denote the orthogonal of 𝑇℘𝑽 (𝐹) in 𝑇℘ M for the fundamental
symplectic
form 𝜛 in M. If 𝑋 is a smooth vector field tangent to 𝑽 (𝐹) we have
𝜛 𝑋, 𝐻 𝑓 𝑗 = 𝑋 𝑓 𝑗 = 0 everywhere in 𝑽 (𝐹); this shows that 𝐻 𝑓1 , ..., 𝐻 𝑓 𝜈 are regular
sections of 𝑇𝑽 (𝐹) ⊥𝜛 , hence of 𝑇𝑽 (𝐹) ∩ 𝑇𝑽 (𝐹) ⊥𝜛 . We have dim 𝑇℘𝑽 (𝐹) =
2𝑛 − 𝜈 and since 𝜛 is nondegenerate dim 𝑇℘𝑽 (𝐹) + dim 𝑇℘𝑽 (𝐹) ⊥𝜛 = 2𝑛 implying
dim 𝑇℘𝑽 (𝐹) ⊥𝜛 = 𝜈. We conclude that the Hamiltonian vector fields 𝐻 𝑓1 , ..., 𝐻 𝑓 𝜈
span the whole vector subbundle 𝑇𝑽 (𝐹) ∩ 𝑇𝑽 (𝐹) ⊥𝜛 of 𝑇𝑽 (𝐹). By (13.4.2) this
vector subbundle is involutive (Definition 12.2.3) and the bicharacteristic leaves are
its integral manifolds.
In the whole of M the vector fields 𝐻 𝑓1 , ..., 𝐻 𝑓 𝜈 generate a Lie algebra 𝔤 (𝐹)
with respect to the commutation bracket; the freezing of 𝔤 (𝐹) at an arbitrary point
℘ ∈ M is a vector subspace of 𝑇℘ M of dimension ≥ 𝜈. In the analytic cases (C 𝜔
when K = R, complex-analytic when K = C) the Nagano Theorem 12.4.2 applies:
the Lie algebra 𝔤 (𝐹) define a foliation of M: every Nagano leaf of 𝔤 (𝐹) is an
immersed regular submanifold of M whose tangent space at any one of its points is
the freezing of 𝔤 (𝐹) at that point. We leave the proof of the following statement as
an exercise.

Proposition 13.4.3 Let M be an analytic manifold. If a Nagano leaf L of 𝔤 (𝐹)

intersects 𝑽 (𝐹) then it is entirely contained in 𝑽 (𝐹) [and therefore it is a leaf of
the Frobenius foliation of 𝑽 (𝐹)].

Remark 13.4.4 When K = R and “regular” means C 𝜔 we may be interested in an

extension of Definition 13.4.1 in which the requirement that 𝑽 (𝐹) be a submanifold
is dropped and the functions 𝑓1 , ..., 𝑓 𝜈 are assumed only to be complex-valued. In
this case we can extend 𝑓1 , ..., 𝑓 𝜈 holomorphically to a neighborhood of 𝑽 (𝐹) in
a complexification M C of M, in which case Definition 13.4.1 ensures that 𝑽 (𝐹)
is the intersection with M of a complex-analytic submanifold of M C to which the
preceding considerations apply.

In the remainder of this section the system 𝐹 = ( 𝑓1 , ..., 𝑓 𝜈 ) will be of principal

type and involutive.

13.4.2 Involutive systems of eikonal equations of principal type

We are now going to reason in an open subset U of the symplectic manifold M

under the following hypothesis: U is the domain of a Darboux system of coordinates
(Definition 13.3.19), 𝑥 1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 , such that
416 13 Elements of Symplectic Geometry

d 𝜉 𝑓1 ∧ · · · ∧ d 𝜉 𝑓 𝜈 ≠ 0 at every point of U. (13.4.3)

Thus 𝑓 𝑗 = 𝑓 𝑗 (𝑥, 𝜉), 𝑗 = 1, ..., 𝜈, 𝐹 (𝑥, 𝜉) = ( 𝑓1 (𝑥, 𝜉) , ..., 𝑓 𝜈 (𝑥, 𝜉)). By the eikonal
equations associated to the system 𝐹 we mean the system of first-order fully non-
linear PDEs,
𝑓 𝑗 (𝑥, 𝜑 𝑥 ) = 0, 𝑗 = 1, ..., 𝜈, (13.4.4)

𝜕𝜑 𝜕𝜑
where 𝜑 (𝑥) is a regular function in U and 𝜑 𝑥 = 𝜕𝑥 1
, ..., 𝜕𝑥 𝑛
(also denoted by 𝜕𝑥 𝜑
or ∇ 𝑥 𝜑).
In what follows, as we seek purely local results we shall be free to contract U
about one of its points, (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑽 (𝐹), as much as needed (but only finitely many
times).
It is convenient to carry out a symplectic change of (regular) coordinates, from
𝑥 𝑗 ’s and 𝜉 𝑘 ’s to 𝑥1 , ..., 𝑥 𝑚 , 𝑡 1 , ..., 𝑡 𝜈 , 𝜉1 , ..., 𝜉 𝑚 , 𝜏1 , ..., 𝜏𝜈 (𝑚 = 𝑛 − 𝜈) such that
the
𝑥 𝑗 , 𝜉 𝑗 ′ = 𝛿 𝑗, 𝑗 ′ , {𝑡 𝑘 , 𝜏𝑘′ } = 𝛿 𝑘,𝑘′ , with all other Poisson brackets between these
coordinates equal to zero, and such, moreover, that

d 𝜏 𝑓1 ∧ · · · ∧ d 𝜏 𝑓 𝜈 ≠ 0 at every point of U. (13.4.5)

What was called 𝜉 ◦ above is now (𝜉 ◦ , 𝜏 ◦ ); we assume that the central point lies on
the “planar” set 𝑡 = 0: what was called 𝑥 ◦ is now (𝑥 ◦ , 0). It is also convenient to
take U = Ω × Θ, Ω=Ω1 × Ω2 , Θ = Θ1 × Θ2 , with open sets Ω1 × Θ1 ⊂ K𝑚 𝑥 × K𝜉 ,
𝑚
◦ ◦
Ω2 ×Θ2 ⊂ K𝑡 ×K 𝜏 (the subscripts indicate what the variables are); (𝑥 , 𝜉 ) ∈ Ω1 ×Θ1 ,
𝜈 𝜈

(0, 𝜏 ◦ ) ∈ Ω2 ×Θ2 . We shall always assume that the geometries of Ω 𝑗 and Θ 𝑗 ( 𝑗 = 1, 2)

are very simple.
We may exploit (13.4.5) and, after due contractions, apply the Implicit Function
Theorem to obtain the following decomposition in U, for each 𝑗 = 1, ..., 𝜈,
𝜈
∑︁
𝑓 𝑗 (𝑥, 𝑡, 𝜉, 𝜏) = 𝐸 𝑗, 𝑗 ′ (𝑥, 𝑡, 𝜉, 𝜏) 𝜏 𝑗 ′ − 𝑞 𝑗 ′ (𝑥, 𝑡, 𝜉) , (13.4.6)
𝑗 ′ =1

with 𝐸 𝑗, 𝑗 ′ , 𝑞 𝑗 ′ regular functions and det 𝐸 𝑗, 𝑗 ′ 1≤ 𝑗, 𝑗 ′ ≤𝜈 nowhere vanishing in U.
[When K = C, i.e., when “regular” means “holomorphic” the Implicit Function
Theorem is an immediate consequence of the Weierstrass Preparation Theorem
14.3.4.] We can now write

𝑽 (𝐹) = (𝑥, 𝑡, 𝜉, 𝜏) ∈ U; 𝜏 𝑗 = 𝑞 𝑗 (𝑥, 𝑡, 𝜉) , 𝑗 = 1, ..., 𝜈 ; (13.4.7)

here 𝜏 ◦ = 𝑞 (𝑥 ◦ , 0, 𝜉 ◦ ).
By (13.4.1) we have, for all 𝑗, 𝑗 ′ = 1, ..., 𝜈 and all (𝑥, 𝑡, 𝜉, 𝜏) ∈ 𝑽 (𝐹),

𝜏 𝑗 − 𝑞 𝑗 (𝑥, 𝑡, 𝜉) , 𝜏 𝑗 ′ − 𝑞 𝑗 ′ (𝑥, 𝑡, 𝜉) = 0.
13.4 Involutive Systems of Functions of Principal Type 417

Explicitly, this reads

𝑚
𝜕𝑞 𝑗 ′ 𝜕𝑞 𝑗 ∑︁ 𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′ 𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′
− = − = 𝑞 𝑗′ , 𝑞 𝑗 . (13.4.8)
𝜕𝑡 𝑗 𝜕𝑡 𝑗 ′ 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝑥 𝑘 𝜕𝜉 𝑘
𝑘=1

We see that (13.4.4) is equivalent to the system of equations

𝜕𝜑
= 𝑞 𝑗 (𝑥, 𝑡, 𝜑 𝑥 ) , 𝑗 = 1, ..., 𝜈. (13.4.9)
𝜕𝑡 𝑗

We seek the regular solution of (13.4.9) that satisfies an initial condition

𝜑| 𝑡=0 = 𝜑◦ (𝑥) (13.4.10)

with 𝜑◦ regular in Ω1 and ∇ 𝑥 𝜑◦ (𝑥) ∈ Θ1 for all 𝑥 ∈ Ω1 as well as ∇ 𝑥 𝜑◦ (𝑥 ◦ ) = 𝜉 ◦ .

13.4.3 The Hamiltonian system and its first integrals

The coordinate projection (𝑥, 𝑡, 𝜉, 𝜏) ↦→ (𝑥, 𝑡, 𝜉) induces a regular diffeomorphism

of 𝑽 (𝐹) onto Ω × Θ1 ; we may, and shall now, restrict our analysis to Ω × Θ1 . The
pushdown of the Hamiltonian vector field of 𝑓 𝑗 ,
𝑚
𝜕 ∑︁ 𝜕𝑞 𝑗 𝜕
𝐻 𝑓𝑗 = − (𝑥, 𝑡, 𝜉) (13.4.11)
𝜕𝑡 𝑗 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘
𝑚 𝑛
∑︁ 𝜕𝑞 𝑗 𝜕 ∑︁ 𝜕𝑞 𝑗 𝜕
+ (𝑥, 𝑡, 𝜉) + (𝑥, 𝑡, 𝜉) ,
𝑘=1
𝜕𝑥 𝑘 𝜕𝜉 𝑘 ℓ=1
𝜕𝑡 ℓ 𝜕𝜏 ℓ

is the vector field

𝑚 𝑚
𝜕 ∑︁ 𝜕𝑞 𝑗 𝜕 ∑︁ 𝜕𝑞 𝑗 𝜕
𝐻𝑗 = − (𝑥, 𝑡, 𝜉) + (𝑥, 𝑡, 𝜉) . (13.4.12)
𝜕𝑡 𝑗 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝑘=1 𝜕𝑥 𝑘 𝜕𝜉 𝑘

By (13.4.2) or (13.4.8) we have, in Ω × Θ1 ,

𝐻 𝑗 , 𝐻 𝑗 ′ = 0, 1 ≤ 𝑗 < 𝑗 ′ ≤ 𝜈.

(13.4.13)
We now proceed to select a complete system of first integrals 𝑍 𝑘 (𝑥, 𝑡, 𝜉),
Ξℓ (𝑥, 𝑡, 𝜉) (1 ≤ 𝑘, ℓ ≤ 𝑚) of the system of vector fields (13.4.12) in Ω×Θ1 . To do this
we carry out a change of variables (𝑥1 , ..., 𝑥 𝑚 , 𝜉1 ., ..., 𝜉 𝑚 ) ⇝ (𝑦 1 , ..., 𝑦 𝑚 , 𝜂1 , ..., 𝜂 𝑚 )
to “rectify” the vector fields (13.4.12). Consider the Hamilton–Jacobi system of
ODEs,
418 13 Elements of Symplectic Geometry

𝜕𝑥𝑖 𝜕𝑞 𝑗 𝜕𝜉𝑖 𝜕𝑞 𝑗
=− (𝑥, 𝑡, 𝜉) , = (𝑥, 𝑡, 𝜉) , (13.4.14)
𝜕𝑡 𝑗 𝜕𝜉𝑖 𝜕𝑡 𝑗 𝜕𝑥𝑖
𝑖 = 1, ..., 𝑚, 𝑗 = 1, ..., 𝜈,

passing through (𝑦, 𝜂) at 𝑡 = 0. The system (13.4.14) of 2𝑛 ODEs is locally solvable

thanks to the following

Lemma 13.4.5 The right-hand sides in the system of equations (13.4.14) satisfy the
natural compatibility conditions.

Proof Applying (13.4.8) yields, for each 𝑖 = 1, ..., 𝑚, and all 𝑗, 𝑗 ′ = 1, ..., 𝜈,
!
𝜕2 𝑞 𝑗′ 𝜕2𝑞 𝑗 𝑚
∑︁ 𝜕𝑞 𝑗 𝜕 2 𝑞 𝑗 ′ 𝜕𝑞 𝑗 𝜕 2 𝑞 𝑗 ′
− = −
𝜕𝑡 𝑗 𝜕𝜉𝑖 𝜕𝑡 𝑗 ′ 𝜕𝜉𝑖 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝜉𝑖 𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝜉𝑖
!
𝑚
∑︁ 𝜕𝑞 𝑗 ′ 𝜕 2 𝑞 𝑗 𝜕𝑞 𝑗 ′ 𝜕 2 𝑞 𝑗
+ − ;
𝑘=1
𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝜉𝑖 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝜉𝑖

!
𝜕2 𝑞 𝑗′ 𝜕2𝑞 𝑗 𝑚
∑︁ 𝜕𝑞 𝑗 𝜕 2 𝑞 𝑗 ′ 𝜕𝑞 𝑗 𝜕 2 𝑞 𝑗 ′
− = −
𝜕𝑡 𝑗 𝜕𝑥𝑖 𝜕𝑡 𝑗 ′ 𝜕𝑥𝑖 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝑥𝑖 𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝑥 𝑖
!
𝑚
∑︁ 𝜕𝑞 𝑗 ′ 𝜕 2 𝑞 𝑗 𝜕𝑞 𝑗 ′ 𝜕 2 𝑞 𝑗
+ − .
𝑘=1
𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝑥 𝑖 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝑥 𝑖

𝜕
We let 𝜕𝑡 𝑗 ′ act on the first set of equations in (13.4.14):
!
𝜕 2 𝑥𝑖 𝜕2𝑞 𝑗 𝑚
∑︁ 𝜕 2 𝑞 𝑗 𝜕𝑥 𝑘 𝜕 2 𝑞 𝑗 𝜕𝜉 𝑘
=− − +
𝜕𝑡 𝑗 𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗 ′ 𝜕𝜉𝑖 𝑘=1 𝜕𝑥 𝑘 𝜕𝜉𝑖 𝜕𝑡 𝑗 ′ 𝜕𝜉 𝑘 𝜕𝜉𝑖 𝜕𝑡 𝑗 ′
!
𝜕2𝑞 𝑗 𝑚
∑︁ 𝜕𝑞 𝑗 ′ 𝜕 2 𝑞 𝑗 𝜕𝑞 𝑗 ′ 𝜕 2 𝑞 𝑗
=− + −
𝜕𝑡 𝑗 ′ 𝜕𝜉𝑖 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝜉𝑖 𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝜉𝑖
!
𝜕 2 𝑞 𝑗 ′ ∑︁ 𝜕𝑞 𝑗 𝜕 2 𝑞 𝑗 ′
𝑚
𝜕𝑞 𝑗 𝜕 2 𝑞 𝑗 ′ 𝜕 2 𝑥𝑖
=− + − = .
𝜕𝑡 𝑗 𝜕𝜉𝑖 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝜕𝜉𝑖 𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝜉𝑖 𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗

2 𝜕2 𝜉𝑖
By letting 𝜕𝑡𝜕𝑗′ act on the second set of equations in (13.4.14) we get 𝜕𝑡𝜕𝑗 𝜕𝑡𝜉𝑖𝑗′ = 𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗 .□

Let us denote by 𝑥 𝑘 (𝑦, 𝑡, 𝜂), 𝜉 𝑘 (𝑦, 𝑡, 𝜂) the solutions of (13.4.14) such that
𝑥 𝑘 (𝑦, 0, 𝜂) = 𝑦 𝑘 , 𝜉 𝑘 (𝑦, 0, 𝜂) = 𝜂 𝑘 (𝑘 = 1, ..., 𝑚) for (𝑦, 𝜂) sufficiently close to
(𝑥 ◦ , 𝜉 ◦ ). Because of the initial values the map (𝑦, 𝑡, 𝜂) ↦→ (𝑥 (𝑦, 𝑡, 𝜂) , 𝑡, 𝜉 (𝑦, 𝑡, 𝜂))
is a regular diffeomorphism from a neighborhood of (𝑥 ◦ , 0, 𝜉 ◦ ) in K𝑛 × K𝑚
onto another such neighborhood which we can assume to contain Ω × Θ1 . The
regular diffeomorphism (𝑥, 𝑡, 𝜉) ↦→ (𝑦, 𝑡, 𝜂) transforms the vector field 𝐻 𝑗 into
𝜕
𝜕𝑡 𝑗 for each 𝑗 = 1, ..., 𝜈: if 𝑔 (𝑥, 𝑡, 𝜉) is a regular function in Ω × Θ1 and if
13.4 Involutive Systems of Functions of Principal Type 419

𝑔˜ (𝑦, 𝑡, 𝜂) = 𝑔 (𝑥 (𝑦, 𝑡, 𝜂) , 𝑡, 𝜉 (𝑦, 𝑡, 𝜂)) is the pullback of 𝑔 then, by (13.4.14),

𝜕 𝑔˜
(𝑦, 𝑡, 𝜂) = 𝐻 𝑗 𝑔 (𝑥 (𝑦, 𝑡, 𝜂) , 𝑡, 𝜉 (𝑦, 𝑡, 𝜂)) .
𝜕𝑡 𝑗

In the coordinates 𝑦 1 , ..., 𝑦 𝑚 , 𝜂1 , ..., 𝜂 𝑚 the integral manifolds of the system of vector
fields (𝐻1 , ..., 𝐻 𝜈 ) in a neighborhood of (𝑥 ◦ , 0, 𝜉 ◦ ) are defined by equations 𝑦 =
const., 𝜂 = const.
At this stage we define 𝑍 𝑘 (resp., Ξℓ ) as the pullback of 𝑦 𝑘 (resp., 𝜂ℓ ) under the
map (𝑥, 𝑡, 𝜉) ↦→ (𝑦, 𝑡, 𝜂); we have 𝐻 𝑗 𝑍 𝑘 ≡ 𝐻 𝑗 Ξℓ ≡ 0 in Ω × Θ1 for every 𝑗 = 1, ..., 𝜈,
and
∀ (𝑥, 𝜉) ∈ Ω1 × Θ1 , 𝑍 𝑘 (𝑥, 0, 𝜉) = 𝑥 𝑘 , Ξℓ (𝑥, 0, 𝜉) = 𝜉ℓ . (13.4.15)
Note that (13.4.15) has the following obvious implication: the Jacobian matrix
satisfies the following “initial” condition

𝜕 (𝑍, Ξ)
= 𝐼2𝑚 . (13.4.16)
𝜕 (𝑥, 𝜉) 𝑡=0

Proposition 13.4.6 If Ω and Θ1 are sufficiently small then the map

(𝑥, 𝑡, 𝜉) ↦→ (𝑍 (𝑥, 𝑡, 𝜉) , 𝑡, Ξ (𝑥, 𝑡, 𝜉))

is a regular diffeomorphism of Ω×Θ1 onto a neighborhood of (𝑥 ◦ , 0, 𝜉 ◦ ) in K𝑛 ×K𝑚 .

Moreover, for each 𝑡 ∈ Θ2 ,

𝑍1 (𝑥, 𝑡, 𝜉) , ..., 𝑍 𝑚 (𝑥, 𝑡, 𝜉) , Ξ1 (𝑥, 𝑡, 𝜉) , ..., Ξ𝑚 (𝑥, 𝑡, 𝜉) ,

form a Darboux system of coordinates in Ω1 × Θ1 .

Proof The diffeomorphism claim is an immediate consequence of (13.4.16). We

have, at every point (𝑥, 𝑡, 𝜉) ∈ Ω × Θ1 ,

{𝑍 𝑘 , Ξℓ } = 𝛿 𝑘,ℓ , {𝑍 𝑘 , 𝑍ℓ } = {Ξ 𝑘 , Ξℓ } = 0, 1 ≤ 𝑘, ℓ ≤ 𝑚,

since this is true when 𝑡 = 0 and

𝐻 𝑗 {𝑍 𝑘 , Ξℓ } = 𝐻 𝑗 {𝑍 𝑘 , 𝑍ℓ } = 𝐻 𝑗 {Ξ 𝑘 , Ξℓ } = 0, 𝑗 = 1, ..., 𝜈. □

The regular functions 𝑍 𝑘 (𝑥, 𝑡, 𝜉), Ξℓ (𝑥, 𝑡, 𝜉) (1 ≤ 𝑘, ℓ ≤ 𝑚) in Ω × Θ1 form the

sought complete system of first integrals of the system of vector fields (𝐻1 , ..., 𝐻 𝜈 ).
We shall use the notation 𝑍 = (𝑍1 , ..., 𝑍 𝜈 ), Ξ = (Ξ1 , ..., Ξ𝜈 ). From their definition
we deduce immediately the following:

Proposition 13.4.7 After adjustment of the sizes and geometries of Ω1 and Θ1 the
functions 𝑍 𝑘 and Ξℓ (1 ≤ 𝑘, ℓ ≤ 𝑚) are constant on every integral manifold of
(𝐻1 , ..., 𝐻 𝜈 ) in Ω × Θ1 . These manifolds are the subsets of Ω × Θ1 in which every
function 𝑍 𝑘 and Ξℓ is constant and which are maximal for this property.
420 13 Elements of Symplectic Geometry

The functions 𝑍 𝑘 (𝑥, 𝑡, 𝜉), Ξℓ (𝑥, 𝑡, 𝜉) can be lifted to 𝑽 (𝐹) ∩ U (U = Ω × Θ)

as functions independent of 𝜏. If Λ (𝑥 ∗ , 𝜉 ∗ ) is the integral manifold of (𝐻1 , ..., 𝐻 𝜈 )
through the point (𝑥 ∗ , 0, 𝜉 ∗ ) ∈ Ω1 × {0} × Θ1 then 𝑍 ≡ 𝑥 ∗ , Ξ ≡ 𝜉 ∗ on Λ (𝑥 ∗ , 𝜉 ∗ ) and
also on the image of Λ (𝑥 ∗ , 𝜉 ∗ ) under the map (𝑥, 𝑡, 𝜉) ↦→ (𝑥, 𝑡, 𝜉, 𝑞 (𝑥, 𝑡, 𝜉)); the said
image is a bicharacteristic leaf of the system 𝐹 = ( 𝑓1 , ..., 𝑓 𝜈 ) in U or, equivalently,
an open and connected subset of a bicharacteristic leaf L (𝑥 ∗ , 𝜉 ∗ ) of 𝐹 in the whole
of 𝑽 (𝐹). For each 𝑘, ℓ = 1, ..., 𝑚, we can define functions 𝑍 𝑘 , Ξℓ in a neighborhood
of L (𝑥 ◦ , 𝜉 ◦ ), precisely in the set made up of all the bicharacteristic leaves of 𝐹 that
intersect U, where the values of 𝑍 𝑘 , Ξℓ are unambiguously determined. This does
not mean, however, that a complete system of first integrals 𝑍 𝑘 , Ξℓ can be defined
globally in 𝑽 (𝐹).

Example 13.4.8 Let M = R2𝑥 \ (0, 0) × R2𝜉 \ (0, 0) with the symplectic structure
induced by 𝑇 ∗ R2𝑥 \ (0, 0) ; the function 𝑓 (𝑥, 𝜉) = 𝜉1 𝑥2 − 𝜉2 𝑥1 is of principal type. If

we use polar coordinates 𝜈, 𝜃 in R2𝑥 \ (0, 0), 𝜌, 𝜔 in R2𝜉 \ (0, 0), we have 𝐻 𝑓 = 𝜕𝜔 𝜕 𝜕
− 𝜕𝜃 .
Smooth solutions of 𝐻 𝑓 ℎ = 0 in the whole of M will all be of the form ℎ (𝜈, 𝜌, 𝜃 + 𝜔);
the vector subbundle of 𝑇 ∗ M spanned by their differentials has rank 3.

A more extreme example is provided in Subsection 12.2.2.

Example 13.4.9 Let M = 𝑇 ∗ S2 \0; we denote by 𝑥1 , 𝑥2 ∈ S1 the coordinates in S2 ,

and 𝜉1 ,𝜉2 ∈ R those in 𝑇𝑥∗ S2 \ {(0, 0)}. The function 𝑓 (𝜉) = 𝜉1 − 𝑎𝜉2 (𝑎 ∈ R) is triv-
ially of principal type; 𝐻 𝑓 = 𝜕𝑥𝜕 1 −𝑎 𝜕𝑥𝜕 2 and 𝑽 (𝐹) = (𝑥, 𝜉) ∈ 𝑇 ∗ S2 \0; 𝜉1 = 𝑎𝜉2 . If
𝑎 ∉ Q every integral curve of 𝐻 𝑓 in S2 is dense in S2 and therefore every 𝑍 ∈ C 1 S2

solution of 𝐻 𝑓 𝑍 = 0 is a constant.

13.4.4 Quasilinear systems of first-order PDEs associated to the

eikonal equations

We continue to let (𝑥, 𝑡) vary in Ω, a domain in K𝑛 ; 𝜉 varies in Θ1 , a domain in K𝑚 ,

𝑚 = 𝑛 − 𝜈. We denote by 𝐿 𝑗 the base projection of the vector field 𝐻 𝑗 [see (13.4.12)];
thus
𝑚
∑︁ 𝜕𝑞 𝑗 𝜕
𝐻𝑗 = 𝐿 𝑗 + (𝑥, 𝑡, 𝜉) . (13.4.17)
𝑘=1
𝜕𝑥 𝑘 𝜕𝜉 𝑘
Throughout this subsection 𝑢 = (𝑢 1 , ..., 𝑢 𝑚 ) : Ω −→ Θ1 shall be a regular map; we
shall use the notation
𝑚
𝜕 ∑︁ 𝜕𝑞 𝑗 𝜕
𝐿 𝑢𝑗 = 𝐿 𝑗 𝜉 =𝑢
= − (𝑥, 𝑡, 𝑢) (13.4.18)
𝜕𝑡 𝑗 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘

where 𝑢 = 𝑢 (𝑥, 𝑡).

13.4 Involutive Systems of Functions of Principal Type 421

Lemma 13.4.10 If 𝜒 is a regular function in Ω × Θ1 then, for every 𝑗 = 1, ..., 𝜈,

𝐿 𝑢𝑗 ( 𝜒 (𝑥, 𝑡, 𝑢)) = 𝐻 𝑗 𝜒 (𝑥, 𝑡, 𝑢)
𝑚
∑︁ 𝜕𝜒 𝜕𝑞 𝑗
+ (𝑥, 𝑡, 𝑢) 𝐿 𝑢𝑗 𝑢 𝑘 (𝑥, 𝑡) − (𝑥, 𝑡, 𝑢) .
𝑘=1
𝜕𝜉 𝑘 𝜕𝑥 𝑘

Proof Since
𝑚
∑︁ 𝜕𝜒
𝐿 𝑢𝑗 ( 𝜒 (𝑥, 𝑡, 𝑢)) = 𝐿 𝑗 𝜒 (𝑥, 𝑡, 𝑢) + (𝑥, 𝑡, 𝑢) 𝐿 𝑢𝑗 𝑢 𝑘 (𝑥, 𝑡)
𝑘=1
𝜕𝜉 𝑘

the claim follows from (13.4.17)–(13.4.18). □

Corollary 13.4.11 If 𝜒 ∈ O (Ω × Θ1 ) is a solution of the system of equations 𝐻 𝑗 𝜒 =

0, 𝑗 = 1, ..., 𝜈, in Ω × Θ1 , then, for the same 𝑗’s and all (𝑥, 𝑡) ∈ Ω,
𝑚
∑︁ 𝜕𝜒 𝜕𝑞 𝑗
𝐿 𝑢𝑗 ( 𝜒 (𝑥, 𝑡, 𝑢)) = (𝑥, 𝑡, 𝑢) 𝐿 𝑢𝑗 𝑢 𝑘 (𝑥, 𝑡) − (𝑥, 𝑡, 𝑢) . (13.4.19)
𝑘=1
𝜕𝜉 𝑘 𝜕𝑥 𝑘

Corollary 13.4.12 If 𝑢 : Ω −→ Θ1 satisfies the system of quasilinear PDEs,

𝜕𝑞 𝑗
𝐿 𝑢𝑗 𝑢 𝑖 (𝑥, 𝑡) = (𝑥, 𝑡, 𝑢) , (13.4.20)
𝜕𝑥𝑖
𝑖 = 1, ..., 𝑚, 𝑗 = 1, ..., 𝜈,

then, for every 𝜒 ∈ O (Ω × Θ1 ), 𝑗 = 1, ..., 𝜈, and all (𝑥, 𝑡) ∈ Ω,

𝐿 𝑢𝑗 ( 𝜒 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡))) = 𝐻 𝑗 𝜒 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) . (13.4.21)

We return to the first integrals 𝑍 𝑘 (𝑥, 𝑡, 𝜉), Ξℓ (𝑥, 𝑡, 𝜉) of the system of vector fields
(13.4.12) determined by the initial values (13.4.15). In the statements that follow we
assume, as before, that the size and geometry of Ω and Θ1 are appropriately adjusted.

Proposition 13.4.13 If 𝑢 satisfies (13.4.20) then the functions 𝑍𝑖 (𝑥, 𝑡, 𝑢) (𝑖 =

1, ..., 𝑚) form a complete system of first integrals of the system of vector fields
𝐿 1𝑢 , ..., 𝐿 𝑢𝜈 in Ω.

Proof The equations

𝐿 𝑢𝑗 (𝑍𝑖 (𝑥, 𝑡, 𝑢)) = 0, 𝑖 = 1, ..., 𝑚, 𝑗 = 1, ..., 𝜈, (13.4.22)

follow from Corollary 13.4.12; (13.4.16) implies 𝜕𝑥 𝑍1 ∧ · · · ∧ 𝜕𝑥 𝑍 𝑚 ≠ 0 at every

point of Ω (Ω suitably small). □
h i
Corollary 13.4.14 If 𝑢 satisfies (13.4.20) then 𝐿 𝑢𝑗 , 𝐿 𝑢𝑗′ ≡ 0 for all 𝑗, 𝑗 ′ = 1, ..., 𝜈.
422 13 Elements of Symplectic Geometry

Proof The vector fields 𝐿 𝑢𝑗 are completely characterized by (13.4.22) and the fact
that 𝐿 𝑢𝑗 𝑡 𝑘 = 𝛿 𝑗,𝑘 for all 𝑗, 𝑘 = 1, ..., 𝜈. This implies that they form an involutive
h i
system which in turn implies that they commute since 𝐿 𝑢𝑗 , 𝐿 𝑢𝑗′ does not involve
𝜕
any partial derivative 𝜕𝑡𝑘 . □

Corollary 13.4.15 If 𝑢 satisfies (13.4.20) then

𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′
𝐿 𝑢𝑗′ (𝑥, 𝑡, 𝑢) = 𝐿 𝑢𝑗 (𝑥, 𝑡, 𝑢) (13.4.23)
𝜕𝑥𝑖 𝜕𝑥𝑖

for all 𝑖 = 1, ..., 𝑚, 𝑗, 𝑗 ′ = 1, ..., 𝜈.

Proof Letting 𝐿 𝑢𝑗′ act on the differential equations (13.4.20) and taking Corollary
13.4.14 into account yields immediately the compatibility conditions (13.4.23). □

Proposition 13.4.16 For 𝑢 : Ω −→ Θ1 to be a solution of the system of equations

(13.4.20) in Ω it is necessary and sufficient that, for all (𝑥, 𝑡) ∈ Ω,

𝐿 𝑢𝑗 (Ξ𝑖 (𝑥, 𝑡, 𝑢)) = 0, 𝑖 = 1, ..., 𝑚, 𝑗 = 1, ..., 𝜈. (13.4.24)

Proof The necessity of the condition is a direct consequence of Corollary 13.4.12.

If (13.4.24) holds then Corollary 13.4.11 implies
𝑚
∑︁ 𝜕Ξ𝑖 𝑢 𝜕𝑞 𝑗
(𝑥, 𝑡, 𝑢) 𝐿 𝑗 𝑢 𝑘 (𝑥, 𝑡) − (𝑥, 𝑡, 𝑢) = 0.
𝑘=1
𝜕𝜉 𝑘 𝜕𝑥 𝑘

If Ω is sufficiently small then, by (13.4.16), the matrix

𝜕Ξ 𝜕Ξ𝑖
(𝑥, 𝑡, 𝑢) = (𝑥, 𝑡, 𝑢)
𝜕𝜉 𝜕𝜉 𝑘 1≤𝑖,𝑘 ≤𝑚

is nonsingular, whence (13.4.20). □

Proposition 13.4.17 If Ω and Θ1 are sufficiently small then, for 𝑢 : Ω −→ Θ1 to be

a solution of the system of equations (13.4.20) in Ω it is necessary and sufficient that

∀ (𝑥, 𝑡) ∈ Ω, Ξ (𝑥, 𝑡, 𝑢) = 𝑢 (𝑍 (𝑥, 𝑡, 𝑢) , 0) .

Proof Proposition 13.4.13 and 13.4.16 have the following consequence: there exist
an open subset 𝑈1 of K𝑚 containing the closure of the range of the map Ω × Θ1 ∋
(𝑥, 𝑡, 𝜉) ↦→ 𝑍 (𝑥, 𝑡, 𝜉) ∈ K𝑚 and a regular map 𝑈1 ∋ 𝑧 ↦→ 𝑔 (𝑧) ∈ K𝑚 whose range
is contained in that of the map Ω × Θ1 ∋ (𝑥, 𝑡, 𝜉) ↦→ Ξ (𝑥, 𝑡, 𝜉) ∈ K𝑚 , such that

Ξ (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) = 𝑔 (𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡))) (13.4.25)

for all (𝑥, 𝑡) ∈ Ω. By (13.4.15) putting 𝑡 = 0 in (13.4.25) yields 𝑢 𝑖 (𝑥, 0) = 𝑔𝑖 (𝑥) for
all 𝑖 = 1, ..., 𝑚 and 𝑥 ∈ 𝑈1 . □
13.4 Involutive Systems of Functions of Principal Type 423

We can rephrase Proposition 13.4.17 by putting the initial value on the forefront.
Regarding 𝑥 as a parameter we apply the Implicit Function Theorem: the equations
(13.4.25) have a unique regular solution 𝑢 (𝑥, 𝑡) in Ω (suitably contracted) such that
𝑢 (𝑥, 0) = 𝑔 (𝑥). We can state:

Proposition 13.4.18 The regular solution 𝑢 of (13.4.25) such that 𝑢 (𝑥, 0) = 𝑔 (𝑥) is
identical to the solution of the Cauchy problem
𝜕𝑞 𝑗
𝐿 𝑢𝑗 𝑢 𝑖 (𝑥, 𝑡) = (𝑥, 𝑡, 𝑢) , 𝑢 𝑖 (𝑥, 0) = 𝑔𝑖 (𝑥) , (13.4.26)
𝜕𝑥 𝑖
𝑖 = 1, ..., 𝑚, 𝑗 = 1, ..., 𝜈.

The next statement connects the results in this subsection to the considerations at
the end of the preceding subsection.

Proposition 13.4.19 Let 𝑢 be a regular solution of (13.4.20) in Ω and (𝑥 ∗ , 0, 𝜉 ∗ ) ∈

Ω × Θ1 be such that 𝜉 ∗ = 𝑢 (𝑥 ∗ , 0). The integral manifold of (𝐻1 , ..., 𝐻 𝜈 ) through
the point (𝑥 ∗ , 0, 𝜉 ∗ ), Λ (𝑥 ∗ , 𝜉 ∗ ), is described by (𝑥 (𝑥 ∗ , 𝑡) , 𝑡, 𝜉 (𝑥 ∗ , 𝑡)), where 𝑥 (𝑥 ∗ , 𝑡)
is the regular solution of the equation 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) = 𝑥 ∗ such that 𝑥 (𝑥 ∗ , 0) = 𝑥 ∗ ,
and 𝜉 (𝑥 ∗ , 𝑡) = 𝑢 (𝑥 (𝑥 ∗ , 𝑡) , 𝑡).

Proof The existence and uniqueness of the regular solutions 𝑥 𝑖 = 𝑥 𝑖 (𝑥 ∗ , 𝑡) (𝑖 =

1, ..., 𝑚) of the equation 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) = 𝑥 ∗ follow from the initial conditions
(13.4.15)–(13.4.16). Likewise for the regular solutions 𝜉𝑖 = 𝜉𝑖 (𝑥 ∗ , 𝑡) (𝑖 = 1, ..., 𝑚)
of the equations

Ξ𝑖 (𝑥 (𝑥 ∗ , 𝑡) , 𝑡, 𝜉 (𝑥 ∗ , 𝑡)) = 𝑢 𝑖 (𝑥 ∗ , 0) , 𝑖 = 1, ..., 𝑚.

The same equation is satisfied by 𝑢 (𝑥 (𝑥 ∗ , 𝑡) , 𝑡) as seen by putting 𝑥 = 𝑥 (𝑥 ∗ , 𝑡) in

(13.4.25). Since Ξ𝑖 (𝑥 (𝑥 ∗ , 0) , 0, 𝜉 (𝑥 ∗ , 0)) = 𝜉𝑖 (𝑥 ∗ , 0) the claim ensues. □

13.4.5 Solving the eikonal equations

In this subsection we show how to construct the (unique) regular solution of the
problem (13.4.9)–(13.4.10). Keep in mind that 𝑥 varies in Ω1 and 𝑡 in Ω2 ; Ω =
Ω1 × Ω2 .
If we let 𝜕𝑥 𝑗 (1 ≤ 𝑖 ≤ 𝑚) act on (13.4.9) we see that 𝑢 = 𝜕𝑥 𝜑 is the regular
solution of the Cauchy problem

𝐿 𝑢𝑗 𝑢 𝑖 (𝑥, 𝑡) = 𝜕𝑥𝑖 𝑞 𝑗 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) , (13.4.27)
𝑖 = 1, ..., 𝑚, 𝑗 = 1, ..., 𝜈,

𝑢 𝑖 (𝑥, 0) = 𝜕𝑥𝑖 𝜑◦ (𝑥) , 𝑖 = 1, ..., 𝑚. (13.4.28)

We need some properties of the solutions of the system of equations (13.4.20).
424 13 Elements of Symplectic Geometry

Proposition 13.4.20 Let 𝑣 = (𝑣 1 , ..., 𝑣 𝜈 ) be a regular solution of the system (13.4.20)

in Ω. If the one-form 𝑣 (𝑥, 𝑡) · d𝑥 in Ω is 𝜕𝑥 -closed when 𝑡 = 0 then it is 𝜕𝑥 -closed for
all 𝑡 ∈ Ω2 .
𝜕 𝜕𝑣𝑖
Proof Let 𝜕𝑥𝑖′ act on the differential equation in (13.4.20); in the notation 𝑣 𝑖,𝑖′ = 𝜕𝑥𝑖′
we get

𝜕𝑣 𝑖,𝑖′ ∑︁ 𝜕 2 𝑞 𝑗 𝜕𝑣 𝑖 ∑︁ 𝜕 2 𝑞 𝑗
𝑚 𝑚
𝜕𝑣 𝑖′
= (𝑥, 𝑡, 𝑣) + (𝑥, 𝑡, 𝑣)
𝜕𝑡 𝑗 𝑘=1
𝜕𝑥𝑖′ 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝑘=1 𝜕𝑥 𝑖 𝜕𝜉 𝑘 𝜕𝑥 𝑘
𝑚
∑︁ 𝜕2𝑞 𝑗 𝜕𝑣 𝑖 𝜕𝑣 𝑖′ 𝜕2𝑞 𝑗
+ (𝑥, 𝑡, 𝑣) + (𝑥, 𝑡, 𝑣) .
𝑘,ℓ=1
𝜕𝜉 𝑘 𝜕𝜉ℓ 𝜕𝑥 𝑘 𝜕𝑥ℓ 𝜕𝑥𝑖 𝜕𝑥𝑖′

𝜕𝑣𝑖,𝑖′
The right-hand side is symmetric with respect to (𝑖, 𝑖 ′) implying directly 𝜕𝑡 𝑗 =
𝜕𝑣𝑖′ ,𝑖
𝜕𝑡 𝑗 , 𝑗 = 1, ..., 𝜈. Since 𝑣 𝑖,𝑖′ = 𝑣 𝑖′ ,𝑖 when 𝑡 = 0 the same is true for all 𝑡 ∈ Ω2 . □

Proposition 13.4.21 Let 𝑣 = (𝑣 1 , ..., 𝑣 𝜈 ) be a regular solution of the system

(13.4.20) in Ω. If the one-form 𝑣 (𝑥, 0) · d𝑥 in Ω1 is 𝜕𝑥 -closed then the one-form
Í𝜈
𝑗=1 𝑞 𝑗 (𝑥, 𝑡, 𝑣 (𝑥, 𝑡)) d𝑡 𝑗 is 𝜕𝑡 -closed in Ω.

Proof We have
𝑚
𝜕 𝜕𝑞 𝑗 ∑︁ 𝜕𝑞 𝑗 𝜕𝑣 ℓ
𝑞 𝑗 (𝑥, 𝑡, 𝑣) = (𝑥, 𝑡, 𝑣) + (𝑥, 𝑡, 𝑣) ,
𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗 ′ ℓ=1
𝜕𝜉 ℓ 𝜕𝑡 𝑗′

where 𝑣 = 𝑣 (𝑥, 𝑡). We derive from (13.4.20):

𝑚
𝜕 𝜕𝑞 𝑗 ∑︁ 𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′
𝑞 𝑗 (𝑥, 𝑡, 𝑣) = (𝑥, 𝑡, 𝑣) + (𝑥, 𝑡, 𝑣) (𝑥, 𝑡, 𝑣)
𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗 ′ ℓ=1
𝜕𝜉 ℓ 𝜕𝑥ℓ
𝑚
∑︁ 𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′ 𝜕𝑣 ℓ
+ (𝑥, 𝑡, 𝑣) (𝑥, 𝑡, 𝑣) ,
𝑘,ℓ=1
𝜕𝜉ℓ 𝜕𝜉 𝑘 𝜕𝑥 𝑘

whence
𝜕 𝜕
𝑞 𝑗 (𝑥, 𝑡, 𝑣) − 𝑞 𝑗 ′ (𝑥, 𝑡, 𝑣)
𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗

𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′
= − − 𝑞 𝑗 , 𝑞 𝑗 ′ (𝑥, 𝑡, 𝑣)
𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗
𝑚 𝑚
∑︁ 𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′ 𝜕𝑣 ℓ ∑︁ 𝜕𝑞 𝑗 ′ 𝜕𝑞 𝑗 𝜕𝑣 ℓ
+ (𝑥, 𝑡, 𝑣) (𝑥, 𝑡, 𝑣) − (𝑥, 𝑡, 𝑣) (𝑥, 𝑡, 𝑣) .
𝑘,ℓ=1
𝜕𝜉 ℓ 𝜕𝜉 𝑘 𝜕𝑥 𝑘 𝑘,ℓ=1
𝜕𝜉 ℓ 𝜕𝜉 𝑘 𝜕𝑥 𝑘

In the last sum we can exchange the indices 𝑘 and ℓ; since 𝑣 (𝑥, 𝑡) · d𝑥 is 𝜕𝑥 -closed
(by Proposition 13.4.20) we get
13.4 Involutive Systems of Functions of Principal Type 425

𝜕 𝜕
𝑞 𝑗 (𝑥, 𝑡, 𝑣) − 𝑞 𝑗 ′ (𝑥, 𝑡, 𝑣)
𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗

𝜕𝑞 𝑗 𝜕𝑞 𝑗 ′
= − − 𝑞 𝑗 , 𝑞 𝑗 ′ (𝑥, 𝑡, 𝑣) ,
𝜕𝑡 𝑗 ′ 𝜕𝑡 𝑗

which vanishes identically by (13.4.8). □

In the following statement “local” means “in a neighborhood of (𝑥 ◦ , 0)”; the
precise terminology would involve germs of regular functions at (𝑥 ◦ , 0). We assume
that the open sets Ω1 ⊂ K𝑚 and Ω2 ⊂ K𝜈 are homeomorphic to open balls.

Theorem 13.4.22 If 𝑢 is the local regular solution of the Cauchy problem (13.4.27)–
(13.4.28) then the local regular solution of the Cauchy problem (13.4.9)–(13.4.10)
is given by
∫ 𝑡 𝜈
∑︁
𝜑 (𝑥, 𝑡) = 𝜑◦ (𝑥) + 𝑞 𝑗 (𝑥, 𝑡 ′, 𝑢 (𝑥, 𝑡 ′)) d𝑡 ′𝑗 , (13.4.29)
0 𝑗=1

where the integration with respect to 𝑡 ′ is carried out on an arbitrary simple smooth
curve in Ω2 joining 0 to 𝑡. We also have
∫ 𝑥
◦
𝜑 (𝑥, 𝑡) = 𝜑◦ (𝑥 ) + 𝑢 (𝑥 ′, 𝑡) · d𝑥 ′ (13.4.30)
𝑥◦
𝜈 ∫
∑︁ 𝑡
+ 𝑞 𝑗 (𝑥 ◦ , 𝑡 ′, 𝑢 (𝑥 ◦ , 𝑡 ′)) d𝑡 ′𝑗 .
𝑗=1 0

Proof By Proposition 13.4.13 and the Poincaré Lemma if 𝑢 satisfies (13.4.27)–

(13.4.28) there are regular functions 𝜓 in a neighborhood of (𝑥 ◦ , 0) in Ω such that
𝜕𝑥 𝜓 = 𝑢 · 𝜕𝑥; 𝜑 is one of them:
∫ 𝑥
𝜑 (𝑥, 𝑡) = 𝜑1 (𝑡) + 𝑢 (𝑥 ′, 𝑡) · d𝑥 ′, (13.4.31)
𝑥◦

where the integration is carried out on an arbitrary simple smooth curve in Ω1 joining
𝑥 ◦ to 𝑥. We derive from (13.4.27)–(13.4.28) and (13.4.31):
𝑚 ∫ 𝑥
∑︁ 𝜕𝑢 𝑘 ′
𝜕𝑡 𝑗 𝜑 (𝑥, 𝑡) = 𝜕𝑡 𝑗 𝜑1 (𝑡) + (𝑥 , 𝑡) d𝑥 𝑘′
𝑘=1 𝑥◦ 𝜕𝑡 𝑗
𝑚 ∫ 𝑥
∑︁ 𝜕𝑞 𝑗 ′
= 𝜕𝑡 𝑗 𝜑1 (𝑡) + (𝑥 , 𝑡, 𝑢 (𝑥 ′, 𝑡)) d𝑥 𝑘′
𝑘=1 𝑥◦ 𝜕𝑥 𝑘
𝑚 ∫ 𝑥
∑︁ 𝜕𝑞 𝑗 ′ 𝜕𝑢 𝑘 ′
+ (𝑥 , 𝑡, 𝑢 (𝑥 ′, 𝑡)) (𝑥 , 𝑡) d𝑥 𝑘′ .
𝑘,ℓ=1 𝑥◦ 𝜕𝜉ℓ 𝜕𝑥ℓ
426 13 Elements of Symplectic Geometry

We have
𝜕𝑞 𝑗 ′ 𝜕
(𝑥 , 𝑡, 𝑢 (𝑥 ′, 𝑡)) = 𝑞 𝑗 (𝑥 ′, 𝑡, 𝑢 (𝑥 ′, 𝑡))

𝜕𝑥 𝑘 𝜕𝑥 𝑘′

𝑚
∑︁ 𝜕𝑞 𝑗 ′ 𝜕𝑢 ℓ ′
− (𝑥 , 𝑡, 𝑢 (𝑥 ′, 𝑡)) (𝑥 , 𝑡) ,
ℓ=1
𝜕𝜉 ℓ 𝜕𝑥 𝑘

whence

𝜕𝑡 𝑗 𝜑 (𝑥, 𝑡) − 𝜕𝑡 𝑗 𝜑1 (𝑡) = 𝑞 𝑗 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) − 𝑞 𝑗 (𝑥 ◦ , 𝑡, 𝑢 (𝑥 ◦ , 𝑡))

𝑚 ∫ 𝑥
∑︁ 𝜕𝑞 𝑗 ′ ′ 𝜕𝑢 𝑘 ′ 𝜕𝑢 ℓ ′
+ (𝑥 , 𝑡, 𝑢 (𝑥 , 𝑡)) (𝑥 , 𝑡) − (𝑥 , 𝑡) d𝑥 𝑘′ .
𝑥 ◦ 𝜕𝜉 ℓ 𝜕𝑥 ℓ 𝜕𝑥 𝑘
𝑘,ℓ=1

Since 𝑢 = 𝜕𝑥 𝜑 the integrals in the preceding equation vanish identically, whence

𝜕𝑡 𝑗 𝜑 (𝑥, 𝑡) − 𝜕𝑡 𝑗 𝜑1 (𝑡) = 𝑞 𝑗 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) − 𝑞 𝑗 (𝑥 ◦ , 𝑡, 𝑢 (𝑥 ◦ , 𝑡)) . (13.4.32)

In (13.4.31) we require

𝜕𝑡 𝑗 𝜑1 (𝑡) = 𝑞 𝑗 (𝑥 ◦ , 𝑡, 𝑢 (𝑥 ◦ , 𝑡)) , 𝑗 = 1, ..., 𝜈, (13.4.33)

as well as 𝜑1 (0) = 𝜑◦ (𝑥 ◦ ). Putting this into (13.4.32) shows that 𝜑 satisfies (13.4.9)
as well as (13.4.29); putting it into (13.4.31) proves (13.4.30). Both (13.4.29) and
(13.4.30) imply (13.4.10). □

Corollary 13.4.23 The differential operator 𝜕𝑥 = 𝜕𝑥1 , ..., 𝜕𝑥𝑚 induces a bijection
of the set of local regular solutions 𝜑 of the Cauchy problem (13.4.9)–(13.4.10)
onto the set of local regular solutions 𝑢 = (𝑢 1 , ..., 𝑢 𝑚 ) of the Cauchy problem
(13.4.27)–(13.4.28).

Proof We have seen that 𝑢 = 𝜕𝑥 𝜑 solves (13.4.27)–(13.4.28) if 𝜑 solves (13.4.9)–

(13.4.10). And if 𝑢 solves (13.4.27)–(13.4.28) then 𝜑 is given by (13.4.29). □

Corollary 13.4.24 The local regular solution of the Cauchy problem (13.4.27)–
(13.4.28) is identical to the regular solution 𝑢 (𝑥, 𝑡) of the system of equations

Ξ𝑖 (𝑥, 𝑡, 𝑢) = 𝜕𝑥𝑖 𝜑◦ (𝑍 (𝑥, 𝑡, 𝑢)) , 𝑖 = 1, ..., 𝑚, (13.4.34)

such that 𝑢 (𝑥, 0) = 𝜕𝑥 𝜑◦ (𝑥).

Proof Combine Proposition 13.4.18 and Theorem 13.4.22. □

Remark 13.4.25 It is permitted to let 𝜑 depend on a parameter 𝜃 varying in an

open subset of K𝑚 (1 ≤ 𝑚 ∈ Z+ ); this does not affect the preceding reasoning;
the dependence on 𝜃 originates with the initial conditions (13.4.10), 𝜑 (𝑥, 0, 𝜃) =
𝜑◦ (𝑥, 𝜃) and 𝑢 (𝑥, 0, 𝜃) = 𝜕𝑥 𝜑◦ (𝑥, 𝜃).
13.4 Involutive Systems of Functions of Principal Type 427

13.4.6 The homogeneous case

In this subsection we take the manifold M to be a conic open subset of the cotangent
bundle 𝑇 ∗ 𝑿 of a regular manifold 𝑿 (see Section 9.3); more precisely, we take
M ⊂ 𝑇 ∗ 𝑿\0, the complement of the null section of 𝑇 ∗ 𝑿. “Conic” means invariant
under the positive dilations (𝑥, 𝜉) ↦→ (𝑥, 𝜆𝜉), 𝜆 > 0; now 𝑥 will be the variable in
the base, 𝜉 the variable in the fibers, meaning the cotangent spaces 𝑇𝑥∗ 𝑿. This allows
us to speak of homogeneous functions in M: functions 𝑓 (𝑥, 𝜉) in a conic subset U
of M such that 𝑓 (𝑥, 𝜆𝜉) = 𝜆 𝑑 𝑓 (𝑥, 𝜉) for some 𝑑 ∈ R, all (𝑥, 𝜉) ∈ U and 𝜆 > 0.
We add now a third fundamental hypothesis about the involutive system of functions
𝐹 = ( 𝑓1 , ..., 𝑓 𝜈 ) of principal type:
(Homgn) The regular functions in M, 𝑓1 , ..., 𝑓 𝜈 , are homogeneous of the same
degree 𝑑 ◦ .
Under this hypothesis we can carry out the operations of the preceding subsections
in a neighborhood of an arbitrary point (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑽 (𝐹) and then extend the results
to the conic span U of said neighborhood by homogeneity. Suppose that the base
projection 𝜋 (U) of U is contained in the domain of regular coordinates 𝑥1 , ..., 𝑥 𝑛 ,
in which case 𝜉1 , ..., 𝜉 𝑛 are the dual coordinates, i.e., the coordinates with respect
to the basis d𝑥 1 , ..., d𝑥 𝑛 in the cotangent spaces 𝑇𝑥∗ 𝑿; of course, 𝑥1 , ..., 𝑥 𝑛 ,𝜉1 , ..., 𝜉 𝑛
is a Darboux coordinates system in U for the standard symplectic structure of 𝑇 ∗ 𝑿
(the latter defined by the fundamental one-form 𝜎 = 𝜉 · d𝑥 and its exterior derivative
𝜛 = d𝜉 ∧ d𝑥; see Subsection 13.3.5). For instance, we shall assume that (13.4.3)
holds and the change of coordinates (𝑥1 , ..., 𝑥 𝑛 ) ⇝ (𝑥 1 , ..., 𝑥 𝑚 , 𝑡1 , ..., 𝑡 𝜈 ) occurs
in the base; 𝜉1 , ..., 𝜉 𝑚 , 𝜏1 , ..., 𝜏𝜈 are the dual coordinates. Thus we can restrict the
factorization (13.4.6) to the intersection of U with 𝜋 (U) × S𝑛−1 , i.e., assume that
|𝜉 | 2 + |𝜏| 2 = 1, and then extend it to the whole of U by positing that the coefficients
𝐸 𝑗,𝑘 are homogeneous of degree 𝑑 ◦ − 1 while the 𝑞 𝑗 are homogeneous of degree 1.
In this set-up the solution of the eikonal equations (13.4.9) is much simpler than in
the general case (cf. preceding subsection).
We take U = Ω × Γ1 , now with Γ1 ⊂ K𝑚 an open cone; the shape and size of Ω
and Γ1 ∩ S2𝑚−1 will be adjusted as needed. As before, 𝑍 𝑘 , Ξℓ (1 ≤ 𝑘, ℓ ≤ 𝑚) are the
first integrals of the system of Hamiltonian vector fields (13.4.12) [cf. also (13.4.17)]
defined by the initial conditions (13.4.15); 𝑍 = (𝑍1 , ..., 𝑍 𝑚 ) maps Ω × Γ1 onto an
open neighborhood of 𝑥 ◦ in K𝑚 . Also as before, 𝑢 is the local regular solution of the
Cauchy problem (13.4.27)–(13.4.28). We assume that the initial datum 𝜑◦ is defined
and regular in the range of the map Ω∋ (𝑥, 𝑡) ↦→ 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)).
Theorem 13.4.26 Suppose that 𝑞 𝑗 (𝑥, 𝑡, 𝜉) is homogeneous of degree 1 for every
𝑗 = 1, ..., 𝜈. For the regular function 𝜑 in Ω to be the local solution of the Cauchy
problem (13.4.9)–(13.4.10) it is necessary and sufficient that

∀ (𝑥, 𝑡) ∈ Ω, 𝜑 (𝑥, 𝑡) = 𝜑◦ (𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡))) . (13.4.35)

Proof The Euler homogeneity identities and Corollary 13.4.23 imply that the system
of equations (13.4.9) can be rewritten as
428 13 Elements of Symplectic Geometry
𝑚
𝜕𝜑 ∑︁ 𝜕𝑞 𝑗 𝜕𝜑
= (𝑥, 𝑡, 𝑢) , 𝑗 = 1, ..., 𝜈,
𝜕𝑡 𝑗 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘

equivalent, in the notation (13.4.18), to 𝐿 𝑢𝑗 𝜑 = 0, 𝑗 = 1, ..., 𝜈. By Proposition 13.4.13

the same equations are satisfied by 𝜑◦ (𝑍 (𝑥, 𝑡, 𝑢)). Since 𝜑◦ (𝑍 (𝑥, 𝑡, 𝑢)) also satisfies
(13.4.10) the identity (13.4.35) ensues. □

Corollary 13.4.27 Assume the same hypotheses as in Theorem 13.4.26. If 𝑥 ∗ ∈ Ω

and 𝜉 ∗ = 𝜕𝜑◦ (𝑥 ∗ ) ∈ Γ1 then 𝜑 (𝑥, 𝑡) = 𝜑◦ (𝑥 ∗ ) for all (𝑥, 𝑡) in the base projection of
the complex bicharacteristic of ( 𝑓1 , ..., 𝑓 𝜈 ) through (𝑥 ∗ , 0, 𝜉 ∗ , 𝑞 (𝑥 ∗ , 0, 𝜉 ∗ )).

Proof According to Proposition 13.4.19 the complex bicharacteristic of ( 𝑓1 , ..., 𝑓 𝜈 )

through (𝑥 ∗ , 0, 𝜉 ∗ ) consists of the points (𝑥 (𝑥 ∗ , 𝑡) , 𝑡, 𝑢 (𝑥 (𝑥 ∗ , 𝑡) , 𝑡)) where 𝑥 (𝑥 ∗ , 𝑡)
is the solution of 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) = 𝑥 ∗ such that 𝑥 (𝑥 ∗ , 0) = 𝑥 ∗ , and 𝜉 (𝑥 ∗ , 𝑡) =
𝑢 (𝑥 (𝑥 ∗ , 𝑡) , 𝑡). Combining this with (13.4.35) yields the desired result. □

Remark 13.4.28 In the absence of homogeneity of degree 1, generally 𝜑 is not

constant on every set where 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) is constant, as shown in the following
(Riccati-type) example.

Example 13.4.29 Take 𝑚 = 𝜈 = 1, 𝑞 (𝑡, 𝜉) = 21 𝜉 2 ; then 𝐻1 = 𝜕𝑡 − 𝜉𝜕𝑥 [cf. (13.4.12)]

and 𝑍 (𝑥, 𝑡, 𝜉) = 𝑥 + 𝑡𝜉; (13.4.9) and (13.4.27) read
1 2 𝜕𝑢 𝜕𝑢
𝜕𝑡 𝜑 = 𝑢 , =𝑢 .
2 𝜕𝑡 𝜕𝑥

We select the initial value 𝜑◦ (𝑥) = 21 𝑥 2 ; then 𝑢 (𝑥, 0) = 𝑥 and 𝑢 (𝑥, 𝑡) = 𝑥

1−𝑡 . It is
readily verified that
𝑥
𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡)) = ,
1−𝑡
1 𝑥2
𝜑 (𝑥, 𝑡) = = (1 − 𝑡) 𝜑◦ (𝑍 (𝑥, 𝑡, 𝑢)) .
21−𝑡
We introduce a parameter 𝜃 varying in a conic open subset Θ of Γ1 and we
call 𝜑 (𝑥, 𝑡, 𝜃) the regular solution of the Cauchy problem (13.4.9)–(13.4.10) where
𝜑◦ (𝑥, 𝜃) = 𝑥 · 𝜃. If 𝑢 (𝑥, 𝑡, 𝜃) = 𝜕𝑥 𝜑 (𝑥, 𝑡, 𝜃) then 𝑢 (𝑥, 0, 𝜃) = 𝜃; Formula (13.4.35)
yields
𝜑 (𝑥, 𝑡, 𝜃) = 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡, 𝜃)) · 𝜃 (13.4.36)
while (13.4.34) becomes
Ξ (𝑥, 𝑡, 𝑢 (𝑥, 𝑡, 𝜃)) = 𝜃. (13.4.37)
In particular, 𝑍 (𝑥, 𝑡, 𝑢 (𝑥, 𝑡, 𝜃)) is homogeneous of degree zero with respect to 𝜃.

Proposition 13.4.30 If 𝜑 ∈ O (Ω × Θ) is the solution of (13.4.9) such that 𝜑 (𝑥, 0, 𝜃) =

𝑥 · 𝜃 then

𝜕𝜃ℓ 𝜑 (𝑥, 𝑡, 𝜃) = 𝑍ℓ (𝑥, 𝑡, 𝑢 (𝑥, 𝑡, 𝜃)) , ℓ = 1, ..., 𝑚, (13.4.38)

13.5 Real and Imaginary Symplectic Structures in C2𝑛 429

for all (𝑥, 𝑡, 𝜃) ∈ Ω × Θ such that 𝑢 (𝑥, 𝑡, 𝜃) ∈ Γ1 .

Proof Letting 𝜕𝜃ℓ (ℓ = 1, ..., 𝑚) act on (13.4.9) yields (by the chain rule), for every
𝑗 = 1, ..., 𝜈,
𝑚
𝜕 ∑︁ 𝜕𝑞 𝜕
𝜕𝜃ℓ 𝜑 = (𝑥, 𝑡, 𝑢) 𝜕𝜃ℓ 𝜑 ,
𝜕𝑡 𝑗 𝑘=1
𝜕𝜉 𝑘 𝜕𝑥 𝑘

which is to say, 𝐿 𝑢𝑗 𝜕𝜃ℓ 𝜑 = 0 [cf. (13.4.18)]; moreover, 𝜕𝜃ℓ 𝜑 𝑡=0 = 𝑥ℓ . We recall
(Proposition 13.4.13) that 𝐿 𝑢𝑗 (𝑍ℓ (𝑥, 𝑡, 𝑢)) = 0, 𝑗 = 1, ..., 𝜈, and 𝑍ℓ (𝑥, 0, 𝑢 (𝑥, 0, 𝜃)) =
𝑥ℓ [cf. (13.4.15)] whence (13.4.38) by uniqueness in the Cauchy problem. □

Remark 13.4.31 Much of the content of this section is valid for C 2 functions when
K = R.

13.5 Real and Imaginary Symplectic Structures in C2𝒏

13.5.1 R and I terminology in complex Euclidean space

In this section, unless specified otherwise, by a vector (or linear) subspace of C2𝑛 𝑧,𝜁
we shall mean a real vector subspace of C2𝑛 𝑧,𝜁 . We identify the (complex) tangent and
cotangent bundles of C𝑛 with C2𝑛 𝑧,𝜁 = C 𝑧 ×C 𝜁 equipped with the complex symplectic
𝑛 𝑛

structure defined by the C-bilinear form

𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) = 𝑧 · 𝜁 ′ − 𝑧 ′ · 𝜁 (13.5.1)

We identify the symplectic linear automorphisms of C2𝑛 𝑧,𝜁 with the matrices in
2𝑛
Sp (𝑛, C) using the canonical basis of C𝑧,𝜁 .
By identifying C2𝑛 2𝑛 2𝑛
𝑧,𝜁 with R 𝑥,𝑦 × R 𝜉 , 𝜂 (𝑧 = 𝑥 + 𝑖𝑦, 𝜁 = 𝜉 + 𝑖𝜂) we see that 𝜛𝑛
C

engenders two real 2-forms on the latter:

Re 𝜛𝑛C ((𝑥, 𝑦, 𝜉, 𝜂) , (𝑥 ′, 𝑦 ′, 𝜉 ′, 𝜂 ′)) (13.5.2)

1 C
= 𝜛𝑛 ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) + 𝜛𝑛C 𝑧¯, 𝜁¯ , 𝑧 ′, 𝜁 ′
2
= 𝑥 · 𝜉 ′ − 𝑥 ′ · 𝜉 − 𝑦 · 𝜂 ′ + 𝑦 ′ · 𝜂,

Im 𝜛𝑛C ((𝑥, 𝑦, 𝜉, 𝜂) , (𝑥 ′, 𝑦 ′, 𝜉 ′, 𝜂 ′)) (13.5.3)

1 C
= 𝜛𝑛 ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) − 𝜛𝑛C 𝑧¯, 𝜁¯ , 𝑧 ′, 𝜁 ′
2𝑖
= 𝑦 · 𝜉 ′ − 𝑦 ′ · 𝜉 + 𝑥 · 𝜂 ′ − 𝑥 ′ · 𝜂.
430 13 Elements of Symplectic Geometry

Exchanging 𝑦 and −𝑦 (resp., 𝑥 and 𝑦) in (13.5.2) [resp., (13.5.3)] transforms Re 𝜛𝑛C

(resp., Im 𝜛𝑛C ) into the standard symplectic form on R2𝑛
( 𝑥,𝑦)
× R2𝑛
( 𝜉 , 𝜂)
, 𝑥 · 𝜉 ′ + 𝑦 · 𝜂′ −
(𝑥 ′ · 𝜉 + 𝑦 ′ · 𝜂). Thus each one of the 2-forms Re 𝜛𝑛C and Im 𝜛𝑛C is nondegenerate;
it is directly checked that

Re 𝜛𝑛C ((𝑧, 𝑖𝜁) , (𝑧 ′, 𝑖𝜁 ′)) = − Im 𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) , (13.5.4)

′ ′ ′ ′
Im 𝜛𝑛C ((𝑧, 𝑖𝜁) , (𝑧 , 𝑖𝜁 )) = Re 𝜛𝑛C ((𝑧, 𝜁) , (𝑧 , 𝜁 )) .

We can make use of the terminology introduced in Definition 13.1.6 for the real
symplectic structure on R2𝑛
( 𝑥,𝑦)
× R2𝑛
( 𝜉 , 𝜂)
defined by either Re 𝜛𝑛C or Im 𝜛𝑛C .

Definition 13.5.1 A real vector subspace of C2𝑛

𝑧,𝜁 is said to be
(1) R-symplectic (resp., I-symplectic),
(2) R-isotropic (resp., I-isotropic),
(3) R-coisotropic (resp., I-coisotropic),
(4) R-Lagrangian (resp., I-Lagrangian),
if the appellative is valid in the symplectic structure on R2𝑛
( 𝑥,𝑦)
× R2𝑛
( 𝜉 , 𝜂)
defined by
Re 𝜛𝑛C (resp., Im 𝜛𝑛C ).

The following properties of a real vector subspace 𝑬 of C2𝑛

𝑧,𝜁 are equivalent:

(1) 𝑬 is isotropic (resp., Lagrangian) for 𝜛𝑛C ;

(2) 𝑬 is both R-isotropic and I-isotropic (resp., R-Lagrangian and I-Lagrangian).

Remark 13.5.2 It follows from (13.5.4) that the map (𝑧, 𝜁) ↦→ (𝑧, 𝑖𝜁) transforms the
prefix R- into I- and vice versa, in each of the adjectives in Definition 13.5.1.

Example 13.5.3 If 𝐹 : R2𝑛 2𝑛

𝑥, 𝜉 −→ R 𝑦, 𝜂 is a linear symplectomorphism (for the
standard 2-forms 𝑥 · 𝜉 ′ − 𝑥 ′ · 𝜉 and 𝑦 · 𝜂 ′ − 𝑦 ′ · 𝜂) then the graph of 𝐹 in C2𝑛
𝑧,𝜁 is
R-Lagrangian.

Example 13.5.4 As real vector subspaces of C2𝑛 2𝑛

𝑧,𝜁 , R 𝑥, 𝜉 is R-symplectic and I-
Lagrangian whereas R 𝑥 ×𝑖R 𝜂 is R-Lagrangian and I-symplectic (cf. Remark 13.5.2).
𝑛 𝑛

Proposition 13.5.5 The following properties of an I-Lagrangian (resp., R-Lagran-

gian) vector subspace 𝑬 of C2𝑛 are equivalent:

(1) 𝑬 is R-symplectic (resp., I-symplectic);

(2) there is a 𝜏 ∈ Sp (𝑛, C) such that 𝜏𝑬 = R2𝑛
𝑥, 𝜉 (resp., 𝜏𝑬 = R 𝑥 × 𝑖R 𝜂 )
𝑛 𝑛

(3) 𝑬 is totally real and dimR 𝑬 = 2𝑛.

√
We recall that a linear subspace 𝑬 of C2𝑛 is said to be totally real if 𝑬 ∩ −1𝑬 =
{0}.
13.5 Real and Imaginary Symplectic Structures in C2𝑛 431

Proof Whether 𝑬 is I-Lagrangian or R-Lagrangian we have dimR 𝑬 = 2𝑛. Suppose

𝑬 is I-Lagrangian and R-symplectic: there is a linear basis 𝒆 1 , ..., 𝒆 𝑛 , 𝜺 1 , ..., 𝜺 𝑛 of
𝑬 such that, for all 𝑗 , 𝑘, 1 ≤ 𝑗, 𝑘 ≤ 𝑛,

𝜛𝑛C 𝒆 𝑗 , 𝒆 𝑘 = 𝜛𝑛C 𝜺 𝑗 , 𝜺 𝑘 = 0, 𝜛𝑛C 𝒆 𝑗 , 𝜺 𝑘 = 𝛿 𝑗,𝑘 .

Let 𝜏 ∈ Sp (𝑛, C) transform 𝒆 𝑗 into the unit vector in the 𝑧 𝑗 -plane and 𝜺 𝑗 into the unit
vector in the 𝜁 𝑗 -plane for each 𝑗 = 1, ..., 𝑛. The basis 𝜏𝒆 1 , ..., 𝜏𝒆 𝑛 , 𝜏𝜺 1 , ..., 𝜏𝜺 𝑛 spans
R2𝑛
𝑥, 𝜉 over R. This shows that (1) implies (2); since (3) is invariant under GL (𝑛, C)
action (3) follows from (2).
Next, we hypothesize that (1) does not hold, i.e., that there is (𝑧, 𝜁) ∈ 𝑬\ {0} or-
thogonal to 𝑬 with respect to the symplectic form Re 𝜛𝑛C . Since 𝑬 is I-isotropic
(𝑧, 𝜁) is orthogonal to 𝑬 with respect to 𝜛𝑛C ; since 𝜛𝑛C ((𝑖𝑧, 𝑖𝜁) , (𝑧 ′, 𝜁 ′)) =
𝑖𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) we conclude that 𝑖 (𝑧, 𝜁) is also orthogonal to 𝑬 with respect
to 𝜛𝑛C and therefore with respect to Im 𝜛𝑛C . Since 𝑬√is I-Lagrangian we reach the
conclusion that 𝑖 (𝑧, 𝜁) ∈ 𝑬, which implies that 𝑬 ∩ −1𝑬 ≠ {0}. This proves that
(3) implies (1).
The analogous claims about R-Lagrangian vector subspaces ensue then from
(13.5.4). □

13.5.2 𝑬 -conjugation and the associated Hermitian form

When 𝑬 is totally real and dimR 𝑬 = 2𝑛 we have C2𝑛 𝑧,𝜁 = 𝑬 ⊕ 𝑖𝑬 and therefore an
arbitrary element of C𝑧,𝜁 has the form (𝑧, 𝜁) + 𝑖 (𝑧 , 𝜁 ′) with (𝑧, 𝜁), (𝑧 ′, 𝜁 ′) ∈ 𝑬.
2𝑛 ′

We can define the 𝑬-conjugation as the map (𝑧, 𝜁) + 𝑖 (𝑧 ′, 𝜁 ′) ↦→ (𝑧, 𝜁) − 𝑖 (𝑧 ′, 𝜁 ′)

of C2𝑛
𝑧,𝜁 into itself; it is the unique antilinear automorphism c𝑬 of C2𝑛 such that
𝑬 = (𝑧, 𝜁) ∈ C2𝑛 ; (𝑧, 𝜁) = (𝑧, 𝜁) c𝑬 . We have

1 C
𝑖𝜛 ((𝑧, 𝜁) + 𝑖 (𝑧 ′, 𝜁 ′) , (𝑧, 𝜁) − 𝑖 (𝑧 ′, 𝜁 ′)) = 𝑧 · 𝜁 ′ − 𝜁 · 𝑧 ′ = 𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) .
2 𝑛
(13.5.5)

Proposition 13.5.6 Let 𝑬 be an I-Lagrangian and R-symplectic vector subspace of

C𝑛 . If 𝜏 ∈ Sp (𝑛, C) is such that 𝜏𝑬 = R2𝑛 𝑥, 𝜉 (cf. Proposition 13.5.5) then, for every
(𝑧, 𝜁) ∈ C2𝑛 ,
(𝑧, 𝜁) c𝑬 = 𝜏 −1 𝜏 (𝑧, 𝜁) (13.5.6)
and
𝜛𝑛C ((𝑧, 𝜁) , (𝑧, 𝜁) c𝑬 ) = 𝜛𝑛C 𝜏 (𝑧, 𝜁) , 𝜏 (𝑧, 𝜁) . (13.5.7)

Proof If (𝑧, 𝜁) ∈ 𝑬, 𝜏 ∈ Sp (𝑛, C) such that 𝜏 (𝑧, 𝜁) ∈ R2𝑛 we have 𝜏 −1 𝜏 (𝑧, 𝜁) =

(𝑧, 𝜁). An arbitrary element of C2𝑛 has the form (𝑧, 𝜁)+𝑖 (𝑧 ′, 𝜁 ′) with (𝑧, 𝜁), (𝑧 ′, 𝜁 ′) ∈
𝑬; it follows that
432 13 Elements of Symplectic Geometry

𝜏 −1 𝜏 ((𝑧, 𝜁) + 𝑖 (𝑧 ′, 𝜁 ′)) = 𝜏 −1 (𝜏 (𝑧, 𝜁) − 𝑖𝜏 (𝑧 ′, 𝜁 ′)) ,

whence (13.5.6). Since 𝜏 ∈ Sp (𝑛, C) we have, for every (𝑧, 𝜁) ∈ C2𝑛 ,

𝜛𝑛C (𝑧, 𝜁) , (𝑧, 𝜁) c𝑬 = 𝜛𝑛C (𝑧, 𝜁) , 𝜏 −1 𝜏 (𝑧, 𝜁) = 𝜛𝑛C 𝜏 (𝑧, 𝜁) , 𝜏 (𝑧, 𝜁) ,

whence (13.5.7). □
When 𝑬 is I-Lagrangian and R-symplectic the left-hand side in (13.5.5) can be
regarded as a Hermitian form in C2𝑛 ; the right-hand side is a real skew-symmetric
bilinear form on 𝑬.

Corollary 13.5.7 If 𝑬 is an I-Lagrangian and R-symplectic vector subspace of C𝑛

then the Hermitian form (13.5.5) on C2𝑛 has 𝑛 eigenvalues > 0 and 𝑛 eigenvalues
< 0 [and therefore (13.5.5) is nondegenerate].

Proof By Proposition 13.5.6 it suffices to prove the claim when 𝑬 = R2𝑛 𝑥, 𝜉 , in which
case (13.5.5) reads 21 𝑖𝜛𝑛C (𝑧, 𝜁) , 𝑧¯, 𝜁¯ = − Im 𝑧 · 𝜁¯ , which is equal to ± |𝑧| 2 when

𝜁 = ±𝑖𝑧. □
When 𝑬 is R-Lagrangian and I-symplectic we replace (13.5.5) by
1 C
𝜛 ((𝑧, 𝜁) + 𝑖 (𝑧 ′, 𝜁 ′) , (𝑧, 𝜁) − 𝑖 (𝑧 ′, 𝜁 ′)) = −𝑖𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) . (13.5.8)
2 𝑛
When 𝑬 = R𝑛𝑥 × 𝑖R𝑛𝜂 (13.5.8) reads

1 C
𝜛 (𝑥 + 𝑖𝑥 ′, 𝑖𝜂 − 𝜂 ′) , (𝑥 − 𝑖𝑥 ′, 𝑖𝜂 + 𝜂 ′) = 𝑥 · 𝜂 ′ − 𝑥 ′ · 𝜂, (13.5.9)
2 𝑛

which is equal to ± |𝑥 + 𝑖𝑥 ′ | 2 when 𝜂 = ∓𝑥 ′, 𝜂 ′ = ±𝑥, and therefore the Hermitian

form (13.5.9) on C2𝑛 has 𝑛 eigenvalues > 0 and 𝑛 eigenvalues < 0. Using 𝜏 ∈
Sp (𝑛, C) such that 𝜏𝑬 = R𝑛𝑥 × 𝑖R𝑛𝜂 (cf. Proposition 13.5.5) we have the analogue of
Proposition 13.5.6 except that the standard complex conjugation in C2𝑛 𝑧,𝜁 , (𝑧, 𝜁) ↦→
𝑧¯, 𝜁 , must be replaced by the conjugation with respect to R 𝑥 ×𝑖R 𝜂 , (𝑧, 𝜁) ↦→ 𝑧¯, −𝜁¯ .
¯
𝑛 𝑛

It implies the analogue of Corollary 13.5.7 which we state here as a proposition, for
reference below.

Proposition 13.5.8 If 𝑬 is an R-Lagrangian and I-symplectic subspace of C2𝑛

𝑧,𝜁 then
2𝑛
the Hermitian form (13.5.8) on C has 𝑛 eigenvalues > 0 and 𝑛 eigenvalues < 0.
13.5 Real and Imaginary Symplectic Structures in C2𝑛 433

13.5.3 Complex Lagrangian subspaces transversal to an R-Lagrangian

subspace

Proposition 13.5.9 A 𝜛𝑛C -Lagrangian subspace 𝑭 of C2𝑛

𝑧,𝜁 is a C-linear subspace of
C2𝑛
𝑧,𝜁 .

Proof Indeed, 𝑭 = 𝑭 ⊤𝜛𝑛 , its orthogonal with respect to 𝜛𝑛C ; the claim then follows
C

from Proposition 13.1.5. □

From now on we refer to any 𝜛𝑛C -Lagrangian subspace of C2𝑛
𝑧,𝜁 as a complex
Lagrangian subspace.

Proposition 13.5.10 To every R-Lagrangian (resp., I-Lagrangian) subspace 𝑬 of

C2𝑛 2𝑛
𝑧,𝜁 there is a complex Lagrangian subspace 𝑭 of C 𝑧,𝜁 such that 𝑬∩ 𝑭 = {0}.

Proof We shall deal with an R-Lagrangian subspace 𝑬, the I-Lagrangian case

ensuing by substitution of (𝑧, 𝑖𝜁) for (𝑧, 𝜁) (cf. Remark 13.5.2). Suppose the pullback
to 𝑬 of Im 𝜛𝑛C has rank 2𝜈, meaning that the maximum real dimension of an Im 𝜛𝑛C -
symplectic linear subspace of 𝑬 is equal to 2𝜈. If 𝜈 = 0, i.e., if 𝑬 is complex
Lagrangian, Proposition 13.1.15 implies that there is a complex Lagrangian subspace
𝑭 of C2𝑛 𝑧,𝜁 such that 𝑬∩ 𝑭 = {0}. Now suppose 1 ≤ 𝜈 ≤ 𝑛. There is a linear basis
𝒆 1 , ..., 𝒆 𝑛 , 𝜺 1 , ..., 𝜺 𝑛 of 𝑬 such that

Im 𝜛𝑛C 𝒆 𝑗 , 𝒆 𝑘 = 0, Im 𝜛𝑛C 𝜺 𝑗 , 𝜺 𝑘 = 0, 𝑗, 𝑘 = 1, ..., 𝑛,

Im 𝜛𝑛C 𝒆 𝑗 , 𝜺 𝑘 = 0 if 𝑗 ≠ 𝑘 or if min ( 𝑗, 𝑘) ≥ 𝜈 + 1,

Im 𝜛𝑛C 𝒆 𝑗 , 𝜺 𝑗 = 1 if 1 ≤ 𝑗 ≤ 𝜈.

The C-linear subspace 𝑭 of C2𝑛 𝑧,𝜁 spanned by 𝒆 1 + 𝑖𝜺 1 , ..., 𝒆 𝑛 + 𝑖𝜺 𝑛 is complex

Lagrangian and 𝑬 ∩ 𝑭 = {0}. Indeed, if 1 ≤ 𝑗 < 𝑘 ≤ 𝑛 then

Re 𝜛𝑛C 𝒆 𝑗 + 𝑖𝜺 𝑗 , 𝒆 𝑘 + 𝑖𝜺 𝑘 = Im 𝜛𝑛C 𝒆 𝑘 , 𝜺 𝑗 − Im 𝜛𝑛C 𝒆 𝑗 , 𝜺 𝑘 = 0,

Im 𝜛𝑛C 𝒆 𝑗 + 𝑖𝜺 𝑗 , 𝒆 𝑘 + 𝑖𝜺 𝑘 = Re 𝜛𝑛C 𝒆 𝑗 , 𝜺 𝑘 − Re 𝜛𝑛C 𝒆 𝑘 , 𝜺 𝑗 = 0,

the last equation because 𝑬 is R-Lagrangian. Now suppose there were real numbers
𝑎 1 , ..., 𝑎 𝑛 , 𝑏 1 , ..., 𝑏 𝑛 and complex numbers 𝑐 1 , ..., 𝑐 𝑛 such that
𝑛
∑︁ 𝑛
∑︁ 𝑛
∑︁
𝑎𝑗𝒆𝑗 + 𝑏 𝑗𝜺𝑗 = 𝑐 𝑗 𝜺 𝑗 + 𝑖𝜺 𝑛+ 𝑗 .
𝑗=1 𝑗=1 𝑗=1

This would require, for every 𝑗 = 1, ..., 𝑛, 𝑐 𝑗 = 𝑎 𝑗 = −𝑖𝑏 𝑗 , hence 𝑐 𝑗 = 0. □

In Proposition 13.5.10 we have dimR 𝑬 = dimR 𝑭 = 2𝑛 and therefore C2𝑛
𝑧,𝜁 =
𝑬 + 𝑭 = 𝑬 ⊕ 𝑭: 𝑬 and 𝑭 are transversal.
434 13 Elements of Symplectic Geometry

Proposition 13.5.11 An arbitrary complex Lagrangian vector subspace 𝑭 of C2𝑛

𝑧,𝜁
has the following properties:
(1) There is an R-linear subspace of C2𝑛𝑧,𝜁 transversal to 𝑭 that is both R-Lagrangian
and I-symplectic (resp., I-Lagrangian and R-symplectic).
(2) There is a 𝜏 ∈ Sp (𝑛, C) such that, given any R-Lagrangian (I-Lagrangian)
subspace 𝑬 of C2𝑛 𝑧,𝜁 transversal to 𝑭, the coordinate projection 𝜏𝑬 ∋ (𝑧, 𝜁) ↦→
𝑧 ∈ C𝑛 is a bijection.

Proof By Proposition 13.1.16 there is a 𝜏 ∈ Sp (𝑛, C) such that 𝜏𝑭 = {0} × C𝑛𝜁 . The
restriction to 𝑬 ◦ = (𝑧, 𝜁) ∈ C2𝑛 ; 𝜁 = 𝑧¯ of the 2-form 𝜛𝑛C is equal to 2𝑖 Im (𝑧 · 𝑧¯ ′)

and thus 𝑬 ◦ is both R-Lagrangian and I-symplectic; it is obviously transversal to 𝜏𝑭

and therefore 𝜏 −1 𝑬 ◦ has the required properties in (1). Given any 2𝑛-dimensional
totally real vector subspace 𝑬 of C2𝑛𝑧,𝜁 , if 𝜏𝑬 and {0} × C 𝜁 are transversal then the
𝑛

map 𝜏𝑬 ∋ (𝑧, 𝜁) ↦→ 𝑧 is injective. □

Proposition 13.5.12 Let 𝑬 be an R-Lagrangian and I-symplectic vector subspace

of C2𝑛 and 𝑭 be a complex Lagrangian subspace of C2𝑛 such that 𝑬 ∩ 𝑭 = {0}. If
𝜏 ∈ Sp (𝑛, C) is such that 𝜏𝑬 = R𝑛𝑥 × 𝑖R𝑛𝜂 (cf. Proposition 13.5.5) then the following
two properties hold:
(1) There is a unique pair of real symmetric 𝑛 × 𝑛 matrices 𝑀, 𝑁 such that 𝜏 𝑭 =
(𝑧, 𝜁) ∈ C2𝑛 ; 𝑧 = 𝑀 𝜁 + 𝑖𝑁 𝜁 ; 𝑀 is nonsingular.

˜ 𝑭 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝑧 = 𝑀◦ 𝜁 with 𝑀◦ a

(2) There is a 𝜏˜ ∈𝑆p (𝑛, C) such that 𝜏𝜏
diagonalization of 𝑀.

Proof Proof of (1). Recall that 𝜏𝑭 is a complex vector subspace of C2𝑛 . If there
existed a complex line Λ ⊂ C𝑛 such that (𝑧, 0) ∈ 𝜏𝑭 for all 𝑧 ∈ Λ then (𝑥, 0)
would belong to 𝜏𝑬 ∩ 𝜏𝑭 for all 𝑥 ∈ Λ ∩ R𝑛 , contradicting 𝑬 ∩ 𝑭 = {0}. Thus the
coordinate projection 𝜏𝑭 ∋ (𝑧, 𝜁) ↦→ 𝜁 is injective hence bijective and therefore there
is a unique 𝑀 + 𝑖𝑁 ∈ GL (𝑛, C), 𝑀, 𝑁 real, such that the equation 𝑧 = (𝑀 + 𝑖𝑁) 𝜁
defines 𝜏𝑭. For 𝑭 and 𝜏𝑭 to be complex Lagrangian it is necessary and sufficient
that 𝑀 + 𝑖𝑁 be symmetric. Moreover, we must have ((𝑀 + 𝑖𝑁) 𝜁, 𝜁) ∉ R𝑛𝑥 × 𝑖R𝑛𝜂
whatever 𝜁 ∈ C𝑛 \ {0} or, equivalently, 𝑀𝜂 ≠ 0 if 0 ≠ 𝜂 ∈ R𝑛 , i.e., det 𝑀 ≠ 0.
Proof of (2). The matrix
𝐼 −𝑖𝑁
𝜏˜1 = 𝑛
0 𝐼𝑛
belongs to Sp (𝑛, C) [cf. (13.1.16)] and 𝜏˜1 𝜏𝑭 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝑧 = 𝑀 𝜁 . If Γ ∈


Γ0
O𝑛 (R) it is immediately checked that Γ ⊗ Γ = ∈ Sp (𝑛, R) and
0Γ

(Γ ⊗ Γ) 𝜏˜1 𝜏 𝑭 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝑧 = Γ𝑀Γ⊤ 𝜁 .

It suffices to select Γ such that Γ𝑀Γ⊤ is diagonal and to take 𝜏˜ = (Γ ⊗ Γ) 𝜏˜1 . □

13.5 Real and Imaginary Symplectic Structures in C2𝑛 435

Remark 13.5.13 The automorphism (𝑧, 𝜁) ↦→ (𝑧, 𝑖𝜁) of C2𝑛 transforms each com-
plex Lagrangian subspace into another such subspace. The analogue of Proposition
13.5.12 is valid when we replace R-Lagrangian by I-Lagrangian, R𝑛𝑥 × 𝑖R𝑛𝜂 by R2𝑛
𝑥, 𝜉
and 𝑀 + 𝑖𝑁 by 𝑁 − 𝑖𝑀.

13.5.4 Real quadratic forms and R-Lagrangian subspaces

In this subsection 𝑄 (𝑧, 𝑧¯) will be a real quadratic form in C𝑛 ; we define

𝑬 𝑄 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝜁 = 𝜕𝑧 𝑄

(13.5.10)

where 𝜕𝑧 is the holomorphic exterior derivative.

Proposition 13.5.14 If a real vector subspace 𝑬 of C2𝑛 𝑧,𝜁 is R-Lagrangian then to

2𝑛
every complex Lagrangian subspace 𝑭 of C𝑧,𝜁 transversal to 𝑬 (cf. Proposition
13.5.10) there exist a linear transformation 𝜏 ∈ Sp (𝑛, C) and a real quadratic form
𝑄 in C𝑛𝑧 satisfying 𝜕𝑧2 𝑄 ≡ 0, such that 𝜏𝑭 = {0} × C𝑛𝜁 , 𝜏𝑬 = 𝑬 𝑄 . Conversely, if
there exist 𝜏 ∈ Sp (𝑛, C) and a real quadratic form 𝑄 in C𝑛𝑧 such that 𝜏𝑬 = 𝑬 𝑄 then
𝑬 is R-Lagrangian.

Proof Let 𝜏1 ∈ Sp (𝑛, C) be such that 𝜏1 𝑭 = {0} × C𝑛𝜁 ; obviously the coordinate
projection 𝜏1 𝑬 ∋ (𝑧, 𝜁) ↦→ 𝑧 is bijective [Proposition 13.5.11, (2)]; in other words,
there are 𝑛 × 𝑛 complex matrices 𝐴, 𝐵, such that the equation 𝜁 = 𝐴𝑧 + 𝐵 𝑧¯ defines
𝜏1 𝑬. As a consequence we have, on 𝜏1 𝑬,

Re 𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)) (13.5.11)

= Re 𝑧 · 𝐴 − 𝐴⊤ 𝑧 ′ + Re (𝑧 · (𝐵 − 𝐵∗ ) 𝑧¯ ′) .

Since 𝑬 is R-Lagrangian we must have ( 𝐴 − 𝐴⊤ ) 𝑧 + (𝐵 − 𝐵∗ ) 𝑧¯ = 0 for every

𝑧 ∈ C𝑛 whence 𝐴 = 𝐴⊤ and 𝐵 = 𝐵∗ . Let us carry out the symplectic transformation
𝜏2 : (𝑧, 𝜁) ↦→ (𝑧, 𝜁 − 𝐴𝑧) of C2𝑛 ; the restriction of 𝜏2 to {0} × C𝑛 is the identity and

𝜏2 𝜏1 𝑬 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝜁 = 𝐵 𝑧¯ .

We have 𝜏2 𝜏1 𝑬 = 𝑬 𝑄 if we take 𝑄 = 𝑧 · 𝐵 𝑧¯. Since 𝐵 is self-adjoint 𝑄 is real.

Conversely, let 𝑄 be a real quadratic form: there are complex 𝑛 × 𝑛 matrices 𝐴,
𝐵, such that
𝑄 (𝑧) = Re ( 𝐴𝑧 · 𝑧) + 𝑧 · 𝐵 𝑧¯, (13.5.12)
𝐴 = 𝐴⊤ , 𝐵∗ = 𝐵; then [cf. (13.5.10)]

𝑬 𝑄 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝜁 = 𝐴𝑧 + 𝐵 𝑧¯ .

The pullback of Re 𝜛𝑛C to 𝑬 𝑄 is equal to (13.5.11), implying that 𝑬 𝑄 is R-

Lagrangian. □
436 13 Elements of Symplectic Geometry

𝜕2 𝑄
Remark 13.5.15 If 𝑄 is given by (13.5.12) we have 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 1≤ 𝑗,𝑘 ≤𝑛 = 𝐵. The
(1, 1)-form in C𝑛 ,
𝑛
∑︁ 𝜕2𝑄
𝑖𝜕𝜕𝑄 = 𝑖 d𝑧 𝑗 ∧ d𝑧¯ 𝑘 , (13.5.13)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

is real; (13.5.13) is often referred to as the Levi 2-form of 𝑄, also when 𝑄 is an

arbitrary real-valued C 2 function in some open subset Ω of C𝑛 .

Proposition 13.5.16 Let 𝑄 be a real quadratic form in C𝑛 . The pullback to 𝑬 𝑄 of

Im 𝜛𝑛C is equal to that of (13.5.13) under the map (𝑧, 𝜁) ↦→ 𝑧.

Proof The pullback to 𝑬 𝑄 of Im 𝜛𝑛C is equal to Im d (𝜁 · d𝑧) = Im d𝜕𝑄; since

𝜕 2 𝑄 = 0 the claim ensues. □
In the following corollary we regard 𝑬 𝑄 as a 2𝑛-dimensional real subspace of
C2𝑛
𝑧,𝜁 .

Corollary 13.5.17 Therank ofthe pullback of Im 𝜛𝑛C to 𝑬 𝑄 is equal to the rank of

𝜕2 𝑄
the self-adjoint matrix 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 1≤ 𝑗,𝑘 ≤𝑛 .

Corollary 13.5.18 If 𝑄 is given by (13.5.12) the R-Lagrangian subspace 𝑬 = 𝜏 −1 𝑬 𝑄

is I-symplectic if and only if det 𝐵 ≠ 0.

Proposition 13.5.19 The properties of an R-Lagrangian subspace 𝑬 of C2𝑛 𝑧,𝜁 in

Proposition 13.5.5 are all equivalent to the existence of a real quadratic form 𝑄
2
in C𝑛 such that det 𝜕𝑧𝜕𝑗 𝜕𝑄𝑧¯ 𝑘 ≠ 0 and a linear transformation 𝜏 ∈ Sp (𝑛, C)
1≤ 𝑗,𝑘 ≤𝑛
such that 𝜏𝑬 = 𝑬 𝑄 .

Proof Combine Proposition 13.5.14 and Corollary 13.5.18. □

For a given R-Lagrangian subspace 𝑬 the choice of 𝑄 and 𝜏 are not unique
(in a fixed coordinate system). Here is an example where we relate 𝑄 to complex
Lagrangian subspaces transversal to 𝑬.

Example 13.5.20 Let 𝑬 = R2𝑥 × R2𝜂 . The symplectic C-linear transformation of C4𝑧,𝜁

1 1
𝜏1 : (𝑧, 𝜁) ↦→ √ (𝑧 + 𝑖𝜁) , √ (𝑧 − 𝑖𝜁)
2 2
n o
transforms 𝑬 into 𝜏1 𝑬 = (𝑧, 𝜁) ∈ C4𝑧,𝜁 ; 𝜁 = 𝑧¯ ; in other words, 𝜏1 𝑬 is defined by
the equation 𝜁 = 𝜕𝑧 |𝑧| 2 . The symplectic C-linear transformation

1 1 1 1
𝜏2 : (𝑧, 𝜁) ↦→ √ (𝑧 1 − 𝜁1 ) , √ (𝑧2 + 𝜁2 ) , √ (𝑧 1 + 𝜁1 ) , √ (𝜁2 − 𝑧 2 )
2 2 2 2
transforms 𝑬 into
13.5 Real and Imaginary Symplectic Structures in C2𝑛 437
n o
𝜏2 𝑬 = (𝑧, 𝜁) ∈ C4𝑧,𝜁 ; 𝜁1 = 𝑧¯1 , 𝜁2 = −𝑧¯2 ;

𝜏2 𝑬 is defined by the equation by 𝜁 = 𝜕𝑧 |𝑧1 | 2 − |𝑧2 | 2 . Note that 𝜏 𝑗 𝑭 𝑗 = {0} × C2𝜁
( 𝑗 = 1, 2) if
n o
𝑭 1 = (𝑧, 𝜁) ∈ C4𝑧,𝜁 ; 𝜁 = −𝑧 ,
n o
𝑭 2 = (𝑧, 𝜁) ∈ C4𝑧,𝜁 ; 𝜁1 = −𝑧1 , 𝜁2 = 𝑧2 .

The following generalization of Example 13.5.20 relates 𝑬-conjugation to the

defining equations of the R-Lagrangian and I-symplectic subspace 𝑬.
n o
Lemma 13.5.21 Let 𝑬 = (𝑧, 𝜁) ∈ C2𝑛 𝑧,𝜁 ; 𝜁 = 𝐵 𝑧¯ with 𝐵 a self-adjoint 𝑛 × 𝑛 matrix,
det 𝐵 ≠ 0. Then 𝑬-conjugation is the map

(𝑧, 𝜁) ↦→ (𝑧, 𝜁) c𝑬 = 𝐵¯ −1 𝜁,
¯ 𝐵 𝑧¯ (13.5.14)

and we have
𝜛𝑛C ((𝑧, 𝜁) , (𝑧, 𝜁) c𝑬 ) = 𝑧 · 𝐵 𝑧¯ − 𝜁 · 𝐵¯ −1 𝜁.
¯ (13.5.15)

Proof The restriction to 𝑬 of (13.5.14) is obviously the identity; (13.5.15) follows

directly from (13.5.1) and (13.5.14). □
Notice that the Hermitian form (13.5.15) has 𝑛 eigenvalues > 0 and 𝑛 eigenvalues
< 0. Since 𝐵¯ −1 = (𝐵⊤ ) −1 (13.5.15) can be rewritten as

𝜛𝑛C (𝑧, 𝜁) , (𝑧, 𝜁) c𝑬 = 𝑧 · 𝐵 𝑧¯ − 𝐵−1 𝜁 · 𝜁.

¯

2
Let 𝑄 1 be the real quadratic form such that 𝐵−1 = 𝜕𝑧𝜕 𝑗 𝑄 1
𝜕 𝑧¯ 𝑘 ; Remark 13.5.15
1≤ 𝑗,𝑘 ≤𝑛
implies
𝑛 𝑛
∑︁ 𝜕2𝑄 ∑︁ 𝜕2𝑄1
𝜛𝑛C (𝑧, 𝜁) , (𝑧, 𝜁) c𝑬 = 𝑧 𝑗 𝑧¯ 𝑘 − 𝜁 𝑗 𝜁¯𝑘 .
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

Since 𝐵¯ is self-adjoint there is a unitary transformation 𝑇 of C𝑛 such that 𝑇 −1 𝐵𝑇

¯
¯
is diagonal and real; since 𝐵 is invertible −1 ¯ −1
𝑇 𝐵 𝑇 is also diagonal and real. The

transformation (𝑧, 𝜁) ↦→ 𝜏 (𝑧, 𝜁) = 𝑇 𝑧, (𝑇 ) −1 𝜁 is symplectic; we have
⊤

−1
𝜛𝑛C 𝜏 (𝑧, 𝜁) , (𝜏 (𝑧, 𝜁)) c𝑬 = 𝑇 𝑧 · 𝐵𝑇¯ 𝑧¯ − 𝐵−1 𝑇 ⊤ 𝜁 · (𝑇 ⊤ ) −1 𝜁

= 𝑇 −1 𝐵𝑇
¯ 𝑧 · 𝑧¯ − 𝜁 · 𝑇 −1 𝐵¯ −1𝑇 𝜁,
¯
438 13 Elements of Symplectic Geometry

whence
𝑛
∑︁ 2 2
𝜛𝑛C 𝜏 (𝑧, 𝜁) , (𝜏 (𝑧, 𝜁)) c𝑬 = 𝜆 𝑗 𝑧 𝑗 − 𝜆−1
𝑗 𝜁 𝑗 , 𝜆 𝑗 ∈ R. (13.5.16)
𝑗=1

13.5.5 A special Abelian subgroup of Sp (𝒏, C)

Let 𝜁 ◦ = (0, ..., 0, −1). We denote by Sym (𝑛, C) the complex vector space of
all complex symmetric 𝑛 × 𝑛 matrices and by Sym∗ (𝑛, C) the linear subspace of
Sym (𝑛, C) consisting of the matrices 𝐴 such that 𝐴𝜁 ◦ = 0. A symmetric matrix
𝐴 = 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 belongs to Sym∗ (𝑛, C) if and only if 𝑎 𝑗,𝑛 = 0 for all 𝑗 = 1, ..., 𝑛.
For application in the next section we will now study, in this subsection, linear
automorphisms of C2𝑛 𝑧,𝜁 of the following kind:

𝜏𝐴 : (𝑧, 𝜁) ↦→ (𝑧 + 𝐴𝜁, 𝜁) , (13.5.17)

where 𝐴 ∈ Sym∗ (𝑛, C). We have 𝜏𝐴 (0, 𝑐𝜁 ◦ ) = (0, 𝑐𝜁 ◦ ) whatever 𝑐 ∈ C.

Proposition 13.5.22 The map 𝐴 ↦→ 𝜏𝐴 is a bijection of Sym∗ (𝑛, C) onto an Abelian

subgroup of Sp (𝑛, C); we have

𝜏𝐴1 ◦ 𝜏𝐴2 = 𝜏𝐴1 +𝐴2 , 𝜏𝐴−1 = 𝜏−𝐴. (13.5.18)

Proof We have
𝑛
∑︁ 𝑛
∑︁

𝜁 𝑗 d 𝑧 𝑗 + 𝐴𝜁 = 𝜁 · d𝑧 + 𝑎 𝑗,𝑘 𝜁 𝑗 d𝜁 𝑘 = 𝜁 · d𝑧 + d𝑄,
𝑗=1 𝑗,𝑘=1

where 𝑄 (𝜁) = 𝐴𝜁 · 𝜁. It follows that 𝜏𝐴∗ (d𝜁 ∧ d𝑧) = d𝜏𝐴∗ (𝜁 · d𝑧) = d𝜁 ∧ d𝑧. The
identities (13.5.18) are self-evident. □

Proposition 13.5.23 If 𝐴 ∈ Sym∗ (𝑛, C) then 𝑭 𝐴 = 𝜏−𝐴 ({0} × C𝑛 ) is a complex

Lagrangian subspace of C2𝑛 containing (0, 𝜁 ◦ ) and such that 𝑭 𝐴 ∩(C𝑛 × {0}) = {0}.
Conversely, to every complex Lagrangian vector subspace 𝑭 of C2𝑛 transversal to
C𝑛 ×{0} and containing (0, 𝜁 ◦ ) there is an 𝐴 ∈ Sym∗ (𝑛, C) such that 𝜏𝐴 𝑭 = {0}×C𝑛 .

Proof The subspace 𝑭 𝐴 = (𝑧, 𝜁) ∈ C2𝑛 ; 𝑧 + 𝐴𝜁 = 0 is obviously transversal to

C𝑛 × {0}; it is complex Lagrangian because 𝜏−𝐴 ∈ Sp (𝑛, C); (0, 𝜁 ◦ ) ∈ 𝑭 𝐴 since

𝜏−𝐴 (0, 𝜁 ◦ ) = (0, 𝜁 ◦ ). If 𝑭 ∩ (C𝑛 × {0}) = {0} the coordinate projection 𝑭 ∋
(𝑧, 𝜁) ↦→ 𝜁 ∈ C𝑛 is bijective. This means that 𝑭 can be defined by a linear equation
𝑧 = 𝐴𝜁. If (𝑧, 𝜁) , (𝑤, 𝜃) ∈ 𝑭 we have

𝜛𝑛C ((𝑧, 𝜁) , (𝑤, 𝜃)) = 𝑧 · 𝜃 − 𝜁 · 𝑤 = 𝐴 − 𝐴⊤ 𝜁 · 𝜃 = 0.

13.5 Real and Imaginary Symplectic Structures in C2𝑛 439

Since 𝜁, 𝜃 ∈ C𝑛 are arbitrary we conclude that 𝐴 = 𝐴⊤ . Since the symplectic

orthogonal of (0, 𝜁 ◦ ) ∈ 𝑭 is the hyperplane 𝑧 𝑛 = 0 in C2𝑛
𝑧,𝜁 we see that (𝑧, 𝜁) ∈
𝑭 = 𝑭 ⊥𝜛𝑛 =⇒ 𝑧 𝑛 = 0, implying 𝐴 ∈ Sym∗ (𝑛, C).
C
□
By Proposition 13.1.16 to every complex Lagrangian subspace 𝑭 of C2𝑛 𝑧,𝜁 there
is a g ∈ Sp (𝑛, C) such that g𝑭 = {0} × C ; if g1 ∈ Sp (𝑛, C) is also such that
𝑛

g1 𝑭 = {0} × C𝑛 then g1 g−1 ({0} × C𝑛 ) = ({0} × C𝑛 ). The subset of Sp (𝑛, C)

preserving {0} × C𝑛 is a subgroup G of Sp (𝑛, C); the set of complex Lagrangian
subspaces of C2𝑛
𝑧,𝜁 can be identified with the quotient Lagr (𝑛, C) = Sp (𝑛, C) /G,
the Lagrangian Grassmannian of C2𝑛 . The latter naturally carries the structure of a
complex-analytic, compact, manifold. It is an easy exercise to prove that, given any
𝑭 ◦ ∈ Lagr (𝑛, C), the set of 𝑭 ∈ Lagr (𝑛, C) such that 𝑭 ∩ 𝑭 ◦ = {0} is open and
dense in Lagr (𝑛, C).

Lemma 13.5.24 The set of all complex Lagrangian vector subspaces of C2𝑛 of the
type 𝑭 𝐴, 𝐴 ∈ Sym∗ (𝑛, C), is open and dense in the set of all 𝑭 ∈ Lagr (𝑛, C)
contained in the hyperplane 𝑧 𝑛 = 0.

Proof We have already seen that (𝑧, 𝜁) ∈ 𝑭 =⇒ 𝑧 𝑛 = 0 is equivalent to (0, 𝜁 ◦ ) ∈ 𝑭.

Define 𝑭 ′ = {(𝑧, 𝜁) ∈ 𝑭; 𝜁 𝑛 = 0}. On the one hand, 𝑭 = 𝑭 ′ ⊕ C (0, 𝜁 ◦ ) and

𝑭 ∩ (C𝑛 × {0}) = 𝑭 ′ ∩ C𝑛−1 × {0} ⊕ C (0, 𝜁 ◦ ) .

On the other hand we can identify 𝑭 ′ with a complex Lagrangian subspace of

C2(𝑛−1) ; the set of complex Lagrangian subspaces of C2(𝑛−1) being open and dense
in Lagr (𝑛 − 1, C), the claim is seen to be a consequence of Proposition 13.5.23. □
In the remainder of this subsection 𝑬 shall be an R-Lagrangian and I-symplectic
vector subspace of C𝑛𝑧,𝜁 such that 𝜁 ◦ ∈ 𝑬. We can select a symplectic basis of 𝑬 for
the symplectic form Im 𝜛𝑛C (= −𝑖𝜛𝑛C on 𝑬), 𝒆 1 , ..., 𝒆 𝑛 , 𝜺 1 , ..., 𝜺 𝑛 with 𝜺 𝑛 = 𝜁 ◦ ; we
have

𝜛𝑛C 𝒆 𝑗 , 𝒆 𝑘 = 𝜛𝑛C 𝜺 𝑗 , 𝜺 𝑘 = 0, 𝜛𝑛C 𝜺 𝑗 , 𝒆 𝑘 = 𝑖𝛿 𝑗,𝑘 (1 ≤ 𝑗, 𝑘 ≤ 𝑛).
Í
A generic element of 𝑬 has the form 𝑛𝑗=1 𝑎 𝑗 𝒆 𝑗 + 𝛼 𝑗 𝜺 𝑗 (𝑎 𝑗 , 𝛼 𝑗 ∈ R) and therefore
a generic element of C2𝑛 will have the form 𝑛𝑗=1 𝑐 𝑗 𝒆 𝑗 + 𝛾 𝑗 𝜺 𝑗 (𝑐 𝑗 , 𝛾 𝑗 ∈ C). The
Í

Hermitian form in C2𝑛 , (13.5.8), reads

𝑛 𝑛 𝑛 𝑛
1 C ©∑︁ ∑︁ ∑︁ ∑︁
𝑐𝑗𝒆𝑗 + 𝛾𝑗𝜺 𝑗, 𝑐¯ 𝑗 𝒆 𝑗 + 𝛾¯ 𝑗 𝜺 𝑗 ® = − Im (𝑐 · 𝛾)
¯ . (13.5.19)
ª
𝜛𝑛 ­
2
« 𝑗=1 𝑗=1 𝑗=1 𝑗=1 ¬
Let 𝑭 ◦ be the complex vector subspace of C2𝑛 spanned by 𝒆 𝑗 −𝑖𝜺 𝑗 (1 ≤ 𝑗 ≤ 𝑛 −1)
and by 𝜺 𝑛 = (0, 𝜁 ◦ ). We have

𝜛 𝒆 𝑗 − 𝑖𝜺 𝑗 , 𝒆 𝑘 − 𝑖𝜺 𝑘 = −𝑖𝜛 𝒆 𝑗 , 𝜺 𝑘 + 𝑖𝜛 𝒆 𝑘 , 𝜺 𝑗 = 0 (1 ≤ 𝑗, 𝑘 ≤ 𝑛 − 1)
440 13 Elements of Symplectic Geometry

and 𝜛 𝒆 𝑗 − 𝑖𝜺 𝑗 , 𝜺 𝑛 = 0: 𝑭 ◦ is complex Lagrangian. We have 𝑬 ∩ 𝑭 ◦ = R𝜺 𝑛 .
Indeed, if 𝑐 𝑗 ∈ C and 𝑎 𝑘 , 𝑏 𝑘 ∈ R (1 ≤ 𝑗, 𝑘 ≤ 𝑛) satisfy
𝑛−1
∑︁ 𝑛
∑︁

𝑐 𝑗 𝒆 𝑗 − 𝑖𝜺 𝑗 + 𝑐 𝑛 𝜺 𝑛 = 𝑎𝑘 𝒆𝑘 + 𝑏𝑘 𝜺𝑘
𝑗=1 𝑘=1

then necessarily 𝑐 𝑗 ∈ R and 𝑖𝑐 𝑗 ∈ R hence 𝑐 𝑗 = 0 for all 𝑗 ≤ 𝑛 − 1, 𝑎 𝑛 = 0, 𝑐 𝑛 = 𝑏 𝑛 .

To restrict (13.5.3) to 𝑭 ◦ we put 𝛾 𝑗 = −𝑖𝑐 𝑗 (1 ≤ 𝑗 ≤ 𝑛 − 1), 𝑐 𝑛 = 0:

𝑛−1 𝑛−1 𝑛−1

1 C ©∑︁ ∑︁ ª ∑︁ 2
𝑐 𝑗 𝒆 𝑗 − 𝑖 𝑗 𝜺 𝑗 + 𝛾𝑛 𝜁 ◦ , 𝑐¯ 𝑗 𝒆 𝑗 + 𝑖𝜺 𝑗 + 𝛾¯ 𝑛 𝜁 ◦ ® =

𝜛𝑛 ­ 𝑐 𝑗 . (13.5.20)
2
« 𝑗=1 𝑗=1 ¬ 𝑗=1
We are now in a position to prove

Proposition 13.5.25 Let 𝑬 be an R-Lagrangian and I-symplectic vector subspace

of C𝑛𝑧,𝜁 such that 𝜁 ◦ ∈ 𝑬. There is a symplectomorphism (13.5.17) such that the
following is true:
(1) 𝜏𝐴 𝑬 ∩ ({0} × C𝑛 ) = (0, R𝜁 ◦ );
(2) the restriction to {0} × C𝑛 of the Hermitian form (13.5.8) is positive semidefinite
with null space (0, C𝜁 ◦ ).

Proof As seen by (13.5.20) the restriction to 𝑭 ◦ of the Hermitian form (13.5.19)

is positive semidefinite with null space C𝜁 ◦ . By Lemma 13.5.24 we know that
there is a Lagrangian subspace of C2𝑛 of the form 𝑭 𝐴 = 𝜏−𝐴 ({0} × C𝑛 ) arbitrarily
close to 𝑭 ◦ , implying that 𝑭 ′ = {(𝑧, 𝜁) ∈ 𝑭 𝐴; 𝜁 𝑛 = 0} is arbitrarily close to 𝑭 ◦′ =
{(𝑧, 𝜁) ∈ 𝑭 ◦ ; 𝜁 𝑛 = 0} so that 𝑬 ∩ 𝑭 ′ = R𝜺 𝑛 and the restriction of (13.5.19) to 𝑭 ′ is
positive definite. It follows that 𝜏𝐴 𝑬 ∩ 𝜏𝐴 𝑭 𝐴 = 𝜏𝐴 𝑬 ∩ ({0} × C𝑛 ) = R𝜺 𝑛 and the
claim (2) is valid. □

13.6 Real and Imaginary Symplectic Structures on Complex

Manifolds

13.6.1 The R and I terminology in a complex-analytic manifold

In this section M will be a complex-analytic manifold; dimC M = 𝑛. We regard the

vector subbundle 𝑇 (1,0) M of C𝑇 ∗ M [see Subsection 9.4.5, in particular (9.4.24)]
as a complex symplectic manifold and denote by 𝜛 the fundamental symplectic
two-form on 𝑇 (1,0) M. If 𝑧 1 , ..., 𝑧 𝑛 are holomorphic coordinates in an open subset 𝑈
of M their differentials d𝑧 𝑗 form a basis of 𝑇𝑧(1,0) M (𝑧 ∈ 𝑈); we denote by 𝜁1 , ..., 𝜁 𝑛
the coordinates in this basis; in these frames,
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 441
𝑛
∑︁
𝜛 = d𝜁 ∧ d𝑧 = d𝜁 𝑗 ∧ d𝑧 𝑗 . (13.6.1)
𝑗=1

We also look at M as a 2𝑛-dimensional real manifold (of class C 𝜔 ) and we identify

𝑇 (1,0) M with the real cotangent bundle of the real manifold M, 𝑇 ∗ M. Holomorphic
local coordinates 𝑧 𝑗 yield real C 𝜔 coordinates 𝑥 𝑗 = Re 𝑧 𝑗 , 𝑦 𝑘 = Im 𝑧 𝑘 , as well as
dual real coordinates 𝜉 𝑗 = Re 𝜁 𝑗 , 𝜂 𝑘 = Im 𝜁 𝑘 , in the real cotangent spaces; then, over
the domain of these coordinates, the vector bundle isomorphism 𝑇 (1,0) M −→ 𝑇 ∗ M
is simply the map (𝑧, 𝜁) ↦→ ((𝑥, 𝑦) , (𝜉, 𝜂)). We may deal with the two real 2-forms
Re 𝜛 and Im 𝜛 on 𝑇 ∗ M given, in the local coordinates 𝑥 𝑗 , 𝑦 𝑘 , 𝜉 𝑗 , 𝜂 𝑘 , by

1
d𝜁 ∧ d𝑧 + d𝜁¯ ∧ d𝑧¯

Re 𝜛 = (13.6.2)
2
= d𝜉 ∧ d𝑥 − d𝜂 ∧ d𝑦,

1
d𝜁 ∧ d𝑧 − d𝜁¯ ∧ d𝑧¯

Im 𝜛 = (13.6.3)
2𝑖
= d𝜉 ∧ d𝑦 + d𝜂 ∧ d𝑥.

Both 2-forms Re 𝜛 and Im 𝜛 are of class C 𝜔 , nondegenerate and closed, and

therefore define symplectic structures on 𝑇 ∗ M: indeed, the (nonsymplectic) change
of variables 𝑦 ⇝ −𝑦 (resp., 𝑥 ↭ 𝑦) transforms Re 𝜛 (resp., Im 𝜛) into the standard
fundamental 2-form on 𝑇 ∗ M, 𝜛 R = d𝜉 ∧ d𝑥 + d𝜂 ∧ d𝑦. The definitions (13.6.2) and
(13.6.3) are invariant under complex symplectic automorphisms of 𝑇 ∗ M.
Each form 𝜛, Re 𝜛, Im 𝜛, defines special Hamiltonian vector fields and Poisson
brackets on 𝑇 ∗ M. If U is an open subset of 𝑇 ∗ M and 𝑓 , 𝑔 ∈ C ∞ (U) we shall
denote by { 𝑓 , 𝑔} the Poisson bracket for 𝜛, and by { 𝑓 , 𝑔}R , { 𝑓 , 𝑔}I those for Re 𝜛,
Im 𝜛 respectively. We deduce from (13.6.2) and (13.6.3) the following respective
expressions for the Hamiltonian vector fields of 𝑓 in U :
𝑛
∑︁ 𝜕𝑓 𝜕 𝜕𝑓 𝜕
𝐻𝑓 = − , (13.6.4)
𝑗=1
𝜕𝜁 𝑗 𝜕𝑧 𝑗 𝜕𝑧 𝑗 𝜕𝜁 𝑗

and
𝑛
∑︁ 𝜕𝑓 𝜕 𝜕𝑓 𝜕 𝜕𝑓 𝜕 𝜕𝑓 𝜕
𝐻 R𝑓 = + − − , (13.6.5)
𝑗=1
𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝜕𝑦 𝑗 𝜕𝜂 𝑗 𝜕𝑥 𝑗 𝜕𝜉 𝑗 𝜕𝜂 𝑗 𝜕𝑦 𝑗
𝑛
∑︁ 𝜕𝑓 𝜕 𝜕𝑓 𝜕 𝜕𝑓 𝜕 𝜕𝑓 𝜕
𝐻 I𝑓 = + − − . (13.6.6)
𝑗=1
𝜕𝜉 𝑗 𝜕𝑦 𝑗 𝜕𝜂 𝑗 𝜕𝑥 𝑗 𝜕𝑦 𝑗 𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝜕𝜂 𝑗

We have{ 𝑓 , 𝑔} = 𝐻 𝑓 𝑔, { 𝑓 , 𝑔}R = 𝐻 R𝑓 𝑔, { 𝑓 , 𝑔}I = 𝐻 I𝑓 𝑔.

Let ℎ = 𝑓 + 𝑖𝑔 be a holomorphic function in U. The Cauchy–Riemann equations
imply directly
442 13 Elements of Symplectic Geometry

𝐻 R𝑓 = 𝐻𝑔I , 𝐻 I𝑓 = −𝐻𝑔R . (13.6.7)

We also get
1 R 1
𝐻ℎ = 𝐻 𝑓 − 𝑖𝐻 I𝑓 = 𝑖 𝐻𝑔R − 𝑖𝐻𝑔I . (13.6.8)
2 2
Proposition 13.6.1 If ℎh = 𝑓 + 𝑖𝑔i is a holomorphic function in an open subset U of
𝑇 ∗ M then { 𝑓 , 𝑔}I and 𝐻 I𝑓 , 𝐻𝑔I vanish identically in U.
h i
Proof By (13.6.7) we have 𝐻 I𝑓 𝑔 = −𝐻𝑔R 𝑔 = 0; also recall that 𝐻 I𝑓 , 𝐻𝑔I = 𝐻 I .□
{ 𝑓 ,𝑔 }I

The linear concepts and results of the preceding section have direct counterparts
in the manifold framework. Using local coordinates 𝑧 𝑗 defined and holomorphic in
a domain Ω of M we can identify 𝑇 ∗ M| Ω with Ω × C𝑛 and the fundamental two
form 𝜛 on 𝑇 ∗ M with 𝜛𝑛C defined in (13.5.1). Many of the proofs of the results that
follow duplicate (or exploit) the linear results of the preceding section.
First of all, we can make use of the terminology introduced in Definition 13.3.15
for the real symplectic structure on 𝑇 ∗ M defined by either Re 𝜛 or Im 𝜛; we get the
manifold version of Definition 13.5.1:

Definition 13.6.2 A C ∞ submanifold of 𝑇 ∗ M is said to be

(1) R-symplectic (resp., I-symplectic),
(2) R-isotropic (resp., I-isotropic),
(3) R-coisotropic (resp., I-coisotropic),
(4) R-Lagrangian (resp., I-Lagrangian)
if the appellative is valid in the real symplectic structure on 𝑇 ∗ M defined by Re 𝜛
(resp., Im 𝜛).

Unless specified otherwise, by a submanifold we shall always mean an immersed

submanifold (in the C ∞ , C 𝜔 or complex-analytic sense) without self-intersections
(the latter, mainly not to complicate the statements). The next two propositions
are statements about the cotangent spaces of the submanifolds and therefore follow
directly from their linear versions.

Proposition 13.6.3 The following properties of a complex-analytic submanifold N

of 𝑇 (1,0) M are equivalent:
(1) N is isotropic (resp., Lagrangian) with respect to 𝜛;
(2) regarded as a submanifold of 𝑇M, N is both R-isotropic and I-isotropic (resp.,
R-Lagrangian and I-Lagrangian).

Proposition 13.6.4 The following properties of an I-Lagrangian (resp., R-Lagran-

gian) C ∞ submanifold N of 𝑇 (1,0) M are equivalent:
(1) N is R-symplectic (resp., I-symplectic);
(2) N is totally real.
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 443

A real submanifold N of the complex manifold M is said to be totally real if

𝑇℘ N is a totally real linear subspace of 𝑇℘ M for every ℘ ∈ N .
The following statement are the analogues of similar ones in the linear set-up
(see the preceding section). The role of the real quadratic form 𝑄 is played here by
a purely imaginary C ∞ function 2𝑖1 𝜑 (thus 𝜑 is real) whose Levi form 2𝑖1 𝜕𝜕𝜑 is key
to the results.
Proposition 13.6.5 Let N be a C ∞ submanifold of 𝑇 (1,0) M such that the restriction
to N of the base projection 𝜋 : (𝑧, 𝜁) ↦→ 𝑧 is a local diffeomorphism (thus dimR N =
2𝑛). For N to be I-Lagrangian it is necessary and sufficient that to every (𝑧◦ , 𝜁 ◦ ) ∈ N
there be a neighborhood 𝑈 of 𝑧◦ in M and a function 𝜑 ∈ C ∞ (𝑈; R) such that the
−1
connected component Λ of 𝜋 (𝑈) ∩ N containing (𝑧 ◦ , 𝜁 ◦ ) is defined by the equation
𝜁 = 2𝑖1 𝜕𝑧 𝜑 in a neighborhood of (𝑧 ◦ , 𝜁 ◦ ).
Proof By hypothesis 𝑧◦ has a neighborhood 𝑈 in which there is a smooth section of
−1
𝑇 (1,0) M over 𝑈, 𝑧 ↦→ (𝑧, ℎ (𝑧)), equal to the connected component Λ of 𝜋 (𝑈) ∩ N
through (𝑧◦ , 𝜁 ◦ ). In 𝑇 ∗ M we have d (Im (𝜁 · d𝑧)) = Im (d𝜁 ∧ d𝑧) = Im 𝜛. It follows
that the pullback of Im (𝜁 · d𝑧) is closed. After contracting 𝑈 about 𝑧 ◦ there is a
𝜑 ∈ C ∞ (𝑈; R) such that Im (𝜁 · d𝑧)| Λ = −𝜋 ∗ d𝜑, implying ℎ = 2𝑖1 𝜕𝑧 𝜑 in 𝑈. The
converse entailment is evident. □
Proposition 13.6.6 Let 𝜑 ∈ C ∞ (M; R) and

1
Λ 𝜑 = (𝑧, 𝜁) ∈ 𝑇 (1,0) M; 𝜁 = 𝜕𝑧 𝜑 (𝑧) .
2𝑖
1
The pullback to Λ 𝜑 of Re 𝜛 is equal to the real 2-form 2𝑖 𝜕𝜕𝜑.
Proof It suffices to prove the claim when M = ΩC , an open subset
of C𝑛 . The
pullback to Λ 𝜑 of Re 𝜛 = Re (d (𝜁d𝑧)) is equal to Re d 2𝑖1 𝜕𝜑 (𝑧) = 2𝑖1 𝜕𝜕𝜑. □
In a coordinate patch,
𝑛
1 1 ∑︁ 𝜕 2 𝜑
𝜕𝜕𝜑 = d𝑧 𝑗 ∧ d𝑧¯ 𝑘 . (13.6.9)
2𝑖 2𝑖 𝑗,𝑘=1 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

Corollary 13.6.7
The submanifold
Λ 𝜑 is totally real, i.e., R-symplectic, if and only
𝜕2 𝜑
if the matrix Im 𝜕𝑧 𝑗 𝜕𝑧¯ 𝑘 is nonsingular.
1≤ 𝑗,𝑘 ≤𝑛

13.6.2 Conormal bundles of nondegenerate smooth real hypersurfaces

Let S be a C ∞ real hypersurface in the complex-analytic manifold M; dimR S = 2𝑛−

1 (𝑛 = dimC M). We shall assume that S is locally closed, meaning that an arbitrary
point ℘◦ ∈ S has a neighborhood 𝑈 in M such that S ∩ 𝑈 = {℘ ∈ 𝑈; 𝜌 (℘) = 0}
with 𝜌 ∈ C ∞ (𝑈; R), d𝜌 (℘) ≠ 0 whatever ℘ ∈ 𝑈.
444 13 Elements of Symplectic Geometry

Lemma 13.6.8 Provided 𝑈 is sufficiently small there are holomorphic coordinates

𝑧 𝑗 = 𝑥 𝑗 + 𝑖𝑦 𝑗 ( 𝑗 = 1, ..., 𝑛), all vanishing at ℘, and a function 𝜌 ∈ C ∞ (𝑈; R) such
that S ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) = 0} and
𝑛−1
1 ∑︁ 𝜕 2 𝜌
𝜌 (𝑧) = −2𝑥 𝑛 + (0) 𝑧 𝑗 𝑧¯ 𝑘 + 𝑂 |𝑧| 3 . (13.6.10)
2 𝑗.𝑘=1 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

𝜕𝜌
∈ 𝑇 (1,0) (M) 𝑈 (see Sub-
Í𝑛
In the sequel we use the notation 𝜕 𝜌 = 𝑗=1 𝜕𝑧 𝑗 d𝑧 𝑗
section 9.4.5); we have d𝜌 = 2 Re 𝜕 𝜌.
Proof We can assume that M = C𝑛 and ℘ is the origin; thus 0 ∈ S ∩ 𝑈 =
{𝑧 ∈ 𝑈; 𝑓 (𝑧) = 0} for some 𝑓 ∈ C ∞ (𝑈; R) such that d 𝑓 (0) ≠ 0. A C-linear
change of variables allows us to posit that d 𝑓 (0) = 2d𝑥 𝑛 . In the notation

𝜕2 𝑓 𝜕2 𝑓
𝑎 𝑗,𝑘 = (0) , 𝑏 𝑗,𝑘 = (0) , 𝑗, 𝑘 = 1, ..., 𝑛,
𝜕𝑧 𝑗 𝜕𝑧 𝑘 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘
𝑛
∑︁ 𝑛
∑︁
𝐴 (𝑧) = 𝑎 𝑗,𝑛 𝑧 𝑗 , 𝐵 (𝑧) = 𝑏 𝑗,𝑛 𝑧 𝑗 ,
𝑗=1 𝑗=1

the Taylor expansion of 𝑓 about 0 reads

𝑛−1
© ∑︁
𝑓 (𝑧) = 2𝑥 𝑛 + Re ­ 𝑎 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + 𝐴 (𝑧) 𝑧 𝑛 ®
ª

« 𝑗,𝑘=1 ¬
𝑛−1
1 ∑︁
+ 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧¯ 𝑘 + Re (𝐵 (𝑧) 𝑧¯𝑛 ) + 𝑂 |𝑧| 3 .
2 𝑗,𝑘=1

Since 𝐵 (𝑧) 𝑧 𝑛 + 𝐵 (𝑧) 𝑧¯𝑛 = 2𝑥 𝑛 𝐵 (𝑧) we get

𝑛−1
© ∑︁
𝑓 (𝑧) = 2𝑥 𝑛 (1 + Re 𝐵 (𝑧)) + Re ­ 𝑎 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + ( 𝐴 (𝑧) − 𝐵 (𝑧)) 𝑧 𝑛 ®
ª

« 𝑗,𝑘=1 ¬
𝑛−1
1 ∑︁
+ 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧¯ 𝑘 + 𝑂 |𝑧| 3 .
2 𝑗,𝑘=1

If 𝑈 is sufficiently small we can define

𝑔 (𝑧) = (1 + Re 𝐵 (𝑧)) −1 𝑓 (𝑧) ∈ C ∞ (𝑈; R) ,

whence S ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝑔 (𝑧) = 0} and

13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 445

𝑛−1
∑︁
𝑔 (𝑧) = 2 Re ­𝑧 𝑛 + 𝑎 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + ( 𝐴 (𝑧) − 𝐵 (𝑧)) 𝑧 𝑛 ®
© ª

« 𝑗,𝑘=1 ¬
𝑛−1
1 ∑︁
+ 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧¯ 𝑘 + 𝑂 |𝑧| 3 .
2 𝑗,𝑘=1

The holomorphic change of variables in 𝑈 (possibly further contracted about 0)

𝑛−1
∑︁
𝑧𝑛 + 𝑎 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + ( 𝐴 (𝑧) − 𝐵 (𝑧)) 𝑧 𝑛 ⇝ 𝑧 𝑛
𝑗,𝑘=1

yields
𝑛−1
1 ∑︁
𝑔 (𝑧) = 2𝑥 𝑛 + 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧¯ 𝑘 + 𝑂 |𝑧| 3 .
2 𝑗,𝑘=1

The function 𝜌 (𝑧) = 𝑔 (−𝑧) fulfills the requirements. □

Remark 13.6.9 When the hypersurface S is of class C 𝜔 in a neighborhood of ℘◦

we can take 𝜌 ∈ C 𝜔 (𝑈; R). The same observation applies to all the forthcoming
developments.
𝜕𝜌 𝜕
Over S ∩ 𝑈 the tangent bundle 𝑇S is spanned by the vectors 𝑍 𝑗,𝑘 = 𝜕𝑧 𝜅 𝜕𝑧 𝑗 −
𝜕𝜌 𝜕
𝜕𝑧 𝑗 𝜕𝑧 𝑘 ( 𝑗, 𝑘 = 1, ..., 𝑛) and their complex conjugates 𝑍 𝑗,𝑘 . For a real 1-form 𝛼 =
Í𝑛 D E
ℓ=1 𝜁ℓ d𝑧 ℓ + 𝜁 ℓ d 𝑧¯ℓ to satisfy 𝛼, 𝑍 𝑗,𝑘 = 𝛼, 𝑍 𝑗,𝑘 = 0 for all 𝑗, 𝑘, it is necessary
𝜕𝜌
and sufficient that 𝜁ℓ = 𝜆 𝜕𝑧 ℓ
for some 𝜆 ∈ R and every ℓ = 1, ..., 𝑛. Here we
identify the conormal bundle 𝑁 ∗ S of S (Definition 9.3.4) with the vector subbundle
of 𝑇 (1,0) M (identified with the real cotangent bundle 𝑇 ∗ M) consisting of the points
(𝑧, 𝜆𝜕 𝜌), 𝑧 ∈ S ∩ 𝑈, 𝜆 ∈ R; we have dimR 𝑁 ∗ S = 2𝑛.
In the sequel 𝜋 : 𝑇 (1,0) M\0 −→ M shall denote the base projection.

Proposition 13.6.10 Let L be a conic real C ∞ submanifold of 𝑇 (1,0) M\0 such that
dimR L = 2𝑛 and let 𝛾 ◦ ∈ L. The following two properties are equivalent:
(a) There exist a conic neighborhood U of 𝛾 ◦ in 𝑇 (1,0) M\0 and a C ∞ real hyper-
surface S ⊂ 𝑈 = 𝜋 (U) such that U ∩ L = U ∩ 𝑁 ∗ S.
(b) There is a conic neighborhood U of ◦ (1,0) M\0 such that L ∩ U is
𝛾 in 𝑇
(1,0)
R-Lagrangian and dimR 𝑇𝛾 ◦ L ∩ 𝑇𝛾 ◦ 𝑇 𝜋 (𝛾 ◦ ) M = 1.

Proof Let 𝛾 ∈ L be arbitrary. Since L is conic, if Λ𝛾 denotes the straight line in

(1,0) (1,0)
𝑇𝛾 L tangent to the ray in 𝑇 𝜋 (𝛾) M through 𝛾 then Λ𝛾 ⊂ 𝑇𝛾 L ∩𝑇𝛾 𝑇 𝜋 (𝛾) M . To say

that Λ𝛾 ◦ = 𝑇𝛾 ◦ L ∩ 𝑇𝛾 ◦ 𝑇 𝜋(1,0) is the same as saying that rank 𝜋| L at 𝛾 ◦ has

◦
(𝛾 )
M
its maximum value, 2𝑛 − 1; this is therefore also true in a conic neighborhood U of
𝛾 ◦ . It is convenient to assume that 𝑈 = 𝜋 (U) is the (geometrically simple) domain
446 13 Elements of Symplectic Geometry

of holomorphic coordinates 𝑧 𝑗 and use these to identify 𝑇 (1,0) M 𝑈 with 𝑈 × C𝑛𝜁 ; we

will then also assume that 𝜁 ◦ = (0, ..., −1). Provided (𝑧, 𝜁) ∈ U =⇒ |𝜁𝜁 | − 𝜁 ◦ < 𝜈,
with 𝜈 > 0 sufficiently small, the set Σ = {(𝑧, 𝜁) ∈ L ∩ U; 𝜁 𝑛 = −1} will be a C ∞
submanifold of U and the fact that rank 𝜋 = 2𝑛 − 1 in 𝑈 will imply that 𝜋| Σ is
a diffeomorphism of Σ onto a real hypersurface S = 𝜋 (U ∩ L). Thus the latter
property holds under both hypotheses (a) and (b). We select 𝜌 ∈ C ∞ (𝑈; R) such
that S = {𝑧 ∈ 𝑈; 𝜌 (𝑧) = 0}, d𝜌 ≠ 0 at every pointÍof 𝑈.
(a)=⇒(b). The pullback of the one-form 𝜎 C = 𝑛𝑗=1 𝜁 𝑗 d𝑧 𝑗 to 𝑁 ∗ S|𝑈 is equal to
𝜆𝜕 𝜌. Since 2 Re 𝜕 𝜌 = d𝜌√we see that the pullback to L of the fundamental two-form
𝜛𝑛C to 𝑁 ∗ S is equal to −1d (𝜆 Im 𝜕 𝜌) and thus is purely imaginary. This proves
that 𝑁 ∗ S is R-Lagrangian.
(b)=⇒(a). Let S ∋ 𝑧 ↦→ 𝜁 (𝑧) ∈ Σ denote the inverse of 𝜋| Σ . Every vector
−1
𝜗 ∈ 𝑇𝑧 S can be lifted under 𝜋 as a vector tangent to L at (𝑧, 𝜆𝜁 (𝑧)) whatever
𝜆 > 0. The hypothesis that L is R-Lagrangian [for the 2-form (13.6.1) on C2𝑛 ]
and conic implies (by Proposition 13.3.26) that the restriction
Í𝑛 of the form Re 𝜎 C to
𝜕
𝑇(𝑧,𝜁 (𝑧)) L vanishes identically. As a consequence, if 𝜗 = 𝑗=1 𝛼 𝑗 𝜕𝑧 𝑗 + 𝛼¯ 𝑗 𝜕𝜕𝑧¯ 𝑗 ∈ 𝑇𝑧 S
Í
is regarded as an element of 𝑇(𝑧,𝜁 (𝑧)) L we have Re 𝑛𝑗=1 𝛼 𝑗 𝜁 𝑗 (𝑧) = 0. This means
Í𝑛
that Re 𝑗=1 𝜁 𝑗 (𝑧) d𝑧 𝑗 is orthogonal to 𝑇𝑧 S and therefore generates the conormal
space 𝑁 𝑧∗ S. □
Throughout the remainder of this section S shall denote a real hypersurface in M
defined in the local chart (𝑈, 𝑧1 , ..., 𝑧 𝑛 ) by the equation 𝜌 (𝑧) = 0, 𝜌 ∈ C ∞ (𝑈; R)
as in (13.6.10); we continue to take 𝜁 ◦ = (0, ..., 0, −1); U is a conic neighborhood
of (0, 𝜁 ◦ ) in C2𝑛
𝑧,𝜁 and 𝑈 = 𝜋 (U). Proposition 13.6.10 implies that the conormal
bundle 𝑁 ∗ S is a (conic, C ∞ ) R-Lagrangian submanifold of 𝑇 (1,0) M\0.

Proposition 13.6.11 The conormal bundle 𝑁 ∗ S is I-symplectic at (0, 𝜁 ◦ ) if and only

Í 𝜕2 𝜌
if the quadratic form 𝑛−1
𝑗,𝑘=1 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 (0) 𝑧 𝑗 𝑧¯ 𝑘 in C
𝑛−1 is nondegenerate at 0.

By saying that the bundle 𝑁 ∗ S is I-symplectic at 𝛾 ◦ = (0, 𝜁 ◦ ) we mean that

𝑇𝛾 ◦ (𝑁 ∗ S) is an I-symplectic linear subspace of 𝑇𝛾 ◦ 𝑇 (1,0) M [identified with
𝑇(0,𝜁 ◦ ) C2𝑛 ]. This implies that 𝑇𝛾 𝑁 ∗ S is I-symplectic for all 𝛾 in a conic neighborhood
of 𝛾 ◦ which we take to be U.
Proof We have (𝑧, 𝜆𝜕 𝜌 (𝑧)) ∈ 𝑁 ∗ S, 𝑧 ∈ S ∩𝑈; we can use 𝑧 𝑗 , 𝑧¯ 𝑘 (1 ≤ 𝑗, 𝑘 ≤ 𝑛 − 1),
𝑦 𝑛 , 𝜆 as coordinates in U ∩ 𝑁 ∗ S; we have

2𝑖d (𝜆 Im 𝜕 𝜌) = d𝜆 ∧ 𝜕 𝜌 − 𝜕¯ 𝜌 + 𝜆 𝜕 + 𝜕¯ 𝜕 𝜌 − 𝜕¯ 𝜌

or, explicitly,
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 447
𝑛
1 ∑︁ 𝜕 𝜌 𝜕𝜌
d (𝜆 Im 𝜕 𝜌) = d𝑥 𝑗 − d𝑦 𝑗 ∧ d𝜆
2 𝑗=1 𝜕𝑦 𝑗 𝜕𝑥 𝑗
𝑛
∑︁ 𝜕2 𝜌
+𝑖 d𝑧 𝑗 ∧ d𝑧¯ 𝑘 .
𝑗.𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

By (13.6.10), at (0, 𝜁 ◦ ) we have

𝑛−1
∑︁ 𝜕2 𝜌
d (𝜆 Im 𝜕 𝜌) = 𝑖 (0) d𝑧 𝑗 ∧ d𝑧¯ 𝑘 + d𝑦 𝑛 ∧ d𝜆.
𝑗.𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

𝜕2 𝜌
This (real) 2-form is nondegenerate if and only if the matrix 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 (0) is
1≤ 𝑗,𝑘<𝑛
nonsingular. □

Definition 13.6.12 We shall say that the hypersurface S is nondegenerate at ℘ ∈ S

if the conormal space 𝑁 ℘∗ S is I-symplectic.

Thus, in the local chart (𝑈, 𝑧1 , ..., 𝑧 𝑛 ), S is nondegenerate at 0 if and only if the
𝜕2 𝜌
matrix 𝜕𝑧 𝑗 𝜕𝑧¯ 𝑘 (0) is nonsingular.
1≤ 𝑗,𝑘<𝑛
Henceforth L will be a conic submanifold of 𝑇 (1,0) M\0, R-Lagrangian and I-
symplectic; this makes L totally real (Proposition 13.5.5). We limit the analysis to
a conic neighborhood U in 𝑇 (1,0) M\0 of 𝛾 ◦ ∈ L, with 𝑈 = 𝜋 (U) the domain
of holomorphic coordinates 𝑧 𝑗 used to identify 𝑇 (1,0) M\0 𝑈 with 𝑈 × (C𝑛 \ {0})
with 𝜁 𝑘 the dual coordinates in the fibers
n {𝑧} × C𝑛 ; we assume 𝛾 ◦ o = (𝑧◦ , 𝜁 ◦ ),
𝜁 ◦ = (0, ..., 0, −1), and U = 𝑈 × Γ, Γ = 𝜁 ∈ C𝑛 ; 𝜁 ≠ 0, |𝜁𝜁 | − 𝜁 ◦ < 𝜈 , 0 < 𝜈 < 1
(implying 𝜁 𝑛 ≠ 0 everywhere in Γ).
Henceforth “complex symplectomorphism” shall mean a symplectic isomor-
phism which is also a biholomorphism.

Lemma 13.6.13 Let L be a conic submanifold of 𝑇 (1,0) M\0, R-Lagrangian and

I-symplectic. If diam 𝑈 and 𝜈 are sufficiently small there is a local, homogeneous,
complex symplectomorphism 𝜒 of U onto another conic neighborhood of 𝛾 ◦ pre-
serving 𝛾 ◦ and fulfilling the following two requirements:

(1) dim 𝑇(𝑧 ◦ ,𝜁 ◦ ) ({0} × C𝑛 ) ∩ D𝜒 𝑇(𝑧 ◦ ,𝜁 ◦ ) L = 1 (D𝜒: differential of 𝜒);
(2) 𝜒 does not modify the functions 𝜁1 , ..., 𝜁 𝑛 in U.

The motivation for Condition (2) is to be found in applications in later chapters.

Proof For simplicity we take 𝑧 ◦ = 0. Since L is conic and 𝜁 ◦ = (0, ..., 0, −1) ∈ L
we have R 𝜕𝜕𝜉𝑛 ⊂ 𝑽 = 𝑇(0,𝜁 ◦ ) L ∩ 𝑇(0,𝜁 ◦ ) ({0} × C𝑛 ); we have 𝜕𝜂𝜕 𝑛 ∉ 𝑽, otherwise
we would have C 𝜕𝜁𝜕𝑛 ⊂ 𝑽, which is impossible since 𝑇(0,𝜁 ◦ ) L is totally real. If
𝑽 = R 𝜕𝜕𝜉𝑛 we take 𝜒 to be the identity map. Now suppose dimR 𝑽 ≥ 2. Condition
(2) demands that 𝜒 not involve any transformation 𝜏 of the variables 𝑧 𝑗 since the
448 13 Elements of Symplectic Geometry

action of the contragredient of D𝜏 would modify 𝜁1 , ..., 𝜁 𝑛 . Thus we are left with
transformations
𝜒 (𝑧, 𝜁) = (𝑧 + 𝑓 (𝜁) , 𝜁) , (13.6.11)
where 𝑓 = ( 𝑓1 , ..., 𝑓𝑛 ) ∈ O (Γ; C𝑛 ) is homogeneous of degree zero, 𝑓 (𝜁 ◦ ) = 0.
Since 𝜒 is a homogeneous symplectomorphism it must preserve the 1-form 𝜁 · d𝑧.
𝜕Φ
The latter is achieved by selecting 𝑓 · d𝜁 = 𝜕Φ, i.e., 𝑓 𝑗 = 𝜕𝜁 𝑗
( 𝑗 = 1, ..., 𝑛), Φ
holomorphic in Γ, homogeneous of degree 1; we shall take Φ (𝜁) = 𝜁 𝑛−1 𝐴 with
1 Í𝑛−1
𝐴 = 2 𝑗=1 𝑐 𝑗,𝑘 𝜁 𝑗 𝜁 𝑘 , 𝑐 𝑗,𝑘 = 𝑐 𝑘, 𝑗 ∈ C.
Suppose 𝑽 is spanned by 𝜕𝜕𝜉𝑛 and the real and imaginary parts of 𝜈 linearly
Í
independent vectors 𝜗 𝑗 = 𝑛−1 𝜕
𝑘=1 𝑏 𝑗,𝑘 𝜕𝜁𝑘 ∈ 𝑇(0,𝜁 ◦ ) ({0} × C ), 𝑗 = 1, ..., 𝜈 ≤ 𝑛 − 1
𝑛

(𝑏 𝑗,𝑘 ∈ C\ {0}). On the one hand we have

𝑛−1
𝜕 𝜕 ∑︁ 𝜕𝐴 𝜕
D𝜒 = − 𝜁 𝑛−2 ,
𝜕𝜁 𝑛 𝜕𝜁 𝑛 𝑗=1
𝜕𝜁 𝑗 𝜕𝑧 𝑗

𝜕 𝜕
whence D𝜒 𝜕𝜁𝑛 = 𝜕𝜁𝑛 ; on the other hand,
𝜁◦

𝑛−1
∑︁ 𝜕
D𝜒 𝜗 𝑗 = 𝜗 𝑗 + 𝜁 𝑛−1

𝑏 𝑗,𝑘 𝑐 𝑘,ℓ .
𝑘,ℓ=1
𝜕𝑧ℓ

We select the matrix 𝑐 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛−1 ∈ Sym (𝑛 − 1, C) such that

𝑛−1
∑︁ 𝜕
𝑏 𝑗,𝑘 𝑐 𝑘,ℓ ≠ 0 for every 𝑗 = 1, ..., 𝜈,
𝑘,ℓ=1
𝜕𝑧ℓ

ensuring that D𝜒 𝜗 𝑗 ∉ 𝑇(0,𝜁 ◦ ) ({0} × C𝑛 ). This means that 𝑇(0,𝜁 ◦ ) ({0} × C𝑛 ) ∩

D𝜒 𝑇(0,𝜁 ◦ ) L = R 𝜕𝜕𝜉𝑛 . □
By combining Lemma 13.6.13 with Propositions 13.6.10, 13.6.11, we obtain
Theorem 13.6.14 Let L be a conic C ∞ submanifold of 𝑇 (1,0) M\0, R-Lagrangian
and I-symplectic, and let 𝛾 ◦ ∈ L be arbitrary. There exist a nondegenerate real
hypersurface S in M, a conic neighborhood U of 𝛾 ◦ in 𝑇 (1,0) M\0 and a homo-
geneous complex symplectomorphism 𝜒 of U onto another conic neighborhood
of 𝛾 ◦ , preserving 𝛾 ◦ , leaving the functions 𝜁1 , ..., 𝜁 𝑛 unchanged and such that
𝜒 (L ∩ U) = 𝜒 (U) ∩ 𝑁 ∗ S.
Note that 𝑁 ∗ S is defined in U by the equations 𝜌 (𝑧) = 0 and

𝜕𝜌
𝜁ℓ − 𝜆 (𝑧) = 0, ℓ = 1, ..., 𝑛.
𝜕𝑧ℓ
𝜕𝜌 𝜕𝜌
Since 𝜕𝑥 𝑛
(0) = −2 [cf. (13.6.10)] we can replace 𝜆 by 2𝜉 𝑛 / 𝜕𝑥 𝑛
(𝑧); the latter
equations become
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 449

𝜕𝜌 𝜕𝜌
𝜁ℓ (𝑧) − 2𝜉 𝑛 (𝑧) = 0, ℓ = 1, ..., 𝑛. (13.6.12)
𝜕𝑥 𝑛 𝜕𝑧ℓ
We introduce the following complex vector fields, in a conic neighborhood U of
(0, 𝜁 ◦ ) in 𝑇 (1,0) M\0,
𝑛−1
∑︁ 𝜕 2 𝜌
𝜕 𝜕
𝐿𝑗 = + (0) , 𝑗 = 1, ..., 𝑛 − 1. (13.6.13)
𝜕 𝑧¯ 𝑗 𝑘=1 𝜕𝑧 𝑘 𝜕 𝑧¯ 𝑗 𝜕𝜁 𝑘

Lemma 13.6.15 The tangent space to 𝑁 ∗ S at (0, 𝜁 ◦ ) is spanned by Re 𝐿 𝑗 , Im 𝐿 𝑗

( 𝑗 = 1, ..., 𝑛 − 1), 𝜕𝑦𝜕𝑛 , 𝜕𝜕𝜉𝑛 .

Proof The vector fields listed in the statement are obviously linearly independent
and therefore the dimension of their span is equal to 2𝑛 = dimR 𝑁 ∗ S. By (13.6.10)
they annihilate the function 𝜌 (𝑧) at (0, 𝜁 ◦ ). Letting 𝐿 𝑗 and 𝐿 𝑗 act on the right-hand
sides in (13.6.12) yields
𝑛−1
!
𝜕2 𝜌

𝜕 ∑︁ 𝜕 𝜕𝜌 𝜕𝜌
+ (0) 𝜁ℓ (𝑧) − 2𝜉 𝑛 (𝑧)
𝜕 𝑧¯ 𝑗 𝑘=1 𝜕𝑧 𝑘 𝜕 𝑧¯ 𝑗 𝜕𝜁 𝑘 𝜕𝑥 𝑛 𝜕𝑧ℓ
𝜕2 𝜌 𝜕2 𝜌 𝜕𝜌 𝜕2 𝜌
= 𝜁ℓ (𝑧) − 2𝜉 𝑛 (𝑧) + (𝑧) (0) ,
𝜕 𝑧¯ 𝑗 𝜕𝑥 𝑛 𝜕𝑧ℓ 𝜕 𝑧¯ 𝑗 𝜕𝑥 𝑛 𝜕𝑧ℓ 𝜕 𝑧¯ 𝑗

𝑛−1
!
𝜕 ∑︁ 𝜕 2 𝜌 𝜕 𝜕𝜌 𝜕𝜌

+ (0) 𝜁ℓ (𝑧) − 2𝜉 𝑛 (𝑧)
𝜕𝑧 𝑗 𝑘=1 𝜕 𝑧¯ 𝑘 𝜕𝑧 𝑗 𝜕 𝜁¯𝑘 𝜕𝑥 𝑛 𝜕𝑧ℓ
𝜕2 𝜌 𝜕2 𝜌
= 𝜁ℓ (𝑧) − 2𝜉 𝑛 (𝑧) .
𝜕𝑧 𝑗 𝜕𝑥 𝑛 𝜕𝑧 𝑗 𝜕𝑧ℓ

The right-hand sides in these equations vanish when 𝑧 = 0, 𝜁 = 𝜁 ◦ since 𝜕𝑧

𝜕𝜌
𝑛
(0) =
𝜕2 𝜌 𝜕
(0, ..., 0, −1) and 𝜕𝑧 𝑗 𝜕𝑧ℓ (0) = 0 if 𝑗, ℓ ≤ 𝑛 − 1. This also implies that 𝜕 𝜉𝑛 annihilates
2
the right-hand sides in (13.6.12) at (0, 𝜁 ◦ ) and so does 𝜕
𝜕𝑦𝑛 since 𝜕 𝜌
𝜕𝑧ℓ 𝜕𝑦𝑛 (0) = 0 if
ℓ ≤ 𝑛 − 1. □

An R-linear basis of the complex tangent space 𝑇(𝑧,𝜁 ) 𝑇 (1,0) M , (𝑧, 𝜁) ∈ U,
is provided by the tangent vectors 𝜕𝑥𝜕 𝑗 , 𝜕𝑦𝜕𝑘 , 𝜕𝜕𝜉ℓ , 𝜕𝜂𝜕𝑚 , and 𝑥 𝑗 ,𝑦 𝑘 , 𝜉ℓ , 𝜂 𝑚 (1 ≤
𝑗, 𝑘, ℓ, 𝑚 ≤ 𝑛) form a Darboux system of coordinates for Im 𝜛. By Lemma 13.6.15
𝑬 = 𝑇(0,𝜁 ◦ ) 𝑁 ∗ S is the set of vectors

𝑛−1 ∑︁𝑛 𝑛−1

∑︁ 𝜕 𝜕 𝜕 ∑︁ 𝜕
𝜗= 𝛼𝑗 + 𝛽𝑗 + 𝐴𝑘 + 𝐵𝑘 (13.6.14)
𝑗=1
𝜕𝑥 𝑗 𝜕𝑦 𝑗 𝑘=1
𝜕𝜉 𝑘 𝑘=1 𝜕𝜂 𝑘

where 𝛼 𝑗 , 𝛽 𝑗 ∈ R ( 𝑗 = 1, ..., 𝑛),

450 13 Elements of Symplectic Geometry

𝑛−1
𝜕2 𝜌
∑︁
𝐴 𝑘 = Re (0) (𝛼ℓ − 𝑖𝛽ℓ ) ,
ℓ=1
𝜕𝑧 𝑘 𝜕 𝑧¯ℓ
𝑛−1 2
∑︁ 𝜕 𝜌
𝐵𝑘 = Im (0) (𝛼ℓ − 𝑖𝛽ℓ ) ,
ℓ=1
𝜕𝑧 𝑘 𝜕 𝑧¯ℓ

(𝑘 = 1, ..., 𝑛 − 1) and 𝛼𝑛 = 0, 𝐵𝑛 = 0. We use the complex coordinates 𝑤 𝑗 = 𝛼 𝑗 +𝑖𝛽 𝑗 ,

𝜃 𝑘 = 𝐴 𝑘 +𝑖𝐵 𝑘 in 𝑇(0,𝜁 ◦ ) 𝑇 (1,0) M ; then 𝑬 is the real linear subspace of C2𝑛
𝑤, 𝜃 defined
by the equations
𝑛−1
∑︁ 𝜕2 𝜌
𝜃𝑗 = (0) 𝑤¯ 𝑘 , 1 ≤ 𝑗 ≤ 𝑛 − 1, Re 𝑤 𝑛 = 0, Im 𝜃 𝑛 = 0; (13.6.15)
𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

(13.6.15) makes it self-evident that 𝑬 is totally real if and only if the matrix
2
𝜕 𝜌
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 1≤ 𝑗,𝑘<𝑛

is nonsingular. It also provides the following expression of the skew-symmetric

R-bilinear form on 𝑬 [cf. ( (13.5.1)]
𝑛−1
∑︁ 𝜕2 𝜌
− 𝑖𝜛𝑛C ((𝑤, 𝜃) , (𝑤 ′, 𝜃 ′)) = 2 Im (0) 𝑤 𝑗 𝑤¯ ′𝑘 + 𝛽𝑛 𝐴𝑛′ − 𝛽𝑛′ 𝐴𝑛 . (13.6.16)
𝑗,𝑘=1
𝜕𝑧 𝑘 𝜕 𝑧¯ 𝑗

If we put 𝑤 ′𝑘 = 𝑖𝑤 𝑘 for 𝑘 = 1, ..., 𝑛 − 1 (hence 𝜃 ′𝑗 = −𝑖𝜃 𝑗 if 𝑗 ≤ 𝑛 − 1), 𝑤 𝑛′ =

−𝑤 𝑛 , 𝜃 𝑛′ = 𝜃 𝑛 , we see that 2𝑖1 𝜛𝑛C ((𝑤, 𝜃) , (𝑤 ′, 𝜃 ′)) is equal to the real quadratic
form on 𝑬,
𝑛−1
∑︁ 𝜕2 𝜌
(0) 𝑤 𝑗 𝑤¯ 𝑘 + Im (𝑤 𝑛 𝜃 𝑛 ) . (13.6.17)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

For reference below we point out that 𝑬 contains the line R 𝜕𝜕𝜉𝑛 in 𝑇(0,𝜁 ◦ ) 𝑇 (1,0) M
tangent to the line (0, R𝜁 ◦ ) (itself contained in the conic set 𝑁 ∗ S); 𝑬 also contains
the line R 𝜕𝑦𝜕𝑛 transversal to 𝑇0 S [regarded as a linear subspace of 𝑇(0,𝜁 ◦ ) (𝑁 ∗ S)].
The union of these two lines, defined in 𝑬 by 𝑤 1 = · · · = 𝑤 𝑛−1 = 0, 𝑤 𝑛 𝜃 𝑛 = 0, is the
null space of (13.6.17).
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 451

13.6.3 Outer conormal bundles of boundaries of strictly pseudoconvex

domains

In this subsection we take the C ∞ hypersurface S of the preceding subsection to

be the boundary 𝜕Ω of a strictly pseudoconvex domain Ω (Definition 11.2.20) in
the complex-analytic manifold M. First we introduce a general concept from [Bony,
1969] useful in a variety of contexts besides the present one.

Definition 13.6.16 Let M be a real C ∞ manifold, 𝑭 a closed subset of M. By the

outer conormal bundle of 𝑭 we shall mean the subset 𝑁out ∗ 𝑭 of the cotangent

bundle 𝑇 M consisting of the points (℘ , 𝜆d𝜌 (℘ )) where ℘◦ ∈ 𝑭, 𝜆 > 0, and 𝜌

∗ ◦ ◦

is a real-valued C ∞ function in a neighborhood U of ℘◦ such that d𝜌 (℘◦ ) ≠ 0 and

𝜌 (℘) ≤ 𝜌 (℘◦ ) for all ℘ ∈ U ∩ 𝑭.

The smoothness assumption (for M and 𝜌) can be replaced by C 1 . Generally

speaking, for arbitrary closed sets 𝑭 and arbitrary points ℘◦ ∈ 𝑭, we have 𝑁out∗ 𝑭∩
∗ 1 ◦
𝑇℘◦ M = ∅, i.e., there are no C functions 𝑓 in a neighborhood U of ℘ such that
d𝜌 (℘◦ ) ≠ 0 and 𝑭 stays on one side of the hypersurface 𝜌 (℘) = 𝜌 (℘◦ ) in U (so the
outer conormal bundle is not a fiber bundle over 𝑭). Here we take 𝑭 = Ω∪𝜕Ω and
∗ 𝜕Ω rather than 𝑁 ∗ Ω; 𝑁 ∗ 𝜕Ω is an open half-line subbundle
use the notation 𝑁out out out
of the conormal bundle 𝑁 ∗ 𝜕Ω. We assume that there is a strictly plurisubharmonic
function (Definition 11.2.11) 𝜌 ∈ C ∞ (M; R) such that Ω = {℘ ∈ M; 𝜌 (℘) < 0};
any such defining function may be used in applying Definition 13.6.16 in this context.
We have dimR 𝑁out∗ 𝜕Ω = 2𝑛.

Proposition 13.6.17 Let Ω be a strictly pseudoconvex domain in C𝑛 . The outer

∗ 𝜕Ω is an R-Lagrangian and I-symplectic submanifold of
conormal bundle 𝑁out
𝑇 (1,0) M.

Proof It suffices to reason in a neighborhood in C𝑛 of an arbitrary point 𝑧◦ ∈ 𝜕Ω.

Í 𝜕2 𝜌
Definition 11.2.20, demands that the quadratic form 𝑛−1𝑗.𝑘=1 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 (0) 𝑧 𝑗 𝑧¯ 𝑘 in C
𝑛−1

be positive definite. Then Propositions 13.6.10 and 13.6.11 imply that the conormal
bundle 𝑁 ∗ 𝜕Ω is an R-Lagrangian and I-symplectic submanifold of 𝑇 (1,0) M. The
∗ 𝜕Ω of 𝑁 ∗ 𝜕Ω.
same is true of the open subset 𝑁out □
∗ 𝜕Ω is a totally real submanifold
Corollary 13.6.18 The outer conormal bundle 𝑁out
of 𝑇 (1,0) M.

Proof Combine Propositions 13.5.5 and 13.6.17. □

Proposition 13.6.19 Let L be a conic C ∞ submanifold of 𝑇 (1,0) M\0, R-Lagrangian

and I-symplectic. For L to be equal, in a conic neighborhood U in 𝑇 (1,0) M\0 of
𝛾 ◦ ∈ L, to the outer conormal bundle 𝑁out ∗ 𝜕Ω of a strictly pseudoconvex domain Ω

in M, it is necessary and sufficient that the following two conditions be satisfied:

452 13 Elements of Symplectic Geometry

(i) 𝑇𝛾 ◦ L ∩ 𝑇𝛾 ◦ 𝑇 𝜋(1,0)
◦
(𝛾 )
M is a straight line Λ𝛾 ◦ ;
(ii) the Hermitian form (13.6.17) on the real vector subspace 𝑬 = 𝑇𝛾 ◦ L of
𝑇𝛾 ◦ 𝑇 𝜋(1,0)
(𝛾 ◦ )
M is positive semidefinite and its null space is the union of two

real lines, Λ𝛾 ◦ and 𝑁 𝛾 ◦ such that 𝑁 𝛾 ◦ ∩ 𝑇𝛾 ◦ 𝑇 𝜋(1,0)
(𝛾 ◦ )
M = {0}.

In this statement 𝑇𝛾 𝑇 𝜋(1,0)
(𝛾)
M is the tangent space at 𝛾 to the fiber of 𝑇 (1,0) M
at 𝜋 (𝛾).
Proof Proposition 13.6.10 states that (i) is necessary and sufficient for the R-
Lagrangian and I-symplectic submanifold L to coincide, in a conic neighborhood
U of 𝛾 ◦ in 𝑇 (1,0) M\0, with the conormal bundle 𝑁 ∗ S of a nondegenerate real hy-
persurface S ⊂ 𝑈 = 𝜋 (U). Now reasoning in the set-up of Lemma 13.6.8 we apply
Proposition 13.6.11: for U ∩ 𝑁 ∗ S to be I-symplectic it is necessary and sufficient
Í 𝜕2 𝜌
that the quadratic form 𝑛−1 𝑗.𝑘=1 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 (0) 𝑧 𝑗 𝑧¯ 𝑘 in C
𝑛−1 be nondegenerate. With this

established, by Definition 11.2.20 and Proposition 13.6.17 (and possibly after further
contracting of 𝑈 about 0) there is a strictly pseudoconvex domain Ω in M such that
S = 𝑈 ∩ 𝜕Ω and Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} if and only if the quadratic form
Í𝑛−1 𝜕2 𝜌
𝑗.𝑘=1 𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 (0) 𝑧 𝑗 𝑧¯ 𝑘 in C
𝑛−1 is positive definite. This last point is equivalent to (ii)

with 𝑁 𝛾 ◦ = R 𝜕𝑦𝜕𝑛 , as seen in (13.6.17). □

The next statement is a strictly pseudoconvex variant of Theorem 13.6.14.
Remark 13.6.20 The outer conormal bundle of a nondegenerate hypersurface S in
C𝑛 which is not pseudoconvex does not make much sense since the two sides are not
distinguishable in the absence of additional requirements.
Theorem 13.6.21 Let L be a conic C ∞ submanifold of 𝑇 (1,0) M\0, R-Lagrangian
and I-symplectic, and let 𝛾 ◦ ∈ L be arbitrary. There exist a strictly pseudoconvex
domain Ω in M, a conic neighborhood U of 𝛾 ◦ in 𝑇 (1,0) M\0 and a homogeneous
complex symplectomorphism 𝜒 of U onto another conic neighborhood of 𝛾 ◦ in
𝑇 (1,0) M\0, preserving 𝛾 ◦ as well as the functions 𝜁1 , ..., 𝜁 𝑛 and transforming U ∩ L
∗ 𝜕Ω.
into 𝜒 (U) ∩ 𝑁out
Proof Using local coordinates we transfer the analysis to Euclidean space: we
assume that M = C𝑛𝑧 and 𝑇 (1,0) M = C2𝑛 ◦ ◦ ◦
𝑧,𝜁 ; 𝛾 = (0, 𝜁 ), 𝜁 = (0, ..., 0, −1)
and 𝑬 = 𝑇(0,𝜁 ◦ ) L. We apply Lemma 13.6.13: a symplectomorphism 𝜒 given
by (13.6.11) preserves 𝛾 ◦ , leaves the functions 𝜁1 , ..., 𝜁 𝑛 unchanged and is such
that D𝜒 (𝑬) ∩ 𝑇(0,𝜁 ◦ ) ({0} × C𝑛 ) = R 𝜕𝜕𝜉𝑛 [D𝜒: differential of 𝜒 at (0, 𝜁 ◦ )]. It is
immediately checked that D𝜒 = 𝜏𝐴, a linear symplectomorphism (13.5.17) with
𝐴 ∈ Sym∗ (𝑛, C) such that

𝜕 −1
𝑓 𝑗 (𝜁) = 𝜁 𝑛 𝐴𝜁 · 𝜁 , 𝑗 = 1, ..., 𝑛.
𝜕𝜁 𝑗

By Proposition 13.5.25 𝐴 can be chosen in such a way that the conditions (i) and (ii)
in Proposition 13.6.19 are satisfied. □
13.6 Real and Imaginary Symplectic Structures on Complex Manifolds 453

In the following
√ corollary we identify 𝑇 (1,0) C𝑛 \0 with C𝑛 × (C𝑛 \ {0}) and re-
gard R𝑛 × −1 (R𝑛 \ {0}) as an R-Lagrangian and I-symplectic submanifold of
C𝑛 × (C𝑛 \ {0}) (cf. Example 13.5.4); the claim is then an immediate application of
Theorem 13.6.21.
√
Corollary 13.6.22 Let (𝑥 ◦ , 𝑖𝜂◦ ) ∈ R𝑛 × −1 (R𝑛 \ {0}) ⊂ C2𝑛 be arbitrary. There
exist an open ball 𝔅𝜈 (𝑥 ◦ ) in C𝑛 centered at 𝑥 ◦ , a strictly pseudoconvex domain
Ω in 𝔅𝜈 (𝑥 ◦ ), a conic neighborhood U of (𝑥 ◦ , 𝑖𝜂◦ ) in 𝔅𝜈 (𝑥 ◦ ) × (C𝑛 \ {0}) and a
homogeneous complex symplectomorphism 𝜒 of U onto another conic neighborhood
of (𝑥 ◦ , 𝑖𝜂◦ ) in C𝑛 × (C𝑛\ {0}), preserving (𝑥 ◦, 𝑖𝜂◦ ) as well as the functions 𝜁1 , ..., 𝜁 𝑛 ,
√ ∗ 𝜕Ω.
and transforming U∩ R𝑛 × −1 (R𝑛 \ {0}) into 𝜒 (U) ∩ 𝑁out
Part IV
Stratification of Analytic Varieties and
Division of Distributions by Analytic
Functions
Chapter 14
Analytic Stratifications

In this part of the book we transition from manifolds to varieties; from this per-
spective, closed submanifolds are subvarieties without singularities. The purpose of
this chapter is to present the Lojasiewicz proof that analytic varieties admit locally
finite partitions made of (embedded) analytic submanifolds with special properties of
adherences (including the so-called Whitney property). Our presentation is intended
for readers who do not know much about the subject, and it stays at the very basic
level where our only tools come from the Weierstrass Preparation and Division The-
orems and the Riemann Extension Theorem (all proved in Section 14.3). It is then
a matter of handling Weierstrass polynomials (monic with respect to 𝑧 𝑛 , with the
nonleading coefficients that are analytic with respect to 𝑧 1 , ..., 𝑧 𝑛−1 and vanish at the
“central point”, usually the origin), factoring them and extracting new such polyno-
mials from their discriminants (after trimming, see Definition 14.4.1). We construct
local stratifications for real-analytic subvarieties (of any codimension) and touch
upon the complex-analytic ones solely when they are complex “hypersurfaces”, i.e.,
their complex codimension is equal to one. Part of our purpose is to convince the
beginner that stratifying varieties is a rather elementary operation, easily done “by
hand”. The main result of the chapter, Theorem 14.5.12, suffices for our needs in the
following chapter, devoted to the proof of the important Lojasiewicz inequality and
the division of distributions by C 𝜔 functions.
The Lojasiewicz results on these and related matters go much farther than what
can be found in this book, dealing with (among other things) the triangulation of
semianalytic sets. This class of sets comes into play, primarily, because the singular
part of a real-analytic variety, in general, is not an analytic variety (see the example
in Subsection 14.2.4). Semianalytic sets are given a brief introduction in the last
section of this chapter; their stratifications are shown to be a fairly straightforward
extension of that of varieties.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 457
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_14
458 14 Analytic Stratifications

14.1 Analytic Stratifications and Stratifiable Sets

14.1.1 The regular and singular parts of a subset of an analytic

manifold

In this section the word “analytic” may be understood in the complex or the real
sense. Thus M will denote an analytic (meaning either C 𝜔 or complex-analytic)
manifold, countable at infinity and often connected. We set dim M = 𝑛 with the
dimension relative to the field K = R in the C 𝜔 case and to K = C in the complex-
analytic case. Unless specified otherwise, in what follows 𝑬 will be an arbitrary
subset of M.
Local charts in M will be denoted by (U, 𝑥1 , ..., 𝑥 𝑛 ) even when the 𝑥𝑖 are complex
coordinates; later, on occasions where distinction or emphasis is needed, we go back
to the notation 𝑧𝑖 for the complex coordinates. Recall that to say that ℎ is an analytic
function in the domain U is to say that each point ℘ ∈ U has a neighborhood in
which ℎ is equal to the sum of its Taylor expansion about ℘ in the coordinates 𝑥 𝑗 .
This implies that if ℘ is a zero of infinite order of ℎ then ℎ ≡ 0 in U. We write A (U)
to mean C 𝜔 (U; R) in the real case and O (U) in the complex case; we underline
the fact that, in the C 𝜔 case, the elements of A (U) will be real-valued.
Here, as in this whole chapter, the meaning of “analytic submanifold” is given by
Definition 9.1.7, i.e., that of “embedded analytic submanifold” (cf. Remark 9.1.10).
By the dimension of an analytic submanifold we shall mean the highest dimension
of its connected components.
We denote by 𝑆 the closure of a subset 𝑆 of a topological space 𝑋. Our use of
the term boundary of 𝑆 will vary. When dealing with an open subset 𝑆 of M or of
one of its submanifolds, L, it will mean 𝜕𝑆 = 𝑆\𝑆. But note that, according to this
definition, 𝜕𝑆 = ∅ means that 𝑆 is closed. In frequent cases, by the boundary of
𝑆 ⊂ L we will mean the boundary of the largest subset of L open in L and contained
in 𝑆 (the interior of 𝑆 relative to L). The usual definition of the boundary, 𝑆 ∩ 𝑋\𝑆,
will be of little use to us, since 𝑋\𝑆 = 𝑋 for most sets 𝑆 we will be interested in.

Definition 14.1.1 A point ℘ ∈ 𝑬 is said to be a regular point1 of 𝑬 if there is a

neighborhood U of ℘ in M such that 𝑬 ∩ U is an analytic submanifold of M. If
℘ ∈ 𝑬 is not a regular point of 𝑬 it is said to be a singular point of 𝑬. The set of
regular points of 𝑬 is called the regular part of 𝑬 and shall be denoted by ℜ (𝑬);
its complement with respect to 𝑬, 𝑬\ℜ (𝑬), is called the singular part of 𝑬 and
shall be denoted by 𝔖 (𝑬).
For each 𝑘 ∈ Z+ we shall denote by ℜ𝑘 (𝑬) the union of all the connected
components of ℜ (𝑬) that have dimension 𝑘.

1 In this part of the book “regular” will have a meaning very feebly related to its meaning in the
preceding part.
14.1 Analytic Stratifications and Stratifiable Sets 459

Obviously, ℜ (𝑬) is an analytic submanifold of M; it is an open subset of 𝑬:

there is an open subset U of M such that U ∩ ℜ (𝑬) = U ∩ 𝑬. If 𝑬 is closed in M
the same is true of 𝔖 (𝑬). [In general ℜ (𝑬) is not closed: e.g., when 𝑬 is the union
of two distinct linear subspaces of R𝑛 .] Needless to say, for a generic subset 𝑬 of M
we have ℜ (𝑬) = ∅.
We now state a property of the regular part repeatedly used in the sequel.

Proposition 14.1.2 Let 𝑬 be an arbitrary subset of M and ℭ1 , ℭ2 two connected

components of ℜ (𝑬); then ℭ1 ∩ 𝜕ℭ2 = ∅.

Proof If ℭ1 = ℭ2 then ℭ1 ∩ ℭ1 \ℭ1 = ∅. Suppose ℭ1 ≠ ℭ2 . There is an open
subset U of M such that 𝑬 ∩ U = ℭ1 . If ℭ1 ∩ ℭ2 ≠ ∅ then U ∩ ℭ2 ≠ ∅ and therefore
ℭ1 ∩ ℭ2 ≠ ∅, an impossibility. □

Corollary 14.1.3 We have ℜ (𝑬) ∩ 𝜕ℜ (𝑬) = ∅.

14.1.2 Basics on local analytic stratifications

As before let 𝑬 be a subset of the analytic manifold M.

Definition 14.1.4 By an analytic stratification of 𝑬 in M we shall mean a partition

Ø
𝑬= S𝜄 (14.1.1)
𝜄 ∈𝐼

where each S 𝜄 is a connected analytic submanifold of M and the following conditions

are satisfied:
(LocFin) Every compact subset of M intersects at most finitely many submanifolds
S𝜄 .
(Strat) Whatever 𝜄, 𝜄′ ∈ 𝐼, 𝜄 ≠ 𝜄′, S 𝜄′ ∩ S 𝜄 ≠ ∅ implies S 𝜄′ ⊂ S 𝜄 \S 𝜄 = 𝜕S 𝜄 and
dim S 𝜄′ < dim S 𝜄 .
We shall refer to the submanifolds S 𝜄 as the strata of the analytic stratification
(14.1.1).
By a local analytic stratification of 𝑬 we shall mean a system (U, {S 𝜄 } 𝜄 ∈𝐼 ) with
U a domain in M, the index set 𝐼 finite and the S 𝜄 connected analytic submanifold
of U satisfying (Strat) and such that
Ø
𝑬∩U = S𝜄 . (14.1.2)
𝜄 ∈𝐼

We shall refer to U as the domain of the local analytic stratification (14.1.2).

In the terminology used in the sequel we do not preclude that all the strata S 𝜄 be
empty, the case 𝑬 ∩ U = ∅.
460 14 Analytic Stratifications

Definition 14.1.5 A subset 𝑬 of M shall be called stratifiable if M admits a

covering by domains of local analytic stratifications of 𝑬.

Stratifiable sets need not be defined by analytic equations.

Example 14.1.6 Take 𝑬 to be the subset of R2 defined by the equation 𝑥2 =

2

exp −1/𝑥1 ; we have 𝔖 (𝑬) = {0}, ℜ (𝑬) = 𝑬\ {0}. Define S± = {𝑥 ∈ 𝑬; 𝑥1 ≷ 0};
{0}, S+ and S− are connected analytic submanifolds of R2 and 𝜕S+ = 𝜕S− = {0}.
Thus {0} ∪ S+ ∪ S− is an analytic stratification of 𝑬 with domain R2 .

In the applications it is often helpful to strengthen the class of stratifications we

deal with. We introduce the following

Definition 14.1.7 The local stratification (14.1.2) shall be said to satisfy the Whitney
condition if the following holds:

(Whit) Whatever 𝜄, 𝜄′ ∈ 𝐼, S 𝜄′ ⊂ S 𝜄 implies 𝑇S 𝜄′ ⊂ 𝑇S 𝜄 .

We have denoted by 𝑇S 𝜄 the closure of the tangent bundle 𝑇S 𝜄 in the tangent

bundle 𝑇M.

Definition 14.1.8 A stratifiable subset 𝑬 of M shall be called Whitney stratifiable

if M admits a covering by domains of local analytic stratifications of 𝑬 all of which
satisfy the Whitney condition.

Proposition 14.1.9 Let (14.1.1) be an analytic stratification of a closed subset 𝑬 of

M. We have 𝑬 = ℜ (𝑬) and 𝔖 (𝑬) = 𝜕ℜ (𝑬).

Proof We reason locally, in the domain U of a local stratification (14.1.2). Let ℜ◦ be

the union of all the local strata S 𝜄 (𝜄 ∈ 𝐼) such that S 𝜄 ∩ S𝜅 = ∅ for every 𝜅 ∈ 𝐼, 𝜅 ≠ 𝜄;
we have ℜ◦ ⊂ ℜ (𝑬) [in general, ℜ◦ ≠ ℜ (𝑬) ∩ U]. If S𝜆 ⊄ ℜ (𝑬) (𝜆 ∈ 𝐼) we have
S𝜆 ∩ S𝜅1 ≠ ∅ for some 𝜅 1 ∈ 𝐼, 𝜅 1 ≠ 𝜆, implying S𝜆 ⊂ S𝜅1 and dim S𝜆 < dim S𝜅1
by (Strat). Assuming that S𝜅1 ⊄ ℜ (𝑬) we can find S𝜅2 (𝜅 2 ∈ 𝐼, 𝜅 2 ≠ 𝜅1 ) such that
S𝜅1 ⊂ S𝜅2 and dim S𝜅1 < dim S𝜅2 . We can repeat this reasoning until we reach (for
dimensional reasons or because the index set 𝐼 is finite) a stratum S𝜅𝑟 ⊂ ℜ◦ (𝜅𝑟 ∈ 𝐼)
such that S𝜅 𝑗 ⊂ S𝜅𝑟 , 𝑗 < 𝑟, whence S𝜆 ⊂ S𝜅𝑟 . □

Lemma 14.1.10 Let 𝑬 be a stratifiable closed subset of M, ℭ a connected component

of ℜ (𝑬) and (U, {S 𝜄 } 𝜄 ∈𝐼 ) an arbitrary local analytic stratification of 𝑬. Given 𝜄 ∈ 𝐼
arbitrarily the following properties hold:
(1) If S 𝜄 ∩ ℭ ≠ ∅ and dim S 𝜄 ≥ dim ℭ then dim S 𝜄 = dim ℭ and S 𝜄 ⊂ ℭ.
(2) If S 𝜄 ∩ ℭ ≠ ∅ then S 𝜄 ⊂ ℭ.

Proof Proof of (1): Suppose S 𝜄 ∩ℭ ≠ ∅ and dim S 𝜄 ≥ dim ℭ. There is an open subset
𝜔 of M such that ℭ = 𝑬 ∩ 𝜔 (by Definition 14.1.1). We deduce that S 𝜄 ∩ 𝜔 ⊂ ℭ and
therefore dim S 𝜄 = dim ℭ. Suppose we had S 𝜄 ∩ S𝜅 ≠ ∅ for some 𝜅 ∈ 𝐼, 𝜅 ≠ 𝜄; we
14.1 Analytic Stratifications and Stratifiable Sets 461

would have S 𝜄 ⊂ 𝜕S𝜅 , dim S𝜅 > dim ℭ, and an open subset of ℭ would be contained
in 𝜕𝑆 𝜅 which is impossible. Thus S 𝜄 ∩ (𝑬 ∩ U) \𝑆 𝜄 = ∅, implying S 𝜄 ⊂ ℜ (𝑬)
hence S 𝜄 ⊂ ℭ.
Proof of (2): Let 𝐼ℭ be the subset of 𝐼 consisting of the indices 𝜅 such that
S𝜅 ∩ ℭ ≠ ∅ and dim S𝜅 ≥ dim ℭ. By (1), 𝜅 ∈ 𝐼ℭ implies S𝜅 ⊂ ℭ. Since 𝐼 is finite
Ð
we have U ∩ ℭ ⊂ 𝜅 ∈𝐼ℭ S𝜅 . If S 𝜄 ∩ ℭ ≠ ∅ there is a 𝜅 ∈ 𝐼ℭ such that S 𝜄 ∩ S𝜅 ≠ ∅,
whence, by (Strat), S 𝜄 ⊂ S𝜅 ⊂ ℭ. □

Theorem 14.1.11 Let 𝑬 be a stratifiable closed subset of M. The natural partition

of the regular part ℜ (𝑬) of 𝑬, meaning the partition of ℜ (𝑬) into its connected
components, is an analytic stratification of ℜ (𝑬) (Definition 14.1.4). The set of
connected components of ℜ (𝑬) is countable.

Proof Let (14.1.2) be an arbitrary local analytic stratification of 𝑬 and ℭ be a

connected component of ℜ (𝑬) such that ℭ ∩ U ≠ ∅. Since there is an open subset
𝜔 of M such that ℭ = 𝑬 ∩𝜔 we have ℭ∩ℜ (𝑬) \ℭ = ∅. Since the index set 𝐼 is finite,
there is a local stratum S 𝜄 such that S 𝜄 ∩ ℭ ≠ ∅ and dim S 𝜄 ≥ dim ℭ. By Lemma
14.1.10 we have S 𝜄 ⊂ ℭ. We deduce that the number of connected components of
ℜ (𝑬) intersecting U does not exceed the number of local strata S 𝜄 . This proves that
the natural partition of ℜ (𝑬) is locally finite which, in turn, implies that the set of
connected components of ℜ (𝑬) is countable since the manifold M is countable at
infinity. □

Remark 14.1.12 The property that ℜ (𝑬) has countably many connected compo-
nents is also a direct consequence of the classical Lindelöf Lemma that every open
covering of a topological space with a countable basis has a countable open subcov-
ering (see [Kelley, 1955], p. 49).

In the sequel 𝑬 will most often be a closed subset of M; the connected components
of ℜ (𝑬) shall be denoted by ℭ 𝑗 , 𝑗 = 0, 1...; thus
Ø
ℜ (𝑬) = ℭ𝑗. (14.1.3)
𝑗 ∈Z+

Keep in mind that

ℭ 𝑗 ∩ ℭ 𝑘 = ∅ if 𝑗 ≠ 𝑘. (14.1.4)
Ð
Corollary 14.1.13 If 𝑬 is a stratifiable closed subset of M then 𝑬 = 𝑗 ∈Z+ ℭ𝑗.

Proof By Proposition 14.1.9 we have 𝑬 = ℜ (𝑬). By Theorem 14.1.11 an arbitrary

point of M has a neighborhood that intersects only finitely many submanifolds ℭ 𝑗 ;
Ð
if U is such a neighborhood then U ∩ ℜ (𝑬) = U ∩ 𝑗 ∈𝐽 ℭ 𝑗 with 𝐽 a finite subset
of Z+ . □

Remark 14.1.14 When 𝑬 is a stratifiable closed subset of M there is a covering of

M, not just of ℜ (𝑬), by domains of local analytic stratifications of ℜ (𝑬).
462 14 Analytic Stratifications

With the notation (14.1.3) we introduce the open sets

Ø
N 𝑗 = M\ ℭ𝑘 (14.1.5)
𝑘 ∈Z+ ,𝑘≠ 𝑗

( 𝑗 ∈ Z+ ); we have
ℭ 𝑗 ⊂ N𝑗 , ℭ 𝑗 ∩ N𝑗 = 𝑬 ∩ N𝑗 . (14.1.6)
In general, ℭ 𝑗 ⊄ N 𝑗 . We can state
Proposition 14.1.15 For each 𝑗 ∈ Z+ the analytic submanifold ℭ 𝑗 is the regular
part of the closed subset ℭ 𝑗 ∩ N 𝑗 of N 𝑗 .

14.2 Analytic Subvarieties

14.2.1 Analytic Subvarieties. Definition. Basic properties

As before, let M denote an analytic (meaning either C 𝜔 or complex-analytic) man-

ifold, countable at infinity. Below we use the notation 𝔅𝑅(𝑛) to denote the open ball
in K𝑛 centered at the origin and having radius 𝑅 > 0.
Definition 14.2.1 A set 𝑽 ⊂ M will be called an analytic subvariety of M if M can
be covered by open sets U with the following property: there is a family Φ (𝑽, U)
of functions defined and analytic in U such that

𝑽 ∩ U = {℘ ∈ U; ∀ 𝑓 ∈ Φ (𝑽, U) , 𝑓 (℘) = 0} .

Occasionally, we abbreviate the terms “analytic (sub)variety” to “(sub)variety”.

A subvariety 𝑽 is a closed subset of M. If 𝑽 𝑗 ( 𝑗 = 1, 2) are two subvarieties, each
defined in U as the set of common zeros of the functions belonging to a family
Φ 𝑽 𝑗 , U ⊂ A (𝑈), then (𝑽 1 ∪ 𝑽 2 ) ∩ U is the set of common zeros of all the
products 𝑓1 𝑓2 with 𝑓 𝑗 ∈ Φ 𝑽 𝑗 , U ; we can take Φ (𝑽 1 ∩ 𝑽 2 , U) = Φ (𝑽 1 , U) ∪
Φ (𝑽 2 , U). It ensues that arbitrary intersections and finite unions of subvarieties are
subvarieties.
Note that the cardinality of the set Φ (𝑽, U) is arbitrary; below, we shall see that
it can always be taken to be finite provided U is suitably small. The case Φ (𝑽, U) =
A (U) or, which amounts to the same, Φ (𝑽, U) = {1}, is not precluded, nor is
the case Φ (𝑽, U) = {0}: the former equates to U ∩ 𝑽 = ∅, the latter to U ⊂ 𝑽,
implying (by Corollary 14.2.5 below) that 𝑽 contains all the connected components
of M that intersect U.
Remark 14.2.2 When K = R and U is sufficiently small we may select Φ (𝑽, U)
to consist of a single function 𝑓 . Indeed, since Φ (𝑽, U) can be a finite set of
functions 𝑓 𝑗 ∈ C 𝜔 (U; R), 𝑗 = 1, ..., 𝑁, as will be shown below, we may then take
𝑓 = 𝑓12 + · · · + 𝑓 𝑁2 .
14.2 Analytic Subvarieties 463

Obviously, Φ (𝑽, U) can also be taken to be an ideal in the ring A (U). We shall
denote by ℑ (𝑽, U) the ideal of all the functions 𝑓 ∈ A (U) such that 𝑓 (℘) = 0
whatever ℘ ∈ 𝑽 ∩ U. In general, it is not true that

𝑽 ∩ U = {℘ ∈ U; ∀ 𝑓 ∈ ℑ (𝑽, U) , 𝑓 (℘) = 0} ; (14.2.1)

(14.2.1) is true, however, provided U is sufficiently small, a direct consequence of

the Noetherian properties of the ring of analytic
functions (see Proposition 14.3.11
below) as are the following claims. Let 𝑽 𝑗 𝑗=1,2,... be a sequence of analytic
subvarieties of M.
∞
Ù
(1) If 𝑽 = 𝑽 𝑗 then every ℘◦ ∈ 𝑽 is contained in an open subset U of M such
𝑗=1
that
𝑁
Ù
∃𝑁 < +∞, 𝑽 ∩ U = 𝑽 𝑗 ∩ U. (14.2.2)
𝑗=1

(2) If 𝑽 1 ⊃ · · · ⊃ 𝑽 𝑗 ⊃ · · · then to each compact subset K of M there is a positive

integer 𝑁 such that 𝑽 𝑗 ∩ K = 𝑽 𝑁 ∩ K for all 𝑗 > 𝑁.
Ù∞
(3) If 𝑽 𝑗 = ∅ then to each compact subset K of M there is a positive integer 𝑁
𝑗=1
𝑁
Ù
such that K ∩ 𝑽 𝑗 = ∅.
𝑗=1

Let 𝔅1 denote the open unit interval (resp., disk) in K = R (resp., K = C).

Lemma 14.2.3 Let M be a connected analytic manifold and 𝑽 ⫋ M a subvariety.

If ℘◦ ∈ 𝑽 there is an analytic embedding 𝔅1 ∋ 𝑡 ↦→ ℘ (𝑡) ∈ M such that ℘ (0) = ℘◦
and ℘ (𝑡) ∈ M\𝑽 for every 𝑡 ≠ 0.

Proof Let (U, 𝑥1 , ..., 𝑥 𝑛 ) be a local chart in M with 𝑥 𝑗 = 0 at ℘◦ for every 𝑗 =

1, ..., 𝑛. Assuming that the domain U is sufficiently small there is an 𝑓 ∈ A (U) that
does not vanish identically and is such that 𝑥 = (𝑥1 , ..., 𝑥 𝑛 ) ∈ 𝑽 ∩ U =⇒ 𝑓 (𝑥) = 0.
Possibly after a linear change of coordinates we can assume that 𝑥 𝑛 = 0 is an isolated
zero of 𝑓 (0, ..., 0, 𝑥 𝑛 ). This means that for a sufficiently small number 𝑟 > 0 the
map 𝑡 ↦→ (0, ..., 0, 𝑟𝑡) satisfies the requirement in the statement. □

Corollary 14.2.4 If no connected component of M is contained in 𝑽 then M\𝑽 is

an open and dense subset of M.

Corollary 14.2.5 Let L be a connected analytic submanifold of M. If there is an

open subset U of M such that ∅ ≠ U ∩ L ⊂ 𝑽 then L ⊂ 𝑽.

Proof If L ⊄ 𝑽 then L ∩ 𝑽 is a proper (i.e., ≠ L) subvariety of L. But no subset

of U ∩ L analytically diffeomorphic to 𝔅1 can intersect L\ (L ∩ 𝑽), contradicting
Lemma 14.2.3. □
464 14 Analytic Stratifications

Provided it is closed, an analytic submanifold (Definition 9.1.7) is a subvariety,

being locally defined by the vanishing of finitely many analytic functions (whose
differentials are linearly independent); as a consequence, the intersection of finitely
many closed analytic submanifolds is also a subvariety.
One can define an analytic variety 𝑽 in the abstract, without any reference to
an “ambient” analytic manifold M, by using coverings by domains of local charts
analytically equivalent to some open ball 𝔅𝑅(𝑛) with the standard coherence conditions
on overlaps (or else by using the sheaf of germs of analytic functions). But classical
theorems allow us, under reasonable hypotheses, to embed 𝑽 in a Euclidean space
K 𝑁 of sufficiently high dimension 𝑁 (see [Grauert, 1958]). In any event, the topics
discussed in this book almost always presume that 𝑽 is a subvariety of an “ambient”
analytic manifold.
Analysis about or on a subvariety 𝑽 frequently requires shrinking the neigh-
borhoods about a point ℘◦ ∈ 𝑽. It is therefore natural to rely on the concept of
the germ of the subvariety 𝑽 at ℘◦ , commonly denoted by (𝑽, ℘◦ ), but often, in
the sequel, if there is no risk of confusion, by bold-faced V℘◦ . In accordance with
Definition 1.2.7 (𝑽, ℘◦ ) is an equivalence class of pairs (𝑽 U , U) consisting
of a
neighborhood U of ℘◦ and of a subvariety 𝑽 U of U, with two such pairs 𝑽 U 𝑗 , U 𝑗
( 𝑗 = 1, 2) being equivalent if there is a third neighborhood U3 ⊂ U1 ∩ U2 such that
𝑽 U1 ∩ U3 = 𝑽 U2 ∩ U3 . If 𝑽 𝑗 , ℘◦ ( 𝑗 = 1, 2) are two germs of subvarieties at the
same point the meaning of the inclusion (𝑽 1 , ℘◦ ) ⊂ (𝑽 2 , ℘◦ ) is obvious: there exist
representatives 𝑽 𝑗, U , U of each of these germs such that 𝑽 1,U ⊂ 𝑽 2, U .
The concept of irreducible varieties is important.

Definition 14.2.6 An analytic variety 𝑽 ⊂ M is said to be reducible in M if there

are two analytic varieties 𝑽 𝑗 ⊂ M, 𝑗 = 1, 2, such that 𝑽 = 𝑽 1 ∪ 𝑽 2 and 𝑽 ≠ 𝑽 1 ,
𝑽 ≠ 𝑽 2 . An analytic variety which is not reducible is said to be irreducible.

The same terminology can be extended to germs of varieties. Explicitly, to say that
a germ of variety 𝑽 at ℘◦ is irreducible is to say that there is a basis of neighborhoods
of ℘◦ in M, {U𝑘 } 𝑘=1,2,... , and for each 𝑘 a representative (U𝑘 , 𝑽 𝑘 ) of the germ 𝑽 ℘◦
such that 𝑽 𝑘 is an irreducible subvariety of U𝑘 .
In the complex-analytic class the following result is useful. For a proof see
[Gunning and Rossi, 1965], end of Ch. III.

Theorem 14.2.7 The following two properties of a complex-analytic subvariety 𝑽

are equivalent: 𝑽 is irreducible; the regular part of 𝑽 (Definition 14.1.1), ℜ (𝑽), is
connected.

The same statement is not true in the real class, as shown in the Whitney umbrella
𝑾 R (see the last subsection of this section).
14.2 Analytic Subvarieties 465

14.2.2 Regular and singular parts of analytic varieties

Let ℜ (𝑽) [resp., 𝔖 (𝑽)] denote the regular (resp., singular) part of the analytic
subvariety 𝑽 of M. We prove directly, for an analytic variety, the property of closed
stratifiable sets stated in Proposition 14.1.9.

Proposition 14.2.8 The regular part ℜ (𝑽) is open and dense in 𝑽 and 𝔖 (𝑽) =
𝜕ℜ (𝑽).

Proof Let U be an open subset of M such that

𝑽 ∩ U = {℘ ∈ U; ∀ 𝑓 ∈ Φ (𝑽, U) , 𝑓 (℘) = 0} ≠ ∅.

There is a ℘◦ ∈ 𝑽 ∩ U such that the maximum of the rank of the map U ∋ ℘ ↦→

( 𝑓1 (℘) , ..., 𝑓𝑟 (℘)) ∈ K𝑟 , as ( 𝑓1 , ..., 𝑓𝑟 ) ranges over all finite subsets of Φ (𝑽, U),
reaches its maximum in 𝑽 ∩ U at ℘◦ ; we have ℘◦ ∈ ℜ (𝑽) and there is an open
neighborhood of ℘◦ in 𝑽 ∩ U where the rank remains constant. The claim ensues.□

Proposition 14.2.9 If we regard a complex-analytic manifold M as a C 𝜔 real mani-

fold and the complex-analytic subvariety 𝑽 of M as a C 𝜔 subvariety then the regular
parts of 𝑽 in the real and in the complex sense are the same.

Proof A regular point in the complex sense is obviously regular in the real sense. Let
U be a domain in M of local complex analytic coordinates 𝑧 𝑗 = 𝑥 𝑗 +𝑖𝑦 𝑗 , 𝑗 = 1, ..., 𝑛 =
dim M, such that 𝑽 ∩U is defined in U by the equations 𝑓1 (𝑥, 𝑦) = · · · = 𝑓𝑚 (𝑥, 𝑦) =
0, where the functions 𝑓 𝑗 ∈ C 𝜔 (U) are real-valued and d 𝑓1 ∧ · · · ∧ d 𝑓𝑚 ≠ 0 at
every point of U; thus dimR 𝑽 ∩ U = 𝑛 − 𝑚. If U is sufficiently small, by applying
the Implicit Function Theorem we can find indices 𝑖1 < · · · < 𝑖 𝑝 , 𝑗 1 < · · · < 𝑗 𝑞
(𝑝 + 𝑞 = 𝑚), such that 𝑽 ∩ U is defined in U by the equations 𝑥𝑖𝜆 = 𝑦 𝑗𝜇 = 0,
𝜆 = 1, ..., 𝑝, 𝜇 = 1, ..., 𝑞. Let ℎ ∈ O (U), ℎ ≡ 0 on 𝑽 ∩ U; we have
𝑝
∑︁ 𝑞
∑︁
ℎ (𝑧) = 𝑐 𝑖𝜆 (𝑧, 𝑧¯) 𝑥 𝑖𝜆 + 𝑐 ′𝑗𝜇 (𝑧, 𝑧¯) 𝑦 𝑗𝜇 (14.2.3)
𝜆=1 𝜇=1

with 𝑐 𝑖𝜆 ,𝑐 ′𝑗𝜇 ∈ C 𝜔 (U) (in general, complex-valued). We let 𝜕𝑧¯𝛼 (𝛼 ∈ Z+𝑛 arbitrary)
act on (14.2.3). With ⟨ℓ⟩ denoting the 𝑛-tuple (𝜄1 , ..., 𝜄𝑛 ) such that 𝜄𝛾 = 0 if 𝛾 ≠ ℓ
and 𝜄ℓ = 1, the Leibniz rule implies
𝑛 𝑝 𝑞
1 ∑︁ 𝛼−⟨ℓ ⟩ ′
∑︁ ∑︁
𝛼ℓ 𝜕𝑧¯ 𝑐 ℓ + 𝑖𝑐 ℓ = − 𝛼
𝑥 𝑖𝜆 𝜕𝑧¯ 𝑐 𝑖𝜆 − 𝑦 𝑗𝜇 𝜕𝑧¯𝛼 𝑐 ′𝑗𝜇 ,
2 ℓ=1 𝜆=1 𝜇=1

𝑛 𝑝 𝑞
1 ∑︁ ∑︁ ∑︁
𝛼ℓ 𝜕𝑧¯𝛼−⟨ℓ ⟩ 𝑐 ℓ + 𝑖𝑐 ℓ′ = − 𝑦 𝑗𝜇 𝜕𝑧¯𝛼 𝑐 ′𝑗𝜇 ,

𝑥 𝑖𝜆 𝜕𝑧¯𝛼 𝑐 𝑖𝜆 −
2 ℓ=1 𝜆=1 𝜇=1
466 14 Analytic Stratifications

where 𝑐 ℓ ≡ 0 (resp., 𝑐 ℓ′ ≡ 0) unless ℓ = 𝑖𝜆 for some 𝜆 (resp., ℓ = 𝑗 𝜇 for some 𝜇).

Induction on |𝛼| yields 𝜕𝑧¯𝛼 𝑐 ℓ + 𝑖𝑐 ℓ′ 𝑽 ∩U ≡ 0 for all 𝛼 ∈ Z+𝑛 . Thus 𝑐 ℓ + 𝑖𝑐 ℓ′ 𝑽 ∩U ≡

0 and, by analyticity with respect to 𝑧¯,

𝜕𝑧¯ 𝑐 ℓ + 𝑖𝑐 ℓ′ ≡ 0.

The first one of these conditions requires, for all 𝑧 ∈ U,

𝑝
∑︁ 𝑞
∑︁
𝑐 ℓ (𝑧, 𝑧¯) + 𝑖𝑐 ℓ′ (𝑧, 𝑧¯) = 𝐴𝑖𝜆 ,ℓ (𝑧, 𝑧¯) 𝑥𝑖𝜆 + 𝐵 𝑗𝜇 ,ℓ (𝑧, 𝑧¯) 𝑦 𝑗𝜇
𝜆=1 𝜇=1
𝑝 𝑞
1 ∑︁ 1 ∑︁
= 𝐴𝑖𝜆 ,ℓ (𝑧, 0) 𝑧𝑖𝜆 + 𝐵 𝑗 ,ℓ (𝑧, 0) 𝑧 𝑗𝜇
2 𝜆=1 2𝑖 𝜇=1 𝜇
𝑛
∑︁
+ 𝑅 𝑘,ℓ (𝑧, 𝑧¯) 𝑧¯ 𝑘 .
𝑘=1

Then the second one requires

𝑝 𝑞
1 ∑︁ 1 ∑︁
𝑐 ℓ (𝑧, 𝑧¯) + 𝑖𝑐 ℓ′ (𝑧, 𝑧¯) = 𝐴𝑖𝜆 ,ℓ (𝑧, 0) 𝑧 𝑖𝜆 + 𝐵 𝑗 ,ℓ (𝑧, 0) 𝑧 𝑗𝜇 .
2 𝜆=1 2𝑖 𝜇=1 𝜇

Thus 𝑐 𝑖′𝜆 ≡ 0 implies 𝑐 𝑖𝜆 (𝑧, 𝑧¯) = 𝑐 𝑖𝜆 (𝑧, 0) ∈ O (U); likewise 𝑐 𝑗𝜇 ≡ 0 implies

𝑐 ′𝑗𝜇 (𝑧, 𝑧¯) = 𝑐 ′𝑗𝜇 (𝑧, 0) ∈ O (U); otherwise, 𝑐 ℓ (𝑧, 𝑧¯) + 𝑖𝑐 ℓ′ (𝑧, 𝑧¯) = 𝐹ℓ (𝑧) ∈ O (U).
Going back to (14.2.3) we derive, after some relabeling,
𝑝 ′ 𝑞 ′ 𝑝+𝑞
∑︁ ∑︁ ∑︁
ℎ (𝑧) = 𝑐 𝑖𝜆 (𝑧, 0) 𝑥𝑖𝜆 + 𝑐 ′𝑗𝜇 (𝑧, 0) 𝑦 𝑗𝜇 + 𝑐 ℓ (𝑧, 𝑧¯) 𝑥ℓ + 𝑐 ℓ′ (𝑧, 𝑧¯) 𝑦 ℓ
𝜆=1 𝜇=1 ℓ= 𝑝′ +𝑞′ +1
𝑝′ 𝑞 ′
1 ∑︁ 1 ∑︁ ′
= 𝑐 𝑖𝜆 (𝑧, 0) 𝑧𝑖𝜆 + 𝑐 (𝑧, 0) 𝑧 𝑗𝜇
2 𝜆=1
2𝑖 𝜇=1 𝑗𝜇
𝑝+𝑞 𝑝+𝑞 𝑝+𝑞
∑︁ ∑︁ 1 ∑︁
+ 𝐹ℓ (𝑧) 𝑧ℓ − 𝑖 𝑐 ℓ′ (𝑧, 0) 𝑧ℓ + 𝑖 𝐺 ℓ (𝑧, 𝑧¯) 𝑧¯ℓ .
ℓ= 𝑝′ +𝑞′ +1 ℓ= 𝑝′ +𝑞′ +1
2 ℓ=1

Since ℎ ∈ O (U) we must have 𝐺 ℓ (𝑧, 𝑧¯) ≡ 0; we have reached the conclusion that
𝑝+𝑞
∑︁
ℎ (𝑧) = 𝐻ℓ (𝑧) 𝑧ℓ ,
ℓ=1

which proves that the ideal in O (U) generated by 𝑧1 , ..., 𝑧 𝑝+𝑞 contains all ℎ ∈ O (U)
that vanish identically on 𝑽 ∩ U. It follows that

𝑽 ∩ U = 𝑧 ∈ U; 𝑧 1 = · · · = 𝑧 𝑝+𝑞 = 0 .
14.2 Analytic Subvarieties 467

This proves that the C 𝜔 regular part of 𝑽 is a complex-analytic submanifold of M.□

Definition 14.2.10 Let 𝑽 be an analytic subvariety of M. By the dimension of 𝑽,
dimK 𝑽 (or simply dim 𝑽 when there is no risk of ambiguity), we shall mean the
dimension of the analytic submanifold ℜ (𝑽). By the dimension of 𝑽 at a point
℘ ∈ 𝑽 we shall mean the dimension of the germ of 𝑽 at ℘, i.e., inf dim (U ∩ 𝑽) as
U
U ranges over the family of all open sets containing ℘.
If 𝑽 and 𝑾 are two analytic subvarieties of M and if 𝑾 ⊂ 𝑽 then dim 𝑾 ≤ dim 𝑽.
When K = R, in general 𝔖 (𝑽) is not an analytic subvariety of M (contrary to
the complex-analytic case, see Proposition 14.3.24 below) but the following is true
(with K = R or C). We recall (Definition 14.1.1) that ℜ𝑘 (𝑽) stands for the union of
all the connected components of ℜ (𝑽) that have dimension 𝑘.
Proposition 14.2.11 If 𝑑 = dim 𝑽 and if 𝑽\ℜ𝑑 (𝑽) ≠ ∅ then 𝑽\ℜ𝑑 (𝑽) is an
analytic subvariety of M of dimension < dim 𝑽.
Proof In a suitably small neighborhood U of an arbitrary point of 𝑽\ℜ𝑑 (𝑽) we
can define 𝑽\ℜ𝑑 (𝑽) in 𝑽 ∩ U by the vanishing of d 𝑓1 ∧ · · · ∧ d 𝑓 𝑑 as the 𝑓 𝑗 range
over ℑ (𝑽, U). □
If 𝑽 1 = 𝑽\ℜ𝑑 (𝑽) and 𝑑1 = dim 𝑽 1 then 𝑽 2 = 𝑽 1 \ℜ𝑑1 (𝑽 1 ) is either empty or
an analytic subvariety of M of dimension
< 𝑑1 . By defining 𝑽 𝑗+1 = 𝑽 𝑗 \ℜ𝑑 𝑗 𝑽 𝑗
if 𝑑 𝑗 = dim 𝑽 𝑗 and 𝑽 𝑗 ≠ ℜ𝑑 𝑗 𝑽 𝑗 for 𝑗 = 1, 2, ..., until we reach an analytic
subvariety of M such that 𝑽 𝜈 = ℜ𝑑𝜈 (𝑽 𝜈 ), i.e., 𝑽 𝜈 is an analytic submanifold, we
get a sequence of nested analytic subvarieties of M,

𝑽 ⊃ 𝑽1 ⊃ · · · ⊃ 𝑽 𝜈, (14.2.4)

with 𝑽 𝑗 \𝑽 𝑗−1 = ℜ𝑑 𝑗 𝑽 𝑗 , dim 𝑽 𝑗 < dim 𝑽 𝑗−1 and 𝑽 𝜈 a closed analytic submanifold
of M (obviously 𝜈 ≤ 𝑑). Note that

ℜ 𝑽 𝑗 \ ℜ 𝑽 𝑗−1 \ℜ𝑑 𝑗−1 𝑽 𝑗−1 ⊂ 𝔖 𝑽 𝑗−1

for every 𝑗 = 1, ..., 𝜈, with the understanding that 𝑽 0 = 𝑽, 𝑑0 = 𝑑.

Proposition 14.2.12 We have the partition of 𝑽 into C 𝜔 submanifolds,
𝜈
Ø
𝑽= ℜ𝑑 𝑗 𝑽 𝑗 . (14.2.5)
𝑗=0

This follows from the facts that 𝑽 𝑗 = 𝑽 𝑗−1 ∪ ℜ𝑑 𝑗 𝑽 𝑗 and 𝑽 𝜈 = ℜ𝑑𝜈 (𝑽 𝜈 ).

In general we may have ℜ𝑑 𝑗 𝑽 𝑗 ∩ ℜ𝑑 𝑗−1 𝑽 𝑗−1 ≠ ∅ and ℜ𝑑 𝑗 𝑽 𝑗 ⊄

ℜ𝑑 𝑗−1 𝑽 𝑗−1 and thus (14.2.5) is not a stratification. Example: the union of two
vector subspaces 𝑬 1 , 𝑬 2 of K𝑛 such that 𝑬 1 ∩ 𝑬 2 = {0}, dim 𝑬 1 < dim 𝑬 2 . This
also shows that a connectedanalytic submanifold of M can be contained in 𝑽 and
not be contained in ℜ𝑑 𝑗 𝑽 𝑗 for any 𝑗.
468 14 Analytic Stratifications

14.2.3 Basic properties of unions of submanifolds and subvarieties

Proposition 14.2.13 Let N be the union of finitely many analytic submanifolds

L1 , ..., L 𝑠 of M satisfying the following conditions, for each 𝑖 = 1, ..., 𝑠:
Ø
(1) No open subset of L𝑖 is contained in L𝑗;
1≤ 𝑗 ≤𝑠
Ø 𝑗≠𝑖
(2) L𝑖 ∩ L 𝑗 is a closed subset of L𝑖 .
1≤ 𝑗 ≤𝑠
𝑗≠𝑖

Under these hypotheses,

Ø
𝔖 (N ) = L𝑖 ∩ L 𝑗 .
1≤𝑖< 𝑗 ≤𝑠

In this statement N need not be an analytic subvariety of M.

Ø
Proof It is clear that L𝑖 \ L 𝑗 ⊂ ℜ (N ) for every 𝑖 ∈ {1, ..., 𝑠}, which proves
1≤ 𝑗 ≤𝑠
Ø 𝑗≠𝑖
that 𝔖 (N ) ⊂ L𝑖 ∩ L 𝑗 .
1≤𝑖< 𝑗 ≤𝑠
We prove the converse inclusion. Every point of ℜ (N ) is contained in some open
subset U of M such that U ∩ ℜ (N ) = U ∩ L. If U ∩ ℜ (N ) ∩ L𝑖 ≠ ∅ for some
𝑖 then either U ∩ L𝑖 ⊂ U ∩ ℜ (N ) is open in U ∩ ℜ (N ) or else the codimension
of U ∩ L𝑖 with respect to ℜ (N ) is positive. In the former case U ∩ L𝑖 cannot
intersect any L 𝑗 otherwise an open subset of L 𝑗 would be contained in L𝑖 . In the
latter case, if L 𝑗 ( 𝑗 ≠ 𝑖) intersects U ∩ L𝑖 then U ∩ L 𝑗 ⊂ U ∩ ℜ (N ) must also
have positive codimension with respect to ℜ (N ) otherwise an open subset of L𝑖
would be contained in L 𝑗 . But U ∩ ℜ (N ) is the union of the finitely many sets
U ∩ L 𝑗 , which is absurd. We have thus proved that ℜ (N ) ∩ L𝑖 ∩ L 𝑗 = ∅ for every
pair of indices 𝑖, 𝑗. □

Proposition 14.2.14 Let 𝑽 be the union of finitely many subvarieties 𝑽 1 , ..., 𝑽 𝑠 of

the analytic manifold
Ø M such that, for every 𝑖 = 1, ..., 𝑠, no open subset of 𝑽 𝑖 is
contained in 𝑽 𝑗 . Under this hypothesis,
1≤ 𝑗 ≤𝑠
𝑗≠𝑖

© Ø ª © Ø
𝔖 (𝑽) = ­ 𝔖 𝑽𝑗 ®∪­ 𝑽 𝑗 ∩ 𝑽𝑘® . (14.2.6)
ª

« 𝑗=1,...,𝑠 ¬ «1≤ 𝑗<𝑘 ≤𝑠 ¬

14.2 Analytic Subvarieties 469

Proof Suppose that a point ℘ ∈ 𝑽 does not belong to the right-hand side of (14.2.6),
a set we provisionally denote by 𝑬. Clearly, if ℘ ∈ 𝑽\𝑬 then ℘ ∈ ℜ (𝑽 𝑖 ) for some 𝑖
and ℘ ∉ 𝑽 𝑗 for every 𝑗 ≠ 𝑖. Since the latter will be automatically true for all points
in a neighborhood U of ℘ we conclude that U ∩ ℜ (𝑽) = U ∩ ℜ (𝑽 𝑖 ), whence
℘ ∈ ℜ (𝑽). This proves that 𝔖 (𝑽) ⊂ 𝑬.
It remains to prove the converse inclusion. We have, for 𝑘 > 1,

𝔖 (𝑽 1 ) ∪ 𝔖 (𝑽 𝑘 ) ∪ (𝑽 1 ∩ 𝑽 𝑘 ) = 𝔖 (𝑽 1 ) ∪ 𝔖 (𝑽 𝑘 ) ∪ (ℜ (𝑽 1 ) ∩ ℜ (𝑽 𝑘 )) .

If ℜ (𝑽 1 ) ∩ ℜ (𝑽 𝑘 ) had a regular point then necessarily one of the two varieties 𝑽 1

or 𝑽 𝑘 would contain an open subset of the other, prohibited by our hypothesis. This
implies
𝔖 (𝑽 1 ) ∪ 𝔖 (𝑽 𝑘 ) ∪ (𝑽 1 ∩ 𝑽 𝑘 ) ⊂ 𝔖 (𝑽 1 ∪ 𝑽 𝑘 ) ,
proving the claim when 𝑠 = 2. Ø
Suppose 𝑠 > 2 and define 𝑽 ♭2 = 𝑽 𝑗 . Our hypothesis demands that no open
𝑗=2,...,𝑠
subset of 𝑽 1 be contained in 𝑽 ♭2 and no open subset of 𝑽 ♭2 be contained in 𝑽 1 . We
derive that

© Ø
𝔖­ 𝑽 𝑗 ® = 𝔖 (𝑽 1 ) ∪ 𝔖 𝑽 ♭2 ∪ 𝑽 1 ∩ 𝑽 ♭2 .
ª

« 𝑗=1,...,𝑠 ¬
Induction on 𝑠 implies

© Ø ª © Ø
𝔖 𝑽 ♭2 = ­ 𝔖 𝑽𝑗 ®∪­ 𝑽 𝑗 ∩ 𝑽𝑘®
ª

« 𝑗=2,...,𝑠 ¬ «2≤ 𝑗<𝑘 ≤𝑠 ¬

whence easily (14.2.6). □

Corollary 14.2.15 Let the subvariety 𝑽 be as in Proposition 14.2.14; then

Ø ©­ Ø ª
ℜ (𝑽) = ℜ 𝑽𝑗 \­ ℜ 𝑽 𝑗 ∩ 𝑽𝑘® . (14.2.7)
®
­ ®
𝑗=1,...,𝑠 1≤𝑘 ≤𝑠
« 𝑘≠ 𝑗 ¬
Proof The right-hand side of (14.2.7) is the complement in 𝑽 of the right-hand side
of (14.2.6). □

14.2.4 The Whitney umbrella

To close this section we describe briefly the natural stratification of a celebrated

example of an analytic subvariety. By the Whitney umbrella in K3 we mean the
variety
470 14 Analytic Stratifications

𝑾 K = 𝑥 ∈ K3 ; 𝑥32 = 𝑥1 𝑥22 .

The variety 𝑾 K is irreducible (Definition 14.2.6) whether K = R or K = C. But

there are important differences between the complex and the real Whitney umbrella.
On the one hand we have

ℜ 𝑾 C = 𝑧 ∈ C3 ; 𝑧23 = 𝑧 1 𝑧 22 , |𝑧 2 | + |𝑧3 | ≠ 0 ,

𝔖 𝑾 C = 𝑧 ∈ C3 ; 𝑧 2 = 𝑧 3 = 0 ,

and ℜ 𝑾 C is connected since 𝑾 C is irreducible (cf. Theorem 14.2.7); 𝔖 𝑾 C is
a complex-analytic subvariety whereas

𝔖 𝑾 R = 𝑥 ∈ R3 ; 𝑥1 ≥ 0, 𝑥2 = 𝑥3 = 0

is a closed half-line; it is not an analytic subvariety (it is a semi-analytic

set, see
last section of this chapter). The analytic submanifold ℜ 𝑾 R has three connected
components:

ℜ+ 𝑾 R = 𝑥 ∈ R3 ; 𝑥22 𝑥1 = 𝑥 32 , 𝑥2 > 0 ,

ℜ− 𝑾 R = 𝑥 ∈ R3 ; 𝑥22 𝑥1 = 𝑥 32 , 𝑥2 < 0 ,

ℜ◦ 𝑾 R = 𝑥 ∈ R3 ; 𝑥1 < 0, 𝑥2 = 𝑥3 = 0 .

Here are the standard stratifications

of the Whitney umbrellas:
In the complex case ℜ 𝑾 is a two-dimensional stratum; in the real case, the
C

two connected components ℜ± 𝑾 R of ℜ 𝑾 R are two-dimensional strata. There
is a single one-dimensional stratum of 𝑾 C : the deleted plane

𝑧 ∈ C3 ; 𝑧1 ≠ 0, 𝑧2 = 𝑧3 = 0 .

There are two one-dimensional strata of 𝑾 R : the open half-axes ℜ◦ 𝑾 R and

𝔖 𝑾 R \ {0} = 𝑥 ∈ R3 ; 𝑥1 > 0, 𝑥2 = 𝑥3 = 0 .

Note that 𝑾 R\ℜ2 𝑾 R is the entire 𝑥 1 -axis; it is the sole stratum of the subvariety
𝑾 R \ℜ2 𝑾 R (cf. Proposition 14.2.11). The origin is the single zero-dimensional
stratum of both 𝑾 R and 𝑾 C .
14.3 The Weierstrass Theorems 471

14.3 The Weierstrass Theorems

Many results in this chapter, and especially in this section, will rely on basic concepts
and results about commutative rings and their ideals (see the Appendix to this
section).

14.3.1 The Weierstrass theorems for functions

Here the focus will be on the case K = C. As usual 𝑧1 , ..., 𝑧 𝑛 denote the coordinates
in C𝑛 ; by a polydisk we mean a subset of C𝑛 ,
n o
Δ𝑟𝑛 (𝑧◦ ) = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 − 𝑧 ◦𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 ; (14.3.1)

we refer to 𝑧 ◦ as the center and to 𝑟 = (𝑟 1 , ..., 𝑟 𝑛 ) as the polyradius; the radii 𝑟 𝑗 are
positive, always finite unless specified otherwise. When 𝑧 ◦ = 0 we write Δ𝑟𝑛 rather
than Δ𝑟𝑛 (0). We shall
often make use of the notation 𝑧 ′ = (𝑧1 , ..., 𝑧 𝑛−1 ) and of the
polydisk Δ𝑟 ′ = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 − 1 .
𝑛−1

Recall (Subsection 1.2.1) that if Ω is an open subset of C𝑛 the ring of holomorphic

functions in Ω is denoted by O (Ω); the linear space O (Ω) is equipped with the
topology of uniform convergence on compact subsets of Ω; O (Ω) is a Fréchet–
Montel space; it is a closed subspace of C ∞ (Ω) and all the reasonable topologies
on O (Ω), including that induced by D ′ (Ω), coincide.

Proposition 14.3.1 Let 𝑓 ∈ O (Ω) and 𝑧◦ ∈ Ω be arbitrary. Unless 𝑓 is constant

the set of vectors u ∈ R𝑛 such that the function 𝜁 ↦→ 𝑓 (𝑧◦ + 𝜁u) − 𝑓 (𝑧◦ ) vanishes
of finite order at 𝜁 = 0 is open and dense in R𝑛 .

Proof Near 𝑧◦ , 𝑓 (𝑧) − 𝑓 (𝑧◦ ) = 𝑃𝑚 (𝑧) + 𝑂 |𝑧| 𝑚+1 for some homogeneous poly-
nomial 𝑃𝑚 of degree 𝑚 ≥ 1. The result ensues then from its obvious validity when
𝑓 = 𝑃𝑚 . □

Corollary 14.3.2 If Ω is connected the ring O (Ω) is an integral domain.

Proof Suppose none of two functions 𝑓 , 𝑔 ∈ O (Ω) is identically equal to zero.

Let 𝑧◦ ∈ Ω be arbitrary. Proposition 14.3.1 has the following consequence: after
a linear change of coordinates, neither 𝑓 (𝑧◦′, 𝑧 𝑛 ) nor 𝑔 (𝑧◦′, 𝑧 𝑛 ) vanish identically.
There are functions 𝑢 (𝑧 𝑛 ) and 𝑣 (𝑧 𝑛 ) that do not vanish at 𝑧 ◦𝑛 and 𝑝, 𝑞 ∈ Z+ such
that 𝑓 (𝑧◦′, 𝑧 𝑛 ) = 𝑢 (𝑧 𝑛 ) 𝑧 𝑛 − 𝑧 ◦𝑛 and 𝑔 (𝑧◦′, 𝑧 𝑛 ) = 𝑣 (𝑧 𝑛 ) 𝑧 𝑛 − 𝑧◦𝑛 , implying that
𝑝 𝑞
◦′ ◦′
𝑓 (𝑧 , 𝑧 𝑛 ) 𝑔 (𝑧 , 𝑧 𝑛 ) cannot vanish identically. □
To state the classical Weierstrass Preparation Theorem it is convenient to intro-
duce the following
472 14 Analytic Stratifications

Definition 14.3.3 By a Weierstrass polynomial of degree 𝑚 (∈ Z+ ) we shall mean

a polynomial in the variable 𝑧 𝑛 of the form
𝑚
∑︁
𝑃 (𝑧 ′; 𝑧 𝑛 ) = 𝑧 𝑚
𝑛 + 𝑎 𝑘 (𝑧 ′) 𝑧 𝑚−𝑘
𝑛 (14.3.2)
𝑘=1

whose coefficients 𝑎 𝑘 (𝑧 ′) are holomorphic functions in a neighborhood of the origin

in C𝑛−1 and vanish at 𝑧 ′ = 0.

Theorem 14.3.4 Let Ω be a neighborhood of the origin in C𝑛 and 𝑓 ∈ O (Ω).

Assume that zero is an isolated root of 𝑓 (0, 𝑧 𝑛 ) and let 𝑚 ∈ Z+ be its multiplicity
( 𝑓 (0) ≠ 0 if 𝑚 = 0). There is a polydisk Δ𝑟𝑛 ⊂⊂ Ω in which 𝑓 = 𝐸 𝑃 with
𝑃 (𝑧 ′; 𝑧 𝑛 ) ∈ O Δ𝑟𝑛−1′ [𝑧 𝑛 ] a Weierstrass polynomial of degree 𝑚 and 𝐸 ∈ O Δ𝑟𝑛
nowhere zero in Δ𝑟 . The pair (𝐸, 𝑃) is unique.
𝑛

By O Δ𝑟𝑛−1 ′ [𝑧 𝑛 ] we mean the ring of polynomials in the variable 𝑧 𝑛 with coef-
ficients in O Δ𝑟𝑛−1 ′ .
Proof The result is trivial if 𝑛 = 1; suppose 𝑛 ≥ 2.
Existence of 𝐸 and 𝑃. Let the circle 𝔠 = {𝑧 𝑛 ∈ C; |𝑧 𝑛 | = 𝑟 𝑛 } be contained in Ω.
By our hypotheses we can select 𝑟 𝑛 > 0 so that 𝑓 (0, 𝑧 𝑛 ) ≠ 0 whatever 𝑧 𝑛 ∈ 𝔠. We
can then select the radii 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 − 1, sufficiently small that | 𝑓 (𝑧)| ≥ 𝑐 ◦ > 0
for all 𝑧 ′ ∈ Δ𝑟𝑛−1
′ and all 𝑧 𝑛 ∈ 𝔠. As a consequence, the integrals
∮
′1 𝜁ℓ 𝜕𝑓
𝜎ℓ (𝑧 ) = (𝑧 ′, 𝜁) d𝜁
2𝜋𝑖 𝑓 (𝑧 ′, 𝜁) 𝜕𝑧 𝑛
𝔠

(ℓ ∈ Z+ ) are holomorphic functions of 𝑧 ′ in Δ𝑟𝑛−1 ′ . By the Rouché Theorem we

know that 𝜎0 is simply the number of zeros 𝜁 𝑘 (𝑧 ′) of the function
Í 𝑧 𝑛ℓ ↦→′ 𝑓
(𝑧 ′, 𝑧 𝑛 )
in the disk |𝑧 𝑛 | < 𝑟 𝑛 , necessarily equal to 𝑚; and 𝜎ℓ (𝑧 ′) = 𝑚 𝜁
𝑘=1 𝑘 (𝑧 ) for each
ℓ = 1, 2, .... The standard symmetric functions of the roots being polynomials with
respect to the 𝜎ℓ we see that the coefficients of the polynomial in the variable 𝑧 𝑛 ,
𝑚
Ö 𝑚
∑︁
𝑃 (𝑧 ′; 𝑧 𝑛 ) = (𝑧 𝑛 − 𝜁 𝑘 (𝑧 ′)) = 𝑧 𝑚 𝑎 𝑗 (𝑧 ′) 𝑧 𝑛
𝑚− 𝑗
𝑛 + , (14.3.3)
𝑘=1 𝑗=1

are holomorphic functions of 𝑧 ′ in Δ𝑟𝑛−1 ′ . Moreover, 𝑎 𝑗 (0) = 0 for every 𝑗 = 1, ..., 𝑚

and |𝑃 (𝑧 ′; 𝑧 𝑛 )| ≥ 𝑐 1 > 0 for |𝑧 𝑛 | = 𝑟 𝑛 . It follows easily [cf. (14.3.5) below] that the
function 𝑓 (𝑧) /𝑃 (𝑧 ′; 𝑧 𝑛 ) is holomorphic and does not vanish anywhere in Δ𝑟𝑛−1 ′ .
Uniqueness of (𝐸, 𝑃). If (𝐸 1 , 𝑃1 ) is another pair like (𝐸, 𝑃) such that 𝐸 𝑃 = 𝐸 1 𝑃1
the set of roots of the two polynomials 𝑧 𝑛 ↦→ 𝑃 (𝑧 ′; 𝑧 𝑛 ) and 𝑧 𝑛 ↦→ 𝑃1 (𝑧 ′, 𝑧 𝑛 ) must
be the same for each 𝑧 ′ in a neighborhood of 0, hence 𝑃1 = 𝑃. By Corollary 14.3.2
(𝐸 − 𝐸 1 ) 𝑃 ≡ 0 entails 𝐸 = 𝐸 1 . □
Next we state and prove the Weierstrass Division Theorem.
14.3 The Weierstrass Theorems 473

Theorem 14.3.5 Let 𝑃 (𝑧 ′; 𝑧 𝑛 ) ∈ O Δ𝑟𝑛−1

′ [𝑧 𝑛 ] be a Weierstrass polynomial of de-
gree 𝑚 and let Δ𝑟𝑛 ⊂⊂ Ω be a polydisk such that 𝑧 ′ ∈ Δ𝑟𝑛−1 ′ , |𝑧 𝑛 | = 𝑟 𝑛 , imply
𝑃 (𝑧 ′; 𝑧 𝑛 ) ≠ 0. Then to each function 𝑔 ∈ O (Ω) there are a function ℎ ∈ O Δ𝑟𝑛 and

a polynomial with respect to 𝑧 𝑛 ,
𝑚
∑︁
𝑅 (𝑧 ′, 𝑧 𝑛 ) = 𝑏 𝑗 (𝑧 ′) 𝑧 𝑛
𝑚− 𝑗
, (14.3.4)
𝑗=1

with 𝑏 𝑗 ∈ O Δ𝑟𝑛−1
′ ( 𝑗 = 1, ..., 𝑚), such that 𝑔 = ℎ𝑃 + 𝑅. The pair (ℎ, 𝑅) is unique.
Proof Existence of ℎ, 𝑅. The integral

𝑔 (𝑧 ′, 𝜁) d𝜁
∮
1
(14.3.5)
2𝜋𝑖 𝑃 (𝑧 ′, 𝜁) 𝜁 − 𝑧 𝑛
|𝜁 |=𝑟𝑛

defines a function ℎ (𝑧) ∈ O Δ𝑟𝑛 . We can write

𝑔 (𝑧 ′, 𝜁) 𝑃 (𝑧 ′; 𝑧 𝑛 ) − 𝑃 (𝑧 ′, 𝜁)
∮
′ 1
ℎ (𝑧) 𝑃 (𝑧 ; 𝑧 𝑛 ) = 𝑔 (𝑧) + d𝜁.
2𝜋𝑖 𝑃 (𝑧 ′, 𝜁) 𝜁 − 𝑧𝑛
|𝜁 |=𝑟𝑛

The integral on the right is a polynomial

of degree < 𝑚 with respect to 𝑧 𝑛 whose
coefficients belong to O Δ𝑟𝑛−1 ′ .
Uniqueness of ℎ, 𝑅. Suppose ℎ1 𝑃 + 𝑅1 = ℎ𝑃 + 𝑅, i.e. (ℎ − ℎ1 ) 𝑃 = 𝑅 − 𝑅1
with ℎ1 (resp., 𝑅1 ) similar to ℎ (resp., 𝑅). The number of roots of the polynomial
𝑧 𝑛 ↦→ (𝑅 − 𝑅1 ) (𝑧 ′, 𝑧 𝑛 ) must be 𝑚, which demands 𝑅 = 𝑅1 . By Corollary 14.3.2
(ℎ − ℎ1 ) 𝑃 ≡ 0 entails ℎ = ℎ1 . □
Corollary 14.3.6 Let 𝑓 , 𝑔 ∈ O (Ω). Assume that 0 ∈ Ω and that 𝑓 (0, 𝑧 𝑛 ) vanishes
to order 𝑚 < +∞ at zero. Then, in a neighborhood of 0 we have 𝑔 = ℎ 𝑓 + 𝑅 with ℎ
and 𝑅 as in Theorem 14.3.5.
Proof Indeed, by Theorem 14.3.4 we know that, in a neighborhood of the origin,
𝑓 = 𝐸 𝑃 with 𝐸 invertible and 𝑃 a Weierstrass polynomial. By Theorem 14.3.5
𝑔 = ℎ◦ 𝑃 + 𝑅 = ℎ◦ 𝐸 −1 𝑓 + 𝑅. □
We are also going to need the Riemann Extension Theorem:
Theorem 14.3.7 Let 𝑓 ∈ O (Ω) and suppose that 𝑽 ( 𝑓 ) = {𝑧 ∈ Ω; 𝑓 (𝑧) = 0} does
not contain any connected component of Ω. Suppose that 𝑔 ∈ O (Ω\𝑽 ( 𝑓 )) has the
following property: every point 𝑧◦ ∈ Ω is the center of an open polydisk Δ𝑟𝑛 (𝑧◦ ) ⊂⊂ Ω
such that |𝑔 (𝑧)| is bounded in (Ω\𝑽 ( 𝑓 )) ∩ Δ𝑟𝑛 (𝑧◦ ). Then 𝑔 can be extended as a
holomorphic function 𝐺 in Ω.
The extension 𝐺 is evidently unique.
Proof Obviously it suffices to prove the claim in the neighborhood of an arbitrary
zero 𝑧◦ of 𝑓 . Since 𝑓 . 0 in every connected component of Ω, after an affine
change of the coordinates we may assume that 𝑧 ◦ = 0 and 𝑓 (0, 𝑧 𝑛 ) . 0 in a
474 14 Analytic Stratifications

neighborhood of 𝑧 𝑛 = 0. We can therefore select the radii 𝑟 𝑗 ( 𝑗 = 1, ..., 𝑛) so that

| 𝑓 (𝑧 ′, 𝑧 𝑛+1 )| ≥ 𝑐 ◦ > 0 for all 𝑧 ∈ C𝑛 such that 𝑧 𝑗 < 𝑟 𝑗 ( 𝑗 = 1, ..., 𝑛 − 1) and
|𝑧 𝑛 | = 𝑟 𝑛 . It follows that the function
∮
1 d𝜁
𝐺 (𝑧) = 𝑔 (𝑧 ′, 𝜁)
2𝜋𝑖 𝜁 − 𝑧𝑛
|𝜁 |=𝑟𝑛

is well defined, and holomorphic, in the polydisk Δ𝑟𝑛 . In order to see that 𝐺 = 𝑔 in
Ω ∩ Δ𝑟𝑛 it suffices to fix 𝑧 ′ ∈ Δ𝑟𝑛−1
′ arbitrarily and apply the Cauchy formula in the
𝑧 𝑛 -plane. □

Corollary
14.3.8 Let 𝑓1 , ..., 𝑓 𝑁 ∈ O (Ω) (1 ≤ 𝑁 < +∞) and let 𝑽 ( 𝑓1 , ..., 𝑓 𝑁 ) =
𝑧 ∈ Ω; 𝑓 𝑗 (𝑧) = 0, 𝑗 = 1.., 𝑁 . Suppose that 𝑔 ∈ O (Ω\𝑽 ( 𝑓1 , ..., 𝑓 𝑁 )) has the fol-
lowing property: every point 𝑧 ◦ ∈ Ω is the center of an open polydisk Δ𝑟𝑛 (𝑧◦ ) ⊂⊂ Ω
such that |𝑔 (𝑧)| is bounded in (Ω\𝑽 ( 𝑓1 , ..., 𝑓 𝑁 )) ∩ Δ𝑟𝑛 (𝑧◦ ). Then 𝑔 can be extended
as a holomorphic function 𝐺 in Ω.
◦ is the center of an open polydisk Δ𝑟𝑛 (𝑧◦ ) ⊂⊂ Ω such that
𝑧 𝑛∈ Ω
Proof Each
◦ 𝑛 ◦
Ω\𝑽 𝑓 𝑗 ∩ Δ𝑟 (𝑧 ) is open and dense in Δ𝑟 (𝑧 )for some 𝑗, 1 ≤ 𝑗 ≤ 𝑁. Since 𝑔 ∈
O Ω\𝑽 𝑓 𝑗 and |𝑔 (𝑧)| is bounded in Ω\𝑽 𝑓 𝑗 ∩ Δ𝑟𝑛 (𝑧◦ ) the desired conclusion
follows directly from Theorem 14.3.7. □
Lastly we wish to discuss briefly the real versions of the Weierstrass theorems:
the functions are restricted to real-space and are real-valued. Thus Ω stands here for
a connected open subset of R𝑛 containing the origin and we consider a C 𝜔 function
𝑓 : Ω −→ R which we assume not to vanish identically. After a linear change of
variables in R𝑛 we can assume that 𝑓 (0, 𝑥 𝑛 ) does not vanish to infinite order at
𝑥 𝑛 = 0. There is an open subset ΩC of C𝑛 containing Ω to which 𝑓 can be extended
as a holomorphic function, which we denote by 𝑓 (𝑧). We can apply Theorem 14.3.4
to 𝑓 (𝑧): in a polydisk Δ𝑟𝑛 we can write 𝑓 = 𝐸 𝑃, where 𝐸 ∈ O Δ𝑟𝑛 does not
vanish anywhere in Δ𝑟𝑛 and 𝑃 (𝑧 ′; 𝑧 𝑛 ) ∈ O Δ𝑟𝑛−1 ′ [𝑧 𝑛 ] is a Weierstrass polynomial
of degree 𝑚. Restriction to Δ𝑟𝑛 ∩ R𝑛 gives us that 𝑓 (𝑥) = 𝐸 (𝑥) 𝑃 (𝑥 ′; 𝑥 𝑛 ). We also
have 𝑓 (𝑥) = 𝐸 (𝑥) 𝑃 (𝑥 ′; 𝑥 𝑛 ). By the uniqueness part in Theorem 14.3.4 we derive
that the holomorphic extensions of 𝐸 (𝑥) and 𝑃 (𝑥 ′; 𝑥 𝑛 ) must be equal to 𝐸 (𝑧) and
𝑃 (𝑧 ′; 𝑧 𝑛 ) respectively, implying that both 𝐸 (𝑥) and 𝑃 (𝑥 ′; 𝑥 𝑛 ) are real. This proves
the real version of Theorem 14.3.4, referred to as the real Weierstrass Preparation
Theorem. A similar argument applies to Theorem 14.3.5 and Corollary 14.3.6: if
𝑓 , 𝑔 ∈ C 𝜔 (Ω) are real-valued then the functions ℎ and 𝑅 such that 𝑔 = ℎ 𝑓 + 𝑅 in
Δ𝑟𝑛 must be real when restricted to Δ𝑟𝑛 ∩ R𝑛 . The latter result is referred to as the real
Weierstrass Division Theorem.
14.3 The Weierstrass Theorems 475

14.3.2 The ring C {𝒛1 , ..., 𝒛 𝒏 }. The Weierstrass theorems for germs

We shall denote by C {𝑧 1 , ..., 𝑧 𝑛 } the set of convergent power series in 𝑛 variables

𝑧1 , ..., 𝑧 𝑛 with complex coefficients,
∑︁
𝑐 𝛼 𝑧 𝛼, (14.3.6)
𝛼=( 𝛼1 ,..., 𝛼𝑛 ) ∈Z+𝑛

where 𝑐 𝛼 ∈ C, 𝑧 𝛼 = 𝑧1𝛼1 · · · 𝑧 𝑛𝛼𝑛 ; C {𝑧1 , ..., 𝑧 𝑛 } is a ring with respect to ordinary

addition and multiplication.
We identify C {𝑧1 , ..., 𝑧 𝑛 } with the ring O0(𝑛) of germs of holomorphic functions
at the origin in C𝑛 . In accordance with the general Definition 1.2.7 such a germ is an
equivalence class of pairs ( 𝑓 , 𝑈) consisting of a function 𝑓 and of a neighborhood
𝑈 ⊂ C𝑛 of 0 in which 𝑓 is defined and holomorphic, i.e., 𝑓 ∈ O (𝑈). Another pair
(𝑔, 𝑉) is equivalent to ( 𝑓 , 𝑈) if there is a neighborhood 𝑊 ⊂ 𝑈 ∩ 𝑉 of the origin
such that 𝑓 = 𝑔 in 𝑊. The correspondence C {𝑧1 , ..., 𝑧 𝑛 } ←→ O0(𝑛) is evident: to the
series (14.3.6), assumed to converge normally (i.e., uniformly on compact subsets)
in a polydisk Δ𝑟𝑛 , we assign its sum 𝑓 . The other way, to the germ at the origin of a
holomorphic function 𝑓 we assign its Taylor expansion about 0. Most often in the
sequel, we do not distinguish between these two concepts, or between C {𝑧1 , ..., 𝑧 𝑛 }
and O0(𝑛) .
It follows directly from Corollary 14.3.2 that O0(𝑛) and C {𝑧1 , ..., 𝑧 𝑛 } are integral
domains.
The ring C {𝑧1 , ..., 𝑧 𝑛 } has a maximal ideal 𝔐𝑛 : the set of series whose zero-order
term is equal to zero: 𝑐 0 = 0 in (14.3.6); we shall express the latter property by saying
that the series (14.3.6) vanishes at the origin. To say that (14.3.6) does not belong to
𝔐𝑛 means that it has an inverse in C {𝑧1 , ..., 𝑧 𝑛 } [the Taylor expansion at 0 of 1/ 𝑓
assuming that (14.3.6) is the Taylor expansion of 𝑓 and that 𝑓 (0) ≠ 0]; then (14.3.6)
is often, in algebra, referred to as a unit. For 𝑘 = 2, 3..., the set 𝔐𝑛𝑘 consisting of the
series (14.3.6) such that 𝑐 𝛼 = 0 if |𝛼| < 𝑘 is also an ideal in C {𝑧1 , ..., 𝑧 𝑛 }; if a series
𝑓 belongs to 𝔐𝑛𝑘 we shall say that 𝑓 vanishes to order 𝑘 at the origin. The decreasing
sequence of ideals 𝔐𝑛 ⊃ 𝔐𝑛2 ⊃ · · · ⊃ 𝔐𝑛𝑘 ⊃ · · · forms a basis of neighborhoods
of 0 in the natural ring topology of C {𝑧1 , ..., 𝑧 𝑛 }. There is also a natural locally
convex topology on C {𝑧1 , ..., 𝑧 𝑛 } distinct from the ring topology, the topology of
convergence of the coefficients.
The Weierstrass theorems 14.3.4 and 14.3.5 imply directly analogous statements
about germs.

Theorem 14.3.9 Let f ∈ C {𝑧 1 , ..., 𝑧 𝑛 } be such that f (0, 𝑧 𝑛 ) = 𝑐 ◦ 𝑧 𝑚 𝑛 + 𝑂 𝑧𝑛
𝑚+1

with 𝑚 ∈ Z+ and 𝑐 ◦ ≠ 0. Then f = EP with E ∈ C {𝑧 1 , ..., 𝑧 𝑛 } invertible and P ∈

C {𝑧 1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ] the germ of a Weierstrass polynomial (14.3.2) with 𝑎 𝑗 ∈ 𝔐𝑛−1 .
The pair (E, P) is unique.
Theorem 14.3.10 Let f ∈ C {𝑧1 , ..., 𝑧 𝑛 } be as in Theorem 14.3.9. Then to each g ∈
C {𝑧 1 , ..., 𝑧 𝑛 } there are h ∈ C {𝑧1 , ..., 𝑧 𝑛 } and a polynomial R ∈ C {𝑧1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ],
deg R < 𝑚, such that g = hf + R. The pair (h, R) is unique.
476 14 Analytic Stratifications

Proposition 14.3.11 The rings C {𝑧 1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ] and C {𝑧1 , ..., 𝑧 𝑛 } are Noethe-
rian unique factorization domains.
Proof We use induction on 𝑛 ≥ 1. Assuming that C {𝑧 1 , ..., 𝑧 𝑛−1 } is a Noetherian
unique factorization domain we derive from Lemmas 14.3.26, 14.3.28 (Appendix)
that C {𝑧 1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ] is a Noetherian unique factorization domain. That the same
is true of C {𝑧1 , ..., 𝑧 𝑛 } follows then directly from Theorem 14.3.9. □
In view of the standard terminology used below it is worthwhile to make more
explicit the unique factorization property: every f ∈ C {𝑧1 , ..., 𝑧 𝑛 } which is not a unit
(i.e., f ∈ 𝔐𝑛 ) admits a factorization

f = f1𝑘1 · · · f𝑟𝑘𝑟 , (14.3.7)

where the 𝑘 𝛼 are positive integers and the f 𝛼 ∈ C {𝑧1 , ..., 𝑧 𝑛 } are prime or irreducible,
in the following sense: no f 𝛼 is a unit and if f 𝛼 = uv, u, v ∈C {𝑧1 , ..., 𝑧 𝑛 }, then
either u or v must be a unit. Moreover, the number 𝑟 of irreducible factors is
minimum: if 𝛼 ≠ 𝛽 then f 𝛼 is not divisible by f𝛽 ; the factorization (14.3.7) is unique
up to permutation and multiplication by units of the factors f 𝛼𝑘 𝛼 . Naturally, to the
factorization (14.3.7) there corresponds the local factorization of representatives of
the germ f.
In connection with Proposition 14.3.11 the unique factorization of Weierstrass
polynomials and of the germs they define will be needed in the sequel. We shall use
the term “irreducible polynomial” in full generality: if 𝑝 (𝑥) is an arbitrary polyno-
mial in the indeterminate 𝑥 with coefficients in a commutative ring A, i.e., 𝑝 ∈ A [𝑥],
we say that 𝑝 is irreducible or prime if the equation 𝑝 = 𝑝 1 𝑝 2 with 𝑝 𝛼 ∈ A [𝑥],
𝛼 = 1, 2, demands (deg 𝑝 1 ) (deg 𝑝 2 ) = 0. Thus the germ of a Weierstrass polyno-
mial P ∈C {𝑧 1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ] is irreducible if P = P1 P2 with P 𝛼 ∈C {𝑧 1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ]
(𝛼 = 1, 2) also germs of Weierstrass polynomials, then either P1 = 1 or P2 = 1. To
say that two irreducible monic (i.e., with leading coefficients equal to 1) polynomials
P1 , P2 are coprime (i.e., do not have a common factor of positive degree) is the same
as saying that they are distinct, i.e., P1 ≠ P2 .
Proposition 14.3.12 Let P ∈C {𝑧 1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ] be the germ of a Weierstrass poly-
nomial of degree ≥ 1. There are distinct germs of irreducible Weierstrass polynomials
P 𝛼 ∈C {𝑧1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ] and positive integers 𝑘 𝛼 (𝛼 = 1, ..., 𝑟) such that

P = P1𝑘1 · · · P𝑟𝑘𝑟 . (14.3.8)

The set of pairs (P 𝛼 , 𝑘 𝛼 ), 𝛼 = 1, ..., 𝑟, is unique up to permutations.

Proof By Proposition 14.3.11 there are germs of irreducible polynomials

P 𝛼 ∈C {𝑧1 , ..., 𝑧 𝑛−1 } [𝑧 𝑛 ]

such that (14.3.8) holds. We can apply Theorem 14.3.9 to each one of them and write
P 𝛼 = E 𝛼b P 𝛼 with E 𝛼 ∈C {𝑧 1 , ..., 𝑧 𝑛 } invertible and b
P 𝛼 the germ of a Weierstrass
polynomial. Since P is monic, (14.3.8) implies E1𝑘1 · · · E𝑟𝑘𝑟 = 1. □
14.3 The Weierstrass Theorems 477

Corollary 14.3.13 Let Δ𝑟𝑛−1 be ′

′
the polydisk of polyradius 𝑟 and center the origin
in C𝑛−1 and let 𝑃 ∈ O Δ𝑟𝑛−1 ′ [𝑧 𝑛 ] be a Weierstrass polynomial. There are distinct
irreducible Weierstrass polynomials 𝑃 𝛼 ∈ O Δ𝑟𝑛−1 ′ [𝑧 𝑛 ] and positive integers 𝑘 𝛼
𝑘1 𝑘𝑟
(𝛼 = 1, ..., 𝑟) such that 𝑃 = 𝑃1 · · · 𝑃𝑟 . The set of pairs (𝑃 𝛼 , 𝑘 𝛼 ), 𝛼 = 1, ..., 𝑟, is
unique up to permutations.

14.3.3 Complex-analytic varieties

We begin by looking at a complex-analytic (or in short, complex) manifold M;

dimC M = 𝑛. Until specified otherwise 𝑽 will be a complex-analytic subvariety of
M (Definition 14.2.1); we shall always assume ∅ ≠ 𝑽 ≠ M.
As announced in Subsection 14.1.1 we are going to make use of the concept of
germ of variety at a point ℘◦ ∈ 𝑽, (𝑽, ℘◦ ) or 𝑽 ℘◦ . (We shall use bold face for germs
of subvarieties and functions and slanted bold face for “true” subvarieties.) To the
germ (𝑽, ℘◦ ) we can associate the set ℑ (𝑽, ℘◦ ) of germs of holomorphic functions at
℘◦ which vanish on (𝑽, ℘◦ ): a representative of a germ f ∈ ℑ (𝑽, ℘◦ ) can be thought
of as a pair ( 𝑓 , U) consisting of an open set U ∋℘◦ and a function 𝑓 ∈ O (U) such
that ℘ ∈ 𝑽 ∩ U =⇒ 𝑓 (℘) = 0. It is clear that ℑ (𝑽, ℘◦ ) is an ideal in the ring O℘(𝑛)◦
of germs at ℘◦ of holomorphic functions of 𝑛 complex variables. This notation is
commonly extended to the case where ℘◦ ∉ 𝑽 and ℑ (𝑽, ℘◦ ) = O℘(𝑛)◦ .
Assuming that U is the domain of complex-analytic coordinates 𝑧1 , ..., 𝑧 𝑛 van-
ishing at ℘◦ we associate to the germ of variety (𝑽, ℘◦ ) the ideal 𝔦𝔡𝑽 of the ring
C {𝑧1 , ..., 𝑧 𝑛 } consisting of the convergent power series which vanish on (𝑽, ℘◦ ): it
is the image of ℑ (𝑽, ℘◦ ) under Taylor expansion about 𝑧 = 0. In what follows we
use the notation f to mean either a germ belonging to O℘(𝑛)◦ or its Taylor expansion
about the origin, an element of C {𝑧 1 , ..., 𝑧 𝑛 }.
By Proposition 14.3.11 the ideal ℑ (𝑽, ℘◦ ) is generated by finitely many germs
f1 , ..., f𝜈 . We can find a neighborhood U of ℘◦ in M in which each germ f 𝑗 has a
representative 𝑓 𝑗 . In practice we think of the germ (𝑽, ℘◦ ) as the subvariety

𝑽 ∩ U = ℘ ∈ U; 𝑓 𝑗 (℘) = 0, 𝑗 = 1, ..., 𝜈 . (14.3.9)

This vindicates in the complex case the claim made at the beginning of this chapter
that, provided the open set U is sufficiently small, the sets 𝚽 (𝑽, U) in Definition
14.2.1 can be taken to be finite. The neighborhood U may be contracted finitely many
times as needed. From this perspective it is sometimes convenient to use the notation
𝑽 ℘◦ (f1 , ..., f𝜈 ) for the germ of variety (𝑽, ℘◦ ). Actually it is immediately checked
that 𝑽 ℘◦ (f1 , ..., f𝜈 ) does not depend on the choice of the generators f1 , ..., f𝜈 . The
Taylor expansions of the functions 𝑓1 , ..., 𝑓 𝜈 in (14.3.9) generate 𝔦𝔡𝑽: every series
belonging to 𝔦𝔡𝑽 is a linear combination with coefficients in C {𝑧1 , ..., 𝑧 𝑛 } of those
Taylor series. Keep in mind, however, that the radii of convergence of the coefficients
might be smaller than those of the Taylor series of 𝑓1 , ..., 𝑓 𝜈 .
478 14 Analytic Stratifications

An arbitrary ideal 𝔍 in O℘(𝑛)◦ defines a germ of subvariety at ℘◦ , commonly called

the locus of ℑ and here denoted by loc ℑ:
Ù
loc ℑ = 𝑽 (f) . (14.3.10)
f ∈𝔍

We now list a few basic facts about the germs of varieties, ideals in C {𝑧1 , ..., 𝑧 𝑛 }
and their loci.

Proposition 14.3.14 If 𝑽 𝑗 ( 𝑗 = 1, 2) are two germs of varieties at ℘◦ then

𝑽 1 ⊂𝑽 2 =⇒ 𝔦𝔡𝑽 2 ⊂ 𝔦𝔡𝑽 1 . If ℑ 𝑗 ( 𝑗 = 1, 2) are two ideals in C {𝑧1 , ..., 𝑧 𝑛 } then
ℑ2 ⊂ ℑ1 =⇒ loc ℑ1 ⊂ loc ℑ2 .

Self-evident.

Proposition 14.3.15 If 𝑽 is a germ of variety at ℘◦ then 𝑽= loc (𝔦𝔡𝑽).

Proof That 𝑽⊂ loc (𝔦𝔡𝑽) is evident. Let the (finitely many) germs f1 , ..., f𝜈 generate
𝔦𝔡𝑽; we have loc (𝔦𝔡𝑽) ⊂𝑽 (f1 ) ∩ · · · ∩𝑽 (f𝜈 ) =𝑽 (f1 , ..., f𝜈 ) =𝑽. □

Corollary 14.3.16 If 𝑽 𝑗 ( 𝑗 = 1, 2) are two germs of varieties at ℘◦ then 𝔦𝔡𝑽 2 ⊂

𝔦𝔡V1 =⇒𝑽 1 ⊂𝑽 2 and 𝔦𝔡𝑽= 𝔦𝔡𝑽 2 ⇐⇒𝑽 1 =𝑽 2 .

√ an ideal of C {𝑧1 , ..., 𝑧 𝑛 } then ℑ ⊂ 𝔦𝔡 (loc ℑ). This

It is also evident that if ℑ is
can be improved. The radical ℑ of the ideal 𝔍 is the set of germs f ∈ C {𝑧1 , ..., 𝑧 𝑛 }
such that f 𝑘 ∈ ℑ for some positive integer 𝑘.
√ √
Proposition 14.3.17 If ℑ is an ideal of C {𝑧1 , ..., 𝑧 𝑛 } then loc ℑ = loc ℑ and ℑ ⊂
𝔦𝔡 (loc ℑ).

The proof is left as an exercise.

Germs of varieties at a point ℘◦ satisfy a descending chain condition:

Proposition 14.3.18 If 𝑽 𝑗 ( 𝑗 = 1, 2, ...) is a monotone decreasing sequence of germs

of varieties at ℘◦ , i.e., 𝑽 𝑗+1 ⊂𝑽 𝑗 , then there is a positive integer 𝑗 ◦ such that 𝑽 𝑗 =𝑽 𝑗◦
for all 𝑗 > 𝑗◦ .

Proof Indeed, the 𝔦𝔡𝑽 𝑗 form a monotone increasing sequence of ideals of C {𝑧 1 , ..., 𝑧 𝑛 }
which must be eventually stationary. It suffices then to apply Propositions 14.3.11,
14.3.14 and Corollary 14.3.16. □
We take a look, now, at irreducible germs of complex-analytic varieties (Definition
14.2.6).

Proposition 14.3.19 For the germ of variety 𝑽 at ℘◦ to be irreducible it is necessary

and sufficient that 𝔦𝔡𝑽 be a prime ideal of C {𝑧 1 , ..., 𝑧 𝑛 }.
14.3 The Weierstrass Theorems 479

Proof Let 𝑽, 𝑽 1 ≠ 𝑽, 𝑽 2 ≠ 𝑽, be germs of varieties such that 𝑽 = 𝑽 1 ∪ 𝑽 2 .

By Corollary 14.3.16 we know that there is f 𝑗 ∈ 𝔦𝔡𝑽 𝑗 , f 𝑗 ∉ 𝔦𝔡𝑽 ( 𝑗 = 1, 2). But
f1 f2 ∈ 𝔦𝔡 (𝑽 1 ∪ 𝑽 2 ) = 𝔦𝔡𝑽 which is therefore not prime. Conversely, suppose
there
are germs f 𝑗 ∉ 𝔦𝔡𝑽 ( 𝑗 = 1, 2) such that f1 f2 ∈ 𝔦𝔡𝑽. Set 𝑽 𝑗 =𝑽∩𝑽 f 𝑗 ; we have
𝑽 𝑗 ≠𝑽 and

(𝑽 ∩ 𝑽 (f1 )) ∪ (𝑽 ∩ 𝑽 (f2 )) = 𝑽 ∩ (𝑽 (f1 ) ∪ 𝑽 (f2 ))

= 𝑽 ∩ 𝑽 (f1 f2 ) = 𝑽. □

Proposition 14.3.20 Let 𝑽 be a germ of variety at ℘◦ . There is a unique set of germs

of irreducible varieties at ℘◦ , 𝑽 𝑗 ( 𝑗 = 1, 2, ..., 𝑟 < +∞), such that
Ø
(1) ∀ 𝑗 = 1, ..., 𝑟, 𝑽 𝑗 ⊄ 𝑽𝑘;
𝑘≠ 𝑗
(2) 𝑽=𝑽 1 ∪ · · · ∪𝑽 𝑟 .
The germs of varieties 𝑽 𝑗 ( 𝑗 = 1, 2, ..., 𝑟) are called the irreducible components
of the germ of variety 𝑽.
Proof Let 𝔉 denote the family of germs of varieties (at ℘◦ ) which are not equal to
a finite union of germs of irreducible varieties. Every totally ordered (for inclusion)
subfamily of 𝔉 is perforce finite, by Proposition 14.3.18, and it has therefore a
minimum element. It ensues that either 𝔉 = ∅ or else 𝔉 contains at least one minimal
member, 𝑾, necessarily reducible. We can write 𝑾=𝑾 1 ∪𝑾 2 with 𝑾 1 ≠𝑾, 𝑾 2 ≠𝑾;
𝑾∈ 𝔉 entails either 𝑾 1 ∈ 𝔉 or 𝑾 2 ∈ 𝔉 contradicting the minimality of 𝑾. We
conclude that 𝔉 = ∅ and that every germ of variety admits a decomposition 𝑽=𝑽 1 ∪
· · · ∪𝑽 𝑟 into irreducible components whose number 𝑟 can be taken to be minimum,
ensuring the validity of Property (1). Now suppose there is another decomposition
𝑽=𝑽 1′ ∪ · · · ∪𝑽 𝑟′ ′ into irreducible components 𝑽 ′𝑗 , also having Property (1). Then
𝑽 𝑖 = 𝑽 𝑖 ∩ 𝑽 1′ ∪ · · · ∪ 𝑽 𝑖 ∩ 𝑽 𝑟′ ′ for each 𝑖 ∈ {1, ..., 𝑟 }. Since 𝑽 𝑖 is irreducible

we must have 𝑽 𝑖 =𝑽 𝑖 ∩𝑽 ′𝑗 (𝑖) for some 𝑗 (𝑖) ∈ {1, ..., 𝑟 ′ }. Conversely, to each 𝑗 ∈
{1, ..., 𝑟 ′ } there is a 𝑘 ( 𝑗) ∈ {1, ..., 𝑟 } such that 𝑽 ′𝑗 =𝑽 ′𝑗 ∩𝑽 𝑘 ( 𝑗) . In particular we
have 𝑽 ′𝑗 (𝑖) =𝑽 ′𝑗 (𝑖) ∩𝑽 𝑘 ( 𝑗 (𝑖)) and thus 𝑽 𝑖 =𝑽 𝑖 ∩𝑽 ′𝑗 (𝑖) ∩𝑽 𝑘 ( 𝑗 (𝑖)) ⊂𝑽 𝑘 ( 𝑗 (𝑖)) , implying
𝑘 ( 𝑗 (𝑖)) = 𝑖 by Property (1), whence 𝑟 ′ = 𝑟 and 𝑽 ′𝑗 (𝑖) =𝑽 ′𝑗 (𝑖) ∩𝑽 𝑖 =𝑽 𝑖 . Thus the set
of irreducible components of 𝑽 is unique. □
The decomposition of varieties in Proposition 14.3.20 is related to that of ideals
in Lemma 14.3.29. Indeed, if 𝑽=𝑽 1 ∪ · · · ∪𝑽 𝑟 then

𝔦𝔡𝑽 =𝔦𝔡𝑽 1 ∩ · · · ∩ 𝔦𝔡𝑽 𝑟 . (14.3.11)

When the varieties 𝑽 𝑖 are irreducible the ideals 𝔓𝑖 = 𝔦𝔡𝑽 𝑖 are prime and (14.3.11) is
the decomposition of 𝔦𝔡𝑽 into its prime components. Conversely let ℑ be an ideal of
C {𝑧 1 , ..., 𝑧 𝑛 } and ℑ = 𝔔1 ∩√· · · ∩ 𝔔𝑟 be a decomposition into √
primary components
(Lemma 14.3.29). We get ℑ=𝔓1 ∩ · · · ∩ 𝔓𝑟 , where 𝔓𝑖 = 𝔔𝑖 is prime. From
Proposition 14.3.14 we derive
480 14 Analytic Stratifications

loc 𝔓1 ∪ · · · ∪ loc 𝔓𝑟 ⊂ loc ℑ,

and therefore, by (14.3.17),
√
ℑ⊂𝔦𝔡 loc ℑ⊂𝔦𝔡 loc 𝔓1 ∩ · · · ∩ 𝔦𝔡 loc 𝔓𝑟 . (14.3.12)

A nontrivial improvement of Proposition 14.3.17 is the classical Nullstellensatz,

which we limit ourselves to stating:
√
Theorem 14.3.21 Let ℑ be an ideal in C {𝑧 1 , ..., 𝑧 𝑛 }; then 𝔦𝔡 (loc ℑ) = ℑ.
We refer the reader to [Gunning and Rossi, 1965], Ch. III for a proof of Theorem
14.3.21; it relies on a thorough study of the local structure of a complex-analytic
variety. Such a study reveals aspects of complex-analytic varieties that have no
analogue for real-analytic ones. One of them has already been stated: Theorem
14.1.11 [𝑽 irreducible⇐⇒ ℜ (𝑽) connected], besides which the most noteworthy
are the following.
Proposition 14.3.22 Let 𝑽 be a complex-analytic subvariety of the complex manifold
M. If the open subset of M, U, is connected then the complement U\ (U∩𝑽) is
connected.
It suffices to prove the claim locally, in which case it is a direct consequence of
the Riemann Extension Theorem 14.3.7.
Proposition 14.3.23 Let 𝑽 be a complex-analytic subvariety of the complex manifold
M. If the germ of 𝑽 at a point ℘◦ ∈ 𝑽 is irreducible then there are finitely many
holomorphic functions 𝑓1 , ..., 𝑓𝑠 in a neighborhood U of ℘◦ whose germs at each
point ℘ ∈ 𝑽 ∩ U generate the ideal ℑ (V, ℘).
Proposition 14.3.24 The singular part 𝔖 (𝑽) of a complex-analytic variety 𝑽 is a
complex-analytic variety of dimension < dim 𝑽.
In the complex-analytic category, given any subvariety 𝑽 of M we can define a
sequence of nested varieties 𝑽 𝑘 (𝑘 = 0, 1...) different from (14.2.4): as in (14.2.4)
we start with 𝑽 0 = 𝑽 but now we take 𝑽 𝑘+1 = 𝔖 (𝑽 𝑘 ) for each 𝑘. It follows
from Proposition 14.3.24 that dim 𝑽 𝑘+1 < dim 𝑽 𝑘 (thus the sequence is finite) and
𝑽 𝑘 \𝑽 𝑘+1 = ℜ (𝑽 𝑘 ) is a submanifold of M which is an open and dense subset of 𝑽 𝑘 .
For each 𝑘 let ℭ 𝑘, 𝛼 (𝛼 = 1, ..., 𝑁 𝑘 with possibly 𝑁 𝑘 = +∞) denote the connected
components of ℜ (𝑽 𝑘 ). We refer to the decomposition
𝑁𝑘
𝑛 Ø
Ø
𝑽= ℭ 𝑘, 𝛼 (14.3.13)
𝑘=0 𝛼=1

as the natural partition of 𝑽. We will prove later (by applying Corollary 14.5.13
below) that (14.3.13) is locally finite.
Remark 14.3.25 The natural partition of the complex Whitney umbrella 𝑾 C is a
stratification coarser than the stratification
described
in Subsection 14.2.3: there are
only two natural strata ℜ 𝑾 C and 𝔖 𝑾 C , the latter being the plane 𝑧 2 = 𝑧3 = 0.
14.3 The Weierstrass Theorems 481

14.3.A Appendix: integral domains, unique factorization domains,

Noetherian rings

All the rings under consideration will be commutative and have a unit element (to
distinguish from units, the invertible elements of 𝑨).
The ring 𝑨 is called an integral domain if for all 𝑎, 𝑏 ∈ 𝑨, 𝑎𝑏 = 0 =⇒ 𝑎 = 0 or
𝑏 = 0.
The ring 𝑨 is said to be a unique factorization domain if every noninvertible
element 𝑎 ∈ 𝑨 can be written as a product 𝑎 1 · · · 𝑎 𝜈 for a finite subset {𝑎 1 , ..., 𝑎 𝜈 }
of 𝑨 unique up to permutation and multiplication by units. (The uniqueness of the
factorization demands that each factor 𝑎 𝑗 be noninvertible and prime or irreducible:
if 𝑎 𝑗 = 𝑏𝑐, 𝑏, 𝑐 ∈ 𝑨, then either 𝑏 or 𝑐 is a unit.)
It is seen immediately that if 𝑨 is an integral domain the same is true of 𝑨 [𝜁],
the ring of polynomials in the indeterminate 𝜁 with coefficients in 𝑨. The analogous
result about unique factorization domains is known as the Gauss lemma (for a proof
we refer the reader to textbooks on ring theory, e.g. [Van der Waerden, 1953], Vol.
I, pp. 70–72):

Lemma 14.3.26 If 𝑨 is a unique factorization integral domain then the same is true
of 𝑨 [𝜁].

The ring 𝑨 is said to be Noetherian if every ideal of 𝑨 is finitely generated. This is

equivalent to saying that any ascending sequence of ideals ℑ1 ⊂ · · · ⊂ ℑ𝑘 ⊂ ℑ𝑘+1 ⊂
· · · is stationary: ℑ𝑘 = ℑ𝑘+1 if 𝑘 is sufficiently large (this is sometimes referred to as
the chain condition). If 𝑨 is Noetherian and if ℑ is an ideal of 𝑨 the quotient ring
𝑨/𝔍 is also Noetherian. Note however that a subring of a Noetherian ring might not
be Noetherian.
We recall that an 𝑨-module 𝑴 is said to be finitely generated if there are finitely
many elements 𝑥1 , ..., 𝑥𝑟 ∈ 𝑴 such that every element of 𝑴 is equal to a linear
combination 𝑎 1 𝑥1 + · · · + 𝑎𝑟 𝑥𝑟 with coefficients 𝑎 𝑗 ∈ 𝑨; the minimum integer 𝑟 such
that this property holds is referred to as the rank of 𝑴 (the representation might not
be unique even when 𝑟 = rank 𝑴 contrary to what happens when 𝑨 is a field, in
which case 𝑴 is a vector space and 𝑟 = dim 𝑴). If 𝑨 is Noetherian then every ideal
of 𝑨 can be regarded as a finitely generated 𝑨-module.

Lemma 14.3.27 Suppose 𝑨 Noetherian and let 𝑴 be a finitely generated 𝑨-module.

Then
(1) every strictly increasing sequence of submodules of 𝑴 is finite;
(2) every submodule of 𝑴 is finitely generated.

Proof First we show that (1) and (2) are equivalent. Let Φ (𝑵) denote the family
(ordered by inclusion) of all finitely generated submodules of a submodule 𝑵 ⊂ 𝑴;
it follows immediately from (1) that there is a maximal element 𝑵 0 in Φ. If there
were an element 𝑥 ∈ 𝑵\𝑵 0 then 𝑵 0 + 𝑨𝑥 would be a finitely generated submodule
of 𝑵 strictly larger than 𝑵 0 . We must therefore have 𝑵 = 𝑵 0 . Conversely suppose
482 14 Analytic Stratifications

(2) is valid and let 𝑵 𝑗 , 𝑗 = 1, 2, ..., be a strictly increasing sequence of submodules

of 𝑴. The union 𝑵 of the 𝑵 𝑗 is a submodule of 𝑴 with finitely many generators
𝑥1 , ..., 𝑥𝑟 . These must perforce belong to 𝑵 𝑗 for some 𝑗 < +∞.
Now let 𝑥 1 , ..., 𝑥𝑟 generate 𝑴 and let 𝑵 be an arbitrary submodule of 𝑴. For each
ℓ, 1 ≤ ℓ ≤ 𝑟, we consider the submodule 𝑵 ℓ consisting of the linear combinations
𝑥 = 𝑎 1 𝑥 1 + · · · + 𝑎 ℓ 𝑥ℓ that belong to 𝑵. We denote by ℑℓ ⊂ 𝑨 the set of all the
coefficients 𝑎 ℓ of such linear combinations; it is obvious that ℑℓ is an ideal; it
follows that ℑℓ has a finite basis 𝛼 (ℓ)𝑗 ( 𝑗 = 1, ..., 𝑁ℓ ). For each 𝑗 there is an element
𝑦 (ℓ)
𝑗 = 𝑏 1 𝑥 1 + · · · + 𝑏 ℓ−1 𝑥
ℓ−1 + 𝛼 (ℓ) 𝑥 ∈ 𝑵 . If 𝑥 = 𝑎 𝑥 + · · · + 𝑎 𝑥 ∈ 𝑵 and if
𝑗 ℓ ℓ 1 1 ℓ ℓ ℓ
𝑎 ℓ = 𝑁𝑗=1 𝑐 𝑗,ℓ 𝛼 (ℓ)
Í ℓ
𝑗 then

𝑁ℓ
∑︁
𝑥− 𝑐 𝑗,ℓ 𝑦 (ℓ) ′ ′
𝑗 = 𝑎 1 𝑥 1 + · · · + 𝑎 ℓ−1 𝑥
ℓ−1
∈ 𝑵 ℓ−1 .
𝑗=1

Descending induction on ℓ = 𝑟, ..., 1, implies directly that the elements 𝑦 (ℓ)

𝑗 (𝑗 =
1, ..., 𝑁ℓ , ℓ = 1, ..., 𝑟) generate 𝑵. □
The next classical result relates to the chain condition.
Lemma 14.3.28 If 𝑨 is a Noetherian ring the same is true of 𝑨 [𝜁].
Proof The leading coefficients (i.e., the coefficients of the highest power of 𝜁) in the
polynomials belonging to an ideal ℑ ⊂ 𝑨 [𝜁] form an ideal in 𝑨, generated by finitely
many elements 𝑎 1 , ..., 𝑎 𝜈 . Among all polynomials belonging to ℑ that have leading
Í
coefficient 𝑎 𝑗 let 𝑓 𝑗 have minimum degree, 𝑑 𝑗 . Let 𝜈𝑗=1 𝑐 𝑗 𝑎 𝑗 , 𝑐 𝑗 ∈ 𝑨, be the leading
coefficient of 𝑓 ∈ ℑ and Í let 𝐽 be the set of indices 𝑗 such that 𝑚 = deg 𝑓 ≥ 𝑑 𝑗 .
The polynomial 𝑓1 = 𝑓 − 𝑗 ∈𝐽 𝑐 𝑗 𝜁 𝑚−𝑑 𝑗 𝑓 𝑗 ∈ ℑ has degree 𝑚 1 < 𝑚. We can repeat
the same argument with 𝑓1 in place of 𝑓 . In finitely many steps we conclude that
ℑ is the direct sum of the ideal 𝔍 in 𝑨 [𝜁] generated by 𝑓1 , ..., 𝑓 𝜈 and of the 𝑨-
module 𝔑 = 𝑔 ∈ ℑ; deg 𝑔 < min 𝑑 𝑗 . The 𝑨-submodule of 𝑨 [𝜁] consisting of
1≤ 𝑗 ≤𝜈
the polynomials of degree < min 𝑑 𝑗 is obviously finitely generated; so is 𝔑 by
1≤ 𝑗 ≤𝜈
Lemma 14.3.27; therefore the same is true of 𝔍 + 𝔑 = ℑ. □
The radical of an ideal ℑ in a commutative ring 𝑨 is the ideal consisting of the
elements
√ 𝑎 ∈ 𝑨 such√that 𝑎 𝑚 ∈ ℑ for some positive integer 𝑚; it will be denoted by
ℑ; obviously ℑ ⊂ ℑ. If ℑ1 , ..., ℑ𝑟 are finitely many ideals in 𝑨 then
√︁ √︁ √︁
ℑ1 ∩ · · · ∩ ℑ𝑟 = ℑ1 ∩ · · · ∩ ℑ𝑟 .

Indeed, if 𝑏 𝑚 𝛼 ∈ ℑ 𝛼 for every 𝛼 = 1, ..., 𝑟 then 𝑏 𝑚 ∈ ℑ1 ∩ · · · ∩ ℑ𝑟 provided

𝑚 ≥ max 𝑚 𝛼 .
1≤ 𝛼 ≤𝑟
Recall that an ideal 𝔓 in a commutative ring 𝑨 is said to be prime if 𝑎𝑏 ∈ 𝔓
(𝑎, 𝑏 ∈ 𝑨) and 𝑎 ∉ 𝔓 entails 𝑏 ∈ 𝔓, which is the same as saying that√the quotient
ring 𝑨/𝔓 is an integral domain. An ideal 𝔔 is said to be primary if 𝔔 is prime:
14.4 Local Partitions of a Complex Hypersurface 483

if 𝑎𝑏 ∈ 𝔔 and 𝑎 ∉ 𝔔 entails 𝑏 𝑚 ∈ 𝔔. An ideal ℑ in 𝑨 is said to be reducible

if 𝔍 = 𝔄 ∩ 𝔔 with ideals 𝔄 ≠ 𝔍 and 𝔔 ≠ 𝔍, irreducible if there is no such
decomposition.

Lemma 14.3.29 Suppose the ring 𝑨 is Noetherian. Then every irreducible ideal
ℑ in 𝑨 is primary. An arbitrary ideal 𝔍 can be decomposed as an intersection
𝔍 = 𝔔1 ∩ · · · ∩ 𝔔𝑟 of primary ideals such that 𝔔 𝑗 ⊄ 𝔔 𝑘 if 𝑗 ≠ 𝑘.

Proof Suppose the ideal ℑ is not primary: there exist 𝑎, 𝑏 ∈ 𝑨 such that 𝑎𝑏 ∈ ℑ,
𝑎 ∉ ℑ, 𝑏 𝑘 ∉ ℑ for every 𝑘 ∈ Z+ . The ideals 𝔄 = ℑ+ 𝑎 𝑨 and ℑ + 𝑏 𝑘 𝑨 are both
distinct from ℑ. Consider the ideals 𝔔 𝑘 = 𝑥 ∈ 𝑨; 𝑥𝑏 ∈ ℑ ; we have 𝔔 𝑘 ⊂ 𝔔 𝑘+1
𝑘

and therefore 𝔔 𝑘+1 = 𝔔 𝑘 for sufficiently large

𝑘, which now we assume to be the
case. Obviously ℑ ⊂ (ℑ + 𝑎 𝑨) ∩ ℑ + 𝑏 𝑘 𝑨 ; let 𝑥 ∈ 𝔄 ∩ ℑ + 𝑏 𝑘 𝑨 . We have 𝑏𝑥 ∈ ℑ
since 𝑎𝑏 ∈ ℑ, as well as 𝑥 −𝑐𝑏 𝑘 ∈ ℑ for some 𝑐 ∈ 𝑨; these two facts imply 𝑐𝑏 𝑘+1 ∈ ℑ
hence 𝑐 ∈ 𝔔 𝑘 , i.e., 𝑐𝑏 𝑘 ∈ ℑ and therefore 𝑥 ∈ ℑ. Thus ℑ is reducible.
Now suppose ℑ = 𝔄1 ∩ 𝔔1 with ideals 𝔄1 ≠ ℑ and 𝔔1 ≠ ℑ. Suppose that we had
a similar a sequence of similar decompositions 𝔔 𝑗 = 𝔄 𝑗+1 ∩ 𝔔 𝑗+1 , 𝑗 = 1, 2, .... For
sufficiently large 𝑗 the ascending sequence of ideals 𝔔 𝑗 must become stationary, in
other words 𝔔 𝑗 must be irreducible. Repeating the same argument with 𝔄1 , ..., 𝔄 𝑗
shows that we can write ℑ = ℑ1 ∩ · · · ∩ ℑ𝑠 with irreducible ideals ℑ𝑘 . With this done
we eliminate the ideals ℑ𝑘 contained in any ideal ℑ 𝑗 , 𝑗 ≠ 𝑘. □
√ √ √
If ℑ = 𝔔1 ∩ · · · ∩ 𝔔𝑟 with primary ideals 𝔔 𝑗 then ℑ = 𝔔1 ∩ · · · ∩ 𝔔𝑟
√︁ √︁
with prime ideals 𝔔 𝑗 ; assuming that redundant ideals, i.e., ideals 𝔔 𝑗 contained
√ √ √ √
in some 𝔔 𝑘 , 𝑘 ≠ 𝑗 , have been deleted the decomposition ℑ = 𝔔1 ∩ · · · ∩ 𝔔𝑟
is unique up to permutations.

14.4 Local Partitions of a Complex Hypersurface

In this subsection we take M = Ω, e a domain in C𝑛 ; 𝑽 e will be a special analytic

(meaning complex-analytic) subvariety of Ω: we shall assume that 𝑽 is defined in Ω
e e e
by a single equation 𝑓 (𝑧) = 0, where 𝑓 ∈ O Ω e . We shall refer to this kind of variety
as a complex hypersurface (with singularities). Dealing with complex hypersurfaces
is considerably simpler than dealing with subvarieties of higher codimension; it will
provide all we need for dealing with a real C 𝜔 subvariety of arbitrary dimension, as
suggested in Remark 14.2.2.
We assume that the origin of C𝑛 lies in 𝑽, e i.e., 𝑓 (0) = 0. We select the coordinates
in C such that 𝑧 𝑛 ↦→ 𝑓 (0, ..., 0, 𝑧 𝑛 ) does not vanish identically (cf. Proposition
𝑛

14.3.1). We apply the Weierstrass Preparation Theorem 14.3.4: in an open polydisk

Δ𝑟𝑛 = Δ𝑟𝑛 (0) ⊂ Ωe [cf. (14.3.1)], 𝑓 (𝑧) = 𝐸 (𝑧) 𝑃 (𝑧 ′; 𝑧 𝑛 ) with 𝐸 ∈ O Δ𝑟𝑛 nowhere
vanishing and 𝑃 (𝑧 ′; 𝑧 𝑛 ) a Weierstrass polynomial (14.3.2) (Definition 14.3.3); here
𝑃 ∈ O Δ𝑟𝑛−1′ [𝑧 𝑛 ] [𝑟 ′ = (𝑟 1 , ..., 𝑟 𝑛−1 )].
484 14 Analytic Stratifications

If 𝑃 ∈ O Δ𝑟𝑛−1
′ [𝑧 𝑛 ] is a Weierstrass polynomial the unique factorization theorem
𝛼𝑁
for Weierstrass polynomials (Corollary 14.3.13) implies 𝑃 = 𝑃1𝛼1 · · · 𝑃 𝑁 , where the

Weierstrass polynomials 𝑃ℓ ∈ O Δ𝑟 ′ [𝑧 𝑛 ] are irreducible and distinct (therefore
𝑛−1

coprime) and 𝛼1 , ..., 𝛼 𝑁 are integers ≥ 1 (see Corollary 14.3.13). To avoid tedious
repetitions we introduce new terminology (also used when the base field is R).
𝛼𝑁
Definition 14.4.1 We shall say that the Weierstrass polynomial 𝑃 = 𝑃1𝛼1 · · · 𝑃 𝑁 is
trim if 𝛼1 = · · · = 𝛼 𝑁 = 1. When 𝑃 is not trim we refer to 𝑃1 · · · 𝑃 𝑁 as its trimming.
We shall say that a system of Weierstrass polynomials 𝑃 𝑘 (𝑘 = 1, ..., 𝜈) is trim if
every polynomial 𝑃 𝑘 is trim.

An equivalent definition is that 𝑃 ∈ O Δ𝑟𝑛−1 ′ [𝑧 𝑛 ] is trim if its discriminant does
not vanish identically in 𝑧 ′-space Δ𝑟𝑛−1
′ . We remind the reader that the discriminant
of 𝑃 is the polynomial 𝐷 (𝑧 ′) in the coefficients of 𝑃 resulting from the elimination
of 𝑧 𝑛 between 𝑃 and 𝜕𝑃/𝜕𝑧 𝑛 ; it vanishes when 𝑃 and 𝜕𝑃/𝜕𝑧 𝑛 have a common
root. For more details we refer to [Van der Waerden, 1953], p. 82. Every irreducible
Weierstrass polynomial is trim. The nullsets of an arbitrary Weierstrass polynomial
𝑃 and of its trimming are identical.

Proposition 14.4.2 If a Weierstrass polynomial 𝑃 is real (meaning real on real

space) so is its trimming.

Proof If an irreducible factor 𝑃ℓ of 𝑃 in O Δ𝑟𝑛−1′ [𝑧 𝑛 ] is not real then 𝑃ℓ and 𝑃ℓ
[𝑃ℓ (𝑧) = 𝑃ℓ ( 𝑧¯)] are coprime and 𝑃ℓ = 𝑃ℓ for some ℓ ′ ≠ ℓ; |𝑃ℓ | 2 is prime as a real
′

polynomial. □
In view of this we may as well posit that 𝑃 (𝑧 ′; 𝑧 𝑛 ) be trim and substitute 𝑃 for 𝑓
as the defining function of 𝑽
e in Δ𝑟𝑛 .

Proposition
14.4.3 The following properties of the Weierstrass polynomial 𝑃 ∈
O Δ𝑟𝑛−1
′ [𝑧 𝑛 ] are equivalent:
(a) deg 𝑃 = 1, i.e., 𝑃 (𝑧 ′; 𝑧 𝑛 ) = 𝑧 𝑛 − 𝑎 (𝑧 ′), 𝑎 ∈ O Δ𝑟𝑛−1

′ , 𝑎 (0) = 0;
(b) the discriminant 𝐷 (𝑧 ′) of 𝑃 does not vanish at any point of Δ𝑟𝑛−1 ′ ;
(c) 𝐷 (𝑧 ′) is identically equal to a nonzero constant (specifically, 1).

Proof If deg 𝑃 > 1 then the origin in C𝑛 is a common root of 𝑃 and 𝜕𝑃/𝜕𝑧 𝑛 , i.e.,
𝐷 (0) = 0. We have
𝜕𝑃
deg 𝑃 = 1 ⇐⇒ ≡ 1 ⇐⇒ 𝐷 (𝑧 ′) ≡ 1. □
𝜕𝑧 𝑛

If 𝑚 = deg 𝑃 = 1 the subvariety 𝑽 e ∩ Δ𝑟𝑛 is an analytic submanifold of Δ𝑟𝑛 defined

′
by the equation 𝑃 (𝑧 ; 𝑧 𝑛 ) = 0, and no further reduction is needed. In what follows
we reason under the assumption that 𝑚 ≥ 2.
Let 𝜌 𝑘 (𝑧 ′) (𝑘 = 1, ..., 𝑚) denote the roots of 𝑃 (𝑧 ′; 𝑧 𝑛 ) for each fixed 𝑧 ′ ∈ Δ𝑟𝑛−1 ′
where 𝑟 ′ = (𝑟 1 , ..., 𝑟 𝑛−1 ); since 𝑃 (0; 𝑧 𝑛 ) = 𝑧 𝑚
𝑛 we can select 𝑟 1 > 0, ..., 𝑟 𝑛−1 > 0, so
small that
14.4 Local Partitions of a Complex Hypersurface 485

sup |𝜌 𝑘 (𝑧 ′)| < 𝑟 𝑛 , 𝑘 = 1, ..., 𝑚. (14.4.1)

𝑧 ′ ∈Δ𝑟𝑛−1
′

By hypothesis the discriminant

Ö
𝐷 (𝑧 ′) = (𝜌 𝑘 (𝑧 ′) − 𝜌ℓ (𝑧 ′)) 2
1≤𝑘<ℓ ≤𝑚

of 𝑃 (𝑧 ′; 𝑧 𝑛 ) does not vanish identically; the origin of C𝑛−1 belongs to the proper
(because 𝑚 ≥ 2) subvariety of Δ𝑟𝑛−1 ′ ,

[1]
𝒁 = 𝑧 ′ ∈ Δ𝑟𝑛−1 ′

′ ; 𝐷 (𝑧 ) = 0 . (14.4.2)
e

[1]
′ \𝒁
In Δ𝑟𝑛−1 e the roots 𝜌 𝑘 (𝑧 ′) (𝑘 = 1, ..., 𝑚) are distinct.

Remark 14.4.4 In general the 𝜌 𝑘 (𝑧 ′) cannot be regarded as functions in Δ𝑟𝑛−1′ \𝒁

e [1]
[1]
as there could be branch points. Example: 𝑛 = 2, 𝑃 (𝑧 1 ; 𝑧2 ) = 𝑧22 − 𝑧 1 , where e
𝒁
reduces to the origin in the 𝑧 1 -plane.

Let 𝜋˜ ′ : Δ𝑟𝑛 −→ Δ𝑟𝑛−1′ be the coordinate projection (𝑧 ′, 𝑧 𝑛 ) ↦→ 𝑧 ′. We shall denote

−1
e [1] and, for

Λ [1]
by e 𝜄 , 𝜄 = 1, ..., 𝜈
[1] , the connected components of 𝑽 e ∩ 𝜋˜ ′ Δ𝑛−1
𝑟′ \𝒁
the time being, refer to them as first-tier leaves of the sought partition of 𝑽 e in Δ𝑟𝑛 .

Λ [1]
Proposition 14.4.5 Each point 𝑧◦ of a first-tier leaf e 𝜄 of 𝑽
e in Δ𝑟𝑛 has an open
neighborhood 𝑈 ⊂ Δ𝑟𝑛 such that

Λ [1] ′ ′ ′
𝑛
𝜄 ∩ 𝑈 = 𝑧 ∈ Δ𝑟 ; 𝑧 ∈ 𝜋˜ (𝑈) , 𝑧 𝑛 = 𝜌 𝑘 (𝑧 ) (14.4.3)
e

e [1] .

for a unique 𝑘, 1 ≤ 𝑘 ≤ 𝑚. Moreover, we have 𝜋˜ ′ eΛ [1]
𝜄 ′ \𝒁
= Δ𝑟𝑛−1

e [1]
Proof The first part of the claim follows immediately from the definition of 𝒁 . The
projection 𝜋˜ ′ is an open mapping and thus 𝜋˜ ′ e Λ [1]
𝜄 is open in Δ𝑟𝑛−1 ′ . If 𝑧
′◦ = 𝜋˜ ′ (𝑧 ◦ )

e [1] there is a sequence of points

belongs to the closure of 𝜋˜ ′ e Λ [1]
𝜄 ′ \𝒁
in Δ𝑟𝑛−1

Λ [1]
𝑧 (𝜈) ∈ e 𝜄 (𝜈 = 1, 2, ...) converging to 𝑧 ◦ such that 𝑧 (𝜈) = 𝜌
𝑛 𝑘 𝑧 ′(𝜈) for some 𝑘 and

[1]
◦
all 𝜈. But then (14.4.3) implies 𝑧 ∈ Λ 𝜄 , proving that 𝜋˜ e
e ′ Λ [1]
𝜄 is closed in, and
e [1] since the latter set is connected (Proposition 14.3.22).□
′ \𝒁
therefore equal to Δ𝑟𝑛−1

Corollary 14.4.6 There are at most 𝑚 first-tier leaves of 𝑽

e in Δ𝑟𝑛 .

The concept of leaf and the procedure we are following are highly coordinate
dependent.
486 14 Analytic Stratifications

Example 14.4.7 If 𝑃 (𝑧 1 ; 𝑧 2 ) = 𝑧22 − 𝑧 1 the subset 𝑽\

e {0} is the single first-tier leaf.
However, if we exchange 𝑧1 and 𝑧 2 the whole of 𝑽 is a first-tier leaf.
e

In the sequel we write

−1 [1]
e [1] = 𝑽
𝑽 e [2] = 𝑽
e ∩ Δ𝑟𝑛 , 𝑽 e [1] ∩ 𝜋˜ ′ e
𝒁 ; (14.4.4)

e [2] is a subvariety of Δ𝑟𝑛 .

𝑽
After a linear change of the variables 𝑧1 , ..., 𝑧 𝑛−1 and a concomitant redefinition
of Δ𝑟𝑛−1′ we may assume that 𝐷 (0, ..., 0, 𝑧 𝑛−1 ) ≠ 0 for some 𝑧 𝑛−1 ≠ 0. We apply
the Weierstrass Preparation Theorem to 𝐷 (𝑧 ′) and replace it, as a defining function
[1] [2]
of e𝒁 in Δ𝑟𝑛−1
′ , by a trim Weierstrass polynomial 𝑃2 (𝑧 ′′; 𝑧 𝑛−1 ) where 𝑧 ′′ =
(𝑧1 , ..., 𝑧 𝑛−2 ) (the reason for the notation will be apparent momentarily). Let 𝑚 2[2] =
deg 𝑃2[2] and let 𝜌2,𝑘 [2]
(𝑧 ′′), 𝑘 = 1, ..., 𝑚 2[2] , denote the roots of 𝑃2[2] (𝑧 ′′; 𝑧 𝑛−1 ); we
denote by 𝐷 2[2] (𝑧 ′′) the discriminant of 𝑃2[2] (𝑧 ′′; 𝑧 𝑛−1 ). We adjust the radii 𝑟 𝑗 ,
𝑗 = 1, ..., 𝑛 − 2, to ensure that the analogue of (14.4.1) holds:
[2]
sup 𝜌2,𝑘 (𝑧 ′′) < 𝑟 𝑛−1 , 𝑘 = 1, ..., 𝑚 2[2] .
𝑧 ′′ ∈Δ𝑟𝑛−2
′′

Next, we form the Weierstrass polynomial

[2]
𝑚2
Ö
[2]
𝑃♭ (𝑧 ′′; 𝑧 𝑛 ) = 𝑃 𝑧 ′′, 𝜌2,𝑘 (𝑧 ′′) ; 𝑧 𝑛 .
𝑘=1

The nonleading coefficients of 𝑃♭ (𝑧 ′′; 𝑧 𝑛 ) are holomorphic functions of 𝑧 ′′ ∈ Δ𝑟𝑛−2 ′′

vanishing at 𝑧 ′′ = 0. But the discriminant of 𝑃♭ (𝑧 ′′; 𝑧 𝑛 ) might vanish identically; we

replace 𝑃♭ by its trimming, 𝑃1[2] (𝑧 ′′; 𝑧 𝑛 ), and we denote by 𝐷 1[2] (𝑧 ′′) the discrimi-
nant of 𝑃1[2] (𝑧 ′′; 𝑧 𝑛−1 ). We define
[2]
n o
[2] [2]
𝒁 = 𝑧 ′′ ∈ Δ𝑟𝑛−2
e ′′ ; 𝐷
1
(𝑧 ′′
) 𝐷 2
(𝑧 ′′
) = 0 . (14.4.5)

At every point 𝑧 ′′ ∈ Δ𝑟𝑛−2 e [2] the roots 𝜌 [2] (𝑧 ′′) of 𝑃 [2] (𝑧 ′′; 𝑧 𝑛−1 ) are distinct and
′′ \ 𝒁 2,𝑘 2
[2] [2]
so are the roots 𝜌1,𝑘 ′′
′ (𝑧 ) of 𝑃1 (𝑧 ′′; 𝑧 𝑛 ) (𝑘 ′ = 1, ..., 𝑚 1[2] = deg 𝑃1[2] ). We adjust
further the radii 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 − 2, to ensure that

[2]
sup 𝜌1,𝑘 (𝑧 ′′) < 𝑟 𝑛 , 𝑘 = 1, ..., 𝑚 2[2] .
𝑧 ′′ ∈Δ𝑟𝑛−2
′′

The following statement is important.

14.4 Local Partitions of a Complex Hypersurface 487

Lemma 14.4.8 Let 𝑈 ′′ be a connected open subset of Δ𝑟𝑛−2 e [2] in which all roots
′′ \ 𝒁

𝜌 [2] [2]
𝑗,𝑘 (1 ≤ 𝑘 ≤ 𝑚 𝑗 ) are pairwise distinct holomorphic functions ( 𝑗 = 1, 2). If

[2] ′′ [2] ′′ ) ; 𝜌 [2] (𝑧 ′′ ) = 0 for some 𝑧 ′′ ∈ 𝑈 ′′
1 ≤ 𝑘 𝑗 ≤ 𝑚 𝑗 , 𝑗 = 1, 2, and 𝑃 𝑧 , 𝜌2,𝑘 2
(𝑧 1,𝑘1

[2] [2]
then 𝑃 𝑧 , 𝜌2,𝑘2 (𝑧 ) ; 𝜌1,𝑘1 (𝑧 ) = 0 for all 𝑧 ∈ 𝑈 ′′.
′′ ′′ ′′ ′′

e [2] every root of 𝑃 𝑧 ′′, 𝜌 [2] (𝑧 ′′) ; 𝑧 𝑛 is a root of

Proof Whatever 𝑧 ′′ ∈ Δ𝑟𝑛−2
′′ \ 𝒁 2,𝑘2
𝑃1[2] (𝑧 ′′; 𝑧 𝑛 ); in 𝑈 ′′ two such roots are everywhere distinct or everywhere equal. □

Let 𝜋˜ ′′ : Δ𝑟(𝑛) −→ Δ𝑟𝑛−2

′′ be the coordinate projection (𝑧 ′′, 𝑧 𝑛−1 , 𝑧 𝑛 ) ↦→ 𝑧 ′′; we
−1 [2] −1 [1]
have 𝜋˜ ′′ e𝒁 ⊂ 𝜋˜ ′ e
𝒁 .

Proposition 14.4.9 We have

−1
−1 −1 [2]
e [2] ∩ 𝜋˜ ′′ Δ𝑟𝑛−2 [2] [1] ′ e [1]

𝑽 ′′ \ 𝒁
e = 𝑽
e ∩ ˜
𝜋 𝒁 \𝜋˜ ′′ e
𝒁 ; (14.4.6)

−1
e [2] ∩ 𝜋˜ ′′ Δ𝑛−2 e [2] is either empty or else a complex-analytic submanifold of

𝑽 𝑟 ′′ \ 𝒁
Δ𝑟(𝑛) of codimension 2.

Proof The left-hand side in (14.4.6) is contained in the right-hand side by the
−1
e [2] . Each point 𝑧◦ ∈ 𝑽
e [2] ∩ 𝜋˜ ′′ Δ𝑛−2 e [2] has an open neighborhood

definition of 𝑽 𝑟 ′′ \ 𝒁

𝑈 in Δ𝑟(𝑛) in which 𝑽e [2] is defined by equations 𝑧 𝑛−1 = 𝜌 [2] (𝑧 ′′) [since 𝑽 e [2] ⊂
2,𝑘2
−1 [1]
e [2] ⊂ 𝑽
e [1] ; 𝑘 2 = 1, ..., 𝑚 [2] ].

[2]
𝜋˜ ′ e
𝒁 ] and 𝑃 𝑧 ′′, 𝜌2,𝑘 2
(𝑧 ′′) ; 𝑧 𝑛 = 0 [since 𝑽 2
The latter equation implies 𝑃1[2] (𝑧 ′′; 𝑧 𝑛 ) = 0 and therefore 𝑧 𝑛 = 𝜌1,𝑘
[2]
1
(𝑧 ′′) for
some 𝑘 1 = 1, ..., 𝑚 1[2] ; in all this 𝑧 ′′ ∈ 𝑈 ′′ ⊂ 𝜋˜ ′′ (𝑈) with 𝑈 ′′ sufficiently small
that 𝜌 [2]
𝑗,𝑘 𝑗 ∈ O
(𝑈 ′′), 𝑗 = 1, 2. This proves the last claim in the statement. Lemma

[2] ′′ ) ; 𝜌 [2] (𝑧 ′′ ) = 0 for some 𝑧 ′′ ∈ 𝑈 ′′ then
14.4.8 tells us that if 𝑃 𝑧 ′′, 𝜌2,𝑘 2
(𝑧 1,𝑘1
[1]

[2] [2]
𝑧 , 𝜌2,𝑘2 (𝑧 ) , 𝜌1,𝑘1 (𝑧 ) ∈ 𝑽 for all 𝑧 ′′ ∈ 𝑈 ′′. By Corollary 14.2.5 this implies
′′ ′′ ′′ e
−1
e [2] ∩ 𝜋˜ ′′ Δ𝑛−2 e [2] that intersects 𝑽
e [1] is

that any connected component of 𝑽 𝑟 ′′ \ 𝒁
[1]
entirely contained in 𝑽
e . □

By a second-tier leaf [implicitly, of the sought partition of 𝑽e in Δ𝑟(𝑛) ] we shall

−1
e [2] ∩ 𝜋˜ ′′ Δ𝑛−2 [2]

mean a connected component eΛ [2]
𝜄 of the submanifold 𝑽 𝑟 ′′ \ 𝒁
e .
488 14 Analytic Stratifications

Example 14.4.10 2Take 𝑃 (𝑧1 , 𝑧2 ; 𝑧 3 ) = 𝑧3 (𝑧3 − 𝑧2 + 𝑧1 ) (𝑧 3 − 2𝑧 2 + 4𝑧 1 ) in C3 ; 𝑽

e
is the union of 3 hyperplanes; 𝑃 is trim. The trimming of the discriminant of 𝑃 is
𝑃2[2] (𝑧1 ; 𝑧2 ) = (𝑧2 − 𝑧1 ) (𝑧2 − 2𝑧1 ) (𝑧2 − 3𝑧 1 ) whose roots are distinct except when
[1]
𝑧1 = 0; e𝒁 ⊂ C2 is the union of the 3 planes 𝑧2 = 𝑘 𝑧1 , 𝑘 = 1, 2, 3. Here we have

𝑃 (𝑧 1 , 𝑧1 ; 𝑧3 ) 𝑃 (𝑧 1 , 2𝑧 1 ; 𝑧 3 ) 𝑃 (𝑧1 , 3𝑧1 ; 𝑧3 ) = 𝑧53 (𝑧 3 − 𝑧 1 ) (𝑧3 − 2𝑧1 ) 2 (𝑧 3 + 2𝑧 1 ) ,

whence 𝑃1[2] (𝑧1 ; 𝑧3 ) = 𝑧3 (𝑧3 − 𝑧 1 ) 𝑧23 − 4𝑧21 , whose roots are also distinct if and

[2]
only if 𝑧 1 ≠ 0. Thus e 𝒁 = {0} ⊂ C. The variety 𝑽 e [2] is the union of 6 one-
dimensional subvarieties of C3 defined by the following equations:

𝑧 2 = 𝑧1 , 𝑧3 (𝑧 3 + 2𝑧 1 ) = 0, (14.4.7)
𝑧 2 = 2𝑧1 , 𝑧 3 (𝑧3 − 𝑧1 ) = 0,
𝑧 2 = 3𝑧1 , 𝑧 3 (𝑧3 − 2𝑧1 ) = 0.

There are 3 first-tier leaves, respectively defined, in the connected open set

𝑧 ∈ C3 ; (𝑧 2 − 𝑧1 ) (𝑧 2 − 2𝑧 1 ) (𝑧2 − 3𝑧1 ) ≠ 0

by the equations 𝑧 3 = 0, 𝑧 3 = 𝑧 2 − 𝑧1 , 𝑧 3 = 2 (𝑧 2 − 2𝑧 1 ). There are 6 second-tier

leaves, defined by the equations (14.4.7) in the connected open set 𝑧 ∈ C3 ; 𝑧 1 ≠ 0 .

We extrapolate the preceding and describe the procedure at every step. We will use
the following notation, for 3 ≤ 𝜈 < 𝑛: 𝑧 (𝜈) = (𝑧1 , ..., 𝑧 𝑛−𝜈 ); 𝜋˜ (𝜈) : Δ𝑟(𝑛) −→ Δ𝑟(𝑛−𝜈)
(𝜈)

(𝜈) (𝜈)
is the coordinate projection 𝑧 , 𝑧 𝑛−𝜈+1 , ..., 𝑧 𝑛 ↦→ 𝑧 (for 𝜈 = 1, 2, 𝑧 (𝜈) = 𝑧 ′ or
𝑧 ′′, 𝜋˜ (𝜈) = 𝜋˜ ′ or 𝜋˜ ′′ respectively).
Let 𝜈◦ ∈ Z+ , 2 ≤ 𝜈◦ ≤ 𝑛 − 1; suppose we have defined, recursively for every
𝜈 = 2, ..., 𝜈◦ , subvarieties 𝑽 e [𝜈 ] ⊂ 𝑽 e [𝜈−1] in Δ𝑟(𝑛) by a system of trim Weierstrass
polynomial equations

𝑃 [𝜈
𝑗
]
𝑧 (𝜈) ; 𝑧 𝑛− 𝑗+1 = 0, 𝑗 = 1, ..., 𝜈, (14.4.8)

Let 𝐷 [𝜈 ]
𝑧 (𝜈) denote the discriminant of 𝑃 [𝜈 ] 𝑧 (𝜈) ; 𝑧
𝑛− 𝑗+1 ; then 𝐷 [𝜈 ] 𝑧 (𝜈) =
𝑗 𝑗
[𝜈 ] (𝜈) [𝜈 ] (𝜈) (𝑛−𝜈)
𝐷1 𝑧 · · · 𝐷𝜈 𝑧 ∈ O Δ𝑟 (𝜈) ; we define

[𝜈 ]
n o
𝒁 = 𝑧 (𝜈) ∈ Δ𝑟(𝑛−𝜈)
e (𝜈) ; 𝐷 [𝜈 ]
𝑧 (𝜈)
= 0 . (14.4.9)

We shall systematically use the following notation

[𝜈 ]
Γ [𝜈 ] = Δ𝑟(𝑛−𝜈)
e (𝜈) 𝒁 ;
\e (14.4.10)

2 I am indebted to A. Bove for this example.

14.4 Local Partitions of a Complex Hypersurface 489

Γ [𝜈 ] is a connected open subset of Δ𝑟(𝑛−𝜈)

e (𝜈) . We shall reason under the following
hypotheses:
−1
[𝜈+1] [𝜈 ]

(1) If 𝜈 < 𝜈◦ then 𝑽
e ⊂ 𝜋˜ (𝜈) e
𝒁 and [cf. (14.4.6)]
!
−1 −1 −1
[𝜈+1] [𝜈 ] [𝜈 ] [𝜈+1]

(𝜈+1) e[𝜈+1] (𝜈) {𝜈+1}
𝑽
e ∩ 𝜋˜ Γ =𝑽
e ∩ 𝜋˜ 𝒁
e \𝜋˜ 𝒁
e . (14.4.11)

(2) Let 𝜌 [𝜈
𝑗,𝑘
]
𝑧 (𝜈) denote the roots of 𝑃 [𝜈 ] , 𝑗 = 1, ..., 𝜈 ≤ 𝜈 , 𝑘 = 1, ..., 𝑚 [𝜈 ] =
𝑗 ◦ 𝑗
deg 𝑃 [𝜈 ]
𝑗 ; then [cf. (14.4.1)]

sup 𝜌 [𝜈
𝑗,𝑘
]
𝑧 (𝜈)
< 𝑟 𝑛− 𝑗+1 . (14.4.12)
(𝑛−𝜈)
𝑧 (𝜈) ∈Δ
𝑟 (𝜈)

[𝜈 ]
𝒁
We underline the fact that e is a proper (possibly empty) subvariety of Δ𝑟(𝑛−𝜈)
(𝜈)

and that, if 𝑧 (𝜈) ∈ e Γ [𝜈 ] and 𝑘 ≠ ℓ then 𝜌 [𝜈 ]
𝑗,𝑘 𝑧 (𝜈) ≠ 𝜌 [𝜈 ] 𝑧 (𝜈) . For 𝜈 ≤ 𝜈 ,
𝑗,ℓ ◦
−1
e [𝜈 ] ∩ 𝜋˜ (𝜈) e

𝑽 Γ [𝜈 ] is a complex-analytic submanifold of Δ𝑟(𝑛) of dimension 𝑛 − 𝜈; we
refer to its connected components e Λ [𝜈
𝜄
]
(𝜄 ∈ 𝐼 [𝜈 ] ) as the 𝜈th-tier leaves [implicitly,
−1
e [𝜈 ] ∩ 𝜋˜ (𝜈) e

e in Δ𝑟(𝑛) ]; 𝑽
of the sought partition of 𝑽 Γ [𝜈 ] is locally defined by systems

of equations 𝑧 𝑛− 𝑗+1 = 𝜌 [𝜈 ]
𝑗,𝑘 𝑗 𝑧
(𝜈) , 𝑗 = 1, ..., 𝜈 (see Proposition 14.4.12 below).

Keepin mind e[𝜈 ]

that globally, i.e., in Γ , there might not be holomorphic functions
[𝜈 ] (𝜈)
𝜌 𝑗,𝑘 𝑧 .
Since every 𝑃 [𝜈 𝑗
◦]
is trim we know that 𝐷 [𝜈◦ ] does not vanish identically. If
[𝜈 ]
𝐷 ◦ is identically equal to a nonzero constant the same must be true of each

𝐷 [𝜈◦ ] 𝑧 (𝜈◦ ) , 𝑗 = 1, ..., 𝜈◦ . By Proposition 14.4.3 this is equivalent to saying that

𝑃 [𝜈
𝑗
◦]
𝑧 (𝜈◦ ) ; 𝑧
𝑛− 𝑗+1 = 𝑧 𝑛− 𝑗+1 − 𝑎 𝑗 𝑧 (𝜈◦ ) , 𝑎 ∈ O Δ (𝑛−𝜈◦ ) . The analytic subman-
𝑗 (𝜈
𝑟 ◦ )

ifold (14.4.11), where 𝜈 = 𝜈◦ , is the graph of the holomorphic map

Γ [𝜈 ] ∋ 𝑧 (𝜈◦ ) → 𝑎 1 𝑧 (𝜈◦ ) , ..., 𝑎 𝜈◦ 𝑧 (𝜈◦ ) ∈ C𝜈◦ .
e

In this case the procedure stops.

Suppose 𝐷 [𝜈◦ ] nonconstant. After a linear change of variables we may as-
sume that 𝐷 [𝜈◦ ] 0, ..., 0, 𝑧 𝑛−𝜈◦ ≠ 0 for some 𝑧 𝑛−𝜈◦ ; it follows that there is a trim
Weierstrass polynomial 𝑃 𝜈[𝜈◦ +1 ◦ +1]
𝑧 (𝜈◦ +1) ; 𝑧 𝑛−𝜈◦ with the same roots 𝜌 𝜈[𝜈◦ +1,ℓ
◦ +1]
𝑧 (𝜈◦ +1)

[ℓ = 1, ..., 𝑚 𝜈[𝜈◦ ◦+1+1] = deg 𝑃 𝜈[𝜈◦ +1
◦ +1]
] as 𝐷 [𝜈◦ ] 𝑧 (𝜈◦ ) . For each 𝑗 = 1, ..., 𝜈◦ , we intro-
duce the symmetrization
490 14 Analytic Stratifications

𝑚𝜈◦◦+1+1]
[𝜈
Ö
𝑃♭[𝜈
𝑗
◦ +1]
𝑧 (𝜈◦ +1) ; 𝑧 𝑛− 𝑗+1 = 𝑃 [𝜈
𝑗
◦]
𝑧 (𝜈◦ +1)
, 𝜌 [𝜈◦ +1]
𝜈◦ +1,ℓ 𝑧 (𝜈◦ +1)
; 𝑧 𝑛− 𝑗+1 ;
ℓ=1

◦ +1] (𝑛−𝜈◦ −1) ◦ +1]
we have 𝑃♭[𝜈
𝑗 ∈ O Δ𝑟 (𝜈◦ +1) 𝑧 𝑛− 𝑗+1 . To each 𝑃♭[𝜈 𝑗 we associate its trim-

[𝜈◦ +1]
ming, 𝑃 𝑗 𝑧 ◦ ; 𝑧 𝑛− 𝑗+1 , and to the latter its discriminant, 𝐷 [𝜈
(𝜈 +1)
𝑗
◦ +1]
𝑧 (𝜈◦ +1) .
We define
e [𝜈◦ +1] = 𝑧 ∈ Δ𝑟(𝑛) ; 𝑃 [𝜈◦ +1] 𝑧 (𝜈◦ +1) ; 𝑧 𝑛− 𝑗+1 = 0, 𝑗 = 1, ..., 𝜈◦ + 1
n o
𝑽 𝑗

and
 ◦ +1
𝜈Ö 
[𝜈◦ +1]

= 𝑧 (𝜈◦ ) ∈ Δ𝑟(𝑛−𝜈 ◦) [𝜈◦ +1]

 (𝜈◦ )


𝒁
e (𝜈◦ ) ; 𝐷 𝑗 𝑧 = 0 .
 
 𝑗=1 
−1
e [𝜈◦ +1] ⊂ 𝜋˜ (𝜈◦ ) e [𝜈◦ ]

The definition of 𝑃 𝜈[𝜈◦ +1
◦ +1]
implies 𝑽 𝒁 . We have the analogue of
Lemma 14.4.8:

Lemma 14.4.11 Let 𝑈 (𝜈◦ +1) be a connected open subset of e Γ [𝜈 ] in which all the
[𝜈◦ +1] [𝜈◦ +1]
roots 𝜌 𝑗,𝑘 (1 ≤ 𝑘 ≤ 𝑚 𝑗 ) are pairwise distinct holomorphic functions

( 𝑗 = 1, ..., 𝜈◦ + 1). Let 𝜌 𝜈◦ +1,𝑘 𝑧 ◦ +1) be a root of 𝑃 𝜈[𝜈◦ +1
[𝜈◦ +1] (𝜈 ◦ +1]
𝑧 (𝜈◦ +1) ; 𝑧 𝑛−𝜈◦ and
1 ≤ 𝑘 𝑗 ≤ 𝑚 [𝜈
𝑗
◦ +1]
, 1 ≤ 𝑗 ≤ 𝜈◦ . If

∃𝑧 (𝜈◦ +1) ∈ 𝑈 (𝜈◦ +1) , 𝑃 [𝜈
𝑗
◦]
𝑧 (𝜈◦ +1) , 𝜌 𝜈[𝜈◦ +1,𝑘
◦ +1]
𝑧 (𝜈◦ +1) ; 𝜌 [𝜈◦ +1]
𝑗,𝑘 𝑗 𝑧 (𝜈◦ +1) = 0,

then

∀𝑧 (𝜈◦ +1) ∈ 𝑈 (𝜈◦ +1) , 𝑃 [𝜈
𝑗
◦]
𝑧 (𝜈◦ +1)
, 𝜌 [𝜈◦ +1]
𝜈◦ +1,𝑘 𝑧 (𝜈◦ +1)
; 𝜌 [𝜈◦ +1]
𝑗,𝑘 𝑗 𝑧 (𝜈◦ +1)
= 0.

Proof Each root of 𝑃 [𝜈 𝑗
◦]
𝑧 (𝜈◦ +1) , 𝜌 [𝜈◦ +1] 𝑧 (𝜈◦ +1) ; 𝑧
𝜈◦ +1,𝑘 𝑛− 𝑗+1 (1 ≤ 𝑗 ≤ 𝜈◦ , 𝑧
(𝜈◦ +1) ∈

Γ [𝜈 ] ) is a root of 𝑃 [𝜈
e
𝑗
◦ +1]
𝑧 (𝜈◦ +1) ; 𝑧 𝑛− 𝑗+1 ; in 𝑈 (𝜈◦ +1) two such roots are everywhere
distinct or everywhere equal. □
The radii 𝑟 𝑖 , 𝑖 = 1, ..., 𝑛 − 𝜈◦ − 1, are adjusted so that (14.4.12) holds for all
𝜈 ≤ 𝜈◦ + 1.
If 1 ≤ 𝑘 𝑗 ≤ 𝑚 [2] , 𝑗 = 1, 2, and 𝑃 𝑧 ′′ , 𝜌 [2] (𝑧 ′′ ) ; 𝜌 [2] (𝑧 ′′ ) ≠ 0 for some
𝑗 2,𝑘2 1,𝑘1
[2] [2]
′′ ′′ ′′
𝑧 ∈ 𝑈 then 𝑃 𝑧 , 𝜌2,𝑘2 (𝑧 ) ; 𝜌1,𝑘1 (𝑧 ) ≠ 0 for all 𝑧 ∈ 𝑈 ′′.
′′ ′′ ′′

The analogue of Proposition 14.4.9 is valid: (14.4.11) holds for 𝜈 = 𝜈◦ and

−1
e [𝜈◦ +1] ∩ 𝜋˜ (𝜈◦ +1) e

𝑽 Γ [𝜈◦ +1] (14.4.13)
14.4 Local Partitions of a Complex Hypersurface 491

is a complex-analytic submanifold of Δ𝑟(𝑛) of dimension 𝑛 − 𝜈◦ − 1. The proof is

essentially the same as that of Proposition 14.4.9 with Lemma 14.4.11 playing the
role of Lemma 14.4.8 to prove that if 𝑽 e [𝜈◦ ] intersects a connected component e Λ [𝜈
𝜄
◦ +1]

[𝜈 ]
of (14.4.13) then e Λ [𝜈
𝜄
◦ +1]
⊂𝑽 e ◦ .
The procedure can be repeated until, upon reaching 𝜈 = 𝜈∗ ≤ 𝑛, we have de-
fined the subvariety 𝑽 e [𝜈∗ ] in 𝑽
e [𝜈∗ −1] by the vanishing of first-degree polynomials

𝑃 [𝜈
𝑗
∗]
𝑧 (𝜈∗ +1) ; 𝑧 𝑛− 𝑗+1 = 𝑧 𝑛− 𝑗+1 − 𝑎 𝑗 𝑧 (𝜈∗ +1) , 𝑎 𝑗 ∈ O Δ𝑟(𝑛−𝜈 ∗ −1)
(𝜈∗ +1) , 𝑗 = 1, ..., 𝜈∗ ,
with the understanding that 𝑧 (𝜈∗ +1) = 0 if 𝜈∗ = 𝑛, when the system of equations is
e [𝑛] = {0}. At this point we have a sequence of nested
𝑧 𝑗 = 0, 𝑗 = 1, ..., 𝑛, and 𝑽
subvarieties
e [1] ⊃ · · · ⊃ 𝑽
Δ𝑟(𝑛) ⊃ 𝑽 e [𝜈 ] ⊃ 𝑽
e [𝜈+1] ⊃ · · · ⊃ 𝑽
e [𝜈∗ ] . (14.4.14)

We have the important properties (14.4.11)–(14.4.12) for all 𝜈. The closures of the
two sides in (14.4.11) are equal and thus we can state:
Proposition 14.4.12 For each 𝜈 < 𝜈∗ ,
−1
e [𝜈+1] = 𝑽
e [𝜈 ] ∩ 𝜋˜ (𝜈) e [𝜈 ]

𝑽 𝒁 . (14.4.15)

A direct consequence of (14.4.15) is that 𝑽 e [𝜈 ] = 𝑽

e [𝜈 ] \𝑽e [𝜈+1] ; the 𝜈th-tier leaves
are the connected components of 𝑽 e [𝜈 ] \𝑽
e [𝜈+1] ; (14.4.11) and (14.4.15) imply that,
as 𝜈 varies, the sets
−1
e [𝜈 ] ∩ 𝜋˜ (𝜈) e
Ø
𝑽 Γ [𝜈 ] = Λ [𝜈
e 𝜄
]
(14.4.16)
𝜄 ∈𝐼 [𝜈 ]

are pairwise disjoint, whence the finite partition

𝑛 Ø
Ø
e ∩ Δ𝑟(𝑛) =
𝑽 Λ [𝜈
e 𝜄 .
]
(14.4.17)
𝜈=1 𝜄 ∈𝐼 [𝜈 ]

The analogues of Proposition 14.4.5 and Corollary 14.4.6 are valid. We content
ourselves with stating them, the proofs being essentially identical to those of the
originals.
Proposition 14.4.13 Each point of a 𝜈th-tier leaf e Λ [𝜈
𝜄
]
of 𝑽e in Δ𝑟(𝑛) has an open
[𝜈 ]
neighborhood 𝑈 ⊂ Δ𝑟(𝑛) such that 𝜋˜ (𝜈) (𝑈) ∩ e𝒁 = ∅ and
n o
Λ [𝜈
e ] (𝑛)
𝜄 ∩ 𝑈 = 𝑧 ∈ Δ𝑟 ; 𝑧
(𝜈)
∈ 𝜋˜ (𝜈) (𝑈) , 𝑧 𝑛− 𝑗+1 = 𝜌 [𝜈 ]
𝑗,𝑘 𝑗 𝑧
(𝜈)
, 𝑗 = 1, ..., 𝜈

for a unique set of indices 𝑘 𝑗 , 1 ≤ 𝑘 𝑗 ≤ 𝑚 [𝜈

𝑗
]
( 𝑗 = 1, ..., 𝜈).

Corollary 14.4.14 For every 𝜄 ∈ 𝐼 [𝜈 ] we have 𝜋˜ [𝜈 ] e Λ [𝜈
𝜄
]
Γ [𝜈 ] [cf. (14.4.10)].
=e
492 14 Analytic Stratifications

Corollary 14.4.15 There are at most 𝑚 1[𝜈 ] + · · · + 𝑚 𝜈[𝜈 ] 𝜈th-tier leaves of 𝑽

e in Δ𝑟(𝑛) .

A rephrasing of Proposition 14.4.13 and Corollary 14.4.14 is that every triple
Λ [𝜈
e ]
𝜄 , 𝜋˜
[𝜈 ] , e
Γ [𝜈 ] is a finitely sheeted analytic cover (see [Gunning and Rossi,
1965], p. 101).

Proposition 14.4.16 If e Λ [𝜈𝜄

]
is a 𝜈th-tier leaf there are Weierstrass polynomials

𝑄 [𝜈 ]
𝑧 (𝜈) ; 𝑧 𝑛− 𝑗+1 ∈ O Δ𝑟(𝑛−𝜈)

𝑗 (𝜈) 𝑧 𝑛− 𝑗+1 , 𝑗 = 1, ..., 𝜈,

such that 𝑄 [𝜈
𝑗
]
𝑧 (𝜈) ; 𝑧
𝑛− 𝑗+1 divides 𝑃 [𝜈 ]
𝑗 𝑧 (𝜈) ; 𝑧
𝑛− 𝑗+1 [see (14.4.8)] for each 𝑗 and
n o
Λ [𝜈
e 𝜄
]
= 𝑧 ∈ Γ
e [𝜈 ]
; 𝑄 [𝜈 ]
𝑗 𝑧 (𝜈)
; 𝑧 𝑛− 𝑗+1 = 0, 𝑗 = 1, ..., 𝜈 . (14.4.18)

Λ [𝜈
The closure of e 𝜄
]
in Δ𝑟(𝑛) is an irreducible
[𝜈 ]analytic
subvariety of Δ𝑟(𝑛) ; its regular
part is a connected component of ℜ𝑛−𝜈 𝑽 e .

Proof Suppose e Λ [𝜈 ]
𝜄 , 𝜋˜
[𝜈 ] , e
Γ [𝜈 ] is 𝑁-sheeted: to every 𝑧 (𝜈) ∈ e Γ [𝜈 ] there are exactly
𝑁 distinct points

𝑧 (𝜈) , 𝜁ℓ,𝑛−𝜈 𝑧 (𝜈) , ..., 𝜁ℓ,𝑛 𝑧 (𝜈) ∈ e Λ [𝜈
𝜄
]

(ℓ = 1, ..., 𝑁). Since 𝑃 [𝜈
𝑗
]
𝑧 (𝜈) ; 𝜁ℓ,𝑛− 𝑗+1 𝑧 (𝜈) = 0 for all ℓ the polynomial

𝑁
Ö
𝑄 [𝜈
𝑗
]
𝑧 (𝜈)
; 𝑧 𝑛− 𝑗+1 = 𝑧 𝑛− 𝑗+1 − 𝜁ℓ,𝑛− 𝑗+1 𝑧 (𝜈)
ℓ=1

divides 𝑃 [𝜈
𝑗
]
𝑧 (𝜈) ; 𝑧 𝑛− 𝑗+1 . Although each 𝜁ℓ,𝑛− 𝑗+1 𝑧 (𝜈) might not represent a
n o
univalued function in e Γ [𝜈 ] the set 𝑬 𝑗 𝑧 (𝜈) = 𝜁1,𝑛− 𝑗+1 𝑧 (𝜈) , ..., 𝜁 𝑁 ,𝑛− 𝑗+1 𝑧 (𝜈)
can be viewed as a set-valued function in e Γ [𝜈 ] that can be extended continuously (for
[𝜈 ]

(𝑛−𝜈) [𝜈 ]
the natural Euclidean metric) to Δ (𝜈) [recall that 𝒁 is the nullset of 𝐷
𝑟
e 𝑧 (𝜈) ;
−1
[𝜈 ]

see (14.4.9)]; its symmetric functions are holomorphic functions in Δ𝑟(𝑛) \𝜋˜ [𝜈 ] e 𝒁

continuous in Δ𝑟(𝑛−𝜈)
(𝜈) : indeed, if 1 ≤ 𝑗 < 𝑘 ≤ 𝜈, 𝑬 𝑗 𝑧 (𝜈) and 𝑬 𝑘 𝑧 (𝜈) only
differ by the order of their components. The Riemann
Extension Theorem 14.3.7
[𝜈 ] (𝜈)
implies that the coefficients of 𝑄 𝑗 𝑧 ; 𝑧 𝑛− 𝑗+1 can be holomorphically extended

to the whole of Δ𝑟(𝑛−𝜈) (𝜈) . Obviously, the extension still divides 𝑃 [𝜈
𝑗
]
𝑧 (𝜈) ; 𝑧
𝑛− 𝑗+1 .

Λ [𝜈
The subvariety e 𝜄
]
Λ [𝜈
is irreducible because e 𝜄
]
is connected (Theorem 14.2.7). □
14.5 Local Stratifications of a Real-Analytic Variety 493

[𝜇]
Corollary 14.4.17 If 1 ≤ 𝜈 < 𝜇 ≤ 𝜈∗ then an arbitrary 𝜇th-tier leaf e Λ 𝜅 is
Λ [𝜈
contained in the boundary 𝜕 e 𝜄
]
Λ [𝜈
of at least one 𝜈th-tier leaf e 𝜄 .
]

Proof It suffices to prove the claim for 𝜇 = 𝜈 + 1 and then use induction on 𝜈. By
(14.4.15) an arbitrary connected component e e [𝜈+1] \𝑽
Λ 𝜅[𝜈+1] of 𝑽 e [𝜈+2] is contained
in 𝑽e [𝜈 ] \𝑽
e [𝜈+1] . Since 𝑽
e [𝜈 ] \𝑽
e [𝜈+1] has finitely many connected components there is
[𝜈 ]
at least one of them, e Λ 𝜄 , whose closure contains an open subset of e Λ 𝜅[𝜈+1] . Since
Λ [𝜈
e 𝜄
]
is an analytic variety (Proposition 14.4.16) we must have e Λ 𝜅[𝜈+1] ⊂ e Λ [𝜈
𝜄
]
by
Corollary 14.2.5. Then the claim follows from the fact that e Λ 𝜅[𝜈+1] ∩ eΛ [𝜈
𝜄
]
= ∅. □

e = ℜ𝑛−1 𝑽
Corollary 14.4.18 We have 𝑽 e .

Proof Direct consequence of Corollary 14.4.17. □

Thus dimC 𝑽e = 𝑛 − 1 (cf. Definition 14.2.10).

Corollary 14.4.19 We have ℜ 𝑽 e = ℜ𝑛−1 𝑽 e ,𝔖 𝑽 e = 𝑽\ℜ
e 𝑛−1 𝑽
e .

Proof If there existed a connected component ℭ of ℜ 𝑽 e such that dimC ℭ < 𝑛 − 1

we would have ℭ ∩ ℜ𝑛−1 𝑽 e = ∅ [since ℭ1 ∩ ℭ2 = ∅ if ℭ1 ≠ ℭ2 are connected

components of ℜ 𝑽 e ], contradicting Corollary 14.4.18. □

Remark 14.4.20 We have not proved that the partition (14.4.17) is a stratification,
[𝜇]
as we have not proved that if an arbitrary 𝜇th-tier leaf e
Λ 𝜅 intersects the closure
[𝜇]
Λ [𝜈
e 𝜄
]
Λ𝜅 ⊂ e
of a 𝜈th-tier leaf then e Λ [𝜈
𝜄 .
]

Remark 14.4.21 Proposition 14.4.16 and the ensuing corollaries are not valid, in
general, for a real variety (even a hypersurface, e.g., the
real Whitney umbrella) nor
for general complex varieties, as in the example 𝑽e = 𝑧 ∈ C3 ; 𝑧1 𝑧3 = 𝑧 2 𝑧3 = 0 , the
union of the 2-plane 𝑧 3 = 0 and the 1-plane 𝑧 1 = 𝑧2 = 0.

14.5 Local Stratifications of a Real-Analytic Variety

14.5.1 Local partition of a real subvariety

Since our viewpoint continues to be purely local we now reason in an open subset Ω
of R𝑛 ; by 𝑽 we shall mean a real-analytic subvariety of Ω; we assume 0 ∈ 𝑽. We are
not assuming here that 𝑽 is a hypersurface, i.e., we may have 𝑑 = dim 𝑽 < 𝑛 − 1;
but, as indicated in Remark 14.2.2, there is no loss of generality in assuming that 𝑽
is the nullset of a single real-valued function 𝑓 ∈ C 𝜔 (Ω).
494 14 Analytic Stratifications

Remark 14.5.1 The first important difference between a real subvariety 𝑽 (even an
irreducible hypersurface) and a complex hypersurface is that, in general, ℜ𝑑 (𝑽) is
not dense in 𝑽 (cf. Corollary 14.4.18). This is the case for the real Whitney umbrella
(Subsection 14.2.4).

We denote by 𝑓 (𝑧) the holomorphic extension of 𝑓 (𝑥) to an open subset Ω

e of
C𝑛 such that 0 ∈ Ω = Ω ∩ R ; we are prepared to contract Ω about Ω as much as
e 𝑛 e
needed. We define n o
𝑽 e 𝑓 (𝑧) = 0 ;
e = 𝑧 ∈ Ω;

e ∩ R𝑛 and 𝑓 (0) = 0; 𝑽
obviously, 𝑽 = 𝑽 e is a complex hypersurface, in general with
singularities.

Example 14.5.2 Take 𝑽 = {0}, the origin in R2 , and 𝑓 (𝑥) = 𝑥12 + 𝑥 22 ; in this case

e = 𝑧 ∈ C2 ; 𝑧 1 = ±𝑖𝑧2

𝑽

is the union of two complex planes in C2 .

We shall assume that Ω contains the multi-interval

𝔔𝑟𝑛 = Δ𝑟𝑛 ∩ R𝑛 = 𝑥 ∈ R𝑛 ; 𝑥 𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 .

Note that 𝑽 ∩ 𝔔𝑟𝑛 = 𝑽e [1] ∩ R𝑛 where 𝑽 e [1] = 𝑽

e ∩ Δ𝑟𝑛 .
The reduction of the defining equation, 𝑓 (𝑥) = 0, to a sequence of systems of
type (14.4.8) proceeds as in the complex case – with some important differences.
To explain the differences we assume that we have reached the stage where the real
versions of the equations (14.4.8),

𝑃 [𝜈
𝑗
]
𝑥 (𝜈) ; 𝑥 𝑛− 𝑗+1 = 0, 𝑗 = 1, ..., 𝜈, (14.5.1)

define a subvariety 𝑽 [𝜈 ] of 𝔔𝑟𝑛 [𝑥 (𝜈) = (𝑥 1 , ..., 𝑥 𝑛−𝜈 ) ∈ 𝔔𝑟𝑛−𝜈

(𝜈) , 𝑟
(𝜈) = (𝑟 , ..., 𝑟
1 𝑛−𝜈 )].
This is true when 𝜈 = 1, in which case the trim Weierstrass polynomial 𝑃 [1] =
𝑃1 · · · 𝑃 𝑁 is real (Proposition 14.4.2); the discriminant 𝐷 [1] of 𝑃 [1] is real. Thus,
in (14.5.1) each Weierstrass polynomial 𝑃 [𝜈 𝑗
]
can be taken to be trim and real; its
discriminant 𝐷 [𝜈 ]
𝑗 is real and so is, therefore, 𝐷
[𝜈 ] = 𝐷 [𝜈 ] · · · 𝐷 [𝜈 ] ; we define
1 𝜈
n o
𝒁 [𝜈 ] = 𝑥 (𝜈) ∈ 𝔔𝑟𝑛−𝜈
(𝜈) ; 𝐷
[𝜈 ]
𝑥 (𝜈) = 0 . (14.5.2)

We are always assuming that (14.4.12) is satisfied for all 𝑗 = 1, ..., 𝜈, 𝑘 = 1, ..., 𝑚 [𝜈
𝑗 .
]

Remark 14.5.3 It should be underlined that 𝐷 [𝜈 𝑗
]
𝑥 (𝜈) ≠ 0 is a statement about

the complex roots of 𝑃 [𝜈
𝑗
]
𝑥 (𝜈) ; 𝜁 , namely that they are pairwise distinct.
14.5 Local Stratifications of a Real-Analytic Variety 495

In carrying out the next step of the procedure we encounter two essential differ-
ences between the real and the complex case:
(1) if dimR 𝒁 [𝜈 ] = 𝑛 − 𝜈 − 1, 𝔔𝑟𝑛−𝜈
(𝜈) \𝒁
[𝜈 ]
may not be connected;
[𝜈 ]
(2) if Γ is a connected component of 𝔔𝑟𝑛−𝜈 \𝒁 [𝜈 ] then, for some 𝑗, 1 ≤ 𝑗 ≤ 𝜈,
(𝜈)
there may not be any real root 𝜌 [𝜈 ]
𝑗,ℓ 𝑥
(𝜈) of

𝑃 [𝜈 ] (𝜈)
𝑗 (𝑥 ; 𝑧 𝑛− 𝑗+1 ) = 0 (14.5.3)

for any 𝑥 (𝜈) ∈ Γ [𝜈 ] .

Let 𝜋 [𝜈 ] denote the coordinate projection 𝔔𝑟𝑛 ∋ 𝑥 ↦→ 𝑥 (𝜈) ∈ 𝔔𝑟𝑛−𝜈
(𝜈) ; thus 𝜋
[𝜈 ] is
[𝜈 ]
the restriction of 𝜋˜ to real space. When (2) happens for some 𝑗, 1 ≤ 𝑗 ≤ 𝜈, then

𝜋 [𝜈 ] 𝑽 [𝜈 ] ∩ Γ [𝜈 ] = ∅. (14.5.4)

The following, however, must be kept in mind.

Lemma 14.5.4 If a root 𝜌 [𝜈 𝑗,ℓ
]
𝑥 (𝜈) of (14.5.3) is real for some 𝑥 (𝜈) ∈ Γ [𝜈 ] then

[𝜈 ]
𝜌 𝑗,ℓ 𝑥 (𝜈) is a simple real root of (14.5.3) for every 𝑥 (𝜈) ∈ Γ [𝜈 ] and 𝜌 [𝜈
𝑗,ℓ
]
𝑥 (𝜈) ∈

C Γ
𝜔 [𝜈 ] .

Proof Direct consequence of the fact that 𝐷 [𝜈 𝑗
]
𝑥 (𝜈) ≠ 0 for all 𝑥 (𝜈) ∈ Γ [𝜈 ] : if a
complex root of a real Weierstrass polynomial becomes real at a point its multiplicity
at that point increases. □

Lemma 14.5.5 Each simple real root 𝜌 [𝜈 ]
𝑗,ℓ 𝑥
(𝜈) ∈ C 𝜔 Γ [𝜈 ] of (14.5.3) can be

extended as a continuous function to the closure Γ [𝜈 ] of Γ [𝜈 ] in 𝔔𝑟𝑛−𝜈

(𝜈) .

Proof Let 𝑥◦(𝜈) ∈ 𝜕Γ [𝜈 ] be arbitrary and 𝜌 [𝜈
𝑗,ℓ
]
𝑥 (𝜈) be a simple real root of (14.5.3)

in Γ [𝜈 ] . There is an 𝜉◦ ∈ R satisfying 𝑃 [𝜈 ] (𝜈)

𝑗 (𝑥 ◦ ; 𝜉◦ ) = 0 with the following property:

there is a sequence of points 𝑥 (𝜈) ∈ Γ [𝜈 ] converging to 𝑥◦(𝜈) such that 𝜌 [𝜈𝑗,ℓ
]
𝑥 (𝜈) ↦→

[𝜈 ] (𝜈)
𝜉◦ . There is an 𝑅 > 0 such that 𝑃 𝑗 𝑥 ◦ ; 𝜁 ≠ 0 if 0 < |𝜁 − 𝜉◦ | < 𝑅 (𝜁 ∈ C) and
therefore there is an integer 𝑚 ◦ ≥ 1 (the multiplicity of the root 𝜉◦ ) such that

1
∮
1 𝜕𝑃 [𝜈 ]
𝑥 ◦(𝜈) ; 𝜁 d𝜁.
𝑗
𝑚◦ =
2𝜋𝑖 𝑃 [𝜈 ]
𝑥◦(𝜈) ; 𝜁 𝜕𝜁
|𝜁 − 𝜉◦ |=𝜀 𝑗
496 14 Analytic Stratifications

whatever 𝜀 ∈ (0, 𝑅). To each 𝜀 there is a 𝛿 > 0 such that 𝑥 (𝜈) − 𝑥◦(𝜈) < 𝛿 im-

plies 𝑃 [𝜈
𝑗
]
𝑥 (𝜈) ; 𝜁 ≠ 0 for all 𝜁 on the circle |𝜁 − 𝜉◦ | = 𝜀. As a consequence,
if 𝑥 (𝜈) − 𝑥◦(𝜈) < 𝛿 there are 𝑚 ◦ roots of (14.5.3) inside the disk |𝜁 − 𝜉◦ | < 𝜀. If

𝜌 [𝜈
𝑗,ℓ
]
𝑥 (𝜈) − 𝜉 < 𝜀 for some 𝑥 (𝜈) ∈ Γ [𝜈 ] such that 𝑥 (𝜈) − 𝑥 (𝜈) < 𝛿 (which we
◦ ◦

know to be true) then the same is true for all 𝑥 (𝜈) ∈ Γ [𝜈 ] such that 𝑥 (𝜈) − 𝑥 ◦(𝜈) < 𝛿.□

−1
Suppose 𝑽 [𝜈 ] contains points outside 𝜋 [𝜈 ] 𝒁 [𝜈 ] and let Γ 𝜄[𝜈 ] (𝜄 ∈ 𝑰 [𝜈 ] ) be the
[𝜈 ]
connected components of 𝔔𝑟𝑛−𝜈 (𝜈) \𝒁 . In this case, for some 𝜄 ∈ 𝑰 [𝜈 ] there is at

least one simple real root 𝜌 [𝜈 ]
𝑗,ℓ 𝑗 𝑥
(𝜈) ∈ C 𝜔 Γ [𝜈 ] of (14.5.3) for every 𝑗 = 1.., 𝜈.
𝜄
Condition (14.4.12) ensures that

∀𝑥 (𝜈) ∈ Γ 𝜄[𝜈 ] , 𝜌 [𝜈 ]
𝑗,ℓ 𝑗 𝑥 (𝜈)
< 𝑟 𝑛− 𝑗+1 , 𝑗 = 1.., 𝜈. (14.5.5)

The equations
𝑥 𝑛− 𝑗+1 = 𝜌 [𝜈 ]
𝑗,ℓ 𝑗 𝑥 (𝜈)
, 𝑥 (𝜈) ∈ Γ 𝜄[𝜈 ] , 𝑗 = 1.., 𝜈, (14.5.6)
−1
define a C 𝜔 submanifold of 𝜋 [𝜈 ] Γ 𝜄[𝜈 ] ∩ 𝔔𝑟𝑛 of dimension 𝑛 − 𝜈, which we denote
by Λ [𝜈 ] [𝜈 ]
𝜄, 𝑝 (𝑝 = 1, ..., 𝑁 𝜄
[𝜈 ]
= number of sets of 𝜈 simple roots 𝜌1,ℓ1
[𝜈 ]
, ..., 𝜌 𝜈,ℓ 𝜈
). As a
consequence of Lemmas 14.5.4 and 14.5.5 we can state:

Proposition 14.5.6 The coordinate projection 𝜋 [𝜈 ] induces a C 𝜔 diffeomorphism of

Λ [𝜈 ] [𝜈 ]
𝜄, 𝑝 onto Γ 𝜄 that extends as a continuous map of the closure of Λ [𝜈 ] 𝑛
𝜄, 𝑝 in 𝔔𝑟 onto
the closure of Γ 𝜄[𝜈 ] in 𝔔𝑟𝑛−𝜈
(𝜈) .

Under the hypotheses that Λ [𝜈 ]

𝜄, 𝑝 ≠ ∅ for some 𝜄 ∈ 𝐼
[𝜈 ] and some 𝑝 ≤ 𝑁 [𝜈 ] we
𝜄
[𝜈 ]
have dimR 𝑽 = 𝑛 − 𝜈 and
[𝜈 ]
−1 Ø 𝑁Ø
𝜄

ℜ𝑛−𝜈 𝑽 [𝜈 ] ∩ 𝜋 [𝜈 ] 𝔔𝑟𝑛−𝜈
(𝜈) \𝒁
[𝜈 ]
= Λ [𝜈 ]
𝜄, 𝑝 . (14.5.7)
𝜄 ∈𝑰 [𝜈 ] 𝑝=1

[𝜈 ]
The set 𝑽 ✠ = 𝑽 [𝜈 ] \ℜ𝑛−𝜈 𝑽 [𝜈 ] ⊂ 𝒁 [𝜈 ] is an analytic subvariety of 𝔔𝑟𝑛 ,
[𝜈 ]
dim 𝑽 ✠ < dim 𝑽 [𝜈 ] = 𝑛 − 𝜈 (Proposition 14.2.11). We obtain the partition

[𝜈 ]
Ø 𝑁Ø
𝜄
[𝜈 ]
𝑽 [𝜈 ]
= 𝑽✠ ∪ Λ [𝜈 ]
𝜄, 𝑝 . (14.5.8)
𝜄 ∈𝑰 [𝜈 ] 𝑝=1

Proposition 14.5.7 The number of connected components of 𝔔𝑟𝑛 \𝑽 [𝜈 ] is finite.

14.5 Local Stratifications of a Real-Analytic Variety 497

Proof If dim 𝑽 [𝜈 ] ≤ 𝑛 − 2 the complement 𝔔𝑟𝑛 \𝑽 [𝜈 ] is connected. Suppose

dim 𝑽 [𝜈 ] = 𝑛 − 1 (hence 𝜈 = 1) in which case the Λ [𝜈 ]
𝜄, 𝑝 are pairwise disjoint
C hypersurfaces. Induction on the dimension allows us to assert that the number
𝜔
[1]
of connected components of 𝔔𝑟𝑛−1′ \𝒁 (i.e., the set of indices 𝑰 [1] ) is finite. If we
1
Ö
note that 𝑁 𝜄[1] ≤ 𝑚 [1] = deg 𝑃 [1]
𝑗 then the number of connected components of
𝑗=1
−1

𝜋1 Γ 𝜄[1] ∩ 𝔔𝑟𝑛 , which does not exceed 𝑁 𝜄[1] + 2, is finite. □

Consider now the case 𝜋 [𝜈 ] 𝑽 [𝜈 ] ⊂ 𝒁 [𝜈 ] , i.e., dim 𝑽 [𝜈 ] < 𝑛 − 𝜈. Here we
proceed as in the preceding section. After an R-linear change of variables we may
assume that 𝐷 [𝜈 ] (0, ..., 0, 𝑧 𝑛−𝜈 ) ≠ 0 for some 𝑧 𝑛−𝜈 ∈ C; it follows that there
[𝜈+1]
is a trim Weierstrass polynomial 𝑃 𝜈+1 𝑧 (𝜈+1) ; 𝑧 𝑛−𝜈 dividing 𝐷 [𝜈 ] 𝑥 (𝜈) whose
[𝜈 ]

[𝜈+1]
𝒁 ; let 𝜌 𝜈+1,ℓ
nullset in Δ𝑟𝑛 is e 𝑧 (𝜈+1) denote its complex roots. We introduce the
symmetrizations [cf. (14.5.1)]
[𝜈+1]
deg 𝑃𝜈+1
Ö
𝑃♭[𝜈+1]
𝑗 𝑧 (𝜈+1)
; 𝑧 𝑛− 𝑗+1 = 𝑃 [𝜈
𝑗
] [𝜈+1]
𝑧 (𝜈+1) , 𝜌 𝜈+1,ℓ 𝑧 (𝜈+1) ; 𝑧 𝑛− 𝑗+1
ℓ=1

belonging to O Δ𝑟𝑛−𝜈−1(𝜈+1) 𝑧 𝑛− 𝑗+1 ( 𝑗 = 1, ..., 𝜈). To each 𝑃♭[𝜈+1]
𝑗 we associate its

trimming, 𝑃 [𝜈+1]
𝑗 𝑧 (𝜈+1) ; 𝑧 𝑛− 𝑗+1 . Now 𝑽 [𝜈 ] is defined in 𝔔𝑟𝑛 by the equations

𝑃 [𝜈+1]
𝑗 𝑥 (𝜈+1)
; 𝑥 𝑛− 𝑗+1 = 0, 𝑗 = 1, ..., 𝜈 + 1. (14.5.9)

The radii 𝑟 𝑖 , 𝑖 = 1, ..., 𝑛 − 𝜈 − 1, are adjusted so that Condition (14.5.5) is satisfied

with 𝜈 + 1 in the place of 𝜈. We define
n o
𝒁 [𝜈+1] = 𝑥 (𝜈+1) ∈ 𝔔𝑟(𝑛−𝜈−1)
(𝜈+1) ; 𝐷 [𝜈+1] 𝑥 (𝜈+1) = 0 ,

where 𝐷 [𝜈+1] 𝑥 (𝜈+1) = 𝐷 1[𝜈+1] 𝑥 (𝜈+1) · · · 𝐷 𝜈+1 [𝜈+1]
𝑥 (𝜈+1) with 𝐷 [𝜈+1]
𝑗 the dis-

[𝜈+1] [𝜈 ] [𝜈+1]
criminant of 𝑃 𝑗 . If 𝜋 [𝜈 ] 𝑽 ⊄𝒁 this stage of the process is completed.

If 𝜋 [𝜈 ] 𝑽 [𝜈 ] ⊂ 𝒁 [𝜈+1] we apply to (14.5.9) the procedure just applied to the equa-

tions (14.5.1) under the hypothesis 𝜋 [𝜈 ] 𝑽 [𝜈 ] ⊂ 𝒁 [𝜈 ] . We repeat this procedure
until we obtain a system of equations

[𝜈+𝜇]
𝑃𝑗 𝑥 (𝜈+𝜇) ; 𝑥 𝑛− 𝑗+1 = 0, 𝑗 = 1, ..., 𝜈 + 𝜇, (14.5.10)

still defining 𝑽 [𝜈 ] in 𝔔𝑟𝑛 but now with the property that

498 14 Analytic Stratifications
n o
(𝑛−𝜈−𝜇)
𝜋 𝜈+𝜇 𝑽 [𝜈 ] ⊄ 𝒁 [𝜈+𝜇] = 𝑥 (𝜈+𝜇) ∈ 𝔔𝑟 (𝜈+𝜇) ; 𝐷 [𝜇] 𝑥 (𝜈+𝜇) = 0 .

The completion of the procedure results in the following situation. We have deter-
mined a sequence of integers 𝜈 𝑘 , 1 ≤ 𝜈1 < · · · < 𝜈 𝑘 < 𝜈 𝑘+1 < · · · ≤ 𝜈 𝑘∗ ≤ 𝑛,
such that the equations (14.5.1), in which 𝜈 = 𝜈 𝑘 , define 𝑽 [𝜈𝑘 ] in 𝔔𝑟𝑛 and
𝜋 𝜈𝑘 𝑽 [𝜈𝑘 ] ⊄ 𝒁 [𝜈𝑘 ] . We have a strictly decreasing sequence of analytic subvarieties
of 𝔔𝑟𝑛 ,
𝑽 [𝜈1 ] ⊋ · · · ⊋ 𝑽 [𝜈𝑘 ] ⊋ 𝑽 [𝜈𝑘+1 ] ⊋ · · · ⊋ 𝑽 [ 𝜈𝑘∗ ] ,
with 𝑽 [ 𝜈𝑘∗ ] a C 𝜔 submanifold of 𝔔 𝑛 and dim 𝑽 [𝜈𝑘 ] = 𝑛 − 𝜈 𝑘 . Another way of stating
𝑟
e [𝜈𝑘 ] ∩ R𝑛 ; 𝜈1 is the largest integer 𝜈 ≥ 1
this is based on the fact that 𝑽 [𝜈𝑘 ] ∩ 𝔔𝑟𝑛 = 𝑽
such that 𝑽e [𝜈 ] ∩ R𝑛 = 𝑽
e [1] ∩ R𝑛 ; more generally, 𝜈 𝑘+1 is the largest integer 𝜈 > 𝜈 𝑘
such that 𝑽e [𝜈−1] ∩ R𝑛 = 𝑽e [𝜈𝑘 ] ∩ R𝑛 .
The partition (14.5.8) now yields
[ 𝜈𝑘 ]
Ø 𝑁Ø
𝜄

𝑽 [𝜈𝑘 ] \𝑽 [𝜈𝑘+1 ] = Λ [𝜈 𝑘]
𝜄, 𝑝 (14.5.11)
𝜄 ∈𝑰 [ 𝜈𝑘 ] 𝑝=1

with the understanding that 𝑽 [ 𝜈𝑘∗ +1 ] = ∅; (14.5.11) yields the partitions

[ 𝜈ℓ ]
𝑘∗ Ø
Ø 𝑁Ø
𝜄

𝑽 [𝜈𝑘 ] = Λ [𝜈 ℓ]
𝜄, 𝑝 (14.5.12)
ℓ=𝑘 𝜄 ∈𝑰 [ 𝜈ℓ ] 𝑝=1

and
[ 𝜈𝑘 ]
𝑘∗ Ø
Ø 𝑁Ø
𝜄

𝑽 ∩ 𝔔𝑟𝑛 =𝑽 [𝜈1 ]
= Λ [𝜈 𝑘]
𝜄, 𝑝 . (14.5.13)
𝑘=1 𝜄 ∈𝑰 [ 𝜈𝑘 ] 𝑝=1

For 1 ≤ 𝑘 ≤ 𝑘 ∗ , 𝜄 ∈ 𝑰 [𝜈𝑘 ] , 1 ≤ 𝑝 ≤ 𝑁 𝜄[𝜈𝑘 ] , the submanifold Λ [𝜈 𝑘]

𝜄, 𝑝 is a connected
[𝜈𝑘 ] [𝜈𝑘+1 ]
component of 𝑽 \𝑽 ; it is the graph of a C map𝜔

Γ 𝜄[𝜈𝑘 ] ∋ 𝑥 (𝜈𝑘 ) ↦→ 𝜌 𝜈[𝜈𝑘 𝑘,ℓ]𝜈 𝑥 (𝜈𝑘 ) , ..., 𝜌1,ℓ
[𝜈𝑘 ]
1
𝑥 (𝜈𝑘 )
∈ R 𝜈𝑘 , (14.5.14)
𝑘

where Γ 𝜄[𝜈𝑘 ] is a connected component of 𝔔𝑟𝑛−𝜈 (𝜈) \𝒁
[𝜈𝑘 ]
and 𝜌 [𝜈 𝑘]
𝑗,ℓ 𝑗 𝑥 (𝜈𝑘 ) is a simple

real root of 𝑃 [𝜈 𝑘] (𝜈𝑘 ) ; 𝑧 𝑥 (𝜈𝑘 ) = 𝑥 1 , ..., 𝑥 𝑛−𝜈𝑘

𝑗 𝑥 𝑛− 𝑗+1 ( 𝑗 = 1, ..., 𝜈 𝑘 ). Recall that
and that (14.5.5) [or (14.4.12)] holds for 𝜈 = 𝜈 𝑘 , 1 ≤ 𝑘 ≤ 𝑘 ∗ .
The next subsection will be entirely devoted to the proof of the following state-
ment:

Proposition 14.5.8 The partition (14.5.13) is a local stratification of 𝑽.

14.5 Local Stratifications of a Real-Analytic Variety 499

14.5.2 Proof that the partition (14.5.13) is a stratification

In view of Proposition 14.5.7 it remains to prove that the partition (14.5.13) satisfies
Condition (Strat) in Definition 14.1.4. We note that, for every 𝑘 = 1, ..., 𝑘 ∗ and every
𝜄 ∈ 𝑰 [𝜈𝑘 ] , 𝑝 = 1, ..., 𝑁 𝜄[𝜈𝑘 ] ,

𝜕Λ [𝜈 𝑘] [𝜈𝑘 ] [𝜈𝑘 ]
𝜄, 𝑝 = Λ 𝜄, 𝑝 \Λ 𝜄, 𝑝 ⊂ 𝑽
[𝜈𝑘+1 ]
. (14.5.15)

Proposition 14.5.9 If Λ 𝜅[𝜈,𝑞ℓ ] ≠ Λ [𝜈 𝑘] [𝜈ℓ ] [𝜈𝑘 ]

𝜄, 𝑝 and Λ 𝜅 ,𝑞 ∩ Λ 𝜄, 𝑝 ≠ ∅ (𝜄 ∈ 𝑰
[𝜈𝑘 ]
, 𝜅 ∈ 𝑰 [𝜈ℓ ] ,
[𝜈𝑘 ] [𝜈ℓ ]
1 ≤ 𝑝 ≤ 𝑁 𝜄 , 1 ≤ 𝑞 ≤ 𝑁 𝜅 ) then 𝑘 < ℓ.
Proof If Λ 𝜅[𝜈,𝑞𝑘 ] ≠ Λ [𝜈 𝑘] [𝜈𝑘 ] [𝜈𝑘 ] [𝜈𝑘 ] [𝜈𝑘 ]
𝜄, 𝑝 then Λ 𝜄, 𝑝 ∩Λ 𝜅 ,𝑞 = ∅ and Λ 𝜄, 𝑝 ∩ 𝜕Λ 𝜄, 𝑝 = ∅ by (14.5.15).

If 𝑘 > ℓ then Λ [𝜈 𝑘]
𝜄, 𝑝 ⊂ 𝑽
[𝜈ℓ+1 ]
and therefore Λ 𝜅[𝜈,𝑞ℓ ] ∩ Λ [𝜈 𝑘]
𝜄, 𝑝 = ∅. □
The next proposition completes the proof that (14.5.13) is a stratification (Defi-
nition 14.1.4); its proof will rely on the following lemma.
Lemma 14.5.10 Let X, Y be two topological Hausdorff spaces, 𝜒 : X −→ Y a
continuous map whose graph 𝑮 = {(𝑥, 𝑦) ∈ X × Y; 𝑦 = 𝜒 (𝑥)} is closed. Let Λ be
a subset of X × Y with the following properties:
(1) the restriction of the projection 𝜋 : (𝑥, 𝑦) ↦→ 𝑥 to Λ is a homeomorphism of Λ
onto 𝑀 = 𝜋 (Λ);
(2) there is an open subset 𝑈 of X × Y such that
−1
∅ ≠ 𝑮 ∩ 𝑈 ∩ 𝜋 (𝑀) ⊂ Λ.

Under these hypotheses Λ ∩ 𝑮 is open and closed in Λ.

Proof Let (𝑥 ◦ , 𝜒 (𝑥 ◦ )) ∈ Λ ∩ 𝑮 be arbitrary and 𝑈 be as in Hypothesis (2). There
is a neighborhood 𝜔◦ of 𝑥 ◦ in X such that (𝑥, 𝜒 (𝑥)) ∈ 𝑮 ∩ 𝑈 for all 𝑥 ∈ 𝜔◦ ; this
implies
−1 −1
𝜋 (𝜔◦ ∩ 𝑀) ⊂ 𝑮 ∩ 𝑈 ∩ 𝜋 (𝑀) .
−1
By Hypothesis (1), Λ ∩ 𝜋 (𝜔◦ ∩ 𝑀) is a neighborhood of 𝑥 ◦ in Λ. By Hypothesis
−1
(2) we have 𝜋 (𝜔◦ ∩ 𝑀) ⊂ Λ. This proves that Λ ∩ 𝑮 is open in Λ; since Λ ∩ 𝑮 is
obviously closed in Λ the claim ensues. □
Proposition 14.5.11 If 1 ≤ 𝑘 < ℓ ≤ 𝑘 ∗ and 𝜄 ∈ 𝑰 [𝜈𝑘 ] , 𝜅 ∈ 𝑰 [𝜈ℓ ] , 1 ≤ 𝑝 ≤ 𝑁 𝜄[𝜈𝑘 ] ,
1 ≤ 𝑞 ≤ 𝑁 𝜅[𝜈ℓ ] , then

Λ 𝜅[𝜈,𝑞ℓ ] ∩ Λ [𝜈 𝑘] [𝜈ℓ ] [𝜈𝑘 ]

𝜄, 𝑝 ≠ ∅ =⇒ Λ 𝜅 ,𝑞 ⊂ Λ 𝜄, 𝑝 .

Proof Let Λ [𝜈 𝑘]
𝜄, 𝑝 be the graph of the map (14.5.14); our basic hypothesis is that

Λ [𝜈 𝑘] [𝜈ℓ ]
𝜄, 𝑝 ∩ Λ 𝜅 ,𝑞 ≠ ∅. We shall use induction on ℓ − 𝑘; in particular, this allows us to
reason under the following hypothesis:
500 14 Analytic Stratifications

(•) There is no integer 𝑘 ′, 𝑘 < 𝑘 ′ ≤ ℓ, such that, for some 𝜆 ∈ 𝑰 [𝜈𝑘′ ] and some
[𝜈 ′ ]
positive integer 𝑟 ≤ 𝑁𝜆 𝑘 ,

[𝜈 ′ ] [𝜈 ′ ]
Λ𝜆,𝑟𝑘 ∩ Λ [𝜈 𝑘] [𝜈ℓ ]
𝜄, 𝑝 ≠ ∅, Λ 𝜅 ,𝑞 ∩ Λ𝜆,𝑟 ≠ ∅.
𝑘
(14.5.16)

Indeed, if there were 𝑘 ′ as in (14.5.16) we would replace ℓ, 𝜅, 𝑞 by 𝑘 ′, 𝜆, 𝑟, and

[𝜈 ′ ]
conclude, by the induction hypothesis, that Λ𝜆,𝑟𝑘 ⊂ Λ [𝜈 𝑘]
𝜄, 𝑝 and then repeat the
argument, now replacing 𝜈 𝑘 , 𝜄, 𝑝 by 𝜈 𝑘′ , 𝜆, 𝑟.
Keep in mind that Λ 𝜅[𝜈,𝑞ℓ ] ⊂ 𝑽 [𝜈ℓ ]\𝑽 [𝜈ℓ+1 ] ; throughout the remainder of the proof
′
we use the notation 𝑀 = 𝜋 𝜈𝑘 Λ 𝜅[𝜈,𝑞ℓ ] . For arbitrary 𝑘 ′ ≥ 𝑘 let 𝑱 [𝑘 ] denote the subset
[𝜈 ′ ]
of 𝑰 [𝜈𝑘′ ] consisting of the indices 𝜆 with the property that there is an 𝑟 ≤ 𝑁𝜆 𝑘 such
that
[𝜈 ′ ] −1 [𝜈 ′ ]
Λ𝜆,𝑟𝑘 ∩ 𝜋 𝜈𝑘 (𝑀) ≠ ∅ and Λ𝜆,𝑟𝑘 ∩ Λ [𝜈 𝑘]
𝜄, 𝑝 ≠ ∅. (14.5.17)

The basic hypothesis implies 𝜄 ∈ 𝑱 [𝑘 ] and 𝜅 ∈ 𝑱 [ℓ ] . According to (•), if 𝑘 < 𝑘 ′ < ℓ

′ [𝜈 ′ ]
then 𝑱 [𝑘 ] = ∅. If 𝑘 = 𝑘 ′ then, by Proposition 14.5.9, Λ𝜆,𝑟𝑘 ∩ Λ [𝜈 𝑘]
𝜄, 𝑝 ≠ ∅ implies
[𝜈𝑘′ ] [𝜈 ′ ]
𝜄′ = 𝜄 and 𝑝 ′ = 𝑝. If 𝑘 ′ > ℓ and 𝜆 ∈ 𝑰 [𝜈𝑘′ ] , 1 ≤ 𝑟 ≤ 𝑁𝜆 , we have Λ𝜆,𝑟𝑘 ⊂ 𝑽 [𝜈ℓ+1 ]
[𝜈 ′ ] −1 ′
and therefore Λ𝜆,𝑟𝑘 ∩ 𝜋 𝜈𝑘 (𝑀) = ∅ contradicting (14.5.17) and implying 𝑱 [𝑘 ] = ∅.
To summarize, aside from the trivial case 𝑘 = 𝑘 ′, 𝜆 = 𝜄, 𝑟 = 𝑝, (14.5.17) entails
𝑘 ′ = ℓ. We derive that there is an open subset 𝑈 of 𝔔𝑟𝑛 such that

−1
∅ ≠ Λ [𝜈 𝑘] [𝜈ℓ ]
𝜄, 𝑝 ∩ 𝜋 𝜈𝑘 (𝑀) ∩ 𝑈 ⊂ Λ 𝜅 ,𝑞 ⊂ 𝑈.

If we prove that 𝑀 ⊂ Γ 𝜄[𝜈𝑘 ] then the proof of Proposition 14.5.11 is completed

by applying Lemma 14.5.10 with 𝑋 = Γ 𝜄[𝜈𝑘 ] , 𝜒 the continuous extension to 𝑋 of
the map (14.5.14) and Λ = Λ 𝜅[𝜈,𝑞ℓ ] (here 𝑮 = Λ [𝜈 𝑘]
𝜄, 𝑝 ; cf. Proposition 14.5.6). Indeed,

since Λ 𝜅[𝜈,𝑞ℓ ] is connected we have Λ 𝜅[𝜈,𝑞ℓ ] ∩ Λ [𝜈 𝑘] [𝜈ℓ ]

𝜄, 𝑝 = Λ 𝜅 ,𝑞 .
This argument settles directly the case Γ 𝜄[𝜈𝑘 ] = 𝔔𝑟𝑛−𝜈 (𝜈) \𝒁
[𝜈𝑘 ]
since, in this case,
[𝜈𝑘 ]
(𝜈) ⊂ Γ 𝜄
𝔔𝑟𝑛−𝜈 .
We look now at the case in which Γ 𝜄[𝜈𝑘 ] is not the sole connected component of
(𝜈) \𝒁
𝔔𝑟𝑛−𝜈 [𝜈𝑘 ]
. This demands that the regular part of 𝜕Γ 𝜄[𝜈𝑘 ] of maximum dimension
be a hypersurface in 𝔔𝑟𝑛−𝜈 (𝜈) , i.e., have dimension 𝑛 − 𝜈 𝑘 − 1. It follows that the

maximum dimension of the connected components of ℜ 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 must be 𝑛−𝜈 𝑘 −1.

Since 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 ⊂ 𝑽
[𝜈𝑘+1 ]
and dim 𝑽 [𝜈𝑘+1 ] = 𝑛 − 𝜈 𝑘+1 < dim 𝑽 [𝜈𝑘 ] we conclude that
𝜈 𝑘+1 = 𝜈 𝑘 + 1. Since
[𝜈 ]
𝑁𝜆 𝑘+1
Ø Ø
[𝜈𝑘+1 ]
𝑽 [𝜈𝑘+1 ] \𝑽 [𝜈𝑘+2 ] = Λ𝜆,𝑟
𝜆∈𝑰 [ 𝜈𝑘+1 ] 𝑟=1
14.5 Local Stratifications of a Real-Analytic Variety 501

is a dense subset of ℜ𝑛−𝜈𝑘+1 𝑽 [𝜈𝑘+1 ] we deduce

[𝜈 ]
𝑁𝜆 𝑘+1
Ø Ø
ℜ𝑛−𝜈𝑘+1 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 ⊂ [𝜈𝑘+1 ]
Λ𝜆,𝑟 . (14.5.18)
𝜆∈𝑰 [ 𝜈𝑘+1 ] 𝑟=1

Case ℓ = 𝑘 + 1 (hence 𝜈 𝑘 = 𝜈ℓ − 1 by the preceding argument). As pointed

out above, the regular part of 𝜕Γ 𝜄[𝜈𝑘 ] contains hypersurfaces in 𝔔𝑟𝑛−𝜈 (𝜈) defined by
equations
𝑥 𝑛−𝜈ℓ +1 = 𝑥 𝑛−𝜈𝑘 = 𝜌 𝑥 (𝜈ℓ ) ,

where the 𝜌 are simple real roots of the equation 𝑃 𝜈[𝜈ℓ ℓ ] 𝑥 (𝜈ℓ ) ; 𝑥 𝑛−𝜈ℓ +1 = 0 with

𝑥 (𝜈ℓ ) varying in a connected component of 𝔔𝑟𝑛−𝜈 (𝜈) \𝒁
[𝜈ℓ ]
. But 𝑀 = 𝜋 𝜈ℓ −1 Λ 𝜅[𝜈,𝑞ℓ ] is
precisely one of these (pairwise disjoint) hypersurfaces; if 𝑀 intersects the closure
of one of them it must be identical to it, which proves the sought result in this case.
Case 2 ≤ 𝜈ℓ − 𝜈 𝑘 ≤ 𝑛 − 1 (and 𝑛 ≥ 3). Let 𝑬 be the subset of pairs of indices
(𝜆, 𝑟), 𝜆 ⊂ 𝑰 [𝜈𝑘+1 ] , 1 ≤ 𝑟 ≤ 𝑁𝜆[𝜈𝑘+1 ] , such that Λ𝜆,𝑟
[𝜈𝑘+1 ]
∩ Λ [𝜈 𝑘]
𝜄, 𝑝 ≠ ∅. Keep in mind
[𝜈𝑘+1 ]
that 𝜈 𝑘+1 = 𝜈 𝑘 + 1. If (𝜆, 𝑟) ∈ 𝑬 then Λ𝜆,𝑟 ⊂ 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 by the result when ℓ = 𝑘 + 1;
we obtain
Ø
[𝜈𝑘+1 ]
Λ𝜆,𝑟 ⊂ 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 , (14.5.19)
(𝜆,𝑟) ∈𝑬
Ø
Λ [𝜈 𝑘]
𝜄, 𝑝 ∩
[𝜈𝑘+1 ]
Λ𝜆,𝑟 = ∅. (14.5.20)
(𝜆,𝑟)∉𝑬

As before let 𝑀 = 𝜋 𝜈𝑘 Λ 𝜅[𝜈,𝑞ℓ ] ; suppose we have proved

−1
Ø
[𝜈𝑘+1 ]
𝜋 𝜈𝑘 (𝑀) ∩ Λ𝜆,𝑟 ≠ ∅, (14.5.21)
(𝜆,𝑟) ∈𝑬

which, incidentally, implies 𝑬 ≠ ∅. Then at least one of the submanifolds Λ 𝜅[𝜈,𝑞ℓ ]′ ⊂

−1
𝜋 𝜈𝑘 (𝑀) [1 ≤ 𝑞 ′ ≤ 𝑁 𝜅[𝜈ℓ ] ; cf. (14.5.17)] must intersect Λ𝜆,𝑟
[𝜈𝑘+1 ]
for some (𝜆, 𝑟) ∈ 𝑬.
By the induction hypothesis on ℓ − 𝑘 this implies Λ 𝜅[𝜈,𝑞ℓ ]′ ⊂ 𝜕Λ𝜆,𝑟
[𝜈𝑘+1 ]
and then (14.5.19)

implies Λ 𝜅 ,𝑞′ ⊂ 𝜕Λ 𝜄, 𝑝 . In turn the latter implies 𝑀 ⊂ 𝜋 𝜈𝑘 𝜕Λ [𝜈
[𝜈ℓ ] [𝜈𝑘 ] 𝑘]
𝜄, 𝑝 , thereby
completing the proof of Proposition 14.5.11.
Proof of (14.5.21). If (14.5.21) were not true there would be an open subset 𝑈 ′
of R𝑛−𝜈𝑘 such that
Ø
[𝜈𝑘+1 ]
𝑀 ⊂ 𝑈 ′ and 𝑈 ′ ∩ 𝜋 𝜈𝑘 Λ𝜆,𝑟 = ∅.
(𝜆,𝑟) ∈𝑬
502 14 Analytic Stratifications

This implies
−1
Ø
[𝜈𝑘+1 ]
𝜋 𝜈𝑘 (𝑈 ′) ∩ Λ𝜆,𝑟 = ∅. (14.5.22)
(𝜆,𝑟) ∈𝑬

We would deduce from (14.5.18) and (14.5.22):

−1
Ø
𝜋 𝜈𝑘 (𝑈 ′) ∩ ℜ𝑛−𝜈𝑘+1 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 ⊂ [𝜈𝑘+1 ]
Λ𝜆,𝑟 .
(𝜆,𝑟)∉𝑬

Since (14.5.20) implies

Ø
𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 ∩
[𝜈𝑘+1 ]
Λ𝜆,𝑟 =∅
(𝜆,𝑟)∉𝑬

we would conclude that

−1

𝜋 𝜈𝑘 (𝑈 ′) ∩ ℜ𝑛−𝜈𝑘+1 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 = ∅. (14.5.23)

Let then Σ be a C 𝜔 hypersurface in R𝑛−𝜈𝑘 contained in 𝑈 ′ ∩ ℜ𝑛−𝜈𝑘+1 𝜕Γ 𝜄[𝜈𝑘 ] . As
𝑥 (𝜈𝑘 ) varies in Σ the point

[𝜈𝑘 ] [𝜈𝑘 ]
𝑥 (𝜈𝑘 ) , 𝜌1,ℓ1
𝑥 (𝜈𝑘 )
, ..., 𝜌 𝜈𝑘 ,ℓ𝜈 𝑥 (𝜈𝑘 )
𝑘

−1
describes a C 𝜔 submanifold Σ ♮ of R𝑛 contained in 𝜋 𝜈𝑘 (𝑈 ′) ∩ 𝜕Λ [𝜈 𝑘]
𝜄, 𝑝 such that
dim Σ ♮ = 𝑛 − 𝜈 𝑘+1 . This contradicts (14.5.23), thus proving (14.5.21). □

14.5.3 Main theorem and direct consequences

Proposition 14.5.8 directly implies the main result of this section:

Theorem 14.5.12 Every real-analytic subvariety of a real-analytic manifold is strat-

ifiable (Definition 14.1.5).

Combining Theorems 14.1.11 and 14.5.12 yields the next statements, where
“analytic” can be taken in the real as well as the complex sense, by Proposition
14.2.9. We continue to assume that the analytic manifold M is countable at infinity.

Corollary 14.5.13 Let 𝑽 be an analytic subvariety of M. The natural partition of

the regular part ℜ (𝑽) of 𝑽, meaning the partition of ℜ (𝑽) into its connected
components, is an analytic stratification of ℜ (𝑽) (Definition 14.1.4). The family of
connected components of ℜ (𝑽) is locally finite and therefore countable.

Corollary 14.5.14 Let 𝑽 be an analytic subvariety of M, 𝑽 ≠ M. The family of

connected components of M\𝑽 is locally finite and therefore countable.
14.5 Local Stratifications of a Real-Analytic Variety 503

Proof There is no loss of generality in assuming that M is connected. When K = C

Proposition 14.3.22 implies that M\𝑽 is connected. Suppose K = R and that there
is a connected component N of M\𝑽, N ≠ M\𝑽. The boundary 𝜕N of the domain
N is a continuous hypersurface; 𝜕N ⊂ 𝑽 and therefore dimR 𝑽 = dimR M − 1. The
claim follows from Corollary 14.5.13. □
Corollary 14.5.15 The critical values of an analytic function 𝑓 in M are locally
finite and therefore countable.
Proof Let 𝑽 = {℘ ∈ M; d 𝑓 (℘) = 0}; 𝑓 is constant on the closure of each connected
component of ℜ (𝑽). Therefore the claim is a direct consequence of Corollary
14.5.13. □
Corollary 14.5.16 If Ω is a domain in M whose closure is compact and 𝑐 ◦ ∈ K
then there is a number 𝛿 > 0 such that, for every 𝑐 ∈ K, 0 < |𝑐 − 𝑐 ◦ | < 𝛿, the sets
{𝑧 ∈ Ω; 𝑓 (𝑧) = 𝑐} are analytic submanifolds of Ω.
Proof The claim follows from Corollary 14.5.15 if Ω is sufficiently small and thereby
in the general case, by the Borel–Lebesgue Lemma. □

14.5.4 The Whitney property

Next we prove that (14.5.13) is a Whitney stratification of 𝑽 (Definition 14.1.7).

Proposition 14.5.17 Let Λ [𝜈 1] [𝜈2 ]

𝜄, 𝑝 and Λ 𝜅 ,𝑞 (0 ≤ 𝜈1 < 𝜈2 < 𝑛, 𝜄 ∈ 𝑰
[𝜈1 ]
, 𝜅 ∈ 𝑰 [𝜈2 ] ,
1 ≤ 𝑝 ≤ 𝑁 𝜄[𝜈1 ] , 1 ≤ 𝑞 ≤ 𝑁 𝜅[𝜈2 ] ) be strata in (14.5.13) such that Λ 𝜅[𝜈,𝑞2 ] ⊂ Λ [𝜈 1]
𝜄, 𝑝 . Every
[𝜈2 ] [𝜈 ]
vector tangent to Λ 𝜅 ,𝑞 at a point 𝑥 ◦ is the limit of vectors tangent to Λ 𝜄, 𝑝1 at points
converging to 𝑥 ◦ .

In other words, Λ 𝜅[𝜈,𝑞2 ] ⊂ Λ [𝜈 1] [𝜈2 ] [𝜈1 ]

𝜄, 𝑝 implies 𝑇Λ 𝜅 ,𝑞 ⊂ 𝑇Λ 𝜄, 𝑝 , the closure of the tangent
bundle 𝑇Λ [𝜈 1]
𝜄, 𝑝 in 𝑇Ω.

Proof Let 𝜋 [𝜈1 ] : 𝑥 ↦→ 𝑥 (𝜈1 ) = 𝑥1 , ..., 𝑥 𝑛−𝜈1 ∈ 𝔔𝑟𝑛 denote the coordinate projection.

By Proposition 14.5.6 there is an open subset 𝑈1 of 𝔔𝑟𝑛 containing Λ [𝜈 1]

𝜄, 𝑝 such that
[𝜈1 ] [𝜈1 ]
Γ𝜄 = 𝜋 [𝜈1 ] (𝑈1 ) is a connected component of 𝔔𝑟𝑛−𝜈 (𝜈) \𝒁 [cf. (14.4.18)] and
Λ [𝜈 1]
𝜄, 𝑝 is defined in 𝑈1 by equations

𝑥 𝑛− 𝑗+1 = 𝜌 [𝜈
𝑗
1]
𝑥 (𝜈1 )
, 𝑗 = 1, ..., 𝜈1 , (14.5.24)

where 𝜌 [𝜈 1] (𝜈1 ) ∈ C 𝜔 Γ [𝜈1 ] is a root of 𝑃 [𝜈1 ] 𝑥 (𝜈1 ) ; 𝑥

𝑥 𝑛− 𝑗+1 = 0 (1 ≤ 𝑗 ≤ 𝜈1 );
𝑗 𝑗

𝜌 [𝜈
𝑗
1]
𝑥 (𝜈1 ) can be extended continuously to the closure of Γ 𝜄[𝜈1 ] in 𝔔𝑟𝑛 (Lemma
14.5.5). We derive from (14.5.24) that the tangent spaces 𝑇𝑥 Λ [𝜈 1]
𝜄. 𝑝 are spanned by the
vector fields
504 14 Analytic Stratifications

𝜕
𝜈1
∑︁ 𝜕 𝜌ℓ[𝜈1 ] (𝜈 ) 𝜕
𝑋𝑘 = + 𝑥 1 , 𝑘 = 1, ..., 𝑛 − 𝜈1 . (14.5.25)
𝜕𝑥 𝑘 ℓ=1 𝜕𝑥 𝑘 𝜕𝑥 𝑛−ℓ+1

Likewise, there is an open subset of 𝔔𝑟𝑛 , 𝑈2 ⊃ Λ 𝜅[𝜈,𝑞2 ] , such that Λ 𝜅[𝜈,𝑞2 ] is defined in
𝑈2 by equations
𝑥 𝑛− 𝑗+1 = 𝜌 [𝜈
𝑗
2]
𝑥 (𝜈2 ) , 𝑗 = 1, ..., 𝜈2 , (14.5.26)

where 𝑥 (𝜈2 ) varies in a connected component Γ𝜅(𝜈2 ) of 𝔔𝑟𝑛 \𝒁 [𝜈2 ] and 𝜌 [𝜈 𝑗
2]
𝑥 (𝜈2 ) is

[𝜈2 ] (𝜈 ) (𝜈 ) (𝜈2 )
a root of 𝑃 𝑗 𝑥 ; 𝑥 𝑛− 𝑗+1 = 0. If 𝑥
2 2 ∈ Γ𝜅 the set of points

𝑥 (𝜈2 ) , 𝜌 𝜈[𝜈2 2 ] 𝑥 (𝜈2 ) , ..., 𝜌 𝜈[𝜈1 +1
2]
𝑥 (𝜈2 ) ∈ 𝔔𝑟(𝑛−𝜈1 )

is contained in the boundary of Γ 𝜄[𝜈1 ] , Γ 𝜄[𝜈1 ] \Γ 𝜄[𝜈1 ] ; we have

𝜌 [𝜈
𝑗
2]
𝑥 (𝜈2 ) = 𝜌 [𝜈
𝑗
1]
𝑥 (𝜈2 ) , 𝜌 𝜈[𝜈2 2 ] 𝑥 (𝜈2 ) , ..., 𝜌 𝜈[𝜈1 +1
2]
𝑥 (𝜈2 ) (14.5.27)

for every 𝑗 = 1, ..., 𝜈1 , 𝑥 (𝜈2 ) ∈ Γ𝜅(𝜈2 ) . We derive from (14.5.26) that the tangent
spaces 𝑇𝑥 Λ 𝜅[𝜈,𝑞2 ] are spanned by the vector fields

𝜕
𝜈2
∑︁ 𝜕 𝜌ℓ[𝜈2 ] (𝜈 ) 𝜕
𝑌𝑘 = + 𝑥 2 , 𝑘 = 1, ..., 𝑛 − 𝜈2 . (14.5.28)
𝜕𝑥 𝑘 ℓ=1 𝜕𝑥 𝑘 𝜕𝑥 𝑛−ℓ+1

There is an open subset 𝑈2′ of 𝔔𝑟𝑛 , Λ 𝜅[𝜈,𝑞2 ] ⊂ 𝑈2′ ⊂ 𝑈2 , in which none of the par-
[𝜈 ]
𝜕𝑃 2

tial derivatives 𝜕𝑥𝑛−𝑗 𝑗+1 𝑥 (𝜈2 ) ; 𝑥 𝑛− 𝑗+1 , 𝑗 = 1, ..., 𝜈2 , vanishes. Differentiating each

equation 𝑃 (𝜈
𝑗
2)
𝑥 (𝜈2 ) ; 𝜌 [𝜈
𝑗
2]
𝑥 (𝜈2 ) = 0 with respect to 𝑥 𝑘 (𝑘 = 1, ..., 𝑛 − 𝜈2 ) yields

𝜕 𝜌 [𝜈2]
𝑥 (𝜈2 )
𝑗
−
𝜕𝑥 𝑘
𝜕𝑃 [𝜈2] 𝜕𝑃 [𝜈2]
,

𝑥 (𝜈2 ) ; 𝜌 [𝜈2]
𝑥 (𝜈2 ) 𝑥 (𝜈2 ) ; 𝜌 [𝜈2]
𝑥 (𝜈2 ) .
𝑗 𝑗
= 𝑗 𝑗
𝜕𝑥 𝑘 𝜕𝑥 𝑛− 𝑗+1

The right-hand side provides a C 𝜔 extension of the left-hand side to 𝜋 [𝜈1 ] 𝑈2′ ,

namely

𝜕𝑃 [𝜈2]
𝜕𝑃 [𝜈2]
,

(𝜈1 ) (𝜈2 )
𝑥 (𝜈2 ) ; 𝑥 𝑛− 𝑗+1 .
𝑗 𝑗
−𝑅 𝑗,𝑘 𝑥 = 𝑥 ; 𝑥 𝑛− 𝑗+1
𝜕𝑥 𝑘 𝜕𝑥 𝑛− 𝑗+1
14.5 Local Stratifications of a Real-Analytic Variety 505

Let 𝑥 ◦ ∈ Λ 𝜅[𝜈,𝑞2 ] ; we can select a suitably small open subset 𝑈2′′ of 𝑈2′ , 𝑈2′′ ∋ 𝑥 ◦ ,
a unit vector 𝑦 (𝜈1 ) ∈ R𝑛−𝜈1 and a number 𝜀 > 0 such that 𝜋 [𝜈1 ] (𝑥) + 𝑡𝑦 (𝜈1 ) ∈ Γ [𝜈1 ]
for all 𝑥 ∈ Λ 𝜅[𝜈,𝑞2 ] ∩ 𝑈2′′ and 𝑡 ∈ (0, 𝜀). We introduce the notation 𝑧 (𝜈1 ) = 𝑧 (𝜈1 ) (𝑡) =
𝑥 (𝜈1 ) + 𝑡𝑦 (𝜈1 ) ∈ R𝑛−𝜈1 and [cf. (14.5.24)]

𝑧 = 𝑧 (𝜈1 ) , 𝜌 𝜈[𝜈1 1 ] 𝑧 (𝜈1 ) , ..., 𝜌1[𝜈1 ] 𝑧 (𝜈1 ) ∈ Λ (𝜈 1)
𝜄, 𝑝 . (14.5.29)

Proposition 14.5.17 will be proved if we prove that, for each 𝑘 = 1, ..., 𝑛 − 𝜈2 , 𝑌𝑘 is

the limit of the vector fields on Λ (𝜈 1) ′′
𝜄, 𝑝 ∩ 𝑈2 ,

𝜈2
∑︁
𝑋˜ 𝑘 𝑧
= 𝑋𝑘 | 𝑧 + 𝑅 𝑗,𝑘 𝑧 (𝜈1 ) 𝑋𝑛− 𝑗+1 𝑧 ,
𝑗=𝜈1 +1

as 𝑧 [defined in (14.5.29)] converges to an arbitrary point 𝑥 ∈ Λ (𝜈

𝜄2
1)
∩ 𝑈2′′. By
(𝜈1 ) ′′
(14.5.25) we obtain, at the point 𝑧 ∈ Λ 𝜄 ∩ 𝑈2 ,

𝜕
𝜈1
∑︁ 𝜕 𝜌ℓ[𝜈1 ] (𝜈 ) 𝜕
𝜈2
∑︁ 𝜕
𝑋˜ 𝑘 𝑧
= + 𝑧 1 + 𝑅 𝑗,𝑘 𝑧 (𝜈1 )
𝜕𝑥 𝑘 ℓ=1 𝜕𝑥 𝑘 𝜕𝑥 𝑛−ℓ+1 𝑗=𝜈 +1 𝜕𝑥 𝑛− 𝑗+1
1

𝜈2 ∑︁
∑︁ 𝜈1 𝜕 𝜌 [𝜈1 ] 𝜕
+ 𝑅 𝑗,𝑘 𝑧 (𝜈1 ) ℓ
𝑧 (𝜈1 ) .
𝑗=𝜈1 +1 ℓ=1
𝜕𝑥 𝑛− 𝑗+1 𝜕𝑥 𝑛−ℓ+1

As 𝑧 ↦→ 𝑥 ∈ Λ 𝜅[𝜈,𝑞2 ] ∩ 𝑈2′′ (i.e., as 𝑡 → 0) 𝑋˜ 𝑘 𝑧

converges to

𝜕
𝜈2
∑︁ 𝜕 𝜌 [𝜈2] 𝜕
𝜈1 𝜕 𝜌 [𝜈1 ]
∑︁ 𝜕
(𝜈2 )
𝑥 (𝜈1 )
𝑗 𝑗
𝑋˜ 𝑘 𝑥
= + 𝑥 +
𝜕𝑥 𝑘 𝑗=𝜈 +1 𝜕𝑥 𝑘 𝜕𝑥 𝑛− 𝑗+1 𝑗=1 𝜕𝑥 𝑘 𝜕𝑥 𝑛− 𝑗+1
1

𝜈1
∑︁ 𝜈2
∑︁ 𝜕 𝜌ℓ[𝜈2 ] 𝜕 𝜌 [𝜈1] 𝜕
𝑥 (𝜈2 ) 𝑥 (𝜈1 )
𝑗
+ .
𝑗=1 ℓ=𝜈1 +1
𝜕𝑥 𝑘 𝜕𝑥 𝑛−ℓ+1 𝜕𝑥 𝑛− 𝑗+1

Making use of (14.5.27) shows that if

𝑥 (𝜈1 ) = 𝑥 (𝜈2 ) , 𝜌 𝜈[𝜈2 2 ] 𝑥 (𝜈2 ) , ..., 𝜌 𝜈[𝜈1 +1
2]
𝑥 (𝜈2 )

then

𝜕 𝜌 [𝜈2] 𝜕 𝜌 [𝜈 1]
𝑥 (𝜈2 ) = 𝑥 (𝜈1 )
𝑗 𝑗
𝜕𝑥 𝑘 𝜕𝑥 𝑘
[𝜈 ]
𝜈2
∑︁ 𝜕 𝜌ℓ[𝜈2 ] (𝜈 ) 𝜕 𝜌 𝑗 1 (𝜈 )
+ 𝑥 2 𝑥 1
ℓ=𝜈 +1
𝜕𝑥 𝑘 𝜕𝑥 𝑛−ℓ+1
1
506 14 Analytic Stratifications

for every 𝑗 = 1, ..., 𝜈1 , thereby proving that 𝑋˜ 𝑘 𝑥

= 𝑌𝑘 given in (14.5.28). □
From the results of this section we reach the following conclusion; here M is a
C 𝜔 manifold countable at infinity.

Theorem 14.5.18 An analytic subvariety 𝑽 of M is Whitney stratifiable (Definition

14.1.8).

14.6 Semianalytic Sets

14.6.1 Semianalytic sets. Definition and basic properties

In this section M shall be a C 𝜔 manifold countable at infinity, dimR M = 𝑛 ≥ 1; all

scalar functions shall be real-valued.

Definition 14.6.1 An open subset U of M shall be called an analytic polyhedron

if U is connected and if there are an open set Ω containing its closure U and finitely
many functions ℎ 𝑗 ∈ C 𝜔 (Ω; R) ( 𝑗 = 1, ..., 𝜈) such that
n o
U = ℘ ∈ Ω; ℎ2𝑗 (℘) < 1, 𝑗 = 1, ..., 𝜈 .

Example 14.6.2 All open balls 𝔅𝑅 (𝑥 ◦ ) = 𝑥 ∈ R𝑛 ; |𝑥 − 𝑥 ◦ | 2 < 𝜌 2 and all multi-

intervals n o
𝔔𝑟𝑛 (𝑥 ◦ ) = 𝑥 ∈ R𝑛 ; 𝑥 𝑗 − 𝑥 ◦𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛

(𝜌 > 0, 𝑟 𝑗 > 0) are analytic polyhedra in R𝑛 .

Finite intersections of analytic polyhedra that are connected are analytic poly-
hedra. Using analytic local charts we see that every point of M has a basis of
neighborhoods consisting of analytic polyhedra.

Proposition 14.6.3 An arbitrary open covering {U 𝜄 } 𝜄 ∈ℑ of M admits a locally finite

refinement consisting of analytic polyhedra.

Proof Since M is countable at infinity the covering {U 𝜄 } 𝜄 ∈ℑ admits a locally finite

refinement {V𝜅 } 𝜅 ∈𝔎 such that, for every 𝜅 ∈ 𝔎, V𝜅 is compact and contained in some
U 𝜄 . Every ℘ ∈ V𝜅 has a neighborhood N℘ ⊂ U 𝜄 which is an analytic polyhedron.
By the Borel–Lebesgue Lemma there is a finite subset 𝑬 𝜅 of V𝜅 such that
Ø
V𝜅 ⊂ N℘ ⊂ U 𝜄 ,
℘∈𝑬 𝜅

whence the claim. □

14.6 Semianalytic Sets 507

Given an analytic polyhedron U of M we introduce 𝑞 pairs of finite (nonempty)

subsets of C 𝜔 (U; R), 𝔉 𝑗 = 𝑓 𝑗,𝑘 𝑘=1,..., 𝜇 𝑗 , 𝔊 𝑗 = 𝑔 𝑗,𝑘 𝑘=1,...,𝜈 𝑗 ( 𝑗 = 1, ..., 𝑞 <
+∞, 1 ≤ 𝜇 𝑗 , 𝜈 𝑗 < +∞). For each 𝑗 we introduce the following pairs of subsets of U:

𝑽 𝔉 𝑗 = ℘ ∈ U; 𝑓 𝑗,𝑘 (℘) = 0, 𝑘 = 1, ..., 𝜇 𝑗 , (14.6.1)

𝑸 𝔊 𝑗 = ℘ ∈ U; 𝑔 𝑗,𝑘 (℘) > 0, 𝑘 = 1, ..., 𝜈 𝑗 ; (14.6.2)

we then form
𝑞
Ø
𝑺U = 𝑽 𝔉𝑗 ∩ 𝑸 𝔊 𝑗 . (14.6.3)
𝑗=1
Í𝜈 𝑗 2
It is not precluded that either 𝑽 𝔉 𝑗 = U (when 𝑘=1 𝑓 𝑗,𝑘 ≡ 0 in U) or 𝑸 𝔊 𝑗 = U
[when 𝑔 𝑗,1 (℘) > 0, ..., 𝑔 𝑗,𝜈 𝑗 (℘) > 0 for all ℘ ∈ U], nor that 𝑺 U = ∅. If U ′ ⊂ U
is an analytic polyhedron then
𝑞
Ø
𝑺U ∩ U′ = 𝑽 𝔉 𝑗′ ∩ 𝑸 𝔊 ′𝑗 = 𝑺 U ′ ,
𝑗=1

where 𝑽 𝔉 𝑗′ = 𝑽 𝔉 𝑗 ∩ U ′, 𝑸 𝔊 ′𝑗 = 𝑸 𝔊 𝑗 ∩ U ′.

Proposition 14.6.4 In the decomposition

(14.6.3) the family of connected compo-
nents of the open set 𝑸 𝔊 𝑗 ⊂ U is locally finite for each 𝑗 = 1, ..., 𝑞.

Proof Using the notation 𝑽 𝑔 𝑗,𝑘 = ℘ ∈ U; 𝑔 𝑗,𝑘 (℘) = 0 and
𝜈𝑗
Ø
♭
𝑽 𝔊𝑗 = 𝑽 𝑔 𝑗,𝑘 = 𝑽 𝑔 𝑗,1 · · · 𝑔 𝑗,𝜈 𝑗 , (14.6.4)
𝑘=1

we see that the connected components of 𝑸 𝔊 𝑗 are those of the set U\𝑽 ♭ 𝔊 𝑗

in which 𝑔 𝑗,𝑘 > 0 for all 𝑘. Since 𝑽 ♭ 𝔊 𝑗 is stratifiable (Theorem 14.5.12) the

connected components of U\𝑽 ♭ 𝔊 𝑗 form a locally finite family, by Proposition
14.5.7. □
Definition 14.6.5 A subset 𝑺 of M is said to be a semianalytic set in M if there is
a covering of 𝑺 by analytic polyhedra U such that 𝑺 ∩ U is of the type (14.6.3).
By Proposition 14.6.3 the covering of 𝑺 by analytic polyhedra in Definition 14.6.5
can be taken to be locally finite. Analytic polyhedra are open semianalytic sets in
M; M is semianalytic in M.
If M ′ ⊂ M is open and if 𝑺 is a semianalytic set in M then 𝑺 ∩M ′ is semianalytic
in M ′. If a subset 𝑺 of M is such that there is a covering of M by open sets U such
that 𝑺 ∩ U is semianalytic in U then 𝑺 is semianalytic in M.
Proposition 14.6.6 Finite intersections and finite unions of semianalytic sets in M
are semianalytic sets in M. If 𝑺 is a semianalytic set in M so is its complement
M\𝑺.
508 14 Analytic Stratifications

Proof It suffices to prove the claims for sets of the type (14.6.3) assuming that
M = U. Then the claim about unions is self-evident. About intersections it follows
from the fact that, if 𝔉, 𝔊, 𝔉 ′, 𝔊 ′ are finite subsets of C 𝜔 (U; R), then

𝑽 (𝔉) ∩ 𝑸 (𝔊) ∩ 𝑽 (𝔉 ′) ∩ 𝑸 (𝔊 ′) = 𝑽 (𝔉 ∪ 𝔉 ′) ∩ 𝑸 (𝔊 ∪ 𝔊 ′) .

About complements it suffices to prove that U\𝑽 𝔉 𝑗 and U\𝑸 𝔊 𝑗 are sets of the
type (14.6.3) since
𝑞
Ù
U\𝑺 U = U\𝑽 𝔉 𝑗 ∪ U\𝑸 𝔊 𝑗 .
𝑗=1

It is convenient to use the notation, in which 𝑓 ∈ C 𝜔 (U; R),

𝑽 ( 𝑓 ) = {℘ ∈ U; 𝑓 (℘) = 0} ,
𝑸 ( 𝑓 ) = {℘ ∈ U; 𝑓 (℘) > 0} .

We have
𝜇𝑗
Ø

U\𝑽 𝔉 𝑗 = 𝑸 𝑓 𝑗,𝑘 ∪ 𝑸 − 𝑓 𝑗,𝑘 ,
𝑘=1
𝜈𝑗
Ø
U\𝑸 𝔊 𝑗 = 𝑽 𝑔 𝑗,𝑘 ∪ 𝑸 −𝑔 𝑗,𝑘 . □
𝑘=1

Corollary 14.6.7 Locally finite unions of semianalytic sets in M are semianalytic

sets in M.
This is obvious, since finite unions of sets of the type (14.6.3) are semianalytic
sets.
Corollary 14.6.8 Let 𝑺 and 𝑺 ′ be two semianalytic sets in M. If 𝑺 ′ ⊂ 𝑺 then 𝑺\𝑺 ′
is a semianalytic set in M.
Proof Indeed, 𝑺\𝑺 ′ = 𝑺 ∩ (M\𝑺 ′). □
Corollary 14.6.9 If 𝑺 is a semianalytic subset of M then its closure 𝑺 as well as
•
𝑺 = 𝑺 ∩ M\𝑺 and 𝜕𝑺 = 𝑺\𝑺 are semianalytic subsets of M.
Proof Indeed, the closure of (14.6.3) in U is
𝑞 𝜈𝑗 !
Ø Ø
𝑽 𝔉 𝑗 ∩ 𝑸 𝔊 𝑗 ∪ 𝑽 𝑔 𝑗,𝑘 .
𝑗=1 𝑘=1

•
The claims about 𝑺 and 𝜕𝑺 then follow from Proposition 14.6.6 and Corollary
14.6.8. □
14.6 Semianalytic Sets 509

Remark 14.6.10 An open set may have a semianalytic closure without

n o being semi-
analytic. Examples: 1) the complement in R of the sequence 𝑝1 ; 2) the
𝑝=1,2,...
complement in the plane of a continuous curve (i.e., the range of a continuous map
[0, 1] −→ R2 ) which is not analytic.
An analytic subvariety 𝑽 of M is a semianalytic subset of M.
Proposition 14.6.11 Let 𝑽 be an analytic subvariety of M. The regular and singular
parts ℜ (𝑽) and 𝔖 (𝑽) of 𝑽 (Definition 14.1.1) are semianalytic sets in M. The same
is true of the submanifolds ℜ𝑘 (𝑽), 𝑘 = 0, 1, ..., 𝑑 = dim 𝑽.
Proof We reason by induction on 𝑑, the result being trivial for 𝑑 = 0. The analytic
subvariety 𝑽 1 = 𝑽\ℜ𝑑 (𝑽) (Proposition 14.2.11) is a semianalytic set in M and
so is ℜ𝑑 (𝑽) by Corollary 14.6.8 and ℜ𝑑 (𝑽) by Corollary 14.6.9. By induction on
dimension, ℜ (𝑽 1 ) is a semianalytic set in M. We claim that

ℜ (𝑽 1 ) \ ℜ (𝑽 1 ) ∩ ℜ𝑑 (𝑽) = ℜ (𝑽) \ℜ𝑑 (𝑽) . (14.6.5)

Indeed, in M\ℜ𝑑 (𝑽) we have 𝑽 1 = 𝑽 hence ℜ (𝑽 1 ) = ℜ (𝑽) = ℜ (𝑽) \ℜ𝑑 (𝑽)

and therefore

ℜ (𝑽 1 ) = ℜ (𝑽 1 ) ∩ ℜ𝑑 (𝑽) ∪ (ℜ (𝑽) \ℜ𝑑 (𝑽)) .

By Proposition 14.1.2 (ℜ (𝑽) \ℜ𝑑 (𝑽)) ∩ ℜ𝑑 (𝑽) = ∅; (14.6.5) ensues. It implies

that ℜ (𝑽) = (ℜ (𝑽) \ℜ𝑑 (𝑽)) ∪ ℜ𝑑 (𝑽) is a semianalytic set in M and so is
𝔖 (𝑽) = 𝑽\ℜ (𝑽) by Corollary 14.6.8. Reasoning once again by induction on 𝑑 we
can assume that the submanifolds ℜ𝑘 (𝑽 1 ), 𝑘 = 1, ..., 𝑑 − 1, are semianalytic sets in
M. Then the same is true of ℜ𝑘 (𝑽) = ℜ𝑘 (𝑽 1 ) ∩ ℜ (𝑽). □

14.6.2 The regular and singular parts of a semianalytic set

Let 𝑺 be a semianalytic set in the C 𝜔 manifold M and ℜ (𝑺) [resp., 𝔖 (𝑺)] denote
its regular (resp., singular) part; let ℜ𝑘 (𝑺) denote the union of the 𝑘-dimensional
connected components of ℜ (𝑺) (Definition 14.1.1). It follows from Corollary 14.1.3
that ℜ (𝑺) ∩ 𝜕ℜ𝑘 (𝑺) = ∅ since 𝜕ℜ𝑘 (𝑺) ⊂ 𝜕ℜ (𝑺) (if 𝑬 ⊂ M, 𝜕𝑬 = 𝑬\𝑬). By
the dimension of 𝑺, dim 𝑺, we shall mean dim ℜ (𝑺).
Proposition 14.6.12 If 𝑺 is a semianalytic set in M then 𝑺 = ℜ (𝑺).
Proof It suffices to show that ℜ (𝑺) ∩ 𝑺 U ≠ ∅ for every set (14.6.3) (U: an analytic
polyhedron of M), a direct consequence of Proposition 14.2.8, according to which

𝑽 𝔉 𝑗 = ℜ 𝑽 𝔉 𝑗 for every 𝑗 = 1, ..., 𝑞. □
Remark 14.6.13 Proposition 14.6.12 does not extend to arbitrary subsets of M.
Example: 𝑬 = Q ∪ (0, +∞) ⊂ R.
510 14 Analytic Stratifications

Corollary 14.6.14 If 𝑺 is a semianalytic set in M then 𝔖 (𝑺) = 𝑺 ∩ 𝜕ℜ (𝑺).

Proof By Corollary 14.1.3 ℜ (𝑺) ∩ 𝜕ℜ (𝑺) = ∅ whence 𝑺 ∩ 𝜕ℜ (𝑺) ⊂ 𝔖 (𝑺). By

Proposition 14.6.12 𝔖 (𝑺) ⊂ ℜ (𝑺) ∪ 𝜕ℜ (𝑺) whence 𝔖 (𝑺) ⊂ 𝜕ℜ (𝑺). The result
ensues. □

Example 14.6.15 Let 𝑺 = (𝑥, 𝑦) ∈ R2 ; 𝑥 ≥ 0, 𝑦 > 0 ; we have

𝔖 (𝑺) = (𝑥, 𝑦) ∈ R2 ; 𝑥 = 0, 𝑦 > 0 ,

𝜕ℜ (𝑺) = (𝑥, 𝑦) ∈ R2 ; 𝑥 ≥ 0, 𝑦 ≥ 0, 𝑥𝑦 = 0 .

Corollary 14.6.16 If 𝑺 is a closed semianalytic subset of M then 𝑺 = ℜ (𝑺) and

𝔖 (𝑺) = 𝜕ℜ (𝑺).

We are going to apply the following elementary result in point-set topology.

Lemma 14.6.17 The interior of a subset 𝑌 of a Hausdorff topological space 𝑋,

• •
meaning the largest open subset of 𝑋 contained in 𝑌 , is equal to 𝑌 \𝑌 where 𝑌 =
𝑌 ∩ 𝑋\𝑌 .
•
Proof Let 𝑥 ∈ 𝑌 . If 𝑥 ∈ 𝑌 \𝑌 there is a neighborhood 𝑈 of 𝑥 such that 𝑈 ∩ 𝑋\𝑌 = ∅,
•
implying 𝑈 ⊂ 𝑌 . This proves that 𝑌 \𝑌 is open in 𝑋. If every neighborhood of 𝑥
•
intersects 𝑋\𝑌 then 𝑥 ∈ 𝑌 ; this also shows that if Ω ⊂ 𝑌 is an open subset of 𝑋 then
•
Ω ∩ 𝑋\𝑌 = ∅ whence Ω ⊂ 𝑌 \𝑌 . □

Proposition 14.6.18 If 𝑺 is a semianalytic subset of M and 𝑑 = dim 𝑺 then ℜ𝑑 (𝑺)

is semianalytic in M.

Proof By Definition 14.6.5 and Proposition 14.6.6 it suffices

to deal with a set 𝑺 U
as in (14.6.3) and, as a matter of fact, with 𝑺 = 𝑽 𝔉 𝑗 ∩ 𝑸 𝔊 𝑗 as a subset of the
analytic polyhedron U. Since 𝑸 𝔊 𝑗 is an open subset of U we have

ℜ𝑑 𝑽 𝔉 𝑗 ∩ 𝑸 𝔊 𝑗 = ℜ𝑑 𝑽 𝔉 𝑗 ∩ 𝑸 𝔊 𝑗 .

Since 𝑽 𝔉 𝑗 \ℜ𝑑 𝑽 𝔉 𝑗 is an analytic subvariety (Proposition 14.2.11)
ℜ𝑑 𝑽 𝔉 𝑗 is semianalytic in U, again by Proposition 14.6.6, whence the claim.□

Corollary 14.6.19 If 𝑺 is a semianalytic subset of M the same is true of 𝑺\ℜ𝑑 (𝑺)

and of 𝜕ℜ𝑑 (𝑺).

Proof Follows directly from Proposition 14.6.18 and Corollaries 14.6.8, 14.6.9. □
Inspection of the proof of Proposition 14.6.18 shows directly that dim (𝑺\ℜ𝑑 (𝑺))
< 𝑑.

Theorem 14.6.20 The regular and singular parts ℜ (𝑺) and 𝔖 (𝑺) of a semianalytic
set 𝑺 in M are semianalytic sets in M and dim 𝔖 (𝑺) < dim 𝑺.
14.6 Semianalytic Sets 511

Proof By Proposition 14.6.18 we know that ℜ𝑑 (𝑺) is semianalytic. This enables

us to reason as in the proof of Proposition 14.6.11: ℜ𝑑 (𝑺) and 𝑺1 = 𝑺\ℜ𝑑 (𝑺) are
semianalytic and so is ℜ (𝑺1 ) by induction on 𝑑. We have, by Proposition 14.1.2,

ℜ𝑑 (𝑺) ∩ (ℜ (𝑺) \ℜ𝑑 (𝑺)) = ∅

and
ℜ (𝑺1 ) = ℜ (𝑺1 ) ∩ ℜ𝑑 (𝑺) ∪ (ℜ (𝑺) \ℜ𝑑 (𝑺)) .

It follows that
ℜ (𝑺) \ℜ𝑑 (𝑺) = ℜ (𝑺1 ) \ ℜ (𝑺) ∩ ℜ𝑑 (𝑺)

is semianalytic; the same is therefore true of ℜ (𝑺) = ℜ𝑑 (𝑺) ∪ (ℜ (𝑺) \ℜ𝑑 (𝑺)) and
of 𝔖 (𝑺) = 𝑺\ℜ (𝑺). That dim 𝔖 (𝑺) < dim 𝑺 is a direct consequence of the fact
that 𝔖 (𝑺) ⊂ 𝑺\ℜ𝑑 (𝑺). □
Chapter 15
Division of Distributions by Analytic Functions

From the viewpoint of this book, one could say that the divisibility of distributions by
analytic functions (part of the ensemble of results due to S. Lojasiewicz) can be seen
as the solvability in distributions of any zero-order PDE with an analytic coefficient
that does not vanish identically. Admittedly, there is some irony in this interpretation
since so little is known (and nothing is discussed in this book) about higher order
linear PDEs in higher dimensions, whose leading part (principal symbol) vanishes
(identically) at some point in the base manifold – in contrast to the state of play
in analytic ODE theory, a very active area of research (going back to Fuchsian
equations, turning points, etc., and now linked to the general theory of asymptotic
expansions).
The proof of the Lojasiewicz inequality (Section 15.1) and the division of a
distribution by a C 𝜔 function (Section 15.2) are presented in all their (sometimes
rather painful) detail. Both proofs exploit the stratification (14.5.13) of the preceding
chapter. Each step, however, can be said to be elementary, no exotic mathematics is
ever needed. And even a superficial reading shows that the strategies of the proofs
are quite natural and help the visualization of the geometry (see regular separation,
Subsections 15.2.5, 15.2.6). The Lojasiewicz inequality confirms the intuition that,
when an analytic function vanishes on a set 𝑽, the rate of its decay as the variable
approaches 𝑽 is slower than some power of the distance to 𝑽.
In passing it should be mentioned that, on a complex-analytic manifold M, the
division of a distribution by a holomorphic function is a simpler affair than in the
real-analytic case (see [Schwartz, 1955]). The same is true of the division of a
hyperfunction by a C 𝜔 function (Theorem 7.1.19).
It must also be admitted that, at this point in time (that is, entering the third decade
of the XXIst century), exploiting stratifications would rather invoke Hironaka’s
desingularization theorem (see [Hironaka, 1965]) and the theory of subanalytic sets
(see [Hironaka, 1973]), a generalization of semi-analytic sets, needed because semi-
analyticity is not preserved under proper C 𝜔 maps (an example of this occurrence
can be found in [Hironaka, 1973], p. 453). As far as I know there is not yet a proof
of the desingularization theorem at the level of simplicity fitting for the intents of

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 513
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_15
514 15 Division of Distributions by Analytic Functions

this text. At the end of the chapter a short section is devoted to the statement of
the desingularization theorem and to a few simple, but interesting, applications that
attest to its power.

15.1 The Lojasiewicz Inequality

15.1.1 Functions with tempered growth or slow decay at the boundary

In this subsection Ω will be a nonempty bounded open subset of R𝑛 ; we use the

notation 𝜕Ω = Ω\Ω.
Definition 15.1.1 We shall say that a function 𝜑 ∈ C ∞ (Ω) has tempered growth
at the boundary of Ω if every point 𝑥 ◦ ∈ 𝜕Ω has a neighborhood 𝑈 in which the
following holds:
(Temp) To every 𝛼 ∈ Z+𝑛 there are positive constants 𝐶 𝛼 , 𝑚 𝛼 such that

∀𝑥 ∈ Ω ∩ 𝑈, |D 𝛼 𝜑 (𝑥)| ≤ 𝐶 𝛼 (dist (𝑥, 𝜕Ω)) −𝑚 𝛼 . (15.1.1)

∞ (Ω) the space of functions 𝜑 ∈ C ∞ (Ω) that have
We shall denote by Ctemp
tempered growth at the boundary.
∞ (Ω):
We list a number of obvious properties of Ctemp
∞ (Ω) is a subalgebra of C ∞ (Ω) (with respect to ordinary multiplication),
(1) Ctemp
stable under all partial derivatives D 𝛼 ;
(2) the restriction to an open set Ω′ ⊂ Ω maps Ctemp
∞ (Ω) into C ∞ (Ω ′ );
temp
∞
(3) the pullback of 𝑓 ∈ Ctemp (Ω) under the coordinate projection Ω × R𝑚 −→ Ω
∞ (Ω × R𝑚 ).
belongs to Ctemp
Let Cb∞ (Ω) denote the space of functions 𝑓 ∈ C ∞ (Ω) whose derivatives of all
orders are bounded in Ω; obviously Cb∞ (Ω) ⊂ Ctemp
∞ (Ω). Moreover, we can state

Proposition 15.1.2 Let Ω′ be an open subset of R𝑚 and 𝑓 ∈ Cb∞ (Ω′). If every

component of the map Φ : Ω −→ Ω′ belongs to Ctemp
∞ (Ω) then 𝑓 ◦ Φ ∈ C ∞ (Ω).
temp

Proof Apply the standard formulas (originally from [Faà di Bruno, 1855], see also
[Johnson, 2002]), for the derivatives of a composite function. □
Definition 15.1.3 Let 𝑺 be a subset of the closure Ω of Ω and 𝑓 a continuous function
in Ω. We shall say that 𝑓 decays slowly (or has slow decay) at 𝑺 if the following
property holds:
(Slow) To every point 𝑥 ◦ ∈ 𝑺 there exist an open set 𝑈 ∋ 𝑥 ◦ and a constant 𝜅 > 0
such that

∀𝑥 ∈ Ω ∩ 𝑈, 𝑥 ∉ 𝑺, (dist (𝑥, 𝑺)) 𝜅 ≲ | 𝑓 (𝑥)| . (15.1.2)

15.1 The Lojasiewicz Inequality 515

We shall denote by Cslow (Ω, 𝑺) the space of continuous functions in Ω that decay
slowly at 𝑺.

Notice that (15.1.2) entails 𝑓 (𝑥) ≠ 0 if 𝑥 ∈ (Ω ∩ 𝑈) \ (Ω ∩ 𝑈 ∩ 𝑺).

∞ (Ω) and if 𝑓 (𝑥) ≠ 0 whatever 𝑥 ∈ Ω
Proposition 15.1.4 If 𝑓 ∈ Cslow (Ω, 𝜕Ω) ∩Ctemp
1 ∞ (Ω).
then 𝑓 ∈ Ctemp

1
Proof For arbitrary 𝛼 ∈ Z+𝑛 , 𝑓 | 𝛼 |+1 𝐷 𝛼 𝑓 is a polynomial of degree |𝛼| with respect
to the derivatives 𝑓 with 𝛽 𝑗 ≤ 𝛼 𝑗 for each 𝑗 = 1, ..., 𝑛. The claim follows easily
𝐷𝛽
from (15.1.1) and (15.1.2). □
∞ (Ω), 𝑔 ∈ C ∞ (Ω). If 𝑓 (𝑥) ≠ 0
Corollary 15.1.5 Let 𝑓 ∈ Cslow (Ω, 𝜕Ω) ∩ Ctemp temp
𝑔 ∞ (Ω).
whatever 𝑥 ∈ Ω then 𝑓 ∈ Ctemp

Proposition 15.1.6 Let Ω and Ω′ be bounded open subsets of R𝑛 and R𝑚 respectively.

Let every component Φ 𝑗 ( 𝑗 = 1, ..., 𝑚) of the map Φ : Ω × Ω′ −→ Ω′ belong to
Cb∞ (Ω × Ω′) and let the C ∞ map 𝜓 : Ω −→ Ω′ satisfy Φ (𝑥, 𝜓 (𝑥)) = 0 for all
𝑥 ∈ Ω. If the Jacobian determinant det 𝜕Φ𝜕𝑦 (𝑥, 𝜓 (𝑥)) belongs to Cslow (Ω, 𝜕Ω) and
𝜕Φ
if det 𝜕𝑦 (𝑥, 𝜓 (𝑥)) ≠ 0 for every 𝑥 ∈ Ω then every component of 𝜓 belongs to
∞ (Ω).
Ctemp

Proof From the equation Φ (𝑥, 𝜓 (𝑥)) = 0 we derive

𝑚
𝜕Φℓ ∑︁ 𝜕Φℓ 𝜕𝜓 𝑘
(𝑥, 𝜓 (𝑥)) + (𝑥, 𝜓 (𝑥)) (𝑥) = 0 (15.1.3)
𝜕𝑥𝑖 𝑘=1
𝜕𝑦 𝑘 𝜕𝑥𝑖

𝜕Φ
for all 𝑖 = 1, ..., 𝑛, ℓ = 1, ..., 𝑚, and all 𝑥 ∈ Ω. Since det 𝜕𝑦 (𝑥, 𝜓 (𝑥)) ≠ 0 for every
𝑥 ∈ Ω we can solve (15.1.3):
𝑚
𝜕𝜓 𝑘 ∑︁ 𝜕Φℓ
(𝑥) = 𝐴 𝑘,ℓ (𝑥) (𝑥, 𝜓 (𝑥)) , (15.1.4)
𝜕𝑥𝑖 ℓ=1
𝜕𝑥𝑖

where the 𝐴 𝑘,ℓ (𝑥) are the entries of the inverse of the Jacobian matrix 𝜕Φ 𝜕𝑦 (𝑥, 𝜓 (𝑥));
−1
𝐴 𝑘,ℓ (𝑥) is of the form det 𝜕Φ
𝜕𝑦 (𝑥, 𝜓 (𝑥)) ×[a polynomial in the first partial deriva-

𝜕𝑦𝛽 (𝑥, 𝜓 (𝑥)), 1 ≤ 𝛼, 𝛽 ≤ 𝑚 ]. Our hypotheses and Corollary 15.1.5 imply that
tives 𝜕Φ 𝛼

the right-hand side in (15.1.4) belongs to Ctemp∞ (Ω); the same is therefore true of the

partial derivatives of all orders of each 𝜓 𝑘 . Since 𝜓 (Ω) ⊂ Ω′ and Ω′ is bounded the
sought result ensues. □
The concepts introduced in Definitions 15.1.1 and 15.1.3 can be easily extended
to a general C 𝜔 manifold M, through the use of local charts (U, 𝑥1 , ..., 𝑥 𝑛 ); in such
a chart the distance (or metric) is that defined by means of the coordinates 𝑥 𝑗 . Now,
supposing that Ω⊂ M is an open set and 𝑓 ∈ C 𝜔 (Ω) we shall say that:
516 15 Division of Distributions by Analytic Functions

(1) 𝑓 has tempered growth at the boundary 𝜕Ω of Ω if given an arbitrary point

℘ ∈ 𝜕Ω, there are a local coordinate chart (U, 𝑥1 , ..., 𝑥 𝑛 ) centered at ℘ and, for
each 𝛼 ∈ Z+𝑛 , positive constants 𝐶 𝛼 , 𝑚 𝛼 such that

∀𝑥 ∈ U ∩ Ω, 𝐷 𝑥𝛼 𝑓 (𝑥) ≤ 𝐶 𝛼 (dist (𝑥, U ∩ 𝜕Ω)) −𝑚 𝛼 ; (15.1.5)

(2) 𝑓 decays slowly (or has slow decay) at a closed subset 𝑺 of Ω if given an arbitrary
point ℘ ∈ 𝑺, there are a local coordinate chart (U, 𝑥1 , ..., 𝑥 𝑛 ) centered at ℘ and
a constant 𝜅 > 0 such that

∀𝑥 ∈ U ∩ Ω, 𝑥 ∉ 𝑺, (dist (𝑥, U ∩ 𝑺)) 𝜅 ≲ | 𝑓 (𝑥)| . (15.1.6)

This section is devoted to the proof of a fundamental property of real-analytic

functions, the Lojasiewicz inequality ([Lojasiewicz, 1959], p. 124 et seq.) which can
be stated as follows:

Theorem 15.1.7 Every C 𝜔 function in a C 𝜔 manifold M decays slowly at its null

set.

As evident in the preceding definition (2) it suffices to prove the result when
M = Ω, a bounded open subset of R𝑛 . It also suffices to prove the result when 𝑓
is real-valued, and then apply it to | 𝑓 | 2 when 𝑓 is complex. Theorem 15.1.7 is a
consequence of

Theorem 15.1.8 Let Ω be an open subset of R𝑛 , 𝑓 ∈ C 𝜔 (Ω) be real-valued and

𝑽 = {𝑥 ∈ Ω; 𝑓 (𝑥) = 0}; assume 0 ∈ 𝑽. There exist a neighborhood 𝑈 of the origin,
𝑈 ⊂ Ω, and 𝜅 > 0 such that

∀𝑥 ∈ 𝑈, (dist (𝑥, 𝑽)) 𝜅 ≲ | 𝑓 (𝑥)| . (15.1.7)

We can assume that 𝑓 extends holomorphically to an open subset Ω e of C𝑛

containing the closed polydisk Δ𝑟 and such that Ω ⊂ Ω
𝑛 e∩R . 𝑛

After division by a nonvanishing factor we may (thanks to Theorem 14.3.4)

′
assume that 𝑓 (𝑧) is a “real” Weierstrass polynomial 𝑃 (𝑧 ; 𝑧 𝑛 ) ∈ O Ω e ′ [𝑧 𝑛 ] of
degree 𝑑 ≥ 1, with Ω e ′ ⊂ C𝑛−1 the image of Ω e under the coordinate projection
′
𝜋˜ : 𝑧 ↦→ 𝑧 = (𝑧1 , ..., 𝑧 𝑛−1 ) (in real 𝑥-space we delete the tildes). The unique
factorization of Weierstrass polynomials (Corollary 14.3.13) yields 𝑃 = 𝑃1𝑝1 · · · 𝑃 𝜈𝑝𝜈 ,

where 𝑃𝑖 ∈ O Δ𝑟𝑛−1 ′ [𝑧 𝑛 ] are distinct irreducible Weierstrass polynomials and the
𝑝 𝛼 are positive integers. If 𝑝 = max 𝑝 𝛼 for each compact subset 𝐾 of Ω then
𝛼=1,...,𝜈

∀ (𝑥 ′; 𝑥 𝑛 ) ∈ 𝐾, |𝑃1 (𝑥 ′; 𝑥 𝑛 ) · · · 𝑃 𝜈 (𝑥 ′; 𝑥 𝑛 )| 𝑝 ≲ |𝑃 (𝑥 ′; 𝑥 𝑛 )| ,

which shows that there is no loss of generality in assuming that 𝑃 is trim (Definition
14.4.1), i.e.,
𝑓 (𝑥) = 𝑃 (𝑥 ′; 𝑥 𝑛 ) = 𝑃1 (𝑥 ′; 𝑥 𝑛 ) · · · 𝑃 𝜈 (𝑥 ′; 𝑥 𝑛 ) (15.1.8)
15.1 The Lojasiewicz Inequality 517

with coprime factors 𝑃 𝑗 . Let 𝐷 (𝑥 ′) denote the discriminant of (15.1.8); the null set
[cf. (14.5.2)]

𝒁 [1] = 𝑥 ′ ∈ 𝔔𝑟𝑛−1 ∩ R𝑛−1 ; 𝐷 (𝑥 ′) = 0

′ = Δ𝑟𝑛−1
′ (15.1.9)

is a proper analytic subvariety of Ω′ that contains the origin. When 𝒁 [1] = ∅ the
origin is a regular point of 𝑽 and the claim in Theorem 15.1.8 is self-evident. In the
proof we shall always assume 𝒁 [1] ≠ ∅; and we shall use induction on 𝑛 ≥ 2, the
result being banal when 𝑛 = 1.
We recall the following properties (see Section 14.4): an arbitrary root 𝜌 (𝑥 ′)
[1]
of 𝑃 (𝑥 ′; 𝑥 𝑛 ) = 0 in a connected component Γ of 𝔔𝑟𝑛−1 ′ \𝒁 is of class C 𝜔 and
either nowhere real in Γ or real at every point of Γ (Lemma 14.5.4). If there are
real roots of 𝑃 (𝑥 ′; 𝑥 𝑛 ) = 0 in Γ we can label them according to their natural order:
𝜌1 (𝑥 ′) < · · · < 𝜌 𝜇 (𝑥 ′) (𝜇 ≤ 𝑑). Each function 𝜌 𝜄 (𝑥 ′) extends as a continuous
function to the closure Γ of Γ (Lemma 14.5.5).

15.1.2 Preparatory lemmas

We regard a general monic polynomial in a single complex variable 𝑧,

𝑝(𝑎; 𝑧) = 𝑧 𝑚 + 𝑎 1 𝑧 𝑚−1 + · · · + 𝑎 𝑚 , (15.1.10)

as a functional of its coefficients,

n that is, of 𝑎 = (𝑎 1 , ..., 𝑎 𝑚 ) ∈ Co𝑚 , 𝑚 ≥ 1.
As before, Δ𝑟 (𝑎 ) = 𝑎 ∈ C ; 𝑎 𝑗 − 𝑎 ◦𝑗 ≤ 𝑟, 𝑗 = 1, ..., 𝑚 and Δ𝑟𝑚 = Δ𝑟𝑚 (0)
𝑚 ◦ 𝑚

(𝑟 > 0 arbitrary).
Lemma 15.1.9 If 𝑎 ∈ Δ𝑟𝑚 and 𝜁 ∈ C satisfies 𝑝(𝑎; 𝜁) = 0 then |𝜁 | < 𝑟 + 1.
Proof Suppose 𝑎 ∈ Δ𝑟𝑚 ; if 𝜁 ∈ C, |𝜁 | ≥ 𝑟 + 1, then
𝑚
∑︁
𝑎 1 𝜁 −1 + · · · + 𝑎 𝑚 𝜁 −𝑚 ≤ 𝑟 (𝑟 + 1) − 𝑗 < 1,
𝑗=1

hence 𝑝 (𝑎; 𝜁) ≠ 0. □
Let us now associate to each subset 𝐴 of C𝑚 the subset of the plane,

Sol ( 𝐴) = {𝑧 ∈ C; ∃𝑎 ∈ 𝐴, 𝑝(𝑎; 𝑧) = 0} .

We recall that a subset 𝑆 of a topological space 𝑿 is said to be connected if there

are no open subsets 𝑈 and 𝑉 of 𝑿 such that 𝑈 ∩ 𝑉 = ∅ and 𝑆 = (𝑆 ∩ 𝑈) ∪ (𝑆 ∩ 𝑉).
Lemma 15.1.10 Assume that 𝐴 ⊂ Δ𝑟𝑚 and let 𝐾 be a compact and connected subset
of Sol ( 𝐴). There is a number 𝜀 > 0 depending on 𝑟 and 𝐾 such that if diam 𝐴 ≤ 𝜀
then 1
diam 𝐾 ≤ 2 (𝑟 + 1) (diam 𝐴) 𝑚 . (15.1.11)
518 15 Division of Distributions by Analytic Functions

Proof If 𝑎 ∈ 𝐴 we denote by 𝜁1 (𝑎) , ..., 𝜁 𝑚 (𝑎) ∈ Sol ( 𝐴) the roots of 𝑝 (𝑎; 𝜁) = 0.

Let 𝑧 be an arbitrary point of Sol ( 𝐴) and 𝑎 ∈ 𝐴 be such that 𝑝 (𝑎; 𝑧) = 0. Lemma
15.1.9 entails |𝑧| < 𝑟 + 1 and if 𝑎 ′ is any other point of 𝐴 the Cauchy–Schwarz
inequality implies
𝑚
Ö
| 𝑝 (𝑎 ′; 𝑧)| = |𝑧 − 𝜁 𝑘 (𝑎 ′)| = | 𝑝 (𝑎 ′; 𝑧) − 𝑝 (𝑎; 𝑧)|
𝑘=1
𝑚
∑︁ √
𝑎 ℓ′ − 𝑎 ℓ 𝑧 𝑚−ℓ ≤ 𝑚 (𝑟 + 1) 𝑚−1 |𝑎 − 𝑎 ′ | .

=
ℓ=1

We conclude that there is a 𝑘 (depending on 𝑧) such that

1
|𝑧 − 𝜁 𝑘 (𝑎 ′)| ≤ 𝐶𝑟 |𝑎 − 𝑎 ′ | 𝑚 , (15.1.12)
√ 1 √
where 𝐶𝑟 = 2 (𝑟 + 1) 1− 𝑚 (since 𝑚 1/2𝑚 ≤ 2).
Let 𝑧1 , 𝑧2 ∈ 𝐾 be such that |𝑧1 − 𝑧2 | = diam 𝐾 and let 𝑎 ′ ∈ 𝐴 be such that
𝑝 (𝑎 ′; 𝑧 1 ) = 0. Let 𝛿 = min
′
𝑧1 − 𝜁 𝑗 (𝑎 ′) . For an arbitrary 𝑧 ∈ 𝐾 let 𝑎 ∈ 𝐴 be
𝜁 𝑗 (𝑎 )≠𝑧1
such that 𝑝 (𝑎; 𝑧) = 0 and let 𝑘 satisfy (15.1.12). If 𝜁 𝑘 (𝑎 ′) = 𝑧 1 then |𝑧 − 𝑧1 | ≤
1
𝐶𝑟 (diam 𝐴) 𝑚 . If 𝜁 𝑘 (𝑎 ′) ≠ 𝑧 1 the triangle inequality implies
1
|𝑧 − 𝑧1 | ≥ |𝑧1 − 𝜁 𝑘 (𝑎 ′)| − |𝑧 − 𝜁 𝑘 (𝑎 ′)| ≥ 𝛿 − 𝐶𝑟 (diam 𝐴) 𝑚 .
1 (1)
require 𝐴 to be so small that 𝐶𝑟 (diam 𝐴) < 𝛿/2; this entails 𝐾 ⊂ Δ𝜌 (𝑧1 ) ∪
We 𝑚

1
C\Δ𝜌(1) (𝑧1 ) , 𝜌 = 2 (𝑟 + 1) (diam 𝐴) 𝑚 . Since 𝐾 is connected one of the sets

(1) (1)
𝐾 ∩ Δ𝜌 (𝑧1 ), 𝐾 ∩ C\Δ𝜌 (𝑧 1 ) must be empty; since 𝑧1 ∈ 𝐾 we must have

𝐾 ⊂ Δ𝜌(1) (𝑧1 ). Since 𝑧2 ∈ 𝐾 we obtain |𝑧1 − 𝑧2 | < 𝜌, whence the claim. □

Corollary 15.1.11 Let the map [0, 1] ∋ 𝑡 ↦→ 𝑎 (𝑡) ∈ C𝑚 satisfy the conditions

∀𝑡 ∈ [0, 1] , |𝑎 (𝑡)| ≤ 𝑟; (15.1.13)

2 𝜅
∀ (𝑡 1 , 𝑡2 ) ∈ [0, 1] , |𝑎 (𝑡1 ) − 𝑎 (𝑡2 )| ≤ 𝑀 |𝑡1 − 𝑡2 |

with 𝑀 > 0 and 0 < 𝜅 ≤ 1. If a continuous function [0, 1] ∋ 𝑡 ↦→ 𝑧 (𝑡) ∈ C satisfies

𝑝 (𝑎 (𝑡) ; 𝑧 (𝑡)) = 0 for every 𝑡 ∈ [0, 1] then there is an 𝜀 > 0 such that the following
is true, for all pairs (𝑡 1 , 𝑡2 ), 0 ≤ 𝑡 1 < 𝑡2 ≤ 1, satisfying |𝑡1 − 𝑡 2 | < 𝜀,
1 𝜅
|𝑧 (𝑡1 ) − 𝑧 (𝑡2 )| ≤ 2 (𝑟 + 1) 𝑀 𝑚 |𝑡1 − 𝑡 2 | 𝑚 . (15.1.14)

Proof Let 𝐴 be the image of the map [𝑡 1 , 𝑡2 ] ∋ 𝑡 ↦→ 𝑎 (𝑡) ∈ C𝑚 ; we have 𝐴 ⊂ Δ𝑟𝑚

and diam 𝐴 ≤ 𝑀 |𝑡 1 − 𝑡 2 | 𝜅 by (15.1.13). Then (15.1.14) is a direct consequence of
Lemma 15.1.10. □
15.1 The Lojasiewicz Inequality 519

In other words, if 𝑎 (𝑡) is Hölder continuous with Hölder exponent 𝜅 ∈ (0, 1] then
𝑧 (𝑡) is Hölder continuous with Hölder exponent 𝑚𝜅 .
In the sequel we shall use the notation

[𝑥1 , 𝑥2 ] = {𝑥 ∈ R𝑛 ; ∃𝑡 ∈ [0, 1] , 𝑥 = 𝑡𝑥1 + (1 − 𝑡) 𝑥2 }

for arbitrary 𝑥1 , 𝑥2 ∈ R𝑛 .

Corollary 15.1.12 Let 𝑥1 , 𝑥2 ∈ R𝑛 and the map [𝑥1 , 𝑥2 ] ∋ 𝑥 ↦→ ℎ (𝑥) ∈ C𝑚 be

Hölder continuous with Hölder exponent 𝜅 ∈ (0, 1]; define

|ℎ (𝑥) − ℎ (𝑦)|
𝑟= max |ℎ (𝑥)| , 𝑀 = sup .
𝑥 ∈ [𝑥1 , 𝑥2 ] ( 𝑥,𝑦) ∈ [𝑥1 , 𝑥2 ] |𝑥 − 𝑦| 𝜅

If a continuous function [𝑥 1 , 𝑥2 ] ∋ 𝑥 ↦→ 𝜁 (𝑥) ∈ C satisfies 𝑝 (ℎ (𝑥) ; 𝜁 (𝑥)) ≡ 0 in

[𝑥1 , 𝑥2 ] and if |𝑥1 − 𝑥2 | ≤ 𝜀, 𝜀 > 0 suitably small, then
1 𝜅
|𝜁 (𝑥1 ) − 𝜁 (𝑥 2 )| ≤ 2 (𝑟 + 1) 𝑀 𝑚 |𝑥1 − 𝑥 2 | 𝑚 . (15.1.15)

Proof It suffices to apply Corollary 15.1.11 to 𝑎 (𝑡) = ℎ(𝑡𝑥1 + (1 − 𝑡) 𝑥 2 ). □

We give right-away a useful application of Corollary 15.1.12 to the polynomial
𝑃 in (15.1.8).

Corollary 15.1.13 If 𝜌 (𝑥 ′) is a real root of 𝑃 (𝑥 ′; 𝑥 𝑛 ) = 0 defined and continuous

in an open set 𝑈 ′ ⊂⊂ 𝔔𝑟𝑛−1 ′ then 𝜌 extends as a Hölder continuous function with
Hölder exponent (deg 𝑃) −1 to the closure 𝑈 ′ of 𝑈 ′.

Proof Let 𝑥1′ ∈ 𝑈 ′, 𝑥2′ ∈ 𝔔𝑟𝑛−1 be arbitrary; there is a 𝜁 (𝑥 ′) ∈ C 𝑥1′ , 𝑥2′ such

′

that 𝑃 (𝑥 ′; 𝜁 (𝑥 ′)) = 0 for every 𝑥 ′ ∈ 𝑥1′ , 𝑥2′ and 𝜁 (𝑥 ′) = 𝜌 (𝑥 ′) for every 𝑥 ′ ∈

′ ′
𝑥 , 𝑥 ∩ 𝑈 ′. It follows from Corollary 15.1.12 that Re 𝜁 is Hölder continuous in
1′ 2′
𝑥1 , 𝑥2 with Hölder exponent (deg 𝑃) −1 ; the same is true in 𝑥1′ , 𝑥2′ ∩ 𝑈 ′.

□
Concerning Hölder continuous maps we shall also make use of the following

Lemma 15.1.14 Let 𝑈 ( 𝛼) be a bounded domain in R𝑛−𝛼 and let G ⊂ R be the

𝑛

graph of a bounded map 𝑈 ( 𝛼) ∋ 𝑥 ( 𝛼) = (𝑥1 , ..., 𝑥 𝑛−𝛼 ) → 𝜉 𝑥 ( 𝛼) ∈ R 𝛼 that has

the following property:
(Höld) There are positive constants 𝜃 ≤ 1 and 𝐶 such that, for every pair of points
𝑥 ( 𝛼) and 𝑦 ( 𝛼) such that 𝑥 ( 𝛼) , 𝑦 ( 𝛼) ⊂ 𝑈 ( 𝛼) ,

𝜃
𝜉 𝑥 ( 𝛼) − 𝜉 𝑦 ( 𝛼) ≤ 𝐶 𝑥 ( 𝛼) − 𝑦 ( 𝛼) .

Under this hypothesis

𝜃
∀𝑥 ( 𝛼) ∈ 𝑈 ( 𝛼) , dist 𝑥 ( 𝛼) , 𝜉 𝑥 ( 𝛼) , 𝜕G ≲ dist 𝑥 ( 𝛼) , 𝜕𝑈 ( 𝛼) . (15.1.16)
520 15 Division of Distributions by Analytic Functions

Proof ( 𝛼) ∈ 𝑈 ( 𝛼) be arbitrary; there is a 𝑦 ( 𝛼) ∈ 𝜕𝑈 such that 𝑥 ( 𝛼) − 𝑦 ( 𝛼) =

Let 𝑥
dist 𝑥 ( 𝛼) , 𝜕𝑈 ( 𝛼) . The segment 𝑥 ( 𝛼) , 𝑦 ( 𝛼) ⊂ 𝑈 ( 𝛼) intersects 𝜕𝑈 only at the point

𝑦 ( 𝛼) . According to (Höld) we have, for 𝜀 > 0 arbitrarily small,

𝜃
𝜉 𝑥 ( 𝛼) − 𝜉 𝜀𝑥 ( 𝛼) + (1 − 𝜀) 𝑦 ( 𝛼) ≤ 𝐶 (1 − 𝜀) 𝜃 𝑥 ( 𝛼) − 𝑦 ( 𝛼) .

As 𝜀 ↘ 0 we have

𝜀𝑥 ( 𝛼) + (1 − 𝜀) 𝑦 ( 𝛼) , 𝜉 𝜀𝑥 ( 𝛼) + (1 − 𝜀) 𝑦 ( 𝛼) → 𝑦 ( 𝛼) , 𝜉 𝑦 ( 𝛼) ∈ 𝜕G

and
2 2 2
dist 𝑥 ( 𝛼) , 𝜉 𝑥 ( 𝛼) , 𝜕G ≤ 𝑥 ( 𝛼) − 𝑦 ( 𝛼) + 𝜉 𝑥 ( 𝛼) − 𝜉 𝑦 ( 𝛼)
2𝜃
≤ 𝐶12 𝑥 ( 𝛼) − 𝑦 ( 𝛼) ,

where 𝐶12 = 𝐶 2 + (diam 𝑈 ( 𝛼) ) 2(1−𝜃) . □

The next statement can be regarded as the generalized mean value theorem:
Lemma 15.1.15 Let 𝑚 ≥ 1 be an integer. Given any bounded segment [𝑎, 𝑏] ⊂ R
and any real-valued function 𝜑 ∈ C 𝑚 ( [𝑎, 𝑏]) there is a 𝑡 ∗ ∈ (𝑎, 𝑏) such that
𝑚 𝑚
∑︁ 𝑚 𝑚− 𝑗 𝑏−𝑎
(−1) 𝑗 𝜑 𝑎+ (𝑏 − 𝑎) = 𝜑 (𝑚) (𝑡 ∗ ) . (15.1.17)
𝑗=0
𝑗 𝑚 𝑚

We are using the notation 𝜑 (𝑚) = 𝜕𝑡𝑚 𝜑.

Proof When 𝑚 = 1 the result is the mean value theorem. We assume 𝑚 ≥ 2 and the
result proved up to 𝑚 − 1. To 𝜑 ∈ C 𝑚 ( [𝑎, 𝑏];
R) we associate the
function, defined
(and obviously, of class C 𝑚 ) in the segment 𝑎, 𝑎 + 𝑚−1
𝑚 (𝑏 − 𝑎) ,

1
𝜓 (𝑡) = 𝜑 𝑡 + (𝑏 − 𝑎) − 𝜑 (𝑡) .
𝑚
𝑚−1
The mean value theorem tells us that to every 𝑡• ∈ [𝑎, 𝑎 + 𝑚 (𝑏 − 𝑎)] there is a

𝑡∗ ∈ 𝑡 • , 𝑡• + 𝑚1 (𝑏 − 𝑎) such that

(𝑏 − 𝑎) (𝑚) 1
𝜑 (𝑡 ∗ ) = 𝜑 (𝑚−1) 𝑡• + (𝑏 − 𝑎) − 𝜑 (𝑚−1) (𝑡• ) (15.1.18)
𝑚 𝑚
= 𝜓 (𝑚−1) (𝑡• ) .
𝑚−1
Induction on 𝑚 and replacing 𝑏 by 𝑎 + 𝑚 (𝑏 − 𝑎), 𝜑 by 𝜓, allows us to select

𝑡• ∈ 𝑎, 𝑎 + 𝑚−1
𝑚 (𝑏 − 𝑎) such that
15.1 The Lojasiewicz Inequality 521
𝑚−1 𝑚−1
𝑏−𝑎 (𝑚−1)
∑︁
𝑗 𝑚−1 𝑗 +1
𝜓 (𝑡• ) = (−1) 𝜓 𝑏+ (𝑎 − 𝑏) .
𝑚 𝑗=0
𝑗 𝑚

The definition of 𝜓 and (15.1.18) then yield

𝑚 𝑚−1
𝑏−𝑎 ∑︁ 𝑚−1 𝑗
𝜑 (𝑚) (𝑡 ∗ ) = (−1) 𝑗 𝜑 𝑏 + (𝑎 − 𝑏)
𝑚 𝑗=0
𝑗 𝑚
𝑚
𝑗 𝑚−1 𝑗
∑︁
+ (−1) 𝜑 𝑏 + (𝑎 − 𝑏)
𝑗=1
𝑗 −1 𝑚
𝑚
𝑗 𝑚 𝑗
∑︁
= (−1) 𝜑 𝑏 + (𝑎 − 𝑏) .
𝑗=0
𝑗 𝑚

Exchanging 𝑗 and 𝑚 − 𝑗 in the sums yields (15.1.17). □

Corollary 15.1.16 Let 𝜑 ∈ C 𝑚 ( [𝑡 ◦ − 𝑇, 𝑡◦ + 𝑇]) (𝑡◦ ∈ R, 𝑇 > 0, 1 ≤ 𝑚 ∈ Z+ ) be

real-valued. If neither 𝜑 (𝑡) nor 𝜑 ′ (𝑡) vanishes and 𝜑 (𝑚) does not change sign in
[𝑡◦ − 𝑇, 𝑡◦ + 𝑇] then
𝑚
𝑇
min 𝜑 (𝑚) (𝑡) ≤ |𝜑 (𝑡◦ )| .
2𝑚 |𝑡−𝑡◦ | ≤𝑇

Proof There is no loss of generality in assuming 𝜑 > 0 and 𝜑 ′ < 0 throughout

[𝑡◦ − 𝑇, 𝑡◦ + 𝑇] since we may substitute either 𝜑 (−𝑡) or −𝜑 (−𝑡) for 𝜑 (𝑡). When 𝜑
is monotone decreasing in [𝑡 ◦ , 𝑡◦ + 𝑇] (15.1.17) yields:
𝑚 𝑚
𝑇 (𝑚)
∑︁ 𝑚 𝑚− 𝑗
min 𝜑 (𝑡) ≤ 𝜑 𝑡◦ + 𝑇 ≤ 2𝑚 𝜑 (𝑡 ◦ ) . □
𝑚 |𝑡−𝑡◦ | ≤𝑇 𝑗 𝑚
𝑗=0

15.1.3 Proof of the Lojasiewicz inequality

In this subsection we complete the proof of Theorem 15.1.8. We shall use the
notation of Ch. 14; in particular, 𝔔𝑟(𝑛) = Δ𝑟𝑛 ∩ R𝑛 ⊂⊂ Ω = Ω e ∩ R𝑛 . We are going
to exploit the properties of a local stratification of the type (14.5.13), but not for the
𝜕𝑓
function 𝑓 itself, instead, for the product 𝑓 𝜕𝑧 𝑛
, with the coordinates in C𝑛 such that
𝑧 𝑛 ↦→ 𝑓 (0, ..., 0, 𝑧 𝑛 ) does not vanish identically. We can also assume that 𝑓 (𝑧) is the
trim Weierstrass polynomial in (15.1.8), 𝑃 (𝑧 ; 𝑧 𝑛 ) ∈ O Ω [𝑧 𝑛 ], 𝑧 ′ = (𝑧 1 , ..., 𝑧 𝑛−1 ).
′ e ′

Actually, we are going to study the nullset 𝑽 ⋄ of the Weierstrass polynomial

𝑃 (𝑥 ′; 𝑥 𝑛 ) 𝜕𝑥
𝜕𝑃
𝑛
(𝑥 ′; 𝑥 𝑛 ), possibly not trim (Definition 14.4.1) as in the example

𝑃 (𝑥 ′; 𝑥 𝑛 ) = 𝑥 𝑛3 − 𝑎 (𝑥 ′). We denote by 𝑄 (𝑧 ′; 𝑧 𝑛 ) ∈ O Ω e ′ [𝑧 𝑛 ] the trimming
522 15 Division of Distributions by Analytic Functions

of 𝑃 (𝑧 ′; 𝑧 𝑛 ) 𝜕𝑧
𝜕𝑃
𝑛
(𝑧 ′; 𝑧 𝑛 ); thus 𝑽 ⋄ = {𝑥 ∈ Ω; 𝑄 (𝑥 ′; 𝑥 𝑛 ) = 0}. We continue to denote
by 𝑽 the nullset of 𝑓 , i.e., of 𝑃 (𝑥 ′; 𝑥 𝑛 ); of course, 𝑽 ⊂ 𝑽 ⋄ . We use the analogue of
(14.5.13), for 𝑽 ⋄ :
𝑛
Ø
𝑽 ⋄ ∩ 𝔔𝑟(𝑛) = 𝑽 ⋄[ 𝛼] , (15.1.19)
𝛼=1
[ 𝛼]
Ø 𝑁Ø
𝜄

𝑽 ⋄[ 𝛼] = Λ [𝜄,𝛼]
𝑝, (15.1.20)
[ 𝛼]
𝜄 ∈𝑱 𝑝=1

where 𝑱 [ 𝛼] is a finite set of indices and 𝑁 𝜄[ 𝛼] a (finite) positive integer. Each Λ [𝜄,𝛼]
𝑝 is
a connected component of 𝑽 ⋄[ 𝛼] , possibly empty: here we might have 𝑽 ⋄[ 𝛼] = 𝑽 ⋄[ 𝛼+1]
since we have not selected the indices 𝛼𝜈 where the jumps 𝑽 ⋄[ 𝛼𝜈 ] ≠ 𝑽 ⋄[ 𝛼𝜈+1 ] occur.
The analogue of 𝑃 [𝑗 𝛼] [see (14.5.1)] will be denoted by 𝑄 [𝑗 𝛼] ( 𝑗 = 1, ..., 𝛼). The
sets Γ 𝜄[ 𝛼] of Section 14.5 will be associated to 𝑽 ⋄ , not to 𝑽: Γ 𝜄[ 𝛼] is a connected
component of the open subset of 𝔔𝑟(𝑛−𝛼) ( 𝛼) in which none of the discriminants of the

[ 𝛼]
polynomials 𝑄 𝑗 𝑧 ; 𝑧 𝑛− 𝑗 vanishes; Λ [𝜄,𝛼]
( 𝛼) [ 𝛼]
𝑝 is the graph of one of the 𝑁 𝜄 maps

Γ 𝜄[ 𝛼] ∋ 𝑥 ( 𝛼) ↦→ 𝜏𝛼,ℓ
[ 𝛼]
𝛼
𝑥 ( 𝛼)
, ..., 𝜏 [ 𝛼]
1,ℓ1 𝑥 ( 𝛼)
∈ R𝛼, (15.1.21)

[ 𝛼] ( 𝛼) is a real root of the equation 𝑄 [ 𝛼] 𝑥 ( 𝛼) ; 𝑥
where 𝜏 𝑗,ℓ 𝑗
𝑥 𝑗 𝑛− 𝑗+1 = 0; the co-
ordinate projection 𝜋 𝛼 : 𝑥 ↦→ 𝑥 ( 𝛼) = (𝑥1 , ..., 𝑥 𝑛−𝛼 ) induces a C 𝜔 diffeomorphism
of Λ [𝜄,𝛼] [ 𝛼]
𝑝 onto Γ 𝜄 that extends continuously to the closures. Keep in mind that the
[ 𝛼]
closure of each Λ 𝜄, 𝑝 contains the origin of R𝑛 . We must also impose conditions of
the type (14.4.4):
1
∀𝑧 ( 𝛼) ∈ Δ𝑟𝑛−𝛼 , 𝑄 [𝑗 𝛼] 𝑧 ( 𝛼) ; 𝑧 𝑛− 𝑗+1 = 0 =⇒ 𝑧 𝑛− 𝑗+1 < 𝑟 𝑛− 𝑗+1 (15.1.22)
2
for every 𝑗 = 1, ..., 𝛼. For 𝑗 = 𝛼 = 1 (15.1.22) is equivalent to

𝜕𝑃 ′ 1
∀𝑧 ′ ∈ Δ𝑟(𝑛−1) , 𝑃 (𝑧 ′; 𝑧 𝑛 )
(𝑧 ; 𝑧 𝑛 ) = 0 =⇒ |𝑧 𝑛 | < 𝑟 𝑛 . (15.1.23)
𝜕𝑧 𝑛 2

Actually, we focus on the submanifolds 𝐿 [𝜄,𝛼] [ 𝛼] (𝑛) [ 𝛼]
𝑝 = Λ 𝜄, 𝑝 ∩ 𝔔𝑟 \𝑽 : 𝐿 𝜄, 𝑝 is the

submanifold of Λ [𝜄,𝛼] ′
𝑝 defined by 𝑃 (𝑥 ; 𝑥 𝑛 ) ≠ 0; thus,

[ 𝛼] (𝑛) ′ 𝜕𝑃 ′
𝐿 𝜄, 𝑝 ⊂ 𝑥 ∈ 𝔔𝑟 ; 𝑃 (𝑥 ; 𝑥 𝑛 ) ≠ 0, (𝑥 ; 𝑥 𝑛 ) = 0 . (15.1.24)
𝜕𝑥 𝑛
15.1 The Lojasiewicz Inequality 523

If an open subset of Λ [𝜄,𝛼] [ 𝛼]

𝑝 is contained in 𝑽 then Λ 𝜄, 𝑝 ⊂ 𝑽 (Corollary 14.2.5) and
𝐿 [𝜄,𝛼] [ 𝛼] [ 𝛼]
𝑝 = ∅. Suppose 𝐿 𝜄, 𝑝 ≠ ∅ and Λ 𝜄, 𝑝 to be the graph
of the map(15.1.21).
The

restriction of 𝑃 to Λ [𝜄,𝛼]
𝑝 , namely 𝑃 𝑥
( 𝛼) , 𝜏 [ 𝛼] 𝑥 ( 𝛼) , ..., 𝜏 [ 𝛼] 𝑥 ( 𝛼) ; 𝜏 [ 𝛼] 𝑥 ( 𝛼)
𝛼,ℓ 𝛼 2,ℓ2 1,ℓ1
with 𝑥 ( 𝛼) ∈ Γ 𝜄[ 𝛼] , divides
[ 𝛼]
𝛼 degÖ
Ö 𝑄𝑗

b[ 𝛼]
𝑃 𝑥 ( 𝛼) = [ 𝛼]
𝑃 𝑥 ( 𝛼) , 𝜏𝛼,ℓ 𝑥 ( 𝛼)
, ..., 𝜏 [ 𝛼]
𝑥 ( 𝛼)
; 𝜏 [ 𝛼]
𝑥 ( 𝛼)
.
1 𝛼 2,ℓ 2 1,ℓ 1
𝑗=1 ℓ 𝑗 =1
(15.1.25)

[ 𝛼]
By the usual argument, (15.1.25) extends as a holomorphic function 𝑃1 b 𝑧 ( 𝛼) in
Δ𝑟𝑛−𝛼 . Induction on the dimension allows us to
apply
(15.1.7) (i.e., Theorem 15.1.8):
there is a 𝜅1 > 0 such that, for all 𝑥 ( 𝛼) ∈ 𝜋 𝛼 𝐿 [𝜄,𝛼] [ 𝛼]
𝑝 (⊂ Γ 𝜄 ),

𝜅1
dist 𝑥 ( 𝛼) , 𝜕𝜋 𝛼 𝐿 [𝜄,𝛼]
𝑝
b[ 𝛼] 𝑥 ( 𝛼) ,
≲ 𝑃 1

whence
𝜅1
∀ (𝑥 ′, 𝑥 𝑛 ) ∈ 𝐿 [𝜄,𝛼]
𝑝 , dist 𝑥 ( 𝛼)
, 𝜕𝜋 𝛼 𝐿 [ 𝛼]
𝜄, 𝑝 ≲ |𝑃 (𝑥 ′; 𝑥 𝑛 )| . (15.1.26)

At this stage we apply Corollary 15.1.13: each 𝜏ℓ[𝑗𝛼] extends as a Hölder continu-
ous function (with Hölder exponent 1/𝑑1[ 𝛼] ) to the closure Γ 𝜄[ 𝛼] of Γ 𝜄[ 𝛼] . Note that

𝜕𝜋 𝛼 𝐿 [𝜄,𝛼]
𝑝 = 𝜋 𝛼 𝜕𝐿 [ 𝛼] (𝑛)
𝜄, 𝑝 (𝜕: boundary in 𝔔𝑟 ) since 𝜋 𝛼 induces a homeomor-

phism of Λ [𝜄,𝛼] [ 𝛼]
𝑝 onto Γ 𝜄 . We then apply Lemma 15.1.14: there is a 𝜃 ∈ (0, 1] such
that
𝜃
∀𝑥 ∈ 𝐿 [𝜄,𝛼] (𝑛) [ 𝛼] [ 𝛼]
𝑝 ∩ 𝔔 1 , dist 𝑥, 𝜕𝐿 𝜄, 𝑝 ≲ dist 𝜋 𝛼 (𝑥) , 𝜋 𝛼 𝜕𝐿 𝜄, 𝑝 . (15.1.27)
2𝑟

Combining (15.1.26) and (15.1.27) shows that, if 𝜅 2 = 𝜅 1 /𝜃 and if 𝑐 > 0 is sufficiently

small, then
𝜅
[ 𝛼] 2
∀𝑥 ∈ 𝐿 [𝜄,𝛼] (𝑛) ′
𝑝 ∩ 𝔔 1 , |𝑃 (𝑥 ; 𝑥 𝑛 )| ≥ 𝑐 dist 𝑥, 𝜕𝐿 𝜄, 𝑝 . (15.1.28)
2𝑟

Since 𝔔𝑟(𝑛) ⊂ Ω there is a constant 𝐶 > 0 such that

∀𝑥, 𝑦 ∈ 𝔔𝑟(𝑛) , |𝑃 (𝑥 ′; 𝑥 𝑛 )| ≥ |𝑃 (𝑦 ′; 𝑦 𝑛 )| − 𝐶 |𝑥 − 𝑦| . (15.1.29)

We shall denote by 𝔗 𝐿 [𝜄,𝛼] (𝑛)
𝑝 the (open) set of all the points 𝑥 ∈ 𝔔𝑟 such that

1 𝜅
[ 𝛼] 2
dist 𝑥, 𝐿 [𝜄,𝛼] −1
𝑝 < 𝐶 𝑐 dist 𝑥, 𝜕𝐿 𝜄, 𝑝 . (15.1.30)
2
524 15 Division of Distributions by Analytic Functions

We derive from (15.1.28) and (15.1.29):

1 𝜅2
|𝑃 (𝑥 ′; 𝑥 𝑛 )| ≥ 𝑐 dist 𝑥, 𝜕𝐿 [𝜄,𝛼]
𝑝
2

for all 𝑥 ∈ 𝔗 𝐿 [𝜄,𝛼] (𝑛) [ 𝛼]
𝑝 ∩ 𝔔 1 . Since 𝜕𝐿 𝜄, 𝑝 ⊂ 𝑽 we obtain
2𝑟

1
|𝑃 (𝑥 ′; 𝑥 𝑛 )| ≥ 𝑐 dist (𝑥, 𝑽) 𝜅2 (15.1.31)
2
𝑛−1 Ø
Ø
for all 𝑥 = (𝑥 ′; 𝑥 𝑛 ) ∈ 𝔔 (𝑛)
1 ∩ 𝔗 𝐿 [𝜄,𝛼] (𝑛) [ 𝛼]
𝑝 . In the complement 𝔔𝑟 \𝔗 𝐿 𝜄, 𝑝
2𝑟
𝛼=0 𝜄 ∈𝑱 [ 𝛼]
we have 1 𝜅
[ 𝛼] 2
dist 𝑥, 𝐿 [𝜄,𝛼] −1
𝑝 ≥ 𝐶 𝑐 dist 𝑥, 𝜕𝐿 𝜄, 𝑝 (15.1.32)
2
whence 1
dist 𝑥, 𝐿 [𝜄,𝛼] −1 𝜅2
𝑝 ≥ 𝐶 𝑐 dist (𝑥, 𝑽) . (15.1.33)
2
In the remainder of the proof we reason under the hypothesis that
𝑛−1 Ø
Ø
(𝑛)
𝑥 ∈ 𝔔𝑟/2 ,𝑥 ∉ 𝔗 𝐿 [𝜄,𝛼]
𝑝 ; (15.1.34)
𝛼=0 𝜄 ∈𝑱 [ 𝛼]

(𝑛)
in other words, we hypothesize that 𝑥 ∈ 𝔔𝑟/2 and that (15.1.32) holds for all
𝛼 ∈ [0, 𝑛 − 1] and all 𝜄 ∈ 𝑱 [ 𝛼] .
′ /2 ⊂ 𝜋 (𝑽 ⋄ ).
Claim 1. There is no loss of generality in assuming that 𝔔𝑟𝑛−1

Proof Indeed, let 𝑥 ′ ∈ 𝔔𝑟𝑛−1 ′ ′

′ /2 , 𝑥 ∉ 𝜋 (𝑽 ⋄ ); this means that neither 𝑃 (𝑥 ; 𝑥 𝑛 ) nor
𝑑
𝜕𝑃
𝜕𝑥𝑛(𝑥 ′; 𝑥 𝑛 ) vanishes for any value of 𝑥 𝑛 ∈ (−𝑟 𝑛 , 𝑟 𝑛 ). Since 𝜕𝜕𝑥𝑃𝑑 = 𝑑! we can apply
𝑛

Corollary 15.1.16 to 𝜑 (𝑡) = 𝑃 (𝑥 ′; 𝑡) with 𝑡 ◦ = 𝑥 𝑛 ∈ − 21 𝑟 𝑛 , 12 𝑟 𝑛 , 𝑚 = 𝑑 = deg 𝑃

and 𝑇 = 12 𝑟 𝑛 ; we get
𝑟 𝑑
|𝑃 (𝑥 ′; 𝑥 𝑛 )| ≥ 𝑑!
𝑛
. (15.1.35)
4𝑑
If the ratios 𝑟 𝑗 /𝑟 𝑛 ( 𝑗 = 1.., 𝑛 − 1) are suitably small then (15.1.35) contradicts (for
|𝑥 𝑛 | ≪ 𝑟 𝑛 ) the basic property of the Weierstrass polynomial 𝑃,
𝑛−1
∑︁ 𝑛−1
∑︁
(𝑛)
∀𝑥 ∈ 𝔔𝑟/2 , |𝑃 (𝑥 ′; 𝑥 𝑛 )| ≲ |𝑥 𝑛 | 𝑑 + 𝑥𝑗 ≲ |𝑥 𝑛 | 𝑑 + 𝑟 𝑗. □
𝑗=1 𝑗=1

In what follows 𝑥 ′ will be an arbitrary point of 𝔔𝑟𝑛−1 ′ /2 .

′
Claim 2. If 𝑃 (𝑥 ; 𝑥 𝑛 ) ≠ 0 for every 𝑥 𝑛 ∈ (−𝑟 𝑛 , 𝑟 𝑛 ) then
15.1 The Lojasiewicz Inequality 525

1 1
∀𝑥 𝑛 ∈ − 𝑟 𝑛 , 𝑟 𝑛 , dist (𝑥, 𝑽) 𝜅1 𝑑 ≤ 𝐶◦ |𝑃 (𝑥 ′; 𝑥 𝑛 )| (15.1.36)
2 2

with 𝐶◦ > 0 independent of 𝑥 ′.

Proof Suppose that 𝑃 (𝑥 ′; 𝑥 𝑛 ) ≠ 0 for every 𝑥 𝑛 ∈ (−𝑟 𝑛 , 𝑟 𝑛 ). The hypothesis that
𝜕𝑃 ′
′ /2 ⊂ 𝜋 (𝑽 ⋄ ) implies that the equation 𝜕𝑥 (𝑥 ; 𝑥 𝑛 ) = 0 must admit real roots
𝔔𝑟𝑛−1 𝑛

𝜏 𝑗 (𝑥 ′), 𝑗 = 1, ..., 𝜈. For each 𝑗 there are 𝛼 and 𝜄 such that 𝑥 ′, 𝜏 𝑗 (𝑥 ′) ∈ 𝐿 [𝜄,𝛼]

𝑝 ; then
(15.1.32)–(15.1.33) imply
1
𝑥 𝑛 − 𝜏 𝑗 (𝑥 ′) ≥ dist 𝑥, 𝐿 [𝜄,𝛼] −1 𝜅1
𝑝 ≥ 𝐶 𝑐 dist (𝑥, 𝑽) .
2
′
We apply Corollary 15.1.16 to 𝜑 (𝑡) = 𝑃 (𝑥 ; 𝑡) with 𝑡 varying1in one of the intervals
𝜏 𝑗 (𝑥 ) , 𝜏 𝑗+1 (𝑥 ) (𝜏0 = −𝑟 𝑛 , 𝜏𝜈+1 = 𝑟 𝑛 ), 𝑡◦ = 𝑥 𝑛 and 𝑇 = 2 min 𝑥 𝑛 − 𝜏 𝑗 (𝑥 ′) ,
′ ′
𝑗=1,...,𝜈
thus getting

|𝑃 (𝑥 ′; 𝑥 𝑛 )| ≥ 𝑑! (4𝑑) −𝑑 min 𝑥 𝑛 − 𝜏 𝑗 (𝑥 ′)
𝑑
𝑗=1,...,𝜈
−𝑑
≥ 𝑑! 8𝑑𝑐𝐶 −1 dist (𝑥, 𝑽) 𝜅1 𝑑 . □

′
In the remainder of the proof we shall assume that 𝑃 (𝑥 ; 𝑥 𝑛 ) = 0 for some
𝑥 𝑛 ∈ − 21 𝑟 𝑛 , 12 𝑟 𝑛 [cf. (15.1.23)]. Once again we subdivide the study into distinct
cases.
Claim 3. If there are no real roots of 𝜕𝑥 𝜕𝑃
𝑛
(𝑥 ′; 𝑥 𝑛 ) = 0 in (−𝑟 𝑛 , 𝑟 𝑛 ) then

1 1 (2𝑑) 𝑑
∀𝑥 𝑛 ∈ − 𝑟 𝑛 , 𝑟 𝑛 , dist (𝑥, 𝑽) 𝑑 ≤ |𝑃 (𝑥 ′; 𝑥 𝑛 )| . (15.1.37)
2 2 𝑑!

Proof The hypothesis demands that there be only one real root 𝜌 (𝑥 ′) of 𝑃 (𝑥 ′; 𝑥 𝑛 )
since, if there were two real roots 𝜌 𝑗 (𝑥 ′) < 𝜌 𝑘 (𝑥 ′), there would be 𝑥 𝑛 , 𝜌 𝑗 (𝑥 ′) < 𝑥 𝑛 <
𝜌 𝑘 (𝑥 ′), satisfying 𝜕𝑥
𝜕𝑃
𝑛
(𝑥 ′; 𝑥 𝑛 ) = 0. Applying Corollary 15.1.16 to 𝜑 (𝑡) = 𝑃 (𝑥 ′; 𝑡)
with

1 ′ ′ 1
𝑡◦ = 𝑥 𝑛 ∈ − 𝑟 𝑛 , 𝜌 (𝑥 ) ∪ 𝜌 (𝑥 ) , 𝑟 𝑛 ,
2 2
𝑇 = |𝑥 𝑛 − 𝜌 (𝑥 ′)| ≥ dist ((𝑥 ′, 𝑥 𝑛 ) , 𝑽) ,

yields directly (15.1.37). □

Claim 4. If there are real roots of 𝜕𝑃
𝜕𝑥𝑛 (𝑥 ′; 𝑥 𝑛 ) = 0 then

1 1
∀𝑥 ′ ∈ 𝔔𝑟𝑛−1
′ /2 , ∀𝑥 𝑛 ∈ − 𝑟 𝑛 , 𝑟 𝑛 , dist (𝑥, 𝑽)
𝜅1 𝑑
≲ |𝑃 (𝑥 ′; 𝑥 𝑛 )| .
2 2

Proof We order the real roots of 𝑃 (𝑥 ′; 𝑥 𝑛 ),

526 15 Division of Distributions by Analytic Functions

1 1
− 𝑟 𝑛 < 𝜌1 (𝑥 ′) < · · · < 𝜌 𝜇 (𝑥 ′) < 𝑟 𝑛 (1 ≤ 𝜇 ≤ 𝑑),
2 2
as well as those of 𝜕𝑃
𝜕𝑥𝑛 (𝑥 ′; 𝑥 𝑛 ):

1 1
− 𝑟 𝑛 < 𝜏1 (𝑥 ′) < · · · < 𝜏𝜈 (𝑥 ′) < 𝑟 𝑛 (1 ≤ 𝜈 ≤ 𝑑 − 1).
2 2
The proof of Claim 2 applies when − 21 𝑟 𝑛 < 𝑥 𝑛 < 𝜌1 (𝑥 ′) < 𝜏1 (𝑥 ′) or when
𝜏𝜈 (𝑥 ′) < 𝜌 𝜇 (𝑥 ′) < 𝑥 𝑛 < 21 𝑟 𝑛 , and leads to (15.1.37).
Next we look at the cases where 𝑥 𝑛 ∈ 𝜌 𝑗 (𝑥 ′) , 𝜏𝑘 (𝑥 ′) or 𝑥 𝑛 ∈ 𝜏𝑘 (𝑥 ′) , 𝜌 𝑗 (𝑥 ′)

under the assumption that neither 𝑃 (𝑥 ′; 𝑥 𝑛 ) nor 𝜕𝑥 𝜕𝑃

𝑛
(𝑥 ′; 𝑥 𝑛 ) vanish in these intervals.
′ ′
Suppose 𝑥 𝑛 − 𝜌 𝑗 (𝑥 ) ≤ |𝜏𝑘 (𝑥 ) − 𝑥 𝑛 |; in this casewe can apply Corollary 15.1.16
to 𝜑 (𝑡) = 𝑃 (𝑥 ′; 𝑡) with 𝑡 • = 𝑥 𝑛 ∈ 𝜌 𝑗 (𝑥 ′) , 𝜏𝑘 (𝑥 ′) or 𝑡 • = 𝑥 𝑛 ∈ 𝜌 𝑗 (𝑥 ′) , 𝜏𝑘 (𝑥 ′)
and 𝑇 = 𝑥 𝑛 − 𝜌 𝑗 (𝑥 ′) ; here also we get (15.1.37).
Suppose now |𝜏𝑘 (𝑥 ′) − 𝑥 𝑛 | < 𝑥 𝑛 − 𝜌 𝑗 (𝑥 ′) . Keep in mind that we are reasoning
under the hypothesis (15.1.32) where 𝐿 [𝜄,𝛼] ′ ′
𝑝 ∋ (𝑥 , 𝜏𝑘 (𝑥 )); this implies

1 −1 𝜅1
|𝜏𝑘 (𝑥 ′) − 𝑥 𝑛 | ≥ 𝐶 𝑐 dist 𝑥, 𝜕𝐿 [𝜄,𝛼]
𝑝 .
2
In this case we can apply Corollary 15.1.16 to 𝜑 (𝑡) = 𝑃 (𝑥 ′; 𝑡) with 𝑡 ◦ = 𝑥 𝑛 ∈
𝜌 𝑗 (𝑥 ′) , 𝜏𝑘 (𝑥 ′) and 𝑇 = |𝜏𝑘 (𝑥 ′) − 𝑥 𝑛 |; here we get (15.1.36).
Lastly we look at the case 𝑥 𝑛 ∈ (𝜏𝑘 (𝑥 ′) ,𝜏𝑘+1 (𝑥 ′)) under the assumption that
neither 𝑃 (𝑥 ′; 𝑥 𝑛 ) nor 𝜕𝑥
𝜕𝑃
𝑛
(𝑥 ′; 𝑥 𝑛 ) vanish in this interval. We can apply Corollary
15.1.16 to 𝜑 (𝑡) = 𝑃 (𝑥 ; 𝑡) with 𝑡 ◦ = 𝑥 𝑛 ∈ 𝜌 𝑗 (𝑥 ′) , 𝜏𝑘 (𝑥 ′) and
′

𝑇 = min (𝑥 𝑛 − 𝜏𝑘 (𝑥 ′) , 𝜏𝑘+1 (𝑥 ′) − 𝑥 𝑛 ) ;

here also (15.1.34) entails (15.1.36). □

The proof of Theorem 15.1.8 is complete.

15.2 Division of Distributions by Analytic Functions

15.2.1 Statement of the division theorem and strategy of the proof

In this section the following result will be proved:

Theorem 15.2.1 Let M be a manifold of class C 𝜔 countable at infinity. If 𝑓 ∈

C 𝜔 (M) does not vanish identically in any open subset of M then to every 𝑢 ∈
D ′ (M) there is a 𝑣 ∈ D ′ (M) such that 𝑢 = 𝑓 𝑣.
15.2 Division of Distributions by Analytic Functions 527

Suppose for a moment that 𝑽 = {℘ ∈ Ω; 𝑓 (℘) = 0} has no singularities, i.e., 𝑽

is a closed C 𝜔 submanifold of M, and that the distribution 𝑢 to be divided is the
function 𝑢 ≡ 1. Finding the divisor 𝑣 is equivalent to showing that the reciprocal
𝑓 −1 ∈ C 𝜔 (M\𝑽) is the restriction to M\𝑽 of a distribution in M. This is not true
for the reciprocal of an arbitrary 𝑓 ∈ C ∞ (M\𝑽). The next subsection is devoted to
characterizing the distributions in a proper open subset of M extendible to the whole
of M; the characterization will show, for example, that the function exp 1/𝑥 2 in

R\ {0} cannot be extended as a distribution in R. It will allow us to make the
first step in the proof of the division theorem: now taking 𝑽 to be an arbitrary
analytic subvariety of M and availing ourselves of Theorem 15.1.8 (the Lojasiewicz
inequality) we will show that there is a 𝑣 ∈ D ′ (M) such that 𝑢 = 𝑓 𝑣 in M\𝑽. From
there on we are allowed to assume that supp 𝑢 ⊂ 𝑽.
It will suffice to prove the successive claims locally and then patch up the local
results by means of a C ∞ partition of unity. This allows us to take M = Ω, a domain
in R𝑛 . It is therefore natural to reason in a neighborhood of a point of Ω ∩ 𝑽; to
lighten the notation it greatly helps to take this central point to be the origin of R𝑛 .
When dim 𝑽 = 0, i.e., 𝑽 is a discrete set, and supposing supp 𝑢 ⊂ 𝑽, the argument
is elementary and we settle it first; this also settles the case 𝑛 = 1.
When dim 𝑽 ≥ 1 (requiring 𝑛 ≥ 2) we exploit the stratification (14.5.13). A useful
feature of (14.5.13) is that each stratum Λ [𝜈 𝑘]
𝜄, 𝑝 is the graph of a C map (14.5.14).
𝜔

We therefore settle the case of the variety 𝑽 defined by an equation 𝑥 𝑛 = 𝜌 (𝑥 ′),

𝑥 ′ = (𝑥 1 , ..., 𝑥 𝑛−1 ) ∈ Γ, a domain in R𝑛−1 , 𝜌 ∈ C 𝜔 (Γ). The reasoning, in this case,
is an adaptation of the case dim 𝑽 = 0, making use of the Dirac distribution attached
to the analytic hypersurface 𝑽.
Lastly we must face the difficulties of a multistrata partition such as (14.5.13)
and take advantage of the strata coming together slowly at the origin – unlike, say,
the 𝑥-axis and the graph of the curve 𝑦 = exp −1/𝑥 2 in the plane. With this aim
we introduce (following Lojasiewicz) the concept of regularly separated sets 𝐴, 𝐵;
roughly speaking, sets that come together no faster than a power of the reciprocal
of dist (𝑥, 𝑦), 𝑥 ∈ 𝐴., 𝑦 ∈ 𝐵. This will be formalized rigorously and shown to be a
feature of the strata Λ [𝜈 𝑘]
𝜄, 𝑝 , thereby enabling us to complete the proof of Theorem
15.2.1.

15.2.2 Extendible distributions

Let M be a connected C 𝜔 manifold, countable at infinity, and let Ω ⊂ M be a

nonempty open set; we write 𝜕Ω = Ω\Ω. The restriction map

D ′ (M) ∋ 𝑢 ↦→ 𝑢| Ω ∈ D ′ (Ω) (15.2.1)

is not surjective unless Ω = M (see Proposition 15.2.28 below). In passing note the
contrast with hyperfunctions (see Proposition 7.1.3).
528 15 Division of Distributions by Analytic Functions

Definition 15.2.2 We shall say that a distribution in Ω is extendible (sometimes,

extendible to M) if it belongs to the range of the map (15.2.1). We shall denote by
′ (Ω) the subspace of D ′ (Ω) consisting of the extendible distributions.
Dext
Extendibility of a distribution in Ω is a property localizable at every point of the
boundary 𝜕Ω:
Proposition 15.2.3 Suppose that 𝑢 ∈ D ′ (Ω) has the following property: there is
an open covering 𝑈 𝑗 𝑗=1,2,... of 𝜕Ω such that, for each 𝑗, there is a 𝑢 𝑗 ∈ D ′ 𝑈 𝑗
whose restriction to 𝑈 𝑗 ∩ Ω is equal to that of 𝑢. Then 𝑢 is extendible.

Proof We can select an open covering 𝑈 𝑗 𝑗=0,1,2,... of Ω with 𝑈0 = Ω to be locally
Í∞ and, for each 𝑗 ∈ Z+ , aÍfunction
finite 𝜒 𝑗 ∈ C ∞ (M) such that supp 𝜒 𝑗 ⊂ 𝑈 𝑗 and
∞ ′
𝑗=0 𝜒 𝑗 ≡ 1 in Ω. Then 𝑢˜ = 𝑗=0 𝜒 𝑗 𝑢 𝑗 ∈ D (M) and 𝑢| ˜ Ω = 𝑢. □
We now focus on the case M = R𝑛 ; we shall make use of the Sobolev spaces
𝐻𝑚 (Ω), with 𝑚 ∈ Z, in an open subset Ω of R𝑛 (see Section 2.2).
Proposition 15.2.4 The following properties hold:
(1) Let 𝑚 ∈ Z+ . Every distribution 𝑢 ∈ 𝐻 −𝑚 (Ω) extends as a distribution 𝑢˜ ∈
𝐻 −𝑚 (R𝑛 ) with supp 𝑢˜ ⊂ Ω.
(2) If Ω is bounded then every extendible distribution 𝑢 ∈ D ′ (Ω) belongs to a
Sobolev space 𝐻 −𝑚 (R𝑛 ) for some 𝑚 ∈ Z+ .
(3) The restriction to an open set Ω′ ⊂⊂ Ω of an arbitrary distribution 𝑢 in Ω has
an extension to R𝑛 whose support is contained in Ω′.
Proof (1) If 𝑢 ∈ 𝐻 −𝑚 (Ω) we have, in Ω,
∑︁
𝑢= D𝛼 𝑓𝛼
|𝑎 | ≤𝑚

with 𝑓 𝛼 ∈ 𝐿 2 (Ω) for every multi-index 𝛼 ∈ Z+𝑛 . Extending each 𝑓 𝛼 as an 𝐿 2 function

in R𝑛 by setting 𝑓 𝛼 ≡ 0 in R𝑛 \Ω produces an extension 𝑢˜ ∈ 𝐻 −𝑚 (R𝑛 ) of 𝑢; obviously
supp 𝑢˜ ⊂ Ω.
(2) If Ω is bounded every 𝑢˜ ∈ D ′ (R𝑛 ) with supp 𝑢˜ ⊂ Ω belongs to a Sobolev
space 𝐻 −𝑚 (R𝑛 ) for some 𝑚 ∈ Z+ ; its restriction to Ω belongs to 𝐻 −𝑚 (Ω).
(3) The restriction of 𝑢 to Ω′ belongs to 𝐻 −𝑚 (Ω′) for some 𝑚 ∈ Z+ ; it suffices
then to apply (1). □
Proposition 15.2.5 For 𝑢 ∈ D ′ (Ω) to be extendible it is necessary and sufficient
that there be a locally finite open covering 𝑈 𝑗 𝑗=1,2,... of 𝜕Ω such that, for each 𝑗,
𝑢| Ω∩𝑈 𝑗 ∈ 𝐻 −𝑚 𝑗 Ω ∩ 𝑈 𝑗 for some 𝑚 𝑗 ∈ Z+ . If 𝑢 is extendible it has an extension

𝑢˜ ∈ D ′ (R𝑛 ) such that supp 𝑢˜ ⊂ Ω.

Proof The condition is necessary since the restriction to a bounded open subset
𝑈 of an arbitrary distribution in R𝑛 belongs to 𝐻 −𝑚 (𝑈) for some 𝑚 ∈ Z. The
condition is sufficient: according to Proposition 15.2.4, 𝑢| Ω∩𝑈 𝑗 ∈ 𝐻 −𝑚 𝑗 Ω ∩ 𝑈 𝑗
can be extended as a distribution 𝑢˜ 𝑗 ∈ 𝐻 −𝑚 𝑗 𝑈 𝑗 such that supp 𝑢˜ 𝑗 ⊂ Ω ∩ 𝑈 𝑗

whence the conclusion by Proposition 15.2.3. □
15.2 Division of Distributions by Analytic Functions 529

Example 15.2.6 The function exp |𝑥| −2 cannot be extended as a distribution in
−𝑚 (R𝑛 \ {0}) whatever
R𝑛 . It is an easy exercise to prove that it does not belong to 𝐻loc
𝑚 ∈ Z+ .
Lemma 15.2.7 Let Ω be a bounded open subset of R𝑛 , 𝑚 ∈ Z+ and 𝜑 ∈ Ctemp∞ (Ω)
−𝑚 −𝑚
(Definition 15.1.1). If 𝑢 ∈ 𝐻 (Ω) then 𝜑𝑢 ∈ 𝐻 1 (Ω) for some 𝑚 1 ∈ Z+ .
Proof Let 𝜒 ∈ Cc∞ (Ω) be arbitrary. According to (15.1.1) and to the Leibniz rule
we have
∑︁
|D 𝛼 (𝜑 𝜒) (𝑥)| ≤ 𝐶 𝛼 (dist (𝑥, 𝜕Ω)) −𝑚𝛽 D 𝛼−𝛽 𝜒 (𝑥) . (15.2.2)
𝛽⪯𝛼

Select 𝑦 ∈ 𝜕Ω such that dist (𝑥, 𝜕Ω) = |𝑥 − 𝑦|; then the segment [𝑥, 𝑦] is entirely
contained in Ω; we have
∫ 1
𝜒 (𝑥) = (𝑥 − 𝑦) · ∇𝜒 (𝑡𝑥 + (1 − 𝑡) 𝑦) d𝑡
0

for some 𝑡 ∈ (0, 1]. We derive

| 𝜒 (𝑥)| ≤ dist (𝑥, 𝜕Ω) max |∇𝜒| .

[𝑥,𝑦 ]

𝜕𝜒
Applying this same inequality with 𝜕𝑥 𝑗
(𝑡𝑥 + (1 − 𝑡) 𝑦) in the place of 𝑥 yields
directly
max |∇𝜒| ≤ (dist (𝑥, 𝜕Ω)) max ∇2 𝜒 ,
[𝑥,𝑦 ] [𝑥,𝑦 ]

whence
| 𝜒 (𝑥)| ≤ dist (𝑥, 𝜕Ω) 2 max ∇2 𝜒 .
[𝑥,𝑦 ]

More generally, given any integer 𝑀 ≥ 1, we have

| 𝜒 (𝑥)| ≤ dist (𝑥, 𝜕Ω) 𝑀 max ∇ 𝑀 𝜒 (15.2.3)

[𝑥,𝑦 ]

with an obvious interpretation of the tensor ∇ 𝑀 𝜒. On the other hand we have

∫ 𝑥1 ∫ 𝑥𝑛
𝜕𝑛 𝜒
𝜒 (𝑥) = ··· (𝜉1 , ..., 𝜉 𝑛 ) d𝜉1 · · · d𝜉 𝑛 ,
−∞ −∞ 𝜕𝑥 1 · · · 𝜕𝑥 𝑛

whence, by the Cauchy–Schwarz inequality,

√︁ 𝜕𝑛 𝜒
| 𝜒 (𝑥)| ≤ diam supp 𝜒 . (15.2.4)
𝜕𝑥 1 · · · 𝜕𝑥 𝑛 𝐿2

Applying (15.2.4) to ∇ 𝑀 𝜒 and combining with (15.2.3) yields

′
√
| 𝜒 (𝑥)| ≤ 𝐶 𝑀 diam Ω dist (𝑥, 𝜕Ω) 𝑀 ∥ 𝜒∥ 𝐻 𝑀+𝑛 , (15.2.5)
530 15 Division of Distributions by Analytic Functions

′ > 0 depends only on 𝑀. Combining

where ∥·∥ 𝐻 𝑘 is the norm in 𝐻 𝑘 (R𝑛 ) and 𝐶 𝑀
(15.2.5) with (15.2.2) yields
√ ∑︁
|D 𝛼 (𝜑 𝜒) (𝑥)| ≤ 𝐶 𝛼′′ diam Ω (dist (𝑥, 𝜕Ω)) 𝑀−𝑚𝛽 D 𝛼−𝛽 𝜒 𝐻 𝑀+𝑛 .
𝛽⪯𝛼

Taking |𝛼| ≤ 𝑚 and 𝑀 ≥ max 𝑚 𝛽 yields directly

|𝛽 | ≤𝑚

′′′
∥𝜑𝜒∥ 𝐻 𝑚 ≤ 𝐶𝑚, 𝑀 (1 + diam Ω) ∥ 𝜒∥ 𝐻 𝑚+𝑀+𝑛 .

We conclude that if 𝑚 1 ≥ 𝑚 + 𝑀 + 𝑛 then the multiplication 𝜒 ↦→ 𝜑 𝜒 defines a

continuous linear map 𝐻◦𝑚1 (Ω) −→ 𝐻 𝑚 (Ω); its transpose, 𝑢 ↦→ 𝜑𝑢, is therefore a
continuous linear map 𝐻 −𝑚 (Ω) −→ 𝐻 −𝑚1 (Ω) (cf. Section 2.2). □
∞ (Ω). If 𝑢 ∈ D ′ (Ω) is ex-
Proposition 15.2.8 Assume 𝜕Ω ≠ ∅ and let 𝜑 ∈ Ctemp
tendible then so is 𝜑𝑢.
′ (Ω) is a C ∞ (Ω)-module with respect to “ordinary” multi-
In other words, Dext temp
plication.
Proof By Proposition 15.2.5 it suffices to deal with the case of Ω bounded and
𝑢 ∈ 𝐻 −𝑚 (Ω), 𝑚 ∈ Z+ , in which case the claim is a direct consequence of Lemma
15.2.7 and Proposition 15.2.4. □
We are now in a position to prove the extendibility of distributions across analytic
varieties.

Theorem 15.2.9 Let 𝑽 be a proper analytic subvariety of a C 𝜔 manifold M. Let

𝑢 ∈ D ′ (M) be arbitrary. If the nullset of 𝑓 ∈ C 𝜔 (M) is contained in 𝑽 then to
every 𝑢 ∈ D ′ (M) there is a 𝑣 ∈ D ′ (M) such that supp (𝑢 − 𝑓 𝑣) ⊂ 𝑽.

Proof Theorem 15.1.7 tells us that 𝑓 ∈ Cslow (M\𝑽, 𝑽); then Proposition 15.1.4
implies that 1𝑓 ∈ Ctemp
∞ (M\𝑽). Finally, Propositions 15.2.5 and 15.2.8 imply that

1
𝑓 𝑢| M\𝑽 is extendible to M. We can take 𝑣 to be an extension of 1𝑓 𝑢| M\𝑽 . □

15.2.3 The division theorem when dim 𝑽 = 0

Proposition 15.2.10 Let M be a manifold of class C 𝜔 , countable at infinity, and 𝑽

a discrete subset of M. Suppose that 𝑓 ∈ C 𝜔 (M) vanishes at every point of 𝑽 and
let 𝑢 ∈ D ′ (M) have the property that 𝑢 = 𝑓 𝑤 in M\𝑽 for some 𝑤 ∈ D ′ (M).
Under these hypotheses there is a 𝑣 ∈ D ′ (M) such that 𝑢 = 𝑓 𝑣 in M.

Proof Since the claim is local we can assume that M is an open subset of R𝑛
containing the origin and that 𝑽 = {0}. It follows that 𝑢 − 𝑓 𝑤 = 𝑃 (D) 𝛿 with
𝛿 the Dirac distribution and 𝑃 ∈ C [𝜉1 , ..., 𝜉 𝑛 ], D = (D1 ...D𝑛 ), D 𝑗 = √1 𝜕𝑥𝜕 𝑗 .
−1
15.2 Division of Distributions by Analytic Functions 531

If we can prove that there is a 𝑄 ∈ C [𝜉1 , ..., 𝜉 𝑛 ] such that 𝑓 𝑄 (D) 𝛿 = 𝑃 (D) 𝛿
then 𝑣 = 𝑣 1 + 𝑄 (D) 𝛿 ∈ D ′ (Ω) will satisfy 𝑓 𝑣 = 𝑢. Thus Proposition 15.2.10
is a consequence of the lemma below. We generalize the framework a little bit by
substituting a formal power series 𝐹 for the (convergent) Taylor expansion of 𝑓
at the origin. Note that 𝛿 is defined as a linear functional on C [[𝑥1 , ..., 𝑥 𝑛 ]] by
⟨𝐹, 𝛿⟩ = 𝐹 (0), the zero-order term in 𝐹; and, naturally,

⟨𝐹, 𝑄 (D) 𝛿⟩ = ⟨𝑄 (−D) 𝐹, 𝛿⟩ . (15.2.6)

As a matter of fact, the bracket (15.2.6) allows us to identify the topological vector
spaces C [𝜉1 , ..., 𝜉 𝑛 ] and C [[𝑥 1 , ..., 𝑥 𝑛 ]] with the dual of each other. (We shall not
make use of the natural topologies on these spaces; on this subject see [Treves, 1967],
pp. 227–231). □
Í
Lemma 15.2.11 Let 𝐹 = 𝛼∈Z+𝑛 𝑐 𝛼 𝑥 𝛼 ∈ C [[𝑥1 , ..., 𝑥 𝑛 ]] be such that 𝑐 𝛼 ≠ 0 for
some 𝛼. To every 𝑃 ∈ C [𝜉1 , ..., 𝜉 𝑛 ] there is a 𝑄 ∈ C [𝜉1 , ..., 𝜉 𝑛 ] such that

𝐹 (𝑥) 𝑄 (D) 𝛿 = 𝑃 (D) 𝛿. (15.2.7)

Proof Eq. (15.2.7) is equivalent to the property that, whatever 𝑆 ∈ C [[𝑥1 , ..., 𝑥 𝑛 ]],

⟨𝑄 (−D) (𝐹𝑆) , 𝛿⟩ = ⟨𝑃 (−D) 𝑆, 𝛿⟩ . (15.2.8)

First, let us find a solution of (15.2.8) when 𝑃 = 1; we have

∑︁ (−1) | 𝛼 | D E
⟨𝑄 (−D) (𝐹𝑆) , 𝛿⟩ = ⟨D 𝛼 𝑆, 𝛿⟩ 𝑄 ( 𝛼) (−D) 𝐹, 𝛿 . (15.2.9)
𝛼∈Z𝑛
𝛼!
+

Let 𝛽Í ∈ Z+𝑛 be such that 𝑐 𝛽 ≠ 0 and if |𝛾| < 𝛽 then necessarily 𝑐 𝛾 = 0; thus
𝐹 = |𝛾 | ≥ |𝛽 | 𝑐 𝛾 𝑥 𝛾 . If 𝑄 (𝜉) = 𝑐−1
𝛽 𝜉 we have
𝛽

D E
𝑄 ( 𝛼) (−D) 𝐹, 𝛿 = 0 if 𝛼 ≠ 0,
⟨𝑄 (−D) 𝐹, 𝛿⟩ = (−1) |𝛽 | 𝛽!,

whence, by (15.2.9), ⟨𝑄 (−D) (𝐹𝑆) , 𝛿⟩ = ⟨𝑆, 𝛿⟩.

Now assume deg 𝑃 = 𝑚 ≥ 1; by the result for 𝑃 = 1 there is a 𝑄 𝑚 ∈ C [𝜉1 , ..., 𝜉 𝑛 ]
satisfying
𝐹 𝑚+1 (𝑥) 𝑄 𝑚 (D) 𝛿 = 𝛿. (15.2.10)
We derive from (15.2.10):

𝑃 (D) 𝛿 = 𝑃 (D) 𝐹 𝑚+1 (𝑥) 𝑄 𝑚 (D) 𝛿
∑︁ 1
= 𝑃 ( 𝛼) (D) 𝐹 𝑚+1 (𝑥) D 𝛼 𝑄 𝑚 (D) 𝛿.
𝛼!
| 𝛼 | ≤𝑚
532 15 Division of Distributions by Analytic Functions

Now, to each 𝛼 there is a 𝐺 𝛼 ∈ C [[𝑥 1 , ..., 𝑥 𝑛 ]] such that 𝑃 ( 𝛼) (D) 𝐹 𝑚+1 (𝑥) =

𝐹 (𝑥) 𝐺 𝛼 (𝑥) (with 𝐺 𝛼 ≡ 0 if |𝛼| > 𝑚), whence
∑︁ 1
𝑃 (D) 𝛿 = 𝐹 (𝑥) 𝐺 𝛼 (𝑥) D 𝛼 𝑄 𝑚 (D) 𝛿. (15.2.11)
𝛼!
| 𝛼 | ≤𝑚

It is readily checked that

∑︁ 1
𝐺 𝛼 (𝑥) D 𝛼 𝑄 𝑚 (D) 𝛿 = 𝑄 (D) 𝛿 (15.2.12)
𝛼!
| 𝛼 | ≤𝑚

for a suitably chosen 𝑄 ∈ C [𝜉1 , ..., 𝜉 𝑛 ]. Indeed, only finite parts of the formal
series 𝐺 𝛼 enter in (15.2.11); when the 𝐺 𝛼 are polynomials the left-hand side is a
distribution supported by the origin. □

Remark 15.2.12 Inspection of the proof of Lemma 15.2.11 shows that the coeffi-
cients of the polynomial 𝑄 in (15.2.12) are linear functions of those of 𝑃. Indeed,
𝑄 𝑚 in (15.2.10) depends solely on finitely many coefficients of the power series 𝐹.
The coefficients of the power series 𝐺 𝛼 depend linearly on those of 𝑃 (as well as on
those of 𝐹).

Theorem 15.2.13 Let M be a manifold of class C 𝜔 , countable at infinity, and let

𝑓 ∈ C 𝜔 (M) have a discrete nullset 𝑽. To every 𝑢 ∈ D ′ (M) there is a 𝑣 ∈ D ′ (M)
such that 𝑢 = 𝑓 𝑣.

Proof Theorem 15.2.9 tells us that there is a 𝑤 ∈ D ′ (M) such that supp (𝑢 − 𝑓 𝑤) ⊂
𝑽. It suffices then to apply Proposition 15.2.10. □

15.2.4 Division of distributions supported by smooth graphs

In this subsection it is convenient to modify our notation for the coordinates in

Euclidean space R𝑛 and write 𝑥 = 𝑥1 , ..., 𝑥 𝑛−𝑞−1 instead of 𝑥 (𝑞+1) , 𝑦 𝑗+1 = 𝑥 𝑛−𝑞+ 𝑗
with 𝑗 = 0, ..., 𝑞, 𝑦 = 𝑦 1 , ..., 𝑦 𝑞+1 ∈ R𝑞+1 and (𝑥, 𝑦) for the generic point in
R𝑛−𝑞−1 × R𝑞+1 ; here Γ will be a bounded open subset of R𝑛−𝑞−1 .
Let Λ ⊂ R𝑛−𝑞−1 × R𝑞+1 be the graph of a C ∞ map

𝜌 : Γ ∋ 𝑥 ↦→ 𝜌 (𝑥) = 𝜌1 (𝑥) , ..., 𝜌𝑞+1 (𝑥) ∈ R𝑞+1 ; (15.2.13)

Λ is a C ∞ submanifold of Γ × R𝑞+1 defined by the equations 𝑦 𝑗 = 𝜌 𝑗 (𝑥), 𝑗 = 0, ..., 𝑞.

We shall assume that 𝜌 extends as a Hölder continuous map Γ −→ R𝑞+1 . The
coordinate projection (𝑥, 𝑦) ↦→ 𝑥 is a C ∞ diffeomorphism of Λ onto Γ; the pullback
induces a (topological vector space) isomorphism of Cc∞ (Γ) onto Cc∞ (Λ) whose
transpose is an isomorphism of D ′ (Λ) onto D ′ (Γ).
15.2 Division of Distributions by Analytic Functions 533

Let U be an open subset of Γ × R𝑞+1 containing Λ; the restriction to Λ,

Cc∞ (U) ∋ 𝜑 (𝑥, 𝑦) ↦→ 𝜑 (𝑥, 𝜌 (𝑥)) ∈ Cc∞ (Λ) , (15.2.14)

is a continuous linear map.

Proposition 15.2.14 The map (15.2.14) is surjective.
Proof We select a function 𝜒 ∈ C ∞ Γ × R𝑞+1 such that supp 𝜒 ⊂ U and 𝜒 (𝑥, 𝑦) =

1 for all (𝑥, 𝑦) in an open subset U ′ ⊂ U containing Λ. Each function 𝜓 (𝑥) ∈ Cc∞ (Λ)
is the restriction to Λ of the function 𝜓 (𝑥) 𝜒 (𝑥, 𝑦) ∈ Cc∞ (U). □
Corollary 15.2.15 The transpose of (15.2.14) is an injective continuous linear map
D ′ (Λ) −→ D ′ (U).
We introduce the Radon measure 𝛿Λ in Γ × R𝑞+1 defined by
∫
⟨𝛿Λ , 𝜑⟩ = 𝜑 (𝑥, 𝜌 (𝑥)) d𝑥, 𝜑 ∈ C Γ × R𝑞+1 ; (15.2.15)
Γ

𝛿Λ can be thought of as the Dirac measure associated to Λ. Given any distribution

𝑤 ∈ D ′ (Γ) we can define a “product” 𝑤 (𝑥) 𝛿Λ by the formula

⟨𝑤 (𝑥) 𝛿Λ , 𝜑⟩ = ⟨𝑤 (𝑥) , 𝜑 (𝑥, 𝜌 (𝑥))⟩ (15.2.16)

for arbitrary 𝜑 ∈ Cc∞ Γ × R𝑞+1 ; 𝑤 (𝑥) 𝛿Λ ∈ D ′ Γ × R𝑞+1 . We can also introduce

𝑤 (𝑥) 𝜕𝑦𝛼 𝛿Λ for any 𝛼 ∈ Z+𝑞+1 :

D E
𝑤 (𝑥) 𝜕𝑦𝛼 𝛿Λ , 𝜑 = (−1) | 𝛼 | 𝑤 (𝑥) , 𝜕𝑦𝛼 𝜑 (𝑥, 𝜌 (𝑥)) ,

for arbitrary 𝜑 ∈ Cc∞ Γ × R𝑞+1 ; obviously, 𝑤 (𝑥) 𝜕𝑦𝛼 𝛿Λ = 𝜕𝑦𝛼 (𝑤 (𝑥) 𝛿Λ ).

We are going to need the following
Lemma 15.2.16 We have for all 𝑚 ∈ Z+ and all 𝜙 ∈ Cc∞ (Γ),
∑︁
∀𝑥 ∈ Γ, |𝜙 (𝑥)| ≤ dist (𝑥, 𝜕Γ) 𝑚 ∥𝜕 𝛼 𝜙∥ 𝐿 ∞ . (15.2.17)
𝑛−𝑞−1
𝛼∈Z+
| 𝛼 |=𝑚

Proof Given 𝑥 ∈ Γ we select arbitrarily 𝜉 ∈ 𝜕Γ such that |𝑥 − 𝜉 | = dist (𝑥, 𝜕Γ);

given 𝜙 ∈ Cc∞ (Γ) we define

𝜙♭ (𝑥, 𝑡) = 𝜙 ((1 − 𝑡) 𝑥 + 𝑡𝜉) , 0 ≤ 𝑡 ≤ 1.

We compute

d𝑚 𝜙♭ ∑︁
(𝑥, 𝑡) = (𝑥 − 𝜉) 𝛼 𝜙 ( 𝛼) ((1 − 𝑡) 𝑥 + 𝑡𝜉) ,
d𝑡 𝑚
| 𝛼 |=𝑚
534 15 Division of Distributions by Analytic Functions

where 𝛼 ∈ Z+𝑛−𝑞−1 , whence

d𝑚 𝜙♭ ∑︁
max (𝑥, 𝑡) ≤ dist (𝑥, 𝜕Γ) 𝑚
𝜙 ( 𝛼) . (15.2.18)
0≤𝑡 ≤1 d𝑡 𝑚 𝐿 ∞ (Γ)
| 𝛼 |=𝑚

We observe that 𝜙♭ (𝑥, 𝑡) vanishes identically in an interval [1 − 𝜀, 1], 𝜀 > 0; as a

consequence, if 𝑚 ≥ 1 then
∫ 1
d𝑚−1 𝜙♭ d𝑚 𝜙♭
(𝑥, 0) = − (𝑥, 𝑠) d𝑠
d𝑡 𝑚−1 𝑡 d𝑠 𝑚
and therefore, by descending induction on 𝑚,

d𝑚 𝜙♭
|𝜙 (𝑥)| = 𝜙♭ (𝑥, 0) ≤ max (𝑥, 𝑡) .
0≤𝑡 ≤1 d𝑡 𝑚

Combining this with (15.2.18) yields (15.2.17). □

Proposition 15.2.17 Assume that the map (15.2.13) is of class C𝜔 in Γ and has a
Hölder continuous extension to Γ. For 𝑤 (𝑥) 𝛿Λ ∈ D ′ Γ × R𝑞+1 to be extendible
to R𝑛 it is necessary and sufficient that 𝑤 ∈ D ′ (Γ) be extendible to R𝑛−𝑞−1 . If 𝑤 is
extendible then 𝑤 (𝑥) 𝛿Λ has an extension 𝑣 ∈ D ′ (R𝑛 ) such that supp 𝑣 ⊂ Λ.

Proof We have, for arbitrary 𝜙 ∈ Cc∞ (Γ), 𝜓 ∈ C R𝑞+1 ,

∫
⟨𝑤 (𝑥) 𝛿Λ , 𝜙 (𝑥) 𝜓 (𝑦)⟩ = ⟨𝑤, 𝜙⟩ 𝜓 (𝜌 (𝑥)) d𝑥. (15.2.19)
Γ

Putting 𝜓 ≡ 1 in (15.2.19) shows that the condition is necessary.

Conversely, suppose that 𝑤 has an extension 𝑤˜ ∈ D ′ R𝑛−𝑞−1 ; by Proposition
15.2.4 we can assume that supp 𝑤˜ ⊂ Γ. Suppose that 𝑤˜ = D𝛾 𝑓 , 𝑓 = 𝐿 2 (Γ),
𝛾 ∈ Z+𝑛−𝑘−1 . We are going to exploit (15.2.16); we may limit our attention to test-
functions 𝜑 ∈ Cc∞ Γ × R𝑞+1 whose support is contained in neighborhood of Λ in
Γ × R𝑞+1 ; in particular, we may select 𝑅 > 0 such that

(𝑥, 𝑦) ∈ Λ =⇒ sup 𝜌 𝑗 (𝑥) < 𝑅

𝑗=1,...,𝑞+1

and then assume that (𝑥, 𝑦) ∈ supp 𝜑 =⇒ sup 𝑦 𝑗 < 𝑅. We have

𝑗=1,...,𝑞+1
∫
⟨𝑤 (𝑥) 𝛿Λ , 𝜑⟩ = (−1) |𝛾 |
𝛾
𝑓 (𝑥) D 𝑥 (𝜑 (𝑥, 𝜌 (𝑥))) d𝑥 (15.2.20)
Γ
∑︁ ∫
= (−1) |𝛾 |
𝛽
𝑓 (𝑥) Π𝜌,𝛽 (𝑥) D 𝑥𝛼 D 𝑦 𝜑 (𝑥, 𝜌 (𝑥)) d𝑥,
𝛼,𝛽 Γ
15.2 Division of Distributions by Analytic Functions 535

where the Π𝜌,𝛽 (𝑥) are polynomials in the partial derivatives of 𝜌 that arise when
we apply repeatedly the chain rule (i.e., according to the formulas in [Faà di Bruno,
1855]; see also [Johnson, 2002]). On the one hand, from the Hölder continuity of
the map 𝜌 we deduce that there are 𝑚 ∈ Z+ and 𝐶 > 0 such that

∀𝑥 ∈ Γ, Π𝜌,𝛽 (𝑥) ≤ 𝐶 dist (𝑥, 𝜕Γ) −𝑚 . (15.2.21)

On the other hand,

∫ 𝜌1 ( 𝑥) ∫ 𝜌𝑞+1 ( 𝑥)
𝛽
D 𝑥𝛼 D 𝑦 𝜑 (𝑥, 𝜌 (𝑥)) = ··· Φ (𝑥, 𝑦) d𝑦 1 · · · d𝑦 𝑞+1 ,
−∞ −∞

where
𝛽 𝜕 𝑞+1 𝜑
Φ (𝑥, 𝑦) = D 𝑥𝛼 D 𝑦 (𝑥, 𝑦) .
𝜕𝑦 1 · · · 𝜕𝑦 𝑞+1
We obtain

𝛽
D 𝑥𝛼 D 𝑦 𝜑 (𝑥, 𝜌 (𝑥)) ≤ 𝑅 𝑞+1 max |Φ (𝑥, 𝑦)| . (15.2.22)
𝑦 ∈R𝑞+1

Lemma 15.2.16 implies

∑︁ ′
𝛽
D 𝑥𝛼 D 𝑦 𝜑 (𝑥, 𝜌 (𝑥)) ≤ 𝑅 𝑞+1 dist (𝑥, 𝜕Γ) 𝑚 D 𝑥𝛼 Φ (𝑥, 𝑦) .
𝐿∞
𝛼′ ∈Z+
𝑛−𝑞−1

| 𝛼′ |=𝑚

Combining this with (15.2.21) we obtain

∑︁
𝛽
|⟨𝑤 (𝑥) 𝛿Λ , 𝜑⟩| ≤ 𝐶1 D 𝑥𝛼 D 𝑦 𝜑 (𝑥, 𝑦)
𝐿∞
𝛼,𝛽

where the summations ranges over a finite set of multi-indices 𝛼 ∈ Z+𝑛−𝑞−1 , 𝛽 ∈ Z+𝑞+1 .
If we apply the Sobolev inequalities (cf. Proposition 2.2.7) we reach the conclusion
that there are numbers 𝑀 ∈ Z+ , 𝐶 > 0 such that

∀𝜑 ∈ Cc∞ Γ × R𝑞+1 , |⟨𝑤 (𝑥) 𝛿Λ , 𝜑⟩| ≤ 𝐶1 ∥𝜑∥ 𝑀 ,

where ∥·∥ 𝑀 is the

norm in the Sobolev space 𝐻 (R ). This proves that 𝑤 (𝑥) 𝛿Λ ∈
𝑀 𝑛

𝐻 −𝑀 Γ×R 𝑞+1 , thereby proving that 𝑤 (𝑥) 𝛿Λ is extendible, by Proposition 15.2.4.

The last part in the statement is a direct consequence of (15.2.20). □
Note that the change of variable 𝑦 − 𝜌 (𝑥) ⇝ 𝑦 transforms Λ into Γ × {0};
after this transformation the meaning of 𝑤 (𝑥) 𝜕𝑦𝛼 𝛿Λ is the standard tensor product

𝑤 (𝑥) ⊗ 𝛿 ( 𝛼) (𝑦) ∈ D ′ (Γ𝑥 ) ⊗ D ′ R𝑞+1
𝑦 , where now 𝛿 (𝑦) is the Dirac measure in
𝑦-space R𝑞+1 .
536 15 Division of Distributions by Analytic Functions

Recall that 𝑢 ∈ D ′ Γ × R𝑞+1 is said to be of finite order if 𝑢 ∈ 𝐻loc

𝑠
Γ × R𝑞+1
for some 𝑠 ∈ R (cf. Section 2.2).

Proposition 15.2.18 Assume that the map (15.2.13) is of class C 𝜔 in Γ and has
′

a Hölder continuous extension to Γ. If 𝑢 ∈ D Γ × R 𝑞+1 is of finite order and
supp 𝑢 ⊂ Λ then the following properties hold:

(i) There exist an integer 𝑚 ∈ Z+ and for each 𝛼 ∈ Z+𝑞+1 a distribution 𝑢 𝛼 (𝑥) of
finite order in Γ such that
∑︁
𝑢 (𝑥, 𝑦) = 𝑢 𝛼 (𝑥) 𝜕𝑦𝛼 𝛿Λ . (15.2.23)
| 𝛼 | ≤𝑚

(ii) If 𝑢 is extendible to R𝑛 then, for each 𝛼 ∈ Z+𝑞+1 , |𝛼| ≤ 𝑚, we can select 𝑢 𝛼 to

have an extension 𝑢˜ 𝛼 to R𝑛−𝑞−1 with supp 𝑢˜ 𝛼 ⊂ Γ. Conversely, if every 𝑢 𝛼 is
extendible to R𝑛−𝑞−1 then 𝑢 is extendible to R𝑛 .

Proof (i) As shown above there is no loss of generality in assuming that Λ = Γ × {0},
in which case (15.2.23) reads
∑︁
𝑢 (𝑥, 𝑦) = 𝑢 𝛼 (𝑥) ⊗ 𝛿 ( 𝛼) (𝑦) . (15.2.24)
| 𝛼 | ≤𝑚

The proof of (15.2.24) is classical; we sketch it here.

′ Γ × R𝑞+1 as hypothesized and for arbitrary 𝜙 ∈ C ∞ (Γ),

∫ With 𝑢 ∈ D c
𝑢 (𝑥, 𝑦) 𝜙 (𝑥) d𝑥 is a distribution in 𝑦-space R𝑞+1 whose support is the origin
(the integral standing for the duality bracket in 𝑥-space). As indicated in the previous
section, we have ∫ ∑︁
𝑢 (𝑥, 𝑦) 𝜙 (𝑥) d𝑥 = 𝑐 𝛼 𝛿 ( 𝛼) (𝑦) (15.2.25)
| 𝛼 | ≤𝑚

for some 𝑚 ∈ Z+ . Since 𝛿 ( 𝛼) (𝑦) , 𝑦 𝛽 = 0 if 𝛽 ≠ 𝛼 and 𝛿 ( 𝛼) (𝑦) , 𝑦 𝛼 = (−1) | 𝛼 | 𝛼!

we derive:
(−1) | 𝛼 |
∫
𝑐𝛼 = ⟨𝑢 (𝑥, 𝑦) , 𝑦 𝛼 ⟩ 𝜙 (𝑥) d𝑥,
𝛼!
meaning that we can take

(−1) | 𝛼 |
𝑢 𝛼 (𝑥) = ⟨𝑢 (𝑥, 𝑦) , 𝑦 𝛼 ⟩ . (15.2.26)
𝛼!
It is readily checked that the distributions 𝑢 𝛼 are of finite order when 𝑢 is of finite
order.
(ii) If 𝑢 is extendible it has an extension 𝑢˜ in R𝑛 whose support is contained in a
compact subset 𝐾 of Γ × R𝑞+1 containing Λ (Proposition 15.2.4). Then, by (15.2.21),

(−1) | 𝛼 |
𝑢˜ 𝛼 (𝑥) = ⟨𝑢˜ (𝑥, 𝑦) , 𝑦 𝛼 ⟩
𝛼!
15.2 Division of Distributions by Analytic Functions 537

is an extension of 𝑢˜ 𝛼 to R𝑛−𝑞−1 such that supp 𝑢˜ 𝛼 ⊂ Γ. If every 𝑢 𝛼 is extendible then

the same is true of every 𝑢 𝛼 (𝑥) 𝜕𝑦𝛼 𝛿Λ = 𝜕𝑦𝛼 (𝑢 𝛼 (𝑥) 𝛿Λ ), according to Proposition
15.2.17. □

Corollary 15.2.19 Every distribution 𝑢 ∈ D ′ Γ × R𝑞+1 such that supp 𝑢 ⊂ Λ is of

the form ∑︁
𝑢 (𝑥, 𝑦) = 𝑢 𝛼 (𝑥) ⊗ 𝜕𝑦𝛼 𝛿Λ , (15.2.27)
𝑞+1
𝛼∈Z+

where only finitely many 𝑢 𝛼 ∈ D ′ (Γ) do not vanish identically in any given open
set Γ ′ ⊂⊂ Γ.

We return to the open subset U of Γ × R𝑞+1 , Λ ⊂ U.

Proposition 15.2.20 Assume that the map (15.2.13) is of class C 𝜔 in Γ and

has a Hölder continuous extension to Γ. Let 𝑓 ∈ C 𝜔 (R𝑛 ) be such that Λ =
{𝑥 ∈ U; 𝑓 (𝑥) = 0} and 𝑢 ∈ D ′ (R𝑛 ) be given by (15.2.23) in U. Under these hy-
potheses there is a distribution 𝑣 ∈ D ′ (U) such that 𝑓 𝑣 = 𝑢 in U and supp 𝑣 ⊂ Λ.
If 𝑢 is extendible to R𝑛 the same is true of 𝑣.

Proof We can carry out the change of variables (𝑥, 𝑦) ↦→ (𝑥, 𝑦 + 𝜌 (𝑥)) in U that
transforms Γ into an open subset of the plane 𝑦 = 0 and gives 𝑢 ∈ D ′ (U) the
expression (15.2.24); we rewrite Formula (15.2.25) as
∫

𝑢 (𝑥, 𝑦) 𝜙 (𝑥) d𝑥 = 𝑃 𝜙, D 𝑦 𝛿 (𝑦) . (15.2.28)

The coefficients of the differential operator 𝑃 𝜙, D 𝑦 are continuous linear func-
tionals of 𝜙 ∈ Cc∞ (Γ). Lemma 15.2.11 (with 𝑦 1 , ..., 𝑦 𝑘+1 replacing 𝑥 1 , ..., 𝑥 𝑛 ) and
Remark 15.2.12 allow us to conclude that

𝑃 𝜙, D 𝑦 𝛿 (𝑦) = 𝑓 (𝑥) 𝑄 𝜙, D 𝑦 𝛿 (𝑦) ,

with the coefficients of 𝑄 depending linearly on those of 𝑃; it ensues that

∑︁
𝑣 𝛽 , 𝜙 D 𝑦 , 𝑣 𝛽 ∈ D ′ (Γ) ,
𝛽
𝑄 𝜙, D 𝑦 =
|𝛽 | ≤𝑚1

and that ∑︁
𝛽
𝑣 (𝑥, 𝑦) = 𝑣 𝛽 (𝑥) ⊗ D 𝑦 𝛿Λ
|𝛽 | ≤𝑚1

satisfies the equation 𝑓 𝑣 = 𝑢 in U. Reverting to the original coordinates 𝑦 𝑗 yields

∑︁
𝛽
𝑣 (𝑥, 𝑦) = 𝑣 𝛽 (𝑥) D 𝑦 𝛿Λ . (15.2.29)
|𝛽 | ≤𝑚1
538 15 Division of Distributions by Analytic Functions

Now suppose 𝑢 is extendible; we apply Proposition 15.2.18: in (15.2.23) we can

select the 𝑢 𝛼 to have an extension 𝑢˜ 𝛼 to R𝑛−𝑞−1 with supp 𝑢˜ 𝛼 ⊂ Γ. Since the 𝑣 𝛽
are linear combinations of the 𝑢 𝛼 they also have an extension 𝑣˜ 𝛽 to R𝑛−𝑞−1 with
supp 𝑣˜ 𝛽 ⊂ Γ. By Proposition 15.2.17 𝑣 itself is extendible. □

15.2.5 Regularly separated sets

Let X be a metric (topological) space; we use the distance function (𝑥, 𝑦) ↦→

dist (𝑥, 𝑦). If 𝐸 is a subset of X then 𝐸 denotes its closure in X; we refer to 𝜕𝐸 = 𝐸\𝐸
as the boundary of 𝐸. The property that 𝜕𝐸 is closed is equivalent to the property
that 𝐸 is locally closed, i.e., that every point 𝑥 ∈ 𝐸 is the center of an open ball
𝔅𝑟 (𝑥) in X (𝑟 > 0) such that 𝐸 ∩ 𝔅𝑟 (𝑥) is closed.
As usual, if 𝐴 and 𝐵 are subsets of X,

dist ( 𝐴, 𝐵) = inf dist (𝑥, 𝑦)

𝑥 ∈ 𝐴,𝑦 ∈𝐵

and dist (𝑥, 𝐵) = dist ({𝑥} , 𝐵) if 𝑥 ∈ X.

Definition 15.2.21 Let 𝐴, 𝐵 be nonempty subsets of X and Σ ≠ ∅ a closed subset

of X. We shall say that 𝐴 and 𝐵 are regularly separated by Σ if there is a covering
of X by open sets 𝑈 having the following property:
• There is a 𝜅 ≥ 1 such that

∀𝑥 ∈ 𝐴 ∩ 𝑈, dist (𝑥, 𝐵 ∩ 𝑈) ≳ dist (𝑥, Σ) 𝜅 , (15.2.30)

∀𝑦 ∈ 𝐵 ∩ 𝑈, dist (𝑦, 𝐴 ∩ 𝑈) ≳ dist (𝑦, Σ) 𝜅 . (15.2.31)

We shall say that 𝐴 is regularly separated from 𝐵 if 𝜕 𝐴 is closed and 𝐴 and 𝐵

are regularly separated by 𝜕 𝐴.

Remark 15.2.22 Actually, (15.2.30) implies (15.2.31) with, possibly, a shrunken 𝑈.

Indeed, assume that (15.2.30) holds; let the open set 𝑉 ⊂ 𝑈 be such that to every
𝑦 ∈ 𝐵 ∩ 𝑉 and every 𝜀 > 0 there is an 𝑥 𝜀 ∈ 𝐴 ∩ 𝑈 such that

dist (𝑥 𝜀 , 𝑦) = dist (𝑦, 𝐴) + 𝜀.

Then we have, for some 𝐶 > 0,

dist (𝑦, Σ) ≤ dist (𝑥 𝜀 , 𝑦) + dist (𝑥 𝜀 , Σ)

−1
≤ dist (𝑥 𝜀 , 𝑦) + 𝐶 dist (𝑥 𝜀 , 𝐵) 𝜅
−1 −1
≤ (1 + 𝐶) dist (𝑥 𝜀 , 𝑦) 𝜅 = (1 + 𝐶) (dist (𝑦, 𝐴) + 𝜀) 𝜅 .

The claim ensues by letting 𝜀 go to zero.

15.2 Division of Distributions by Analytic Functions 539

Remark 15.2.23 If 𝐴 and 𝐵 are regularly separated by Σ and 𝐵 ⊂ 𝐴 (e.g., when

𝐵 = 𝜕 𝐴) then dist (𝑦, 𝐴) = dist (𝑦, Σ) = 0 for all 𝑦 ∈ 𝐵 and 𝐵 ⊂ Σ: both (15.2.30)
and (15.2.31) are trivially true (with 𝜅 = 1).

Remark 15.2.24 If dist ( 𝐴, 𝐵) > 0 (e.g., when 𝐴 and 𝐵 are compact and 𝐴 ∩ 𝐵 = ∅)
then 𝐴 and 𝐵 are regularly separated by any closed subset of X.

Proposition 15.2.25 If the subsets 𝐴 and 𝐵 of X are regularly separated by the

closed set Σ the same is true of their closures, 𝐴 and 𝐵.

This is self-evident.

Remark 15.2.26 Two subsets 𝐴 and 𝐵 of X can be regularly separated and yet
intersect. Example: the closed half-lines [0, +∞) and (−∞, 0] in R.

Proposition 15.2.27 If the subsets 𝐴 and 𝐵 of X are regularly separated by the

closed set Σ the same is true of any pair of subsets, 𝐴 ′ ⊂ 𝐴 and 𝐵 ′ ⊂ 𝐵; it is also
true if Σ is replaced by a closed set Σ ′ ⊃ Σ.

Proof Suppose (15.2.30) holds; then dist (𝑥, 𝐵 ′) ≥ dist (𝑥, 𝐵) for every 𝑥 ∈ 𝐴 ∩ 𝑈,
a fortiori for every 𝑥 ∈ 𝐴 ′ ∩ 𝑈. Likewise for (15.2.31). If Σ ⊂ Σ ′ then dist (𝑥, Σ) ≥
dist (𝑥, Σ ′). □

Proposition 15.2.28 If 𝐴 and 𝐵 are regularly separated by Σ then 𝐴 ∩ 𝐵 ⊂ Σ.

Proof Indeed, 𝑥 ∈ 𝐴 ∩ 𝐵 implies dist (𝑥, 𝐵) = 0 and then Condition (15.2.30)

demands dist (𝑥, Σ) = 0; the conclusion ensues since Σ is closed. □
In the sequel we are going to speak of regularly separated subsets of a C 𝜔 manifold
M although, strictly speaking, M is not a metric space. The reader may assume that
M has been equipped with a C 𝜔 Riemannian metric, for instance that inherited from
an ambient Euclidean space in which M (which is countable at infinity) has been
embedded; or use a covering of M by the domains of local embeddings.
If 𝑽 is a proper analytic subvariety of M the regular part ℜ (𝑽) is regularly sepa-
rated from 𝔖 (𝑽) (Definition 15.2.21), simply because 𝔖 (𝑽) = 𝜕ℜ (𝑽) (Proposition
14.2.8, Remark 15.2.23).
In the case X = R𝑛 we will need the following later.

Lemma 15.2.29 If the bounded subsets of R𝑛 , 𝐴 and 𝐵, are regularly separated by

the compact subset Σ then there exists a function 𝜓 ∈ C ∞ (R𝑛 \Σ) with the following
properties:
(1) 0 ≤ 𝜓 (𝑥) ≤ 1 for every 𝑥 ∈ R𝑛 \Σ; 𝜓 ≡ 1 in an open set containing 𝐴\ ( 𝐴 ∩ Σ);
𝜓 ≡ 0 in an open set containing 𝐵\ (𝐵 ∩ Σ);
(2) there are numbers 𝜅 ≥ 1 and 𝐶 𝛼 > 0 to each 𝛼 ∈ Z+𝑛 such that

∀𝑥 ∈ R𝑛 \Σ, |𝐷 𝛼 𝜓 (𝑥)| ≤ 𝐶 𝛼 dist (𝑥, Σ) −𝜅 | 𝛼 | .

540 15 Division of Distributions by Analytic Functions

Proof Possibly after a translation in R𝑛 we may assume that 0 ∈ Σ. Let 𝑅 > 1

be such that Σ ∪ 𝐴 ∪ 𝐵 ⊂ 𝔅𝑅 = {𝑥 ∈ R𝑛 ; |𝑥| < 𝑅} and set Ω = 𝔅2𝑅 \Σ; we have
𝜕Ω = Ω\Ω = Σ ∪ 𝜕𝔅2𝑅 and therefore

∀𝑥 ∈ 𝔅𝑅 \Σ, dist (𝑥, 𝜕Ω) = dist (𝑥, Σ) . (15.2.32)

We introduce the partition of unity 𝜑 (℘,𝑞) constructed in the Appendix to this chapter
[see (15.A.21)]. In the Appendix we have attached to each point ℘ ∈ Zodd𝑛 the “cube”

(15.A.10),
(𝑞)
𝔔♯ (℘) = 𝑥 ∈ R𝑛 ; 𝑥 − 2−𝑞−1 ℘ < 3 × 2−𝑞−2 ,

(𝑞)
and focused on the family 𝔉♯ (Ω, 𝜅, 𝑀) of those cubes 𝔔♯ (℘) ⊂ Ω that satisfy the
inequalities (15.A.12):
𝜅
(𝑞) (𝑞) (𝑞)
𝑀 diam 𝔔♯ (℘) ≤ dist 𝔔♯ (℘) , Σ < 22𝜅+1 𝑀 diam 𝔔♯ (℘) . (15.2.33)

(𝜅 ≥ 1, 𝑀 ≥ 2, submitted to all the requirements in the Appendix). We shall assume

that (15.2.30)–(15.2.31) hold.
(𝑞) (𝑞)
Let 𝔔♯ (℘) ∈ 𝔉♯ (Ω, 𝜅, 𝑀) be such that there is an 𝑥 ∈ 𝔔♯ (℘) ∩ ( 𝐴\Σ); we
have, by (15.2.30) and (15.2.33),

(𝑞)
dist 𝔔♯ (℘) , 𝐵 ≥ dist (𝑥, 𝐵) − 2−𝑞−1
≥ 𝑐 dist (𝑥, Σ) 𝜅 − 2−𝑞−1
(𝑞)
≥ 𝑐𝑀 diam 𝔔♯ (℘) − 2−𝑞−1 .

It suffice to require 𝑀 ≥ 𝑐−1 to obtain

(𝑞)
dist 𝔔♯ (℘) , 𝐵 ≥ 2−𝑞−1 . (15.2.34)

𝑛 × Z such that 𝔔 (𝑞) (℘) ∈ 𝔉 (Ω, 𝜅, 𝑀) and

Let 𝑆 be the set of pairs (℘, 𝑞) ∈ Zodd + ♯ ♯
𝔔 (𝑞) (℘) ∩ ( 𝐴\Σ) ≠ ∅. The function
∑︁
𝜓 (𝑥) = 𝜑 ℘,𝑞 (𝑥)
(℘,𝑞) ∈𝑆

is well defined in R𝑛 \Σ = (R𝑛 \𝔅𝑅 ) ∪ Ω and has all the required properties, as a
consequence of (15.2.34) and of the properties of the functions 𝜑 ℘,𝑞 established in
the Appendix, in particular (15.A.23):
∑︁
∀𝑥 ∈ R𝑛 , 𝜕𝑥𝛼 𝜑 ℘,𝑞 (𝑥) ≤ 𝐶 𝛼′′′ (𝜃) dist (𝑥, Σ) −𝜅 | 𝛼 | . □
(℘,𝑞) ∈𝑆
15.2 Division of Distributions by Analytic Functions 541

15.2.6 Regular separation and extendibility of distributions

Proposition 15.2.30 If two compact subsets of R𝑛 , 𝐾1 and 𝐾2 , are regularly sep-

arated by 𝐾1 ∩ 𝐾2 then every distribution 𝑢 in R𝑛 with supp 𝑢 ⊂ 𝐾1 ∪ 𝐾2 can be
decomposed as a sum 𝑢 1 + 𝑢 2 with 𝑢 𝑗 ∈ D ′ (R𝑛 ) and supp 𝑢 𝑗 ⊂ 𝐾 𝑗 , 𝑗 = 1, 2.
Proof Let 𝜓 ∈ C ∞ (R𝑛 \Σ) be the function in Lemma 15.2.29 in which we take
𝐴 = 𝐾1 , 𝐵 = 𝐾2 and Σ = 𝐾1 ∩ 𝐾2 . According to Proposition 15.2.8 the distribution
𝜓𝑢 in R𝑛 \ (𝐾1 ∩ 𝐾2 ) is extendible. Let 𝑢 1 be an extension of 𝜓𝑢 to R𝑛 . According to
Lemma 15.2.29 there are open subsets of R𝑛 \ (𝐾1 ∩ 𝐾2 ), 𝑈1 ⊃ 𝐾1 \ (𝐾1 ∩ 𝐾2 ) and
𝑈2 ⊃ 𝐾2 \ (𝐾1 ∩ 𝐾2 ), such that 𝑢 1 = 𝑢 in 𝑈1 and 𝑢 1 ≡ 0 in 𝑈2 . From this and from
Proposition 15.2.4 it ensues that supp 𝑢 1 is contained in the closure of 𝐾1 \ (𝐾1 ∩ 𝐾2 )
hence in 𝐾1 . Repeating the same argument with 1 − 𝜓 substituted for 𝜓 yields an
extension 𝑢 2 of (1 − 𝜓) 𝑢 to R𝑛 such that 𝑢 2 = 𝑢 in an open subset of R𝑛 \ (𝐾1 ∩ 𝐾2 )
containing 𝐾2 \ (𝐾1 ∩ 𝐾2 ) and supp 𝑢 2 ⊂ 𝐾2 . □
Remark 15.2.31 The converse of Proposition 15.2.30 is true: if every distribution 𝑢
in R𝑛 with supp 𝑢 ⊂ 𝐾1 ∪ 𝐾2 can be decomposed as a sum 𝑢 1 +𝑢 2 with 𝑢 𝑗 ∈ D ′ (R𝑛 )
and supp 𝑢 𝑗 ⊂ 𝐾 𝑗 ( 𝑗 = 1, 2) then 𝐾1 and 𝐾2 are regularly separated by 𝐾1 ∩ 𝐾2 . For
a proof see [Lojasiewicz, 1959], pp. 97–98.
Proposition 15.2.32 Let Ω1 , Ω2 be bounded open subsets of R𝑛 and let Ω = Ω1 ∪Ω2 .
If the boundaries 𝜕Ω1 and 𝜕Ω2 are regularly separated by 𝜕Ω then every distribution
𝑢 in Ω whose restrictions to Ω1 and Ω2 are extendible is itself extendible.
Proof It is evident by Definition 15.2.21 that if 𝜕Ω1 and 𝜕Ω2 are regularly separated
by 𝜕Ω then the same is true of 𝐾1 = 𝜕Ω ∪ 𝜕Ω1 and 𝐾2 = 𝜕Ω ∪ 𝜕Ω2 . Let 𝑢 𝑗 be
an extension to R𝑛 of the restrictions to Ω 𝑗 of 𝑢 ∈ D ′ (Ω) such that supp 𝑢 𝑗 ⊂ Ω 𝑗
( 𝑗 = 1, 2). Every restriction of 𝑢 to an open subset 𝑈 𝑗 of Ω 𝑗 has an extension
whose support is contained
in 𝑈 𝑗 (Proposition
15.2.5). Thus the restriction of 𝑢 to
Ω1 \ Ω1 ∩ Ω2 = Ω1 \ Ω1 ∩ Ω1 ∩ Ω2 admits an extension 𝑤 1 to R𝑛 with

supp 𝑤 1 ⊂ Ω1 \ Ω1 ∩ Ω2 ;

similarly, the restriction of 𝑢 to Ω2 \ Ω1 ∩ Ω2 admits an extension 𝑤 2 to R𝑛 with

supp 𝑤 2 ⊂ Ω2 \ Ω1 ∩ Ω2 ;

the restriction of 𝑢 to Ω1 ∩Ω2 has an extension 𝑤 1,2 to R𝑛 with supp 𝑤 1,2 ⊂ Ω1 ∩ Ω2 .

We have 𝑢 1 = 𝑤 1 = 𝑢 and 𝑤 1,2 = 0 in Ω1 \Ω1 ∩ Ω2 , 𝑢 1 = 𝑤 1,2 = 𝑢 in Ω1 . As a
consequence,

supp 𝑢 1 − 𝑤 1 − 𝑤 1,2 ⊂ Ω1 \ Ω1 \Ω1 ∩ Ω2 ∪ (Ω1 ∩ Ω2 )
⊂ 𝜕Ω1 ∪ 𝜕Ω2 ⊂ 𝐾1 ∪ 𝐾2 ;
542 15 Division of Distributions by Analytic Functions

likewise,

supp 𝑢 2 − 𝑤 2 − 𝑤 1,2 ⊂ 𝜕Ω2 ∪ 𝜕 (Ω2 ∩ Ω1 ) ⊂ 𝐾1 ∪ 𝐾2 .

We apply Proposition 15.2.30: 𝑢 1 − 𝑤 1 − 𝑤 1,2 = 𝑢 1,1 + 𝑢 1,2 , 𝑢 2 − 𝑤 2 − 𝑤 1,2 =

𝑢 2,1 + 𝑢 2,2 , all distributions being defined in R𝑛 and supp 𝑢 𝑖, 𝑗 ⊂ 𝐾 𝑗 . We note that
Ω1 ∩ 𝐾1 = ∅; it follows that 𝑢 1,1 = 𝑢 2,1 = 0 in Ω1 whence 𝑤 1 + 𝑤 1,2 + 𝑢 1,2 = 𝑢 1 = 𝑢
in Ω1 . We also have 𝑤 2 = 0 in Ω1 since

Ω1 ∩ Ω2 \ Ω1 ∩ Ω2 = ∅.

As a consequence of all this the restriction to Ω1 of

𝑓 = 𝑤 1 + 𝑤 2 + 𝑤 1,2 + 𝑢 1,2 + 𝑢 2,1

is equal to that of 𝑢; likewise, the restriction of 𝑓 to Ω2 is equal to that of 𝑢. We

conclude that the restriction of 𝑓 to Ω is equal to 𝑢, which proves the claim since
𝑓 ∈ D ′ (R𝑛 ). □

Remark 15.2.33 The separability of the boundaries 𝜕Ω1 and 𝜕Ω2 is necessary to
reach the conclusion in Proposition 15.2.32 (see [Lojasiewicz, 1959]).

15.2.7 Regular separation in analytic stratifications

We go back to Section 14.5, whose notation, such as 𝑥 ′ = (𝑥 1 , ..., 𝑥 𝑛−1 ), 𝑥 ( 𝛼) =

(𝑥1 , ..., 𝑥 𝑛−𝛼 ), 𝜋 𝛼 : 𝑥 ↦→ 𝑥 ( 𝛼) , we follow. We focus on the stratification (14.5.13) of
the nullset 𝑽 of the trimmed Weierstrass polynomial

𝑓 (𝑥) = 𝑃 (𝑥 ′; 𝑥 𝑛 ) = 𝑃1 (𝑥 ′; 𝑥 𝑛 ) · · · 𝑃 𝜈 (𝑥 ′, 𝑥 𝑛 )

where the 𝑃 𝑗 extend as distinct, irreducible Weierstrass polynomials 𝑃 𝑗 (𝑧 ′; 𝑧 𝑛 ) to

a neighborhood of 𝔔𝑟(𝑛) = 𝑥 ∈ R𝑛 ; 𝑥 𝑗 ≤ 𝑟 𝑗 , 1 ≤ 𝑗 ≤ 𝑛 in C𝑛 . We adapt the

definition (14.5.12): for each 𝛼 = 1, ..., 𝛼∗ ≤ 𝑛, we introduce the analytic subvariety

of 𝔔𝑟(𝑛)
𝛼∗ Ø 𝑁 [𝛽 ]
Ø Ø
[ 𝛼] [𝛽 ]
𝑽 = Λ 𝜄, 𝑝 .
𝛽=𝛼 𝜄 ∈𝑰 [𝛽 ] 𝑝=1

Keep in mind that 𝑽 [1] = 𝑽 ∩ 𝔔𝑟(𝑛) = 𝑽 [𝛼1 ] with 𝛼1 = 𝑛 − 𝑑, 𝑑 = dim 𝑽 [1] . We

return to the Weierstrass polynomials 𝑃 [𝑗 𝛼] 𝑥 ( 𝛼) ; 𝑥 𝑛− 𝑗+1 ( 𝑗 = 1, ..., 𝛼) and to the

product of discriminants 𝐷 [ 𝛼] 𝑥 ( 𝛼) whose nullset is 𝒁 [ 𝛼] [see (14.5.1), (14.5.2)].
15.2 Division of Distributions by Analytic Functions 543

We focus on a stratum Λ [𝜄,𝛼]
𝑝 (𝛼1 ≤ 𝛼 ≤ 𝛼∗ , 𝜄 ∈ 𝑰
[ 𝛼]
); its projection 𝜋 𝛼 Λ [𝜄,𝛼]
𝑝

is a connected component Γ 𝜄[ 𝛼] of 𝔔𝑟(𝑛−𝛼)( 𝛼) \𝒁 [ 𝛼] ; Λ [𝜄,𝛼] (𝑛)

𝑝 ⊂ 𝔔𝑟 is the graph in

𝔔𝑟 ( 𝛼) ×R 𝛼 of a map (14.5.14), whose components are simple real roots 𝜌 [𝑗,ℓ
(𝑛−𝛼) 𝛼]
𝑗
𝑥 ( 𝛼)

of the Weierstrass polynomials 𝑃 [𝑗 𝛼] 𝑥 ( 𝛼) ; 𝑥 𝑛− 𝑗+1 (for each 𝑗 = 0, ..., 𝛼) in Γ 𝜄[ 𝛼] .

(𝑛)
Proposition 15.2.34 In 𝔔𝑟/2 an arbitrary stratum Λ [𝜄,𝛼]
𝑝 is regularly separated from
𝑽 [ 𝛼] \Λ [𝜄,𝛼]
𝑝 (Definition 15.2.21).

Proof By the Lojasiewicz inequality (Theorem 15.1.8) there are constants 𝑐 1 > 0,
𝜅1 ≥ 1 such that 𝜅1
𝑐 1 dist 𝑥 ( 𝛼) , 𝒁 [ 𝛼] ≤ 𝐷 [ 𝛼] 𝑥 ( 𝛼)

for every 𝑥 ( 𝛼) ∈ 𝔔𝑟(𝑛−𝛼)

( 𝛼) /2 . We observe that the C
𝜔 function in Γ [ 𝛼]
𝜄

𝛼 𝜕𝑃 [ 𝛼]
Ö
𝐷 𝜌[ 𝛼] ( 𝛼)
𝑥 ( 𝛼) ; 𝜌 [𝑗,ℓ ( 𝛼)
𝑗 𝛼]
𝑥 = 𝑗
𝑥
𝑗=1
𝜕𝑥 𝑛− 𝑗+1

extends as a Hölder continuous function in Γ 𝜄[ 𝛼] (Corollary 15.1.13); 𝐷 𝜌[ 𝛼] 𝑥 ( 𝛼)
[
𝜕𝑃 𝛼]

divides 𝐷 [ 𝛼] 𝑥 ( 𝛼) since 𝜕𝑥𝑛−𝑗 𝑗+1 𝑥 ( 𝛼) ; 𝜌 [𝑗,ℓ
𝛼]
𝑗
𝑥 ( 𝛼) divides the discriminant of

[ 𝛼]
𝑃𝑗 𝑧 ( 𝛼) ; 𝑧 𝑛− 𝑗 for each 𝑗. Let 𝜕Γ 𝜄 mean the boundary of Γ 𝜄[ 𝛼] in 𝔔𝑟(𝑛−𝛼)
[ 𝛼]
( 𝛼) ; then
𝑥 ( 𝛼) ∈ Γ 𝜄[ 𝛼] ∩ 𝔔𝑟(𝑛−𝛼)
( 𝛼) /2 implies

dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] = dist 𝑥 ( 𝛼) , 𝒁 [ 𝛼] .

We derive that 𝜅1
𝑐 2 dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] ≤ 𝐷 𝜌[ 𝛼] 𝑥 ( 𝛼) (15.2.35)

for some 𝑐 2 > 0 and all 𝑥 ( 𝛼) ∈ Γ 𝜄[ 𝛼] ∩ 𝔔𝑟(𝑛−𝛼)

( 𝛼) /2 . We have

𝛼 𝜕𝑃 [ 𝛼]
Ö 𝛼
∑︁
𝐷 𝜌[ 𝛼] ( 𝛼)
𝑥 ( 𝛼) ; 𝑥 𝑛− 𝑗+1 𝑥 𝑛− 𝑗+1 − 𝜌 [𝑗,ℓ ( 𝛼)
𝑗 𝛼]
𝑥 − ≤ 𝐶1 𝑗
𝑥
𝑗=2
𝜕𝑥 𝑛− 𝑗+1 𝑗=1

−1

for some 𝐶1 > 0 and all 𝑥 ∈ 𝜋 𝛼 Γ 𝜄[ 𝛼] . It follows directly from this and from
−1

(15.2.35) that there is an 𝜀 > 0 such that if 𝑥 ∈ 𝜋 𝛼 Γ 𝜄[ 𝛼] ∩ 𝔔𝑟(𝑛−𝛼)
( 𝛼) /2 satisfies

𝛼 𝜅
[ 𝛼] 1
∑︁
𝑥 𝑛− 𝑗+1 − 𝜌 [𝑗,ℓ
𝛼]
𝑗
𝑥 ( 𝛼)
< 𝜀 dist 𝑥 ( 𝛼)
, 𝜕Γ 𝜄 (15.2.36)
𝑗=1
544 15 Division of Distributions by Analytic Functions

then

1 𝜅1 Ö 𝛼 𝜕𝑃 [𝑗 𝛼]
𝑐 2 dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] ≤ 𝑥 ( 𝛼) ; 𝑥 𝑛− 𝑗+1 . (15.2.37)
2 𝑗=1
𝜕𝑥 𝑛− 𝑗+1

−1

Let 𝜔𝜌, 𝜀 denote the open subset of 𝜋 𝛼 Γ 𝜄[ 𝛼] ∩ 𝔔𝑟/2
(𝑛)
defined by (15.2.36); we have

Λ [𝜄,𝛼] (𝑛)
𝑝 ∩ 𝔔𝑟/2 = 𝑽
[ 𝛼]
∩ 𝜔𝜌, 𝜀 . (15.2.38)

Indeed, suppose that for some 𝑥 ( 𝛼) ∈ Γ 𝜄[ 𝛼] ∩ 𝔔𝑟(𝑛−𝛼)

( 𝛼) /2 and some 𝑗 ◦ ∈ [1, ..., 𝛼]

there were a (necessarily simple) real root 𝜌 [𝑗◦𝛼],ℓ 𝑗 𝑥 ( 𝛼) ≠ 𝜌 [𝑗◦𝛼],ℓ 𝑗 𝑥 ( 𝛼) of
◦

𝑃 [𝑗◦𝛼] 𝑥 ( 𝛼) ; 𝑥 𝑛− 𝑗◦ +1 such that

𝜅1
𝜌 [𝑗◦𝛼],ℓ 𝑗 𝑥 ( 𝛼) − 𝜌 [𝑗◦𝛼],ℓ 𝑥 ( 𝛼) < 𝜀 dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] .
◦

[
𝜕𝑃 𝛼]

Then there would be a root 𝜏 𝑗◦ 𝑥 ( 𝛼) of 𝜕𝑥𝑛−𝑗◦𝑗 +1 𝑥 ( 𝛼) ; 𝑥 𝑛− 𝑗◦ +1 between 𝜌 [𝑗◦𝛼],ℓ 𝑗 𝑥 ( 𝛼)
◦
and 𝜌 [𝑗◦𝛼],ℓ 𝑗 𝑥 ( 𝛼) , contradicting (15.2.37) as one sees by putting 𝑥 𝑛− 𝑗+1 = 𝜌 [𝑗,ℓ𝛼]
𝑥 ( 𝛼)
◦ 𝑗

for every 𝑗 ≠ 𝑗 ◦ and 𝑥 𝑛− 𝑗◦ +1 = 𝜏 𝑗◦ 𝑥 ( 𝛼) .

Now let 𝑥 ∈ Λ 𝜄, 𝑝 ∩𝔔𝑟/2 and 𝑦 ∈ 𝔔𝑟/2 be such that |𝑥 − 𝑦| ≤ 21 dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] ;
[ 𝛼] (𝑛) (𝑛)

we have 1
dist 𝑦 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] ≥ dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] .
2
Provided the segment joining 𝑥 ( 𝛼) to 𝑦 ( 𝛼) is contained in Γ 𝜄[ 𝛼] we can avail ourselves
of Corollary 15.1.13 to obtain
𝛼
∑︁ 𝛼
∑︁
𝑦 𝑛− 𝑗+1 − 𝜌 [𝑗,ℓ
𝛼]
𝑗
𝑦 ( 𝛼)
≤ 𝑦 𝑛− 𝑗+1 − 𝜌 [ 𝛼]
𝑗,ℓ 𝑗 𝑥 ( 𝛼)

𝑗=0 𝑗=0
𝛼
∑︁
+ 𝜌 [𝑗,ℓ
𝛼]
𝑗
𝑥 ( 𝛼)
− 𝜌 [ 𝛼]
𝑗,ℓ 𝑗 𝑦 ( 𝛼)

𝑗=0
𝜃
≤ |𝑥 − 𝑦| + 𝐶2 𝑥 ( 𝛼) − 𝑦 ( 𝛼)

(𝐶2 > 1 and 0 < 𝜃 ≤ 1 independent of 𝑥, 𝑦). We derive

𝛼
∑︁
𝑦 𝑛− 𝑗+1 − 𝜌 [𝑗,ℓ
𝛼]
𝑗
𝑦 ( 𝛼)
≤ 𝐶3 |𝑥 − 𝑦| 𝜃
𝑗=0
15.2 Division of Distributions by Analytic Functions 545
𝜅1
for some 𝐶3 > 𝐶2 , also independent of 𝑥, 𝑦. Thus 𝐶3 |𝑥 − 𝑦| 𝜃 ≤ 𝜀 dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼]
implies 𝑦 ∈ 𝜔𝜌, 𝜀 . In other words, 𝑦 ∈ 𝜔𝜌, 𝜀 whenever
𝜅 /𝜃
1
( 𝛼) [ 𝛼]

( 𝛼) [ 𝛼] 1
|𝑥 − 𝑦| ≤ min dist 𝑥 , 𝜕Γ 𝜄 , 𝜀∗ dist 𝑥 , 𝜕Γ 𝜄
2
1/𝜃
where 𝜀∗ = 𝐶3−1 𝜀 . We reach the conclusion that there are constants 𝜀◦ > 0,
𝜅2 ≥ 1, such that
𝜅2
|𝑥 − 𝑦| < 𝜀◦ dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] =⇒ 𝑦 ∈ 𝜔𝜌, 𝜀 ,

whence, for every 𝑥 ∈ Λ [𝜄,𝛼] (𝑛)

𝑝 ∩ 𝔔𝑟/2 = 𝑽
[ 𝛼]
∩ 𝜔𝜌, 𝜀 [cf. (15.2.38)],
𝜅2
(𝑛)
dist 𝑥, 𝔔𝑟/2 \𝜔𝜌, 𝜀 ≥ 𝜀◦ dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼] ,

and therefore, since 𝑽\Λ [𝜄,𝛼] (𝑛)

𝑝 ⊂ 𝔔𝑟/2 \𝜔 𝜌, 𝜀 ,

𝜅2
dist 𝑥, 𝑽\Λ [𝜄,𝛼]
𝑝 ≥ 𝜀 ◦ dist 𝑥
( 𝛼)
, 𝜕Γ 𝜄[ 𝛼] . (15.2.39)

According to Lemma 15.1.14 there are constants 𝑐 > 0, 𝜅 ≥ 1, such that, for every
𝑥 ∈ Λ [𝜄,𝛼] (𝑛)
𝑝 ∩ 𝔔𝑟/2 ,

𝜅
𝑐 dist 𝑥, 𝜕Λ [𝜄,𝛼]
𝑝 ≤ dist 𝑥 ( 𝛼) , 𝜕Γ 𝜄[ 𝛼]

and therefore, by (15.2.39),

𝜅 𝜅
[ 𝛼] 2
dist 𝑥, 𝑽 [ 𝛼] \Λ [𝜄,𝛼]
𝑝 ≥ 𝜀 ◦ 𝑐 dist 𝑥 ( 𝛼)
, 𝜕Λ 𝜄, 𝑝 .

The proof of Proposition 15.2.34 is complete. □

(𝑛)
Corollary 15.2.35 In 𝔔𝑟/2 an arbitrary stratum Λ [𝜄,𝛼]
𝑝 is regularly separated from
𝑽\Λ [𝜄,𝛼]
𝑝.

(𝑛) (𝑛)
Proof Indeed, 𝑽 ∩ 𝔔𝑟/2 = 𝑽 [ 𝛼1 ] ∩ 𝔔𝑟/2 since 𝑽 [ 𝛼1 ] = 𝑽 ∩ 𝔔𝑟(𝑛) . □

15.2.8 End of the proof of the division theorem

Thanks to Theorem 15.2.9 it suffices to solve the division problem when supp 𝑢 ⊂
−1
𝑽 = 𝑓 (0); Theorem 15.2.13 settles the case of dim 𝑽 = 0; henceforth we suppose
𝑑 = dim 𝑽 ≥ 1 (and 𝑑 < 𝑛 = dim M).
546 15 Division of Distributions by Analytic Functions

As in the proof of Proposition 15.2.10 it suffices to deal with M = Ω, a

bounded, open subset of R𝑛 such that 0 ∈ Ω ∩ 𝑽. We shall assume that Ω contains
the closure of the multi-interval 𝔔𝑟(𝑛) = 𝑥 ∈ R𝑛 ; 𝑥 𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛 where the

stratification (14.5.13) is valid:

[ 𝛼𝜈 ]
𝑁Ø
Ø
𝑽 [ 𝛼𝜈 ]
=𝑽 [ 𝛼𝜈+1 ]
∪ Λ [𝜄,𝛼𝑝𝜈 ] (15.2.40)
𝜄 ∈𝑰 [ 𝛼𝜈 ] 𝑝=1

(𝜈 = 1, ..., 𝜈∗ ) . Recall that 𝑽 ∩ 𝔔𝑟(𝑛) = 𝑽 [ 𝛼1 ] with 𝛼1 = 𝑛 − 𝑑 [cf. (14.4.17)] and that

dim 𝑽 [ 𝛼𝜈+1 ] < dim 𝑽 [ 𝛼𝜈 ] , implying Λ [𝜄,𝛼𝑝𝜈 ] ≠ ∅ for some 𝜄 ∈ 𝑰 [ 𝛼𝜈 ] .
Consider the following statement, in which 𝛼𝜈 ∈ Z+ , 1 ≤ 𝜈 ≤ 𝜈∗ ≤ 𝑛:

(∗) 𝜈 Given any distribution 𝑢 ∈ D ′ 𝔔2(𝑛) −𝜈 𝑟 such that supp 𝑢 ⊂ 𝑽 [ 𝛼𝜈 ] there is a

𝑣 𝜈 ∈ D ′ 𝔔2(𝑛) −𝜈−1 𝑟 such that supp (𝑢 − 𝑓 𝑣 𝜈 ) ⊂ 𝑽
[ 𝛼𝜈 ]
∩ 𝔔2(𝑛)
−𝜈−1 𝑟 .

Suppose (∗) 𝜈 is proved for every 𝜈 ∈ [1, 𝜈∗ ]. Consider 𝑢 ∈ D ′ 𝔔𝑟(𝑛) with

supp 𝑢 ⊂ 𝑽 [ 𝛼𝜈 ] ; according to (∗) 1 there is a 𝑤 𝛼1 ∈ D ′ 𝔔2(𝑛) −2 𝑟 such that

supp (𝑢 − 𝑓 𝑤 1 ) ⊂ 𝑽 [ 𝛼𝜈 ] ∩𝔔2−2 𝑟 . Now (∗) 2 implies that there is a 𝑤 𝛼2 ∈ D ′ 𝔔2(𝑛)
(𝑛)
−3 𝑟

such that
supp (𝑢 − 𝑓 𝑤 1 − 𝑓 𝑤 2 ) ⊂ 𝑽 [ 𝛼𝜈 ] ∩ 𝔔2(𝑛)
−3 𝑟 .

We can repeat this procedure until 𝜈 = 𝜈∗ ≤ 𝑛; since 𝑽 [ 𝛼𝜈∗ +1] = ∅ we get 𝑢 = 𝑓 𝑣 in

𝔔2(𝑛)
Í𝜈∗
−𝑛 𝑟 with 𝑣 = 𝜈=1 𝑤 𝛼𝜈 .
We use once more the notation of Section 14.4: 𝑥 ( 𝛼𝜈 ) = 𝑥1 , ..., 𝑥 𝑛−𝛼𝜈 , 𝜋 𝛼𝜈 :

𝑥 ↦→ 𝑥 ( 𝛼𝜈 ) ; we will also use the notation 𝑦 = 𝑥 𝑛−𝛼𝜈 +1 , ..., 𝑥 𝑛 . To simplify matters,
in what follows we write 𝑟 rather than 2−𝜈 𝑟. Let Γ 𝜄[ 𝛼𝜈 ] be a connected component
of 𝔔𝑟(𝑛−𝛼
( 𝛼𝜈 )
𝜈)
\𝒁 [ 𝛼𝜈 ] and Λ [𝜄,𝛼𝑝𝜈 ] be the graph of the map (14.5.14). Corollary 15.1.13
(𝑛)
and the fact that 𝑢 is of finite order in 𝔔 (1−𝜀)𝑟 whatever 𝜀 > 0 allow us to apply
Proposition 15.2.18: there are distributions 𝑢 𝜈 in R𝑛−𝛼𝜈 with supp 𝑢 𝜈 ⊂ Γ 𝜄[ 𝛼𝜈 ] ,
satisfying ∑︁
𝑢 𝜈 𝑥 ( 𝛼𝜈 ) 𝜕𝑦 𝛿Λ [ 𝛼𝜈 ] ,
𝛾
𝑢 (𝑥) = (15.2.41)
𝜄, 𝑝
𝛾 ∈Z+𝑛 , |𝛾 | ≤𝑚

−1

in an open subset 𝑈 𝜄[ 𝛼𝜈 ] of 𝔔𝑟/2
(𝑛)
∩ 𝜋 𝛼𝜈 Γ 𝜄[ 𝛼𝜈 ] such that Λ [𝜄,𝛼𝑝𝜈 ] = 𝑽 [ 𝛼𝜈 ] ∩ 𝑈 𝜄[ 𝛼𝜈 ] ;
moreover, there is an extension 𝑢 [𝜄,𝛼𝑝𝜈 ] to R𝑛 of the right-hand side in (15.2.41) such
that supp 𝑢 [𝜄,𝛼𝑝𝜈 ] = Λ [𝜄,𝛼𝑝𝜈 ] . We apply once again Proposition 15.2.18 and now also
Proposition 15.2.20: there is a distribution 𝑣 [𝜄,𝛼𝑝𝜈 ] ∈ D ′ (R𝑛 ) satisfying 𝑓 𝑣 [𝜄,𝛼𝑝𝜈 ] =
−1

𝑢 [𝜄,𝛼𝑝𝜈 ] in 𝔔𝑟/2
(𝑛)
∩ 𝜋 𝛼𝜈 Γ 𝜄[ 𝛼𝜈 ] and such that supp 𝑣 [𝜄,𝛼𝑝𝜈 ] ⊂ Λ [𝜄,𝛼𝑝𝜈 ] . We avail ourselves
of Propositions 15.2.32 and 15.2.32 to conclude that
15.2 Division of Distributions by Analytic Functions 547
[ 𝛼𝜈 ]
𝑁∑︁
∑︁
𝑣𝜈 = 𝑣 [𝜄,𝛼𝑝𝜈 ] ∈ D ′ (R𝑛 ) .
𝜄 ∈𝑰 [ 𝛼𝜈 ] 𝑝=1

(𝑛)
Taking (15.2.40) into account shows that 𝑓 𝑣 𝜈 = 𝑢 in 𝔔𝑟/2 \𝑽 [ 𝛼𝜈 ] . This proves (∗) 𝜈
and thereby Theorem 15.2.1.

15.2.9 Application: Tempered solutions of PDEs with constant

coefficients

The solvability in tempered distributions of linear partial differential equations with

constant coefficients is a direct consequence of Theorem 15.2.1. Tempered distribu-
tions in R𝑛 are the elements of the space S ′ (R𝑛 ), the dual of the space S (R𝑛 ) of
C ∞ functions whose derivatives of all orders decay at infinity rapidly, meaning faster
than any power of |𝑥| −1 (Subsection 2.1.2). We make use of the following result, in
which S𝑛 is the 𝑛-dimensional sphere.

Proposition 15.2.36 The natural diffeomorphism of R𝑛 onto S𝑛 \ {∞} induces an

isomorphism of the space S (R𝑛 ) of rapidly decaying C ∞ functions at ∞ onto the
space of C ∞ functions in S𝑛 that vanish to infinite order at ∞, C◦∞ (S𝑛 ).

Proof We use the coordinates 𝑦 𝑖 = 𝑥 𝑖 /|𝑥| 2 , 𝑖 = 1, ..., 𝑛, in S𝑛 \ {0}. If 𝑓 (𝑥) ∈ S (R𝑛 )

to every 𝛼 ∈ Z+𝑛 there is a 𝐶 𝛼 > 0 such that

∀𝑦 ∈ R𝑛 \ {0} , 𝑦 𝛼 𝑓 𝑦/|𝑦| 2 ≤ 𝐶 𝛼 |𝑦| 2 | 𝛼 | . (15.2.42)

This proves that 𝑓 vanishes to infinite order as 𝑦 → 0, i.e., as 𝑥 −→ ∞. The same is

true of 𝐷 𝑥 𝑓 ∈ S (R𝑛 ), 𝛽 ∈ Z+𝑛 arbitrary. If 𝑓 ∈ C ∞ (S𝑛 ) vanishes to infinite order
𝛽

at ∞ the restriction of 𝑓 to R𝑛 decays faster than any power of |𝑥| −1 as |𝑥| → +∞,
and the same is true of 𝐷 𝛽 𝑓 whatever 𝛽 ∈ Z+𝑛 . This map S (R𝑛 ) −→ C ∞ (S𝑛 ) is an
injective linear map onto the closed subspace C◦∞ (S𝑛 ) of C ∞ (S𝑛 ); it is continuous,
say by the Closed Graph Theorem (all topological vector spaces involved are Fréchet
spaces) or by the inequalities (15.2.42) for all the partial derivatives of 𝑓 . □
The dual of C◦∞ (S𝑛 ) is the quotient space D ′ (S𝑛 ) /D∞ ′ (S𝑛 ), where D ′ (S𝑛 ) is
∞
the linear span of all the partial derivatives of the Dirac distribution at ∞. By trans-
position the natural isomorphism S (R𝑛 ) −→ C◦∞ (S𝑛 ) leads to the exact sequence
′
0 −→ D∞ (S𝑛 ) −→ D ′ (S𝑛 ) −→ D ′ (S𝑛 ) /D∞
′
(S𝑛 ) −→ S ′ (R𝑛 ) → 0. (15.2.43)

Corollary 15.2.37 The composite D ′ (S𝑛 ) −→ D ′ (S𝑛 ) /D∞

′ (S𝑛 ) −→ S ′ (R𝑛 ) in

(15.2.43) is a continuous linear surjection.

In the statements that follow 𝑃 ∈ C [𝜉1 , ..., 𝜉 𝑛 ], 𝑃 (𝜉) . 0, and 𝑃 (D) is the
corresponding linear partial differential operator with constant coefficients in R𝑛 .
548 15 Division of Distributions by Analytic Functions

Theorem 15.2.38 To every tempered distribution 𝑓 ∈ S ′ (R𝑛 ) there is a 𝑢 ∈ S ′ (R𝑛 )

such that 𝑃 (D) 𝑢 = 𝑓 .
−𝑘
Proof Suppose 2𝑘 ≥ deg 𝑃; the function 1 + |𝜉 | 2 𝑃 (𝜉) extends as an analytic
function on the sphere S𝑛 . We extend the Fourier transform of 𝑓 , b 𝑓 , as a distribution
in S𝑛 (which is permitted by Corollary 15.2.37). Theorem 15.2.1 implies that there
−𝑘 −𝑘
𝑢 ∈ D ′ (S𝑛 ) such that 1 + |𝜉 | 2
is a b 𝑢 = 1 + |𝜉 | 2
𝑃 (𝜉) b 𝑓 ∈ D ′ (S𝑛 ). The
b
𝑢 to R𝑛 belongs to S ′ (R𝑛 ) and satisfies 𝑃 (𝜉) b
restriction of b 𝑓 . Applying the
𝑢 = b
inverse Fourier transform yields 𝑃 (D) 𝑢 = 𝑓 . □
Corollary 15.2.39 There exists a distribution 𝐸 ∈ S ′ (R𝑛 ) such that 𝑃 (D) 𝐸 = 𝛿.
The distribution 𝐸 in Corollary 15.2.39 (first proved, by different methods, in
[Hörmander, 1958]) is a tempered fundamental solution of 𝑃 (D).

15.3 Desingularization and Applications

15.3.1 The Hironaka Theorem. Statement for C 𝝎 varieties

In this subsection we content ourselves with stating the Hironaka Theorem on the
resolution of singularities for a real-analytic variety; the C 𝜔 version suffices for
our needs in the applications that follow. The original, very difficult, proof in the
case of an arbitrary base field of characteristic zero can be found in [Hironaka,
1965]. There have been a number of simplified proofs, mainly from the viewpoint
of Algebraic Geometry. For full explanations and proofs in the complex-analytic
set-up see [Aroca-Hironaka-Vicente, 2018]; or [Kollar, 2007] for an overview with
historical context. The statement selected here comes from [Hironaka, 1973], pp.
459–460 (under the title Desingularization II).
In this section Ω will be a domain in R𝑛 (𝑛 ≥ 1). Generally speaking, the results
will be local and it suffices to reason in neighborhoods of the origin, 0 ∈ Ω.
Theorem 15.3.1 Let 𝐹 ∈ C 𝜔 (Ω) be real-valued, 𝐹 . 0, 𝐹 (0) = 0. There exist a
neighborhood 𝑈 of 0 in Ω, a C 𝜔 manifold M of (real) dimension 𝑛, countable at
infinity, and a proper analytic map 𝜒 : M −→ 𝑈 such that 𝑽 = {𝑥 ∈ 𝑈; 𝐹 (𝑥) = 0}
has the following properties:
−1
(1) The restriction of 𝜒 to M\ 𝜒 (𝑽) is a C 𝜔 diffeomorphism onto 𝑈\𝑽.
−1
(2) For each ℘ ∈ 𝜒 (𝑽) there are real-analytic coordinates 𝑦 1 , ..., 𝑦 𝑛 in a neighbor-
hood N℘ of ℘ in M, all vanishing at ℘, such that

∀ 𝑦 ∈ N℘ , (𝐹 ◦ 𝜒) (𝑦) = 𝐸 (𝑦) 𝑦 1𝛼1 · · · 𝑦 𝑛𝛼𝑛 , (15.3.1)

with 𝐸 ∈ C 𝜔 N℘ ; R nowhere vanishing in N℘ and 𝛼 𝑘 ∈ Z+ , 𝑘 = 1, ..., 𝑛.
15.3 Desingularization and Applications 549

Remark 15.3.2 The map 𝜒 is surjective by Property (1) and the obvious fact that 𝜒
−1
is a surjection of 𝜒 (𝑽) onto 𝑽.
−1
Property (2) is often formulated as “ 𝜒 (𝑽) has normal crossings”.
−1 −1
Evidently, 𝜒 (𝑽) is a C 𝜔 subvariety of M, the nullset of 𝜒∗ 𝐹 = 𝐹 ◦ 𝜒; 𝜒 (𝑽) is
−1
compact if and only if 𝑽 ⊂⊂ 𝑈. The set N℘ ∩ 𝜒 (𝑽) is the analytic subvariety of N℘
defined by the equation 𝑦 𝜄1 · · · 𝑦 𝜄𝜈 = 0 where 1 ≤ 𝜄1 < · · · < 𝜄𝜈 ≤ 𝑛 and, in (15.3.1),
−1
𝛼 𝜄1 · · · 𝛼 𝜄𝜈 ≠ 0 whereas 𝛼 𝜄 = 0 if 𝜄 ≠ 𝜄 𝑗 for every 𝑗 = 1, ..., 𝜈. For each 𝑥 ∈ 𝑽, 𝜒 (𝑥)
−1 −1
is a compact analytic subvariety of M; we have 𝜒 (𝑥 ◦ ) ∩ 𝜒 (𝑥 ∗ ) = ∅ if 𝑥 ◦ ≠ 𝑥 ∗ .

15.3.2 Application: a variant of the Lojasiewicz inequality

We can derive from Theorem 15.3.1 a variant of the Lojasiewicz inequality (Theorem
15.1.8). In what follows 𝑈, 𝑽, M, 𝜒, N℘ are the objects in Theorem 15.3.1.
Unless the proper, surjective map 𝜒 is injective, it is not open; more precisely, the
following can be said:

Lemma 15.3.3 Let 𝑥 ◦ ∈ 𝑽. The following properties hold:

−1
(1) If there are two distinct points ℘ 𝑗 ∈ 𝜒 (𝑥 ◦ ), 𝑗 = 1, 2, and if the neighborhoods
U 𝑗 of ℘ 𝑗 in M are such that U1 ∩ U2 = ∅ then there is no neighborhood of 𝑥 ◦
contained in 𝜒 (U1 ).
−1
(2) Given any open subset N of M containing 𝜒 (𝑥 ◦ ), there is a neighborhood 𝑈 ′
−1
of 𝑥 ◦ in 𝑈 such that 𝜒 (𝑈 ′) ⊂ N .

Proof If 𝜒 (U1 ) contained a neighborhood of 𝑥 ◦ then the open set 𝜒 (U1 ) ∩ 𝜒 (U2 ) ∩
(𝑈\𝑽) could not be empty since 𝑥 ◦ belongs to the closure of 𝜒 (U2 ) ∩ (𝑈\𝑽).
By
−1
Property (1) in Theorem 15.3.1 we would have U1 ∩ U2 ∩ M\ 𝜒 (𝑽) ≠ ∅, a
contradiction. This proves (1).
Assume there were a sequence 𝑆 = 𝑥 (𝜈) 𝜈=1,2,... ⊂ 𝑈 converging to 𝑥 ◦ such


that to each 𝜈 there is a ℘ (𝜈) ∉ N , 𝜒 ℘ (𝜈) = 𝑥 (𝜈) . Since 𝜒 is proper and 𝑆 ∪ {𝑥 ◦ }
is compact there is a subsequence of ℘ (𝜈) 𝜈=1,2,... converging to some point ℘ ∈

−1
𝜒 (𝑥 ◦ ), implying ℘ (𝜈) ∈ N for some 𝜈, a contradiction. This proves (2). □
−1
Corollary 15.3.4 There are finitely many points ℘ ( 𝑗) ∈ 𝜒 (𝑥 ◦ ), 𝑗 = 1, ..., 𝑟, such
that there is a neighborhood 𝑈 ′ of 𝑥 ◦ in 𝑈 such that

𝑈 ′ ⊂ 𝜒 N℘(1) ∪ · · · ∪ 𝜒 N℘(𝑟 ) (15.3.2)

where the neighborhoods N℘( 𝑗) have the property in (2), Theorem 15.3.1.
550 15 Division of Distributions by Analytic Functions

−1
Proof Indeed, there are finitely many points ℘ ( 𝑗) ∈ 𝜒 (𝑥 ◦ ), 𝑗 = 1, ..., 𝑟, such that
−1
𝜒 (𝑥 ◦ ) ⊂ N℘(1) ∪ · · · ∪ N℘(𝑟 ) . □

Theorem 15.3.5 Let Ω be a domain in R𝑛 and 𝐹, 𝐺 ∈ C 𝜔 (Ω) be such that 𝐹 (𝑥) = 0

entails 𝐺 (𝑥) = 0 whatever 𝑥 ∈ Ω. Then each point 𝑥 ◦ ∈ Ω has a neighborhood 𝑈 ′ ⊂
Ω such that, for some constants 𝑐 > 0, 𝑠 > 0 and all 𝑥 ∈ 𝑈 ′, |𝐹 (𝑥) | 𝑠 ≥ 𝑐|𝐺 (𝑥) |.

Proof It suffices to prove the claim in a neighborhood of an arbitrary point 𝑥 ◦ ∈ Ω

such that 𝐹 (𝑥 ◦ ) = 0. Let then 𝑈 ∋ 𝑥 ◦ , 𝑽, M, 𝜒, N℘ be as in Theorem 15.3.1 and
℘ (1) , ..., ℘ (𝑟) be such that (15.3.2) holds. The claim will follow if we prove that, for
each 𝑗 and some 𝑐 𝑗 > 0, 𝑠 𝑗 > 0,

∀𝑦 ∈ N℘( 𝑗) , |𝐹 ( 𝜒 (𝑦)) | 𝑠 𝑗 ≥ 𝑐 𝑗 |𝐺 ( 𝜒 (𝑦)) |, (15.3.3)

since then we can take 𝑠 = min 𝑠 𝑗 , 𝑐 = min 𝑐 𝑗 .

1≤ 𝑗 ≤𝑟 1≤ 𝑗 ≤𝑟
−1
Let us prove the analogue of (15.3.3) in N℘ for arbitrary ℘ ∈ 𝜒 (𝑥 ◦ ). Let 𝜄ℓ be the
integers such that 𝛼 𝜄ℓ ≠ 0 in (15.3.1) (1 ≤ 𝜄1 < · · · < 𝜄 𝑘 ≤ 𝑛); the hypothesis implies
𝐺 ( 𝜒 (𝑦)) = 0 if 𝑦 𝜄1 · · · 𝑦 𝜄𝑘 = 0 whence, for 𝜇 = max 𝛼 𝜄ℓ and suitably small 𝑠 > 0,
ℓ=1,...,𝑘
𝑐 > 0,
∀𝑦 ∈ N℘ , 𝑐|𝐺 ( 𝜒 (𝑦)) | ≤ |𝑦 𝜄1 · · · 𝑦 𝜄𝑘 | 𝜇𝑠 ≤ 𝑐−1 |𝐹 ( 𝜒 (𝑦)) | 𝑠 . □

Remark 15.3.6 In view of Theorem 15.3.5 the proof of the Lojasiewicz inequality
(Theorem 15.1.8) is reduced to finding one function 𝐺 ∈ C 𝜔 (𝑈) (𝑈: a neighborhood
of 𝑥 ◦ ∈ Ω) and 𝜅 > 0, such that

∀𝑥 ∈ 𝑈, (dist (𝑥, 𝑽)) 𝜅 ≲ |𝐺 (𝑥)| ,

where 𝑽 = {𝑥 ∈ Ω; 𝐹 (𝑥) = 0}, 𝐹 ∈ C 𝜔 (Ω). This is quite simple, by using again

Theorem 15.3.1 and the representations (15.3.1).

15.3.3 Application: reciprocals of analytic functions as distributions

Theorem 15.3.1 provides a new proof of a special case of the division theorem,
Theorem 15.2.1.
Let 𝐹 ∈ C 𝜔 (Ω), 𝑈, M, 𝜒 be as in Theorem 15.3.1. An important feature in
Theorem 15.3.1 is that, generally speaking, the manifold M is not orientable. This
compels us to deal with densities in M, concepts defined and discussed in Section
11.1. In this subsection the space of distributions D ′ (M) is identified with the
dual of the space Cc∞ (M; |d𝑦|) of compactly supported C ∞ one-densities 𝜑 |d𝑦|,
𝜑 ∈ Cc∞ (M). In other words, we regard the elements of D ′ (M) as distribution
zero-densities.
15.3 Desingularization and Applications 551

Let 𝑥1 , ..., 𝑥 𝑛 be the canonical coordinates in 𝑈 ⊂ R𝑛 . We can use the coordinates

𝑦 1 , ..., 𝑦 𝑛 in the neighborhood N℘ in Property (2), Theorem 15.3.1, to express the
restriction 𝜒℘ of the map 𝜒 to N℘ by equations 𝑥 𝑗 = 𝑥 𝑗 (𝑦), 𝑗 = 1, ..., 𝑛. This leads
to the pullback map of C ∞ one-densities

D𝑥
𝜒℘∗ (𝜑 (𝑥) |d𝑥|) = 𝜑 (𝑥 (𝑦)) |d𝑦| .
D𝑦

There is a locally finite covering of M consisting of neighborhoods N℘ (℘ ∈ 𝑬, a

−1
discrete set) and M\ 𝜒 (𝑽); we can select a C ∞ partition of unity in M, 𝑔℘ ℘∈𝑬

Í Í
and 𝑔◦ = 1 − ℘∈𝑬 𝑔℘ , with 𝑔℘ ∈ Cc𝜔 N℘ , and ℘∈𝑬 𝑔℘ ≡ 1 in an open subset
−1 −1 ∗
of M containing 𝜒 (𝑽). Let 𝜒−1 ◦ denote the restriction of 𝜒 to M\ 𝜒 (𝑽) and 𝜒◦ :
C ∞ (𝑈\𝑽, |d𝑥|) −→ C ∞ M\ 𝜒 (𝑽) , |d𝑦| the associated pullback map. We can
define the global pullback map as
∑︁
𝜒∗ = 𝑔◦ 𝜒◦∗ + 𝑔℘ 𝜒℘∗ : C ∞ (𝑈, |d𝑥|) −→ C ∞ (M, |d𝑦|) . (15.3.4)
℘∈𝑬

Since 𝜒 is proper we have 𝜒∗ Cc∞ (𝑈, |d𝑥|) ⊂ Cc∞ (M, |d𝑦|). Condition (1) in

Theorem 15.3.1 ensures
that this definition does not depend on the choice of the
partition of unity, 𝑔℘ ℘∈𝑬 and 𝑔◦ . The transpose of (15.3.4) is the pushforward
map 𝜒∗ : D ′ (M) −→ D ′ (𝑈).
We need the following result in Euclidean space.

Lemma 15.3.7 Let 𝑄 𝜌(𝑛) = 𝑦 ∈ R𝑛 ; 𝑦 𝑗 < 𝜌, 𝑗 = 1...𝑛 , 𝜌 > 0, and (𝛼1 , ..., 𝛼𝑛 ) ∈


Z+𝑛 . The subspace 𝑦 𝛼 C ∞ 𝑄 𝜌(𝑛) is closed in C ∞ 𝑄 𝜌(𝑛) .

Proof For 𝜓 ∈ C ∞ 𝑄 𝜌(𝑛) to belong to 𝑦 𝑗 𝑗 C ∞ 𝑄 𝜌(𝑛) (1 ≤ 𝑗 ≤ 𝑛, 𝛼 𝑗 ≥ 1) it is
𝛼

necessary and sufficient that

𝜕𝛽𝑗 𝜓
∀𝛽 𝑗 ∈ Z+ , 𝛽 𝑗 < 𝛼 𝑗 , 𝛽
≡ 0,
𝜕𝑦 𝑗 𝑗
𝑦 𝑗 =0

implying that 𝑦 𝑗 𝑗 C ∞ 𝑄 𝜌(𝑛) is closed in C ∞ 𝑄 𝜌(𝑛) ; the same is true of
𝛼

Ù𝑛
𝑦 𝛼 C ∞ 𝑄 𝜌(𝑛) = 𝑦 𝑗 𝑗 C ∞ 𝑄 𝜌(𝑛) .
𝛼
□
𝑗=1

Proposition 15.3.8 The linear subspace ( 𝜒∗ 𝐹) C ∞ (M) is closed in C ∞ (M).

552 15 Division of Distributions by Analytic Functions

Proof Let {𝜑 𝜈 } 𝜈=1,2,... be a sequence in C ∞ (M) such that 𝜑 𝜈 ( 𝜒∗ 𝐹) converges

−1
to 𝜙 in C ∞ (M). For every ℘ ∈ 𝜒 (𝑽) we select a neighborhood N℘ as in (2),
Theorem 15.3.1; we can adjust it so that N℘ = 𝑄 𝜌(𝑛) for the coordinates 𝑦 1 , ...,∗ 𝑦 𝑛 .
∞ N
Lemma 15.3.7 implies that there is a 𝜓 ℘ ∈ C ℘ such that 𝜙 = 𝜓 ℘ ( 𝜒 𝐹).
We use the partition of unity 𝑔℘ ℘∈𝑬 and 𝑔◦ introduced above. We have 𝜙 =
Í
∗ −1
℘∈𝑬 𝑔 ℘ 𝜓 ℘ + 𝑔◦ 𝜙 ( 𝜒 𝐹) ( 𝜒∗ 𝐹). □

Corollary 15.3.9 The subspace ( 𝜒∗ 𝐹) Cc∞ (M, |d𝑦|) is closed in Cc∞ (M, |d𝑦|).

Proof If 𝜓 ∈ C ∞ (M, |d𝑦|) and 𝜓 ( 𝜒∗ 𝐹) ∈ ( 𝜒∗ 𝐹) Cc∞ (M, |d𝑦|) then 𝜓 ∈

Cc∞ (M, |d𝑦|).
□

Proposition 15.3.10 We have ( 𝜒∗ 𝐹) D ′ (M) = D ′ (M).

Proof The map

Cc∞ (M, |d𝑦|) ∋ 𝜑 ↦→ ( 𝜒∗ 𝐹) 𝜑 ∈ Cc∞ (M, |d𝑦|)

−1
is injective since M\ 𝜒 (𝑽) is open and dense in M. Therefore the image of its
transpose, ( 𝜒∗ 𝐹) D ′ (M), is dense in D ′ (M) (by the Hahn–Banach Theorem);
( 𝜒∗ 𝐹) D ′ (M) is closed in D ′ (M) by Corollary 15.3.9 and the Banach–Steinhaus
Theorem. □

Corollary 15.3.11 There is a 𝑢 ∈ D ′ (M) such that ( 𝜒∗ 𝐹) 𝑢 = 1.

We come now to the result we were seeking:

Theorem 15.3.12 Let X be a C 𝜔 manifold, connected and countable at infinity, and

let 𝐹 ∈ C 𝜔 (X), 𝐹 . 0. There is a 𝑣 ∈ D ′ (X) such that 𝐹𝑣 = 1.

Proof It suffices to prove the statement locally and then patch the local results by
means of a locally finite C ∞ partition of unity. Using local coordinates the problem
reduces to the Euclidean set-up of X = 𝑈, 𝐹 and 𝜒 : M −→ 𝑈 as above. If
𝑢 ∈ D ′ (M) satisfies ( 𝜒∗ 𝐹) 𝑢 = 1 (Corollary 15.3.11) then 𝐹 𝜒∗ 𝑢 = 1. □

Remark 15.3.13 We have stated Theorem 15.3.1 for real-valued functions 𝐹 ∈

C 𝜔 (X). The preceding statements, starting with Proposition 15.3.8, extend
easily
2
to complex 𝐹. For instance, Theorem 15.3.12 applied to |𝐹 | yields 𝐹 𝐹𝑣 = 1.

Remark 15.3.14 One can deduce from Proposition 15.3.10 more than what is stated
in Theorem 15.3.12, namely that 𝐹D ′ (𝑈) = 𝜒∗ D ′ (𝑈). Theorem 15.2.1 states that
𝐹D ′ (𝑈) = D ′ (𝑈), implying that 𝜒∗ is surjective. The latter is equivalent to proving
that the range of 𝜒∗ : Cc∞ (𝑈, |d𝑥|) −→ Cc∞ (M, |d𝑦|) is closed, which we have not
proved.
15.3 Desingularization and Applications 553

15.3.4 Reciprocal of analytic functions. An alternate proof

We now give an alternate, quite different, proof of Theorem 15.3.12 (see [Bernstein-
Gelfand, 1969], [Atiyah, 1970]).

Proposition 15.3.15 Let 𝐹 ∈ C 𝜔 (Ω) be as in Theorem 15.3.1, 𝐹 ≥ 0 in 𝑈. If

𝑓 = 𝐹 ◦ 𝜒 ∈ C 𝜔 (M) then 𝑠 ↦→ 𝑓 𝑠 is a holomorphic function in the half-plane
Re 𝑠 > 0 valued in D ′ (M).

Proof It suffices to prove the claim when M is replaced by the domain U of analytic
local coordinates 𝑦 1 , ..., 𝑦 𝑛 . For every 𝜑 ∈ Cc∞ (U),
∫ ∫
𝑠
𝜑 (𝑦) 𝑓 (𝑦) d𝑦 ≤ |𝜑 (𝑦)| d𝑦 max | 𝑓 | Re 𝑠
supp 𝜑

as well as ∫
𝜕
𝜑 (𝑦) e𝑠 ln 𝑓 ( 𝑦) d𝑦 = 0
𝜕 𝑠¯
by the Lebesgue theorem on differentiation under the integral sign. □
−1
Remark 15.3.16 We can regard 𝑓 0 as the characteristic function of the set M\ 𝜒 (𝑽).
−1
However, regarded as a distribution in M and since the Lebesgue measure of 𝜒 (𝑽)
is zero, 𝑓 0 is the (locally integrable) function 1.

We now take a look at the one-variable case.

Proposition 15.3.17 Let U = (−1, 1) ⊂ R and 𝛼 ∈ Z+ , 𝛼 ≥ 1, and 𝑓 (𝑦) =

𝑦 2𝛼 . Then the map 𝑠 ↦→ 𝑓 𝑠 ∈ D ′ (U) can be extended as a distribution-valued
meromorphic function with simple poles at the points −𝑘/2𝛼, 𝑘 = 1, 2, ....

Proof Since, for arbitrary 𝑘 ∈ Z, 𝑘 ≥ 1,

𝜕 𝑘 2𝛼𝑠+𝑘
𝑦 = (2𝛼𝑠 + 1) · · · (2𝛼𝑠 + 𝑘) 𝑦 2𝛼𝑠
𝜕𝑦 𝑘

we get, if Re 𝑠 > 0 and 𝜑 ∈ Cc∞ (U),

(−1) 𝑘
∫ ∫
2𝛼𝑠 𝜕𝑘 𝜑
𝜑 (𝑦) 𝑦 d𝑦 = 𝑦 2𝛼𝑠+𝑘 (𝑦) d𝑦.
(2𝛼𝑠 + 1) · · · (2𝛼𝑠 + 𝑘) 𝜕𝑦 𝑘
We note that the right-hand side integral is finite provided 2𝛼 Re 𝑠 + 𝑘 > −1. The
claim ensues. □

Theorem 15.3.18 Let 𝛼 ∈ Z+ be arbitrary. There is a distribution 𝑢 in (−1, 1) such

that 𝑦 2𝛼 𝑢 = 1.
554 15 Division of Distributions by Analytic Functions

Proof The statement being trivial when 𝛼 = 0 we assume 𝛼 ≥ 1; let 𝑓 (𝑦) = 𝑦 2𝛼 . We

look at the Laurent expansion of the meromorphic function C ∋ 𝑠 ↦→ 𝑓 𝑠 ∈ D ′ (−1, 1)
about 𝑠 = −1:
+∞
1 ∑︁
𝑓𝑠 = 𝑢+ 𝑢 𝑘 (𝑠 + 1) 𝑘 , 𝑢, 𝑢 𝑘 ∈ D ′ (−1, 1) ,
𝑠+1 𝑘=0

whence, by the Cauchy formula,

∮
0 1
𝑓 = 𝑓 𝑠+1 d𝑠 = 𝑢 𝑓
2𝑖𝜋
|𝑠+1|=𝜀

with 0 < 𝜀 ≪ 1/2𝛼. As pointed out in Remark 15.3.16, 𝑓 0 = 1 is a distribution in

(−1, 1). □

Corollary 15.3.19 Let 𝛼 ∈ Z+𝑛 be arbitrary (𝑛 ≥ 1). There is a 𝑢 ∈ D ′ ((−1, 1) 𝑛 )

such that 𝑦 2𝛼1 2𝛼𝑛
1 · · · 𝑦 𝑛 𝑢 = 1.

2𝛼
Proof For each 𝑗 = 1, ..., 𝑛 there is a 𝑢 ( 𝑗) ∈ D ′ (−1, 1) such that 𝑦 𝑗 𝑗 𝑢 ( 𝑗) = 1. Then
𝑢 (𝑦) = 𝑢 (1) (𝑦 1 ) ⊗ · · · ⊗ 𝑢 (𝑛) (𝑦 𝑛 ) satisfies the requisites of the corollary. □

Corollary 15.3.20 Let 𝐹 ∈ C 𝜔 (Ω), 𝐹 ≥ 0 in 𝑈 and let 𝜒 : M −→ 𝑈 be as in

Theorem 15.3.1. There is a 𝑢 ∈ D ′ (M) such that (𝐹 ◦ 𝜒) 𝑢 = 1.

Proof The proof is the same as that of Proposition 15.3.8 except that we use Corollary
15.3.19 in lieu of Lemma 15.3.7. □
Theorem 15.3.12 for 𝐹 ≥ 0 follows directly from Corollary 15.3.20 and then for
complex 𝐹 by the result for |𝐹 | 2 (cf. Remark 15.3.14).

15.3.5 Tempered fundamental solutions of linear PDEs with constant

coefficients

In this subsection we return to the space of tempered distributions in R𝑛 , S ′ (R𝑛 ).

We apply Proposition 15.2.36 and make use of the isomorphism of the space S (R𝑛 )
of rapidly decaying C ∞ functions at ∞ onto the space of C ∞ functions in S𝑛 that
vanish to infinite order at ∞, C◦∞ (S𝑛 ), defined by the natural diffeomorphism of
R𝑛 onto S𝑛 \ {∞}. The transpose of the resulting map S (R𝑛 ) −→ C ∞ (S𝑛 ) is a
surjection of D ′ (S𝑛 ) onto S ′ (R𝑛 ), whose kernel is the orthogonal of C◦∞ (S𝑛 ) in
D ′ (S𝑛 ), namely the linear span of all the partial derivatives of the Dirac distribution
at infinity.
15.A Appendix 555
−𝑘
Let 𝑃 ∈ C [𝜉1 , ..., 𝜉 𝑛 ], 𝑃 (𝜉) . 0; if 𝑘 ∈ Z, 2𝑘 ≥ deg 𝑃, 1 + |𝜉 | 2 𝑃 (𝜉) can be
extended to S as a C function. By Theorem 15.3.12 there is a 𝑣 ∈ D ′ (S𝑛 ) such
𝑛 𝜔
−𝑘 −𝑘
that 𝑃 (𝜉) 1 + |𝜉 | 2 𝑣 = 1. Let 𝐸 ∈ S ′ (R𝑛 ) be the transform of 1 + |𝜉 | 2 𝑣
′
under the map D (S ) −→ S (R ) and 𝐸
𝑛 ′ 𝑛 e denote the inverse Fourier transform of
𝐸; we have 𝑃 (D) 𝐸 e = 𝛿, which provides a new proof of Corollary 15.2.39.

Remark 15.3.21 To the knowledge of the author, no explicit formula of a tempered

fundamental solution of a linear PDE with constant coefficients has been proposed
in the general case.

15.A Appendix

15.A.1 Coverings by cubes

In this appendix we shall use systematically the norm ∥𝑥∥ = max 𝑥 𝑗 in R𝑛 and
𝑗=1,...,𝑛
the notation 𝔔 𝑅 = {𝑥 ∈ R𝑛 ; ∥𝑥∥ < 𝑅}; we refer to 𝔔 𝑅 as an open “cube”. In this
section dist and diam are to be understood in the sense of the metric ∥𝑥 − 𝑦∥.
(𝑞)
For each integer 𝑞 ∈ Z+ let 𝔈 (𝑞) (resp., 𝔈clos ) denote the family of open (resp.,
closed) cubes with vertices at the points of the lattice 2−𝑞 Z𝑛 whose edge length
is exactly equal to 2−𝑞 . Every cube belonging to 𝔈 (𝑞) is of the form 𝔔 (𝑞) (℘) =
𝔔2𝑛−𝑞−1 + 2−𝑞−1 ℘ with ℘ ∈ Zodd
𝑛 (meaning all the components of ℘ are odd integers);

explicitly,
𝔔 (𝑞) (℘) = 𝑥 ∈ R𝑛 ; 𝑥 − 2−𝑞−1 ℘ < 2−𝑞−1 ;

(15.A.1)

𝔔 (𝑞) (℘) shall denote the closure of 𝔔 (𝑞) (℘). The following statement is readily
proved.
′
𝑛 then either 𝔔 (𝑞) (℘)∩𝔔 (𝑞 ) (℘ ′ ) = ∅
Lemma 15.A.1 If 0 ≤ 𝑞 ≤ 𝑞 ′ and ℘, ℘′ ∈ Zodd
′
[equivalent to 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞 ) (℘′) = ∅] or else 𝔔 (𝑞 ) (℘′) ⊂ 𝔔 (𝑞) (℘).
′

Let Ω be a bounded domain in R𝑛 ; if 𝔔 (𝑞) (℘) ⊂ Ω, then

dist 𝔔 (𝑞) (℘) , 𝜕Ω = dist 𝜕𝔔 (𝑞) (℘) , 𝜕Ω (15.A.2)

= dist 2−𝑞−1 ℘, 𝜕Ω − 2−𝑞−1 .

(𝑞)
We introduce two numbers 𝜅 ≥ 1, 𝑀 ≥ 2; we define 𝔈clos (Ω, 𝜅, 𝑀) to be the
(𝑞)
family of all closed cubes 𝔔 (𝑞) (℘) ∈ 𝔈clos such that 𝔔 (𝑞) (℘) ⊂ Ω [and therefore

dist 𝔔 (𝑞) (℘) , 𝜕Ω > 0] and
𝜅
𝑀 diam 𝔔 (𝑞) (℘) ≤ dist 𝔔 (𝑞) (℘) , 𝜕Ω ; (15.A.3)
556 15 Division of Distributions by Analytic Functions

by (15.A.2) this is equivalent to

𝑀 1/𝜅 2−𝑞/𝜅 + 2−𝑞−1 ≤ dist 2−𝑞−1 ℘, 𝜕Ω . (15.A.4)

Lemma 15.A.2 Let ℘, ℘1 ∈ Zodd

𝑛 be such that 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞−1) (℘1 ) ≠ ∅. If
𝔔 (𝑞) (℘) ∈ 𝔈clos (Ω, 𝜅, 𝑀) then 𝔔 (𝑞−1) (℘1 ) ⊂ Ω.

Proof By Lemma 15.A.1 the hypothesis is equivalent to 𝔔 (𝑞) (℘) ⊂ 𝔔 (𝑞−1) (℘1 )
and implies

dist 2−𝑞−1 ℘, 𝜕𝔔 (𝑞−1) (℘1 ) = dist 2−𝑞−1 ℘, 𝜕𝔔 (𝑞) (℘) = 2−𝑞−1 .

Suppose there were 𝑥 ∈ 𝔔 (𝑞−1) (℘1 ), 𝑥 ∉ Ω; we would have, by (15.A.3),

3 × 2−𝑞−1 = diam 𝔔 (𝑞−1) (℘1 ) − dist 2−𝑞−1 ℘, 𝜕𝔔 (𝑞−1) (℘1 )
𝑞 1
≥ 𝑥 − 2−𝑞−1 ℘ ≥ dist 𝔔 (𝑞) (℘) , 𝜕Ω ≥ 2− 𝜅 𝑀 𝜅 ,

which is impossible since 𝜅 ≥ 1, 𝑀 ≥ 2. □

We denote by Ω1 the complement with respect to Ω of the union of all 𝔔 (0) (℘) ∈
(0) (1) (1)
𝔈clos (Ω, 𝜅, 𝑀) and by 𝔉clos (Ω1 , 𝜅, 𝑀) the family of all 𝔔 (1) (℘) ∈ 𝔈clos (Ω, 𝜅, 𝑀)
(1)
such that 𝔔 (℘) ⊂ Ω1 . Recursively for 𝑞 = 1, 2, ..., we define Ω𝑞 to be the
(𝑞′ )
complement with respect to Ω of the union of all 𝔔 (𝑞′ ) (℘) ∈ 𝔉clos Ω𝑞′ , 𝜅, 𝑀 ,

(𝑞) (𝑞)
𝑞 ′ < 𝑞, and 𝔉clos Ω𝑞 , 𝜅, 𝑀 to be the family of all 𝔔 (𝑞) (℘) ∈ 𝔈clos (Ω, 𝜅, 𝑀) such

that 𝔔 (𝑞) (℘) ⊂ Ω𝑞 . Obviously, Ω𝑞+1 ⊂ Ω𝑞 for all integers 𝑞 ≥ 1. The union
∞
Ø
(𝑞)
𝔉clos (Ω, 𝜅, 𝑀) = 𝔉clos Ω𝑞 , 𝜅, 𝑀
𝑞=0

is a partition, meaning that 𝔔 (𝑞) (℘) ∈ 𝔉clos Ω𝑞 , 𝜅, 𝑀 is equivalent to 𝔔 (𝑞) (℘) ∈
(𝑞)
𝔉clos Ω𝑞 , 𝜅, 𝑀 .

Lemma 15.A.3 If 𝔔 (𝑞) (℘) and 𝔔 (𝑞′ ) (℘′) belong to 𝔉clos (Ω, 𝜅, 𝑀) then
′
𝔔 (𝑞) (℘) ≠ 𝔔 (𝑞 ) (℘′) =⇒ 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞′ ) (℘′) = ∅. (15.A.5)
′
Proof If 𝑞 ′ = 𝑞 then 𝔔 (𝑞) (℘) ≠ 𝔔 (𝑞 ) (℘′) is equivalent to ℘ ≠ ℘′ and the
entailment (15.A.5) is evident. If 𝑞 ′ < 𝑞 it follows from the definition of Ω𝑞 and
(𝑞)
of 𝔉clos Ω𝑞 , 𝜅, 𝑀 . If 𝑞 ′ > 𝑞 the same is true since 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞′ ) (℘′) = ∅ is

′
equivalent to 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞 ) (℘′) = ∅. □
15.A Appendix 557

(𝑞)
Lemma 15.A.4 Every closed cube 𝔔 (𝑞) (℘) ∈ 𝔈clos (Ω, 𝜅, 𝑀) is contained in a
(𝑞′ )
unique closed cube 𝔔 (𝑞′ ) (℘′) ∈ 𝔉clos Ω𝑞′ , 𝜅, 𝑀 , 0 ≤ 𝑞 ′ ≤ 𝑞.

(𝑞)
Proof If 𝔔 (𝑞) (℘) ∉ 𝔉clos (Ω, 𝜅, 𝑀), i.e., 𝔔 (𝑞) (℘) ⊄ Ω𝑞 , there is a 𝔔 (𝑞′ ) (℘′) ∈
(𝑞′ )
𝔉clos Ω𝑞′ , 𝜅, 𝑀 , 0 ≤ 𝑞 ′ < 𝑞, such that 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞′ ) (℘′) ≠ ∅, implying

′
𝔔 (𝑞) (℘) ⊂ 𝔔 (𝑞 ) (℘′) (Lemma 15.A.1). The uniqueness follows from (15.A.5). □
Lemma 15.A.5 If 𝑀 ≥ 2−𝜅−1 (diam Ω) 𝜅 then 𝔔 (𝑞) (℘) ∈ 𝔉clos (Ω, 𝜅, 𝑀) entails
𝜅
dist 𝔔 (𝑞) (℘) , 𝜕Ω ≤ 2 𝜅+1 𝑀 diam 𝔔 (𝑞) (℘) . (15.A.6)

Proof Since dist 𝔔 (𝑞) (℘) , 𝜕Ω ≤ diam Ω if

𝔔 (𝑞) (℘) ∈ 𝔉clos (Ω, 𝜅, 𝑀) and diam 𝔔 (0) (℘) = 1

the hypothesis on 𝑀 directly implies (15.A.6) when 𝑞 = 0.

Assume 𝑞 ≥ 1. The negation of (15.A.6) reads
𝑞−1 1
dist 𝔔 (𝑞) (℘) , 𝜕Ω > 21− 𝜅 𝑀 𝜅 . (15.A.7)

Let ℘1 be one of the vertices of 𝔔 (𝑞) (℘); we have 𝔔 (𝑞) (℘) ⊂ 𝔔 (𝑞−1) (℘1 ); the
triangle inequality implies

dist 𝔔 (𝑞) (℘) , 𝜕Ω ≤ dist 𝜕𝔔 (𝑞−1) (℘1 ) , 𝜕Ω + dist 2−𝑞−1 ℘, 𝜕𝔔 (𝑞−1) (℘1 )

= dist 𝜕𝔔 (𝑞−1) (℘1 ) , 𝜕Ω + dist 2−𝑞−1 ℘, 𝜕𝔔 (𝑞) (℘)

= dist 𝜕𝔔 (𝑞−1) (℘1 ) , 𝜕Ω + 2−𝑞−1 .

Combining this with (15.A.7) yields

dist 𝜕𝔔 (𝑞−1) (℘1 ) , 𝜕Ω ≥ dist 𝔔 (𝑞) (℘) , 𝜕Ω − 2−𝑞−1
𝑞−1 1
> 21− 𝜅 𝑀 𝜅 − 2−𝑞−1 .

Since 𝜅 ≥ 1, 𝑀 ≥ 2, we get
𝑞−1 1
𝜅1
dist 𝜕𝔔 (𝑞−1) (℘1 ) , 𝜕Ω ≥ 2− 𝜅 𝑀 𝜅 = 𝑀 diam 𝔔 (𝑞−1) (℘1 ) .

This means that (15.A.3) holds with 𝑞 − 1 substituted for 𝑞 and ℘1 for ℘. Since
𝔔 (𝑞−1) (℘1 ) ⊂ Ω (Lemma 15.A.2) this implies 𝔔 (𝑞−1) (℘1 ) ∈ 𝔈 (𝑞−1) Ω𝑞−1 , 𝜅, 𝑀 .

𝑛 such that 𝔔 (𝑞′ ) (℘ ′ ) ∈
By Lemma 15.A.3 there is 𝑞 ′ ∈ Z+ , 𝑞 ′ ≤ 𝑞 − 1, and ℘′ ∈ Zodd
(𝑞−1) (𝑞 ′) ′
𝔉clos (Ω, 𝜅, 𝑀) and 𝔔 (℘1 ) ⊂ 𝔔 (℘ ), implying 𝔔 (𝑞) (℘) ⊂ 𝔔 (𝑞 ) (℘′) and
′

thereby contradicting (15.A.5). □

558 15 Division of Distributions by Analytic Functions

Proposition 15.A.6 Let Ω ≠ ∅ be a bounded open subset of R𝑛 . Whatever the real

numbers 𝜅 ≥ 1, 𝑀 ≥ 2 the family 𝔉clos (Ω, 𝜅, 𝑀) is a closed covering of Ω.
(𝑞)
Proof Let 𝑥 ◦ ∈ Ω be arbitrary. Since 𝔈clos is a closed covering of R𝑛 whatever 𝑞 ∈ Z+
(𝑞)
there is at least one cube 𝔔 (𝑞) (℘) ∈ 𝔈clos , ℘ ∈ Zodd
𝑛 , such that 𝑥 ◦ ∈ 𝔔 (𝑞) (℘). For

sufficiently large 𝑞 we will have 𝔔 (𝑞) (℘) ⊂ Ω and (15.A.3) will be true, meaning
(𝑞) (𝑞′ )
𝔔 (𝑞) (℘) ∈ 𝔈clos (Ω, 𝜅, 𝑀). There is a 𝔔 (𝑞′ ) (℘′) ∈ 𝔉clos (Ω, 𝜅, 𝑀), 𝑞 ′ ≤ 𝑞, such
′
that 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞′ ) (℘′) ≠ ∅. The latter implies 𝔔 (𝑞) (℘) ⊂ 𝔔 (𝑞 ) (℘′) (Lemma
15.A.1), whence 𝑥 ◦ ∈ 𝔔 (𝑞′ ) (℘′). □

Proposition 15.A.7 Let Ω ≠ ∅ be a bounded open subset of R𝑛 and 𝜅 ≥ 1. If 𝑀 ≥ 2 𝜅

there is a number Φ (𝑛, 𝜅) depending only on Ω and 𝜅 such that at most Φ (𝑛, 𝜅)
cubes 𝔔 (𝑞) (℘) ∈ 𝔉clos (Ω, 𝜅, 𝑀) intersect.

Proof Let 𝔔 (𝑞) (℘) ∈ 𝔉clos (Ω, 𝜅, 𝑀) be arbitrary and assume 𝑞 ≤ 𝑞 ♭ < 𝑞 + 𝜅 + 1.
Since the map 𝑥 ↦→ 2−𝑚 𝑥 is a lattice isomorphism 2−𝑞 Z𝑛 −→ 2−𝑞−𝑚 Z𝑛 the number
( 𝑞♭ ) ♭
Φ◦ (𝑛, 𝜅) of cubes 𝔔 ( 𝑞 ) ℘♭ ∈ 𝔈 (℘ ∈ Z𝑛 ) such that 𝔔 (𝑞) (℘)∩𝔔 ( 𝑞 ) ℘♭ ≠
♭ ♭
clos odd
∅ is independent of 𝑞. If 0 ≤ 𝑞 ′ < 𝑞 ♭ then 𝔔 (𝑞′ ) (℘′) (℘′ ∈ Zodd
𝑛 ) is contained in a

(𝑞 )
♭
unique cube 𝔔 ( 𝑞 ) (℘) ∈ 𝔈
♭
(Lemma 15.A.1). It follows directly that there are at
clos
′ (𝑞 ) ′
most Φ◦ (𝑛, 𝜅) cubes 𝔔 (𝑞 ) (℘′) ∈ 𝔉clos (Ω, 𝜅, 𝑀) such that 𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞′ ) (℘′) ≠
∅.
The remainder of the proof is devoted to showing that 𝑞 ′ ≥ 𝑞+𝜅+1 =⇒ 𝔔 (𝑞) (℘)∩
(𝑞′ )
𝔔 (𝑞′ ) (℘′) = ∅. We reason by contradiction. Let 𝔔 (𝑞′ ) (℘′) ∈ 𝔉clos (Ω, 𝜅, 𝑀) be
′ ′
such that 𝔔 (𝑞) (℘)∩𝔔 (𝑞 ) (℘′) ≠ ∅. Since 𝔔 (𝑞 ) (℘′) ⊂ Ω𝑞+1 and 𝔔 (𝑞) (℘)∩Ω𝑞+1 =
∅ we have
′
𝔔 (𝑞) (℘) ∩ 𝔔 (𝑞′ ) (℘′) = 𝜕𝔔 (𝑞) (℘) ∩ 𝜕𝔔 (𝑞 ) (℘′) ,
whence
′ ′
2−𝑞−1 ℘ − 2−𝑞 −1 ℘′ = 2−𝑞−1 + 2−𝑞 −1 .

From this and from (15.A.3) we derive

(2−𝑞 𝑀) 1/𝜅 ≤ dist 2−𝑞−1 ℘, 𝜕Ω
′ ′
≤ dist 2−𝑞 −1 ℘′, 𝜕Ω + 2−𝑞−1 + 2−𝑞 −1 ,

whence
′
′
(2−𝑞 𝑀) 1/𝜅 − 2−𝑞−1 − 2−𝑞 −1 ≤ dist 2−𝑞 −1 ℘′, 𝜕Ω . (15.A.8)

Let ℘′′ ∈ Zodd

𝑛 be such that

′ ′ ′
𝔔 (𝑞 ) (℘′) ⊂ 𝔔 (𝑞 −1) (℘′′) , 𝔔 (𝑞 −1) (℘′′) ∩ 𝔔 (𝑞) (℘) = ∅; (15.A.9)
15.A Appendix 559
′
this means that ℘′′ is a vertex of 𝔔 (𝑞 ) (℘′) and ℘′′ ∉ 𝔔 (𝑞) (℘). We derive from
(15.A.8):
′ ′ ′
(2−𝑞 𝑀) 1/𝜅 − 2−𝑞−1 − 2−𝑞 −1 ≤ dist (℘′′, 𝜕Ω) + 2−𝑞 ℘′′ − 2−𝑞 −1 ℘′
′
≤ dist (℘′′, 𝜕Ω) + 2−𝑞 −1 ,

whence
′
dist (℘′′, 𝜕Ω) ≥ (2−𝑞 𝑀) 1/𝜅 − 2−𝑞−1 − 2−𝑞
and
′
′
dist 𝔔 (𝑞 −1) (℘′′) , 𝜕Ω ≥ dist (℘′′, 𝜕Ω) − 2−𝑞 ≥ (2−𝑞 𝑀) 1/𝜅 − 2−𝑞 .

From our hypotheses 𝜅 ≥ 1, 𝑀 ≥ 2 𝜅 , 𝑞 ′ ≥ 𝑞 + 𝜅 + 1, we derive

(2−𝑞 𝑀) 1/𝜅 − 2−𝑞 ≥ (2−𝑞 𝑀) 1/𝜅 − 2−𝑞−1 𝑀 1/𝜅

𝑞
≥ 2 (2−𝑞−𝜅 𝑀) 1/𝜅 − 2− 𝜅 −1 𝑀 1/𝜅
′ 1/𝜅
≥ (2−𝑞−𝜅 𝑀) 1/𝜅 ≥ 2−𝑞 +1 𝑀 .

We conclude that
′
𝜅 ′
dist 𝔔 (𝑞 −1) (℘′′) , 𝜕Ω ≥ 𝑀 diam 𝔔 (𝑞 −1) (℘′′) .

This proves that (15.A.9) entails 𝔔 (𝑞′ −1) (℘′′) ∈ 𝔈𝑞′ −1 (Ω, 𝜅, 𝑀). By Lemma 15.A.4
( 𝑞♭ )
′

𝔔 (𝑞 −1) (℘′′) is contained in a unique closed cube 𝔔 ( 𝑞 ) ℘♭ ∈ 𝔉clos Ω𝑞♭ , 𝜅, 𝑀 ,
♭

′ ′

0 ≤ 𝑞 ♭ ≤ 𝑞 ′ −1. This entails 𝔔 (𝑞 ) (℘′) ⊂ 𝔔 (𝑞 −1) (℘′′) ⊂ 𝔔 ( 𝑞 ) ℘♭ , contradicting
♭

the hypothesis that 𝔔 (𝑞′ ) (℘′) ∈ 𝔉 (Ω, 𝜅, 𝑀) (by Proposition 15.A.6). □

(𝑞)
We modify our cubes: we take 𝔔 (𝑞) (℘)∈ 𝔉clos (Ω, 𝜅, 2 𝜅 𝑀), 𝑞 ≥ 1, 𝑀 ≥
𝜅
(diam Ω) ; the latter requirement allows us to apply Lemma 15.A.5. We replace
each 𝔔 (𝑞) (℘) by a slightly larger cube:
(𝑞)
𝑥 − 2−𝑞−1 ℘ < 3 × 2−𝑞−2 ;

𝔔♯ (℘) = 𝑥 ∈ R𝑛 ; (15.A.10)

we have
(𝑞)
diam 𝔔♯ (℘) = 3 × 2−𝑞−1 . (15.A.11)

(𝑞)
Lemma 15.A.8 If 𝔔 (𝑞) (℘)∈ 𝔉clos (Ω, 𝜅, 2 𝜅 𝑀) then
𝜅
(𝑞)
𝑀 diam 𝔔♯ (℘) ≤ dist 𝔔♯(𝑞) (℘) , 𝜕Ω ≤ 22𝜅+1 𝑀 diam 𝔔♯(𝑞) (℘) . (15.A.12)

Proof From (15.A.4) and (15.A.11) we derive

560 15 Division of Distributions by Analytic Functions

(𝑞)
dist 𝔔♯ (℘) , 𝜕Ω ≥ dist 𝔔 (𝑞) (℘) , 𝜕Ω − 2−𝑞−2
𝑞 1
≥ 2− 𝜅 +1 𝑀 𝜅 − 2−𝑞−2 .

≥ 23 , we have
𝜅
Since 𝜅 ≥ 1, 𝑀 ≥ 1 and therefore 2 − 2−2
𝑞 1 𝑞
1
2− 𝜅 +1 𝑀 𝜅 − 2−𝑞−2 ≥ 2− 𝜅 2 − 2−2 𝑀 𝜅
1 𝑞+1 1
≥ 3 𝜅 2− 𝜅 𝑀𝜅;

the first inequality in (15.A.12) is a consequence of this and of (15.A.11).

Replacing 𝑀 by 2 𝜅 𝑀 in (15.A.6) yields
𝜅
dist 𝔔 (𝑞) (℘) , 𝜕Ω ≤ 22𝜅+1 𝑀 diam 𝔔 (𝑞) (℘) .

We then see that the second inequality in (15.A.12) is a direct consequence of the
obvious inequalities

(𝑞)
dist 𝔔♯ (℘) , 𝜕Ω ≤ dist 𝔔 (𝑞) (℘) , 𝜕Ω ,
(𝑞)
diam 𝔔 (𝑞) (℘) < diam 𝔔♯ (℘) . □

(𝑞)
Thus the cubes 𝔔♯ (℘) satisfy the inequalities (15.A.3); they form a family
𝔉♯ (Ω, 𝜅, 𝑀). The following statement is a direct consequence of Propositions 15.A.6
and 15.A.7:
Proposition 15.A.9 The family 𝔉♯ (Ω, 𝜅, 𝑀) is an open covering of Ω. There is a
number Φ♯ (𝑛, 𝜅), depending only on 𝑛 and 𝜅, such that every point of Ω belongs to
(𝑞)
at most Φ♯ (𝑛, 𝜅) cubes 𝔔♯ (℘) ∈ 𝔉♯ (Ω, 𝜅, 𝑀).

15.A.2 Partitions of unity subordinate to an open covering by cubes

We denote by 𝜒𝑅 the characteristic ∫function of the cube 𝔔 𝑅𝑛 . Let 𝜌 (𝑡) ∈ C ∞ (R),

𝜌 > 0 if |𝑡| < 1, 𝜌 (𝑡) = 0 if |𝑡| > 1, R 𝜌 (𝑡) d𝑡 = +1; we shall write
𝑛
Ö
𝜌 𝑛, 𝜀 (𝑥) = 𝜀 −𝑛 𝜌 𝜀 −1 𝑥 𝑗 . (15.A.13)
𝑗=1

We select 𝜀 < 𝑅/2; the convolution

∫
(𝜌 𝑛. 𝜀 ∗ 𝜒𝑅−𝜀 ) (𝑥) = 𝜌 𝑛, 𝜀 (𝑥 − 𝑦) 𝜒𝑅−𝜀 (𝑦) d𝑦 (15.A.14)

belongs to Cc∞ (R𝑛 ) and we have

15.A Appendix 561

𝜌 𝑛, 𝜀 ∗ 𝜒𝑅−𝜀 (𝑥) = 1 if ∥𝑥∥ < 𝑅 − 2𝜀, (15.A.15)

0 < 𝜌 𝑛, 𝜀 ∗ 𝜒𝑅−𝜀 (𝑥) ≤ 1 if ∥𝑥∥ < 𝑅,

𝜌 𝑛, 𝜀 ∗ 𝜒𝑅−𝜀 (𝑥) = 0 if ∥𝑥∥ ≥ 𝑅.

Moreover, if 𝛼 = (𝛼1 , ..., 𝛼𝑛 ) ∈ Z+𝑛 ,

𝑛 ∫
Ö +𝑅
−|𝛼|
𝜌 ( 𝛼 𝑗 ) 𝜀 −1 𝑥 𝑗 − 𝑦 𝑗 d𝑦 𝑗 ,

𝜕𝑥𝛼 𝜌 𝑛, 𝜀 ∗ 𝜒𝑅 (𝑥) = 𝜀
𝑗=1 −𝑅

whence
sup 𝜕𝑥𝛼 𝜌 𝑛, 𝜀 ∗ 𝜒𝑅 (𝑥) ≤ 𝐶 𝛼 𝜀 −| 𝛼 | ,

(15.A.16)
𝑥 ∈R𝑛
𝑛 ∫
+∞
Ö
𝜌 ( 𝛼 𝑗 ) 𝑦 𝑗 d𝑦 𝑗 .

where 𝐶 𝛼 = −∞
𝑗=1
(𝑞)
Now we return to the cubes 𝔔♯ (℘) ∈ 𝔉♯ (Ω, 𝜅, 𝑀). We select 𝜀 = 2−𝑞−2 and
define, for (℘, 𝑞) ∈ Zodd
𝑛 ×Z ,
+

𝜙 ℘,𝑞 (𝑥) = 𝜌 𝑛, 𝜀 ∗ 𝜒2−𝑞−1 +𝜀 𝑥 − 2−𝑞−1 ℘ . (15.A.17)

(𝑞)
We deduce from the definitions (15.A.1), (15.A.10) of 𝔔 (𝑞) (℘), 𝔔♯ (℘), and from
(15.A.15) where 𝑅 = 2−𝑞−1 + 𝜀:

𝜙 ℘,𝑞 (𝑥) = 1 if 𝑥 ∈ 𝔔 (𝑞) (℘) , (15.A.18)

(𝑞)
0 < 𝜙 ℘,𝑞 (𝑥) ≤ 1 if 𝑥 ∈ 𝔔♯ (℘) ,
(𝑞)
𝜙 ℘,𝑞 (𝑥) = 0 if 𝑥 ∉ 𝔔♯ (℘) .

(𝑞)
In particular, supp 𝜙 ℘,𝑞 = 𝔔♯ (℘). We derive from Lemma 15.A.3 and Proposition
15.A.9 that
∑︁
∀𝑥 ∈ Ω, 1 ≤ 𝜙 ℘,𝑞 (𝑥) ≤ Φ♯ (𝑛, 𝜅) , (15.A.19)
(℘,𝑞)

𝑛 × Z such that 𝔔 (𝑞) (℘) ∈

with the summation ranging over the pairs (℘, 𝑞) ∈ Zodd + ♯
𝔉♯ (Ω, 𝜅, 𝑀). We apply (15.A.16):

sup 𝜕𝑥𝛼 𝜙 ℘,𝑞 (𝑥) ≤ 𝐶 𝛼 𝜀 − | 𝛼 | = 2 (𝑞+2) | 𝛼 | 𝐶 𝛼 .

𝑥 ∈R𝑛

(𝑞)
Let 𝑥 ∈ 𝔔♯ (℘) be arbitrary; it follows from (15.A.11) that
−1
(𝑞)
2𝑞+1 ≤ 3 diam 𝔔♯ (℘)
562 15 Division of Distributions by Analytic Functions

and from Lemma 15.A.5 that

(𝑞) (𝑞)
dist (𝑥, 𝜕Ω) ≤ dist 𝔔♯ (℘) , 𝜕Ω + diam 𝔔♯ (℘)
1
𝜅1
(𝑞)
≤ 4 (2𝑀) 𝜅 + 1 diam 𝔔♯ (℘) ,

whence
−1
(𝑞)
𝜀 −1 = 2𝑞+2 ≤ 6 diam 𝔔♯ (℘)
1
𝜅
≤ 6 4 (2𝑀) 𝜅 + 1 dist (𝑥, 𝜕Ω) −𝜅 .

We reach the conclusion that to every 𝛼 ∈ Z+𝑛 there is a constant 𝐶 𝛼′ (𝜅) > 0,
(𝑞)
independent of 𝑥 ∈ 𝔔♯ (℘), such that

𝜕𝑥𝛼 𝜙 ℘,𝑞 (𝑥) ≤ 𝐶 𝛼′ (𝜅) dist (𝑥, 𝜕Ω) −𝜅 | 𝛼 | . (15.A.20)

(𝑞)
This is obviously also true for all 𝑥 ∈ R𝑛 since supp 𝜙 ℘,𝑞 = 𝔔♯ (℘) by (15.A.18).
Availing ourselves of (15.A.19) we define
𝜙 ℘,𝑞
𝜑 ℘,𝑞 = Í . (15.A.21)
(℘,𝑞) 𝜙 ℘,𝑞 (𝑥)

Here and below the summation ranges over the set of pairs (℘, 𝑞) ∈ Z+ × (Z𝑛 \2Z𝑛 )
(𝑞)
such that 𝔔♯ (℘) ∈ 𝔉♯ (Ω, 𝜅, 𝑀). We have 𝜑 ℘,𝑞 ∈ C ∞ (Ω), 0 ≤ 𝜑 ℘,𝑞 ≤ 1,
(𝑞) Í
supp 𝜑 ℘,𝑞 = 𝔔♯ (℘) and (℘,𝑞) 𝜑 ℘,𝑞 (𝑥) = 1 for all 𝑥 ∈ Ω. We leave it as an
exercise to derive from (15.A.20) that to every 𝛼 ∈ Z+𝑛 there is a constant 𝐶 𝛼′′ (𝜅) > 0
independent of (℘, 𝑞) such that

∀𝑥 ∈ R𝑛 , 𝜕𝑥𝛼 𝜑 ℘,𝑞 (𝑥) ≤ 𝐶 𝛼′′ (𝜅) dist (𝑥, 𝜕Ω) −𝜅 | 𝛼 | . (15.A.22)

(𝑞)
The fact that each point of Ω belongs to at most Φ♯ (𝑛, 𝜅) cubes 𝔔♯ (℘) ∈
𝔉clos♯ (Ω, 𝜅, 𝑀) and (15.A.22) imply
∑︁
∀𝑥 ∈ R𝑛 , 𝜕𝑥𝛼 𝜑 ℘,𝑞 (𝑥) ≤ Φ♯ (𝑛, 𝜅) 𝐶 𝛼′′ (𝜅) dist (𝑥, 𝜕Ω) −𝜅 | 𝛼 | . (15.A.23)
(℘,𝑞)
Part V
Analytic Pseudodifferential Operators and
Fourier Integral Operators
Chapter 16
Elementary Pseudodifferential Calculus in the
𝑪 ∞ Class

The present chapter rapidly sketches the rudiments of pseudodifferential calculus in

the C ∞ category, an important step in expanding the analysis of linear PDEs from
ordinary space (the base manifold M) to phase-space (the cotangent bundle 𝑇 ∗ M)
– what is called microlocal analysis. The motivation is also to serve as a guide to
the needed refinements in the C 𝜔 category developed in the following chapter. For
more details we refer to textbooks on the topic, e.g., [Hörmander, 1983, III], [Shubin,
1978], [Taylor, 1981], [Treves, 1980].
In Subsection 2.1.3 we introduced one of the basic concepts of microlocal analysis
in the C ∞ category, the wave-front set. Handling the wave-front set of a distribution
in phase-space requires new tools, pseudodifferential operators constructed from
symbols. Moving between differential operators is constrained by what can be done
with polynomials in the covariables 𝜉 𝑗 with functions of the coordinates 𝑥 𝑘 as
coefficients. The only available transformations in the ring (with respect to addition
and composition) of differential operators are (regular) changes of independent
variables, substitutions of unknowns, division by nonvanishing functions. The lift to
phase-space and the techniques of pseudodifferential calculus immediately free the
mathematician to move within (earlier, unimaginably wide) classes of the functions
of (𝑥, 𝜉) that are their symbols. Substitution, in the symbols, of D 𝑥 𝑗 = −𝑖𝜕/𝜕𝑥 𝑗 for
𝜉 𝑗 (the factor −𝑖 is due to our reliance on the Fourier transform) is often referred to
as quantization (indeed going back to the quantum mechanics of the first decades of
the XXth century) and mathematically formalized as the functor Op.
In the whole algebra of pseudodifferential operators we can move smoothly out
of, and then back to, the thin subset of the polynomials with respect to the partial
derivatives D 𝑥 . The possibilities afforded by this new freedom were spectacularly
demonstrated, in the early 1960s, by the proof of the Atiyah–Singer Index Theorem
([Atiyah-Singer, 1963]) – within the framework of elliptic pseudodifferential oper-
ators. That these form a group with respect to composition underlines the contrast
with elliptic differential operators. This group can be given the structure of a (reg-
ular, infinite dimensional) manifold; it incorporates the parametrices of the elliptic
differential operators constructed using the Fourier transform.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 565
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_16
566 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

The last section is an introduction to Weyl’s quantization; it ends with an example,

the application of the Weyl calculus to a transformation of the Schrödinger equation
when the potential is smooth and satisfies suitable growth conditions at infinity.

16.1 Standard Pseudodifferential Operators

16.1.1 Amplitudes and the quantizing functor Op

Throughout this section Ω1 and Ω2 will be two open subsets of R𝑛 ; unless specified
otherwise 𝑚 will be an arbitrary real number.

Definition 16.1.1 By a standard amplitude (or simply, an amplitude) of order 𝑚

in Ω1 × Ω2 we shall mean a function 𝑎 (𝑥, 𝑦, 𝜉) ∈ C ∞ (Ω1 × Ω2 × R𝑛 ) endowed with
the following property:
Given any compact subset K of Ω1 ×Ω2 and any triplet of multi-indices 𝛼,𝛽,𝛾 ∈ Z+𝑛
there is a constant 𝐶 K, 𝛼,𝛽,𝛾 > 0 such that

∀ (𝑥, 𝑦, 𝜉) ∈ K × R𝑛 , 𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) ≤ 𝐶 K, 𝛼,𝛽,𝛾 (1 + |𝜉 |) 𝑚−|𝛾 | .

𝛽 𝛾
(16.1.1)

We shall denote by 𝑆 𝑚 (Ω1 × Ω2 ) the linear space of amplitudes of order ≤ 𝑚 in

Ω1 × Ω2 ; it is agreed that the null amplitude (the amplitude that vanishes identically)
belongs to 𝑆 𝑚 (Ω1 × Ω2 ) whatever 𝑚 ∈ R. The seminorms

(1 + |𝜉 |) |𝛾 |−𝑚 𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉)
𝛽 𝛾
𝑎 ↦→ sup
( 𝑥,𝑦) ∈K, 𝜉 ∈R𝑛

define a Fréchet space structure on 𝑆 𝑚 (Ω1 × Ω2 ). We denote by 𝑆 (Ω1 × Ω2 ) the

union of all the spaces 𝑆 𝑚 (Ω1 × Ω2 ), 𝑚 ∈ R; and by 𝑆 −∞ (Ω1 × Ω2 ) their in-
tersection; 𝑆 0 (Ω1 × Ω2 ), 𝑆 (Ω1 × Ω2 ), 𝑆 −∞ (Ω1 × Ω2 ) are algebras with respect to
addition and ordinary multiplication.
With any 𝑎 ∈ 𝑆 𝑚 (Ω1 × Ω2 ) one associates the oscillatory integral
∫
((Op𝑎) 𝜓) (𝑥) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝜓 (𝑦) d𝑦d𝜉, (16.1.2)
R2𝑛

in which 𝜓 ∈ Cc∞ (Ω2 ). Unless 𝑚 < −𝑛 the integral (16.1.2) is not, generally
speaking, absolutely convergent. But it is well-defined if we agree to carry out the
integration with respect to 𝑦 first, for 𝑎 (𝑥,
∫ 𝑦, 𝜉) 𝜓 (𝑦) is smooth and has compact
support with respect to 𝑦, and therefore R𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝜓 (𝑦) d𝑦 is rapidly
decreasing as |𝜉 | → +∞, whatever 𝑥 ∈ Ω1 . Actually this is true of a class of
“amplitudes” much larger than the standard ones.
16.1 Standard Pseudodifferential Operators 567

Definition 16.1.2 By a substandard amplitude of order 𝑚 in Ω1 ×Ω2 we shall mean

a function 𝑎 (𝑥, 𝑦, 𝜉) ∈ C ∞ (Ω1 × Ω2 × R𝑛 ) endowed with the following property:
Given any compact subset K of Ω1 × Ω2 and any triplet of multi-indices 𝛼,𝛽,𝛾 ∈
Z+𝑛 there is a constant 𝐶 K, 𝛼,𝛽,𝛾 > 0 such that

𝛽 𝛾
∀ (𝑥, 𝑦, 𝜉) ∈ K × R𝑛 , 𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) ≤ 𝐶 K, 𝛼,𝛽,𝛾 (1 + |𝜉 |) 𝑚 . (16.1.3)

𝑚 (Ω × Ω ) the vector space of the substandard amplitudes

We shall denote by 𝑆sub 1 2
of order 𝑚 in Ω1 × Ω2 . Obviously, 𝑆 𝑚 (Ω1 × Ω2 ) ⊂ 𝑆sub 𝑚 (Ω × Ω ) for every 𝑚 ∈ R.
1 2
The seminorms

(1 + |𝜉 |) −𝑚 𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉)
𝛽 𝛾
𝑎 ↦→ sup (16.1.4)
( 𝑥,𝑦) ∈K, 𝜉 ∈R𝑛

𝑚 (Ω × Ω ). We shall write 𝑆
define
Ø a Fréchet space structureÙ on 𝑆sub 1 2 sub (Ω1 × Ω2 ) =
𝑆sub (Ω1 × Ω2 ). Note that
𝑚 −∞
𝑆sub (Ω1 × Ω2 ) = 𝑆 (Ω1 × Ω2 ).
𝑚

𝑚∈𝑹 𝑚∈𝑹
The following lemma will be helpful.
Lemma 16.1.3 Let 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆sub𝑚 (Ω × Ω ) [resp., 𝑆 𝑚 (Ω × Ω )]. Given any
1 2 1 2
𝑚−2𝑁 (Ω
1 × Ω2 ) [resp., 𝑆
𝑁 ∈ Z+ there are amplitudes 𝑏 𝛼 ∈ 𝑆sub 𝑚−2𝑁 (Ω × Ω )]
1 2
with 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 2𝑁, such that
∑︁
Op𝑎 = (Op𝑏 𝛼 ) 𝐷 𝛼 .
| 𝛼 | ≤2𝑁

Proof It suffices to prove the claim when 𝑁 = 1 and then apply induction on 𝑁. We
have, for arbitrary 𝜓 ∈ Cc∞ (Ω2 ),
∫
𝑛 𝑎 (𝑥, 𝑦, 𝜉) 𝑖 ( 𝑥−𝑦) 𝜉
(2𝜋) (Op𝑎) 𝜓 = 𝜓 (𝑦) 1 − 𝚫 𝑦 e d𝑦
1 + |𝜉 | 2
∫
𝑎 (𝑥, 𝑦, 𝜉)
= e𝑖 ( 𝑥−𝑦) 𝜉 1 − 𝚫 𝑦 𝜓 (𝑦) d𝑦
1 + |𝜉 | 2
(Δ: the Laplacian in R𝑛 ). The Leibniz rule (1.1.5) implies

𝑎
(Op𝑎) 𝜓 = Op 2
1 − 𝚫𝑦 𝜓
1 + |𝜉 |
𝑛
∑︁ 𝐷 𝑦𝑗 𝑎 Δ𝑦 𝑎
+2 Op 𝐷 𝑦 𝑗 𝜓 − Op 𝜓,
𝑗=1 1 + |𝜉 | 2 1 + |𝜉 | 2

whence the claim. □

𝑚 (Ω × Ω ) Formula
Theorem 16.1.4 If 𝑎 ∈ 𝑆sub (16.1.2) defines Op𝑎 as a continu-
1 2
ous linear operator Cc (Ω2 ) −→ C ∞ (Ω1 ).
∞
568 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Proof It suffices to show that Op𝑎 is a continuous linear operator Cc∞ (Ω2 ) −→
𝐿 ∞ (Ω1 ), since

D 𝑥𝛼 ((Op𝑎) 𝜓) (𝑥)
∑︁ 𝛼 ∫
= (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 D 𝑥 D 𝑦 (𝑎 (𝑥, 𝑦, 𝜉) 𝜓 (𝑦)) d𝑦d𝜉
𝛼−𝛽 𝛽

𝛽⪯𝛼
𝛽 R2𝑛

∑︁ ∑︁ 𝛼 𝛽 ′
𝛼−𝛽 𝛽−𝛽′
= (2𝜋) −𝑛 ′
OpD 𝑥 D 𝑦 𝑎 D𝛽 𝜓 (𝑥)
′
𝛽 ⪯ 𝛼 𝛽 ⪯𝛽
𝛽 𝛽

𝛼−𝛽 𝛽−𝛽′ 𝑚 (Ω × Ω ) also, which would imply

and D 𝑥 D𝑦 𝑎 ∈ 𝑆sub 1 2
′
𝛽−𝛽′
𝑎 D𝛽 𝜓 ∈ 𝐿 ∞ (Ω1 ) .
𝛼−𝛽
OpD 𝑥 D𝑦

According to Lemma 16.1.3 it suffices to consider the case Op𝑎 = (Op𝑏) D 𝛼 ,

𝑚−2𝑁 (Ω 𝛼 ∞ ∞
𝑏 ∈ 𝑆sub 1 × Ω2 ), 𝑁 ∈ Z+ . Since D Cc (Ω2 ) ⊂ Cc (Ω2 ) we may as well take
𝑚−2𝑁 (Ω
𝑎 ∈ 𝑆sub 1 × Ω2 ) and select 2𝑁 > 𝑚 + 𝑛 + 1. In this case we deduce from
(16.1.3) and (16.1.2),
|((Op𝑎) 𝜓) (𝑥)| ≤ 𝐶 ∥𝜓∥ 𝐿 1
with 𝐶 > 0 depending solely on 𝑛, 𝑎 (𝑥, 𝑦, 𝜉) and supp 𝜓. □
𝑚 (Ω × Ω ) ∋ 𝑎 ↦→ Op𝑎 is not injective: for
Remark 16.1.5 The linear map 𝑆sub 1 2
example, Op𝑎 = 0 if Ω1 = Ω2 and 𝑎 (𝑥, 𝑦, 𝜉) = (𝑥 − 𝑦) 𝛼 𝜉 𝛽 , 𝛼, 𝛽 ∈ Z+𝑛 , |𝛼| > |𝛽|.

Proposition 16.1.6 If 𝑎 ∈ 𝑆sub𝑚 (Ω × Ω ) the functions 𝑎 ⊤ (𝑥, 𝑦, 𝜉) = 𝑎(𝑦, 𝑥, −𝜉)

1 2
∗ 𝑚 (Ω × Ω ); Op𝑎 ⊤ and Op𝑎 ∗ are continu-
and 𝑎 (𝑥, 𝑦, 𝜉) = 𝑎(𝑦, 𝑥, 𝜉) belong to 𝑆sub 2 1
ous linear operators Cc∞ (Ω1 ) −→ C ∞ (Ω2 ) such that
∫ ∫
𝜓 (𝑦) Op𝑎 ⊤ 𝜑 (𝑦) d𝑦,

𝜑 (𝑥) ((Op𝑎) 𝜓) (𝑥) d𝑥 = (16.1.5)
Ω Ω
∫ 1 ∫ 2
𝜑 (𝑥) ((Op𝑎) 𝜓) (𝑥) d𝑥 = 𝜓 (𝑦) ((Op𝑎 ∗ ) 𝜑) (𝑦)d𝑦 (16.1.6)
Ω1 Ω2

for all 𝜑 ∈ Cc∞ (Ω1 ), 𝜓 ∈ Cc∞ (Ω2 ).

If 𝑎 ∈ 𝑆 𝑚 (Ω1 × Ω2 ) then 𝑎 ⊤ and 𝑎 ∗ belong to 𝑆 𝑚 (Ω2 × Ω1 ).

The meaning of (16.1.5) and (16.1.6) is, of course, that Op𝑎 ⊤ = (Op𝑎) ⊤ , the
transpose of Op𝑎, and Op𝑎 ∗ = (Op𝑎) ∗ , the adjoint of Op𝑎.
Proof The claims about Op𝑎 ⊤ and Op𝑎 ∗ are direct consequences of the identities
∫
𝜑 (𝑥) ((Op𝑎) 𝜓) (𝑥) d𝑥
Ω1
∫ ∫
−𝑛
= (2𝜋) e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝜑 (𝑥) 𝜓 (𝑦) d𝑥d𝑦d𝜉,
R𝑛 Ω1 ×Ω2
16.1 Standard Pseudodifferential Operators 569

∫
𝜑 (𝑥) ((Op𝑎) 𝜓) (𝑥) d𝑥
Ω1
∫ ∫
= (2𝜋) −𝑛 e𝑖 ( 𝑦−𝑥) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉)𝜑 (𝑥)𝜓 (𝑦) d𝑥d𝑦d𝜉,
R𝑛 Ω1 ×Ω2

and of Theorem 16.1.4. The last claim, concerning 𝑆 𝑚 , i.e., standard amplitudes
(Definition 16.1.1), is self-evident. □
𝑚 (Ω × Ω ) the operator Op𝑎 extends as a continuous
Corollary 16.1.7 If 𝑎 ∈ 𝑆sub 1 2
linear operator E (Ω2 ) −→ D ′ (Ω1 ).
′

Proof Theorem 16.1.4 and Proposition 16.1.6 imply directly that the transpose of
Op𝑎 ⊤ is a continuous linear operator E ′ (Ω2 ) −→ D ′ (Ω1 ). Formula (16.1.5) shows
that the restriction of this transpose to Cc∞ (Ω1 ) coincides with Op𝑎. □

Remark 16.1.8 Let 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆sub𝑚 (Ω × Ω ) [resp., 𝑆 𝑚 (Ω × Ω )]. According

1 2 1 2
to Lemma 16.1.3 and to Proposition 16.1.6, given any 𝑁 ∈ Z+ there are amplitudes
𝑚−2𝑁 (Ω
2 × Ω1 ) [resp., 𝑆
𝑏 𝛼 ∈ 𝑆sub 𝑚−2𝑁 (Ω × Ω )] with 𝛼 ∈ Z𝑛 , |𝛼| ≤ 𝑁, such that
2 1 +
∑︁
Op𝑎 ⊤ = (Op𝑏 𝛼 ) D 𝛼 .
| 𝛼 | ≤2𝑁

By transposing this formula we obtain

∑︁
(−1) | 𝛼 | D 𝛼 Op𝑏 ⊤𝛼 .

Op𝑎 = (16.1.7)
| 𝛼 | ≤2𝑁

The distribution kernel, belonging to D ′ (Ω1 × Ω2 ), that corresponds to the linear

operator Op𝑎 is the oscillatory integral
∫
−𝑛
𝐴 (𝑥, 𝑦) = (2𝜋) e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) d𝜉. (16.1.8)
R𝑛

The meaning of Theorem 16.1.4 and Corollary 16.1.7 is that 𝐴 (𝑥, 𝑦) is a C ∞ function
of 𝑥 ∈ Ω1 valued in the space of distributions with respect to 𝑦 ∈ Ω2 , as well as a
C ∞ function of 𝑦 ∈ Ω2 valued in the space of distributions with respect to 𝑥 ∈ Ω1 . In
other words, 𝐴 (𝑥, 𝑦) is a semiregular distribution kernel (Definition 2.3.1). When
𝑎 is a standard amplitude 𝐴 (𝑥, 𝑦) has an all important property not shared by the
substandard amplitudes.

Theorem 16.1.9 Let 𝑎 ∈ 𝑆 (Ω1 × Ω2 ); the distribution kernel (16.1.8) is a C ∞

function of (𝑥, 𝑦) in the open set {(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 ≠ 𝑦}.

Proof If 𝑥 ≠ 𝑦 we can write, for any integer 𝑁 ≥ 0,

570 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class
∫
𝐴 (𝑥, 𝑦) = (2𝜋) −𝑛 |𝑥 − 𝑦| −2𝑁 Δ 𝑁
𝜉e
𝑖 ( 𝑥−𝑦) · 𝜉
𝑎 (𝑥, 𝑦, 𝜉) d𝜉
𝑛
∫R
= (2𝜋) −𝑛 |𝑥 − 𝑦| −2𝑁 e𝑖 ( 𝑥−𝑦) · 𝜉 Δ 𝑁
𝜉 𝑎 (𝑥, 𝑦, 𝜉) d𝜉.
R𝑛

We take advantage of the fact that Δ 𝑁 𝜉 𝑎 is an amplitude of degree 𝑚 − 2𝑁. Selecting

2𝑁 > 𝑚 + 𝑛 we see that 𝐴 (𝑥, 𝑦) ∈ C 0 (Ω1 × Ω2 ). But the same argument applies to
each term in the Leibniz rule expansion of the sum at the right in
∫
D 𝑥𝛼 D 𝑦 𝐴 (𝑥, 𝑦) = (2𝜋) −𝑛 D 𝑥𝛼 D 𝑦 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) d𝜉.
𝛽 𝛽
□
R𝑛

Corollary 16.1.10 If 𝑎 ∈ 𝑆 (Ω1 × Ω2 ) and Ω1 ∩Ω2 = ∅ then 𝐴 (𝑥, 𝑦) ∈ C ∞ (Ω1 × Ω2 ).

Corollary 16.1.11 If 𝑎 ∈ 𝑆 (Ω1 × Ω2 ) then

singsupp (Op𝑎) 𝑢 ⊂ Ω1 ∩ singsupp 𝑢

for all 𝑢 ∈ E ′ (Ω2 ).

Corollary 16.1.11 is a direct consequence of Theorems 2.3.2 and 16.1.9: thus, the
operator Op𝑎 is pseudolocal.
When forming the Volterra composition of distribution kernels it was convenient
to introduce the requirement that they be properly supported (Definition 2.3.6); the
same applies to amplitudes, in the following strong sense:

Definition 16.1.12 We shall say that the amplitude 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆sub

𝑚 (Ω × Ω ) is
1 2
properly supported if it satisfies the following conditions:
To every compact set 𝐾1 ⊂ Ω1 there is a compact set 𝐾2 ⊂ Ω2 such that

∀𝑥 ∈ 𝐾1 , ∀ (𝑦, 𝜉) ∈ Ω2 × R𝑛 , 𝑎 (𝑥, 𝑦, 𝜉) ≠ 0 =⇒ 𝑦 ∈ 𝐾2 . (16.1.9)

To every compact set 𝐾2 ⊂ Ω2 there is a compact set 𝐾1 ⊂ Ω1 such that

∀𝑦 ∈ 𝐾2 , ∀ (𝑥, 𝜉) ∈ Ω1 × R𝑛 , 𝑎 (𝑥, 𝑦, 𝜉) ≠ 0 =⇒ 𝑥 ∈ 𝐾1 . (16.1.10)

If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆sub
𝑚 (Ω × Ω ) is properly supported the same is evidently true of
1 2
the distribution kernel (16.1.8) and of Op𝑎.
𝑚1 𝑚2
Let 𝑎 1 ∈ 𝑆sub (Ω1 × Ω2 ) and suppose that 𝑎 2 ∈ 𝑆sub (Ω2 × Ω3 ) is properly sup-
∞
ported; we can write, for arbitrary 𝜓 ∈ Cc (Ω3 ),

((Op𝑎 1 ◦ Op𝑎 2 ) 𝜓) (𝑥) = ((Op𝑎 1 ) ((Op𝑎 2 ) 𝜓)) (𝑥)

∫ ∫ ∫
′ ′ ′
= (2𝜋) −2𝑛 e𝑖 ( 𝑥−𝑦 ) · 𝜉 e𝑖 ( 𝑦 −𝑦) · 𝜉
R2𝑛 𝑦′ ∈Ω2 𝑦 ∈Ω3
× 𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉) 𝜓 (𝑦) d𝑦d𝑦 ′d𝜉d𝜉 ′
16.1 Standard Pseudodifferential Operators 571

whence

((Op𝑎 1 ◦ Op𝑎 2 ) 𝜓) (𝑥) (16.1.11)

∫ ∫
= (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 1,2 (𝑥, 𝑦, 𝜉) 𝜓 (𝑦) d𝑦d𝜉
R𝑛 𝑦 ∈Ω3

where

𝑎 1,2 (𝑥, 𝑦, 𝜉) (16.1.12)

∫ ∫
−𝑖 ( 𝑥−𝑦 ′ ) · ( 𝜉 − 𝜉 ′ )
= (2𝜋) −𝑛 e 𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉) d𝑦 ′d𝜉 ′.
R𝑛 Ω2

𝑚1 𝑚2
Proposition 16.1.13 If both 𝑎 1 ∈ 𝑆sub (Ω1 × Ω2 ) and 𝑎 2 ∈ 𝑆sub (Ω2 × Ω3 ) are prop-
𝑚1 +𝑚2
erly supported then 𝑎 1,2 belongs to 𝑆sub (Ω1 × Ω3 ) and it too is properly sup-
ported. If both 𝑎 1 ∈ 𝑆 𝑚1 (Ω1 × Ω2 ) and 𝑎 2 ∈ 𝑆 𝑚2 (Ω2 × Ω3 ) are properly supported
then 𝑎 1,2 ∈ 𝑆 𝑚1 +𝑚2 (Ω1 × Ω3 ).
𝑚1 +𝑚2
Proof To prove that 𝑎 1,2 ∈ 𝑆 𝑚1 +𝑚2 (Ω1 × Ω3 ) [resp., 𝑎 1,2 ∈ 𝑆sub (Ω1 × Ω3 )] one
must prove estimates of the type (16.1.1) [resp., (16.1.3)]. In (16.1.12) we restrict the
variation of (𝑥, 𝑦) to a compact subset K of Ω1 × Ω3 . Since 𝑎 1 and 𝑎 2 are properly
supported this has the effect of limiting the integration we respect to 𝑦 ′ to a compact
subset 𝐾2 of Ω2 . This also shows that 𝑎 1,2 is properly supported.
We use the identity
𝑁 ′ ′ ′ ′)
1 + |𝜉 − 𝜉 ′ | 2 e−𝑖 ( 𝑥−𝑦 ) · ( 𝜉 − 𝜉 ) = (1 − Δ 𝑦′ ) 𝑁 e−𝑖 ( 𝑥−𝑦 ) · ( 𝜉 − 𝜉

whence

(2𝜋) 𝑛 𝑎 1,2 (𝑥, 𝑦, 𝜉)

∫ ∫
′ ′
= e−𝑖 ( 𝑥−𝑦 ) · ( 𝜉 − 𝜉 ) (1 − Δ 𝑦′ ) 𝑁 (𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉)) d𝑦 ′
R𝑛 𝐾2
d𝜉 ′
× 𝑁 .
1 + |𝜉 − 𝜉 ′ | 2

We apply (16.1.1):

(1 − Δ 𝑦′ ) 𝑁 (𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉)) ≤ 𝐶 𝑁 (1 + |𝜉 ′ |) 𝑚1 (1 + |𝜉 |) 𝑚2

for some 𝐶 > 0 and all (𝑥, 𝑦) ∈ K, 𝑦 ′ ∈ 𝐾2 and 𝜉 ∈ R𝑛 . Taking 𝑁 suitably large we
obtain
(1 + |𝜉 − 𝜉 ′ |) 𝑚1 d𝜉 ′
∫
′ 𝑚2
𝑎 1,2 (𝑥, 𝑦, 𝜉) ≤ 𝐶 𝑁 (1 + |𝜉 |) 𝑁 . (16.1.13)
R𝑛
1 + |𝜉 ′ | 2
572 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

We use the inequality

(1 + |𝜉 − 𝜉 ′ |) 𝑚1 ≤ (1 + |𝜉 ′ |) |𝑚1 | (1 + |𝜉 |) 𝑚1 . (16.1.14)

Proof (of (16.1.14)) Since the claim is unchanged if we exchange 𝜉 and 𝜉 − 𝜉 ′ there
is no loss of generality in assuming 𝑚 1 ≥ 0, in which case (16.1.14) is equivalent to
the evident consequence of the triangle inequality, 1+|𝜉 − 𝜉 ′ | ≤ (1 + |𝜉 ′ |) (1 + |𝜉 |).□
We deduce from (16.1.13), where we take 𝑁 ≥ 21 (|𝑚 1 | + 𝑛 + 1), and from
(16.1.14):
′′
𝑎 1,2 (𝑥, 𝑦, 𝜉) ≤ 𝐶 𝑁 (1 + |𝜉 |) 𝑚1 +𝑚2 . (16.1.15)
We have, for 𝑗 = 1, ..., 𝑛,

(2𝜋) 𝑛 𝜕𝑥 𝑗 𝑎 1,2 (𝑥, 𝑦, 𝜉) (16.1.16)

∫ ∫
′ ′
= e−𝑖 ( 𝑥−𝑦 ) · ( 𝜉 − 𝜉 ) 𝜕𝑦′𝑗 (𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉)) d𝑦 ′ d𝜉 ′
R𝑛 𝐾2
∫ ∫
′ ′
+ e−𝑖 ( 𝑥−𝑦 ) · ( 𝜉 − 𝜉 ) 𝜕𝑥 𝑗 𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉) d𝑦 ′ d𝜉 ′,
R𝑛 𝐾2

(2𝜋) 𝑛 𝜕 𝜉 𝑗 𝑎 1 (𝑥, 𝑦, 𝜉) (16.1.17)

∫ ∫
′ ′
= e−𝑖 ( 𝑥−𝑦 ) · ( 𝜉 − 𝜉 ) 𝜕 𝜉 𝑗′ 𝑎 1 (𝑥, 𝑦 ′, 𝜉 ′) 𝑎 2 (𝑦 ′, 𝑦, 𝜉) d𝑦 ′ d𝜉 ′
R𝑛 𝐾
∫ ∫ 2
−𝑖 ( 𝑥−𝑦′ ) · ( 𝜉 − 𝜉 ′ )
+ e 𝑎 1 (𝑥, 𝑦 , 𝜉 ) 𝜕 𝜉 𝑗 𝑎 2 (𝑦 , 𝑦, 𝜉) d𝑦 d𝜉 ′.
′ ′ ′ ′
R𝑛 𝐾2

If we apply to the right-hand sides in (16.1.16) and (16.1.17) the argument that has
𝑚1 𝑚2
led us to (16.1.15) we obtain, when 𝑎 1 ∈ 𝑆sub (Ω1 × Ω2 ) and 𝑎 2 ∈ 𝑆sub (Ω2 × Ω3 ),

∇ 𝑥 𝑎 1,2 (𝑥, 𝑦, 𝜉) ≤ 𝑀1 (1 + |𝜉 |) 𝑚1 +𝑚2 ;

∇ 𝜉 𝑎 1,2 (𝑥, 𝑦, 𝜉) ≤ 𝑀1 (1 + |𝜉 |) 𝑚1 +𝑚2 ;

likewise,
∇ 𝑦 𝑎 1,2 (𝑥, 𝑦, 𝜉) ≤ 𝑀1 (1 + |𝜉 |) 𝑚1 +𝑚2 .
The constant 𝑀1 > 0 is independent of (𝑥, 𝑦, 𝜉) ∈ K ×R𝑛 . When 𝑎 1 ∈ 𝑆 𝑚1 (Ω1 × Ω2 )
and 𝑎 2 ∈ 𝑆 𝑚2 (Ω2 × Ω3 ) the same argument yields

∇ 𝜉 𝑎 1,2 (𝑥, 𝑦, 𝜉) ≤ 𝑀1 (1 + |𝜉 |) 𝑚1 +𝑚2 −1 .

𝛽 𝛾
It suffices to replace 𝑎 1,2 by 𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 1,2 in the preceding estimates to see that
induction on |𝛼 + 𝛽 + 𝛾| proves the claim. □
16.1 Standard Pseudodifferential Operators 573

16.1.2 Standard pseudodifferential operators. Smoothing operators

Definition 16.1.14 A linear operator Op𝑎 : E ′ (Ω2 ) −→ D ′ (Ω1 ) with 𝑎 ∈

𝑆 𝑚 (Ω1 × Ω2 ), 𝑚 ∈ R, will be called a standard pseudodifferential operator
of order 𝑚 from Ω2 to Ω1 . We shall denote by ØΨ (Ω1 × Ω2 ) the linear space of
𝑚

these pseudodifferential operators. The union Ψ𝑚 (Ω1 × Ω2 ) will be denoted by

𝑚∈𝑹
Ψ (Ω1 × Ω2 ).
If Ω1 = Ω2 = Ω we shall simply say that Op𝑎 is a standard pseudodifferential
operator in Ω and write Ψ (Ω) rather than Ψ (Ω × Ω) and Ψ𝑚 (Ω) rather than
Ψ𝑚 (Ω × Ω).

n o
Proposition 16.1.15 If there are open coverings Ω′𝑗,𝜈 of Ω 𝑗 ( 𝑗 = 1, 2)
𝜈=1,2,...
such that the restriction of a continuous linear
operator 𝑨 : E ′ (Ω2 )
−→ D ′ (Ω1 ) to
′
every product Ω1,𝜈 ′
× Ω2,𝜈 ′
belongs to Ψ𝑚 Ω1,𝜈 ′
× Ω2,𝜈 then 𝑨 ∈ Ψ𝑚 (Ω1 × Ω2 ).
1 2 1 2

n o
Proof We may assume that the coverings Ω′𝑗,𝜈 ( 𝑗 = 1, 2) are locally
𝜈=1,2,...
finite. Let the functions 𝜑 𝑗,𝜈 ∈ Cc∞ Ω′𝑗,𝜈 (𝜈 ∈ Z+ ) form partitions of unity in Ω 𝑗
n o
subordinate to the covering Ω′𝑗,𝜈 . For each pair of indices 𝜈1 , 𝜈2 there is an
𝜈=1,2,...

′
amplitude 𝑎 𝜈1 ,𝜈2 ∈ 𝑆 𝑚 Ω1,𝜈 ′
× Ω2,𝜈 such that the restriction of 𝑨 to Ω1,𝜈′ ′
× Ω2,𝜈
1 2 1 2
(Definition 2.3.9) is equal to Op𝑎 𝜈1 ,𝜈2 . The sum
∑︁
𝜑1,𝜈1 (𝑥) 𝜑2,𝜈2 (𝑦) 𝑎 𝜈1 ,𝜈2 (𝑥, 𝑦, 𝜉)
𝜈1 ,𝜈2 ∈Z+

being locally finite defines a standard amplitude 𝑎Í∈ 𝑆 𝑚 (Ω1 × Ω2 ). For each 𝑢 ∈
E ′ (Ω2 ) all terms, except finitely many, in the sum 𝜈1 ,𝜈2 ∈Z+ 𝜑1,𝜈1 𝑨 𝜑2,𝜈2 𝑢 vanish
identically and said sum is obviously equal to 𝑨𝑢 = (Op𝑎) 𝑢. □
Smoothing operators (Definition 2.3.3) are standard pseudodifferential operators:

Proposition 16.1.16 For the continuous linear operator 𝑹: E ′ (Ω2 ) −→ D ′ (Ω1 )

to be smoothing it is necessary and sufficient that there be an amplitude 𝑎 ∈
𝑆 −∞ (Ω1 × Ω2 ) such that 𝑹= Op𝑎.

Proof The sufficiency is a direct consequence of the definition of 𝑆 −∞ (Ω1 × Ω2 ) and

differentiation under the integral sign. We prove the necessity: the distribution kernel
corresponding to the smoothing operator 𝑹 is ∫a C ∞ function 𝑅 (𝑥, 𝑦) in Ω1 × Ω2
(Proposition 2.3.4). If 𝜒 ∈ Cc∞ (R𝑛 ) satisfies R𝑛 𝜒 (𝜉) d𝜉 = 1 then 𝑎 (𝑥, 𝑦, 𝜉) =
e−𝑖 ( 𝑥−𝑦) · 𝜉 𝑅 (𝑥, 𝑦) 𝜒 (𝜉) fulfills our requirements. □
574 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Thus smoothing operators make up the linear space

Ù
Ψ−∞ (Ω1 × Ω2 ) = Ψ𝑚 (Ω1 × Ω2 ) . (16.1.18)
𝑚∈𝑹

We write Ψ−∞ (Ω) rather than Ψ−∞ (Ω × Ω).

The transpose 𝑨⊤ of 𝑨 ∈ Ψ𝑚 (Ω1 × Ω2 ) belongs to Ψ𝑚 (Ω2 × Ω1 ) since 𝑎 ∈
𝑆 (Ω1 × Ω2 ) implies 𝑎 ⊤ ∈ 𝑆 𝑚 (Ω2 × Ω1 ) (cf. Proposition 16.1.6).
𝑚

For the composition of pseudodifferential operators we require that they be prop-

erly supported (Definition 2.3.6). In this connection the following result is notewor-
thy:

Proposition 16.1.17 Given any operator 𝑨∈ Ψ𝑚 (Ω1 × Ω2 ) there is a smoothing

operator 𝑹∈ Ψ−∞ (Ω1 × Ω2 ) such that 𝑨−𝑹 is properly supported.
∫
Proof Let 𝐴 (𝑥, 𝑦) be the distribution kernel such that 𝑨𝑢 (𝑥) = 𝐴 (𝑥, 𝑦) 𝑢 (𝑦) d𝑦;
𝐴 (𝑥, 𝑦) is a C ∞ function in {(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 ≠ 𝑦} (Theorem 16.1.9). Sup-
pose {(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 = 𝑦} ≠ ∅, otherwise 𝑨 ∈ Ψ−∞ (Ω1 × Ω2 ). Select at
random a properly supported function 𝜒 ∈ C ∞ (Ω1 × Ω2 ), 𝜒 ≡ 1 in a neighbor-
hood of {(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 = 𝑦}. We have 𝑅 (𝑥, 𝑦) = (1 − 𝜒 (𝑥, 𝑦)) 𝐴 (𝑥, 𝑦) ∈
C ∞ (Ω1 × Ω2 ) and 𝐴 (𝑥, 𝑦) − 𝑅 (𝑥, 𝑦) is properly supported. □
We refer to Definition 2.3.9 for the restriction of an operator 𝑨: E ′ (Ω2 ) −→
D′ (Ω1 ) to open sets 𝑈 𝑗 ⊂ Ω 𝑗 ( 𝑗 = 1, 2). Such a restriction defines a natural
map Ψ𝑚 (Ω1 × Ω2 ) −→ Ψ𝑚 (𝑈1 × 𝑈2 ). In particular, when 𝑈1 = 𝑈2 = Ω1 ∩ Ω2
= Ω1,2 ≠ ∅, this yields a natural map Ψ𝑚 (Ω1 × Ω2 ) −→ Ψ𝑚 Ω1,2 (cf. Definition
16.1.14) and therefore a natural map Ψ (Ω1 × Ω2 ) −→ Ψ Ω1,2 .
𝑚 (Ω × Ω ) −→
16.1.18 If Ω1,2 = Ω1 ∩ Ω2 ≠ ∅ the restriction map Ψ
Proposition 1 2
−∞ (Ω × Ω ) −→
Ψ𝑚 Ω1,2 (𝑚 ∈ R) induces a linear injection Ψ (Ω1 × Ω 2 ) /Ψ 1 2
Ψ Ω1,2 /Ψ−∞ Ω1,2 .

Proof We go back to the cutoff 𝜒 (𝑥, 𝑦) in the proof of Proposition 16.1.17; we can
select it so that 𝜒 (𝑥, 𝑦) = 0 if

dist (𝑥, 𝑦) , (Ω1 × Ω2 ) \ Ω1,2 × Ω1,2

exceeds one-half the distance between {(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 = 𝑦} ⊂ Ω1,2 × Ω1,2

and (Ω1 × Ω2 ) \ Ω1,2 × Ω1,2 . In this case
𝜒 (𝑥, 𝑦) 𝐴 (𝑥,
𝑦) defines an element of
Ψ Ω1,2 and therefore a coset in Ψ Ω1,2 /Ψ−∞ Ω1,2 which vanishes when 𝑨 ∈
Ψ−∞ (Ω1 × Ω2 ).
□
In the analysis of C ∞ singularities absolute precision in the use of pseudodiffer-
ential operators is not required: most often, smoothing operators can be neglected. In
reality, one deals with equivalence classes of operators, that is to say, elements of the
quotient spaces Ψ𝑚 (Ω1 × Ω2 ) /Ψ−∞ (Ω1 × Ω2 ) and Ψ (Ω1 × Ω2 ) /Ψ−∞ (Ω1 × Ω2 ).
16.1 Standard Pseudodifferential Operators 575
•
Proposition 16.1.17 states that every coset 𝑨 ∈ Ψ𝑚 (Ω1 × Ω2 ) /Ψ−∞ (Ω1 × Ω2 ) con-
tains one representative 𝑨 (and actually, as many as there are cutoffs 𝜒 as in the
above proof) belonging to the set of properly supported elements of Ψ𝑚 (Ω1 × Ω2 ),
𝑚 (Ω × Ω ). If 𝑩 ∈ Ψ 𝑚′ (Ω × Ω ) is a representative
a set we shall denote by Ψprop 1 2 prop 2 3
• ′
of a coset 𝑩 ∈ Ψ𝑚 (Ω2 × Ω3 ) /Ψ−∞ (Ω2 × Ω3 ) the composition
• • ′
𝑨 ◦ 𝑩 ∈ Ψ𝑚+𝑚 (Ω1 × Ω3 ) /Ψ−∞ (Ω1 × Ω3 )

is defined as the equivalence class mod Ψ−∞ (Ω1 × Ω3 ) of the composite operator
𝑚+𝑚′ (Ω × Ω ). This is a valid definition: if 𝑨 ≠ 𝑨 and 𝑩 ≠ 𝑩 are also
𝑨 ◦ 𝑩 ∈ Ψprop 1 3 1 1
• •
properly supported representatives of 𝑨 and 𝑩 respectively then 𝑹 = 𝑨 − 𝑨1 and
𝑹 1 = 𝑩 − 𝑩1 are both smoothing and properly supported, and therefore the same is
true of 𝑨 ◦ 𝑩 − 𝑨1 ◦ 𝑩1 = 𝑨 ◦ 𝑹 1 + 𝑹 ◦ 𝑩 − 𝑹 ◦ 𝑹 1 .
Proposition 16.1.18 makes it clear that in the analysis of C ∞ singularities, in
essence we are dealing with elements of the quotient space

Ψ𝑚 (Ω1 × Ω2 ) /Ψ−∞ (Ω1 × Ω2 )

and therefore, we may as well take Ω1 = Ω2 = Ω. In connection with this, we point

−∞ (Ω) is a right and left ideal in the composition algebra Ψ
out that Ψprop
Ø prop (Ω) =
Ψprop (Ω) and thus
𝑚

𝑚∈R

•
−∞
Ψ (Ω) = Ψprop (Ω) /Ψprop (Ω)

is an algebra with respect to the composition just defined.

16.1.3 Action on Sobolev spaces

Standard pseudodifferential operators have important continuity properties. The con-

tinuity properties in the scale of the Sobolev spaces (Section 2.2) are actually valid
for the more general operators Op𝑎, 𝑎 ∈ 𝑆sub𝑚 (Ω × Ω ) (Definition 16.1.2, 𝑚 ∈ R).
1 2

𝑚 (Ω × Ω ) then the map Op𝑎 : C ∞ (Ω ) −→ C ∞ (Ω )

Theorem 16.1.19 If 𝑎 ∈ 𝑆sub 1 2 c 2 1
(Theorem 16.1.4) extends as a continuous linear map 𝐻c𝑠 (Ω2 ) −→ 𝐻loc
𝑠−𝑚 (Ω )
1
whatever 𝑠 ∈ R.

Proof We shall write 𝑨 = Op𝑎, 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆sub 𝑚 (Ω × Ω ). Let 𝐾 ⊂ Ω be a

1 2 2 2
∞
compact set; let 𝜓 ∈ Cc (Ω2 ), 𝜓 (𝑦) = 1 if dist (𝑦, 𝐾2 ) < 𝜀 < dist (𝐾2 , R𝑛 \Ω2 );
and let 𝜑 ∈ Cc∞ (Ω1 ) be arbitrary. We must show that 𝑢 ∈ 𝐻 𝑠 (𝐾2 ) =⇒ 𝜑 𝑨 (𝜓𝑢) =
𝜑 𝑨𝑢 ∈ 𝐻 𝑠−𝑚 (R𝑛 ) and that ∥𝜑 𝑨 (𝜓𝑢)∥ 𝑠−𝑚 ≤ 𝐶 ∥𝑢∥ 𝑠 with 𝐶 > 0 independent of
𝑢. We have 𝜑 𝑨𝜓 = Op (𝜑 (𝑥) 𝑎 (𝑥, 𝑦, 𝜉) 𝜓 (𝑦)) and 𝜑 (𝑥) 𝑎 (𝑥, 𝑦, 𝜉) 𝜓 (𝑦) = 0 unless
(𝑥, 𝑦) ∈ 𝐾1 × 𝐾2 with 𝐾1 = supp 𝜑. We may as well assume that (𝑥, 𝑦) ∉ 𝐾1 × 𝐾2 =⇒
576 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

𝑎 (𝑥, 𝑦, 𝜉) = 0. We then have

∫
e−𝑖 𝑥· 𝜃 𝑨𝑢 (𝑥) d𝑥
∫ ∫ ∫
−𝑛
= (2𝜋) e−𝑖 𝑥· ( 𝜃− 𝜉 ) e−𝑖𝑦· 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑢 (𝑦) d𝑥d𝑦d𝜉
R𝑛 𝐾1 𝐾2
∫ ∫
= (2𝜋) −𝑛 e−𝑖𝑦· 𝜉 b
𝑎 (𝜃 − 𝜉, 𝑦, 𝜉) 𝑢 (𝑦) d𝑦d𝜉,
R𝑛 𝐾2
∫
where b𝑎 (𝜃, 𝑦, 𝜉) = 𝐾 e−𝑖 𝑥· 𝜃 𝑎 (𝑥, 𝑦, 𝜉) d𝑥. (The integrals with respect to 𝑥 and 𝑦
1
are duality brackets between C ∞ functions and compactly supported distributions.)
Since the Fourier transform of a product is equal to the convolution of the Fourier
transforms of the factors we have
∫
c (𝜃) =
𝑨𝑢 e−𝑖 𝑥· 𝜃 𝑨𝑢 (𝑥) d𝑥
∫
−2𝑛
= (2𝜋) b (𝜃 − 𝜉, 𝜉 − 𝜂, 𝜉) b
𝑎b 𝑢 (𝜂) d𝜉d𝜂
R2𝑛
∫
b (𝜃, 𝜂, 𝜉) =
where 𝑎b R𝑛
e−𝑖𝑦· 𝜂 b
𝑎 (𝜃, 𝑦, 𝜉) d𝑦. Thus
∫
c (𝜃) ≤ (2𝜋) −2𝑛
𝑨𝑢 b (𝜃 − 𝜉, 𝜉 − 𝜂, 𝜉) |b
𝑎b 𝑢 (𝜂)| d𝜉d𝜂.
R2𝑛

We use now the inequality [cf. (16.1.14)]:

𝑡 |𝑡 | 𝑡
1 + |𝜃| 2 ≤ 2 |𝑡 | 1 + |𝜃 − 𝜉 | 2 1 + |𝜉 | 2 (16.1.19)

valid for all 𝑡 ∈ R and all (𝜃, 𝜉) ∈ R2𝑛 . It implies

21 (𝑠−𝑚) 1
12 |𝑠−𝑚 | 12 (𝑠−𝑚)
1 + |𝜃| 2 ≤ 2 2 |𝑠−𝑚| 1 + |𝜃 − 𝜉 | 2 1 + |𝜉 | 2 ,
12 𝑠 1
1
2 |𝑠 |
1
2𝑠
1 + |𝜉 | 2 ≤ 2 2 |𝑠 | 1 + |𝜉 − 𝜂| 2 1 + |𝜂| 2 ,

and therefore
1 1
1 (𝑠−𝑚)
c (𝜃) 1 + |𝜃| 2 2
(2𝜋) 2𝑛 2− 2 |𝑠−𝑚|− 2 |𝑠 | 𝑨𝑢
∫ 2 − 21 𝑚 12 |𝑠−𝑚 |
≤ b (𝜃 − 𝜉, 𝜉 − 𝜂, 𝜉) 1 + |𝜉 | 2
𝑎b 1 + |𝜃 − 𝜉 | 2
R2𝑛
12 |𝑠 | 21 𝑠
× 1 + |𝜉 − 𝜂| 2 𝑢 (𝜂)| 1 + |𝜂| 2
|b d𝜉d𝜂.

Since 𝑎 (𝑥, 𝑦, 𝜉) has compact support with respect to (𝑥, 𝑦) the estimates (16.1.3)
imply that there is a 𝐶𝑚,𝑠 > 0 such that, for all (𝜃, 𝜉, 𝜂) ∈ R3𝑛 ,
16.1 Standard Pseudodifferential Operators 577
−𝑚 2
1 + |𝜉 | 2 b (𝜃 − 𝜉, 𝜉 − 𝜂, 𝜉)
𝑎b
−𝑛− 21 |𝑠−𝑚| −𝑛− 12 |𝑠 |
≤ 𝐶𝑚,𝑠 1 + |𝜃 − 𝜉 | 2 1 + |𝜉 − 𝜂| 2 .

We obtain
1 (𝑠−𝑚)
c (𝜃) 1 + |𝜃| 2 2
𝑨𝑢
∫ −𝑛 −𝑛 21 𝑠
≤ 𝐶𝑚,𝑠 1 + |𝜃 − 𝜉 | 2 1 + |𝜉 − 𝜂| 2 𝑢 (𝜂)| 1 + |𝜂| 2
|b d𝜉d𝜂.
R2𝑛

We view the right-hand side as a double convolution 𝑓 ∗ 𝑓 ∗ 𝑔 with

−𝑛 21 𝑠
𝑓 = 1 + |𝜉 | 2 𝑢 (𝜉)| 1 + |𝜉 | 2
∈ 𝐿 1 (R𝑛 ) , 𝑔 = |b ∈ 𝐿 2 (R𝑛 )

(indeed, 2𝑛 ≥ 𝑛 + 1). The classical property of 𝐿 2 (R𝑛 ), that it is an 𝐿 1 (R𝑛 )-module

for convolution, and the inequality

∥ 𝑓 ∗ 𝑓 ∗ 𝑔∥ 𝐿 2 ≤ ∥ 𝑓 ∥ 2𝐿 1 ∥𝑔∥ 𝐿 2

yield the desired result. □

Remark 16.1.20 Inspection of the proof of Theorem 16.1.19 shows that a stronger,
and sometimes useful, result could have been claimed, namely: If a set 𝑨 of sub-
𝑚 (Ω × Ω ) (𝑚 ∈ R) is bounded, meaning that every
standard amplitudes 𝑎 ∈ 𝑆sub 1 2
seminorm (16.1.4) is bounded on 𝑨, then the restrictions to 𝐻c𝑠 (Ω2 ) of the maps
Op𝑎, 𝑎 ∈ 𝑨, form an equicontinuous set of linear maps 𝐻c𝑠 (Ω2 ) −→ 𝐻loc 𝑠−𝑚 (Ω )
1
(𝑠 ∈ R arbitrary). On the topic of equicontinuous sets of linear maps we refer the
reader to [Treves, 1967].

Next we state, without proof, an important result about pseudodifferential oper-

ators in the C ∞ class, originally proved in [Hõrmander, 1966, 2] and known as the
sharp Gårding Inequality. Fairly elementary proofs are available in various treatises
on pseudodifferential operators, e.g. [Hörmander, 1983, III], pp. 76–79, [Treves,
1980], Vol. 1, pp. 233–236. In [Hörmander, 1983, III] the reader will also find
significant refinements.

Theorem 16.1.21 Let 𝑎 (𝑥, 𝑦, 𝜉) be a standard amplitude of order 2𝑚 + 1 in Ω1 × Ω2 .

If Re 𝑎 ≥ 0 in Ω1 × Ω2 then, to every compact subset 𝐾 of Ω2 there is a 𝐶 > 0 such
that
Re ((Op𝑎) 𝑢, 𝑢) 𝐿 2 + 𝐶 ∥𝑢∥ 2𝑚 ≥ 0
for every 𝑢 ∈ Cc∞ (𝐾).
578 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Corollary 16.1.22 Let Ω be an open subset of R𝑛 and 𝑏 (𝑥, 𝑦, 𝜉) be a standard

amplitude of order 𝑚 in Ω × Ω. To every compact subset 𝐾 of Ω there is a 𝐶 > 0
such that
∥(Op𝑏) 𝑢∥ 2𝐿 2 + 𝐶 ∥𝑢∥ 2𝑚− 1 ≥ 0
2

for every 𝑢 ∈ Cc∞ (𝐾).

Proof Anticipating the next section, it is not difficult to check that if 𝑏¯ is the complex
conjugate of 𝑏 ∈ 𝑆 𝑚 (Ω × Ω) and Op 𝑏 ∗ is the adjoint of Op𝑏 then

Op𝑏 ∗ Op𝑏 − Op |𝑏| 2 = Op𝑅, 𝑅 ∈ 𝑆 2𝑚−1 (Ω × Ω) .

We derive:

∥ (Op𝑏) 𝑢∥ 2𝐿 2 ≥ Re Op |𝑏| 2 𝑢, 𝑢 − |((Op𝑅) 𝑢, 𝑢) 𝐿 2 |

≥ Re Op |𝑏| 2 𝑢, 𝑢 − 𝐶 ′ ∥𝑢∥ 2𝑚− 1
2

for every 𝑢 ∈ Cc∞ (𝐾). Then the claim follows directly from Theorem 16.1.21. □

16.1.4 Effect of diffeomorphisms on pseudodifferential operators

In this subsection we limit our attention to the case where Ω1 = Ω2 = Ω. Let Ω′ be

another open subset of R𝑛 and let 𝑥 ′ ↦→ 𝑥 = 𝐹 (𝑥 ′) be a C ∞ diffeomorphism of Ω′
onto Ω. We recall that the pullback under 𝐹 defines an isomorphism 𝐹 ∗ : 𝜑 ↦→ 𝜑◦𝐹 of
Cc∞ (Ω) onto Cc∞ (Ω) and yields by transposition an isomorphism 𝐹∗ : D ′ (Ω′) −→
D ′ (Ω). Given a continuous linear operator 𝑨 : Cc∞ (Ω) −→ D ′ (Ω) we can form
the composite 𝐹∗−1 𝑨 (𝐹 ∗ ) −1 , the lower horizontal arrow in the commutative diagram
𝑨
Cc∞ (Ω) −→ D ′ (Ω)
𝐹∗ ↓ ↑ 𝐹∗ (16.1.20)
Cc (Ω′) → D ′ (Ω′) .
∞

The map

L Cc∞ (Ω) ; D ′ (Ω) ∋ 𝑨 ↦→ 𝐹∗−1 𝑨 (𝐹 ∗ ) −1 ∈ L Cc∞ (Ω′) ; D ′ (Ω′)

is an isomorphism. We are going to show that if 𝑨 = Op𝑎 ∈ Ψ𝑚 (Ω), 𝑎 ∈ 𝑆 𝑚 (Ω),

then 𝐹∗−1 𝑨 (𝐹 ∗ ) −1 ∈ Ψ𝑚 (Ω′).
We derive from (16.1.2), after a change of the variable of integration,

((Op𝑎) 𝜑) (𝐹 (𝑥 ′)) = (16.1.21)

∫ ∫
′ ′ 𝜕𝐹 ′
(2𝜋) −𝑛 e𝑖 (𝐹 ( 𝑥 )−𝐹 ( 𝑦 )) · 𝜉 𝑎 (𝐹 (𝑥 ′) , 𝐹 (𝑦 ′) , 𝜉) 𝜑 (𝐹 (𝑦 ′)) det (𝑦 ) d𝑦 ′d𝜉,
R𝑛 Ω′ 𝜕𝑥 ′
16.1 Standard Pseudodifferential Operators 579

recalling that 𝜑 ∈ Cc∞ (Ω) is arbitrary; 𝜕𝑥 𝜕𝐹

′ is the Jacobian matrix of the map 𝐹. It
is not difficult to see that there is a nonsingular 𝑛 × 𝑛 real matrix 𝐽 (𝑥 ′, 𝑦 ′) with C ∞
entries in a neighborhood of diag (Ω′ × Ω′) in Ω′ × Ω′ such that

∀ (𝑥 ′, 𝑦 ′) ∈ Ω′ × Ω′, 𝐹 (𝑥 ′) − 𝐹 (𝑦 ′) = 𝐽 (𝑥 ′, 𝑦 ′) (𝑥 ′ − 𝑦 ′) . (16.1.22)

[In a neighborhood of the diagonal 𝑥 ′ = 𝑦 ′ we have 𝐽 (𝑥 ′, 𝑦 ′) = 𝜕𝑥 𝜕𝐹 ′

′ (𝑦 ) +
⊤
𝑂 (|𝑥 ′ − 𝑦 ′ |).] In (16.1.21) we carry out the change of variable 𝜉 ↦→ 𝐽 (𝑥 ′, 𝑦 ′) −1 𝜉 ′,
whence

((Op𝑎) 𝜑) (𝐹 (𝑥 ′)) (16.1.23)

∫ ∫
𝑖 ( 𝑥 ′ −𝑦 ′ ) · 𝜉 ′
= (2𝜋) −𝑛 e 𝑎 𝐹 (𝑥 ′, 𝑦 ′, 𝜉 ′) 𝜑 (𝐹 (𝑦 ′)) d𝑦 ′d𝜉 ′,
R𝑛 Ω′

where
𝜕𝐹 ′
det 𝜕𝑥 ′ (𝑦 )
⊤
′ ′ ′ ′ ′ ′ ′ −1 ′
𝑎 𝐹 (𝑥 , 𝑦 , 𝜉 ) = 𝑎 𝐹 (𝑥 ) , 𝐹 (𝑦 ) , 𝐽 (𝑥 , 𝑦 ) 𝜉 . (16.1.24)
|det 𝐽 (𝑥 ′, 𝑦 ′)|

Proposition 16.1.23 If 𝐹 : Ω′ −→ Ω is a diffeomorphism then the maps defined by

(16.1.24),

↦ 𝑎 𝐹 ∈ 𝑆 𝑚 (Ω′ × Ω′) ,
𝑆 𝑚 (Ω × Ω) ∋ 𝑎 →
𝑚
𝑆sub (Ω × Ω) ∋ 𝑎 → 𝑚
↦ 𝑎 𝐹 ∈ 𝑆sub (Ω′ × Ω′) ,

(𝑚 ∈ R ∪ {−∞}) are Fréchet space isomorphisms. In turn, these maps define

isomorphisms Ψ𝑚 (Ω) Ψ𝑚 (Ω′), Ψprop
𝑚 (Ω) Ψ 𝑚 (Ω ′ ).
prop

Proof It suffices to prove that, if 𝑎 ∈ 𝑆 𝑚 (Ω × Ω) [resp., 𝑎 ∈ 𝑆sub

𝑚 (Ω × Ω)] then

𝑎 𝐹 ∈ 𝑆 𝑚 (Ω′ × Ω′) [resp., 𝑎 𝐹 ∈ 𝑆sub

𝑚 (Ω × Ω)], a straightforward exercise. □

Corollary 16.1.24 The diffeomorphism 𝐹 : Ω′ −→ Ω induces a linear bijection

Ψ𝑚 (Ω) /Ψ−∞ (Ω) −→ Ψ𝑚 (Ω) /Ψ−∞ (Ω) for each 𝑚 ∈ R and a ring isomorphism
of Ψ (Ω) /Ψ−∞ (Ω) onto Ψ (Ω) /Ψ−∞ (Ω).

16.1.5 Pseudodifferential operators on a manifold, between vector

bundles

Corollary 16.1.24 is the key to the definition of standard pseudodifferential operators

on a C ∞ manifold M. Indeed, given an arbitrary local coordinate chart in M,
(U, 𝑥1 , ..., 𝑥 𝑛 ) (𝑛 = dim M), we can define Ψ𝑚 (U) by transfer from an open
subset of R𝑛 using the coordinates 𝑥 𝑗 ; a change Ø of these coordinates induces an
automorphism of Ψ𝑚 (U). We write Ψ (U) = Ψ𝑚 (U).
𝑚∈R
580 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Definition 16.1.25 Let M be a C ∞ manifold countable at infinity. By a pseudodif-

ferential operator in M we shall mean a continuous linear operator 𝑨 : E ′ (M) −→
D ′ (M) whose restriction to an arbitrary domain U of local coordinates belongs to
Ψ (U). We denote by Ψ (M) the set of all pseudodifferential operators in M. The
operator 𝑨 is said to be of order 𝑚 (∈ R) if its restriction to an arbitrary domain U
of local coordinates belongs to Ψ𝑚 (U); in this case we write 𝑨 ∈ Ψ𝑚 (M).

The essence of Definition 16.1.25 is best expressed in the language of sheaves (see
Subsection
1.2.2).
Using domains of local coordinates U, V we obtain a presheaf
U
Ψ (U) , 𝜌 V with the natural restriction map (cf. Definition 2.3.9): assuming V ⊂
U, this is the composite

𝐿 (E ′ (U) ; D ′ (U)) −→ 𝐿 (E ′ (V) ; D ′ (U)) → 𝐿 (E ′ (V) ; D ′ (V))

where the first arrow is the restriction of maps from E ′ (U) to E ′ (V) and the second
arrow is composition of the map with the restriction D ′ (U) −→ D ′ (V). We could
as well have used Cc∞ in the place of E ′ and C ∞ in that of D ′.
If U ∩ V ≠ ∅ two operators 𝑨 ∈ Ψ (U) and 𝑩 ∈ Ψ (V) equal in U ∩ V define
an element of Ψ (U ∪ V); in other words, a coherent system of locally defined
operators can be patched together to define a global element. The presheaf gives
rise to the sheaf of germs of pseudodifferential operators in M whose continuous
sections are the pseudodifferential operators in open subsets of M.
Now let 𝑽 1 , 𝑽 2 be two complex vector bundles on M (see Section 9.2). We need to
define pseudodifferential operators E ′ (M; 𝑽 1 ) −→ D ′ (M; 𝑽 2 ). For convenience
we define these as continuous linear operators Cc∞ (M; 𝑽 1 ) −→ C ∞ (M; 𝑽 2 ). (If
U is an open subset of M, C ∞ (U; 𝑽 ℓ ) [resp., Cc∞ (U; 𝑽 ℓ )] denotes the space
of (resp., compactly supported) C ∞ sections of 𝑽 ℓ over U.) Simplest is to use
local trivializations of the vector bundles. Let (U, 𝑥1 , ..., 𝑥 𝑛 ) be a local chart such
that there exist vector bundle isomorphisms 𝜒ℓ : 𝑽 ℓ | U −→ 𝑈×C 𝑁ℓ , ℓ = 1, 2,
where the open subset 𝑈 of R𝑛 is the image of U under the map ℘ ↦→ 𝑥 (℘) =
(𝑥1 (℘) , . . . , 𝑥 𝑛 (℘)) ∈ R𝑛 . By a pseudodifferential operator 𝑨 : Cc∞ (𝑈) 𝑁1 −→
C ∞ (𝑈) 𝑁2 we mean an 𝑁1 × 𝑁2 matrix whose entries are (scalar) pseudodifferential
operators 𝑨 𝑗,𝑘 : Cc∞ (𝑈) −→ C ∞ (𝑈). The isomorphisms 𝜒ℓ define pullback maps
𝜒ℓ∗ : C ∞ (𝑈) 𝑁ℓ −→ C ∞ (U; 𝑽 ℓ ) and 𝜒ℓ∗ : Cc∞ (𝑈) 𝑁ℓ −→ Cc∞ (U; 𝑽 ℓ ); 𝜒ℓ∗ is an
isomorphism. Then the sought pseudodifferential operator is the map represented by
the lower horizontal arrow in the commutative diagram

Cc∞ (𝑈) 𝑁1 −→ C ∞ (𝑈) 𝑁2

𝜒1∗ ↓ 𝜒2∗ ↓
𝑨
Cc∞ (U; 𝑽 1 ) −→ C ∞ (U; 𝑽 2 ) .
16.2 Symbolic Calculus 581

Using coverings of M by domains of local trivializations of both vector bundles we

patch up the local definitions in the same manner as is done when 𝑽 1 = 𝑽 2 = M × C.
Or better, by using the local definitions we can define the sheaf of germs of pseudod-
ifferential operators 𝑽 1 −→ 𝑽 2 whose continuous sections are the pseudodifferential
operators 𝑽 1 −→ 𝑽 2 in open subsets of M.

16.2 Symbolic Calculus

16.2.1 Standard symbols

In this section Ω will be an arbitrary open subset of R𝑛 .

Definition 16.2.1 By a standard symbol (or simply, a symbol) of order 𝑚 ∈ R in

Ω we shall mean an amplitude 𝑎 (𝑥, 𝜉) of order 𝑚 in Ω × R𝑛 independent of 𝑦.

The standard symbols 𝑎 (𝑥, 𝜉) of order 𝑚 in Ω are characterized by the following

property:
• Given any compact subset 𝐾 of Ω and any pair of multi-indices 𝛼, 𝛽 ∈ Z+𝑛 there
is a 𝐶𝐾 , 𝛼,𝛽 > 0 such that

∀ (𝑥, 𝜉) ∈ 𝐾 × R𝑛 , 𝜕𝑥𝛼 𝜕 𝜉 𝑎 (𝑥, 𝜉) ≤ 𝐶𝐾 , 𝛼,𝛽 (1 + |𝜉 |) 𝑚−|𝛽 | .

𝛽
(16.2.1)

We shall denote by 𝑆 𝑚 (Ω) the linear space of standard symbols of order 𝑚 in Ω;

by 𝑆 (Ω) the union of all the spaces 𝑆 𝑚 (Ω), 𝑚 ∈ R; by 𝑆 −∞ (Ω) their intersection.
Note that 𝑆 (Ω) is a ring with respect to addition and ordinary multiplication; 𝑆 0 (Ω)
is a subring and 𝑆 −∞ (Ω) is an ideal in 𝑆 (Ω).
We also introduce the substandard symbols: they are substandard amplitudes
in Ω × R𝑛 (Definition 16.1.2) independent of 𝑦. The set ofØ substandard symbols of
order 𝑚 will be denoted by 𝑆sub (Ω); we write 𝑆sub (Ω) =
𝑚 𝑚 (Ω).
𝑆sub
𝑚∈R
𝑚 (Ω) then, for 𝑢 ∈ C ∞ (R𝑛 ) and 𝑥 ∈ Ω,
If 𝑎 ∈ 𝑆sub c
∫
(Op𝑎) 𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝜉) 𝑢 (𝑦) d𝑦d𝜉 (16.2.2)
R2𝑛
∫
= (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 𝑎 (𝑥, 𝜉) b𝑢 (𝜉) d𝜉 = 𝑎 (𝑥, D) 𝑢 (𝑥) ,
R𝑛

𝑎 (𝑥, D) being the commonly used notation for Op𝑎 when 𝑎 is a symbol, in analogy
with differential operators, when 𝑎 (𝑥, 𝜉) is a polynomial in the 𝜉 variables.
Dealing with symbols rather than amplitudes allows us to make Theorem 16.1.19
more “uniform” on the 𝑦-side.
582 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Proposition 16.2.2 Let 𝑠, 𝑚 ∈ R and 𝑎 (𝑥, 𝜉) ∈ 𝑆sub

𝑚 (Ω) be arbitrary. The linear

operator 𝑎 (𝑥, D) : 𝐻c𝑠 (R𝑛 ) −→ 𝐻loc

𝑠−𝑚 (Ω)
(Theorem 16.1.19) can be extended as
a continuous linear operator 𝐻 𝑠 (R𝑛 ) −→ 𝐻loc
𝑠−𝑚 (Ω).

Proof It suffices to duplicate step by step the proof of Theorem 16.1.19 after elim-
inating 𝑦 and 𝜂 at every step. We end up with the estimate of a single convolution
𝑓 ∗ 𝑔. □
The distribution kernel corresponding to the operator 𝑎 (𝑥, D) is e 𝑎 (𝑥, 𝑥 − 𝑦)
𝑎 (𝑥, 𝑦) = (2𝜋) −𝑛 R𝑛 e𝑖𝑦· 𝜂 𝑎 (𝑥, 𝜂) d𝜂, the inverse Fourier transform of the C ∞
∫
if e
function (with tempered growth) 𝜂 ↦→ 𝑎 (𝑥, 𝜂). Notice that the distribution kernel
𝑎 (𝑥, 𝑥 − 𝑦) ∈ D ′ (Ω × R𝑛 ) cannot be properly supported (Definition 2.3.6).
e
In the sequel, unless otherwise specified, we shall regard 𝑎 (𝑥, D) as an element of
Ψ (Ω), i.e., a linear operator E ′ (Ω) −→ D ′ (Ω), rather than an operator E ′ (R𝑛 ) −→
D ′ (Ω). This is convenient when dealing with the composition of operators. Recall
that Ψ−∞ (Ω) is the ring of smoothing operators, (16.1.18).

Proposition 16.2.3 If 𝑎 ∈ 𝑆 (Ω), 𝑎 (𝑥, D) ∈ Ψ−∞ (Ω) implies 𝑎 ∈ 𝑆 −∞ (Ω).

Proof If 𝑎 (𝑥, D) ∈ Ψ−∞ (Ω) then e 𝑎 (𝑥, 𝑥 − 𝑦) ∈ C ∞ (Ω). Let Ω′ ⊂⊂ Ω be open and
𝜑 ∈ Cc (Ω), 𝜑 (𝑥) = 1 for all 𝑥 ∈ Ω′. Consider the integral
∞

∫
𝑏 (𝑥, 𝜉) = e𝑖𝑦· 𝜉 e
𝑎 (𝑥, 𝑥 − 𝑦) 𝜑 (𝑦) d𝑦
R𝑛
∫
= (2𝜋) −𝑛 e𝑖 𝑥· 𝜂+𝑖𝑦· ( 𝜉 −𝜂) 𝑎 (𝑥, 𝜂) 𝜑 (𝑦) d𝑦d𝜂.
R2𝑛

On the one hand, 𝑏 (𝑥, 𝜉) ∈ 𝑆 −∞ (Ω) since we have, for any 𝛼, 𝛽 ∈ Z+𝑛 and 𝑘 ∈ Z+ ,
𝑘 ∫
2 𝛼 𝛽 𝛽
1 + |𝜉 | 𝜕𝑥 𝜕𝑦 𝑏 (𝑥, 𝜉) ≤ (1 − Δ 𝑦 ) 𝑘 𝜕𝑥𝛼 𝜕𝑦 (e
𝑎 (𝑥, 𝑥 − 𝑦) 𝜑 (𝑦)) d𝑦.
R𝑛

As a consequence e−𝑖 𝑥· 𝜉 𝑏 (𝑥, 𝜉) ∈ 𝑆 −∞ (Ω). On the other hand, using finite Taylor
expansions of 𝑎 (𝑥, 𝜂) about 𝜉, we get for any integer 𝑁 ≥ 0,
∫
(2𝜋) e 𝑛 −𝑖 𝑥· 𝜉
𝑏 (𝑥, 𝜉) = e𝑖 ( 𝑥−𝑦) · ( 𝜉 −𝜂) 𝑎 (𝑥, 𝜂) 𝜑 (𝑦) d𝑦d𝜂
R2𝑛
∑︁ 1 ∫
(𝜂 − 𝜉) 𝛾 e𝑖 ( 𝑥−𝑦) · ( 𝜉 −𝜂) 𝜑 (𝑦) d𝑦d𝜂
𝛾
= 𝜕 𝜉 𝑎 (𝑥, 𝜉)
𝛾! R 2𝑛
|𝛾 | ≤ 𝑁
∑︁ 1 ∫
+ (𝑁 + 1) (𝜂 − 𝜉) 𝛾 e𝑖 ( 𝑥−𝑦) · ( 𝜉 −𝜂) e
𝑎 𝛾 (𝑥, 𝜉, 𝜂) 𝜑 (𝑦) d𝑦d𝜂,
𝛾! R2𝑛
|𝛾 |=𝑁 +1

where ∫ 1
𝛾
𝑎 𝛾 (𝑥, 𝜉, 𝜂) =
e 𝑡 𝑁 𝜕 𝜉 𝑎 (𝑥, 𝜂 + 𝑡 (𝜉 − 𝜂)) d𝑡.
0
16.2 Symbolic Calculus 583

We have
∫
(𝜂 − 𝜉) 𝛾 e𝑖 ( 𝑥−𝑦) · ( 𝜉 −𝜂) 𝜑 (𝑦) d𝑦d𝜂
R2𝑛
∫
= (−1) |𝛾 | e𝑖 ( 𝑥−𝑦) · ( 𝜉 −𝜂) D𝛾 𝜑 (𝑦) d𝑦d𝜂 = (2𝜋) 𝑛 D𝛾 𝜑 (𝑥) ,
R2𝑛

which vanishes if 𝑥 ∈ Ω′ and |𝛾| ≥ 1. Thus, if 𝑥 ∈ Ω′ we obtain

(2𝜋) 𝑛 e−𝑖 𝑥· 𝜉 𝑏 (𝑥, 𝜉) = 𝑎 (𝑥, 𝜉)

∑︁ (−1) |𝛾 | ∫
𝛾
+ (𝑁 + 1) e𝑖 𝑥· 𝜂 e
𝑎 𝛾 (𝑥, 𝜉, 𝜉 − 𝜂) D
d
𝑦 𝜑 (𝜂) d𝜂.
𝛾! R 2𝑛
|𝛾 |=𝑁 +1

To every 𝛼 ∈ Z+𝑛 there is a constant 𝐶 𝛼 > 0 such that, for all 𝑥 ∈ Ω and all
(𝜉, 𝜂) ∈ R2𝑛 ,

∑︁ ∫ 1 𝑚−𝑁 −1
𝜕𝑥𝛼 e
𝑎 𝛾 (𝑥, 𝜉, 𝜂) ≤ 𝐶 𝛼 1 + |𝜂 + 𝑡 (𝜉 − 𝜂)| d𝑡 .
|𝛾 |=𝑁 +1 0

Given any 𝛼 ∈ Z+𝑛 and any 𝑘 ∈ Z+ there is a 𝐶 𝛼,𝑘 > 0 such that

𝛾 𝛾 −𝑘
𝜕𝑥𝛼 e𝑖 𝑥· 𝜂 D
d 𝛼d
𝑦 𝜑 (𝜂) = 𝜂 D 𝑦 𝜑 (𝜂) ≤ 𝐶 𝛼,𝑘 (1 + |𝜂|) .

We leave it as an exercise to show that these two estimates combined imply that to
each 𝛼 ∈ Z+𝑛 there is a 𝐶 𝛼′ > 0 such that, for all 𝑥 ∈ Ω′,
∫
𝛾 ′ 𝑚−𝑁 −1
𝜕𝑥𝛼 e𝑖 𝑥· 𝜂 e
𝑎 𝛾 (𝑥, 𝜉, 𝜉 − 𝜂) D
d
𝑦 𝜑 (𝜂) d𝜂 ≤ 𝐶 𝛼 (1 + |𝜉 |) .
R2𝑛

This proves that 𝑎 (𝑥, 𝜉)−e−𝑖 𝑥· 𝜉 𝑏 (𝑥, 𝜉) ∈ 𝑆 𝑚−𝑁 −1 (Ω′) hence 𝑎 (𝑥, 𝜉) ∈ 𝑆 𝑚−𝑁 −1 (Ω′),
whatever the integer 𝑁. □

Corollary 16.2.4 The map 𝑆 (Ω) ∋ 𝑎 (𝑥, 𝜉) ↦→ 𝑎 (𝑥, D) ∈ Ψ (Ω) induces a linear
injection 𝑆 (Ω) /𝑆 −∞ (Ω) −→ Ψ (Ω) /Ψ−∞ (Ω).

In the next subsection we are going to show that the induced injection in Corollary
16.2.4 is a bijection.

16.2.2 Formal symbols

Definition 16.2.5 Let 𝑚 ∈ R. By a formal symbol of order 𝑚 in Ω we shall mean a

formal series ∞
Í
𝑗=0 𝑎 𝑗 (𝑥, 𝜉) with 𝑎 𝑗 ∈ 𝑆
𝑚− 𝑗 (Ω). We denote by 𝑆 𝑚 (Ω) the linear
form
space of formal symbols of order 𝑚 in Ω.
584 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

𝑚2
We point out that if 𝑚 1 < 𝑚 2 then 𝑆form (Ω) can
Øbe viewed as a linear subspace of
𝑚1
𝑆form (Ω). This allows us to define 𝑆form (Ω) = 𝑚 (Ω). Obviously, 𝑆
𝑆form form (Ω)
𝑚∈R
𝑚 (Ω) (−∞ < 𝑚 ≤ 0) are commutative rings with respect to addition and
and 𝑆form
𝑚 (Ω) an ideal in 𝑆 0
ordinary multiplication, with 𝑆form (Ω).
form
The relevance of formal symbols to pseudodifferential operators is rooted in
the twin facts that any standard amplitude 𝑎 ∈ 𝑆 𝑚 (Ω × Ω) determines unambigu-
ously a formal symbol, and that any formal symbol determines an equivalence class
mod 𝑆 −∞ (Ω) of “true” standard symbols (Definition 16.2.1).
We introduce straightaway the map amplitudes in Ω × Ω −→formal symbols in Ω:
to an arbitrary 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 (Ω × Ω) it assigns the following element of 𝑆form
𝑚 (Ω):

∑︁ 1
𝛽 𝛽
exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥
= 𝜕𝑦 D 𝜉 𝑎 (𝑥, 𝑥, 𝜉) , (16.2.3)
𝛽 ∈Z𝑛
𝛽!
+

where
𝑛
1 ∑︁ 𝜕 2
D𝑦 · 𝜕𝜉 = √ .
−1 𝑗=1 𝜕𝑦 𝑗 𝜕𝜉 𝑗
When 𝑎 itself is a symbol, the right-hand side in (16.2.3) reduces to 𝑎 (𝑥, 𝜉).
Remark 16.2.6 We recall (see Ch. 13) that a linear symplectic transformation in
phase-space R2𝑛 is a linear automorphism (𝑥, 𝜉) ↦→ (𝑦, 𝜂) such that d𝑦∧d𝜂 = d𝑥 ∧d𝜉
and therefore such that D 𝑦 · 𝜕𝜂 = D 𝑥 · 𝜕 𝜉 .
Now consider the finite sums
∑︁ 1
𝛽 𝛽
𝑎 𝑁 (𝑥, 𝜉) = 𝜕𝑦 D 𝜉 𝑎 (𝑥, 𝑥, 𝜉) ∈ 𝑆 𝑚 (Ω) . (16.2.4)
𝛽!
|𝛽 | ≤𝑁

Proposition 16.2.7 If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 (Ω × Ω) (𝑚 ∈ R) and 𝑁 ∈ Z+ then Op𝑎 −

𝑎 𝑁 (𝑥, D) ∈ Ψ𝑚−𝑁 −1 (Ω).
Proof Let 𝜑 𝑗 𝑗=1,2,... ⊂ Cc∞ (Ω) be a locally finite partition of unity . It suffices to

prove the claim when we replace 𝑎 (𝑥, 𝑦, 𝜉) by 𝜑 𝑗 (𝑦) 𝑎 (𝑥, 𝑦, 𝜉) and then sum over
𝑗. We can select the cutoffs 𝜑 𝑗 so that supp 𝜑 𝑗 is convex. We may as well assume
that Ω itself is convex, in which case
∑︁ 1
𝛽
𝑎 (𝑥, 𝑦, 𝜉) = (𝑦 − 𝑥) 𝛽 𝜕𝑦 𝑎 (𝑥, 𝑥, 𝜉) (16.2.5)
𝛽!
|𝛽 | ≤ 𝑁
∑︁ 1
+ (𝑁 + 1) (𝑦 − 𝑥) 𝛽 e
𝑎 𝛽 (𝑥, 𝑦, 𝜉) ,
𝛽!
|𝛽 |=𝑁 +1

where
∫ 1
𝛽
𝑎 𝛽 (𝑥, 𝑦, 𝜉) =
e 𝑡 𝑁 𝜕𝑦 𝑎 (𝑥, 𝑡𝑥 + (1 − 𝑡) 𝑦, 𝜉) d𝑡 ∈ 𝑆 𝑚 (Ω × Ω) .
0
16.2 Symbolic Calculus 585

We derive from (16.2.5), for 𝑢 ∈ E ′ (Ω),

∑︁ 1 ∫
−D 𝜉 e𝑖 ( 𝑥−𝑦) · 𝜉 𝜕𝑦 𝑎 (𝑥, 𝑥, 𝜉) 𝑢 (𝑦) d𝑦d𝜉
𝛽 𝛽
(2𝜋) 𝑛 (Op𝑎) 𝑢 (𝑥) =
𝛽! R2𝑛
|𝛽 | ≤𝑁
∑︁ 1 ∫
−D 𝜉 e𝑖 ( 𝑥−𝑦) · 𝜉 e
𝛽
+ (𝑁 + 1) 𝑎 𝛽 (𝑥, 𝑦, 𝜉) 𝑢 (𝑦) d𝑦d𝜉;
𝛽! R2𝑛
|𝛽 |=𝑁 +1

and then, after integrations by parts with respect to 𝜉,

(2𝜋) 𝑛 (Op𝑎) 𝑢 (𝑥)

∑︁ 1 ∫
𝛽 𝛽
= e𝑖 𝑥· 𝜉 𝜕𝑦 D 𝜉 𝑎 (𝑥, 𝑥, 𝜉) b 𝑢 (𝜉) d𝜉
𝛽! R2𝑛
|𝛽 | ≤ 𝑁
∑︁ 1 ∫
e𝑖 ( 𝑥−𝑦) · 𝜉 D 𝜉 e
𝛽
+ (𝑁 + 1) 𝑎 𝛽 (𝑥, 𝑦, 𝜉) 𝑢 (𝑦) d𝑦d𝜉.
𝛽! R𝑛
|𝛽 |=𝑁 +1

We have
∑︁ 1 𝛽
𝑏 𝑁 (𝑥, 𝑦, 𝜉) = (𝑁 + 1) D e𝑎 𝛽 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚−𝑁 −1 , (16.2.6)
𝛽! 𝜉
|𝛽 |=𝑁 +1

and obviously Op𝑏 𝑁 = Op𝑎 − 𝑎 𝑁 (𝑥, D). □

If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 (Ω × Ω) (𝑚 ∈ R) then 𝑎 ⊤ (𝑥, 𝑦, 𝜉) = 𝑎(𝑦, 𝑥, −𝜉) and
𝑎∗ (𝑥, 𝑦, 𝜉) = 𝑎(𝑦, 𝑥, 𝜉) also belong to 𝑆 𝑚 (Ω × Ω) (Proposition 16.1.6); Op𝑎 ⊤ is
the transpose of Op𝑎, and Op𝑎 ∗ is the adjoint of Op𝑎. According to (16.2.3) the
formal symbols associated to these amplitudes are, respectively,
∑︁ (−1) |𝛽 |
𝛽 𝛽
exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑦, 𝑥, −𝜉) 𝑦=𝑥
= 𝜕𝑥 D 𝜉 𝑎 (𝑥, 𝑥, −𝜉) , (16.2.7)
𝛽 ∈Z+𝑛
𝛽!
∑︁ 1
𝛽 𝛽
exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑦, 𝑥, 𝜉) = 𝜕𝑥 D 𝜉 𝑎¯ (𝑥, 𝑥, 𝜉) . (16.2.8)
𝑦=𝑥
𝛽 ∈Z𝑛
𝛽!
+

Let us now go the other way, from formal symbols to true, standard symbols. In a
generally divergent series (this is also true in the C 𝜔 category) 𝑎 = ∞
Í
𝑗=0 𝑎 𝑗 (𝑥, 𝜉) ∈
𝑚 (Ω) we shall insert appropriate cutoffs to construct a convergent analogue.
𝑆form
In the C 𝜔 category the cutoffs will have to be selected with much greater care.
One begins by selecting a sequence of compact subsets 𝐾 𝑗 of Ω, 𝑗 ∈ Z+ , such
∞
◦ Ø
that 𝐾 𝑗 ⊂ 𝐾 𝑗+1 , the interior of 𝐾 𝑗+1 ; and such that Ω = 𝐾 𝑗 (such a sequence
𝑗=0
is commonly referred to as an exhausting sequence of compact subsets in Ω). By
(16.2.1), given any 𝑗 ∈ Z+ and any pair (𝛼, 𝛽) ∈ Z2𝑛
+ there is a constant 𝐶 𝑗;𝛼,𝛽 > 0
such that
586 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

∀ (𝑥, 𝜉) ∈ 𝐾 𝑗 × R𝑛 , 𝜕𝑥𝛼 𝜕 𝜉 𝑎 𝑗 (𝑥, 𝜉) ≤ 𝐶 𝑗;𝛼,𝛽 (1 + |𝜉 |) 𝑚− 𝑗− |𝛽 | .

𝛽
(16.2.9)

Let the function 𝜒 ∈ C ∞ (R𝑛 ) be such that 𝜒 (𝜉) = 0 for |𝜉 | < 1, 𝜒 (𝜉) = 1 for
|𝜉 | > 2. Given any sequence of numbers 𝑅 𝑗 ↗ +∞, 𝑗 = 1, 2, ..., the series
∞
∑︁
𝑎♭ (𝑥, 𝜉) = 𝜒 𝜉/𝑅 𝑗 𝑎 𝑗 (𝑥, 𝜉) (16.2.10)
𝑗=0

is locally finite in 𝜉-space: it terminates when 𝑗 is so large that |𝜉 | < 𝑅 𝑗 . Thus it

defines a function 𝑎♭ ∈ C ∞ (Ω × R𝑛 ). If the numbers 𝑅 𝑗 increase sufficiently fast
as 𝑗 ↗ +∞ then 𝑎♭ (𝑥, 𝜉) will be a standard symbol of order 𝑚. This claim is a
straightforward consequence of the inequalities (16.2.9) and its proof is left as an
exercise.

Remark 16.2.8 Constructing a series such as (16.2.10) does not require that 𝑎 𝑗 (𝑥, 𝜉)
be known, or even defined, for |𝜉 | < 𝑅 𝑗 . This is the base on which the classical
symbolic calculus is built (see Subsection 16.2.5 below).

The construction (16.2.10) applied to the formal symbol (16.2.3) produces a

standard symbol of order 𝑚,
∞
∑︁ ∑︁ 1 𝛼 𝛼
𝑎♭ (𝑥, 𝜉) = 𝜒 𝜉/𝑅 𝑗 𝜕 D 𝑎 (𝑥, 𝑥, 𝜉) . (16.2.11)
𝛼! 𝑦 𝜉
𝑗=0 | 𝛼 |= 𝑗

Proposition 16.2.9 If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 (Ω × Ω) and if 𝑎♭ (𝑥, 𝜉) ∈ 𝑆 𝑚 (Ω) is given by

(16.2.11) then Op𝑎 − 𝑎♭ (𝑥, D) is smoothing.

Proof On the one hand, using the notation of Proposition 16.2.7 we know that
Op𝑎 − 𝑎 𝑁 (𝑥, D) ∈ Ψ𝑚−𝑁 −1 (Ω) for every 𝑁 ∈ Z+ . On the other hand it is immediate
that 𝑎♭ (𝑥, D) − 𝑎 𝑁 (𝑥, D) ∈ Ψ𝑚−𝑁 −1 (Ω), whence Op𝑎 − 𝑎♭ (𝑥, D) ∈ Ψ𝑚−𝑁 −1 (Ω).
Letting 𝑁 → +∞ proves the claim. □
It is useful to summarize what we have obtained so far.

Theorem 16.2.10 The map 𝑆 (Ω) ∋ 𝑎 (𝑥, 𝜉) ↦→ 𝑎 (𝑥, D) ∈ Ψ (Ω) induces a linear
bijection 𝔒𝔭 : 𝑆 (Ω) /𝑆 −∞ (Ω) −→ Ψ (Ω) /Ψ−∞ (Ω).

Proof By Corollary 16.2.4 we know that the map 𝔒𝔭 is injective. Consider an

arbitrary operator Op𝑎 ∈ Ψ (Ω), 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 (Ω × Ω). By composing the maps

𝑆 𝑚 (Ω × Ω) ∋ 𝑎 (𝑥, 𝑦, 𝜉) ↦→ exp 𝐷 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 ∈ 𝑆form (Ω) ,
𝑆form (Ω) ∋ 𝑎 (𝑥, 𝜉) ↦→ 𝑎♭ (𝑥, 𝜉) ∈ 𝑆 𝑚 (Ω) ,
𝑆 𝑚 (Ω) ∋ 𝑎♭ (𝑥, 𝜉) ↦→ [𝑎♭ ] ∈ 𝑆 𝑚 (Ω) /𝑆 −∞ (Ω) ,

we obtain a map 𝑇 : Ψ (Ω) −→ 𝑆 𝑚 (Ω) /𝑆 −∞ (Ω). If Op𝑎 ∈ Ψ−∞ (Ω × Ω) then

𝑎♭ (𝑥, D) ∈ Ψ−∞ (Ω) by Proposition 16.2.9, which proves that 𝑇 induces a map
16.2 Symbolic Calculus 587

𝑇˜ : Ψ (Ω) /Ψ−∞ (Ω) −→ 𝑆 𝑚 (Ω) /𝑆 −∞ (Ω) .

It is easily checked that 𝔒𝔭 ◦ 𝑇˜ is equal to the identity map of Ψ (Ω) /Ψ−∞ (Ω). □

Corollary 16.2.11 Let Ω and Ω′ be two open subsets of R𝑛 . A diffeomorphism

𝐹 : Ω′ −→ Ω induces a linear bijection 𝑆 𝑚 (Ω) /𝑆 −∞ (Ω) −→ 𝑆 𝑚 (Ω′) /𝑆 −∞ (Ω′)
for each 𝑚 ∈ R.

Proof Combine Theorem 16.2.10 with Corollary 16.1.24. □

Formula (16.1.24) assigns to 𝑎 ∈ 𝑆 𝑚 (Ω) the amplitude
𝜕𝐹 ′
det 𝜕𝑥 ′ (𝑦 )
⊤
′ ′ ′ ′ ′ ′ −1 ′
𝑎 𝐹 (𝑥 , 𝑦 , 𝜉 ) = 𝑎 𝐹 (𝑥 ) , 𝐽 (𝑥 , 𝑦 ) 𝜉 , (16.2.12)
|det 𝐽 (𝑥 ′, 𝑦 ′)|

which determines a formal symbol 𝑎 𝐹 (𝑥 ′, 𝜉 ′) by applying (16.2.3); in turn, by

(16.2.9) 𝑎 𝐹 (𝑥 ′, 𝜉 ′) determines a coset of standard symbols mod 𝑆 −∞ (Ω′).
If 𝑎 ∈ 𝑆 𝑚 (Ω) the transpose (Op𝑎) ⊤ of Op𝑎 is equal to Op𝑎 ⊤ where 𝑎 ⊤ is the
amplitude 𝑎 (𝑦, −𝜉), which determines, according to (16.2.7), the formal symbol
∑︁ (−1) |𝛽 |
𝛽 𝛽
exp D 𝑥 · 𝜕 𝜉 𝑎 (𝑥, −𝜉) = 𝜕𝑥 D 𝜉 𝑎 (𝑥, −𝜉) . (16.2.13)
𝛽 ∈Z𝑛
𝛽!
+

𝑚 (Ω).
This can also be taken as the effect of transposition on a formal symbol 𝑎 ∈ 𝑆form
∗ ∗
The amplitude of the adjoint of Op𝑎 is equal to Op𝑎 , where 𝑎 is the amplitude
𝑎 (𝑦, 𝜉), which determines, according to (16.2.8), the formal symbol
∑︁ 1
𝛽 𝛽
exp D 𝑥 · 𝜕 𝜉 𝑎 (𝑥, 𝜉) = 𝜕𝑥 D 𝜉 𝑎 (𝑥, 𝜉) . (16.2.14)
𝛽 ∈Z 𝑛 𝛽!
+

We associate to (16.2.13) and (16.2.14) unique cosets

Í∞of symbols Í∞mod 𝑆 −∞ (Ω).
It is obvious that two different formal symbols 𝑗=0 𝑎 𝑗 and 𝑗=0 𝑏 𝑗 may define
the same equivalence class of standard pseudodifferential operators mod Ψ−∞ (Ω);
in this case, 𝑎 𝑗 − 𝑏 𝑗 ∈ 𝑆 −∞ (Ω) for every 𝑗 (cf. Propositions 16.2.3, 16.2.7). The
most convenient way to handle such equivalent pairs of formal symbols is to extend
Definition 16.2.5 to the case 𝑚 = −∞, in other words to introduce “formal series”
(essentially, sequences) of symbols belonging to 𝑆 −∞ (Ω) and call 𝑆form −∞ (Ω) the
−∞
space they form, and then reason mod 𝑆form (Ω). This makes sense of the following
statement, a direct consequence of the preceding considerations.

Proposition 16.2.12 Let 𝑎 ∈ 𝑆 (Ω × Ω). For Op𝑎 to be congruent mod Ψ−∞ (Ω) to
a self-adjoint pseudodifferential operator in Ω it is necessary and sufficient that
∑︁ 1 ∑︁ 1
𝛽 𝛽 𝛽 𝛽 −∞
𝜕𝑦 𝐷 𝜉 𝑎 (𝑥, 𝑥, 𝜉) 𝜕𝑥 𝐷 𝜉 𝑎 (𝑥, 𝑥, 𝜉) mod 𝑆form (Ω) .
𝛽 ∈Z 𝑛 𝛽! 𝛽 ∈Z 𝑛 𝛽!
+ +
(16.2.15)
588 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

16.2.3 Composition of symbols

In this subsection we seek the formula for the composition of symbols corresponding
to the composition of pseudodifferential operators. There is no problem with the latter
when the operators are properly supported. Otherwise, rather than thinking of the
operators we must think of their equivalence classes modulo smoothing operators,
the elements of Ψ (Ω) /Ψ−∞ (Ω), whose composition is defined by that of properly
supported representatives (see Proposition 16.1.17 and following remarks). It is then
natural to look at formal symbols, or at cosets in 𝑆 (Ω) /𝑆 −∞ (Ω).
We go back to the composite Op𝑎 1,2 = Op𝑎 1 ◦ Op𝑎 2 , 𝑎 𝑗 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 𝑗 (Ω × Ω),
𝑗 = 1, 2. Since we are willing to reason mod Ψ−∞ (Ω) we are allowed by Proposition
16.2.9 to assume that both amplitudes 𝑎 1 and 𝑎 2 are (true) symbols. In the integrals
below we can assume that the base projection of supp 𝑎 2 is contained in a compact
subset of Ω; in the end all symbolic formulas can be extended to general symbols.
Under these hypotheses we have [see (16.1.12)]
∫
′ ′
𝑎 1,2 (𝑥, 𝜉) = (2𝜋) −𝑛 e−𝑖 ( 𝑥−𝑥 ) · ( 𝜉 − 𝜉 ) 𝑎 1 (𝑥, 𝜉 ′) 𝑎 2 (𝑥 ′, 𝜉) d𝑥 ′d𝜉 ′,
R2𝑛

where (𝑥, 𝜉) ∈ Ω × R𝑛 . Formal Taylor expansion yields

(2𝜋) 𝑛 𝑎 1,2 (𝑥, 𝜉)

∑︁ 1 ∫
′ ′
(𝑖 (𝜉 ′ − 𝜉)) 𝛾 e−𝑖 ( 𝑥−𝑥 ) · ( 𝜉 − 𝜉 ) 𝑎 2 (𝑥 ′, 𝜉) d𝑥 ′d𝜉 ′
𝛾
𝐷 𝜉 𝑎 1 (𝑥, 𝜉)
𝛾 ∈Z+𝑛
𝛾! R 2𝑛

∑︁ 1 ∫
′ ′
(−𝜕𝑥′ ) 𝛾 e−𝑖 ( 𝑥−𝑥 ) · ( 𝜉 − 𝜉 ) 𝑎 2 (𝑥 ′, 𝜉) d𝑥 ′d𝜉 ′
𝛾
𝐷 𝜉 𝑎 1 (𝑥, 𝜉)
𝛾 ∈Z+𝑛
𝛾! R2𝑛
∑︁ 1 ∫
′ ′
e−𝑖 ( 𝑥−𝑥 ) · ( 𝜉 − 𝜉 ) 𝜕𝑥 𝑎 2 (𝑥 ′, 𝜉) d𝑥 ′d𝜉 ′.
𝛾 𝛾
= D 𝜉 𝑎 1 (𝑥, 𝜉)
𝛾 ∈Z𝑛
𝛾! R 2𝑛
+

We can carry out the integration with respect to 𝜉 ′ in each integral in the last formal
series, to get 𝑎 1,2 𝑎 1 #𝑎 2 , where
∑︁ 1
𝛾 𝛾
𝑎 1 #𝑎 2 = 𝜕𝜉 𝑎1 D𝑥 𝑎2 . (16.2.16)
𝛾 ∈Z𝑛
𝛾!
+

This can be rewritten as [cf. (16.2.3)]

(𝑎 1 #𝑎 2 ) (𝑥, 𝜉) = exp D 𝑦 · 𝜕 𝜉 𝑎 1 (𝑥, 𝜉) 𝑎 2 (𝑦, 𝜂) 𝑦=𝑥, 𝜂= 𝜉
. (16.2.17)

It is immediately checked that the right-hand side in Formula (16.2.16) is a well

defined formal symbol for any pair of formal (not just “true”) symbols
∞
∑︁ 𝑚
𝑎 𝑗 (𝑥, 𝜉) = 𝑗
𝑎 𝑗,𝑘 (𝑥, 𝜉) ∈ 𝑆form (Ω) , 𝑗 = 1, 2. (16.2.18)
𝑘=0
16.2 Symbolic Calculus 589

The composition law # is not commutative; its associativity can be easily verified (it
also ensues directly from that of the composition of operators); it has a unit element,
the symbol identically equal to 1, here denoted by 1. Equipped with the composition
# the vector space 𝑆form (Ω) becomes an associative ring.
We can then go from formal symbols to (classes of) true symbols. If 𝑎 𝑗 ∈ 𝑆 𝑚 𝑗 (Ω),
𝑗 = 1, 2, we can form their composite (16.2.16) but in general it is a formal symbol,
not a true one. We can then form a symbol (𝑎 1 #𝑎 2 ) ♭ ∈ 𝑆 𝑚1 +𝑚2 (Ω) according to
the prescription (16.2.10) and from there go to the equivalence class of (𝑎 1 #𝑎 2 ) ♭
mod 𝑆 −∞ (Ω) (cf. the proof of Theorem 16.2.10). Note that if either 𝑎 1 or 𝑎 2 belongs
to 𝑆 −∞ (Ω) the same is true of (𝑎 1 #𝑎 2 ) ♭ . We can therefore define a composition law,
also denoted by #, on the quotient space 𝑆 (Ω) /𝑆 −∞ (Ω), turning the latter into a
ring. We easily reach the following conclusion.
Theorem 16.2.13 The map 𝔒𝔭 : 𝑆 (Ω) /𝑆 −∞ (Ω) −→ Ψ𝑚 (Ω) /Ψ−∞ (Ω) is a ring
isomorphism.
𝑚1 +𝑚2
We derive from (16.2.16) and (16.2.18) that 𝑎 1 #𝑎 2 ∈ 𝑆form (Ω) and
𝑚1 +𝑚2 −1
[𝑎 1 , 𝑎 2 ] # = 𝑎 1 #𝑎 2 − 𝑎 2 #𝑎 1 ∈ 𝑆form (Ω) , (16.2.19)

implying
[𝔒𝔭𝑎 1 , 𝔒𝔭𝑎 2 ] ∈ Ψ𝑚1 +𝑚2 −1 (Ω) /Ψ−∞ (Ω) .
If the operators 𝑨 𝑗 ∈ Ψ𝑚 𝑗 (Ω) ( 𝑗 = 1, 2) are properly supported then

[ 𝑨1 , 𝑨2 ] ∈ Ψ𝑚1 +𝑚2 −1 (Ω) . (16.2.20)

16.2.4 Elliptic symbols and their inverses

In accordance with Definition 1.3.2 we posit

Definition 16.2.14 An amplitude 𝑎 ∈ 𝑆 𝑚 (Ω × Ω) and the pseudodifferential opera-
tor Op𝑎 ∈ Ψ𝑚 (Ω) are said to be elliptic of order 𝑚 if to every compact set 𝐾 ⊂ Ω
there are positive constants 𝑅 (large), 𝑐 (small) such that, for all (𝑥, 𝑦) ∈ 𝐾 × 𝐾,

∀𝜉 ∈ R𝑛 , |𝜉 | > 𝑅, |𝑎 (𝑥, 𝑦, 𝜉)| ≥ 𝑐 |𝜉 | 𝑚 .

This terminology extends naturally toÍstandard symbols (Definition 16.2.1). We

shall also say that a formal symbol 𝑎 = ∞ 𝑗=0 𝑎 𝑗 , 𝑎 𝑗 ∈ 𝑆
𝑚− 𝑗 (Ω), is elliptic of order

𝑚 in Ω if this is true of its leading symbol 𝑎 0 .

Proposition 16.2.15 If an amplitude 𝑎 ∈ 𝑆 𝑚 (Ω × Ω) is elliptic
of degree 𝑚 then the
Í 1 𝛽 𝛽
associated formal symbol 𝑎 (𝑥, 𝜉) = 𝛽 ∈Z+𝑛 𝛽! 𝜕𝑦 D 𝜉 𝑎 (𝑥, 𝑥, 𝜉) is also elliptic of
♭

degree 𝑚.
Proof Obvious, since the leading symbol of 𝑎♭ is 𝑎 (𝑥, 𝑥, 𝜉). □
590 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

𝑚 (Ω) is elliptic of order 𝑚 there is a unique 𝑏 ∈

Proposition 16.2.16 If 𝑎 ∈ 𝑆form
−𝑚
𝑆form (Ω) such that 𝑎#𝑏 = 𝑏#𝑎 = 1. Moreover, 𝑏 is elliptic of degree −𝑚.
Proof We duplicate the construction of the parametrix in Subsection 4.1.2. Clearly
we must have 𝑎 0 𝑏 0 = 1 and therefore deg 𝑎 + deg 𝑏 = 0. Furthermore, all the
homogeneous parts of strictly negative degree of 𝑎#𝑏 must vanish identically. The
degree of D 𝛼𝜉 𝑎 𝑘 being 𝑚 − 𝑘 − |𝛼| and that of 𝜕𝑥𝛼 𝑏 ℓ being −𝑚 − ℓ we require
∑︁ 1 𝛼 𝛼
D 𝜉 𝑎 𝑘 𝜕𝑥 𝑏 ℓ = 0
𝛼!
𝑘+ℓ+| 𝛼 |=𝑁

for every 𝑁 = 1, 2, .... We see that the last equations can be rewritten as
∑︁ 1 𝛼 𝛼
𝑏 𝑁 = −𝑏 0 D 𝜉 𝑎 𝑘 𝜕𝑥 𝑏 ℓ , (16.2.21)
𝛼!
𝑘+ℓ+| 𝛼 |=𝑁
ℓ<𝑁

which allows the unambiguous determination of each 𝑏 𝑁 recursively. □

Thus the elliptic formal symbols form a (noncommutative) group with respect to
the composition law #.
Corollary 16.2.17 If 𝑨 ∈ Ψ𝑚 (Ω) is elliptic of order 𝑚 there is a 𝑩 ∈ Ψ−𝑚 (Ω) such
that 𝑨 ◦𝑩 − 𝐼 ∈ Ψ−∞ (Ω) and 𝑩◦𝑨 −𝐼 ∈ Ψ−∞ (Ω).
An example of an elliptic pseudodifferential operator of order zero is the expo-
nential of a (standard) pseudodifferential operator 𝑨 of order zero. Let us first look
at a possible definition of exp𝑨 from the viewpoint of Operator Theory. Let Ω′ be an
open subset of Ω, Ω′ ⊂⊂ Ω, and 𝜒Ω′ its characteristic function; by Theorem 16.1.19
we know that 𝑨 defines a continuous linear operator 𝐿 2 (Ω′) −→ 𝐿 loc 2 (Ω) and there-
2 ′
fore 𝑓 ↦→ 𝜒Ω′ 𝑨 𝑓 is a bounded linear operator of 𝐿 (Ω ) into itself, 𝑨Ω′ , and the
same is true of the series ∞ 1 𝑘 ′′
Í
𝑘=0 𝑘! 𝑨Ω′ = exp𝑨Ω′ . Let Ω be another open subset of Ω,
Ω ⊂ Ω ⊂⊂ Ω; if 𝑓 ∈ 𝐿 (Ω ) is viewed as an element of 𝐿 2 (Ω′′) vanishing a.e. in
′ ′′ 2 ′

Ω′′\Ω′, we have 𝑨Ω′ 𝑓 = 𝜒Ω′ 𝑨Ω′′ 𝑓 . Generally speaking, however, 𝑨 Ω′′ 𝑓 is smooth
but not identically zero in Ω′′\Ω′; it follows that 𝑨Ω′′ 𝜒Ω′′ \Ω′ 𝑨Ω′′ 𝑓 is smooth but
not identically zero in Ω′; and therefore the same is true of 𝑨2Ω′ 𝑓 − 𝜒Ω′ 𝑨2Ω′′ 𝑓 and
more generally of 𝑨Ω 𝑘 𝑓 − 𝜒 ′ 𝑨 𝑘 𝑓 , 𝑘 = 2, 3.... In this approach it is not clear that
′ Ω Ω′′
there exists an operator 𝑩 ∈ Ψ0 (Ω) whose restriction to 𝐿 2 (Ω′) is equal to exp𝑨Ω′
for every Ω′ ⊂⊂ Ω.
We shall approach the exponentiation of 𝑨∈ Ψ0 (Ω) from the symbolic calculus
side. Let 𝑎 (𝑥, 𝜉) ∈ 𝑆 0 (Ω) be such that 𝑨−𝑎 (𝑥, D) ∈ Ψ−∞ (Ω). With the one-
parameter group of operators exp (𝑡𝑎 (𝑥, D)) (𝑡 ∈ R) in mind, we seek a symbol
0
𝐸 (𝑡; 𝑥, 𝜉) ∈ 𝑆form (Ω) depending smoothly on 𝑡 such that [cf. (16.2.16)]

d𝐸 ∑︁ 1
𝛾

𝛾
= 𝐸#𝑎 = D 𝜉 𝐸 𝜕𝑥 𝑎, (16.2.22)
d𝑡 𝛾 ∈Z𝑛
𝛾!
+

𝐸 (0; 𝑥, 𝜉) = 1.
16.2 Symbolic Calculus 591
Í∞ −𝑚 (Ω) a solution of the linear ODE
We set 𝐸 = 𝑚=0 𝐸 𝑚 with 𝐸 𝑚 ∈ Ψform
𝑚
d𝐸 𝑚 ∑︁ ∑︁ 1 𝛾
𝛾
= D 𝜉 𝐸 𝑚−𝑘 𝜕𝑥 𝑎, (16.2.23)
d𝑡 𝑘=0 |𝛾 |=𝑘
𝛾!

1 if 𝑚 = 0
𝐸 𝑚 (0; 𝑥, 𝜉) = . (16.2.24)
0 if 𝑚 ≥ 1

Both sides in (16.2.23) have order ≤ −𝑚. The solvability of this initial value problem
is a direct consequence of the following

Lemma 16.2.18 If 𝑎 (𝑥, 𝜉) ∈ 𝑆 0 (Ω) then exp 𝑎 (𝑥, 𝜉) ∈ 𝑆 0 (Ω).

Proof It suffices to show that, given a compact subset 𝐾 of Ω and 𝛼, 𝛽 ∈ Z+𝑛

arbitrarily, there is a 𝐶𝐾 , 𝛼,𝛽 > 0 such that, for every 𝑝 = 1, 2, ...,

∀ (𝑥, 𝜉) ∈ 𝐾 × R𝑛 , 𝜕𝑥𝛼 𝜕 𝜉 (𝑎 𝑝 (𝑥, 𝜉)) ≤ 𝐶𝐾𝑝 , 𝛼,𝛽 (1 + |𝜉 |) − |𝛽 | .

𝛽
(16.2.25)

The result is true when 𝑝 = 1 (Definition 16.2.1). There is no loss of generality in

assuming that 𝐶𝐾 , 𝛼,𝛽 ≤ 𝐶𝐾 , 𝛼′ ,𝛽′ if 𝛼𝑖 ≤ 𝛼𝑖′, 𝛽𝑖 ≤ 𝛽𝑖′ for every 𝑖 = 1, ..., 𝑛. By the
Leibniz rule we have
𝛽
𝜕𝑥𝛼 𝜕 𝜉 (𝑎 𝑝 ) (16.2.26)
𝑝
𝛼! 𝛽! 𝛽 ( 𝑗)
∑︁ ∑︁ Ö ( 𝑗)
= 𝜕𝑥𝛼 𝜕 𝜉 𝑎.
𝛼 (1) ! · · · 𝛼 ( 𝑝) ! 𝛽 (1) ! · · · 𝛽 ( 𝑝) !
𝛼 (1) +···+𝛼 ( 𝑝) =𝛼 𝛽 (1) +···+𝛽 ( 𝑝) =𝛽 𝑗=1

Since ∑︁ 𝛼!
= 𝑝 |𝛼|
𝛼 (1) ! · · · 𝛼 ( 𝑝) !
𝛼 (1) +···+𝛼 ( 𝑝) =𝛼

we derive, from (16.2.25) with 𝑝 = 1 and (16.2.26),

𝜕𝑥𝛼 𝜕 𝜉 (𝑎 𝑝 ) ≤ 𝑝 | 𝛼+𝛽 | 𝐶𝐾𝑝 , 𝛼,𝛽 (1 + |𝜉 |) −|𝛽 | ,

𝛽

whence (16.2.25) for all 𝑝 ≥ 1 if e | 𝛼+𝛽 | 𝐶𝐾 , 𝛼,𝛽 is substituted for 𝐶𝐾 , 𝛼,𝛽 . □

Thanks to Lemma 16.2.18 we can take 𝐸 0 (𝑡; 𝑥, 𝜉) = exp 𝑡𝑎 (𝑥, 𝜉) and rewrite
(16.2.23) for 𝑚 ≥ 1 as
𝑚 ∑︁
d −𝑡 𝑎 ∑︁ 1 𝛾
e 𝐸 𝑚 = e−𝑡 𝑎
𝛾
D 𝜉 𝐸 𝑚−𝑘 𝜕𝑥 𝑎,
d𝑡 𝛾!
𝑘=1 |𝛾 |=𝑘

with the natural solution

𝑚 ∑︁ ∫ 𝑡
∑︁ 1 𝛾 ′
e (𝑡−𝑡 ) 𝑎 D 𝜉 𝐸 𝑚−𝑘 (𝑡 ′; 𝑥, 𝜉)d𝑡 ′.
𝛾
𝐸 𝑚 (𝑡; 𝑥, 𝜉) = 𝜕𝑥 𝑎(𝑥, 𝜉) (16.2.27)
𝛾! 0
𝑘=1 |𝛾 |=𝑘
592 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Induction on 𝑚 implies directly that 𝐸 𝑚 (𝑡; ∞

Í 𝑥, 𝜉) is a C function of 𝑡 ∈ R valued in
𝑆 −𝑚 (Ω). Using the formal symbol 𝐸 = ∞ 𝐸
𝑚=0 𝑚 we can construct “true” standard
symbols by the method described in Subsection 16.2.2; the cutoff function 𝜒 ∈
C ∞ (R𝑛 ) in (16.2.10) can be chosen not to depend on 𝑡. If 𝑒 (𝑡; 𝑥, 𝜉) is any of the true
symbols so constructed we can regard 𝑒 (𝑡; 𝑥, D 𝑥 ) as the exponential exp 𝑡𝑎 (𝑥, D 𝑥 ).
Regarding the latter as a coset mod Ψ−∞ (Ω) it is not difficulty to check that it forms
a one-parameter group, meaning

exp 𝑡1 𝑎 (𝑥, D 𝑥 ) exp 𝑡2 𝑎 (𝑥, D 𝑥 ) − exp (𝑡1 + 𝑡2 ) 𝑎 (𝑥, D 𝑥 ) ∈ Ψ−∞ (Ω) . (16.2.28)

In applications we might have to deal with a pseudodifferential operator that depends

on 𝑡 and therefore, with a symbol 𝑎 (𝑡, 𝑥, 𝜉). In this case, the sought
∫𝑡 solution of
−∞
(16.2.22) would define the cosets of operators mod Ψ (Ω), exp 0 𝑎 (𝑠, 𝑥, D 𝑥 ) d𝑠.

16.2.5 Pseudodifferential operators decrease the wave-front set

Let U be a conic open subset of Ω × (R𝑛 \ {0}). We define 𝑆 −∞ (Ω, U) as the

vector subspace of 𝑆 (Ω) consisting of the symbols 𝑎 (𝑥, 𝜉) that satisfy the following
condition:
(SU) 𝜒𝑎 ∈ 𝑆 −∞ (Ω) whatever 𝜒 ∈ 𝑆 (Ω) such that supp 𝜒 ⊂ U.
Since, obviously, 𝑆 −∞ (Ω) ⊂ 𝑆 −∞ (Ω, U) we may introduce the quotient lin-
ear subspace 𝑆 −∞ (Ω, U) /𝑆 −∞ (Ω) of 𝑆 (Ω) /𝑆 −∞ (Ω). It is readily checked that
𝑆 −∞ (Ω, U) /𝑆 −∞ (Ω) is a two-sided ideal in the ring (with respect to the composi-
tion law #) 𝑆 (Ω) /𝑆 −∞ (Ω) (cf. Theorem 16.2.13).

Definition 16.2.19 Let U be a conic open subset of Ω × (R𝑛 \ {0}). We shall say
that a pseudodifferential operator 𝑨: E ′ (Ω) −→ D ′ (Ω) is smoothing in U if there
is a symbol 𝑏 ∈ 𝑆 −∞ (Ω, U) such that 𝑨−Op𝑏 ∈ Ψ−∞ (Ω). We shall denote by
Ψ−∞ (Ω, U) the set of operators 𝑨∈ Ψ (Ω) smoothing in U.
The complement in Ω × (R𝑛 \ {0}) of the union of all the conic open subsets
in which 𝑨 is smoothing shall be called the microsupport of 𝑨 and denoted by
microsupp𝑨 .

The property that 𝑨 is smoothing in U is a microlocal property: it is equivalent

to the property that 𝑨 is smoothing in every conic open subset of U.
The microsupport of 𝑨 is a closed conic subset of Ω × (R𝑛 \ {0}). Obviously the
property of 𝑨∈ Ψ (Ω) to be smoothing in U is a property of the equivalence class
of 𝑨 mod Ψ−∞ (Ω).
Concerning the wave-front set of a distribution 𝑢 ∈ D ′ (Ω), 𝑊 𝐹 (𝑢), we refer the
reader to Definition 2.1.5.

Theorem 16.2.20 If 𝑨∈ Ψ (Ω) then 𝑊 𝐹 ( 𝑨𝑢) ⊂ 𝑊 𝐹 (𝑢) ∩ microsupp𝑨 for every

𝑢 ∈ E ′ (Ω).
16.2 Symbolic Calculus 593

Proof In what follows 𝑢 ∈ E ′ (Ω) is arbitrary. It is permitted to reason mod 𝑆 −∞ (Ω)

and therefore, to suppose either that 𝑨 = 𝑎(𝑥, 𝐷), 𝑎 (𝑥, 𝜉) ∈ 𝑆 (Ω), or 𝑨 = Op𝑎
with 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 (Ω × Ω) properly supported, the latter by Proposition 16.1.17.
These two choices are mutually exclusive. We choose 𝑨 = 𝑎 (𝑥, D) and let (𝑥 ◦ , 𝜉 ◦ ) ∈
Ω × S𝑛−1 . In the proof 𝑈 ⊂⊂ Ω shall be a neighborhood of 𝑥 ◦ in R𝑛 and Γ ⊂ R𝑛 \ {0}
an open cone containing 𝜉 ◦ , with diam 𝑈 and diam Γ ∩ S𝑛−1 as small as needed.
Let 𝑔 ∈ Cc∞ (𝑈), 𝑔 ≡ 1 in a neighborhood of 𝑥 ◦ , and 𝜒 ∈ C ∞ (R𝑛 ), supp 𝜒 ⊂ Γ,
homogeneous of degree zero for |𝜉 | > 1/2, 𝜒 ≡ 1 in a neighborhood of 𝜉 ◦ , 𝜒 (𝜉) = 0
if 𝜉 ∉ Γ ′, Γ ′ an open cone in R𝑛 \ {0}, Γ ′ ∩ S𝑛−1 ⊂⊂ Γ, also 𝜒 (𝜉) = 0 if |𝜉 | < 1/4.
When (𝑥 ◦ , 𝜉 ◦ ) ∉ microsupp𝑨 we select 𝑈 and Γ such that

(𝑈 × Γ) ∩ microsupp 𝑨 = ∅;

then the formal symbol 𝜒 (𝜉) #𝑔 (𝑥) 𝑎 (𝑥, 𝜉) belongs to 𝑆form −∞ (R𝑛 ), implying

𝜒 (D) (𝑔 𝑨𝑢) ∈ C (R ). By Definition 2.1.5 this implies (𝑥 , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨𝑢).

∞ 𝑛 ◦

Now assume (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 (𝑢). We can select 𝑈 and Γ such that (𝑈 × Γ) ∩

𝑊 𝐹 (𝑢) = ∅. With 𝑔 and 𝜒 as before we have 𝑔 (𝑥) 𝜒 (D) 𝑢 ∈ C ∞ (R𝑛 ) hence
𝑨 (𝑥, D) 𝑔 (𝑥) 𝜒 (D) 𝑢 ∈ C ∞ (R𝑛 ). We prove that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨𝑔 (1 − 𝜒 (D)) 𝑢);
this would imply (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨𝑔𝑢) and the latter implies (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨𝑢)
since 𝑥 ◦ ∉ singsupp ( 𝑨 (𝑔 𝑓 ) − 𝑨 𝑓 ) whatever 𝑓 ∈ E ′ (Ω).
Let 𝜙 ∈ C ∞ (R𝑛 ) satisfy the following conditions:
(1) 𝜙 is homogeneous of degree zero for |𝜉 | > 3/4,
(2) 𝜙 ≡ 1 in a neighborhood of 𝜉 ◦ ,
(3) (1 − 𝜒) 𝜙 ≡ 0.
We apply (16.2.16) and Corollary 16.2.4: the formal symbol
∑︁ 1
(1 − 𝜒 (𝜉)) D 𝜉 𝜙 𝜕𝑥 𝑔 2 (𝑥) 𝑎 (𝑥, 𝜉)
𝛾 𝛾

𝛾 ∈Z𝑛
𝛾!
+

vanishes identically, implying 𝜙 (D) 𝑔 𝑨 (𝑔 (1 − 𝜒 (D))) ∈ Ψ−∞ (Ω). The sought

result ensues easily. □

Corollary 16.2.21 If 𝑨∈ Ψ𝑚 (Ω) (𝑚 ∈ R) is elliptic of order 𝑚 then 𝑊 𝐹 ( 𝑨𝑢) =

𝑊 𝐹 (𝑢) for every 𝑢 ∈ E ′ (Ω).

Proof By Corollary 16.2.11 there is a 𝑩 ∈ Ψ−𝑚 (Ω) such that 𝑩 ◦ 𝑨 − 𝐼 ∈ Ψ−∞ (Ω).
We have 𝑊 𝐹 (𝑢) = 𝑊 𝐹 (𝑩 𝑨𝑢) ⊂ 𝑊 𝐹 ( 𝑨𝑢) ⊂ 𝑊 𝐹 (𝑢). □

Corollary 16.2.22 Let U be a conic open subset of Ω × (R𝑛 \ {0}). If 𝑨∈ Ψ𝑚 (Ω)

is elliptic of order 𝑚 and if 𝑢 ∈ E ′ (Ω) is such that 𝑊 𝐹 ( 𝑨𝑢) ∩ U = ∅ then
𝑊 𝐹 (𝑢) ∩ U = ∅; in particular, 𝑨𝑢 ∈ C ∞ (Ω) entails 𝑢 ∈ C ∞ (Ω).

Proposition 16.2.23 An operator 𝑨∈ Ψ (Ω) is smoothing in the conic open set U if

and only if 𝑊 𝐹 ( 𝑨𝑢) ∩ U = ∅ whatever 𝑢 ∈ E ′ (Ω).
594 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Proof The condition is necessary since 𝑊 𝐹 ( 𝑨𝑢) ⊂ microsupp𝑨 for every 𝑢 ∈

E ′ (Ω), by Theorem 16.2.20. Let us prove that the condition is sufficient; we assume
that 𝑨 = 𝑎 (𝑥, D) with 𝑎(𝑥, 𝜉) ∈ 𝑆 (Ω). Let (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω×S𝑛−1 be arbitrary. We select
the open set 𝑈 ⊂⊂ Ω and the open cone Γ ⊂ R𝑛 \ {0} such that (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑈 ×Γ ⊂ U;
the functions 𝑔 ∈ Cc∞ (𝑈) and 𝜒 ∈ C ∞ (R𝑛 ) are chosen exactly as in the proof of
Theorem 16.2.20. We know by Theorem 16.2.20 that 𝑊 𝐹 ( 𝑨 (𝑔 (𝑥) 𝜒 (D) 𝑢)) ⊂ 𝑈 ×
Γ ⊂ U. Since we are hypothesizing that 𝑊 𝐹 ( 𝑨 (𝑔 (𝑥) 𝜒 (D) 𝑢)) ∩ U = ∅ we derive
that 𝑨 (𝑔 (𝑥) 𝜒 (D) 𝑢) ∈ C ∞ (Ω) for all 𝑢 ∈ E ′ (Ω), i.e., 𝑨 (𝑔 (𝑥) 𝜒 (D)) ∈ Ψ−∞ (Ω).
By Proposition 16.2.3 this implies
∑︁ 1
𝛾 −∞
𝜒 (𝜉) D 𝜉 𝑎 (𝑥, 𝜉) 𝜕𝑥𝛼 𝑔 (𝑥) ∈ 𝑆form (Ω) ,
𝛾 ∈Z 𝑛 𝛾!
+

whence 𝜒 (𝜉) 𝑔 (𝑥) 𝑎 (𝑥, 𝜉) ∈ 𝑆 −∞ (Ω). In turn the latter implies that 𝑎 (𝑥, D) is
smoothing in a conic neighborhood of (𝑥 ◦ , 𝜉 ◦ ) contained in 𝑈 × Γ. □

Remark 16.2.24 Proposition 16.2.23 suggests an alternate definition of “smooth-

ing”: a continuous linear operator 𝑨: E ′ (Ω) −→ D ′ (Ω) is smoothing in the conic
set U if 𝑊 𝐹 ( 𝑨𝑢) ∩ U = ∅ whatever 𝑢 ∈ E ′ (Ω).

Theorem 16.2.20 is the gateway to the microlocal theory of pseudodifferential

operators (in the C ∞ category). The first step towards C ∞ microlocalization is to mod
off smooth functions and 𝑆 −∞ (Ω) as well as smoothing operators, meaning Ψ−∞ (Ω):
in essence, to deal with equivalence classes in D ′ (Ω) /C ∞ (Ω), in 𝑆 (Ω) /𝑆 −∞ (Ω)
and in Ψ (Ω) /Ψ−∞ (Ω).
The next step is to replace, wherever it makes sense, 𝑇 ∗ Ω\0 by one of its conic
open subsets U. Thus, a distribution 𝑢 in Ω will be said to be smooth in U if
U ∩ 𝑊 𝐹 (𝑢) = ∅; we denote by C ∞ (Ω, U) the linear space of such distributions. A
restatement of the fact that, if 𝑨∈ Ψ (Ω) and 𝑢 ∈ D ′ (Ω) then 𝑊 𝐹 ( 𝑨𝑢) ⊂ 𝑊 𝐹 (𝑢), is
that 𝑢 ∈ C ∞ (Ω, U) =⇒ 𝑨𝑢 ∈ C ∞ (Ω, U): not only are pseudodifferential operators
pseudolocal (i.e., they decrease the singular support) but they are microlocal.
We can go further and microlocalize the operators themselves: a pseudodiffer-
ential of order 𝑚 in U is an equivalence class in Ψ𝑚 (Ω) /Ψ−∞ (Ω, U). Such an
“operator”, e𝑨, is elliptic in U if there is an elliptic operator 𝑩 ∈ Ψ𝑚 (Ω) mapped
𝑨 by the quotient map Ψ𝑚 (Ω) −→ Ψ𝑚 (Ω) /Ψ−∞ (Ω, U). Etc.
into e
Needless to say, the above definitions and results that are invariant under coordi-
nates changes can be generalized to pseudodifferential operators on a C ∞ manifold.
16.3 Classical symbols and classical pseudodifferential operators 595

16.3 Classical symbols and classical pseudodifferential operators

16.3.1 Classical symbols

Let 𝑚 ∈ R be arbitrary. We continue to use the adjective “homogenous” as before:

a function 𝑓 defined in Ω × (R𝑛 \ {0}) is homogeneous of degree 𝑚 if 𝑓 (𝑥, 𝜆𝜉) =
𝜆 𝑚 𝑓 (𝑥, 𝜉) for all (𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}) and all 𝜆 > 0. If moreover 𝑓 is smooth then
𝛾
𝜕 𝜉 𝑓 is homogeneous of degree 𝑚 − |𝛾| (𝛾 ∈ Z+𝑛 ). We shall also say that a function
𝑓 defined in Ω × R𝑛 is homogeneous of degree 𝑚 for large |𝜉 | if there is an 𝑅 > 0,
𝑅 < +∞, such that 𝑓 (𝑥, 𝜆𝜉) = 𝜆 𝑚 𝑓 (𝑥, 𝜉) for all (𝑥, 𝜉) ∈ Ω × R𝑛 , |𝜉 | > 𝑅, and all
𝜆 ≥ 1.
The following statement is self-evident:

Proposition 16.3.1 If 𝑓 ∈ C ∞ (Ω × R𝑛 ) is homogeneous of degree 𝑚 for large |𝜉 |

then 𝑓 (𝑥, 𝜉) is a standard symbol of order 𝑚. This symbol is elliptic in a conic subset
𝑈 × Γ, where 𝑈 is an open subset of Ω and Γ an open cone in R𝑛 \ {0}, if and only if,
whatever 𝑥 ∈ 𝑈, the function 𝜉 ↦→ lim 𝜆−𝑚 𝑓 (𝑥, 𝜆𝜉) does not vanish at any point
𝜆→+∞
of Γ ∩ S𝑛−1 . In this case 𝑓 ∉ 𝑆 𝑚−𝜀 (Ω) whatever 𝜀 > 0.

Definition 16.3.2 By a classical symbol of order 𝑚(∈ R) in Ω weÍshall mean a

symbol 𝑃 ∈ 𝑆 𝑚 (Ω) equal to the sum of a convergent series in 𝑆 𝑚 (Ω), ∞
𝑗=0 𝑝 𝑗 (𝑥, 𝜉),
with 𝑝 𝑗 ∈ C ∞ (Ω × R𝑛 ) homogeneous of degree 𝑚 − 𝑗 for large |𝜉 |, for each 𝑗.
𝑚 (Ω) the linear space of all classical symbols of order
We shall denote by 𝑆class
≤ 𝑚, and by 𝑆class (Ω) the union of all these linear spaces as 𝑚 ranges over R. The
set 𝑆class (Ω) is a subring of 𝑆 (Ω) with respect to Í
ordinary addition and multiplica-
tion. Definition 16.3.2 does not exclude symbols ∞ 𝑗=0 𝑝 𝑗 (𝑥, 𝜉) with the following
property: to every 𝑗 there is an 𝑅 𝑗 > 0 such that 𝑝 𝑗 (𝑥, 𝜉) = 0 if |𝜉 | > 𝑅 𝑗 . In fact,
this is a requirement if we want 𝑆class (Ω) to be a group with respect to addition. It
follows that 𝑆class (Ω) ∩ 𝑆 −∞ (Ω) ≠ {0}.

Proposition 16.3.3 We have

𝑆 −∞ (Ω) = 𝑆class (Ω) ∩ 𝑆 −∞ (Ω) . (16.3.1)

Proof Let 𝑎 (𝑥, 𝜉) ∈ 𝑆 −∞ (Ω) be arbitrary

Í∞and let 𝜒 𝑗 ∈ Cc
∞ (R), 𝑗 = 1, 2, ..., be such

that 0 ≤ 𝜒 𝑗 ≤ 1, supp 𝜒 𝑗 ⊂ ( 𝑗 − 2, 𝑗), 𝑗=1 𝜒 𝑗 (𝑡) = 1 for all 𝑡 ≥ 0. Then

∞
∑︁
𝑎 (𝑥, 𝜉) = 𝜒 𝑗 (|𝜉 |) 𝑎 (𝑥, 𝜉) .
𝑗=1

The right-hand side is a classical symbol. □

The concept of a formal symbol (Definition 16.2.5) is very precise in the “classi-
cal” context.
596 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Definition 16.3.4 By a formal classical symbol of order 𝑚 in Ω we shall mean a

formal series
∞
∑︁
𝑃 (𝑥, 𝜉) = 𝑝 𝑗 (𝑥, 𝜉) , (16.3.2)
𝑗=0

where 𝑝 𝑗 ∈ C∞ (Ω × (R𝑛 \ {0}))

is homogeneous of degree 𝑚 − 𝑗 for each 𝑗. For
any 𝑗 ∈ Z+ , 𝑝 𝑗 will be referred to as the homogeneous part of degree 𝑚 − 𝑗 of 𝑝;
𝑝 0 is called the principal (or leading) part of the symbol 𝑃.

The linear space of all symbols (16.3.2) (with 𝑚 ranging over R) will be denoted
by 𝑆class,form (Ω); the linear subspace of 𝑆class,form (Ω) consisting of the symbols of
degree ≤ 𝑚 will be denoted by 𝑆class,form
𝑚 (Ω). We shall also talk of formal classical
amplitudes if the homogeneous parts depend smoothly on (𝑥, 𝑦) ∈ Ω × Ω.
𝑚𝑗 𝑚1 +𝑚2
If 𝑃 𝑗 ∈ 𝑆class,form (Ω) ( 𝑗 = 1, 2) then 𝑃1 #𝑃2 ∈ 𝑆class,form (Ω). For 𝑚 ≤ 0,
𝑆class,form (Ω) is a subalgebra of 𝑆class,form (Ω) with respect to addition and #. If
𝑚

𝑃 ∈ 𝑆class,form
𝑚 (Ω) the same is true of its transpose and its adjoint [(16.2.13),
(16.2.14)]. The following statements are direct consequences of (16.2.16).
𝑗 𝑚
Proposition 16.3.5 If 𝑝 𝑗,0 is the principal part of 𝑃 𝑗 ∈ 𝑆class,form (Ω) ( 𝑗 = 1, 2) the
following properties hold:
(1) The principal part of the composite 𝑃1 #𝑃2 is the (ordinary) product of the
principal parts, 𝑝 1,0 𝑝 2,0 .
(2) The principal part of the commutator [𝑃1 , 𝑃2 ] # = 𝑃1 #𝑃2 − 𝑃2 #𝑃1 is equal to
√1 𝑝 1,0 , 𝑝 2,0 , where
−1

𝑛
∑︁ 𝜕 𝑝 1,0 𝜕 𝑝 2,0 𝜕 𝑝 2,0 𝜕 𝑝 1,0
𝑝 1,0 , 𝑝 2,0 = −
ℓ=1
𝜕𝜉ℓ 𝜕𝑥ℓ 𝜕𝜉ℓ 𝜕𝑥ℓ

is the Poisson bracket (Definition 13.3.8) of the two functions 𝑝 1,0 and 𝑝 2,0 in
𝑚1 +𝑚2 −1
(𝑥, 𝜉)-space [cf. (13.3.19)]. Thus [𝑃1 , 𝑃2 ] # ∈ 𝑆class,form (Ω).
𝑚 (Ω) −→ 𝑆 𝑚 Í∞
There is a natural map 𝑆class class,form (Ω): if 𝑃 (𝑥, 𝜉) = 𝑗=0 𝑝 𝑗 (𝑥, 𝜉) ∈
𝑆class (Ω) we define the homogeneous functions in Ω × (R𝑛 \ {0}),
𝑚

𝑝˜ 𝑗 (𝑥, 𝜉) = lim 𝜆 𝑗−𝑚 𝑝 𝑗 (𝑥, 𝜆𝜉) , (16.3.3)

0<𝜆→+∞

and assign to 𝑃 (𝑥, 𝜉) the formal classical symbol

∞
∑︁
𝑃˜ (𝑥, 𝜉) = 𝑚
𝑝˜ 𝑗 (𝑥, 𝜉) ∈ 𝑆class,form (Ω) .
𝑗=0

We shall refer to 𝑃˜ (𝑥, 𝜉) as the formal symbol defined by 𝑃 (𝑥, 𝜉) and to 𝑃 (𝑥, 𝜉) as
a true classical symbol (or true symbol) associated to the formal symbol 𝑃˜ (𝑥, 𝜉).
We must stress the fact that there are no formal classical symbols of order −∞
other than zero. It follows that (16.3.3) defines a linear map, for each 𝑚 ∈ Z+𝑚 ,
16.3 Classical symbols and classical pseudodifferential operators 597
∞
∑︁ ∞
∑︁
𝑚
𝑆class (Ω) ∋ 𝑃 (𝑥, 𝜉) = 𝑝 𝑗 (𝑥, 𝜉) ↦→ 𝑃˜ (𝑥, 𝜉) = 𝑚
𝑝˜ 𝑗 (𝑥, 𝜉) ∈ 𝑆class,form (Ω) ,
𝑗=0 𝑗=0
(16.3.4)
𝑚 (Ω) /𝑆 −∞ (Ω) −→ 𝑆 𝑚
and that, in turn, (16.3.4) defines an injection 𝑆class class,form (Ω).
Actually, (16.3.4) induces a bijection: indeed, given an arbitrary (nonzero) 𝑃form (𝑥, 𝜉)
= ∞ ∞
Í
𝑗=0 𝑝 𝑗 (𝑥, 𝜉) ∈ 𝑆 class,form (Ω) we can select 𝜒 ∈ C (R ), 𝜒 (𝜉) = 0 if
𝑚 𝑛

|𝜉 | (𝜉) |𝜉 | numbers 𝑅 𝑗 ↗ ∞ so that

Í∞ < 1, 𝜒 = 1 if 𝑚 > 2, and the positive ˜
𝑗=0 𝜒 𝜉/𝑅 𝑗 𝑝 𝑗 (𝑥, 𝜉) ∈ 𝑆 class (Ω); obviously, 𝑃 = 𝑃form .
A formal classical symbol (16.3.2) and the true symbols associated to it are elliptic
(Definition 16.2.14) if the principal part 𝑝 0 does not vanish at any point of Ω × S𝑛−1 .
It is seen directly from the equations (16.2.21) that the inverse (with respect to #) of
an elliptic formal classical symbol of order 𝑚 is an elliptic formal classical symbol
of order −𝑚. When 𝑚 ranges over R (or Z, or any subgroup of R for addition) these
symbols form a group with respect to #.

16.3.2 Classical pseudodifferential operators

We switch our attention to the Op map 𝑆 𝑚 (Ω) ∋ 𝑃 (𝑥, 𝜉) ↦→ 𝑃 (𝑥, D) ∈ Ψ𝑚 (Ω).

𝑚 (Ω) the image of 𝑆 𝑚 (Ω)
Definition 16.3.6 Let 𝑚 ∈ R. We shall denote by Ψclass class
under the Op map and refer to the elements of Ψclass (Ω)
𝑚
Ø as classical pseudodiffer-
ential operators of order 𝑚. We define Ψclass (Ω) = 𝑚 (Ω).
Ψclass
𝑚∈R

The simplest examples of classical pseudodifferential operators are the differential

operators, their symbols being polynomials with respect to 𝜉 whose coefficients are
smooth functions of 𝑥 ∈ Ω.
If we follow up the natural map 𝑆class,form
𝑚 𝑚 (Ω) /𝑆 −∞ (Ω) with the
(Ω) −→ 𝑆class
bijection
𝑚
𝑆class (Ω) /𝑆 −∞ (Ω) −→ Ψclass
𝑚
(Ω) /Ψ−∞ (Ω)
induced by the natural bijection 𝑆 𝑚 (Ω) /𝑆 −∞ (Ω) −→ Ψ𝑚 (Ω) /Ψ−∞ (Ω) (itself
defined by the Op map) we obtain a linear bijection
𝑚
𝑆class,form 𝑚
(Ω) → Ψclass (Ω) /Ψ−∞ (Ω) . (16.3.5)

Letting 𝑚 range over R yields a bijection

𝑆class,form (Ω) → Ψclass (Ω) /Ψ−∞ (Ω) . (16.3.6)

The map (16.3.6) is an algebra isomorphism, assuming that 𝑆class,form (Ω) is equipped
with the composition law # [see (16.2.16)] and Ψclass (Ω) /Ψ−∞ (Ω) is equipped
with the composition of (cosets of) operators law inherited from Ψ𝑚 (Ω) /Ψ−∞ (Ω).
Following up the quotient map Ψclass (Ω) −→ Ψclass (Ω) /Ψ−∞ (Ω) with the inverse
of (16.3.6) produces the symbol map
598 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Ψclass (Ω) ∋ 𝑃 ↦→ symb (𝑃) ∈ 𝑆class,form (Ω) . (16.3.7)

𝑚 (Ω) we shall refer to

Definition 16.3.7 If 𝑃 ∈ Ψclass
∞
∑︁
𝑚
symb (𝑃) = 𝑝 𝑗 (𝑥, 𝜉) ∈ 𝑆class,form (Ω)
𝑗=0

as the total symbol and to 𝑝 0 as the principal symbol of the operator 𝑃.

In Definition 16.3.7 it is tacitly assumed that 𝑝 0 does not vanish identically. The
symbol of the leading part of a differential operator is its principal symbol (Definition
1.3.1). The total symbol of the transpose 𝑃⊤ of 𝑃 is easily derived from the general
formula (16.2.13):
∞
∑︁ ∑︁ (−1) |𝛽 | 𝛽 𝛽
symb 𝑃⊤ = 𝜕𝑥 D 𝜉 𝑝 𝑗−|𝛽 | (𝑥, −𝜉) . (16.3.8)
𝑗=0 𝛽 ∈Z+𝑛 , |𝛽 | ≤ 𝑗
𝛽!

The sum with respect to 𝛽 in (16.3.8) is the homogeneous part of degree 𝑚 − 𝑗 of

symb (𝑃⊤ ).

Next we wish to explore the effect of a change of variables on a pseudodifferential

𝑚 (Ω) with total symbol 𝑃 (𝑥, 𝜉) = Í∞ 𝑝 (𝑥, 𝜉) ∈ 𝑆 𝑚
operator 𝑃 ∈ Ψclass 𝑗=0 𝑗 class,form (Ω).
Let Ω and Ω′ be two open subsets of R𝑛 and 𝐹 : Ω′ −→ Ω a diffeomorphism. We
look at the amplitude (16.2.12) when 𝑎 = 𝑝 𝑗 . We see that 𝑝 𝑗 𝐹 (𝑥 ′, 𝑦 ′, 𝜉 ′) is also

homogeneous of degree 𝑚 − 𝑗. By applying (16.2.3) to 𝑎 𝐹 we obtain a formal classi-
cal symbol 𝑃 𝑗 𝐹 (𝑥 ′, 𝜉 ′), a series of homogeneous symbols (of decreasing degrees
≤ 𝑚 − 𝑗), in general not a single homogeneous symbol. It is readily checked that
summing over 𝑗 yields a formal classical symbol of order 𝑚, 𝑃𝐹 (𝑥 ′, 𝜉 ′). Through
the bijection (16.3.5) (also with Ω′ in the place of Ω) the map 𝐹 induces a lin-
ear bijection Ψclass𝑚 (Ω) /Ψ−∞ (Ω) −→ Ψ 𝑚 (Ω ′ ) /Ψ−∞ (Ω ′ ) and when 𝑚 ranges
class
over R, an algebra isomorphism of Ψclass (Ω) /Ψ−∞ (Ω) onto Ψclass (Ω′) /Ψ−∞ (Ω′).
This observation enables us to define classical pseudodifferential operators on a
C ∞ manifold M. Indeed, the notion makes sense in the domain U of local coor-
dinates 𝑥1 , ..., 𝑥 𝑛 , since the preceding isomorphism (applied when Ω is the image
of U under the coordinate map) shows that one can define Ψclass 𝑚 (U) /Ψ−∞ (U)

independently of these coordinates. Since the same property is trivially true for
Ψ−∞ (U) [isomorphic to C ∞ (U × U) by the Schwartz Kernels Theorem] we have
a coordinate-independent definition of Ψclass 𝑚 (U).

Definition 16.3.8 A continuous linear operator 𝑃 : E ′ (M) −→ D ′ (M) will be

called a classical pseudodifferential operator in M if its restriction [cf. Definition
2.3.9] to an arbitrary domain U of local coordinates is a classical pseudodifferential
operator in U. The linear space of these operators will be denoted by Ψclass (M). If
the restriction of 𝑃 ∈ Ψclass (M) to every domain U of local coordinates has order
𝑚 ∈ R we shall say that 𝑃 has order 𝑚 and we write 𝑃 ∈ Ψclass𝑚 (M).
16.3 Classical symbols and classical pseudodifferential operators 599

Definition 16.3.8 is consistent with Definition 16.1.25: Ψclass (M) ⊂ Ψ (M).

16.3.3 Characteristic set of a classical pseudodifferential operator

Definition 1.3.1 is extended to classical pseudodifferential operators in the obvious

manner:

Definition 16.3.9 Let 𝑝 0 (𝑥, 𝜉) be the principal symbol of 𝑃 = 𝑃 (𝑥, D) ∈ Ψclass (Ω);
the nullset
Char 𝑃 = {(𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}) ; 𝑝 0 (𝑥, 𝜉) = 0}
is called the characteristic set of 𝑃.

Char 𝑃 is a closed conic subset of Ω × (R𝑛 \ {0}).

We return to the diffeomorphism 𝐹 : Ω′ −→ Ω where Ω and Ω′ are two open
subsets of R𝑛 . As indicated above, if we apply (16.2.12) to 𝑃 we obtain an amplitude
in Ω′ × Ω′ to which the operation (16.1.23) assigns a classical formal symbol
𝑃𝐹 (𝑥 ′, 𝜉 ′) whose principal part is
⊤
𝜕𝐹 ′ −1
𝑝 0 𝐹 (𝑥 ′) , (𝑥 ) 𝜉′ (16.3.9)
𝜕𝑥 ′

since 𝐽 (𝑥 ′, 𝑥 ′) = 𝜕𝐹
𝜕𝑥 ′ (𝑥 ′). We point out that
⊤
𝜕𝐹 ′ −1
(𝑥 ′, 𝜉 ′) ↦→ 𝐹 (𝑥 ′) , (𝑥 ) 𝜉′ (16.3.10)
𝜕𝑥 ′

is the map between cotangent bundles, 𝑇 ∗ Ω′ −→ 𝑇 ∗ Ω, defined by 𝐹; (16.3.10) is

a homeomorphism and maps fiber to fiber (as a linear isomorphism): it is a vector
bundle isomorphism; (16.3.9) states the equivariance of principal symbols under
diffeomorphisms in the base.
Going back to a C ∞ manifold M and to a classical pseudodifferential operator 𝑃 of
order 𝑚 in M (Definition 16.3.9) we derive from (16.3.9) that the principal symbol 𝑝 0
−1
of 𝑃 is a well-defined (i.e., coordinate-free) function in 𝜋 (U)\0 ⊂ 𝑇 ∗ M\0, whatever
the domain U of local coordinates (here 𝜋 is the base projection 𝑇 ∗ M\0 −→ M).
By (16.3.9) these local definitions coincide in overlaps, thereby providing a global
definition of 𝑝 0 ∈ C ∞ (𝑇 ∗ M\0).
The preceding statements and definitions will also make sense in the C 𝜔 category
(see Ch. 17).
From Definition 16.2.14 we derive easily
𝑚 (M) is elliptic if and only if Char 𝑃 =
Proposition 16.3.10 An operator 𝑃 ∈ Ψclass
∅.
600 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

The composite of two elliptic classical pseudodifferential operators is an elliptic

𝑚 (M) is elliptic there is a 𝑄 ∈
classical pseudodifferential operator. If 𝑃 ∈ Ψclass
𝑚 (M) such that both 𝑃𝑄 − 𝐼 and 𝑄𝑃 − 𝐼 belong to Ψ−∞ (M) (𝐼: the identity
Ψclass class
map; see Corollary 16.2.17).

16.3.4 The subprincipal symbol of a classical pseudodifferential

operator

Let 𝑃 (𝑥, D) be a classicalÍ

pseudodifferential operator of order 𝑚 in the open subset Ω
of R𝑛 and let symb (𝑃) = ∞ 𝑗=0 𝑝 𝑗 be its total formal symbol. The Euler homogeneity
identities imply
𝑛
∑︁ 𝜕𝑝𝑗
(𝑚 − 𝑗) 𝑝 𝑗 (𝑥, 𝜉) = 𝜉𝑘 (𝑥, 𝜉) , 𝑗 ∈ Z+ . (16.3.11)
𝑘=1
𝜕𝜉 𝑘

Definition 16.3.11 We say that (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑇 ∗ Ω\0 is a doubly characteristic point

of 𝑃 if (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 and if, moreover, 𝜕𝜕 𝜉𝑝𝑘0 (𝑥 ◦ , 𝜉 ◦ ) = 0 for every 𝑘 = 1, ..., 𝑛.
We shall denote by Char (2) 𝑃 the set of doubly characteristic points of 𝑃.

When 𝑚 ≠ 0 it follows from (16.3.11) that d 𝜉 𝑝 0 (𝑥 ◦ , 𝜉 ◦ ) = 0 =⇒ 𝑝 0 (𝑥 ◦ , 𝜉 ◦ ) = 0.

Since the principal symbol 𝑝 0 of a classical pseudodifferential operator 𝑃 on a
manifold M is a well-defined function in 𝑇 ∗ M\0 its differential with respect to the
fiber variables, d 𝜉 𝑝 0 , is also well-defined and so is the nullset of the “jet” 𝑝 0 , d 𝜉 𝑝 0 ,
Char (2) 𝑃.

Proposition 16.3.12 If (𝑥 ◦ , 𝜉 ◦ ) ∈ Char (2) 𝑃 then the complex number

𝑛
1 ∑︁ 𝜕 2 𝑝 0
𝜎sub (𝑃) (𝑥 ◦ , 𝜉 ◦ ) = 𝑝 1 (𝑥 ◦ , 𝜉 ◦ ) − (𝑥 ◦ , 𝜉 ◦ ) (16.3.12)
2𝑖 𝑗=1 𝜕𝑥 𝑗 𝜕𝜉 𝑗

is independent of the choice of coordinates in a neighborhood of 𝑥 ◦ .

Proof Let 𝑈 and 𝑈 ′ be two open subsets of Ω containing 𝑥 ◦ and 𝐹 a diffeomorphism

of 𝑈 ′ onto 𝑈 such that 𝐹 (𝑥 ◦ ) = 𝑥 ◦ . We want to show that 𝜎sub (𝑃) (𝑥 ◦ , 𝜉 ◦ ) is invariant
under the transformation induced by 𝐹 [see (16.2.12)]. The result is immediate if
𝐹 (𝑥 ′) = 𝑥 ◦ + 𝑇 (𝑥 ′ − 𝑥 ◦ ), where 𝑇 is a nonsingular linear transformation of R𝑛 . This
𝜕𝐹 ◦
means that, when 𝐹 is nonlinear, we may assume 𝜕𝑥 ′ (𝑥 ) = 𝐼 𝑛 , the 𝑛 × 𝑛 identity
matrix.
Under this assumption we apply (16.2.12) with 𝑎 (𝑥, 𝜉) = 𝑝 0 (𝑥, 𝜉) + 𝑝 1 (𝑥, 𝜉);
we get
16.3 Classical symbols and classical pseudodifferential operators 601
⊤ det 𝜕𝐹′ (𝑦 ′)
𝑎 𝐹 (𝑥 ′, 𝑦 ′, 𝜉 ′) = 𝑝 0 𝐹 (𝑥 ′) , 𝐽 (𝑥 ′, 𝑦 ′) −1 𝜉 ′ 𝜕𝑥
|det 𝐽 (𝑥 ′, 𝑦 ′)|
⊤ det 𝜕𝐹′ (𝑦 ′)
′ ′ ′ −1
+ 𝑝 1 𝐹 (𝑥 ) , 𝐽 (𝑥 , 𝑦 ) 𝜉′ 𝜕𝑥
.
|det 𝐽 (𝑥 ′, 𝑦 ′)|

We then apply (16.2.3) to 𝑎 𝐹 in the coordinate systems 𝑥 ′ = 𝐹 (𝑥), 𝑦 ′ = 𝐹 (𝑦). It

suffices to look at the terms of degree ≥ 𝑚 − 1 in the series on the right in (16.2.3):
⊤
′ 𝜕𝐹 ′ −1
𝑝 0 𝐹 (𝑥 ) , (𝑥 ) 𝜉′ (16.3.13)
𝜕𝑥 ′
⊤
𝜕𝐹 ′ −1
+𝑝 1 𝐹 (𝑥 ′) , (𝑥 ) 𝜉 ′ − 𝑖𝑞 (𝑥 ′, 𝜉 ′) ,
𝜕𝑥 ′

where
!
𝑛
∑︁ 𝜕2
⊤ det 𝜕𝐹′ (𝑦 ′)
′ ′ ′ ′ ′ −1 ′ 𝜕𝑥
𝑞 (𝑥 , 𝜉 ) = 𝑝 0 𝐹 (𝑥 ) , 𝐽 (𝑥 , 𝑦 ) 𝜉 .
𝜕𝑦 ′𝑗 𝜕𝜉 ′𝑗 |det 𝐽 (𝑥 ′, 𝑦 ′)|
𝑗,𝑘=1 𝑦′ =𝑥 ′

At (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 the function (16.3.13) is equal to 𝑝 1 (𝑥 ◦ , 𝜉 ◦ ) − 𝑖𝑞 (𝑥 ◦ , 𝜉 ◦ ) and

𝑛 ⊤
◦ ◦
∑︁ 𝜕 ′ 𝜕 𝑝0 ◦ ◦ ′ −1
𝑞 (𝑥 , 𝜉 ) = 𝑐 𝑘, 𝑗 (𝑦 ) 𝑥 , 𝐽 (𝑥 , 𝑦 ) 𝜉′ ,
𝑗,𝑘=1
𝜕𝑦 ′𝑗 𝜕𝜉 𝑘 𝑦′ =𝑥 ◦
𝜉 ′= 𝜉 ◦

where we have used the notation

⊤ 𝑛
∑︁
◦ ′ −1
𝐽 (𝑥 , 𝑦 ) 𝜉 = 𝑐 𝑘,ℓ (𝑦 ′) 𝜉ℓ .
𝑘 ℓ=1

The hypotheses that d 𝜉 𝑝 0 (𝑥 ◦ , 𝜉 ◦ ) = 0 and 𝐽 (𝑥 ◦ , 𝑥 ◦ ) = 𝜕𝐹

𝜕𝑥 ′ (𝑥 ◦ ) = 𝐼𝑛 imply
𝑛 ⊤
◦ ◦
∑︁ 𝜕 𝜕 𝑝0 ◦ ◦ ′ −1
𝑞 (𝑥 , 𝜉 ) = ′ 𝑥 , 𝐽 (𝑥 , 𝑦 ) 𝜉◦ . (16.3.14)
𝜕𝑦 𝜕𝜉 𝑗
𝑗=1 𝑗 𝑦′ =𝑥 ◦

These hypotheses also imply

𝑛 ⊤
𝜕2

∑︁
′ 𝜕𝐹 ′ −1
𝑝 0 𝐹 (𝑥 ) , (𝑥 ) 𝜉′
𝑗=1
𝜕𝑥 ′𝑗 𝜕𝜉 ′𝑗 𝜕𝑥 ′ 𝑥 ′ =𝑥 ◦
𝜉 ′= 𝜉 ◦
𝑛 ⊤
∑︁ 𝜕 𝜕 𝑝0 ′ 𝜕𝐹 ′ −1
= ′ 𝜕𝜉 𝐹 (𝑥 ) , ′
(𝑥 ) 𝜉◦
𝜕𝑥 𝑗 𝜕𝑥
𝑗=1 𝑗 𝑥 ′ =𝑥 ◦
602 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class
𝑛
∑︁ 𝜕2 𝑝0
= (𝑥 ◦ , 𝜉 ◦ )
𝑗=1
𝜕𝑥 𝑗 𝜕𝜉 𝑗

𝑛 ⊤
∑︁ 𝜕 𝜕 𝑝 0 ◦ 𝜕𝐹 ′ −1
+ ′ 𝜕𝜉 𝑥 , ′
(𝑥 ) 𝜉◦ .
𝜕𝑥 𝑗 𝜕𝑥
𝑗=1 𝑗 𝑥 ′ =𝑥 ◦

We have shown above that the homogeneous part of degree 𝑚 − 1 in (16.3.13) at

(𝑥 ◦ , 𝜉 ◦ ) is equal to 𝑝 1 (𝑥 ◦ , 𝜉 ◦ ) − 𝑖𝑞 (𝑥 ◦ , 𝜉 ◦ ). We obtain

𝜎sub (𝑃𝐹 ) (𝑥 ◦ , 𝜉 ◦ ) = 𝑝 1 (𝑥 ◦ , 𝜉 ◦ ) − 𝑖𝑞 (𝑥 ◦ , 𝜉 ◦ ) (16.3.15)

𝑛
1 ∑︁ 𝜕 2 𝑝 0
+ 𝑖 (𝑥 ◦ , 𝜉 ◦ )
2 𝑗=1 𝜕𝑥 𝑗 𝜕𝜉 𝑗
𝑛 𝑛 ⊤
1 ∑︁ ∑︁ 𝜕 𝜕 𝑝 0 ◦ 𝜕𝐹 ′ −1
+ 𝑖 𝑥 , (𝑥 ) 𝜉◦ .
2 𝑗=1 𝑗=1 𝜕𝑥 ′𝑗 𝜕𝜉 𝑗 𝜕𝑥 ′ ′ ◦ 𝑥 =𝑥

At this point we go back to the defining equation of the matrix 𝐽 (𝑥 ′, 𝑦 ′), (16.1.22):

𝐹 (𝑥 ′) − 𝐹 (𝑦 ′) = 𝐽 (𝑥 ′, 𝑦 ′) (𝑥 ′ − 𝑦 ′) ;

we derive the matrix equation

𝜕𝐽 1 𝜕 𝜕𝐹 ′
(𝑥 ′, 𝑦 ′) = (𝑥 )
𝜕𝑦 ′𝑗 2 𝜕𝑥 ′𝑗 𝜕𝑥 ′
𝑦′ =𝑥 ′

whence, by (16.3.14),
𝑛 𝑛 ⊤
◦ ◦ 1 ∑︁ ∑︁ 𝜕 𝜕 𝑝 0 ◦ 𝜕𝐹 ′ −1
𝑞 (𝑥 , 𝜉 ) = 𝑥 , (𝑥 ) 𝜉◦ .
2 𝑗=1 𝑗=1 𝜕𝑥 ′𝑗 𝜕𝜉 𝑗 𝜕𝑥 ′
𝑥 ′ =𝑥 ◦

Putting this into (16.3.15) proves the desired result. □

Proposition 16.3.12 is the basis for the following

Definition 16.3.13 Let 𝑃 be a classical pseudodifferential operator on a C ∞ manifold

M. The function 𝜎sub (𝑃) on Char (2) 𝑃 is called the subprincipal symbol of 𝑃.

The subprincipal symbol plays a very important role in the analysis of the solutions
of a PDE in neighborhoods of its multiple characteristics.
16.4 The Weyl Calculus in Euclidean Space 603

16.4 The Weyl Calculus in Euclidean Space

16.4.1 Tempered symbols and the Op𝑾 functor

In this section we deal with a special class of symbols in R𝑛 (Definition 16.2.1).

Definition 16.4.1 We shall say that a function 𝑎 (𝑥, 𝜉) ∈ C ∞ R2𝑛 is a tempered

symbol of order 𝑚 ∈ R in R𝑛 if it satisfies the following condition, for some 𝜅 ∈ R:
(TempSymb) To every pair of multi-indices 𝛽, 𝛾 ∈ Z+𝑛 there is a 𝐶𝛽,𝛾 > 0 such that

∀ (𝑥, 𝜉) ∈ R2𝑛 , 𝜕𝑥 𝜕 𝜉 𝑎 (𝑥, 𝜉) ≤ 𝐶𝛽,𝛾 (1 + |𝑥|) 𝜅+|𝛽 | (1 + |𝜉 |) 𝑚−|𝛾 | .

𝛽 𝛾

(16.4.1)
The linear space of tempered symbols of order 𝑚 in R𝑛 shall be denoted by
𝑚
𝑆temp (R𝑛 ); the union of the sets 𝑆temp
𝑚 (R𝑛 ) as 𝑚 ranges over R shall be denoted by
𝑆temp (R ).
𝑛

Let 𝑎 ∈ 𝑆temp
𝑚 (R𝑛 ); it is evident that the symmetrization 𝑎 𝑥+𝑦 2 , 𝜉 is a standard
amplitude of order 𝑚 in R𝑛 × R𝑛 (Definition 16.1.1) and we can form the oscillatory
integral
∫ 𝑥 + 𝑦
𝑎 𝑤 (𝑥, D) 𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉 (16.4.2)
R2𝑛 2

for arbitrary 𝑢 ∈ E ′ (R𝑛 ); 𝑎 𝑤 (𝑥, D) is a standard pseudodifferential operator of order

𝑚 in R𝑛 (Definition 16.1.14).

Proposition 16.4.2 Let 𝑎 ∈ 𝑆temp (R𝑛 ). The restriction of 𝑎 𝑤 (𝑥, D) to Cc∞ (R𝑛 )
extends as a continuous linear map of the Schwartz space S (R𝑛 ) into itself.

Proof We assume that 𝑎 satisfies (TempSymb). Let 𝑣 = (2𝜋) 𝑛 𝑎 𝑤 (𝑥, D) 𝑢, 𝑢 ∈

Cc∞ (R𝑛 ); we know that 𝑣 ∈ C ∞ (R𝑛 ) by Theorem 16.1.4. We begin by proving
that 𝑎 𝑤 (𝑥, D) induces a continuous linear map S (R𝑛 ) −→ 𝐿 ∞ (R𝑛 ). We have, by
(16.4.2), for any 𝑁 ∈ Z+ ,
∫ −𝑁 𝑥 + 𝑦
1 + |𝜉 | 2 , 𝜉 𝑢 (𝑦) 1 − Δ 𝑦 e𝑖 ( 𝑥−𝑦) · 𝜉 d𝑦d𝜉
𝑁
𝑣 (𝑥) = 𝑎
2𝑛 2
∫R −𝑁
𝑁 𝑥 + 𝑦
= e𝑖 ( 𝑥−𝑦) · 𝜉 1 + |𝜉 | 2 1 − Δ𝑦 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉.
R2𝑛 2
According to (16.4.1) there is a 𝐶 𝑁 > 0 such that, for all 𝑥, 𝑦 ∈ R𝑛 ,
𝑁 𝑥 + 𝑦
1 − Δ𝑦 𝑎 , 𝜉 𝑢 (𝑦)
2 ∑︁
≤ 𝐶 𝑁 (1 + |𝑥 + 𝑦|) 𝜅+2𝑁 (1 + |𝜉 |) 𝑚 |D 𝛼 𝑢 (𝑦)| .
| 𝛼 | ≤2𝑁
604 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

If |𝑥 − 𝑦| ≤ 1, after increasing 𝐶 𝑁 we get

𝑁 𝑥 + 𝑦
1 − Δ𝑦 𝑎 , 𝜉 𝑢 (𝑦) (16.4.3)
2 ∑︁
𝜅+2𝑁
≤ 𝐶 𝑁 (1 + |𝑦|) (1 + |𝜉 |) 𝑚 |D 𝛼 𝑢 (𝑦)| .
| 𝛼 | ≤2𝑁

We define
∫ ∫ −𝑁
𝑁 𝑥 + 𝑦
𝑣 ◦ (𝑥) = e𝑖 ( 𝑥−𝑦) · 𝜉 1 + |𝜉 | 2 1 − Δ𝑦 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉,
R𝑛 | 𝑥−𝑦 |<1 2
∫ ∫ −𝑁
𝑁 𝑥 + 𝑦
𝑣 1 (𝑥) = e𝑖 ( 𝑥−𝑦) · 𝜉 1 + |𝜉 | 2 1 − Δ𝑦 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉.
R𝑛 | 𝑥−𝑦 |>1 2

Selecting 2𝑁 > 𝑚 + 𝑛 yields, for some 𝐶◦ > 0 independent of 𝑢 ∈ Cc∞ (R𝑛 ),

∑︁ ∫
sup |𝑣 ◦ (𝑥)| ≤ 𝐶◦ (1 + |𝑦|) 𝜅+2𝑁 |D 𝛼 𝑢 (𝑦)| d𝑦. (16.4.4)
𝑥 ∈R𝑛 | 𝛼 | ≤2𝑁 R𝑛

Setting
−𝑁 𝑁 𝑥 + 𝑦
𝑓 (𝑥, 𝑦, 𝜉) = 1 + |𝜉 | 2 1 − Δ𝑦 𝑎 , 𝜉 𝑢 (𝑦)
2
we have, for 𝑁 ∗ ∈ Z+ ,
∫ ∫
∗
𝑣 1 (𝑥) = |𝑥 − 𝑦| −2𝑁 𝑓 (𝑥, 𝑦, 𝜉) Δ 𝑁 𝜉e
𝑖 ( 𝑥−𝑦) · 𝜉
d𝑦d𝜉
R𝑛 | 𝑥−𝑦 |>1
∫ ∫
∗ ∗
= e𝑖 ( 𝑥−𝑦) · 𝜉 |𝑥 − 𝑦| −2𝑁 Δ 𝑁
𝜉 𝑓 (𝑥, 𝑦, 𝜉) d𝑦d𝜉.
R𝑛 | 𝑥−𝑦 |>1

We observe that
−𝑁 𝑥 + 𝑦
∗ 𝑁 ∗ 2
Δ𝑁
𝜉 𝑓 (𝑥, 𝑦, 𝜉) = 1 − Δ 𝑦 𝑢 (𝑦) Δ 𝑁
𝜉 1 + |𝜉 | 𝑎 ,𝜉
2

whence, by (16.4.1),
∗ ∗
|𝑥 − 𝑦| −2𝑁 Δ 𝑁
𝜉 𝑓 (𝑥, 𝑦, 𝜉)

(1 + |𝑥 + 𝑦|) 𝜅+2𝑁 ∗
∑︁
′
≤ 𝐶𝑁 2𝑁 ∗
(1 + |𝜉 |) 𝑚−2𝑁 −2𝑁 |D 𝛼 𝑢 (𝑦)| .
|𝑥 − 𝑦| | 𝛼 | ≤2𝑁

If |𝑥 − 𝑦| > 1 we have
′′
(1 + |𝑥 + 𝑦|) 𝜅+2𝑁 ≤ 𝐶 𝑁 (1 + 2 |𝑦|) 𝜅+2𝑁 |𝑥 − 𝑦| 𝜅+2𝑁 ,
16.4 The Weyl Calculus in Euclidean Space 605

whence
∗ ∗
|𝑥 − 𝑦| −2𝑁 Δ 𝑁
𝜉 𝑓 (𝑥, 𝑦, 𝜉)
∗ ∗
∑︁
′′′
≤ 𝐶𝑁 |𝑥 − 𝑦| 𝜅+2𝑁 −2𝑁 (1 + |𝜉 |) 𝑚−2𝑁 −2𝑁 (1 + |𝑦|) 𝜅+2𝑁 |D 𝛼 𝑢 (𝑦)| .
| 𝛼 | ≤2𝑁

By taking 2𝑁 ∗ ≥ 𝜅 + 2𝑁 we obtain
∑︁ ∫
sup |𝑣 1 (𝑥)| ≤ 𝐶◦′ (1 + |𝑦|) 𝜅+2𝑁 |D 𝛼 𝑢 (𝑦)| d𝑦
𝑥 ∈R𝑛 | 𝛼 | ≤2𝑁 R𝑛

with 𝐶◦′ > 0 independent of 𝑢. By combining this last estimate with (16.4.4) we
conclude that
∑︁ ∫
∥𝑣∥ 𝐿 ∞ ≤ 𝐶◦′′ (1 + |𝑦|) 𝜅+2𝑁 |D 𝛼 𝑢 (𝑦)| d𝑦,
| 𝛼 | ≤2𝑁 R𝑛

where 𝐶◦′′ = 𝐶◦ + 𝐶◦′ , proving our claim.

For each 𝑗 ∈ Z+ , 1 ≤ 𝑗 ≤ 𝑛, we have, by (16.4.2),
∫ 𝑥 + 𝑦
D 𝑥 𝑗 𝑣 (𝑥) = e𝑖 ( 𝑥−𝑦) · 𝜉 𝜉 𝑗 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉
R2𝑛 2
∫ 𝑥 + 𝑦
+ e𝑖 ( 𝑥−𝑦) · 𝜉 D 𝑥 𝑗 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉.
R2𝑛 2

Both 𝜉 𝑗 𝑎 (𝑥, 𝜉) and D 𝑥 𝑗 𝑎 (𝑥, 𝜉) are tempered symbols, whence D 𝑥 𝑗 𝑣 ∈ 𝐿 ∞ (R𝑛 )
by the first part of the proof. We also have
∫ 𝑥 + 𝑦
1
𝑥 𝑗 − 𝑦 𝑗 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎

𝑥 𝑗 𝑣 (𝑥) = , 𝜉 𝑢 (𝑦) d𝑦d𝜉
2 R2𝑛 2
∫ 𝑥 + 𝑦
1
𝑥 𝑗 + 𝑦 𝑗 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎

+ , 𝜉 𝑢 (𝑦) d𝑦d𝜉
2 2𝑛 2
∫ R
𝑥+𝑦
= 𝑎 , 𝜉 𝑢 (𝑦) D 𝜉 𝑗 e𝑖 ( 𝑥−𝑦) · 𝜉 d𝑦d𝜉
R2𝑛 2
∫ 𝑥𝑗 + 𝑦𝑗 𝑥 + 𝑦
+ e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉
2𝑛 2 2
∫R 𝑥 + 𝑦
=− e𝑖 ( 𝑥−𝑦) · 𝜉 D 𝜉 𝑗 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉
2𝑛 2
∫R 𝑥𝑗 + 𝑦𝑗 𝑥 + 𝑦
+ e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 , 𝜉 𝑢 (𝑦) d𝑦d𝜉.
R2𝑛 2 2

Since 𝑥 𝑗 𝑎 (𝑥, 𝜉) is also a tempered symbol we see that 𝑥 𝑗 𝑣 ∈ 𝐿 ∞ (R𝑛 ). Using

induction on |𝛽| + |𝛾|, 𝛽, 𝛾 ∈ Z+𝑛 , we conclude that the map S (R𝑛 ) ∋ 𝑢 ↦→
𝑥 𝛽 D 𝑥 𝑎 𝑤 (𝑥, D 𝑥 ) 𝑢 ∈ 𝐿 ∞ (R𝑛 ) is continuous, which is what needed to be proved. □
𝛾
606 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

Definition 16.4.3 Let 𝑎 ∈ 𝑆temp (R𝑛 ). The pseudodifferential operator 𝑎 𝑤 (𝑥, D) ∈

Ψ𝑚 (R𝑛 ) will be referred to as the Weyl pseudodifferential operator defined by
𝑎 and denoted by Op𝑊 𝑎. We shall denote by Ψ𝑊 ,𝑚 (R𝑛 ) the set of operators
𝑎 𝑤 (𝑥, D) ∈ 𝑆temp (R𝑛 ), and by Ψ𝑊 (R𝑛 ) the union of the sets Ψ𝑊 ,𝑚 (R𝑛 ) as 𝑚
ranges over R.

For each 𝑚 ∈ R, Ψ𝑊 ,𝑚 (R𝑛 ) is a vector subspace of Ψ𝑚 (R𝑛 ); Ψ𝑊 (R𝑛 ) is a

vector subspace of Ψ (R𝑛 ).
By (16.4.2) the distribution kernel corresponding to the operator 𝑎 𝑤 (𝑥, D) is
∫ 𝑥 + 𝑦
𝐴𝑊 (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 , 𝜉 d𝜉. (16.4.5)
R𝑛 2

This has the direct consequence that the transpose of 𝑎 𝑤 (𝑥, D) is 𝑎 𝑤 (𝑥, −D).

Proposition 16.4.4 Let 𝑎 ∈ 𝑆temp (R𝑛 ); 𝑎 𝑤 (𝑥, D) extends as a continuous linear

map of the Schwartz space of tempered distributions, S ′ (R𝑛 ), into itself.

Proof Indeed, according to Proposition 16.4.2 𝑎 𝑤 (𝑥, −D) maps S (R𝑛 ) contin-
uously into itself and therefore its transpose, which is equal to 𝑎 𝑤 (𝑥, D), maps
S ′ (R𝑛 ) continuously into itself. □
The adjoint of 𝑎 𝑤 (𝑥, D) is 𝑎¯ 𝑤 (𝑥, D); this allows us to state one of the most
important properties of the Weyl calculus:

Proposition 16.4.5 Let 𝑎 ∈ 𝑆temp (R𝑛 ). For 𝑎 𝑤 (𝑥, D) to be self-adjoint it is neces-

sary and sufficient that 𝑎 be real.

The Schwartz Kernel Theorem is valid for S ′ as well as for D ′ (because S is

nuclear, see [Treves, 1967], p. 525 and p. 530): from the fact that 𝑎 𝑤 (𝑥, D 𝑥 ) defines
continuous
endomorphisms
of bothS (R𝑛 ) and S ′ (R 𝑛 ) we derive that 𝑨 𝑊
(𝑥, 𝑦) ∈
S R𝑛𝑥 ; S ′ R𝑛𝑦 ∩ S R𝑛𝑦 ; S ′ R𝑛𝑥 ; S R𝑛𝑥 ; S ′ R𝑛𝑦 ⊗S ′ R𝑛𝑦 is the

S R𝑛𝑥 b
space of C ∞ functions rapidly decaying at infinity in R𝑛 valued in the space of
tempered distributions in R𝑛 (or vice versa).
We are going to exploit the following consequence of (16.4.5):
∫
1 1 ′
𝐴𝑊 𝑥 + 𝑥 ′, 𝑥 − 𝑥 ′ = (2𝜋) −𝑛 e𝑖 𝑥 · 𝜉 𝑎 (𝑥, 𝜉) d𝜉. (16.4.6)
2 2 R 𝑛

Proposition 16.4.6 Let 𝑎 ∈ 𝑆temp𝑚 (R𝑛 ). The distribution kernel (16.4.6) is a C ∞

′
function of (𝑥, 𝑥 ) in R × (R \ {0}) having the following property:
𝑛 𝑛

Given any pair of multi-indices 𝛽, 𝛾 ∈ Z+𝑛 and any integer 𝑁 > 𝑚 + 𝑛 + |𝛾| there
is a constant 𝐶𝛽,𝛾, 𝑁 > 0 such that

1 1
𝜕𝑥 𝜕𝑥′ 𝐴𝑊 𝑥 + 𝑥 ′, 𝑥 − 𝑥 ′ ≤ 𝐶𝛽,𝛾, 𝑁 (1 + |𝑥|) 𝜅+|𝛽 | |𝑥 ′ | −𝑁
𝛽 𝛾
(16.4.7)
2 2

for all 𝑥, 𝑥 ′ ∈ R𝑛 .
16.4 The Weyl Calculus in Euclidean Space 607
𝛽
Proof It suffices to prove (16.4.7) when 𝛽 = 𝛾 = 0 since 𝜉 𝛾 𝜕𝑥 𝑎 satisfies (16.4.1)
with 𝑚 replaced by 𝑚 + |𝛾| and 𝜅 replaced by 𝜅 + |𝛽|. To simplify matters we limit
our attention to even integers 𝑁; we have
∫
1 1 ′ /2
|𝑥 ′ | 𝑁 𝐴𝑊 𝑥 + 𝑥 ′, 𝑥 − 𝑥 ′ = (2𝜋) −𝑛 e𝑖 𝑥 · 𝜉 Δ 𝑁
𝜉 𝑎 (𝑥, 𝜉) d𝜉.
2 2 R 𝑛

By (16.4.1) we know that there is a constant 𝐶 𝑁 > 0 such that

/2
∀𝜉 ∈ R𝑛 , Δ 𝑁 𝜅
𝜉 𝑎 (𝑥, 𝜉) ≤ 𝐶 𝑁 (1 + |𝑥|) (1 + |𝜉 |)
𝑚−𝑁
,

whence the claim if 𝑁 > 𝑚 + 𝑛. □

Proposition 16.4.6 enables us to recover the symbol 𝑎 from the distribution
kernel
1 ′ 1 ′
𝐴𝑊 . Indeed, Proposition 16.4.6 entails that 𝐴 𝑥 + 2 𝑥 , 𝑥 − 2 𝑥 is a tempered
𝑊

distribution with respect to 𝑥 ′ (depending smoothly on 𝑥); by (16.4.6) it is equal to

the inverse
Fourier transform of 𝜉 ↦→ 𝑎 (𝑥, 𝜉), i.e., 𝑎 (𝑥, 𝜉) is the Fourier transform
of 𝐴𝑊 𝑥 + 21 𝑥 ′, 𝑥 − 21 𝑥 ′ with respect to 𝑥 ′, a property expressed by the formula (to
be interpreted in the distribution sense, to make sense of the singularity at 𝑥 ′ = 0)
∫
−𝑖 𝜉 · 𝑥 ′ 𝑊 1 ′ 1 ′
𝑎 (𝑥, 𝜉) = e 𝐴 𝑥 + 𝑥 , 𝑥 − 𝑥 d𝑥 ′. (16.4.8)
R𝑛 2 2

The formal symbol determined by the amplitude 𝑎 𝑥+𝑦 2 , 𝜉 [see (16.2.3)] is

𝑥 + 𝑦 ∑︁ 1
2−|𝛽 | 𝜕𝑥 D 𝜉 𝑎 (𝑥, 𝜉) .
𝛽 𝛽
exp D 𝑦 · 𝜕 𝜉 𝑎 ,𝜉 = (16.4.9)
2 𝑦=𝑥
𝛽 ∈Z 𝑛 𝛽!
+

Now consider a classical symbol

Í∞of order 𝑚 in R (Definition
𝑛 16.3.2), 𝑃 (𝑥, 𝜉),
and its associated formal symbol 𝑗=0 𝑝 𝑗 (𝑥, 𝜉); 𝑝 𝑗 ∈ C ∞ (R𝑛 × (R𝑛 \ {0})) is ho-
mogeneous of degree 𝑚 − 𝑗 for each 𝑗. The formal symbol associated to 𝑃 𝑤 (𝑥, D),
𝑥 + 𝑦 ∞ ∑︁
∑︁ 1 −|𝛽 | 𝛽 𝛽
exp D 𝑦 · 𝜕 𝜉 𝑃 ,𝜉 = 2 𝜕𝑥 D 𝜉 𝑝 𝑗 (𝑥, 𝜉) , (16.4.10)
2 𝑦=𝑥
𝑗=0 𝛽 ∈Z𝑛
𝛽!
+

is classical (Definition 16.3.4). A direct calculation based on (16.4.10) and on Defi-

nition 16.3.13 yields another important property of the Weyl calculus:

Proposition 16.4.7 The subprincipal symbol of 𝑃 𝑤 (𝑥, D) is equal to 𝑝 1 (𝑥, 𝜉).

608 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

16.4.2 Composition of Weyl operators

Let 𝑎 1 (𝑥, 𝜉) and 𝑎 2 (𝑥, 𝜉) be two tempered symbols in R𝑛 and let 𝐴𝑊 𝑗 denote the
𝑊
distribution kernel corresponding to Op 𝑎 𝑗 ( 𝑗 = 1, 2). Thanks to Proposition 16.4.2
the Volterra product
∫
𝐴1𝑊 (𝑥, 𝑧) 𝐴2𝑊 (𝑧, 𝑦) d𝑧 (16.4.11)
R𝑛
∫ 𝑥 + 𝑧 𝑧 + 𝑦
= (2𝜋) −2𝑛 e𝑖 ( 𝑥−𝑧) ·𝜁 +𝑖 (𝑧−𝑦) · 𝜏 𝑎 1 , 𝜁 𝑎2 , 𝜏 d𝜁d𝜏d𝑧
R3𝑛 2 2
∫
makes sense in tempered distributions, since 𝑣 (𝑧) = R𝑛 𝐴2𝑊 (𝑧, 𝑦) 𝑢 (𝑦) d𝑦 ∈
∫
S ′ (R𝑛 ) if 𝑢 ∈ S ′ (R𝑛 ) and then R𝑛 𝐴1𝑊 (𝑥, 𝑧) 𝑣 (𝑧) d𝑧 ∈ S ′ (R𝑛 ).
𝑚
𝑗
Proposition 16.4.8 Let 𝑎 𝑗 (𝑥, 𝜉) ∈ 𝑆temp (R𝑛 ), 𝑗 = 1, 2. There is a tempered symbol
𝑚1 +𝑚2
𝑎 1,2 (𝑥, 𝜉) ∈ 𝑆temp (R ) such that
𝑛

Op𝑊 𝑎 1 ◦ Op𝑊 𝑎 2 = Op𝑊 𝑎 1,2 . (16.4.12)

Proof By applying (16.4.8) to (16.4.11) we define

∫
−𝑖 𝜉 ·𝑡 𝑊 1 𝑊 1
𝑎 1,2 (𝑥, 𝜉) = e 𝐴1 𝑥 + 𝑡, 𝑧 𝐴2 𝑧, 𝑥 − 𝑡 d𝑧d𝑡 (16.4.13)
R2𝑛 2 2
∫
𝑥+𝑧 1 𝑥+𝑧 1
= (2𝜋) −2𝑛 e𝑖Φ 𝑎 1 + 𝑡, 𝜁 𝑎 2 − 𝑡, 𝜏 d𝜁d𝜏d𝑧d𝑡
R4𝑛 2 4 2 4

where

1 1
Φ = −𝜉 · 𝑡 + 𝑥 + 𝑡−𝑧 ·𝜁 + 𝑧−𝑥+ 𝑡 ·𝜏
2 2

1 1
= 𝑥+ 𝑡−𝑧 · (𝜁 − 𝜉) + 𝑧 − 𝑥 + 𝑡 · (𝜏 − 𝜉) .
2 2

The integrals in (16.4.13) make sense in tempered distributions (as will be shown
explicitly below). If we make use of (16.4.8) we see that (16.4.13) implies
∫ ∫ 𝑥 + 𝑦
𝑊 𝑊
𝐴1 (𝑥, 𝑧) 𝐴2 (𝑧, 𝑦) d𝑧 = e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 1,2 , 𝜉 d𝜉. (16.4.14)
R𝑛 R2𝑛 2

It remains to show that 𝑎 1,2 (𝑥, 𝜉) is a tempered symbol. We carry out the change of
variables

𝜁 = 𝜆 + 𝜉, 𝜏 = 𝜇 + 𝜉,
1
𝑧 = 𝑝 + 𝑞 + 𝑥, 𝑡 = 𝑝 − 𝑞,
2
16.4 The Weyl Calculus in Euclidean Space 609

whose Jacobian determinant is equal to 22𝑛 . We have

1 1 1 1
𝑝= 𝑧−𝑥+ 𝑡 ,𝑞 = 𝑧−𝑥− 𝑡 ,
2 2 2 2
𝑥+𝑧 1 𝑥+𝑧 1
+ 𝑡 = 𝑝 + 𝑥, − 𝑡 = 𝑞 + 𝑥,
2 4 2 4
whence Φ = 2 ( 𝑝 · 𝜇 − 𝑞 · 𝜆) and therefore

𝜋 2𝑛 𝑎 1,2 (𝑥, 𝜉)
∫
= e2𝑖 ( 𝑝·𝜇−𝑞·𝜆) 𝑎 1 (𝑥 + 𝑝, 𝜉 + 𝜆) 𝑎 2 (𝑥 + 𝑞, 𝜉 + 𝜇) d𝜆d𝜇d𝑝d𝑞.
R4𝑛

Given arbitrary 𝛽, 𝛾 ∈ Z+𝑛 the Leibniz rule yields

𝜋 2𝑛 𝜕𝑥 𝜕 𝜉 𝑎 1,2 (𝑥, 𝜉)
𝛽 𝛾

∑︁ ∑︁ 𝛽 𝛾 ∫
𝛽′ 𝛾′
= ′ 𝛾′
e2𝑖 ( 𝑝· 𝜇−𝑞·𝜆) 𝜕𝑥 𝜕 𝜉 𝑎 1 (𝑥 + 𝑝, 𝜉 + 𝜆)
𝛽′ ⪯𝛽 𝛾 ⪯𝛾 ′
𝛽 R 4𝑛

𝛽−𝛽′ 𝛾−𝛾 ′
×𝜕𝑥 𝜕𝜉 𝑎2 (𝑥 + 𝑞, 𝜉 + 𝜇) d𝜆d𝜇d𝑝d𝑞.

We must estimate integrals of the type

∫
e2𝑖 ( 𝑝· 𝜇−𝑞·𝜆) 𝜕𝑥 1 𝑏 1 (𝑥 + 𝑝, 𝜉 + 𝜆)
𝛽
𝐼 𝛽,𝛾 =
R4𝑛
𝛽
× 𝜕𝑥 2 𝑏 2 (𝑥 + 𝑞, 𝜉 + 𝜇) d𝜆d𝜇d𝑝d𝑞.
𝑚′
𝑗
where 𝑏 𝑗 ∈ 𝑆temp (R𝑛 ), 𝑗 = 1, 2, with 𝑚 1′ + 𝑚 2′ = 𝑚 1 + 𝑚 2 − |𝛾|, and 𝛽1 + 𝛽2 = 𝛽.
Given arbitrary 𝑁, 𝑁 ∗ ∈ Z+ we have
∫ 𝑁 𝑁 𝑁∗ 𝑁∗
1 1 1 1
𝐼 𝛽,𝛾 = 1 − Δ𝑝 1 − Δ𝑞 1 − Δ𝜆 1 − Δ𝜇 e2𝑖 ( 𝑝·𝜇−𝑞·𝜆)
R4𝑛 4 4 4 4
−𝑁 −𝑁 ∗
× 1 + |𝜆| 2 1 + | 𝑝| 2
𝛽
𝜕𝑥 1 𝑏 1 (𝑥 + 𝑝, 𝜉 + 𝜆)
−𝑁 −𝑁 ∗
× 1 + |𝜇| 2 1 + |𝑞| 2
𝛽
𝜕𝑥 2 𝑏 2 (𝑥 + 𝑞, 𝜉 + 𝜇) d𝜆d𝜇d𝑝d𝑞
∫ −𝑁 𝑁
1
= e2𝑖 ( 𝑝· 𝜇−𝑞·𝜆) 1 + |𝜆| 2 1 − Δ𝑝 𝑩1 (𝑥, 𝜉, 𝑝, 𝜆)
R4𝑛 4
−𝑁 𝑁

2 1
× 1 + |𝜇| 1 − Δ𝑞 𝑩2 (𝑥, 𝜉, 𝑞, 𝜇) d𝜆d𝜇d𝑝d𝑞,
4

where
610 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

−𝑁 ∗ 𝑁∗
1
𝑩1 (𝑥, 𝜉, 𝑝, 𝜆) = 1 + | 𝑝| 2
𝛽
1 − Δ𝜆 𝜕𝑥 1 𝑏 1 (𝑥 + 𝑝, 𝜉 + 𝜆) ,
4
−𝑁 ∗ 𝑁∗

2 1 𝛽
𝑩2 (𝑥, 𝜉, 𝑞, 𝜇) = 1 + |𝑞| 1 − Δ𝜇 𝜕𝑥 2 𝑏 2 (𝑥 + 𝑞, 𝜉 + 𝜇) .
4

Our assumptions are that

−𝑁 𝑁
1
1 + |𝜆| 2 1 − Δ𝑝 𝑩1 (𝑥, 𝜉, 𝑝, 𝜆)
4
′
(1 + |𝑥 + 𝑝|) 𝜅1 +|𝛽1 |+2𝑁 (1 + |𝜉 + 𝜆|) 𝑚1
≲ ∗
(1 + | 𝑝|) 2𝑁 (1 + |𝜆|) 2𝑁
and
−𝑁 𝑁

2 1 𝛽
1 + |𝜇| 1 − Δ𝑞 𝜕𝑥 2 𝑏 2 (𝑥 + 𝑞, 𝜉 + 𝜆)
4
′
(1 + |𝑥 + 𝑞|) 𝜅2 +|𝛽2 |+2𝑁 (1 + |𝜉 + 𝜇|) 𝑚2
≲ ∗ .
(1 + |𝑞|) 2𝑁 (1 + |𝜇|) 2𝑁

We select 2𝑁 ≥ max 𝑚 1′ , 𝑚 2′ + 𝑛 + 1 and 2𝑁 ∗ ≥ 𝜅 2 + |𝛽2 | + 2𝑁 + 𝑛 + 1, thereby

obtaining
𝐼 𝛽,𝛾 ≤ 𝐶3 (1 + |𝑥|) 𝜅1 +𝜅2 +4𝑁 +|𝛽 | (1 + |𝜉 |) 𝑚1 +𝑚2 − |𝛾 | .
1 𝑚 +𝑚2 − |𝛾 |
Since 𝑁 can be selected independently of 𝛽 and 𝛾 this proves 𝐼 𝛽,𝛾 ∈ 𝑆temp (R𝑛 ).□

Corollary 16.4.9 The set Ψ𝑊 (R𝑛 ) is a subalgebra of Ψ (Ω) with respect to operator
composition.

Formal Taylor expansion yields

𝜋 2𝑛 𝑎 1,2 (𝑥, 𝜉)
∫
∑︁ 1
𝜆 𝛼 𝜇 𝛽 e2𝑖 ( 𝑝·𝜇−𝑞·𝜆) 𝜕 𝜉𝛼 𝑎 1 (𝑥 + 𝑝, 𝜉) 𝜕 𝜉 𝑎 2 (𝑥 + 𝑞, 𝜉) d𝜆d𝜇d𝑝d𝑞.
𝛽

𝛼,𝛽 ∈Z 𝑛 𝛼!𝛽! R 4𝑛
+

We have
∫
(2𝜆) 𝛼 (2𝜇) 𝛽 e2𝑖 ( 𝑝·𝜇−𝑞·𝜆) 𝜕 𝜉𝛼 𝑎 1 (𝑥 + 𝑝, 𝜉) 𝜕 𝜉 𝑎 2 (𝑥 + 𝑞, 𝜉) d𝜆d𝜇d𝑝d𝑞
𝛽
R4𝑛
∫
(−D𝑞 ) 𝛼 D 𝑝 e2𝑖 ( 𝑝· 𝜇−𝑞·𝜆) 𝜕 𝜉𝛼 𝑎 1 (𝑥 + 𝑝, 𝜉) 𝜕 𝜉 𝑎 2 (𝑥 + 𝑞, 𝜉) d𝜆d𝜇d𝑝d𝑞
𝛽 𝛽
=
R4𝑛
∫
= (−1) |𝛽 | e2𝑖 ( 𝑝·𝜇−𝑞·𝜆) 𝜕 𝜉𝛼 D 𝑝 𝑎 1 (𝑥 + 𝑝, 𝜉) 𝜕 𝜉 D𝑞𝛼 𝑎 2 (𝑥 + 𝑞, 𝜉) d𝜆d𝜇d𝑝d𝑞
𝛽 𝛽
R 4𝑛

= 𝜋 2𝑛 (−1) |𝛽 | 𝜕 𝜉𝛼 D 𝑥 𝑎 1 (𝑥, 𝜉) 𝜕 𝜉 D 𝑥𝛼 𝑎 2 (𝑥, 𝜉) ,
𝛽 𝛽
16.4 The Weyl Calculus in Euclidean Space 611

𝑤 (𝑥,
which proves that the formal symbol associated to 𝑎 1,2 D) is

1
exp D 𝑥 · 𝜕 𝜉 − D 𝑦 · 𝜕𝜂 (𝑎 1 (𝑦, 𝜉) 𝑎 2 (𝑥, 𝜂)) . (16.4.15)
2 𝑦=𝑥, 𝜂= 𝜉

𝑚1 +𝑚2
As a finite sum of elements of 𝑆temp (R𝑛 ) the symbol

∑︁ (−1) |𝛽 | −| 𝛼+𝛽 | 𝛼 𝛽 𝛽
𝑐 𝑁 (𝑥, 𝜉) = 2 𝜕 𝜉 D 𝑥 𝑎 1 (𝑥, 𝜉) 𝜕 𝜉 D 𝑥𝛼 𝑎 2 (𝑥, 𝜉)
𝛼!𝛽!
| 𝛼+𝛽 | ≤𝑁

𝑚1 +𝑚2
belongs to 𝑆temp (R𝑛 ). We leave the proof of the following statement as an exercise
(cf. the proof of Proposition 16.2.7).
𝑚
𝑗
Proposition 16.4.10 Let 𝑎 𝑗 (𝑥, 𝜉) ∈ 𝑆temp (R𝑛 ), 𝑗 = 1, 2. We have

𝑊 ,𝑚1 +𝑚2 −𝑁
𝑎 1𝑤 (𝑥, D) 𝑎 2𝑤 (𝑥, D) − 𝑐 𝑤
𝑁 (𝑥, D) ∈ Ψ (R𝑛 ) .

16.4.3 An example related to the Schrödinger equation

Consider the following Schrödinger equation in R𝑛+1 :

D𝑡 𝑢 = Δ 𝑥 𝑢 + 𝑉 (𝑥) 𝑢, (16.4.16)
√
where Δ 𝑥 = 𝑛𝑗=1 𝜕 2 /𝜕𝑥 2𝑗 , D𝑡 = − −1𝜕/𝜕𝑡 . We select the potential 𝑉 = 𝑄 + 𝑉1 with
Í
𝑄 (𝑥) a real quadratic form in R𝑛 , 𝑉1 ∈ C ∞ (R𝑛 ) also real-valued and submitted to
the following tempered growth condition:
(Temp) To every 𝛼 ∈ Z+𝑛 there is a 𝐶 𝛼 > 0 such that
1− | 𝛼 |
∀𝑥 ∈ R𝑛 , 𝜕𝑥𝛼𝑉1 (𝑥) ≤ 𝐶 𝛼 1 + |𝑥| 2 . (16.4.17)

We can form exp (𝑖𝑡Δ 𝑥 ), 𝑡 ∈ R, as continuous linear “unitary” transformations of

S (R𝑛 ) and S ′ (R𝑛 ); they form a one-parameter group with respect to composition.
We regard the conjugate 𝑨 (𝑡) = exp (−𝑖𝑡Δ 𝑥 ) 𝑉 (𝑥) exp (𝑖𝑡Δ 𝑥 ) (sometimes referred
to as the interaction representation of the potential) as an operator acting on S (R𝑛 ).
Note that, formally speaking, 𝑨 (𝑡) is the solution of the initial value problem

D𝑡 𝑨 + [Δ 𝑥 , 𝑨] = 0, 𝑨 (0) = 𝑉 (𝑥) , (16.4.18)

where 𝑉 (𝑥) is regarded as a multiplication operator, interpreted in the Weyl calculus

thanks to the Fourier inversion formula and to (16.4.17):
∫ 𝑥 + 𝑦
𝑉 (𝑥) 𝑢 (𝑥) = 𝑉 𝑤 (𝑥, D) 𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑉 𝑢 (𝑦) d𝑦d𝜉.
R2𝑛 2
612 16 Elementary Pseudodifferential Calculus in the 𝐶 ∞ Class

If we write, for 𝑢 ∈ S (R𝑛 ),

∫ 𝑥+𝑦
𝑨𝑢 (𝑥) = 𝑎 𝑤 (𝑥, D) 𝑢 (𝑥) = e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 𝑡, , 𝜉 𝑢 (𝑦) d𝑦d𝜉 (16.4.19)
R2𝑛 2
then (13.3.25) is equivalent to
𝑥+𝑦 𝑥 + 𝑦
D𝑡 𝑎 𝑡, , 𝜉 = −e−𝑖 ( 𝑥−𝑦) · 𝜉 Δ 𝑥 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 𝑡, ,𝜉
2 𝑥 +2 𝑦
+ e−𝑖 ( 𝑥−𝑦) · 𝜉 Δ 𝑦 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 𝑡, ,𝜉 ,
𝑥+𝑦 𝑥 + 𝑦 2
𝑎 0, ,𝜉 =𝑉 .
2 2
Since
𝑥 + 𝑦 𝑥+𝑦
e−𝑖 ( 𝑥−𝑦) · 𝜉 Δ 𝑥 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 𝑡, , 𝜉 = Δ 𝑥 𝑎 𝑡, ,𝜉
2 2
𝑛
∑︁ 𝜕𝑎 𝑥 + 𝑦 𝑥+𝑦
+𝑖 𝜉𝑗 𝑡, , 𝜉 − |𝜉 | 2 𝑎 𝑡, ,𝜉
𝑗=1
𝜕𝑥 𝑗 2 2

we have
𝑥 + 𝑦 𝑥 + 𝑦
Δ 𝑥 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 𝑡, , 𝜉 − Δ 𝑦 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 𝑡, ,𝜉
2 2
𝑛
∑︁ 𝜕𝑎 𝑥 + 𝑦
= 2𝑖e𝑖 ( 𝑥−𝑦) · 𝜉 𝜉𝑗 𝑡, ,𝜉 .
𝑗=1
𝜕𝑥 𝑗 2

After the change of variables

𝑥+𝑦 𝑥−𝑦
⇝ 𝑥, ⇝ 𝑦,
2 2
we see that (13.3.25) is equivalent to
𝑛
∑︁ 𝜕𝑎
𝜕𝑡 𝑎 (𝑡, 𝑥, 𝜉) − 2 𝜉𝑗 (𝑡, 𝑥, 𝜉) = 0, (16.4.20)
𝑗=1
𝜕𝑥 𝑗
𝑎 (0, 𝑥, 𝜉) = 𝑉 (𝑥) , (16.4.21)

whose solution is immediate: 𝑎 (𝑡, 𝑥, 𝜉) = 𝑉 (𝑥 + 2𝑡𝜉). Thus the solution of (16.4.18)

is ∫ 𝑥 + 𝑦
𝑨𝑢 (𝑥) = e𝑖 ( 𝑥−𝑦) · 𝜉 𝑉 + 2𝑡𝜉 𝑢 (𝑦) d𝑦d𝜉. (16.4.22)
R2𝑛 2
By (16.4.2) this means that 𝑨 = 𝑉 (𝑥 + 2𝑡D 𝑥 ) regarded as a standard pseudodif-
ferential operator of order 2 in R𝑛𝑥 . By Proposition 16.4.5 𝑨 is self-adjoint since
𝑉 (𝑥 + 2𝑡𝜉) is real; and since
16.4 The Weyl Calculus in Euclidean Space 613

exp (−𝑖𝑡Δ 𝑥 ) (D𝑡 − Δ 𝑥 ) exp (𝑖𝑡Δ 𝑥 ) = D𝑡

conjugation with exp (𝑖𝑡Δ 𝑥 ) transforms (16.4.17) into

D𝑡 𝑢 = 𝑉 (𝑥 + 2𝑡D 𝑥 ) 𝑢 (16.4.23)
= 𝑄 (𝑥 + 2𝑡D 𝑥 ) 𝑢 + 𝑉1 (𝑥 + 2𝑡D 𝑥 ) .

Note that D𝑡 −𝑄 (𝑥 + 2𝑡D 𝑥 ) is a differential operator of order 2 whereas 𝑉1 (𝑥 + 2𝑡D 𝑥 )

is a pseudodifferential operator of order 1. For the construction of a (sort of)
parametrix based on the reduction (16.4.23) see [Treves, 1995]. The presence of
the operator 𝑥 + 2𝑡D 𝑥 is significant in quantum mechanics, related to the propagation
set (see [Sigal-Soffer, 1987]).
Chapter 17
Analytic Pseudodifferential Calculus

Pseudodifferential calculus in the C ∞ category, as described (at the most elementary

level) in the preceding chapter, needs substantial refinements when translated into
the C 𝜔 category. That the amplitudes 𝑎 (𝑥, 𝑦, 𝜉) be required to be analytic with
respect to the base variables (𝑥, 𝑦) goes without saying. But the need for cut-offs
in frequency space in balls of ever larger radius or, even worse, in whole open
cones, makes it difficult to deal solely with amplitudes that are analytic also with
respect to 𝜉. We shall therefore content ourselves with asymptotic analyticity, as
explained in Subsection 17.1.1 in the simple case of amplitudes depending on 𝜉
alone, i.e., multipliers. This is generalized in Subsection 17.1.2 by the reinsertion of
the position variables, producing the concept of pseudoanalytic amplitudes (and later,
pseudoanalytic symbols). In Section 20.1, we will introduce the standard analytic
symbols and, when they are of finite order, establish their relation to pseudoanalytic
symbols.
Analytic pseudodifferential operators are (at least, semiglobally) operators of the
form Op𝑎, with 𝑎 a pseudoanalytic amplitude. In Subsection 17.2.3 we introduce
cutoff functions in frequency space akin to Ehrenpreis’ sequences (Section 3.2) –
designed to play the role of C ∞ cutoffs 𝑔 (𝜉) homogeneous for large |𝜉 | and supported
in cones, and constructed for this purpose by L. Hörmander. These special cutoffs
are not quite pseudoanalytic multipliers but they admit asymptotically holomorphic
extensions and, either alone or composed with pseudoanalytic operators, they are
an important tool in circumscribing in phase-space the analytic wave-front sets
(introduced in Section 3.4) of distribution solutions of PDEs. However, it does not
seem to be practical to adapt this tool to deal with hyperfunctions or microfunctions.
In Section 17.4 we describe the action of analytic pseudodifferential operators on
singularity hyperfunctions (see Subsection 11.3.2) through their action on analytic
functionals, itself defined by duality to their action on holomorphic functions. We
show that these actions decrease the analytic wave-front set (defined in Euclidean
space in Section 7.5, extended to manifolds in Section 11.3) of a singularity hy-
perfunction. The natural next step is taken in Section 17.5: to microlocalize the
operators themselves, which is to say: to define microdifferential operators acting on

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 615
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_17
616 17 Analytic Pseudodifferential Calculus

microdistributions (on the latter, see Subsection 11.3.3). This is a very simple matter
as our initial Definition 17.1.12 (for distributions) is already essentially microlocal
in nature.

17.1 Analytic Pseudodifferential Operators

17.1.1 Asymptotically analytic multipliers

Throughout this chapter ℭ will be an open cone in 𝜉-space R𝑛 \ {0}, possibly equal
to R𝑛 \ {0}; generally speaking Ω shall be a domain in 𝑥-space R𝑛 , ΩC a domain in
𝑧-space C𝑛 such that Ω ⊂ ΩC ∩ R𝑛 . In this subsection 𝑎 will denote a C ∞ function
in ℭ of tempered growth at infinity. We investigate conditions on 𝑎 ensuring the
existence of extensions e𝑎 (𝜉, 𝜂) to a conic closed subset of C𝑛 of the type ℭ + 𝑖R𝑛 ,
𝑎 decays exponentially fast as |𝜉 | → +∞ (𝜁 = 𝜉 + 𝑖𝜂, 𝜉 ∈ ℭ). This will
such that 𝜕 𝜁 e
prepare the introduction of pseudoanalytic amplitudes (Definition 17.1.4) below.
We use the notation e 𝑎 (𝜁) in place of e
𝑎 (𝜉, 𝜂) even when e𝑎 is not a holomorphic
function of 𝜁; we write
1
𝑛 𝑛 2 2
∑︁ 𝜕e
𝑎 ©∑︁ 𝜕e 𝑎 ª
𝑎=
𝜕𝜁 e 𝑑𝜁 𝑗 , 𝜕 𝜁 e
𝑎 =­ ® .
𝑗=1 𝜕𝜁 𝑗 𝑗=1 𝜕𝜁 𝑗
« ¬
Lemma 17.1.1 Let 𝑎 (𝜉) be a C ∞ function in ℭ. Suppose that there are numbers
𝑚 ∈ R and 𝐶◦ > 0, 𝜌◦ > 0 such that, for all 𝛾 ∈ Z+𝑛 and all 𝜉 ∈ ℭ,
|𝛾 |+1
𝛾! |𝜉 | 𝑚−|𝛾 | .
𝛾
|𝜉 | ≥ 𝜌◦ max (1, |𝛾|) =⇒ 𝜕 𝜉 𝑎 (𝜉) ≤ 𝐶◦ (17.1.1)

Then there are constants 𝐶 𝑗 > 0, 𝑗 = 1, 2, such that the following is true. To every
number 𝑅 ≥ 2𝜌◦ there is a C ∞ function e 𝑎 𝑅 (𝜁) in ℭ + 𝑖R𝑛 satisfying the following
inequalities

𝑎 𝑅 (𝜁)| ≤ 𝐶◦ (1 + |𝜉 |) 𝑚 exp (𝐶1 |𝜂| /𝑅) ,

|e (17.1.2)
𝑎 𝑅 (𝜁) ≤ 𝐶2 (1 + |𝜉 |) 𝑚 exp ((𝐶1 |𝜂| − |𝜉 |) /𝑅) ,
𝜕𝜁 e (17.1.3)

𝑎 𝑅 (𝜉) = 𝑎 (𝜉) if |𝜉 | > 𝑅 (𝜉 ∈ ℭ).

and such that e

Remark 17.1.2 The inequalities (17.1.1), (17.1.2), (17.1.3) extend in 𝜉-space to the
closure of ℭ in R𝑛 \ {0}, ℭ.

It will be convenient to say that 𝑎 (𝜉) is asymptotically extended as the function

𝑎 𝑅 (𝜁) in ℭ +𝑖R𝑛 when the properties in Lemma 17.1.1 hold. This same terminology
e
will be used even when 𝑎 depends on other variables than 𝜉.
17.1 Analytic Pseudodifferential Operators 617

Proof Let 𝐶◦ and 𝜌◦ be the positive constants in (17.1.1); we may as well assume
𝜌◦ ≥ 1. We select a function 𝜒 ∈ C ∞ (R), 0 ≤ 𝜒 ≤ 1, 𝜒 (𝜏) = 0 if 𝜏 < 21 , 𝜒 (𝜏) = 1
if 𝜏 > 1. We define, for 𝑅 ≥ 𝜌◦ and 𝜁 = 𝜉 + 𝑖𝜂 ∈ ℭ + 𝑖R𝑛 ,

𝑎 𝑅 (𝜁) = 𝜒 (|𝜉 | /𝑅) 𝑎 (𝜉)

e (17.1.4)
∑︁ (−𝜂) 𝛾 |𝜉 |
𝛾
+ 𝜒 D 𝜉 𝑎 (𝜉) .
0≠𝛾 ∈Z𝑛
𝛾! 𝑅 |𝛾|
+

|𝜉 | 1
Note that 𝛾 ≠ 0 and 𝜒 𝑅 |𝛾 | ≠ 0 =⇒ |𝜉 | > 2𝑅 |𝛾| ≥ 𝜌◦ |𝛾|, allowing us to apply
(17.1.1):

|e
𝑎 𝑅 (𝜁)| ∑︁ |𝜉 | |𝛾 |+1
≤ 𝐶◦ 𝜒 (|𝜉 | /𝑅) + |𝜂 𝛾 | 𝜒 𝐶 |𝜉 | −|𝛾 |
(1 + |𝜉 |) 𝑚 𝑅 |𝛾| ◦
𝛾 ∈Z+𝑛 , |𝛾 | ≥1
|𝛾 |
∑︁ 2𝐶◦ |𝜂|
≤ 𝐶◦ ­ 𝜒 (|𝜉 | /𝑅) +
© ª
|𝛾|
®
𝑅
« 𝛾 ∈Z+𝑛 , |𝛾 | ≥1 ¬
∞
∑︁ 1
≤ 𝐶◦ (𝑀𝐶◦ |𝜂| /𝑅) 𝑘 ,
𝑘=0
𝑘!

the last inequality a consequence of Stirling’s formula; 𝑀 > 0 is a universal constant.

We obtain (17.1.2) with 𝐶1 ≥ 𝑀𝐶◦ . We also have

𝜕e
𝑎𝑅 𝜕e𝑎𝑅 𝜕𝑎
+𝑖 (𝜁) = (𝜉)
𝜕𝜉 𝑗 𝜕𝜂 𝑗 𝜕𝜉 𝑗
√ ∑︁ 1 𝛾+ ⟨ 𝑗 ⟩
+ −1 (−𝜂) 𝛾 𝜒 (|𝜉 | /𝑅 |𝛾|) D 𝜉 𝑎 (𝜉)
𝛾!
𝛾 ∈Z+ , |𝛾 | ≥1
𝑛

√ ∑︁ 1
(−𝜂) 𝛾− ⟨ 𝑗 ⟩ 𝜒 (|𝜉 | /𝑅 |𝛾|) D 𝜉 (𝑎 (𝜉))
𝛾
− −1
𝛾 ∈Z+ ,𝛾 𝑗 ≠0
𝑛 (𝛾 − ⟨ 𝑗⟩)!
∑︁ 1
𝛾
𝜕
+ (−𝜂) 𝛾 D 𝜉 𝑎 (𝜉) 𝜒 (|𝜉 | /𝑅 |𝛾|) ,
𝛾! 𝜕𝜉 𝑗
𝛾 ∈Z+ , |𝛾 | ≥1
𝑛

where ⟨ 𝑗⟩ is the multi-index (𝛼1 , ..., 𝛼𝑛 ) such that 𝛼 𝑗 = 1 and 𝛼 𝑘 = 0 if 𝑘 ≠ 𝑗. We

derive

𝜕 𝜕 𝜕𝑎
+𝑖 𝑎 𝑅 (𝜁) ≤ (1 − 𝜒 (|𝜉 | /𝑅))
e (𝜉)
𝜕𝜉 𝑗 𝜕𝜂 𝑗 𝜕𝜉 𝑗

∑︁ |𝜂 𝛾 | |𝜉 | |𝜉 | 𝛾+ ⟨ 𝑗 ⟩
+ 𝜒 −𝜒 D𝜉 𝑎 (𝜉)
𝛾! 𝑅 |𝛾| 𝑅 (|𝛾| + 1)
𝛾 ∈Z+𝑛 , |𝛾 | ≥1

∑︁ |𝜂 𝛾 | 𝜕 |𝜉 | 𝛾
+ 𝜒 D 𝜉 𝑎 (𝜉) .
𝛾! 𝜕𝜉 𝑗 𝑅 |𝛾|
𝛾 ∈Z+ , |𝛾 | ≥1
𝑛
618 17 Analytic Pseudodifferential Calculus

We know that 1 − 𝜒 (|𝜉 | /𝑅) ≠ 0 =⇒ |𝜉 | ≤ 𝑅 and that

|𝜉 | |𝜉 | 1 1
𝜒 −𝜒 ≠ 0 =⇒ |𝛾| ≤ |𝜉 | ≤ |𝛾| + 1,
𝑅 |𝛾| 𝑅 (|𝛾| + 1) 2 𝑅

𝜕 |𝜉 | 1 1
𝜒 ≠ 0 =⇒ |𝛾| ≤ |𝜉 | ≤ |𝛾| .
𝜕𝜉 𝑗 𝑅 |𝛾| 2 𝑅

Thanks to (17.1.1) and to Stirling’s formula we get

1 |𝜉 | |𝜉 | 𝛾+ ⟨ 𝑗 ⟩
𝜒 −𝜒 D𝜉 𝑎 (𝜉)
2 𝑅 |𝛾| 𝑅 (|𝛾| + 1)

1 |𝛾 |+2 |𝜉 | |𝜉 |
≤ 𝐶◦ 𝜒 −𝜒 |𝜉 | 𝑚−|𝛾 |−1
2 𝑅 |𝛾| 𝑅 (|𝛾| + 1)
|𝛾 |+2
≤ 2 |𝛾 | 𝐶◦ (𝛾 + ⟨ 𝑗⟩)! (𝑅 |𝛾|) −|𝛾 |−1 |𝜉 | 𝑚

|𝛾 | |𝛾 |+2 1
≤ 2 𝐶◦ 1+ 𝑅 −|𝛾 |−1 e−|𝛾 | |𝜉 | 𝑚
|𝛾|
≤ 𝐶◦ (𝑀𝐶◦ /𝑅) |𝛾 | |𝜉 | 𝑚 e− | 𝜉 |/𝑅 ,

1 𝜕 |𝜉 | 𝛾
𝜒 D 𝜉 𝑎 (𝜉)
2 𝜕𝜉 𝑗 𝑅 |𝛾|

1 |𝛾 |+1 𝜕 |𝜉 |
≤ 𝐶◦ 𝛾! 𝜒 |𝜉 | 𝑚−|𝛾 |
2 𝜕𝜉 𝑗 𝑅 |𝛾|
|𝛾 |+1
≤ 2 |𝛾 | 𝐶◦ 𝛾! (𝑅 |𝛾|) − |𝛾 | |𝜉 | 𝑚
≤ 𝐶◦ (𝑀𝐶◦ /𝑅) |𝛾 | |𝜉 | 𝑚 e−|𝛾 |
≤ 𝐶◦ 𝑒 (𝑀𝐶◦ /𝑅) |𝛾 | |𝜉 | 𝑚 e− | 𝜉 |/𝑅 ,

again with 𝑀 > 0 a universal constant. We reach the conclusion that, for some
positive constants 𝐶1 , 𝐶2 and all 𝜁 ∈ Ω + 𝑖R𝑛 ,
∞
1 𝜕 𝜕 ∑︁ 1
+𝑖 𝑎 𝑅 (𝜁) ≤ 𝐶2 |𝜉 | 𝑚 e− | 𝜉 |/𝑅
e (𝐶1 |𝜂| /𝑅) 𝑘 ,
2 𝜕𝜉 𝑗 𝜕𝜂 𝑗 𝑘=0
𝑘!

whence (17.1.3). □

Remark 17.1.3 The conclusion in Lemma 17.1.1 is not equivalent to the hypothesis.
Indeed, the hypothesis implies that 𝜕 𝛼 𝑎 (0 ≠ 𝛼 ∈ Z+𝑛 ) has properties similar to those
of 𝑎 but with possibly increased 𝐶◦ and 𝜌◦ . It is obviously possible to prove estimates
similar to (17.1.2)–(17.1.3) for 𝜕 𝛼 𝑎˜ 𝑅 . A careful control of the constants 𝐶 𝑗 ( 𝑗 = 1, 2)
in those estimates would likely produce a property equivalent to the hypothesis. We
shall not pursue this matter further.
17.1 Analytic Pseudodifferential Operators 619

17.1.2 Pseudoanalytic amplitudes

The following definition is crucial to the approach to analytic microlocal analysis in

this part of the book;1 Ω1 and Ω2 are open subsets of R𝑛 ; 𝑚 is a real number.

Definition 17.1.4 A function 𝑎 (𝑥, 𝑦, 𝜉) ∈ C ∞ (Ω1 × Ω2 × ℭ) will be called a pseu-

doanalytic amplitude of order 𝑚 in Ω1 × Ω2 × ℭ if there are open subsets of C𝑛 ,
ΩC𝑗 ⊃ Ω 𝑗 ( 𝑗 = 1, 2), such that 𝑎 (𝑧, 𝑤, 𝜉) can be extended to ΩC1 × ΩC2 × ℭ as a C ∞
function holomorphic with respect to (𝑧, 𝑤) and having the following property:
To each compact set 𝐾 C ⊂ ΩC1 × ΩC2 there are two positive numbers 𝐶, 𝜌, such
that, for all 𝛾 ∈ Z+𝑛 , all (𝑧, 𝑤) ∈ 𝐾 C and all 𝜉 ∈ ℭ,

|𝜉 | ≥ 𝜌 max (1, |𝛾|) =⇒ 𝜕 𝜉 𝑎 (𝑧, 𝑤, 𝜉) ≤ 𝐶 |𝛾 |+1 𝛾! |𝜉 | 𝑚−|𝛾 | .

𝛾
(17.1.5)

A pseudoanalytic amplitude of order 𝑚 in Ω1 ×Ω2 ×ℭ will be called a pseudoanalytic

symbol of order 𝑚 in Ω1 × Ω2 × ℭ if it is independent of 𝑦.

Remark 17.1.5 The concept of pseudoanalytic amplitude is local in the following

sense. Suppose there are open coverings 𝑈 𝑗, 𝜄 𝜄 ∈𝐼 𝑗 of Ω 𝑗 , 𝑗 = 1, 2, and for each
pair of indices 𝜄1 ∈ 𝐼1 , 𝜄2 ∈ 𝐼2 , a pseudoanalytic amplitude 𝑎 𝜄1 , 𝜄2 of order 𝑚 in
𝑈1, 𝜄1 × 𝑈2, 𝜄2 × ℭ, and that these amplitudes coincide in overlaps. There is then a
pseudoanalytic amplitude 𝑎 of order 𝑚 in Ω1 ×Ω2 ×ℭ such that 𝑎|𝑈1, 𝜄 ×𝑈2, 𝜄 ×ℭ = 𝑎 𝜄1 , 𝜄2
1 2
for every (𝜄1 , 𝜄2 ) ∈ 𝐼1 × 𝐼2 .

𝜓a (Ω1 × Ω2 × ℭ) the linear space of pseudoanalytic ampli-

We shall denote by 𝑆 𝑚
tudes of order 𝑚 in Ω1 × Ω2 × ℭ. If Ω′𝑗 ⊂ Ω 𝑗 ( 𝑗 = 1, 2), ℭ ′ ⊂ ℭ a cone in R𝑛 \ {0}
′
and 𝑚 ′ > 𝑚, the restriction map induces a linear map 𝑆 𝑚 ′ ′ ′ →

𝜓a Ω1 × Ω2 × ℭ
𝑆𝑚𝜓a (Ω1 × Ω2 × ℭ). We shall also use the following notation:
Ø
𝑆 𝜓a (Ω1 × Ω2 × ℭ) = 𝑆𝑚
𝜓a (Ω1 × Ω2 × ℭ) , (17.1.6)
𝑚∈R
Ù
𝑆 −∞
𝜓a (Ω1 × Ω2 × ℭ) = 𝑆𝑚
𝜓a (Ω1 × Ω2 × ℭ) . (17.1.7)
𝑚∈R

When ℭ = R𝑛 \ {0} we omit ×ℭ in the notation. In what follows ΩC1 and ΩC2 will
be related to the amplitude 𝑎 (𝑥, 𝑦, 𝜉) as in Definition 17.1.4.

𝜓a (Ω1 × Ω2 × ℭ) then to each compact set 𝐾 ⊂ Ω1 ×

Proposition 17.1.6 If 𝑎 ∈ 𝑆 𝑚 C C
C
Ω2 there are two positive numbers 𝐶, 𝜌, such that

𝜕𝑧𝛼 𝜕𝑤 𝜕 𝜉 𝑎 (𝑧, 𝑤, 𝜉) ≤ 𝐶 | 𝛼+𝛽+𝛾 |+1 𝛼!𝛽!𝛾! |𝜉 | 𝑚−|𝛾 |

𝛽 𝛾
(17.1.8)

for all 𝛼, 𝛽, 𝛾 ∈ Z+𝑛 , all (𝑧, 𝑤) ∈ 𝐾 C and all 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌 max (1, |𝛾|).
1 A more directly complex-analytic approach will be described in Part VI.
620 17 Analytic Pseudodifferential Calculus
𝛾
Proof Apply the Cauchy inequalities to the holomorphic function 𝜕 𝜉 𝑎 (𝑧, 𝑤, 𝜉) of
(𝑧, 𝑤) ∈ ΩC1 × ΩC2 depending on the parameter 𝜉 and satisfying (17.1.5). □

Corollary 17.1.7 Every pseudoanalytic amplitude of order 𝑚 in Ω1 × Ω2 × R𝑛 is a

standard amplitude of order 𝑚 in Ω1 × Ω2 (Definition 16.1.1).

Proof Let the compact set 𝐾 C ⊂ ΩC1 ×ΩC2 , the constants 𝜌 and 𝐶 be as in Proposition
17.1.6. For 𝛾 ∈ Z+𝑛 denote by 𝑀 𝛼,𝛽,𝛾 the maximum of

(1 + |𝜉 |) −𝑚+|𝛾 | 𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉)

𝛽 𝛾

for (𝑥, 𝑦) ∈ 𝐾 = 𝐾 C ∩ R2𝑛 , |𝜉 | ≤ 𝜌 max (1, |𝛾|). Taking (17.1.8) into account we
get, for all (𝑥, 𝑦, 𝜉) ∈ 𝐾 × ℭ,

𝜕𝑥𝛼 𝜕𝑦 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) ≤ 𝐶 𝛼,𝛽,𝛾 (1 + |𝜉 |) 𝑚−|𝛾 | ,

𝛽 𝛾

where 𝐶 𝛼,𝛽,𝛾 = max 𝑀 𝛼,𝛽,𝛾 , 𝐶 | 𝛼+𝛽+𝛾 |+1 , whence the claim [cf. (16.1.1)]. □

Thus 𝑆 𝜓a (Ω1 × Ω2 ) ⊂ 𝑆 (Ω1 × Ω2 ) and 𝑆 −∞

𝜓a (Ω1 × Ω2 ) ⊂ 𝑆
−∞ (Ω × Ω ).
1 2
The next proposition states that pseudoanalytic amplitudes have extensions that
are asymptotically holomorphic.

Proposition 17.1.8 Let 𝑎 ∈ 𝑆 𝑚 𝜓a (Ω1 × Ω2 × ℭ) and let U ⊂⊂ Ω1 × Ω2 be an open

C C

set such that (17.1.5) holds for all (𝑧, 𝑤) ∈ U and all 𝜉 ∈ ℭ. There are constants
𝐶 𝑗 > 0, 𝑗 = 0, 1, 2, such that the following is true. Given any sufficiently large
positive number 𝑅, 𝑎 (𝑥, 𝑦, 𝜉) can be asymptotically extended as a C ∞ function
𝑎 𝑅 (𝑧, 𝑤, 𝜁) in U × (ℭ + 𝑖R𝑛 ), holomorphic with respect to (𝑧, 𝑤) and satisfying the
e
following estimates, for all (𝑧, 𝑤, 𝜁) ∈ U × (ℭ + 𝑖R𝑛 ),

𝑎 𝑅 (𝑧, 𝑤, 𝜁)| ≤ 𝐶0 (1 + |Re 𝜁 |) 𝑚 exp (𝐶1 |Im 𝜁 | /𝑅) ;

|e (17.1.9)
𝑎 𝑅 (𝑧, 𝑤, 𝜁) ≤ 𝐶2 (1 + |Re 𝜁 |) 𝑚 exp ((𝐶1 |Im 𝜁 | − |Re 𝜁 |) /𝑅) .
𝜕𝜁 e (17.1.10)

Proof Same as that of Lemma 17.1.1 but for the insertion of the parameters (𝑧, 𝑤)
on which 𝑎 depends holomorphically. □

Remark 17.1.9 It is important to note that the inequalities (17.1.5), (17.1.8), (17.1.9),
(17.1.10), extend in 𝜉-space to ℭ (cf. Remark 17.1.2).

We introduce right-away an important subclass of amplitudes.

Definition 17.1.10 By an exponentially decaying amplitude in Ω1 ×Ω2 ×ℭ we shall

mean a function 𝑟 (𝑥, 𝑦, 𝜉) ∈ C ∞ (Ω1 × Ω2 × ℭ) that has an extension 𝑟˜ (𝑧, 𝑤, 𝜉) to
ΩC1 × ΩC2 × ℭ (ΩC𝑗 ⊃ Ω 𝑗 open subsets of C𝑛 , 𝑗 = 1, 2) holomorphic with respect to
(𝑧, 𝑤) and satisfying the following condition:
17.1 Analytic Pseudodifferential Operators 621

• To each compact set 𝐾 C ⊂ ΩC1 × ΩC2 there are positive numbers 𝐶, 𝜌 and 𝑅 such
that, for all 𝛾 ∈ Z+𝑛 , all (𝑧, 𝑤) ∈ 𝐾 C and all 𝜉 ∈ ℭ,

|𝜉 | ≥ 𝜌 max (1, |𝛾|) =⇒ 𝜕 𝜉 𝑟˜ (𝑧, 𝑤, 𝜉) ≤ 𝐶 |𝛾 |+1 exp (− |𝜉 | /𝑅) .

𝛾
(17.1.11)

We shall denote by 𝑆 −𝜔
𝜓a (Ω1 × Ω2 × ℭ) the set of exponentially decaying ampli-
tudes.

We write 𝑆 −𝜔 −𝜔
𝜓a (Ω1 × Ω2 ) rather than 𝑆 𝜓a (Ω1 × Ω2 × R ).
𝑛

Proposition 17.1.11 We have 𝑆 −𝜔 −∞

𝜓a (Ω1 × Ω2 ) ⊂ 𝑆 𝜓a (Ω1 × Ω2 ) [see (17.1.7)].

Proof Let 𝑟 and 𝑟˜ be as in Definition 17.1.10. We note that (17.1.11) remains valid
if we increase 𝜌 𝐾 , 𝑅 and 𝐶𝐾 . We have

|𝜉 | |𝛾 | ≤ (2𝑛𝑅) |𝛾 | 𝛾! exp (|𝜉 | /2𝑅)

and for every 𝑚 ∈ Z+ ,

|𝜉 | 𝑚 ≤ (2𝑅) 𝑚 𝑚! exp (|𝜉 | /2𝑅) .

We conclude that

𝜕 𝜉 𝑟˜ (𝑧, 𝑤, 𝜉) ≤ 𝐶 (2𝑅) 𝑚 (2𝐶 𝑅) |𝛾 | 𝑚!𝛾! |𝜉 | −𝑚−|𝛾 |

𝛾

for all (𝑧, 𝑤) ∈ 𝐾 C and 𝜉 ∈ R𝑛 , |𝜉 | ≥ 𝜌 𝐾 max (1, |𝛾|). This proves the claim [cf.
(17.1.5)]. □
If to every compact subset 𝐾 of Ω1 × Ω2 there is a number 𝜌 > 0 such that
𝑟 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝜓a (Ω1 × Ω2 × ℭ) vanishes identically for |𝜉 | > 𝜌 then 𝑟 (𝑥, 𝑦, 𝜉) ∈
𝑆 −𝜔
𝜓a (Ω1 × Ω2 × ℭ).
From (17.1.11) and the Cauchy inequalities, possibly after some contraction of
𝐾 C ⊂ ΩC1 × ΩC2 about 𝐾 C ∩ R2𝑛 we derive, for all 𝛼, 𝛽, 𝛾 ∈ Z+𝑛 , (𝑧, 𝑤) ∈ 𝐾 C , 𝜉 ∈ ℭ,
|𝜉 | ≥ 𝜌 max (1, |𝛾|), and possibly increased 𝐶 > 0,

𝜕𝑧𝛼 𝜕𝑤 𝜕 𝜉 𝑟˜ (𝑧, 𝑤, 𝜉) ≤ 𝐶 | 𝛼+𝛽+𝛾 |+1 𝛼!𝛽! exp (− |𝜉 | /𝑅) .

𝛽 𝛾
(17.1.12)

Example 17.1.23 below shows that 𝑆 −𝜔 −∞

𝜓a (Ω1 × Ω2 ) ≠ 𝑆 𝜓a (Ω1 × Ω2 ).

17.1.3 Analytic pseudodifferential operators acting on distributions

If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝜓a (Ω1 × Ω2 ) the operator Op𝑎 : E ′ (Ω2 ) −→ D ′ (Ω1 ) has already

been defined, in (17.1.13) (see also Corollary 16.1.7). We must now take the cone ℭ
into account. We introduce the operator
622 17 Analytic Pseudodifferential Calculus
∫
E ′ (Ω2 ) ∋ 𝑓 ↦→ 𝑨ℭ 𝑓 (𝑥) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑓 (𝑦) d𝑦d𝜉,
Ω2 ×ℭ
(17.1.13)
where the integration with respect to 𝑦 stands for a duality bracket.

Definition 17.1.12 Assuming that 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 𝜓a (Ω1 × Ω2 × ℭ) (Definition

17.1.4; 𝑚 ∈ R) we shall refer to (17.1.13) as an analytic pseudodifferential opera-
tor of order 𝑚 from Ω2 to Ω1 associated to ℭ and denote it by Op𝑎.

Occasionally, we will write 𝑨ℭ instead of Op𝑎 (in the latter notation ℭ is regarded
as part of the definition of 𝑎).
The analytic pseudodifferential operators of order 𝑚 from Ω2 to Ω1 associated to
Ø space that shall be denoted by Ψa (Ω1 × Ω2 , ℭ); we define
the cone ℭ form a vector 𝑚

Ψa (Ω1 × Ω2 , ℭ) = Ψa𝑚 (Ω1 × Ω2 , ℭ). If Ω1 = Ω2 = Ω we shall refer to 𝑨ℭ as

𝑚∈R
an analytic pseudodifferential operator in Ω (associated to the cone ℭ) and write
Ψa (Ω, ℭ) and Ψa𝑚 (Ω, ℭ) rather than Ψa (Ω × Ω, ℭ) and Ψa𝑚 (Ω × Ω, ℭ). In all this
there will be no mention of ℭ when ℭ = R𝑛 \ {0}.
For each 𝑚 ∈ R we can identify Ψa𝑚 (Ω1 × Ω2 ) with a vector subspace of
Ψ (Ω1 × Ω2 ).
𝑚

If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚
𝜓a (Ω1 × Ω2 × ℭ) then

𝑎 ∗ (𝑥, 𝑦, 𝜉) = 𝑎(𝑦, 𝑥, 𝜉) ∈ 𝑆 𝑚
𝜓a (Ω2 × Ω1 × ℭ) and
𝑎 ⊤ (𝑥, 𝑦, 𝜉) = 𝑎(𝑦, 𝑥, −𝜉) ∈ 𝑆 𝑚
𝜓a (Ω2 × Ω1 × (−ℭ))

(cf. Proposition 16.1.6); Op𝑎 ⊤ ∈ Ψa𝑚 (Ω2 × Ω1 , −ℭ) is the transpose of Op𝑎, and
Op𝑎 ∗ ∈ Ψa𝑚 (Ω2 × Ω1 , ℭ) is the adjoint of Op𝑎.

17.1.4 Analyticity of the distribution kernel off the diagonal

Proposition 17.1.8 enables us to prove the analytic analogue of Theorem 16.1.9. We

shall avail ourselves of the pseudoanalytic version of Lemma 16.1.3

Lemma 17.1.13 Let 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 𝜓a (Ω1 × Ω2 × ℭ). Given any 𝑁 ∈ Z+ there are
amplitudes 𝑏 1, 𝛼 , 𝑏 2, 𝛼 ∈ 𝑆 𝑚−𝑁
𝜓a (Ω1 × Ω2 × ℭ) with 𝛼 ∈ Z+ , |𝛼| ≤ 2𝑁, such that
𝑛

∑︁
Op𝑎 = Op𝑏 1, 𝛼 D 𝛼 (17.1.14)
| 𝛼 | ≤2𝑁
∑︁
= D 𝛼 Op𝑏 2, 𝛼 . (17.1.15)
| 𝛼 | ≤2𝑁

Proof The proof of (17.1.14) is the same as that of Lemma 16.1.3, combined with
the observation that 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚
𝜓a (Ω1 × Ω2 × ℭ) implies
17.1 Analytic Pseudodifferential Operators 623
−𝑁
1 + |𝜉 | 2 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚−2𝑁
𝜓a (Ω1 × Ω2 × ℭ) ;

(17.1.15) follows by applying (17.1.14) to Op𝑎 ⊤ and then transposing (cf. Remark
16.1.8). □

In the next statement, we are going to use the following notation: if 𝐾 ⊂ Ω1 × Ω2

and 𝜀 > 0, 𝐾 𝜀 = (𝑧, 𝑤) ∈ C2𝑛 ; dist ((𝑧, 𝑤) , 𝐾) < 𝜀 .

Theorem 17.1.14 Let 𝑎 ∈ C ∞ (Ω1 × Ω2 × ℭ) have the following property:

Given an arbitrary compact set 𝐾 ⊂ Ω1 × Ω2 , if 𝑅 > 0 is sufficiently large then 𝑎
can be asymptotically extended as a function e 𝑎 𝑅 (𝑧, 𝑤, 𝜁) ∈ C ∞ 𝐾1/𝑅 × (ℭ + 𝑖R𝑛 ) ,
holomorphic with respect to (𝑧, 𝑤) and satisfying the inequalities (17.1.9)–(17.1.10)
for all (𝑧, 𝑤, 𝜁) ∈ 𝐾1/𝑅 × (ℭ + 𝑖R𝑛 ) (with positive constants 𝐶◦ , 𝐶1 , 𝐶2 depending
on 𝑎 and 𝐾 but not on 𝑅).
Under these hypotheses the distribution kernel
∫
𝐴ℭ (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝑦, 𝜉) d𝜉 (17.1.16)
ℭ

is analytic in the open set {(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 ≠ 𝑦}.

𝑁
Proof By Lemma 17.1.13 it suffices to prove the result when 𝑎 = 1 + |𝜉 | 2 𝑏
with 𝑏 ∈ 𝑆 𝑚−2𝑁𝜓a (Ω1 × Ω2 × ℭ). If 𝐵 (𝑥, 𝑦) denotes the distribution kernel de-
fined by the amplitude 𝑏, the analyticity of 𝐵 (𝑥, 𝑦) in an open subset of Ω1 × Ω2
entails that of 𝐴ℭ (𝑥, 𝑦). In other words, it suffices to prove the result when
𝑎 ∈ 𝑆 𝑚−2𝑁
𝜓a (Ω1 × Ω2 × ℭ). Taking 2𝑁 > 𝑚 + 𝑛 allows us to assume that 𝑚 < −𝑛
and that the integral in (17.1.16) is absolutely convergent.
Let e𝑎 𝑅 (𝑧, 𝑤, 𝜁) be an extension of 𝑎 to ΩC1 × ΩC2 × C𝑛 as per Proposition 17.1.8
(𝑅 > 0 will be chosen below) and let (17.1.9)–(17.1.10) hold with

U ⊂⊂ (𝑧, 𝑤) ∈ ΩC1 × ΩC2 ; 𝑧 ≠ 𝑤 .

We have, for (𝑥, 𝑦) ∈ U ∩ R2𝑛 ,

∫
𝐴ℭ (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 e
𝑎 𝑅 (𝑥, 𝑦, 𝜉) d𝜉.
ℭ

We deform the domain of integration from ℭ to the image of ℭ under the map

𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜀 |𝜉 | (𝑥 − 𝑦)

with 𝜀 > 0 small, to be chosen below. Stokes’ theorem implies

624 17 Analytic Pseudodifferential Calculus

(2𝜋) 𝑛 𝐴ℭ (𝑥, 𝑦) (17.1.17)

∫
2
= e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜀 | 𝑥−𝑦 | |𝜉 |
𝑎 𝑅 (𝑥, 𝑦, 𝜉 + 𝑖𝜀 (𝑥 − 𝑦) |𝜉 |) Δ 𝜀 (𝑥 − 𝑦, 𝜉) d𝜉
e
ℭ
∫
+ e𝑖 ( 𝑥−𝑦) ·𝜁 𝜕 𝜁 e
𝑎 𝑅 (𝑥, 𝑦, 𝜁) ∧ d𝜁,
Λ 𝜀 ( 𝑥−𝑦)

where Δ 𝜀 (𝑥, 𝜉) is the Jacobian determinant of the map 𝜉 ↦→ 𝜉 + 𝑖𝜀𝑥 |𝜉 | and

Λ 𝜀 (𝑥) = {𝜁 ∈ C𝑛 ; ∃ (𝜉, 𝑡) ∈ R𝑛 × (0, 1) , 𝜁 = 𝜉 + 𝑖𝜀𝑡𝑥 |𝜉 |} .

Note that Δ 𝜀 (𝑥, 𝜉) is homogeneous of degree zero. We take advantage of (17.1.9):

𝑎 𝑅 (𝑥, 𝑦, 𝜉 + 𝑖𝜀 (𝑥 − 𝑦) |𝜉 |)| ≤ 𝐶◦ (1 + |𝜉 |) 𝑚 exp (𝐶1 𝜀 |𝑥 − 𝑦| |𝜉 | /𝑅) .

|e

We derive
2
e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜀 | 𝑥−𝑦 | |𝜉 |
𝑎 𝑅 (𝑥, 𝑦, 𝜉 + 𝑖𝜀 (𝑥 − 𝑦) |𝜉 |)
e
𝑚
≤ 𝐶◦ (1 + |𝜉 |) exp (−𝜀 (|𝑥 − 𝑦| − 𝐶1 /𝑅) |𝑥 − 𝑦| |𝜉 |) .

Requiring
1
𝑅 −1 < inf |𝑥 − 𝑦|
2𝐶1 ( 𝑥,𝑦) ∈U
leads to
2
e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜀 |𝑥−𝑦 | |𝜉 |
𝑎 𝑅 (𝑥, 𝑦, 𝜉 + 𝑖𝜀 (𝑥 − 𝑦) |𝜉 |)
e (17.1.18)

𝑚 1
≤ 𝐶◦ (1 + |𝜉 |) exp − 𝜀 |𝑥 − 𝑦| |𝜉 |
2
≤ 𝐶◦ (1 + |𝜉 |) 𝑚 exp (−𝐶1 𝜀 |𝜉 | /𝑅) .

Next we take advantage of (17.1.10):

𝑎 𝑅 (𝑥, 𝑦, 𝜉 + 𝑖𝜀𝑡 (𝑥 − 𝑦) |𝜉 |)
𝜕𝜁 e
≤ 𝐶2 (1 + |𝜉 |) 𝑚 exp (− (1 − 𝐶1 𝜀𝑡 |𝑥 − 𝑦|) |𝜉 | /𝑅) .

We require 𝜀 > 0 to be sufficiently small that

1
𝐶1 𝜀𝑡 |𝑥 − 𝑦| ≤ 𝐶1 𝜀 sup |𝑥 − 𝑦| ≤ .
( 𝑥,𝑦) ∈U 2

We get, for all 𝜁 ∈ Λ 𝜀 (𝑥 − 𝑦),

2
e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜀𝑡 | 𝑥−𝑦 | |𝜉 |
𝑎 𝑅 (𝑥, 𝑦, 𝜁) ≤ 𝐶2 (1 + |𝜉 |) 𝑚 e− | 𝜉 |/2𝑅 .
𝜕𝜁 e
17.1 Analytic Pseudodifferential Operators 625

This last estimate together with (17.1.18) allows the holomorphic extension to U
𝜀
of the integrands in (17.1.17) provided (𝑧, 𝑤) ∈ U =⇒ |Im 𝑧| + |Im 𝑤| ≪ 2𝑅 .
2 2 Í 𝑛 2
[Naturally, the holomorphic extension of |𝑥 − 𝑦| is ⟨𝑧 − 𝑤⟩ = 𝑗=1 𝑧 𝑗 − 𝑤 𝑗 .] □

Corollary 17.1.15 If 𝑎 ∈ 𝑆 𝜓a (Ω1 × Ω2 × ℭ) the distribution kernel 𝐴ℭ (𝑥, 𝑦) as-

sociated to the operator Op𝑎 [i.e., given by (17.1.16)] is analytic in the open set
{(𝑥, 𝑦) ∈ Ω1 × Ω2 ; 𝑥 ≠ 𝑦}.

Proof Apply Theorem 17.1.14 taking Proposition 17.1.8 into account. □

The next statement is meaningful in the case Ω1 = Ω2 = Ω.

Corollary 17.1.16 If 𝑨𝕮 ∈ Ψa (Ω, ℭ) then

singsuppa 𝑨ℭ 𝑢 ⊂ supp 𝑢

whatever 𝑢 ∈ E ′ (Ω).

Proof We must show that if 𝑢 ≡ 0 in an open subset 𝑉 then 𝑨𝕮 𝑢| 𝑉 ∈ C 𝜔 (𝑉). By

Theorem 17.1.14 the kernel 𝐴ℭ (𝑥, 𝑦) is real-analytic in 𝑉 × Ω\𝑉 : if 𝐾1 ⊂ 𝑉 is a
compact set and 𝐾2 = supp 𝑢 ⊂ Ω\𝑉 there is a constant 𝐶 > 0 such that

∀𝛼 ∈ Z+𝑛 , max D 𝑥𝛼 𝐴ℭ (𝑥, 𝑦) ≤ 𝐶 | 𝛼 |+1 𝛼!.

( 𝑥,𝑦) ∈𝐾1 ×𝐾2

Differentiation with respect to 𝑥 under the integral sign implies directly that
∫
𝐴ℭ (𝑥, 𝑦) 𝑢 (𝑦) d𝑦 ∈ C 𝜔 (𝑉) . □
ℭ 𝑉

17.1.5 Analytic regularizing operators

We are going to need the C 𝜔 analogue of the regularizing (or smoothing) operators
E ′ (Ω2 ) −→ C ∞ (Ω1 ) (Definition 2.3.3); recall that these were characterized by the
fact that the associated distribution kernels belonged to C ∞ (Ω1 × Ω2 ). This justifies
the following

Definition 17.1.17 A continuous linear operator 𝑹 : Cc∞ (Ω2 ) −→ D ′ (Ω1 ) is said

to be analytic (or C 𝜔 ) regularizing if the distribution kernel 𝑅 (𝑥, 𝑦) associated to
𝑹 belongs to C 𝜔 (Ω1 × Ω2 ). These operators form a vector space which we denote
by Ψa−𝜔 (Ω1 × Ω2 ).

The property of a continuous linear operator 𝑹 : Cc∞ (Ω2 ) −→ D ′ (Ω1 ) to be

analytic regularizing is local.
626 17 Analytic Pseudodifferential Calculus

Proposition 17.1.18 If 𝑹 ∈ Ψa−𝜔 (Ω1 × Ω2 ) then 𝑹 maps Cc∞ (Ω2 ) into C 𝜔 (Ω1 ),
its transpose 𝑹 ⊤ maps Cc∞ (Ω1 ) into C 𝜔 (Ω2 ) and they extend as linear operators
E ′ (Ω2 ) −→ C 𝜔 (Ω1 ) and E ′ (Ω1 ) −→ C 𝜔 (Ω2 ) respectively.

The proof is straightforward since 𝑅 (𝑥, 𝑦) extends holomorphically to an open

subset of C2𝑛 that contains Ω1 × Ω2 . As a matter of fact, 𝑹 maps O ′ (Ω2 ) into
C 𝜔 (Ω1 ) [O ′ (Ω2 ): the space of analytic functionals carried by compact subsets of
Ω2 ] and its transpose 𝑹 ⊤ maps O ′ (Ω1 ) into C 𝜔 (Ω2 ) (see Definition 6.1.3 and
Corollary 6.3.9).
All the linear maps under consideration here are continuous; this is not difficult
to see by looking closely at the topology of C 𝜔 (Ω1 ), which is the inductive limit of
the Fréchet spaces O Ω1 as Ω1 ranges over the family of open subsets of C𝑛 that
C C

contain Ω1 . Another way of looking at this is to observe that the graphs of all the
maps under consideration are closed, since they are continuous maps into D ′ (Ω1 ).
The closed graph theorem applies to all the locally convex topological vector spaces
we encounter (see [De Wilde, 1978]).
We have the analogue of Proposition 16.1.16:

Proposition 17.1.19 Suppose 𝑟 (𝑧, 𝑤, 𝜉) ∈ C ∞ ΩC1 × ΩC2 × ℭ is holomorphic with

respect to (𝑧, 𝑤) and satisfies the following condition [cf. (17.1.11)]:
To each compact set 𝐾 C ⊂ ΩC1 × ΩC2 there are positive numbers 𝐶, 𝜌 and 𝑅 such
that |𝑟 (𝑧,∫ 𝑤, 𝜉)| ≤ 𝐶 exp (− |𝜉 | /𝑅) for all (𝑧, 𝑤) ∈ 𝐾 C and all 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌.
Then ℭ e𝑖 ( 𝑥−𝑦) · 𝜉 𝑟 (𝑥, 𝑦, 𝜉) d𝜉 ∈ C 𝜔 (Ω1 × Ω2 ).
Conversely, if a distribution kernel 𝑅 (𝑥, 𝑦) extends as a holomorphic function
in ΩC1 × ΩC2 then the associated operator Cc∞ (Ω2 ) −→ D ′ (Ω1 ) is equal to Op𝑟 for
some 𝑟 ∈ 𝑆a−𝜔 (Ω1 × Ω2 × ℭ) (Definition 17.1.10).

Proof The first assertion∫ is self-evident. Let 𝑅 (𝑧, 𝑤) ∈ O ΩC1 × ΩC2 and select
𝜑 ∈ Cc∞ (R𝑛 ) such that ℭ 𝜑 (𝜉) d𝜉 = 1; the amplitude e−𝑖 ( 𝑥−𝑦) · 𝜉 𝑅 (𝑥, 𝑦) 𝜑 (𝜉) is
exponentially decaying in Ω1 × Ω2 × ℭ (Definition 17.1.10). □

Remark 17.1.20 Proposition 17.1.19 implies that in dealing with operators Op𝑟,
𝑟 ∈ 𝑆a−𝜔 (Ω1 × Ω2 × ℭ), we may omit all mentions of ℭ.

Corollary 17.1.21 We have Ψa−𝜔 (Ω1 × Ω2 ) ⊂ Ψa−∞ (Ω1 × Ω2 ).

Proof Follows directly from Propositions 17.1.11, 17.1.19. □

Remark 17.1.22 The reader will notice that the condition in Proposition 17.1.19
𝛾
involves no bound on the derivatives 𝜕 𝜉 𝑟 (𝑧, 𝑤, 𝜉), 0 ≠ 𝛾 ∈ Z+𝑛 , and it is not clear,
a priori, that we can derive Ψa−𝜔 (Ω1 × Ω2 ) ⊂ Ψ−∞ (Ω1 × Ω2 ) from it. But the
converse part of Proposition 17.1.19 implies that, if said condition holds, then there
is a possibly different amplitude that belongs to 𝑆a−𝜔 (Ω1 × Ω2 × ℭ) and defines the
same distribution kernel 𝑅 (𝑥, 𝑦).

The following example shows that Ψa−𝜔 (Ω1 × Ω2 ) ≠ Ψa−∞ (Ω1 × Ω2 ) and there-
fore 𝑆 −𝜔 −∞
𝜓a (Ω1 × Ω2 ) ≠ 𝑆 𝜓a (Ω1 × Ω2 ).
17.1 Analytic Pseudodifferential Operators 627

Example 17.1.23 Let 𝑎 ∈ C ∞ (R) satisfy the following conditions:

√︁
𝑎 (𝜉) = exp − 𝜉 for 𝜉 > 1,
𝑎 (𝜉) = 0 for 𝜉 ≤ 0.

There is a constant 𝐶 > 0 such that, for all 𝑚 ∈ Z+ , 𝛾 ∈ Z+ \ {0} and all 𝜉 > 1,
1 √
|𝜕 𝛾 𝑎 (𝜉)| ≤ 𝐶 𝛾+1 𝛾!𝜉 −𝛾− 2 e− 𝜉

1 √
= 𝐶 𝛾+1 𝛾!𝜉 −𝑚−𝛾− 2 𝜉 𝑚 e− 𝜉

1
≤ (2𝑚)!𝐶 𝛾+1 𝛾!𝜉 −𝑚−𝛾− 2 .

This shows that, if we regard 𝑎 as an amplitude in R × R then 𝑎 ∈ 𝑆 −𝑚 𝜓a (R × R)

whatever 𝑚 ∈ R. But 𝑎 ∉ 𝑆 −𝜔
𝜓a (R × R). Indeed, the distribution kernel
∫ 1 ∫ +∞ √
1 1
𝐴 (𝑥 − 𝑦) = e𝑖 𝜉 ( 𝑥−𝑦) 𝑎 (𝜉) d𝜉 + e𝑖 𝜉 ( 𝑥−𝑦)− 𝜉
d𝜉
2𝜋 0 2𝜋 1

is analytic in the complement of the diagonal in R×R but not at points of diag (R × R).
It is of Gevrey 2 class at those points.

17.1.6 Composition of analytic pseudodifferential operators

Defining the composition of two operators

𝑨ℭ,1 ∈ Ψa (Ω1 × Ω2 , ℭ) , 𝑨ℭ,2 ∈ Ψa (Ω2 × Ω3 × ℭ) ,

is delicate since they are not properly supported (Definition 2.3.6). We can palliate
these difficulties by exploiting Theorem 17.1.14 and modding off analytic functions
and analytic regularizing operators (Definition 17.1.17). From this perspective it is
important to keep in mind the following consequence of Theorem 17.1.14:

Proposition 17.1.24 The restriction from Ω1 ×Ω2 to (Ω1 ∩ Ω2 ) × (Ω1 ∩ Ω2 ) induces

a linear injection

Ψa (Ω1 × Ω2 , ℭ) /Ψa−𝜔 (Ω1 × Ω2 ) −→ Ψa (Ω1 ∩ Ω2 , ℭ) /Ψa−𝜔 (Ω1 ∩ Ω2 ) .

Proof Let the restriction of 𝑨ℭ ∈ Ψa (Ω1 × Ω2 , ℭ) to (Ω1 ∩ Ω2 ) × (Ω1 ∩ Ω2 ) × ℭ

belong to Ψa−𝜔 (Ω1 ∩ Ω2 , ℭ); this means that the associated distribution ker-
nel, 𝐴ℭ (𝑥, 𝑦) ∈ D ′ (Ω1 × Ω2 ), is analytic in (Ω1 ∩ Ω2 ) × (Ω1 ∩ Ω2 ). Since
singsuppa 𝐴ℭ (𝑥, 𝑦) is a closed subset of (Ω1 ∩ Ω2 ) × (Ω1 ∩ Ω2 ) it must be empty.□

Corollary 17.1.25 If Ω1 ∩ Ω2 = ∅ then Ψa (Ω1 × Ω2 , ℭ) /Ψa−𝜔 (Ω1 × Ω2 ) = {0}.

628 17 Analytic Pseudodifferential Calculus

If we wish to compose cosets belonging to Ψa (Ω1 × Ω2 , ℭ) /Ψa−𝜔 (Ω1 × Ω2 ) we

might as well compose cosets belonging to Ψa (Ω1 ∩ Ω2 , ℭ) /Ψa−𝜔 (Ω1 ∩ Ω2 ). Or,
put more simply, we can assume from the start that Ω 𝑗 = Ω, an open subset of
R𝑛 , for 𝑗 = 1, 2. Thus we deal with a pair of analytic pseudodifferential operators
𝑚
𝑨ℭ, 𝑗 ∈ Ψa 𝑗 (Ω, ℭ), 𝑗 = 1, 2, and their associated distribution kernels 𝐴ℭ,1 (𝑥, 𝑦),
𝐴ℭ,2 (𝑦, 𝑧).
Consider an open set Ω′ ⊂⊂ Ω and a function 𝜑 ∈ Cc∞ (Ω) such that 𝜑 ≡ 1 in Ω′.
We can form the Volterra product
∫

𝑨ℭ,1 𝜑 𝑨ℭ,2 (𝑥, 𝑧) = 𝐴ℭ,1 (𝑥, 𝑦) 𝜑 (𝑦) 𝐴ℭ,2 (𝑦, 𝑧) d𝑦.
ℭ

It is a distribution kernel in Ω × Ω, defining a standard pseudodifferential operator

𝑨ℭ,1 𝜑 𝑨ℭ,2 : E ′ (Ω) −→ D ′ (Ω) of order 𝑚 1 +𝑚 2 . The restriction of 𝜑 (𝑦) 𝐴ℭ,2 (𝑦, 𝑧)
to Ω′ × Ω′ coincides with that of 𝐴ℭ,2 (𝑥, 𝑦).
Let 𝑢 ∈ E ′ (Ω) be arbitrary. If 𝐴ℭ,1 (𝑥, 𝑦) ∈ C 𝜔 (Ω × Ω) then we have
𝑨ℭ,1 𝜑 𝑨ℭ,2 𝑢 Ω′ ∈ C 𝜔 (Ω′). If 𝐴ℭ,2 (𝑥, 𝑦) ∈ C 𝜔 (Ω × Ω) then we have 𝜑 𝑨ℭ,2 𝑢 Ω′ ∈
C 𝜔 (Ω′); it then follows from Theorem 17.2.13 below that, in this case too,
𝑨ℭ,1 𝜑 𝑨ℭ,2 𝑢 Ω′ ∈ C 𝜔 (Ω′). This means that if we are willing to reason modulo C 𝜔
functions it is permitted to mod off analytic regularizing operators. In dealing with
the composite 𝑨ℭ,1 𝜑 𝑨ℭ,2 it is convenient to view the operator 𝑨ℭ, 𝑗 whose distribu-
𝑚
tion kernel is 𝐴ℭ, 𝑗 as the representative of a coset 𝑨ℭ, 𝑗 in Ψa 𝑗 (Ω, ℭ) /Ψa−𝜔 (Ω).

If 𝜓 ∈ Cc∞ (Ω) is also identically equal to 1 in Ω′ then, by Corollary 17.1.16, we see

that
𝐴ℭ,1 (𝜑 − 𝜓) 𝐴ℭ,2 (𝑥, 𝑧) Ω′ ×Ω′ ∈ C 𝜔 (Ω′ × Ω′) .

The restriction (Definition 2.3.9) to Ω′ of the operator 𝑨ℭ,1 (𝜑 − 𝜓) 𝑨ℭ,2 belongs to

𝑚1 +𝑚2
Ψa−𝜔 (Ω′). We conclude that 𝑨ℭ,1 𝜑 𝑨ℭ,2 defines a coset in Ψa (Ω′, ℭ) /Ψa−𝜔 (Ω′)
independent of 𝜑, which we shall denote by 𝑨ℭ,1 𝑨ℭ,2 Ω′ .
•𝑚
This said we form the direct limit Ψ 𝑎 (Ω, ℭ) of the spaces Ψa𝑚 (Ω′, ℭ) /Ψa−𝜔 (Ω′),
namely the subset of the (infinite) product
Ö
Ψa𝑚 (Ω′, ℭ) /Ψa−𝜔 (Ω′)
Ω′ ⊂ ⊂Ω

consisting of the elements whose components [ 𝑨ℭ ] satisfy the following “coherence”

condition: if Ω′′ is an open subset of R𝑛 , Ω′′ ⊂ Ω′ ⊂⊂ Ω, then the restriction of
[ 𝑨ℭ | Ω′ ] to Ω′′ is equal to [ 𝑨ℭ | Ω′′ ]. Here “restriction” means the map

Ψa𝑚 (Ω′, ℭ) /Ψa−𝜔 (Ω′) −→ Ψa𝑚 (Ω′′, ℭ) /Ψa−𝜔 (Ω′′) .

defined by the restriction maps Ψa𝑚 (Ω′, ℭ) −→ Ψa𝑚 (Ω′′, ℭ), Ψa−𝜔 (Ω′) −→
•
Ψa−𝜔 (Ω′′). The sets Ψ𝑎𝑚 (Ω, ℭ) carry a natural vector space structure. There is a nat-
ural (linear) injection Ψa𝑝 (Ω′, ℭ) /Ψa−𝜔 (Ω′) ↩→ Ψa𝑞 (Ω′, ℭ) /Ψa−𝜔 (Ω′) if 𝑞 < 𝑝;
•𝑝 •𝑞
this defines a natural injection Ψ 𝑎 (Ω, ℭ) ↩→ Ψ 𝑎 (Ω, ℭ). We define
17.1 Analytic Pseudodifferential Operators 629
• Ø •𝑚
Ψa (Ω, ℭ) = Ψ 𝑎 (Ω, ℭ) . (17.1.19)
𝑚∈R

The composite element 𝑨ℭ,1 𝑨ℭ,2 Ω′ ∈ Ψa𝑚1 +𝑚2 (Ω′, ℭ) /Ψa−𝜔 (Ω′) defined above

•
is the Ω′-component of the composite of two elements of Ψa (Ω, ℭ); composition
•
turns Ψa (Ω, ℭ) into a noncommutative (but associative) algebra. The elements of
•
Ψa (Ω, ℭ) can be thought of as linear operators E ′ (Ω) −→ D ′ (Ω) /C 𝜔 (Ω). Since
C 𝜔 (Ω) is dense in D ′ (Ω) the target space does not carry a Hausdorff topology
compatible with its linear structure.
In the preceding paragraph the open sets Ω′ were supposed to expand and fill
•𝑚
Ω. Selecting arbitrarily a point 𝑥 ◦ ∈ Ω we now consider the direct limit Ψ 𝑎 (𝑥 ◦ , ℭ)
of the spaces Ψa𝑚 (𝑈, ℭ) /Ψa−𝜔 (𝑈, ℭ) as the neighborhoods 𝑈 of 𝑥 ◦ contract to 𝑥 ◦ ,
namely the subset of the (infinite) product
Ö
Ψa𝑚 (𝑈, ℭ) /Ψa−𝜔 (𝑈, ℭ)
𝑈 ∋𝑥 ◦

consisting of the elements whose components 𝑨𝑈,ℭ satisfy the following “co-
herence” condition: if 𝑈1 and 𝑈2 are two neighborhoods
h ofi 𝑥 ◦h there is ai third
neighborhood 𝑈3 of 𝑥 ◦ such that 𝑈3 ⊂ 𝑈1 ∩ 𝑈2 and 𝑨𝑈1 ,ℭ 𝑈3 = 𝑨𝑈2 ,ℭ 𝑈3 . This
•𝑚
defines the stalk A 𝑥 ◦ ,ℭ of a sheaf of complex vector spaces over R𝑛 , Ψa,ℭ . The union
of these sheaves as 𝑚 ranges over R is the sheaf in R𝑛 , which we shall denote by
•
𝚿a,ℭ , of germs of analytic pseudodifferential operators associated to the cone ℭ, mod
analytic regularizing operators.
•
The definition of 𝚿a,ℭ could have been approached differently: by introducing
the presheaf (see Subsection 1.2.2) Ψa𝑚 (𝑈, ℭ) , 𝜌𝑈 𝑉 with 𝑚 ∈ R as well as 𝑚 =
−∞ or 𝑚 = −𝜔, and 𝑈 ranging over the family of open subsets of R𝑛 ; 𝜌𝑈 𝑉 :
Ψa𝑚 (𝑈, ℭ) −→ Ψa𝑚 (𝑉, ℭ), 𝑉 ⊂ 𝑈, is the natural restriction map (Definition 2.3.9).
This presheaf defines the sheaf 𝚿a,ℭ
𝑚 of germs of analytic pseudodifferential operators

in R associated to the cone ℭ. If 𝑚 ∈ R we can form the quotient sheaf 𝚿a,ℭ

𝑛 𝑚 /𝚿−𝜔
Ø a
and the sheaves 𝚿a,ℭ = 𝚿a,ℭ
𝑚 (all the stalks are complex vector spaces); we have

𝑚∈R
•
𝚿a,ℭ = 𝚿a,ℭ /𝚿−𝜔
Recall (Remark 17.1.20) that 𝚿−𝜔
a . a can be identified with the
sheaf of germs of C functions in R𝑛 as well as with a subsheaf of 𝚿a,ℭ
𝜔 𝑚 whatever
•
ℭ and 𝑚. The space of continuous sections of 𝚿a,ℭ over the open set Ω can be
• •
identified with Ψa (Ω, ℭ). The composition of elements of Ψa (Ω, ℭ) defines a sheaf
•
map 𝚿a,ℭ 𝑚1
/𝚿−𝜔 × 𝚿 𝑚2
/𝚿−𝜔 −→ 𝚿𝑚1 +𝑚2 /𝚿 and turns 𝚿
a a,ℭ a a,ℭ a a,ℭ into a sheaf of
rings.
630 17 Analytic Pseudodifferential Calculus

We can also contract the cone ℭ about one of its rays, {𝜆𝜉 ◦ }𝜆>0 , 𝜉 ◦ ∈ ℭ ∩ S𝑛−1 .
If combined with contraction about 𝑥 ◦ this would lead to a sheaf over R𝑛 × S𝑛−1 . We
shall approach such a sheaf from the standpoint of singularity hyperfunctions and
microfunctions, later in this chapter.

17.1.7 The effect of analytic diffeomorphisms

Adapting to the C 𝜔 case the observations in Subsection 16.1.4, is straightforward.

To lighten the notation we limit our attention to the case Ω1 = Ω2 = Ω, an open
subset of R𝑛 (although the extension to the general case Ω1 ≠ Ω2 is routine).
We deal with two open subsets Ω and Ω′ of R𝑛 and a C 𝜔 diffeomorphism of Ω′
onto Ω, 𝑥 ′ ↦→ 𝑥 = 𝐹 (𝑥 ′). (In the general case we would have to deal with two
diffeomorphisms Ω′𝑗 −→ Ω 𝑗 , 𝑗 = 1, 2.)

Proposition 17.1.26 The conjugation 𝑨ℭ −→ 𝐹∗−1 𝑨ℭ (𝐹 ∗ ) −1 derived from the di-

agram (16.1.20) in which 𝑨=𝑨ℭ induces an isomorphism of Ψa𝑚 (Ω × ℭ) onto
Ψa𝑚 (Ω′ × ℭ) (𝑚 ∈ R or 𝑚 = −𝜔).
Proof It suffices (cf. Remark 17.1.5) to deal with an operator Op𝑎 where 𝑎 ∈
𝜓a (𝑈1 × 𝑈2 × ℭ), with 𝑈1 and 𝑈2 open subsets of Ω contracted as needed about
𝑆𝑚
points 𝑥 ◦ , 𝑦 ◦ ∈ Ω. We refer the reader to Formula (16.1.21): if 𝐴ℭ (𝑥, 𝑦) is the
distribution kernel associated with Op𝑎 then the distribution kernel associated with
𝐹∗−1 (Op𝑎) (𝐹 ∗ ) −1 is

𝜕𝐹 ′
𝐴ℭ (𝐹 (𝑥 ′) , 𝐹 (𝑦 ′)) det (𝑦 ) ∈ D ′ 𝑈1′ × 𝑈2′ ,

𝜕𝑥 ′

−1
where 𝑈 ′𝑗 = 𝐹 𝑈 𝑗 and 𝜕𝑥
𝜕𝐹
′ is the Jacobian matrix of the map 𝐹.

The claim that 𝐹∗ (Op𝑎) (𝐹 ∗ ) −1 ∈ 𝑆 −𝜔

−1 ′ ′ ◦ ◦

𝜓a 𝑈1 × 𝑈2 is trivially true when 𝑥 ≠ 𝑦
since, in this case, we can assume that 𝑈1 ∩ 𝑈2 = ∅ and therefore 𝐴ℭ (𝑥, 𝑦) ∈
C 𝜔 (𝑈1 × 𝑈2 ) by Theorem 17.1.14 and, as a consequence,

𝜕𝐹 ′
𝐴ℭ (𝐹 (𝑥 ′) , 𝐹 (𝑦 ′)) det (𝑦 ) ∈ C 𝜔 𝑈1′ × 𝑈2′ ,

𝜕𝑥 ′

implying 𝐹∗−1 Op𝑎 (𝐹 ∗ ) −1 ∈ 𝑆 −𝜔 ′ ′

𝜓a 𝑈1 × 𝑈2 ⊂ 𝑆 𝜓a (𝑈1 × 𝑈2 ) (Corollary 17.1.21).
𝑚

In the case 𝑥 = 𝑦 we can take 𝑈1 = 𝑈2 = 𝑈 = 𝐹 (𝑈 ′) sufficiently small that the

◦ ◦

𝑛 × 𝑛 matrix 𝐽 (𝑥 ′, 𝑦 ′) with entries in C 𝜔 (𝑈 ′ × 𝑈 ′), defined by the formula

𝐹 (𝑥 ′) − 𝐹 (𝑦 ′) = 𝐽 (𝑥 ′, 𝑦 ′) (𝑥 ′ − 𝑦 ′) ,

is nonsingular [since 𝐽 (𝑥 ′, 𝑥 ′) = 𝜕𝑥
𝜕𝐹 ′
′ (𝑥 ) ]. This allows us to duplicate verba-
tim the argument in Subsection 16.1.4, based on the change of variables 𝜉 ↦→
−1
𝐽 (𝑥 ′, 𝑦 ′) ⊤ 𝜉 in the integral
17.1 Analytic Pseudodifferential Operators 631
∫
′ ′
(2𝜋) 𝑛 𝐴ℭ (𝐹 (𝑥 ′) , 𝐹 (𝑦 ′)) = e𝑖 (𝐹 ( 𝑥 )−𝐹 ( 𝑦 )) · 𝜉 𝑎 (𝐹 (𝑥 ′) , 𝐹 (𝑦 ′) , 𝜉) d𝜉.
ℭ

We conclude that 𝐹∗−1 (Op𝑎) (𝐹 ∗ ) −1 = Op𝑎 𝐹 with 𝑎 𝐹 the amplitude (16.1.24). It is

′ ′

𝜓a (𝑈1 × 𝑈2 × ℭ) implies 𝑎 𝐹 ∈ 𝑆 𝜓a 𝑈1 × 𝑈2 × ℭ .
readily seen that 𝑎 ∈ 𝑆 𝑚 □
𝑚

17.1.8 Analytic pseudodifferential operators in an analytic manifold M

Proposition 17.1.26 enables us to define analytic pseudodifferential operators on

a C 𝜔 manifold M. The role of the cone ℭ ⊂ R𝑛 \ {0} is played by a conic fiber
subbundle F of 𝑇 ∗ M\0 with typical fiber ℭ and structure group GL (𝑛, R) (see
Section 9.2; 𝑛 = dim M). In practical terms this means that, given arbitrarily local
C 𝜔 coordinates 𝑥1 , ..., 𝑥 𝑛 in a sufficiently small domain U in M, there is a C 𝜔 map
U ∋ 𝑥 ↦→ 𝒈 (𝑥) ∈ GL (𝑛, R) such that ∀𝑥 ∈ U, 𝒈 (𝑥) F 𝑥 = ℭ in the coordinates
𝜉1 , ..., 𝜉 𝑛 with respect to the basis d𝑥 1 , ..., d𝑥 𝑛 of the cotangent space at 𝑥. We can
then define Ψa𝑚 (U, F) by transfer of Ψa𝑚 (𝑈, ℭ) from the domain Ω = 𝑥 (U) of R𝑛 .
By Proposition 17.1.26 a change of these coordinates induces an automorphism of
Ψa𝑚 (U, F) (𝑚 ∈ R or 𝑚 = −𝜔; in the latter case all the statements are trivial since
Øcorresponding distribution kernels are analytic functions). We write Ψa (U, F) =
the
Ψa𝑚 (U, F). Given another domain of local C 𝜔 coordinates V ⊂ U there are
𝑚∈R
Ψa (U, F) −→
natural restriction maps Ψa (V, F) and Ψa (U, F) −→ Ψa (V, F),
𝑚 𝑚

U U
whence a presheaf Ψa (U, F) , 𝜌 V and its subpresheaves Ψa𝑚 (U, F) , 𝜌 V . If
U ∩ V ≠ ∅ two operators 𝑨∈ Ψa (U, F) and 𝑩 ∈ Ψa (V, F) equal in U ∩ V
define an element of Ψa (U ∪ V); in other words, a coherent system of locally
defined
operators
can be patched together to define a global element. The presheaf
U
Ψa (U) , 𝜌 V gives rise to the sheaf of germs of analytic pseudodifferential
operators in M associated to the conic fiber bundle F, which we shall denote by
𝚿a,F , whose global pseudodifferential operators in M associated to
sections are the
U
F. The presheaf Ψa𝑚 (U, F) , 𝜌 V gives rise to the sheaf 𝚿a,F
𝑚 . For our purposes, the

important sheaves (of noncommutative rings for composition) will be the quotients
𝑚 /𝚿−𝜔 , 𝚿 /𝚿−𝜔 . Keep in mind that 𝚿−𝜔 is another notation for the sheaf of
𝚿a,F a a,F a a
germs of C 𝜔 functions in M × M.
632 17 Analytic Pseudodifferential Calculus

17.2 Symbolic Calculus

17.2.1 Pseudoanalytic symbols in Euclidean space, true and formal

As before let Ω be an open subset of R𝑛 and ℭ an open cone in R𝑛 \ {0}. We recall

(Definition 17.1.4) that a pseudoanalytic symbol of order 𝑚 ∈ R in Ω × ℭ is a
pseudoanalytic amplitude 𝑎 (𝑥, 𝜉) of order 𝑚 in Ω 𝑥 × Ω 𝑦 × ℭ independent of 𝑦. We
𝜓a (Ω × ℭ) the vector space of pseudoanalytic symbols of order 𝑚 in
denote by 𝑆 𝑚
Ω × ℭ and define
Ø
𝑆 𝜓a (Ω × ℭ) = 𝑆𝑚𝜓a (Ω × ℭ) , (17.2.1)
𝑚∈R
Ù
𝑆 −∞
𝜓a (Ω × ℭ) = 𝑆𝑚
𝜓a (Ω × ℭ) . (17.2.2)
𝑚∈R

We shall also use the notation (cf. Definition 17.1.10)

𝑆 −𝜔 −𝜔
𝜓a (Ω × ℭ) = 𝑆 𝜓a (Ω × ℭ) ∩ 𝑆 𝜓a (Ω × Ω × ℭ) . (17.2.3)

There shall be no mention of ℭ when ℭ = R𝑛 \ {0}; a pseudoanalytic symbol 𝑎 of

order 𝑚 in Ω is a standard symbol of order 𝑚 in the C ∞ sense (Definition 16.2.1,
Corollary 17.1.7).
A symbol 𝑎 ∈ 𝑆 𝑚𝜓a (Ω × ℭ) of order 𝑚 in Ω × ℭ can be asymptotically extended
∞
as a C function of (𝑧, 𝜉) in ΩC × ℭ with ΩC an open subset of C𝑛 , Ω ⊂ ΩC ∩ R𝑛 ,
holomorphic with respect to 𝑧 and having the following property:
• To each compact set 𝐾 C ⊂ ΩC there are three positive numbers 𝜌◦ , 𝐶◦ , 𝐶, such
that, for all 𝛾 ∈ Z+𝑛 , all 𝑧 ∈ 𝐾 C and all 𝜉 ∈ ℭ,
|𝛾 |
|𝜉 | ≥ 𝜌◦ max (1, |𝛾|) =⇒ 𝜕 𝜉 𝑎 (𝑧, 𝜉) ≤ 𝐶𝐶◦ 𝛾! |𝜉 | 𝑚−|𝛾 | .
𝛾
(17.2.4)

The Cauchy inequalities imply directly, for some 𝐶1 > 0, all 𝛽, 𝛾 ∈ Z+𝑛 and all
(𝑧, 𝜉) ∈ 𝐾 C × ℭ,
|𝛽+𝛾 |
𝛽!𝛾! |𝜉 | 𝑚−|𝛾 | .
𝛽 𝛾
|𝜉 | ≥ 𝜌◦ max (1, |𝛾|) =⇒ 𝜕𝑧 𝜕 𝜉 𝑎 (𝑧, 𝜉) ≤ 𝐶𝐶1 (17.2.5)

There is a natural notion of convergence of a sequence 𝑎 𝑘 (𝑘 = 1, 2, ...) to 𝑎

𝜓a (Ω × ℭ): for every 𝑘 the functions 𝑎 𝑘 − 𝑎 must satisfy (•) with 𝜌◦ , 𝐶◦ ,
in 𝑆 𝑚
independent of 𝑘 and 𝐶 ↘ 0 as 𝑘 ↗ +∞.
The analogue of Proposition 17.1.8 is valid:
′C ⊂⊂ ΩC [see above] be open.
Proposition 17.2.1 Let 𝑎 ∈ 𝑆 𝑚 𝜓a (Ω × ℭ) and let Ω
There are constants 𝐶◦ > 0, 𝐶 𝑗 > 0, 𝑗 = 1, 2, such that the following is true. Given
any sufficiently large positive number 𝑅, 𝑎 can be asymptotically extended as a C ∞
𝑎 𝑅 (𝑧, 𝜁) in Ω′C × (ℭ + 𝑖R𝑛 ), holomorphic with respect to 𝑧 and satisfying
function e
the following estimates, for all (𝑧, 𝜁) ∈ Ω′C × (ℭ + 𝑖R𝑛 ),
17.2 Symbolic Calculus 633

𝑎 𝑅 (𝑧, 𝜁)| ≤ 𝐶◦ (1 + |Re 𝜁 |) 𝑚 exp (𝐶1 |Im 𝜁 | /𝑅) ;

|e (17.2.6)
𝑎 𝑅 (𝑧, 𝜁) ≤ 𝐶2 (1 + |Re 𝜁 |) 𝑚 exp ((𝐶1 |Im 𝜁 | − |Re 𝜁 |) /𝑅) .
𝜕𝜁 e

Remark 17.2.2 The inequalities (17.2.4), (17.2.5), (17.2.6) extend in 𝜉-space to the
closure of ℭ in R𝑛 \ {0}, ℭ (cf. Remark 17.1.9).

Í pseudoanalytic symbol of order 𝑚 in Ω × ℭ we

Definition 17.2.3 By a formal
shall mean a formal series ∞
𝑗=0 𝑎 𝑗 (𝑥, 𝜉) submitted to the following condition:

(FA) There is an open subset ΩC of C𝑛 such that Ω ⊂ ΩC ∩ R𝑛 and to which

every 𝑎 𝑗 can be asymptotically extended as a C ∞ function of (𝑧, 𝜉) in ΩC × ℭ,
holomorphic with respect to 𝑧 and having the following property:
To each compact set 𝐾 C ⊂ ΩC there are two positive numbers 𝜌◦ , 𝐶◦ , 𝐶, such
that, for all 𝑗 ∈ Z+ , 𝛾 ∈ Z+𝑛 and (𝑧, 𝜉) ∈ 𝐾 C × ℭ,
𝑗+|𝛾 |
𝑗!𝛾! |𝜉 | 𝑚− 𝑗− |𝛾 | .
𝛾
|𝜉 | ≥ 𝜌◦ max (1, 𝑗 + |𝛾|) =⇒ 𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) ≤ 𝐶𝐶◦
(17.2.7)

In particular, 𝑎 𝑗 is a pseudoanalytic symbol of order 𝑚 − 𝑗 in Ω × ℭ. The Cauchy

inequalities imply directly, for an increased 𝐶◦ > 0 and all 𝑗 ∈ Z+ , 𝛼, 𝛾 ∈ Z+𝑛 ,
(𝑧, 𝜉) ∈ 𝐾 C × ℭ,
𝑗+| 𝛼+𝛾 |
𝑗!𝛼!𝛾! |𝜉 | 𝑚− 𝑗−|𝛾 | .
𝛾
|𝜉 | ≥ 𝜌◦ max (1, 𝑗 + |𝛾|) =⇒ 𝜕𝑧𝛼 𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) ≤ 𝐶𝐶◦
(17.2.8)
We shall denote by 𝑆 𝑚
𝜓a,form (Ω × ℭ) the linear space of formal pseudoanalytic
Ø
symbols of order 𝑚 and define 𝑆 𝜓a,form (Ω × ℭ) = 𝑆𝑚𝜓a,form (Ω × ℭ). A formal
𝑚∈R
pseudoanalytic symbol 𝑎 of order 𝑚 in Ω is a formal symbol of order 𝑚 in the C ∞
sense (Definition 16.2.5).
The following definition simplifies some of the forthcoming statements.

Definition 17.2.4 By Í a formal exponentially decaying symbol in Ω × ℭ we shall

mean a formal series ∞ 𝑗=0 𝑟 𝑗 (𝑥, 𝜉) submitted to the following condition:
There is an open subset ΩC of C𝑛 , Ω ⊂ ΩC ∩ R𝑛 , to which every 𝑟 𝑗 can be
asymptotically extended as a C ∞ function of (𝑧, 𝜉) in ΩC × ℭ, holomorphic with
respect to 𝑧 and having the following property:
To each compact set 𝐾 C ⊂ ΩC there are positive numbers 𝐶◦ , 𝜌◦ , 𝑅 such that, for
all 𝑗 ∈ Z+ , 𝛾 ∈ Z+𝑛 and (𝑧, 𝜉) ∈ 𝐾 C × ℭ,
𝛾 𝑗+|𝛾 |+1
|𝜉 | ≥ 𝜌◦ max (1, 𝑗 + |𝛾|) =⇒ 𝜕 𝜉 𝑟 𝑗 (𝑧, 𝜉) ≤ 𝐶◦ 𝑗! exp (− |𝜉 | /𝑅) . (17.2.9)

We shall denote by 𝑆 −𝜔𝜓a,form (Ω × ℭ) the set of formal exponentially decaying

symbols in Ω × ℭ. The argument in the proof of Proposition 17.1.11 serves also to
prove that 𝑆 −𝜔 −∞
𝜓a,form (Ω × ℭ) ⊂ 𝑆 𝜓a,form (Ω × ℭ).
634 17 Analytic Pseudodifferential Calculus

We
Í now show how to produce a true pseudoanalytic symbol out of a formal one.
Let ∞ 𝑎
𝑗=0 𝑗 (𝑥, 𝜉) ∈ 𝑆 𝜓a,form (Ω × ℭ) satisfy Condition (FA) in Definition 17.2.3.
𝑚

We introduce a sequence 𝜑 𝑗 ( 𝑗 = 1, 2, ...) of Ehrenpreis’ cutoffs relative to the

open sets {𝜉 ∈ R𝑛 ; |𝜉 | ≥ 2}, {𝜉 ∈ R𝑛 ; |𝜉 | ≥ 1} (Definition 3.2.2); in particular,
𝜑 𝑗 (𝜉) = 0 if |𝜉 | ≤ 1 and 𝜑 𝑗 (𝜉) = 1 if |𝜉 | ≥ 2. Let the number 𝜌 > 0 be arbitrary
and 𝜉 ∈ ℭ; the series
∞
∑︁
𝑎 0 (𝑧, 𝜉) + 𝜑 𝑗 (𝜉/ 𝑗 𝜌) 𝑎 𝑗 (𝑧, 𝜉) (17.2.10)
𝑗=1

converges to a function 𝑎 (𝑧, 𝜉) ∈ C ∞ (Ω × ℭ) holomorphic with respect to

𝑧. This is simply a consequence of the fact that the series is finite in a set
ΩC × {𝜉 ∈ ℭ; |𝜉 | ≤ 𝑘 𝜌}, 𝑘 = 1, 2, ...: in this set 𝜑 𝑗 (𝜉/ 𝑗 𝜌) = 0 for all 𝑗 ≥ 𝑘.

Proposition 17.2.5 Let ∞

Í
𝑗=0 𝑎 𝑗 (𝑥, 𝜉) ∈ 𝑆 𝜓a,form (Ω × ℭ) satisfy Condition (FA) in
𝑚

′
Definition 17.2.3 and let the open set Ω ⊂⊂ Ω be arbitrary. Let the compact set 𝐾 C
in (FA) be such that Ω′ ⊂ 𝐾 C ∩ R𝑛 and let 𝜌◦ be the number in (17.2.7).
I. There is a number 𝜌 > 𝜌◦ such that the series (17.2.10) converges in
′ ′
𝑆𝑚𝜓a (Ω × ℭ) to a pseudoanalytic symbol in Ω .
II. Let 𝜓 𝑗 ( 𝑗 = 1, 2, ...) be another sequence of Ehrenpreis’ cutoffs relative to
{𝜉 ∈ R𝑛 ; |𝜉 | ≥ 2}, {𝜉 ∈ R𝑛 ; |𝜉 | ≥ 1}, and let the number 𝜌1 > 𝜌◦ be sufficiently
large for
∞
∑︁
𝑎♭ (𝑥, 𝜉) = 𝑎 0 (𝑥, 𝜉) + 𝜓 𝑗 (𝜉/ 𝑗 𝜌1 ) 𝑎 𝑗 (𝑥, 𝜉)
𝑗=1
′
𝜓a (Ω × ℭ). Provided min (𝜌, 𝜌1 ) is sufficiently
to converge in 𝑆 𝑚 large there is an
open set Ω′C ⊂ C𝑛 , Ω′ ⊂ Ω′C , such that the following is true:
(1) 𝑎 (𝑥, 𝜉) and 𝑎♭ (𝑥, 𝜉) extend as C ∞ functions 𝑎 (𝑧, 𝜉), 𝑎♭ (𝑧, 𝜉) in Ω′C × ℭ,
holomorphic with respect to 𝑧;
(2) there are positive constants 𝐶, 𝜌, 𝑅, such that, for all 𝛾 ∈ Z+𝑛 and (𝑧, 𝜉) ∈ 𝐾 C ×ℭ,

|𝜉 | ≥ 𝜌 max (1, |𝛾|) =⇒ 𝜕 𝜉 𝑎 (𝑧, 𝜉) − 𝑎♭ (𝑧, 𝜉) ≤ 𝐶 |𝛾 |+1 exp (− |𝜉 | /𝑅) .
𝛾

(17.2.11)
Proof Let 𝜑 𝑗 ( 𝑗 = 1, 2, ...) be a sequence of Ehrenpreis’ cutoffs relative to

{𝜉 ∈ R𝑛 ; |𝜉 | ≥ 2} , {𝜉 ∈ R𝑛 ; |𝜉 | ≥ 1} .

We apply (3.2.2): there is a constant 𝐶1 > 0 independent of 𝑗 and of 𝜌 > 0 such that

𝜕 𝜉𝛼 𝜑 𝑗 (𝜉/ 𝑗 𝜌) = ( 𝑗 𝜌) − | 𝛼 | 𝜕 𝜉𝛼 𝜑 𝑗 (𝜉/ 𝑗 𝜌) ≤ (𝐶1 /𝜌) | 𝛼 |

(17.2.12)

for all 𝜉 ∈ R𝑛 , 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑗. Keep in mind that 𝜑 𝑗 (𝜉/ 𝑗 𝜌) = 0 if |𝜉 | ≤ 𝑗 𝜌 and

𝜑 𝑗 (𝜉/ 𝑗 𝜌) = 1 if |𝜉 | ≥ 2 𝑗 𝜌.
17.2 Symbolic Calculus 635

Let ΩC and 𝐾 C be as in (FA), Definition 17.2.3, 𝐾 C ⊃ Ω′. We have, for all 𝑧 ∈ 𝐾 C ,

𝜉 ∈ ℭ and 𝛾 ∈ Z+𝑛 ,
∞
1 𝛾 𝛾 ∑︁ ∑︁
𝜉 𝜕 𝜉 (𝑎 (𝑧, 𝜉) − 𝑎 0 (𝑧, 𝜉)) = 𝑏 𝑗,𝛾,𝛽 (𝑧, 𝜉)
𝛾! 𝑗=1 𝛽 ⪯𝛾

where
1 𝛾−𝛽 𝛽
𝑏 𝑗,𝛾,𝛽 (𝑧, 𝜉) = 𝜉 𝛾 𝜕𝜉 𝜑 𝑗 (𝜉/ 𝑗 𝜌) 𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) .
(𝛾 − 𝛽)!𝛽!
In accordance with Definition 17.1.4 we must show that ∞
Í Í
𝑗=1 𝛽 ⪯𝛾 𝑏 𝑗,𝛾,𝛽 converges
absolutely uniformly in 𝐾 × ℭ to a function that satisfies the correct estimates; these
C

allow us to assume |𝜉 | ≥ 𝜌 max (1, |𝛾|). We relate this last property to (17.2.7): if
we select 𝜌 ≥ 2𝜌◦ and if 𝑗 ≥ |𝛾| then |𝜉 | ≥ 𝜌 𝑗 entails |𝜉 | ≥ 𝜌◦ ( 𝑗 + |𝛾|) and thereby
𝑗+|𝛽 |+1
𝑗!𝛽! |𝜉 | 𝑚− 𝑗− |𝛽 | .
𝛽
∀𝜉 ∈ ℭ, ∀𝛽 ⪯ 𝛾, 𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) ≤ 𝐶◦

Combining this with (17.2.12) yields

𝑗! 𝑗+|𝛽 |+1 𝛾−𝛽
|𝜉 | −𝑚 𝑏 𝑗,𝛾,𝛽 (𝑧, 𝜉) ≤ 𝜑 𝑗 (𝜉/ 𝑗 𝜌) |𝜉 | |𝛾−𝛽 |− 𝑗 . (17.2.13)

𝐶 𝜕𝜉
(𝛾 − 𝛽)! ◦

First we get, under the assumption that |𝜉 | ≥ 𝜌 max (1, |𝛾|),

|𝛾 |+1
𝑏 𝑗,𝛾,𝛾 (𝑧, 𝜉) ≤ 𝐶◦ (𝜅/𝑒) 𝑗 |𝜉 | 𝑚 . (17.2.14)
𝛾−𝛽
Now assume 𝛽 ≺ 𝛾, entailing 𝜕 𝜉 𝜑 𝑗 (𝜉/ 𝑗 𝜌) = 0 unless 1 < |𝜉 | /𝜌 𝑗 < 2 and
therefore
𝑗!
𝑗− |𝛾−𝛽 |
|𝜉 | −𝑚 𝑏 𝑗,𝛾,𝛽 (𝑧, 𝜉) ≤ 𝐶◦ (𝐶1 /𝜌) |𝛾−𝛽 | |𝜉 | |𝛾−𝛽 |− 𝑗
(𝛾 − 𝛽)!
𝑗 |𝛾−𝛽 |
≤ (𝐶◦ /𝜌) 𝑗 (2𝐶1 /𝐶◦ ) |𝛾−𝛽 | 𝑗! 𝑗 − 𝑗
(𝛾 − 𝛽)!
𝑗
≤ 𝐶◦ e2𝐶1 /𝐶◦ /𝜌 𝑗! 𝑗 − 𝑗 .

We√ select 𝜌 such that 𝐶◦ e2𝐶1 /𝐶◦ ≤ 𝜅𝜌, 0 < 𝜅 < 1. Stirling’s Formula implies
𝑗! ≲ 𝑗 ( 𝑗/e) 𝑗 , whence

|𝜉 | −𝑚 𝑏 𝑗,𝛾,𝛽 (𝑧, 𝜉) ≲ 𝑗 𝜅 𝑗 .
√︁

Combining this with (17.2.14) leads to the conclusion that |𝜉 | ≥ 𝜌 max (1, |𝛾|)
implies
∞ ∑︁
∑︁
|𝛾 |+1
𝑏 𝑗,𝛾,𝛽 (𝑧, 𝜉) ≤ 𝐶◦ |𝜉 | 𝑚 ,
𝑗=1 𝛽 ⪯𝛾
636 17 Analytic Pseudodifferential Calculus

whence [cf. (17.2.4)]

|𝛾 |+1
𝛾! |𝜉 | 𝑚−|𝛾 | .
𝛾
𝜕 𝜉 (𝑎 (𝑧, 𝜉) − 𝑎 0 (𝑧, 𝜉)) ≤ 𝐶◦

Since 𝑎 0 (𝑧, 𝜉) satisfies similar estimates this proves Part I of the statement.
To prove Part II suppose 𝜌1 ≥ 𝜌 > max (1, 2𝜌◦ ); if 𝑗 ≥ 1 we obtain, then,

𝜑 𝑗 (𝜉/ 𝑗 𝜌) ≠ 𝜓 𝑗 (𝜉/ 𝑗 𝜌1 ) =⇒ 𝑗 𝜌 < |𝜉 | < 2 𝑗 𝜌1 .

This allows us to duplicate the reasoning in the part 𝛽 ≺ 𝛾 of the proof of Part I
and prove that (17.2.11) holds for suitably large 𝐶, 𝜌, 𝑅. We leave the details as an
exercise. □

17.2.2 From amplitudes to symbols

As we have seen in the C ∞ setup each amplitude 𝑎 in Ω × Ω canonically defines a

formal symbol [see (16.2.6)].
Proposition 17.2.6 If 𝑎 (𝑥, 𝑦, 𝜉) is a pseudoanalytic amplitude in Ω × Ω × ℭ then
∞
∑︁

exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥
= 𝑎 𝑗 (𝑥, 𝜉) , (17.2.15)
𝑗=0
∑︁ 1
𝛽 𝛽
with 𝑎 𝑗 (𝑥, 𝜉) = 𝜕𝑦 D 𝜉 𝑎 (𝑥, 𝑥, 𝜉) ,
𝛽!
|𝛽 |= 𝑗

is a formal pseudoanalytic symbol in Ω × ℭ.

Proof A straightforward check that the properties of 𝑎 posited in Definition 17.1.4
imply the properties required of ∞
Í
𝑗=0 𝑎 𝑗 (𝑥, 𝜉) in Definition 17.2.3. □
Remark 17.2.7 If 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝜓a (Ω1 × Ω2 × ℭ) we can define

exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 ∈ 𝑆 𝜓a,form ((Ω1 ∩ Ω2 ) × ℭ) .

Proposition 17.2.8 If 𝑟 (𝑥, 𝑦, 𝜉) ∈ 𝑆 −𝜔

𝜓a (Ω × Ω × ℭ) then

∈ 𝑆 −𝜔

exp D 𝑦 · 𝜕 𝜉 𝑟 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 𝜓a,form (Ω × ℭ)

(Definition 17.2.4).
Proof Let 𝑟, 𝑟˜ and 𝐾 C be as in Definition 17.1.10 where Ω1 = Ω2 = Ω. By (17.1.12)
we have, possibly after some contraction of 𝐾 C ⊂ ΩC × ΩC about 𝐾 C ∩ R2𝑛 , for
some 𝐶 > 0 and all 𝛼, 𝛾 ∈ Z+𝑛 , all (𝑧, 𝑤) ∈ 𝐾 C and all 𝜉 ∈ R𝑛 , |𝜉 | ≥ 𝜌◦ max (1, |𝛼|),

1 𝛼 𝛼
𝜕 𝜕 𝑟˜ (𝑧, 𝑤, 𝜉) ≤ 𝐶 2| 𝛼 |+1 exp (− |𝜉 | /𝑅) . □
𝛼! 𝑤 𝜉
17.2 Symbolic Calculus 637

Let 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚
𝜓a (Ω × Ω × ℭ); we can follow the prescription (17.2.10) to
construct a “true” pseudoanalytic
symbol in Ω defined by the formal pseudoanalytic
symbol exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 :

∞
∑︁ ∑︁ 1
𝑎♭ (𝑥, 𝜉) = 𝑎 (𝑥, 𝑥, 𝜉) + 𝜑 𝑗 (𝜉/ 𝑗 𝜌) 𝜕𝑦𝛼 D 𝛼𝜉 𝑎 (𝑥, 𝑥, 𝜉) . (17.2.16)
𝛼!
𝑗=1 | 𝛼 |= 𝑗

Proposition 17.2.5 implies that if Ω′ ⊂⊂ Ω is open and if the number 𝜌 is sufficiently

large then the restriction of 𝑎♭ (𝑥, 𝜉) to Ω′ × ℭ belongs to 𝑆 𝑚 ′
𝜓a (Ω × ℭ).
Before going further we show how exponentially decaying amplitudes produce
exponentially decaying symbols.

Proposition 17.2.9 Let Ω′ ⊂⊂ Ω be open. If 𝑟 (𝑥, 𝑦, 𝜉) ∈ 𝑆 −𝜔 𝜓a (Ω × Ω × ℭ) and

𝑟 (𝑥, 𝜉) is defined by (17.2.16) with 𝑎 = 𝑟, and if the number 𝜌 is sufficiently large,
♭

then 𝑟 ♭ (𝑥, 𝜉) ∈ 𝑆 −𝜔 ′
𝜓a (Ω × ℭ).

Proof Let 𝑟˜ be as in Definition 17.1.10 where Ω1 = Ω2 = Ω and ΩC1 = ΩC2 = ΩC ; let

𝐾 C be a compact subset of ΩC such that Ω′ ⊂ 𝐾 C ∩ R𝑛 and select an open set 𝑈 C ⊂⊂
ΩC containing 𝐾 C ; let 𝑑 = dist 𝐾 C , 𝜕𝑈 C . There are numbers 𝐶◦ > 0, 𝜌◦ ≥ 1 and
𝑅 > 0 such that, for all 𝛼, 𝛽 ∈ Z+𝑛 , 𝑧 ∈ 𝐾 C and 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌◦ max (1, |𝛼 + 𝛽|),
| 𝛼+𝛽 |+1
𝑟˜ (𝑧, 𝑧, 𝜉) ≤ 𝑑 − | 𝛼 | 𝐶◦
𝛼+𝛽
𝜕𝑤𝛼 𝜕 𝜉 𝛼! exp (− |𝜉 | /𝑅) . (17.2.17)

First of all, we note that 𝑟 ♭ admits an extension 𝑟 ♭ (𝑧, 𝜉) ∈ C ∞ ΩC × ℭ holomorphic

with respect to 𝑧,
∞ ∑︁
∑︁ 1
𝑟 ♭ (𝑧, 𝜉) = 𝑟˜ (𝑧, 𝑧, 𝜉) + 𝜑 𝑗 (𝜉/ 𝑗 𝜌) 𝜕𝑦𝛼 D 𝛼𝜉 𝑟˜ (𝑧, 𝑧, 𝜉) , (17.2.18)
𝛼!
𝑗=1 | 𝛼 |= 𝑗

where we take 𝜌 ≥ 𝜌◦ . We recall that 0 ≤ 𝜑 𝑗 ≤ 1 and 𝜑 𝑗 (𝜉/ 𝑗 𝜌) ≠ 0 =⇒ |𝜉 | > 𝑗 𝜌

( 𝑗 ≥ 1; cf. Proposition 3.2.1). From (17.2.17) we deduce, for all 𝑧 ∈ 𝐾 C and 𝜉 ∈ ℭ,
∞ ∑︁
∑︁
𝑟 ♭ (𝑧, 𝜉) ≤ 𝐶◦ e− | 𝜉 |/𝑅 + 𝐶◦ 𝑑 − 𝑗 𝜑 𝑗 (𝜉/ 𝑗 𝜌) e− | 𝜉 |/𝑅
𝑗+1

𝑗=1 | 𝛼 |= 𝑗
∞
∑︁ (𝑛 + 𝑗)! 𝑗
≤ 𝐶◦ 𝐶◦ 𝑑 −1 e−𝜌/2𝑅 e− | 𝜉 |/2𝑅 .
𝑗=0
𝑛! 𝑗!

We require 𝜌 to be large enough that 𝐶◦ 𝑑 −1 e−𝜌/2𝑅 < 41 , whence

𝑟 ♭ (𝑧, 𝜉) ≤ 2𝑛+1 𝐶◦ e− | 𝜉 |/2𝑅 .

Now suppose 𝛾 ≠ 0; we shall require |𝜉 | ≥ 2𝜌 |𝛾|. By the Leibniz rule we derive

from (17.2.17)–(17.2.18):
638 17 Analytic Pseudodifferential Calculus

𝛾 𝛾
𝜕 𝜉 𝑟 ♭ (𝑧, 𝜉) ≤ 𝜕 𝜉 𝑟˜ (𝑧, 𝑧, 𝜉)
∞ ∑︁
∑︁
𝑗+|𝛾 |+1 − | 𝜉 |/𝑅
+ 𝜑 𝑗 (𝜉/ 𝑗 𝜌) 𝑑 − | 𝛼 | 𝐶◦ e
𝑗=1 | 𝛼 |= 𝑗
∞
∑︁ 𝛾 ∑︁ (𝑛 + 𝑗)! 𝛽 𝑗+|𝛾−𝛽 |+1 −| 𝜉 |/𝑅
𝜕 𝜉 𝜑 𝑗 (𝜉/ 𝑗 𝜌) 𝑑 −| 𝛼 | 𝐶◦

+ e .
0≠𝛽 ⪯𝛾
𝛽 𝑗=1
𝑛! 𝑗!

𝛽
We know that |𝜉 | > 2 𝑗 𝜌 =⇒ 𝜑 𝑗 (𝜉/ 𝑗 𝜌) = 1 hence 𝜕 𝜉 𝜑 𝑗 (𝜉/ 𝑗 𝜌) = 0 when 𝛽 ≠ 0.
The nonzero terms in the second sum on the right correspond to 2 𝑗 𝜌 < |𝜉 | ≤ 2𝜌 |𝛾|,
hence to 𝑗 ≤ |𝛾|. If |𝛽| ≤ 𝑗 (17.2.12) implies

max 𝜕 𝜉 𝜑 𝑗 (𝜉/ 𝑗 𝜌) ≤ (𝐶1 /𝜌) |𝛽 | .

𝛽

By requiring 𝜌 to be suitably large and reasoning as in the case 𝛾 = 0 we obtain

easily
|𝛾 |+1
𝜕 𝜉 𝑟 ♭ (𝑧, 𝜉) ≤ 2𝑛+1 𝐶◦ e−| 𝜉 |/2𝑅 .
𝛾
□

The analytic analogue of Proposition 16.2.9, is valid:

Proposition 17.2.10 Given 𝑎 (𝑥, 𝑦, 𝜉) ∈ 𝑆 𝑚 ′

𝜓a (Ω × Ω × ℭ), let Ω ⊂⊂ Ω be open
and let us denote by Op𝑎| Ω′ the restriction of Op𝑎 as a linear operator E ′ (Ω′) −→
D ′ (Ω′). If 𝑎♭ (𝑥, 𝜉) ∈ 𝑆 𝑚 ′
𝜓a (Ω × ℭ) is given by (17.2.16) then Op𝑎| Ω′ − 𝑎 (𝑥, D) ∈
♭
−𝜔
Ψa (Ω).

Proof By Proposition 17.1.19 the claim is equivalent to the claim that the distribu-
tion kernel associated with Op𝑎| Ω′ − 𝑎♭ (𝑥, D) belongs to C 𝜔 (Ω′ × Ω′); the latter
property is purely local. It suffices therefore to prove it in the neighborhood of an
arbitrary point of Ω′ or, equivalently, under the hypothesis that Ω′ is convex. Using
the same notation as in (16.2.5) we write Op𝑎 = Op𝑎 𝑁 + Op𝑏 𝑁 , where 𝑁 can be
any positive integer and where (for 𝑥, 𝑦 ∈ Ω′ × ℭ)
1 𝛼 𝛼
∑︁
𝑎 𝑁 (𝑥, 𝜉) = 𝜕 D 𝑎 (𝑥, 𝑥, 𝜉) ,
𝛼! 𝑦 𝜉
|𝛼|≤𝑁
∑︁ 𝑁 + 1 ∫ 1
𝑏 𝑁 (𝑥, 𝑦, 𝜉) = 𝜕𝑦𝛼 D 𝛼𝜉 𝑎 (𝑥, (1 − 𝑡)𝑥 + 𝑡𝑦, 𝜉) (1 − 𝑡) 𝑁 d𝑡.
𝛼! 0
| 𝛼 |=𝑁 +1

From this and from (17.2.16) we derive that Op𝑎| Ω′ − 𝑎♭ (𝑥, D) = Op𝑟 with
17.2 Symbolic Calculus 639

𝑁
∑︁ ∑︁ 1 𝛼 𝛼
𝑟 (𝑥, 𝑦, 𝜉) = 1 − 𝜑 𝑗 (𝜉/ 𝑗 𝑅) 𝜕 D 𝑎 (𝑥, 𝑥, 𝜉)
𝛼! 𝑦 𝜉
𝑗=1 | 𝛼 |= 𝑗

+ (1 − 𝜑 𝑁 (𝜉/𝑁 𝑅)) 𝑏 𝑁 (𝑥, 𝑦, 𝜉)

∑︁ ∑︁ 1
− 𝜑 𝑗 (𝜉/ 𝑗 𝑅) 𝜕𝑦𝛼 D 𝛼𝜉 𝑎 (𝑥, 𝑥, 𝜉) + 𝜑 𝑁 (𝜉/𝑁 𝑅) 𝑏 𝑁 (𝑥, 𝑦, 𝜉) .
𝛼!
𝑗>𝑁 | 𝛼 |= 𝑗

Let ΩC be an open subset of C𝑛 such that Ω ⊂ ΩC ∩ R𝑛 such that 𝑎 can be

asymptotically extended as a C ∞ function of (𝑧, 𝑤, 𝜉) in ΩC × ΩC × ℭ, holomorphic
with respect to (𝑧, 𝑤) and having the following property: If 𝐾 C is a compact subset
of ΩC then there are three positive numbers 𝜌◦ , 𝐶◦ , 𝐶, such that, for all 𝛾 ∈ Z+𝑛 ,
(𝑧, 𝑤) ∈ 𝐾 C × 𝐾 C , 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌◦ max (1, |𝛾|), we have
|𝛾 |+1
𝛾! |𝜉 | 𝑚−|𝛾 | .
𝛾
𝜕 𝜉 𝑎 (𝑧, 𝑤, 𝜉) ≤ 𝐶◦ (17.2.19)

We select 𝐾 C convex and such that Ω′ ⊂ 𝐾 C ∩ R𝑛 ; we also select a convex open

set 𝑈 C ⊂⊂ ΩC containing 𝐾 C . The function in 𝑈 C × 𝑈 C × ℭ,
𝑁
∑︁ ∑︁ 1 𝛼 𝛼
𝑓 𝑁 (𝑧, 𝑤, 𝜉) = 1 − 𝜑 𝑗 (𝜉/ 𝑗 𝑅) 𝜕 D 𝑎 (𝑧, 𝑧, 𝜉)
𝛼! 𝑤 𝜉
𝑗=1 | 𝛼 |= 𝑗

+ (1 − 𝜑 𝑁 (𝜉/(𝑁 + 1) 𝑅)) 𝑏 𝑁 (𝑧, 𝑤, 𝜉)

vanishes if |𝜉 | > 2𝑅𝑁. In dealing with the following function in 𝑈 C × 𝑈 C × ℭ,

∞
∑︁ ∑︁ 1
𝑔 𝑁 (𝑧, 𝑤, 𝜉) = 𝜑 𝑗 (𝜉/ 𝑗 𝑅) 𝜕𝑤𝛼 D 𝛼𝜉 𝑎 (𝑧, 𝑤, 𝜉) ,
𝑗=𝑁 +1
𝛼!
| 𝛼 |= 𝑗

we avail ourselves of the fact that 𝜑 𝑗 (𝜉/ 𝑗 𝑅) ≠ 0 =⇒ |𝜉 | ≥ 𝑗 𝑅. Thanks to (17.2.19)

and to Stirling’s formula we derive, for suitably large positive constants 𝐶1 , 𝐶2 ,
independent of 𝑁, for all (𝑧, 𝑤, 𝜉) ∈ 𝐾 C × 𝐾 C × ℭ and all 𝑗 > 𝑁,
∑︁ 1
𝑗+1
𝜑 𝑗 (𝜉/ 𝑗 𝑅) 𝜕 𝛼 𝜕 𝛼 𝑎 (𝑧, 𝑤, 𝜉) ≤ 𝐶1 𝑗! |𝜉 | 𝑚− 𝑗
𝛼! 𝑤 𝜉
| 𝛼 |= 𝑗

𝑗𝑗
≤ 𝐶1 𝐶2 |𝜉 | 𝑚−𝑁 ≤ 𝐶1 (𝐶2 /𝑅) 𝑗 (𝑅 𝑗) 𝑁 |𝜉 | 𝑚−𝑁
|𝜉 | 𝑗−𝑁
≤ 𝐶1 𝑅 𝑁 (𝐶2 e/𝑅) 𝑗 𝑁! |𝜉 | 𝑚−𝑁 .

It suffices to select 𝑅 > 𝐶2 e to obtain

|𝑔 𝑁 (𝑧, 𝑧, 𝜉)| ≤ 𝐶1 (𝐶2 e) 𝑁 𝑁! |𝜉 | 𝑚−𝑁 .

What we have just proved shows that for some 𝐶3 > 0 independent of 𝑁 and all
(𝑧, 𝑤) ∈ 𝐾 C × 𝐾 C ,
640 17 Analytic Pseudodifferential Calculus

|𝜑 𝑁 (𝜉/𝑁 𝑅) 𝑏 𝑁 (𝑧, 𝑤, 𝜉)|

∫ 1
∑︁ 𝑁 +1
≤ 𝜕𝑤𝛼 D 𝛼𝜉 𝑎 (𝑧, (1 − 𝑡)𝑧 + 𝑡𝑤, 𝜉) (1 − 𝑡) 𝑁 d𝑡 ≤ 𝐶3𝑁 +1 𝑁! |𝜉 | 𝑚−𝑁 .
𝛼! 0
| 𝛼 |=𝑁 +1

+
We conclude that |𝜉 | > 2𝑅𝑁 =⇒ |𝑟 (𝑧, 𝑤, 𝜉)| ≤ 𝐶4𝑁 +1 𝑁! |𝜉 | 𝑚 −𝑁 , where 𝑚 + =
max (𝑚, 0) and 𝐶4 = 3 max (𝐶1 , 𝐶2 e, 𝐶3 ). If 2𝑅𝑁 < |𝜉 | ≤ 2𝑅 (𝑁 + 1) then, by
Stirling’s formula and provided 𝑅 is sufficiently large,
+
|𝑟 (𝑧, 𝑤, 𝜉)| ≤ 𝐶4𝑁 𝑁! |𝜉 | 𝑚 −𝑁
√
𝑁
𝐶4 +
≤ 𝐶5 𝑁 |𝜉 | 𝑚 e−𝑁
2𝑅
++ 1
≤ 𝐶6 |𝜉 | 𝑚 2 e− | 𝜉 |/2𝑅 ≤ 𝐶7 e− | 𝜉 |/3𝑅 ,

the constants 𝐶 𝑗 ( 𝑗 = 5, 6, 7) being independent

∫ of 𝑁 and of 𝜉. The first part of Propo-
sition 17.1.19 allows us to conclude that ℭ e𝑖 ( 𝑥−𝑦) · 𝜉 𝑟 (𝑥, 𝑦, 𝜉) d𝜉 ∈ C 𝜔 (Ω × Ω). □
Consider an operator 𝑨ℭ ∈ Ψa (Ω, ℭ) [cf. (17.1.13)]. Every (𝑥 ◦ , 𝑦 ◦ ) ∈ Ω × Ω
is contained in a product 𝑈1 × 𝑈2 , 𝑈 𝑗 ⊂ Ω open, such that the restriction
of 𝑨ℭ as a linear operator E ′ (𝑈2 ) −→ D ′ (𝑈1 ) is equal to Op𝑎 for some
𝑎 ∈ 𝑆 𝜓a (𝑈1 × 𝑈2 ×ℭ). If 𝑥 ◦ ≠ 𝑦 ◦ we can select the sets 𝑈 𝑗 such that 𝑈1 ∩ 𝑈2 = ∅,
implying exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 = 0 for all 𝑥 ∈ 𝑈1 : in this case the formal sym-
bol in 𝑈1 associated with the amplitude 𝑎 vanishes identically. [This suggests that
the symbol associated to 𝑨ℭ through Formula (17.2.16) is, in reality, associated to
the germ of the distribution kernel 𝐴ℭ (𝑥, 𝑦) of 𝑨ℭ at the diagonal of Ω × Ω.] If
𝑈 ⊂⊂ Ω is open the restriction of 𝑨ℭ as a linear operator E ′ (𝑈) −→ D ′ (𝑈) is
equal to Op𝑎 for some 𝑎 ∈ 𝑆 𝜓a (𝑈 × 𝑈×ℭ). In this case the formal pseudoanalytic
symbol in 𝑈, exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 [cf. (17.2.15)], does not vanish identically.
The procedure provided by Formula (17.2.16) produces a class of symbols in 𝑈 that
are equivalent mod 𝑆 −𝜔𝜓a (𝑈×ℭ). In this manner we have associated to 𝑨ℭ a coset
in 𝑆 𝜓a (𝑈×ℭ) /𝑆 −𝜔
𝜓a (𝑈×ℭ). This leads naturally to the sheaf of germs of pseudo-
analytic symbols at points of R𝑛 × ℭ, S 𝜓a,ℭ , and its subsheaf, the sheaf S−𝜔 𝜓a,ℭ
of
−𝜔
germs at points (𝑥, 𝜉) ∈ R × ℭ of symbols 𝑎 (𝑥, 𝜉) ∈ 𝑆 𝜓a (𝑈×ℭ) [see (17.2.3)],
𝑛

and thence to the quotient sheaf S 𝜓a,ℭ /S−𝜔𝜓a,ℭ

(all stalks are complex vector spaces).
The counterpart of S 𝜓a,ℭ /S−𝜔𝜓a,ℭ
on the operators side is the sheaf (defined at the end
of Subsection 17.1.6) 𝚿a,ℭ /𝚿−𝜔 a of germs (at points of R𝑛 ) of analytic pseudodif-
ferential operators associated to the cone ℭ modulo analytic regularizing operators.
Combining Propositions 17.1.19, 17.2.5 and 17.2.10 allows us to state

Theorem 17.2.11 The map 𝑎 (𝑥, 𝜉) ↦→ 𝑎 (𝑥, D 𝑥 ) induces an isomorphism of

S 𝜓a,ℭ /S−𝜔
𝜓a,ℭ
onto 𝚿a,ℭ /𝚿−𝜔
a in the sense of sheaves of vector spaces.

We could also introduce the sheaf of germs of formal pseudoanalytic symbols.

This, however, is better done within the framework of classical symbols (see Sub-
section 17.2.7 below).
17.2 Symbolic Calculus 641

Let now Ω and Ω′ be two open subsets of R𝑛 and suppose there is a C 𝜔 dif-
feomorphism 𝐹 : Ω′ −→ Ω. By Proposition 17.1.26 we know that 𝐹 induces an
isomorphism of Ψa𝑚 (Ω × ℭ) onto Ψa𝑚 (Ω′ × ℭ) (𝑚 ∈ R or 𝑚 = −∞ or 𝑚 = −𝜔). Re-
garding the effect of 𝐹 on symbols, we know by Formula (16.1.24) that if 𝑎 ∈ 𝑆 𝑚𝜓a (Ω)
then 𝑎 𝐹 is an amplitude but, in general, not a symbol; it gives rise to a formal sym-
bol and thence to an equivalence class mod 𝑆 −𝜔 ′
𝜓a (Ω × ℭ) of symbols belonging to
𝑆𝑚 (Ω ′ × ℭ).
𝜓a

Proposition 17.2.12 For each 𝑚 ∈ R the C 𝜔 diffeomorphism 𝐹 : Ω′ −→ Ω induces

−𝜔 ′ −𝜔 ′
𝜓a (Ω × ℭ) /𝑆 𝜓a (Ω × ℭ) onto 𝑆 𝜓a (Ω × ℭ) /𝑆 𝜓a (Ω × ℭ).
a linear bijection of 𝑆 𝑚 𝑚

As a consequence of Proposition 17.2.12 we see that 𝐹 induces an isomorphism of

the sheaf S 𝜓a,ℭ /S−𝜔
𝜓a,ℭ
over Ω onto its analogue over Ω′. This enables us to associate
a sheaf S 𝜓a,F /S−𝜔
𝜓a,F to a fiber bundle F on a C manifold M with typical fiber ℭ
𝜔

and structure group GL (𝑛, R), in the same manner that we have defined the sheaf
𝚿a,F /𝚿−𝜔
a in Subsection 17.1.8. A consequence of Theorem 17.2.11 is a natural
sheaf isomorphism S 𝜓a,F /S−𝜔 −𝜔 over M.
𝜓a,F 𝚿a,F /𝚿a

17.2.3 Analytic pseudodifferential operators are analytic pseudolocal

Using symbols makes it rather simple to strengthen Corollary 17.1.16 and thus extend
the nontrivial part of Theorem 3.1.4, namely the fact that analytic parametrices of
analytic differential operators decrease the analytic singular support. As before Ω is
an open subset of R𝑛 .
Theorem 17.2.13 If 𝑨ℭ ∈ Ψa (Ω, ℭ) then

singsuppa 𝑨ℭ 𝑢 ⊂ singsuppa 𝑢

whatever 𝑢 ∈ E ′ (Ω).
Proof Since the property we want to establish is local we may as well assume
𝑨ℭ = Op𝑎 with 𝑎 ∈ 𝑆 𝑚 𝜓a (Ω × ℭ). Thanks to Lemma 17.1.13 there is no loss of
generality in assuming 𝑚 ≤ −𝑛 − 1. By availing ourselves of Proposition 17.2.10
we can take 𝑨ℭ = 𝑎 (𝑥, D), 𝑎 (𝑥, 𝜉) ∈ 𝑆 𝑚 𝜓a (Ω × ℭ). In this case the associated
distribution kernel is 𝑨ℭ (𝑥, 𝑦) = 𝑎ˆ (𝑥, 𝑦 − 𝑥) where
∫
−𝑛
𝑎ˆ (𝑥, 𝑦) = (2𝜋) e−𝑖𝑦· 𝜉 𝑎 (𝑥, 𝜉) d𝜉.
ℭ

Since 𝑎 (𝑥, 𝜉) is an 𝐿 1 function in 𝜉-space ℭ, ∥ 𝑎ˆ (𝑥, 𝑦) ∥ 𝐿 ∞ ( R𝑛𝑦 ) is uniformly bounded

as 𝑥 stays in an arbitrary compact subset 𝐾 of Ω. Moreover, by the defining properties
of pseudoanalytic symbols (Definition 17.1.4) there is a 𝐶𝐾 > 0 such that, for all
𝛾 ∈ Z+𝑛 ,
𝛾 |𝛾 |+1
max D 𝑥 𝑎ˆ (𝑥, 𝑦) 𝐿 ∞ ( R𝑛 ) ≤ 𝐶𝐾 𝛾!. (17.2.20)
𝑥 ∈𝐾 𝑦
642 17 Analytic Pseudodifferential Calculus

Let 𝑈 be a relatively compact open subset of Ω such that 𝑢 is a C 𝜔 function in an

open set containing the closure of 𝑈. We propose to show that 𝑎 (𝑥, D) 𝑢 is analytic
in an arbitrarily given open set 𝑉 ⊂⊂ 𝑈 hence in 𝑈. Let 𝜑 𝑁 (𝑁 = 1, 2, ...) be an
Ehrenpreis sequence relative to the pair (𝑈, 𝑉) (Definition 3.2.2). We have

𝑎 (𝑥, D) 𝑢 = 𝑎 (𝑥, D) (𝜑 𝑁 𝑢) + 𝑎 (𝑥, D) ((1 − 𝜑 𝑁 ) 𝑢) .

Since (1 − 𝜑 𝑁 ) 𝑢 ≡ 0 in 𝑉 Corollary 17.1.16 implies that 𝑎 (𝑥, D) ((1 − 𝜑 𝑁 ) 𝑢) is

analytic in 𝑉. It remains to show that the same is true of 𝑎 (𝑥, D) (𝜑 𝑁 𝑢).
For 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑁, we have

(2𝜋) 𝑛 D 𝑥𝛼 𝑎 (𝑥, D) (𝜑 𝑁 𝑢)
∫
∑︁ 𝛼
(−1) |𝛽 |
𝛽 𝛼−𝛽
= (𝜑 𝑁 𝑢) (𝑦) D 𝑦 D 𝑥 𝑎ˆ (𝑥, 𝑦 − 𝑥) d𝑦
𝛽⪯𝛼
𝛽 R𝑛
∫
∑︁ 𝛼! 𝛼−𝛽 𝛽
= D 𝑥 𝑎ˆ (𝑥, 𝑦 − 𝑥) D 𝑦 (𝜑 𝑁 𝑢) (𝑦) d𝑦
𝛽⪯𝛼
𝛽! (𝛼 − 𝛽)! 𝑈
∫
∑︁ ∑︁ 𝛼! 𝛼−𝛽
= D 𝑥 𝑎ˆ (𝑥, 𝑦 − 𝑥) (D𝛾 𝜑 𝑁 (𝑦)) D𝛽−𝛾 𝑢 (𝑦) d𝑦.
𝛽 ⪯ 𝛼 𝛾 ⪯𝛽
𝛾! (𝛼 − 𝛽)! (𝛽 − 𝛾)! 𝑈

We restrict the variation of 𝑥 to the closure of 𝑉, which we call 𝐾; we can apply

(17.2.20). We also take advantage of the analyticity of 𝑢 in 𝑈 and of the crucial
property of Ehrenpreis’ sequences, Property (3), Proposition 3.2.1: there is a 𝐶◦ > 0,
independent of 𝑁, such that

∀𝛾 ∈ Z+𝑛 , 𝛾 ⪯ 𝛼, max |D𝛾 𝜑 𝑁 | ≤ (𝐶◦ 𝑁) |𝛾 | .

Finally there is a 𝐶1 > 0, independent of 𝛽 − 𝛾, such that

|𝛽−𝛾 |+1
sup D𝛽−𝛾 𝑢 (𝑦) ≤ 𝐶1 (𝛽 − 𝛾)!.
𝑦 ∈𝑈

Putting all this together we obtain, in 𝑉,

1
(2𝜋) 𝑛 D 𝛼 𝑎 (𝑥, D) (𝜑 𝑁 𝑢)
𝛼! 𝑥
∑︁ ∑︁ 1
| 𝛼−𝛽 |+1 |𝛽−𝛾 |+1
≤ |𝑈| (𝐶◦ 𝑁) |𝛾 | 𝐶𝐾 𝐶1 ,
𝛽 ⪯ 𝛼 𝛾 ⪯𝛽
𝛾!

where |𝑈| is the Lebesgue measure of 𝑈. If 𝐶2 ≥ max (𝐶◦ , 𝐶𝐾 , 𝐶1 ) we have

|𝛾 | | 𝛼−𝛽 | |𝛽−𝛾 |
𝐶◦ 𝐶𝐾 𝐶1 ≤ 𝐶 | 𝛼 | , whence, for all 𝑥 ∈ 𝑉,

1
D 𝛼 𝑎 (𝑥, D) (𝜑 𝑁 𝑢) ≤ 4𝑛 | 𝛼 | 𝐶2| 𝛼 |+2 |𝑈| e 𝑁 .
𝛼! 𝑥
17.2 Symbolic Calculus 643

The proof is completed by taking |𝛼| = 𝑁: we have shown that there is a 𝐶 > 0,
depending on the symbol 𝑎, on (𝑈, 𝑉) and 𝐶◦ but not on 𝛼 ∈ Z+𝑛 , such that

max D 𝑥𝛼 𝑎 (𝑥, D) (𝜑 𝑁 𝑢) ≤ 𝐶 | 𝛼 |+1 𝛼!. □

𝐾

The microlocal version of Theorem 17.2.13 (which implies the local version) will
be proved in greater generality in the next section.

17.2.4 Symbolic calculus

The symbolic calculus in the C ∞ category as outlined in Section 16.2 extends

(or, rather, restricts) with no significant modification to pseudoanalytic symbols,
be they true or formal. We now restrict the composition law # [see (16.2.16)] to
𝑆 𝜓a,form (Ω × ℭ) × 𝑆 𝜓a,form (Ω × ℭ). The following rough estimates will be useful.

Lemma 17.2.14 We have, for all positive integers 𝑝, 𝑞,

∑︁ 𝛼1 ! · · · 𝛼𝑞 !
𝑁0 ( 𝑝, 𝑞) = ≤ 3𝑞−1 , (17.2.21)
𝛼1 +···+𝛼𝑞 = 𝑝
𝑝!
∑︁ 𝛼1 ! · · · 𝛼𝑞 !
𝑁1 ( 𝑝, 𝑞) = ≤ 1. (17.2.22)
𝛼1 +···+𝛼𝑞 = 𝑝 𝑝!
𝛼1 ··· 𝛼𝑞 ≠0

In (17.2.21)–(17.2.22) 𝛼1 , ..., 𝛼𝑞 ∈ Z+ .
Proof We have 𝑁0 ( 𝑝, 1) = 𝑁1 ( 𝑝, 1) = 1 for all 𝑝 ≥ 1 and 𝑁0 (1, 𝑞) = 𝑞 ≤ 2𝑞−1 for
all 𝑞 ≥ 1; also, 𝑁1 (1, 𝑞) ≠ 0 implies 𝑞 = 1 and 𝑁1 (1, 𝑞) = 1. Therefore it suffices
to consider the cases 𝑝 ≥ 2, 𝑞 ≥ 2. We note that
𝑝
∑︁ 𝛼! ( 𝑝 − 𝛼)!
𝑁0 ( 𝑝, 2) = = 2 + 𝑁1 ( 𝑝, 2) .
𝛼=0
𝑝!

𝑝 −1 1
Since 1 ≤ 𝛼 < 𝑝 =⇒ 𝛼 ≤ 𝑝, we have

𝑝−1
∑︁ 𝛼! ( 𝑝 − 𝛼)! 1
𝑁1 ( 𝑝, 2) = ≤ 1− (17.2.23)
𝛼=1
𝑝! 𝑝

whence 𝑁0 ( 𝑝, 2) ≤ 3 − 𝑝1 . With the understanding that 𝑁0 (0, 𝑞) = 1 for all 𝑞 ≥ 1

we can write, for 𝑞 ≥ 3,

𝑝
∑︁ 𝛼𝑞 ! 𝑝 − 𝛼𝑞 ! © ∑︁ 𝛼1 ! · · · 𝛼𝑞−1 ! ª
𝑁0 ( 𝑝, 𝑞) = ­ ®,
𝛼𝑞 =0
𝑝! 𝛼1 +···+𝛼𝑞−1 = 𝑝−𝛼𝑞 𝑝 − 𝛼𝑞 !
« ¬
644 17 Analytic Pseudodifferential Calculus

whence

1 𝑞−2
𝑁0 ( 𝑝, 𝑞) ≤ 𝑁0 ( 𝑝, 2) max 𝑁0 (𝑚, 𝑞 − 1) ≤ 3 − 3 ,
0≤𝑚≤ 𝑝 𝑝

the last inequality ensuing from (17.2.23) and induction on 𝑞. This proves (17.2.21).
The conditions 𝛼1 + · · · + 𝛼𝑞 = 𝑝, 𝛼1 · · · 𝛼𝑞 ≠ 0 in (17.2.22) imply 𝑞 ≤ 𝑝. For
𝑝 ≥ 𝑞 ≥ 2 we can write

𝑝−1
∑︁ 𝛼𝑞 ! 𝑝 − 𝛼𝑞 ! ©­ ∑︁ 𝛼1 ! · · · 𝛼𝑞−1 ! ª®
𝑁1 ( 𝑝, 𝑞) = ­ ®
𝛼𝑞 =1
𝑝! ­ 𝛼 +···+𝛼 = 𝑝−𝛼
1 𝑞−1 𝑞
𝑝 − 𝛼𝑞 ! ®
« 𝛼1 ··· 𝛼𝑞−1 ≠0 ¬
≤ 𝑁1 ( 𝑝, 2) max 𝑁1 (𝑚, 𝑞 − 1) .
𝑞−1≤𝑚≤ 𝑝

Induction on 𝑞 combined with (17.2.23) proves (17.2.22). □

Proposition 17.2.15 If 𝑎 𝑖 ∈ 𝑆 𝑚𝑖
𝜓a,form
(Ω × ℭ), 𝑖 = 1, 2, then

1 +𝑚2
𝑎 1 #𝑎 2 ∈ 𝑆 𝑚
𝜓a,form
(Ω × ℭ) .

Proof Let
∞
∑︁
𝑎𝑖 = 𝑎 𝑖, 𝑗 (𝑥, 𝜉) , 𝑖 = 1, 2,
𝑗=0

be two formal pseudoanalytic symbols in Ω × ℭ that extend as C ∞ functions 𝑎 𝑖 (𝑧, 𝜉)

to ΩC × R𝑛 , holomorphic with respect to 𝑧 (with ΩC ⊂ C𝑛 open, Ω ⊂ ΩC ) having
the following property: given an arbitrary compact set 𝐾 C ⊂ ΩC , there are two
positive numbers 𝐶◦ , 𝜌◦ such that, for all 𝑗 ∈ Z+ , all 𝛾 ∈ Z+𝑛 , all (𝑧, 𝜉) ∈ 𝐾 C × ℭ,
|𝜉 | ≥ 𝜌◦ max (1, 𝑗 + |𝛾|),
𝑗+|𝛾 |+1
𝑗!𝛾! |𝜉 | 𝑚𝑖 − 𝑗− |𝛾 | .
𝛾
𝜕 𝜉 𝑎 𝑖, 𝑗 (𝑧, 𝜉) ≤ 𝐶◦

Actually we shall exploit this inequality after replacing 𝐾 C by a larger compact

subset of ΩC , n o
𝐾1C = 𝑥 ∈ ΩC ; dist 𝑧, 𝐾 C ≤ 𝑑

with 𝑑 < dist 𝐾 C , 𝜕ΩC (and suitably increased 𝐶◦ , 𝜌◦ ). This allows us to apply the
Cauchy inequalities:

max 𝜕𝑧𝛼 𝜕 𝜉 𝑎 𝑖, 𝑗 (𝑧, 𝜉) ≤ 𝛼!𝑑 − | 𝛼 | max 𝜕 𝜉 𝑎 𝑖, 𝑗 (𝑧, 𝜉)

𝛾 𝛾
𝑧 ∈𝐾 C 𝑧 ∈𝐾1C
𝑗+|𝛾 |+1
≤ 𝐶◦ 𝑑 − | 𝛼 | 𝑗!𝛼!𝛾! |𝜉 | 𝑚𝑖 − 𝑗− |𝛾 | .
Í∞
We can rewrite (16.2.16) as 𝑎 1 #𝑎 2 = ℓ=0 𝑏 ℓ , where
17.2 Symbolic Calculus 645
∑︁ 1 𝛼
𝑏ℓ = D 𝜉 𝑎 1, 𝑗 𝜕𝑥𝛼 𝑎 2,𝑘 ∈ 𝑆 𝑚1 +𝑚2 −ℓ (Ω × ℭ) .
𝛼!
𝑗+𝑘+| 𝛼 |=ℓ

By the Leibniz rule we obtain, for all 𝛾 ∈ Z+𝑛 and (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | ≥

𝜌◦ max (1, ℓ + |𝛾|),
𝛾
𝜕 𝜉 𝑏 ℓ (𝑧, 𝜉)
∑︁ ∑︁ 1 𝛾! 𝛼+𝛽 𝛾−𝛽
≤ 𝜕 𝜉 𝑎 1, 𝑗 (𝑧, 𝜉) 𝜕𝑧𝛼 𝜕 𝜉 𝑎 2,𝑘 (𝑧, 𝜉)
𝛽 ⪯𝛾
𝛼! 𝛽! (𝛾 − 𝛽)!
𝑗+𝑘+| 𝛼 |=ℓ
∑︁ ∑︁
𝑗+𝑘+ | 𝛼+𝛾 |+2 (𝛼 + 𝛽)! 𝑚1 +𝑚2 − 𝑗−𝑘− | 𝛼+𝛾 |
≤ 𝛾! 𝑑 − | 𝛼 | 𝐶◦ 𝑗!𝑘! |𝜉 |
𝛽!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽 ⪯𝛾

ℓ+ |𝛾 |+2
∑︁ ∑︁ (𝛼 + 𝛽)! 𝑚1 +𝑚2 −ℓ−|𝛾 |
≤ 𝑑 −ℓ 𝐶◦ 𝛾!ℓ! 𝑗!𝑘! |𝜉 | .
ℓ!𝛽!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽 ⪯𝛾

Applying (17.2.21) we obtain

∑︁ ∑︁ (𝛼 + 𝛽)! ∑︁ 𝑗!𝑘!𝛼! ∑︁ (𝛼 + 𝛽)!
𝑗!𝑘! =
ℓ!𝛽! ℓ! 𝛽 ⪯𝛾 𝛼!𝛽!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽 ⪯𝛾 𝑗+𝑘+| 𝛼 |=ℓ
∑︁ 𝑗!𝑘!𝛼! | 𝛼+𝛾 |
≤ 2 ≤ 3𝑛+1 2ℓ+|𝛾 | ,
ℓ!
𝑗+𝑘+| 𝛼 |=ℓ

whence, for all ℓ ∈ Z+ , all 𝛾 ∈ Z+𝑛 , (𝑧, 𝜉) ∈ 𝐾 C × R𝑛 , 𝜉 ∈ ℭ such that |𝜉 | ≥

𝜌◦ max (1, 𝑗 + |𝛾|),

𝜕 𝜉 𝑏 ℓ (𝑧, 𝜉) ≤ 3𝑛+1 𝐶◦2 𝑑 −ℓ (2𝐶◦ ) ℓ+|𝛾 | 𝛾!ℓ! |𝜉 | 𝑚1 +𝑚2 −ℓ− |𝛾 | ,

𝛾

1 +𝑚2
Í∞
which proves that ℓ=0 𝑏 ℓ (𝑧, 𝜉) ∈ 𝑆 𝑚
𝜓a,form
(Ω × ℭ). □

The case ℭ = R𝑛 \ {0} is worth mentioning:

Corollary 17.2.16 The set of formal pseudoanalytic symbols in Ω, 𝑆 𝜓a,form (Ω), is a

subalgebra of 𝑆form (Ω) [with respect to addition and the composition law #].

Proposition 17.2.17 If
∞
∑︁ ∞
∑︁
𝑎= 𝑎 𝑗 ∈ 𝑆 𝜓a,form (Ω × ℭ) and 𝑟 = 𝑟 𝑘 ∈ 𝑆 −𝜔
𝜓a,form (Ω × ℭ)
𝑗=0 𝑘=0

(Definitions 17.1.10, 17.2.4) then 𝑎#𝑟 and 𝑟#𝑎 belong to 𝑆 −𝜔

𝜓a,form (Ω × ℭ).

In other words, 𝑆 −𝜔
𝜓a,form (Ω × ℭ) is a two-sided ideal in the ring 𝑆 𝜓a,form (Ω × ℭ)
with respect to the composition law #.
646 17 Analytic Pseudodifferential Calculus

Proof We suppose that 𝑎 𝑗 and 𝑟 𝑘 extend as C ∞ functions of (𝑧, 𝜉) in ΩC × R𝑛 ,

holomorphic with respect to 𝑧, and that the following holds:
⊛ To each compact set 𝐾 C ⊂ ΩC there are positive numbers 𝐶◦ , 𝜌◦ and 𝑅 such
that, for all 𝛽 ∈ Z+𝑛 , all 𝑧 ∈ 𝐾 C and all 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌◦ max (1, 𝑗 + |𝛽|),
𝑗+|𝛽 |+1
𝑗!𝛽! |𝜉 | 𝑚− 𝑗− |𝛽 | ,
𝛽
𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) ≤ 𝐶◦
𝛽 𝑗+|𝛽 |+1
𝜕 𝜉 𝑟 𝑗 (𝑧, 𝜉) ≤ 𝐶◦ exp (− |𝜉 | /𝑅) .

Consider an open set 𝑈 C ⊂⊂ ΩC containing 𝐾 C . Taking 𝑑 = dist 𝐾 C , 𝜕𝑈 C we
derive from the Cauchy inequalities
𝑘+|𝛽 |+1
𝜕𝑧𝛼 𝜕 𝜉 𝑟 𝑘 (𝑧, 𝜉) ≤ 𝑑 − | 𝛼 | 𝐶◦
𝛽
𝛼! exp (− |𝜉 | /𝑅)

∈ 𝐾 C ∩ R𝑛 ⊂ Ω and all 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌◦ max (1, 𝑘 + |𝛽|).

for all 𝛽 ∈ Z+𝑛 , all 𝑧 Í
∞
We write 𝑎#𝑟 = ℓ=0 𝑏 ℓ where
∑︁ 1 𝛼 𝛼
𝑏ℓ = D 𝜉 𝑎 𝑗 𝜕𝑧 𝑟 𝑘 .
𝛼!
𝑗+𝑘+| 𝛼 |=ℓ

The estimate of 𝑏 0 is simple enough; we assume ℓ ≥ 1. By the Leibniz rule we

obtain, for all 𝛾 ∈ Z+𝑛 and (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | ≥ 𝜌◦ (ℓ + |𝛾|),
𝛾
𝜕 𝜉 𝑏 ℓ (𝑧, 𝜉)
𝛾
∑︁ ∑︁ 1 𝛾! 𝛼+𝛽 𝛾−𝛽
≤ 𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) 𝜕𝑧𝛼 𝜕 𝜉 𝑟 𝑘 (𝑧, 𝜉)
𝛼! 𝛽! (𝛾 − 𝛽)!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽=0
𝛾
∑︁ ∑︁
𝑗+𝑘+| 𝛼+𝛾 |+2 (𝛼 + 𝛽)!𝛾! 𝑚−ℓ−|𝛽 | −| 𝜉 |/𝑅
≤ 𝑑 −| 𝛼 | 𝐶◦ 𝑗! |𝜉 | e
𝛽! (𝛾 − 𝛽)!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽=0
𝛾
ℓ+|𝛾 |+2
∑︁ ∑︁ (𝛼 + 𝛽)!𝛾! 𝑗! −ℓ− |𝛽 |
≤ 𝑑 −ℓ 𝐶◦ |𝜉 | 𝑚 e−| 𝜉 |/𝑅 |𝜉 | .
𝛽! (𝛾 − 𝛽)!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽=0

Stirling’s formula implies, for a suitably large 𝐶1 > 0, depending on 𝜌◦ but not on ℓ,
𝛽, 𝜉,
ℓ+|𝛽 |+1 1
|𝜉 | −ℓ− |𝛽 | ≤ (ℓ + |𝛽|) −ℓ− |𝛽 | ≤ 𝐶1 ,
(ℓ + |𝛽|)!
whence
17.2 Symbolic Calculus 647

e | 𝜉 |/𝑅 𝜕 𝜉 𝑏 ℓ (𝑧, 𝜉)
𝛾

𝛾
ℓ+|𝛾 |+2
∑︁ ∑︁ (𝛼 + 𝛽)! 𝑗! 𝛾!
≤𝑑 −ℓ
(𝐶◦ 𝐶1 ) |𝜉 | 𝑚
|𝜉 | − |𝛽 |
(ℓ + |𝛽|)! 𝛽! (𝛾 − 𝛽)!
𝑗+𝑘+| 𝛼 |=ℓ 𝛽=0
|𝛾 |
𝜌 −1

≤3 𝑑𝑛 −ℓ
(𝐶◦ 𝐶1 ) ℓ+|𝛾 |+2
|𝜉 | 𝑚
1+ ◦ ,
ℓ + |𝛾|

the last inequality a consequence of (17.2.21). Since

|𝜉 | 𝑚 ≤ (2𝑅) 𝑚 |𝑚|! exp (− |𝜉 | /2𝑅)

we conclude that there is a 𝐶2 > 0 such that, for all ℓ ∈ Z+ , 𝛾 ∈ Z+𝑛 , all 𝑧 ∈ 𝐾 C and
all 𝜉 ∈ ℭ, |𝜉 | ≥ 𝜌◦ max (1, |𝛾|),
𝛾 ℓ+ |𝛾 |+1
𝜕 𝜉 𝑏 ℓ (𝑧, 𝜉) ≤ 𝐶2 exp (− |𝜉 | /2𝑅) ,

which is what had to be proved. A similar argument proves that 𝑟#𝑎 ∈ 𝑆 −𝜔

𝜓a,form (Ω).□

In the next statement we identify 𝑆 𝜓a (Ω × ℭ) with a subset of 𝑆 𝜓a,form (Ω × ℭ),

Í
namely the set of finite series 𝑎 = 𝑗 𝑎 𝑗 (𝑥, 𝜉).

Corollary 17.2.18 If 𝑎 ∈ 𝑆 𝜓a (Ω × ℭ) and 𝑟 ∈ 𝑆 −𝜔

𝜓a (Ω × ℭ) then 𝑎#𝑟 and 𝑟#𝑎
belong to 𝑆 −𝜔
𝜓a,form (Ω × ℭ).

Proposition 17.2.9 assigns to each element of 𝑆 −𝜔 𝜓a,form (Ω × ℭ) a family of true

symbols belonging to 𝑆 −𝜔
𝜓a (Ω × ℭ) [characterized by the Ehrenpreis’ cutoffs 𝜑 𝑗 , see
(17.2.16)]; Corollary 17.2.18 associates to 𝑎#𝑟 and 𝑟#𝑎 such a family.
Thanks to Propositions 17.2.15 and 17.2.17 we can complement Theorem 17.2.11
and state the analogue of Theorem 16.2.13. Recall that 𝚿a,ℭ /𝚿−𝜔 a (as defined in
Subsection 17.1.7) is a sheaf of rings with respect to composition of operators. It is
readily checked that S 𝜓a,ℭ /S−𝜔
𝜓a,ℭ
is a sheaf of rings with respect to the composition
law #.

Theorem 17.2.19 The map 𝑎 (𝑥, 𝜉) ↦→ 𝑎 (𝑥, D 𝑥 ) induces a sheaf of ring isomor-
phisms S 𝜓a,ℭ /S−𝜔
𝜓a,ℭ
−→ 𝚿a,ℭ /𝚿−𝜔
a .

This statement extends directly to a C 𝜔 manifold M and to a conic fiber bundle

F with typical fiber ℭ and structure group GL (𝑛, R) (see the remarks following
Proposition 17.2.12): the natural isomorphism S 𝜓a,F /S−𝜔 −𝜔 is a sheaf
𝜓a,F 𝚿a,F /𝚿a
of ring isomorphisms.
648 17 Analytic Pseudodifferential Calculus

17.2.5 Elliptic analytic pseudodifferential operators

A pseudoanalytic amplitude 𝑎 (resp., symbol) is said to be elliptic if it is elliptic

as a standard amplitude (resp., symbol; Definition 16.2.14); the same applies to
analytic pseudodifferential operators defined by elliptic pseudoanalytic amplitudes
or symbols. Let Ω ⊂ R𝑛 be open and 𝑚 ∈ R. A formal pseudoanalytic symbol
𝑎 = ∞
Í
𝑗=0 𝑗 (𝑥, 𝜉) ∈ 𝑆 𝜓a,form (Ω) is said to be elliptic if its leading symbol 𝑎 0 ∈
𝑎 𝑚

𝜓a (Ω × ℭ) is elliptic, meaning that to each compact subset 𝐾 of Ω there are positive

𝑆𝑚
constants 𝑐 and 𝑅 such that

(𝑥, 𝜉) ∈ 𝐾 × ℭ, |𝜉 | > 𝑅, |𝑎 0 (𝑥, 𝜉)| ≥ 𝑐 (1 + |𝜉 |) 𝑚 , (17.2.24)

Under this hypothesis there exists a 𝑏 = ∞ −𝑚

Í
𝑗=0 𝑏 𝑗 (𝑥, 𝜉) ∈ 𝑆 form (Ω × ℭ) such that
𝑎#𝑏 = 𝑏#𝑎 = 1 (Proposition 16.2.16); the terms 𝑏 𝑗 are determined recursively by
the equations
∑︁ 1 𝛾 𝛾
𝑏 0 = 𝑎 −1
0 , 𝑏 𝑗 = −𝑏 0 𝜕 𝑎 𝑘 D 𝜉 𝑏 ℓ , 𝑗 = 1, 2, .... (17.2.25)
𝛾! 𝑧
𝑘+ℓ+|𝛾 |= 𝑗
ℓ< 𝑗

𝜓a,form (Ω × ℭ) is elliptic then the equations (17.2.25)

Proposition 17.2.20 If 𝑎 ∈ 𝑆 𝑚
define a formal pseudoanalytic symbol 𝑏 in Ω × ℭ, elliptic of order −𝑚.
𝑚− 𝑗
Proof We assume that the 𝑎 𝑗 ∈ 𝑆 𝜓a (Ω × ℭ) satisfy Condition (FA) in Definition
17.2.3 and that (17.2.24) holds; we must show that the same is true of the 𝑏 𝑗 after
substituting −𝑚 for 𝑚. We can select the open subset ΩC of C𝑛 in (FA) so that, to
each compact set 𝐾 C ⊂ ΩC there are positive constants 𝑐 and 𝑅 such that

∀ (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | > 𝑅, |𝑎 0 (𝑧, 𝜉)| ≥ 𝑐 (1 + |𝜉 |) 𝑚 . (17.2.26)

This, together with (17.2.7), implies immediately, for all (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | > 𝑅,

𝐶◦−1 (1 + |𝜉 |) −𝑚 ≤ |𝑏 0 (𝑧, 𝜉)| ≤ 𝑐−1 (1 + |𝜉 |) −𝑚 (17.2.27)

which, in turn, implies the ellipticity of 𝑏.

Now suppose that 0 ≠ 𝛼 ∈ Z+𝑛 ; we have, for all (𝑧, 𝜉) ∈ 𝐾 C × ℭ,
∑︁ 𝛼!
𝛽 𝛼−𝛽
𝑎 0 (𝑧, 𝜉) D 𝛼𝜉 𝑏 0 (𝑧, 𝜉) = − D 𝜉 𝑎 0 (𝑧, 𝜉) D 𝜉 𝑏 0 (𝑧, 𝜉) .
0≠𝛽 ⪯ 𝛼
(𝛼 − 𝛽)!𝛽!

If |𝜉 | > 𝜌◦ |𝛼| then, by (17.2.7) and (17.2.26),

∑︁ 𝛼! |𝛽 |+1
|𝜉 | 𝑚−|𝛽 | D 𝜉 𝑏 0 (𝑧, 𝜉) .
𝛼−𝛽
𝑐 |𝜉 | 𝑚 D 𝛼𝜉 𝑏 0 (𝑧, 𝜉) ≤ 𝐶
0≠𝛽 ⪯ 𝛼
(𝛼 − 𝛽)! ◦

We reason by induction on 𝛼: we assume that, if 0 ≠ 𝛽 ⪯ 𝛼, then

17.2 Symbolic Calculus 649

| 𝛼−𝛽 |+1
(𝛼 − 𝛽)! |𝜉 | −𝑚−| 𝛼−𝛽 |
𝛼−𝛽
D𝜉 𝑏 0 (𝑧, 𝜉) ≤ 𝐵◦

for all (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | ≥ 𝜌◦ max (1, |𝛼 − 𝛽|). By (17.2.27) this is true when
𝛼 = 𝛽 provided 𝐵◦ ≥ 𝑐−1 . If 𝐵◦ > 𝐶◦ we obtain
∑︁
|𝜉 | 𝑚+| 𝛼 | D 𝛼𝜉 𝑏 0 (𝑧, 𝜉) ≤ 𝐵◦| 𝛼 |+1 𝑐−1 𝐶◦ 𝛼! (𝐶◦ /𝐵◦ ) |𝛽 |
0≠𝛽 ⪯ 𝛼

𝐶◦𝑛+1
≤ 𝐵◦| 𝛼 |+1 𝛼!𝑐−1 .
(𝐵◦ − 𝐶◦ ) 𝑛

If (𝐵◦ − 𝐶◦ ) 𝑛 ≥ 𝑐−1 𝐶◦𝑛+1 we can conclude that D 𝛼𝜉 𝑏 0 (𝑧, 𝜉) ≤ 𝐵◦| 𝛼 |+1 𝛼! |𝜉 | −𝑚−| 𝛼 |
for all (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | ≥ 𝜌◦ |𝛼|.
The same type of argument applied to the equations
∑︁ 1 𝛾 𝛾
𝑎0 𝑏 𝑗 = − 𝜕 𝑎 𝑘 D 𝜉 𝑏 ℓ , 𝑗 = 1, 2, ...,
𝛾! 𝑧
𝑘+ℓ+|𝛾 |= 𝑗
ℓ< 𝑗

now exploiting (17.2.8), allows us to conclude that

| 𝛼 |+ 𝑗+1
D 𝛼𝜉 𝑏 𝑗 (𝑧, 𝜉) ≤ 𝐵◦ 𝛼! 𝑗! |𝜉 | −𝑚−| 𝛼 |

for all (𝑧, 𝜉) ∈ 𝐾 C × ℭ, |𝜉 | ≥ 𝜌◦ ( 𝑗 + |𝛼|). We leave the details as an exercise. □

Thus the elliptic formal pseudoanalytic symbols in Ω associated to ℭ form a group
with respect to the composition law #; the elliptic formal pseudoanalytic symbols of
order zero associated to ℭ form a subgroup of this group.

Proposition 17.2.21 If 𝑨ℭ ∈ Ψa𝑚 (Ω, ℭ) is elliptic of order 𝑚 then, to every open set
Ω′ ⊂⊂ Ω there is an analytic pseudodifferential operator 𝑩ℭ ∈ Ψa−𝑚 (Ω′, ℭ) such
that the composites 𝑨ℭ | Ω′ 𝑩ℭ and 𝑩ℭ ( 𝑨ℭ | Ω′ ) are equivalent mod Ψa−𝜔 (Ω′, ℭ) to
the identity 𝑰.

For the definition of the composites in Proposition 17.2.21 we refer to Subsection

17.1.6.
Proof Given any open set Ω′ ⊂⊂ Ω we can find a pseudoanalytic amplitude 𝑎 of
order 𝑚 in Ω′ × Ω′ × ℭ such that 𝑨ℭ [restricted to E ′ (Ω′ × ℭ)] is equivalent to
−𝜔 ′

Op𝑎 mod Ψa (Ω × ℭ). Let exp D 𝑦 · 𝜕 𝜉 𝑎 (𝑥, 𝑦, 𝜉) 𝑦=𝑥 be the formal pseudoan-
alytic symbol associated to 𝑎. By Proposition 17.2.20 this symbol has an inverse
−𝑚 (Ω) with respect to the composition law #; we can construct a “true” pseu-
𝑏 ∈ 𝑆a,form
doanalytic symbol 𝑏 ♮ in Ω′ × ℭ out of 𝑏. By Proposition 17.2.10, given any open set
Ω′′ ⊂⊂ Ω containing the closure of Ω′ there is a pseudoanalytic symbol 𝑎♭ in Ω′′
such that Op𝑎 − 𝑎♭ (𝑥, D) ∈ Ψa−𝜔 (Ω′ × ℭ). It follows directly that 𝑩ℭ | Ω′ = 𝑏 ♮ (𝑥, D)
fulfills the requirements in the statement. □
650 17 Analytic Pseudodifferential Calculus

It is obvious that we can define the ellipticity (of order 𝑚) of an element

•
𝑨ℭ = ( 𝑨ℭ | Ω′ ) Ω′ ⊂ ⊂Ω of Ψa (Ω, ℭ) [see (17.1.19)]: it simply demands that each
g
coset 𝑨ℭ | Ω′ mod Ψa−𝜔 (Ω′) be elliptic of order 𝑚 (which in turn simply means that
one, or equivalently every, representative of 𝑨ℭ | Ω′ be elliptic of order 𝑚 in Ω′).
A restatement of Proposition 17.2.21 is that every elliptic element of the algebra
•
Ψa (Ω, ℭ) is invertible.

17.2.6 Resolvent and functionals of analytic pseudodifferential

operators of order zero

We shall deal with a symbol 𝑎 ∈ 𝑆 0𝜓a (Ω × ℭ) (Definition 17.1.4) extendible as a

C ∞ function of (𝑧, 𝜉) in ΩC × ℭ with ΩC an open subset of C𝑛 containing Ω ⊂ ℭ,
holomorphic with respect to 𝑧 and having the property (•) in Subsection 17.2.1
with 𝑚 = 0. In particular, to each compact set 𝐾 C ⊂ ΩC there corresponds a
smallest simply connected compact subset of C, which we shall denote by 𝑎 𝐾 C ,
such that 𝑎 (𝑧, 𝜉) ∈ 𝑎 𝐾 C for all (𝑧, 𝜉) ∈ 𝐾 C × ℭ. For each fixed 𝜆 ∈ C\𝑎 𝐾 C ,
𝜆 − 𝑎 (𝑥, 𝜉) is an elliptic pseudoanalytic symbol of order zero in a neighborhood
Ω′ ⊂ Ω of 𝐾 = 𝐾 C ∩ R𝑛 ; the function C\𝑎 𝐾 C ∋ 𝜆 ↦→ (𝜆 − 𝑎 (𝑧, 𝜉)) −1 is well
defined and holomorphic. By Proposition 17.2.20 𝜆 − 𝑎 (𝑥, 𝜉) has an inverse in the
0
(noncommutative) ring 𝑆a,form (Ω′ × ℭ) (with respect to the composition law #). As
′
Ω expands, the domain in the 𝜆-plane in which the #-inverse of 𝜆 − 𝑎 (𝑥, 𝜉) is well
defined “shrinks” to ∞; we can give it a meaning as a germ of holomorphic function
at infinity in the Riemann sphere S2 valued in 𝑆 0𝜓a,form (Ω), which we denote by
𝑅 𝑎 (𝜆; 𝑥, 𝜉) and call the resolvent of 𝑎. This is given a more precise description in
the next proposition.
Let us denote by C [𝑎, 𝜕𝑎, ..., ] (resp., C [[𝑎, 𝜕𝑎, ..., ]]) the ring of polynomials
(resp., formal power series) with complex coefficients, in the partial derivatives of
all orders of 𝑎 with respect to (𝑥, 𝜉).
Proposition 17.2.22 For (𝑥, 𝜉) ∈ Ω′ and 𝜆 ∈ C\𝑎 𝐾 C we have

∞
∑︁
𝑅 𝑎 (𝜆; 𝑥, 𝜉) = (𝜆 − 𝑎 (𝑥, 𝜉)) −𝑚−1 𝑔𝑚 [𝜕𝑎 (𝑥, 𝜉)] , (17.2.28)
𝑚=0

where 𝑔𝑚 [𝜕𝑎] ∈ C [[𝑎, 𝜕𝑎, ..., ]] and 𝑔0 = 1. The formal series 𝑔𝑚 [𝜕𝑎] is homo-
geneous of degree 𝑚 with respect to 𝑎.
Note that the formal series 𝑔𝑚 [𝜕𝑎] can be homogeneous of degree 𝑚 with
𝛽
respect to 𝑎 without being a polynomial with respect to the 𝜕𝑥𝛼 𝜕 𝜉 𝑎. Example:
Í 1 𝛼 𝛽
𝛼,𝛽 ∈Z+𝑛 𝛼!𝛽! 𝑐 𝛼,𝛽 𝜕𝑥 𝜕 𝜉 𝑎, 𝑐 𝛼,𝛽 ∈ C.

Proof We may view 𝑅 𝑎 (𝜆; 𝑥, 𝜉) as a formal series in the powers of 𝜆−1 with coeffi-
cients in 𝑆 0𝜓a,form (Ω × ℭ) and write
17.2 Symbolic Calculus 651
∞
∑︁
𝑅 𝑎 (𝜆; 𝑥, 𝜉) = 𝑏 𝑗 (𝜆; 𝑥, 𝜉) (17.2.29)
𝑗=0

−𝑗
with 𝑏 𝑗 ∈ 𝑆 𝜓a (Ω × ℭ) determined by the equations (17.2.25) in which 𝑎 0 = 𝜆 − 𝑎,
𝑎 𝑘 ≡ 0 whatever 𝑘 ≥ 1. Thus 𝑏 0 = (𝜆 − 𝑎 (𝑥, 𝜉)) −1 . It suffices to prove that we have,
for each 𝑗 = 1, 2, ...,
2𝑗
∑︁
𝑏 𝑗 (𝜆; 𝑥, 𝜉) = (𝜆 − 𝑎 (𝑥, 𝜉)) −𝑘−1 𝑔 𝑗,𝑘 [𝜕𝑎 (𝑥, 𝜉)] , (17.2.30)
𝑘=0

−𝑗
where 𝑔 𝑗,𝑘 ∈ 𝑆 𝜓a (Ω) ∩ C [𝑎, 𝜕𝑎, ..., ]. First of all, (17.2.30) is true when 𝑗 = 0 (with
𝑔0,0 = 1). Notice that, if (17.2.30) is true then, whatever 𝛾 ∈ Z+𝑛 ,
𝑗+ |𝛾 |
2 ∑︁
(𝜆 − 𝑎 (𝑥, 𝜉)) −𝑘−1 𝑔 𝑗,𝑘;𝛾 [𝜕𝑎 (𝑥, 𝜉)] ,
𝛾
D 𝜉 𝑏 𝑗 (𝜆; 𝑥, 𝜉) =
𝑘=0

− 𝑗−|𝛾 |
where 𝑔 𝑗,𝑘;𝛾 (𝑥, 𝜉) ∈ 𝑆 𝜓a (Ω × ℭ) ∩ C [𝑎, 𝜕𝑎, ..., ]. Induction on 𝑗 and the equa-
tions (17.2.25), here
𝑗−1 ∑︁
∑︁ 1 𝛾 𝛾
𝑏 𝑗 = − (𝜆 − 𝑎) −1 𝜕 𝑎 D𝜉 𝑏𝑘, (17.2.31)
𝑘=0 |𝛾 |= 𝑗−𝑘
𝛾! 𝑥

prove (17.2.30) for all 𝑗.

For 𝑡 > 0 we have

1 = (𝜆 − 𝑡𝑎 (𝑥, 𝜉)) #𝑅𝑡 𝑎 (𝜆; 𝑥, 𝜉)

= 𝑡 𝑡 −1 𝜆 − 𝑎 (𝑥, 𝜉) #𝑅𝑡 𝑎 (𝜆; 𝑥, 𝜉)

= 𝑡 −1 𝜆 − 𝑎 (𝑥, 𝜉) #𝑅 𝑎 𝑡 −1 𝜆; 𝑥, 𝜉 ,

whence 𝑅𝑡 𝑎 (𝜆; 𝑥, 𝜉) = 𝑡 −1 𝑅 𝑎 𝑡 −1 𝜆; 𝑥, 𝜉 . By (17.2.28) we derive

∞
∑︁
𝑅𝑡 𝑎 (𝜆; 𝑥, 𝜉) = (𝜆 − 𝑡𝑎 (𝑥, 𝜉)) −𝑚−1 𝑔𝑚 [𝑡𝜕𝑎 (𝑥, 𝜉)]
𝑚=0
∞
∑︁ −𝑚−1
= 𝑡 −1 𝑡 −1 𝜆 − 𝑎 (𝑥, 𝜉) 𝑡 −𝑚 𝑔𝑚 [𝑡𝜕𝑎 (𝑥, 𝜉)]
𝑚=0
∞
∑︁ −𝑚−1
= 𝑡 −1 𝑡 −1 𝜆 − 𝑎 (𝑥, 𝜉) 𝑔𝑚 [𝜕𝑎 (𝑥, 𝜉)] ,
𝑚=0

proving that 𝑔𝑚 [𝑡𝜕𝑎] = 𝑡 𝑚 𝑔𝑚 [𝜕𝑎]. □

652 17 Analytic Pseudodifferential Calculus

Formulas (17.2.30) show that we can extend each 𝑏 𝑗 as a function 𝑏 𝑗 (𝜆; 𝑧, 𝜉)

holomorphic with respect to (𝑧, 𝜆) in the region 𝑧 ∈ ΩC , 𝜆 ∈ C\𝑎(𝑧, 𝜉), 𝜉 ∈ ℭ. It is
in this sense that we can introduce the formal series 𝑅 𝑎 (𝜆; 𝑧, 𝜉).

Remark 17.2.23 Inspection of the proof of Proposition 17.2.22 shows that no mono-
mial in 𝑔𝑚 [𝜕𝑎] ∈ C [[𝑎, 𝜕𝑎, ..., ]] has a factor 𝑎: 𝑔𝑚 [𝜕𝑎] is a formal series in the
powers of the partial derivatives of 𝑎 with respect to (𝑥, 𝜉) of order ≥ 1. Using
(17.2.31) and induction on 𝑗 shows that, in each monomial of each “coefficient” 𝑔 𝑗,𝑘
in (17.2.30), the total order of differentiation with respect to either 𝑥 or 𝜉 is equal
to 𝑗. It follows that in each monomial of each “coefficient” 𝑔𝑚 in (17.2.28) the total
order of differentiation with respect to 𝑥 is equal to that with respect to 𝜉.

Let us restrict
the variation of (𝑧, 𝜉) to 𝐾 × S ; if we select a smooth curve
C 𝑛−1

𝔠 ⊂ C\𝑎 𝐾 winding (once) around 𝑎 𝐾 then, whatever the function 𝐹 ∈ O (C),

C C

the integral ∫
1
𝐹 (𝜆) 𝑅 𝑎 (𝜆; 𝑧, 𝜉) d𝜆 (17.2.32)
2𝜋𝑖 𝔠
0
defines an element of 𝑆a,form (Ω′ × ℭ) (𝐾 ⊂ Ω′ ⊂ Ω). The Cauchy Integral Theorem
entails that (17.2.32) is independent of 𝔠. By pulling 𝔠 towards ∞ we allow Ω′ ↗ Ω;
0
in this sense we can say that (17.2.32) defines an element of 𝑆a,form (Ω × ℭ), which
#
we shall denote by 𝐹 (𝑎 (𝑥, 𝜉)). It follows from (17.2.28) that
∞
∑︁ 1 (𝑚)
𝐹 # (𝑎) = 𝐹 (𝑎) 𝑔𝑚 [𝜕𝑎] . (17.2.33)
𝑚=0
𝑚!

The fundamental property of the contour integral (17.2.32) is that

(𝜆𝐹 (𝜆)) # (𝑎) = 𝑎#𝐹 # (𝑎) , (17.2.34)

an immediate
∫ consequence of the fact that 𝜆𝑅 𝑎 (𝜆; 𝑥, 𝜉) = 𝑎 (𝑥, 𝜉) #𝑅 𝑎 (𝜆; 𝑥, 𝜉) + 1
and 𝔠 𝐹 (𝜆) d𝜆 = 0. Proposition 17.2.22 implies
∫
1
𝑅 𝑎 (𝜆; 𝑧, 𝜉) d𝜆 = 1. (17.2.35)
2𝜋𝑖 𝔠

Then (17.2.34) and induction on 𝑚 ≥ 1 implies

𝑚
∫
1
z }| {
𝑎 (𝑥, 𝜉) # · · · #𝑎 (𝑥, 𝜉) = 𝜆 𝑚 𝑅 𝑎 (𝜆; 𝑧, 𝜉) d𝜆. (17.2.36)
2𝜋𝑖 𝔠
As expected we obtain, for an arbitrary entire function 𝐹,
𝑚
∞
1 (𝑚)
∑︁ z }| {
𝐹 # (𝑎) = 𝐹 (0) 𝑎 (𝑥, 𝜉) # · · · #𝑎 (𝑥, 𝜉) (17.2.37)
𝑚=0
𝑚!
17.2 Symbolic Calculus 653

0
z }| {
with the understanding that 𝑎 (𝑥, 𝜉) # · · · #𝑎 (𝑥, 𝜉) = 1. It is natural to denote by
𝐹 (𝑎 (𝑥, D)) the coset mod Ψa−𝜔 (Ω) of analytic pseudodifferential operators corre-
sponding to the formal pseudoanalytic symbol (17.2.37) – or one of its representa-
tives.
We apply (17.2.33) with 𝐹 (𝜆) replaced by 𝐸 (𝑡, 𝜆) = e𝑡𝜆 where 𝑡 ∈ C; (17.2.33)
implies
∞ 𝑚
∑︁ 𝑡
𝐸 # (𝑡, 𝑎) = e𝑡 𝑎 𝑔𝑚 [𝜕𝑎] , (17.2.38)
𝑚=0
𝑚!
while (17.2.37) implies
𝑚
∞ 𝑚z }| {
∑︁ 𝑡
𝐸 # (𝑡, 𝑎) = 𝑎 (𝑥, 𝜉) # · · · #𝑎 (𝑥, 𝜉). (17.2.39)
𝑚=0
𝑚!

We deduce from (17.2.37) that the 𝐸 # (𝑡, 𝑎) form a one-parameter subgroup of the
group of formal elliptic pseudoanalytic symbols of order zero associated to ℭ, and
that
d𝐸 #
= 𝑎#𝐸 # , 𝐸 # (0, 𝑎) = 1. (17.2.40)
d𝑡
The map 𝑡 ↦→ 𝐸 # (𝑡, 𝑎) can be extended as a holomorphic function in C valued in
the vector space 𝑆 0𝜓a,form (Ω × ℭ) equipped with its natural locally convex topology.

Definition 17.2.24 By the exponential of the analytic pseudodifferential operator

𝑎 (𝑥, D) ∈ Ψa0 (Ω, ℭ) we shall mean the coset mod Ψa−𝜔 (Ω) (Definitions 17.1.12,
17.1.17) or, sometimes, any one of its representatives, defined by the formal pseudo-
analytic symbol 𝐸 # (1, 𝑎 (𝑥, 𝜉)). We shall denote said exponential by exp 𝑎 (𝑥, D).

17.2.7 Classical analytic pseudodifferential operators

Let ΩC ⊂ C𝑛 be open and 𝑎 (𝑧, 𝜉) ∈ C ∞ ΩC × ℭ be holomorphic with respect to

𝑧 and homogeneous of degree 𝑚 (∈ R), meaning that 𝑎 (𝑧, 𝜆𝜉) = 𝜆 𝑚 𝑎 (𝑧, 𝜉) for all
𝜆 > 0, (𝑧, 𝜉) ∈ ΩC × ℭ. As we will now see the symbol 𝑎 (𝑥, 𝜉) has true, not just
asymptotic (cf. Proposition 17.1.8), holomorphic extendibility to cones

ℭe𝜅 = {𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 ; 𝜉 ∈ ℭ, |𝜂| < 𝜅 |𝜉 |} .

Lemma 17.2.25 Suppose that 𝑎 (𝑧, 𝜉) ∈ C ∞ ΩC × ℭ is holomorphic with respect

to 𝑧 and homogeneous of degree 𝑚. Let 𝑈 C ⊂⊂ ΩC be an open set. The following
properties are equivalent:
654 17 Analytic Pseudodifferential Calculus

(a) there are positive constants 𝜌, 𝐶, such that

∀𝛾 ∈ Z+𝑛 , 𝜕 𝜉 𝑎 (𝑧, 𝜉) ≤ 𝐶 |𝛾 |+1 𝛾! |𝜉 | 𝑚−|𝛾 |

𝛾
(17.2.41)

for all 𝑧 ∈ 𝑈 C , 𝜉 ∈ ℭ, |𝜉 | > 𝜌;

(b) there is a 𝐶 > 0 such that (17.2.41) holds for all 𝑧 ∈ 𝑈 C , 𝜉 ∈ ℭ;
(c) there exists a 𝜅 > 0 such that 𝑎 can be asymptotically extended as a holomorphic
function 𝑎 (𝑧, 𝜁) in 𝑈 C × ℭ e𝜅 satisfying |𝑎 (𝑧, 𝜁)| ≤ 𝜅 −1 |Re 𝜁 | 𝑚 .
𝛾
Proof Indeed, (a)⇐⇒(b) trivially since 𝜕 𝜉 𝑎 (𝑧, 𝜉) is homogeneous of degree 𝑚−|𝛾|;
(b)=⇒(c) since (17.2.41) implies that the power series
∑︁ 1
𝛾
(𝑖𝜂) 𝛾 𝜕 𝜉 𝑎 (𝑧, 𝜉)
𝛾 ∈Z𝑛
𝛾!
+

converges uniformly absolutely in 𝑈 C × ℭ1/2𝐶 to a holomorphic function 𝑎 (𝑧, 𝜁)

satisfying the desired inequality; (c)=⇒(b) as a consequence of the homogeneity of
e𝜅 ∩ R𝑛 .
𝑎 and of the Cauchy inequalities in balls centered at points of ℭ □
The preceding proof shows that if 𝐶 is the constant in (a) we can take 𝜅 = 1/2𝐶
in (c); conversely, if (c) holds then we can take 𝐶 > 1/𝜅 in (17.2.41). This has the
following consequence:
Lemma 17.2.26 For each 𝑗 ∈ Z+ let the function 𝑎 𝑗 (𝑧, 𝜉) ∈ C ∞ ΩC × ℭ be

holomorphic with respect to 𝑧 and homogeneous of degree 𝑚 − 𝑗. Let 𝑈 C ⊂⊂ ΩC be
an open set and suppose that there is a 𝐶 > 0 such that

∀ 𝑗 ∈ Z+ , ∀𝛾 ∈ Z+𝑛 , 𝜕 𝜉 𝑎 𝑗 (𝑧, 𝜉) ≤ 𝐶 𝑗+|𝛾 |+1 𝑗!𝛾! |𝜉 | 𝑚− 𝑗− |𝛾 | ,

𝛾
(17.2.42)

whatever 𝑧 ∈ 𝑈 C , 𝜉 ∈ ℭ. Then there exist positive numbers 𝜅, 𝐶1 , such that, for each
𝑗 ∈ Z+ , 𝑎 𝑗 can be asymptotically extended as a holomorphic function 𝑎 𝑗 (𝑧, 𝜁) in
𝑈C × ℭe𝜅 such that 𝑎 𝑗 (𝑧, 𝜉) ≤ 𝐶 𝑗+1 𝑗! |𝜉 | 𝑚− 𝑗 . The converse is true: if such numbers
1
𝜅, 𝐶1 exist then so does 𝐶 > 0, satisfying (17.2.42) for all 𝑧 ∈ 𝑈 C , 𝜉 ∈ ℭ.
We can also state:
Lemma 17.2.27 Let the functions 𝑎 𝑗 ( 𝑗 ∈ Z+ ) and 𝑈 C be as in Lemma 17.2.26 and
suppose (17.2.42) is valid for all 𝑧 ∈ 𝑈 C , 𝜉 ∈ ℭ. Let 𝜒 𝑗 𝑗 ∈Z+ be a sequence of
Ehrenpreis cutoffs relative to the open sets {𝜉 ∈ R𝑛 ; |𝜉 | ≥ 2}, {𝜉 ∈ R𝑛 ; |𝜉 | ≥ 1}
(Definition 3.2.2). If 𝜌 > 0 is sufficiently large the series [cf. (17.2.10)]
∞
∑︁
𝑎♭ (𝑥, 𝜉) = 𝜒1 (𝜉) 𝑎 0 (𝑥, 𝜉) + 𝜒 𝑗 (𝜉/ 𝑗 𝜌) 𝑎 𝑗 (𝑥, 𝜉) (17.2.43)
𝑗=1

converges in C ∞ (𝑈 × ℭ) to a symbol 𝑎 ∈ 𝑆 𝑚 𝜓a (𝑈 × ℭ) (𝑈 = 𝑈 ∩ R ). Modifying

C 𝑛
−𝜔
the cutoffs 𝜒 𝑗 leads to a symbol 𝑎 + 𝑟, with 𝑟 ∈ 𝑆 𝜓a (𝑈) (Definition 17.1.10).
17.2 Symbolic Calculus 655

Proof Let 𝜀 ≪ 1; the hypotheses imply that ∞

Í
𝑗=0 𝜒1 (𝜉/𝜀) 𝑎 𝑗 (𝑥, 𝜉) ∈ 𝑆 a,form (𝑈)
𝑚

(Definition 17.2.3). We have, for all (𝑥, 𝜉) ∈ 𝑈 × ℭ,

𝜒1 (𝜉) 𝑎 0 (𝑥, 𝜉) = 𝜒1 (𝜉) 𝜒1 (𝜉/𝜀) 𝑎 0 (𝑥, 𝜉) ,

𝜒 𝑗 (𝜉/ 𝑗 𝜌) 𝑎 𝑗 (𝑥, 𝜉) = 𝜒 𝑗 (𝜉/ 𝑗 𝜌) 𝜒1 (𝜉/𝜀) 𝑎 𝑗 (𝑥, 𝜉) , 𝑗 = 1, 2, ....

It suffices then to apply Proposition 17.2.5. □

Definition 17.2.28 ByÍa classical analytic symbol of order 𝑚 ∈ R in Ω we shall
mean a formal series +∞ ∞
𝑗=0 𝑎 𝑗 (𝑥, 𝜉) with 𝑎 𝑗 ∈ C (Ω × ℭ) homogeneous of degree
𝑚 − 𝑗, with the following property: there is an open subset ΩC of C𝑛 containing
Ω and
𝜅 > 0 such that every 𝑎 𝑗 has an extension 𝑎 𝑗 (𝑧, 𝜁) ∈ O ΩC × ℭ e𝜅 homogeneous

𝑚 − 𝑗 satisfying
of degree the following condition: to each compact subset 𝐾 C of
ΩC × ℭ e𝜅 ∩ S2𝑛−1 there is a 𝐶 > 0 such that

∀ 𝑗 ∈ Z+ , max 𝑎 𝑗 ≤ 𝐶 𝑗+1 𝑗!. (17.2.44)

𝐾C

We shall denote by 𝑆a,class

𝑚 (Ω × ℭ) the linear space of classical analytic symbols
Ø
of order 𝑚 in Ω and we write 𝑆a,class (Ω × ℭ) = 𝑚
𝑆a,class (Ω × ℭ); obviously,
Ù 𝑚∈R
𝑚
𝑆a,class (Ω × ℭ) = {0}. We omit mentioning ℭ when ℭ = R𝑛 \ {0}.
𝑚∈R
Let the functions 𝑎 𝑗 (𝑧, 𝜁) be as in Definition 17.2.28. The Cauchy inequalities
and (17.2.44) lead directly to estimates of the kind

∀ 𝑗 ∈ Z+ , ∀𝛽, 𝛾 ∈ Z+𝑛 , ∀𝜆 > 0, 𝜕𝑧 𝜕𝜁 𝑎 𝑗 (𝑧, 𝜆𝜁) ≤ 𝐶 𝑗+|𝛽+𝛾 |+1 𝑗!𝛽!𝛾!𝜆 𝑚− 𝑗− |𝛾 | ,

𝛽 𝛾

(17.2.45)
valid for an increased constant 𝐶 and all (𝑧, 𝜁) ∈ 𝐾 C .
For the proof of the next statement (in a somewhat wider context) we refer the
reader to Theorem 19.1.17.
Proposition 17.2.29 If 𝑎 𝑖 = +∞ 𝑚𝑖
Í
𝑗=0 𝑎 𝑖, 𝑗 ∈ 𝑆 a,class (Ω × ℭ) (𝑖 = 1, 2) then

∑︁ 1
𝛾 𝛾
𝑎 1 #𝑎 2 = D 𝜉 𝑎 1 𝜕𝑥 𝑎 2 (17.2.46)
𝛾 ∈Z 𝑛 𝛾!
+

𝑚1 +𝑚2
is an element of 𝑆a,class (Ω × ℭ).
Thus 𝑆a,class (Ω × ℭ) is an associative (noncommutative) algebra with respect to
# and ordinary addition.
A classical analytic symbol +∞
Í
𝑗=0 𝑎 𝑗 is said to be elliptic if its principal symbol is
elliptic, i.e., 𝑎 0 (𝑧, 𝜉) ≠ 0 for all (𝑧, 𝜉) ∈ C ∞ (Ω × ℭ).
Proposition 17.2.30 If 𝑎 = +∞
Í
𝑗=0 𝑎 𝑗 ∈ 𝑆 a,class (Ω × ℭ) is elliptic there is a unique
𝑚
Í+∞ −𝑚
𝑏 = 𝑗=0 𝑏 𝑗 ∈ 𝑆a,class (Ω × ℭ) such that 𝑎#𝑏 = 𝑏#𝑎 = 1; 𝑏 is elliptic of order −𝑚.
656 17 Analytic Pseudodifferential Calculus

Proof The 𝑏 𝑗 are determined recursively by (17.2.25); then one must deduce from
this formula that the 𝑏 𝑗 satisfy the requirements in Definition 17.2.28 when the 𝑎 𝑗
do. We will not give the details at this point. The analogous result will be proved
in the “pure” complex-analytic class later (Theorem 19.1.17). To prove Proposition
17.2.30 simply combine Theorem 19.1.17 with the Cauchy inequalities and introduce
the appropriate cutoff functions. (The complete proof can also be found in [Treves,
1980], Vol. I, pp. 271–275.) □
We are going to use the following terminology.

Definition 17.2.31 We shall say that a classical analytic symbol (Definition 17.2.28)
𝑎 = +∞
Í
𝑗=0 𝑎 𝑗 ∈ 𝑆 a,class (Ω × ℭ) is the total symbol of a classical analytic pseu-
𝑚

dodifferential operator (Definition 17.2.32 below) 𝑨ℭ : E ′ (Ω) −→ D ′ (Ω) if

𝑨ℭ − 𝑎♭ (𝑥, D) ∈ Ψa−𝜔 (Ω) where 𝑎♭ (𝑥, 𝜉) is one of the “true” symbols (17.2.43).

We continue to refer to 𝑎 0 as the principal symbol of 𝑨ℭ or of 𝑎♭ (𝑥, D).

Definition 17.2.32 An analytic pseudodifferential operator 𝑨ℭ : E ′ (Ω) −→ D ′ (Ω)

[cf. (17.1.3)] will be referred to as a classical analytic pseudodifferential operator
of order 𝑚 in Ω associated to the cone ℭ if the following is true: there is an open
covering 𝑈 ( 𝜄) of Ω such that, whatever the index 𝜄, a classical analytic symbol

𝑎 ( 𝜄) ∈ 𝑆a,class
𝑚 𝑈 ( 𝜄) × ℭ is the total symbol of the restriction of 𝑨ℭ to E ′ 𝑈 ( 𝜄) .

The classical analytic pseudodifferential operators of order 𝑚 in Ω associ-

ated to ℭ form a vector space which will be denoted by Ψa,class
𝑚 (Ω, ℭ); we write
Ø
Ψa,class (Ω, ℭ) = ⊤ ∗
Ψa,class (Ω, ℭ). Let 𝑨ℭ (resp., 𝑨ℭ ) denote the transpose (resp.,
𝑚

𝑚∈R
adjoint) of 𝑨ℭ ∈ Ψa,class
𝑚 (Ω, ℭ). The classical analytic symbols associated locally
⊤ ∗
to 𝑨ℭ and 𝑨ℭ are related to those of 𝑨ℭ by the same formulas as those for formal
standard symbols, (16.2.7), (16.2.8); from these formulas it is easy to derive the
following.

Proposition 17.2.33 If 𝑨ℭ ∈ Ψa,class

𝑚 (Ω, ℭ) then 𝑨⊤
ℭ
∈ Ψa,class
𝑚 (Ω, −ℭ) and 𝑨∗ℭ ∈
Ψa,class (Ω, ℭ).
𝑚

The following assertions are self-evident:

(1) If 𝑨ℭ ∈ Ψa,class (Ω, ℭ) has a total symbol 𝑎 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) the latter is
unique.
(2) If 𝑨ℭ and 𝑩ℭ belong to Ψa,class (Ω, ℭ) and if they have the same total symbol
𝑎 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) then 𝑨ℭ − 𝑩ℭ ∈ Ψa−𝜔 (Ω).

These properties have the following consequence. Suppose that, given any open
set Ω′ ⊂⊂ Ω, the restriction of 𝑨ℭ ∈ Ψa (Ω, ℭ) to Ω′ has a total symbol 𝑎 Ω′ (𝑥, 𝜉) ∈
𝑆a,class (Ω × ℭ). Then there is an 𝑎 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) such that 𝑎 (𝑥, 𝜉) =
𝑎 Ω′ (𝑥, 𝜉) for all (𝑥, 𝜉) ∈ Ω′ × ℭ and the restriction of 𝑨ℭ − 𝑎 (𝑥, D) to Ω′ belongs
to Ψa−𝜔 (Ω′).
17.3 Analytic Microlocalization In Distribution Theory 657

To define the composition of two classical analytic pseudodifferential operators

𝑚𝑗
we can proceed as follows: if 𝑎 𝑗 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) is the total symbol of the
classical analytic pseudodifferential operator 𝑨ℭ, 𝑗 in Ω ( 𝑗 = 1, 2) we can define
𝑚1 +𝑚2
𝑨ℭ,1 𝑨ℭ,2 as an operator (𝑎 1 #𝑎 2 ) (𝑥, D) + 𝑹 ∈ Ψa,class (Ω × ℭ), with 𝑹 ∈ Ψa−𝜔 (Ω).
For a general classical analytic pseudodifferential operator 𝑨ℭ 𝑗 in Ω we can do
this locally. The ambiguity about the choice of 𝑹 ∈ Ψa−𝜔 (Ω) can be eliminated
• •
by using the associative algebra Ψa (Ω, ℭ) [see (17.1.19)]. We define Ψa,class (Ω, ℭ)
•𝑚 •
[resp., Ψa,class (Ω, ℭ)] as the subset of Ψa (Ω, ℭ) that have a representative of the
form 𝑎♭ (𝑥, D), with 𝑎 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) [resp., 𝑆a,class
𝑚 (Ω × ℭ)]. In the next
• •
statement we denote by 𝑎 (𝑥, D) the element of Ψa,class (Ω, ℭ) determined by 𝑎 (𝑥, 𝜉).
Proposition 17.2.34 The map
• •
𝑆a,class (Ω × ℭ) ∋ 𝑎 (𝑥, 𝜉) ↦→ 𝑎 (𝑥, D) ∈ Ψa,class (Ω, ℭ)

is an algebra isomorphism with respect to the law (17.2.46) on the symbol side and
composition on the “operator” side.
𝑚
𝑗
Proposition 17.2.35 If 𝑎 𝑗,0 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) is the principal symbol of 𝑨ℭ, 𝑗

( 𝑗 = 1, 2) then the principal symbol of the commutation bracket 𝑨ℭ,1 , 𝑨ℭ,2 is

equal to the Poisson bracket 𝑎 1,0 (𝑥, 𝜉) , 𝑎 2,0 (𝑥, 𝜉) and
𝑚1 +𝑚2 −1
𝑨ℭ,1 , 𝑨ℭ,2 ∈ Ψa,class (Ω × ℭ) .

0
Given 𝑎 (𝑥, 𝜉) ∈ 𝑆a,class (Ω × ℭ) and an arbitrary entire function 𝐹 in C the
formulas in the preceding subsection, (17.2.28), (17.2.32), etc., make it evident that
we can define 𝐹 # (𝑎 (𝑥, 𝜉)) as an element of 𝑆a,class
0 (Ω × ℭ); we have 𝐹 (𝑎 (𝑥, D)) ∈
0
Ψa,class (Ω, ℭ).
By combining Propositions 17.2.21 and 17.2.30 we can state:
Proposition 17.2.36 If 𝑨ℭ ∈ Ψa,class
𝑚 (Ω, ℭ) is elliptic of order 𝑚 then to every open
set Ω′ ⊂⊂ Ω there is a pseudodifferential operator 𝑩ℭ ∈ Ψa,class −𝑚 (Ω′, ℭ) such that
𝑨ℭ | Ω′ 𝑩ℭ and 𝑩ℭ ( 𝑨ℭ | Ω′ ) are equivalent mod Ψa−𝜔 (Ω′) to the identity operator.

17.3 Analytic Microlocalization In Distribution Theory

In the present section we introduce new symbols that are not properly pseudoanalytic,
and pseudodifferential operators that are not properly analytic, but can serve as cutoffs
in frequency space – when dealing with distributions. We show how they can be
used, either alone or composed with analytic pseudodifferential operators, to delimit
the analytic wave-front set (Definition 7.4.7) of distribution solutions of analytic
PDEs. These results are special cases of the analogous results for hyperfunctions
and microfunctions, established in the last two sections of this chapter.
658 17 Analytic Pseudodifferential Calculus

17.3.1 Asymptotically analytic cutoff multipliers

As announced, we describe a construction of cutoff functions in frequency space

when 𝑛 ≥ 2, of the Ehrenpreis’ type (Section 3.1). Ideally, one would like to have
at one’s disposal a smooth function 𝑔 (𝜉) homogeneous of degree zero for |𝜉 | large,
equal to 1 in an open cone Γ ⊂ R𝑛 \ {0} and to zero in the complement of a larger
cone Γ♭ . The convolution operator 𝑔 (D 𝑥 ) would be the tool enabling us to limit the
study of the wave-front set of a distribution to 𝑈 ×Γ (𝑈 an open set in position space).
Obviously dealing with cutoffs of this type, a simple matter in the C ∞ setup, would
not do if we insist on getting holomorphic extensions to complex frequency space.
But even if we content ourselves with asymptotic holomorphy it is not possible to
preserve homogeneity; we must accept a certain amount of degradation as |𝜉 | ↗ +∞.
This is achieved by relying on special partitions of unity (of Ehrenpreis’ type) in the
half-line in which |𝜉 | ranges.

Lemma 17.3.1 There is a partition of unity on the real line consisting of C ∞ func-
tions ℎ 𝜈 ≥ 0 (𝜈 ∈ Z+ ) that satisfy the following conditions:
(1) ℎ0 (𝑡) = 1 if 𝑡 < 1, ℎ0 (𝑡) = 0 if 𝑡 > 2;
1
(2) if 𝜈 ≥ 1, ℎ 𝜈 (𝑡) ≠ 0 =⇒ 4𝜈−1 < 𝑡 < 4𝜈+ 2 ;
(3) there is a 𝐶◦ > 0 such that ℓ ∈ Z+ , 2ℓ ≤ 𝑡 =⇒ ℎ 𝜈(ℓ) (𝑡) ≤ 𝐶◦ℓ .

Note that perforce ℎ 𝜈 ≤ 1; and that, by Property (2), any point 𝑡 ∈ R can belong
to supp ℎ 𝜈 for at most two values of 𝜈.
Proof Let { 𝜒 𝑁 } 𝑁 =1,2,... be a sequence of Ehrenpreis’ cutoffs in a single real variable,
relative to the open half-lines (−∞, 0) and (−∞, 1) (Definition 3.2.2). We are going
to avail ourselves of (3.2.2):
( 𝑗)
∀𝑡 ∈ R, 𝜒 𝑁 (𝑡) ≤ (𝐶◦ 𝑁) 𝑗 if 𝑗 ≤ 𝑁. (17.3.1)

Define ℎ0 (𝑡) = 𝜒1 (𝑡 − 1) and for 𝜈 ≥ 1,

1
ℎ 𝜈 (𝑡) = 1 − 𝜒4𝜈−1 41−𝜈 𝑡 − 1 if 𝑡 < 4𝜈 , (17.3.2)
2
1
ℎ 𝜈 (𝑡) = 1 if 4𝜈 < 𝑡 < 4𝜈 ,
2
ℎ 𝜈 (𝑡) = 𝜒4𝜈 (4−𝜈 𝑡 − 1) if 𝑡 > 4𝜈 .

Thus, according to (17.3.2),

ℎ 𝜈−1 (𝑡) + ℎ 𝜈 (𝑡) = 1 if 4𝜈−1 < 𝑡 < 2 × 4𝜈−1 ,

ℎ 𝜈−1 (𝑡) = 0 if 𝑡 > 2 × 4𝜈−1 ,
ℎ 𝜈 (𝑡) = 0 if 𝑡 < 4𝜈−1 ,
17.3 Analytic Microlocalization In Distribution Theory 659
Í∞
implying Property (2) in Lemma 17.3.1, and also that 𝜈=1 ℎ 𝜈 (𝑡) = 1 for every
𝑡 ∈ R. Moreover, (17.3.2) entails
1
ℎ 𝜈(ℓ) (𝑡) = 4ℓ (1−𝜈) 𝜒4(ℓ) 4 1−𝜈
𝑡 − 1 if 4𝜈−1 ≤ 𝑡 ≤ 4𝜈 ,
𝜈−1
2
ℎ 𝜈(ℓ) (𝑡) = 4−ℓ 𝜈 𝜒4(ℓ)
𝜈 (4 −𝜈
𝑡 − 1) if 4 𝜈
≤ 𝑡 ≤ 2 × 4 𝜈
,

and ℎ 𝜈(ℓ) (𝑡) = 0 for all other values of 𝑡. Suppose 2ℓ ≤ 𝑡; (17.3.1) implies ℎ 𝜈(ℓ) (𝑡) ≤
𝐶◦ℓ either when 𝑡 ≤ 21 4𝜈 or when 4𝜈 < 𝑡 ≤ 2 × 4𝜈 . □

Lemma 17.3.2 Let ℭ, ℭ♭ be two open cones in R𝑛 \ {0} such that ℭ♭ ∩ S𝑛−1 ⊂⊂
ℭ ∩ S𝑛−1 ≠ S𝑛−1 . There are positive constants 𝐶 𝑗 ( 𝑗 = 1, 2, 3) such that the following
√
is true. Given any number 𝑅 > 0 there is a C ∞ function 𝑔 𝑅 (𝜁) in C𝑛 \ −1R𝑛 having
the following properties:
(1) ∀𝜉 ∈ R𝑛 \ {0}, 0 ≤ 𝑔 𝑅 (𝜉) ≤ 1;
(2) ∀𝜉 ∈ R𝑛 \ {0}, ∀𝛼 ∈ Z+𝑛 , |𝜉 | ≥ 2𝑅 |𝛼| =⇒ |D 𝛼 𝑔 𝑅 (𝜉)| ≤ 2 (𝐶1 /𝑅) | 𝛼 | ;
(3) ∀𝜁 𝜉 ∈ ℭ♭ =⇒ 𝑔 𝑅 (𝜁) = 1 and 𝜉 ∉ ℭ =⇒ 𝑔 𝑅 (𝜁) = 0;
= 𝜉 + 𝑖𝜂 ∈ C𝑛 , √
(4) ∀𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 \ −1R𝑛 , we have

|𝑔 𝑅 (𝜁)| ≤ exp (𝐶1 |𝜂| /𝑅) , (17.3.3)

𝜕 𝜁 𝑔 𝑅 (𝜁) ≤ 𝐶2 exp ((𝐶3 |𝜂| − |𝜉 |) /𝑅) . (17.3.4)

Proof We make use of a sequence {𝜑 𝑁 } 𝑁 =1,2,... of Ehrenpreis’ cutoffs relative to

the open subsets of S𝑛−1 , 𝑈 = ℭ♭ ∩ S𝑛−1 and 𝑉 = ℭ ∩ S𝑛−1 . To define the 𝜑 𝑁 we
use coordinates pulled back from R𝑛−1 under the stereographic projection from a
point in S𝑛−1 \𝑉 taken as the North Pole. It is convenient (but not necessary) to take
0 ≤ 𝜑 𝑁 ≤ 1 for all 𝑁. Actually we shall retain only the subsequence corresponding
to 𝑁 = 4𝜈−1 (𝜈 = 1, 2, ...) and set 𝜓 𝜈 = 𝜑4𝜈−1 . According to (3.2.2), we have, for any
𝛾 ∈ Z+𝑛−1 ,
|𝛾 |
𝛾
𝜕𝜃 𝜓 𝜈 (𝜃) ≤ 4𝜈−1 𝐶◦ if |𝛾| ≤ 4𝜈−1 . (17.3.5)

Making use of the functions ℎ 𝜈 of Lemma 17.3.1 we define

∞
∑︁
𝑔 𝑅 (𝜉) = ℎ 𝜈 (|𝜉 | /𝑅) 𝜓 𝜈+1 (𝜉/|𝜉 |) , (17.3.6)
𝜈=0

where 𝑅 > 0. Note that 0 ≤ 𝑔 𝑅 (𝜉) ≤ 1 for all 𝜉 ∈ R𝑛 \ {0} and

𝜉 ∈ ℭ♭ =⇒ 𝑔 𝑅 (𝜉) = 1, 0 ≠ 𝜉 ∉ ℭ =⇒ 𝑔 𝑅 (𝜉) = 0, (17.3.7)

since ∞
Í
𝜈=0 ℎ 𝜈 ≡ 1. By Property (2) in Lemma 17.3.1 the only terms not equal to
1
zero in the series (17.3.6) are those for which 4𝜈−1 𝑅 ≤ |𝜉 | ≤ 4𝜈+ 2 𝑅. There are at
most two such terms.
660 17 Analytic Pseudodifferential Calculus

We have, for 𝛼 ∈ Z+𝑛 ,

∞ ∑︁
𝛼
∑︁ 𝛼 𝛽 𝛼−𝛽
D 𝑔 𝑅 (𝜉) = D 𝜉 ℎ 𝜈 (|𝜉 | /𝑅) D 𝜉 (𝜓 𝜈+1 (𝜉/|𝜉 |)) .
𝜈=0 𝛽 ⪯ 𝛼
𝛽

Suppose |𝜉 | ≥ 2𝑅 |𝛼|. By Property (3) in Lemma 17.3.1,

𝛽 ⪯ 𝛼 =⇒ D 𝜉 ℎ 𝜈 (|𝜉 | /𝑅) ≤ (𝐶/𝑅) |𝛽 |

𝛽

with 𝐶 > 0 independent of 𝜈, 𝛼 and 𝑅. We also have |𝛼| ≤ 4𝜈 , in which case, taking
(17.3.5) into account:

(𝜓 𝜈+1 (𝜉/|𝜉 |)) ≤ (4𝜈 𝐶/|𝜉 |) | 𝛼−𝛽 | ≤ (4𝐶/𝑅) | 𝛼−𝛽 |

𝛼−𝛽
D𝜉

possibly with 𝐶 increased. We obtain

∑︁ 𝛼
D 𝜉 ℎ 𝜈 (|𝜉 | /𝑅) D 𝜉 (𝜓 𝜈+1 (𝜉/|𝜉 |)) ≤ (5𝐶/𝑅) | 𝛼 | ,
𝛽 𝛼−𝛽

𝛽⪯𝛼
𝛽

whence Property (2) provided

√ 𝐶1 ≥ 5𝐶.
We extend 𝑔 𝑅 to C𝑛 \ −1R𝑛 by an expansion of the type (17.1.4): let 𝜒 ∈ C ∞ (R),
0 ≤ 𝜒 ≤ 1, 𝜒 (𝜏) = 0 if 𝜏 < 1, 𝜒 (𝜏) = 1 if 𝜏 > 2, and define for 𝜁 = 𝜉 + 𝑖𝜂 ∈ R2𝑛 ,
𝜉 ≠ 0,
∑︁ (−𝜂) 𝛾 𝛾
𝑔 𝑅 (𝜁) = 𝑔 𝑅 (𝜉) + 𝜒 (|𝜉 | /2𝑅 |𝛾|) D 𝜉 𝑔 𝑅 (𝜉) .
𝛾!
𝛾 ∈Z+𝑛 , |𝛾 | ≥1

This is a convergent series, thanks to the fact that 0 ≤ 𝜒 (|𝜉 | /2𝑅 |𝛾|) ≤ 1 and to
Property (2). Property (3) is an immediate consequence of (17.3.7); (17.3.3) is a
direct consequence of Property (2).
The proof of (17.3.4) mimics that of (17.1.3) taking into account the difference
between the properties of 𝑔 𝑅 (𝜉) here and those of 𝑎 (𝜉) in Lemma 17.1.1. We leave
the details as an exercise. □

Remark 17.3.3 If we compare the properties of the cutoff 𝑔 𝑅 in Lemma 17.3.2 to

those of the amplitude 𝑎 as stated in Proposition 17.1.8 we see that changing 𝑅
entails changing 𝑔 𝑅 itself, whereas in the case of 𝑎 it only entails modifying the
asymptotically holomorphic extension 𝑎˜ 𝑅 . Moreover, the properties of 𝑔 𝑅 do not
entail any decay of the type |𝜕 𝛼 𝑔 𝑅 (𝜉)| ≤ 𝐶 |𝜉 | 𝑚−| 𝛼 | . It follows that if 𝑎 (𝑥, 𝑦, 𝜉) ∈
𝑆 𝜓a (Ω × Ω) then, generally, 𝑎 (𝑥, 𝑦, 𝜉) 𝑔 𝑅 (𝜉) ∉ 𝑆 𝜓a (Ω × Ω).

If 𝑔 𝑅′ is a function with the same properties as those of 𝑔 𝑅 listed in Lemma 17.3.2

but relative to a different pair of cones ℭ ′, ℭ ′♭ , ℭ ′♭ ∩ S𝑛−1 ⊂⊂ ℭ ′ ∩ S𝑛−1 ≠ S𝑛−1 ,
then 𝑔 𝑅 𝑔 𝑅′ has those properties too, relative to ℭ ′ ∩ ℭ ′♭ and ℭ ′ ∪ ℭ ′♭ (supposing
17.3 Analytic Microlocalization In Distribution Theory 661

ℭ ′ ∪ ℭ ′♭ ≠ R𝑛 \ {0}). It must also be pointed out that the function 1 − 𝑔 𝑅 has the
same properties as 𝑔 𝑅 but relative to the open cones R𝑛 \ℭ♭ ⊂ R𝑛 \ℭ (where the bars
signify closures in R𝑛 ).
These observations allow us to construct partitions of unity consisting of functions
like 𝑔 𝑅 subordinate to a finite covering of R𝑛 \ {0} by open cones ℭ1 , ..., ℭ 𝜈 . We
mimic the construction at the end of Subsection 17.1.2. For each 𝑗 = 1, ..., 𝜈 we can
select a pair of open cones ℭ ′𝑗 , ℭ ′′𝑗 such that

ℭ ′′𝑗 ∩ S𝑛−1 ⊂⊂ ℭ ′𝑗 ∩ S𝑛−1 ⊂⊂ ℭ 𝑗 ∩ S𝑛−1 ≠ S𝑛−1

and
R𝑛 \ {0} = ℭ1′′ ∪ · · · ∪ ℭ 𝜈′′. (17.3.8)
First, for each 𝑗 = 1, ..., 𝜈 −1, we define a function 𝑔 𝑅, 𝑗 as in Lemma 17.3.2 where we
take ℭ = ℭ ′𝑗 and ℭ♭ = ℭ ′′𝑗 . Setting 𝑔 𝑅(1) = 𝑔 𝑅,1 we then define 𝑔 𝑅(2) = 1 − 𝑔 𝑅,1 𝑔 𝑅,2

with the understanding that 𝑔 𝑅,2 = 1 if 𝜈 = 2. We have supp 𝑔 𝑅(1) ⊂ ℭ1′ . When 𝜈 = 2
we have 𝑔 𝑅(1) + 𝑔 𝑅(2) = 1 and supp 𝑔 𝑅(2) ⊂ R𝑛 \ ℭ1′′ ∪ {0} ⊂ ℭ2′′ by (17.3.8). If 𝜈 ≥ 3

we define 𝑔 𝑅(3) = 1 − 𝑔 𝑅,1 1 − 𝑔 𝑅,2 𝑔 𝑅,3 with the understanding that 𝑔 𝑅,3 = 1 if

𝜈 = 3. When 𝜈 = 3 we have 𝑔 𝑅(2) + 𝑔 𝑅(3) = 1 − 𝑔 𝑅,1 , hence 𝑔 𝑅(1) + 𝑔 𝑅(2) + 𝑔 𝑅(3) = 1; we

also have ′′
′′

supp 𝑔 𝑅(2) ⊂ R𝑛 \ℭ 1 ∩ R𝑛 \ℭ 2 ⊂ ℭ3′′.

If 𝜈 ≥ 4 we define 𝑔 𝑅(4) = 1 − 𝑔 𝑅,1 1 − 𝑔 𝑅,2 1 − 𝑔 𝑅,3 𝑔 𝑅,4 with the understand-

( 𝑗)
ing that 𝑔 𝑅,4 = 1 if 𝜈 = 4. And so forth. For every 𝜈 we end up with functions 𝑔 𝑅 ,
( 𝑗) (1) (𝜈)
𝑗 = 1, ..., 𝜈, such that supp 𝑔 𝑅 ⊂ ℭ ′𝑗 for each 𝑗, and 𝑔 𝑅 + · · · + 𝑔 𝑅 ≡ 1 in R𝑛 \ {0}.

17.3.2 Inverse Fourier transform of the cutoff 𝒈 𝑹

Let 𝑔 𝑅 be the function in Lemma 17.3.2; we define the distribution in R𝑛 ,

∫
𝐺 𝑅 (𝑥) = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 𝑔 𝑅 (𝜉) d𝜉.
R𝑛

If 𝑢 ∈ E ′ (R𝑛 ) the convolution 𝐺 𝑅 ∗ 𝑢 ∈ D ′ (R𝑛 ) will be rather denoted by 𝑔 𝑅 (D) 𝑢,

in the spirit of pseudodifferential calculus notation.

Proposition 17.3.4 There is a constant 𝐶 > 0, independent of 𝑅 > 0, such that the
distribution 𝐺 𝑅 (𝑥) is analytic in the region |𝑥| > 𝐶 𝑅 −1 .

Proof We have 𝐺 𝑅 = (1 − Δ 𝑥 ) 𝑁 ℎ 𝑅, 𝑁 with

∫
−𝑛 d𝜉
ℎ 𝑅, 𝑁 (𝑥) = (2𝜋) e𝑖 𝑥· 𝜉 𝑔 𝑅 (𝜉) 𝑁 , (17.3.9)
1 + |𝜉 | 2
𝑛 R
662 17 Analytic Pseudodifferential Calculus

a continuous function decaying at infinity provided 2𝑁 > 𝑛 (we recall that 0 ≤ 𝑔 𝑅 ≤

1). It suffices to show that the function ℎ 𝑅, 𝑁 is analytic in a region |𝑥| > 𝐶 𝑅 −1 . We
are allowed to deform the domain of integration in (17.3.9) from R𝑛 to the image of
R𝑛 under the map 𝜉 ↦→ 𝜉 + 𝑖𝜅𝑥 |𝜉 | with 𝜅 > 0 suitably small (in particular, we must
have 𝜅 |𝑥| < 1):

(2𝜋) 𝑛 ℎ 𝑅, 𝑁 (𝑥) (17.3.10)

∫
2 𝐷 (𝜅𝑥, 𝜉) d𝜉
= e𝑖 𝑥· 𝜉 −𝜅 | 𝑥 | |𝜉 |
𝑔 𝑅 (𝜉 + 𝑖𝜅𝑥 |𝜉 |) 𝑁
R𝑛
1 + ⟨𝜉 + 𝑖𝜅𝑥 |𝜉 |⟩ 2
∫
d𝜁
+ e𝑖 𝑥·𝜁 𝜕 𝜁 𝑔 𝑅 (𝜁) ∧ 𝑁 ,
Λ 𝜅 ( 𝑥) 2
1 + ⟨𝜁⟩

where 𝐷 (𝜅𝑥, 𝜉) is the Jacobian determinant of the map 𝜉 ↦→ 𝜉 + 𝑖𝜅𝑥 |𝜉 | and

Λ 𝜅 (𝑥) = {𝜁 ∈ C𝑛 ; ∃ (𝜉, 𝑡) ∈ R𝑛 × (0, 1) , 𝜁 = 𝜉 + 𝑖𝜅𝑡𝑥 |𝜉 |} .

We are using the notation ⟨𝜁⟩ 2 = 𝜁 · 𝜁 = 𝜁12 + · · · + 𝜁 𝑛2 . We take advantage of

(17.3.3)–(17.3.4):

|𝑔 𝑅 (𝜉 + 𝑖𝜅𝑥 |𝜉 |)| ≤ 𝐶1 exp (𝐶1 𝜅 |𝑥| |𝜉 | /𝑅) ;

𝜕 𝜁 𝑔 𝑅 (𝜉 + 𝑖𝜅𝑡𝑥 |𝜉 |) ≤ 𝐶1 exp (− (1 − 𝐶1 𝜅𝑡 |𝑥|) |𝜉 | /𝑅) ,

We get

𝑖 𝑥· 𝜉 −𝜅 | 𝑥 | 2 | 𝜉 | 1
e 𝑔 𝑅 (𝜉 + 𝑖𝜅𝑥 |𝜉 |) ≤ 𝐶1 exp −𝜅 |𝑥| − 𝐶1 /𝑅 |𝑥| |𝜉 | ,
2
2
e𝑖 𝑥· 𝜉 −𝜅 |𝑥 | |𝜉 |
𝜕 𝜁 𝑔 𝑅 (𝜉 + 𝑖𝜅𝑥 |𝜉 |) ≤ 𝐶1 exp (− (1 − 𝜅 (𝐶1 − |𝑥|) |𝑥|) |𝜉 | /𝑅) .

1
If we require 𝜅 to be sufficient small that 𝜅 (𝐶1 − |𝑥|) |𝑥| ≤ 2 we get, for all 𝑥 ∈ R𝑛 ,
1 −1
2 𝐶1 /𝑅 + 𝛿 < |𝑥| < 𝜅 , 𝛿 > 0,

2
e𝑖 𝑥· 𝜉 −𝜅 | 𝑥 | |𝜉 |
𝑔 𝑅 (𝜉 + 𝑖𝜅𝑥 |𝜉 |) ≤ 𝐶1 e− 𝛿 | 𝜉 | ,
2
e𝑖 𝑥· 𝜉 −𝜅 | 𝑥 | |𝜉 |
𝜕 𝜁 𝑔 𝑅 (𝜉 + 𝑖𝜅𝑥 |𝜉 |) ≤ 𝐶1 e−| 𝜉 |/2𝑅 .

Estimates of this type will remain valid after replacement of 𝑥 by 𝑧 ∉ R𝑛 provided

|𝑧 − Re 𝑧| is suitably small. By letting 𝜅 and 𝛿 go to zero we see that each one of the
two integrals on the right in (17.3.10) admits a well-defined holomorphic extension
to an open subset of C𝑛 containing the region |𝑥| > 21 𝐶1 /𝑅. □
17.3 Analytic Microlocalization In Distribution Theory 663

Corollary 17.3.5 Let 𝑢 ∈ E ′ (R𝑛 ) be an analytic function in an open set Ω ⊂ supp 𝑢.

Given arbitrary 𝜀 > 0 the convolution 𝐺 𝑅 ∗ 𝑢 is an analytic function in the open set
Ω 𝜀 = {𝑥 ∈ Ω; dist (𝑥, R𝑛 \Ω) > 𝜀} provided 𝑅𝜀 > 𝐶, the constant in the statement
of Proposition 17.3.4.

We leave the proof of Corollary 17.3.5 as an exercise.

17.3.3 Using 𝒈 𝑹 (D) to delimit the analytic wave-front set of a

distribution

In this subsection we apply to distributions the concepts of microlocal analyticity and

analytic wave-front set introduced in Section 3.4. We are going to need the following

Lemma 17.3.6 Let ℭ♭ be a cone closed in R𝑛 \ {0} and 𝑔 be a measurable function

in R𝑛 such that
sup (1 + |𝜉 |) −𝑚 |𝑔 (𝜉)| < +∞ (17.3.11)
𝜉 ∈R𝑛

for some 𝑚 ∈ R. Under this hypothesis, if supp 𝑔 ⊂ ℭ♭ then 𝑊 𝐹a (𝑔 (D) 𝑢) ⊂ R𝑛 ×ℭ♭

for every 𝑢 ∈ E ′ (R𝑛 ).

Proof The FBI transform (3.4.8), (with 𝜅 = 21 ) of 𝑔 (D) 𝑢 at a point (𝑥, 𝜃) ∈ R𝑛 ×

(R𝑛 \ {0}) is equal to
∫ ∫
1 2
𝐹 (𝑥, 𝜃) = (2𝜋) −𝑛 e𝑖 𝜃 · ( 𝑥−𝑦)− 2 | 𝜃 | | 𝑥−𝑦 | e𝑖𝑦· 𝜉 𝑔 (𝜉) b𝑢 (𝜉) d𝜉d𝑦
R 𝑛 ℭ ♭
∫ ∫
1 2
= (2𝜋) −𝑛 e𝑖 ( 𝜃− 𝜉 ) ·𝑦− 2 | 𝜃 | |𝑦 | e𝑖 𝑥· 𝜉 𝑔 (𝜉) b
𝑢 (𝜉) d𝜉d𝑦
R𝑛 ℭ ♭
∫
1 1 −1 2
= (2𝜋 |𝜃|) − 2 𝑛 e𝑖 𝑥· 𝜉 e− 2 | 𝜃 | | 𝜃− 𝜉 | 𝑔 (𝜉) b 𝑢 (𝜉) d𝜉.
ℭ♭

If 𝜃 ∉ ℭ♭ we select 𝛿 ∈ 0, 21 such that | 𝜃𝜃 | − | 𝜉𝜉 | ≥ 3𝛿 for all 𝜉 ∈ ℭ♭ . In this case
we contend that
1
|𝜃| −1 |𝜃 − 𝜉 | 2 ≥ 𝛿2 (|𝜃| + |𝜉 |) . (17.3.12)
2
Indeed, if |𝜃| ≥ (1 − 𝛿) −1 |𝜉 | then |𝜃| −1 |𝜃 − 𝜉 | 2 ≥ 𝛿2 |𝜃|, whence (17.3.12). If
|𝜉 | ≥ (1 − 𝛿) −1 |𝜃| then
2
−1 𝛿
|𝜃| 2
|𝜃 − 𝜉 | ≥ |𝜃| −1 |𝜉 | 2 ≥ 𝛿2 |𝜉 | ,
1−𝛿
|𝜉 |
whence (17.3.12) here too. Finally, suppose 1 − 𝛿 < |𝜃 | < (1 − 𝛿) −1 = 1 + 𝛿
1− 𝛿 ≤
1 + 2𝛿; in this case,
664 17 Analytic Pseudodifferential Calculus

𝜃 𝜉 𝜃 𝜉 𝜉 𝜉
− ≥ − − −
|𝜃| |𝜃| |𝜃| |𝜉 | |𝜃| |𝜉 |
|𝜉 |
≥ 3𝛿 − −1 ≥ 𝛿
|𝜃|
2
1 2
implies |𝜃| | 𝜃𝜃 | − | 𝜉𝜃 | ≥ 𝛿2 |𝜃| ≥ 2𝛿 (|𝜃| + (1 − 𝛿) |𝜉 |), whence (17.3.12) once
again. This bound implies
∫
1 1 2|𝜃 | 1 2|𝜉 |
|𝐹 (𝑥, 𝜃)| ≤ (2𝜋 |𝜃|) − 2 𝑛 e− 2 𝛿 e− 2 𝛿 |𝑔 (𝜉) b
𝑢 (𝜉)| d𝜉. (17.3.13)
ℭ♭

To complete the proof of Lemma 17.3.6 it suffices to apply Theorem 3.4.7 taking
into account (17.3.11) as well as Theorem 2.1.2. □
Proposition 17.3.7 For 𝑢 ∈ E ′ (R𝑛 ) to be microanalytic at (𝑥 ◦ , 𝜉 ◦ ) ∈ R𝑛 ×(R𝑛 \ {0})
(Definition 7.4.7) it is necessary and sufficient that there be an open subset 𝑈 of R𝑛
containing 𝑥 ◦ , two open cones ℭ, ℭ♭ ⊂ R𝑛 \ {0}, 𝜉 ◦ ∈ ℭ♭ ∩S𝑛−1 ⊂⊂ ℭ∩S𝑛−1 ≠ S𝑛−1 ,
and a number 𝑅 > 0 such that 𝑔 𝑅 (D) 𝑢 is an analytic function in 𝑈, where 𝑔 𝑅 is the
cutoff function associated with ℭ, ℭ♭ , 𝑅, in Lemma 17.3.2.
−𝑘
Proof If the positive integer 𝑘 is sufficiently large b 𝑓 (𝜉) = 1 + |𝜉 | 2 𝑢 (𝜉) will
b
belong to 𝐿 1 (R𝑛 ) and 𝑓 (𝑥) = (2𝜋) −𝑛 ♭ e𝑖 𝑥· 𝜉 b
∫
𝑓 (𝜉) d𝑥 will be a continuous func-
ℭ
tion. If 𝑔 𝑅 (D) 𝑓 is analytic in 𝑈 the same is true of 𝑔 𝑅 (D) 𝑢 = (1 − Δ) 𝑘 𝑔 𝑅 (D) 𝑓 ;
and vice versa by Theorem 4.1.1, since the differential operator (1 − Δ) 𝑘 is elliptic.
We can even replace 𝑓 by 𝜒 𝑓 with 𝜒 ∈ Cc∞ (R𝑛 ), 𝜒 (𝑥) = 1 for all 𝑥 such that
dist (𝑥, 𝑈) ≥ 𝛿 > 0. Indeed, according to Corollary 17.3.5 𝑔 𝑅 (D) ((1 − 𝜒) 𝑓 ) will
be analytic in 𝑈 provided 𝑅 is sufficiently large. In other words, there is no loss of
generality in assuming that 𝑢 ∈ Cc0 (R𝑛 ).
I. Necessity of the condition.
We shall avail ourselves of Theorem 3.5.13. Since (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢) we can find
a pair of open subsets 𝑈, 𝑉 of R𝑛 with 𝑥 ◦ ∈ 𝑈 ⊂⊂ 𝑉, an Ehrenpreis sequence 𝜑 𝑁
(𝑁 = 1, 2, ...) relative to the pair (𝑈, 𝑉) (Definition 3.2.2) and positive numbers 𝜌,
𝐶, such that

∀𝑁, 𝑘 ∈ Z+ , 𝑘 ≤ 𝑁, ∀𝜉 ∈ ℭ𝜌 (𝜉 ◦ ) , |𝜉 | 𝑘 | 𝜑d
𝑁 𝑢 (𝜉)| ≤ 𝐶
𝑁 +1
𝑘!, (17.3.14)
n o
where ℭ𝜌 (𝜉 ◦ ) = 𝜉 ∈ R𝑛 ; 𝜉 ≠ 0, | 𝜉𝜉 | − 𝜉 ◦ < 𝜌 , 𝜌 > 0 suitably small. We take
ℭ = ℭ𝜌 (𝜉 ◦ ), ℭ♭ = ℭ𝜌/2 (𝜉 ◦ ). Let 𝑔 𝑅 (𝜉) denote the function in Lemma 17.3.2
relative to this pair of cones and to a number 𝑅 > 0 to be determined. We have, for
each 𝛼 ∈ Z+𝑛 ,
∫
𝛼 −𝑛
D 𝑔 𝑅 (D) (𝜑 𝑁 𝑢) (𝑥) = (2𝜋) e𝑖 𝑥· 𝜉 𝜉 𝛼 𝑔 𝑅 (𝜉) 𝜑d
𝑁 𝑢 (𝜉) d𝜉.
ℭ

Applying (17.3.14) we derive, for 𝑁 = |𝛼| + 𝑛 + 1,

17.3 Analytic Microlocalization In Distribution Theory 665

sup |D 𝛼 𝑔 𝑅 (D) (𝜑 𝑁 𝑢)| ≲ (1 + |𝜉 |) 𝑁 𝜑d

𝑁𝑢 𝐿∞
≲ 𝐶 𝑁 𝑁!. (17.3.15)
R𝑛

We let 𝜑 𝑁 be any element of a partition of unity 𝜑 𝑁 , 𝜄 𝜄=1,...,𝑟 subordinate to an
open covering {𝑉 𝜄 } 𝜄=1,...,𝑟 of the closure 𝐾 of 𝑈, meaning that 𝜑 𝑁 ,1 + · · · + 𝜑 𝑁 ,𝑟 = 1
in 𝑈 ′ = 𝑈1′ ∪ · · · ∪ 𝑈𝑟′ ⊃ 𝐾, where 𝑈 𝜄′ ⊂⊂ 𝑉 𝜄 for each 𝜄; and that for each 𝜄,

𝜑 𝑁 , 𝜄 𝑁 =1,2,... is an Ehrenpreis sequence associated to the pair (𝑈 𝜄 , 𝑉 𝜄 ). Adding up
all the estimates (17.3.15) for 𝜑 𝑁 = 𝜑 𝑁 , 𝜄 , 𝜄 = 1, ..., 𝑟, leads to the estimate
𝑟
∑︁
sup D 𝛼 𝑔 𝑅 (D) 𝜑 𝑁 , 𝜄 𝑢 ≲ 𝐶 𝑁 𝑁!, (17.3.16)
R𝑛 𝜄=1

valid for all 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑁 −𝑛−1. Fixing 𝑁 we define 𝜓 𝑁 = 1− 𝜑 𝑁 ,1 + · · · + 𝜑 𝑁 ,𝑟 .
We claim that there are positive constants 𝑅◦ and 𝐶◦ such that, for all 𝑅 > 𝑅◦ , all
𝑁 = 1, 2, ..., and all 𝛼 ∈ Z+𝑛 ,

max |D 𝛼 𝑔 𝑅 (D) (𝜓 𝑁 𝑢)| ≤ 𝐶◦| 𝛼 |+1 𝛼!. (17.3.17)

𝐾

Proof of (17.3.17): As before let 𝐺 𝑅 (𝑥) stand for the inverse Fourier transform
of 𝑔 𝑅 (𝜉). We have
∫
𝑔 𝑅 (D) (𝜓 𝑁 𝑢) (𝑥) = 𝐺 𝑅 (𝑥 − 𝑦) (𝜓 𝑁 𝑢) (𝑦) d𝑦, (17.3.18)
𝐾′

where 𝐾 ′ = supp 𝑢 ∩ (R𝑛 \𝑈 ′). If 𝑥 ∈ 𝐾 and 𝑦 ∈ 𝐾 ′ then 𝑥 − 𝑦 ∈ 𝐾 ′′, a compact

subset of R𝑛 on which |𝑥 − 𝑦| ≥ dist (𝐾, R𝑛 \𝑈). By Proposition 17.3.4 we know
that if 𝑅 and 𝐶 are sufficiently large then max ′′ |D 𝛼 𝐺 𝑅 (𝑥 − 𝑦)| ≤ 𝐶 | 𝛼 |+1 𝛼! for all
𝑥−𝑦 ∈𝐾
𝛼 ∈ Z+𝑛 . Since 𝑢 ∈ Cc0 (R𝑛 ) and 0 ≤ 𝜓 𝑁 ≤ 1 differentiation under the integral sign
in (17.3.18) leads directly to (17.3.17).
Finally, combining (17.3.16) and (17.3.17) yields max |D 𝛼 𝑔 𝑅 (D) 𝑢| ≤ 𝐶1| 𝛼 |+1 𝛼!
𝐾
for all 𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝑁 − 𝑛 − 1, with a constant 𝐶1 > 0 independent of 𝑁.
II. Sufficiency of the condition. Here we avail ourselves of the fact that both 𝑔 𝑅
and 1 − 𝑔 𝑅 have the properties of the multiplier 𝑔 in Lemma 17.3.6. We derive from
this that
𝑊 𝐹a ((1 − 𝑔 𝑅 (D)) 𝑢) ∩ R𝑛 × ℭ♭ = ∅.

Our hypothesis is that

𝑊 𝐹a (𝑔 𝑅 (D) 𝑢) ∩ (𝑈 × (R𝑛 \ {0})) = ∅.

We conclude that 𝑊 𝐹a (𝑢) ∩ 𝑈 × ℭ♭ = ∅. □

Remark 17.3.8 Inspection of the proof of Proposition 17.3.7 shows that if 𝑔 𝑅 (D) 𝑢
is analytic in 𝑈 then the same is true of 𝑔 𝑅1 (D) 𝑢 whatever the number 𝑅1 > 𝑅.
666 17 Analytic Pseudodifferential Calculus

17.3.4 Analytic pseudodifferential operators decrease the analytic

wave-front set of distributions

Combined with Theorem 17.2.13 the next lemma will enable us to prove the analytic
version of the important part of Theorem 16.2.20. As before Ω is an open subset of
R𝑛 , ℭ is an open cone in R𝑛 \ {0}.

Lemma 17.3.9 Let 𝑎 (𝑥, 𝜉) ∈ C ∞ (Ω × ℭ) satisfy the following condition, for some
𝑚 ∈ R:
(A𝑚 ) To every compact subset 𝐾 of Ω there is a 𝐶 > 0 such that, for all 𝛾 ∈ Z+𝑛 ,
(𝑥, 𝜉) ∈ 𝐾 × ℭ,
𝜕𝑥 𝑎 (𝑥, 𝜉) ≤ 𝐶 |𝛾 |+1 𝛾! (1 + |𝜉 |) 𝑚 .
𝛾
(17.3.19)

Assume furthermore that there is an open cone in R𝑛 \ {0}, ℭ♭ ⊂ ℭ, such that

𝑎 (𝑥, 𝜉) ≡ 0 when 𝜉 ∈ ℭ♭ .
Under these hypotheses, Ω × ℭ♭ ∩ 𝑊 𝐹a ((Op𝑎) 𝑢) = ∅ whatever 𝑢 ∈ E ′ (R𝑛 ).

It is a direct consequence of Theorem 16.1.4 that Property (A𝑚 ) ensures that Op𝑎
is a well-defined continuous linear operator E ′ (R𝑛 ) −→ D ′ (Ω).
Proof Inspection of the proof of Lemma 17.1.13 shows that we can replace the
hypothesis 𝑎 (𝑥, 𝜉) ∈ 𝑆 𝑚 ∞
𝜓a (Ω × ℭ) by the hypothesis that 𝑎 ∈ C (Ω × ℭ) satisfies
Condition (A𝑚 ). It therefore suffices to deal with the case 𝑚 < −𝑛 − 1 and 𝑢 ∈
𝐿 c1 (R𝑛 ); we set 𝑣 = (Op𝑎) 𝑢. Consider two open subsets 𝑉 ⊂⊂ 𝑈 ⊂⊂ Ω and let
{𝜑 𝜈 } 𝜈=1,2,... be an Ehrenpreis sequence relative to the pair (𝑈, 𝑉) (Definition 3.2.2).
We have
∫ ∫
−𝑛
𝜑d 𝜈 𝑣 (𝜉) = (2𝜋) e𝑖 𝑥· ( 𝜉 −𝜂) 𝜑 𝜈 (𝑥) 𝑎(𝑥, 𝜂)b
𝑢 (𝜂) d𝑥d𝜂,
R𝑛 \ℭ 𝑈

whence

(2𝜋) 𝑛 𝜑
d 𝜈 𝑣 (𝜉)

(1 − Δ 𝑥 ) 𝑁 (𝜑 𝜈 (𝑥) 𝑎(𝑥, 𝜂))

∫ ∫
= e𝑖 𝑥· ( 𝜉 −𝜂) 𝑁 𝑢 (𝜂) d𝑥d𝜂.
b
R𝑛 \ℭ 𝑈 2
1 + |𝜉 − 𝜂|

The definition of an Ehrenpreis sequence implies that there is a 𝐶◦ > 0 independent

of 𝜈, such that
∀𝛼 ∈ Z+𝑛 , |𝛼| ≤ 𝜈, max |𝜕 𝛼 𝜑 𝜈 | ≤ (𝐶◦ 𝜈) | 𝛼 | . (17.3.20)
We select 𝜈 = 2𝑁 − 1 or 𝜈 = 2𝑁. Applying the Leibniz and Sterling formulas we
derive from (17.3.19) and (17.3.20) that there is a 𝐶1 > 0 independent of 𝜈 such that

(1 − Δ 𝑥 ) 𝑁 (𝜑 𝜈 (𝑥) 𝑎(𝑥, 𝜂)) ≤ 𝐶1𝜈+1 𝜈! (1 + |𝜂|) −𝑛−1 .

17.3 Analytic Microlocalization In Distribution Theory 667

We derive

(1 + |𝜉 |) 𝜈 | 𝜑
d 𝜈 𝑣 (𝜉)| (17.3.21)
(1 + |𝜉 |) 𝜈
∫
≤ 𝐶1𝜈+1 𝜈! |𝑈| ∥b
𝑢 ∥ 𝐿∞ 𝑁 d𝜂
R𝑛 \ℭ 2 −𝑛−1
1 + |𝜉 − 𝜂| (1 + |𝜂|)

where |𝑈| is the Lebesgue measure of 𝑈.

Let 𝜉 ◦ ∈ ℭ♭ be arbitrary and select an open cone ℭ∗ ∋ 𝜉 ◦ such that ℭ∗ ∩ S𝑛−1 ⊂⊂
ℭ ∩ S𝑛−1 . The latter implies that there is a 𝛿 > 0 such that
♭

𝜉 ∈ ℭ∗ , 𝜂 ∉ ℭ♭ =⇒ |𝜉 − 𝜂| ≥ 𝛿 (|𝜉 | + |𝜂|) (17.3.22)

since this is true when |𝜉 | = |𝜂| = 1. Taking (17.3.22) into account in (17.3.21)
yields, for 𝐶2 > 0 independent of 𝜈 and of 𝜉 ∈ ℭ∗ ,
𝜈+1 −𝜈
|𝜑
d 𝜈 𝑣 (𝜉)| ≤ 𝐶2 𝜈! (1 + |𝜉 |) . (17.3.23)

By applying Theorem 3.5.13, we conclude from (17.3.23) that 𝑣 ∈ D ′ (Ω) is micro-

analytic at (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑈 × ℭ∗ whatever 𝑥 ◦ ∈ 𝑈. □

Theorem 17.3.10 If 𝑨ℭ ∈ Ψa (Ω, ℭ) and if 𝑢 ∈ E ′ (Ω) then 𝑊 𝐹a ( 𝑨ℭ 𝑢) ⊂ 𝑊 𝐹a (𝑢).

Proof We must show that if 𝑢 ∈ E ′ (Ω) is microanalytic at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × ℭ

(Definition 3.5.1) then the same is true of 𝑨ℭ 𝑢. According to Proposition 17.3.7
there exist a neighborhood 𝑈 ⊂⊂ Ω of 𝑥 ◦ , an open cone ℭ♭ , 𝜉 ◦ ∈ ℭ♭ ∩ S𝑛−1 ⊂⊂
ℭ, and a number 𝑅 > 0 such that 𝑔 𝑅 (D) 𝑢|𝑈 ∈ C 𝜔 (𝑈), 𝑔 𝑅 being the cutoff
function associated with ℭ, ℭ♭ , 𝑅 as in Lemma 17.3.2. According to Proposition
17.2.9 there is no loss of generality in assuming that 𝑨ℭ = 𝑎 (𝑥, D), 𝑎 (𝑥, 𝜉) ∈
𝑆 𝜓a (Ω × ℭ). Let 𝜒 ∈ Cc∞ (Ω), 𝜒 (𝑥) = 1 for all 𝑥 ∈ 𝑈; we have 𝜒 (𝑥) 𝑔 𝑅 (D) 𝑢|𝑈 ∈
C 𝜔 (𝑈) hence 𝑎 (𝑥, D) ( 𝜒 (𝑥) 𝑔 𝑅 (D) 𝑢)|𝑈 ∈ C 𝜔 (𝑈) by Theorem 17.2.13. We avail
ourselves of Proposition 16.2.2: the pseudodifferential operator 𝑎 (𝑥, D) extends
as a continuous linear operator 𝐻 𝑠 (R𝑛 ) −→ D ′ (Ω) whatever 𝑠 ∈ R𝑛 . If 𝑢 ∈
E ′ (R𝑛 ) then 𝑔 𝑅 (D) 𝑢 ∈ 𝐻 𝑠 (R𝑛 ) for some 𝑠 ∈ R𝑛 ; a simple adaptation of the proof
of Theorem 17.1.14 shows that 𝑎 (𝑥, D) ((1 − 𝜒 (𝑥)) 𝑔 𝑅 (D) 𝑢)|𝑈 ∈ C 𝜔 (𝑈). Thus
𝑎 (𝑥, D) 𝑔 𝑅 (D) 𝑢|𝑈 ∈ C 𝜔 (𝑈).
We have
∫
𝑎 (𝑥, D) (1 − 𝑔 𝑅 (D)) 𝑢 = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 ℎ 𝑅 (𝑥, 𝜉) b
𝑢 (𝜉) d𝜉, (17.3.24)
R𝑛

where
ℎ 𝑅 (𝑥, 𝜉) = 𝑎 (𝑥, 𝜉) (1 − 𝑔 𝑅 (𝜉)) .
It follows directly from the definition of 𝑔 𝑅 (see Lemma 17.3.2) and from Def-
inition 17.2.3 that ℎ 𝑅 satisfies the hypotheses in Lemma 17.3.9 with 𝑚 = order of
𝑎 (𝑥, 𝜉). We conclude that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑎 (𝑥, D) (1 − 𝑔 𝑅 (D)) 𝑢) and therefore,
that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑎 (𝑥, D) 𝑢). □
668 17 Analytic Pseudodifferential Calculus

Corollary 17.3.11 If 𝑨ℭ ∈ Ψa (Ω, ℭ) is elliptic then 𝑊 𝐹a ( 𝑨ℭ 𝑢) = 𝑊 𝐹a (𝑢) for

every 𝑢 ∈ E ′ (Ω) .
Proof Let Ω′ ⊂⊂ Ω be an open set. By Proposition 17.2.21 we know that there
is an elliptic pseudodifferential operator 𝑩ℭ ∈ Ψa (Ω′, ℭ) such that 𝑰 − 𝑩ℭ 𝑨ℭ ∈
Ψa−𝜔 (Ω′, ℭ). By Theorem 17.3.10 we derive that if 𝑢 ∈ E ′ (Ω′) then 𝑊 𝐹a (𝑢) =
𝑊 𝐹a (𝑩ℭ 𝑨ℭ 𝑢) ⊂ 𝑊 𝐹a ( 𝑨ℭ 𝑢) ⊂ 𝑊 𝐹a (𝑢). □
Corollary 17.3.12 If 𝑃 (𝑥, D) is an elliptic differential operator with analytic coef-
ficients in Ω and if 𝑢 ∈ E ′ (Ω) then 𝑊 𝐹a (𝑃 (𝑥, D) 𝑢) = 𝑊 𝐹a (𝑢).
Corollary 17.3.12 is a refinement of Theorem 4.1.2.
Corollary 17.3.13 Let U be a conic open subset of Ω × ℭ. If 𝑨ℭ ∈ Ψa (Ω, ℭ) is
elliptic and if 𝑢 ∈ E ′ (Ω) is such that 𝑊 𝐹a ( 𝑨ℭ 𝑢) ∩ U = ∅ then 𝑊 𝐹a (𝑢) ∩ U = ∅.

17.4 Action on Singularity Hyperfunctions

17.4.1 Action on analytic functionals

As before Ω is an open subset of R𝑛 and ℭ an open cone in R𝑛 \ {0}; 𝑎 (𝑥, 𝜉) will

be a pseudoanalytic symbol of order 𝑚 in Ω × ℭ (Definition 17.1.4), extendible as a
C ∞ function 𝑎 (𝑧, 𝜉) in ΩC × ℭ holomorphic with respect to 𝑧, with ΩC open in C𝑛 ,
Ω ⊂ ΩC ∩ R𝑛 . The distribution kernel 𝐴ℭ (𝑥, 𝑦) corresponding to Op𝑎 = 𝑎 (𝑥, D) is
given by the oscillatory integral [cf. (17.1.16)]
∫
𝐴ℭ (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 𝑎 (𝑥, 𝜉) d𝜉. (17.4.1)
ℭ

We introduce an open subset Ω′C of C𝑛 such that Ω′C ⊂⊂ ΩC and we define

Ω′ = Ω′C ∩ R𝑛 ; 𝜒Ω′ will denote the characteristic function of Ω′. We avail ourselves
of Proposition 16.2.2: given an arbitrary ℎ ∈ O (C𝑛 ), the distribution
∫
𝑨ℭ,Ω′ ℎ (𝑦) = 𝐴ℭ (𝑥, 𝑦) ℎ (𝑥) d𝑥 (17.4.2)
Ω′
−𝑚 (R𝑛 ).
belongs to 𝐻loc
Lemma 17.4.1 To every compact subset 𝐾 of Ω′ there is an open subset 𝑉 C of Ω′C
such that 𝐾 ⊂ 𝑉 = 𝑉 C ∩ R𝑛 ⊂⊂ Ω′ and the following property holds:
⊛ Given an arbitrary ℎ ∈ O (C𝑛 ), the restriction of (17.4.2) to 𝑉 is equal to the
restriction of a holomorphic function 𝑨ℭ,Ω′ ℎ in 𝑉 C such that

sup 𝑨ℭ,Ω′ ℎ ≤ 𝐶sup |ℎ| (17.4.3)

𝑉C Ω′C

for some 𝐶 > 0 independent of ℎ.

17.4 Action on Singularity Hyperfunctions 669

Proof It is convenient to deal with an approximation of 𝐴ℭ (𝑥, 𝑦),

∫
2
𝐴ℭ( 𝜀) (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜀 | 𝜉 | 𝑎 (𝑥, 𝜉) d𝜉, 𝜀 > 0. (17.4.4)
ℭ

It is obvious that 𝐴ℭ( 𝜀) can be extended as a holomorphic function 𝐴ℭ( 𝜀) (𝑧, 𝑤) in

ΩC × C𝑛 . As 𝜀 ↘ 0 the function 𝐴ℭ( 𝜀) (𝑥, 𝑦) converges to 𝐴ℭ (𝑥, 𝑦) in D ′ (Ω × R𝑛 ).
We shall avail ourselves of Proposition 17.2.1: there are positive constants 𝑅◦ , 𝐶◦ ,
𝐶1 , 𝐶2 , such that, given any number 𝑅 ≥ 𝑅◦ , 𝑎 (𝑥, 𝜉) can be asymptotically extended
as a C ∞ function e 𝑎 𝑅 (𝑧, 𝜁) in Ω′C × (ℭ + 𝑖R𝑛 ), holomorphic with respect to 𝑧 and
satisfying the following estimates, for all (𝑧, 𝜁) ∈ Ω′C × (ℭ + 𝑖R𝑛 ),

𝑎 𝑅 (𝑧, 𝜁)| ≤ 𝐶◦ (1 + |Re 𝜁 |) 𝑚 exp (𝐶1 |Im 𝜁 | /𝑅) ;

|e (17.4.5)
𝑎 𝑅 (𝑧, 𝜁) ≤ 𝐶2 (1 + |Re 𝜁 |) 𝑚 exp ((𝐶1 |Im 𝜁 | − |Re 𝜁 |) /𝑅) .
𝜕𝜁 e (17.4.6)

Let 𝜒 ∈ Cc∞ (Ω′), 0 ≤ 𝜒 ≤ 1, 𝜒 (𝑥) = 1 if 𝑥 ∈ 𝑈, 𝑈 open, 𝐾 ⊂ 𝑈. Let 𝜆◦ > 0 be

sufficiently small that supp 𝜒 + 𝑖𝜆◦ 𝔅1(𝑛) ⊂⊂ Ω′C (𝔅1(𝑛) : open unit ball in R𝑛 ). By the
Cauchy Integral Theorem we have, for every ℎ ∈ O (C𝑛 ) and every 𝜆 ∈ [0, 𝜆◦ ],
∫
(2𝜋) 𝑛
𝐴ℭ( 𝜀) (𝑥, 𝑦) ℎ (𝑥) d𝑥 (17.4.7)
Ω′
∫ ∫
2 𝜉 𝜉
= e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜆𝜒 ( 𝑥) | 𝜉 |−𝜀 | 𝜉 | 𝑎 𝑥 + 𝑖𝜆 𝜒 (𝑥) , 𝜉 ℎ 𝑥 + 𝑖𝜆 𝜒 (𝑥) d𝑥d𝜉.
ℭ Ω ′ |𝜉 | |𝜉 |

In the right-hand side of (17.4.7) we deform the domain of 𝜉-integration from ℭ to

the image ℭ (𝑡◦ ) of ℭ under the map 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝑡◦ |𝜉 | (𝑥 − 𝑦), where 0 < 𝑡◦ ≪ 1.
We have
𝑛
∑︁ √
⟨𝜁⟩ 2 − |𝜉 | 2 = 𝜁 2𝑗 − 𝜉 2𝑗 ≤ 5𝑡 ◦ |𝜉 | 2 |𝑥 − 𝑦| 2 .
𝑗=1

We select 𝑡◦ sufficiently small that 𝑡◦ (diam Ω′) 2 ≪ 1, whence

2
𝜉
∀𝑡 ∈ [0, 𝑡◦ ], + 𝑖𝑡 (𝑥 − 𝑦) − 1 ≪ 1. (17.4.8)
|𝜉 |

It follows that we can use the main branch of the square-root ⟨𝜁⟩ of ⟨𝜁⟩ 2 .
Stokes’ theorem implies
∫
𝐴ℭ( 𝜀) (𝑥, 𝑦) ℎ (𝑥) d𝑥 = 𝐼1𝜀 (𝑦, 𝜆) + 𝐼2𝜀 (𝑦, 𝜆) , (17.4.9)
Ω′

where
670 17 Analytic Pseudodifferential Calculus

𝐼1𝜀 (𝑦, 𝜆)
∫ ∫
2 𝜁 𝜁
= e𝑖Φ( 𝑥,𝑦, 𝜉 ,𝑡◦ )−𝜀 ⟨𝜁 ⟩ e𝑎 𝑅 𝑥 + 𝑖𝜆 𝜒 (𝑥) , 𝜁 ℎ 𝑥 + 𝑖𝜆 𝜒 (𝑥) d𝑥d𝜁,
Ω′ ℭ ⟨𝜁⟩ ⟨𝜁⟩

𝐼2𝜀 (𝑦, 𝜆)
∫ ∫ 𝑡◦ ∫ 2
𝜁

= e𝑖Φ( 𝑥,𝑦, 𝜉 ,𝑡)−𝜀 ⟨𝜁 ⟩ 𝜕 𝜁 e
𝑎 𝑅 𝑥 + 𝑖𝜆 𝜒 (𝑥) ,𝜁
Ω′ 0 ℭ ⟨𝜁⟩

𝜁
×ℎ 𝑥 + 𝑖𝜆 𝜒 (𝑥) d𝑥d𝜁d𝑡
⟨𝜁⟩

with 𝜉 + 𝑖𝑡 ◦ |𝜉 | (𝑥 − 𝑦) and

Φ (𝑥, 𝑦, 𝜉, 𝑡) = (𝑥 − 𝑦) · 𝜉 + 𝑖 𝜆 𝜒 (𝑥) ⟨𝜁⟩ + 𝑡 |𝜉 | |𝑥 − 𝑦| 2 .

By (17.4.8) we can assume, for 𝑘 = 1, 2, all 𝑥, 𝑦 ∈ Ω′, all 𝜉 ∈ ℭ,

Im Φ (𝑥, 𝑦, 𝜉, 𝑡) + 𝜀 Re ⟨𝜉 + 𝑖𝑡 |𝜉 | (𝑥 − 𝑦)⟩ 2
1 1
≥ 𝜆 𝜒 (𝑥) |𝜉 | + 𝑡 |𝜉 | |𝑥 − 𝑦| 2 + 𝜀 |𝜉 | 2 .
2 2
We restrict the variation of 𝑦 to 𝐾 ⊂ 𝑈; since either 𝑥 ∈ 𝑈, in which case 𝜒 (𝑥) = 1,
or else |𝑥 − 𝑦| ≥ 𝑑 = dist (𝐾, R𝑛 \𝑈) > 0, we derive from the preceding inequality:

1
Im Φ (𝑥, 𝑦, 𝜉, 𝑡) + 𝜀 Re ⟨𝜉 + 𝑖𝑡 |𝜉 | (𝑥 − 𝑦)⟩ 2 ≥ |𝜉 | min 𝜆, 𝑡𝑑 2 (17.4.10)
2

for all 𝑥 ∈ Ω′, 𝑦 ∈ 𝐾, 𝜆 ∈ [0, 𝜆◦ ], 𝑡 ∈ [0, 𝑡◦ ].

Recalling that 𝑧 = 𝑥 + 𝑖𝜆 𝜒 (𝑥) ⟨𝜁𝜁 ⟩ ∈ Ω′C for all 𝑥 ∈ Ω′, 𝜆 ∈ [0, 𝜆◦ ], we estimate
the integral 𝐼1𝜀 (𝑦, 𝜆) in the right-hand side of (17.4.9) by exploiting (17.4.5):

𝜁
𝑎𝑅
e 𝑥 + 𝑖𝜆 𝜒 (𝑥) ,𝜁 ≤ 𝐶◦ (1 + |Re 𝜁 |) 𝑚 exp (𝐶1 |Im 𝜁 | /𝑅) ,
⟨𝜁⟩

where 𝜁 = 𝜉 + 𝑖𝑡◦ |𝜉 | (𝑥 − 𝑦). Thus

𝜁 𝜁
𝑎 𝑅 𝑥 + 𝑖𝜆 𝜒 (𝑥)
e , 𝜁 ℎ 𝑥 + 𝑖𝜆 𝜒 (𝑥) (17.4.11)
⟨𝜁⟩ ⟨𝜁⟩

≤ 𝐶◦ sup |ℎ| (1 + |𝜉 |) 𝑚 exp (𝐶1 𝑡◦ |𝑥 − 𝑦| |𝜉 | /𝑅) .
Ω′C

We select 𝑅 sufficiently large that 𝐶1 𝑅 −1 < 𝑐 ◦ < min 21 𝜆, 𝑡 ◦ 𝑑 (recall that
𝑡◦ diam Ω′ ≪ 1). By combining (17.4.10) and (17.4.11) we see that the absolute
value of the integrand in 𝐼1𝜀 (𝑦, 𝜆) does not exceed a constant times exp (−𝑐 ◦ |𝜉 |) and
17.4 Action on Singularity Hyperfunctions 671

this remains true if we let 𝑦 = 𝑤 vary in an open subset 𝑉 C of C𝑛 , 𝐾 ⊂ 𝑉 C ⊂⊂ Ω′C .

This allows us to let 𝜀 go to zero, in which case 𝐼1𝜀 (𝑤, 𝜆) converges in O 𝑉 C to a
holomorphic function 𝐼10 (𝑤, 𝜆) that satisfies

sup 𝐼10 (𝑤, 𝜆) ≤ 𝐶sup |ℎ|

𝑤 ∈𝑉 C Ω′C

for some 𝐶 > 0 independent of ℎ and of 𝜆 ∈ [0, 𝜆◦ ].

In dealing with 𝐼2𝜀 (𝑦, 𝜆) we exploit (17.4.6) where now 𝜁 = 𝜉 + 𝑖𝑡 |𝜉 | (𝑥 − 𝑦),
0 ≤ 𝑡 ≤ 𝑡◦ :

𝜁
𝑎 𝑅 𝑥 + 𝑖𝜆 𝜒 (𝑥)
𝜕𝜁 e , 𝜁 ≤ 𝐶2 (1 + |𝜉 |) 𝑚 exp (− (1 − 𝐶1 𝑡◦ ) |𝜉 | /𝑅) .
⟨𝜁⟩

Here we require 𝑡 ◦ < (2𝐶1 ) −1 , whence

𝜁 𝜁
ℎ 𝑥 + 𝑖𝜆 𝜒 (𝑥) 𝑎 𝑅 𝑥 + 𝑖𝜆 𝜒 (𝑥)
𝜕𝜁 e ,𝜁
⟨𝜁⟩ ⟨𝜁⟩

1
≤ 𝐶◦ sup |ℎ| (1 + |𝜉 |) 𝑚 exp − |𝜉 | .
Ω′C 2𝑅

Noting that since 𝑡 can vanish so can the right-hand side in (17.4.10), we see that
the absolute value of the integrand in 𝐼2𝜀 (𝑦, 𝜆) does not exceed a constant times

1
exp − 4𝑅 |𝜉 | and this too remains true if we let 𝑦 vary in a neighborhood of 𝐾 in
C𝑛 ; we can take 𝑉 C to be such a neighborhood. This allows us to let 𝜀 go to zero, in
which case 𝐼2𝜀 (𝑤, 𝜆) converges in O 𝑉 C to a holomorphic function 𝐼20 (𝑤, 𝜆) that

satisfies
sup 𝐼20 (𝑤, 𝜆) ≤ 𝐶 ′sup |ℎ|
𝑤 ∈𝑉 C Ω′C

for some constant 𝐶 ′ > 0 independent of ℎ. We fix 𝜆 > 0 suitably small and conclude
that 𝐼1𝜀 (𝑤, 𝜆) + 𝐼2𝜀 (𝑤, 𝜆) converges in O 𝑉 C as 𝜀 ↘ 0 to a holomorphic function,
denoted by 𝑨ℭ,Ω′ ℎ in the statement, satisfying (17.4.3). □
It makes sense to use the notation
∫
𝑨ℭ,Ω′ ℎ (𝑤) = 𝐴ℭ (𝑥, 𝑤) ℎ (𝑥) d𝑥, 𝑤 ∈ 𝑉 C . (17.4.12)
Ω′

Recall that O ′ (𝐾) is the set of analytic functionals in C𝑛 carried by the compact
set 𝐾 ⊂ R𝑛 (see Definition 6.1.3, also Corollary 6.2.15). Taking 𝑉 C ⊃ 𝐾 in (17.4.12)
yields the following immediate consequence of Lemma 17.4.1.

Corollary 17.4.2 Let 𝐾 be a compact subset of Ω′, Ω′C an open subset of C𝑛 such
that Ω′C ⊂⊂ ΩC and Ω′ = Ω′C ∩ R𝑛 ; let 𝑨ℭ,Ω′ ℎ be as in (17.4.12). If 𝜇 ∈ O ′ (𝐾)
then O (C𝑛 ) ∋ ℎ ↦→ 𝜇, 𝑨ℭ,Ω′ ℎ defines an analytic functional in C𝑛 (henceforth
denoted by 𝑨⊤
ℭ,Ω′ 𝜇 ) carried by Ω .
′C
672 17 Analytic Pseudodifferential Calculus

Actually, in the preceding statement Ω′C can be replaced by Ω′ ⊂⊂ R𝑛 , as we

show now.

Lemma 17.4.3 Let Ω′, Ω′C , 𝐾, be as in Corollary 17.4.2. If 𝜇 ∈ O ′ (𝐾) then the
analytic functional 𝑨⊤ ′
ℭ,Ω′ 𝜇 is carried by Ω .

Proof Let an open subset Ω′′C of C𝑛 have the same properties as Ω′C : 𝐾 ⊂ Ω′′ =
R𝑛 ∩ Ω′′C and Ω′′C ⊂⊂ ΩC . We have
∫
𝑨ℭ,Ω′ ℎ (𝑦) − 𝑨ℭ,Ω′′ ℎ (𝑦) = 𝐴ℭ (𝑥, 𝑦) ℎ (𝑥) d𝑥,
𝐾′

where 𝐾 ′ is the closure of Ω′\Ω′′. The distance between 𝐾 and 𝐾 ′ is positive and
𝐴ℭ (𝑥, 𝑦) is an analytic function off the diagonal of Ω × Ω (Theorem 17.1.14). It
follows that there is an open subset 𝑊 C of ΩC such that

𝐾 ′ ⊂ 𝑊 = 𝑊 C ∩ R𝑛 , 𝐾 ∩ 𝑊 C = ∅,
∫
to which 𝑦 ↦→ 𝐾 ′ 𝐴ℭ (𝑥, 𝑦) ℎ (𝑥) d𝑥 extends as a holomorphic function. It follows
that if 𝜇 ∈ O ′ (C𝑛 ) is carried by 𝐾 we have

𝜇, 𝑨ℭ,Ω′ ℎ − 𝜇, 𝑨ℭ,Ω′′ ℎ ≤ 𝐶 ′max |ℎ| (17.4.13)

𝐾

for some 𝐶 ′ > 0 dependent on 𝑎 (𝑥, 𝑦, 𝜉) and 𝜇 but not on ℎ. We apply Corollary
17.4.2 with Ω′ replaced by Ω′′: 𝑨⊤ ℭ,Ω′′ 𝜇 is carried by Ω
′′C ; since 𝐾 ⊂ Ω ′′ the same

is true of 𝑨⊤
ℭ,Ω′ 𝜇. By letting Ω
′′C contract to 𝐾 we reach the conclusion that the

analytic functional 𝑨⊤ ′
ℭ,Ω′ 𝜇 is carried by Ω . □
The Laplace–Borel transform L𝜇 (Definition 6.2.1) provides a different interpre-
⊤ . From Theorem 6.2.2, and the fact that supp 𝜇 ⊂ 𝐾 ⊂ R𝑛
tation of the operator 𝐴ℭ,Ω′
we deduce that to every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that

∀𝜉 ∈ R𝑛 , |L𝜇 (𝑖𝜉)| ≤ 𝐶 𝜀 e 𝜀 | 𝜉 | . (17.4.14)

This shows that, for each 𝜀 > 0 fixed and 𝑧 in an arbitrary compact subset of ΩC , the
integral on the right in
D E ∫
2
𝜇 𝑤 , 𝐴ℭ( 𝜀) (𝑧, 𝑤) = (2𝜋) −𝑛 e𝑖𝑧· 𝜉 −𝜀 | 𝜉 | 𝑎 (𝑧, 𝜉) L𝜇 (𝑖𝜉) d𝜉 (17.4.15)
ℭ

is absolutely convergent and defines a holomorphic function of 𝑧 in ΩC . It follows

that
D E
𝑨⊤
ℭ,Ω′ 𝜇, ℎ = 𝜇, 𝑨ℭ,Ω ℎ
′
∫ ∫
2
−𝑛
= (2𝜋) lim e𝑖 𝑥· 𝜉 −𝜀 | 𝜉 | 𝑎 (𝑥, 𝜉) ℎ (𝑥) L𝜇 (𝑖𝜉) d𝑥d𝜉,
𝜀↘0 ℭ Ω′
17.4 Action on Singularity Hyperfunctions 673

i.e.,
∫
2
−𝑛
𝑨⊤
ℭ,Ω′ 𝜇 = (2𝜋) 𝜒Ω′ (𝑥) lim e𝑖 𝑥· 𝜉 −𝜀 | 𝜉 | 𝑎 (𝑥, 𝜉) L𝜇 (𝑖𝜉) d𝜉. (17.4.16)
𝜀↘0 ℭ

We can abbreviate (17.4.16) by the formula

∫ D E
−𝑛
𝑨⊤ℭ,Ω′ 𝜇 = (2𝜋) 𝜒Ω′ (𝑥) 𝜇 𝑤 , e𝑖 ( 𝑥−𝑤) · 𝜉 𝑎 (𝑥, 𝜉) d𝜉, (17.4.17)
ℭ

an equality between analytic functionals in R𝑛 carried by Ω′ (𝜒Ω′ : the characteristic

function of Ω′).

Remark 17.4.4 Formula (17.4.17) does not mean that the right-hand side is an
𝐿 c1 -function of 𝑥; it is merely short-hand for the right-hand side in (17.4.16).

17.4.2 Action on singularity hyperfunctions

The argument in the preceding subsection shows that the analytic functional 𝑨⊤ ℭ,Ω′ 𝜇
is the limit, inDthe sense of O ′ (C
E
𝑛 ), as 𝜀 ↘ 0, of the compactly supported functions

in Ω, 𝜒Ω′ (𝑥) 𝜇 𝑤 , 𝐴ℭ( 𝜀) (𝑥, 𝑤) [cf. (17.4.4)] viewed as analytic functionals. These
functions are analytic in Ω′, which leads us to the next type of results.
We refer the reader to Section 7.1: 𝑨⊤ ℭ,Ω′ 𝜇 defines a hyperfunction 𝑓 𝜇, 𝐴,ℭ,Ω in Ω
′

(Definition 7.1.1).

Lemma 17.4.5 Let 𝐾 and Ω′ be as in Lemma 17.4.1 and let 𝜇 ∈ O ′ (𝐾). The
restriction to Ω′\𝐾 of the hyperfunction 𝑓 𝜇, 𝐴,ℭ,Ω′ defined by 𝑨⊤
ℭ,Ω′ 𝜇 [cf.(17.4.16)]
is a C 𝜔 function.

Proof Let 𝑊 ⊂⊂ Ω′\𝐾 be an open subset of ℭ. We know that the restriction of
𝐴ℭ (𝑥, 𝑦) to 𝑊 × Ω′\𝑊 is a C 𝜔 function (Theorem 17.1.14); it follows that there
is an open subset 𝑉 C of C𝑛 such that
(1) Ω′\𝑊 ⊂ 𝑉 = 𝑉 C ∩ R𝑛 ⊂ Ω and 𝑉 ∩ 𝑊 = ∅;
(2) for each 𝑥 ∈ 𝑊, 𝑦 ↦→ 𝐴ℭ (𝑥, 𝑦) can be extended as a holomorphic function
𝑤 ↦→ 𝐴ℭ (𝑥, 𝑤) in 𝑉 C;
(3) given any 𝜈 ∈ O ′ 𝑉 C , 𝑥 ↦→ ⟨𝜈 𝑤 , 𝐴ℭ (𝑥, 𝑤)⟩ is a C 𝜔 function in 𝑊.

We have 𝐾 ⊂ 𝑉 and therefore the restriction of ⟨𝜇 𝑤 , 𝐴ℭ (𝑥, 𝑤)⟩ to 𝑊 is easily

seen to be equal to that of 𝑓 𝜇, 𝐴,ℭ,Ω′ . □

Lemma 17.4.6 Let Ω′ be an open subset of R𝑛 , 𝐾 be a compact subset of Ω′ and let

𝜇 ∈ O ′ (𝐾). Let 𝑈 be an open subset of R𝑛 contained in 𝐾. The restriction to 𝑈 of
the hyperfunction 𝑓 𝜇, 𝐴,ℭ,Ω′ in Ω defined by 𝑨⊤ ′
ℭ,Ω′ 𝜇 is independent of Ω .
674 17 Analytic Pseudodifferential Calculus

Proof The analogue of Lemma 17.4.3 is valid for Ω′′ ⊃ 𝐾 in the place of Ω′,
yielding a hyperfunction 𝑓 𝜇, 𝐴,ℭ,Ω′′ in Ω. By (17.4.13) we know that the analytic
functional
O (C𝑛 ) ∋ ℎ ↦→ 𝜇, 𝑨ℭ,Ω′ ℎ − 𝜇, 𝑨ℭ,Ω′′ ℎ
is carried by the closure of (Ω′ ∪ Ω′′) \ (Ω′ ∩ Ω′′). From this we derive that
𝑨⊤ ⊤
ℭ,Ω′ 𝜇 = 𝑨ℭ,Ω′′ 𝜇 + 𝜈, supp 𝜈 ⊂ R \𝑈. The restriction to 𝑈 of the hyperfunction in
𝑛

Ω defined by 𝜈 vanishes identically (Proposition 7.1.4). □

Let us denote by 𝑓 𝜇, 𝐴,ℭ 𝑈 the restriction of 𝑓 𝜇, 𝐴,ℭ,Ω′ to 𝑈 ⊂ 𝐾.

Lemma 17.4.7 Let Ω′, 𝐾 ⊂⊂ Ω′, 𝜇 ∈ O ′ (𝐾) and 𝑈 ⊂ 𝐾 be as in Lemma 17.4.6. If

the hyperfunction 𝑢 represented by 𝜇 in 𝑈 is an analytic function the same is true of
𝑓 𝜇, 𝐴,ℭ 𝑈 .

Proof Let 𝜒𝑉 stand for the characteristic function of an open set 𝑉 ⊂⊂ 𝑈; we

view the function 𝜒𝑉 𝑢 as an analytic functional in C𝑛 carried by 𝑉. According to
Theorems 6.3.6, 6.3.11, we can write 𝜇 = 𝜇˜ + 𝜒𝑉 𝑢 with 𝜇˜ ∈ O ′ (𝐾\𝑉). It is readily
checked that
˜ 𝐴,ℭ 𝑈 + 𝑎 (𝑥, D) ( 𝜒𝑉 𝑢) .
𝑓 𝜇, 𝐴,ℭ 𝑈 = 𝑓 𝜇,

˜ 𝐴,ℭ 𝑈 is a C function, by Lemma 17.4.5; the same is true

The restriction to 𝑉 of 𝑓 𝜇, 𝜔

of the restriction to 𝑉 of 𝑎 (𝑥, D) ( 𝜒𝑉 𝑢), by Theorem 17.2.13. □

We apply what precedes to the case 𝐾 = 𝑈 where 𝑈 is an arbitrary relatively
compact open subset of Ω. Let 𝑢 ∈ B (𝑈) and 𝜇 ∈ O ′ 𝑈 represent 𝑢 in 𝑈. If

𝜇1 ∈ O ′ 𝑈 is another representative of 𝑢 in 𝑈 we have 𝜇1 − 𝜇 ∈ O ′ (𝜕𝑈) whence,
by Lemma 17.4.7, 𝑓 𝜇1 , 𝐴,ℭ 𝑈 − 𝑓 𝜇, 𝐴,ℭ 𝑈 ∈ C 𝜔 (𝑈). This property suggests that we
look at the singularity hyperfunction in 𝑈 (Definition 11.3.4) defined by 𝑓 𝜇, 𝐴,ℭ , i.e.,
the coset in B sing (𝑈) that contains 𝑓 𝜇, 𝐴,ℭ 𝑈 [denoted by 𝑓 𝜇, 𝐴,ℭ
♮
in the sequel].
𝑈
By Lemma 17.4.7 we know that if 𝑢 ∈ C 𝜔 (𝑈) then 𝑓 𝜇, 𝐴,ℭ 𝑈 ∈ C 𝜔 (𝑈). Thus
we have defined a bona fide linear map of B sing (𝑈) into itself, assigning the coset
𝑓 𝜇, 𝐴,ℭ to 𝑢 ♮ ∈ B sing (𝑈) (represented by 𝑢, itself represented by 𝜇); we define
♮
𝑈
♮
𝑎 (𝑥, D) 𝑢 ♮ = 𝑓 𝜇, 𝐴,ℭ . This defines the action of the operator 𝑎 (𝑥, D) ∈ Ψa (Ω, ℭ)
𝑈
on the singularity hyperfunctions in Ω; Ψa (Ω, ℭ) can be regarded as a subring of
the ring (with respect to ordinary addition and composition) of endomorphisms of
B sing (𝑈). The following inclusion is easily proved:

∀𝑢 ♮ ∈ B sing (Ω) , supp 𝑎 (𝑥, D) 𝑢 ♮ ⊂ supp 𝑢 ♮ . (17.4.18)

In other words, when acting on singularity hyperfunctions analytic pseudodifferential

operators are local, like differential operators.
17.4 Action on Singularity Hyperfunctions 675

Remark 17.4.8 We have not defined the action of 𝑎 (𝑥, D) on hyperfunctions but only
′
on singularity hyperfunctions. This is due to the fact that if 𝜇1 , 𝜇2 ∈ O 𝑈 are such
that 𝜇1 − 𝜇2 ∈ O ′ (𝜕𝑈) then, by Lemma 17.4.7, 𝑓 𝜇1 , 𝐴,ℭ 𝑈 − 𝑓 𝜇2 , 𝐴,ℭ 𝑈 ∈ C 𝜔 (𝑈),
but, generally speaking, 𝑓 𝜇1 , 𝐴,ℭ 𝑈 ≠ 𝑓 𝜇2 , 𝐴,ℭ 𝑈 .

Theorem 17.4.9 If the analytic pseudodifferential operator 𝑎 (𝑥, D) in Ω is ana-

lytic regularizing (Definition 17.1.17) then 𝑎 (𝑥, D) 𝑓 ♮ = 0 for every singularity
hyperfunction 𝑓 ♮ in Ω.

Proof If 𝑎 (𝑥, D) is analytic regularizing there is an open subset ΩC of C𝑛 such

that the distribution kernel (17.4.1) can be extended as a holomorphic function
𝐴ℭ (𝑧, 𝑤) in ΩC × ΩC . If 𝜇 ∈ O ′ (R𝑛 ) is carried by a compact subset of Ω then
⟨𝜇 𝑤 , 𝐴ℭ (𝑧, 𝑤)⟩ ∈ O ΩC , implying directly the claim. □

17.4.3 Analytic pseudodifferential operators decrease the analytic

wave-front set of singularity hyperfunctions

We continue to reason in the set-up of the preceding subsections: Ω (resp., ΩC ) is an

open subset of R𝑛 (resp., C𝑛 ); Ω ⊂ ΩC ∩ R𝑛 ; ℭ is an open cone in 𝜉-space R𝑛 \ {0}.
The symbol 𝑎 (𝑥, 𝜉) ∈ 𝑆 𝑚 ∞
𝜓a (Ω × ℭ) extends as a C function 𝑎 (𝑧, 𝜉) in Ω × ℭ,
C

holomorphic with respect to 𝑧 and satisfying the inequalities (17.2.5). We also need
the open subset Ω′C of C𝑛 , Ω′C ⊂⊂ ΩC ; we now assume Ω′ ⊂⊂ Ω′C ∩ R𝑛 . As in
the proof of Lemma 17.4.1 we shall avail ourselves of Proposition 17.2.1: Given any
sufficiently large positive number 𝑅, 𝑎 (𝑥, 𝜉) can be asymptotically extended as a C ∞
function e𝑎 𝑅 (𝑧, 𝜁) in Ω′C × (ℭ + 𝑖R𝑛 ), holomorphic with respect to 𝑧 and satisfying
(17.4.5)–(17.4.6). In what follows (𝑥 ◦ , 𝜉 ◦ ) is an arbitrary point of Ω′ × ℭ.
This subsection is entirely devoted to the proof of the following statement.

Theorem 17.4.10 We have 𝑊 𝐹a 𝑎(𝑥, 𝐷)𝑢 ♮ ⊂ 𝑊 𝐹a 𝑢 ♮ for all 𝑢 ♮ ∈ B sing (Ω).

The singularity hyperfunction 𝑎(𝑥, 𝐷)𝑢 ♮ is defined in the preceding subsection.

We will prove Theorem 17.4.10 by focusing on a hyperfunction 𝑢, the boundary
value in Ω′ of a holomorphic function 𝑓 (Definition 7.2.3) in a wedge

W𝛿 (Ω′, Γ) = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; 𝑥 ∈ Ω′, 𝑦 ∈ Γ, |𝑦| < 𝛿} ,

with Γ ⊂ R𝑛 \ {0} a convex open cone; 𝛿 > 0 is sufficiently small that W𝛿 (Ω′, Γ) ⊂⊂
Ω′C .
676 17 Analytic Pseudodifferential Calculus

17.4.3.1 Preparatory results

We assume that 𝑢| Ω′ = 𝑏 Ω′ 𝑓 is represented

in a neighborhood 𝑈 ⊂⊂ Ω′ of 𝑥 ◦ in R𝑛
′
by an analytic functional 𝜇 ∈ O 𝑈 which is the “limit” as Γ ∋ v ↘ 0, in the sense
of Lemma 7.2.2, of the functionals
∫
𝜇𝑈𝑓,v : O (C𝑛 ) ∋ 𝜑 ↦→ 𝑓 (𝑠 + 𝑖v) 𝜑 (𝑠 + 𝑖v) d𝑠.
𝑈

⊤ 𝜇𝑈 [see (17.4.17)]. Actually, for technical reasons, when

We propose to study 𝐴ℭ,Ω ′ 𝑓 ,v
ℭ ≠ R𝑛 \ {0} we must replace ℭ by a smaller open cone ℭ ′ in R𝑛 \ {0}; we require
𝜉 ◦ ∈ ℭ ′ ∩ S𝑛−1 ⊂⊂ ℭ; we take 𝜕ℭ ′∩S𝑛−1 to be the intersection of S𝑛−1 with a
hyperplane in R𝑛 and diam ℭ ′∩S𝑛−1 as small as needed. By Lemma 17.4.3 the
analytic functional 𝑨⊤ ′
ℭ′ ,Ω′ 𝜇 𝑓 ,v is carried by Ω and the following duality bracket [cf.
𝑈

(17.4.12), (17.4.16)] makes sense:

D E
(2𝜋) 𝑛 𝜇𝑈𝑓,v , 𝐴ℭ′ ,Ω′ ℎ (17.4.19)
∫ ∫
2
= lim 𝜇𝑈𝑓,v , e𝑖 ( 𝑥−𝑤) · 𝜉 −𝜀 | 𝜉 | 𝑎 (𝑥, 𝜉) ℎ (𝑥) d𝑥d𝜉
𝜀↘0 ′
∫ ∫ ℭ∫ 𝑈
2
= lim e𝑖 ( 𝑥−𝑠−𝑖v) · 𝜉 −𝜀 | 𝜉 | 𝑓 (𝑠 + 𝑖v) 𝑎 (𝑥, 𝜉) ℎ (𝑥) d𝑥d𝜉d𝑠.
𝜀↘0 𝑠 ∈𝑈 ℭ′ 𝑥 ∈𝑈

The first step is to deform the domain of 𝜉-integration from ℭ ′ to the image of ℭ ′
under the map 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜅 |𝜉 | (𝑧 − 𝑠 − 𝑖v). With this meaning of 𝜁 we define
∫ ∫
2
𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v) = e𝑖 (𝑧−𝑠−𝑖v) ·𝜁 −𝜀 ⟨𝜁 ⟩ 𝑓 (𝑠 + 𝑖v) e
𝑎 𝑅 (𝑧, 𝜁) d𝜁d𝑠 (17.4.20)
𝑠 ∈𝑈 𝜉 ∈ℭ′

D𝜁 D𝜁
where d𝜁 = D 𝜉 d𝜉, D 𝜉 = Δ 𝜅 (𝑧 − 𝑠 − 𝑖v, 𝜉), the Jacobian determinant of 𝜁 with
respect to 𝜉 [cf. (3.4.4)]. We select the positive numbers 𝛿 and 𝜅 sufficiently small
Re 𝜁 𝑛−1 if 𝑧 ∈ Ω ′C , 𝑠 + 𝑖v ∈ W (𝑈, Γ),
that |Re 𝜁 | stays in a compact subset of ℭ ∩ S 𝛿
𝜉 ∈ ℭ ′. (This is where we need ℭ ′ ∩ S𝑛−1 ⊂⊂ ℭ ∩ S𝑛−1 .) In this notation,
D E ∫
𝜇𝑈𝑓,v , 𝐴ℭ′ ,Ω′ ℎ = lim (2𝜋) −𝑛 𝐽ℭ𝜀′ ,𝑅,0 (𝑥, v) ℎ (𝑥) d𝑥. (17.4.21)
𝜀↘0 𝑥 ∈𝑈

We derive from (17.4.5)

𝑎 𝑅 (𝑧, 𝜁)| ≤ 𝐶◦ (1 + |𝜉 |) 𝑚 exp (𝐶1 𝜅 |𝜉 | |𝑥 − 𝑠| /𝑅)

|e (17.4.22)

while (17.4.6) yields

𝑎 𝑅 (𝑧, 𝜁) ≤ 𝐶2 (1 + |𝜉 |) 𝑚 exp (− (1 − 𝐶1 𝜅) |𝜉 | /𝑅)

𝜕𝜁 e (17.4.23)

for suitably large 𝐶◦ , 𝐶1 , 𝐶2 .

17.4 Action on Singularity Hyperfunctions 677

Crucial in the ongoing argument are estimates for

2 2
e𝑖 (𝑧−𝑠−𝑖v) ·𝜁 −𝜀 ⟨𝜁 ⟩ e
𝑎 𝑅 (𝑧, 𝜁) and e𝑖 (𝑧−𝑠−𝑖v) ·𝜁 −𝜀 ⟨𝜁 ⟩ 𝜕 𝜁 e
𝑎 𝑅 (𝑧, 𝜁) ;

in other words, lower bounds for the real parts of the exponents,

𝐸 1 = Im (𝑧 − 𝑠 − 𝑖v) · 𝜁 − 𝜀 Re ⟨𝜁⟩ 2 + 𝐶1 𝜅 |𝜉 | |𝑥 − 𝑠| /𝑅,

𝐸 2 = Im (𝑧 − 𝑠 − 𝑖v) · 𝜁 − 𝜀 Re ⟨𝜁⟩ 2 − (1 − 𝐶1 𝜅) |𝜉 | /𝑅.

We have

Im ((𝑧 − 𝑠 − 𝑖v) · 𝜁)
= −𝜉 · Im (𝑧 − 𝑠 − 𝑖v) − 𝜅 |𝜉 | Re ⟨𝑧 − 𝑠 − 𝑖v⟩ 2

= − (𝑦 − v) · 𝜉 − 𝜅 |𝜉 | |𝑥 − 𝑠| 2 − |𝑦 − v| 2 ,

Re ⟨𝜁⟩ 2 = Re ⟨(𝜉 + 𝜅 |𝜉 | (v − 𝑦) + 𝑖𝜅 |𝜉 | (𝑥 − 𝑠))⟩ 2

2
𝜉
= |𝜉 | 2 + 𝜅 (v − 𝑦) − 𝜅 |𝜉 | 2 |𝑥 − 𝑠| 2 .
|𝜉 |

Throughout the reasoning we will have |v| < 𝛿 and bounds on |𝑥| 2 + |𝑦| 2 since
𝑧 ∈ Ω′C , on |𝑠| since 𝑠 ∈ 𝑈. We may therefore choose, right away, 𝜅 > 0 sufficiently
small that Re ⟨𝜁⟩ 2 ≥ |𝜉 | 2 /2, 𝐶1 𝜅 < 1/2. We obtain

𝐸 1 ≤ − (𝑦 − v) · 𝜉 − 𝜅 |𝜉 | |𝑥 − 𝑠| 2 − |𝑦 − v| 2 + |𝜉 | |𝑥 − 𝑠| /2𝑅 − 𝜀 |𝜉 | 2 /2,
(17.4.24)

𝐸 2 ≤ − (𝑦 − v) · 𝜉 − 𝜅 |𝜉 | |𝑥 − 𝑠| 2 − |𝑦 − v| 2 − |𝜉 | /2𝑅 − 𝜀 |𝜉 | 2 /2. (17.4.25)

We need to compare 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v) for different values of v.

Lemma 17.4.11 Suppose that the boundary of 𝑈 is smooth and let 𝑉 ⊂⊂ 𝑈 be an

open subset of R𝑛 . If 𝛿 and 𝑅 −1 are sufficiently small then, given any pair of vectors
v◦ , v∗ ∈ Γ, v∗ ≠ v◦ , max (|v◦ | , |v∗ |) < 𝛿, 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v∗ ) converges,
as 𝜀 ↘ 0, to a holomorphic function of 𝑧 in 𝑉 𝛿C = {𝑧 ∈ C𝑛 ; 𝑥 ∈ 𝑉, |𝑦| < 𝛿}.

Proof We use the notation

v (𝜏) = (1 − 𝜏) v◦ + 𝜏v∗ (∈ Γ) and

𝜁 = 𝜉 + 𝑖𝜅 |𝜉 | (𝑧 − 𝑠 − 𝑖v (𝜏)) ;

Stokes’ Theorem implies

678 17 Analytic Pseudodifferential Calculus

𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v∗ )

∫ 𝜏=1 ∫ ∫
2
= e𝑖 (𝑧−𝑠−𝑖v( 𝜏)) ·𝜁 −𝜀 ⟨𝜁 ⟩ 𝑓 (𝑠 + 𝑖v (𝜏)) e
𝑎 𝑅 (𝑧, 𝜁) d𝜁d𝑆𝑈 d𝜏
𝜏=0 𝑠 ∈𝜕𝑈 𝜉 ∈ℭ′

with the appropriate density d𝑆𝑈 on 𝜕𝑈 (in 𝑠-space). From (17.4.24) we derive that
we have, in the integrand of (17.4.20),

|𝜉 | −1 𝐸 1 ≤ −𝜅 |𝑥 − 𝑠| 2 + |𝜉 | |𝑥 − 𝑠| /2𝑅 + |𝑦 − v (𝜏)| (1 + 𝜅 |𝑦 − v (𝜏)|) 𝑉.

We set 𝑑 = dist (𝑉, 𝜕𝑈); then, if 𝑥 ∈ 𝑉 and |𝑦| < 𝛿,

|𝜉 | −1 𝐸 1 ≤ −𝜅𝑑 2 + 𝑅 −1 (diam 𝑈) + 2𝛿 (1 + 2𝜅𝛿) .

We require 𝛿 (1 + 2𝜅𝛿) + 𝑅 −1 (diam 𝑈) < 𝜅𝑑 2 /2 whence, in (17.4.20),

2 1 2|𝜉 |
e𝑖 (𝑧−𝑠−𝑖v( 𝜏)) ·𝜁 −𝜀 ⟨𝜁 ⟩ e
𝑎 𝑅 (𝑧, 𝜁) ≲ e− 2 𝜅 𝑑 .

By the Lebesgue Dominated Convergence Theorem 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v∗ )
converges, as 𝜀 ↘ 0, to a holomorphic function of 𝑧 in 𝑉 𝛿C . □
Given arbitrarily v ∈ R𝑛 \ {0} and 𝜌 > 0 we introduce the cone

𝔠𝜌 (v) = {𝜉 ∈ R𝑛 ; 𝜉 · v ≤ −𝜌 |𝜉 | |v|} . (17.4.26)

As before let v◦ ∈ Γ, |v| < 𝛿; we can find finitely many cones (17.4.26), 𝔠𝜌 𝑗 v ( 𝑗) ,
v ( 𝑗) ∈ Γ, v ( 𝑗) < 𝛿, 𝑗 = 1, ..., 𝑁, such that

𝑁
Ø 𝑁
Ø
R𝑛 \𝔠𝜌 (−v◦ ) ⊂ 𝔠𝜌 𝑗 v ( 𝑗) , 𝔠𝜌 𝑗 v ( 𝑗) ∩ 𝔠2𝜌 (−v◦ ) = ∅.
𝑗=1 𝑗=1

Note that 𝔠𝜌 (−v◦ ) = {𝜉 ∈ R𝑛; 𝜉 ·v◦ ≥ 𝜌 |𝜉 |} and that 𝔠2𝜌 (−v◦ ) ⊂ 𝔠𝜌 (−v◦ ). Actu-
ally we replace the cones 𝔠𝜌 𝑗 v ( 𝑗) by the following open (but generally not convex)
cones 𝔠 𝑗 :

𝔠1 = 𝔠𝜌1 v (1) ∩ R𝑛 \𝔠𝜌 (−v◦ ) ,


𝔠2 = interior of 𝔠𝜌2 v (2) \ 𝔠1 ∩ 𝔠𝜌2 v (2) ∩ R𝑛 \𝔠𝜌 (−v◦ ) ,


𝔠3 = interior of 𝔠𝜌3 v (3) \ 𝔠2 ∩ 𝔠𝜌3 v (3) ∩ R𝑛 \𝔠𝜌 (−v◦ ) , etc.

𝑁
Ø
The cones 𝔠 𝑗 are pairwise disjoint and R𝑛 \𝔠𝜌 (−v◦ ) \ 𝔠 𝑗 has Lebesgue measure
𝑗=1
zero; as a consequence, we have
17.4 Action on Singularity Hyperfunctions 679

𝑁
∑︁
𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) = 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ),𝑅,𝜅 (𝑧, v◦ ) + 𝐽𝔠𝜀𝑗 ∩ℭ′ ,𝑅,𝜅 (𝑧, v◦ ) .
𝑗=1

We apply Lemma 17.4.11 with 𝔠 𝑗 ∩ℭ ′ substituted for ℭ ′ and v∗ = v( 𝑗) (1 ≤ 𝑗 ≤ 𝑁): if

𝑅 −1 and 𝛿 are sufficiently small then 𝐽𝔠𝜀𝑗 ∩ℭ′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽𝔠𝜀𝑗 ∩ℭ′ ,𝑅,𝜅 𝑧, v ( 𝑗) converges
(as 𝜀 ↘ 0) to a holomorphic function 𝐹 𝑗 (𝑧) of 𝑧 in 𝑉 𝛿C = {𝑧 ∈ C𝑛 ; 𝑥 ∈ 𝑉, |𝑦| < 𝛿}.
It follows that
𝑁
∑︁ 𝑁
∑︁
𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐹 𝑗 (𝑧) = 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ) ,𝑅,𝜅 (𝑧, v◦ ) + 𝐽𝔠𝜀𝑗 ∩ℭ′ ,𝑅,𝜅 𝑧, v ( 𝑗) .
𝑗=1 𝑗=1
(17.4.27)

Lemma 17.4.12 Let 𝑉 ⊂⊂ 𝑈 be an open subset of R𝑛 . If 𝑅 −1 and 𝛿 are sufficiently

small then

𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ),𝑅,𝜅 (𝑧, v◦ ) (17.4.28)

∫ ∫
◦ ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 2
= e𝑖 (𝑧−𝑠−𝑖v 𝑓 (𝑠 + 𝑖v◦ ) e
𝑎 𝑅 (𝑧, 𝜁) d𝜁d𝑠
𝑠 ∈𝑈 𝜉 ∈ℭ′ ∩𝔠 𝜌 (−v◦ )

converges, as 𝜀 ↘ 0, to a holomorphic function in the wedge W𝛿 𝑉, Γ𝜌♯ , where

v◦

𝑦
Γ𝜌♯ = 𝑦 ∈ R𝑛 \ {0} ; − ◦ <𝜌 . (17.4.29)
|𝑦| |v |

In (17.4.28) 𝜁 = 𝜉 + 𝑖𝜅 |𝜉 | (𝑧 − 𝑠 − 𝑖v◦ ) [cf. (17.4.20)].

Proof We apply Lemma 17.4.11 with ℭ ′∩𝔠 𝜌 (−v◦ ) substituted for ℭ ′: if 𝛿 and
𝑅 −1 are sufficiently small then 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ),𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ),𝑅,𝜅 𝑧, 𝑝1 v◦ (𝑝 =
2, 3...) converges (as 𝜀 ↘ 0) to a holomorphic function 𝜑 𝑝 (𝑧) of 𝑧 in 𝑉 C =
{𝑧 ∈ C𝑛 ; 𝑥 ∈ 𝑉, |𝑦| < 𝛿}. When 𝑤 ∈ 𝑈 + 𝑝1 𝑖v◦ and 𝜉 ∈ 𝔠𝜌 (−v◦ ) we have, by
(17.4.24),

1 ◦ 1
Re (𝑖 (𝑧 − 𝑤) · 𝜉) = − 𝑦 − v · 𝜉 ≥ 𝜌 |𝜉 | |v◦ | − 𝑦 · 𝜉.
𝑝 𝑝

It is not difficult to derive from −1

this that, again provided 𝛿 and 𝑅 are sufficiently
1
small, 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ),𝑅,𝜅 𝑧, 𝑝 v◦ converges (as 𝜀 ↘ 0) to a holomorphic function
𝜓 𝑝 (𝑧) of 𝑧 in

1
Ω′ × 𝑦 ∈ R𝑛 ; ∀𝜉 ∈ 𝔠𝜌 (−v◦ ) , 𝑦 · 𝜉 > 𝜌 |𝜉 | |v◦ | .
𝑝

This implies that 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ),𝑅,𝜅 (𝑧, v◦ ) converges to 𝜑 𝑝 (𝑧) + 𝜓 𝑝 (𝑧) in
680 17 Analytic Pseudodifferential Calculus

𝑛 1 ◦ ◦
𝑉 × 𝑦 ∈ R ; ∀𝜉 ∈ 𝔠𝜌 (−v ) , 𝑦 · 𝜉 > 𝜌 |𝜉 | |v | , |𝑦| < 𝛿 .
𝑝

Since 𝐽ℭ𝜀′ ∩𝔠 𝜌 (−v◦ ) ,𝑅,𝜅 (𝑧, v◦ ) is independent of 𝑝 we have 𝜑 𝑝+1 (𝑧) + 𝜓 𝑝+1 (𝑧) =
𝜑 𝑝 (𝑧) + 𝜓 𝑝 (𝑧) if 𝑦 · 𝜉 > 𝑝1 𝜌 |𝜉 | |v◦ |, 𝜉 ∈ 𝔠𝜌 (−v◦ ). We can therefore let 𝑝 go to +∞;
𝜑 𝑝 + 𝜓 𝑝 converges to a holomorphic function of 𝑧 in

𝑉 × 𝑦 ∈ R𝑛 ; ∀𝜉 ∈ 𝔠𝜌 (−v◦ ) , 𝑦 · 𝜉 > 0 .

(17.4.30)

If 𝑦 ∈ Γ𝜌♯ and 𝜉 ∈ 𝔠𝜌 (−v◦ ) then

v◦ · 𝜉 v◦

𝑦·𝜉 𝜉 𝑦
= + · − > 0.
|𝜉 | |𝑦| |𝜉 | |v◦ | |𝜉 | |𝑦| |v◦ |

It follows that (17.4.30) contains W𝛿 𝑉, Γ𝜌♯ , whence the desired result. □
Lemma 17.4.13 If 𝜅 and 𝛿 ′ > 0 are sufficiently small then, as 𝜀 ↘ 0, the integral
∫ ∫
( 𝑗) 2

𝐽𝔠𝜀𝑗 ∩ℭ′ ,𝑅,𝜅 𝑧, v ( 𝑗) = e𝑖 ( 𝑧−𝑠−𝑖v ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 𝑓 𝑠 + 𝑖v ( 𝑗) e𝑎 𝑅 (𝑧, 𝜁) d𝜁d𝑠
𝑠 ∈𝑈 𝜉 ∈𝔠 𝑗 ∩ℭ′
(17.4.31)
converges to a holomorphic function in 𝑉 𝛿C′ ( 𝑗 = 1, ..., 𝑁).

In the integral (17.4.31) we have 𝜁 = 𝜉 + 𝑖𝜅 |𝜉 | 𝑧 − 𝑠 − 𝑖v ( 𝑗) .
Proof We avail ourselves of (17.4.24) and (17.4.26):
2
|𝜉 | −1 𝐸 1 ≤ −𝜌 𝑗 v ( 𝑗) + |𝑦| + 𝜅 𝑦 − v ( 𝑗) − 𝜀 |𝜉 | /2. (17.4.32)

It suffices to require |𝑦| < 𝛿 ′ ≪ 𝜌 𝑗 v ( 𝑗) and 𝜅 ≪ 𝜌 𝑗 to see that we can let 𝜀 go to

zero in (17.4.31). □
In the remainder of the proof we assume 𝜅 ≪ min 𝜌 𝑗 . By taking Lemmas
1≤ 𝑗 ≤ 𝑁
17.4.12, 17.4.13 into account in (17.4.27) we can state
Proposition 17.4.14 If the positive numbers 𝑅 −1 , 𝛿(≪ 𝑅 −1 ), 𝛿 ′(< 𝛿) and 𝜌 ′ are
sufficiently small then, as 𝜀 ↘ 0, 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) converges (uniformly on compact

subsets) to a holomorphic function in W𝛿′ 𝑉, Γ𝜌♯ ′ .
Next we estimate the difference

𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ,𝑅,0 (𝑧, v◦ ) (17.4.33)

∫ 𝜅∫ ∫
◦ ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 2
= e𝑖 (𝑧−𝑠−𝑖v 𝑓 (𝑠 + 𝑖v◦ ) d𝑠 𝜕 𝜁 𝑎 (𝑧, 𝜁) ∧ d𝜁
0 𝜉 ∈ℭ′ 𝑈
∫ 𝜅 ∫ ∫
◦ ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 2 D𝜁
+ e𝑖 (𝑧−𝑠−𝑖v 𝑓 (𝑠 + 𝑖v◦ ) 𝑎 (𝑧, 𝜁) d𝑠d𝑆ℭ′ d𝜆,
0 𝜉 ∈𝜕ℭ′ 𝑈 D𝜉
17.4 Action on Singularity Hyperfunctions 681

where d𝑆ℭ′ is the appropriate volume element on 𝜕ℭ ′ (in 𝜉-space). In both integrals in
the right-hand side of (17.4.33) we have Re 𝜁 = 𝜉+𝜆 |𝜉 | (𝑥 − 𝑠), Im 𝜁 = 𝜆 |𝜉 | (𝑦 − v◦ ),
0 ≤ 𝜆 ≤ 𝜅. We underline the fact that 𝜉 ∈ ℭ ′ in the first integral whereas 𝜉 ∈ 𝜕ℭ ′ in
the second one; in both there is integration with respect to 𝜆 over the segment [0, 𝜅].
Lemma 17.4.15 If the positive numbers 𝛿 and 𝛿 ′(< 𝛿) are sufficiently small then, as
𝜀 ↘ 0, 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ,𝑅,0 (𝑧, v◦ ) converges to a holomorphic function in 𝑉 𝛿C′ .
Proof In dealing with the first integral in the right-hand side of (17.4.33) we take
advantage of (17.4.23) [cf. (17.4.25)]:
◦ ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 2
e𝑖 (𝑧−𝑠−𝑖v 𝜕 𝜁 𝑎 (𝑧, 𝜁) ≲ e−𝑐𝐸3

where [cf. (17.4.25)]

|𝜉 | −1 𝐸 3 ≤ |𝑦 − v◦ | + 𝜅 |𝑦 − v◦ | 2 − 𝜅 |𝑥 − 𝑠| 2 − (2𝑅) −1 − 𝜀 |𝜉 | /2
≤ 2𝛿 (1 + 𝜅𝛿) − (2𝑅) −1 − 𝜀 |𝜉 | /2

if |𝑦| < 𝛿. At this point we require 𝛿 ≪ 𝑅 −1 (𝑅 and 𝛿 were unrelated so far), ensuring
that ∫ ∫
◦ 2
e𝑖 (𝑧−𝑠−𝑖v ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 𝑓 (𝑠 + 𝑖v◦ ) d𝑠 𝜕 𝜁 𝑎 (𝑧, 𝜁) ∧ d𝜁
ℭ′ (𝜅 ,𝑧−𝑠) 𝑈

converges to a holomorphic function in 𝑉 𝛿C as 𝜀 ↘ 0. In the second integral in the

right-hand side of (17.4.33) we observe that
𝑁
Ø
𝜕ℭ ′ ⊂ 𝔠 𝑗 ∩ ℭ ′.
𝑗=1

The estimate (17.4.32) in the proof of Lemma 17.4.13 is valid in 𝔠 𝑗 ∩ ℭ ′. It follows

that
∫ 𝜅∫ ∫
◦ 2 D𝜁
e𝑖 (𝑧−𝑠−𝑖v ) ·𝜁 −𝜀 ⟨𝜁 ⟩ 𝑓 (𝑠 + 𝑖v◦ ) 𝑎 (𝑧, 𝜁) d𝑠d𝑆ℭ′ d𝜆
0 ′
𝜉 ∈𝜕ℭ 𝑈 D𝜉

converges to a holomorphic function in 𝑉 𝛿C′ as 𝜀 ↘ 0. □

Combining Proposition 17.4.14 with Lemma 17.4.15 yields
Proposition 17.4.16 Let 𝑓 ∈ O (W𝛿 (Ω′, Γ)), v◦ ∈ Γ, |v◦ | < 𝛿 ′ < 𝛿. If the positive
numbers 𝑅 −1 , 𝛿(≪ 𝑅 −1 ), 𝛿 ′(< 𝛿) and 𝜌 ′ are sufficiently small then, as 𝜀 ↘ 0, the
integral
∫ ∫
◦ 2
(2𝜋) −𝑛 e𝑖 (𝑧−𝑠−𝑖v ) · 𝜉 −𝜀 | 𝜉 | 𝑓 (𝑠 + 𝑖v◦ ) 𝑎 (𝑧, 𝜉) d𝜉d𝑠
𝑠 ∈𝑈 𝜉 ∈ℭ′

converges
(uniformly
on compact subsets) to a holomorphic function 𝐹 (𝑧, v◦ ) in
W𝛿′ 𝑉, Γ𝜌♯ ′ [cf. (17.4.29)].
682 17 Analytic Pseudodifferential Calculus

17.4.3.2 Completing the proof of Theorem 17.4.10

We continue to deal with vectors

v◦ , v∗ ∈ Γ, v∗ ≠ v◦ , max (|v◦ | , |v∗ |) < 𝛿.

By Proposition 17.4.14 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) and 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v∗ ) converge, as 𝜀 ↘ 0, to

holomorphic functions in W𝛿′ 𝑉, Γ𝜌♯ ′ , 𝐹 (𝑧, v◦ ) and 𝐹 (𝑧, v∗ ) respectively. By
Lemma 17.4.11 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v◦ ) − 𝐽ℭ𝜀′ ,𝑅,𝜅 (𝑧, v∗ ) converges to a holomorphic func-
tion 𝐺 (𝑧, v◦ , v∗ ) in 𝑉 𝛿C . It follows directly that the hyperfunction boundary values of
𝐹 (𝑧, v◦ ) and 𝐹 (𝑧, v∗ ) in 𝑉 differ by 𝐺 (𝑥, v◦ , v∗ ) and thus define one and the same
singularity hyperfunction in 𝑉. We apply this result to the case v∗ = 𝑡v◦ , 0 < 𝑡 < 1;
Lemma 7.2.2 states that there is an analytic functional 𝜇𝑈𝑓 ∈ O ′ 𝑈 having the fol-
lowing property: to every compact neighborhood 𝐾 of 𝜕𝑈 in C𝑛 there is a 𝑡 𝐾 ∈ (0, 1)
such that the analytic functional 𝜇𝑈𝑓,𝑡v◦ − 𝜇𝑈𝑓 is carried by 𝐾 if 0 < 𝑡 < 𝑡 𝐾 . We take

𝐾 ⊂ Ω′; Lemma 17.4.5 entails that 𝐴ℭ⊤′ ,Ω′ 𝜇𝑈𝑓,𝑡v◦ − 𝜇𝑈𝑓 is a C 𝜔 function in Ω′\𝐾.
Back to (17.4.19) we see that (2𝜋) 𝑛 𝐴ℭ⊤′ ,Ω′ 𝜇𝑈𝑓,𝑡v◦ is equal to the limit, as 𝜀 ↘ 0,
of ∫ ∫
◦ 2
lim e𝑖 ( 𝑥+𝑖𝑦−𝑠−𝑖𝑡v ) · 𝜉 −𝜀 | 𝜉 | 𝑓 (𝑠 + 𝑖𝑡v◦ ) 𝑎 (𝑥 + 𝑖𝑦, 𝜉) d𝜉d𝑠,
Γ∋𝑦→0 𝑠 ∈𝑈 ℭ′
which is to say, the limit, as 𝜀 ↘ 0, of the boundary value of 𝐽ℭ𝜀′ ,𝑅,0 (𝑧, 𝑡v◦ ) in 𝑈. By
Proposition 17.4.16 𝐴ℭ⊤′ ,Ω′ 𝜇𝑈𝑓,𝑡v◦ is equal in 𝑉 to the boundary value of 𝐹 (𝑧, 𝑡v◦ ).
Assuming that 𝑉 ⊂⊂ Ω′\𝐾 we reach the conclusion that 𝐴ℭ⊤′ ,Ω′ 𝜇𝑈𝑓 is equal to the

boundary value of a function 𝑔 ∈ O W𝛿′ 𝑉, Γ𝜌♯ ′ (𝛿 ′, 𝜌 ′ suitably small).
Now suppose (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑉 × ℭ ′ and select the cone Γ so that v · 𝜉 ◦ < 0 whatever
v ∈ Γ; this implies that 𝑏 Ω′ 𝑓 is microlocally analytic at (𝑥 ◦ , 𝜉 ◦ ) (Definition 7.4.7). We
have just shown that 𝐴ℭ⊤′ ,Ω′ 𝜇𝑈𝑓 defines the singularity hyperfunction in 𝑉 represented
in 𝑉 by 𝑏 𝑉 𝑔; it follows from (17.4.29) that the latter singularity hyperfunction is
microlocally analytic at (𝑥 ◦ , 𝜉 ◦ ). In general, a hyperfunction 𝑢 microlocally analytic
at (𝑥 ◦ , 𝜉 ◦ ) is equal, in a sufficiently small neighborhood Ω′ of 𝑥 ◦ , to a finite sum
of hyperfunctions
Ð𝑟 𝑏 Ω′ 𝑓 𝑗 ( 𝑗 = 1, ..., 𝑟), 𝑓 𝑗 ∈ O W𝛿 Ω′, Γ 𝑗 , v · 𝜉 ◦ < 0 whatever
v ∈ 𝑗=1 Γ 𝑗 . If 𝑢 is the singularity hyperfunction defined by 𝑢 then 𝑎 (𝑥, D) 𝑢 ♮ (see
♮

the preceding subsection) is microlocally analytic at (𝑥 ◦ , 𝜉 ◦ ).

Theorem 17.4.10 has the familiar consequence (cf. Corollaries 17.3.11, 17.3.13):

Theorem 17.4.17 If 𝑨ℭ ∈ Ψa (Ω, ℭ) is elliptic and 𝑢 ♮ ∈ B sing (Ω) then

𝑊 𝐹a 𝑨ℭ 𝑢 ♮ = 𝑊 𝐹a 𝑢 ♮ .

Needless to say, Theorems 17.4.10 and 17.4.17 imply their respective distribution
analogues, Theorem 17.4.9 and Corollary 17.3.11.
17.5 Microdifferential Operators 683

17.5 Microdifferential Operators

The results of the preceding section enable us to define the action of an operator
𝑨ℭ ∈ Ψa (Ω, ℭ) on microfunctions. We recall once again how the latter are defined
(cf. Section 7.5). We may as well identify the cotangent bundle 𝑇 ∗ Ω with Ω × R𝑛 and
the cosphere bundle 𝑆 ∗ Ω with Ω × S𝑛−1 ; let 𝜋 ∗ : 𝑇 ∗ Ω\0 −→ 𝑆 ∗ Ω denote the natural
projection.
♮ Given 𝑢 ♮ ∈ B sing (Ω) and (𝑥, 𝜉) ∈ Ω × ℭ ∩ S𝑛−1 arbitrarily we denote
by 𝑢 ( 𝑥, 𝜉 ) the coset of the germ at 𝑥 of 𝑢 ♮ mod the linear space of the germs at
𝑥 of singularity hyperfunctions microanalytic at (𝑥, 𝜉). These cosets form the stalk
at (𝑥, 𝜉) (now regarded as a point in 𝑆 ∗ R𝑛 ) of the sheaf of microfunctions over R𝑛 ,
Bmicro (R𝑛 ). Let U † denote an open subset of 𝑆 ∗ R𝑛 . We micro U † the
denote by B micro
†
linear space of continuous sections U ∋ (𝑥, 𝜉) ↦→ 𝑢 ( 𝑥, 𝜉 ) of the sheaf B
♮ (R𝑛 ).
the action of 𝑨ℭ on this section is defined as the continuous
Naturally, section
𝑨ℭ 𝑢 ♮ ( 𝑥, 𝜉 ) : (𝑥, 𝜉) ↦→ 𝑨ℭ 𝑢 ♮ ( 𝑥, 𝜉 ) over U † = Ω × ℭ ∩ S𝑛−1 . By transitioning

to the germs of such continuous sections this defines the linear sheaf endomorphism
𝑨ℭ of Bmicro (𝑆 ∗ R𝑛 ) over ℭ ∩ S𝑛−1 .
Let (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × ℭ be arbitrary. It has been clear, from the start (cf. Definitions
17.1.4, 17.1.12) as well as throughout the presentation in the preceding sections
of this chapter, that there are no obstructions to contracting Ω about 𝑥 ◦ and ℭ ∩
S𝑛−1 about 𝜉 ◦ . The, by now routine, procedure of germification allows us to define
the germ of operator [ 𝑨] ( 𝑥 ◦ , 𝜉 ◦ ) , thereby defining the sheaf of microdifferential
operators over R𝑛 , which we shall denote by 𝚿micro a (R𝑛 ). We are thinking of these
operators as acting on microfunctions, thus implicitly “modding off” the germs of
analytic regularizing operators (Definition 17.1.17). We denote by Ψamicro U † the
algebra (with respect to addition and composition) of continuous sections of the sheaf
𝚿micro
a (R𝑛 ) over the open subset U † of 𝑆 ∗ R𝑛 , i.e. the algebra of microdifferential
operators over U † .
The following statements reformulate the results established in the preceding
subsection.

If [ 𝑨] ∈ Ψamicro U † then supp [ 𝑨 𝑓 ] ⊂ supp [ 𝑓 ] for every [ 𝑓 ] ∈

Theorem 17.5.1
B micro U † .

When we refer to the order 𝑚 of [ 𝑨] this presumes that [ 𝑨] ( 𝑥 ◦ , 𝜉 ◦ ) has a repre-

sentative 𝑨 ∈ Ψ𝑚 (Ω, ℭ), (𝑥 ◦ , 𝜉 ◦ ) ∈ Ω × ℭ.
By saying that [ 𝑨] ∈ Ψa (𝜋 ∗ (Ω × ℭ)) is a classical microdifferential operator of
order 𝑚 we mean that it has a representative in Ω × ℭ whose symbol is classical of
order 𝑚 (Definition 17.2.28). If 𝑎 0 (𝑥, 𝜉) ∈ 𝑆 𝑚
𝜓a (Ω × ℭ) is its principal symbol we
write
Char 𝑨 = {(𝑥, 𝜉) ∈ Ω × ℭ ; 𝑎 0 (𝑥, 𝜉) = 0} .
Then we say that [ 𝑨] is elliptic of order 𝑚 in the open subset 𝜋 ∗ (Ω × ℭ\ Char 𝑨) of
𝑆 ∗ R𝑛 . These definitions extend routinely to open subsets U † of 𝑆 ∗ R𝑛 not necessarily
of the form 𝜋 ∗ (Ω × ℭ).
684 17 Analytic Pseudodifferential Calculus

Since the analytic regularizing operators have been modded off we can state:

Theorem 17.5.2 If [ 𝑨] ∈ Ψamicro U † is an elliptic classical microdifferential op-

erator of order 𝑚 then [ 𝑨] has an inverse [ 𝑨] −1 which is an elliptic classical

microdifferential operator of order −𝑚.

Combining Theorems 17.5.1 and 17.5.2 has the following immediate conse-
quence:

Theorem 17.5.3 If [ 𝑨] ∈ Ψamicro U † is an elliptic classical microdifferential op-

erator in U † then

∀ [ 𝑓 ] ∈ B micro U † , supp [ 𝑨 𝑓 ] = supp [ 𝑓 ] .

Theorem 17.5.3 directly implies Sato’s fundamental principle (cf. Theorem

7.4.19):

Theorem 17.5.4 If [ 𝑨] ∈ Ψamicro U † is a classical microdifferential operator then

∀ [ 𝑓 ] ∈ B micro U † , supp [ 𝑓 ] ⊂ supp [ 𝑨 𝑓 ] ∪ Char 𝑨. (17.5.1)

Corollary 17.5.5 Let Ω be a domain in R𝑛 and 𝑨 a classical analytic pseudodiffer-

ential operator in Ω. We have

∀ 𝑓 ∈ B (Ω) , 𝑊a ( 𝑓 ) ⊂ 𝑊a ( 𝑨 𝑓 ) ∪ Char 𝑨. (17.5.2)

All the definitions and results of this section can be extended to a general para-
compact C 𝜔 manifold by using local affinizations of the cotangent and cosphere
bundles (see Section 11.3).
Chapter 18
Fourier Integral Operators

Within the whole algebra of pseudodifferential operators there is a profusion of tools

that enable us to transform a pseudodifferential operator 𝑨 into another one, 𝐵, better
suited to our needs: foremost, conjugations (or similarities) 𝑨 ↦→ Γ−1 𝑨Γ with Γ
in large groups of invertible operators that are not (elliptic) pseudodifferential. The
evident analogue, in nonmicrolocal analysis, of the conjugate Γ−1 𝑨Γ is that of a
bounded linear operator 𝑨 on a complex Hilbert space 𝑯, conjugated with a unitary
operator Γ. The analogy is well-founded since, in important applications, that is how
the pseudodifferential operator 𝑨 is made to act and Γ is selected (with 𝑯 a Hilbert
space of distributions, say a Sobolev space).
At the time of writing the class of candidates for the role of Γ is so large that it
is impossible to circumscribe. This chapter is an introduction to one of the simpler
and more natural classes of such operators, the Fourier Integral Operators (FIOs),
generally defined by oscillatory integrals
∫ ∫
e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (𝑥, 𝑦, 𝜃) 𝑢 (𝑦) d𝑦d𝜃 (18.1)

where 𝑎 (𝑥, 𝑦, 𝜃) is an amplitude (Section 16.1) and the phase-function 𝜑 (𝑥, 𝑦, 𝜃) is

smooth and homogeneous of degree 1 with respect to 𝜃 ∈ R 𝑁 \ {0} (𝑁: a positive
integer possibly unrelated to the dimension 𝑛 of the base M in which 𝑥 and 𝑦
vary). In this chapter 𝜑 shall always be real-valued. Crucially, the presence of 𝜑
opens a link with the symplectic structure of 𝑇 ∗ M, through the symplectic maps
(𝑥, 𝜕𝑥 𝜑) ↦→ 𝑦, −𝜕𝑦 𝜑 [under the condition that 𝜕𝜃 𝜑 = 0].
An important class of FIOs are the parametrices in the Cauchy problem for
hyperbolic equations (see [Treves, 1980], Vol. II). We begin the chapter with a
detailed description of the natural parametrix of the wave equation, the archetypical
hyperbolic equation. In close connection with this topic, the reader might profit from
taking a look at the Appendix on the Stationary Phase expansion.
In the last section of the chapter we rapidly sketch the properties of analytic
Fourier Integral Operators. In the terminology of M. Sato and his school (see [Sato-
Kawai-Kashiwara, 1973]) the transformations defined by analytic FIOs acting on

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 685
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_18
686 18 Fourier Integral Operators

hyperfunctions or microfunctions (cf. Section 11.3) are referred to as quantized

contact transformations. This terminology is widely used, especially in the sheaf-
cohomology approach to microfunctions (e.g, [Schapira, 1985], [Hörmander, 1994]).
Part VI will extend the pool of integrals (18.1) by allowing 𝜑 to be complex-
valued, in particular of the FBI type (cf. Section 3.4). In Part VII we will encounter
many uses of the transforms Γ−1 𝑨Γ where Γ is an FIO, with real or complex phase-
function.
Most of the contents of this chapter follow closely the approach of L. Hörmander
to FIOs (see for instance [Hörmander, 1983, IV], Ch. XXV). They have deep roots in
the mid-XXth century works of Russian mathematicians and theoretical physicists
(see [Maslov, 1965]; for a more straightforward presentation, [Leray, 1981]).

18.1 Fourier Distribution Kernels in Euclidean Space

18.1.1 The prototype

Consider the wave operator in R𝑛𝑥 × R𝑡 :

𝜕2
□= − Δ𝑥 , (18.1.1)
𝜕𝑡 2
𝜕2
is the Laplacian in 𝑥-space. The natural problem for □ is the
Í𝑛
where Δ 𝑥 = 𝑗=1 𝜕𝑥 2
𝑗
Cauchy problem in R𝑛𝑥 × R𝑡 :

□𝑢 = 𝑓 (𝑥, 𝑡) , (18.1.2)
𝜕𝑢
𝑢| 𝑡=0 = 𝑢 ◦ (𝑥) , = 𝑢 1 (𝑥) , (18.1.3)
𝜕𝑡 𝑡=0

where 𝑓 ∈ D ′ R𝑛𝑥 × R𝑡 , 𝑢 ◦ , 𝑢 1 ∈ D ′ (R𝑛 ). At first we reason under the hypothesis

that the data are compactly supported with respect to 𝑥. Fourier transform in 𝑥-space
transforms (18.1.2)–(18.1.3) into the ODE initial value problem:

𝜕2b 𝑢
+ |𝜉 | 2 b
𝑢= b𝑓 (𝜉, 𝑡) , (18.1.4)
𝜕𝑡 2
𝜕b
𝑢
𝑢 | 𝑡=0 = b
b 𝑢 ◦ (𝜉) , =b 𝑢 1 (𝜉) , (18.1.5)
𝜕𝑡 𝑡=0

whose solution is elementary:

18.1 Fourier Distribution Kernels in Euclidean Space 687

sin (|𝜉 | (𝑡 − 𝑡 ′)) b

∫ 𝑡
𝑢 (𝜉, 𝑡) =
b 𝑓 (𝜉, 𝑡 ′) d𝑡 ′ (18.1.6)
0 |𝜉 |
sin (|𝜉 | 𝑡)
+ cos (|𝜉 | 𝑡) b
𝑢 ◦ (𝜉) + 𝑢 1 (𝜉) .
b
|𝜉 |
Inverse Fourier transform yields

sin (|𝜉 | (𝑡 − 𝑡 ′))

∫ ∫ ∫ 𝑡
1 𝑖 𝜉 · ( 𝑥−𝑦)
𝑢 (𝑥, 𝑡) = e 𝑓 (𝑦, 𝑡 ′) d𝑡 ′d𝑦d𝜉 (18.1.7)
(2𝜋) 𝑛 R𝑛 R𝑛 0 |𝜉 |
∫ ∫
1
+ e𝑖 𝜉 · ( 𝑥−𝑦) cos (|𝜉 | 𝑡) 𝑢 ◦ (𝑦) d𝑦d𝜉
(2𝜋) 𝑛 R𝑛 R𝑛
∫ ∫
1 sin (|𝜉 | 𝑡)
+ e𝑖 𝜉 · ( 𝑥−𝑦) 𝑢 1 (𝑦) d𝑦d𝜉.
(2𝜋) 𝑛 R𝑛 R𝑛 |𝜉 |

None of the integrals on the right in (18.1.7) defines a pseudodifferential operator

because sin(| 𝜉| 𝜉| |𝑡) is not a standard amplitude (Definition 16.1.1). It is, however, a
substandard amplitude of order −1 (Definition 16.1.2) and therefore, by Theorem
16.1.4, it defines a continuous linear operator E ′ (R𝑛 ) −→ D ′ (R𝑛 ). Actually, the
hypothesis that the data are compactly supported in 𝑥-space (for 𝑡 ∈ R fixed, arbitrary)
is equivalent to the fact that the functions of 𝜉, 𝑓 (𝜉, 𝑡), b 𝑢 ◦ (𝜉), b
𝑢 1 (𝜉), can be extended
to C𝑛 as entire functions of exponential type (Paley–Wiener–Schwartz Theorem
2.1.2); the same is then true of b 𝑢 (𝜉, 𝑡): each integral on the right in (18.1.7) is
compactly supported in 𝑥-space, and so is 𝑢 (𝑥, 𝑡). It is well known that there is a
direct proportionality between supp 𝑢 (·, 𝑡) and the supports of the data, determined
by the so-called domain of influence [see (18.1.11) below]. This also shows that the
operator defined by the distribution kernel
∫
−𝑛 sin (|𝜉 | 𝑡)
(2𝜋) e𝑖 𝜉 · ( 𝑥−𝑦) d𝜉 (18.1.8)
R 𝑛 |𝜉 |

can be extended as a continuous linear operator D ′ (R𝑛 ) ←↪. For more details on
this topic see, e.g., [Treves, 1975], Sections 7 and 14.
When 𝑛 ≥ 2 it is convenient to expand sin(| 𝜉| 𝜉| |𝑡) = 2𝑖1 |𝜉 | −1 e𝑖 | 𝜉 |𝑡 − e−𝑖𝑡 | 𝜉 | ; the
following, ∫
−𝑛 d𝜉
𝑈 (𝑥, 𝑡) = (2𝜋) e𝑖 ( 𝜉 · 𝑥+𝑡 | 𝜉 |) , (18.1.9)
R 𝑛 |𝜉 |
is a bona fide oscillatory integral as defined in Subsection 18.1.4 below, and therefore
a distribution in R𝑛𝑥 depending smoothly on 𝑡 ∈ R; 𝑈 (𝑥 − 𝑦, 𝑡) is the basic example
of a Fourier distribution kernel as defined
and studied in this section; it defines a
continuous linear operator 𝑼 (𝑡) : D R𝑛𝑦 −→ D ′ R𝑛𝑥 .
′

Of course, 𝑈 (𝑥, 𝑡) has special properties. Select at random 𝜒 ∈ Cc∞ (R𝑛 ), 𝜒 (𝜉) =
1 if |𝜉 | < 𝜀; we see that
688 18 Fourier Integral Operators
∫
1 − 𝜒 (𝜉)
𝑈 (𝑥, 𝑡) = (2𝜋) −𝑛 e𝑖 ( 𝜉 · 𝑥+𝑡 | 𝜉 |) d𝜉 + 𝑅 (𝑥, 𝑡)
R𝑛 |𝜉 |
where ∫
−𝑛 d𝜉
𝑅 (𝑥, 𝑡) = (2𝜋) e𝑖 ( 𝜉 ·𝑥+𝑡 | 𝜉 |) 𝜒 (𝜉)
R𝑛 |𝜉 |
can be extended (for 𝑡 fixed) as an entire function of exponential type 𝑅 (𝑧, 𝑡) in C𝑛 .
Formula (18.1.7) can be rewritten as
∫ 𝑡∫
1
𝑢 (𝑥, 𝑡) = (𝑈 (𝑥 − 𝑦, 𝑡 − 𝑡 ′) − 𝑈 (𝑥 − 𝑦, 𝑡 ′ − 𝑡)) 𝑓 (𝑦, 𝑡 ′) d𝑦d𝑡 ′
2𝑖 0 R𝑛
(18.1.10)
∫
1 𝜕𝑈 𝜕𝑈
+ (𝑥 − 𝑦, 𝑡) + (𝑥 − 𝑦, −𝑡) 𝑢 ◦ (𝑦) d𝑦
2𝑖 R𝑛 𝜕𝑡 𝜕𝑡
∫
1
− (𝑈 (𝑥 − 𝑦, 𝑡) − 𝑈 (𝑥 − 𝑦, −𝑡)) 𝑢 1 (𝑦) d𝑦.
2𝑖 R𝑛
In passing notice that all the integrals with respect to 𝑦 are convolutions. The phase-
function 𝜑◦ (𝑥, 𝑦, 𝑡, 𝜉) = 𝜉 · (𝑥 − 𝑦) + 𝑡 |𝜉 | is strongly nondegenerate (Definition
18.2.10 below).
The equation d 𝜉 𝜑◦ = 0 in R2𝑛 × R2𝑛 defines the graph of the map
(𝑥, 𝜉) ↦→ 𝑥 + 𝑡 | 𝜉𝜉 | , 𝜉 , obviously a symplectomorphism. If we take d𝜉 ∧ d𝑥 − d𝜂 ∧ d𝑦
as the fundamental symplectic form in R2𝑛 2𝑛
𝑥, 𝜉 × R 𝑦, 𝜂 we see that the graph of said
map is a Lagrangian submanifold (Definition 13.3.15). The map reveals directly the
expansion of the supports under the action of 𝑼 (𝑡):

supp 𝑼 (𝑡) 𝑓 ⊂ supp 𝑓 + 𝑡S𝑛−1 . (18.1.11)

It can be checked on (18.1.9) that 𝑼 (𝑡) = |D 𝑥 | −1 e𝑖𝑡 |D 𝑥 | , whence

𝑼 ′ (𝑡) = 𝑖 |D 𝑥 | 𝑼 (𝑡) ; (18.1.12)

D𝑡 𝑼 (𝑡) = e𝑖𝑡 |D 𝑥 | defines a unitary transformation of 𝐿 2 (R𝑛 ).

18.1.2 Amplitudes in Euclidean space

In this section we enlarge the class of distribution kernels introduced in Ch. 16. As
in Ch. 16 we let Ω1 and Ω2 be open subsets of one and the same Euclidean space R𝑛
(actually, many concepts and results in this section can be extended routinely to the
case where dim Ω1 ≠ dim Ω2 ). One novelty is that instead of limiting the auxiliary
variable 𝜉 to be a frequency, i.e. the variable in the dual of R𝑛 , we replace it by
the variable point 𝜃 = (𝜃 1 , ..., 𝜃 𝑁 ) in an open cone Γ in R 𝑁 \ {0} with 𝑁 a priori
unrelated to 𝑛. Our purpose is to study oscillatory integrals of the following kind:
18.1 Fourier Distribution Kernels in Euclidean Space 689
∫
𝐴 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (𝑥, 𝑦, 𝜃) d𝜃, (18.1.13)
Γ

where 𝜑 (𝑥, 𝑦, 𝜃) ∈ C ∞ (Ω1 × Ω2 × Γ) is real-valued and homogeneous of de-

gree one, meaning that 𝜑 (𝑥, 𝑦, 𝜆𝜃) = 𝜆𝜑 (𝑥, 𝑦, 𝜃) for 𝜆 > 0, and 𝑎 (𝑥, 𝑦, 𝜃) ∈
C ∞ (Ω1 × Ω2 × Γ) is an amplitude in the following sense (cf. Definition 16.1.1):

Definition 18.1.1 By an amplitude of order 𝑚 (∈ R) in Ω1 × Ω2 × Γ we shall mean

a function 𝑎 (𝑥, 𝑦, 𝜃) ∈ C ∞ (Ω1 × Ω2 × Γ) endowed with the following property:
Given any compact subset 𝐾 of Ω1 × Ω2 , any cone Γ♭ in R 𝑁 \ {0} such that
Γ ∩ S 𝑁 −1 ⊂⊂ Γ and any triplet of multi-indices 𝛼, 𝛽 ∈ Z+𝑛 , 𝛾 ∈ Z+𝑁 we have
♭

sup (1 + |𝜃|) −𝑚+|𝛾 | 𝜕𝑥𝛼 𝜕𝑦 𝜕𝜃 𝑎 (𝑥, 𝑦, 𝜃) < +∞.

𝛽 𝛾
(18.1.14)
𝐾×Γ♭

We denote by 𝑆 𝑚 (Ω1 × Ω2 , Γ) the vector space of amplitudes of order 𝑚 in Ω1 ×

Ω2 × Γ.

Actually, in order to spare ourselves the complications arising from the behavior
of the integrand in (18.1.13) at the boundary of Γ, we shall limit ourselves to dealing
with a special class of amplitudes, which could be said to have a conically compact
support.

Definition 18.1.2 We denote by 𝑆c𝑚 (Ω1 × Ω2 , Γ) the subspace of 𝑆 𝑚 (Ω1 × Ω2 , Γ)

consisting of the amplitudes 𝑎 (𝑥, 𝑦, 𝜃) that have the following property:
(K) There is a cone Γ♭ in R 𝑁 \ {0}, Γ♭ ∩S 𝑁 −1 ⊂⊂ Γ, such that supp 𝑎 ⊂ Ω1 ×Ω2 ×Γ♭ .

When Γ = R 𝑁 \ {0} we use the notation 𝑆 𝑚 Ω1 × Ω2 , R 𝑁 instead of

𝑆 𝑚 Ω1 × Ω2 , R 𝑁 \ {0} = 𝑆c𝑚 Ω1 × Ω2 , R 𝑁 \ {0} .

We also define
Ø
𝑆 (Ω1 × Ω2 , Γ) = 𝑆 𝑚 (Ω1 × Ω2 , Γ) , (18.1.15)
𝑚∈R
Ù
𝑆 −∞ (Ω1 × Ω2 , Γ) = 𝑆 𝑚 (Ω1 × Ω2 , Γ) . (18.1.16)
𝑚∈R

We use an analogous notation with 𝑆c in the place of 𝑆 or with R 𝑁 in the place

of Γ. The set 𝑆 (Ω1 × Ω2 , Γ) is a commutative ring with respect to addition and
ordinary multiplication; 𝑆 −∞ (Ω1 × Ω2 , Γ) and 𝑆c (Ω1 × Ω2 , Γ) are ideals in the ring
𝑆 (Ω1 × Ω2 , Γ). The vector space 𝑆 𝑚 (Ω1 × Ω2 , Γ) carries a natural Fréchet space
structure: its topology is defined by the seminorms on the left in (18.1.14). We have
′
𝑆 𝑚 (Ω1 × Ω2 , Γ) ⊂ 𝑆 𝑚 (Ω1 × Ω2 , Γ) if 𝑚 ′ < 𝑚 and the corresponding injection is
continuous. Each 𝑆c𝑚 (Ω1 × Ω2 , Γ) carries the topology defined as follows: a convex
subset 𝔘 of 𝑆c𝑚 (Ω1 × Ω2 , Γ) is open if and only if, given an arbitrary cone Γ♭
690 18 Fourier Integral Operators

in R 𝑁 \ {0} such that Γ♭ ⊂⊂ Γ, the intersection 𝔘 ∩ 𝑆c𝑚 Ω1 × Ω2 , Γ♭ is open

in 𝑆c𝑚 Ω1 × Ω2 , Γ♭ equipped with the topology inherited from the Fréchet space

𝑆 𝑚 Ω1 × Ω2 , Γ♭ .
Evidently one can extend the concept of amplitude to an arbitrary C ∞ real vector
bundle B of rank 𝑁 over a C ∞ manifold M (see Section 9.2) by applying Definition
18.1.1 (or Definition 18.1.2) in every local affinization of B over a coordinate chart
of M.

18.1.3 Real phase-functions in Euclidean space

Turning our attention to the exponent 𝜑 in (18.1.13) we introduce an all-important

concept. As before Ω1 and Ω2 are open subsets of R𝑛 , Γ an open cone in R 𝑁 \ {0}
(never empty).

Definition 18.1.3 By the critical set of 𝜑 we shall mean the set

Σ 𝜑 = {(𝑥, 𝑦, 𝜃) ∈ Ω1 × Ω2 × Γ; 𝜕𝜃 𝜑 (𝑥, 𝑦, 𝜃) = 0} . (18.1.17)

Notice that 𝜑 ≡ 0 on Σ 𝜑 since 𝜑 is homogeneous of degree 1 and therefore

𝜑 = 𝜃 · 𝜑 𝜃 by Euler’s homogeneity identity. We write, for (𝑥, 𝑦, 𝜃) ∈ Σ 𝜑 ,

𝑝 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥, 𝜕𝑥 𝜑 (𝑥, 𝑦, 𝜃)) , 𝑞 𝜑 (𝑥, 𝑦, 𝜃) = 𝑦, −𝜕𝑦 𝜑 (𝑥, 𝑦, 𝜃) . (18.1.18)

Definition 18.1.4 The function 𝜑 ∈ C ∞ (Ω1 ×Ω2 ×Γ) will be called a phase-function
in Ω1 × Ω2 × Γ if 𝜑 is real-valued, homogeneous of degree one, and if the following
conditions are satisfied:
(1) Σ 𝜑 is a (closed, conic) C ∞ submanifold
(Definition 9.3.7) of Ω1 × Ω2 × Γ;
(2) the rank of the map 𝑝 𝜑 , 𝑞 𝜑 is equal to 2𝑛 at every point of Σ 𝜑 ;
(3) ∀ (𝑥, 𝑦, 𝜃) ∈ Σ 𝜑 , |𝜕𝑥 𝜑 (𝑥, 𝑦, 𝜃)| × 𝜕𝑦 𝜑 (𝑥, 𝑦, 𝜃) ≠ 0.

Identifying Ω 𝑗 × (R𝑛 \ {0}) with the cotangent bundle 𝑇 ∗ Ω 𝑗 \0 we shall be dealing

with the pair of maps
Σ𝜑
𝑝𝜑 ↙ ↘ 𝑞𝜑 (18.1.19)
𝑇 ∗ Ω1 \0 𝑇 ∗ Ω2 \0.

Sometimes it may be convenient to use polar coordinates in 𝜃-space. The homo-

geneity of 𝜑 implies 𝜑 (𝑥, 𝑦, 𝜃) = 𝜆𝜓 (𝑥, 𝑦, 𝜔) where 𝜆 = |𝜃|, 𝜔 = 𝜃/|𝜃| ∈ S 𝑁 −1 and
𝜓 ∈ C ∞ Ω1 × Ω2 × S 𝑁 −1 . We have 𝜕𝜔 𝜕𝜃 𝑗 = 𝜆
𝑘 −1 𝛿
𝑗,𝑘 − 𝜔 𝑗 𝜔 𝑘 whence

𝑁
𝜕 𝜕𝜓 ∑︁ 𝜕𝜓
(𝜆𝜓 (𝑥, 𝑦, 𝜔)) = 𝜔 𝑗 𝜓 (𝑥, 𝑦, 𝜔) + (𝑥, 𝑦, 𝜔) − 𝜔 𝑗 𝜔𝑘 (𝑥, 𝑦, 𝜔) .
𝜕𝜃 𝑗 𝜕𝜔 𝑗 𝑘=1
𝜕𝜔 𝑘
18.1 Fourier Distribution Kernels in Euclidean Space 691

We see that Σ 𝜑 is defined by the equations

𝑁
𝜕𝜓 ∑︁ 𝜕𝜓
𝜓 (𝑥, 𝑦, 𝜔) = 0, (𝑥, 𝑦, 𝜔) = 𝜔 𝑗 𝜔𝑘 (𝑥, 𝑦, 𝜔) , 𝑗 = 1, ..., 𝑛. (18.1.20)
𝜕𝜔 𝑗 𝑘=1
𝜕𝜔 𝑘

The following is standard terminology:

Definition 18.1.5 The phase-function 𝜑 in Ω1 × Ω2 × Γ is said to be clean when

the conormal
bundle (see Subsection 9.3.2) of Σ 𝜑 is spanned by the differentials
𝜕𝜑
d 𝜕𝜃𝑘 , 𝑘 = 1, ..., 𝑁; the number dim Σ 𝜑 − 2𝑛 is then called the excess of 𝜑.

Definition 18.1.6 The phase-function 𝜑 in Ω1 ×Ω2 ×Γ is said to be nondegenerate

at
◦ ◦ ◦ 𝜕𝜑
the point (𝑥 , 𝑦 , 𝜃 ) ∈ Σ 𝜑 if 𝜑 is clean and if the differentials d 𝜕𝜃𝑘 , 𝑘 = 1, ..., 𝑁,
are linearly independent at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). We shall say that 𝜑 is nondegenerate if 𝜑 is
nondegenerate at every point of Σ 𝜑 .

Example 18.1.7 If 𝑁 = 𝑛, Ω1 = Ω2 = Ω, and 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃, then

Σ 𝜑 = {(𝑥, 𝑦, 𝜃) ∈ Ω × Ω × Γ; 𝑥 = 𝑦} = (diag Ω) × Γ

and 𝑝 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥, 𝜃), 𝑞 𝜑 (𝑥, 𝑦, 𝜃) = (𝑦, 𝜃); 𝜑 is a nondegenerate phase-function.

Proposition 18.1.8 If the phase-function 𝜑 is nondegenerate

then Σ 𝜑 is a conic,
closed, smooth submanifold of Ω1 × Ω2 × Γ and 𝑝 𝜑 , 𝑞 𝜑 is a local diffeomorphism
of Σ 𝜑 onto an immersed submanifold (Definition 9.3.7) Λ 𝜑 of (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0).
We have dim Σ 𝜑 = dim Λ 𝜑 = 2𝑛.

Proof Indeed, if 𝜑 is nondegenerate then codim Σ 𝜑 = 𝑁, dim Σ 𝜑 = 2𝑛; by Con-

ditions (2) and (3), Definition 18.1.4, rank 𝑝 𝜑 , 𝑞 𝜑 = 2𝑛 and ((𝑥, 𝜉) , (𝑦, 𝜂)) ∈
Λ 𝜑 =⇒ |𝜉 | × |𝜂| ≠ 0. □
𝜕𝜑 𝜕𝜑
If the phase-function 𝜑 is nondegenerate we can use 𝜕𝜃1 , ..., 𝜕𝜃 𝑁 as local coordi-
nates transversal to Σ 𝜑 .

Example Assume
18.1.9 𝑁 = 2𝑛, Ω1 = Ω2 = Ω, (𝑥 ◦ , 𝑦 ◦ ) ∈ Ω2 , and 𝜑 (𝑥, 𝑦, 𝜃) =
Í𝑛 ◦ Í𝑛 ◦
𝑗=1 𝑥 𝑗 − 𝑥 𝑗 𝜃 𝑗 − 𝑗=1 𝑦 𝑗 − 𝑦 𝑗 𝜃 𝑛+ 𝑗 . We have

n o
Σ 𝜑 = (𝑥, 𝑦, 𝜃) ∈ Ω × Ω × R2𝑛 \ {0} ; 𝑥 = 𝑥 ◦ , 𝑦 = 𝑦 ◦ ,

𝜕𝜑 𝜕𝜑
d 𝜕𝜃 𝑗
= d𝑥 𝑗 if 1 ≤ 𝑗 ≤ 𝑛, d 𝜕𝜃 𝑗 = d𝑦 𝑗 if 𝑛 + 1 ≤ 𝑗 ≤ 2𝑛; 𝜑 is a nondegenerate
phase-function. Coordinates on Σ 𝜑 are 𝜃 1 ...,𝜃 2𝑛 , and

𝑝 𝜑 (𝜃) = 𝑥1◦ , ..., 𝑥 𝑛◦ , 𝜃 1 , ..., 𝜃 𝑛 ,

𝑞 𝜑 (𝜃) = 𝑦 ◦1 , ..., 𝑦 ◦𝑛 , 𝜃 𝑛+1 , ..., 𝜃 2𝑛 ;

692 18 Fourier Integral Operators

in other words, 𝑝 𝜑 Σ 𝜑 = 𝑇𝑥 ◦ Ω\ {0}, 𝑞 𝜑 Σ 𝜑 = 𝑇𝑦 ◦ Ω\ {0}. We have Λ 𝜑 =
𝑇𝑥∗◦ Ω\ {0} × 𝑇𝑦∗◦ Ω\ {0} ; Λ 𝜑 can be identified with an embedded submanifold of

𝑇 ∗ Ω2 \ {0} but not, when 𝑥 ◦ ≠ 𝑦 ◦ , with a submanifold of the squared vector bundle
(𝑇 ∗ Ω) 2 .

18.1.4 Making sense of oscillatory integrals

Notwithstanding the precise definitions in the preceding two subsections, at this

point the integrals (18.1.13) do not make much sense. In this subsection we show
how to interpret (18.1.13) as a distribution kernel in Ω1 × Ω2 .
We assume throughout that 𝜑 is a phase-function in Ω1 ×Ω2 ×Γ (Definition 18.1.4)
and that 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) (Definition 18.1.2). We are going to give (18.1.13) a
meaning as the limit in D ′ (Ω1 × Ω2 ), when 𝜀 ↘ 0, of the integrals
∫
2
−𝑁
𝐴 𝜀 (𝑥, 𝑦) = (2𝜋) e𝑖 𝜑 ( 𝑥,𝑦, 𝜃)−𝜀 | 𝜃 | 𝑎 (𝑥, 𝑦, 𝜃) d𝜃. (18.1.21)
Γ

Since all the partial derivatives of 𝑎 have tempered growth as |𝜃| ↗ +∞ we see
immediately that 𝐴 𝜀 ∈ C ∞ (Ω1 × Ω2 ) for fixed 𝜀 > 0.

Proposition 18.1.10 If 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) with 𝑚 < −𝑁 − 𝜈 (𝜈 ∈ Z+ ) then, as

𝜀 ↘ 0, 𝐴 𝜀 converges to a function 𝐴 in C 𝜈 (Ω1 × Ω2 ).

Proof Let 𝛼, 𝛽 ∈ Z+𝑛 be such that |𝛼 + 𝛽| ≤ 𝜈. Straightforward calculation shows

that
∫
2
D 𝑥𝛼 D 𝑦 𝐴 𝜀 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃)−𝜀 | 𝜃 | 𝑎 𝛼,𝛽 (𝑥, 𝑦, 𝜃) d𝜃
𝛽
(18.1.22)
Γ

with 𝑎 𝛼,𝛽 ∈ 𝑆c𝑚+𝜈 (Ω1 × Ω2 , Γ). If 𝑚 + 𝜈 = −𝑁 − 𝛿 (𝛿 > 0) then, by (18.1.14), to

every compact set 𝐾 ⊂ Ω1 × Ω2 there is a constant 𝐶𝐾 , 𝛼,𝛽 > 0, independent of 𝜀
and 𝜃 ∈ Γ, such that

e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 𝛼,𝛽 (𝑥, 𝑦, 𝜃) ≤ 𝐶𝐾 , 𝛼,𝛽 (1 + |𝜃|) −𝑁 − 𝛿 .

The Lebesgue Dominated Convergence Theorem implies that (18.1.22) converges

uniformly on compact subsets of Ω1 × Ω2 . □

Corollary 18.1.11 If 𝑎 (𝑥, 𝑦, 𝜃) ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) vanishes identically when |𝜃| >
𝑅 (0 < 𝑅 < +∞) then 𝐴 𝜀 converges in C ∞ (Ω1 × Ω2 ) as 𝜀 ↘ 0.

Proof The hypothesis implies that 𝑎 ∈ 𝑆c𝑚−𝜈 (Ω1 × Ω2 , Γ) whatever 𝜈 ∈ Z+ . The

claim therefore follows from Proposition 18.1.10. □
18.1 Fourier Distribution Kernels in Euclidean Space 693

The critical set (18.1.17) plays a crucial role in the action of the integral operators
defined starting from the distribution kernel (18.1.13).

Proposition 18.1.12 If 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) vanishes identically in a conic neigh-

borhood U of Σ 𝜑 in Ω1 × Ω2 × Γ then 𝐴 𝜀 converges in C ∞ (Ω1 × Ω2 ) as 𝜀 ↘ 0.

Proof Thanks to Corollary 18.1.11 it suffices to reason under the hypothesis that
𝑎 (𝑥, 𝑦, 𝜃) = 0 when |𝜃| < 1, whatever (𝑥, 𝑦). We have
𝑁 ∫ 𝑎 (𝑥, 𝑦, 𝜃) d𝜃
2
∑︁
(2𝜋) 𝑁
𝐴 𝜀 (𝑥, 𝑦) = D 𝜃𝑘 e𝑖 𝜑−𝜀 | 𝜃 |
𝑘=1 Γ
𝜑 𝜃𝑘 + 2𝑖𝜀𝜃 𝑘
𝑁 ∫
∑︁ 𝑎 (𝑥, 𝑦, 𝜃)
=− e𝑖 𝜑 D 𝜃 𝑘 d𝜃,
𝑘=1 Γ
𝜑 𝜃𝑘 + 2𝑖𝜀𝜃 𝑘

where 𝜑 𝜃𝑘 = 𝜕𝜑/𝜕𝜃 𝑘 . To every compact subset 𝐾 of Ω1 × Ω2 there is a constant

𝑐 𝐾 > 0 such that |𝜑 𝜃 (𝑥, 𝑦, 𝜃)| ≥ 𝑐 𝐾 if 𝑥 ∈ 𝐾, 𝑎 (𝑥, 𝑦, 𝜃) ≠ 0 and (𝑥, 𝑦, 𝜃) ∉ U. This
entails
𝑁
∑︁ 𝑎 (𝑥, 𝑦, 𝜃)
D 𝜃𝑘 ∈ 𝑆c𝑚−1 (Ω1 × Ω2 , Γ) .
𝑘=1
𝜑 𝜃 𝑘 + 2𝑖𝜀𝜃 𝑘

We can repeat this procedure indefinitely, which shows that we may replace 𝑎 in
(18.1.22) by an element of 𝑆c𝑚−𝜈 (Ω1 × Ω2 , Γ), 𝜈 ∈ Z+ arbitrarily large. The claim
then follows from Proposition 18.1.10. □

Theorem 18.1.13 If 𝜑 is a phase-function and 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) (𝑚 ∈ R ar-

bitrary) the integrals (18.1.21) converge in D ′ (Ω1 × Ω2 ) as 𝜀 ↘ 0 to a distribu-
tion kernel 𝐴 [represented by (18.1.13)] which is semiregular (Definition 2.3.1).

Proof Let the open sets Ω′𝜄 ⊂⊂ Ω 𝜄 , 𝜄 = 1, 2, be arbitrary and set 𝐾 = Ω1′ × Ω2′ .
Thanks to Corollary 18.1.11 and Proposition 18.1.12 there is no loss of generality
in assuming that 𝑎 ≡ 0 in a ball |𝜃| < 𝑅 and outside a conic neighborhood U of Σ 𝜑
in Ω1 × Ω2 × Γ. Condition (3), Definition 18.1.4, enables us to select U such that
|𝜑 𝑥 (𝑥, 𝑦, 𝜃)| ≥ 𝑐 |𝜃|, for some 𝑐 > 0 and all (𝑥, 𝑦, 𝜃) ∈ U , (𝑥, 𝑦) ∈ 𝐾, |𝜃| > 2𝑅
(we are using the notation 𝜑 𝑥 = 𝜕𝑥 𝜑). We shall reason by induction on 𝑚. Under our
hypotheses
−1
𝑏 = 1 + |𝜑 𝑥 | 2 𝑎 ∈ 𝑆c𝑚−2 (Ω1 × Ω2 , Γ) .
2
As 𝜀 ↘ 0 the amplitudes 𝑏 𝜀 = 𝑏e−𝜀 | 𝜃 | converge in 𝑆c𝑚−2 Ω1′ × Ω2′ × Γ . We see

that
694 18 Fourier Integral Operators
∫
(2𝜋) 𝑁 𝐴 𝜀 (𝑥, 𝑦) = e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 𝜀 (𝑥, 𝑦, 𝜃) 1 + |𝜑 𝑥 (𝑥, 𝑦, 𝜃)| 2 d𝜃 (18.1.23)
Γ
∫ 𝑛 ∫
∑︁
𝑖 𝜑 ( 𝑥,𝑦, 𝜃)
= e 𝑏 𝜀 (𝑥, 𝑦, 𝜃) d𝜃 + 𝜑 𝑥 𝑗 D 𝑥 𝑗 e𝑖 𝜑 𝑏 𝜀 d𝜃
Γ 𝑗=1 Γ
∫ 𝑛
∑︁ ∫
= e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 𝜀 (𝑥, 𝑦, 𝜃) d𝜃 + D𝑥 𝑗 e𝑖 𝜑 𝜑 𝑥 𝑗 𝑏 𝜀 d𝜃
Γ 𝑗=1 Γ
𝑛 ∫
∑︁
− e𝑖 𝜑 D 𝑥 𝑗 𝜑 𝑥 𝑗 𝑏 𝜀 d𝜃.
𝑗=1 Γ

It follows directly from Definitions 18.1.1 and 18.1.4 that 𝜑 𝑥 𝑗 𝑏 is an amplitude of

order ≤ 𝑚 − 1 in Ω1′ × Ω2′ × Γ. It is not difficult to check that the amplitudes 𝜑 𝑥 𝑗 𝑏 𝜀

and D 𝑥 𝑗 𝜑 𝑥 𝑗 𝑏 𝜀 form convergent sequences in 𝑆c𝑚−1 Ω1′ × Ω2′ , Γ as 𝜀 ↘ 0. If we

assume that the claim has been proved for amplitudes of order ≤ 𝑚 − 1 we conclude
that the integrals
∫ ∫ ∫
e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 𝜀 (𝑥, 𝑦, 𝜃) d𝜃, e𝑖 𝜑 𝜑 𝑥 𝑗 𝑏 𝜀 d𝜃, e𝑖 𝜑 D 𝑥 𝑗 𝜑 𝑥 𝑗 𝑏 𝜀 d𝜃,
Γ Γ Γ

converge in D ′ Ω1′ × Ω2 . The same is therefore true of 𝐴 𝜀 , proving the claim for
′

amplitudes of order ≤ 𝑚. Letting Ω′𝜄 ↗ Ω 𝜄 , 𝜄 = 1, 2, we conclude that 𝐴 𝜀 converges
to 𝐴 in D ′ (Ω1 × Ω2 ).
The operation leading to (18.1.23) can be repeated to prove that, whatever the
integer 𝑘 ≥ 0, we have
∑︁ ∫
𝑁
(2𝜋) 𝐴 𝜀 (𝑥, 𝑦) = D𝑥𝛼
e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 (𝜀𝛼) (𝑥, 𝑦, 𝜃) d𝜃
| 𝛼 | ≤𝑘 Γ

where, for each 𝛼, 𝑏 (𝜀𝛼) form a convergent sequence in 𝑆c𝑚−𝑘 (Ω1 × Ω2 , Γ); let 𝑏 ( 𝛼)
be its limit. We can take 𝑘 > 𝑁 + 𝑚 + 𝜈 with 𝜈 ∈ Z+ arbitrarily large; we obtain (cf.
Proposition 18.1.10)
∫
(2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 ( 𝛼) (𝑥, 𝑦, 𝜃) d𝜃 ∈ C 𝜈 (Ω1 × Ω2 ) .
Γ
∫
For arbitrary 𝑢 ∈ Cc∞ (Ω1 ) the duality bracket R𝑛 𝐴 (𝑥, 𝑦) 𝑢 (𝑥) d𝑥 is equal (in Ω2 )
to
∑︁ ∫ ∫
(2𝜋) −𝑁 (−1) | 𝛼 | e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 ( 𝛼) (𝑥, 𝑦, 𝜃) D 𝑥𝛼 𝑢 (𝑥) d𝜃d𝑥 ∈ C 𝜈 (Ω2 ) .
|𝛼|≤𝑝 R𝑛 Γ

∫
This proves that R𝑛 𝐴 (𝑥, 𝑦) 𝑢 (𝑥) d𝑥 ∈ C ∞ (Ω2 ). Exchanging 𝑥 and 𝑦 leads to the
∫
result that R𝑛 𝐴 (𝑥, 𝑦) 𝑣 (𝑦) d𝑦 ∈ C ∞ (Ω1 ) whatever 𝑣 ∈ Cc∞ (Ω2 ). The conjunction
of the last two properties proves that 𝐴 is semiregular. □
18.1 Fourier Distribution Kernels in Euclidean Space 695

We shall avail ourselves repeatedly of the following observation (cf. Proposition

16.1.16):

Proposition 18.1.14 If 𝐴 ∈ C ∞ (Ω1 × Ω2 ) there is an 𝑎 ∈ 𝑆c−∞ (Ω1 × Ω2 , Γ) [cf.

Definition 18.1.1 and (18.1.16)] such that (18.1.13) is valid.
−𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝐴 (𝑥, 𝑦) 𝜒 (𝜃) where 𝜒 ∈ C ∞ (Γ) is such that
−𝑁
∫ 𝑎 (𝑥, 𝑦, 𝜃) = e
Proof Take c
(2𝜋) 𝜒 (𝜃) d𝜃 = 1. □
In the sequel we shall refer to an oscillatory integral such as (18.1.13) as a Fourier
distribution kernel.

18.1.5 Action of pseudodifferential operators on Fourier

distribution kernels

We adapt to the present situation Definition 16.1.12:

Definition 18.1.15 The amplitude 𝑎 (𝑥, 𝑦, 𝜃) ∈ 𝑆 (Ω1 × Ω2 , Γ) will be said to be

properly supported if it satisfies the following conditions:
To each compact set 𝐾1 ⊂ Ω1 there is a compact set 𝐾2 ⊂ Ω2 such that

∀𝑥 ∈ 𝐾1 , ∀ (𝑦, 𝜃) ∈ Ω2 × Γ, 𝑎 (𝑥, 𝑦, 𝜃) ≠ 0 =⇒ 𝑦 ∈ 𝐾2 ,

and to each compact set 𝐾2 ⊂ Ω2 there is a compact set 𝐾1 ⊂ Ω1 such that

∀𝑦 ∈ 𝐾2 , ∀ (𝑥, 𝜃) ∈ Ω1 × Γ, 𝑎 (𝑥, 𝑦, 𝜃) ≠ 0 =⇒ 𝑥 ∈ 𝐾1 .

Remark 18.1.16 Obviously, if 𝑎 (𝑥, 𝑦, 𝜃) ∈ 𝑆c (Ω1 × Ω2 , Γ) is properly supported

the same is true of the distribution kernel (18.1.13) (Definition 2.3.6). The converse
is not true, as shown by
∫
−𝑛
𝛿 (𝑥 − 𝑦) = (2𝜋) e𝑖 ( 𝑥−𝑦) · 𝜃 d𝜃. (18.1.24)
R𝑛

What is true is that, if 𝐴 (𝑥, 𝑦) given by (18.1.13) is properly supported, then

there is a Fourier integral representation of the same type as (18.1.13) with a
properly supported amplitude: it suffices to select a properly supported function
𝜒 ∈ C ∞ (Ω1 × Ω2 ) such that 𝜒 (𝑥, 𝑦) = 1 in an open set containing supp 𝐴 and to
replace 𝑎 (𝑥, 𝑦, 𝜃) by 𝜒 (𝑥, 𝑦) 𝑎 (𝑥, 𝑦, 𝜃).

Assuming that 𝑎 is properly supported we shall let a standard pseudodifferential

′
operator in Ω1 with symbol 𝑃 (𝑥, 𝜉) ∈ 𝑆 𝑚 (Ω1 ) (Definitions 16.1.14, 16.2.1) act on
𝐴 (𝑥, 𝑦):
∫ ∫
′ d𝜉
𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) = e𝑖 ( 𝑥−𝑥 ) · 𝜉 𝑃 (𝑥, 𝜉) 𝐴 (𝑥 ′, 𝑦) d𝑥 ′ . (18.1.25)
R𝑛 Ω1 (2𝜋) 𝑛
696 18 Fourier Integral Operators

The integral
∫ ∫ ∫
′ ′ d𝜉d𝜃
𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) = e𝑖 ( 𝑥−𝑥 ) · 𝜉 +𝑖 𝜑 ( 𝑥 ,𝑦, 𝜃) 𝑃 (𝑥, 𝜉) 𝑎 (𝑥 ′, 𝑦, 𝜃) d𝑥 ′
Γ R𝑛 Ω1 (2𝜋) 𝑛+𝑁
(18.1.26)
makes sense, i.e., 𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) ∈ D ′ (Ω1 × Ω2 ).

Theorem 18.1.17 Let 𝐴 (𝑥, 𝑦) be given by (18.1.13) with

𝑎 (𝑥, 𝑦, 𝜃) ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ)

properly supported. If 𝑷 is a pseudodifferential operator in Ω1 with symbol 𝑃 (𝑥, 𝜉) ∈

′
𝑆 𝑚 (Ω1 × R𝑛 ) then
∫
𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 (𝑥, 𝑦, 𝜃) d𝜃 (18.1.27)
Γ
′
is a Fourier distribution kernel with phase 𝜑 and amplitude 𝑏 ∈ 𝑆c𝑚+𝑚 (Ω1 × Ω2 , Γ).

Proof We begin by reasoning formally, disregarding the convergence questions in

the Taylor expansions:
∑︁ 1
𝑎 (𝑥 ′, 𝑦, 𝜃) = (𝑥 ′ − 𝑥) 𝛼 𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃)
𝛼∈Z𝑛
𝛼!
+

and
∫ ∫
′ ′
e𝑖 ( 𝑥−𝑥 ) · 𝜉 +𝑖 𝜑 ( 𝑥 ,𝑦, 𝜃) 𝑃 (𝑥, 𝜉) 𝑎 (𝑥 ′, 𝑦, 𝜃) d𝑥 ′d𝜉
R𝑛 Ω1
∑︁ (−1) | 𝛼 | ∫ ∫ ′ ′

= 𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃) 𝑃 (𝑥, 𝜉) D 𝛼𝜉 e𝑖 ( 𝑥−𝑥 ) · 𝜉 +𝑖 𝜑 ( 𝑥 ,𝑦, 𝜃) d𝑥 ′d𝜉
𝛼∈Z+𝑛
𝛼! R𝑛 Ω1
∑︁ 1 ∫ ∫
′ ′
= 𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃) e𝑖 ( 𝑥−𝑥 ) · 𝜉 +𝑖 𝜑 ( 𝑥 ,𝑦, 𝜃) D 𝛼𝜉 𝑃 (𝑥, 𝜉) d𝑥 ′d𝜉.
𝛼∈Z𝑛
𝛼! R𝑛 Ω1
+

We can also write

𝜑 (𝑥 ′, 𝑦, 𝜃) = 𝜑 (𝑥, 𝑦, 𝜃) − (𝑥 − 𝑥 ′) · ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃) (18.1.28)

with ℎ1 ∈ C ∞ (Ω1 × Γ) real-valued and homogeneous of degree 1. We obtain the

formal expansions
∫ ∫
−𝑖 𝜑 ( 𝑥,𝑦, 𝜃) ′ ′
e e𝑖 ( 𝑥−𝑥 ) · 𝜉 +𝑖 𝜑 ( 𝑥 ,𝑦, 𝜃) D 𝛼𝜉 𝑃 (𝑥, 𝜉) d𝑥 ′d𝜉
R𝑛 Ω1
∫ ∫
′
= e𝑖 ( 𝑥−𝑥 ) · 𝜉 D 𝛼𝜉 𝑃 (𝑥, 𝜉 + ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃)) d𝑥 ′d𝜉
R𝑛 Ω1
18.1 Fourier Distribution Kernels in Euclidean Space 697
∫ ∫
∑︁ 1 ′
e𝑖 ( 𝑥−𝑥 ) · 𝜉 𝜉 𝛽 D 𝜉 𝑃 (𝑥, ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃)) d𝑥 ′d𝜉
𝛼+𝛽
=
𝛼,𝛽 ∈Z+𝑛
𝛽! R𝑛 Ω1
∫ ∫
∑︁ 1 ′

e𝑖 ( 𝑥−𝑥 ) · 𝜉 D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃)) d𝑥 ′d𝜉.
𝛽 𝛼+𝛽
=
𝛼,𝛽 ∈Z+𝑛
𝛽! R𝑛 Ω1

′
We know that (2𝜋) −𝑛
∫
R𝑛
e𝑖 ( 𝑥−𝑥 ) · 𝜉 d𝜉 = 𝛿 (𝑥 − 𝑥 ′), the Dirac distribution. We con-
clude that the integral
∫ ∫
−𝑛 −𝑖 𝜑 ( 𝑥,𝑦, 𝜃) ′ ′
(2𝜋) e e𝑖 ( 𝑥−𝑥 ) · 𝜉 +𝑖 𝜑 ( 𝑥 ,𝑦, 𝜃) 𝑃 (𝑥, 𝜉) 𝑎 (𝑥 ′, 𝑦, 𝜃) d𝑥 ′d𝜉
R𝑛 Ω1

is associated with the formal symbol (cf. Subsection 16.2.2)

∑︁ 1
D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃))
𝛽 𝛼+𝛽
𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃) . (18.1.29)
𝛼!𝛽! 𝑥 ′ =𝑥
𝛼,𝛽 ∈Z+𝑛

Notice that, for fixed 𝛼, 𝛽 ∈ Z+𝑛 ,

′
D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃)) 𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃) ∈ 𝑆 𝑚+𝑚 − | 𝛼+𝛽 | (Ω1 × Γ) .
𝛽 𝛼+𝛽
𝑥 ′ =𝑥

In order to produce a true symbol out of (18.1.29) we follow the prescription in

Subsection 16.2.2, culminating in Formula (16.2.11), the only difference being that
𝜃 ∈ Γ must now replace 𝜉 ∈ R𝑛 \ {0}. We end up with a series
∞
∑︁ ∑︁ 1 𝛽 𝛼+𝛽
𝜒 𝜃/𝑅 𝑗 D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃)) 𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃)
𝛼!𝛽! 𝑥 ′ =𝑥
𝑗=0 | 𝛼+𝛽 |= 𝑗

′
that converges in 𝑆c𝑚+𝑚 (Ω1 × Γ), to an amplitude 𝑏 ♮ (𝑥, 𝑦, 𝜃). Proposition 16.2.16,
allows us to conclude from this and from (18.1.26) that
∫
𝑃 𝐴 (𝑥, 𝑦) − (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑏 ♮ (𝑥, 𝑦, 𝜃) d𝜃 = 𝑔 (𝑥, 𝑦) ∈ C ∞ (Ω1 ) .
Γ

We can write ∫
𝑔 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑐 (𝑥, 𝑦, 𝜃) d𝜃,
Γ

where 𝑐 (𝑥, 𝑦, 𝜃) = e−𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑔

(𝑥, 𝑦) 𝜒 (𝜃) with 𝜒 ∈ Cc∞ (Γ) as in the proof of
Proposition 18.4.23. We may then take 𝑏 = 𝑏 ♮ + 𝑐 in (18.1.27). □
It is interesting to look at the amplitudes of the Fourier distribution kernels of the
type 𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) or, at least, at their formal expressions. From (18.1.29) we
derive that a formal amplitude for 𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) is
698 18 Fourier Integral Operators
∑︁ 1
𝑏˜ (𝑥, 𝑦, 𝜃) = D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃))
𝛽 𝛼+𝛽
𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃) ,
𝛼!𝛽! 𝑥 ′ =𝑥
𝛼,𝛽 ∈Z+𝑛
(18.1.30)
where [cf. (18.1.28)]
∫ 1
ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃) = 𝜑 𝑥 (𝑡𝑥 + (1 − 𝑡) 𝑥 ′, 𝑦, 𝜃) d𝑡 (18.1.31)
0

in a neighborhood of the “diagonal” 𝑥 = 𝑥 ′. Note that ℎ1 (𝑥, 𝑥, 𝑦, 𝜃) = 𝜑 𝑥 (𝑥, 𝑦, 𝜃).

18.2 The Lagrangian Manifold Associated to a Phase-function

18.2.1 The Lagrangian submanifold associated to a phase-function.

Definition

In this subsection we go back to the critical set Σ 𝜑 of the phase-function 𝜑 (Definition

18.1.3) and to Diagram (18.1.19). We introduce the symplectic one-form 𝜎1 = 𝜉 · d𝑥
on 𝑇 ∗ Ω1 and 𝜎2 = 𝜂 · d𝑦 on 𝑇 ∗ Ω2 . We equip 𝑇 ∗ Ω1 × 𝑇 ∗ Ω2 𝑇 ∗ (Ω1 × Ω2 ) with the
symplectic structure defined by the one-form 𝝈 = 𝜎1 − 𝜎2 or, if one prefers, by the
two-form
𝑛
∑︁
𝝕 = 𝑑𝝈 = d𝜉 𝑗 ∧ d𝑥 𝑗 − d𝜂 𝑗 ∧ d𝑦 𝑗 .
𝑗=1

In the sequel we refer to 𝝈 (resp., 𝝕) as the twisted symplectic one-form (resp.,

two-form) on 𝑇 ∗ (Ω1 × Ω2 ). According to (18.1.15) the pullback of 𝜉 · d𝑥 to Σ 𝜑 via
𝑝 𝜑 is the one-form d 𝑥 𝜑 whereas the pullback of 𝜂 · d𝑦 via 𝑞 𝜑 is the one-form −d 𝑦 𝜑.
The pullback to 𝚺 𝜑 via 𝑝 𝜑 , 𝑞 𝜑 of the one-form 𝜎 is equal to d 𝑥 𝜑 + d 𝑦 𝜑 = d𝜑 since
d 𝜃 𝜑 ≡ 0 on 𝚺 𝜑 .
We turn our attention to the image Λ 𝜑 = 𝑝 𝜑 , 𝑞 𝜑 Σ 𝜑 (Proposition 18.1.8).
Proposition 18.2.1 If 𝜑 is a phase-function in Ω1 × Ω2 × Γ (Definition 18.1.4) then
Λ 𝜑 is an immersed Lagrangian submanifold (Definition 13.3.15) of (𝑇 ∗ Ω1 \0) ×
(𝑇 ∗ Ω2 \0).
Proof Let (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ Λ 𝜑 be arbitrary. If 𝜃 ◦ ∈ R 𝑁 satisfies the equations
𝜑 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 and

𝜉 ◦ = 𝜑 𝑥 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) , 𝜂◦ = −𝜑 𝑦 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ,

then, by Property (2) in Definition 18.1.4, there is a 2𝑛-dimensional submanifold L of

Σ 𝜑 such that (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈ L and the restriction of 𝑝 𝜑 , 𝑞 𝜑 to L is a diffeomorphism

onto a neighborhood N of (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) in Λ 𝜑 ; N is a 2𝑛-dimensional smooth
submanifold of (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0). It ensues that the pushforward under 𝑝 𝜑 , 𝑞 𝜑
of d𝜑| L is a closed one-form on N where it coincides with 𝝈| Λ 𝜑 , whence 𝝕 ≡ 0 in
N. □
18.2 The Lagrangian Manifold Associated to a Phase-function 699

Remark 18.2.2 The proof of Proposition 18.2.1 does not rely on the fact that 𝚲 𝜑 is
conic nor on the fact that the pullback of d𝜑 to 𝚺 𝜑 vanishes identically (since 𝜑 ≡ 0
on Σ 𝜑 ).

Definition 18.2.3 We shall refer to Λ 𝜑 as the Lagrangian submanifold defined by

(or associated to) the phase-function 𝜑.

Proposition 18.2.4 Let 𝜑 be a phase-function in Ω1 × Ω2 × Γ. If 𝐾 is a conic subset

of Σ 𝜑 such that 𝐾 ∩ Ω1 × Ω2 × S𝑁 −1 is compact then the restriction of 𝑝 𝜑 , 𝑞 𝜑

to 𝐾 is a proper map onto 𝑝 𝜑 , 𝑞 𝜑 (𝐾).

Proof We use polar coordinates in R 𝑁 \ {0} and write 𝜑 (𝑥, 𝑦, 𝜃) = |𝜃| 𝜓 (𝑥, 𝑦, 𝜔),
𝜔 = 𝜃/|𝜃|. The hypothesis and Condition (3), Definition 18.1.4, imply

|𝜓 𝑥 (𝑥, 𝑦, 𝜔)| ≥ 𝑐, 𝜓 𝑦 (𝑥, 𝑦, 𝜔) ≥ 𝑐,

𝑁 −1 . Let the sequence

n some 𝑐 > 0 oand all (𝑥, 𝑦, 𝜔) ∈ 𝐾 ∩ Ω1 × Ω2 × S
for
𝑥 (𝜈) , 𝑦 (𝜈) , 𝜔 (𝜈) in 𝐾 be such that 𝑝 𝜑 , 𝑞 𝜑 𝑥 (𝜈) , 𝑦 (𝜈) , 𝜔 (𝜈) converges
𝜈=1,2,... n o
to (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) and a subsequence of 𝑥 (𝜈) , 𝑦 (𝜈) , 𝜔 (𝜈) be such that 𝜔 (𝜈)
𝜈=1,2,...
converges in S 𝑁 −1 to 𝜔◦ ; it follows that the equations (18.1.20) are satisfied at
(𝑥 ◦ , 𝑦 ◦ , 𝜔◦ ) which must therefore belong to Σ 𝜑 hence to 𝐾 ∩ Ω1 × Ω2 × S 𝑁 −1 . □

Proposition 18.2.5 Suppose 𝜑 is a nondegenerate phase-function in Ω1 × Ω2 × Γ. If

−1
z }| {

𝐾 ⊂ Σ 𝜑 is as in Proposition 18.2.4 then the set 𝐾 ∩ 𝑝 𝜑 , 𝑞 𝜑 (𝑥, 𝜉, 𝑦, 𝜂) is finite for
every (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑 .

Proof If the phase-function 𝜑 is nondegenerate 𝑝 𝜑 , 𝑞 𝜑 is a local diffeomorphism

of Σ 𝜑 onto Λ 𝜑 (Proposition 18.1.8). By Proposition 18.2.4 𝑝 𝜑 , 𝑞 𝜑 𝐾 is a proper
map and, as a consequence, the preimages of points of 𝑝 𝜑 , 𝑞 𝜑 (𝐾) under this map,
being discrete, must be finite. □

18.2.2 The wave-front set of a Fourier distribution kernel

The theorem proved in this subsection is the first indication of the relevance of the
Lagrangian manifold Λ 𝜑 . Stronger evidence will be presented in Section 18.4.
The C ∞ wave-front set of a distribution was introduced in Definition 2.1.5. To
throw some light on the theorem proved in this subsection let us describe the wave-
front set of the distribution kernel of the identity transformation in R𝑛 , (18.1.24).
Obviously 𝑊 𝐹 (𝛿 (𝑥 − 𝑦)) = {0} off the diagonal of R𝑛𝑥 × R𝑛𝑦 . The Fourier transform
of 𝛿 (𝑥 − 𝑦) is the oscillatory integral
700 18 Fourier Integral Operators
∫
𝐹 (𝜉, 𝜂) = e−𝑖 ( 𝜉 +𝜂) ·𝑥 d𝑥
R𝑛
∫
2
= lim e−𝑖 ( 𝜉 +𝜂) · 𝑥−𝜀 | 𝑥 | d𝑥
𝜀↘0 R𝑛

𝑛/2 1 2
= lim (𝜋/𝜀) exp − |𝜉 + 𝜂| ,
𝜀↘0 4𝜀

obviously equal to zero if 𝜉 + 𝜂 ≠ 0. It is easily proved that

n o
𝑊 𝐹 (𝛿 (𝑥 − 𝑦)) = (𝑥, 𝑦, 𝜉, 𝜂) ∈ R2𝑛 𝑥,𝑦 × R 2𝑛
𝜉,𝜂 ; 𝑥 = 𝑦, 𝜉 = −𝜂 ≠ 0 .

Contrast this with the definition of Λ 𝜑 when 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃:

n o
Λ 𝜑 = (𝑥, 𝑦, 𝜉, 𝜂) ∈ R2𝑛 2𝑛
𝑥,𝑦 × R 𝜉 , 𝜂 \ {0} ; 𝑥 = 𝑦, 𝜉 = 𝜂 .

We now turn to the general case.

Theorem 18.2.6 Let Ω1 and Ω2 be two open subsets of R𝑛 and 𝐴 (𝑥, 𝑦) ∈

D ′ (Ω1 × Ω2 ) be given by (18.1.13) with 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) (Definition 18.1.2).
Then
(𝑥, 𝑦, 𝜉, 𝜂) ∈ 𝑊 𝐹 ( 𝐴) =⇒ (𝑥, 𝑦, 𝜉, −𝜂) ∈ Λ 𝜑 .

Proof By Corollary 18.1.11 there is no loss of generality if we assume 𝑎 (𝑥, 𝑦, 𝜃) ≡ 0

if |𝜃| < 𝑅, 𝑅 > 0 large to be chosen below. It is convenient to√change notation for
points in (𝑇 ∗√Ω1 \0) × (𝑇 ∗ Ω2 \0): we set 𝑧 = 𝑥 + 𝑖𝑦 ∈ Ω1 + −1Ω2 = ΩC , 𝜁 =
𝜉 +𝑖𝜂 ∈ R𝑛 + −1R𝑛 , |𝜉 | |𝜂| ≠ 0, and we identify Ω1 × Ω2 with ΩC , ((𝑥, 𝜉) , (𝑦, 𝜂)) ∈
(𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0) with (𝑧, 𝜁). We write (𝑧, 𝜃) for (𝑥, 𝑦, 𝜃); thus
∫
𝐴 (𝑧) = (2𝜋) −𝑁 e𝑖 𝜑 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃) d𝜃 (18.2.1)
Γ

(with no suggestion of holomorphy with respect to 𝑧!). We identify (𝑥, 𝜑 𝑥 ) , 𝑦, −𝜑 𝑦
with (𝑧, 2𝜑 𝑧 ) where 𝜑 𝑧 = 𝜕𝑧 𝜑 = 21 𝜑 𝑥 − 𝑖𝜑 𝑦 . In this notation

Σ 𝜑 = (𝑧, 𝜃) ∈ ΩC × Γ; 𝜑 𝜃 (𝑧, 𝜃) = 0 .
◦ ◦ ◦
Let (𝑧 , 𝜁 ) ∈ 𝑊 𝐹 ( 𝐴), |𝜁 | = 1. We are going to show that the hypothesis
𝑧◦ , 𝜁 ◦ ∉ Λ 𝜑 leads to a contradiction. This hypothesis is equivalent to the following:

∀𝜃 ∈ Γ, 2𝜑 𝑧 (𝑧◦ , 𝜃) = 𝜁 ◦ =⇒ (𝑧◦ , 𝜃) ∉ Σ 𝜑 . (18.2.2)

In turn, (18.2.2) is equivalent to the existence of a conic neighborhood U of Σ 𝜑 in

ΩC × Γ such that (𝑧◦ , 𝜃) ∈ U =⇒ 2𝜑 𝑧 (𝑧 ◦ , 𝜃) ≠ 𝜁 ◦ . Obviously, the latter remains
true if we contract U about Σ 𝜑 . Then Property (3), Definition 18.1.4, allows us to
select U such that
18.2 The Lagrangian Manifold Associated to a Phase-function 701

∀ (𝑧, 𝜃) ∈ U, |𝜑 𝑥 (𝑧, 𝜃)| × 𝜑 𝑦 (𝑧, 𝜃) ≠ 0. (18.2.3)

Let 𝜒 ∈ C ∞ ΩC × Γ be homogeneous of degree zero, supp 𝜒 ⊂ U, 𝜒 ≡ 1 in a

neighborhood of Σ 𝜑 . Proposition 18.1.12 implies
∫
e𝑖 𝜑 (𝑧, 𝜃) (1 − 𝜒 (𝑧, 𝜃)) 𝑎 (𝑧, 𝜃) d𝜃 ∈ C ∞ ΩC .
Γ

As a consequence, (𝑧◦ , 𝜁 ◦ ) ∈ 𝑊 𝐹 ( 𝐴) ⇐⇒ (𝑧◦ , 𝜁 ◦ ) ∈ 𝑊 𝐹 𝐴 𝜒 where

∫
𝐴 𝜒 (𝑧) = (2𝜋) −𝑁 e𝑖 𝜑 (𝑧, 𝜃) 𝜒 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃) d𝜃.
R𝑁

In the remainder of the proof we shall assume (𝑧, 𝜃) ∈ supp ( 𝜒𝑎), 𝑧 ∈ 𝐾 C , 𝜃 ∈ Γ,

with 𝐾 C a compact subset of ΩC whose interior contains 𝑧 ◦ and whose size can be
decreased as needed.
By (18.2.3) there is a 𝑐 ◦ ∈ (0, 1) such that |𝜑 𝑧 (𝑧◦ , 𝜃)| ≥ 𝑐 ◦ |𝜃|, implying, if
|𝜃| ≥ 𝑐−1
◦ ,

2𝜑 𝑧 (𝑧 ◦ , 𝜃) − 𝜁 ◦ ≥ 2 |𝜑 𝑧 (𝑧◦ , 𝜃)| − 1 ≥ |𝜑 𝑧 (𝑧◦ , 𝜃)| ≥ 𝑐 ◦ |𝜃| .

At this stage we select 𝑅 ≥ max 𝑐−1 ◦

◦ , 1 , ensuring that 𝜑 𝑧 (𝑧 , 𝜃) < 𝑐 ◦ |𝜃| =⇒
◦
𝑎 (𝑧 , 𝜃) = 0.
Let ℭ be an open cone in C𝑛 \ {0} such that 𝜁 ◦ ∈ ℭ ∩ S2𝑛−1 ; if diam ℭ ∩ S2𝑛−1

is sufficiently small then
𝜁 1
− 𝜁 ◦ < 𝑐◦ .
|𝜁 | 2
If we combine the last two inequalities we get (for |𝜃| ≥ 𝑅 ≥ 1)

𝜁 1
− 2𝜑 𝑧 (𝑧 ◦ , 𝜃) ≥ 𝑐 ◦ |𝜃| . (18.2.4)
|𝜁 | 2

Let then 𝑓 ∈ C ∞ ΩC , supp 𝑓 ⊂ 𝐾 C , 𝑓 (𝑧) = 1 if |𝑧 − 𝑧◦ | < 𝜀, 0 < 𝜀 <

dist(𝑧 ◦ , 𝜕𝐾 C ), and consider the Fourier transform
∫
◦
𝐹 (𝜁) = (2𝜋) −𝑁
e−𝑖 Re ( 𝜁 · (𝑧−𝑧 ) ) 𝑓 (𝑧) 𝐴 𝜒 (𝑧) d𝑥d𝑦 (18.2.5)
2𝑛
∫R ∫
◦
= (2𝜋) −𝑁 e𝑖 𝜑 (𝑧, 𝜃)−𝑖 Re ( 𝜁 · (𝑧−𝑧 ) ) 𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜃) d𝜃d𝑥d𝑦.
R2𝑛 Γ

Taking 𝐾 C convex we can write

𝜑 (𝑧, 𝜃) = 𝜑 (𝑧◦ , 𝜃) + Re ((𝑧 − 𝑧◦ ) · ℎ (𝑧, 𝜃))

where
702 18 Fourier Integral Operators
∫ 1
ℎ (𝑧, 𝜃) = 2 𝜑 𝑧 (𝑧 ◦ + 𝑡 (𝑧 − 𝑧◦ ) , 𝜃) d𝑡 ∈ C𝑛 .
0
We derive from (18.2.3) that

Re ℎ (𝑧◦ , 𝜃) = 𝜑 𝑥 (𝑧◦ , 𝜃) ≠ 0, Im ℎ (𝑧◦ , 𝜃) = −𝜑 𝑦 (𝑧 ◦ , 𝜃) ≠ 0,

and from (18.2.4) that, if diam 𝐾 C and diam ℭ∩S2𝑛−1 are sufficiently small, then

𝜁 1
− ℎ (𝑧, 𝜃) ≥ 𝑐 ◦ |𝜃| (18.2.6)
|𝜁 | 4

for all (𝑧, 𝜃) ∈ 𝐾 C ×Γ such that (𝑧, 𝜃) ∈ supp ( 𝜒𝑎), and all 𝜁 ∈ ℭ.
After the change of variables 𝜃 ⇝ 𝜆𝜃, 𝜆 = |𝜁 |, the integral (18.2.5) becomes
∫ ∫
◦
𝐹 (𝜁) = 𝜆 𝑁
e𝑖𝜆𝜓 (𝑧,𝜁 , 𝜃) 𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜆𝜃) e𝑖𝜆𝜑 (𝑧 , 𝜃) d𝜃d𝑥d𝑦 (18.2.7)
R2𝑛 Γ

where
𝜁¯

𝜓 (𝑧, 𝜁, 𝜃) = Re ℎ (𝑧, 𝜃) − · (𝑧 − 𝑧◦ ) .
|𝜁 |
We have
𝜁¯
𝜓 𝑧 (𝑧, 𝜁, 𝜃) = ℎ (𝑧, 𝜃) − + 𝑂 (|𝑧 − 𝑧 ◦ |) .
|𝜁 |
From (18.2.6) we deduce the following: there is a 𝐶 > 0 such that, provided 𝜁 ∈ ℭ
and (𝑧, 𝜃) ∈ R2𝑛 × Γ satisfy 𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜆𝜃) ≠ 0, then
1
|𝜓 𝑧 (𝑧, 𝜁, 𝜃)| ≥ 𝑐 ◦ |𝜃| − 𝐶 |𝑧 − 𝑧◦ | |𝜃| .
4
We can contract 𝐾 C about 𝑧 ◦ (and therefore also supp 𝑓 and decrease 𝜀) so that we
have, for those same 𝑧, 𝜃, 𝜁,
1
|𝜓 𝑧 (𝑧, 𝜁, 𝜃)| ≥ 𝑐 ◦ |𝜃| . (18.2.8)
8
We introduce the following differential operator in 𝐾 C :
𝑛
∑︁ 𝜕𝜓 𝜕 𝜕𝜓 𝜕
𝐿 (𝑧 , 𝜁, 𝜃, D𝑧 ) = 1 + (𝑧, 𝜁, 𝜃) + (𝑧, 𝜁, 𝜃) .
𝑗=1
𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝑦 𝑗 𝜕𝑦 𝑗

We derive from (18.2.7):

18.2 The Lagrangian Manifold Associated to a Phase-function 703

𝜆−𝑁 𝐹 (𝜁)
∫ ∫ 𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜆𝜃) ◦
= 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 ) e𝑖𝜆𝜓 (𝑧,𝜁 , 𝜃) 2
e𝑖𝜆𝜑 (𝑧 , 𝜃) d𝜃d𝑥d𝑦
R2𝑛 Γ 1 + 𝜆 |𝜓 𝑧 (𝑧, 𝜁, 𝜃)|
∫ ∫
𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜆𝜃) 𝑖𝜆𝜑 (𝑧 ◦ , 𝜃)
= e𝑖𝜆𝜓 (𝑧,𝜁 , 𝜃) 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 ) ⊤ e d𝜃d𝑥d𝑦
R2𝑛 Γ 1 + 𝜆 |𝜓 𝑧 (𝑧, 𝜁, 𝜃)| 2

where 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 ) ⊤ is the transpose of 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 ). The inequality (18.2.8) and

the homogeneity of degree 1 with respect to 𝜃 of the leading part of 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 )
imply that
⊤ 𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜆𝜃)
𝑏 1 (𝑧, 𝜃, 𝜆) = 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 )
1 + 𝜆 |𝜓 𝑧 (𝑧, 𝜁, 𝜃)| 2
is a symbol of the same type as 𝑓 (𝑧) ( 𝜒𝑎) (𝑧, 𝜆𝜃) but of degree ≤ 𝑚 − 1 with respect
to 𝜆𝜃. We can repeat the operation as many times as we wish; for every 𝜈 = 2, 3...,
we define the symbol 𝑏 𝜈 (𝑧, 𝜃, 𝜆) by the formula

𝑏 𝜈−1 (𝑧, 𝜆𝜃)
𝑏 𝜈 (𝑧, 𝜃, 𝜆) = 𝐿 (𝑧, 𝜁, 𝜃, D𝑧 ) ⊤ .
1 + 𝜆 |𝜓 𝑧 (𝑧, 𝜁, 𝜃)| 2
We have
∫ ∫
◦ , 𝜃)
𝜆−𝑁 𝐹 (𝜁) = e𝑖𝜆𝜓 (𝑧,𝜁 , 𝜃) e𝑖𝜆𝜑 (𝑧 𝑏 𝜈 (𝑧, 𝜃, 𝜆) d𝜃d𝑥d𝑦.
R2𝑛 Γ

Since |𝑏 𝜈 (𝑧, 𝜃, 𝜆)| ≤ const.𝜆 𝑚 (1 + 𝜆 |𝜃|) −𝜈 we conclude that to every 𝜈 ≥ 𝑁 + 1

𝑁 +1−𝜈
there is a 𝐶𝜈 > 0 such that |𝐹 (𝜁)| ≤ 𝐶𝜈 |𝜁◦ | ◦ if 𝜁 ∈ ℭ.
This proves that
𝑊 𝐹 𝑓 𝐴 𝜒 ∩ 𝐾 × ℭ = ∅ and therefore (𝑧 , 𝜁 ) ∉ 𝑊 𝐹 𝐴 𝜒 , contradicting our
C

hypothesis. □

Remark 18.2.7 Of course we can have (𝑥, 𝑦, 𝜉, −𝜂) ∈ Λ 𝜑 while (𝑥, 𝑦, 𝜉, 𝜂) ∉

𝑊 𝐹 ( 𝐴); for instance, the amplitude 𝑎 (𝑥, 𝑦, 𝜃) could vanish identically in an open
subset of Ω1 × Ω2 .

18.2.3 Strongly nondegenerate phase-functions. Local symplectic

graphs

Clearly the fact that (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ (𝑇 ∗ Ω1 \0)×(𝑇 ∗ Ω2 \0) belongs to Λ 𝜑 establishes

an equivalence relation between points (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑇 ∗ Ω1 \0 and (𝑦 ◦ , 𝜂◦ ) ∈ 𝑇 ∗ Ω2 \0. In
general, the correspondence (𝑦 ◦ , 𝜂◦ ) ↦→ (𝑥 ◦ , 𝜉 ◦ ) cannot be equated to a map in the
usual sense, only to a multivalued map: there might be different points (𝑦, 𝜃) ∈ Ω2 ×Γ
such that

𝜉 ◦ = 𝜑 𝑥 (𝑥 ◦ , 𝑦, 𝜃) , 𝜂◦ = −𝜑 𝑦 (𝑥 ◦ , 𝑦, 𝜃) , 𝜑 𝜃 (𝑥 ◦ , 𝑦, 𝜃) = 0. (18.2.9)
704 18 Fourier Integral Operators

In this subsection we introduce a class of phase-functions such that said correspon-

dence is a map, at least locally.
𝜕𝜑
Let us denote by 𝜑 𝑥,𝑦 the 𝑛 × 𝑛 matrix whose columns are the vectors ∇ 𝑦 𝜕𝑥 ,
𝑗
𝜕𝜑
𝑗 = 1, ..., 𝑛; by 𝜑 𝑥, 𝜃 the 𝑁 × 𝑛 matrix whose columns are the 𝑛-vectors ∇ 𝑥 𝜕𝜃 ,
𝑘
𝜕𝜑
𝑘 = 1, ..., 𝑁; by 𝜑 𝑦, 𝜃 the 𝑛×𝑁 the matrix whose columns are the 𝑁-vectors ∇ 𝜃 𝜕𝑦 𝑗 ,
2
𝑗 = 1, ..., 𝑛; and finally by 𝜑 𝜃 , 𝜃 the Hessian matrix 𝜕𝜃𝜕𝑘 𝜕𝜃
𝜑
. We introduce
ℓ 1≤𝑘,ℓ ≤ 𝑁
the (𝑛 + 𝑁) × (𝑛 + 𝑁) matrix
𝜑 𝑥,𝑦 𝜑 𝑥, 𝜃
. (18.2.10)
𝜑 𝜃 ,𝑦 𝜑 𝜃 , 𝜃
Note that the determinant of (18.2.10) is homogeneous (with respect to 𝜃) of degree
zero.

Proposition 18.2.8 The following properties of the phase-function 𝜑 in Ω1 × Ω2 × Γ

are equivalent:
(a) the matrix (18.2.10) is nonsingular at every point of Σ 𝜑 ;
−1
(b) the phase-function 𝜑 is nondegenerate (Definition 18.1.6) and 𝑞 𝜑 ◦ 𝑝 𝜑 is a local
diffeomorphism of 𝑝 𝜑 Σ 𝜑 ⊂ 𝑇 ∗ Ω1 \0 onto 𝑞 𝜑 Σ 𝜑 ⊂ 𝑇 ∗ Ω2 \0.

Proof The matrix (18.2.10) can be regarded as the Jacobian matrix of (𝜑 𝑥 , 𝜑 𝜃 ) with
respect to (𝑦, 𝜃) or as the Jacobian matrix of 𝜑 𝑦 , 𝜑 𝜃 with respect to (𝑥, 𝜃). To
say that these Jacobian matrices are nonsingular at every point of Σ 𝜑 is equivalent
𝜕𝜑 𝜕𝜑
to saying that 𝑥 𝑗 , 𝜉 𝑘 = 𝜕𝑥 𝑘
, 𝜕𝜃ℓ
(1 ≤ 𝑗, 𝑘 ≤ 𝑛, 1 ≤ ℓ ≤ 𝑁) can be taken as local
coordinates in a neighborhood in Ω1 ×Ω2 ×Γ of each point of Σ 𝜑 ; and the same is true
𝜕𝜑
with (𝑦, 𝜂) replacing (𝑥, 𝜉). The functions 𝜕𝜃 𝑗
( 𝑗 = 1, ..., 𝑁) become the coordinates

𝜕𝜑
transversal to Σ 𝜑 in a neighborhood of Σ 𝜑 , implying that the differentials d 𝜕𝜃 ℓ
,
1 ≤ ℓ ≤ 𝑁, are linearly independent (thus 𝜑 is nondegenerate) and the 𝑥 𝑗 , 𝜉 𝑘 (or
the 𝑦 𝑗 , 𝜂 𝑘 ) can be regarded as local coordinates in Σ 𝜑 . The latter is equivalent to
saying that the maps 𝑝 𝜑 and 𝑞 𝜑 have (separately) rank 2𝑛 at every point of Σ 𝜑 ,
−1
hence 𝑞 𝜑 ◦ 𝑝 𝜑 is a local diffeomorphism of 𝑝 𝜑 Σ 𝜑 onto 𝑞 𝜑 Σ 𝜑 . This proves that
(a)=⇒(b). Reversing the preceding argument shows that (b)=⇒(a). □

Remark 18.2.9 As stated Proposition 18.2.8 relies decidedly on the fact that
dim Ω1 = dim Ω2 .

Definition 18.2.10 The phase-function 𝜑 shall be said to be strongly nondegenerate

if the (equivalent) conditions (a) and (b) in Proposition 18.2.8 are satisfied.

Example 18.2.11 The phase-function (𝑥 − 𝑦) ·𝜃 in R𝑛 ×R𝑛 × (R𝑛 \ {0}) (cf. Example

18.1.7 ) is strongly nondegenerate. The phase-function 𝜑 (𝑥, 𝑦, 𝜃) = 𝜃 1 𝑥 − 𝜃 2 𝑦 in R ×
R × R2 \ {0} is nondegenerate (cf. Example 18.1.9) but not strongly nondegenerate.

18.2 The Lagrangian Manifold Associated to a Phase-function 705

Let Λ be a Lagrangian submanifold of (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0) for the twisted

symplectic form 𝝈 = 𝜉 · d𝑥 − 𝜂 · d𝑦. Suppose that the following property holds:
(LSG) Given arbitrarily (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ Λ there are conic open sets U1 ⊂
𝑇 ∗ Ω1 \0, U2 ⊂ 𝑇 ∗ Ω2 \0, with (𝑥 ◦ , 𝜉 ◦ ) ∈ U1 , (𝑦 ◦ , 𝜂◦ ) ∈ U2 , such that
Λ ∩ (U1 × U2 ) is the graph of a diffeomorphism of U1 onto U2 .
The diffeomorphism of U1 onto U2 in (LSG) is perforce a symplectomorphism
since d𝜉 ∧ d𝑥 = d𝜂 ∧ d𝑦 on Λ.

Definition 18.2.12 If (LSG) holds we say that Λ is a local symplectic graph.

If (LSG) holds we can form the commutative diagram

(𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ
↙ ↘ (18.2.11)
𝑇 ∗ Ω1 \0 ∋ (𝑥, 𝜉) ↦−→ (𝑦, 𝜂) ∈ 𝑇 ∗ Ω2 \0.

where the lower horizontal arrow is the local diffeomorphism made up of all the
diffeomorphisms U1 −→ U2 in (LSG) as (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ranges over Λ.

Proposition 18.2.13 For the nondegenerate phase-function 𝜑 to be strongly non-

degenerate it is necessary and sufficient that Λ 𝜑 (Definition 18.2.3) be a local
symplectic graph.

Proof Since the map 𝑝 𝜑 , 𝑞 𝜑 is a local diffeomorphism of Σ 𝜑 onto Λ 𝜑 (Proposi-
tion 18.2.4) the lower horizontal arrow in (18.2.11) represents the (locally defined)
−1
composite 𝑞 𝜑 ◦ 𝑝 𝜑 . □

Example 18.2.14 If 𝑁 = 𝑛 and 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃 (cf. Example 18.1.7) then

the relation (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑 is the equality (𝑥, 𝜉) = (𝑦, 𝜂). We shall refer to this
phase-function 𝜑 as the phase-function of the identity.

Example 18.2.15 Assume that 𝑁 = 𝑛 and let 𝐻 ∈ C ∞ (Γ) be homogeneous of

degree 1. The phase-function 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃 − 𝐻 (𝜃) in R2𝑛 × Γ is strongly
nondegenerate. We have

Σ 𝜑 = (𝑥, 𝑦, 𝜃) ∈ R2𝑛 × Γ; 𝑥 − 𝑦 = ∇𝐻 (𝜃) ,

𝑝 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥, 𝜃) , 𝑞 𝜑 (𝑥, 𝑦, 𝜃) = (𝑦, 𝜃) .

The Lagrangian submanifold associated with 𝜑,

n o
Λ 𝜑 = (𝑥, 𝜉, 𝑦, 𝜂) ∈ (R𝑛 × Γ) 2 ; 𝑥 = 𝑦 + ∇𝐻 (𝜉) , 𝜉 = 𝜂 ,

is the graph of a symplectomorphism of R𝑛 × Γ onto itself.

The following simple result will be useful later.

706 18 Fourier Integral Operators

Lemma 18.2.16 If Λ 𝜑 is a local symplectic graph (Definition 18.2.12) then, for each
(𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑇 ∗ Ω1 \0, the set of points (𝑦, 𝜃) ∈ Ω2 × Γ satisfying

𝜑 𝑥 ◦ (𝑥 ◦ , 𝑦, 𝜃) = 𝜉 ◦ , 𝜑 𝜃 (𝑥 ◦ , 𝑦, 𝜃) = 0 (18.2.12)

is discrete.

Proof When Λ 𝜑 is a local symplectic graph an arbitrary solution (𝑦 ◦ , 𝜃 ◦ ) ∈ Ω2 × Γ

of (18.2.12) is isolated since 𝑞 𝜑 is a local diffeomorphism. □

18.3 Fourier Integral Operators. Basics

18.3.1 Fourier integral operators. Definition

We continue to assume 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) (Definition 18.1.2). According to the

Schwartz Kernels Theorem and to Theorem 18.1.13, 𝐴 given by (18.1.13) defines a
continuous linear operator 𝑨 : Cc∞ (Ω2 ) −→ C ∞ (Ω1 ):
∫
𝑨𝑢 (𝑥) = 𝐴 (𝑥, 𝑦) 𝑢 (𝑦) d𝑦 (18.3.1)
Ω2
∫ ∫
= (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (𝑥, 𝑦, 𝜃) 𝑢 (𝑦) d𝑦d𝜃,
Γ Ω2

where 𝑢 ∈ Cc∞ (Ω2 ).

Definition 18.3.1 The operator 𝑨 is called a Fourier integral operator (abbreviated

to FIO) in Ω1 × Ω2 (or simply in Ω when Ω1 = Ω2 = Ω).

The transpose of 𝑨 is given by

∫
⊤
𝑨 𝑢 (𝑥) = 𝐴 (𝑥, 𝑦) 𝑢 (𝑥) d𝑥 (18.3.2)
Ω1
∫ ∫
= (2𝜋) −𝑁 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (𝑥, 𝑦, 𝜃) 𝑢 (𝑥) d𝑥d𝜃,
Γ Ω1

where 𝑢 ∈ Cc∞
(Ω1 ). Theorem 18.1.13 implies that 𝑨C ∞ ∞
c (Ω2 ) ⊂ C (Ω1 ) and
⊤ ∞ ∞ ⊤⊤
𝑨 Cc (Ω1 ) ⊂ C (Ω2 ), the latter meaning that 𝑨 = 𝑨 extends as a continuous
linear operator E ′ (Ω2 ) −→ D ′ (Ω1 ). The adjoint of 𝑨 is given by
∫
𝑨∗ 𝑢 (𝑥) = 𝐴 (𝑥, 𝑦)𝑢 (𝑥) d𝑥 (18.3.3)
Ω1
∫ ∫
= (2𝜋) −𝑁 e−𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (𝑥, 𝑦, 𝜃)𝑢 (𝑥) d𝑥d𝜃.
Γ Ω1
18.3 Fourier Integral Operators. Basics 707

The Lagrangian submanifold Λ∗𝜑 defined by the phase-function of the FIO 𝑨∗ is

the image of Σ 𝜑 under the map (𝑥, 𝑦) ↦→ 𝑦, −𝜑 𝑦 , (𝑥, 𝜑 𝑥 ) ; when 𝜑 is strongly
nondegenerate Λ∗𝜑 is the local symplectic graph of the inverse of the map defined by
𝜑.
If 𝜑 is the phase-function of the identity (see Example 18.2.14) 𝑨 is a pseudod-
ifferential operator (Definition 16.1.14).
Proposition 18.1.14 shows that kernels 𝐴 (𝑥, 𝑦) ∈ C ∞ (Ω1 × Ω2 ) define FIOs; in
other words, smoothing operators (Definition 2.3.3) are Fourier integral operators
(with an arbitrary real phase-function); we continue to denote by Ψ−∞ (Ω1 × Ω2 )
the ring (with respect to addition and composition) of the smoothing operators in
Ω1 × Ω2 .
Proposition 18.3.2 Let 𝑨 be a Fourier integral operator in Ω1 × Ω2 defined by
the distribution kernel (18.1.13). There is a properly supported distribution kernel
(Definition 2.3.6) 𝐴♭ (𝑥, 𝑦) in Ω1 × Ω2 defining an operator 𝑨♭ in Ω1 × Ω2 such that
𝑨 − 𝑨♭ ∈ Ψ−∞ (Ω1 × Ω2 ).
Proof Indeed, select a properly supported amplitude (Definition 18.1.15)

𝜒 ∈ 𝑆 0 (Ω1 × Ω2 × Γ)

such that 𝜒 ≡ 1 in a conic neighborhood U of Σ 𝜑 in Ω1 × Ω2 × Γ. We take 𝑨♭ to be

defined by the distribution kernel (18.1.13) after we substitute 𝜒𝑎 for 𝑎. It follows
from Proposition 18.1.12 that (1 − 𝜒) 𝑎 defines a smoothing operator in Ω1 × Ω2 ,
obviously equal to 𝑨 − 𝑨♭ . □
Needless to say, the operator 𝑨♭ in Proposition 18.3.2 is not unique. We have
defined 𝑨♭ by an integral (18.1.13) with the same phase function 𝜑. It will often
be convenient to deal with equivalence classes of FIOs mod smoothing operators,
rather than with representatives of such cosets.

18.3.2 The effect of FIOs on wave-front sets

In the following statements 𝐴 (𝑥, 𝑦) is given by (18.1.13) with 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ)

properly supported.
′′
Proposition 18.3.3 Let 𝑄 (𝑦, 𝜂) ∈ 𝑆 𝑚 (Ω2 × R𝑛 ) be the symbol of a pseudodiffer-
ential operator 𝑸 in Ω2 . Then 𝑨𝑸 : E ′ (Ω2 ) −→ D ′ (Ω1 ) is also an FIO with phase
′′
𝜑 and amplitude 𝑏♭ ∈ 𝑆 𝑚+𝑚 (Ω1 × Ω2 , Γ).
⊤
Proof It suffices to transpose 𝑸 𝑦, D 𝑦 𝑨⊤ : E ′ (Ω1 ) −→ D ′ (Ω2 ) which is an
FIO of the same type as (18.3.2) according to Theorem 18.1.17. □
Corollary 18.3.4 If 𝑷 is a pseudodifferential operator in Ω1 of order 𝑚 ′ and 𝑸 is a
pseudodifferential operator in Ω2 of order 𝑚 ′′ then 𝑷 𝑨𝑸 : E ′ (Ω2 ) −→ D ′ (Ω1 ) is
′ ′′
also an FIO with phase 𝜑 and amplitude 𝑏 ♮ ∈ 𝑆 𝑚+𝑚 +𝑚 (Ω1 × Ω2 , Γ).
708 18 Fourier Integral Operators

Corollary 18.3.4 allows us to microlocalize the handling of FIOs and thus gain
some insight into the effect of the operator 𝑨 given by (18.3.1) on the C ∞ wave-front
set of 𝑢 ∈ D ′ (Ω2 ).
If 𝐸 ⊂ 𝑇 ∗ Ω2 \0 we denote by ℜ 𝜑 (𝐸) the set of points (𝑥, 𝜉) ∈ 𝑇 ∗ Ω1 \0 such
that (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑 for some (𝑦, 𝜂) ∈ 𝐸; we write ℜ 𝜑 (𝑦, 𝜂) if 𝐸 = {(𝑦, 𝜂)}. This
subsection is entirely devoted to the proof of the following result.

Theorem 18.3.5 Let 𝐴 ∈ D ′ (Ω1 × Ω2 ) be as in (18.1.13), 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ). If

the phase-function function 𝜑 is strongly nondegenerate (Definition 18.2.10) then

∀𝑢 ∈ E ′ (Ω2 ) , 𝑊 𝐹 ( 𝑨𝑢) ⊂ ℜ 𝜑 (𝑊 𝐹 (𝑢)) . (18.3.4)

Since the Lagrangian submanifold Λ 𝜑 is a local symplectic graph (Definition

18.2.12; see also Proposition 18.2.13) the correspondence (𝑦, 𝜂) → ℜ 𝜑 (𝑦, 𝜂) is a
local symplectomorphism. Extension to degenerate phase-functions is possible, see
[Hörmander, 1983, III], Ch. XXI (also [Treves, 1980], p. 417).
As we are talking about wave-front sets our conclusions will not be affected
by “modding off” C ∞ functions, thanks to Theorem 18.1.13. Proposition 18.1.12
and Corollary 18.1.11 allow us to assume that the support of the amplitude 𝑎 ∈
𝑆c𝑚 (Ω1 × Ω2 , Γ) in (18.3.1) is contained in a conic neighborhood of Σ 𝜑 in Ω1 ×Ω2 ×Γ
as “thin” as needed and that 𝑎 (𝑥, 𝑦, 𝜃) = 0 if |𝜃| < 1.
By Γ♭ we shall mean a cone in R 𝑁 \ {0}, Γ♭ ∩ S 𝑁 −1 ⊂⊂ Γ, such that supp 𝑎 ⊂
Ω1 ×Ω2 × Γ♭ . Since we are interested in the properties of 𝑨𝑢 in a conic neighborhood
of a given point (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑇 ∗ Ω1 \0 we might as well assume that the projection of
supp 𝑎 in (𝑥, 𝑦)-space is compact. Then the equation

𝜉 ◦ = 𝜑 𝑥 (𝑥 ◦ , 𝑦, 𝜃) (18.3.5)

imposes lower and upper bounds on |𝜃|. Since Λ 𝜑 is a local symplectic graph, Lemma
18.2.16 implies that there are finitely many points (𝑦, 𝜃) ∈ Ω2 × Γ♭ (and perhaps
none) such that (𝑥 ◦ , 𝑦, 𝜃) ∈ supp 𝑎 and (18.3.5) holds. We can select finitely many
C ∞ functions
Í 𝜓 𝜄 (𝑦, 𝜃) (𝜄 = 1, ..., 𝑟) in Ω2 × Γ, homogeneous of degree zero, such
that 𝑟𝜄=1 𝜓 𝜄 (𝑦, 𝜃) = 1 in a neighborhood of (supp 𝑢) × Γ♭ and such that there is at
most one point (𝑥 ◦ , 𝑦, 𝜃) ∈ Σ 𝜑 ∩ supp 𝜓 𝜄 satisfying (18.3.5). It suffices therefore to
prove (18.3.4) when 𝑎 (𝑥, 𝑦, 𝜃) = 𝑎 (𝑥, 𝑦, 𝜃) 𝜓 𝜄 (𝑦, 𝜃) for some 𝜄.
We begin by settling the cases where there are no points (𝑥 ◦ , 𝑦, 𝜃) ∈ Σ 𝜑 ∩ supp 𝑎
satisfying (18.3.5).

Lemma 18.3.6 If (𝑥 ◦ , 𝑦 ◦ ) ∈ Ω1 ×Ω2 is such that 𝜑 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃) ≠ 0 for all 𝜃 ∈ Γ then

there is an open set 𝑈2 ⊂ Ω2 containing 𝑦 ◦ such that 𝑥 ◦ ∉ singsupp 𝑨𝑢 whatever
𝑢 ∈ E ′ (𝑈2 ).

Proof The hypothesis implies that there are functions 𝑔 𝑗 ∈ Cc∞ Ω 𝑗 ( 𝑗 = 1, 2),

𝑔1 ≡ 1 in a neighborhood 𝑈1 of 𝑥 ◦ , 𝑔2 ≡ 1 in a neighborhood 𝑈2 of 𝑦 ◦ , such
that 𝑔1 (𝑥) 𝑎 (𝑥, 𝑦, 𝜃) 𝑔2 (𝑦) vanishes identically in a conic neighborhood of Σ 𝜑 in
Ω1 ×Ω2 × Γ. Proposition 18.1.12 implies that 𝑔1 (𝑥) 𝐴 (𝑥, 𝑦) 𝑔2 (𝑦) ∈ C ∞ (Ω1 × Ω2 ),
whence the claim. □
18.3 Fourier Integral Operators. Basics 709

Lemma 18.3.7 Assume 𝑨𝑢 is given by (18.3.1) and (𝑥, 𝑦, 𝜃) ∈ supp 𝑎 ⇒ 𝜃 ∈ Γ♭ .

Let (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 ∩ supp 𝑎 and ◦ ◦ ◦ ◦ ◦
𝜉 ∈ R \ {0} be such that 𝜉 ≠ 𝜑 𝑥 (𝑥 , 𝑦 , 𝜃 ).
𝑛

Then, provided diam Γ♭ ∩ S 𝑁 −1 is sufficiently small, there is an open set 𝑈2 ⊂ Ω2

containing 𝑦 ◦ such that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨𝑢) whatever 𝑢 ∈ E ′ (𝑈2 ).

Proof If diam Γ♭ ∩ S 𝑁 −1 is sufficiently small there are neighborhoods 𝑈1 ⊂ Ω1
of 𝑥 ◦ , 𝑈2 ⊂ Ω2 of 𝑦 ◦ , an open cone ℭ1 ⊂ R𝑛 \ {0} containing 𝜉 ◦ and a constant
𝑐 ◦ > 0 such that
𝜉 𝜑 𝑥 (𝑥, 𝑦, 𝜃)
− ≥ 𝑐◦ (18.3.6)
|𝜉 | |𝜑 𝑥 (𝑥, 𝑦, 𝜃)|
for all (𝑥, 𝑦) ∈ 𝑈1 × 𝑈2 , 𝜃 ∈ Γ♭ , 𝜉 ∈ ℭ1 . We take the open set 𝑈1 to be convex. If ℎ1
is the function (18.1.31) contractions of 𝑈1 × 𝑈2 about (𝑥 ◦ , 𝑦 ◦ ) and of ℭ1 about 𝜉 ◦
enable us to deduce from (18.3.6):

∀𝑥, 𝑥 ′ ∈ 𝑈1 , 𝑦 ∈ 𝑈2 , 𝜃 ∈ Γ♭ =⇒ ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃) ∉ ℭ1 . (18.3.7)

Let then 𝜒1 ∈ C ∞ (R𝑛 \ {0}) be a homogeneous function of degree zero, equal to 1

in a neighborhood of 𝜉 ◦ and to zero off ℭ1 ; we derive from (18.3.6):

∀𝑥, 𝑥 ′ ∈ 𝑈1 , 𝑦 ∈ 𝑈2 , 𝜃 ∈ Γ♭ =⇒ 𝜒1 (ℎ1 (𝑥, 𝑥 ′, 𝑦, 𝜃)) = 0. (18.3.8)

We apply (18.1.30) with 𝑃 (𝑥, 𝜉) = 𝑔1 (𝑥) 𝜒1 (𝜉), 𝑔1 ∈ Cc∞ (𝑈), 𝑔1 ≡ 1 in a neigh-

borhood of 𝑥 ◦ . The argument at the end of the proof of Theorem 18.1.17 allows us
to conclude that

supp 𝑢 ⊂ 𝑈2 =⇒ 𝑔1 (𝑥) 𝜒1 (D 𝑥 ) 𝑨𝑢 ∈ Cc∞ (𝑈1 ) .

Indeed, Formula (18.1.30) yields

∑︁ 1
( 𝛼+𝛽)

𝑏˜ (𝑥, 𝑦, 𝜃) = (ℎ (𝑥, 𝑥 ′, 𝑦, 𝜃))
𝛽
D 𝑥′ 𝜒1 𝑔1 (𝑥) 𝜕𝑥𝛼 𝑎 (𝑥, 𝑦, 𝜃) ,
𝛼!𝛽! 𝑥 ′ =𝑥
𝛼,𝛽 ∈Z+𝑛

which vanishes identically for 𝑦 ∈ Ω2 due to (18.3.8). We avail ourselves of the

principle that any true amplitude constructed (following the prescription in Subsec-
tion 16.2.2) out of a formal amplitude that vanishes identically, defines a smoothing
operator. Our choice of 𝑔1 and 𝜒1 implies (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨𝑢) by Theorem 16.2.20.□
Thanks to the Borel–Lebesgue Lemma, when there is no point (𝑥 ◦ , 𝑦, 𝜃) ∈ Σ 𝜑 ∩
supp 𝑎 satisfying (18.3.5) we can devise a C ∞ partition of unity 𝜒 𝜄′ (𝑦, 𝜃) (𝜄′ =
1, ..., 𝑟 ′) in a neighborhood of (supp 𝑢) × Γ♭ , homogeneous of degree zero, such that
either Lemma 18.3.6 or Lemma 18.3.7 applies, proving (18.3.4) in this case.
Lastly we deal with the case in which (18.3.5) has a (unique) solution (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈
Σ𝜑 .
710 18 Fourier Integral Operators

Lemma 18.3.8 Suppose that (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 and

𝜉 ◦ = 𝜑 𝑥 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) , 𝜂◦ = −𝜑 𝑦 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) . (18.3.9)

Let 𝑢 ∈ E ′ (Ω2 ) be such that (𝑦 ◦ , 𝜂◦ ) ∉ 𝑊 𝐹 (𝑢). Under these hypotheses [and

provided the cone Γ♭ in (18.3.1) is sufficiently contracted about the ray through 𝜃 ◦ ]
there is an open set 𝑈2 ⊂ Ω2 containing 𝑦 ◦ such that (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹 ( 𝑨 (𝑔2 𝑢))
whatever 𝑔2 ∈ Cc∞ (𝑈2 ).

Proof Let the open sets 𝑈 𝑗 ⊂ Ω 𝑗 and the open cone ℭ2 ⊂ R𝑛 \ {0} be such that
𝑥 ◦ ∈ 𝑈1 , (𝑦 ◦ , 𝜂◦ ) ∈ U2 = 𝑈2 × (−ℭ2 ); thus 𝜑 𝑦 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈ ℭ2 . We take 𝑈2 to be
convex and define
∫ 1
ℎ2 (𝑥, 𝑦, 𝑦 ′, 𝜃) = 𝜑 𝑦 (𝑥, 𝑡𝑦 + (1 − 𝑡) 𝑦 ′, 𝜃) d𝑡.
0

Let the open cone ℭ2′ ⊂⊂ ℭ2 ( 𝑗 = 1, 2) be such that −𝜂◦ ∈ ℭ2′ . Since ℎ2 (𝑥, 𝑦, 𝑦 ′, 𝜃) ≈
𝜕𝑦 𝜑 (𝑥, 𝑦, 𝜃) after contracting 𝑈1 about 𝑥 ◦ , 𝑈2 about 𝑦 ◦ and Γ♭ about 𝜃 ◦ , we obtain

∀ (𝑥, 𝑦, 𝜃) ∈ 𝑈1 × 𝑈2 × Γ♭ , ℎ2 (𝑥, 𝑦, 𝑦 ′, 𝜃) ∈ ℭ2′ .

We can select the real-valued, homogeneous of degree zero, function 𝜒2 ∈

C ∞ (R𝑛 \ {0}) to be such that 𝜒2 (𝜂) = 0 for all 𝜂 ∉ ℭ2 and 𝜒2 (𝜂) = 1 for all
𝜂 ∈ ℭ2′ . Again provided 𝑈1 × 𝑈2 is suitably contracted about (𝑥 ◦ , 𝑦 ◦ ) and Γ♭ about
𝜃 ◦ , we will have

∀ (𝑥, 𝑦, 𝜃) ∈ 𝑈1 × 𝑈2 × Γ♭ , 𝜒 (ℎ2 (𝑥, 𝑦, 𝑦 ′, 𝜃)) = 1. (18.3.10)

Then consider the adjoint of the operator 𝑨, which is to say, oscillatory integrals
∫ ∫
𝑨∗ 𝑣 (𝑦) = (2𝜋) −𝑁 e−𝑖 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (𝑥, 𝑦, 𝜃)𝑣 (𝑥) d𝑥d𝜃
Γ♭

where 𝑣 ∈ E ′ (Ω1 ) and the integral with respect to 𝑥 is a duality bracket. We mimic
the reasoning in the proof of Lemma 18.3.7 exchanging 𝑥 and 𝑦 as well as 𝜒1 and
1 − 𝜒2 : (18.3.10) implies that the formal amplitude

(1 − 𝜒2 (ℎ2 (𝑥, 𝑦, 𝑦 ′, 𝜃))) 𝑔2 (𝑦) 𝜕𝑦𝛼 𝑎 (𝑥, 𝑦, 𝜃)

∑︁ 1
( 𝛼+𝛽)

(ℎ2 (𝑥, 𝑦, 𝑦 ′, 𝜃)) ′ 𝑔2 (𝑦) 𝜕𝑦𝛼 𝑎 (𝑥, 𝑦, 𝜃)
𝛽
− D 𝑦′ 𝜒2
𝛼!𝛽! 𝑦 =𝑦
𝛼,𝛽 ∈Z𝑛+
𝛼+𝛽≠0

vanishes identically when (𝑥, 𝑦, 𝜃) ∈ 𝑈1 × 𝑈2 × Γ♭ . It follows that the FIO operator

E ′ (𝑈1 ) ∋ 𝑣 ↦→ 1 − 𝜒2 D 𝑦 (𝑔2 (𝑦) 𝑨∗ 𝑣 (𝑦)) ∈ D ′ (𝑈2 )

is smoothing, and therefore so is its adjoint:

18.3 Fourier Integral Operators. Basics 711

∀𝑢 ∈ E ′ (𝑈2 ) , 𝑨 𝑔2 (𝑦) 1 − 𝜒2 −D 𝑦 𝑢 ∈ C ∞ (𝑈1 ) .

If (𝑦 ◦ , 𝜂◦ ) ∉ 𝑊 𝐹 (𝑢) we can
choose∞U2 = 𝑈2 × (−ℭ2 ) so that ∞
𝑊 𝐹 (𝑢) ∩ U2 =
∅, implying 𝑔2 (𝑦) 𝜒2 −D 𝑦 𝑢 ∈ Cc (𝑈2 ) whatever 𝑔 2 ∈ C c (𝑈2 ), whence
𝑨 𝑔2 (𝑦) 𝜒2 −D 𝑦 𝑢 ∈ C ∞ (𝑈1 ) by Theorem 18.1.13.

□
The proof of Theorem 18.3.5 is complete.

18.3.3 Exponentiation of a skew-symmetric pseudodifferential operator

We are now going to generalize Equation (18.1.12). Let 𝑃 be a selfadjoint pseudod-

ifferential∫operator in an open∫ subset Ω of R𝑛 (Subsection 16.2.5); thus 𝑃 = 𝑃∗ ,
meaning 𝑣¯ (𝑥) 𝑃𝑢 (𝑥) d𝑥 = 𝑢 (𝑥) 𝑃𝑣 (𝑥)d𝑥 for all 𝑢, 𝑣 ∈ Cc∞ (Ω). We consider
the following initial value problem:

d𝑼
= 𝑖𝑃𝑼, 𝑼| 𝑡=0 = 𝑰, (18.3.11)
d𝑡
where 𝑰 is the identity operator on E ′ (Ω); note that

d𝑼 ∗
= −𝑖𝑼 ∗ 𝑃, 𝑼 ∗ | 𝑡=0 = 𝑰. (18.3.12)
d𝑡
If we presume that 𝑼 (𝑡) is an operator E ′ (Ω) −→ E ′ (Ω) depending smoothly on
𝑡 ∈ R we get d𝑡d (𝑼 ∗𝑼) = 0, implying 𝑼 ∗ (𝑡) 𝑼 (𝑡) = 𝑰 for all 𝑡: one may say that
𝑼 (𝑡) is a unitary operator.
We shall assume that 𝑃 = 𝑃 (𝑥, D) where 𝑃 (𝑥, 𝜉) = ∞
Í
𝑗=0 𝑝 𝑗 (𝑥, 𝜉) is a classical
symbol of order 1 in Ω: 𝑝 𝑗 (𝑥, 𝜆𝜉) = 𝜆1− 𝑗 𝑝 𝑗 (𝑥, 𝜉) for large |𝜉 | (Definition 16.3.2).
Actually, in the forthcoming argument we are going to deal with ∞
Í
𝑝
𝑗=0 𝑗 (𝑥, 𝜉) as a
formal classical symbol (Definition 16.3.4), meaning that each term is homogenous
in the whole R𝑛 \ {0}. We know (see the end of Subsection 16.2.5) that the princi-
pal symbol 𝑝 0 (𝑥, 𝜉) is a well-defined (in particular, independent of the choice of
coordinates) C ∞ function in the whole of 𝑇 ∗ Ω\0.
We assume from the start that the solution 𝑼 (𝑡) of (18.3.11) is an FIO; the
distribution kernel corresponding to 𝑼 (𝑡) is
∫
−𝑛
𝑈 (𝑡, 𝑥, 𝑦) = (2𝜋) e𝑖 𝜑 (𝑡 , 𝑥,𝑦, 𝜃) 𝑎 (𝑡, 𝑥, 𝑦, 𝜃) d𝜃, (18.3.13)
R𝑛

where 𝑎 (𝑡, 𝑥, 𝑦, 𝜃) is a smooth function of 𝑡 ∈ R valued in 𝑆 𝑚 (Ω × Ω, R𝑛 ). In

accordance with the initial condition in (18.3.11) we take

𝜑 (0, 𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃, 𝑎 (0, 𝑥, 𝑦, 𝜃) = 1. (18.3.14)

712 18 Fourier Integral Operators

On the one hand we have, by Theorem 18.1.17,

∫
𝑃 (𝑥, D 𝑥 ) 𝑈 (𝑡, 𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 𝜑 (𝑡 , 𝑥,𝑦, 𝜃) 𝑏 (𝑡, 𝑥, 𝑦, 𝜃) d𝜃, (18.3.15)
R𝑛

where 𝑏 (𝑡, 𝑥, 𝑦, 𝜃) is also a smooth function of 𝑡 ∈ R valued in 𝑆 𝑚 (Ω × Ω, R𝑛 ) that

can be built out of the formal amplitude
∑︁ 1
𝑏˜ (𝑡, 𝑥, 𝑦, 𝜃) = D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃))
𝛽 𝛼+𝛽
𝜕𝑥𝛼 𝑎 (𝑡, 𝑥, 𝑦, 𝜃)
𝛼!𝛽! 𝑥 ′ =𝑥
𝛼,𝛽 ∈Z+𝑛

[cf. (18.1.30)], where ℎ1 is defined by the equation

𝜑 (𝑡, 𝑥 ′, 𝑦, 𝜃) = 𝜑 (𝑡, 𝑥, 𝑦, 𝜃) − (𝑥 − 𝑥 ′) · ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃) (18.3.16)

[cf. (18.1.28), (18.1.31)]; in particular, ℎ1 (𝑡, 𝑥, 𝑥, 𝑦, 𝜃) = 𝜑 𝑥 (𝑡, 𝑥, 𝑦, 𝜃). The latter

implies
∑︁ 1
𝑏˜ (𝑡, 𝑥, 𝑦, 𝜃) = D 𝛼𝜉 𝑃 (𝑥, 𝜑 𝑥 (𝑡, 𝑥, 𝑦, 𝜃)) 𝜕𝑥𝛼 𝑎 (𝑡, 𝑥, 𝑦, 𝜃) (18.3.17)
𝛼∈Z+𝑛
𝛼!
∑︁ 1
D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃)) ′ 𝑎 (𝑡, 𝑥, 𝑦, 𝜃)
𝛽 𝛽
+
𝛽! 𝑥 =𝑥
|𝛽 | ≥1
∑︁ ∑︁ 1
D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃)) ′ 𝜕𝑥𝛼 𝑎 (𝑡, 𝑥, 𝑦, 𝜃) .
𝛽 𝛼+𝛽
+
𝛼!𝛽! 𝑥 =𝑥
| 𝛼 | ≥1 |𝛽 | ≥1

On the other hand,

∫
−𝑛
D𝑡 𝑈 (𝑡, 𝑥, 𝑦) = (2𝜋) e𝑖 𝜑 (𝑡 , 𝑥,𝑦, 𝜃) (𝜑𝑡 𝑎 + D𝑡 𝑎) (𝑡, 𝑥, 𝑦, 𝜃) d𝜃,
Γ

whence, by (18.3.11),
˜
𝜑𝑡 𝑎 + D𝑡 𝑎 = 𝑏. (18.3.18)
It makes sense to assume that 𝑎 is a classical amplitude of order zero (cf. Definition
16.3.4): this is consistent with the initial conditions (18.3.14) and with the fact that
the order of 𝑏˜ is 1, this being the case for 𝑃 and ℎ1 , the latter due to the homogeneity
of degree 1 of 𝜑.
We also want to use the hypothesis that 𝑃 (𝑥, D) is selfadjoint. By (16.2.14), this
means that
∞ ∞ ∑︁
∑︁ ∑︁ 1 𝛼 𝛼
𝑝 𝑗 (𝑥, 𝜉) = 𝜕 D 𝑝 𝑗 (𝑥, 𝜉) , (18.3.19)
𝑗=0 𝑗=0 𝛼∈Z𝑛
𝛼! 𝑥 𝜉
+

whose first consequence is that the principal symbol 𝑝 0 is real. Taking this into
account in (18.3.18) we get 𝜑𝑡 𝑎 0 = 𝑝 0 (𝑥, 𝜑 𝑥 ) 𝑎 0 , whence the eikonal equation

𝜑𝑡 (𝑡, 𝑥, 𝑦, 𝜃) = 𝑝 0 (𝑥, 𝜑 𝑥 (𝑡, 𝑥, 𝑦, 𝜃)) . (18.3.20)

18.3 Fourier Integral Operators. Basics 713

Given a domain Ω′ ⊂⊂ Ω and a cone Γ♭ in R𝑛 \ {0}, Γ♭ ∩ S𝑛−1 ⊂⊂ Γ, there is a

unique C ∞ function 𝜑 in (−𝑇, 𝑇) ×Ω′ ×Ω′ × Γ♭ (for some 𝑇 > 0) satisfying (18.3.20)
and the initial condition in (18.3.14), 𝜑 (0, 𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃; 𝜑 is real-valued. On
this subject we refer the reader to Section 13.4 (particularly Subsection 13.4.5).
Once the phase-function is determined we can proceed with the determination of
the amplitude. First note that (18.3.18)–(18.3.20) imply
˜
D𝑡 𝑎 + 𝑝 0 (𝑥, 𝜑 𝑥 (𝑡, 𝑥, 𝑦, 𝜃)) 𝑎 = 𝑏.

We return to (18.3.17) and use the following consequence of (18.3.16)

𝜕ℎ1,𝑘 1 𝜕2 𝜑
′ (𝑡, 𝑥, 𝑥, 𝑦, 𝜃) = (𝑡, 𝑥, 𝑦, 𝜃) .
𝜕𝑥ℓ 2 𝜕𝑥 𝑘 𝜕𝑥ℓ

We derive
𝑛 ∞
∑︁ 𝜕𝑃 𝜕𝑎 ∑︁
D𝑡 𝑎 + 𝑖 (𝑥, 𝜑 𝑥 ) = 𝑝 𝑗 (𝑥, 𝜑 𝑥 ) 𝑎
𝑗,𝑘=1
𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝑗=1
𝑛
1 ∑︁ 𝜕 2 𝑃 𝜕2 𝜑 ∑︁ 1
− (𝑥, 𝜑 𝑥 ) 𝑎+ D 𝛼𝜉 𝑃 (𝑥, 𝜑 𝑥 ) 𝜕𝑥𝛼 𝑎
2 𝑘,ℓ=1 𝜕𝜉 𝑘 𝜕𝜉ℓ 𝜕𝑥 𝑘 𝜕𝑥ℓ 𝛼!
| 𝛼 | ≥2
∑︁ 1
D ′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃)) ′ 𝑎
𝛽 𝛽
+
𝛽! 𝑥 𝑥 =𝑥
|𝛽 | ≥2
∑︁ ∑︁ 1
D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃)) ′ 𝜕𝑥𝛼 𝑎.
𝛽 𝛼+𝛽
+
𝛼!𝛽! 𝑥 =𝑥
| 𝛼 | ≥1 |𝛽 | ≥1

This can be rewritten as follows

𝔏𝑎 + 𝑖𝑞𝑎 = 𝑖𝐹 [𝑎] , (18.3.21)

where
𝑛
𝜕 ∑︁ 𝜕 𝑝 0 𝜕
𝔏= − (𝑥, 𝜑 𝑥 (𝑡, 𝑥, 𝑦, 𝜃)) , (18.3.22)
𝜕𝑡 𝑘=1 𝜕𝜉 𝑘 𝜕𝑥 𝑘
𝑛
1 ∑︁ 𝜕 2 𝑝 0 𝜕2 𝜑
𝑞 (𝑡, 𝑥, 𝑦, 𝜃) = (𝑥, 𝜑 𝑥 (𝑡, 𝑥, 𝑦, 𝜃)) (𝑡, 𝑥, 𝑦, 𝜃) , (18.3.23)
2 𝑘,ℓ=1 𝜕𝜉 𝑘 𝜕𝜉ℓ 𝜕𝑥 𝑘 𝜕𝑥ℓ
714 18 Fourier Integral Operators
∞
∑︁
𝐹 [𝑎] (𝑡, 𝑥, 𝑦, 𝜃) = 𝑝 𝑗 (𝑥, 𝜑 𝑥 ) 𝑎 (18.3.24)
𝑗=1
∞
1 ∑︁ ∑︁ 𝜕 2 𝑝 𝑗
𝑛
𝜕2 𝜑 ∑︁ 1
− (𝑥, 𝜑 𝑥 ) 𝑎+ D 𝛼 𝑃 (𝑥, 𝜑 𝑥 ) 𝜕𝑥𝛼 𝑎
2 𝑗=1 𝑘,ℓ=1 𝜕𝜉 𝑘 𝜕𝜉ℓ 𝜕𝑥 𝑘 𝜕𝑥ℓ 𝛼! 𝜉
| 𝛼 | ≥2
∑︁ 1
D ′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃)) ′ 𝑎
𝛽 𝛽
+
𝛽! 𝑥 𝑥 =𝑥
|𝛽 | ≥2
∑︁ ∑︁ 1
D 𝑥′ D 𝜉 𝑃 (𝑥, ℎ1 (𝑡, 𝑥, 𝑥 ′, 𝑦, 𝜃)) ′ 𝜕𝑥𝛼 𝑎.
𝛽 𝛼+𝛽
+
𝛼!𝛽! 𝑥 =𝑥
| 𝛼 | ≥1 |𝛽 | ≥1

The coefficients of 𝔏 as well as 𝑞 (𝑡, 𝑥, 𝑦, 𝜃) are homogeneous of degree zero and

real-valued. Since 𝔏 is a nowhere vanishing C ∞ real vector field in (−𝑇, 𝑇) × Ω′
(depending smoothly on the parameters 𝑦 ∈ Ω′, 𝜃 ∈ Γ♭ ) there is a unique C ∞ solution
𝜓 of the initial value problem 𝔏𝜓 = 𝑞, 𝜓| 𝑡=0 = 0, in an open set (−𝑇 ′, 𝑇 ′) × Ω′′,
Ω′′ ⊂⊂ Ω′, 0 < 𝑇 ′ < 𝑇; 𝜓 is real-valued and homogeneous of degree zero (with
respect to 𝜃); the latter is true of e𝑖 𝜓 . Equation (18.3.21) is equivalent to

𝔏 e𝑖 𝜓 𝑎 = 𝑖e𝑖 𝜓 𝐹 [𝑎] . (18.3.25)

It is natural to take 𝑎 = ∞
Í
𝜈=0 𝑎 𝜈 with 𝑎 𝜈 homogeneous of degree −𝜈. To solve
(18.3.25) we equate the terms in both sides that have the same homogeneity
degree.
Since the order of 𝐹 [𝑎] is ≤ −1 the first equation is 𝔏 e𝑖 𝜓 𝑎 0 = 0. The initial
condition in (18.3.14) being 𝑎 0 | 𝑡=0 = 1 we conclude that 𝑎 0 ≡ e−𝑖 𝜓 . Thus 𝑎 =
e−𝑖 𝜓 + ∞
Í
𝜈=1 𝑎 𝜈 is elliptic of degree zero (Definition 16.2.14).
For 𝜈 ≥ 1 we get

𝔏 e𝑖 𝜓 𝑎 𝜈 = 𝑖e𝑖 𝜓 𝐹𝜈 [𝑎 0 , ..., 𝑎 𝜈−1 ] (18.3.26)
𝜈−1
∑︁ ∑︁
= 𝑓 𝜈,𝑘, 𝛼 𝜕𝑥𝛼 𝑎 𝑘 ,
𝑘=0 | 𝛼 | ≤𝑚𝜈,𝑘

where the coefficients 𝑓 𝜈,𝑘, 𝛼 (𝑡, 𝑥, 𝑦, 𝜃) ∈ C ∞ (−𝑇 ′′, 𝑇 ′′) × Ω′′ × Ω′′ × Γ♭ can be
calculated, if need be, from the right-hand side of (18.3.24). These equations are
commonly referred to as the transport equations. They can be solved recursively.
The initial condition in (18.3.14) demands 𝑎 𝜈 (0, 𝑥, 𝑦, 𝜃) = 0 for all 𝜈 = 1, 2, ....
The operator defined by the distribution kernel 𝑈 (𝑡, 𝑥, 𝑦) will be denoted by
exp (𝑖𝑡𝑃 (𝑥, D 𝑥 )).

Remark 18.3.9 The construction in this subsubsection can be exactly duplicated

if we
∫allow 𝑃 to depend on 𝑡. In this case we would end up with the FIO
𝑡
exp 𝑖 0 𝑃 (𝑠, 𝑥, D 𝑥 ) d𝑠 .
18.4 Reduction of the Fiber Variables 715

18.4 Reduction of the Fiber Variables

18.4.1 Reducing 𝑵 . General phase-functions

A natural question is to what extent the number 𝑁 of the “integration variables” 𝜃 𝑘

in a representation of the type (18.1.13) can be reduced. Generally speaking, this
can be done if we are willing to contract the cone Γ about a “central” ray (𝜆𝜃 ◦ ) 𝜆>0
as well as the neighborhood of the point (𝑥 ◦ , 𝑦 ◦ ) ∈ Ω1 × Ω2 where one wishes to
study (18.1.13). As we have seen in Subsection 18.2.2 such contractions are also
helpful in the microlocal analysis of Fourier distribution kernels, focusing on a conic
neighborhood of a point (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ Λ 𝜑 with (𝜉 ◦ , 𝜂◦ ) given by (18.3.9).
In the reduction of 𝑁 the rank of the Hessian matrix of the phase-function 𝜑
(Definition 18.1.4) with respect to the 𝜃 𝑘 ’s, rank 𝜑 𝜃 , 𝜃 (𝑥, 𝑦, 𝜃), is important, and so
is the rank at the point (𝑥, 𝜉, 𝑦, 𝜂) = 𝑝 𝜑 , 𝑞 𝜑 (𝑥, 𝑦, 𝜃) ∈ Λ 𝜑 of the restriction to Λ 𝜑
of the base projection 𝜋 ∗ : (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0) −→ Ω1 × Ω2 , 𝜋 ∗ | Λ 𝜑 . [The map

𝑝 𝜑 , 𝑞 𝜑 : Σ 𝜑 −→ Λ 𝜑 is defined in (18.1.18).] Actually, these numbers are related,
as we now show. Note that rank 𝜑 𝜃 , 𝜃 < 𝑁, otherwise the equations 𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0
would admit a smooth solution 𝜃 = 𝜃 (𝑥, 𝑦) in some open subset of Σ 𝜑 and Σ 𝜑 could
not be conic. Likewise, rank 𝜋 ∗ | Λ 𝜑 < 2𝑛 since dim Λ 𝜑 = 2𝑛 and Λ 𝜑 is conic.

Proposition 18.4.1 If 𝜑 is a phase-function in Ω1 × Ω2 × Γ then, for every (𝑥, 𝑦, 𝜃) ∈

Σ𝜑 ,

𝑁 − rank 𝜑 𝜃 , 𝜃 (𝑥, 𝑦, 𝜃) = dim Σ 𝜑 − rank 𝜋 ∗ | Λ 𝜑 (𝑥, 𝜉, 𝑦, 𝜂) , (18.4.1)

where (𝑥, 𝜉, 𝑦, 𝜂) = 𝑝 𝜑 , 𝑞 𝜑 (𝑥, 𝑦, 𝜃).

Proof Let 𝜋 𝜑 denote the coordinate projection Σ 𝜑 ∋ (𝑥, 𝑦, 𝜃) → (𝑥, 𝑦) and d𝜋 𝜑 :

𝑇 Σ 𝜑 −→ 𝑇 (Ω1 × Ω2 ) its differential; we have

dim ker d𝜋 𝜑 (𝑥, 𝑦, 𝜃) = dim Σ 𝜑 − rank 𝜋 𝜑 Σ 𝜑 . (18.4.2)

The null space ker d𝜋 𝜑 (𝑥, 𝑦, 𝜃) consists of the vectors tangent to Σ 𝜑 at (𝑥, 𝑦, 𝜃) of
Í𝑁
the form 𝑘=1 𝑏 𝑘 𝜕𝜃𝜕 𝑘 with b = (𝑏 1 .., 𝑏 𝑁 ) ∈ ker 𝜑 𝜃 , 𝜃 (𝑥, 𝑦, 𝜃). We conclude that

dim ker d𝜋 𝜑 (𝑥, 𝑦, 𝜃) = dim ker 𝜑 𝜃 , 𝜃 (𝑥, 𝑦, 𝜃) = 𝑁 − rank 𝜑 𝜃 , 𝜃 (𝑥, 𝑦, 𝜃) . (18.4.3)

We have 𝜋 𝜑 = 𝜋 ∗ | Λ 𝜑 ◦ 𝑝 𝜑 , 𝑞 𝜑 . Since rank 𝑝 𝜑 , 𝑞 𝜑 = 2𝑛 we see that rank 𝜋 𝜑 =

rank 𝜋 ∗ | Λ 𝜑 . Combining (18.4.2) and (18.4.3) yields (18.4.1). □

Example 18.4.2 Suppose 𝑁 = 𝑛, Ω1 = Ω2 = Ω and let 𝜑 be the phase of the identity

(Example 18.2.14). Both sides in (18.4.1) are equal to 𝑛 whatever (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈
diag Ω × (R𝑛 \ {0}).
716 18 Fourier Integral Operators

Example 18.4.3 Suppose 𝑁 = 𝑛, Ω1 = Ω2 = R2𝑛 and 𝜑 (𝑥, 𝑦, 𝜃) = (𝑥 − 𝑦) · 𝜃 − 21 |𝜃|;

then
𝜃
Σ 𝜑 = (𝑥, 𝑦, 𝜃) ∈ R2𝑛 × (R𝑛 \ {0}) ; 𝑥 − 𝑦 =
|𝜃|
and
𝜉
Λ 𝜑 = (𝑥, 𝜉, 𝑦, 𝜂) ∈ (𝑇 ∗ Ω\0) 2 ; 𝑥 − 𝑦 = ,𝜉 =𝜂 .
|𝜉 |
∗

2see that 𝜋 Λ 𝜑 is the hypersurface |𝑥 − 𝑦| = 1. Since the rank of the matrix
We
𝜕 𝜑
𝜕𝜃 𝑗 𝜕𝜃𝑘 is equal to 𝑛 − 1 both sides in (18.4.1) are equal to 1 whatever
1≤ 𝑗,𝑘 ≤𝑛
(𝑥, 𝑦, 𝜃) ∈ Σ 𝜑 .

We propose to modify the phase-function 𝜑 (with a corresponding modification

of the amplitude 𝑎) in the representation (18.1.13) with the aim of decreasing the
number 𝑟 = rank 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ), down to and including 𝑟 = 0. Actually, when
𝑟 = 0 we leave (18.1.13) unchanged. We shall now deal with the case 𝑟 ≥ 1 (as
pointed out above, we always have 𝑟 < 𝑁).
An orthogonal transformation in 𝜃-space allows us to assume that

𝜕2 𝜑
(𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 𝜒𝑘 ≠ 0 if 𝑘 = ℓ = 𝑁 − 𝑟 + 1, ..., 𝑁, (18.4.4)
𝜕𝜃 𝑘 𝜕𝜃 ℓ
𝜕2 𝜑
(𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 for all other pairs (𝑘, ℓ) . (18.4.5)
𝜕𝜃 𝑘 𝜕𝜃 ℓ

By (18.4.1) we have 𝑁 − 𝑟 = dim Σ 𝜑 − rank 𝜋 ∗ | Λ 𝜑 (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ≤ dim Σ 𝜑 .

Lemma 18.4.4 If (18.4.4)–(18.4.5) hold with 1 ≤ 𝑟 < 𝑁 then 𝜃 ◦𝑁 −𝑟+1 +· · ·+ 𝜃 ◦𝑁 = 0

and therefore 𝜃 1◦ + · · · + 𝜃 ◦𝑁 −𝑟 ≠ 0.

Proof By the Euler homogeneity formula we have

𝑁
∑︁ 𝜕𝜑 𝜕𝜑 ◦ ◦ ◦
𝜃 ◦𝑘 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = (𝑥 , 𝑦 , 𝜃 ) = 0,
𝑘=𝑁 −𝑟+1
𝜕𝜃 𝑗 𝜕𝜃 𝑘 𝜕𝜃 𝑗

for every 𝑗 = 𝑁 − 𝑟 + 1, ..., 𝑁; by (18.4.4) this implies 𝜃 ◦𝑁 −𝑟+1 + · · · + 𝜃 ◦𝑁 = 0. □

2 2
In the remainder of this subsection we assume that |𝜃 ◦ | 2 = 𝜃 1◦ +· · ·+ 𝜃 ◦𝑁 −𝑟 = 1.
We select arbitrarily an integer 𝜈, 𝑁 − 𝑟 ≤ 𝜈 ≤ 𝑁 − 1, and use the notation

𝜃 ′ = (𝜃 1 , ..., 𝜃 𝜈 ) , 𝜃 ′′ = (𝜃 𝜈+1 , ..., 𝜃 𝑁 ) . (18.4.6)

Note that 𝜃 ◦′ = 𝜃 1◦ , ..., 𝜃 ◦𝑁 −𝑟 , 0, ..., 0 , |𝜃 ◦′ | = 1, 𝜃 ◦′′ = 0. We contract Γ about 𝜃 ◦

and take Γ = Γ ′ × Γ ′′ with
18.4 Reduction of the Fiber Variables 717

𝜃′

Γ ′ = 𝜃 ′ ∈ R𝜈 \ {0} ; − 𝜃 ◦′ < 𝜀 ′ , (18.4.7)
|𝜃 ′ |
Γ ′′ = 𝜃 ′′ ∈ R 𝑁 −𝜈 \ {0} ; |𝜃 ′′ | < 𝜀 ′′ |𝜃 ′ | .

(18.4.8)

Lemma 18.4.5 Suppose that (18.4.4)–(18.4.5) hold with 1 ≤ 𝑟 < 𝑁. If the neigh-
borhoods 𝑈1 ⊂ Ω1 of 𝑥 ◦ , 𝑈2 ⊂ Ω2 of 𝑦 ◦ and the positive numbers 𝜀 ′′, 𝜀 ′/𝜀 ′′ are
suitably small there is a unique C ∞ map, homogeneous of degree 1,

𝑈1 × 𝑈2 × Γ ′ ∋ (𝑥, 𝑦, 𝜃 ′) → 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) = 𝜃˜𝜈+1 (𝑥, 𝑦, 𝜃 ′) , ..., 𝜃˜ 𝑁 (𝑥, 𝑦, 𝜃 ′) ∈ Γ ′′,

satisfying the equations 𝜑 𝜃𝑘 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) = 0, 𝑘 = 𝜈 + 1, ..., 𝑁, and such

that 𝜃˜′′ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′) = 0.

Proof Apply the Implicit Function Theorem taking (18.4.4) into account. □

Proposition 18.4.6 Suppose that (18.4.4)–(18.4.5) hold with 1 ≤ 𝑟 < 𝑁 and let
the integer 𝜈 be such that 𝑁 − 𝑟 ≤ 𝜈 ≤ 𝑁 − 1. Let 𝑈1 , 𝑈2 , 𝜀 ′ and the function
𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) be as in Lemma 18.4.5. Then 𝜑♭ (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) is

a phase-function in 𝑈1 × 𝑈2 × Γ ′ (Definition 18.1.4) whose associated Lagrangian
submanifold (Definition 18.2.3) Λ 𝜑♭ coincides with Λ 𝜑 in a conic neighborhood of
(𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) in (𝑇 ∗𝑈1 \0) × (𝑇 ∗𝑈2 \0). We have

rank 𝜑♭𝜃 ′ , 𝜃 ′ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′) = rank 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) − 𝑁 + 𝜈. (18.4.9)

Proof We apply Lemma 18.4.5: we have, for 𝑗 = 1, ..., 𝜈,

𝜑♭𝜃 𝑗 (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝜃 𝑗 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′)

(18.4.10)
𝑁
∑︁ 𝜕 𝜃˜ℓ
𝜑 𝜃ℓ 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝜃 𝑗 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) .

+
ℓ=𝜈+1
𝜕𝜃 𝑗

Since 𝜑 𝜃 𝑗 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) vanishes identically in 𝑈1 × 𝑈2 × Γ ′ for all 𝑗 > 𝜈

we conclude that
𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0 ⇐⇒ 𝜑♭𝜃 ′ (𝑥, 𝑦, 𝜃 ′) = 0,
i.e.,
Σ 𝜑♭ ∋ (𝑥, 𝑦, 𝜃 ′) ↦→ 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) ∈ Σ 𝜑

(18.4.11)
is a diffeomorphism. We have

𝜑♭𝑥,𝑦 (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝑥,𝑦 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′)

𝑁
∑︁
𝜑 𝜃ℓ 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) 𝜕𝑥,𝑦 𝜃˜ℓ (𝑥, 𝑦, 𝜃 ′) ,

+
ℓ=𝜈+1

implying

∀ (𝑥, 𝑦, 𝜃 ′) ∈ Σ 𝜑♭ , 𝜑♭𝑥,𝑦 (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝑥,𝑦 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) ,

(18.4.12)
718 18 Fourier Integral Operators

which proves that there is an open cone ℭ ⊂ (R𝑛 \ {0}) 2 containing (𝜉 ◦ , 𝜂◦ ), such
that
Λ 𝜑♭ ∩ (𝑈1 × 𝑈2 × ℭ) = Λ 𝜑 ∩ (𝑈1 × 𝑈2 × ℭ) .
Formula (18.4.9) is proved by comparing (18.4.1) to the same formula with 𝜑♭
substituted for 𝜑 and therefore 𝑁 replaced by 𝜈. □

Corollary 18.4.7 Suppose that (18.4.4)–(18.4.5) hold with 1 ≤ 𝑟 < 𝑁. The phase-
function 𝜑♭ in Proposition 18.4.6 can be selected so that 𝜑♭𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′) = 0.

Proof It suffices to select 𝜈 = 𝑛 − 𝑟 in (18.4.9). □

We are going to apply the Stationary Phase Formula (18.A.6) (Appendix to the
present chapter) to the integral (18.1.13) with 𝑞 = 𝑁 − 𝜈. To do this we take 𝜆 = |𝜃 ′ |
and, as integration variables, 𝑠 𝑗 = 𝜆−1 𝜃 𝜈+ 𝑗 − 𝜃˜𝜈+ 𝑗 (𝑥, 𝑦, 𝜃 ′) ( 𝑗 = 1, ..., 𝑁 − 𝜈).
The parameters will be 𝑥 𝑗 , 𝑦 𝑘 , 𝑡ℓ = 𝜆−1 𝜃 ℓ ∈ Γ ′ ∩ S𝑛−1 ( 𝑗, 𝑘, ℓ = 1, ..., 𝑛). Note that
𝜃 ◦′ = 𝜃 1◦ , ..., 𝜃 ◦𝑁 −𝑟 , 0, ..., 0 , |𝜃 ◦′ | = 1, 𝜃 ◦′′ = 0.
We define 𝜙 (𝑥, 𝑦, 𝑠, 𝑡) = 𝜑 𝑥, 𝑦, 𝑡, 𝑠 + 𝜆−1 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) ; thus

𝜙 (𝑥, 𝑦, 0, 𝑡) = 𝜆−1 𝜑 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) = 𝜆−1 𝜑♭ (𝑥, 𝑦, 𝜃 ′) .

(18.4.13)

We have, by (18.4.4),
det 𝜙 𝑠,𝑠 (𝑥 ◦ , 𝑦 ◦ , 0, 0) ≠ 0. (18.4.14)
We select the neighborhood 𝑈1 × 𝑈2 of (𝑥 ◦ , 𝑦 ◦ ) and the number 𝜀 ′ > 0 sufficiently
small that det 𝜕𝑠𝜕𝜙
𝑗 𝜕𝑠𝑘
(𝑥, 𝑦, 0, 𝑡) ≠ 0 for all (𝑥, 𝑦) ∈ 𝑈1 × 𝑈2 and 𝑡 ∈ Γ ′ ∩ S𝜈−1 . The
variation of 𝑠 is limited to the ball 𝔅 ′′ ′′
𝜀′′ of center 0 and radius 𝜀 in R
𝑁 −𝜈 ; keep in

mind that 𝑠 and 𝑡 are homogeneous of degree zero with respect to 𝜃. Taking (18.4.9)
into account and keeping (𝑥, 𝑦) ∈ 𝑈1 × 𝑈2 we rewrite (18.1.13) as follows
∫ ∫
−𝑁 𝑖 𝜑 ( 𝑥,𝑦, 𝜃)
𝐴 (𝑥, 𝑦) = (2𝜋) e 𝑎 (𝑥, 𝑦, 𝜃) d𝜃 d𝜃 ′
′′
(18.4.15)
Γ′ Γ′′
!
∫ ∞ ∫ ∫
= (2𝜋) −𝑁 𝜆 𝑁 −1 e𝑖𝜆𝜙 ( 𝑥,𝑦,𝑠,𝑡) 𝑎˜ (𝑥, 𝑦, 𝜆, 𝑠, 𝑡) d𝑠 d𝜇 (𝑡) d𝜆
0 Γ′ ∩S𝜈−1 𝔅′′𝜀′′

where 𝜙 is defined in (18.4.13),

𝑎˜ (𝑥, 𝑦, 𝜆, 𝑠, 𝑡) = 𝑎 𝑥, 𝑦, 𝜆𝑡, 𝜆𝑠 + 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′)

and d𝜇 (𝑡) is the volume element on S𝜈−1 such that d𝜃 ′ = 𝜆 𝜈−1 d𝜆d𝜇 (𝑡). We apply
(18.A.6) with 𝑞 = 𝑁 − 𝜈 to deduce the asymptotic expansion of
∫
𝐼𝜆 (𝑥, 𝑦, 𝑡) = e𝑖𝜆𝜙 ( 𝑥,𝑦,𝑠,𝑡) 𝑎˜ (𝑥, 𝑦, 𝜆, 𝑠, 𝑡) d𝑠.
𝔅′′𝜀′′

By (18.4.13) 𝜆𝜙 (𝑥, 𝑦, 0, 𝑡) = 𝜑♭ (𝑥, 𝑦, 𝜃 ′); with the notation

18.4 Reduction of the Fiber Variables 719

𝐻 𝜑 = det 𝜑 𝜃 ′′ , 𝜃 ′′ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ )

and taking (18.4.14) into account we get

12 ( 𝑁 −𝜈) 𝜋
sign 𝐻 𝜑 +𝑖 𝜑 ♭ ( 𝑥,𝑦, 𝜃 ′ ) ∞
2𝜋 e𝑖 4 ∑︁ (2𝑖) ℓ
𝐼𝜆 (𝑥, 𝑦, 𝑡) ≈ √︃ 𝑎 ℓ (𝑥, 𝑦, 𝜃 ′) (18.4.16)
𝜆 ℓ!
det 𝐻 𝜑 ℓ=0

where
ℓ 𝜃′

𝑎 ℓ (𝑥, 𝑦, 𝜃 ′) = D𝑠 · 𝐻 −1𝜑 D 𝑠 ˜
𝑎 𝑥, 𝑦, 𝜆, 𝑠, − 𝜃 ◦′
𝜆 𝑠=0
ℓ
D 𝜃 ′′ · 𝐻 𝜑 D 𝜃 ′′ 𝑎 𝑥, 𝑦, 𝜃 , 𝜃˜ (𝑥, 𝑦, 𝜃 ′)
−1 ′ ′′
ℓ
=𝜆

with 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) the function in Lemma 18.4.5. Since 𝑎 ∈ 𝑆c𝑚 (𝑈1 × 𝑈2 , Γ) we have
ℓ
−1
𝑎 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) ∈ 𝑆c𝑚−2ℓ (𝑈1 × 𝑈2 ; Γ ′) ,

D𝜃 · 𝐻𝜑 D𝜃
′′ ′′

whence 𝑎 ℓ (𝑥, 𝑦, 𝜃 ′) ∈ 𝑆c𝑚−ℓ (𝑈1 × 𝑈2 ; Γ ′). This allows us to build a true amplitude
Í∞ 1
𝑎♭ (𝑥, 𝑦, 𝜃 ′) ∈ 𝑆c𝑚 (𝑈1 × 𝑈2 ; Γ ′) out the formal amplitude ℓ=0 ℓ ′
ℓ! (2𝑖) 𝑎 ℓ (𝑥, 𝑦, 𝜃 )
′
by inserting suitably chosen cut-off functions in 𝜃 -space as described in (16.2.11)(see
also Proposition 16.2.9). We can summarize the results of this subsection:
Theorem 18.4.8 Let 𝐴 (𝑥, 𝑦) be the distribution kernel (18.1.13). After possibly
decreasing 𝜀 ′ and 𝜀 ′′ in (18.4.7)–(18.4.8) we have, in 𝑈1 ×𝑈2 sufficiently contracted
about (𝑥 ◦ , 𝑦 ◦ ),
𝜋
e𝑖 4 sign 𝐻𝜑
∫
1 ♭ ( 𝑥,𝑦, 𝜃 ′ ) 1
𝐴 (𝑥, 𝑦) (2𝜋) − 2 ( 𝑁 +𝜈) √︃ e𝑖 𝜑 𝑎♭ (𝑥, 𝑦, 𝜃 ′) |𝜃 ′ | 2 ( 𝑁 −𝜈) d𝜃 ′
det 𝐻 𝜑 Γ′

(18.4.17)
mod C ∞ (𝑈1 × 𝑈2 ).

18.4.2 Reducing 𝑵 . Nondegenerate phase-functions

When the phase-function 𝜑 in Ω1 × Ω2 × Γ is nondegenerate we have dim Σ 𝜑 = 2𝑛.

By putting this into (18.4.1) we obtain
Proposition 18.4.9 If 𝑁 = 𝑛 and 𝜑 is nondegenerate at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) then

rank 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = rank 𝜋 ∗ | Λ 𝜑 (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) − 𝑛. (18.4.18)

The point (𝜉 ◦ , 𝜂◦ ) is defined in (18.3.9); 𝜋 ∗ is the base projection (𝑇 ∗ Ω1 \0) ×

(𝑇 ∗ Ω
2 \0) −→ Ω1 × Ω2 .
720 18 Fourier Integral Operators

Proposition 18.4.10 If 𝜑 is strongly nondegenerate at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) (Definition 18.2.10)

and 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 then necessarily 𝑁 ≤ 𝑛 and the differentials d 𝑥 𝜑 𝜃1 ,...,d 𝑥 𝜑 𝜃 𝑁
are linearly independent at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ); the same is true of d 𝑦 𝜑 𝜃1 , ..., d 𝑦 𝜑 𝜃 𝑁 .

Proof Since Λ 𝜑 is a local symplectic graph it can be defined locally by equations
𝑦 = 𝑦 (𝑥, 𝜉), 𝜂 = 𝜂 (𝑥, 𝜉) and therefore 𝜋 ∗ Λ 𝜑 contains the range of the composite
map
Λ 𝜑 ∋ (𝑥, 𝜉, 𝑦, 𝜂) ↦→ (𝑥, 𝑦 (𝑥, 𝜉)) ∈ Ω1 × Ω2
whose rank is obviously ≥ 𝑛. From (18.4.1) it follows that

𝑁 = 2𝑛 − rank 𝜋 ∗ | Λ 𝜑 (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ≤ 𝑛.

Since 𝜑 is strongly nondegenerate we must have 𝐷𝐷( 𝜑( 𝑦, 𝑥 , 𝜑𝜃 ) ◦ ◦ ◦

𝜃) (𝑥 , 𝑦 , 𝜃 ) ≠ 0 (Definition
18.2.10); since 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 this is only possible if

𝜕𝜑 𝜃ℓ 𝜕𝜑 𝜃ℓ
rank = rank =𝑁
𝜕𝑥 𝑗 1≤ 𝑗 ≤𝑛,1≤ℓ ≤ 𝑁 𝜕𝑦 𝑗 1≤ 𝑗 ≤𝑛,1≤ℓ ≤ 𝑁
at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). □
The case 𝑁 < 𝑛 is possible, as shown in the following
3
3 2
Consider the case 𝑛 = 3, 𝑁 = 2, and the phase-function in R ×
Example 18.4.11
R × R \ {0} ,

𝜑 (𝑥, 𝑦, 𝜃) = 𝜃 1 (𝑥 1 − 𝑦 1 ) + 𝜃 2 (𝑥2 − 𝑦 2 ) + 𝑥3 𝑦 3 |𝜃| .

The set Σ 𝜑 is defined by the equations

𝜃𝑗
𝑦 𝑗 = 𝑥 𝑗 + 𝑥3 𝑦 3 , 𝑗 = 1, 2;
|𝜃|

Σ 𝜑 is a 6-dimensional smooth submanifold of R3 ×R3 × R2 \ {0} . If 𝑥 ◦ = 𝑦 ◦ = 0 ∈ R3

and 𝜃 ◦ = (1, 0) then
𝜑 𝜑
det 𝑥,𝑦 𝑥, 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 1,
𝜑 𝑦, 𝜃 𝜑 𝜃 , 𝜃
implying that 𝜑 is strongly nondegenerate. The Lagrangian submanifold Λ 𝜑 is the
set of points (𝑥, 𝜉, 𝑦, 𝜂) such that 𝜂12 + 𝜂22 ≠ 0 and

𝜉3 𝑥 3
𝑦𝑗 = 𝑥𝑗 + 𝜉𝑗 , 𝜂 𝑗 = 𝜉 𝑗 , 𝑗 = 1, 2,
𝜉12+ 𝜉22
√︃
𝜉3
𝑦 3 = √︃ , 𝜂3 = −𝑥 3 𝜉12 + 𝜉22 .
𝜉12 + 𝜉22
18.4 Reduction of the Fiber Variables 721

Proposition 18.4.12 Suppose that (18.4.4)–(18.4.5) hold with 1 ≤ 𝑟 < 𝑁 and let the
integer 𝜈 be such that 𝑁 − 𝑟 ≤ 𝜈 ≤ 𝑁 − 1. If the phase-function 𝜑 is nondegenerate
at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) then the phase-function 𝜑♭ (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) (see

Lemma 18.4.5) is nondegenerate at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′).

Proof Indeed, (18.4.4)–(18.4.5) imply d 𝜃 𝜑 𝜃1 ∧ · · · ∧ d 𝜃 𝜑 𝜃 𝑁 −𝑟 = 0 and d 𝜃 𝜑 𝜃 𝑁 −𝑟+1 ∧

· · ·∧d 𝜃 𝜑 𝜃 𝑁 ≠ 0 at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). The nondegeneracy of
𝜑 requires d𝜑 𝜃1 ∧· · ·∧d𝜑 𝜃 𝑁 −𝑟 ≠
0 and therefore also d 𝑥 + d 𝑦 𝜑 𝜃1 ∧ · · · ∧ d 𝑥 + d 𝑦 𝜑 𝜃 𝑁 −𝑟 ≠ 0 at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ), which
settles the case 𝜈 = 𝑁 − 𝑟 by (18.4.12 ). When 𝜈 > 𝑁 − 𝑟 we avail ourselves of
(18.4.10). We get, for 𝑗 = 𝑁 − 𝑟 + 1, ..., 𝜈, 𝑘 = 1, ..., 𝜈,

𝜕 ♭
𝜑 𝜃 𝑗 (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝜃 𝑗 , 𝜃𝑘 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′)

𝜕𝜃 𝑘
∑︁𝑁
𝜕 𝜃˜ℓ
+ 𝜑 𝜃 𝑗 , 𝜃ℓ 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) (𝑥, 𝑦, 𝜃 ′) .
ℓ=𝜈+1
𝜕𝜃 𝑘

By (18.4.4)–(18.4.5) we know that 𝜑 𝜃 𝑗 , 𝜃𝑘 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 𝜆 𝑘 𝛿 𝑗,𝑘 if 𝑁 − 𝑟 < 𝑘 ≤ 𝜈 and

𝜑 𝜃 𝑗 , 𝜃ℓ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 if ℓ > 𝜈, whence d 𝜃 ′ 𝜑♭𝜃 𝑗 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′) = d 𝜃 𝜑 𝜃 𝑗 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) and
d 𝜃 ′ 𝜑♭𝜃 𝑁 −𝑟+1 ∧ · · · ∧ d 𝜃 ′ 𝜑♭𝜃𝜈 ≠ 0 at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′). □

Remark 18.4.13 When 𝜑 is nondegenerate we have dim Σ 𝜑♭ = dim Σ 𝜑 = 2𝑛 and we

can regard Σ 𝜑 as the graph over Σ 𝜑♭ of the map (18.4.11).

Corollary 18.4.14 Suppose that (18.4.4)–(18.4.5) hold with 1 ≤ 𝑟 < 𝑁 and let the
integer 𝜈 be such that 𝑁 −𝑟 ≤ 𝜈 ≤ 𝑁 −1. If the phase-function 𝜑 is strongly nondegen-
erate at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) then the phase-function 𝜑♭ (𝑥, 𝑦, 𝜃 ′) is strongly nondegenerate
at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′).

Proof Combine Propositions 18.2.13, 18.4.6, 18.4.9. □

Proposition 18.4.15 Let the phase-function 𝜑 be nondegenerate at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). Pos-

sibly after decreasing 𝜀 ′ and 𝜀 ′′ in (18.4.7)–(18.4.8) and contracting 𝑈1 × 𝑈2 about
(𝑥 ◦ , 𝑦 ◦ ) the expression (18.4.17) is valid in 𝑈1 × 𝑈2 with 𝜑♭𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 and

𝜈 = 2𝑛 − rank 𝜋 ∗ | Λ 𝜑 (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) . (18.4.19)

Proof If we apply (18.4.1) with 𝜑♭ in the place of 𝜑 (and take Remark 18.4.13 into
account) we see that

𝜈 − rank 𝜑♭𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦′) = 2𝑛 − rank 𝜋 ∗ | Λ 𝜑 (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) .

If 𝜑♭𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0 this reduces to (18.4.19). □

Remark 18.4.16 Formula (18.4.19) is equivalent to 𝜈 = 𝑁 − rank 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ),

by (18.4.9).
722 18 Fourier Integral Operators

18.4.3 Equivalent phase-functions

We discuss now possible changes of the phase-function (Definition 18.1.4) in the

integral (18.1.13) through changes, depending on (𝑥, 𝑦) ∈ Ω1 ×Ω2 , of the integration
variables 𝜃 𝑘 (but without changing their number 𝑁; the amplitudes, however, will
have to be modified). We recall that a subset V of R2𝑛 × R 𝑁 \ {0} is said to be
conic if (𝑥, 𝑦, 𝜃) ∈ V =⇒ (𝑥, 𝑦, 𝜆𝜃) ∈ V for all 𝜆 > 0, and that a function 𝐹 defined
in V and valued in some vector space is said to be homogeneous of degree 𝑑 if
𝐹 (𝑥, 𝑦, 𝜆𝜃) = 𝜆 𝑑 𝐹 (𝑥, 𝑦, 𝜃) for all (𝑥, 𝑦, 𝜃) ∈ V, 𝜆 > 0. As before Ω 𝑗 ( 𝑗 = 1, 2)
will be open subsets of R𝑛 , with “central” points 𝑥 ◦ ∈ Ω1 , 𝑦 ◦ ∈ Ω2 ; we will identify
Ω 𝑗 × (R𝑛 \ {0}) with 𝑇 ∗ Ω 𝑗 \0.
Definition 18.4.17 Let Γ ( 𝜄) (𝜄 = 1, 2) be two open cones in R 𝑁 \ {0}. Two phase-
functions 𝜑 ( 𝜄) (𝑥, 𝑦, 𝜃) in Ω1 × Ω2 × Γ ( 𝜄) (𝜄 = 1, 2) will be said to be equivalent at
(𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0) if there are vectors 𝜃 ( 𝜄) ∈ Γ ( 𝜄) satisfying

𝜑 (𝜃𝜄) 𝑥 ◦ , 𝑦 ◦ , 𝜃 ( 𝜄) = 0, 𝜑 𝑥( 𝜄) 𝑥 ◦ , 𝑦 ◦ , 𝜃 ( 𝜄) = 𝜉 ◦ , 𝜑 𝑦( 𝜄) 𝑥 ◦ , 𝑦 ◦ , 𝜃 ( 𝜄) = −𝜂◦ , (18.4.20)

and a conic open subset V ( 𝜄) of Ω1 × Ω2 × Γ ( 𝜄) containing 𝑥 ◦ , 𝑦 ◦ , 𝜃 ( 𝜄) such that

the following is valid:
There is a diffeomorphism (𝑥, 𝑦, 𝜃) ↦→ (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) of V (1) onto V (2) ,
homogeneous of degree 1, such that 𝜃 (2) = ℎ 𝑥 ◦ , 𝑦 ◦ , 𝜃 (1) and

𝜑 (1) (𝑥, 𝑦, 𝜃) = 𝜑 (2) (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) . (18.4.21)

Dℎ
The Jacobian determinant D𝜃 does not vanish at any point of V (1) . Consider a
Fourier distribution kernel
∫
(2) ( 𝑥,𝑦, 𝜃)
𝐴 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜑 𝑎 (2) (𝑥, 𝑦, 𝜃) d𝜃, (18.4.22)
Γ

where 𝑎 (2) ∈ 𝑆 𝑚 Ω1 × Ω2 , R 𝑁 , supp 𝑎 (2) ⊂ V (2) and |𝜃| < 1 =⇒ 𝑎 (2) (𝑥, 𝑦) = 0.

By (18.4.21) we obtain
∫
(2) Dℎ
𝐴 (𝑥, 𝑦) = (2𝜋) −𝑁 e 𝜄 𝜑 ( 𝑥,𝑦,ℎ ( 𝑥,𝑦, 𝜃)) 𝑎 (2) (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) (𝑥, 𝑦, 𝜃) d𝜃
Γ D𝜃
(18.4.23)
∫
(1)
= (2𝜋) −𝑁 e 𝜄 𝜑 ( 𝑥,𝑦, 𝜃) 𝑎 (1) (𝑥, 𝑦, 𝜃) d𝜃,
Γ

where 𝑎 (1) (𝑥, 𝑦, 𝜃) ∈ 𝑆 𝑚 Ω1 × Ω2 , R 𝑁 , supp 𝑎 (1) ⊂ V (1) .

Below we use the notation
n o
Σ 𝜑 ( 𝜄) = (𝑥, 𝑦, 𝜃) ∈ Ω1 × Ω2 × Γ ( 𝜄) ; 𝜑 (𝜃𝜄) (𝑥, 𝑦, 𝜃) = 0 , 𝜄 = 1, 2.

Dℎ
The nonvanishing of D𝜃 implies directly
18.4 Reduction of the Fiber Variables 723

Proposition 18.4.18 Suppose that two phase-functions 𝜑 (1) and 𝜑 (2) are equivalent
at (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) and let 𝜃 (1) and 𝜃 (2) be as in Definition 18.4.17. The diffeomorphism
(𝑥, 𝑦, 𝜃) ↦→ (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) of V (1) onto V (2) in Definition 18.4.17 induces a
diffeomorphism of Σ 𝜑 ( 𝜄) ∩ V (1) onto Σ 𝜑 ( 𝜄) ∩ V (1) ; thus dim Σ 𝜑 (1) = dim Σ 𝜑 (2) and
the rank of d𝜑 (1) ◦ ◦ (1) and that of d𝜑 (2) at 𝑥 ◦ , 𝑦 ◦ , 𝜃 (2) are equal. In

𝜃 at 𝑥 , 𝑦 , 𝜃 𝜃
particular, 𝜑 (1) is nondegenerate at 𝑥 ◦ , 𝑦 ◦ , 𝜃 (1) if and only if 𝜑 (2) is nondegenerate

at 𝑥 ◦ , 𝑦 ◦ , 𝜃 (2) .

We recall that the signature of a real symmetric 𝑁 × 𝑁 matrix 𝑀 is the difference

sgn 𝑀 = #eigenvalues> 0 − #eigenvalues< 0. It is a well-known fact (easy to prove)
that if 𝐴 ∈ GL (𝑁, R) then sgn ( 𝐴⊤ 𝑀 𝐴) = sgn 𝑀.

Proposition 18.4.19 Let 𝜑 ( 𝜄) ∈ C ∞ Ω1 × Ω2 × Γ ( 𝜄) (𝜄 = 1, 2) be phase-functions
as in Definition 18.4.17 and let Λ 𝜑 ( 𝜄) denote their associated Lagrangian submanifold
(Definition 18.2.3). If 𝜑 (1) and 𝜑 (2) are equivalent at (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ Λ 𝜑 (1) ∩ Λ 𝜑 (2)
then the following conditions are satisfied:
(i) There is a conic neighborhood V of (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) in (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0)
such that Λ 𝜑 (1) ∩ V = Λ 𝜑 (2) ∩ V.
(ii) There are vectors 𝜃 ( 𝜄) ∈ Σ 𝜑 ( 𝜄) ∩ Γ ( 𝜄) satisfying (18.4.20) and

sgn 𝜑 (1)
𝜃,𝜃 𝑥 ◦ ◦ (1)
, 𝑦 , 𝜃 = sgn 𝜑 (2)
𝜃,𝜃 𝑥 ◦ ◦ (2)
, 𝑦 , 𝜃 . (18.4.24)

Proof Let V ( 𝜄) be a conic open subset of Ω1 × Ω2 × Γ ( 𝜄) as in Definition 18.4.17

and suppose (18.4.21) holds. We apply Proposition 18.4.18: if (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) ∈
V (2) ∩ Σ 𝜑 (2) then (𝑥, 𝑦, 𝜃) ∈ V (1) ∩ Σ 𝜑 (1) and

𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) = 𝜕𝑥 𝜑 (2) (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃))

= 𝜑 𝑥(2) (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) + (𝜕𝜃 ℎ (𝑥, 𝑦, 𝜃)) 𝜑 (2)

𝜃 (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃))
= 𝜑 𝑥(2) (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) ,

and likewise for the partial derivatives with respect to 𝑦, thereby proving Property
(i). We also obtain

𝜑 (1) (2)
𝜃 , 𝜃 (𝑥, 𝑦, 𝜃) = (𝜕 𝜃 ℎ (𝑥, 𝑦, 𝜃)) 𝜑 𝜃 , 𝜃 (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) 𝜕 𝜃 ℎ (𝑥, 𝑦, 𝜃) ,

proving Property (ii). [It is understood, here and throughout the sequel, that the
“multiplication” of the various vectors (i.e., gradients) and matrices (of second
partial derivatives), in short, tensors, are carried out as prescribed by the chain-
rule.] □

Actually the conjunction of (i) and (ii) implies that the phase-functions 𝜑 (1)
and 𝜑 (2) are equivalent at (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) provided they are clean and have the
same excess (Definition 18.1.5), implying dim Σ 𝜑 (1) = dim Σ 𝜑 (2) . For the sake of
724 18 Fourier Integral Operators

simplicity we shall prove this result for nondegenerate phase-functions, in which

case dim Σ 𝜑 (1) = dim Σ 𝜑 (2) = 2𝑛. For the general case see [Hörmander, 1983, III],
Ch. XXI, or [Treves, 1980], pp. 421–427.
We need a couple of lemmas; Sym (𝑁, R) is the linear space of real symmetric
𝑁 × 𝑁 matrices.
Lemma 18.4.20 Let 𝐺 ∈ Sym (𝑁, R) and define

ℜ𝐺 = {𝑆 ∈ Sym (𝑁, R) ; det (𝐼 + 𝑆𝐺) ≠ 0} .

The following properties of a pair of elements of ℜ𝐺 , 𝑆0 and 𝑆1 , are equivalent:

(1) 𝑆0 and 𝑆1 belong to the same connected component of ℜ𝐺 ;
(2) sgn (𝐺 + 𝐺𝑆0 𝐺) = sgn (𝐺 + 𝐺𝑆1 𝐺).
Proof That (1) entails (2) is evident, since sgn (𝐺 + 𝐺𝑆𝐺) is a Z-valued continu-
ous function of 𝑆 ∈ Sym (𝑁, R). We now show that (2) entails (1). We shall not
distinguish between matrices and the linear maps they define (in fixed Euclidean co-
ordinates). Call 𝑬 = (ker 𝐺) ⊥ = Im 𝐺. Since (𝐼 + 𝑆𝐺) = 𝐼 on ker 𝐺, for 𝐼 + 𝑆𝐺 to be
nonsingular it is necessary and sufficient that (𝐼 + 𝑆𝐺)| 𝑬 be injective, in turn equiv-
alent to (𝐺 + 𝐺𝑆𝐺)| 𝑬 being an automorphism of 𝑬. Let 𝑃 𝑬 denote the orthogonal
projection onto 𝑬; we have 𝐺 = 𝑃 𝑬 𝐺 = 𝐺𝑃 𝑬 , hence

(𝐺 + 𝐺𝑆𝐺)| 𝑬 = (𝐺 + 𝐺𝑃 𝑬 𝑆𝑃 𝑬 𝐺)| 𝑬 .

Since 𝑃 𝑬 𝑆𝑃 𝑬 is a symmetric (with respect to the scalar product inherited from R 𝑁 )

linear transformation of 𝑬 we see that we may as well deal with a nonsingular matrix
𝐺. It suffices then to find a continuous path

[0, 1] ∋ 𝑡 ↦→ 𝑀𝑡 = 𝐺 + 𝐺𝑆𝑡 𝐺 ∈ Sym (𝑁, R) ∩ GL (𝑁, R) ;

𝑆𝑡 = 𝐺 −1 𝑀𝑡 𝐺 −1 − 𝐺 −1 would connect continuously 𝑆0 to 𝑆1 . For each 𝑗 = 0, 1 we

have the orthogonal sum decomposition R 𝑁 = 𝑬 +𝑗 ⊕ 𝑬 −𝑗 with 𝑬 +𝑗 (resp., 𝑬 −𝑗 ) the
eigenspace of 𝑀 𝑗 on which 𝑀 𝑗 > 0 (resp., 𝑀 𝑗 < 0); necessarily dim 𝑬 ±0 = dim 𝑬 ±1 .
There is an orthogonal transformation 𝑇 of R 𝑁 transforming 𝑬 ±0 into 𝑬 ±1 ; we have
𝑀1 = 𝑇 𝑀0𝑇 −1 . If det 𝑇 > 0 there is a curve [0, 1] ∋ 𝑡 ↦→ 𝑇𝑡 ∈ O+ (𝑁, R), the
connected component of the identity in the group O (𝑁, R) of real orthogonal 𝑁 × 𝑁
matrices, such that 𝑇0 = 𝐼 and 𝑇1 = 𝑇; the matrices 𝑀𝑡 = 𝑇𝑡 𝑀0𝑇𝑡−1 satisfy our
requirements. If det 𝑇 < 0 a change of the orientation of R 𝑁 brings us to the
previous situation. □
Lemma 18.4.21 Let Γ be an open cone in R 𝑁 \ {0}. Let 𝜑 ( 𝜄) ∈ C ∞ (Ω1 × Ω2 × Γ)
denote two nondegenerate phase-functions such that

Σ 𝜑 (1) = Σ 𝜑 (2) = Σ ⊂ Ω1 × Ω2 × Γ

and
(∗) 𝜑 (1) − 𝜑 (2) vanishes to second order on Σ.
18.4 Reduction of the Fiber Variables 725

If sgn 𝜑 (1) (2) ◦ ◦ ◦

𝜃 , 𝜃 = sgn 𝜑 𝜃 , 𝜃 at a point (𝑥 , 𝑦 , 𝜃 ) ∈ Σ ⊂ Ω1 × Ω2 × Γ then 𝜑
(1)
(2) ◦ ◦ ◦ ◦ ◦ ◦
and 𝜑 are equivalent at (𝑥 , 𝜉 , 𝑦 , 𝜂 ) ∈ Ω1 × Ω2 × (R \ {0}) with 𝜉 , 𝜂 as in
𝑛

(18.4.20).

Proof Throughout the argument we allow Ω1 × Ω2 × Γ to be contracted as needed,

about (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). We can use 𝜑 (2) (2)
𝜃1 , ..., 𝜑 𝜃 𝑁 as coordinates transversal to Σ; then
Taylor expansion in a conic neighborhood of (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) yields, by the hypothesis:
1
𝜑 (1) = 𝜑 (2) + 𝜑 (2) · 𝐺𝜑 (2)
𝜃 , (18.4.25)
2 𝜃
where 𝐺 = 𝐺 (𝑥, 𝑦, 𝜃) is a Sym (𝑁, R)-valued C ∞ function in Ω1 × Ω2 × Γ, homo-
geneous of degree 1. We derive the following equations in Σ:

d𝜑 (1) (2) (2)
𝜃 = 𝐼 𝑁 + 𝐺𝜑 𝜃 , 𝜃 d𝜑 𝜃 , (18.4.26)

where 𝐼 𝑁 is the 𝑁 × 𝑁 identity matrix (and the matrix multiplication must be carried
out according to the chain-rule); (18.4.26) and the fact that d𝜑 (𝜃𝜄)1 ∧ · · · ∧ d𝜑 (𝜃𝜄)𝑁 ≠ 0
(𝜄 = 1, 2) implies
det 𝐼 𝑁 + 𝐺𝜑 (2)
𝜃,𝜃 ≠ 0 (18.4.27)

everywhere in Ω1 × Ω2 × Γ.
Case 𝐺 small.
Let ∥·∥ stand for the matrix norm. By “𝐺 small” we mean that ∥𝐺 (𝑥, 𝑦, 𝜃) ∥ ≤ 𝜀 |𝜃|
for some small 𝜀 > 0 and all (𝑥, 𝑦, 𝜃) ∈ Ω1 × Ω2 × Γ. We seek a 𝐶 ∞ function 𝑊 in
Ω1 × Ω2 × Γ valued in Sym (𝑁, R), homogeneous of degree 1, such that (18.4.21)
will be valid if we put

ℎ (𝑥, 𝑦, 𝜃) = 𝜃 + 𝑊 (𝑥, 𝑦, 𝜃) 𝜑 (2)

𝜃 (𝑥, 𝑦, 𝜃) . (18.4.28)

Since 𝑊 𝑥, 𝑦, 𝜃
|𝜃 | 𝜑 (2)
𝜃 (𝑥, 𝑦, 𝜃) ≪ 1 for 𝜀 small, the map

(𝑥, 𝑦, 𝜃) ↦→ (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃))

will be a diffeomorphism
of Ω1 × Ω2 × Γ onto a conic open subset V of Ω1 × Ω2 ×
R 𝑁 \ {0} such that V ∩ Σ = (Ω1 × Ω2 × Γ) ∩ Σ. Taylor expansion yields:

𝜑 (2) (𝑥, 𝑦, ℎ (𝑥, 𝑦, 𝜃)) = 𝜑 (2) (𝑥, 𝑦, 𝜃) + 𝜑 (2) (2)

𝜃 (𝑥, 𝑦, 𝜃) · 𝑊 (𝑥, 𝑦, 𝜃) 𝜑 𝜃 (𝑥, 𝑦, 𝜃)
(18.4.29)
+𝑊 (𝑥, 𝑦, 𝜃) 𝜑 (2) (2)
𝜃 (𝑥, 𝑦, 𝜃) · 𝐻 (𝑥, 𝑦, 𝜃, 𝑊) 𝑊 (𝑥, 𝑦, 𝜃) 𝜑 𝜃 (𝑥, 𝑦, 𝜃) ,

where 𝐻 (𝑥, 𝑦, 𝜃, 𝑊) ∈ Sym (𝑁, R) is a C ∞ function in Ω1 × Ω2 × Γ × Sym (𝑁, R),

homogeneous of degree 1. We require 𝑊 to be a solution of the equation
1
𝑊 + 𝑊 𝐻 (𝑥, 𝑦, 𝜃, 𝑊) 𝑊 = 𝐺. (18.4.30)
2
726 18 Fourier Integral Operators

The differential of the left-hand side with respect to 𝑊 is equal to 𝐼 𝑁 at 𝑊 = 0.

Provided 𝜀 > 0 is sufficiently small the Implicit Function Theorem implies the
existence and uniqueness of a solution 𝑊 ∈ Sym (𝑁, R) such that 𝑊 = 0 when
𝐺 = 0. Putting (18.4.30) into (18.4.29) and taking (18.4.25) into account yields
(18.4.21) and thus, that 𝜑 (1) and 𝜑 (2) are equivalent at (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ).
General Case.
What we have just proved has the following corollary. If we can find a continuous
map [0, 1] ∋ 𝑡 ↦→ 𝜑 (1+𝑡) into the space (equipped with its natural topology) of
nondegenerate phase-functions in Ω1 ×Ω2 ×Γ (suitably contracted) such that 𝜑 (1+𝑡) −
𝜑 (1) vanishes to second order on Σ for every 𝑡 then 𝜑 (1) and 𝜑 (2) are equivalent at
(𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ). Indeed, we have proved that the subset of [0, 1] in which 𝜑 (1+𝑡) is
equivalent to 𝜑 (2) is open and closed.
There is no loss of generality in assuming that |𝜃 ◦ | = 1, Ω1 ×Ω2 is a biball centered
at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) and Γ is a cone of revolution around the ray R𝜃 ◦ . According to Lemma
18.4.20 and (18.4.26) we can find a continuous map 𝐺 (𝑡) : [0, 1] −→ Sym (𝑁, R)
such that 𝐺 (0) = 0, 𝐺 (1) = 𝐺 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) and

det 𝐼 𝑁 + 𝐺 (𝑡) 𝜑 (2)
𝜃,𝜃 (𝑥 ◦ ◦ ◦
, 𝑦 , 𝜃 ) ≠ 0.

The phase-function Φ (𝑡) = 𝜑 (2) + 21 |𝜃| 𝐺 (𝑡) 𝜑 (2) (2)

𝜃 · 𝜑 𝜃 is nondegenerate; since Φ
(0) =
(𝑡) (2)
𝜑 (2) and Φ 𝜃 = 𝜑 𝜃 on Σ we see that Φ (𝑡) − 𝜑 (2) vanishes to second order on Σ for
every 𝑡. We conclude that 𝜑 (2) and Φ (1) = 𝜑 (2) + 21 |𝜃| 𝐺 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) 𝜑 (2) (2)
𝜃 · 𝜑 𝜃 are
equivalent at (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ). The only thing left to prove is that Φ (1) is equivalent to
𝜑 (1) at (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ). For 0 ≤ 𝑡 ≤ 1 define

(1+𝑡) ◦ ◦ 𝜃
𝐺 (𝑥, 𝑦, 𝜃) = 𝐺 (1 − 𝑡) 𝑥 + 𝑡𝑥, 𝑦, (1 − 𝑡) 𝜃 + 𝑡 ;
|𝜃|

then 𝐺 (1) = 𝐺 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) as before, and |𝜃| 𝐺 (2) (𝑥, 𝑦, 𝜃) = 𝐺 (𝑥, 𝑦, 𝜃) for all
(𝑥, 𝑦, 𝜃) ∈ Γ. The phase-functions Φ (1+𝑡) = 𝜑 (2) + 21 |𝜃| 𝐺 (1+𝑡) 𝜑 (2) (2)
𝜃 · 𝜑 𝜃 are nonde-
generate and Φ (1+𝑡) − 𝜑 (1) vanishes to second order on Σ for every 𝑡. They connect
Φ (1) to 𝜑 (1) , proving what we wanted. □

Proposition 18.4.22 Let 𝜑 ( 𝜄) ∈ C ∞ Ω1 × Ω2 × Γ ( 𝜄) (𝜄 = 1, 2) be two phase-
functions as in Definition 18.4.17 and let Λ 𝜑 ( 𝜄) denote their associated Lagrangian
submanifolds. If 𝜑 (1) and 𝜑 (2) are nondegenerate and if both Conditions (i) and
(ii) in Proposition 18.4.19 are satisfied then 𝜑 (1) and 𝜑 (2) are equivalent at
(𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) ∈ Λ 𝜑 (1) ∩ Λ 𝜑 (2) .

Proof Let 𝜃 (1) , 𝜃 (2) be the vectors in Condition

(ii) in Proposition 18.4.19 and 𝑔 be
a linear automorphism of R such that 𝑔 𝜃
𝑁 (1) = 𝜃 (2) . After replacing 𝜑 (2) (𝑥, 𝑦, 𝜃)
by 𝜑 (2) (𝑥, 𝑦, 𝑔 (𝜃)) we can assume that 𝜃 (1) = 𝜃 (2) = 𝜃 ◦ ; we select an open cone Γ
such that 𝜃 ◦ ∈ Γ ⊂ Γ (1) ∩ Γ (2) . There is a conic neighborhood V of (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ )
18.4 Reduction of the Fiber Variables 727

in (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0) such that Λ 𝜑 (1) ∩ V = Λ 𝜑 (2) ∩ V = Λ [Condition (i),

Proposition 18.4.19]. Here
n o
Σ 𝜑 ( 𝜄) = (𝑥, 𝑦, 𝜃) ∈ Ω1 × Ω2 × Γ; 𝜑 (𝜃𝜄) (𝑥, 𝑦, 𝜃) = 0 ;

dim Σ 𝜑 ( 𝜄) = 2𝑛. We exploit the hypothesis that 𝜑 ( 𝜄) is nondegenerate (Proposition

18.1.8):
(1) For each 𝜄 = 1, 2, there is a conic neighborhood U ( 𝜄) of (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) in Ω1 ×Ω2 ×Γ
such that the map

𝐽 ( 𝜄) : (𝑥, 𝑦, 𝜃) ↦→ 𝑥, 𝑦, 𝜑 𝑥( 𝜄) (𝑥, 𝑦, 𝜃) , −𝜑 𝑦( 𝜄) (𝑥, 𝑦, 𝜃)

is a diffeomorphism of Σ 𝜑 ( 𝜄) ∩ U ( 𝜄) onto Λ, homogeneous of degree 1. The

−1
composite 𝐽 (2) ◦ 𝐽 (1) is a diffeomorphism of Σ 𝜑 (1) ∩ U (1) onto Σ 𝜑 (2) ∩ U (2) .
(2) For each 𝜄 = 1, 2, the functions 𝜑 (𝜃𝜄)𝑗 (𝑥, 𝑦, 𝜃), 𝑗 = 1, ..., 𝑁, can be used as
coordinates transversal to Σ 𝜑 ( 𝜄) in U ( 𝜄) suitably contracted.

The equations 𝜑 𝑥(2) (𝑥, 𝑦, 𝜃) = 𝜉, 𝜑 𝑦(2) (𝑥, 𝑦, 𝜃) = −𝜂, 𝜑 (2)

𝜃 (𝑥, 𝑦, 𝜃) = 0, have a
unique smooth solution 𝜃 = 𝜃 (2) (𝑥, 𝑦, 𝜉, 𝜂) in V such that 𝜃 (2) (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) = 𝜃 ◦ .
−1
Thus 𝐽 (2) is the map

Λ ∋ (𝑥, 𝜉, 𝑦, 𝜂) ↦→ 𝑥, 𝑦, 𝜃 (2) (𝑥, 𝜉, 𝑦, 𝜂) ∈ Σ 𝜑 (2) ∩ V

and the composite

−1
𝐽 (2) ◦ 𝐽 (1) : (𝑥, 𝑦, 𝜃) ↦→ 𝑥, 𝑦, 𝜃 (2) 𝑥, 𝑦, 𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) , −𝜑 𝑦(1) (𝑥, 𝑦, 𝜃) (18.4.31)

is a diffeomorphism of Σ 𝜑 (1) ∩ U (1) onto Σ 𝜑 (2) ∩ U (2) (after suitable contractions

of U (1) and U (2) ). This entails

𝜑 (1)
𝜃 (𝑥, 𝑦, 𝜃) = 𝜑 (2)
𝜃 𝑥, 𝑦, 𝜃 (2)
𝑥, 𝑦, 𝜑 (1)
𝑥 (𝑥, 𝑦, 𝜃) , −𝜑 (1)
𝑦 (𝑥, 𝑦, 𝜃) = 0. (18.4.32)

We derive that 𝜑 (1)

𝜃 (𝑥, 𝑦, 𝜃) = 0 implies

𝜕 (2)
𝜑 𝑥, 𝑦, 𝜃 (2) 𝑥, 𝑦, 𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) , −𝜑 𝑦(1) (𝑥, 𝑦, 𝜃) = 𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) ,
𝜕𝑥
𝜕 (2)
𝜑 𝑥, 𝑦, 𝜃 (2) 𝑥, 𝑦, 𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) , −𝜑 𝑦(1) (𝑥, 𝑦, 𝜃) = 𝜑 𝑦(1) (𝑥, 𝑦, 𝜃) ,
𝜕𝑦
𝜕 (2)
𝜑 𝑥, 𝑦, 𝜃 (2) 𝑥, 𝑦, 𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) , −𝜑 𝑦(1) (𝑥, 𝑦, 𝜃) = 0,
𝜕𝜃
728 18 Fourier Integral Operators

and we reach the conclusion that

𝜑 (1) (𝑥, 𝑦, 𝜃) − 𝜑 (2) 𝑥, 𝑦, 𝜃 (2) 𝑥, 𝑦, 𝜑 𝑥(1) (𝑥, 𝑦, 𝜃) , −𝜑 𝑦(1) (𝑥, 𝑦, 𝜃)

vanishes to second order on Σ 𝜑 (1) . This property in conjunction with Property (ii)
in Proposition 18.4.19 enables us to apply Lemma 18.4.21 to reach the desired
conclusion. □

18.4.4 FIOs that are pseudodifferential operators

Proposition 18.4.23 Suppose that Ω1 = Ω2 = Ω and 𝑛 ≤ 𝑁. If the phase-function

𝜑 is nondegenerate and the Lagrangian submanifold Λ 𝜑 associated with 𝜑 is the
diagonal of (𝑇 ∗ Ω\0) × (𝑇 ∗ Ω\0) (i.e., Λ 𝜑 is the graph of the identity transformation
of 𝑇 ∗ Ω\0) then 𝑨 defined in (18.3.1) is a pseudodifferential operator in Ω.

Proof Since 𝜑 is nondegenerate we have dim Σ 𝜑 = 2𝑛 and 𝜑 is strongly nondegen-

erate (Proposition 18.2.13). The base projection of Λ 𝜑 is the diagonal in Ω × Ω,
diag Ω × Ω, whence rank 𝜋 ∗ | Λ 𝜑 = 𝑛. If 𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0 for some 𝜃 ∈ Γ then neces-
sarily 𝑥 = 𝑦; thus Σ 𝜑 ⊂ (diag Ω × Ω) × Γ. This said, it suffices to prove the claim
locally, i.e., in a conic neighborhood of an arbitrary point (𝑥 ◦ , 𝜉 ◦ , 𝑥 ◦ , 𝜉 ◦ ) ∈ Λ 𝜑 ; we
select 𝜃 ◦ ∈ Γ such that (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 and 𝜑 𝑥 (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) = 𝜉 ◦ . As usual we allow
ourselves to contract Γ about the ray (𝜆𝜃 ◦ ) 𝜆>0 as much as needed.
We avail ourselves of Corollary 18.4.14: there is no loss of generality in as-
suming that 𝑁 = 𝑛, 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) = 0, and therefore (see Definition 18.2.10)
det 𝜑 𝑥, 𝜃 (𝑥, 𝑦, 𝜃) ≠ 0 whatever (𝑥, 𝑦, 𝜃) in a suitable conic neighborhood of
(𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ), which we take of the form 𝑈 × 𝑈 × Γ. The Euler homogeneity for-
Í 𝜕𝜑
mula entails 𝜑 = 𝑛𝑘=1 𝜃 𝑘 𝜕𝜃 𝑘
; by hypothesis we have
𝑛
𝜕𝜑 ∑︁
(𝑥, 𝑦, 𝜃) = 𝑥 𝑗 − 𝑦 𝑗 𝑏 𝑗,𝑘 (𝑥, 𝑦, 𝜃) , 𝑘 = 1, ..., 𝑛.
𝜕𝜃 𝑘 𝑗=1

We may assume that the matrix 𝑏 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 is nonsingular at every point of
𝑈 × 𝑈 × Γ. The following change of the variables of integration in (18.1.13):
𝑛
∑︁
𝜉𝑗 = 𝑏 𝑗,𝑘 (𝑥, 𝑦, 𝜃) 𝜃 𝑘 (18.4.33)
𝑘=1

yields
∫
𝐴 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜉 · ( 𝑥−𝑦) 𝑎 (𝑥, 𝑦, 𝜃 (𝑥, 𝑦, 𝜉)) det 𝜕 𝜉 𝜃 (𝑥, 𝑦, 𝜉) d𝜉,
Γ
18.5 Composition and Continuity of Fourier Integral Operators 729

where 𝜃 (𝑥, 𝑦, 𝜉) is the solution with respect to 𝜃 of the equations (18.4.33) such that
𝜃 (𝑥 ◦ , 𝑥 ◦ , 𝜉 ◦ ) = 𝜃 ◦ and 𝜕 𝜉 𝜃 is the corresponding Jacobian matrix. We are assuming
that supp 𝑎 ⊂ 𝑈 × 𝑈 × Γ♭ , 𝜃 ◦ ∈ Γ♭ ∩ S𝑛−1 ⊂⊂ Γ. Thanks to Corollary 18.1.11
and Proposition 18.1.12 we may also assume that 𝑎 ≡ 0 in a ball |𝜃| < 𝑅. Since 𝜃
and 𝜉 are homogeneous functions (of degree 1) of one another it is readily checked
that 𝑎 (𝑥, 𝑦, 𝜃 (𝑥, 𝑦, 𝜉)) det 𝜕 𝜉 𝜃 (𝑥, 𝑦, 𝜉) is a standard amplitude of order 𝑚. By
Definition 16.1.14 this completes the proof of Proposition 18.4.23. □

18.5 Composition and Continuity of Fourier Integral Operators

18.5.1 Composition of FIOs

Let 𝑨 be the FIO defined by the distribution kernel (18.1.13); as before the phase-
function 𝜑 ∈ C ∞ (Ω1 × Ω2 × Γ), Γ an open cone in R 𝑁 \ {0}, 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ),
𝑚 ∈ R. We introduce a second phase-function 𝜓 in Ω2 × Ω3 × Γ♭ , where Ω3 is a
♭
third open subset of R𝑛 and Γ♭ an open cone in R 𝑁 \ {0}. We shall assume that both
phase-functions 𝜑 and 𝜓 are strongly nondegenerate (Definition 18.2.10). We let 𝑨
act on the distribution kernel in Ω2 × Ω3 ,
∫
𝐵 (𝑦, 𝑧) = (2𝜋) −𝑁
♭
e𝑖 𝜓 ( 𝑦,𝑧, 𝜔) 𝑏 (𝑦, 𝑧, 𝜔) d𝜔, (18.5.1)
Γ♭

with 𝑏 ∈ 𝑆c𝑚1 Ω2 × Ω3 , Γ♭ , 𝑚 1 ∈ R. Unless specified otherwise, we assume that
the amplitude 𝑏 (and therefore the distribution kernel 𝐵) is properly supported
(Definition 18.1.15). We shall reason in conic neighborhoods of (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑
and (𝑦 ◦ , 𝑧◦ , 𝜔◦ ) ∈ Σ 𝜓 where

Σ 𝜑 = {(𝑥, 𝑦, 𝜃) ∈ Ω1 × Ω2 × Γ; 𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0} ,
n o
Σ 𝜓 = (𝑦, 𝑧, 𝜔) ∈ Ω2 × Ω3 × Γ♭ ; 𝜓 𝜔 (𝑦, 𝑧, 𝜔) = 0 .

Under these hypotheses we form the oscillatory integral

∫
( 𝐴 ◦ 𝐵) (𝑥, 𝑧) = 𝐴 (𝑥, 𝑦) 𝐵 (𝑦, 𝑧) d𝑦 (18.5.2)
Ω2
∫ ∫ ∫
= (2𝜋) −𝑁 −𝑁
♭
e𝑖 𝜑 ( 𝑥,𝑦, 𝜃)+𝑖 𝜓 ( 𝑦,𝑧, 𝜔) 𝑎 (𝑥, 𝑦, 𝜃) 𝑏 (𝑦, 𝑧, 𝜔) d𝑦d𝜃d𝜔.
Ω2 Γ Γ♭

Proposition 18.5.1 Let 𝜑 (𝑥, 𝑦, 𝜃) be a strongly nondegenerate phase-function in

Ω1 × Ω2 × Γ and 𝜓 (𝑦, 𝑧, 𝜔) be a phase-function in Ω2 × Ω3 × Γ♭ that has the
following property:
730 18 Fourier Integral Operators

𝜕𝜓 𝜕𝜓
(i) the differentials d 𝑦 𝜕𝜔1 , ..., d 𝑦 𝜕𝜔 𝑁 ♭ are linearly independent at every point
of Σ 𝜓 .
Suppose, moreover, that the following condition is satisfied:
(ii) to every (𝑥, 𝑦, 𝜃) ∈ Σ 𝜑 there is a (𝑦, 𝑧, 𝜔) ∈ Σ 𝜓 satisfying

𝜓 𝑦 (𝑦, 𝑧, 𝜔) = −𝜑 𝑦 (𝑥, 𝑦, 𝜃) . (18.5.3)

Under these hypotheses, if 𝐴 is given by (18.1.13) and 𝐵 by (18.5.1) then 𝐴 ◦ 𝐵

is a Fourier distribution kernel in Ω1 × Ω3 .
♭
Proof Let Γ and Γ♭ be the open cones in√︃R 𝑁 \ {0} and R 𝑁 \ {0} respectively,

introduced above; we use the notation 𝜆 = |𝜃| 2 + |𝜔| 2 and 𝜏 = 𝜏 (1) , 𝜏 (2) , 𝜏 (3) ,
where
𝜏 (1) = 𝜆𝑦, 𝜏 (2) = 𝜃, 𝜏 (3) = 𝜔. (18.5.4)
We define, for 𝜏 (1) ∈ 𝜆Ω2 , 𝜏 (2) ∈ Γ, 𝜏 (3) ∈ Γ♭ ,

Φ (𝑥, 𝑧, 𝜏) = 𝜑 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) + 𝜓 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3) (18.5.5)
= 𝜑 (𝑥, 𝑦, 𝜃) + 𝜓 (𝑦, 𝑧, 𝜔) .
Í 12 1
Keeping in mind that 𝜆 = 𝑁 +𝑁 ♭ 2
𝜏 = 𝜏 (2) 2 + 𝜏 (3) 2 2 we rewrite (18.5.2)
𝑗=1 𝑛+ 𝑗
as
∫
(2𝜋) 𝑁 +𝑁
♭
𝐴 (𝑥, 𝑦) 𝐵 (𝑦, 𝑧) d𝑦 (18.5.6)
Ω2
∫ ∫ ∫
= e𝑖Φ( 𝑥,𝑧, 𝜏) 𝑎 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) 𝑏 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3) 𝜆−𝑛 d𝜏.
Γ♭ Γ 𝜏 (1) ∈𝜆Ω2
√︃
We set 𝜏 ◦ = (𝜆◦ 𝑦 ◦ , 𝜃 ◦ , 𝜔◦ ), 𝜆◦ = |𝜃 ◦ | 2 + |𝜔◦ | 2 , and
n o
Γ = 𝜏 ∈ R𝑛 × Γ × Γ♭ ; 𝜆−1 𝜏 (1) ∈ Ω2 .
e

We are going to show that Φ (𝑥, 𝑧, 𝜏) is a nondegenerate phase-function in a conic

neighborhood of (𝑥 ◦ , 𝑧◦ , 𝜏 ◦ ) in Ω1 × Ω3 × e
Γ. We compute

𝜕Φ 𝜕𝜑 𝜕𝜓
𝜆 = (𝑥, 𝑦, 𝜃) + (𝑦, 𝑧, 𝜔)
𝜕𝜏 𝑗 𝜕𝑦 𝑗 𝜕𝑦 𝑗

where 𝑗 = 1, ..., 𝑛;
𝑛
𝜕Φ 𝜕𝜑 −2
∑︁ 𝜕𝜑 𝜕𝜓
= (𝑥, 𝑦, 𝜃) − 𝜆 𝜏𝑛+ 𝑗 𝑦𝑘 (𝑥, 𝑦, 𝜃) + (𝑦, 𝑧, 𝜔)
𝜕𝜏𝑛+ 𝑗 𝜕𝜃 𝑗 𝑘=1
𝜕𝑦 𝑘 𝜕𝑦 𝑘
18.5 Composition and Continuity of Fourier Integral Operators 731

where 𝑗 = 1, ..., 𝑁;
𝑛
𝜕Φ 𝜕𝜓 −2
∑︁ 𝜕𝜑 𝜕𝜓
= (𝑦, 𝑧, 𝜔) − 𝜆 𝜏𝑛+𝑁 + 𝑗 𝑦𝑘 (𝑥, 𝑦, 𝜃) + (𝑦, 𝑧, 𝜔)
𝜕𝜏𝑛+𝑁 + 𝑗 𝜕𝜔 𝑗 𝑘=1
𝜕𝑦 𝑘 𝜕𝑦 𝑘

where 𝑗 = 1, ..., 𝑁 ♭ . It follows that the submanifold

n o
ΣΦ = (𝑥, 𝑧, 𝜏) ∈ Ω1 × Ω3 × e Γ; Φ 𝜏 (𝑥, 𝑧, 𝜏) = 0

is defined by the equations (18.5.3) and

𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0, 𝜓 𝜔 (𝑦, 𝑧, 𝜔) = 0. (18.5.7)

Of course, the meaning of (18.5.7) is that (𝑥, 𝑦, 𝜃) ∈ Σ 𝜑 and (𝑦, 𝑧, 𝜔) ∈ Σ 𝜓 . By

hypothesis the determinant of the matrix (18.2.10) does not vanish at any point of Σ 𝜑 .
It follows that the differentials with respect to (𝑥, 𝜃) of the functions 𝜑 𝑦 𝑗 (𝑥, 𝑦, 𝜃) +
𝜓 𝑦 𝑗 (𝑦, 𝑧, 𝜔) and 𝜑 𝜃𝑘 (𝑥, 𝑦, 𝜃) are linearly independent at every point of Σ 𝜑 × Ω3 ×
Γ♭ ⊃ ΣΦ . Thanks to Hypothesis (i) we have, at every point of ΣΦ ,

∧𝑛𝑗=1 (d 𝑥 + d 𝜃 ) 𝜑 𝑦 𝑗 (𝑥, 𝑦, 𝜃) + 𝜓 𝑦 𝑗 (𝑦, 𝑧, 𝜔) ∧
𝑁♭
∧𝑁𝑘=1 (d 𝑥 + d 𝜃 ) 𝜑 𝜃 𝑘 (𝑥, 𝑦, 𝜃) ∧ ∧ℓ=1 d 𝑦 𝜓 𝜔ℓ (𝑦, 𝑧, 𝜔) ≠ 0,

which shows that Φ is a nondegenerate phase-function. □

The conclusion in Proposition 18.5.1 can be restated as follows: The composite
𝑨𝑩 is a Fourier integral operator with nondegenerate phase-function (18.5.5).

Proposition 18.5.2 Under the hypotheses of Proposition 18.5.1 the Lagrangian

submanifold associated with the Fourier integral operator 𝑨𝑩 is the submanifold
Λ 𝜓 ◦ Λ 𝜑 of (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω3 \0) consisting of the elements (𝑥, 𝜉, 𝑧, 𝜁) such that

∃ (𝑦, 𝜂) ∈ 𝑇 ∗ Ω2 \0, (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑 , (𝑦, 𝜂, 𝑧, 𝜁) ∈ Λ 𝜓 . (18.5.8)

The Lagrangian submanifold Λ 𝜑 ◦ Λ 𝜓 is locally the graph of a symplectomorphism

(𝑥, 𝜉) ↦→ (𝑧, 𝜁), the composite of the symplectomorphism (𝑥, 𝜉) ↦→ (𝑦, 𝜂) defined by
𝜑, followed by (𝑦, 𝜂) ↦→ (𝑧, 𝜁) defined by 𝜓.

The order of the compositions 𝑨𝑩, Λ 𝜓 ◦ Λ 𝜑 and the sequence (𝑥, 𝜉) ↦→ (𝑦, 𝜂) ↦→
(𝑧, 𝜁) conform with the meaning of the lower horizontal arrow in Diagram (18.2.11).
Proof Let us use, in the representation of 𝑨𝑩, the phase-function Φ (𝑥, 𝑧, 𝜏) de-
fined in (18.5.5). The submanifold ΣΦ ⊂ Ω1 × Ω3 × e Γ is defined by the equations
(18.5.3) and (18.5.7). If these equations hold and 𝜉 = Φ 𝑥 (𝑥, 𝑧, 𝜏) = 𝜑 𝑥 (𝑥, 𝑦, 𝜃),
𝜁 = −Φ𝑧 (𝑥, 𝑧, 𝜏) = −𝜓 𝑧 (𝑦, 𝑧, 𝜔), then (𝑥, 𝜉, 𝑧, 𝜁) ∈ ΛΦ . If moreover 𝜂 =
𝜓 𝑦 (𝑦, 𝑧, 𝜔) = −𝜑 𝑦 (𝑥, 𝑦, 𝜃) then (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑 and (𝑦, 𝜂, 𝑧, 𝜁) ∈ Λ 𝜓 . □
732 18 Fourier Integral Operators

Remark 18.5.3 Under the hypotheses of Proposition 18.5.1 the correspondence

Σ 𝜑 ∋ (𝑥, 𝑦, 𝜃) ↦→ (𝑦, 𝑧, 𝜔) ∈ Σ 𝜓 in Condition (ii) is a local diffeomorphism, as is the
correspondence Λ 𝜓 ∋ (𝑦, 𝜂, 𝑧, 𝜁) ↦→ (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑 in (18.5.8). If (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ 𝜑
there is a 𝜃 ∈ Γ such that (𝑥, 𝑦, 𝜃) ∈ Σ 𝜑 , 𝜉 = 𝜑 𝑥 (𝑥, 𝑦, 𝜃), 𝜂 = −𝜑 𝑦 (𝑥, 𝑦, 𝜃); if
(𝑦, 𝜂, 𝑧, 𝜁) ∈ Λ 𝜓 there is an 𝜔 ∈ Γ♭ such that (𝑦, 𝑧, 𝜔) ∈ Σ 𝜓 , 𝜂 = 𝜓 𝑦 (𝑦, 𝑧, 𝜔),
𝜁 = −𝜓 𝑧 (𝑦, 𝑧, 𝜔). Thus (18.5.8) implies (ii).
Concerning the amplitude in (18.5.6) we can state the following: to every set of
multi-indices 𝛼,𝛽, 𝛾 ′ ∈ Z+𝑛 , 𝛾 ′′ ∈ Z+𝑁 , 𝛾 ′′′ ∈ Z+𝑁 , and to every compact subset 𝐾 of
♭

Ω1 × Ω3 there is a constant 𝐶𝐾 > 0 such that

𝛾′ 𝛾′′ 𝛾′′′
𝜕𝜏 (1) 𝜕𝜏 (2) 𝜕𝜏 (3) 𝜕𝑥𝛼 𝑎 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) 𝜕𝑧 𝑏 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3)
𝛽

≤ 𝐶𝐾 (1 + |𝜏|) 𝑚+𝑚1 −|𝛾 |

for all 𝜏 (1) ∈ 𝜆Ω2 , 𝜏 (2) ∈ Γ, 𝜏 (3) ∈ Γ♭ . We reach the conclusion that if we use the
phase-function (18.5.5) the amplitude of 𝑨𝑩,

𝑎 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) 𝑏 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3) 𝜆−𝑛 ,

has order 𝑚 + 𝑚 1 − 𝑛.
We avail ourselves of Proposition 18.4.15 (see also Remark 18.4.16): it allows us
to select the phase-functions 𝜑 and 𝜓 so that

𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0, 𝜓 𝜔, 𝜔 (𝑦 ◦ , 𝑧◦ , 𝜔◦ ) = 0; (18.5.9)

this implies 𝑁 ≤ 𝑛, 𝑁 ♭ ≤ 𝑛 (Proposition 18.4.10).

Lemma 18.5.4 Let Φ (𝑥, 𝑧, 𝜏) be the phase-function (18.5.5). If (18.5.9) holds then

rank Φ 𝜏, 𝜏 (𝑥 ◦ , 𝑧◦ , 𝜏 ◦ ) = 𝑁 + 𝑁 ♭ . (18.5.10)

Proof After a translation in 𝑦-space we may ◦

suppose 𝑦 = 0. We use the notation
𝜏 (2)◦ = 𝜃 ◦ , 𝜏 (3)◦ = 𝜔◦ , 𝜏 ◦ = 0, 𝜏 (2)◦ , 𝜏 (3)◦ . If we combine (18.5.3), (18.5.7) and
(18.5.9) we obtain, at the point (𝑥 ◦ , 𝑧◦ , 𝜏 ◦ ):

|𝜃 ◦ | 2 + |𝜔◦ | 2 Φ 𝜏 (1) , 𝜏 (1) = 𝜑 𝑦,𝑦 (𝑥 ◦ , 0, 𝜃 ◦ ) + 𝜓 𝑦,𝑦 (0, 𝑧◦ , 𝜔◦ ) , (18.5.11)
√︃
|𝜃 ◦ | 2 + |𝜔◦ | 2 Φ 𝜏 (1) , 𝜏 (2) = 𝜑 𝑦, 𝜃 (𝑥 ◦ , 0, 𝜃 ◦ ) ,
√︃
|𝜃 ◦ | 2 + |𝜔◦ | 2 Φ 𝜏 (1) , 𝜏 (3) = 𝜓 𝑦, 𝜔 (0, 𝑧◦ , 𝜔◦ ) ,
Φ 𝜏 (2) , 𝜏 (2) = 0, Φ 𝜏 (3) , 𝜏 (3) = 0, Φ 𝜏 (2) , 𝜏 (3) = 0,

implying that rank Φ 𝜏, 𝜏 (𝑥 ◦ , 𝑧◦ , 𝜏 ◦ ) = rank 𝜑 𝑦, 𝜃 (𝑥 ◦ , 0, 𝜃 ◦ ) + rank 𝜓 𝑦, 𝜔 (0, 𝑧◦ , 𝜔◦ ) =

𝑁 + 𝑁 ♭ by Proposition 18.4.12. □
∗
◦ ◦ ◦ ◦
Corollary 18.5.5 If (18.5.9) holds then rank 𝜋 | ΛΦ (𝑥 , 𝜉 , 𝑦 , 𝜂 ) = 𝑛.
18.5 Composition and Continuity of Fourier Integral Operators 733

Proof Since Φ (𝑥, 𝑧, 𝜏) is nondegenerate (see the end of the proof of Proposition
18.5.1) dim ΣΦ = 2𝑛. Then the claim follows directly from Lemma 18.5.4 and from
(18.4.1) with Φ substituted for 𝜑. □

We apply Proposition 18.4.6 to the phase-function Φ (𝑥, 𝑧, 𝜏) selecting 𝜈 = 𝑛.

Since the dimension of 𝜏-space is equal to 𝑁 + 𝑁 ♭ + 𝑛, Lemma18.5.4 implies the fol-
lowing: after a linear change of coordinates 𝜏1 , ..., 𝜏𝑛+𝑁 +𝑁 ♭ ⇝ 𝜏˜1 , ..., 𝜏˜𝑛+𝑁 +𝑁 ♭
we can introduce the split 𝜏˜ = ( 𝜏˜ ′, 𝜏˜ ′′) ∈ Γ ′ × Γ ′′, with Γ ′ ⊂ R𝑛 \ {0}, Γ ′′ ⊂
R 𝑁 +𝑁 \ {0} open cones, such that Φ 𝜏˜ ′ , 𝜏˜ ′ (𝑥 ◦ , 𝑧◦ , 𝜏˜ ◦ ) = 0, Φ 𝜏˜ ′ , 𝜏˜ ′′ (𝑥 ◦ , 𝑧◦ , 𝜏˜ ◦ ) = 0,
♭

det Φ 𝜏˜ ′′ , 𝜏˜ ′′ (𝑥 ◦ , 𝑧◦ , 𝜏˜ ◦ ) ≠ 0. As a consequence there is a unique smooth solution

𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′) of the equation Φ 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′) = 0 in a conic neighborhood of
(𝑥 ◦ , 𝑧◦ , 𝜏˜ ′◦ ) such that 𝜏˜ ′′ (𝑥 ◦ , 𝑧◦ , 𝜏˜ ′◦ ) = 𝜏˜ ′′◦ . We then define the new phase-function

Φ♭ (𝑥, 𝑧, 𝜏˜ ′) = Φ (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) . (18.5.12)

According to (18.4.9) we have Φ♭𝜏˜ ′ , 𝜏˜ ′ (𝑥 ◦ , 𝑧◦ , 𝜏˜ ′◦ ) = 0; Corollary 18.4.14 tells us that

Φ♭ is strongly nondegenerate.
Returning to the amplitudes in (18.5.6) we write

𝑐 (𝑥, 𝑧, 𝜏˜ ′) = 𝑎 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) 𝑏 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3) 𝜆−𝑛 (18.5.13)

1
2 2 2
where 𝜏 (1) , 𝜏 (2) , 𝜏 (3) , 𝜆 = 𝜏 (2) + 𝜏 (3) are now regarded as homogeneous
functions of 𝜏˜ = ( 𝜏˜ ′, 𝜏˜ ′′) of degree 1 and 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′) is as in (18.5.12). We have
shown earlier that 𝑐 (𝑥, 𝑧, 𝜏˜ ′) is an amplitude of order ≤ 𝑚 + 𝑚 1 − 𝑛; then
ℓ
𝑐 ℓ (𝑥, 𝑧, 𝜏˜ ′) = 𝜆ℓ D 𝜏˜ ′′ · 𝐻Φ
−1
D 𝜏˜ ′′ 𝑐 (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) (18.5.14)

is an amplitude of order ≤ 𝑚 + 𝑚 1 − 𝑛 − ℓ. WeÍconstruct a true amplitude 𝑐♭ (𝑥, 𝑧, 𝜏˜ ′)

∞
in Ω1 × Ω1 × Γ ′ from the formal amplitude ℓ=0 (2𝑖) ℓ 𝑐 ℓ (𝑥, 𝑧, 𝜏˜ ′) /ℓ!. By (18.5.6)
and Theorem 18.4.8, where we put 𝜈 = 𝑛 and 𝑁 + 𝑁 ♭ + 𝑛 in the place of 𝑁, we obtain

Theorem 18.5.6 Let 𝐴 (𝑥, 𝑦) and 𝐵 (𝑦, 𝑧) be the distribution kernels (18.4.1) in
Ω1 ×Ω2 and (18.5.1) in Ω2 ×Ω3 respectively, with the phase-functions and amplitudes
specified in this subsection. Let Φ (resp., Φ♭ ) be the phase-function given by (18.5.5)
[resp., (18.5.12)]. If diam Ω 𝑗 ( 𝑗 = 1, 2, 3) and diam Γ ′ ∩ S𝑛−1 are sufficiently small
then, mod C ∞ (Ω1 × Ω3 ),
∫
1
e𝑖Φ ( 𝑥,𝑧, 𝜏˜ ) 𝑐♭ (𝑥, 𝑧, 𝜏˜ ′) | 𝜏˜ ′ | 2 ( 𝑁 +𝑁 ) d𝜏˜ ′
′ ♭
( 𝐴 ◦ 𝐵) (𝑥, 𝑧) 𝐶 𝑁 +𝑁 ♭ (Φ) (2𝜋) −𝑛
♭

Γ′
(18.5.15)
where 𝜋
e𝑖 4 sign 𝐻Φ
𝐶 𝑁 +𝑁 ♭ (Φ) = √︃ . (18.5.16)
(2𝜋) 𝑁 +𝑁 |det 𝐻Φ |
♭
734 18 Fourier Integral Operators

We recall that 𝐻Φ = Φ 𝜏˜ ′′ , 𝜏˜ ′′ (𝑥 ◦ , 𝑧◦ , 𝜏˜ ◦ ). The following result, proved above, is

worth stating.

Proposition
18.5.7 The order of the amplitude in (18.5.15) does not exceed 𝑚 + 𝑚 1 +
1
2 𝑁 + 𝑁 ♭ − 𝑛.

Remark 18.5.8 If 𝑨 and 𝑩 are pseudodifferential operators with the phase-functions

of the identity, (𝑥 − 𝑦) · 𝜃 and (𝑦 − 𝑧) ·𝜔 (thus 𝑁 = 𝑁 ♭ = 𝑛), we see that the amplitude
of their composite, using the phase-function (𝑥 − 𝑧)· 𝜏˜ ′, has order 𝑚+𝑚 1 , as expected.

What can be said if we wish to remove the requirement that 𝑩 is properly sup-
ported? By Proposition 18.3.2 there is a smoothing operator 𝑹 in Ω2 × Ω3 such that
𝑩 − 𝑹 is properly supported (and associated with the same phase-function as 𝑩). We
can then form the composite 𝑨 (𝑩 − 𝑹). Let 𝑹 1 in Ω2 × Ω3 also be such that 𝑩 − 𝑹 1
is properly supported. It follows that 𝑹 1 − 𝑹 is smoothing and properly supported;
therefore
𝑨 (𝑩 − 𝑹) − 𝑨 (𝑩 − 𝑹 1 ) = 𝑨 ( 𝑹 1 − 𝑹)
is well defined and smoothing. We can state:

Proposition 18.5.9 If [ 𝑨] (resp., [𝑩]) denotes the equivalence class of 𝑨 (resp., 𝑩)

mod Ψ−∞ (Ω1 × Ω2 ) [resp., mod Ψ−∞ (Ω2 × Ω3 ) ] then the composite [ 𝑨] [𝑩] is an
equivalence class of FIOs mod Ψ−∞ (Ω1 × Ω3 ).

18.5.2 Composites with opposite phase-functions I

A special case of composition of FIOs is of great importance: when Ω1 = Ω3 ,

𝑁 = 𝑁 ♭ , Γ = Γ♭ , and
𝜓 (𝑦, 𝑧, 𝜔) = −𝜑(𝑧, 𝑦, 𝜔). (18.5.17)
This subsection is devoted to analyzing the integral (18.5.15) under the hypothesis
(18.5.17). Thus, let the distribution kernels 𝐴, 𝐵, be given respectively by (18.1.13)
and (18.5.1) with 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ), 𝑏 ∈ 𝑆c𝑚1 (Ω2 × Ω1 , Γ), 𝑚, 𝑚 1 ∈ R, 𝑏 properly
supported; then
∫
( 𝐴 ◦ 𝐵) (𝑥, 𝑧) = 𝐴 (𝑥, 𝑦) 𝐵 (𝑦, 𝑧) d𝑦 (18.5.18)
Ω2
∫ ∫ ∫
= (2𝜋) −2𝑛 e𝑖 𝜑 ( 𝑥,𝑦, 𝜃)−𝑖 𝜑 (𝑧,𝑦, 𝜔) 𝑎 (𝑥, 𝑦, 𝜃) 𝑏 (𝑦, 𝑧, 𝜔) d𝑦d𝜃d𝜔.
Γ Γ Ω2

We shall always assume that 𝜑 is a strongly nondegenerate phase-function in

Ω1 × Ω2 × Γ (Definition 18.2.10); thus the associated Lagrangian manifold Λ 𝜑 is a
local symplectic graph. We shall also assume that

𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0. (18.5.19)
18.5 Composition and Continuity of Fourier Integral Operators 735

By Proposition 18.4.10 the strong nondegeneracy of 𝜑 and (18.5.19) demand 𝑁 ≤ 𝑛;

matters will be considerably simplified if we add the hypothesis that 𝑁 = 𝑛; we
shall do so. We select “central” points, 𝑥 ◦ ∈ Ω1 and 𝑦 ◦ ∈ Ω2 , and a central ray in
Γ generated by 𝜃 ◦ such that 𝜑 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0. Definition 18.2.10 combined with
(18.5.19) implies that both matrices 𝜑 𝑥, 𝜃 and 𝜑 𝑦, 𝜃 are nonsingular at (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ).
This allows us to apply the Implicit Function Theorem, with the tacit understanding
that, each time we do, this may require a contraction of the domains Ω1 and Ω2 about
𝑥 ◦ and 𝑦 ◦ respectively, and of Γ about the ray generated by 𝜃 ◦ .
We can apply Proposition 18.5.1: Condition (i) is obviously satisfied if 𝑁 ♭ =
𝑛 and (18.5.17) holds; Condition (ii) is also satisfied since (18.5.3) here reads
𝜑 𝑦 (𝑧, 𝑦, 𝜔) = 𝜑 𝑦 (𝑥, 𝑦, 𝜃), which is a tautology if we take 𝑧 = 𝑥, 𝜔 = 𝜃. As in
√︃
the proof of Proposition 18.5.1 we define 𝜏 (1) = 𝜆𝑦, 𝑦 ∈ Ω2 (𝜆 = |𝜃| 2 + |𝜔| 2 ),
𝜏 (2) = 𝜃 ∈ Γ, 𝜏 (3) = 𝜔 ∈ Γ, as fiber variables. At this stage it is convenient to assume
that 𝑦 ◦ is the origin in R𝑛 and, therefore, that 𝜏 varies in e Γ = R𝑛 × Γ × Γ ⊂ R3𝑛 \ {0};
the central point becomes (𝑥 ◦ , 𝑧◦ , 𝜏 ◦ ), 𝜏 ◦ = (0, 𝜃 ◦ , 𝜃 ◦ ). Then the phase-function in
Ω1 × Ω1 × e Γ [cf. (18.5.5)]

Φ (𝑥, 𝑧, 𝜏) = 𝜑 (𝑥, 𝑦, 𝜃) − 𝜑 (𝑧, 𝑦, 𝜔) (18.5.20)

= 𝜑 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) − 𝜑(𝑧, 𝜆−1 𝜏 (1) , 𝜏 (3) )

is nondegenerate. It follows that the critical set

n o
ΣΦ = (𝑥, 𝑧, 𝜏) ∈ Ω1 × Ω1 × e Γ; Φ 𝜏 (𝑥, 𝑧, 𝜏) = 0 (18.5.21)

is defined by the equations

𝜑 𝑦 (𝑥, 𝑦, 𝜃) = 𝜑 𝑦 (𝑧, 𝑦, 𝜔) , (18.5.22)

𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0, (18.5.23)
𝜑 𝜔 (𝑧, 𝑦, 𝜔) = 0. (18.5.24)

Lemma 18.5.10 After suitable contractions of Ω1 and Γ ∩ S𝑛−1 there is a unique

solution 𝑦 (𝑥, 𝜃) ∈ C ∞ (Ω1 × Γ) of (18.5.23) such that 𝑦 (𝑥 ◦ , 𝜃 ◦ ) = 0. Then 𝑧 = 𝑥,
𝜔 = 𝜃, are the unique C ∞ solutions in Ω1 × Γ of the system of equations

𝜑 𝑦 (𝑥, 𝑦, 𝜃) 𝑦=𝑦 ( 𝑥, 𝜃)
= 𝜑 𝑦 (𝑧, 𝑦, 𝜔) 𝑦=𝑦 ( 𝑥, 𝜃)
,
𝜑 𝜔 (𝑧, 𝑦 (𝑥, 𝜃) , 𝜔) = 0.

that are equal to 𝑥 ◦ , 𝜃 ◦ respectively, when 𝑥 = 𝑥 ◦ , 𝜃 = 𝜃 ◦ .

Proof Since det 𝜑 𝑦, 𝜃 and det 𝜑 𝑦, 𝜃 do not vanish in Ω1 × Ω1 × Γ × Γ the claims ensue
by the Implicit Function Theorem. □
736 18 Fourier Integral Operators

Thus

ΣΦ = {(𝑥, 𝑦, 𝑧, 𝜃, 𝜔) ∈ Ω1 × Ω2 × Ω1 × Γ × Γ; 𝑦 = 𝑦 (𝑥, 𝜃) , 𝑧 = 𝑥, 𝜔 = 𝜃} .
(18.5.25)
The coordinate projection (𝑥, 𝑦, 𝑧, 𝜃, 𝜔) ↦→ (𝑥, 𝑦, 𝜃) induces a diffeomorphism of
ΣΦ onto Σ 𝜑 . The Lagrangian submanifold ΛΦ = ( 𝑝 Φ , 𝑞 Φ ) (ΣΦ ) is the set of points
((𝑥, 𝜉) , (𝑧, 𝜁)) ∈ (𝑇 ∗ Ω1 \0) 2 such that

𝜉 = Φ 𝑥 (𝑥, 𝑦 (𝑥, 𝜃) , 𝑥, 𝜃, 𝜃) ,
𝜁 = −Φ𝑧 (𝑥, 𝑦 (𝑥, 𝜃) , 𝑥, 𝜃, 𝜃) .

By (18.5.20) the right-hand sides in these equations are both equal to 𝜑 𝑥 (𝑥, 𝑦 (𝑥, 𝜃) , 𝜃).
It follows that ΛΦ is the graph of the identity of 𝑇 ∗ Ω1 \0, i.e., ΛΦ is the diagonal of
(𝑇 ∗ Ω1 \0) 2 .
Proposition 18.5.11 The phase-function Φ is strongly nondegenerate,
Proof Since Φ is nondegenerate Proposition 18.2.13 implies that Φ is strongly
nondegenerate. □
Remark 18.5.12 In dealing with Φ (𝑥, 𝑧, 𝜏) we have 𝑁 = 3𝑛. This does not contradict
Proposition 18.4.10 as Φ 𝜏, 𝜏 (𝑥 ◦ , 𝑧◦ , 𝜏 ◦ ) ≠ 0 (see below).
Corollary 18.5.13 The composite 𝑨𝑩 is a pseudodifferential operator in Ω1 .
Proof Combine Propositions 18.4.23 and 18.5.11. □
We continue to adapt the argument in the preceding subsection to take (18.5.17)
into account. We have, for 𝑗 = 1, ..., 𝑛,
𝜕Φ 𝜕𝜑 −1 (1) 𝜕𝜑 −1 (1)
𝜆 (𝑥, 𝑧, 𝜏) = 𝑥, 𝜆 𝜏 , 𝜃 − 𝑧, 𝜆 𝜏 , 𝜔 , (18.5.26)
𝜕𝜏 𝑗 𝜕𝑦 𝑗 𝜕𝑦 𝑗
𝑛
𝜕Φ 𝜕𝜑 −1 (1) ∑︁ 𝜕Φ
(𝑥, 𝑧, 𝜏) = 𝑥, 𝜆 𝜏 , 𝜃 − 𝜆−1 𝜏𝑛+ 𝑗 𝜏𝑘 (𝑥, 𝑧, 𝜏) ,
𝜕𝜏𝑛+ 𝑗 𝜕𝜃 𝑗 𝑘=1
𝜕𝜏𝑘
𝑛
𝜕Φ 𝜕𝜑 −1 (1) ∑︁ 𝜕Φ
(𝑥, 𝑧, 𝜏) = − 𝑧, 𝜆 𝜏 , 𝜔 + 𝜆−2 𝜏𝑛+𝑁 + 𝑗 𝜏𝑘 (𝑥, 𝑧, 𝜏) .
𝜕𝜏2𝑛+ 𝑗 𝜕𝜔 𝑗 𝑘=1
𝜕𝜏𝑘

We take advantage of (18.5.19) and of the fact that 𝜏𝑘◦ = 0 and 𝜕Φ

𝜕𝜏𝑘 Σ = 0 if 𝑘 ≤ 𝑛.
Φ
As a consequence we have, at (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ), for 1 ≤ 𝑗, 𝑘 ≤ 𝑛,

𝜕2Φ
(𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ) = 0,
𝜕𝜏 𝑗 𝜕𝜏𝑘
𝜕2Φ 𝜕2 𝜑
𝜆 (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ) = (𝑥 ◦ , 0, 𝜃 ◦ ) ,
𝜕𝜏 𝑗 𝜕𝜏𝑛+𝑘 𝜕𝑦 𝑗 𝜕𝜃 𝑘
𝜕2Φ 𝜕2 𝜑
𝜆 (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ) = − (𝑥 ◦ , 0, 𝜃 ◦ ) ;
𝜕𝜏 𝑗 𝜕𝜏2𝑛+𝑘 𝜕𝑦 𝑗 𝜕𝜃 𝑘
18.5 Composition and Continuity of Fourier Integral Operators 737

all other partial second derivatives with respect to 𝜏 vanish at (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ). In block

notation we see that
0 𝜑 𝑦, 𝜃 (𝑥 ◦ , 0, 𝜃 ◦ ) −𝜑 𝑦, 𝜃 (𝑥 ◦ , 0, 𝜃 ◦ )
◦ ◦ ◦ ◦ ◦
Φ 𝜏, 𝜏 (𝑥 , 𝑥 , 𝜏 ) = 𝜆 ­ 𝜑 𝑦, 𝜃 (𝑥 , 0, 𝜃 ) 0 0
© ª
®
−𝜑 (𝑥 ◦ , 0, 𝜃 ◦ ) 0 0
« 𝑦, 𝜃 ¬
whose rank is 2𝑛 since 𝜑 𝑦, 𝜃 (𝑥 ◦ , 0, 𝜃 ◦ ) is nonsingular. We carry out an orthogonal
change of variables from 𝜏 to 𝜏˜ defined by the equations
1 1
𝜏 𝑗 = − √ 𝜏˜𝑛+ 𝑗 + √ 𝜏˜2𝑛+ 𝑗 , (18.5.27)
2 2
1 1 1
𝜏𝑛+ 𝑗 = √ 𝜏˜ 𝑗 − 𝜏˜𝑛+ 𝑗 − 𝜏˜2𝑛+ 𝑗 ,
2 2 2
1 1 1
𝜏2𝑛+ 𝑗 = √ 𝜏˜ 𝑗 + 𝜏˜𝑛+ 𝑗 + 𝜏˜2𝑛+ 𝑗 .
2 2 2

We have
𝜕 1 𝜕 1 𝜕
=√ +√ ,
𝜕 𝜏˜ 𝑗 2 𝜕𝜏𝑛+ 𝑗 2 𝜕𝜏2𝑛+ 𝑗
𝜕 1 𝜕 1 𝜕 1 𝜕
= −√ − + ,
𝜕 𝜏˜𝑛+ 𝑗 2 𝜕𝜏 𝑗 2 𝜕𝜏𝑛+ 𝑗 2 𝜕𝜏2𝑛+ 𝑗
𝜕 1 𝜕 1 𝜕 1 𝜕
=√ − + ,
𝜕 𝜏˜2𝑛+ 𝑗 2 𝜕𝜏 𝑗 2 𝜕𝜏𝑛+ 𝑗 2 𝜕𝜏2𝑛+ 𝑗
where 𝑗 = 1, ..., 𝑛. If we let these vector fields act on (18.5.26) and evaluate the result
at (𝑥 ◦ , 𝑦 ◦ , 𝜏 ◦ ) we get, for 𝑘 = 1, ..., 𝑛,

𝜕2Φ 𝜕2Φ 𝜕2Φ

= = = 0,
𝜕 𝜏˜ 𝑗 𝜕 𝜏˜𝑘 𝜕 𝜏˜𝑛+ 𝑗 𝜕 𝜏˜𝑘 𝜕 𝜏˜2𝑛+ 𝑗 𝜕 𝜏˜𝑘
𝜕2Φ
2
𝜕2 𝜑

1 𝜕 𝜑
=√ + (𝑥 ◦ , 0, 𝜃 ◦ ) ,
𝜕 𝜏˜𝑛+ 𝑗 𝜕 𝜏˜𝑛+𝑘 2 𝜕𝑦 𝑗 𝜕𝜃 𝑘 𝜕𝑦 𝑘 𝜕𝜃 𝑗
𝜕2Φ
2
𝜕2 𝜑

1 𝜕 𝜑
= −√ − (𝑥 ◦ , 0, 𝜃 ◦ ) ,
𝜕 𝜏˜2𝑛+ 𝑗 𝜕 𝜏˜𝑛+𝑘 2 𝜕𝑦 𝑗 𝜕𝜃 𝑘 𝜕𝑦 𝑘 𝜕𝜃 𝑗
𝜕2Φ
2
𝜕2 𝜑

1 𝜕 𝜑
= −√ + (𝑥 ◦ , 0, 𝜃 ◦ ) .
𝜕 𝜏˜2𝑛+ 𝑗 𝜕 𝜏˜2𝑛+𝑘 2 𝜕𝑦 𝑗 𝜕𝜃 𝑘 𝜕𝑦 𝑘 𝜕𝜃 𝑗
In block notation we see that the Hessian matrix of Φ with respect to 𝜏˜ at (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ )
is
0 0 0
1 © ⊤ ⊤ ª
Hess Φ = √ ­ 0 𝜕𝑦 𝜕𝜃 𝜑 + 𝜕𝑦 𝜕𝜃 𝜑 𝜕𝑦 𝜕 𝜃 𝜑 − 𝜕𝑦 𝜕 𝜃 𝜑 ® . (18.5.28)
2 0 −𝜕 𝜕 𝜑 + 𝜕 𝜕 𝜑 ⊤ −𝜕 𝜕 𝜑 − 𝜕 𝜕 𝜑 ⊤

« 𝑦 𝜃 𝑦 𝜃 𝑦 𝜃 𝑦 𝜃 ¬
738 18 Fourier Integral Operators

If then we set 𝜏˜ ′ = 𝜏˜ (1) , 𝜏˜ ′′ = 𝜏˜ (2) , 𝜏˜ (3) , we see that Φ 𝜏˜ ′ , 𝜏˜ ′ (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ) = 0,
Φ 𝜏˜ ′ , 𝜏˜ ′′ (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ) = 0 and

𝐻Φ = Φ 𝜏˜ ′′ , 𝜏˜ ′′ (𝑥 ◦ , 𝑥 ◦ , 𝜏 ◦ ) (18.5.29)
⊤ ⊤
1 𝜕𝑦 𝜕 𝜃 𝜑 + 𝜕𝑦 𝜕 𝜃 𝜑 𝜕 𝜕 𝜑 − 𝜕𝑦 𝜕 𝜃 𝜑
= ⊤ 𝑦 𝜃 ⊤ (𝑥 ◦ , 0, 𝜃 ◦ ) .
2 −𝜕𝑦 𝜕𝜃 𝜑 + 𝜕𝑦 𝜕𝜃 𝜑 −𝜕𝑦 𝜕𝜃 𝜑 − 𝜕𝑦 𝜕𝜃 𝜑

We now use the property that det 𝜕𝑦 𝜕𝜃 𝜑 (𝑥 ◦ , 0, 𝜃 ◦ ) ≠ 0.

Proposition 18.5.14 We have det 𝐻Φ ≠ 0. If 𝜒 is an eigenvalue of 𝐻Φ so is −𝜒 and

1 2
4 𝜒 is an eigenvalue of both matrices
⊤ ◦
(𝑥 , 0, 𝜃 ◦ ) ,

𝜕𝑦 𝜕 𝜃 𝜑 𝜕𝑦 𝜕 𝜃 𝜑
⊤
𝜕𝑦 𝜕 𝜃 𝜑 𝜕𝑦 𝜕𝜃 𝜑 (𝑥 ◦ , 0, 𝜃 ◦ ) .

This follows directly from the next two elementary lemmas.

Lemma 18.5.15 Let the real 𝑛×𝑛 matrices 𝑆 and

𝑇 be symmetric and skew-symmetric
𝑆 𝑇
respectively. If 𝜒 is an eigenvalue of so is −𝜒.
−𝑇 −𝑆

𝑆 𝑇
Proof The matrix is symmetric, hence semisimple. If 𝜒 ≠ 0 and u, v ∈
−𝑇 −𝑆
S𝑛−1 are such that
𝑆 𝑇 u u
=𝜒
−𝑇 −𝑆 v v
then
𝑆 𝑇 v v
= −𝜒 . □
−𝑇 −𝑆 u u

Lemma 18.5.16 If 𝑀 is a nonsingular real 𝑛 × 𝑛 matrix then

𝑀 + 𝑀⊤ 𝑀 − 𝑀⊤

(18.5.30)
−𝑀 + 𝑀 ⊤ − (𝑀 + 𝑀 ⊤ )

is nonsingular. If 𝜒 is an eigenvalue of (18.5.30) then 41 𝜒2 is an eigenvalue of 𝑀 𝑀 ⊤

and of 𝑀 ⊤ 𝑀.

Proof Let u, v ∈ R𝑛 be such that

𝑀 + 𝑀 ⊤ u + 𝑀 − 𝑀 ⊤ v = 𝜒u,

− 𝑀 − 𝑀 ⊤ u − 𝑀 + 𝑀 ⊤ v = 𝜒v.

By adding and subtracting these two equations we get

18.5 Composition and Continuity of Fourier Integral Operators 739

1
𝑀 ⊤ (u − v) = 𝜒 (u + v) ,
2
1
𝑀 (u + v) = 𝜒 (u − v) .
2
If 𝜒 = 0 then u = v = 0 since det 𝑀 ≠ 0, proving the first claim. Now suppose
𝜒 ≠ 0 and |u| = |v| = 1. We see that
1 1
𝑀 𝑀 ⊤ (u − v) = 𝜒𝑀 (u + v) = 𝜒2 (u − v) ,
2 4
1 1
𝑀 ⊤ 𝑀 (u + v) = 𝜒𝑀 ⊤ (u − v) = 𝜒2 (u + v) ,
2 4
proving the second claim. □

Corollary 18.5.17 We have Tr 𝐻Φ = 0, the signature of 𝐻Φ is equal to zero and

2
det 𝐻Φ = 22𝑛 det 𝜑 𝑦, 𝜃 .

We are now in a position to fully exploit the results of the preceding

subsection.

′ ′′ ′ (1) ′′
We continue to use the notation 𝜏˜ = ( 𝜏˜ , 𝜏˜ ), 𝜏˜ = 𝜏˜ , 𝜏˜ = 𝜏˜ , 𝜏˜ (2) (3) ; we can
′
assume that 𝜏˜ varies in Γ. We write Φ (𝑥, 𝑧, 𝜏)˜ = Φ (𝑥, 𝑧, 𝜏) with 𝜏, 𝜏˜ related by
e
𝜕Φ 𝜕Φ
(18.5.27); we have d 𝜕 𝜏˜1 ∧ · · · ∧ 𝜕 𝜏˜3𝑛 ≠ 0 on ΣΦ . The pair of equations 𝑦 = 0,
e e
√
𝜔 = 𝜃, is equivalent to the pair 𝜏˜ ′ = 2𝜃, 𝜏˜ ′′ = 0, by (18.5.27). Combining (18.5.25)
and (18.5.27) yields [cf. (18.5.21)]
n
′
√ ′′
o
ΣΦ
e = (𝑥, 𝑧, ˜
𝜏) ∈ Ω1 × Ω1 × Γ;
e 𝑧 = 𝑥, ˜
𝜏 = 2𝜃, ˜
𝜏 = |𝜃| (−𝑦 (𝑥, 𝜃) , 𝑦 (𝑥, 𝜃))

with 𝑦 (𝑥, 𝜃) as in Lemma√18.5.10. The central point (𝑥 ◦ , 0, 𝑥 ◦ , 𝜃 ◦ , 𝜃 ◦ ) is now defined

by 𝑥 = 𝑧 = 𝑥 ◦ , 𝜏˜ ′ = 𝜏˜ ′◦ = 2𝜃 ◦ , 𝜏˜ ′′ = 0. We know that ΣΦ ∞
e (= ΣΦ ) is a C submanifold
of Ω1 × Ω1 × e Γ, dim ΣΦ e = 2𝑛.
Since det 𝐻Φ ≠ 0 (Proposition 18.5.14), if diam Ω1 and diam Γ ∩ S𝑛−1 are
sufficiently small there is a unique C ∞ function 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′) in Ω1 × Ω1 × Γ valued
in R2𝑛 , homogeneous of degree 1 and such that
e (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) = 0, 𝜏˜ ′′ (𝑥 ◦ , 𝑥 ◦ , 𝜏˜ ′◦ ) = 0.
𝜕𝜏˜ ′′ Φ (18.5.31)

We define the new phase-function in Ω1 × Ω1 × Γ,

Φ♭ (𝑥, 𝑧, 𝜏˜ ′) = Φ
e (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) . (18.5.32)

Proposition 18.5.18 The following properties hold:

(1) the map ΣΦ♭ ∋ (𝑥, 𝑧, 𝜏˜ ′) ↦→ (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) ∈ ΣΦ
e is a diffeomorphism;
(2) the Lagrangian submanifold ΛΦ♭ is the graph of the identity of 𝑇 ∗ Ω1 \0;
(3) the phase-function Φ♭ is strongly nondegenerate;
(4) (𝑥, 𝑧, 𝜏 ′) ∈ ΣΦ♭ ⇐⇒ 𝑥 = 𝑧 and Φ♭ (𝑥, 𝑥, 𝜏 ′) = 0 for all 𝜏 ′ ∈ Γ.
740 18 Fourier Integral Operators

Proof It follows directly from (18.5.31)–(18.5.32) that, for all 𝑗 = 1, ..., 𝑛,

𝜕Φ♭ 𝜕Φ e
(𝑥, 𝑧, 𝜏˜ ′) = (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) , (18.5.33)
𝜕 𝜏˜ 𝑗 𝜕 𝜏˜ 𝑗
𝜕Φ♭ 𝜕Φ e
(𝑥, 𝑧, 𝜏˜ ′) = (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) ,
𝜕𝑥 𝑗 𝜕𝑥 𝑗
𝜕Φ♭ 𝜕Φ e
(𝑥, 𝑧, 𝜏˜ ′) = (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) .
𝜕𝑧 𝑗 𝜕𝑧 𝑗

This has the following consequences:

(i) 𝜕𝜏˜ ′ Φ♭ (𝑥, 𝑧, 𝜏˜ ′) = 0 and 𝜕𝜏˜ Φ e (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) = 0 are equivalent;
♭ ♭
(ii) d 𝜕Φ 𝜕Φ
𝜕 𝜏˜1 ∧ · · · ∧ d 𝜕 𝜏˜ 𝑛 ≠ 0 on ΣΦ♭ , i.e., Φ is nondegenerate;
♭
′
(iii) if (𝑥, 𝑧, 𝜏˜ ) ∈ ΣΦ♭ and (𝑥, 𝑧, 𝜏) ˜ ∈ ΣΦ e then the maps

𝑥, Φ♭𝑥 (𝑥, 𝑧, 𝜏˜ ′) ↦→ 𝑧, −Φ♭𝑧 (𝑥, 𝑧, 𝜏˜ ′) ,

e 𝑥 ↦→ 𝑧, −Φ
𝑥, Φ e 𝑧 (𝑥, 𝑧, 𝜏)˜

∗
map of 𝑇 Ω 1 \0;
are both the identity
(iv) ΣΦe ∋ (𝑥, 𝑧, 𝜏)
˜ ↦→ 𝑥, Φ ˜ and ΣΦ♭ ∋ (𝑥, 𝑧, 𝜏˜ ′) ↦→ 𝑥, Φ♭𝑥 (𝑥, 𝑧, 𝜏˜ ′) are
e 𝑥 (𝑥, 𝑧, 𝜏)
local diffeomorphisms.
Combining Proposition 18.2.13 with (ii) and (iii) yields Property (3); (i) and (iv)
imply Property (1); (ii) entails Property (2); (iii) and the fact that Φ♭ ≡ 0 on ΣΦ♭
entail Property (4). □
We make use of the amplitudes (18.5.13), (18.5.14) and of the true amplitude
Í∞ (2𝑖) ℓ
𝑐♭ (𝑥, 𝑧, 𝜏˜ ′) in Ω1 × Ω1 × Γ derived from the formal amplitude ℓ=0 ′
ℓ! 𝑐 ℓ (𝑥, 𝑧, 𝜏˜ );
−1
𝐻Φ is defined in (18.5.29); here 𝐶 (Φ) = |det 𝐻Φ | = 2𝑛 det 𝐻 𝜑 [cf. (18.5.16) and
Corollary 18.5.13]. We get the version of Theorem 18.5.6 in the present context:
Theorem 18.5.19 Assume that 𝜑 is a strongly nondegenerate phase-function in Ω1 ×
Ω2 ×Γ and that (18.5.17) holds. Let the distribution kernels 𝐴, 𝐵, be given respectively
by (18.1.13) and (18.5.1) with the amplitudes as in (18.5.18). Under these hypotheses

(2𝜋) −2𝑛
∫
♭ ( 𝑥,𝑧, 𝜏˜ ′ )
( 𝐴 ◦ 𝐵) (𝑥, 𝑧) e𝑖Φ 𝑐♭ (𝑥, 𝑧, 𝜏˜ ′) | 𝜏˜ ′ | 𝑛 d𝜏˜ ′ (18.5.34)
det 𝐻 𝜑 Γ

mod C ∞ (Ω1 × Ω1 ).
We derive from Proposition 18.5.7:
Corollary 18.5.20 Assume the same hypotheses as in Theorem 18.5.19. The ampli-
tude of 𝑨𝑩 relative to the phase-function Φ♭ is 𝑐♭ (𝑥, 𝑧, 𝜏˜ ′) | 𝜏˜ ′ | 𝑛 ; its order does not
exceed 𝑚 + 𝑚 1 .
18.5 Composition and Continuity of Fourier Integral Operators 741

By Property (4) in Proposition 18.5.18 we have

Φ♭ (𝑥, 𝑧, 𝜏˜ ′) = (𝑥 − 𝑧) · 𝑓 (𝑥, 𝑧, 𝜏˜ ′)

with 𝑓 (𝑥, 𝑥, 𝜏˜ ′) = 𝜕𝑥 Φ♭ (𝑥, 𝑥, 𝜏˜ ′). After contracting Ω1 about 𝑥 ◦ and Γ ∩ S𝑛−1
about 𝜃 ◦ we can carry out the change of variables 𝜏˜ ′ ⇝ 𝜉 = 𝑓 (𝑥, 𝑧, 𝜏˜ ′), transforming
(18.5.34) into

( 𝐴 ◦ 𝐵) (𝑥, 𝑧) (18.5.35)
−2𝑛 ∫ ′
(2𝜋) D𝜏˜
e𝑖 ( 𝑥−𝑧) · 𝜉 𝑐♭ (𝑥, 𝑧, 𝜏˜ ′ (𝑥, 𝑧, 𝜉)) | 𝜏˜ ′ (𝑥, 𝑧, 𝜉)| 𝑛 (𝑥, 𝑧, 𝜉) d𝜉
det 𝐻 𝜑 ℭ D𝜉

where the map 𝜉 ↦→ 𝜏˜ ′ (𝑥, 𝑧, 𝜉) is the inverse of 𝜏˜ ′ ↦→ 𝑓 (𝑥, 𝑧, 𝜏˜ ′) and ℭ is a cone in

R𝑛 \ {0} containing 𝜉 ◦ = 𝜑 𝑥 (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ).

18.5.3 Continuity of FIOs. Elliptic FIOs

We continue to deal with the operator 𝑨 given by (18.3.1). Here we assume that the
phase-function 𝜑 is strongly nondegenerate, and therefore Λ 𝜑 is a local symplectic
graph; and, for the sake of simplicity, that (18.5.19) holds. We begin by assuming that
the amplitude 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) is properly supported. This allows us to apply
the results of the preceding subsection to 𝑩 = 𝑨∗ , the adjoint of 𝑨 [cf. (18.3.3)].
The hypothesis (18.5.17) is obviously satisfied. The following statement is now easy
to prove.

Theorem 18.5.21 The operator 𝑨 defined by (18.3.1) with 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ)

properly supported induces continuous linear maps
𝑠 𝑠−𝑚
𝐻loc (Ω2 ) −→ 𝐻loc (Ω1 ) ,
𝐻c (Ω2 ) −→ 𝐻c𝑠−𝑚 (Ω1 ) ,
𝑠

whatever 𝑠 ∈ R.

Proof From Corollaries 18.5.13, 18.5.20, we derive that 𝑨∗ 𝑨 is a𝑠 pseudodifferential

operator of order ≤ 2𝑚 in Ω1 . We introduce the operator 𝑠
(1 − Δ) 2 (Δ: the Laplacian
in R𝑛 ). Since 𝑨 and 𝑨∗ are properly supported
𝑠
(1 − Δ) 2 𝑨 is a well-defined operator

D ′ (Ω2 ) −→ D ′ (Ω1 ) and 𝑨∗ (1 − Δ) 2 is a well-defined operator D ′ (Ω1 ) −→

D ′ (Ω2 ). We derive that 𝑨∗ (1 − Δ) 𝑠 𝑨 is a pseudodifferential operator in Ω1 of order
≤ 2 (𝑚 + 𝑠). It follows from Theorem 16.1.19 that to every open subset Ω1′ ⊂⊂ Ω1
there is a 𝐶 > 0 such that

∥ 𝑨𝑢∥ 2𝑠 = ( 𝑨∗ (1 − Δ) 𝑠 𝑨𝑢, 𝑢) ≤ 𝐶 ∥𝑢∥ 2𝑠−𝑚

𝐿2
742 18 Fourier Integral Operators

whatever 𝑢 ∈ Cc∞ Ω2′ . This shows that 𝑨 maps 𝐻c𝑠 (Ω2 ) continuously into

𝐻 𝑠−𝑚 (Ω1 ), hence into 𝐻c𝑠−𝑚 (Ω1 ) since 𝑨 is properly supported. □
If we want to remove the requirement that 𝑎 be properly supported we may
replace 𝑨 by a properly supported FIO 𝑨◦ in Ω1 × Ω2 such that 𝑻 = 𝑨 − 𝑨◦ is
𝑠′ (Ω )
smoothing. Since 𝑻 defines a continuous linear operator 𝐻c𝑠 (Ω2 ) −→ 𝐻loc 1
whatever (𝑠, 𝑠 ′) ∈ R2 we can state
Corollary 18.5.22 The FIO 𝑨 defined by (18.3.1) with 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) (not
necessarily properly supported) induces a continuous linear map 𝐻c𝑠 (Ω2 ) −→
𝑠−𝑚 (Ω ),
𝐻loc 1 whatever 𝑠 ∈ R.

Definition 18.5.23 The Fourier integral operator 𝑨 : E ′ (Ω2 ) −→ D ′ (Ω1 ) defined

by (18.3.1) is said to be elliptic of order 𝑚 if there is a 𝑻 ∈ Ψ−∞ (Ω1 × Ω2 ) such that
𝑨◦ = 𝑨 − 𝑻 is properly supported and the pseudodifferential operator in Ω1 , 𝑨◦ 𝑨∗◦ ,
is elliptic of order 2𝑚 (Definition 16.2.14).
We introduce the natural generalization of Definition 16.2.14:
Definition 18.5.24 The amplitude 𝑎 ∈ 𝑆 𝑚 (Ω1 × Ω2 , Γ) is said to be elliptic of order
𝑚 at a point (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ∈ Ω1 × Ω2 × Γ if there are neighborhoods in R𝑛 , 𝑈1 ⊂ Ω1
of 𝑥 ◦ , 𝑈2 ⊂ Ω2 of 𝑦 ◦ , an open cone Γ♭ , 𝜃 ◦ ∈ Γ♭ ⊂ Γ, and positive constants 𝑅 and 𝑐
such that

∀ (𝑥, 𝑦, 𝜃) ∈ 𝑈1 × 𝑈2 × Γ♭ , |𝜃| > 𝑅 =⇒ |𝑎 (𝑥, 𝑦, 𝜃)| ≥ 𝑐 |𝜃| 𝑚 .

We shall say that 𝑎 is elliptic in a conic open subset U of Ω1 × Ω2 × Γ if 𝑎 is elliptic

at every point of U.
Remark 18.5.25 If 𝑎 is elliptic in Ω1 × Ω2 × Γ it cannot be properly supported (see
Definition 18.1.15).
Proposition 18.5.26 Let 𝐴 (𝑥, 𝑦) be defined by (18.1.13) with the phase-function 𝜑
strongly nondegenerate and 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) properly supported. For the FIO
𝑨 given by (18.3.1) to be elliptic of order 𝑚 it is necessary and sufficient that the
amplitude 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) be elliptic of order 𝑚 in a conic open subset U of
Ω1 × Ω2 × Γ containing Σ 𝜑 .
Proof Follows directly from Definition 18.5.23, (18.5.36). □
Properly supported elliptic FIOs of order zero define isomorphisms of 𝐿 c2 spaces.
To simplify matters suppose that Ω1 = Ω2 = Ω and Γ = R𝑛 \ {0}. With 𝐴 given
by (18.1.13), amplitude 𝑎 ∈ 𝑆 𝑚 (Ω × Ω) properly supported and phase-function 𝜑
strongly nondegenerate, we will investigate the conditions under which the FIO 𝑨 is
unitary, in the following sense:
Definition 18.5.27 We shall say that the properly supported FIO 𝑨 in Ω × Ω is
unitary if both 𝑨 and 𝑨∗ are continuous linear operators of 𝐿 c2 (Ω) into itself and

𝑨 𝑨∗ = 𝑨∗ 𝑨 = Identity map of 𝐿 c2 (Ω) .

18.5 Composition and Continuity of Fourier Integral Operators 743

Proposition 18.5.28 If 𝑨 is unitary then 𝑨 is elliptic of order zero.

Proof Follows directly from Definitions 18.5.23, 18.5.27. □

∗
We take a closer look at the representation (18.5.34) when 𝑩 = 𝑨 with 𝑎 ∈
𝑆c𝑚 (Ω1 × Ω2 , Γ) properly supported. Formula (18.5.13) reads, here,

2 −𝑛/2

2
′ −1 (1) (2) (2) (3)
𝑐 (𝑥, 𝑧, 𝜏˜ ) = 𝑎 𝑥, 𝜆 𝜏 ,𝜏 𝑎 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3) 𝜏 + 𝜏
(18.5.36)
where 𝜏 is related to 𝜏˜ = ( 𝜏˜ ′, 𝜏˜ ′′) by (18.5.27) and 𝜏˜ ′′ = 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′) is determined
by (18.5.31); by Corollary 18.5.20 𝑐 (𝑥, 𝑧, 𝜏˜ ′) is an amplitude of order ≤ 2𝑚 − 𝑛.
For 𝑨 to be unitary, in view of (18.5.35) it is necessary and sufficient that 𝑚 = 0
and
(2𝜋) −𝑛 D𝜏˜ ′
𝑐 (𝑥, 𝑧, 𝜏˜ ′ (𝑥, 𝑧, 𝜉)) | 𝜏˜ ′ (𝑥, 𝑧, 𝜉)| 𝑛 (𝑥, 𝑧, 𝜉) = 1. (18.5.37)
det 𝐻 𝜑 D𝜉

This demands that 𝑐 (𝑥, 𝑧, 𝜏˜ ′) be homogeneous of degree −𝑛; by (18.5.36) 𝑎 (𝑥, 𝑦, 𝜃)

must be homogeneous of degree zero.

18.5.4 Composites with opposite phase-functions II. Amplitudes that

have leading terms

For our present purpose it is convenient to deal with amplitudes and symbols that
have leading terms (or a principal symbol). We state this requirement more precisely:
′
Definition 18.5.29 Let 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) be such that 𝑎 ∉ 𝑆c𝑚 (Ω1 × Ω2 , Γ) if
𝑚 ′ < 𝑚. We shall say that the amplitude 𝑎 0 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) is a leading term of
𝑎 if 𝑎 0 is homogeneous of degree 𝑚 for large |𝜃| and 𝑎 − 𝑎 0 ∈ 𝑆c𝑚−𝜅 (Ω1 × Ω2 , Γ)
for some 𝜅 > 0.

A leading term𝑎 0 of 𝑎 is not unique: an integration by parts in (18.3.1) shows that

𝜕𝜑
1 + 𝑖 𝜕𝜃1
(𝑥, 𝑦, 𝜃) 𝑎 0 (𝑥, 𝑦, 𝜃) is also a leading term of 𝑎; the number 𝜅 > 0 (if 𝜅 ≤ 1)
in Definition 18.5.29 is independent of 𝑎 0 . For a more meaningful interpretation of
the ongoing discussion we ought to transition from amplitudes to symbols, following
the procedure starting from Formula (16.2.3). For this we would need the concept
of symbol of an FIO, which we introduce in the next section. For the time being
we investigate leading terms for the composites 𝑨𝑩 under the hypothesis that the
amplitudes 𝑎 and 𝑏 have leading terms, 𝑎 0 and 𝑏 0 respectively.
We continue to reason under Hypothesis (18.5.17) (with 𝑁 = 𝑁1 = 𝑛). We seek
a leading term of the amplitude in (18.5.35). To simplify the notation we write

(2𝜋) −𝑛 ♭ D𝜏˜ ′
𝑘 (𝑥, 𝑧, 𝜉) = 𝑐 (𝑥, 𝑧, 𝜏˜ ′ (𝑥, 𝑧, 𝜉)) | 𝜏˜ ′ (𝑥, 𝑧, 𝜉)| 𝑛 (𝑥, 𝑧, 𝜉) ;
det 𝐻 𝜑 D𝜉
744 18 Fourier Integral Operators

then ∫
( 𝐴 ◦ 𝐵) (𝑥, 𝑧) (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑧) · 𝜉 𝑘 (𝑥, 𝑧, 𝜉) d𝜉.
ℭ

By Corollary 18.5.20 𝑘 ∈ 𝑆c𝑚+𝑚1 (Ω1 × Ω2 , ℭ); if we write 𝑘 (𝑥, 𝑧, 𝜉) = 𝑘 (𝑥, 𝑥, 𝜉) +

(𝑥 − 𝜉) · 𝑘 1 (𝑥, 𝑧, 𝜉) integration by parts yields
∫ ∫ ∫
𝑖 ( 𝑥−𝑧) · 𝜉 𝑖 ( 𝑥−𝑧) · 𝜉
e 𝑘 (𝑥, 𝑧, 𝜉) d𝜉 = e 𝑘 (𝑥, 𝑥, 𝜉) d𝜉− e𝑖 ( 𝑥−𝑧) · 𝜉 D 𝜉 𝑘 1 (𝑥, 𝑧, 𝜉) d𝜉.
ℭ ℭ ℭ

Since D 𝜉 𝑘 ∈ 𝑆c𝑚+𝑚1 −1 (Ω1 × Ω2 , ℭ) we see that 𝑘 (𝑥, 𝑥, 𝜉) is a leading term of

𝑘 (𝑥, 𝑧, 𝜉).
By (18.5.13) 𝑐 (𝑥, 𝑧, 𝜏˜ ′) has the leading term

𝑐 0 (𝑥, 𝑧, 𝜏˜ ′) = 𝑎 0 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) 𝑏 0 𝜆−1 𝜏 (1) , 𝑧, 𝜏 (3) 𝜆−𝑛 , (18.5.38)

where 𝜏 is defined in terms of 𝜏˜ by (18.5.27), 𝜏˜ (2) , 𝜏˜ (3) = 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′) [cf.
(18.5.31)] and √︃
2 2
𝜆 = 𝜆 (𝑥, 𝑧, 𝜏˜ ′) = 𝜏 (2) + 𝜏 (3) .
We conclude that a leading term of 𝑘 (𝑥, 𝑧, 𝜉) is

(2𝜋) −𝑛 D𝜏˜ ′
𝑘 0 (𝑥, 𝜉) = 𝑐 0 (𝑥, 𝑧, 𝜏˜ ′) (| 𝜏˜ ′ (𝑥, 𝑥, 𝜉)|) 𝑛 (𝑥, 𝑥, 𝜉) . (18.5.39)
det 𝐻 𝜑 D𝜉

Recall that 𝜉 ↦→ 𝜏˜ ′ (𝑥, 𝑥, 𝜉) is the inverse of 𝜏˜ ′ ↦→ 𝑓 (𝑥, 𝑥, 𝜏˜ ′) = Φ♭𝑥 (𝑥, 𝑥, 𝜏˜ ′).

Lemma 18.5.30 If 𝑧 = 𝑥 and 𝜏˜ ′′ = 𝜏˜ (𝑥, 𝑥, 𝜏˜ ′) the following properties hold:
(a) 𝜏 (2) = 𝜏 (3) = √𝜃; √
(b) 𝜏˜ ′ (𝑥, 𝑥, 𝜉) = 2𝜃, 𝜆 = 2 |𝜃|;
(c) 𝜆−1 𝜏 (1) = 𝑦 (𝑥, 𝜃), the function in Lemma 18.5.10;
(d) 𝑓 (𝑥, 𝑥, 𝜏˜ ′) = 𝜑 𝑥 (𝑥, 𝑦 (𝑥, 𝜃) , 𝜃).
Proof From (18.5.32) and (18.5.33) we derive
e 𝑥 (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′)) + 𝜕𝑥 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′) 𝜕𝜏˜ ′′ Φ
𝑓 (𝑥, 𝑧, 𝜏˜ ′) = Φ e (𝑥, 𝑧, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑧, 𝜏˜ ′))

and then, from (18.5.31) and (18.5.20),

𝑓 (𝑥, 𝑥, 𝜏˜ ′) = Φ
e 𝑥 (𝑥, 𝑥, 𝜏˜ ′, 𝜏˜ ′′ (𝑥, 𝑥, 𝜏˜ ′)) (18.5.40)

= 𝜑 𝑥 𝑥, 𝜆−1 𝜏 (1) , 𝜏 (2) .
𝑧=𝑥, 𝜏˜ ′′ = 𝜏˜ ′′ ( 𝑥, 𝑥, 𝜏˜ ′ )

𝜏˜ (2)√+ 𝜏˜ (3) 𝜏˜ (2)√− 𝜏˜ (3)

In the notation 𝛼 = ,𝛽= , (18.5.27) reads
2 2

1 1
𝜏 (1) = −𝛽, 𝜏 (2) = √ ( 𝜏˜ ′ − 𝛼) , 𝜏 (3) = √ ( 𝜏˜ ′ + 𝛼) ,
2 2
18.5 Composition and Continuity of Fourier Integral Operators 745
√︃
and entails 𝜆 = | 𝜏˜ ′ | 2 + |𝛼| 2 . Taking the homogeneity of 𝜑 into account we see that
√
e (𝑥, 𝑥, 𝜏˜ ′, 𝜏˜ ′′) = 𝜑 (𝑥, −𝛽/𝜆, 𝜏˜ ′ − 𝛼) − 𝜑 (𝑥, −𝛽/𝜆, 𝜏˜ ′ + 𝛼)
2Φ
e (𝑥, 𝑥, 𝜏˜ ′, 𝜏˜ ′′) = 0 implies 𝜕𝛽 Φ
and that 𝜕𝜏˜ ′′ Φ e (𝑥, 𝑥, 𝜏˜ ′, 𝜏˜ ′′) = 0, whence

𝜑 𝑦 (𝑥, −𝛽/𝜆, 𝜏˜ ′ − 𝛼) − 𝜑 𝑦 (𝑥, −𝛽/𝜆, 𝜏˜ ′ + 𝛼) = 0. (18.5.41)

Since 𝜑 𝑦, 𝜃 is nonsingular in a neighborhood of the central point we derive 𝛼 = 0.

This implies √ √
𝜏˜ ′ = 2𝜏 (2) = 2𝜏 (3) , 𝜆 = | 𝜏˜ ′ | .
√ √
Since 𝜏 (2) = 𝜃, 𝜏 (3) = 𝜔, this proves (a); we also deduce that 𝜏˜ ′ = 2𝜃, 𝜆 = 2 |𝜃|
(and 𝜔 = 𝜃), which proves (b).
e (𝑥, 𝑥, 𝜏˜ ′, 𝜏˜ ′′) = 0 also implies 𝜕𝛼 Φ
We see that 𝜕𝜏˜ ′′ Φ e (𝑥, 𝑥, 𝜏˜ ′, 𝜏˜ ′′) = 0, whence,
since 𝛼 = 0,
𝜑 𝜃 (𝑥, −𝛽/|𝜃| , 𝜃) = 0,
entailing 𝛽 = − |𝜃| 𝑦 (𝑥, 𝜃). Putting all this into (18.5.40) proves (d). □
Lemma 18.5.30 and (18.5.38) lead to the following formula:

(2𝜋) −𝑛
𝑘 0 (𝑥, 𝜉) d𝜉 = 𝑎 0 (𝑥, 𝑦 (𝑥, 𝜃) , 𝜃) 𝑏 0 (𝑦 (𝑥, 𝜃) , 𝑥, 𝜃) d𝜃. (18.5.42)
det 𝐻 𝜑

We underline the fact that 𝑎 0 (𝑥, 𝑦 (𝑥, 𝜃) , 𝜃) 𝑏 0 (𝑦 (𝑥, 𝜃) , 𝑥, 𝜃) is the restriction

of the function 𝑎 0 (𝑥, 𝑦, 𝜃) 𝑏 0 (𝑦, 𝑥, 𝜃) ∈ C ∞ (Ω1 × Ω2 × Γ) to Σ 𝜑 . If we use
𝑥 𝑗 , 𝜉 𝑘 (1 ≤ 𝑗, 𝑘 ≤ 𝑛) as coordinates in ΣΦ ⊂ Ω1 × Ω2 × Ω1 × Γ, then
(𝑥, 𝜉) ↦→ (𝑥, 𝑦 (𝑥, 𝜃 (𝑥, 𝜉)) , 𝜃 (𝑥, 𝜉)) is the restriction to ΣΦ of the map (𝑥, 𝑦, 𝑧, 𝜃) ↦→
(𝑥, 𝑦, 𝜃) ∈ 𝑇 ∗ Ω1 while (𝑥, 𝜉) ↦→ (𝑦 (𝑥, 𝜃 (𝑥, 𝜉)) , 𝑥, 𝜃 (𝑥, 𝜉)) is the restriction to ΣΦ
of the map (𝑥, 𝑦, 𝑧, 𝜃) ↦→ (𝑦, 𝑧, 𝜃) ∈ 𝑇 ∗ Ω1 .

18.5.5 Egorov’s Theorem

Next we consider the case of 𝑩 = 𝑃 𝑨∗ where 𝑃 is a pseudodifferential operator in

Ω2 . Thus we assume that 𝐵 is the distribution kernel in Ω2 × Ω1 ,
∫
−𝑁
e−𝑖 𝜑 (𝑧,𝑦, 𝜔) 𝑎 (𝑧, 𝑦, 𝜔)d𝜔.

𝐵 (𝑦, 𝑧) = (2𝜋) 𝑃 𝑦, D 𝑦 (18.5.43)
Γ

Keeping in mind that 𝑨𝑩 = 𝑨𝑃 𝑨∗ is a pseudodifferential operator in Ω1 (Corollary

18.5.13) it can be regarded as the conjugate of 𝑃 by the FIO 𝑨.
We avail ourselves of the formulas (18.1.30) and (18.1.31). We find that a formal
amplitude of 𝑩 is the formal series
746 18 Fourier Integral Operators
∑︁ 1
D 𝑦′ D 𝜉 𝑃 (𝑦, −ℎ1 (𝑧, 𝑦, 𝑦 ′, 𝜔))
𝛽 𝛼+𝛽
𝜕𝑦𝛼 𝑎 (𝑧, 𝑦, 𝜔), (18.5.44)
𝛼!𝛽! 𝑦′ =𝑦
𝛼,𝛽 ∈Z+𝑛

where ∫ 1
′
ℎ1 (𝑧, 𝑦, 𝑦 , 𝜔) = 𝜑 𝑦 (𝑧, 𝑡𝑦 + (1 − 𝑡) 𝑦 ′, 𝜔) d𝑡
0
in a neighborhood of the “diagonal” 𝑦 = 𝑦 ′. Note that ℎ1 (𝑧, 𝑦, 𝑦, 𝜔) = 𝜑 𝑦 (𝑧, 𝑦, 𝜔).
Let us assume that 𝑎 0 is a leading term of 𝑎 and that 𝑃 is a classical pseudod-
ifferential operator of order 𝑑, with principal symbol 𝑃 𝑑 (𝑦, 𝜂) (Definitions 16.3.6,
16.3.7). Then
𝑃 𝑑 𝑦, −𝜑 𝑦 (𝑧, 𝑦, 𝜔) 𝑎 0 (𝑧, 𝑦, 𝜔)
is a leading term of (18.5.44); Formula (18.5.42) implies that

(2𝜋) −𝑛
𝑃 𝑑 𝑦 (𝑥, 𝜉) , −𝜑 𝑦 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) (18.5.45)
det 𝐻 𝜑
D𝜃
× |𝑎 0 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉))| 2 (𝑥, 𝜃)
D𝜉

is a leading term in the amplitude of 𝑨𝑩 = 𝑨𝑃 𝑨∗ restricted to ΣΦ . We refer the

reader to Diagram (18.2.11): 𝑃 𝑑 𝑦 (𝑥, 𝜉) , −𝜑 𝑦 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) is the pullback
to 𝑇 ∗ Ω1 of the symbol 𝑃 𝑑 (𝑦, 𝜂) under the symplectomorphism (𝑥, 𝜉) ↦→ (𝑦, 𝜂) whose
graph is the Lagrangian submanifold Λ 𝜑 .
If we apply this to a unitary FIO 𝑨 (Definition 18.5.23) we derive directly
from (18.5.37) and (18.5.45) a first version of the celebrated Egorov’s theorem (see
[Egorov, 1969]):
Theorem 18.5.31 Let the operator 𝑨 in Ω1 × Ω2 be defined by the distribution ker-
nel (18.1.13) with the phase-function 𝜑 strongly nondegenerate and the amplitude
𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) properly supported. Let 𝑃 𝑦, D 𝑦 be a classical pseudodif-
ferential operator in Ω2 of order 𝑑 ∈ R. If 𝑨 is unitary then 𝑨𝑃 𝑦, D 𝑦 𝑨∗ is a

pseudodifferential operator in Ω1 with principal symbol

𝑃 𝑑 𝑦 (𝑥, 𝜉) , −𝜑 𝑦 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) .

18.6 Globally Defined Fourier Integral Operators

18.6.1 From Lagrangian submanifolds to FIOs

So far we have regarded a Fourier distribution kernel as determined by the choice

of a phase-function and an amplitude. But in Section 18.4 we have seen that related
but different phase-functions 𝜑, with correspondingly modified amplitudes, can be
used in a representation (18.1.13) of one and the same distribution kernel 𝐴 [see
(18.4.23)]. One object, however, that never changes is the Lagrangian submanifold
18.6 Globally Defined Fourier Integral Operators 747

associated with the changing phase-functions. This is important if we need to handle

globally defined Fourier integral operators. To do this we introduce a pair of smooth
manifolds M1 and M2 such that dim M1 = dim M2 = 𝑛, and a local symplectic
graph (Definition 18.2.12) Λ ⊂ (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0); the FIOs will act from
D ′ (M1 ) to D ′ (M2 ) and be associated with Λ in the manner described below. In
many applications of FIOs, in particular those originating with geometric optics,
the phase-function might be defined only microlocally (as a solution of one or more
eikonal equations; see Section 13.4) while the Lagrangian submanifold Λ is defined
globally, meaning in the whole of (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0).
Therein lies the motivation for reversing our approach: from a local symplectic
graph Λ in (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) to (equivalence) classes of phase-functions 𝜑
whose associated Lagrangian submanifold (Definition 18.2.3) coincides with Λ in a
conic neighborhood U in (𝑇 ∗ M1 \0)×(𝑇 ∗ M2 \0) of one of its points, (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ),
to the equivalence classes mod smooth functions of the Fourier distribution kernels
in the base-projection that can be represented by an integral (18.1.13) in some
local coordinates systems 𝑥1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 [these coordinates allow us to identify
𝜋 ∗ (U) with the open subset Ω1 × Ω2 of R𝑛 × R𝑛 in which (18.1.13) makes sense].
As frequently evidenced in the preceding sections, microlocalization is achieved
by restricting the Fourier distribution kernel 𝐴 under study to a neighborhood 𝑈1 ×𝑈2
of a point (𝑥 ◦ , 𝑦 ◦ ) ∈ Ω1 × Ω2 and contracting the domain of 𝜃-integration, the open
cone Γ ⊂ R 𝑁 \ {0}, about a ray (𝜆𝜃 ◦ ) 𝜆>0 . By using cutoff functions it is always
possible to assume that the projection into 𝜃-space of the support of the amplitude is
contained in a cone Γ1 closed in R 𝑁 \ {0}. In phase-space (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0)
this has the effect of focusing the analysis to the image 𝑝 𝜑 , 𝑞 𝜑 (𝑈1 × 𝑈2 × Γ) ⊂ Λ
[cf. Diagram (18.1.19)]. Obviously each globally defined Fourier distribution can
be decomposed as a microlocally finite sum of “pieces” of the type (18.1.13). And
equally evident is the need for criteria that allow us to “patch” together pieces of said
type. We tackle this question in the remainder of this section.
We begin with the following observation:

Lemma 18.6.1 Let M be a C ∞ manifold, Λ be a Lagrangian submanifold of

𝑇 ∗ M\ {0} and ℘ ∈ Λ be arbitrary. There exist a local chart (𝑈, 𝑥1 , ..., 𝑥 𝑛 ) in
M, such that ℘ ∈ 𝑇 ∗ M|𝑈 , 𝑥 𝑗 (℘) = 0, 𝜉 𝑗 (℘) = 𝜉 ◦𝑗 , 𝑗 = 1, ..., 𝑛, with 𝜉1 , ..., 𝜉 𝑛 the
coordinates with respect to the basis d𝑥 1 , ..., d𝑥 𝑛 in the cotangent spaces to M at
points of 𝑈, a neighborhood Ξ of 𝜉 ◦ in R𝑛 \ {0} and a function 𝐻 ∈ C ∞ (Ξ; R) such
that if 𝜑 (𝑥, 𝜉) = 𝑥 · 𝜉 − 𝐻 (𝜉), then Λ ∩ (𝑈 × Ξ) is the nullset of 𝜕 𝜉 𝜑. When Λ is
conic we can take Ξ to be a convex open cone and 𝐻 homogeneous of degree 1.

We have identified 𝑈 × Ξ with an open subset of 𝑇 ∗ M\ {0}.

Proof We select local coordinates 𝑥 𝑗 in 𝑈 and the dual coordinates 𝜉 𝑘 such that
℘ = (0, 𝜉 ◦ ) and 𝜉 ◦ = (0, ..., 0, 1). We assume that 𝑈 is an open ball in R𝑛𝑥 centered at
the origin and 𝑇 ∗ M|𝑈 to 𝑈 × R𝑛𝜉 ; 𝑇0∗ M is identified with {0} × R𝑛𝜉 and Λ ∩ 𝑇 ∗ M|𝑈
with a Lagrangian submanifold of 𝑈 ×R𝑛𝜉 which we denote by Λ𝑈 . Let us regard both
𝑇(0, 𝜉 ◦ ) 𝑇0∗ M and 𝑇(0, 𝜉 ◦ ) Λ as affine subspaces of R𝑛𝑥 × R𝑛𝜉 (⊃ 𝑈 × R𝑛𝜉 ) intersecting

at (0, 𝜉 ◦ ); both are Lagrangian affine subspaces of the symplectic space R𝑛𝑥 × R𝑛𝜉 . We
748 18 Fourier Integral Operators

apply Proposition 13.1.15: there is a Lagrangian subspace 𝐸 of R𝑛𝑥 × R𝑛𝜉 containing

(0, 𝜉 ◦ ) and transversal to both 𝑇(0, 𝜉 ◦ ) 𝑇0∗ M and 𝑇(0, 𝜉 ◦ ) Λ. Such an affine

space is
defined in 𝑇(0, 𝜉 ◦ ) (𝑇 ∗ M) by an equation 𝜉 = 𝜉 ◦ + 𝐴𝑥 with 𝐴 = 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝑛 an
𝑛 × 𝑛 real matrix. The fact that 𝐸 is Lagrangian means that 𝐴 is symmetric.
Let 𝑓 (𝑥) = 𝑥 𝑛 + 21 𝐴𝑥 · 𝑥. After decreasing the radius of 𝑈 we carry out the change
of coordinates 𝑥 𝑛 ⇝ 𝑥˜ 𝑛 = 𝑓 (𝑥). Since 𝜉 · d𝑥 = 𝜉˜ · d𝑥˜ (where 𝑥˜ 𝑗 = 𝑥 𝑗 if 𝑗 < 𝑛) the
coordinates in the fibers of 𝑇 ∗ M|𝑈 with respect to the basis d𝑥˜1 , ..., d𝑥˜ 𝑛 are
𝑛
∑︁
𝜉˜ 𝑗 = 𝜉 𝑗 − 𝜉˜𝑛 𝑎 𝑗,𝑘 𝑥 𝑘 , 𝑗 = 1, ..., 𝑛 − 1,
𝑘=1
𝑛
! −1
∑︁
𝜉˜𝑛 = 1 + 𝑎 𝑛,𝑘 𝑥 𝑘 𝜉𝑛 .
𝑘=1

Note that 𝜉˜ = (0, ..., 0, 1) entails 𝜉𝑖 = 𝑛𝑗=1 𝑎 𝑖, 𝑗 𝑥 𝑗 (1 ≤ 𝑖 ≤ 𝑛 − 1), 𝜉 𝑛 = 1 +

Í
Í𝑛 ˜
𝑘=1 𝑎 𝑛,𝑘 𝑥 𝑘 : in the new coordinates 𝐸 is defined by the equation 𝜉 = (0, ..., 0, 1),
i.e., coincides with the section 𝑥 ↦→ ( 𝑥, ◦ ∗
˜ 𝜉 ) of 𝑇 M over 𝑈.
Let us omit the tildes. Since Λ is transversal to 𝐸 at (0, 𝜉 ◦ ) it can be defined by
equations 𝑥 𝑗 = ℎ 𝑗 (𝜉), 𝑗 = 1, ..., 𝑛, in 𝑈 × Ξ, with 𝑈 possibly contracted about 0, Ξ a
convex neighborhood of 𝜉 ◦ in R𝑛 \ {0} and Í𝑛ℎ 𝑗 ∈ C
∞ (Ξ). Since Λ is Lagrangian the

pullback of 𝑥 · d𝜉 to Λ𝑈 , i.e., the 1-form 𝑗=1 ℎ 𝑗 (𝜉) d𝜉 𝑗 , is closed and therefore, by

the Poincaré Lemma, there is an 𝐻 ∈ C ∞ (Γ) such that ℎ 𝑗 = 𝜕𝜕𝐻 𝜉 𝑗 , 𝑗 = 1, ..., 𝑛; thus

Λ ∩ U = (𝑥, 𝜉) ∈ 𝑈 × Γ; 𝜕 𝜉 𝜑 (𝑥, 𝜉) = 0

if 𝜑 (𝑥, 𝜉) = 𝑥 · 𝜉 − 𝐻 (𝜉). When Λ is conic we take the ℎ 𝑗 to be homogeneous of

degree zero for every 𝑗 and 𝐻 homogeneous of degree 1, all well defined in the cone
spanned by Ξ. □

Corollary 18.6.2 Let Λ be a conic local symplectic graph in (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0)

(Definition 18.2.12) and let ℘ ∈ Λ be arbitrary. There are open sets U1 ⊂ 𝑇 ∗ Ω1 \0,
U2 ⊂ 𝑇 ∗ Ω2 \0, such that ℘ ∈ U1 × U2 , and a strongly nondegenerate phase-
function (Definition 18.2.10) 𝜑 ∈ C ∞ (U1 × U2 ; R) whose associated Lagrangian
submanifold is Λ ∩ (U1 × U2 ).

Proof It suffices to modify the proof of Lemma 18.6.1 to take into account that the
fundamental 2-form is now the twisted symplectic two-form d𝜉 ∧ d𝑥 − d𝜂 ∧ d𝑦 in
a local coordinates chart. This produces a phase-function in an appropriate conic
open subset U1 × U2 of (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) containing ℘, in appropriate local
coordinates 𝑥 𝑗 , 𝑦 𝑘 , and cocoordinates 𝜉 𝑗 , 𝜂 𝑘 . We find

𝜑 (𝑥, 𝑦, 𝜉, 𝜂) = 𝑥 · 𝜉 − 𝑦 · 𝜂 − 𝐻 (𝜉, 𝜂) . (18.6.1)

18.6 Globally Defined Fourier Integral Operators 749

The critical submanifold Σ𝜑 is defined by 𝑥 = 𝜕 𝜉 𝐻 (𝜉, 𝜂), 𝑦 = 𝜕𝜂 𝐻 (𝜉, 𝜂); clearly,
the differentials d 𝑥 𝑗 𝜕 𝜉𝑘 𝜑 , d 𝑦 𝑗 𝜕𝜂𝑘 𝜑 , are linearly independent at every point of
Σ 𝜑 , proving that 𝜑 is nondegenerate (Definition 18.1.6). Since Λ is a conic local
symplectic graph the phase-function 𝜑 is strongly nondegenerate, by Proposition
18.2.13. □

Remark 18.6.3 The phase-function 𝜑 (𝑥, 𝑦, 𝜉) = (𝑥 − 𝑦) · 𝜉 − 𝐻 (𝜉) in R𝑛 × R𝑛 ×

(R𝑛 \ {0}) is clearly strongly nondegenerate at every point of (diag R𝑛 × R𝑛 ) ×
(R𝑛 \ {0}) where 𝑥 − 𝑦 = 𝜕 𝜉 𝐻 (𝜉).

18.6.2 Densities attached to a local symplectic graph

Let Λ ⊂ (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) be a local symplectic graph and the point

℘ ∈ Λ be arbitrary. Assume that there are local coordinates charts (Ω1 , 𝑥1 , ..., 𝑥 𝑛 )
and (Ω2 , 𝑦 1 , ..., 𝑦 𝑛 ) in M1 and M2 respectively, such that 𝑥 ◦ ∈ Ω1 , 𝑦 ◦ ∈ Ω2 ,
℘ = (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ), an open cone Γ in R 𝑁 \ {0} and a phase-function 𝜑 ∈
C ∞ (Ω1 × Ω2 × Γ) whose associated Lagrangian submanifold Λ 𝜑 (Definition 18.2.3)
represents Λ in the coordinate chart

𝑇 ∗ M1 | Ω1 × 𝑇 ∗ M2 | Ω2 , 𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 , 𝜂1 , ..., 𝜂 𝑛 .

If this is so we shall say that the phase-function 𝜑 defines Λ at ℘. Then there is a

𝜃 ◦ ∈ Γ satisfying

𝜑 𝑥 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 𝜉 ◦ , 𝜑 𝑦 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = −𝜂◦ , 𝜑 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0. (18.6.2)

Notation 18.6.4 We denote by 𝚽℘ (Λ) the set of nondegenerate phase-functions

(Definition 18.1.6) that define Λ at ℘.

Proposition 18.6.5 The set 𝚽℘ (Λ) is not empty whatever ℘ ∈ Λ. Every phase-
function 𝜑 ∈ 𝚽℘ (Λ) is strongly nondegenerate.

Proof That 𝚽℘ (Λ) ≠ ∅ is a direct consequence of Corollary 18.6.2. The second

claim follows from Proposition 18.2.13. □
We denote by 𝜋 ∗ the base projection (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) −→ M1 × M2 .

Notation 18.6.6 We denote by 𝚽◦℘ (Λ) the subset of 𝚽℘ (Λ) consisting of the phase-
functions 𝜑 ∈ 𝚽℘ (Λ) defined in a set Ω1 × Ω2 × Γ ∋ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) such that

𝑁 = 2𝑛 − rank ( 𝜋 ∗ | Λ ) (℘) . (18.6.3)

We begin by focusing on the elements of 𝚽◦℘ (Λ).

Proposition 18.6.7 For the phase-function 𝜑 ∈ 𝚽℘ (Λ) to belong to 𝚽◦℘ (Λ) it is

necessary and sufficient that 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0.
750 18 Fourier Integral Operators

Proof Since 𝜑 is nondegenerate dim Σ 𝜑 = 2𝑛; by (18.4.1), (18.6.3) is equivalent to

𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0. □

Corollary 18.6.8 The set 𝚽◦℘ (Λ) is not empty whatever ℘ ∈ Λ and consists of phase-
functions that are pairwise equivalent at ℘ (in the sense of Definition 18.4.17).

Proof Proposition 18.6.5 shows that there is a phase function 𝜑 ∈ 𝚽℘ (Λ). By

Proposition 18.4.6 and Corollary 18.4.7 we know that there is a nondegenerate
phase-function equivalent to 𝜑 at ℘ such that 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0. Every pair of
phase-functions belonging to 𝚽◦℘ (Λ) satisfy Conditions (i) and (ii) in Proposition
18.4.19; that they are equivalent then follows from Proposition 18.4.22. □
Let 𝜑 ∈ 𝚽◦℘ (Λ). We introduce the determinant Δ 𝜑 of the nonsingular (𝑁 + 𝑛) ×
(𝑁 + 𝑛) matrix (18.2.10); Proposition 18.6.7 entails

𝜑 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) 𝜑 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ )

Δ 𝜑 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = det 𝑥,𝑦 ◦ ◦ ◦ 𝑥, 𝜃 . (18.6.4)
𝜑 𝑦, 𝜃 (𝑥 , 𝑦 , 𝜃 ) 0

The maps in Diagram (18.2.11) are local diffeomorphisms: with an appropriate

choice of the conic neighborhood U of ℘ in (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) we can use the
coordinates (𝑥 1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 ) in Λ ∩ U or, alternatively, (𝑦 1 , ..., 𝑦 𝑛 , 𝜂1 , ..., 𝜂 𝑛 ).
We can pullback (𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 ) via the map 𝑝 𝜑 , 𝑞 𝜑 as local coordinates
in Σ 𝜑 in a neighborhood of (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). This is the same as saying that there are
unique smooth (homogeneous of degree 0 and 1 respectively) solutions 𝑦 = 𝑦 (𝑥, 𝜉),
𝜃 = 𝜃 (𝑥, 𝜉) ∈ Γ in U (possibly contracted about ℘) of the equations

𝜑 𝑥 (𝑥, 𝑦, 𝜃) = 𝜉, 𝜑 𝑦 (𝑥, 𝑦, 𝜃) = −𝜂, 𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0. (18.6.5)

This allows us to introduce the 1-density on Λ ∩ U (see Section 11.1):

−1
𝑑 𝜑 (𝑥, 𝜉) = Δ 𝜑 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) |d𝑥d𝜉 | . (18.6.6)

The density 𝑑 𝜑 is homogeneous of degree 𝑛 with respect to 𝜉.

Remark 18.6.9 The densities |d𝑥d𝜉 | 𝜅 (𝜅 ∈ Z) are invariant under symplectomor-

phisms since the Jacobian determinant of a symplectomorphism is equal to 1 (Corol-
lary 13.1.23).

We continue to reason in the local charts (Ω1 , 𝑥1 , ..., 𝑥 𝑛 ) and (Ω2 , 𝑦 1 , ..., 𝑦 𝑛 ) in
M1 and M2 respectively, assuming that 𝑥 ◦ ∈ Ω1 , 𝑦 ◦ ∈ Ω2 , ℘ = (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) and
𝜑 ∈ C ∞ (Ω1 × Ω2 × Γ) with Γ an open cone in R 𝑁 \ {0}. Consider now a second
phase-function 𝜓 ∈ 𝚽◦℘ (Λ) defined in Ω1 × Ω2 × Γ1 , where Γ1 ⊂ R 𝑁 \ {0} is an open
cone [𝑁 as in (18.6.3)]. Thus (18.6.3) is satisfied with 𝜓 in the place of 𝜑 and some
𝜃˜◦ ∈ Γ1 in that of 𝜃 ◦ ; and Λ 𝜓 ∩ U = Λ ∩ U, possibly after contracting U about ℘.
We avail ourselves of Corollary 18.6.8 (cf. also Definition 18.4.17): after contracting
Ω1 × Ω2 about (𝑥 ◦ , 𝑦 ◦ ) and ◦
Γ about the ray (𝜆𝜃 ) 𝜆>0 there is a diffeomorphism
(𝑥, 𝑦, 𝜃) ↦→ 𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) , homogeneous of degree 1, of Ω1 × Ω2 × Γ onto a conic
open subset of Ω1 × Ω2 × Γ1 such that 𝜃˜ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 𝜃˜◦ and
18.6 Globally Defined Fourier Integral Operators 751

𝜑 (𝑥, 𝑦, 𝜃) = 𝜓 𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) .

(18.6.7)

The chain-rule formulas

𝜕 𝜃˜
𝜓 𝜃 𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) ,

𝜑 𝜃 (𝑥, 𝑦, 𝜃) = (18.6.8)
𝜕𝜃
𝜕 𝜃˜
𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) + 𝜓 𝜃 𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) ,

𝜑 𝑥 (𝑥, 𝑦, 𝜃) = 𝜓 𝑥
𝜕𝑥
𝜕 𝜃˜
𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) + 𝜓 𝜃 𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) ,

𝜑 𝑦 (𝑥, 𝑦, 𝜃) = 𝜓 𝑦
𝜕𝑦
entail, after restriction to Σ 𝜑 ,

𝜑 𝑥, 𝜃 𝜑 𝜃 , 𝜃
det
𝜑 𝑥,𝑦 𝜑 𝑦, 𝜃
2
𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜
𝜓
𝜕𝜃 𝑥, 𝜃 ˜ + 𝜕𝜃 𝜕𝑥 𝜃 , 𝜃𝜓 ˜ ˜ 𝜓 𝜃˜ , 𝜃˜
= det ­ 𝜕𝜃
© ª
𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜ 𝜕 𝜃˜
®
« 𝜓 𝑥,𝑦 + 𝜕𝑦 𝜓 𝑥, ˜
𝜃 + 𝜕𝑥 𝜓 𝑦, 𝜃˜ + 𝜕𝑦 𝜕𝑥 𝜓 ˜
𝜃 , ˜
𝜃 𝜕𝜃 𝜓 𝑦, ˜
𝜃 + 𝜕𝜃 𝜕𝑦 𝜓 ˜
𝜃 , ˜
𝜃 ¬!
2 ˜
D𝜃˜ 𝜓 𝑥, 𝜃˜ + 𝜕𝜕𝑥𝜃 𝜓 𝜃˜ , 𝜃˜ 𝜓 𝜃˜ , 𝜃˜
= det ˜ ˜ ˜ ˜ ˜
D𝜃 𝜓 𝑥,𝑦 + 𝜕𝜕𝑦𝜃 𝜓 𝑥, 𝜃˜ + 𝜕𝜕𝑥𝜃 𝜓 𝑦, 𝜃˜ + 𝜕𝜕𝑥𝜃 𝜕𝜕𝑦𝜃 𝜓 𝜃˜ , 𝜃˜ 𝜓 𝑦, 𝜃˜ + 𝜕𝜕𝑦𝜃 𝜓 𝜃˜ , 𝜃˜
2
!
D𝜃˜ 𝜓 𝑥, 𝜃˜ 𝜓 𝜃˜ , 𝜃˜
= det ˜ ˜
D𝜃 𝜓 𝑥,𝑦 + 𝜕𝜕𝑦𝜃 𝜓 𝑥, 𝜃˜ 𝜓 𝑦, 𝜃˜ + 𝜕𝜕𝑦𝜃 𝜓 𝜃˜ , 𝜃˜
2
D𝜃˜

𝜓 𝑥, 𝜃˜ 𝜓 𝜃˜ , 𝜃˜
= det ,
D𝜃 𝜓 𝑥,𝑦 𝜓 𝑦, 𝜃˜
˜
D 𝜃˜
where D𝜃 = det 𝜕𝜕𝜃𝜃𝑘𝑗 . To summarize we have, in Σ 𝜑 ,
1≤ 𝑗,𝑘 ≤ 𝑁

2
D𝜃˜
Δ𝜑 = Δ𝜓 .
D𝜃

In the notation of (18.6.6) we get

√︃ 𝐷 𝜃˜ √︃
𝑑 𝜓 (𝑥, 𝜉) =
(𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) 𝑑 𝜑 (𝑥, 𝜉). (18.6.9)
𝐷𝜃

Formula (18.6.9) suggests that 𝑑 𝜑 (𝑥, 𝜉) may represent a local 21 -density section
√︁

of a globally defined fiber bundle over Λ with the 𝑁 th linear group LG (𝑁, R) as
structure bundle (Sections 9.1, 9.2). But rather than explore this aspect we proceed
to exploit (18.6.9) in terms of the amplitudes of the FIOs associated with Λ.
752 18 Fourier Integral Operators

18.6.3 The principal symbol of an FIO

Let the phase-functions 𝜑, 𝜓 ∈ 𝚽◦℘ (Λ) be as in the preceding subsection (cf.

Notation 18.6.6); (Ω1 , 𝑥1 , ..., 𝑥 𝑛 ) and (Ω2 , 𝑦 1 , ..., 𝑦 𝑛 ) are local charts in M1 and M2
respectively. We shall assume that the Fourier integral operator under consideration,
𝑨, acts on Cc∞ (Ω2 ) as in (18.3.1) with 𝑎 ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ); in this subsection we
write 𝑎 𝜑 rather than 𝑎, to dispel any ambiguity. We derive from (18.6.7) that
∫
˜
𝐴 (𝑥, 𝑦) = (2𝜋) −𝑁 e𝑖 𝜓 ( 𝑥,𝑦, 𝜃 ) 𝑎 𝜓 𝑥, 𝑦, 𝜃˜ d𝜃,˜

(18.6.10)
Γ1

with 𝑎 𝜓 ∈ 𝑆c𝑚1 (Ω1 × Ω2 , Γ1 ); we have [cf. (18.4.23)]

D𝜃˜
𝑎 𝜑 (𝑥, 𝑦, 𝜃) 𝑎 𝜓 𝑥, 𝑦, 𝜃˜ (𝑥, 𝑦, 𝜃) (𝑥, 𝑦, 𝜃) (18.6.11)
D𝜃

mod 𝑆 −∞ (Ω1 × Ω2 , Γ) and therefore 𝑚 = 𝑚 1 . Comparing (18.6.11) to (18.6.9) yields

√︃
𝑎 𝜑 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) 𝑑 𝜑 (𝑥, 𝜉)
√︃
𝑎 𝜓 𝑥, 𝑦 (𝑥, 𝜉) , 𝜃˜ (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) 𝑑 𝜓 (𝑥, 𝜉)

or, equivalently,
√︁ √︁
𝑝𝜑, 𝑞𝜑 ∗ 𝑎𝜑 Σ𝜑
𝑑𝜑 𝑝𝜓, 𝑞𝜓 ∗ 𝑎𝜓 Σ𝜓
𝑑𝜓, (18.6.12)

where the subscript star means pushforward from Σ 𝜑 to Λ; both sides are 21 -densities
valued amplitudes in Λ ∩ U and the congruence is mod 21 -density-valued amplitudes
of order −∞. The meaning of (18.6.12) is obvious: we can patch together the cosets
represented in each side of the congruence and, after modding off amplitudes of
order −∞, obtain a globally defined (in Λ) 21 -density-valued amplitude attached to 𝑨
independent of the phase-function used in the local (actually, the microlocal – taking
the “modding off” into account) representation – provided said phase-function stays
in 𝚽◦℘ (Λ).
We now turn our attention to a phase-function 𝜑 ∈ 𝚽℘ (Λ) \𝚽◦℘ (Λ) (cf. Notation
18.6.4, 18.6.6)
defined
in Ω1 × Ω2 × Γ as before but with dim Γ = 𝑁 > 𝜈 =
2𝑛 − rank 𝜋 ∗ | Λ 𝜑 (℘) [cf. (18.6.3)]. We adopt fully the notation of Subsection
18.4.1: Γ = Γ ′ × Γ ′′, with the cones Γ ′ ⊂ R𝜈 \ {0}, Γ ′′ ⊂ R 𝑁 −𝜈 \ {0} given by
(18.4.7)–(18.4.8); 𝜃 ′ (resp., 𝜃 ′′) is the variable in Γ ′ (resp., Γ ′′). We make use of
the local coordinates 𝑥 𝑗 , 𝑦 𝑘 as before; the vector 𝜃 ◦ = (𝜃 ◦′, 0) (cf. Lemma 18.4.4)
satisfies (18.6.2). We avail ourselves of Theorem 18.4.8 with 𝜈 as specified above,
implying 𝜑 𝜃 ′ , 𝜃 ′ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 0, det 𝜑 𝜃 ′′ , 𝜃 ′′ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) ≠ 0 [see (18.4.4)–(18.4.5)];
we write 𝐻 𝜑 = 𝜑 𝜃 ′′ , 𝜃 ′′ (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ). Recall Proposition 18.4.6:
18.6 Globally Defined Fourier Integral Operators 753

𝜑♭ (𝑥, 𝑦, 𝜃 ′) = 𝜑 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) ,

where 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′) is the function in Lemma 18.4.5, where we replace 𝜑 by 𝜑♭ . We

have 𝜑♭ ∈ 𝚽◦℘ (Λ) [while 𝜑 ∉ 𝚽◦℘ (Λ)]. The analogue of the amplitude in (18.4.17)
is 𝜋
1 e−𝑖 4 sign 𝐻𝜑
𝑎♭𝜑♭ (𝑥, 𝑦, 𝜃 ′) (2𝜋) − 2 ( 𝑁 −𝜈) √︃ 𝑎 𝜑 𝑥, 𝑦, 𝜃 ′, 𝜃˜′′ (𝑥, 𝑦, 𝜃 ′)

det 𝐻 𝜑

mod 𝑆c𝑚−1 (𝑈1 × 𝑈2 , Γ ′), whence

𝜋
e−𝑖 4 sign 𝐻𝜑

− 12 ( 𝑁 −𝜈)
𝑝 𝜑♭ , 𝑞 𝜑♭ 𝑎 𝜑♭ Σ
(2𝜋) √︃ 𝑝𝜑, 𝑞𝜑 ∗ 𝑎𝜑 Σ𝜑
(18.6.13)
∗ 𝜑♭
det 𝐻 𝜑

mod amplitudes (in Λ ∩ U) of order ≤ 𝑚 − 1. Keep in mind that 𝜈 is independent of

𝜑 ∈ 𝚽℘ (Λ); 𝜈 only depends on Λ.
Now assume that (18.6.10) is also valid, with dim Γ1 = 𝑁1 arbitrary and 𝑎 𝜓 ∈
𝑆c𝑚1 (Ω1 × Ω2 , Γ1 ). We denote by 𝜓 ♭ ∈ 𝚽◦℘ (Λ) the analogue of 𝜑♭ and 𝑎 𝜓♭ that of
𝑎 𝜑♭ ; we can apply (18.6.12):
√︃ √︃

𝑝𝜑, 𝑞𝜑 ∗ 𝑎 𝜑♭ Σ
𝑑 𝜑♭ 𝑝 𝜓 , 𝑞 𝜓 ∗ 𝑎 𝜓♭ Σ 𝜓♭
𝑑 𝜓♭ . (18.6.14)
𝜑♭

By (18.6.13) and its analogue for 𝜓 we obtain

𝜋
e−𝑖 4 sign 𝐻𝜑
√︃
1
(2𝜋) − 2 𝑁 √︃

𝑝 𝜑 , 𝑞 𝜑 ∗ 𝑎 𝜑♭ Σ 𝑑 𝜑♭ (18.6.15)
𝜑♭
det 𝐻 𝜑
−𝑖 𝜋 sign 𝐻 𝜓
√︃
− 12 𝑁1 e 4
(2𝜋) √︃ 𝑝 𝜓 , 𝑞 𝜓 ∗ 𝑎 𝜓♭ Σ 𝑑 𝜓♭ .
𝜓♭
det 𝐻 𝜓

In passing note that these congruences imply

1 1
𝑚+ 𝑁 = 𝑚 1 + 𝑁1 . (18.6.16)
2 2
Notation 18.6.10 Assuming 𝜑 ∈ 𝚽℘ (Λ) and det 𝐻 𝜑 ≠ 0, i.e., 𝜑 ∈ 𝚽℘ (Λ) \𝚽◦℘ (Λ),
we shall use the notation 𝑎 𝜑 to mean the 21 -density in Λ ∩ 𝑈

√︃
1 1
(2𝜋) − 2 ( 𝑁 −𝜈) √︃

𝑝𝜑, 𝑞𝜑 ∗ 𝑎 𝜑♭ Σ
𝑑 𝜑♭ . (18.6.17)
𝜑♭
det 𝐻 𝜑

If 𝜑 ∈ 𝚽◦℘ (Λ), implying 𝑁 = 𝜈 and 𝐻 𝜑 = 0, by 𝑎 𝜑 we shall mean the 21 -density

in Λ ∩ U √︁
𝑝 𝜑 , 𝑞 𝜑 ∗ 𝑎 𝜑 Σ𝜑 𝑑𝜑. (18.6.18)
754 18 Fourier Integral Operators
√︁
If we write 𝑎 𝜑 = 𝑎˜ 𝜑 (𝑥, 𝜉) |d𝑥d𝜉 | then, in all cases, 𝑎˜ 𝜑 (𝑥, 𝜉) can be viewed as a
scalar amplitude in Λ ∩ U of order 𝑚 + 21 (𝑁 − 𝜈).
In Notation 18.6.10 the congruence (18.6.15) reads
𝜋 𝜋
e−𝑖 4 sign 𝐻𝜓 𝑎 𝜓 e−𝑖 4 sign 𝐻𝜑 𝑎 𝜑

(18.6.19)

mod 12 -density-valued amplitudes of order ≤ 𝑚 + 21 (𝑁 − 𝜈) − 1 in Λ ∩ U.

So far we have kept the point (𝑥 ◦ , 𝜉 ◦ , 𝑦 ◦ , 𝜂◦ ) fixed. We now let it vary in Λ ∩
U. If U is sufficiently small, to every (𝑥, 𝜉) ∈ Λ ∩ U there are unique points
(𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) ∈ Σ 𝜑 and 𝑥, 𝑦˜ (𝑥, 𝜉) , 𝜃˜ (𝑥, 𝜉) ∈ Σ 𝜓 such that

𝜑 𝑥 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) = 𝜓 𝑥 𝑥, 𝑦˜ (𝑥, 𝜉) , 𝜃˜ (𝑥, 𝜉) = 𝜉.

On the one hand, at those points we have

sign 𝜑 𝜃 , 𝜃 = rank 𝜑 𝜃 , 𝜃 − 2𝑛 𝜑 ,

where 𝑛 𝜑 is the number of strictly negative eigenvalues of 𝜑 𝜃 , 𝜃 ; and likewise for 𝜓.

On the other hand, we have, by (18.4.1),

rank 𝜑 𝜃 , 𝜃 = 𝑁 − 2𝑛 + rank ( 𝜋 ∗ | Λ ) , (18.6.20)

whence

sign 𝜑 𝜃 , 𝜃 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) − sign 𝜓 𝜃˜ , 𝜃˜ 𝑥, 𝑦˜ (𝑥, 𝜉) , 𝜃˜ (𝑥, 𝜉) = 2𝜅 (𝑥, 𝜉) ,

(18.6.21)
where 𝜅 = 𝑛 𝜓 − 𝑛 𝜑 ∈ Z; (18.6.19) now reads

𝑎𝜓 𝑖𝜅 𝑎𝜑 . (18.6.22)

When 𝑁 = 𝑁1 = 𝜈 then 𝜑 𝜃 , 𝜃 (𝑥 ◦ , 𝑦 ◦ , 𝜃 ◦ ) = 𝜓 𝜃˜ , 𝜃˜ 𝑥 ◦ , 𝜃˜◦ = 0; in this case we set

𝜅 (𝑥 ◦ , 𝜉 ◦ ) = 0.

Proposition 18.6.11 If rank ( 𝜋 ∗ | Λ ) is constant in Λ ∩ U the same is true of 𝜅.

Proof Indeed, if rank ( 𝜋 ∗ | Λ ) is locally constant in Λ ∩ U then, by (18.6.20), so

is rank 𝜑 𝜃 , 𝜃 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)). Since both the numbers of strictly negative and
strictly positive eigenvalues of 𝜑 𝜃 , 𝜃 (𝑥, 𝑦 (𝑥, 𝜉) , 𝜃 (𝑥, 𝜉)) are lower semicontinuous
functions in Λ ∩ U they must be continuous and therefore locally constant. □

Remark 18.6.12 Obviously, it is possible for 𝜅 to remain constant at points where

rank ( 𝜋 ∗ | Λ ) changes. In geometric optics it is customary to refer to 𝜋 ∗ (Λ) in the
neighborhood of points in the base above which rank ( 𝜋 ∗ | Λ ) changes as caustics.

The above congruence are clearer when the amplitudes 𝑎 𝜑 and 𝑎 𝜓 have
leading
terms (see Subsection 18.5.4) 𝑎 𝜑 0 (𝑥, 𝑦, 𝜃) ∈ 𝑆c𝑚 (Ω1 × Ω2 , Γ) and
𝑎 𝜓 0 (𝑥, 𝑦, 𝜃) ∈ 𝑆c𝑚1 (Ω1 × Ω2 , Γ1 ) respectively. In this case (18.6.22) is equiva-
lent to
18.7 Principles of Analytic Fourier Integral Operators 755

𝑎𝜓 0 = 𝑖𝜅 𝑎𝜑 0 . (18.6.23)
Keep in mind that 𝑎 𝜑 0 and 𝑎 𝜓 0 are 21 -densities in Λ ∩ U, i.e., elements of


C ∞ Λ ∩ U, |d𝑥d𝜉 | 1/2 according to the notation of Subsection 11.1.1. If 𝑎 𝜑 0


and 𝑎 𝜓 0 were scalar (18.6.23) would indicate that they are sections (over Λ ∩ U)
of a line bundle L (Λ) over Λ called the Keller–Maslov line bundle. Actually the
complex line bundle L (Λ) is trivial, i.e., there is a vector bundle isomorphism
of L (Λ) onto Λ × C (for a proof see [Duistermaat, 1973], Lemma 4.1.3). What
is important for us is its structure group (see Section 9.1) which is taken to be
Z4 = Z/4Z with the action 𝑧 ↦→ 𝑖 𝜅 𝑧, 𝜅 ∈ Z4 : two sections 𝑠1 , 𝑠2 of L (Λ) over an
open subset Λ ∩ U of Λ are viewed as equal if there is a 𝜅 ∈ Z4 (or 𝜅 one of the
integers 0, 1, 2, 3) such that 𝑠2 (𝑥, 𝜉) = 𝑖 𝜅 𝑠1 (𝑥, 𝜉) for all (𝑥, 𝜉) ∈ Λ ∩ U. Formula
(18.6.23) suggests that we define the principal symbol of a globally defined FIO
1
𝑨 as a section of the bundle L (Λ) ⊗C 𝛀 2 (Λ); we shall denote it by 𝜎 ( 𝑨). We
1
are using accepted notation: 𝛀 2 (Λ) is the complex line bundle whose sections over
Λ ∩ U are the 21 -densities in Λ ∩ U.
The Keller–Maslov line bundle (and the Maslov index, here 𝜅) can be interpreted
in terms of the Lagrangian Grassmannian of the tangent spaces to (𝑇 ∗ Ω1 \0) ×
(𝑇 ∗ Ω2 \0) at points of the submanifold Λ; on this topic we refer the reader to
[Duistermaat, 1973] or to [Hörmander, 1983, III] (pp. 332 et seq; see also [Treves,
1980], pp. 408 et seq).

18.7 Principles of Analytic Fourier Integral Operators

To devise the C 𝜔 version of the theory described in the previous sections of this
chapter is a straightforward matter. The basic ingredients of FIO theory must be
reinterpreted in the analytic category, in a manner analogous to what was done for
pseudodifferential operators (see Ch. 17). Let Ω1 , Ω2 be two open subsets of R𝑛
and let Γ ⊂ R 𝑁 \ {0} be an open cone. In this chapter, by a real-analytic (or C 𝜔 )
phase-function 𝜑 in Ω1 × Ω2 × Γ we shall mean a real-valued C 𝜔 phase-function in
Ω1 × Ω2 × Γ (Definition 18.1.4).
Thus there are complex neighborhoods ΩC𝑗 of Ω 𝑗 ( 𝑗 = 1, 2) in C𝑛 and a conic
neighborhood ΓC of the closure of Γ in C 𝑁 \ {0} such that 𝜑 extends as a holomorphic
function in ΩC1 × ΩC2 × ΓC , here denoted by 𝜑 (𝑧, 𝑤, 𝜃), such that 𝜑 (𝑧, 𝑤, 𝜆𝜃) =
𝜆𝜑 (𝑧, 𝑤, 𝜃) whatever 𝜆 > 0. We can assume that 𝜑 𝑧 and 𝜑 𝑤 do not vanish at any
point of ΩC1 × ΩC2 × ΓC . We can define the complex version of (18.1.17).

ΣC𝜑 = (𝑥, 𝑦, 𝜃) ∈ ΩC1 × ΩC2 × ΓC ; 𝜑 𝜃 (𝑥, 𝑦, 𝜃) = 0 . (18.7.1)
756 18 Fourier Integral Operators

The maps 𝑝 𝜑 and 𝑞 𝜑 in (18.1.18) extend holomorphically to ΩC1 × ΩC2 × ΓC and

we can assume that the rank of the extended holomorphic map 𝑝 𝜑 , 𝑞 𝜑 : ΣC𝜑 −→
𝑇 ∗ ΩC1 × 𝑇 ∗ ΩC2 is equal to 2𝑛 everywhere (naturally, 𝑇 ∗ ΩC𝑗 is the complex cotangent
bundle).
Due to the possible presence of singularities at 𝜃 = 0 we cannot use analytic
amplitudes without some modification. In this section, as we have done in the theory
of analytic pseudodifferential operators, we circumvent this type of difficulty by
making use of pseudoanalytic amplitudes: we adapt Definition 17.1.4, to bring into
play the variable 𝜃 in R 𝑁 (in place of the covector 𝜉) and the holomorphic extension
to 𝜃 ∈ ΓC (the latter to simplify matters).

Definition 18.7.1 An amplitude 𝑎 (𝑥, 𝑦, 𝜃) ∈ 𝑆 𝑚 (Ω1 × Ω2 × Γ) will be called a

pseudoanalytic amplitude of order (or of degree) 𝑚 in Ω1 × Ω2 × Γ if 𝑎 can be
extended as a C ∞ function of (𝑧, 𝑤, 𝜃) in ΩC1 × ΩC2 × Γ, with ΩC𝑗 ⊃ Ω 𝑗 ( 𝑗 = 1, 2)
open subsets of C𝑛 , holomorphic with respect to (𝑧, 𝑤) and having the following
property:
To each compact subset 𝐾 of ΩC1 ×ΩC2 and open cone Γ♭ ⊂ Γ in R 𝑁 \ {0} such that
Γ♭ ∩S 𝑁 −1 ⊂⊂ Γ there are two positive numbers 𝐶◦ , 𝑅◦ , such that, for all 𝛾 ∈ Z+𝑁 ,
♭
all (𝑧, 𝑤) ∈ 𝐾 and all 𝜃 ∈ Γ ,
|𝛾 |+1
𝛾! |𝜃| 𝑚−|𝛾 | .
𝛾
|𝜃| ≥ 𝑅◦ max (1, |𝛾|) =⇒ 𝜕𝜃 𝑎 (𝑧, 𝑤, 𝜃) ≤ 𝐶◦ (18.7.2)

We shall denote by 𝑆 𝑚 𝜓a (Ω1 × Ω2 × Γ) the linear space of pseudoanalytic ampli-

tudes of order 𝑚 in Ω1 × Ω2 × Γ, and by 𝑆 𝜓a (Ω1 × Ω2 × Γ) the union of all spaces
𝑆𝑚𝜓a (Ω1 × Ω2 × Γ) as 𝑚 ranges over R.
For pseudoanalytic amplitudes the analogue of 𝑆c𝑚 (Ω1 × Ω2 × Γ) (Definition
18.1.1) is precluded by Condition (18.7.2). To cut down the support of these am-
plitudes to subcones Γ♭ such that Γ♭ ∩S 𝑁 −1 ⊂⊂ Γ one must use asymptotically
analytic cutoff multipliers of the type introduced in Section 17.3: the functions 𝑔 𝑅
constructed in Lemma 17.3.2 with a change of notation: take 𝑛 = 𝑁, let 𝜃 denote the
variable (rather than 𝜉) and exchange Γ and Γ♭ . With these changes the procedures
of microlocalization are the same as those described in Ch. 17, loc. cit.
An operator defined as in (18.3.1) where the phase-function 𝜑 is analytic and the
amplitude 𝑎 is pseudoanalytic will be referred to as an analytic Fourier integral
operator. Mostly, the properties of C ∞ FIO have analogues for analytic FIOs. More
precise results that relate to concepts in the C 𝜔 category can be proved. Perhaps the
most important is the analytic version of Theorem 18.3.5, which we now state in a
more global set-up.
Let M1 and M2 be two manifolds of class C 𝜔 of the same dimension, 𝑛, and Λ a
Lagrangian submanifold of (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) also of class C 𝜔 . We consider
an FIO 𝑨 : E ′ (M1 ) −→ D ′ (M2 ) associated with Λ (see Subsection 18.3.1); it
will be analytic, meaning that every one of its microlocal representations will be
analytic (needless to say, with the understanding that the local trivializations of the
cotangent bundles in which the analysis is carried out are themselves analytic).
18.A Appendix: Stationary Phase Formal Expansion 757

We shall assume that the Lagrangian analytic submanifold Λ is a local symplectic

graph (Definition 18.2.12); we denote by ℜΛ the relation between points of 𝑇 ∗ M1 \0
and 𝑇 ∗ M2 \0 defined by Λ: we write that (𝑥, 𝜉) ∈ ℜΛ (𝑦, 𝜂) if (𝑥, 𝜉, 𝑦, 𝜂) ∈ Λ. Locally
the correspondence (𝑦, 𝜂) ↦→ (𝑥, 𝜉) ∈ ℜΛ (𝑦, 𝜂) is a C 𝜔 symplectomorphism.

Theorem 18.7.2 Let the analytic FIO 𝑨 : D ′ (M1 ) −→ D ′ (M2 ) be associated

with the analytic local symplectic graph Λ. For every 𝑢 ∈ E ′ (M2 ) we have

𝑊 𝐹a ( 𝑨𝑢) ⊂ ℜΛ (𝑊 𝐹a (𝑢)) .

The proof is a duplication of that of Theorem 18.3.5 with the due modifications:
in the base M1 × M2 the cutoffs must be of the Ehrenpreis type (Section 3.2) while
those in phase-space (𝑇 ∗ M1 \0) × (𝑇 ∗ M2 \0) must be of the type constructed in
Lemma 17.3.2. See also the proof of Theorem 17.3.10.
We have the analogue of Proposition 18.4.23:

Proposition 18.7.3 Let the phase-function 𝜑 be C 𝜔 and nondegenerate. If Λ 𝜑 (Def-

inition 18.2.3) is the diagonal of (𝑇 ∗ Ω1 \0) × (𝑇 ∗ Ω2 \0) then 𝑨 is an analytic pseu-
dodifferential operator.

The proof of Proposition 18.7.3 is similar to that of Proposition 18.4.23 with the
added requirement that the amplitudes be pseudoanalytic. All this will be revised in
a truly analytic framework when we deal with complex phase-functions.
The action of analytic FIOs on analytic pseudodifferential operators and the
important Theorem 18.5.31 extends practically verbatim to the analytic class.

18.A Appendix: Stationary Phase Formal Expansion

In this appendix we prove the classical asymptotic expansion formula for an oscilla-
tory integral ∫
𝐼𝜆 (𝑡) = e𝑖𝜆𝜑 (𝑠,𝑡) 𝑎 (𝑠, 𝑡) d𝑠, (18.A.1)
R𝑞
with 𝑡 = (𝑡1 , ..., 𝑡 𝜈 ) a parameter varying in a smooth manifold M; 𝜑 ∈ C ∞ (R𝑞 × M)
is real-valued, 𝑎 ∈ C ∞ (R𝑞 × M) and 𝜆 → +∞. We shall assume that the support of
the function 𝑠 ↦→ 𝑎 (𝑠, 𝑡), denoted in the sequel by supp𝑠 𝑎, is contained in a bounded
open subset Ω of R𝑞 independent of 𝑡. We denote by 𝜑 𝑠 the gradient of 𝜑 and by
𝜑 𝑠𝑠 the Hessian of 𝜑 with respect to 𝑠 (𝜑 𝑠𝑠 is the 𝑞 × 𝑞 matrix of the second partial
2𝜑
derivatives 𝜕𝑠𝜕𝑗 𝜕𝑠 𝑘
).

Proposition 18.A.1 If 𝜑 𝑠 does not vanish at any point of supp𝑠 𝑎 then

sup (𝜆 𝜅 |𝐼𝜆 (𝑡)|) < +∞

𝜆>0

for all 𝑡 ∈ M, 𝜅 ∈ R+ .
758 18 Fourier Integral Operators

The proof is essentially the same as that of Proposition 18.1.12.

We are interested in the case where 𝜑 has critical points (i.e., at which 𝜑 𝑠 vanishes)
provided the critical points are nondegenerate:

(NDGN) ∀ (𝑠, 𝑡) ∈ Ω × M, 𝜑 𝑠 (𝑠, 𝑡) = 0 =⇒ det 𝜑 𝑠,𝑠 (𝑠, 𝑡) ≠ 0.

Proposition 18.A.2 If (NDGN) holds then

Crit 𝜑 = {(𝑠, 𝑡) ∈ Ω × M; 𝜑 𝑠 (𝑠, 𝑡) = 0} (18.A.2)

is a C ∞ submanifold of Ω × M. Moreover, given any 𝑡 ∗ ∈ M the set of points

(𝑠∗ , 𝑡 ∗ ) ∈ Crit 𝜑 such that 𝑠∗ ∈ supp𝑠 𝑎 is finite.

Proof By the Implicit Function Theorem the (NDGN) hypothesis implies that Crit 𝜑
can be defined locally by equations 𝑠 𝑗 = 𝑠 𝑗 (𝑡) ∈ C ∞ (M), 𝑗 = 1, ..., 𝑞. Furthermore,
it implies that the intersection points of Crit 𝜑 with a subspace 𝑡 = 𝑡 ∗ are isolated;
since supp𝑠 𝑎 is compact their number is finite. □
The forthcoming reasoning will be local. We select a point 𝑡 ◦ ∈ M and we
reason in a neighborhood N of 𝑡 ◦ in M which will be contracted about 𝑡 ◦ as
needed. We can Íselect a finite C ∞ partition of unity 𝜒 𝜄 (𝜄 = 1, ..., 𝜈) in Ω and
decompose 𝑎 = 𝜈𝜄=1 𝜒 𝜄 𝑎 in such a way that, to each 𝑡 ∗ ∈ N there is at most one
point (𝑠∗ , 𝑡 ∗ ) ∈ Crit 𝜑, 𝑠∗ ∈ supp𝑠 𝜒 𝜄 . We may as well assume that 𝑎 itself has this
property:

• There is a C ∞ map 𝑠∗ : N −→ Ω such that (Ω × N ) ∩ Crit 𝜑 is the set of points

(𝑠, 𝑡) ∈ Ω × N such that 𝑠 = 𝑠∗ (𝑡).

To simplify the notation we carry out a translation in 𝑠-space allowing us to

assume that 𝑠∗ (𝑡) = 0 ∈ Ω for all 𝑡 ∈ N . We need a version “with parameters” of
the Morse lemma.

Lemma 18.A.3 Suppose that 𝜑 ∈ C ∞ (Ω × N ) satisfies the following conditions:

(1) 𝜑 𝑠 (𝑠, 𝑡) ≠ 0 for all (𝑠, 𝑡) ∈ Ω × N , 𝑠 ≠ 0;
(2) 𝜑 𝑠 (0, 𝑡) = 0 and det 𝜑 𝑠,𝑠 (0, 𝑡) ≠ 0 for all 𝑡 ∈ N .

Then, provided N is sufficiently small, there are open subsets 𝑈, 𝑈 ′ of Ω, 0 ∈

𝑈 ∩ 𝑈 ′, and a diffeomorphism 𝑠 ↦→ 𝑥 = 𝑠 + 𝑔 (𝑠, 𝑡) of 𝑈 onto an open subset 𝑉𝑡 of Ω
such that 𝑈 ′ ⊂ 𝑉𝑡 for all 𝑡 ∈ N , with 𝑔 ∈ C ∞ (𝑈 × N ) and sup |𝑔 (𝑠, 𝑡)| ≤ const. |𝑠| 2 ,
𝑡 ∈N
such that
1
𝜑 (𝑥, 𝑡) = 𝜑 (0, 𝑡) + 𝜑 𝑠,𝑠 (0, 𝑡 ◦ ) 𝑥 · 𝑥
2
in 𝑈 ′ × N .

Proof We can write 𝜑 (𝑠, 𝑡) = 𝜑 (0, 𝑡) + 12 𝑄 (𝑠, 𝑡) 𝑠 · 𝑠, where 𝑄 (𝑠, 𝑡) is a C ∞

function in N × Ω valued in the space Sym (𝑞, R) of 𝑞 × 𝑞 real symmetric matrices;
by Hypothesis (2) 𝑄 (0, 𝑡) = 𝜑 𝑠,𝑠 (0, 𝑡) is nonsingular. Let 𝑆◦ ∈ Sym (𝑞, R) be
18.A Appendix: Stationary Phase Formal Expansion 759

nonsingular and consider the quadratic map Γ −→ Γ⊤ 𝑆◦ Γ from the algebraic group
GL (𝑞, R) of nonsingular 𝑞 × 𝑞 real matrices into Sym (𝑞, R). The differential of
this map is the linear map 𝑀 −→ ℓ (𝑀) = 𝑀 ⊤ 𝑆◦ + 𝑆◦ 𝑀, which is a surjection of
the linear space M𝑞 (R) of 𝑞 × 𝑞 real matrices onto its linear subspace Sym (𝑞, R):
indeed,
1 −1
∀𝑆 ∈ Sym (𝑞, R) , ℓ 𝑆◦ 𝑆 = 𝑆.
2
From the Constant Rank Theorem we derive that there exist a neighborhood Σ of 𝑆◦
in Sym (𝑞, R) and a C 𝜔 map 𝑆 ↦→ Γ (𝑆) of Σ into GL (𝑞, R) such that Γ (𝑆◦ ) = 𝐼𝑞
(𝐼𝑞 : the identity 𝑞 × 𝑞 matrix) and

∀𝑆 ∈ Σ, Γ (𝑆) ⊤ 𝑆◦ Γ (𝑆) = 𝑆.

The lemma is proved by taking 𝑆◦ = 𝑄 (0, 𝑡 ◦ ) and 𝑥 = Γ (𝑄 (𝑠, 𝑡)) 𝑠 with 𝑈 × N

appropriately small. □

Remark 18.A.4 Since the map Σ ∋ 𝑆 ↦→ Γ (𝑆) ∈ GL (𝑞, R) is C 𝜔 the C 𝜔 version

of Lemma 18.A.3 is valid.

Let us denote by 𝑎♭ (𝑥, 𝑡) d𝑥 the transform of the density 𝑎 (𝑠, 𝑡) d𝑠 under the
diffeomorphism
Ù 𝑈 ∋ 𝑠 ↦→ 𝑥 ∈ 𝑉𝑡 of Lemma 18.A.3. Let us assume that supp 𝑥 𝑎♭ ⊂⊂
′
𝑈 ⊂ 𝑉𝑡 . In the integral (18.A.1) we carry out the change of variables 𝑠 ↦→ 𝑥;
𝑡 ∈N
setting 𝐻 = 𝜑 𝑠,𝑠 (0, 𝑡 ◦ ) we obtain
∫
1
e−𝑖𝜆𝜑 (0,𝑡) 𝐼𝜆 (𝑡) = e 2 𝑖𝜆𝐻 𝑥· 𝑥 𝑎♭ (𝑥, 𝑡) d𝑥, 𝑡 ∈ N . (18.A.3)
R𝑞

The sought asymptotic expansion is obtained by carrying out a Fourier transform

with respect to 𝑥 and applying the Parseval formula. We avail ourselves of the
following lemma (whose proof is left as an exercise):

Lemma 18.A.5 If 𝛼 ∈ R, 𝛼 ≠ 0, the Fourier transform of the tempered distribution

1 2
e− 2 𝑖 𝛼𝑥 in R is equal to √︄
2𝜋 −𝑖 𝜋4 | 𝛼𝛼| − 12 𝑖 𝛼−1 𝜉 2
e . (18.A.4)
|𝛼|

Let us denote by sgn 𝐻 the signature (i.e., #eigenvalues > 0 − #eigenvalues < 0)
of the real symmetric matrix 𝐻.

Corollary 18.A.6 If 𝐻 ∈ Sym (𝑞, R) is nonsingular then the Fourier transform of

1
the tempered distribution e− 2 𝑖𝐻 𝑥·𝑥 in R𝑞 is equal to
1
sgn 𝐻− 12 𝑖𝐻 −1 𝜉 · 𝜉
𝑞
(2𝜋) 2 |det 𝐻| − 2 e−𝑖 4
𝜋
. (18.A.5)

Proof After an orthogonal transformation we may assume that 𝐻 is diagonal, in

which case the claim follows directly from Lemma 18.A.5. □
760 18 Fourier Integral Operators

We apply Corollary 18.A.6 to (18.A.3); we obtain, by Parseval’s formula,

∫
1
e−𝑖𝜆𝜑 (0,𝑡) 𝐼𝜆 (𝑡) = e− 2 𝑖𝜆𝐻 𝑥·𝑥 𝑎♭ (𝑥, 𝑡) d𝑥
R𝑞
∫
𝑞
− 12 𝑖 𝜋4 sgn 𝐻 −𝑞 −1
= (2𝜋/𝜆) |det 𝐻| e
2 (2𝜋) e−𝑖𝐻 𝜉 · 𝜉 /2𝜆 𝑎ˆ (𝜉, 𝑡) d𝜉,
R𝑞

where 𝑎ˆ (𝜉, 𝑡) is the Fourier transform of 𝑎♭ (𝑥, 𝑡) with respect to 𝑥. The Fourier
transform inversion formula yields, for each ℓ ∈ Z+ ,
∫ ℓ ℓ
e𝑖 𝜉 ·𝑥 𝐻 −1 𝜉 · 𝜉 𝑎ˆ (𝜉, 𝑡) d𝜉 = (2𝜋) 𝑞 D 𝑥 · 𝐻 −1 D 𝑥 𝑎♭ (𝑥, 𝑡) ,
R𝑞

where
𝑞
∑︁
D 𝑥 · 𝐻 −1 D 𝑥 𝑓 (𝑥) = 𝛾 𝑗,𝑘 D 𝑥 𝑗 D 𝑥𝑘 𝑓 (𝑥) ,
𝑗,𝑘=1

assuming that 𝐻 −1

= 𝛾 𝑗,𝑘 𝑗,𝑘=1,...,𝑞 . We get the announced asymptotic (i.e., formal)
expansion:
1 𝜋
sgn 𝐻−𝑖𝜆𝜑 (0,𝑡)
|det 𝐻| 2 e−𝑖 4 𝐼𝜆 (𝑡) (18.A.6)
𝑞2 ∑︁
∞
2𝜋 1 ℓ
≈ (2𝑖𝜆) −ℓ D 𝑥 · 𝐻 −1 D 𝑥 𝑎♭ .
𝜆 ℓ=0
ℓ! 𝑥=0

In the analytic class we might want a finite expansion with a good estimate of
the resulting remainder. We shall look at this question within the complex (i.e.,
holomorphic) framework, in Section 19.3.
Part VI
Complex Microlocal Analysis
Chapter 19
Classical Analytic Formalism

Prior to this chapter and in the chapters

Í that follow the predominant class in symbolic
calculus is that of classical symbols ∞ 𝑗=0 𝑎 𝑗 (𝑥, 𝜉), generally formal series in which
each term 𝑎 𝑗 is homogeneous (with respect to 𝜉) of degree 𝑚 − 𝑗, 𝑚 ∈ Z+ . The
emphasis will now be on such symbols where 𝑥 is allowed to vary in a domain Ω in
C𝑛 . By giving a special role to the radial variable in 𝜉-space, |𝜉 | (in the sequel often
denoted by 𝜆 and regarded as a large parameter or, in other texts, as the reciprocal of
the Planck constant), we deal with series ∞
Í
𝑗=0 𝑎 𝑗 (𝑥, 𝜃) 𝜆
𝑚− 𝑗 , 𝜃 a variable in open

subsets of the unit sphere in preceding chapters and in domains of R 𝑁 or C 𝑁 , in the

chapters immediately following the present one. But there is a large and clearly visible
area of overlapping in the structures and procedures in the diverse usage of the series,
for instance the frequent resurgence of the composition law # [understandable since
it symbolizes the composition of pseudodifferential operators; see e.g. (16.2.16)
and (19.1.12), (20.1.6) below] and the bounds on the absolute values, or norms,
of coefficients: the series almost never converge but the applications to the theory
of analytic PDEs require growth conditions of the standard kind as 𝑗 ↗ +∞;
see e.g. (17.2.42) or (19.1.1) just below. We gather the basic facts as a rather loose
formalization, at first series with coefficients in the algebra of holomorphic functions
in Ω ⊂ C𝑛 but, soon after, with coefficients in a locally convex topological vector
space 𝑬, mainly with the prospect of taking 𝑬 = O ′ (R𝑛 ), the space of analytic
functionals carried by compact subsets of R𝑛 (Ch. 6). This opens the way to dealing
with series whose coefficients are hyperfunctions or microfunctions; although the
latter are not elements of a Hausdorff TVS, the needed “analytic” constraints are
imposed on the series with coefficients in O ′ (R𝑛 ) that represent them.
Section 19.3 is devoted to the statement and proof of the complex-analytic version
of the Stationary Phase expansion formula (cf. the Appendix), not overtly used in the
remainder of the book but underlying much of the treatment of the FBI transform,
the latter somewhat redefined and applied in crucial instances.
The last section is a departure from the main themes of the book; it is devoted
to what may seem to many analysts a puzzling curiosity: the intimate relationship,
discovered by I.M. Gel’fand and L.A. Dickey in the late 1970s, between the classical
symbolic calculus in one dimension and special hierarchies of nonlinear evolution

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 763
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_19
764 19 Classical Analytic Formalism

equations, justifiably called Hamiltonian and less clearly completely integrable, the
model of which is the well-known Korteweg–De Vries (KdV) equation. We show that
the KdV hierarchy admits an infinite sequence of first integrals (also called constants
of motions or conserved quantities) and how to define them recursively. The section
is a rewriting of well-known basic material; it does not attempt to step into what
is a vast field of research (see for instance [Kappeler-Pöschel, 2003], [Kupershmidt
2000]). The only analytic feature, easily laid over the algebraic structure as outlined
in Ch. 1, [Dickey, 2003], is that we keep track of the “analytic bounds” on the
coefficients in the formal series. The reader interested in this fascinating and still
evolving subject is referred to Dickey’s excellent monograph.

19.1 Formal Analytic Series

19.1.1 Formal analytic series and their realizations

Throughout this chapter 𝜆 is a large positive parameter. Let Ω be an open subset of

C𝑛 .

DefinitionÍ19.1.1 By a formal analytic series in Ω we shall mean a formal series

ℎ (𝑧, 𝜆) = ∞𝑗=0 ℎ 𝑗 (𝑧) 𝜆
− 𝑗 in the powers of 𝜆 −1 whose coefficients ℎ ∈ O (Ω) satisfy
𝑗
the following condition:
(FA) To every compact subset 𝐾 of Ω there are positive constants 𝑀𝐾 , 𝐶𝐾 such that
𝑗
∀ 𝑗 ∈ Z+ , max ℎ 𝑗 (𝑧) ≤ 𝑀𝐾 𝐶𝐾 𝑗!. (19.1.1)
𝑧 ∈𝐾

We shall denote by Aform (Ω) the set of formal analytic series in Ω.

If 𝑈, 𝑉 are open subsets of C𝑛 such that 𝑉 ⊂ 𝑈 ⊂ Ω the restriction map

𝑉:
𝜌𝑈 Aform (𝑈) −→ Aform (𝑉) is defined coefficientwise; the pairs Aform (𝑈) , 𝜌𝑈 𝑉
define a presheaf over Ω and the latter defines a sheaf Aform (Ω).
In a number of instances
Í (e.g., in Subsection 19.1.3) it is preferable to deal with
formal series ℎ (𝑧, 𝜆) = ∞ 𝑗=0 ℎ 𝑗 (𝑧) 𝜆 𝑚− 𝑗 , 𝑚 ∈ R; we shall denote by 𝜆 𝑚 A
form (Ω)
the linear space of these series and use the notation
Ø
Aeform (Ω) = 𝜆 𝑚 Aform (Ω) . (19.1.2)
𝑚∈R

The linear spaces Aform (Ω) and Aeform (Ω) are rings with respect to ordinary addition
and multiplication, admitting various functional interpretations. We use the notation
{𝜆 ∈ S2 ; |𝜆| > 𝑅} to mean the subset {𝜆 ∈ C; |𝜆| > 𝑅} ∪ {∞} of the Riemann
sphere S2 .
19.1 Formal Analytic Series 765

Proposition 19.1.2 To every ℎ ∈ Aform (Ω) there is a unique holomorphic function

ℎ of (𝑧, 𝜆) ∈ Ω × 𝜆 ∈ S2 ; |𝜆| > 𝑅 (𝑅 > 0 depending on ℎ) such that ℎ is the

e
restriction to Ω × (𝑅, +∞) of the Taylor expansion of e
ℎ with respect to 𝜆 about ∞.

This is self-evident. There is an analogous statement for A eform (Ω) where holo-
morphic is replaced by holomorphic with respect to 𝑧 and meromorphic (witha single
pole at ∞), Laurent instead of Taylor expansion, with respect to 𝜆. Let O∞ S2 denote
the ring of germs at ∞ of holomorphic functions in 𝜆-space S2 .

Corollary 19.1.3 The map ℎ ↦→ eℎ is a ring isomorphism of Aform (Ω) onto the ring
of holomorphic functions in Ω valued in O∞ S2 [or vice versa, onto the ring of

germs at ∞ of holomorphic functions in 𝜆-space S2 valued in O (Ω)].

Corollary 19.1.3 states that there is a natural ring isomorphism

Aform (Ω) O Ω; O∞ S2 O (Ω) b ⊗O∞ S2

where the second isomorphism applies to the TVS structures, O (Ω) and O∞ S2

being nuclear.

Corollary 19.1.4 The natural locally convex topology on Aform (Ω) is the topology
⊗O∞ S2 .

transferred from O (Ω) b

Remark 19.1.5 The topology of Aform (Ω) can be described “concretely”. We select
arbitrarily a sequence of compact subsets 𝐾 𝜈 (𝜈 = 1, 2, ...) of Ω such that 𝐾 𝜈 ⊂
𝐾 𝜈+1 and Ω = ∞
Ð
𝜈=1 𝜈 . For each increasing sequence of positive integers, 𝑵 =
𝐾
(𝑵 )
{𝑁 𝜈 } 𝜈=1,2,... , we denote by Aform (Ω) the subspace of Aform (Ω) consisting of the
Í∞
series ℎ (𝑧, 𝜆) = 𝑗=0 ℎ 𝑗 (𝑧) 𝜆− 𝑗 that satisfy

1 −𝑗
∥ℎ∥ (𝑵 ) = sup 𝑁 max ℎ 𝑗 (𝑧) < +∞.
𝑗 ∈Z+ 𝑗! 𝜈 𝑧 ∈𝐾𝜈

(𝑵)
Each space Aform (Ω) is a Banach space with respect to the norm ∥·∥ (𝑵 ) . By defini-
(𝑵 ) (𝑵 )
tion, a convex subset 𝔘 of Aform (Ω) is open if 𝔘∩Aform (Ω) is open in Aform (Ω)
whatever the sequence 𝑵. Thus Aform (Ω) is an uncountable “inductive limit” of
Banach spaces. It is not difficult to check that this definition is independent of the
choice of the sequence of compact sets 𝐾 𝜈 .

Corollary 19.1.6 For ℎ (𝑧, 𝜆) = ∞ − 𝑗 to be invertible in A

Í
𝑗=0 ℎ 𝑗 (𝑧) 𝜆 form (Ω) it is
necessary and sufficient that ℎ0 (𝑧) ≠ 0 whatever 𝑧 ∈ Ω.

It is a direct consequence of the Cauchy inequalities that the partial derivatives

𝜕
𝜕𝑧 𝑗 ( 𝑗 = 1, ..., 𝑛) define continuous linear maps of Aform (Ω) into itself.
766 19 Classical Analytic Formalism

Definition
Í∞ 19.1.7 By a finite realization of the formal analytic series in Ω, ℎ (𝑧, 𝜆) =
ℎ
𝑗=0 𝑗 (𝑧) 𝜆− 𝑗 , we mean a finite series
∑︁
ℎ (𝑅) (𝑧, 𝜆) = ℎ 𝑗 (𝑧) 𝜆− 𝑗 . (19.1.3)
𝑗 ≤𝜆/𝑅

Except in trivial cases, (0, +∞) ∋ 𝜆 ↦→ ℎ (𝑅) (𝑧, 𝜆) is not a Í continuous function:
if 𝑚𝑅 < 𝜆 < (𝑚 + 1) 𝑅 (𝑚 ∈ Z+ ) we have ℎ (𝑅) (𝑧, 𝜆) = 𝑚 −𝑗
𝑗=0 ℎ 𝑗 (𝑧) 𝜆 . The
foundation for the use of finite realizations lies in the following two lemmas.

Lemma 19.1.8 Let ℎ ∈ Aform (Ω). Given an arbitrary compact subset 𝐾 of Ω let
𝑀𝐾 , 𝐶𝐾 be the constants in (19.1.1). If 𝑅 ≥ 𝐶𝐾 then
∑︁
∀𝜆 ≥ 1, 𝜆− 𝑗 max ℎ 𝑗 (𝑧) ≲ 𝑀𝐾 .
𝑧 ∈𝐾
𝑗 ≤𝜆/𝑅

Proof By (19.1.1), (19.1.3) and Stirling’s inequality, 𝜆 ≥ 𝑅 𝑗 implies

𝑗
−𝑗
√︁ 𝐶𝐾
𝜆 max ℎ 𝑗 ≲ 𝑗 + 1𝑀𝐾 , (19.1.4)
𝐾 e𝑅

whence 𝑗
∑︁ ∑︁ √︁ 𝐶𝐾
𝜆− 𝑗 max ℎ 𝑗 (𝑧) ≲ 𝑀𝐾 𝑗 +1 . (19.1.5)
𝑧 ∈𝐾 e𝑅
𝑗 ≤𝜆/𝑅 𝑗 ≤𝜆/𝑅
Í∞ √︁
If 𝑅 ≥ 𝐶𝐾 the claims follow from the fact that 𝑗=0 𝑗 + 1e− 𝑗 < +∞. □

Lemma 19.1.9 Let ℎ ∈ Aform (Ω). Given an arbitrary compact subset 𝐾 of Ω let
𝑀𝐾 , 𝐶𝐾 be the constants in (19.1.1). If 𝐶𝐾 ≤ 𝑅 < 𝑅 ′ then
[𝜆/𝑅]
∑︁ √︁ ′
∀𝜆 ≥ 1, 𝜆− 𝑗 max ℎ 𝑗 (𝑧) ≲ 𝑀𝐾 1 + 𝜆/𝑅e−𝜆/𝑅 . (19.1.6)
𝑧 ∈𝐾
𝑗=[𝜆/𝑅′ ]

Proof We suppose that (19.1.1) and (19.1.4) hold. If 𝑅 ′ > 𝑅 ≥ 𝐶𝐾 we derive [cf.
(19.1.5)]
[𝜆/𝑅]
∑︁ [𝜆/𝑅]
∑︁ √︁
𝜆− 𝑗 max ℎ 𝑗 (𝑧) ≲ 𝑀𝐾 𝑗 + 1e− 𝑗
𝑧 ∈𝐾
𝑗=[𝜆/𝑅′ ]+1 𝑗=[𝜆/𝑅′ ]+1
√︁ ′
≲ 𝑀𝐾 1 + 𝜆/𝑅e−𝜆/𝑅 ,

whence the claim. □

19.1 Formal Analytic Series 767

19.1.2 Auxiliary classes. Globalization and germification

We can exploit Lemmas 19.1.8 and 19.1.9 to construct global (as well as sheaves
of germs of) function-like objects of interest to the topic of this Part. For this we
need to introduce two rings (with respect to ordinary addition and multiplication) of
holomorphic functions in Ω depending on 𝜆 ≥ 1.
In the sequel we shall often mod off functions and germs of functions that decay
exponentially as 𝜆 → +∞.
Definition 19.1.10 We shall denote by O (−𝜔) (Ω) the algebra of measurable func-
tions ℎ (𝑧, 𝜆) in Ω × [1, +∞) that are holomorphic with respect to 𝑧 ∈ Ω and satisfy
the following condition:
(EXP DECAY) To every compact subset 𝐾 of Ω there are positive constants 𝐶, 𝜅
such that
∀𝜆 ≥ 1, max |ℎ (𝑧, 𝜆)| ≤ 𝐶e−𝜅𝜆 . (19.1.7)
𝑧 ∈𝐾

We shall denote by O𝑧(−𝜔)

◦ the set of germs at 𝑧 ◦ ∈ Ω of the functions ℎ belonging
to O (−𝜔) (𝑈) for some neighborhood 𝑈 of 𝑧◦ , and by O (−𝜔) the sheaf over C𝑛 whose
stalks are the rings O𝑧(−𝜔)
◦ .
There is a natural locally convex topology on O (−𝜔) (Ω) that makes it an LF
space, the inductive limit of the Fréchet subspaces 𝑬 𝜈 (𝜈 = 1, 2, ...) defined by the
property
∀𝐾 ⊂⊂ Ω, ∀𝜆 ≥ 1, max e𝜆/𝜈 ℎ (𝑧, 𝜆) < +∞.
𝑧 ∈𝐾

Definition 19.1.11 We shall denote by Ob (Ω) the ring (with respect to ordinary
addition and multiplication) of measurable functions ℎ (𝑧, 𝜆) in Ω × [1, +∞) that are
holomorphic with respect to 𝑧 ∈ Ω and satisfy the following condition: whatever the
compact subset 𝐾 of Ω,

sup max |ℎ (𝑧, 𝜆)| < +∞.
𝜆≥1 𝑧 ∈𝐾

We shall denote by Ob,𝑧 ◦ the ring of the germs at 𝑧◦ ∈ C𝑛 of functions belonging to

Ob (𝑈) for some neighborhood 𝑈 of 𝑧 ◦ , and by Ob the sheaf over C𝑛 whose stalks
are the rings Ob,𝑧 ◦ .
In Definition 19.1.11 “measurable” can be taken to mean that the scalar function
𝜆 ↦→ ⟨𝜇 𝑧 , ℎ (𝑧, 𝜆)⟩ is Lebesgue measurable whatever 𝜇 ∈ O ′ (Ω). Clearly, O (−𝜔) (Ω)
(resp., O𝑧(−𝜔)
◦ ) is an ideal in the ring Ob (Ω) (resp., Ob,𝑧 ◦ ).
Obvious consequences of the Cauchy inequalities are
(1) Given an arbitrary ℎ ∈ Ob (Ω), to every compact set 𝐾 ⊂ Ω there is a 𝐶 > 0
such that, for all 𝛼 ∈ Z+𝑛 ,

∀𝜆 ≥ 1, max 𝜕𝑧𝛼 ℎ (𝑧, 𝜆) ≤ 𝐶 | 𝛼 |+1 𝛼!. (19.1.8)

𝑧 ∈𝐾
768 19 Classical Analytic Formalism

(2) Given an arbitrary ℎ ∈ O (−𝜔) (Ω), to every compact set 𝐾 ⊂ Ω there are positive
constants 𝐶, 𝜅 such that, for all 𝛼 ∈ Z+𝑛 ,

∀𝜆 ≥ 1, max 𝜕𝑧𝛼 ℎ (𝑧, 𝜆) ≤ 𝐶 | 𝛼 |+1 𝛼!e−𝜅𝜆 . (19.1.9)

𝑧 ∈𝐾

We can state

Proposition 19.1.12 For every 𝛼 ∈ Z+𝑛 , 𝜕𝑧𝛼 maps each one of the rings Ob (Ω),
O (−𝜔) (Ω), Ob,𝑧 ◦ , O𝑧(−𝜔)
◦ , as well as each one of the corresponding sheaves of rings
Ob , O (−𝜔) , into itself.

The following can be directly deduced from Lemmas 19.1.8 and 19.1.9:

Proposition 19.1.13 Given arbitrarily ℎ ∈ Aform (Ω) and an open subset Ω′ of C𝑛 ,

Ω′ ⊂⊂ Ω, there is a number 𝑅◦ > 0 such that ℎ (𝑅) ∈ Ob (Ω′) [cf. (19.1.3)] and
′
ℎ (𝑅) − ℎ (𝑅 ) ∈ O (−𝜔) (Ω′) whatever 𝑅 ′ > 𝑅 ≥ 𝑅◦ .

We exploit Proposition 19.1.13 to construct global function-like objects of in-

terest. We select arbitrarily a sequence of open subsets of Ω, Ω𝜈 ⊂⊂ Ω𝜈+1
(𝜈 = 1, 2, ...), Ω = ∞
Ð
𝜈=1 𝜈 For each ℎ ∈ A form (Ω) we can select a sequence
Ω .
of positive numbers 𝑅 𝜈 < 𝑅 𝜈+1 such that, if 𝜈 ′ > 𝜈 then ℎ (𝑅𝜈 ) ∈ Ob (Ω𝜈 ) and
ℎ (𝑅𝜈′ ) Ω𝜈 − ℎ (𝑅𝜈 ) ∈ O (−𝜔) (Ω𝜈 ); restriction from Ω𝜈′ to Ω𝜈 defines a natural map

𝜌 𝜈𝜈′ : Ob (Ω𝜈′ ) /O (−𝜔) (Ω𝜈′ ) −→ Ob (Ω𝜈 ) /O (−𝜔) (Ω𝜈 ) .

eb (Ω) of the quotient rings Ob (Ω𝜈 ) /O (−𝜔) (Ω𝜈 ): by

We can form the direct limit O
this we mean the quotient of the disjoint union
∞
•

+𝜈=1
Ob (Ω𝜈 ) /O (−𝜔) (Ω𝜈 ) (19.1.10)

for the restriction relation

· (𝜈′ ) · (𝜈)
𝜌 𝜈𝜈′ ℎ =ℎ , 𝜈 < 𝜈′
· (𝜈)
[here ℎ is a coset in Ob (Ω𝜈 ) /O (−𝜔) (Ω𝜈 ) ]. The set O
eb (Ω) is independent of
the choice of the sequence Ω𝜈 in the following sense. Supposewe select a different
“exhaustive” sequence of relatively compact subsets of Ω, Ω′𝜇 𝜇=1,2,... . Given
(𝜈)
·
ℎ ∈O eb (Ω) we can arbitrarily select representatives ℎ (𝜈) ∈ Ob (Ω𝜈 ) of
𝜈=1,2,...
· (𝜈)
the coset ℎ for each 𝜈; given any integer 𝜇 such that Ω′𝜇 ⊂ Ω𝜈 the restriction of

ℎ (𝜈) to Ω′𝜇 determines unambiguously a coset in Ob Ω′𝜇 /O (−𝜔) Ω′𝜇 ; obviously,
· (𝜈)
the latter does not change if we choose a different representative of ℎ ; nor does
19.1 Formal Analytic Series 769

it change if we replace 𝜈 by a different positive integer 𝜈 ′ provided Ω′𝜇 ⊂ Ω𝜈′ since

′

ℎ (𝜈 ) Ω′ − ℎ (𝜈) Ω′ ∈ O (−𝜔) Ω′𝜇 . By letting 𝜇 ↗ +∞ we determine unambiguously
𝜇 𝜇
eb (Ω) based on the sequence Ω′

an element of the analogue of O . This
𝜇 𝜇=1,2,...
mapping has an obvious inverse and is therefore bijective. Since O (−𝜔) (Ω𝜈 ) is
an ideal of the ring Ob (Ω𝜈 ) the quotients Ob (Ω𝜈 ) /O (−𝜔) (Ω𝜈 ) are commutative
rings and since the restriction mappings are ring homomorphisms O eb (Ω) inherits a
commutative ring structure through (19.1.10).
Rather than expand outward we may want to contract inward and select the open
sets Ω𝜈 to form a basis of neighborhoods of 𝑧 ◦ ∈ Ω, now such that Ω𝜈+1 ⊂⊂ Ω𝜈 ⊂⊂
Ω0 ⊂⊂ Ω. If ℎ ∈ Aform (Ω𝜈 ) but ℎ ∉ Aform (Ω𝜈−1 ) (𝜈 = 1, 2, ...) we can attach to ℎ a
′
number 𝑅 𝜈 > 0 such that ℎ (𝑅) Ω𝜈 ∈ Ob (Ω𝜈 ) and ℎ (𝑅) Ω𝜈 − ℎ (𝑅 ) Ω𝜈 ∈ O (−𝜔) (Ω𝜈 )
whatever 𝑅 ′ > 𝑅 ≥ 𝑅 𝜈 . If 𝑅 ≥ 𝑅 𝜈 we will have ℎ (𝑅) Ω𝜈+𝑘 ∈ Ob (Ω𝜈+𝑘 ) for every
𝑘 = 1, 2, ..., thereby defining a germ at 𝑧◦ , h𝑧 ◦ ∈ Ob,𝑧 ◦ /O𝑧(−𝜔)
◦ , independent of the
choice of Ω𝜈 or 𝑅 𝜈 .

19.1.3 Classical analytic symbols

We extend to complex space the notion of a classical symbol (Definition 16.3.2); Ω

and Θ are domains in C𝑛 and C 𝑁 respectively.

Definition 19.1.14 Let 𝑚 ∈ R. By aÍclassical analytic symbol of order 𝑚 in Ω × Θ

we shall mean a series 𝑎 (𝑧, 𝜃, 𝜆) = ∞
𝑗=0 𝜆
𝑚− 𝑗 𝑎 (𝑧, 𝜃) in Ω×Θ ∈ 𝜆 𝑚 A
𝑗 form (Ω × Θ).
We shall denote by 𝑆class (Ω × Θ) the linear space of classical analytic symbols of
𝑚

order 𝑚 in Ω × Θ.

We shall denote by 𝑆class (Ω × Θ) the ring (with respect to ordinary addition

and multiplication) of all classical analytic symbols of all orders in Ω × Θ; thus
𝑆class (Ω × Θ) = A eform (Ω × Θ) [cf. (19.1.2)]. Classical analytic amplitudes are
simply classical analytic symbols where 𝑧 is replaced by (𝑧, 𝑤) ∈ Ω1 × Ω2 , Ω 𝑗 an
open subset of C𝑛 , and 𝑛 by 2𝑛.

Remark 19.1.15 In the classical framework the notion of order of a symbol does not
have much meaning since it can be modified by multiplication by a power of 𝜆. Such
a multiplication does not have any effect on the operations we shall be interested in.
In most instances one can assume 𝑚 = 0.

It is important to keep in mind that a classical analytic symbol, generally speak-

ing, is not a function since the series 𝑎 (𝑧, 𝜃, 𝜆) = ∞
Í 𝑚− 𝑗 𝑎 (𝑧, 𝜃) is generally
𝑗=0 𝜆 𝑗
divergent. To palliate this we can exploit the procedure introduced at the end of the
previous subsection. We introduce a sequence of open subsets Ω𝜈 × Θ𝜈 (𝜈 = 1, 2, ...)
in C𝑛+𝑁 such that

Ω𝜈 × Θ𝜈 ⊂⊂ Ω𝜈+1 × Θ𝜈+1 ⊂⊂ Ω × Θ, 𝜈 = 1, 2, ...,

770 19 Classical Analytic Formalism
+∞
Ø
and Ω × Θ = Ω𝜈 × Θ𝜈 . We associate to the formal series 𝑎 (𝑧, 𝜃, 𝜆) a (unique)
𝜈=1
· (𝜈)
sequence of cosets 𝑎 ∈ Ob (Ω𝜈 × Θ𝜈 ) /O (−𝜔) (Ω𝜈 × Θ𝜈 ) with the following prop-
·
erty: if we choose arbitrarily a representative 𝑎 (𝜈) ∈ Ob (Ω𝜈 × Θ𝜈 ) of 𝑎 𝜈 for each
𝜈 = 1, 2, ..., then

𝑎 (𝜈+1) − 𝑎 (𝜈) ∈ O (−𝜔) (Ω𝜈 × Θ𝜈 ) .

Ω𝜈 ×Θ𝜈

These representatives have the form

∑︁
𝑎 (𝜈) (𝑧, 𝜃, 𝜆) = 𝜆 𝑚− 𝑗 𝑎 𝑗 (𝑧, 𝜃) (19.1.11)
𝑗 ≤𝜆/𝑅𝜈

for a suitable choice of the numbers 𝑅 𝜈 > 0, 𝑅 𝜈 < 𝑅 𝜈+1 ↗ +∞; 𝑎 (𝜈) shall be
referred to as a finite realization of the formal symbol 𝑎 (Definition 19.1.7).
Throughout the remainder of this subsection we assume 𝑁 = 𝑛; not to √ lose sight
of this fact we write 𝜁 instead of 𝜃 and Ξ instead of Θ. As usual, 𝐷 = − −1𝜕.
Proposition 19.1.16 If 𝑎 = ∞
Í 𝑚− 𝑗 𝑎 (𝑧, 𝜁) and 𝑏 = Í∞ 𝜆 𝑚′ − 𝑗 𝑏 (𝑧, 𝜁) are two
𝑗=0 𝜆 𝑗 𝑗=0 𝑗
classical analytic symbols in Ω × Ξ then
∞
∑︁ ′
∑︁ 1 𝛾 𝛾
𝑎#𝑏 = 𝜆 𝑚+𝑚 −ℓ 𝜕 𝑎 𝑗 D𝑧 𝑏 𝑘 (19.1.12)
𝛾! 𝜁
ℓ=0 𝑗+𝑘+|𝛾 |=ℓ

is also a classical analytic symbol in Ω × Ξ.

Proof We know (Definitions 19.1.1, 19.1.14) that to every compact subset 𝐾 of
Ω × Ξ there is a 𝐶𝐾 > 0 such that
𝑗+1
∀ 𝑗 ∈ Z+ , max 𝑎 𝑗 (𝑧, 𝜁) ≤ 𝐶𝐾 𝑗!. (19.1.13)
𝑧 ∈𝐾

Let 𝐾1 (resp., 𝐾2 ) be a compact subset of Ω (resp., Ξ). We select two positive

numbers 𝑑1 < dist (𝐾1 , 𝜕Ω), 𝑑2 < dist (𝐾2 , 𝜕Ξ). From (19.1.13) (and the analogue
for 𝑏) and the Cauchy inequalities we get
∑︁ 1 𝛾 𝛾
max 𝜕 𝑎 𝑗 𝜕𝑧 𝑏 𝑘
𝐾1 ×𝐾2 𝛾! 𝜁
𝑗+𝑘+|𝛾 |=ℓ
∑︁
(𝑑1 𝑑2 ) − |𝛾 | 𝐶𝐾1 𝐶𝐾𝑘+1
𝑗+1
≤ 2
𝑗!𝑘!𝛾!
𝑗+𝑘+|𝛾 |=ℓ
ℓ
𝐶𝐾1 𝐶𝐾2 © ∑︁ 𝑗!𝑘!𝛾! ª
≤ 𝐶𝐾1 𝐶𝐾2 ­ ® ℓ!
𝑑1 𝑑2 ℓ!
« 𝑗+𝑘+|𝛾 |=ℓ ¬
ℓ+1
2𝑛 𝐶𝐾1 𝐶𝐾2
≤ 3 𝑑1 𝑑2 ℓ!,
𝑑1 𝑑2
19.1 Formal Analytic Series 771

the last inequality a direct consequence of Lemma 17.2.14. From it the claim follows
immediately. □
The classical symbol ∞
Í
𝑗=0 𝑎 𝑗 (𝑧, 𝜁) 𝜆
𝑚− 𝑗 is said to be elliptic (of order 𝑚) in

Ω × Ξ if 𝑎 0 (𝑧, 𝜁) ≠ 0 for all (𝑧, 𝜁) ∈ Ω × Ξ (cf. Definition 16.2.14 and Proposition

16.3.10).

Theorem 19.1.17 If the classical analytic symbol 𝑎 in Ω × Ξ is elliptic of order 𝑚

then there is a classical analytic symbol 𝑏 in Ω × Ξ elliptic of order −𝑚, such that
𝑎#𝑏 = 1.

Theorem 19.1.17 is the analytic version of the “pseudoanalytic” result, Proposition

17.2.30.
Proof We assume 𝑚 = 0 (see Remark 19.1.15). We refer the reader to the
proof of Proposition 16.2.16. The equations that determine the symbol 𝑏 (𝑧, 𝜁) =
Í∞ − 𝑗 are first 𝑎 𝑏 ≡ 1 and therefore deg 𝑎 = deg 𝑏 = 0 and then the
𝑗=0 𝑏 𝑗 (𝑧, 𝜁) 𝜆 0 0
analogues of (16.2.21); the latter imply directly that 𝑏 𝑗 ∈ O (Ω × Ξ) as well as the
estimate ∑︁ 1 𝛼
𝑏 𝑗 ≤ |𝑏 0 | 𝜕 𝑎 𝑘 𝜕𝑧𝛼 𝑏 ℓ (19.1.14)
𝛼! 𝜁
𝑘+ℓ+ | 𝛼 |= 𝑗
ℓ< 𝑗

( 𝑗 = 1, 2, ...). Assuming that (19.1.13) holds we must prove that, given an arbitrary
compact subset 𝐾 of Ω × Ξ, there is a constant 𝐶𝐾 > 0 such that
𝑗+1
∀ 𝑗 ∈ Z+ , max 𝑏 𝑗 (𝑧, 𝜁) ≤ 𝐶𝐾 𝑗!. (19.1.15)
(𝑧,𝜁 ) ∈𝐾

We are going to apply Lemma 23.2.6 and make use of the notation in that lemma
(replacing 𝑛 by 2𝑛), in particular the norm (23.2.16). Let 𝑟 ∈ (0, 1) be such that
n o
𝐾𝑟 = (𝑧, 𝜁) ∈ C𝑛 × C𝑛 ; ∃ (𝑧∗ , 𝜁 ∗ ) ∈ 𝐾, 𝑧 𝑗 − 𝑧∗𝑗 ≤ 𝑟, 𝜁 𝑗 − 𝜁 ∗𝑗 ≤ 𝑟, 𝑗 = 1, ..., 𝑛

is a compact subset of Ω × Ξ. Our claim is that the following holds, for some constant
𝐶1 > 0, every 𝑘 ∈ Z+ and every 𝑠, 0 < 𝑠 < 1,
e 𝑘
𝑁 𝑠 (𝑏 𝑘 ) ≤ 𝐶1𝑘+1 𝑘!. (19.1.16)
1−𝑠
Obviously (19.1.16) holds for 𝑘 = 0 if 𝐶1 ≥ 𝑁1 (𝑏 0 ) = 𝑁1 𝑎 −1

0 . For use below we
define 𝐶◦ = max 2, 𝑟 −1 , 𝑁1 𝑎 −1

0 .
We shall reason by induction on 𝑗 ≥ 1 and assume that (19.1.16) holds for all
𝑘 < 𝑗. Lemma 23.2.6 entails, for every 𝛼 ∈ Z+𝑛 ,

(𝑘 + |𝛼|)! e 𝑘+| 𝛼 |
𝑁 𝑠 𝜕𝑧𝑎 𝑏 𝑘 ≤ 𝐶◦| 𝛼 | 𝐶1𝑘+1

.
𝑘! 1−𝑠
772 19 Classical Analytic Formalism

In dealing with the 𝑎 𝑘 we make use of (19.1.13) and of the Cauchy inequalities;
we get
1
𝑁1 𝜕𝜁𝑎 𝑎 𝑘 ≤ 𝐶2| 𝛼 |+𝑘+1 𝑘!
𝛼!
for some 𝐶2 > 0 independent of 𝑘 and 𝛼. We put both these last estimates in (19.1.14):
∑︁ 𝑘! (ℓ + |𝛼|)! e ℓ+| 𝛼 |
𝐶◦| 𝛼 |+1 𝐶1ℓ+1 𝐶2| 𝛼 |+𝑘+1

𝑁𝑠 𝑏 𝑗 ≤
ℓ! 1−𝑠
𝑘+ℓ+| 𝛼 |= 𝑗
ℓ< 𝑗
𝑗−1 ∑︁
𝑗
𝑗+1
e 𝑗 ∑︁ 𝑘! ( 𝑗 − 𝑘)!
≤ 𝐶1 𝑗! 𝐶◦ 𝐶2 (𝐶◦ 𝐶2 /𝐶1 ) 𝑗−ℓ .
1−𝑠 ℓ=0 𝑘=0
𝑗!ℓ!

For (19.1.16) to hold when 𝑘 = 𝑗 we need 𝐶1 to be sufficiently large that

𝑗−1 ∑︁
𝑗−ℓ
∑︁ 𝑘! ( 𝑗 − 𝑘)! 1−𝑘
𝐵𝑗 = 𝐶◦ 𝐶2 (𝐶◦ 𝐶2 /𝐶1 ) 𝑗−ℓ ≤ 1 (19.1.17)
ℓ=0 𝑘=0
𝑗!ℓ!

whatever the integer 𝑗 ≥ 1. Lemma 17.2.14, implies

𝑗−ℓ
∑︁ 𝑘! ( 𝑗 − 𝑘)! 1−𝑘
𝐶◦ ≤ 3𝐶◦
𝑘=0
𝑗!ℓ!

whence
𝑗−1
∑︁
𝐵 𝑗 ≤ 3𝐶◦ 𝐶2 (𝐶◦ 𝐶2 /𝐶1 ) 𝑗−ℓ ;
ℓ=0

(19.1.17) ensues if we take 𝐶1 ≥ 𝐶◦ 𝐶2 + 3𝐶◦2 𝐶22 . □

19.1.4 General formal analytic series

In this subsection 𝑬 shall be a locally convex, Hausdorff topological vector space

(TVS).

Í By a formal
Definition 19.1.18 −1 series in 𝑬 we shall mean a formal power
analytic
series 𝒉 (𝜆) = ∞ 𝒂
𝑗=0 𝑗 𝜆 −𝑗 ∈ 𝑬 𝜆 whose coefficients 𝒂 𝑗 ∈ 𝑬 satisfy the
following condition:
(GFA) To every continuous seminorm ℘ on 𝑬 there are positive constants 𝑀℘ , 𝐶℘ ,
such that 𝑗
∀ 𝑗 ∈ Z+ , ℘ 𝒂 𝑗 ≤ 𝑀℘𝐶℘ 𝑗!. (19.1.18)
The linear space of formal analytic series in 𝑬 shall be denoted by Aform 𝑬.
19.1 Formal Analytic Series 773

Definition 19.1.18 can be applied directly to all the TVSs encountered in this
book; if Ω is an open subset of C𝑛 we have Aform O (Ω) = Aform (Ω). The analogues
of Proposition 19.1.2 and Corollary 19.1.3 are valid:
Proposition 19.1.19 To every ℎ ∈ Aform 𝑬 there is a 𝑬-valued holomorphic function
ℎ of 𝜆 ∈ S2 , |𝜆| > 𝑅 (𝑅 > 0 depending on ℎ) such that ℎ is the restriction of the Taylor
e
expansion of e ℎ about ∞ to (𝑅, +∞). The map ℎ ↦→ e ℎ induces a linear bijection of
Aform 𝑬 onto O∞ S2 ; 𝑬 , the space of germs at ∞ of 𝑬-valued holomorphic functions
in S2 .
Remark 19.1.20 We describe the locally convex topology on Aform 𝑬 transferred
from the natural topology of O∞ S2 ; 𝑬 . We select a family 𝔉 of continuous semi-
norms ℘ on 𝑬 that define the topology of 𝑬: the sets 𝒆 ∈ 𝑬; ℘ (𝒆) ≤ 𝑚1 form a

basis of neighborhoods of 0 in 𝑬 as ℘ ranges over 𝔉 and 𝑚 = 1, 2, .... Consider

a map 𝔉 ∋ ℘ ↦→ 𝐶℘ > 0 submitted to the following condition (and otherwise
arbitrary): if ℘, ℘′ ∈ 𝔉, ℘ ′
≤ ℘ implies 𝐶℘ ≤ 𝐶℘ . We introduce the Í
′ vector sub-
space Aform 𝑬 𝐶℘ ℘∈𝔉 of Aform 𝑬 consisting of the series 𝒉 (𝜆) = ∞ −𝑗

𝑗=0 𝒂 𝑗 𝜆
that satisfy
(19.1.18) for some 𝑀℘ ≥ 0; we define the following seminorm on
Aform 𝑬 𝐶℘ ℘∈𝔉 :

1 −𝑗
e (𝒉) = sup
℘ 𝐶℘ ℘ 𝒂 𝑗 . (19.1.19)
𝑗 ∈Z+ 𝑗!

We equip Aform 𝑬 𝐶℘ ℘∈𝔉 with the topology defined by the seminorms ℘, e ℘ ∈ 𝔉.

The following facts are self-evident: each TVS Aform 𝑬 𝐶℘ ℘∈𝔉 is Hausdorff and

Aform 𝑬 is the union of all the subspaces Aform 𝑬 𝐶℘ ℘∈𝔉 as we let the map
𝔉 ∋ ℘ ↦→ 𝐶℘ vary freely. We define the strong locally convex topology on Aform 𝑬
of Aform
by requiring that a convex subset 𝑬 is open if and only if its intersection
with every subspace Aform 𝑬 𝐶℘ ℘∈𝔉 is open in it.

If 𝑬 is a topological algebra, meaning that the multiplication is a continuous

bilinear map 𝑬 × 𝑬 −→ 𝑬, then Aform 𝑬 is also an algebra with respect to coeffi-
cientwise multiplication and the strong locally convex topology turns Aform 𝑬 into a
topological algebra.
We can generalize Definition 19.1.7:
Definition
Í∞ 19.1.21 By a finite realization of the formal analytic series 𝒉 (𝜆) =
𝒂 −𝑗 ∈ A
𝑗=0 𝑗 𝜆 form 𝑬 we mean a finite series
∑︁
𝒉 (𝑅) (𝜆) = 𝒂 𝑗 𝜆− 𝑗 . (19.1.20)
𝑗 ≤𝜆/𝑅

Lemmas 19.1.8 and 19.1.9 can be generalized to the present set-up, simply by
replacing the seminorm O (Ω) ∋ ℎ ↦→ max |ℎ (𝑧)| by an arbitrary continuous semi-
𝐾
norm on 𝑬.
774 19 Classical Analytic Formalism

Lemma 19.1.22 Let 𝒉 = ∞ −𝑗 ∈ A

Í
𝑗=0 𝒂 𝑗 𝜆 form 𝑬. To every continuous seminorm ℘
on 𝑬 let 𝑀℘ , 𝐶℘ be the constants in (19.1.18). If 𝑅 ≥ 𝐶℘ then
∑︁
𝜆 − 𝑗 ℘ 𝒂 𝑗 ≤ 𝑀℘ ;

∀𝜆 ≥ 1,
𝑗 ≤𝜆/𝑅

and if 𝑅 ′ > 𝑅 ≥ 𝐶℘ then

[𝜆/𝑅]
∑︁ √︁ ′
𝜆− 𝑗 ℘ 𝒂 𝑗 ≲ 𝑀℘ 1 + 𝜆/𝑅e−𝜆/𝑅 .

∀𝜆 ≥ 1,
𝑗=[𝜆/𝑅′ ]

If we apply the above to 𝑬 = C 𝜔 (𝑈), 𝑈 an open subset 𝑈 of R𝑛 , we get the

analogue of Definition 19.1.1 in real space.
Definition 19.1.23 By a formal analytic series in an open subset 𝑈 of R𝑛 we shall
mean a formal series ℎ (𝑥, 𝜆) = ∞ − 𝑗 having the following property:
Í
𝑗=0 ℎ 𝑗 (𝑥) 𝜆
Every coefficient ℎ 𝑗 extends holomorphically to one and the same open subset
𝑈 C of C𝑛 , 𝑈 = 𝑈 C ∩ R𝑛 , so that ℎ (𝑧, 𝜆) = ∞ −𝑗 ∈ A
Í
𝑗=0 ℎ 𝑗 (𝑧) 𝜆 form 𝑈 .
C

We shall denote by Aform (𝑈) the ring of formal analytic series in 𝑈. By the sheaf
Aform (R𝑛 ) of germs of formal analytic series in R𝑛 we shall mean the pullback (or
restriction) to R𝑛 of the sheaf Aform (C𝑛 ) (see Subsection 1.1).
We can take 𝑬 = O ′ (Ω) with Ω an open subset of C𝑛 as before; then Aform 𝑬 is the
vector space of formal power series 𝜇 = ∞ −𝑗 ′
Í
𝑗=0 𝜆 𝜇 𝑗 with coefficients 𝜇 𝑗 ∈ O (C )
𝑛

that satisfy the following condition:

(GFA∗ ) There exist a compact subset 𝐾 of Ω and, to every 𝜀 > 0 such that
𝐾 𝜀 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝐾) ≤ 𝜀} ⊂ Ω, positive constants 𝐶 𝜀 , 𝑅 𝜀 such that
𝑗
∀ℎ ∈ O (Ω) , 𝜇𝑗, ℎ ≤ 𝐶 𝜀 𝑅 𝜀 𝑗!max |ℎ (𝑧)| . (19.1.21)
𝑧 ∈𝐾 𝜀

This leads to the following

Definition 19.1.24 By a formal analytic functional series in Ω we mean a series
belonging to Aform 𝑬, 𝑬 = O ′ (Ω). We shall denote by Oform
′ (Ω) the vector space of
formal analytic functional series in Ω.
′
There is a natural locally convex topology on Oform (Ω) that makes it isomorphic
2 ′

to O∞ S ; O Í(Ω) (cf. Remark 19.1.20).
Let ℎ = ∞ −𝑗 Í∞ − 𝑗
𝑗=1 𝜆 ℎ 𝑗 ∈ A form (Ω) and 𝜇 = 𝑗=0 𝜆 𝜇 𝑗 with 𝜇 𝑗 satisfying
(19.1.21). Consider the bracket
∞
∑︁ 𝑚
∑︁
⟨𝜇, ℎ⟩ = 𝜆−𝑚 𝜇 𝑗 , ℎ 𝑚− 𝑗 . (19.1.22)
𝑚=0 𝑗=0

We apply (19.1.1) and (19.1.21) [also (17.2.21)]: there is a 𝐶 > 0 such that, for all
𝑗 ∈ Z+ ,
19.1 Formal Analytic Series 775

𝑗
∑︁
𝜇 𝑘 , ℎ 𝑗−𝑘 ≤ 𝐶 𝑗+1 𝑗!. (19.1.23)
𝑘=0

It makes sense to shorten Aform C to Í

Aform ; Aform can be regarded as the subring
of Aform (Ω) consisting of the series ∞ 𝑗=0 𝑎 𝑗 𝜆
− 𝑗 with constant coefficients; thus

𝑎 𝑗 ≤ 𝐶 𝑗+1 𝑗! for some 𝐶 > 0 and all 𝑗 ∈ Z+ ; in other words, these series are
restrictions to some half-line (𝑅, +∞) of functions defined and holomorphic in a
neighborhood of ∞ in the Riemann sphere: we can identify 2
2
Aform with O∞ S and
transfer to Aform the locally convex topology of O∞ S (cf. Remark 19.1.5).

Proposition 19.1.25 The bracket (19.1.22) defines a separately continuous bilinear

′
map Aform (Ω) × Oform (Ω) −→ Aform .

Remark 19.1.26 The bracket (19.1.22) is not a duality bracket in the usual sense: it
′
is not a bilinear map of Aform (Ω) × Oform (Ω) into the scalar field, here C. The linear
′
space Oform (Ω) is not the dual of the locally convex space Aform (Ω) (cf. Remark
19.1.5).

From analytic functionals we proceed, in straightforward fashion, to hyperfunc-

tions and thence to microfunctions (on these concepts see Section 7.3).

Definition 19.1.27 By a formal hyperfunction series in an open subset 𝑈 of R𝑛

we shall mean a formal power series 𝑓 = ∞ − 𝑗 𝑓 whose coefficients 𝑓 ∈ B (𝑈)
Í
𝑗=0 𝜆 𝑗 𝑗
satisfy the following condition:
(AHF) Given
Í∞ −arbitrarily a relatively compact open subset 𝑈 ′ of 𝑈 there is a 𝜇 =
′ ′
𝑗=0 𝜆 𝜇 𝑗 ∈ Oform (𝑈) whose coefficients 𝜇 𝑗 ∈ O (R ) have the following
𝑗 𝑛

′
properties: 1) supp 𝜇 ⊂ 𝑈 ; 2) 𝜇 𝑗 represents 𝑓 𝑗 in 𝑈 .′

We shall denote by Bform (𝑈) the vector space of all formal hyperfunction series
in 𝑈.

We may now deal with the sheaf Bform (R𝑛 ) of germs of formal hyperfunction
series in R𝑛 defined by the presheaf Bform (𝑈) , 𝜌𝑈 𝑉 , where 𝑈, 𝑉 are open subsets
of R such that 𝑉 ⊂ 𝑈 and 𝜌Í 𝑉 : Bform (𝑈) −→ Bform (𝑉) is the restriction map
𝑛 𝑈

applied coefficientwise to 𝑓 = ∞ 𝑗=0 𝜆

− 𝑗 𝑓 , 𝑓 ∈ B (𝑈).
𝑗 𝑗

Definition 19.1.28 By a formal singularity hyperfunction series in an open subset

· ·
−𝑗
𝑈 of R𝑛 we shall mean a formal power series 𝑓 = ∞
Í
𝑗=0 𝜆 𝑗 𝑓 𝑗 whose coefficients
·
𝑓 𝑗 ∈ B sing (𝑈) (Definition 11.3.4) satisfy the following condition: for every 𝑗 ∈ Z+
·
there is an 𝑓 𝑗 ∈ B (𝑈) in the coset 𝑓 𝑗 such that ∞ −𝑗 𝑓 ∈ B
Í
𝑗=0 𝜆 𝑗 form (𝑈). We shall
sing
denote by Bform (𝑈) the vector space of formal singularity hyperfunction series in
𝑈.
776 19 Classical Analytic Formalism
sing
Through the presheaf on R𝑛 defined by the linear spaces Bform (𝑈) and the natural
sing sing
restriction maps Bform (𝑈) −→ Bform (𝑉) (𝑉 ⊂ 𝑈) we can define the sheaf of germs
sing
of formal singularity hyperfunction series in R𝑛 , Bform (R𝑛 ); the latter can be
identified with the quotient sheaf Bform (R ) /Aform (R ).
𝑛 𝑛

Lastly we come to the following

Definition 19.1.29 Let 𝑈 be an open subset of R𝑛 and Σ an open subset of S𝑛−1 .

By a formal microfunction series in 𝑈 × Σ we shall mean a formal power series
[𝑓] = ∞ −𝑗 𝑓 micro (𝑈 × Σ) satisfy the follow-
Í
𝑗=0 𝜆 𝑗 whose coefficients 𝑓 𝑗 ∈ B
ing condition: there is a formal hyperfunction series 𝑓 = ∞ − 𝑗 𝑓 in 𝑈 such
Í
𝑗=0 𝜆 𝑗
that 𝑓 𝑗 ∈ B (𝑈) represents 𝑓 𝑗 for each 𝑗 (cf. Definition 11.3.11). We denote by
micro (𝑈 × Σ) the linear space of formal microfunction series in 𝑈 × Σ.
Bform

We will make use of the sheaf Bmicroform (R ) of germs at points of R × S

𝑛 𝑛 𝑛−1

(identified with the cosphere bundle of R ) of formal microfunction series.

𝑛

19.2 Classical Analytic Differential Operators of Infinite Order

19.2.1 Classical analytic differential operators of infinite order.

Definition. Action on formal analytic series

As before, Ω is an open subset of C𝑛 .

Definition 19.2.1 By a classical analytic differential operator of infinite order in

Ω we shall mean a formal power series
∑︁
𝐿 (𝑧, D, 𝜆) = 𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 ∈ Aform (Ω) [[D1 , ..., D𝑛 ]]
𝛼∈Z+𝑛

Í∞ −𝑗𝑐
where the formal analytic series 𝑐 𝛼 (𝑧, 𝜆) = 𝑗=0 𝜆 𝛼, 𝑗 (𝑧) satisfy the following
condition:
(D 𝜔 ) To every compact subset 𝐾 of Ω there is a 𝐶𝐾 > 0 such that
| 𝛼 |+ 𝑗
∀𝛼 ∈ Z+𝑛 , ∀ 𝑗 ∈ Z+ , max 𝑐 𝛼, 𝑗 (𝑧) ≲ 𝐶𝐾 𝑗!. (19.2.1)
𝑧 ∈𝐾

We shall denote by 𝔇𝔦𝔣𝔣∞ class (Ω) the linear space of classical analytic differential
operators of infinite order in Ω.

Strictly speaking, in Definition 19.2.1 D 𝑗 can be viewed as an indeterminate but

√
we will want to interpret it in the standard manner, D 𝑗 = − −1 𝜕𝑧𝜕 𝑗 . This is how we
let 𝐿 (𝑧, D, 𝜆) act on Aform (Ω).
19.2 Classical Analytic Differential Operators of Infinite Order 777

𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D𝑧𝛼 is
Í
Remark 19.2.2 The differential operator 𝐿 (𝑧, D𝑧 , 𝜆) = 𝛼∈Z+𝑛
completely determined by its action on O (Ω) since
1
𝑐 𝛼 (𝑧, 𝜆) = (𝑖𝜆) | 𝛼 | 𝐿 (𝑧, D𝑧 , 𝜆) (𝑧 − 𝑤) 𝛼 | 𝑤=𝑧 .
𝛼!
Hyperdifferential operators in Ω (Definition 8.1.14) are special classical analytic
differential operators of infinite order in Ω.
Proposition 19.2.3 Every classical analytic differential operator of infinite order in
Ω defines a linear map of Aform (Ω) into itself.
Proof
Í∞ −Let 𝐿 (𝑧, D, 𝜆) ∈ Aform (Ω) [[D1 , ..., D𝑛 ]] satisfy (D 𝜔 ) and let ℎ (𝑧, 𝜆) =
𝑗=0 𝜆 ℎ 𝑗 (𝑧) ∈ A form (Ω). We have directly
𝑗

∞
∑︁ ∑︁
𝐿 (𝑧, D, 𝜆) ℎ = 𝜆− 𝑗−𝑘−| 𝛼 | 𝑐 𝛼, 𝑗 (𝑧) D 𝛼 ℎ 𝑘 (𝑧) .
𝛼∈Z+𝑛 𝑗,𝑘=0

Let 𝐾 be a compact subset of Ω. From (FA) (Definition 19.1.1) and the Cauchy
inequalities we derive that there is a constant 𝑀𝐾 such that

∀𝛼 ∈ Z+𝑛 , ∀𝑘 ∈ Z+ , max |D 𝛼 ℎ 𝑘 (𝑧)| ≲ 𝑀𝐾𝑘+| 𝛼 | 𝑘!𝛼!.

𝑧 ∈𝐾

Combining this with (19.2.1) yields

∀𝛼 ∈ Z+𝑛 , ∀𝑘 ∈ Z+ , max 𝑐 𝛼, 𝑗 (𝑧) D 𝛼 ℎ 𝑘 (𝑧) ≲ (𝐶𝐾 𝑀𝐾 ) | 𝛼 |+ 𝑗+𝑘 𝑗!𝑘!𝛼!.

𝑧 ∈𝐾

Since, by (17.2.21),
∑︁ 𝑗!𝑘!𝛼!
≤ 3𝑛+1
𝑚!
𝑗+𝑘+| 𝛼 |=𝑚

we conclude that, for some 𝐵 𝐾 > 0,

∑︁
∀𝑚 ∈ Z+ , max 𝑐 𝛼, 𝑗 (𝑧) D 𝛼 ℎ 𝑘 (𝑧) ≲ 𝐵 𝐾
𝑚
𝑚!. □
𝑧 ∈𝐾
𝑗+𝑘+| 𝛼 |=𝑚

Remark 19.2.4 The endomorphism of Aform (Ω) defined by 𝐿 (𝑧, D, 𝜆) is continu-

ous with respect to the natural locally convex topology on Aform (Ω) (cf. Remark
19.1.5). This is easily seen by a closer examination of the estimates in the proof of
Proposition 19.2.3 – or by applying
the
closed graph theorem since the natural
in-
jection of Aform (Ω) into O (Ω) 𝜆−1 is trivially continuous when O (Ω) 𝜆−1

is equipped with the topology of convergence of the individual coefficients.

We can now view 𝔇𝔦𝔣𝔣∞
class (Ω) as a linear subspace of the ring of linear operators
on Aform (Ω).
Proposition 19.2.5 Classical analytic differential operators of infinite order in Ω
form a ring with respect to composition.
778 19 Classical Analytic Formalism

Proof Let 𝐿 ℓ (𝑧, D, 𝜆) = 𝛼∈Z+𝑛 𝑎 ℓ, 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 ∈ 𝔇𝔦𝔣𝔣∞

Í
class (Ω), ℓ = 1, 2, ℎ ∈
Aform (Ω). The Leibniz rule entails
∑︁
𝐿 1 (𝑧, D, 𝜆) 𝐿 2 (𝑧, D, 𝜆) ℎ = 𝜆− | 𝛼+𝛽 | 𝑎 1, 𝛼 D 𝛼 𝑎 2,𝛽 D𝛽 ℎ
𝛼,𝛽 ∈Z+𝑛
∑︁ ∑︁ 𝛼!
𝜆− | 𝛼+𝛽 | 𝑎 1, 𝛼 D 𝛼−𝛾 𝑎 2,𝛽 D𝛽+𝛾 ℎ,

=
𝛼,𝛽 ∈Z+𝑛 𝛾 ⪯ 𝛼
𝛾! (𝛼 − 𝛾)!

where 𝛾 ⪯ 𝛼 means 𝛾 𝑗 ≤ 𝛼 𝑗 , 𝑗 = 1, ..., 𝑛. With a change of summation indices we

get ∑︁
𝐿 1 (𝑧, D, 𝜆) 𝐿 2 (𝑧, D, 𝜆) ℎ = 𝜆− |𝜅 | 𝑏 𝜅 (𝑧, 𝜆) D 𝜅 ℎ,
𝜅 ∈Z+𝑛

where ∑︁ ∑︁ 𝛼!
𝑏 𝜅 (𝑧, 𝜆) = 𝜆− | 𝛼−𝛾 | 𝑎 1, 𝛼 D 𝛼−𝛾 𝑎 2,𝛽 .
𝛼∈Z+𝑛 𝛾 ⪯𝜅 ,𝛾 ⪯ 𝛼
𝛾! (𝛼 − 𝛾)!
Í∞ −𝑚 𝑏
It remains to prove that the formal analytic series 𝑏 𝜅 (𝑧, 𝜆) = 𝑚=0 𝜆 𝜅 ,𝑚 (𝑧)
satisfy (D 𝜔 ). We have
∞
∑︁ ∑︁ ∑︁ 𝛼!
𝑏 𝜅 (𝑧, 𝜆) = 𝜆− | 𝛼−𝛾 |− 𝑗−𝑘 𝑎 1, 𝛼, 𝑗 (𝑧) D 𝛼−𝛾 𝑎 2,𝜅−𝛾,𝑘 (𝑧) .
𝛼∈Z+𝑛 𝛾 ⪯𝜅 ,𝛾 ⪯ 𝛼 𝑗,𝑘=0
𝛾! (𝛼 − 𝛾)!

If we take into account the hypothesis (D 𝜔 ) for 𝐿 1 (𝑧, D, 𝜆) and 𝐿 2 (𝑧, D, 𝜆) as well
as the Cauchy inequalities we see that to every compact subset 𝐾 of Ω there is a
constant 𝐶𝐾 such that
∑︁ 𝛼! |𝜅 |+2 | 𝛼−𝛾 |+ 𝑗+𝑘
max 𝑏 𝜅 ,𝑚 (𝑧, 𝜆) ≲ 𝑗!𝑘!𝐶𝐾 .
𝑧 ∈𝐾 𝛾!
| 𝛼−𝛾 |+ 𝑗+𝑘=𝑚

Exploiting once again (17.2.21) it is not difficult to see that there is a 𝐵 𝐾 > 0
𝑚+ |𝜅 |
independent of 𝜅 and 𝑚 such that max 𝑏 𝜅 ,𝑚 (𝑧, 𝜆) ≲ 𝐵 𝐾 𝑚!. □
𝑧 ∈𝐾

If now 𝜁 𝑗 ( 𝑗 = 1, ..., 𝑛) are complex variables we may refer to the power series
∑︁
𝐿 (𝑧, 𝜁, 𝜆) = 𝜆− | 𝛼 | 𝑐 𝛼 (𝑧, 𝜆) 𝜁 𝛼 ∈ Aform (Ω) [[𝜁1 , ..., 𝜁 𝑛 ]] (19.2.2)
𝛼∈Z+𝑛

as the (total) symbol of 𝐿 (𝑧, D, 𝜆). The symbol of the composite

𝐿 1 (𝑧, D, 𝜆) 𝐿 2 (𝑧, D, 𝜆)

is
∑︁ 1
𝐿 1 (𝑧, 𝜁, 𝜆) #𝐿 2 (𝑧, 𝜁, 𝜆) = D 𝜁𝛼 𝐿 1 (𝑧, 𝜁, 𝜆) 𝜕𝑧𝛼 𝐿 2 (𝑧, 𝜁, 𝜆) (19.2.3)
𝛼∈Z𝑛
𝛼!
+
19.2 Classical Analytic Differential Operators of Infinite Order 779

since this is true for differential operators of finite order.

The next statement is a direct consequence of the Leibniz rule and the Cauchy
inequalities.

Proposition 19.2.6 If 𝐿 (𝑧, D𝑧 , 𝜆) = 𝛼∈Z+𝑛 𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 ∈ 𝔇𝔦𝔣𝔣∞

Í
class (Ω) then
its transpose 𝐿 (𝑧, D𝑧 , 𝜆) ⊤ , defined by the standard formula
∑︁
𝐿 (𝑧, D𝑧 , 𝜆) ⊤ ℎ (𝑧, 𝜆) = (−1) | 𝛼 | 𝜆− | 𝛼 | D𝑧𝛼 (𝑐 𝛼 (𝑧, 𝜆) ℎ (𝑧)) , ℎ ∈ O (Ω) ,
𝛼∈Z+𝑛

also belongs to 𝔇𝔦𝔣𝔣∞

class (Ω) .

To define finite realizations of a differential operator 𝐿 (𝑧, D, 𝜆) ∈ 𝔇𝔦𝔣𝔣∞

class (Ω) is
simple: it suffices to regard its symbol (19.2.2) as a formal analytic series in Ω × C𝑛
and use its finite realizations (Definition 19.1.7)
∑︁
𝐿 (𝑅) (𝑧, 𝜁, 𝜆) = 𝜆− | 𝛼 |− 𝑗 𝑐 𝛼, 𝑗 (𝑧) 𝜁 𝛼 (19.2.4)
( 𝛼, 𝑗) ∈Z+𝑛+1
| 𝛼 |+ 𝑗 ≤𝜆/𝑅

where 𝑅 > 0 is large. Evidently, for fixed 𝜆 ≥ 1, 𝐿 (𝑅) (𝑧, D, 𝜆) is a differential

operator of finite order with holomorphic coefficients in Ω. Its natural domain of
action is the space Ob (Ω) (Definition 19.1.11).

Proposition 19.2.7 To every compact subset 𝐾 of Ω and to every 𝜀 > 0 such that

𝐾 𝜀 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝐾) ≤ 𝜀} ⊂ Ω

> 0 such that 𝑅 > max 𝐶𝐾 , 𝜀 −1 implies, for all ℎ ∈ Ob (Ω), 𝜆 ≥ 1

there is a 𝐶𝐾
and 𝑧 ∈ 𝐾,
𝐿 (𝑅) (𝑧, D, 𝜆) ℎ (𝑧, 𝜆) ≤ 𝐶𝐾 max |ℎ (𝑧, 𝜆)| . (19.2.5)
𝑧 ∈𝐾 𝜀

Proof Let ℎ ∈ Ob (Ω); we apply the Cauchy inequalities:

D𝑧𝛼 ℎ (𝑧, 𝜆) ≤ 𝛼!𝜀 −| 𝛼 | max |ℎ (𝑧, 𝜆)| .

𝑧 ∈𝐾 𝜀

Taking (19.2.1) and (19.2.4) into account we get, for 𝑧 ∈ 𝐾,

∑︁
𝐿 (𝑅) (𝑧, D𝑧 , 𝜆) ℎ (𝑧, 𝜆) ≤ 𝜆− | 𝛼 |− 𝑗 𝑐 𝛼, 𝑗 (𝑧) D𝑧𝛼 ℎ (𝑧, 𝜆)
( 𝛼, 𝑗) ∈Z+𝑛+1
| 𝛼 |+ 𝑗 ≤𝜆/𝑅
∑︁
| 𝛼 |+ 𝑗 −| 𝛼 |
≤ 𝐶𝐾 𝐶𝐾 𝜀 𝑗!𝛼!𝜆− | 𝛼 |− 𝑗 max |ℎ (𝑧, 𝜆)| .
𝑧 ∈𝐾 𝜀
( 𝛼, 𝑗) ∈Z+𝑛+1
| 𝛼 |+ 𝑗 ≤𝜆/𝑅

We exploit Stirling’s inequality as in the proof of Lemma 19.1.8:

780 19 Classical Analytic Formalism

∞
1 ∑︁ © ∑︁ 𝛼! 𝑗! ª √
ℓ
∑︁
| 𝛼 |+ 𝑗 𝐶𝐾
𝐶𝐾 𝜀 − | 𝛼 | 𝑗!𝛼!𝜆− | 𝛼 |− 𝑗 ≤ ­ ® ℓ+1 .
2 ℓ=0 ℓ! e𝑅𝜀
( 𝛼, 𝑗) ∈Z+𝑛+1 « | 𝛼 |+ 𝑗=ℓ ¬
| 𝛼 |+ 𝑗 ≤𝜆/𝑅

If 𝑅 ≥ 𝑅𝐾 , 𝜀 > max 𝐶𝐾 , 𝜀 −1 then, by Lemma 17.2.14,

𝐿 (𝑅) (𝑧, D𝑧 , 𝜆) ℎ (𝑧, 𝜆) ≤ 3𝑛 𝐶𝐾 max |ℎ (𝑧, 𝜆)| . □

𝑧 ∈𝐾 𝜀

Corollary 19.2.8 To every compact subset 𝐾 of Ω there is an 𝑅𝐾 > 0 such that the
following is true, if 𝑅 > 𝑅𝐾 : to each ℎ ∈ O (−𝜔) (Ω) there are positive constants 𝐶,
𝜅 such that
∀𝜆 ≥ 1, max 𝐿 (𝑅) (𝑧, D, 𝜆) ℎ (𝑧, 𝜆) ≤ 𝐶e−𝜅𝜆 .
𝑧 ∈𝐾

Proof Combine (19.2.5) with (19.1.7). □

In other words, if Ω′ ⊂⊂ Ω is an open set and 𝑅 is sufficiently large then
𝐿 (𝑅) (𝑧, D, 𝜆) ℎ ∈ O (−𝜔) (Ω′) for every ℎ (𝑧, 𝜆) ∈ O (−𝜔) (Ω).
The same sort of argument, now mimicking the proof of Lemma 19.1.9, enables
us to prove

Proposition 19.2.9 To every compact subset 𝐾 of Ω there is an 𝑅𝐾 > 0 such that

the following is true, if 𝑅 ′ > 𝑅 > 𝑅𝐾 : to each ℎ ∈ Ob (Ω) there is a 𝐶 > 0 such that
′ ′
∀𝜆 ≥ 1, max 𝐿 (𝑅) (𝑧, D, 𝜆) ℎ (𝑧, 𝜆) − 𝐿 (𝑅 ) (𝑧, D, 𝜆) ℎ (𝑧, 𝜆) ≤ 𝐶e−𝜆/𝑅 .
𝑧 ∈𝐾

From Propositions 19.2.7 and 19.2.9 we derive the following (cf. Subsection
19.1.2): select arbitrarily aÐsequence of relatively compact subsets of Ω, Ω𝜈 ⊂⊂
∞
Ω𝜈+1 (𝜈 = 1, 2, ...), Ω = ∞ 𝜈=1 Ω 𝜈 . For each 𝐿 (𝑧, D, 𝜆) ∈ 𝔇𝔦𝔣𝔣 class (Ω) we can
select a sequence of positive numbers 𝑅 𝜈 < 𝑅 𝜈+1 such that, for every ℎ ∈ Ob (Ω),
𝐿 (𝑅𝜈 ) (𝑧, D, 𝜆) ℎ Ω𝜈 ∈ Ob (Ω𝜈 ), and

𝐿 (𝑅𝜈+1 ) (𝑧, D, 𝜆) ℎ − 𝐿 (𝑅𝜈 ) (𝑧, D, 𝜆) ℎ ∈ O (−𝜔) (Ω𝜈 ) .

Ω𝜈 Ω𝜈

In analogy with what was done at the end of Subsection 1 we can introduce the
eb (Ω) of the quotient linear spaces Ob (Ω𝜈 ) /O (−𝜔) (Ω𝜈 ) and conclude
direct limit O
that 𝐿 (𝑧, D, 𝜆) defines a linear operator on O
eb (Ω); different choices of the Ω𝜈 and
𝑅 𝜈 yield equivalent definitions of such an operator, which we shall also denote by
𝐿 (𝑧, D, 𝜆).
In the inward direction we let the Ω𝜈 form a basis of neighborhoods of a point
𝑧◦ ∈ Ω, Ω𝜈+1 ⊂⊂ Ω𝜈 ⊂⊂ Ω0 ⊂⊂ Ω. As described at the end of Subsection 1 this
enables us to select numbers 𝑅 𝜈 such that the finite realizations 𝐿 (𝑅𝜈 ) (𝑧, D, 𝜆) define
a germ linear operator

L𝑧 ◦ (𝑧, D, 𝜆) : Ob,𝑧 ◦ /O𝑧(−𝜔)

◦ −→ Ob,𝑧 ◦ /O𝑧(−𝜔)
◦ . (19.2.6)
19.2 Classical Analytic Differential Operators of Infinite Order 781

Changes in the choices of the Ω𝜈 and 𝑅 𝜈 do not modify this definition. As 𝑧◦ ranges
over Ω, the (noncommutative) algebras (with respect to addition and composition)
of germ operators L𝑧 ◦ (𝑧, D, 𝜆) define a sheaf of algebras Diff ∞
a (Ω). The continuous
sections of Diff ∞
a (Ω) define sheaf maps of O b /O (−𝜔) into itself.

19.2.2 Elliptic classical analytic differential operators of infinite order

Definition 19.2.10 We shallÍsay that Í the classical analytic differential operator of

infinite order 𝐿 (𝑧, D, 𝜆) = ∞ 𝑗=0 𝛼∈Z+ 𝑛 𝑐 𝛼, 𝑗 (𝑧, 𝜆) 𝜆 − 𝑗−| 𝛼 | D 𝛼 is elliptic at a point

𝑧◦ ∈ Ω if 𝑐 0,0 (𝑧 ◦ ) ≠ 0. We shall say that 𝐿 (𝑧, D, 𝜆) is elliptic in an open subset Ω′

of Ω if it is elliptic at every point of Ω′.

If 𝐿 (𝑧, D, 𝜆) is elliptic at 𝑧◦ it is elliptic in a neighborhood of 𝑧◦ .

Theorem 19.2.11 If 𝐿 (𝑧, D, 𝜆) ∈ 𝔇𝔦𝔣𝔣∞ class (Ω) is elliptic in Ω then there is an

𝐿 −1 (𝑧, D, 𝜆) ∈ 𝔇𝔦𝔣𝔣∞
class (Ω) such that

𝐿 −1 (𝑧, D, 𝜆) 𝐿 (𝑧, D, 𝜆) = 𝐿 (𝑧, D, 𝜆) 𝐿 −1 (𝑧, D, 𝜆) = 1.

Proof Follows directly from Theorem 19.1.17 and Proposition 19.2.5. □

Corollary 19.2.12 Every elliptic classical analytic differential operator of infinite

order in Ω defines a linear automorphism of Aform (Ω).

Proof Combine Lemma 19.1.8 with Theorem 19.2.11. □

Likewise we can state

Theorem 19.2.13 If 𝐿 (𝑧, D, 𝜆) ∈ 𝔇𝔦𝔣𝔣∞ ◦

class (Ω) is elliptic at 𝑧 then the germ map
(19.2.6) is an automorphism.

The continuous sections of the sheaf Diff ∞

a (Ω) elliptic at every point of Ω define
sheaf automorphisms of Ob /O (−𝜔) .

19.2.3 Action on formal analytic functional series and on formal

microfunction series

In this subsection we assume that the domain Ω is Runge. Í Consider a classical ana-
lytic differential operator of infinite order 𝐿 (𝑧, D𝑧 , 𝜆) = 𝛼∈Z+𝑛 𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 ∈
𝔇𝔦𝔣𝔣∞ 𝜆) ⊤ be its transpose (cf. Proposition 19.2.6). For arbi-
class (Ω) and let 𝐿 (𝑧, D 𝑧 , Í
trary ℎ ∈ Aform (Ω) and 𝜇 = ∞ −𝑗 ′
𝑗=0 𝜆 𝜇 𝑗 ∈ Oform (Ω) (Definitions 19.1.1, 19.1.18)
we define 𝐿 (𝑧, D𝑧 , 𝜆) 𝜇 by the formula

⟨𝐿 (𝑧, D𝑧 , 𝜆) 𝜇, ℎ⟩ = 𝜇, 𝐿 (𝑧, D𝑧 , 𝜆) ⊤ ℎ (19.2.7)

782 19 Classical Analytic Formalism

where the brackets are as per (19.1.18). The right-hand side makes sense by Propo-
sitions 19.1.25, 19.2.3, 19.2.6 (cf. also Remarks 19.1.5, 19.1.26). The following
statement ensues:
Proposition 19.2.14 Every differential operator 𝐿 (𝑧, D𝑧 , 𝜆) ∈ 𝔇𝔦𝔣𝔣∞
class (Ω) defines
a continuous endomorphism of the topological vector space Oform′ (Ω).
′
We denote by Oform ′
(Ω ∩ R𝑛 ) the subspace of Oform (Ω) consisting of the formal
Í∞ − 𝑗
analytic functional series 𝜇 = 𝑗=0 𝜆 𝜇 𝑗 that satisfy (19.1.21) with 𝐾 ⊂ Ω ∩ R𝑛 .
′
We equip Oform (Ω ∩ R𝑛 ) with the weakest locally convex topology that renders the
′
embedding Oform ′
(Ω ∩ R𝑛 ) ↩→ Oform (Ω1 ) continuous for every domain Ω1 in C𝑛
such that Ω ∩ R𝑛 ⊂ Ω1 .
Proposition 19.2.15 Every differential operator 𝐿 (𝑧, D𝑧 , 𝜆) ∈ 𝔇𝔦𝔣𝔣∞ class (Ω) defines
an endomorphism of the topological vector space Oform ′ (Ω ∩ R𝑛 ).

Now consider 𝑓 = ∞ −𝑗 𝑓 ∈ B
Í
form (Ω ∩ R ) (Definition 19.1.23). Given
𝑛
𝑗=0 𝜆 𝑗
arbitrarily an open subset 𝑈 of R , 𝑈 ⊂⊂ Ω ∩ R , there is a 𝜇 = ∞ −𝑗
Í
𝑗=0 𝜆 𝜇 𝑗 ∈
𝑛 𝑛
′
Oform (Ω ∩ R ) representing 𝑓 in 𝑈, i.e., such that 𝜇 𝑗 represents 𝑓 𝑗 in 𝑈 for every
𝑛

𝑗. We define 𝐿 (𝑧, D𝑧 , 𝜆) 𝑓 as the formal hyperfunction series in Ω represented by

′
𝐿 (𝑧, D𝑧 , 𝜆) 𝜇 ∈ Oform (Ω ∩ R𝑛 ) in 𝑈 (cf. Proposition 19.2.15); this uniquely defines
𝐿 (𝑧, D𝑧 , 𝜆) 𝑓 in Ω (independently of the choice of the representative 𝜇 in each
𝑈 ⊂⊂ Ω∩R𝑛 ). In other words, every classical analytic differential operator of infinite
order 𝐿 (𝑧, D𝑧 , 𝜆) ∈ 𝔇𝔦𝔣𝔣∞class (Ω) defines a linear endomorphism of Bform (Ω ∩ R ).
𝑛

If we combine this last fact with Proposition 19.2.3 we reach the following
conclusion: if 𝑈 is an open subset of R𝑛 contained in the open subset Ω of C𝑛 ,
sing
𝐿 (𝑧, D𝑧 , 𝜆) defines an endomorphism of Bform (𝑈) (cf. Definition 19.1.28).
In turn, each one of the above maps defines stalk-preserving endomorphisms
of the corresponding sheaves. We shall denote by L 𝑥 ◦ = L 𝑥 ◦ (𝑧, D𝑧 , 𝜆) the linear
sing
operator on the stalk of Bform (R𝑛 ) [resp., Bform (R𝑛 )] at 𝑥 ◦ ∈ R𝑛 , Bform,𝑥 ◦ [resp.,
sing
Bform,𝑥 ◦ ].
Definition 19.2.16 The linear operators

L 𝑥 ◦ : Bform,𝑥 ◦ −→ Bform,𝑥 ◦
sing sing
[resp., Bform,𝑥 ◦ −→ Bform,𝑥 ◦ ]

will be referred to as germ classical analytic differential operators of infinite order

at 𝑥 ◦ ∈ R𝑛 acting on germs of formal hyperfunction (resp., singularity hyperfunction)
series at the same point (Definitions 19.1.23 and 19.1.28).

Í∞Let −Σ𝑗 be
an open subset of S . Consider a formal microfunction series [ 𝑓 ] =
𝑛−1

𝑗=0 𝜆 𝑓 𝑗 in 𝑈 × Σ (Definition 19.1.29) and let a formal hyperfunction series

𝑓 = ∞ − 𝑗 𝑓 in 𝑈 be such that 𝑓 ∈ B (𝑈) represents 𝑓
Í
𝑗=0 𝜆 𝑗 𝑗 𝑗 for each 𝑗. The series

∞
∑︁
𝐿 (𝑧, D𝑧 , 𝜆) 𝑓 = 𝜆− 𝑗 𝐿 (𝑧, D𝑧 , 𝜆) 𝑓 𝑗
𝑗=0
19.2 Classical Analytic Differential Operators of Infinite Order 783

belongs to Bform (𝑈) and represents an element of Bform micro (𝑈 × Σ) which we de-

note by 𝐿 (𝑧, D𝑧 , 𝜆) [ 𝑓 ]. In this manner 𝐿 (𝑧, D𝑧 , 𝜆) defines an endomorphism of

micro (𝑈 × Σ) and therefore also one, henceforth denoted by L ◦ ◦ , of the stalk of
Bform 𝑥 ,𝜃
Bform (R𝑛 ) at (𝑥 ◦ , 𝜃 ◦ ) ∈ R𝑛 ×S𝑛−1 , Bmicro
micro
◦
form,𝑥 , 𝜃 ◦ (cf. Definition 11.3.11 and subsequent
considerations).

Definition 19.2.17 The linear operators

L 𝑥 ◦ , 𝜃 ◦ : Bmicro micro
form,𝑥 ◦ , 𝜃 ◦ −→ Bform,𝑥 ◦ , 𝜃 ◦

will be referred to as germ classical analytic differential operators of infinite

order at (𝑥 ◦ , 𝜃 ◦ ) ∈ R𝑛 × S𝑛−1 acting on germs of formal microfunction series at the
same point (Definition 19.1.29).
micro,∞
We can regard L 𝑥 ◦ and L 𝑥 ◦ , 𝜃 ◦ as stalks of sheaves D∞
form (R ) and Dform
𝑛 (R𝑛 )
respectively. If action is needed not on germs at single points but in open subsets of
the base (either R𝑛 or R𝑛 × S𝑛−1 ) we would let continuous sections of the sheaves
micro,∞
D∞form (R ) and Dform
𝑛 (R𝑛 ) act on continuous sections of the sheaves Bform (R𝑛 )
micro
and Bform (R ) respectively.
𝑛

There are obvious embeddings B (𝑈) ↩→ Bform (𝑈) (𝑈 ⊂ R𝑛 open), B (R𝑛 ) ↩→

Bform (R𝑛 ) and Bmicro (R𝑛 ) ↩→ Bmicro form (R ): we simply identify an element in each
𝑛

one of the source sets with a formal series in the powers 𝜆− 𝑗 (in the corresponding
target set) consisting only of the zero power term. As a consequence of what has
been done above we have well-defined germ operators

L 𝑥 ◦ : B 𝑥 ◦ −→ Bform,𝑥 ◦ , (19.2.8)
L𝑥◦ , 𝜃 ◦ : Bmicro
𝑥◦ , 𝜃 ◦ −→ Bmicro
form,𝑥 ◦ , 𝜃 ◦ .

19.2.4 Conjugacy of classical analytic differential operators of infinite

order

Let 𝐿 (𝑧, D, 𝜆) = 𝛼∈Z+𝑛 𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 be a classical analytic differential opera-

Í
tor of infinite order in Ω and 𝜑 ∈ O (Ω). In this subsection we are interested in the
transformation

Aform (Ω) ∋ 𝑎 ↦→ 𝐿 ( 𝜑) (𝑧, D, 𝜆) 𝑎 = e−𝑖𝜆𝜑 𝐿 (𝑧, D, 𝜆) e𝑖𝜆𝜑 𝑎 . (19.2.9)

It is evident that if 𝐿 (𝑧, D) is a differential operator of finite order the same is true of
𝐿 ( 𝜑) (𝑧, D, 𝜆). The following example shows that, if 𝐿 (𝑧, D, 𝜆) is of infinite order,
𝐿 ( 𝜑) (𝑧, D, 𝜆) is not necessarily a classical analytic differential operator of infinite
order.
784 19 Classical Analytic Formalism

Example 19.2.18 Take Ω = C𝑛 and 𝜑 (𝑧) = 𝑧 · 𝜁, 𝜁 ∈ C𝑛 . We have

∑︁
𝐿 ( 𝜑) (𝑧, D, 𝜆) 𝑎 = 𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | e−𝑖𝜆𝑧·𝜁 D𝑧𝛼 e𝑖𝜆𝑧·𝜁 𝑎
𝛼∈Z+𝑛
∑︁ 𝛼
= 𝑐 𝛼 (𝑧, 𝜆) 𝜆−1 D𝑧 + 𝜁 𝑎
𝛼∈Z+𝑛
∑︁ ∑︁ 𝛼!
𝑐 𝛼 (𝑧, 𝜆) 𝜁 𝛼−𝛽 𝜆− |𝛽 | D𝑧 𝑎,
𝛽
=
𝛼∈Z+𝑛 𝛽⪯𝛼
𝛽! (𝛼 − 𝛽)!

where 𝛽 ⪯ 𝛼 means 𝛽 𝑗 ≤ 𝛼 𝑗 , 𝑗 = 1,Í ..., 𝑛. We make the following choice of the

coefficients of 𝐿 (𝑧, D, 𝜆), 𝑐 𝛼 (𝑧, 𝜆) = ∞ 𝑗=0 𝑗!𝜆
− 𝑗 [thus 𝑐
𝛼, 𝑗 (𝑧) = 𝑗!]; they satisfy
(19.1.1) trivially. We get
∑︁
𝐿 ( 𝜑) (𝑧, D, 𝜆) 𝑎 = 𝑐♭𝛽 𝜆−|𝛽 | D𝑧 𝑎,
𝛽

𝛽 ∈Z+𝑛

where
∞
∑︁ © ∑︁ (𝛼 + 𝛽)! 𝛼 ª − 𝑗
𝑐♭𝛽 = 𝑗! ­ 𝜁 ®𝜆 .
𝑗=0 𝛼∈Z 𝑛 𝛼!𝛽!
« + ¬
The series
∑︁ (𝛼 + 𝛽)! 1 𝛽 © ∑︁ 𝛼+𝛽 ª
𝜁 𝛼 = 𝜕𝜁 ­ 𝜁 ®
𝛼∈Z+𝑛
𝛽!𝛼! 𝛽! 𝑛
« 𝛼∈Z+ ¬
converges if and only if 𝜁 𝑗 < 1 for every 𝑗 = 1, ..., 𝑛.
In the general case the following can be said.
Proposition 19.2.19 To every open subset Ω′ ⊂⊂ Ω there is a 𝛿 = 𝛿 (𝐿, Ω′) > 0
such |𝜑 (𝑧) − 𝜑 (𝑤)| ≤ 𝛿 |𝑧 − 𝑤| for all (𝑧, 𝑤) ∈ Ω then 𝐿 ( 𝜑) (𝑧, D, 𝜆) =
Í that♭ if − | 𝛼 | D 𝛼 is a classical analytic differential operator of infinite order
𝛼∈Z+𝑛 𝑐 𝛼 (𝑧, 𝜆) 𝜆
′
in Ω .
Proof The Leibniz rule implies
1 ( 𝜑)
𝑐♭𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | = 𝐿 (𝑧, D𝑤 , 𝜆) (𝑤 − 𝑧) 𝛼
𝛼! 𝑤=𝑧
1
= 𝐿 (𝑧, D𝑤 , 𝜆) (𝑤 − 𝑧) 𝛼 e𝑖𝜆( 𝜑 (𝑤)−𝜑 (𝑧))
𝛼! 𝑤=𝑧
1 ∑︁
𝑐 𝛽 (𝑧, 𝜆) 𝜆− |𝛽 | D𝑤 (𝑤 − 𝑧) 𝛼 e𝑖𝜆( 𝜑 (𝑤)−𝜑 (𝑧))
𝛽
=
𝛼! 𝛽 ∈Z𝑛 𝑤=𝑧
+
𝛽⪰𝛼
∑︁ 𝛽!
𝑐 𝛽 (𝑧, 𝜆) 𝜆−|𝛽 | D𝑤 e𝑖𝜆( 𝜑 (𝑤)−𝜑 (𝑧))
𝛽−𝛼
= .
𝛽 ∈Z+𝑛
𝛼! (𝛽 − 𝛼)! 𝑤=𝑧
𝛽⪰𝛼
19.2 Classical Analytic Differential Operators of Infinite Order 785

Thus
∑︁ (𝛼 + 𝛽)!
𝜆−|𝛽 | 𝑐 𝛼+𝛽 (𝑧, 𝜆) D𝑤 e𝑖𝜆( 𝜑 (𝑤)−𝜑 (𝑧))
𝛽
𝑐♭𝛼 (𝑧, 𝜆) = . (19.2.10)
𝛽 ∈Z𝑛
𝛼!𝛽! 𝑤=𝑧
+

Let 𝐾 denote the closure of Ω′ and 𝑟 > 0 be sufficiently small that

Ø
𝐾𝑟 = Δ𝑟(𝑛) (𝑧) ⊂ Ω.
𝑧 ∈𝐾

In what follows 𝑧 ∈ 𝐾 and 𝑤 ∈ 𝐾𝑟 ; we apply the Cauchy formula:

1 𝛽 𝑖𝜆( 𝜑 (𝑤)−𝜑 (𝑧))
D e
𝛽! 𝑤 𝑤=𝑧

e𝑖𝜆( 𝜑 (𝑤)−𝜑 (𝑧))

∮ ∮
−𝑛
= (2𝑖𝜋) ··· Î𝑛 𝛽 𝑗 +1
d𝑤
|𝑤1 −𝑧1 |=𝑟 |𝑤𝑛 −𝑧𝑛 |=𝑟 𝑗=1 (𝑤 − 𝑧)
∑︁ 1
= (2𝑖𝜋) −𝑛 𝜆 𝑘 𝐸 𝛽,𝑘 (𝑧, 𝑟) ,
𝑘!
𝑘 ≤ |𝛽 |

where

(𝜑 (𝑤) − 𝜑 (𝑧)) 𝑘
∮ ∮
𝐸 𝛽,𝑘 (𝑧, 𝑟) = ··· Î𝑛 𝛽 𝑗 +1
d𝑤. (19.2.11)
|𝑤1 −𝑧1 |=𝑟 |𝑤𝑛 −𝑧𝑛 |=𝑟 𝑗=1 (𝑤 − 𝑧)

Note that if 𝑘 > |𝛽| then necessarily 𝐸 𝛽,𝑘 (𝑧, 𝑟) ≡ 0. Recalling that
∞
∑︁
𝑐 𝛼+𝛽 (𝑧, 𝜆) = 𝜆− 𝑗 𝑐 𝛼+𝛽, 𝑗 (𝑧)
𝑗=0

we obtain
∑︁ ∑︁ (𝛼 + 𝛽)!
𝑐♭𝛼 (𝑧, 𝜆) = (2𝑖𝜋) −𝑛 𝜆 𝑘− |𝛽 |− 𝑗 𝑐 𝛼+𝛽, 𝑗 (𝑧) 𝐸 𝛽,𝑘 (𝑧, 𝑟) . (19.2.12)
𝛽 ∈Z 𝑛 𝛼!𝑘!
+ 𝑘 ≤ |𝛽 |

Í∞ −𝑚 𝑐♭
If we write 𝑐♭𝛼 (𝑧, 𝜆) = 𝑚=0 𝜆 𝛼,𝑚 (𝑧) we derive
∞
∑︁ ∑︁ (𝛼 + 𝛽)!
𝑐♭𝛼,𝑚 (𝑧) = (2𝑖𝜋) −𝑛 𝑐 𝛼+𝛽,𝑚− |𝛽 |+𝑘 (𝑧) 𝐸 𝛽,𝑘 (𝑧, 𝑟) .
𝑘=0 𝛽 ∈Z+𝑛
𝛼!𝑘!
0≤ |𝛽 |−𝑘 ≤𝑚

By (19.2.1) there is a 𝐶𝐾 > 0 independent of 𝑚, 𝑘, 𝛼, 𝛽 (𝑘 ≤ |𝛽| ≤ 𝑚 + 𝑘) such

that
max 𝑐 𝛼+𝛽,𝑚−|𝛽 |+𝑘 (𝑧) ≤ 𝐶𝐾| 𝛼 |+𝑚+𝑘+1 (𝑚 + 𝑘 − |𝛽|)!.
𝑧 ∈𝐾

By our hypothesis on 𝜑 we deduce from (19.2.11), for all 𝑘 ≤ |𝛽|,

786 19 Classical Analytic Formalism

𝐸 𝛽,𝑘 (𝑧, 𝑟) 𝑤=𝑧

≤ 𝛿 𝑘 𝑟 𝑘−|𝛽 | .

Taking (19.2.12) and the last two estimates into account, as well as the inequality
(𝛼 + 𝛽)! ≤ 𝛼!𝛽!2 | 𝛼+𝛽 | , we derive
∞
∑︁ ∑︁ (𝛼 + 𝛽)! (𝑚 + 𝑘 − |𝛽|)! 𝑘 𝑘 𝑘−|𝛽 |
𝑐♭𝛼,𝑚 (𝑧) ≤ 𝐶𝐾| 𝛼 |+𝑚+1 𝐶𝐾 𝛿 𝑟
𝑘=0 𝛽 ∈Z+𝑛
𝛼!𝑘!
0≤ |𝛽 |−𝑘 ≤𝑚
∞
∑︁ ∑︁ (𝑚 + 𝑘 − |𝛽|)!𝛽! |𝛽 | 𝑘 𝑘
≤ 2 | 𝛼 | 𝑟 −𝑚 𝐶𝐾| 𝛼 |+𝑚+1 2 𝐶𝐾 𝛿 .
𝑘=0 𝛽 ∈Z+𝑛
𝑘!
0≤ |𝛽 |−𝑘 ≤𝑚

We apply once again Lemma 17.2.14:

𝑚
∑︁ (𝑚 + 𝑘 − |𝛽|)!𝛽! ∑︁ (𝑚 − ℓ)! (𝑘 + ℓ)! ∑︁ 𝛽!
=
𝛽 ∈Z+𝑛
𝑘! 𝑘! (𝑘 + ℓ)!
ℓ=0 |𝛽 |=𝑘+ℓ
0≤ |𝛽 |−𝑘 ≤𝑚
𝑚
∑︁ (𝑚 − ℓ)!ℓ! 𝑘+ℓ
≤ 3𝑛−1 𝑚! 2 ≤ 3𝑛 2𝑚+𝑘 𝑚!,
ℓ=0
𝑚!

whence
∞
∑︁
𝑐♭𝛼,𝑚 (𝑧) ≤ 3𝑛 22𝑚+| 𝛼 | 𝑟 −𝑚 𝐶𝐾| 𝛼 |+𝑚+1 𝑚! (8𝐶𝐾 𝛿) 𝑘 .
𝑘=0
−1
Selecting 𝛿 < (16𝐶𝐾 ) implies

∀𝛼 ∈ Z+𝑛 , ∀𝑚 ∈ Z+ , max 𝑐♭𝛼,𝑚 (𝑧) ≲ 𝑀𝐾| 𝛼 |+𝑚 𝑚!

𝑧 ∈𝐾

for some 𝑀𝐾 > 0. □

Corollary 19.2.20 If 𝑧◦ ∈ Ω is such that 𝜕𝜑Í(𝑧◦ ) = 0 then there is a neighborhood

Ω′ ⊂⊂ Ω of 𝑧◦ such that 𝐿 ( 𝜑) (𝑧, D, 𝜆) = 𝛼∈Z+𝑛 𝑐♭𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 is a classical
analytic differential operator of infinite order in Ω′.

Proposition 19.2.21 Suppose that 𝐿 (𝑧, D, 𝜆) = 𝛼∈Z+𝑛 𝑐 𝛼 (𝑧, 𝜆) 𝜆−| 𝛼 | D 𝛼 is elliptic

Í

in Ω (Definition 19.2.10) and that 𝛿 = sup | 𝜑 (𝑧)−𝜑 (𝑤) |

|𝑧−𝑤 | is sufficiently small for
𝑧,𝑤 ∈Ω
𝐿 ( 𝜑) (𝑧, D, 𝜆) = ∞ − 𝑗− | 𝛼 | D 𝛼 to be a classical analytic differ-
Í Í
𝑗=0 𝛼∈Z+𝑛 𝑐 𝛼, 𝑗 (𝑧, 𝜆) 𝜆
♭

ential operator of infinite order in Ω. Under these hypotheses 𝐿 ( 𝜑) (𝑧, D, 𝜆) is also

elliptic in Ω.

Proof We derive from (19.2.10) that 𝑐♭0,0 (𝑧, 𝜆) = 𝑐 0,0 (𝑧). □

Corollary 19.2.22 Under the same hypotheses as in Proposition 19.2.21,

𝐿 ( 𝜑) (𝑧, D, 𝜆) defines an automorphism of the vector space Aform (Ω).
19.2 Classical Analytic Differential Operators of Infinite Order 787

Proof Combine Proposition 19.2.21 and Theorem 19.2.11: given an arbitrary 𝑏 ∈

Aform (Ω), 𝑎 = (𝐿 𝜑 ) −1 (𝑧, D, 𝜆) 𝑏 satisfies 𝐿 ( 𝜑) (𝑧, D, 𝜆) 𝑎 = 𝑏 in Ω. □

19.2.5 Correspondence between classical symbols and classical analytic

differential operators of infinite order

Consider a classical symbol of order 𝑚 ∈ R in Ω × Ξ (Ξ ⊂ C𝑛 open)

∞
∑︁
𝑝 (𝑧, 𝜁, 𝜆) = 𝜆 𝑚− 𝑗 𝑝 𝑗 (𝑧, 𝜁) (19.2.13)
𝑗=1

and let 𝜁 ◦ ∈ Ξ be arbitrary. If 𝑟 > 0 is sufficiently small that

n o
Δ𝑟(𝑛) (𝜁 ◦ ) = 𝜁 ∈ Ξ; 𝜁 𝑗 − 𝜁 ◦𝑗 < 𝑟, 𝑗 = 1, ..., 𝑛 ⊂⊂ Ξ (19.2.14)

then the Taylor expansion

∑︁ 1
(𝜁 − 𝜁 ◦ ) 𝛼 𝜕𝜁𝛼 𝑝 𝑗 (𝑧, 𝜁 ◦ )
𝛼∈Z𝑛
𝛼!
+

converges to 𝑝 𝑗 (𝑧, 𝜁) uniformly on compact subsets of Ω×Δ𝑟(𝑛) (𝜁 ◦ ). More precisely,

by Definition 19.1.14 and (19.1.1), given an arbitrary compact subset 𝐾 of Ω there
is an 𝑀𝐾 > 0 such that, for all 𝑗 ∈ Z+ , 𝑧 ∈ 𝐾, 𝜁 ∈ Δ𝑟(𝑛) (𝜁 ◦ ), we have 𝑝 𝑗 (𝑧, 𝜁) ≤
𝑗+1
𝑀𝐾 𝑗!. Applying the Cauchy inequalities yields

max 𝜕𝜁𝛼 𝑝 𝑗 (𝑧, 𝜁 ◦ ) ≤ 𝑀𝐾 𝑟 −| 𝛼 | 𝑗!𝛼!.

𝑗+1
(19.2.15)
𝑧 ∈𝐾

In the remainder of this subsection we take 𝑚 = 0; generality is recovered by

multiplication by 𝜆 𝑚 . To the formal series
∞
∑︁ ∑︁ 1
𝑝 (𝑧, 𝜁, 𝜆) = 𝜆− 𝑗 (𝜁 − 𝜁 ◦ ) 𝛼 𝜕𝜁𝛼 𝑝 𝑗 (𝑧, 𝜁 ◦ )
𝑗=0 𝛼∈Z𝑛
𝛼!
+

we associate the series

∞ ∑︁
∑︁ 1 − 𝑗− | 𝛼 | 𝛼
𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) = 𝜆 𝜕𝜁 𝑝 𝑗 (𝑧, 𝜁 ◦ ) D𝑧𝛼 . (19.2.16)
𝑗=0 𝛼∈Z𝑛
𝛼!
+

Note that
◦ ◦
𝑝 (𝑧, 𝜁, 𝜆) = e−𝑖𝜆𝑧· ( 𝜁 −𝜁 ) 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e𝑖𝜆𝑧· (𝜁 −𝜁 ) . (19.2.17)
788 19 Classical Analytic Formalism

Proposition 19.2.23 The series (19.2.16) is a classical analytic differential operator

of infinite order in Ω.

Proof If we write 𝐿 (𝑧, D, 𝜆) = 𝛼∈Z+𝑛 𝑐 𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D 𝛼 with

Í

∞
1 ∑︁ − 𝑗 𝛼
𝑐 𝛼 (𝑧, 𝜆) = 𝜆 𝜕𝜁 𝑝 𝑗 (𝑧, 𝜁 ◦ )
𝛼! 𝑗=0

1 𝛼
then 𝑐 𝛼, 𝑗 (𝑧) = 𝛼! 𝜕𝜁 𝑝 𝑗 (𝑧, 𝜁 ◦ ) satisfies (19.2.1) as a consequence of (19.2.15) □
We can reverse the procedure: to a classical analytic differential operator of infinite
order in Ω,
∑︁ ∑︁ ∞
𝐿 (𝑧, D, 𝜆) = 𝑐 𝛼, 𝑗 (𝑧) 𝜆−| 𝛼 |− 𝑗 D 𝛼 , (19.2.18)
𝛼∈Z+𝑛 𝑗=0

we associate the formal series

∞
∑︁ ∑︁ ◦ ◦
𝑐 𝛼, 𝑗 (𝑧) 𝜆−| 𝛼 |− 𝑗 (𝜁 − 𝜁 ◦ ) 𝛼 = e−𝑖𝜆𝑧· (𝜁 −𝜁 ) 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e𝑖𝜆𝑧· (𝜁 −𝜁 ) .
𝛼∈Z+𝑛 𝑗=0
(19.2.19)
From (D 𝜔 ) we derive that the series
∑︁
𝑐 𝛼, 𝑗 (𝑧) 𝜆− | 𝛼 | (𝜁 − 𝜁 ◦ ) 𝛼
𝛼∈Z+𝑛

converges normally to a holomorphic function 𝑝 𝑗 (𝑧, 𝜁, 𝜆) in Ω × Δ𝑟(𝑛) (𝜁 ◦ ) provided

−1 [𝐶 : the constant in (19.2.1)]. Moreover, if 𝜁 ∈ Δ (𝑛) (𝜁 ◦ ),
𝑟 ≤ 𝐶𝐾 𝐾 𝑟

𝑗+1 𝑟
max 𝑝 𝑗 (𝑧, 𝜁, 𝜆) ≤ 𝐶𝐾 𝑗!,
𝑧 ∈𝐾 𝑟 − |𝜁 − 𝜁 ◦ |

which proves that 𝑝 (𝑧, 𝜁, 𝜆) as defined in (19.2.13) is a classical analytic symbol in

Ω′ × Δ𝑟(𝑛) (𝜁 ◦ ), Ω′ ⊂ 𝐾.

19.3 The Complex Stationary Phase Formula

19.3.1 Complex Stationary Phase Formula. Statement

We revisit the asymptotic stationary phase formula (18.A.6) with the purpose of
proving a finite expansion formula with an estimate of the remainder, under the
hypotheses that the integrand is analytic and the phase-function is complex. Since our
viewpoint is microlocal the reasoning will be valid in a complex neighborhood of a
point 𝑥 ◦ ∈ R𝑛 . Let Ω be an open subset of C𝑛 (where the variable is 𝑧 = 𝑥+𝑖𝑦), 𝑥 ◦ ∈ Ω,
19.3 The Complex Stationary Phase Formula 789

and let Θ be an open subset of C 𝑁 (where the variable is 𝜃). Let 𝜙 ∈ O (Ω × Θ)

have the following properties: 𝜕𝑧 𝜙 (𝑥 ◦ , 𝜃 ◦ ) = 0 for some 𝜃 ◦ ∈ Θ; 𝜕𝑧 𝜙 (𝑧, 𝜃) =
0 =⇒ det 𝜕𝑧2 𝜙 (𝑧, 𝜃) ≠ 0 for all 𝑧 ∈ Ω × Θ. The holomorphic Implicit Function
Theorem applies: possibly after contracting Θ about 𝜃 ◦ there is a holomorphic
map Θ ∋ 𝜃 ↦→ 𝑧 (𝜃) ∈ Ω such that 𝜕𝑧 𝜙 (𝑧 (𝜃) , 𝜃) = 0 whatever 𝜃 ∈ Θ and
𝑧 (𝜃 ◦ ) = 𝑥 ◦ . We carry out the change of variables 𝑧 ⇝ 𝑧 + 𝑧 (𝜃) − 𝑥 ◦ and set
𝜙e (𝑧, 𝜃) = 𝜙 (𝑧 + 𝑧 (𝜃) − 𝑥 ◦ , 𝜃). After some further contracting of Ω×Θ about (𝑥 ◦ , 𝜃 ◦ )
we can assume that 𝜙e ∈ O (Ω × Θ) and that, for all (𝑧, 𝜃) ∈ Ω × Θ,

𝜕𝑧 𝜙e (𝑧, 𝜃) = 0 ⇐⇒ 𝑧 = 𝑥 ◦ ; det 𝜕𝑧2 𝜙e (𝑥 ◦ , 𝜃) ≠ 0. (19.3.1)

The Morse lemma with parameters (Lemma 18.A.3) extends to the complex set-up:
Lemma 19.3.1 Let 𝜙e ∈ O (Ω × Θ) satisfy (19.3.1). There exist neighborhoods Ω′of
𝑥 ◦ in Ω, Θ′ of 𝜃 ◦ in Θ, and for each 𝜃 ∈ Θ′ a biholomorphism 𝑧 ↦→ 𝑧 + 𝑔 (𝑧, 𝜃) of
Ω′ onto an open subset of Ω, with 𝑔 ∈ O (Ω′ × Θ′), sup |𝑔 (𝑧, 𝜃)| ≲ |𝑧 − 𝑥 ◦ | 2 , such
𝜃 ∈Θ′
that if 𝜑 (𝑧, 𝜃) = 𝜙e (𝑧 + 𝑔 (𝑧, 𝜃) , 𝜃) then
1
𝜑 (𝑧, 𝜃) = 𝜑 (𝑥 ◦ , 𝜃) + 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) (𝑧 − 𝑥 ◦ ) · (𝑧 − 𝑥 ◦ )
2
in Ω′ × Θ′.
The proof duplicates exactly that of Lemma 18.A.3, except that the base field R is
replaced by C and C ∞ is replaced by holomorphic. We use the fact that Sym (𝑛, C)
is a complex vector space and GL (𝑛, C) is a complex Lie group; we make use of
the complex Constant Rank Theorem.
Lemma 19.3.2 If the matrix Im 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) is positive definite then there are 𝑇 ∈
GL (𝑛, R) and real numbers 𝜒1 , ..., 𝜒𝑛 be such that

Im 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) 𝑇𝑥 · 𝑇𝑥 = |𝑥| 2 ,
Re 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) 𝑇𝑥 · 𝑇𝑥 = 𝜒1 𝑥 12 + · · · + 𝜒𝑛 𝑥 𝑛2

for all 𝑥 ∈ R𝑛 .
Proof If Im 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) is positive definite then there is a 𝑇1 ∈ GL (𝑛, R) such that
Im 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) 𝑇1 𝑥 · 𝑇1 𝑥 = |𝑥| 2 . Since 𝑆 = 𝑇1⊤ Re 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) 𝑇1 ∈ Sym (𝑛, R) there
is an orthogonal transformation of R𝑛 , 𝑥 ↦→ 𝑇2 𝑥, such that 𝑇2⊤ 𝑆𝑇2 is diagonal; we
can take 𝑇 = 𝑇2𝑇1 . □
With 𝑇 as in Lemma 19.3.2 we have
𝑛 2
1 ©∑︁
𝜑 (𝑥 ◦ + 𝑇 (𝑧 − 𝑥 ◦ ) , 𝜃) = 𝜑 (𝑥 ◦ , 𝜃) + 𝑖 ­ 1 − 𝑖 𝜒 𝑗 𝑧 𝑗 − 𝑥 ◦𝑗 ® . (19.3.2)
ª
2 𝑗=1
« ¬
Under these assumptions we seek a finite series approximation of the type (18.A.6),
with good estimate of the “error”, of integrals
790 19 Classical Analytic Formalism
∫
𝐼𝑈 (𝜃, 𝜆) = e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃) d𝑥, (19.3.3)
𝑈

where 𝑎 (𝑧, 𝜃) ∈ O (Ω × Θ) and 𝑈 ⊂⊂ Ω ∩ R𝑛 . The remainder of this section is

devoted to the proof of the following statement, in which

𝑎♭ (𝑧, 𝜃) = 𝑎 (𝑥 ◦ + 𝑇 (𝑧 − 𝑥 ◦ ) , 𝜃) .

Theorem 19.3.3 Let the phase-function 𝜑 be as in Lemma 19.3.1 with 𝜙e satisfying

(19.3.1). Let 𝑇 ∈ GL (𝑛, R) be such that (19.3.2) holds. If 𝑈 and Θ are sufficiently
small and if 𝑅 > 0 is sufficiently large then
𝑛2 ∫
𝜆 −𝑖𝜆𝜑 ( 𝑥 ◦ , 𝜃)
|det 𝑇 | e e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃) d𝑥 (19.3.4)
2𝜋 𝑈
(2𝜆) − |𝛽 | Ö
𝑛
∑︁ −𝛽 − 1 2𝛽
1 + 𝑖 𝜒 𝑗 𝑗 2 𝜕𝑥 𝑎♭ (𝑥 ◦ , 𝜃)
𝛽!
|𝛽 |<𝜆/𝑅 𝑗=1

mod O (−𝜔) (Θ) (Definition 19.1.10).

We refer to [Hörmander, 1983, I], Theorem 7.7, for a finer result (under weaker
hypotheses on the phase-function!).

19.3.2 Complex Stationary Phase Formula. Proof

An immediate consequence of (19.3.2) is the following (cf. Proposition 20.4.8). If

𝑈 𝑗 , 𝑗 = 1, 2, are sufficiently small open subsets of Ω ∩ R𝑛 such that 𝑥 ◦ ∈ 𝑈1 ∩ 𝑈2
then there are positive constants 𝐶 and 𝑐 such that, for all 𝜃 ∈ Θ,

𝐼𝑈1 ,𝜆 (𝜃) − 𝐼𝑈2 ,𝜆 (𝜃) ≤ 𝐶e−𝑐𝜆 .

This last estimate allows us to modify the domain of integration 𝑈 as needed,

provided 𝑥 ◦ ∈ 𝑈 ⊂⊂ Ω.
To lighten the notation in the remainder of the proof, we take 𝑥 ◦ = 0. We have
∫
1 2
𝐼𝑈 (𝜃, 𝜆) = |det 𝑇 | −1 e𝑖𝜆𝜑 (0, 𝜃) e− 2 𝜆| 𝑥 | +𝑖𝜆𝑄 ( 𝑥) 𝑎♭ (𝑥, 𝜃) d𝑥,
𝑇𝑈

1 Í𝑛
where 𝑄 (𝑥) = 2 𝑗=1 𝜒 𝑗 𝑥 2𝑗 . We select 𝑈 such that 𝑇𝑈 = 𝔅𝑟 = {𝑥 ∈ R𝑛 ; |𝑥| < 𝑟}
(𝑟 > 0); thus we will be studying the integral
∫
1 Í𝑛 2
−1 𝑖𝜆𝜑 (0, 𝜃)
𝐼𝔅𝑟 (𝜃, 𝜆) = |det 𝑇 | e e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑎♭ (𝑥, 𝜃) d𝑥.
| 𝑥 |<𝑟

We have 𝐼𝔅𝑟 (𝜃, 𝜆) ∈ O (Θ) for all 𝜆 > 0.

19.3 The Complex Stationary Phase Formula 791

We are going to make use of a simple estimate of the remainder in a finite Taylor
expansion. Let us use the notation 𝔅𝑟C = {𝑧 ∈ C𝑛 ; |𝑧| < 𝑟}.

Lemma 19.3.4 If ℎ ∈ O 𝔅𝑟C′ for some 𝑟 ′ > 𝑟 > 0 then, for all 𝑧 ∈ 𝔅𝑟C ,

2𝑁
∑︁ 1 ( 𝛼) |𝑧|
ℎ (𝑧) − ℎ (0) 𝑧 𝛼 ≤ (2𝑁 + 1) sup |ℎ| .
𝛼! 𝑟 𝔅𝑟C
| 𝛼 |<2𝑁

Proof After a division we may assume sup |ℎ (𝑧)| = 1. With 𝑥 ∈ R𝑛 , 0 < |𝑥| < 𝑟,
𝑧 ∈𝔅𝑟C
𝜁 ∈ C, |𝜁 | < 1, (hence 𝜁𝑥 ∈ 𝔅𝑟C ) we note that
∑︁ 𝑥𝛼 1 𝜕𝑘
ℎ ( 𝛼) (0) = ℎ (𝜁𝑥) .
𝛼! 𝑘! 𝜕𝜁 𝑘 𝜁 =0
| 𝛼 |=𝑘

Since 𝜁 ↦→ ℎ (𝜁𝑥) extends holomorphically to the disk |𝜁 | < 𝑟 ′ |𝑥| −1 the Cauchy
inequalities in the 𝜁-plane imply

∑︁ 𝑥𝛼
ℎ ( 𝛼) (0) ≤ 𝑟 −𝑘 |𝑥| 𝑘 max |ℎ (𝜁𝑥)| ≤ 𝑟 −𝑘 |𝑥| 𝑘 .
𝛼! |𝜁 |=𝑟/ | 𝑥 |
| 𝛼 |=𝑘

Assuming that |𝜁𝑥| < 𝑟 we derive

2𝑁 −1
!
∑︁ 𝜁 | 𝛼 | ∑︁
( 𝛼) 𝛼 −𝑘 𝑘
ℎ (𝜁𝑥) − ℎ (0) 𝑥 ≤ 1 + 𝑟 |𝜁𝑥| ≤ 2𝑁 + 1.
𝛼!
| 𝛼 |<2𝑁 𝑘=0

The maximum principle in the 𝜁-plane implies

∑︁ 𝜁 | 𝛼 | 2𝑁
−2𝑁 ( 𝛼) 𝛼ª |𝑥|
­ ℎ (𝜁𝑥) − ℎ (0) 𝑥 ® ≤ (2𝑁 + 1) ,
©
𝜁
𝛼! 𝑟
« | 𝛼 |<2𝑁 ¬
which is the sought result. □
We write
∫
1 Í𝑛 2
e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑎♭ (𝑥, 𝜃) d𝑥 (19.3.5)
𝑥 ∈𝔅𝑟
∑︁ 1 ∫
1 Í𝑛 2
= 𝜕𝑥𝛼 𝑎♭ (0, 𝜃) e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 𝛼 d𝑥 + 𝐽 ( 𝑁 ) (𝜃, 𝜆)
𝛼! 𝑥 ∈𝔅𝑟
| 𝛼 |<2𝑁

where
792 19 Classical Analytic Formalism

𝐽 ( 𝑁 ) (𝜃, 𝜆) (19.3.6)
∫ ∑︁ 𝑥 𝛼
1 Í𝑛 2
e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗
© ♭
= ­𝑎 (𝑥, 𝜃) − 𝜕 𝛼 𝑎♭ (0, 𝜃) ® d𝑥.
ª
𝑥 ∈𝔅𝑟 𝛼! 𝑥
« | 𝛼 |<2𝑁 ¬
In the remainder
of this subsection we select 𝑟 ′ > 𝑟 sufficiently small that 𝑎♭ (𝑧, 𝜃) ∈
O 𝔅𝑟 ′ × Θ . We shall also assume that
C

𝑀𝑟 (𝑎) = sup 𝑎♭ (𝑧, 𝜃) < +∞. (19.3.7)

(𝑧, 𝜃) ∈𝔅𝑟C ×Θ

Lemma 19.3.5 If 𝑈 ⊂ 𝔅𝑟 then, for all 𝜃 ∈ Θ, 𝜆 > 0, 𝑁 = 1, 2, ...,

−1
S𝑛−1 𝑟 2𝑁 𝐽 ( 𝑁 ) (𝜃, 𝜆)

1 𝑛 𝑛
≤ 𝑁+ Γ 𝑁+ (2/𝜆) 𝑁 + 2 𝑀𝑟 (𝑎) ,
2 2

where S𝑛−1 is the measure of the unit sphere.

Here and in the sequel Γ (·) is the Euler gamma function.

Proof We derive from Lemma 19.3.4, for all 𝑥 ∈ 𝔅𝑟 ,

∑︁ 𝑥 𝛼
sup 𝑎♭ (𝑥, 𝜃) − 𝜕 𝛼 𝑎♭ (0, 𝜃) ≤ (2𝑁 + 1) 𝑟 −2𝑁 |𝑥| 2𝑁 𝑀𝑟 (𝑎) ,
𝛼! 𝑥
| 𝛼 |<2𝑁

whence

𝐽 ( 𝑁 ) (𝜃, 𝜆) / sup 𝑎♭ (𝑧, 𝜃)

𝑧 ∈𝔅𝑟C
∫
1 2
≤ (2𝑁 + 1) 𝑟 −2𝑁 e− 2 𝜆 | 𝑥 | |𝑥| 2𝑁 d𝑥
R𝑛
∫ ∞
1 2
≤ (2𝑁 + 1) S𝑛−1 𝑟 −2𝑁 e− 2 𝜆𝑡 𝑡 2𝑁 +𝑛−1 d𝑡 . □
0

Next we look at the finite sum in the right-hand side of (19.3.5). We use the fact
that, if 𝛼 𝑗 is odd for some 𝑗 then
∫
1 Í𝑛 2
e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 𝛼 d𝑥 = 0,
𝑥 ∈𝔅𝑟

whence
19.3 The Complex Stationary Phase Formula 793
∫
1
∑︁
2𝛽 1 Í𝑛 2
𝜕𝑥 𝑎♭ (0, 𝜃) e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 𝛼 d𝑥
(2𝛽)! 𝑥 ∈𝔅𝑟
|𝛽 |<𝑁
∫
∑︁ 1 2𝛽 1 Í𝑛 2
= 𝜕𝑥 𝑎♭ (0, 𝜃) e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 2𝛽 d𝑥 − 𝐾 ( 𝑁 ) (𝜃, 𝜆) ,
(2𝛽)! R 𝑛
|𝛽 |<𝑁

where
∫
∑︁ 1 2𝛽 1 Í𝑛 2
𝐾 (𝑁)
(𝜃, 𝜆) = 𝜕 𝑎♭ (0, 𝜃) e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 2𝛽 d𝑥.
(2𝛽)! 𝑥 | 𝑥 |>𝑟
|𝛽 |<𝑁

Lemma 19.3.6 We have, for all 𝜃 ∈ Θ, 𝜆 ≥ 1, 𝑁 = 1, 2, ...,

−1 1 2
S𝑛−1 e 4 𝜆𝑟 𝐾 ( 𝑁 ) (𝜃, 𝜆) (19.3.8)
∑︁ 𝑛 |𝛽 |+ 𝑛2
≤ Γ |𝛽| + 4/𝜆𝑟 2 𝑟 𝑛 𝑀𝑟 (𝑎) .
2
|𝛽 |<𝑁

Proof We have
∫
𝑛−1 −1 1 2
Í𝑛
S e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 2𝛽 d𝑥
|𝑥 |>𝑟
∫ ∞ ∫ ∞
− 21 𝜆𝑡 2 2 |𝛽 |+𝑛−1 − 41 𝜆𝑟 2 1 2 d𝑡
≤ e 𝑡 d𝑡 ≤ e e− 4 𝜆𝑡 𝑡 2 |𝛽 |+𝑛
𝑟 𝑟 𝑡
∫ ∞
|𝛽 |+ 𝑛2 − 14 𝜆𝑟 2 |𝛽 |+ 𝑛2 −1
≤ (4/𝜆) e e−𝑠 𝑠 d𝑠,
0

whence
∫
𝑛−1 −1 1 2
Í𝑛
S e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 2𝛽 d𝑥
| 𝑥 |>𝑟
𝑛 𝑛 1 2
≤ Γ |𝛽| + (4/𝜆) |𝛽 |+ 2 e− 4 𝜆𝑟 .
2
We derive from this and the definition of 𝐾 ( 𝑁 ) :

𝐾 ( 𝑁 ) (𝜃, 𝜆)
∑︁ 1 𝑛 𝑛 1 2 2𝛽
≤ Γ |𝛽| + (4/𝜆) |𝛽 |+ 2 e− 4 𝜆𝑟 𝜕𝑥 𝑎♭ (0, 𝜃) .
(2𝛽)! 2
|𝛽 |<𝑁

The Cauchy inequalities imply

1 2𝛽
𝜕 𝑎♭ (0, 𝜃) ≤ 𝑟 −2|𝛽 | sup 𝑎♭ (𝑧, 𝜃) ,
(2𝛽)! 𝑥 𝑧 ∈𝔅𝑟C

whence (19.3.8). □
794 19 Classical Analytic Formalism

We compute
∫
1 2
∫ ∞
2𝛽 1 1
e− 2 𝜆(1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 𝑗 𝑗 d𝑥 𝑗 = e− 2 𝜆(1+𝑖 𝜒 𝑗 )𝑡 𝑡 𝛽 𝑗 − 2 d𝑡
R 0
𝛽 𝑗 + 21
2 1
= Γ 𝛽𝑗 + ,
1 + 𝑖 𝜒𝑗 𝜆 2
21
where 1 + 𝑖 𝜒 𝑗 stands for the main branch of the square root, whence

1
∫
1 Í 𝑛 2
|𝛽 |+ 𝑛2 Ö
2
𝑛 Γ 𝛽 𝑗 + 2
e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 2𝛽 d𝑥 = 𝛽 𝑗 + 21 .
R 𝑛 𝜆 𝑗=1 1 + 𝑖 𝜒 𝑗

1 √
Since Γ 2 = 𝜋 and

√

1
Γ 𝑘+ Γ (𝑘) = 21−2𝑘 𝜋Γ (2𝑘) , 𝑘 = 1, 2, ...,
2
we get
𝑛2 ∫
𝜆 1 Í𝑛 2
e− 2 𝜆 𝑗=1 (1+𝑖 𝜒 𝑗 ) 𝑥 𝑗 𝑥 2𝛽 d𝑥
2𝜋 R𝑛
𝑛
Ö −𝛽 𝑗 − 21 2𝛽 𝑗 − 1 !
= 2𝑛 (2𝜆) − |𝛽 | 1 + 𝑖 𝜒𝑗
𝑗=1
𝛽𝑗 − 1 !

( 2𝛽 𝑗 −1) !
with the understanding that = 1 if 𝛽 𝑗 = 0. Since
( 𝛽 𝑗 −1) !

2𝛽 𝑗 − 1 ! 1 1
=
𝛽 𝑗 − 1 ! 2𝛽 𝑗 ! 2 𝛽 𝑗 !

this yields
𝑛

𝜆 2
|det 𝑇 | e𝑖𝜆𝜑 (0, 𝜃) 𝐼𝔅𝑟 (𝜃, 𝜆)
2𝜋
∑︁ (2𝜆) − |𝛽 | Ö𝑛
−𝛽 − 1 2𝛽
= 1 + 𝑖 𝜒 𝑗 𝑗 2 𝜕𝑥 𝑎♭ (0, 𝜃)
𝛽!
|𝛽 |<𝑁 𝑗=1

+ 𝐽 ( 𝑁 ) (𝜃, 𝜆) + 𝐾 ( 𝑁 ) (𝜃, 𝜆) ,

where 𝜃 ∈ Θ and 𝜆 > 0. Keep in mind that 𝜑 satisfies (19.3.2).

We exploit Lemmas 19.3.5 and 19.3.6 to get the needed estimates of the “error”
terms 𝐽 ( 𝑁 ) and 𝐾 ( 𝑁 ) .
19.3 The Complex Stationary Phase Formula 795

Lemma 19.3.7 There is a constant 𝐶𝑛 > 0 depending only on 𝑛 such that

𝑁 + 21 𝑛 − 𝜆𝑟 2 < 21 implies
1 2
𝐽 ( 𝑁 ) (𝜃, 𝜆) + 𝐾 ( 𝑁 ) (𝜃, 𝜆) ≤ 𝐶𝑛 𝑟 𝑛 e− 4 𝜆𝑟 𝑀𝑟 (𝑎) (19.3.9)

for all 𝜃 ∈ Θ.

Proof We repeatedly use Stirling’s formula to get a (rough) majorant of the gamma
function: √︂
𝜋 𝑥 𝑥
|Γ (𝑥)| ≤ 2 , 𝑥 > 0 large. (19.3.10)
𝑥 e
We derive from (19.3.10)

1 𝑛 𝑛
𝑁+ Γ 𝑁+ (2/𝜆) 𝑁 + 2 𝑟 −2𝑁
2 2
𝑁 + 𝑛2
𝑛 2𝑁 + 𝑛
√︁
≤ 𝜋 (2𝑁 + 𝑛)𝑟 .
e𝜆𝑟 2

When 𝑁 + 12 𝑛 ≤ 12 𝜆𝑟 2 + 1
2 we have

1 𝑛 𝑛
𝑁+Γ 𝑁+ (2/𝜆) 𝑁 + 2 𝑟 −2𝑁
2 2
𝑁 + 𝑛2
1
e− ( 𝑁 + 2 )
√︁ 𝑛
𝑛
≤ 𝜋 (2𝑁 + 𝑛)𝑟 1 +
2𝑁 + 𝑛 − 1
√︃
≤ 𝜋e 𝜆𝑟 2 + 1 𝑟 𝑛 e− ( 𝑁 + 2 ) .
𝑛

1 2
Since 𝜆𝑟 2 < 4e 4 𝜆𝑟 we obtain, when 𝑁 + 21 𝑛 ≥ 12 𝜆𝑟 2 − 21 ,

1 𝑛 𝑛 1 2
𝑁+ Γ 𝑁+ (2/𝜆) 𝑁 + 2 𝑟 −2𝑁 ≤ 𝐶∗ 𝑟 𝑛 e− 4 𝜆𝑟 ,
2 2

where 𝐶∗ > 0 is a universal constant. Putting this into (19.3.6) yields

1 2
𝐽 ( 𝑁 ) (𝜃, 𝜆) ≤ 𝐶∗ S𝑛−1 𝑟 𝑛 e− 4 𝜆𝑟 𝑀𝑟 (𝑎) . (19.3.11)

Next we use (19.3.10) to obtain

√︂ 𝑛 |𝛽 |+ 𝑛2
|𝛽| +

∑︁ 𝑛 |𝛽 |+ 𝑛2 −2|𝛽 | 2𝜋 𝑛 ∑︁ 2
Γ |𝛽| + (4/𝜆) 𝑟 ≤2 𝑟 4 .
|𝛽 |<𝑁
2 𝑛
|𝛽 |<𝑁
e𝜆𝑟 2

When 𝑁 + 12 𝑛 ≤ 12 𝜆𝑟 2 + 1
2 the right-hand side does not exceed
796 19 Classical Analytic Formalism
𝑛
|𝛽| + 𝑛2 |𝛽 |+ 2
√︂ |𝛽 |
2𝑛+1 𝜋 𝑛 ∑︁ 2
2 𝑟
𝑛e e 𝑁 − 1 + 𝑛2
|𝛽 |<𝑁
√︂
2𝑛+1 𝜋 e 𝑛 𝑛 − 1 𝜆𝑟 2
≤2 𝑟 e 4 .
𝑛e e−2
Combining this with (19.3.11) yields (19.3.9). □

At this stage we have proved, by combining (19.3.8) and (19.3.9), that if

𝑁 + 21 𝑛 − 𝜆𝑟 2 < 21 then
𝑛2
𝜆
|det 𝑇 | e𝑖𝜆𝜑 (0, 𝜃) 𝐼𝔅𝑟 (𝜃, 𝜆)
𝜋
∑︁ (2𝜆) −|𝛽 | Ö𝑛
−𝛽 − 1 2𝛽
1 + 𝑖 𝜒 𝑗 𝑗 2 𝜕𝑥 𝑎♭ (0, 𝜃)
𝛽!
|𝛽 |<𝑁 𝑗=1

mod O (−𝜔) (Θ). Keeping 𝑟 fixed and selecting 𝜆 sufficiently large we see that to each
𝑁 ∈ Z+ there is an 𝑅 > 0 such that 𝑁 + 12 𝑛 − 𝜆𝑟 2 < 21 implies 𝑁 = [𝜆/𝑅] + 1, in
which case the preceding congruence reads
𝑛2
𝜆
|det 𝑇 | e𝑖𝜆𝜑 (0, 𝜃) 𝐼𝔅𝑟 (𝜃, 𝜆) (19.3.12)
𝜋
∑︁ (2𝜆) −|𝛽 | Ö 𝑛
−𝛽 − 1 2𝛽
1 + 𝑖 𝜒 𝑗 𝑗 2 𝜕𝑥 𝑎♭ (0, 𝜃) .
𝛽!
|𝛽 | ≤ [𝜆/𝑅] 𝑗=1

Using once again the Cauchy inequalities we see that

2− |𝛽 | Ö
𝑛
−𝛽 𝑗 − 21 2𝛽
1 + 𝑖 𝜒𝑗 𝜕𝑥 𝑎♭ (0, 𝜃)
𝛽! 𝑗=1
(2𝛽)! − |𝛽 |
≤ 2−|𝛽 | 𝑟 𝑀𝑟 (𝑎)
𝛽!
≤ (2/𝑟) |𝛽 | 𝛽!𝑀𝑟 (𝑎)

with 𝑀𝑟 (𝑎) as in (19.3.7). This proves that

∑︁ (2𝜆) − |𝛽 | Ö𝑛
−𝛽 − 1 2𝛽
1 + 𝑖 𝜒 𝑗 𝑗 2 𝜕𝑥 𝑎♭ (0, 𝜃)
𝛽 ∈Z𝑛
𝛽! 𝑗=1
+

is a formal analytic series (cf. Definition 19.1.1) and the right-hand side in (19.3.12)
is one of its finite realizations (Definition 19.1.7). In view of this we can exploit
Lemma 19.1.9: provided 𝑅 is large enough the congruence (19.3.12) remains valid
if we replace 𝑅 by 𝑅 ′ > 𝑅.
This completes the proof of Theorem 19.3.3.
19.3 The Complex Stationary Phase Formula 797

19.3.3 Complex stationary phase formula and classical analytic

differential operators of infinite order

Let Ω and Θ be domains in C𝑛 ; 𝑥 ◦ ∈ Ω ∩ R𝑛 , 𝜃 ◦ ∈ Θ; we continue to deal with

𝜑 ∈ O (Ω × Θ) satisfying the following hypotheses:
(1) ∀ (𝑧, 𝜃) ∈ Ω × Θ, 𝜕𝑧 𝜑 (𝑧, 𝜃) = 0 ⇐⇒ 𝑧 = 𝑥 ◦ ;
(2) Im 𝜕𝑧2 𝜑 (𝑥 ◦ , 𝜃 ◦ ) is positive definite;
(3) 𝑇 is the linear transformation in Lemma 19.3.2.
Let 𝑈 be an open subset of Ω ∩ R𝑛 and ℎ ∈ O (Ω). We define
∞
∑︁
♭
𝑎 (𝑥, 𝜃, 𝜆) = 𝜆−𝑘 𝑎♭𝑘 (𝑧, 𝜃) , 𝑎♭𝑘 (𝑧, 𝜃) = 𝑎 𝑘 (𝑥 ◦ + 𝑇 (𝑧 − 𝑥 ◦ ) , 𝜃) .
𝑘=0

If we substitute 𝑎 (𝑥, 𝜃, 𝜆) ℎ (𝑥) for 𝑎 (𝑥, 𝜃) in (19.3.4) the Leibniz rule implies
𝑛 ∫
𝜆 2 −𝑖𝜆𝜑 ( 𝑥 ◦ , 𝜃)
|det 𝑇 | e e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) ℎ (𝑥) d𝑥 (19.3.13)
2𝜋 𝑈
∞
∑︁ ∑︁ 2− |𝛽 | ∑︁ 2𝛽
2𝛽−𝛾
Π (𝛽) 𝜆−𝑘− |𝛽 | 𝑎 𝑘 (𝑥 ◦ , 𝜃) 𝜕𝑥 ℎ (𝑥 ◦ ) ,
𝛾
𝜕𝑥
𝛽! 𝛾 ⪯2𝛽
𝛾
𝑘=0 |𝛽 |<𝜆/𝑅

where stands for congruent mod O (−𝜔) (Θ) and

𝑛
Ö −𝛽 𝑗 − 21
Π (𝛽) = 1 + 𝑖 𝜒𝑗 .
𝑗=1

We are interested in the series on the right-hand side of (19.3.13) without the
limitation |𝛽| < 𝜆/𝑅, and in the differential operator
∑︁
𝑏♭𝛾 (𝑧, 𝜃, 𝜆) 𝜆− |𝛾 | 𝜕𝑧 ,
𝛾
𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) = (19.3.14)
𝛾 ∈Z+𝑛

where D𝑧 = −𝑖𝜕𝑧 and

∞ ∑︁ − |𝛽 |
∑︁ 2 2𝛽 2𝛽−𝛾 ♭

𝑏♭𝛾 (𝑧, 𝜃, 𝜆) = Π (𝛽) 𝜆−𝑘− |𝛽 |+|𝛾 | 𝜕𝑧 𝑎𝑘 .
𝑘=0 𝛽 ∈Z𝑛
𝛽! 𝛾
+
2𝛽 ⪰𝛾

If we write 𝑏♭𝛾 (𝑧, 𝜃, 𝜆) =

Í − 𝑗 𝑏♭ (𝑧, 𝜃) then
𝑗 ≥ |𝛾 |/2 𝜆 𝛾, 𝑗

∑︁ (2𝛽)! 2𝛽−𝛾 ♭
𝑏♭𝛾, 𝑗 (𝑧, 𝜃) = 2−|𝛽 | Π (𝛽) 𝜕 𝑎 𝑗− |𝛽 |+ |𝛾 | (𝑧, 𝜃) . (19.3.15)
𝛽!𝛾! (2𝛽 − 𝛾)! 𝑧
|𝛽 | ≤ 𝑗+|𝛾 |
2𝛽 ⪰𝛾
798 19 Classical Analytic Formalism

Proposition 19.3.8 The differential operator (19.3.14) satisfies Condition (D 𝜔 )

(Definition 19.2.1) uniformly with respect to 𝜃 on compact subsets of Θ; 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆)
is elliptic if and only if the classical symbol 𝑎 (𝑧, 𝜃, 𝜆) is elliptic (of order zero) in
Ω × Θ.

Proof To every pair of open sets Ω′ ⊂⊂ Ω and Θ′ ⊂⊂ Θ there is a constant 𝐶 > 0

such that
2𝛽−𝛾 ♭
sup 𝜕𝑧 𝑎 𝑗−|𝛽 |+|𝛾 | (𝑧, 𝜃) ≲ 𝐶 𝑗+|𝛽 |+|𝛾 | (2𝛽 − 𝛾)! ( 𝑗 − |𝛽| + |𝛾|)!.
(𝑧, 𝜃) ∈Ω′ ×Θ′
(19.3.16)
By combining (19.3.15) and (19.3.16), and using the fact that (2𝛽)! ≤ 4 |𝛽 | (𝛽!) 2 and
Π (𝛽) ≤ 1, we see that there is a constant 𝐶1 > 0 such that

𝑗+|𝛾 |+1
∑︁ 𝛽!
sup 𝑏♭𝛾, 𝑗 (𝑧, 𝜃) ≤ 𝐶1 ( 𝑗 − |𝛽| + |𝛾|)!.
(𝑧, 𝜃) ∈Ω′ ×Θ′ 𝛾!
|𝛽 | ≤ 𝑗+|𝛾 |
2𝛽 ⪰𝛾

We have
1 ∑︁ 𝛽!
( 𝑗 − |𝛽| + |𝛾|)!
𝑗! 𝛾!
|𝛽 | ≤ 𝑗+|𝛾 |
2𝛽 ⪰𝛾
|𝛾|! ( 𝑗 + |𝛾|)! ∑︁ ( 𝑗 + |𝛾| − |𝛽|)! |𝛽|! 𝛽! ∑︁ 𝛽!
≤ ≤ 2 𝑗+2|𝛾 |
𝛾! 𝑗! |𝛾|! ( 𝑗 + |𝛾|)! |𝛽|! |𝛽|!
|𝛽 | ≤ 𝑗+|𝛾 | |𝛽 | ≤ 𝑗+|𝛾 |

since ( 𝑗 + |𝛾| − |𝛽|)! |𝛽|! ≤ ( 𝑗 + |𝛾|)!. We apply (17.2.21):

∑︁ 𝛽! ∑︁ | ∑︁ 𝛽1 ! · · · 𝛽𝑛 !
𝑗+|𝛾
= ≤ 3𝑛−1 ( 𝑗 + |𝛾| + 1) ,
|𝛽|! 𝑝!
|𝛽 | ≤ 𝑗+|𝛾 | 𝑝=0 |𝛽 |= 𝑝

whence ∑︁ 𝛽! 𝑗+|𝛾 |
( 𝑗 − |𝛽| + |𝛾|)! ≲ 𝐶2 𝑗!.
𝛾!
|𝛽 | ≤ 𝑗+|𝛾 |
2𝛽 ⪰𝛾

Putting this into (19.3.16) proves the first part of the statement. The last part follows
from the fact that 𝑏♭0,0 (𝑧, 𝜃) = 𝑎♭0 (𝑧, 𝜃). □
An alternate statement of Proposition 19.2.3 is that 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) can be viewed
as a classical analytic differential operator of infinite order in (𝑧, 𝜃)-space which
happens not to involve partial derivatives with respect to 𝜃. All the consequences of
Definition 19.2.1 apply:

Proposition 19.3.9 For every ℎ ∈ Aform (Ω × Θ) we have

𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ℎ ∈ Aform (Ω × Θ) .
19.3 The Complex Stationary Phase Formula 799

Formally, we have, for 𝑎, ℎ ∈ Aform (Ω × Θ),

∑︁ 2−|𝛽 |
2𝛽
𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ℎ = Π (𝛽) 𝜆−|𝛽 | 𝜕𝑧 𝑎♭ ℎ (19.3.17)
𝛽 ∈Z𝑛
𝛽!
+
∞ ∑︁ −|𝛽 |
∑︁ 2 2𝛽

= Π (𝛽) 𝜆− 𝑗−𝑘− |𝛽 | 𝜕𝑧 𝑎♭𝑘 ℎ 𝑗 .
𝑗,𝑘=0 𝛽 ∈Z𝑛
𝛽!
+

We can rewrite (19.3.13) as follows:

𝑛2 ∫
𝜆
e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) ℎ (𝑥) d𝑥 (19.3.18)
𝜋 𝑈
◦ , 𝜃)
|det 𝑇 | −1 e𝑖𝜆𝜑 ( 𝑥 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ℎ (𝑥 ◦ , 𝜃,𝜆) mod O (−𝜔) (Θ) ,

As before 𝐿 (𝑅) (𝑧, 𝜃, D𝑧𝑧 , 𝜆) denotes a finite realization of (19.3.14). The results
of Subsection 19.2.1 have the following consequences.

Proposition 19.3.10 Given arbitrarily open subsets of C𝑛 and C 𝑁 respectively,

Ω′ ⊂⊂ Ω and Θ′ ⊂⊂ Θ, the following is true:
(1) To every pair of open subsets Ω′′ and Θ′′ of C𝑛 and C 𝑁 respectively, such that
Ω′ ⊂⊂ Ω′′ ⊂⊂ Ω and Θ′ ⊂⊂ Θ′′ ⊂⊂ Θ, and to every ℎ ∈ Ob (Ω × Θ) there are
positive constants 𝑅, 𝐶 such that if 𝑅 ′ > 𝑅 then
′
∀𝜆 ≥ 1, sup 𝐿 (𝑅 ) (𝑧, 𝜃, D𝑧 , 𝜆) ℎ (𝑧, 𝜆) ≤ 𝐶 sup |ℎ (𝑧, 𝜆)| .
Ω′ ×Θ′ Ω′′ ×Θ′′

(2) To every ℎ ∈ Ob (Ω × Θ) there are positive constants 𝑅 ℎ , 𝐶 such that if 𝑅 ′ >

𝑅 > 𝑅 ℎ then
′ ′
∀𝜆 ≥ 1, sup 𝐿 (𝑅) (𝑧, 𝜃, D𝑧 , 𝜆) ℎ (𝑧, 𝜆) − 𝐿 (𝑅 ) (𝑧, 𝜃, D𝑧 , 𝜆) ℎ (𝑧, 𝜆) ≤ 𝐶e−𝜆/𝑅 .
Ω′ ×Θ′

(3) To every ℎ ∈ O (−𝜔) (Ω × Θ) there is an 𝑅 ℎ > 0 such that 𝐿 (𝑅) (𝑧, 𝜃, D𝑧𝑧 , 𝜆) ℎ
belongs to O (−𝜔) (Ω′ × Θ′) whatever 𝑅 > 𝑅 ℎ .

By contracting Ω arbitrarily about 𝑥 ◦ ∈ Ω ∩ R𝑛 we can use Proposition 19.3.8 to

define a germ map
L 𝑥 ◦ (𝜃) : Aform, 𝑥 ◦ −→ Aform, 𝑥 ◦ , (19.3.19)
depending on 𝜃 ∈ Θ holomorphically, meaning that if ℎ ∈ Aform (Ω′) (Ω′: a neigh-
borhood of 𝑥 ◦ in Ω) represents a germ h 𝑥 ◦ then L 𝑥 ◦ (𝜃) h 𝑥 ◦ is represented by
𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ℎ ∈ Aform (Ω′ × Θ). We also have the analogue of (19.2.6): the finite
realizations 𝐿 (𝑅) (𝑧, 𝜋, D𝑧 , 𝜆) define a germ linear operator

L (𝑧 ◦ , 𝜃 ◦ ) : Ob, (𝑧 ◦ , 𝜃 ◦ ) /O (−𝜔)
(𝑧 ◦ , 𝜃 ◦ )
−→ Ob, (𝑧 ◦ , 𝜃 ◦ ) /O (−𝜔)
(𝑧 ◦ , 𝜃 ◦ )
, (19.3.20)
800 19 Classical Analytic Formalism

or, equivalently, a sheaf map of Ob (Ω × Θ) /O (−𝜔) (Ω × Θ) into itself.

19.3.4 Action of 𝑳 (𝒛, 𝜽, D𝒛 , 𝝀) on formal microfunction series

Here we look at the formal transpose of (19.3.14) regarded as a linear operator of

Aform (Ω) into O (Ω × Θ) 𝜆−1 :

∑︁
𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ⊤ ℎ = |det 𝑇 | −1 (−1) |𝛾 | 𝜆−𝑁𝛾 𝜕𝑧 𝑏♭𝛾 ℎ .
𝛾
(19.3.21)
𝛾 ∈Z+𝑛

Proposition 19.3.11 For every ℎ ∈ Aform (Ω) we have

𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ⊤ ℎ ∈ Aform (Ω × Θ) .

Proof It suffices to deal with ℎ ∈ O (Ω). Going back to (19.3.13) we see that
∞
∑︁ ∑︁ (2𝛽)!
2𝛽−𝛾 ♭

𝜆 𝑁𝛾 −𝑚 2−|𝛽 | Π (𝛽)
𝛾 𝛾
𝜕𝑧 ℎ𝑏♭𝛾 = 𝜕𝑧 ℎ𝜕𝑧 𝑎 𝑚−|𝛽 | ,
𝛽!𝛾! (2𝛽 − 𝛾)!
𝑚=𝑁𝛾 |𝛽 | ≤𝑚
2𝛽 ⪰𝛾

whence
∞
∑︁
|det 𝑇 | 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ⊤ ℎ = 𝜆−𝑚 𝑐 𝑚
′
(𝑧, 𝜃) ,
𝑚=0

where
∑︁ ∑︁ (2𝛽)!
2𝛽−𝛾 ♭

′
(−1) |𝛾 | 2−|𝛽 | Π (𝛽)
𝛾
𝑐𝑚 = 𝜕𝑧 ℎ𝜕𝑧 𝑎 𝑚−|𝛽 | .
𝛾 ∈Z+𝑛
𝛽!𝛾! (2𝛽 − 𝛾)!
|𝛽 | ≤𝑚
𝑁𝛾 ≤𝑚 2𝛽 ⪰𝛾

We reverse the argument in the proof of Proposition 19.3.8. Let Ω′ ⊂⊂ Ω and

Θ′ ⊂⊂ Θ be open subsets of C𝑛 and C 𝑁 respectively; we have
1 𝛾

2𝛽−𝛾 ♭

sup 𝜕𝑧 ℎ𝜕𝑧 𝑎 𝑚−|𝛽 |
𝛾! (𝑧, 𝜃) ∈Ω′ ×Θ′
2𝛽−𝛾 ♭
≤ 𝐶 |𝛾 |+1 sup ℎ𝜕𝑧 𝑎 𝑚−|𝛽 |
(𝑧, 𝜃) ∈Ω′′ ×Θ′′
2𝛽−𝛾 ♭
≤ 𝐶 |𝛾 |+1 𝑀 (ℎ) sup 𝜕𝑧 𝑎 𝑚−|𝛽 | ,
(𝑧, 𝜃) ∈Ω′′ ×Θ′′

where Ω′′ and Θ′′ are open subsets of C𝑛 and C 𝑁 such that Ω′ ⊂⊂ Ω′′ ⊂⊂ Ω and
Θ′ ⊂⊂ Θ′′ ⊂⊂ Θ; 𝑀 (ℎ) = sup |ℎ (𝑧)|. We derive from (19.3.16):
𝑧 ∈Ω′′
19.3 The Complex Stationary Phase Formula 801

2𝛽−𝛾 ♭
sup 𝜕𝑧 𝑎 𝑚−|𝛽 | (𝑧, 𝜃) ≤ 𝐶 𝑚+|2𝛽−𝛾 |+1 (2𝛽 − 𝛾)! (𝑚 − |𝛽|)!,
(𝑧, 𝜃) ∈Ω′ ×Θ′

whence
∑︁ (2𝛽)!
2𝛽−𝛾 ♭

2− |𝛽 | Π (𝛽)
𝛾
𝜕𝑧 ℎ𝜕𝑧 𝑎 𝑚−|𝛽 |
𝛽!𝛾! (2𝛽 − 𝛾)!
|𝛽 | ≤𝑚
2𝛽 ⪰𝛾
∑︁ (2𝛽)!
≤ 𝐶 𝑚+1 𝑀 (ℎ) 2− |𝛽 | 𝐶 2 |𝛽 | (𝑚 − |𝛽|)!
𝛽!
|𝛽 | ≤𝑚
2𝛽 ⪰𝛾
∑︁
≤ 𝐶 𝑚+1 𝑀 (ℎ) 2 |𝛽 | 𝐶 2|𝛽 | 𝛽! (𝑚 − |𝛽|)!
|𝛽 | ≤𝑚
2𝛽 ⪰𝛾
∑︁ 𝛽! (𝑚 − |𝛽|)!
≤ 2𝑚 𝐶 3𝑚+1 𝑀 (ℎ) 𝑚!
𝑚!
|𝛽 | ≤𝑚
𝑛 𝑚 3𝑚+1
≤3 2 𝐶 𝑀 (ℎ) 𝑚!. □

Proposition 19.3.11 enables us to duplicate verbatim the procedure in Subsec-

tion 19.2.3 [based on Formula (19.2.7)] extending the action of differential opera-
tors of the type (19.3.14) to formal analytic functional, hyperfunction, singularity
hyperfunction and microfunction series. The only difference is the holomorphic
dependence on 𝜃 ∈ Θ, which must be appropriately defined:
As derived from the estimates in the preceding proofs, or from the closed
graph theorem, 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) ⊤ defines a continuous linear map Aform (Ω) −→
Aform (Ω × Θ) with respect to the natural locally convex topologies on these vector
spaces (cf. Remark 19.1.5). The transpose of this map defines an endomorphism of
′ (Ω). If 𝜇 = ∞ −𝑗 ′ ′
Í
Oform 𝑗=0 𝜆 𝜇 𝑗 ∈ Oform (Ω ∩ R ), meaning that 𝜇 ∈ Oform (Ω) and,
𝑛

moreover, that there is a compact subset 𝐾 of Ω ∩ R such that supp 𝜇 𝑗 ⊂ 𝐾 for

𝑛

every 𝑗, then 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) 𝜇 has the same properties. Restriction of 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆)

defines an endomorphism of Oform ′ (Ω); 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) 𝜇 depends holomorphically
on 𝜃 ∈ Θ, i.e., it is a holomorphic function of 𝜃 ∈ Θ valued in the TVS Oform ′ (Ω)
(see Remark 19.1.26).
This concept is then extended to formal hyperfunction series 𝑓 = ∞ −𝑗 𝑓 ∈
Í
𝑗=0 𝜆 𝑗
Bform (Ω ∩ R ) (Definition 19.1.23).
𝑛
Í∞ − 𝑗 Given arbitrarily an open subset 𝑈 of R 𝑛,

𝑈 ⊂⊂ Ω ∩ R , there is a 𝜇 = 𝑗=0 𝜆 𝜇 𝑗 ∈ Oform

𝑛 ′ (Ω) representing 𝑓 in 𝑈, i.e.,
such that 𝜇 𝑗 represents 𝑓 𝑗 in 𝑈 for every 𝑗. We define 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) 𝑓 as the formal
hyperfunction series in Ω represented by 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) 𝜇 in 𝑈; the holomorphic
dependence on 𝜃 of 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) 𝑓 is simply synonymous to that of 𝐿 (𝑧, 𝜃, D𝑧 , 𝜆) 𝜇.
Then every formal microfunction series [ 𝑓 ] ∈ Bform micro (𝑈 × Σ) (Σ: an open subset

of S ) represented by 𝑓 ∈ Bform (Ω ∩ R ) is said to depend holomorphically on

𝑛−1 𝑛

𝜃 ∈ Θ.
By contracting Ω arbitrarily about 𝑥 ◦ we get a linear operator acting on germs of
formal hyperfunction series,
802 19 Classical Analytic Formalism

L 𝑥 ◦ (𝜃) : Bform,𝑥 ◦ −→ Bform,𝑥 ◦ . (19.3.22)

Then, taking (19.3.19) into account, we can define a linear operator on germs of
formal singularity hyperfunction series

L 𝑥 ◦ (𝜃) : Bform,𝑥 ◦ /Aform, 𝑥 ◦ −→ Bform,𝑥 ◦ /Aform, 𝑥 ◦ . (19.3.23)

Likewise, we can define a linear operator acting on germs of formal microfunction

series. All these maps depend holomorphically on 𝜃 ∈ Θ as explained for (19.3.19).

19.4 Symbolic Calculus and the KdV Hierarchy

19.4.1 The differential algebra of polynomials in an infinite sequence of

variables

Let C [𝑢 0 , 𝑢 1 , ..., 𝑢 𝜈 ...] denote the algebra of polynomials in the infinite sequence of
indeterminates 𝑢 𝑗 ; we turn C [𝑢 0 , 𝑢 1 , ..., 𝑢 𝜈 ...] into a differential algebra (denoted
by 𝔓 in this section) for the chain-rule derivation 𝔡:
∞
∑︁ 𝜕𝑃
𝔡𝑃 = 𝑢 𝑗+1 ; (19.4.1)
𝑗=0
𝜕𝑢 𝑗

𝔡 is a linear operator and obeys the Leibniz rule.

We denote by 𝔓◦ the subset of 𝔓 consisting of the polynomials 𝑃 such that 𝑃 (0) =
0. Note that 𝔡𝔓 ⊂ 𝔓◦ ; and that 𝔡 preserves the standard degree of polynomials
whereas it increases the order of nonconstant polynomials by 1. This has the following
direct consequence:

Proposition 19.4.1 The chain-rule derivation 𝔡 restricted to 𝔓◦ is injective.

The transpose Leibniz rule (1.1.5) entails, for all 𝑃, 𝑄 ∈ 𝔓 and all 𝑘 ∈ Z+ :

𝑃𝔡 𝑘 𝑄 − (−1) 𝑘 𝑄𝔡 𝑘 𝑃 ∈ 𝔡𝔓. (19.4.2)

Another formula frequently applied and valid for all 𝑃 ∈ 𝔓 and all 𝑗, 𝑘 ∈ Z+ is the
following:
min(
∑︁𝑗,𝑘) 𝑘
𝜕 𝑘 𝑘−ℓ 𝜕𝑃
𝔡 𝑃= 𝔡 . (19.4.3)
𝜕𝑢 𝑗 ℓ=0
ℓ 𝜕𝑢 𝑗−ℓ

In particular, 𝜕𝑢𝜕 0 and 𝔡 commute. The proof of (19.4.3), by induction on ( 𝑗, 𝑘), is

straightforward.
The chain-rule derivation is a special case of the following derivations associated
to an arbitrary polynomial 𝑃 ∈ 𝔓◦ :
19.4 Symbolic Calculus and the KdV Hierarchy 803
∞
∑︁ 𝜕𝑄
𝔓 ∋ 𝑄 ↦→ 𝜗𝑃 𝑄 = 𝔡𝑗𝑃 . (19.4.4)
𝑗=0
𝜕𝑢 𝑗

Indeed, 𝜗𝑃 = 𝔡 if 𝑃 (𝑢) = 𝑢 1 . For any 𝑃 ∈ 𝔓 the operator 𝜗𝑃 is the derivation of 𝔓

such that 𝜗𝑃 𝑢 0 = 𝑃; as a matter of fact, 𝜗𝑃 𝑢 𝑚 = 𝔡 𝑚 𝑃 for every 𝑚 ∈ Z+ , whence

∀𝑃 ∈ 𝔓, 𝔡𝜗𝑃 = 𝜗𝑃 𝔡. (19.4.5)

The degree of 𝑃 ∈ 𝔓 shall have the standard meaning for polynomials in several
variables and will be denoted by deg 𝑃. We define the order of 𝑃 ∈ 𝔓 as the highest
integer 𝑗 such that 𝜕𝑃/𝜕𝑢 𝑗 ≠ 0 and denote it by order 𝑃. We have order 𝔡𝑃 =
1 + order 𝑃.
It is convenient to introduce a notion of weight in 𝔓:
Definition 19.4.2 If 𝑃 is a monomial 𝑢 𝑝𝑗11 , ..., 𝑢 𝑝𝑗𝜈𝜈 (0 ≤ 𝑗 1 < · · · < 𝑗 𝜈 < +∞) then its
weight shall be 𝑗 1 𝑝 1 + · · · + 𝑗 𝜈 𝑝 𝜈 ; the weight of an arbitrary 𝑃 ∈ C [𝑢 0 , 𝑢 1 , ..., 𝑢 𝜈 ...],
wght 𝑃, shall be the largest weight of its monomials.
Note that wght 𝔡𝑃 = 1 + wght 𝑃; deg 𝑃 ≤ wght 𝑃, order 𝑃 ≤ wght 𝑃.
There is a natural locally convex topology on 𝔓: For each 𝑚 ∈ Z+ the vector
subspace 𝔓 (𝑚) of 𝑃 ∈ 𝔓 such that wght 𝑃 ≤ 𝑚 is a finite-dimensional complex
vector space. A convex subset of 𝔓 is open if its intersection with 𝔓 (𝑚) is open in
𝔓 (𝑚) for every 𝑚; this turns 𝔓 into an LF space (see [Treves, 1967], Ch. 13). A
subset 𝑩 of 𝔓 is bounded if and only if 𝑩 ⊂ 𝔓 (𝑚) for some 𝑚 and 𝑩 is bounded
in 𝔓 (𝑚) ; it follows that every closed bounded subset of 𝔓 (𝑚) is compact: 𝔓 (𝑚) is a
Montel space. For a linear map of 𝔓 into some locally convex TVS to be continuous
it is (necessary and) sufficient that its restriction to each 𝔓 (𝑚) be continuous. Since
𝔡𝔓 (𝑚) ⊂ 𝔓 (𝑚+1) the derivation 𝔡 is a continuous linear operator on 𝔓.
Let 𝑨 be a differential (associative) algebra with differential D. If 𝑃 (𝑢 0 , ..., 𝑢 𝑚 ) ∈
𝔓 and ℎ ∈ 𝑨 we shall use the notation

𝑃 [ℎ] = 𝑃 (ℎ, Dℎ, ..., D𝑚 ℎ) (19.4.6)

(𝑚 is the order of 𝑃 [ℎ]); we have D𝑃 [ℎ] = 𝔡𝑃 [ℎ]. We shall denote by 𝔓 [ℎ] the
set of elements 𝑃 [ℎ] of 𝑨 as 𝑃 ranges over 𝔓.
The following example motivates the reference to (19.4.2) as an abstract refor-
mulation of the standard “integration by parts” formula.
Example 19.4.3 Let 𝑨 = C ∞ S1 and 𝜃 be 1

∞ 1
the variable in S . Let 𝑃, 𝑄 ∈ 𝔓 and
𝑘 = 1, 2, ...; we have, for all 𝜑, 𝜓 ∈ C S ,
∫ 2𝜋 ∫ 2𝜋
𝜕𝑘 𝜕𝑘
𝑃 [𝜑 (𝜃)] 𝑄 [𝜓 (𝜃)] d𝜃 = (−1) 𝑘 𝑄 [𝜓 (𝜃)] 𝑃 [𝜑 (𝜃)] d𝜃.
0 𝜕𝜃 𝑘 0 𝜕𝜃 𝑘
𝜕𝑘
𝑃 [𝜑 (𝜃)] = 𝔡 𝑘 𝑃 [𝜑 (𝜃)]. Given an arbitrary 𝑓 ∈ C ∞ S1 , Fourier

We have 𝜕𝜃 𝑘 ∫ 2𝜋 𝜕𝑔
series expansion implies directly 0 𝑓 (𝜃) d𝜃 = 0 if and only if 𝑓 = 𝜕𝜃 for some
∞ 1

𝑔∈C S .
804 19 Classical Analytic Formalism

19.4.2 The variational derivative

Definition 19.4.4 Let 𝑃 ∈ 𝔓. The polynomial

∞
𝛿𝑃 ∑︁ 𝑗 𝑗 𝜕𝑃
= (−1) 𝔡 (19.4.7)
𝛿𝑢 𝑗=0
𝜕𝑢 𝑗

is called the variational derivative of 𝑃.

The importance of the operator 𝛿/𝛿𝑢 was first recognized in the works of
I.M. Gelfand and L.A. Dickey (see for instance [Gelfand-Dickey, 1978]); they named
𝛿𝑃 𝛿
𝛿𝑢 the “variational derivative” and used the notation 𝛿𝑢 . A simple example shows
why that choice of name was appropriate.
Take 𝑨 = S (R), the algebra of C ∞ functions in R rapidly decaying at ±∞, with
the differential 𝜕𝑥 = 𝜕/𝜕𝑥. If 𝑃 ∈ 𝔓◦ the map 𝑢 ↦→ 𝑃 [𝑢] is a continuous (nonlinear)
map of S (R) into itself. We associate to 𝑃 the nonlinear functional (often called an
action integral) ∫ +∞
S (Ω) ∋𝑢 ↦→ 𝐼 𝑃 (𝑢) = 𝑃 [𝑢 (𝑥)] d𝑥.
−∞

One could equally well take 𝑨 = C ∞ S1 and integration over the unit circle, as in

Example 19.4.3. ∫ +∞
We use the fact that −∞ 𝑓 (𝑥) d𝑥 = 0, 𝑓 ∈ S (R), if and only if there is a
𝑔 ∈ S (R) such∫ that 𝑓 = 𝜕𝑥 𝑔. The “if” part is obvious. To prove the “only if” part
𝑥
define 𝑔 (𝑥) = −∞ 𝑓 (𝑦) d𝑦. We have 𝑔 (𝑥) → 0 as 𝑥 → +∞; then 𝜕𝑥 𝑔 = 𝑓 ∈ S (R)
entails 𝑔 ∈ S (R). If 𝐼 𝑃 (𝑢) = 0 for every 𝑢 ∈ S (R) it is not difficult to show
that there is a polynomial 𝑄 ∈ 𝔓 such that 𝑃 = 𝔡𝑄. Conversely 𝑃 = 𝔡𝑄 implies
𝐼 𝑃 (𝑢) = 0 for every 𝑢 ∈ S (R). Thus, going to the quotient 𝔓◦ /𝔡𝔓 is the algebraic
equivalent of identifying those differential polynomials 𝑃 [·] that yield the same
values of 𝐼 𝑃 (𝑢) for each 𝑢 ∈ S (R).
∫ That is why the quotient map 𝔓◦ → 𝔓◦ /𝔡𝔓 is
often denoted by the integral sign .
The next observation belongs to the “calculus of variations”, in which case we
assume that the coefficients of 𝑃 and 𝑢 are real. A necessary condition for 𝑢 to be
𝛿
a local minimizer of 𝐼 𝑃 (𝑢) is that the Fréchet derivative 𝛿𝑢 𝐼 𝑃 (𝑢) of 𝐼 𝑃 (𝑢) at the
𝛿
“point” 𝑢 of S (R) vanishes. By definition 𝛿𝑢 𝐼 𝑃 (𝑢) is the following linear functional
on S (R):
∫ +∞ ∞ ∫ +∞
𝛿 d ∑︁
𝑗
𝜕𝑃
𝑣 ↦→ ⟨ 𝐼 𝑃 (𝑢) , 𝑣⟩ = 𝑃 [𝑢 + 𝜆𝑣] d𝑥 = 𝜕𝑥 𝑣 [𝑢] d𝑥.
𝛿𝑢 d𝜆 −∞ 𝜆=0 𝑗=0 −∞ 𝜕𝑢 𝑗

In the last expression the summation with respect to 𝑗 effectively stops at the order
of 𝑃. Integration by parts shows that
∞ ∫ +∞
𝛿 ∑︁
𝑗 𝜕𝑃
𝐼 𝑃 (𝑢) = (−1) 𝑗 𝑣 (𝑥) 𝜕𝑥 [𝑢 (𝑥)] d𝑥, 𝑢, 𝑣 ∈ S (R) ,
𝛿𝑢 𝑗=0 −∞ 𝜕𝑢 𝑗
19.4 Symbolic Calculus and the KdV Hierarchy 805

which can be abbreviated to

𝛿
𝐼 𝑃 (𝑢) = 𝐼 𝛿 𝑃/ 𝛿𝑢 (𝑢) .
𝛿𝑢

We return to the differential algebra 𝔓.

𝛿
Proposition 19.4.5 We have 𝔡𝔓 = 𝔓◦ ∩ ker 𝛿𝑢 .

𝛿
Proof First we show that 𝛿𝑢 𝔡𝑃 ≡ 0 for every 𝑃 ∈ 𝔡𝔓. Indeed,

𝜕 𝜕𝑃
(𝔡𝑃) = 𝔡 ,
𝜕𝑢 0 𝜕𝑢 0

𝑗 𝜕 𝑗 𝜕𝑃 𝑗+1 𝜕𝑃
(−𝔡) (𝔡𝑃) = (−𝔡) − (−𝔡) if 𝑗 ≥ 1.
𝜕𝑢 𝑗 𝜕𝑢 𝑗−1 𝜕𝑢 𝑗
𝛿 𝛿
Next we show that 𝔓◦ ∩ ker 𝛿𝑢 ⊂ 𝔡𝔓. Call 𝑚 the order of 𝑃 ∈ 𝔓◦ ∩ ker 𝛿𝑢 , 𝑃 ≠ 0.
If 𝑚 = 0 then
𝛿𝑃 𝜕𝑃
= = 0 ⇐⇒ 𝑃 = 0
𝛿𝑢 𝜕𝑢 0
since 𝑃(0) = 0.
Suppose 𝑚 ≥ 1. If 𝑃 is linear with respect to 𝑢 𝑚 there is a 𝑄 ∈ 𝔓 such
𝛿
that order (𝑃 − 𝔡𝑄) < 𝑚, 𝛿𝑢 (𝑃 − 𝔡𝑄) = 𝛿𝛿𝑢𝑃 and we can use induction on 𝑚.
If 𝑃 contains a monomial of the form 𝑐𝑢 0𝑝0 · · · 𝑢 𝑚
𝑝𝑚
, 𝑝 𝑚 ≥ 2, 0 ≠ 𝑐 ∈ C, then

𝑝0 𝑝𝑚 −1
𝑐 𝑝 𝑚 𝑢 0 · · · 𝑢 𝑚 𝑢 2𝑚 is a monomial in 𝔡 𝜕𝑢𝑚 and therefore 𝛿𝛿𝑢𝑃 ≠ 0.
𝑚 𝜕𝑃
□

Proposition 19.4.5 tells us that

𝛿
𝔡 𝛿𝑢
0 → C → 𝔓 → 𝔓◦ → 𝔓

is a (short) exact sequence (C stands for the constant polynomials).

𝛿
We point out, for use below, that 𝛿𝑢 transforms homogeneous polynomials of
degree 𝑚 into homogeneous polynomials of degree 𝑚 − 1.

19.4.3 Range of the variational derivative

𝛿 𝛿
We will need a characterization of 𝛿𝑢 𝔓 = 𝛿𝑢 𝔓◦ . This will now be achieved by
means of the formal linear differential operators
∞
𝛿𝑃 ∑︁ 𝑘 𝑘 𝑘− 𝑗 𝜕𝑃
= (−1) 𝔡 (19.4.8)
𝛿𝑢 𝑗 𝑘= 𝑗 𝑗 𝜕𝑢 𝑘
806 19 Classical Analytic Formalism

𝛿 𝛿 𝛿
( 𝑗 = 0, 1, 2, ...); 𝛿𝑢0 = 𝛿𝑢 . Each operator 𝛿𝑢 𝑗 maps 𝔓 into itself: the summation on
𝛿𝑃
the right in (19.4.8) stops at 𝑘 = order 𝑃 and 𝛿𝑢𝑘 ≡ 0 if 𝑘 > order 𝑃.
𝛿
Proposition 19.4.6 For a polynomial 𝑃 ∈ 𝔓 to belong to 𝛿𝑢 𝔓 it is necessary and
sufficient that
𝜕𝑃 𝛿𝑃
∀ 𝑗 ∈ Z+ , = . (19.4.9)
𝜕𝑢 𝑗 𝛿𝑢 𝑗
Proof I. Condition (19.4.9) is necessary. Suppose
∞
𝛿𝑄 ∑︁ 𝜕𝑄
𝑃= = (−1) ℓ 𝔡ℓ .
𝛿𝑢 ℓ=0 𝜕𝑢 ℓ

for some 𝑄 ∈ 𝔓. Then, by (19.4.3) applied to 𝜕𝑄/𝜕𝑢 ℓ ,

∞ min(𝑘,ℓ)
𝜕2𝑄

𝜕𝑃 ∑︁ ∑︁ ℓ ℓ−𝛼
= (−1) ℓ 𝔡
𝜕𝑢 𝑘 ℓ=0 𝛼=0
𝛼 𝜕𝑢 𝑘−𝛼 𝜕𝑢 ℓ

whence, by (19.4.8),
∞ ∞ min(𝑘,ℓ)
∑︁ ℓ 𝜕2𝑄

𝛿𝑃 ∑︁ ∑︁ 𝑘+ℓ 𝑘 𝑘+ℓ− 𝑗−𝛼
= (−1) 𝔡 .
𝛿𝑢 𝑗 𝑘= 𝑗 ℓ=0 𝑗 𝛼=0
𝛼 𝜕𝑢 𝑘−𝛼 𝜕𝑢 ℓ

𝜕2 𝑄
To shorten the notation we write 𝑄 𝑗,ℓ,𝑚 = 𝔡ℓ+𝑚− 𝑗 𝜕𝑢ℓ 𝜕𝑢𝑚 . Making the change of
summation index 𝛼 = 𝑘 − 𝑚 we obtain
∞ ∞ 𝑘
𝛿𝑃 ∑︁ ∑︁ 𝑘+ℓ 𝑘
∑︁ ℓ
= (−1) 𝑄 𝑗,ℓ,𝑚
𝛿𝑢 𝑗 ℓ=0 𝑘= 𝑗 𝑗 𝑘 −𝑚
𝑚=(𝑘−ℓ) +
𝑗 ∑︁∞ 𝑘
∑︁
𝑘+ℓ 𝑘 ℓ
∑︁
= (−1) 𝑄 𝑗,ℓ,𝑚
ℓ=0 𝑘= 𝑗
𝑗 𝑚=𝑘−ℓ 𝑘 − 𝑚
∞ ∑︁
∞ 𝑘
∑︁
𝑘+ℓ 𝑘 ∑︁ ℓ
+ (−1) 𝑄 𝑗,ℓ,𝑚 .
𝑗 𝑘 −𝑚
ℓ= 𝑗+1 𝑘= 𝑗 𝑚=(𝑘−ℓ) +

We have used the notation 𝑛+ = max (𝑛, 0) (𝑛 ∈ Z).

For ℓ ≤ 𝑗 we interchange summations with respect to 𝑘 and 𝑚:
∞ 𝑘
∑︁ ∞ ℓ+𝑚
∑︁
𝑘𝑘 ℓ ∑︁ ∑︁ 𝑘𝑘 ℓ
(−1) 𝑄 𝑗,ℓ,𝑚 = (−1) 𝑄 𝑗,ℓ,𝑚 ;
𝑘= 𝑗
𝑗 𝑚=𝑘−ℓ 𝑘 − 𝑚 𝑚= 𝑗−ℓ
𝑗 𝑘 −𝑚
𝑘=max( 𝑗,𝑚)

and we use the fact that

19.4 Symbolic Calculus and the KdV Hierarchy 807

ℓ+𝑚
∑︁ 𝑘 𝑘ℓ ℓ+𝑚 𝑚
(−1) = (−1) .
𝑗 𝑘 −𝑚 𝑗 −ℓ
𝑘=max( 𝑗,𝑚)

For ℓ > 𝑗; we have

∞ 𝑘
∑︁ 𝑘 ∑︁ ℓ
(−1) 𝑘+ℓ 𝑄 𝑗,ℓ,𝑚
𝑗 𝑘 −𝑚
𝑘= 𝑗 𝑚=(𝑘−ℓ) +
∞ ℓ+𝑚
∑︁ ∑︁ 𝑘 ℓ
= (−1) 𝑘+ℓ 𝑄 𝑗,ℓ,𝑚
𝑚=0 𝑘=max( 𝑗,𝑚)
𝑗 𝑘 −𝑚
∞ ℓ
∑︁
ℓ+𝑚
∑︁
𝑘 𝑘 +𝑚 ℓ
= (−1) (−1) 𝑄 𝑗,ℓ,𝑚 .
𝑗 𝑘
𝑚=0 𝑘=( 𝑗−𝑚) +

Here we use the fact that

ℓ
∑︁
𝑘 𝑘 +𝑚 ℓ
(−1) = 0.
𝑗 𝑘
𝑘=( 𝑗−𝑚) +

In summary we have shown that

𝑗 ∞
𝜕2𝑄

𝛿𝑃 ∑︁ ∑︁ 𝑚
= (−1) 𝑚 𝔡ℓ+𝑚− 𝑗 .
𝛿𝑢 𝑗 ℓ=0 𝑚= 𝑗−ℓ 𝑗 −ℓ 𝜕𝑢 ℓ 𝜕𝑢 𝑚

The change of summation variables ℓ = 𝑗 − 𝛼 combined with (19.4.3) yields

𝑗 ∞
𝜕2𝑄

𝛿𝑃 ∑︁ ∑︁ 𝑚 𝑚 𝑚−𝛼
= (−1) 𝔡
𝛿𝑢 𝑗 𝛼=0 𝑚=𝛼 𝛼 𝜕𝑢 𝑗−𝛼 𝜕𝑢 𝑚
∞ min( 𝑗,𝑚)
𝜕2𝑄

∑︁
𝑚
∑︁ 𝑚 𝑚−𝛼
= (−1) 𝔡
𝑚=0 𝛼=0
𝛼 𝜕𝑢 𝑚 𝜕𝑢 𝑗−𝛼
∞
𝜕 ∑︁ 𝜕𝑄 𝜕𝑃
= (−1) 𝑚𝔡 𝑚 = .
𝜕𝑢 𝑗 𝑚=0 𝜕𝑢 𝑚 𝜕𝑢 𝑗

II. Condition (19.4.9) is sufficient. First consider the case of a homogeneous poly-
nomial 𝑃 of degree 𝜅 ≥ 0. We make use of the Euler homogeneity vector field
∞
𝜕 ∑︁ 𝜕
= 𝑢𝑗 .
𝜕 𝜌 𝑗=0 𝜕𝑢 𝑗
808 19 Classical Analytic Formalism

If (19.4.9) is valid we have

∞
𝜕𝑃 ∑︁ 𝛿𝑃
= 𝑢𝑗
𝜕𝜌 𝑗=0
𝛿𝑢 𝑗
∞ ∑︁
∞
∑︁ 𝑘 𝜕𝑃
= (−1) 𝑘 𝔡 𝑗 𝑢 0 𝔡 𝑘− 𝑗
𝑗=0 𝑘= 𝑗
𝑗 𝜕𝑢 𝑘
∞ 𝑘
∑︁
𝑘
∑︁ 𝑘 𝑗 𝑘− 𝑗 𝜕𝑃
= (−1) 𝔡 𝑢0 𝔡
𝑘=0 𝑗=0
𝑗 𝜕𝑢 𝑘
∞ ∑︁∞
∑︁ 𝜕𝑃 𝜕
= (−1) 𝑘 𝔡 𝑘 𝑢 0 = (−1) 𝑘 𝔡 𝑘 (𝑢 0 𝑃) − 𝑃,
𝑘=0
𝜕𝑢 𝑘 𝑘=0
𝜕𝑢 𝑘

i.e.,
𝛿 𝜕𝑃
𝑃= (𝑢 0 𝑃) − .
𝛿𝑢 𝜕𝜌
𝜕𝑃
By the Euler homogeneity relation 𝜕𝜌 = 𝜅𝑃 we derive

𝛿 1
𝑃= 𝑢0 𝑃 .
𝛿𝑢 𝜅 + 1

If 𝑃 is arbitrary (ie, not necessarily homogeneous) we decompose 𝑃 into its homo-

geneous components: 𝑃 = 𝑑𝜅=0 𝑃 𝜅 , which allows us to write 𝑃 = 𝛿𝑄
Í
𝛿𝑢 with

𝑑
∑︁ 1
𝑄 = 𝑢0 𝑃𝜅 . (19.4.10)
𝜅=0
𝜅+1

□
𝛿𝔓
If 𝑃 ∈𝛿𝑢 the algebraic functional 𝑃 [𝑢], 𝑢 in some function differential algebra,
is often called a hamiltonian (noun). The evolution equation 𝜕𝑡 𝑢 = 𝔡𝑃 [𝑢] is then
said to be Hamiltonian (adjective).
We list a few simple formulas about the differential operators (19.4.8). The proofs
are routine and we skip them. First, the symmetry between the partial derivatives
𝜕 𝛿
𝜕𝑢 𝑗 and the operators 𝛿𝑢 𝑗 :

∞
𝜕𝑃 ∑︁ 𝑘 𝑘 𝑘− 𝑗 𝛿𝑃
= (−1) 𝔡 , (19.4.11)
𝜕𝑢 𝑗 𝑘= 𝑗 𝑗 𝛿𝑢 𝑘

for all 𝑗 ∈ Z+ and 𝑃 ∈ 𝔓.

Next, what replaces the Leibniz rule for 𝛿/𝛿𝑢:
∞
𝛿 ∑︁ 𝛿𝑄 𝛿𝑃
(𝑃𝑄) = (𝔡 𝑗 𝑃) + (𝔡 𝑗 𝑄) ; (19.4.12)
𝛿𝑢 𝑗=0
𝛿𝑢 𝑗 𝛿𝑢 𝑗
19.4 Symbolic Calculus and the KdV Hierarchy 809

for all 𝑃, 𝑄 ∈ 𝔓.
Last, we observe that the property
𝛿𝑄
𝜗𝑃 𝑄 − 𝑃 ∈ 𝔡𝔓 (19.4.13)
𝛿𝑢
is a direct consequence of the “integration by parts” formula (19.4.4).
𝛿 𝛿
Keep in mind that 𝛿𝑢 𝔡 ≡ 0 by Proposition 19.4.5 and that 𝛿𝑢 and 𝜕𝑢𝜕 0 commute
since 𝜕𝑢𝜕 0 commutes with 𝔡 and 𝜕𝑢𝜕 𝑗 , 𝑗 ≥ 1.

19.4.4 Formal analytic series with coefficients in 𝕻

We shall make use of the following norm, obviously continuous with respect to the
natural topology of 𝔓:
∞
∑︁ ∑︁ 1 𝜕𝛼𝑃
∥𝑃∥ 𝔓 = |𝑃 (0)| + (0) . (19.4.14)
𝑚=0 𝛼∈Z+𝑚+1 , 𝛼𝑚 ≥1
𝛼! 𝜕𝑢 𝛼

𝛼0 𝛼𝑚
𝜕𝛼 𝜕 𝜕
Keep in mind that 𝛼 = (𝛼0 , ..., 𝛼𝑚 ) and 𝜕𝑢 𝛼 = 𝜕𝑢0 · · · 𝜕𝑢𝑚 . It is not true
that the derivation 𝔡 is continuous with respect to the norm (19.4.14):

sup ∥𝔡𝑃∥ 𝔓 /∥𝑃∥ 𝔓 ↗ +∞ as 𝑚 ↗ +∞.

0≠𝑃 ∈𝔓 (𝑚)

Indeed, take 𝑃 (𝑢) = 𝑢 𝑘𝑝 ; then ∥𝑃∥ 𝔓 = 1 while ∥𝔡𝑃∥ 𝔓 = 𝑝. This last equation,
however, has the following immediate, yet important, consequence:

∀𝑃 ∈ 𝔓, ∥𝔡𝑃∥ 𝔓 ≤ (deg 𝑃) ∥𝑃∥ 𝔓 . (19.4.15)

(𝑚)
Definition 19.4.7 WeÍshall denote by Aform (𝔓) (𝑚 ∈ Z) the linear space of formal
power series 𝔞 (𝜆) = ∞𝑗=0 𝑃 𝑗 𝜆
𝑚− 𝑗 , 𝑃 ∈ 𝔓, that satisfy the following condition:
𝑗

(FA𝔓) There are positive constants 𝑐 ◦ , 𝐶◦ , such that, for every 𝑗 ∈ Z+ , deg 𝑃 𝑗 ≤
𝑗+1
𝑐 ◦ ( 𝑗 + 1) and 𝑃 𝑗 𝔓 ≤ 𝐶◦ 𝑗!.

If 𝑃0 ≠ 0 we refer to 𝑚 as the order of 𝔞. We shall use the notation Aform (𝔓) =

Ð (𝑚)
𝑚∈Z A form (𝔓).

Proposition 19.4.8 For every 𝑚 ∈ Z the derivations 𝜕/𝜕𝜆, 𝔡, 𝜗𝑃 (𝑃 ∈ 𝔓 arbitrary),

(𝑚)
extend as derivations in Aform (𝔓).
Í∞
Proof Let 𝔞 (𝜆) = 𝑗=0 𝑃 𝑗 𝜆 𝑚− 𝑗 ∈ Aform (𝔓) be arbitrary. By (19.4.15) we have,
for every 𝑗,
810 19 Classical Analytic Formalism
𝑗+1
𝔡𝑃 𝑗 𝔓
≤ wght 𝑃 𝑗 𝑃𝑗 𝔓
≤ 𝑐 ◦ ( 𝑗 + 1) 𝐶◦ 𝑗!,
the last inequality a consequence of (FA𝔓). The argument for 𝜕/𝜕𝜆, 𝜗𝑃 is similar.□
The following concept is the key to the characterization of the range of 𝜕/𝜕𝜆.

Definition 19.4.9 By the residue of a series 𝔞 = ∞

Í
form (𝔓) we
𝑚− 𝑗 ∈ A
𝑗=0 𝑃 𝑗 𝜆
−1
mean the coefficient of 𝜆 , 𝑃𝑚+1 ∈ 𝔓; 𝑃𝑚+1 shall be denoted by Res 𝔞.

Proposition 19.4.10 The two following properties of 𝔞 ∈ Aform (𝔓) are equivalent:
(1) Res 𝔞 = 0;
(2) there is a 𝔟 ∈ Aform (𝔓) such that 𝔞 = 𝜕𝔟/𝜕𝜆.

Proof That (2) implies (1) is obvious. If 𝔞 = ∞

Í 𝑚− 𝑗 , 𝑃
𝑗=0 𝑃 𝑗 𝜆 𝑚+1 = 0, then 𝔞 = 𝜕𝜆 𝔟,
Í∞ 1
𝔟 = 𝑗=0 𝑚+1− 𝑗 𝑃 𝑗 𝜆 𝑚+1− 𝑗 , and it is immediate that 𝔟 satisfies (FA𝔓). □

Let us denote by 𝔓 𝜆−1 the ring of formal series in the powers of 𝜆−1 with

coefficients in 𝔓. Given two formal series 𝔞 (𝜆) = ∞ 𝑚− 𝑗 ∈ 𝜆 𝑚 𝔓 𝜆 −1 ,

Í
𝑗=0 𝑃 𝑗 𝜆
𝔟 (𝜆) = +∞
Í 𝑚′ − 𝑗 ∈ 𝜆 𝑚′ , 𝑚, 𝑚 ′ ∈ Z, we introduce the composition law [cf.
𝑗=0 𝑄 𝑗 𝜆
(19.1.12)]:
∞
∑︁ 1 𝑘
(𝔞#𝔟) (𝜆) = 𝜕𝜆 𝔞 (𝜆) 𝔡 𝑘 𝔟 (𝜆) (19.4.16)
𝑘=0
𝑘!
∞ 𝑗 ∑︁
𝑗−𝑘
!
∑︁ ∑︁
𝑚+𝑚′ − 𝑗
= 𝜆 𝑁 (𝑚 − ℓ, 𝑘) 𝑃ℓ 𝔡 𝑘 𝑄 𝑗−𝑘−ℓ ,
𝑗=0 𝑘=0 ℓ=0

1 −𝑚+𝑘+ℓ

where 𝑁 (𝑚 − ℓ, 𝑘) = 𝑘! 𝜆 𝜕𝜆𝑘 𝜆 𝑚−ℓ ∈ Z. It is useful to reformulate (19.4.16)
in the case 𝑚 = 𝑚 ′ = 0:
∞ 𝑗 ∑︁
𝑗−𝑘
∑︁
−𝑗
∑︁ (𝑘 + ℓ − 1)!
(𝔞#𝔟) (𝜆) = 𝑃0 𝑄 0 + 𝜆 (−1) 𝑘 𝑃ℓ 𝔡 𝑘 𝑄 𝑗−𝑘−ℓ . (19.4.17)
𝑗=1 𝑘=0 ℓ=1
𝑘! (ℓ − 1)!

(𝑚) (𝑚)
Proposition 19.4.11 Given two formal series 𝔞 ∈ Aform (𝔓), 𝔟 ∈ Aform (𝔓), then
(𝑚)
(𝔞#𝔟) (𝜆) ∈ Aform (𝔓).
′
Proof If we define 𝔞◦ = 𝜆−𝑚𝔞 (𝜆), 𝔟◦ = 𝜆−𝑚 𝔟 (𝜆), then
′

𝔞#𝔟 = (𝔞◦ 𝜆 𝑚 ) # 𝔟◦ 𝜆 𝑚
′
= (𝔞◦ # (𝜆 𝑚 #𝔟◦ )) 𝜆 𝑚 .
19.4 Symbolic Calculus and the KdV Hierarchy 811

Suppose 𝑚 ≥ 1; in this case,

𝑚
∑︁ 𝑚!
𝜆 𝑚 #𝔟◦ (𝜆) = 𝜆 𝑚−𝑘 𝔡 𝑘 𝑏 ◦ (𝜆) ,
𝑘=0
𝑘! (𝑚 − 𝑘)!

whence
𝑚
∑︁ 𝑚 ′
𝔞#𝔟 = 𝜆 𝑚+𝑚 −𝑘 𝔞◦ #𝔡 𝑘 𝑏 ◦ .
𝑘=0
𝑘
(0)
Since 𝔡 𝑘 𝑏 ◦ ∈ Aform (𝔓) for all 𝑘 we see that it suffices to prove that 𝔞, 𝔟 ∈Aform (𝔓)
(0)
implies 𝔞#𝔟 ∈Aform (𝔓), 𝔞#𝔟 as in (19.4.17).
By (FA𝔓) and (19.4.15), since

deg 𝔡 𝑘 𝑄 𝑗−𝑘−ℓ = deg 𝑄 𝑗−𝑘−ℓ ≤ wght 𝑄 𝑗−𝑘−ℓ ≤ 𝛾 ( 𝑗 − 𝑘 − ℓ)

for some 𝛾 > 0, we derive

𝑘
∥𝑃ℓ ∥ 𝔓 𝔡 𝑘 𝑄 𝑗−𝑘−ℓ 𝔓
≤ 𝐶◦ℓ+1 ℓ! deg 𝑄 𝑗−𝑘−ℓ 𝑄
𝑗−𝑘−ℓ 𝔓
𝑘 𝑗−𝑘−ℓ+1
≤ 𝐶◦ℓ+1 𝛾 𝑘 ( 𝑗 − 𝑘 − ℓ) 𝐶1 ℓ! ( 𝑗 − 𝑘 − ℓ)!
𝑗−𝑘
≤ 𝐶2 ( 𝑗 − 𝑘 − ℓ) 𝑘 ℓ! ( 𝑗 − 𝑘 − ℓ)!

with 𝐶1 < 𝐶2 suitably large. Thus, for 𝑗 ≥ 1,

𝑗 ∑︁
𝑗−𝑘
∑︁ (𝑘 + ℓ − 1)!
∥𝑃ℓ ∥ 𝔓 𝔡 𝑘 𝑄 𝑗−𝑘−ℓ 𝔓
𝑘=0 ℓ=1
𝑘! (ℓ − 1)!
𝑗 𝑗−𝑘
∑︁ ∑︁ (𝑘 + ℓ − 1)!
𝐶2−𝑘
𝑗
≤ 𝐶2 ℓ ( 𝑗 − 𝑘 − ℓ) 𝑘 ( 𝑗 − 𝑘 − ℓ)!
𝑘=0 ℓ=1
𝑘!
𝑗−1 𝑗
∑︁ ∑︁ (ℓ − 1)!
𝐶2−𝑘
𝑗
≤ 𝐶2 (ℓ − 𝑘) ( 𝑗 − ℓ) 𝑘 ( 𝑗 − ℓ)!
𝑘=0 ℓ=𝑘+1
𝑘!
𝑗 ℓ−1
𝑗
∑︁ ∑︁ 1 −𝑘
≤ 𝐶2 ( 𝑗 − ℓ)!ℓ! 𝐶2 ( 𝑗 − ℓ) 𝑘
ℓ=1 𝑘=0
𝑘!
𝑗
∑︁
( 𝑗 − ℓ)!ℓ!e ( 𝑗−ℓ)/𝐶2 ≤ 𝐶3
𝑗 𝑗+1
≤ 𝐶2 𝑗!,
ℓ=1

the last inequality a direct consequence of Lemma 17.2.14. □

Thus Aform (𝔓) is an associative (but not commutative) ring with respect to
(0)
addition and #; Aform (𝔓) is a subring of Aform (𝔓).
812 19 Classical Analytic Formalism

Proposition 19.4.12 For every 𝑚 ∈ Z the derivations 𝜕/𝜕𝜆, 𝔡, 𝜗𝑃 (𝑃 ∈ 𝔓 arbitrary),

extend as derivations in the algebra Aform (𝔓) with respect to the composition law
#.

This follows directly from the fact that 𝜕/𝜕𝜆, 𝔡, 𝜗𝑃 pairwise commute [cf.
(19.4.5)].
In connection with residues and the law # the following proposition expresses an
important property of commutators:

Proposition 19.4.13 For all 𝔞1 , 𝔞2 ∈ Aform (𝔓), Res [𝔞1 , 𝔞2 ] # ∈ 𝔡𝔓.

Proof It suffices to deal with 𝔞𝑖 = 𝜆 𝑚𝑖 𝑃𝑖 (𝑃𝑖 ∈ 𝔓, 𝑚 𝑖 ∈ Z, 𝑖 = 1, 2). We have

∞
∑︁ 1 𝜕𝑘 𝜕𝑘
[𝔞1 , 𝔞2 ] # = 𝑃1 𝜕 𝑘 𝑃2 𝜆 𝑚2 𝑘 𝜆 𝑚1 − 𝜕 𝑘 𝑃1 𝑃2 𝜆 𝑚1 𝑘 𝜆 𝑚2 .
𝑘=1
𝑘! 𝜕𝜆 𝜕𝜆

But clearly, by the Leibniz rule,

𝑘
𝑘

𝑚1 𝜕 𝑚2 𝑘 𝑚2 𝜕 𝑚1
Res 𝜆 𝜆 = (−1) Res 𝜆 𝜆 ,
𝜕𝜆 𝑘 𝜕𝜆 𝑘

𝑃1 𝜕 𝑘 𝑃2 − (−1) 𝑘 𝜕 𝑘 𝑃1 𝑃2 ∈ 𝜕𝔓,

whence the claim. □

We now introduce a special subset of Aform (𝔓):

Definition 19.4.14 Let 𝑚 ∈ Z. We shall say that 𝔞 = ∞

Í 𝑚− 𝑗 ∈ A (𝑚) (𝔓) is
𝑗=0 𝑃 𝑗 𝜆 form
purely monic of order 𝑚 if 𝑃0 = 1 and if 𝑃 𝑗 ∈ 𝔓◦ for all 𝑗 ≥ 1.

Proposition 19.4.15 If 𝔞 ∈ Aform (𝔓) is purely monic of order 𝑚 then it has an

inverse in Aform (𝔓), 𝔞 #−1 ; 𝔞 #−1 is purely monic of order −𝑚.

Proof The proof is the same as that of Theorem 19.1.17 after replacing absolute
values by the norm (19.4.14) and taking (FA𝔓) into account. □
(0)
The purely monic symbols in Aform (𝔓) [resp., Aform (𝔓)] form a group with
(0)
respect to # which we shall denote by GAform (𝔓) [resp., GAform (𝔓)]. We shall
(𝑚)
denote by GAform (𝔓) the set of purely monic symbols of order 𝑚.

19.4.5 Formal analytic series with coefficients in 𝕻 [𝒉]

Let Ω be an open subset of the complex plane, ℎ ∈ O (Ω) and 𝑃 (𝑢 0 , ..., 𝑢 𝑚 ) ∈ 𝔓,

we use the notation (19.4.6) with D = D𝑧 = −𝑖𝜕/𝜕𝑧:

𝑃 [ℎ] = 𝑃 ℎ, D𝑧 ℎ, ..., D𝑚
𝑧 ℎ (19.4.18)
19.4 Symbolic Calculus and the KdV Hierarchy 813

and we denote by 𝔓 [ℎ] the set of functions 𝑃 [ℎ] ∈ O (Ω). We have 𝔡𝑃 [ℎ] =
D𝑧 𝑃 [ℎ]; 𝔓 [ℎ] is a differential algebra with respect to addition and ordinary mul-
tiplication, with derivation D𝑧 and unit element the function identically equal to
1.

Lemma 19.4.16 Let 𝑃 (𝑢 0 , ..., 𝑢 𝑚 ) ∈ 𝔓 and ℎ ∈ O (Ω). To every compact subset 𝐾

of Ω there is a 𝐶𝐾 > 0 such that
𝑘
max |𝑃 [ℎ]| ≤ 𝐶𝐾 𝜀 −1 max |ℎ (𝑧)| ∥𝑃∥ 𝔓 (19.4.19)
𝐾 dist(𝑧,𝐾) ≤ 𝜀

if 𝑘 = wght 𝑃, 𝜀 > 0, 𝜀 < min (1, dist (𝐾, 𝜕Ω)).

Proof We have
∞
∑︁ ∑︁ 1 𝜕𝛼𝑃 𝛼
𝑃 [ℎ] = (0) ℎ 𝛼0 (D𝑧 ℎ) 𝛼1 · · · D𝑧𝑝 ℎ 𝑝 . (19.4.20)
𝑝=0 𝛼∈Z 𝑝+1
𝛼! 𝜕𝑢 𝛼
+

The Cauchy inequalities imply, for some 𝐵 𝐾 ≥ 1,

𝛼𝑝
max D𝑧𝑝 ℎ 𝑝 ≤ 𝐵 𝐾 𝜀 − 𝑝 𝑝!max |ℎ|
𝛼
,
𝐾 𝐾𝜀

whence
|𝛼|
𝐵 𝐾| 𝛼 | 𝜛 ( 𝛼)
𝛼
max |ℎ| 𝛼0
|D𝑧 ℎ| 𝛼1
··· D𝑧𝑝 ℎ 𝑝 ≤ ( 𝑝/𝜀𝑒) max |ℎ| ,
𝐾 𝐾𝜀

implying easily (19.4.19). □

Í∞
Corollary 19.4.17 Suppose 𝔞 (𝜆) = 𝑗=0 𝑃 𝑗 𝜆 𝑚− 𝑗 satisfies (FA𝔓). Let Ω be an open
subset of the complex plane, ℎ ∈ O (Ω). To every compact subset 𝐾 of Ω there is a
𝐶𝐾 > 0 such that
𝑗+1
∀ 𝑗 ∈ Z+ , max 𝑃 𝑗 [ℎ] ≤ 𝐶𝐾 𝑗!. (19.4.21)
𝐾

Proof Combining (FA𝔓) and Lemma 19.4.16 yields

𝜅 𝑗
max 𝑃 𝑗 [ℎ] ≤ 𝐶𝐾 𝜀 −1
𝑗+1
max |ℎ (𝑧)| 𝐶◦ 𝑗!. □
𝐾 dist(𝑧,𝐾) ≤ 𝜀

Í∞
The meaning of (19.4.21) is that 𝑎 (𝑧, 𝜆) = 𝑗=0 𝑃 𝑗 [ℎ (𝑧)] 𝜆 𝑚− 𝑗 belongs to
eform (Ω) [cf. (19.1.2)].
A

Definition 19.4.18 We shall denote 𝑚 (Ω; 𝔓 [ℎ]) the subset of A eform (Ω) con-
Í∞ by 𝑆class
sisting of symbols 𝑎 (𝑧, 𝜆) = 𝑗=0 𝑎 𝑗 (𝑧) 𝜆 𝑚− 𝑗 such that 𝑎 𝑗 ∈ 𝔓 [ℎ] for every 𝑗 ∈ Z+ .
We shall denote by 𝑆class (Ω; 𝔓 [ℎ]) the union of the linear spaces 𝑆class
𝑚 (Ω; 𝔓 [ℎ]),

𝑚 ∈ Z.
814 19 Classical Analytic Formalism

We underline the fact that 𝑎 (𝑧, 𝜆) = ∞

Í
𝑗=0 𝑃 𝑗 [ℎ (𝑧)] 𝜆 class (Ω; 𝔓 [ℎ])
𝑚− 𝑗 ∈ 𝑆

must still satisfy (FA). In dealing with 𝑎 (𝑧, 𝜆) we may allow 𝜆 to vary in a sector
Γ 𝜀 = {𝜁 = ℎ + 𝑖𝜂 ∈ C; |𝜂| < 𝜀ℎ} and thus view 𝑎 (𝑧, 𝜁) = ∞
Í
𝑃
𝑗=0 𝑗 [ℎ (𝑧)] 𝜁 𝑚− 𝑗 as
a classical analytic symbol in Ω × Γ 𝜀 ; but it must be kept in mind that 𝜀 > 0 depends
on 𝑎 (𝑧, 𝜆), meaning on the constant 𝐶𝐾 in (19.4.21). Another way of looking at
𝑚 (Ω; 𝔓 [ℎ]) is to regard them as classical analytic symbols
the elements of 𝑆class
in (Ω ∩ R) × (0, +∞) (cf. Definition 17.2.28; also the present chapter, Definition
19.1.14). In both these interpretations the composition law (19.4.16) induced on
𝑆class (Ω; 𝔓 [ℎ]) is the same as that induced by (16.2.16).

Proposition 19.4.19 If 𝑎 and 𝑏 belong to 𝑆class (Ω; 𝔓 [ℎ]) so do 𝑎#𝑏, D𝑧 𝑎, 𝜕𝜆 𝑎, 𝜗𝑃 𝑎

(𝑃 ∈ 𝔓). Moreover, D𝑧 , 𝜕𝜆 , 𝜗𝑃 , are pairwise commuting derivations of the algebra
𝑆class (Ω; 𝔓 [ℎ]) with respect to the composition law #.

This follows directly from

Í the properties of Aform (𝔓) (cf. Proposition 19.4.12).
We shall say that 𝑎 = ∞ 𝑗=0 𝜆 𝑚− 𝑗 𝑃 [ℎ] ∈ 𝑆 𝑚 (Ω; 𝔓 [ℎ]) is purely monic if
𝑗 class
Í∞ 𝑚− 𝑗
𝑗=0 𝜆 𝑃 𝑗 is purely monic in the sense of Definition 19.4.9. Proposition 19.4.15
yields directly

Proposition 19.4.20 If 𝑎 ∈ 𝑆class

𝑚 (Ω; 𝔓 [ℎ]) is purely monic then it has an inverse
𝑎 #−1 −𝑚
∈ 𝑆class (Ω; 𝔓 [ℎ]); 𝑎 #−1 is also purely monic.

The purely monic symbols in 𝑆class (Ω; 𝔓 [ℎ]) form a group for # which shall be
denoted by G𝑆class (Ω; 𝔓 [ℎ]); we define, for each 𝑚 ∈ Z,
𝑚 𝑚
G𝑆class (Ω; 𝔓 [ℎ]) = 𝑆class (Ω; 𝔓 [ℎ]) ∩ G𝑆class (Ω; 𝔓 [ℎ]) ; (19.4.22)
0
G𝑆class (Ω; 𝔓 [ℎ]) is a subgroup of G𝑆class (Ω; 𝔓 [ℎ]). Keep in mind that, through
all of this, ℎ ∈ O (Ω) is arbitrary.

19.4.6 Evolution equations and conserved quantities

We relate some of the material presented in the preceding subsections to an evolution

equation
𝜕𝑡 ℎ = 𝑃 [ℎ] (19.4.23)
where 𝑃 ∈ 𝔓 and ℎ = ℎ (𝑡) is a C 1 and, most of the time, a C ∞ function of the
“time” 𝑡, valued in a differential algebra 𝑨 (commutative, with the derivation D); we
assume that 𝑨 is equipped with a locally convex topology for which D is continuous
and the multiplication is a continuous bilinear map 𝑨×𝑨 −→ 𝑨. Mostly we think
of 𝑡 as a real variable but it could also be a complex variable in a domain Ω of C and
ℎ ∈ O (Ω).
It is traditional to think of the map 𝑡 → ℎ(𝑡) as a trajectory inside the algebra
𝑨. Keep in mind that, a priori, ℎ, Dℎ, D2 ℎ, ... is an infinite sequence of unknowns.
The finite-dimensional analogue is a system of ODEs
19.4 Symbolic Calculus and the KdV Hierarchy 815

dx
= 𝐹 (x) (19.4.24)
d𝑡
where 𝐹 is a map R𝑛 −→ R𝑛 and R ∋𝑡 ↦→ x (𝑡) is valued in R𝑛 . Now suppose that the
trajectories of (19.4.24) can be defined by “implicit” equations 𝐹 (𝑥 1 , ..., 𝑥 𝑛 ) = const.
The classical way of expressing this is by saying that the function 𝐹 (𝑥1 , ..., 𝑥 𝑛 ) is a
constant of motion or alternatively, a first integral of the ODE (19.4.24).
If there are first integrals of the infinite-dimensional “system” (19.4.23) they must
be functionals not only of the solution 𝜑 but also of its derivatives D 𝑗 𝜑. Let us go
back again to the special cases 𝑨= S (R) ∫or 𝑨= C ∞ S1 and form, for a polynomial
𝑄 ∈ 𝔓, the definite integral 𝐼𝑄 (𝜑; 𝑡) = X 𝑄 [𝜑 (𝑡, 𝑥)] d𝑥 (with X = R or X = S1 ).
At first sight the right thing to do might seem to call 𝐼𝑄 (𝜑; 𝑡) a constant of motion
or conserved quantity of the evolution equation (19.4.23) when

d
𝐼𝑄 (𝜑; 𝑡) = 0 (19.4.25)
d𝑡
for every solution 𝜑 ∈ C ∞ (X; 𝑨) of (19.4.23). Explicitly (19.4.25) means that
𝑛 ∫
∑︁ 𝜕𝑄
𝑗
𝜕𝑥 𝜕𝑡 𝜑 [𝜑 (𝑡, 𝑥)] d𝑥 = 0
𝑗=0 X 𝜕𝑢 𝑗

or equivalently, by letting 𝜕𝑥 act on both sides of (19.4.23), that

𝑛 ∫
∑︁ 𝜕𝑄
𝑗
𝜕𝑥 𝑃 [𝜑 (𝑡, 𝑥)] [𝜑 (𝑡, 𝑥)] d𝑥 = 0. (19.4.26)
𝑗=0 X 𝜕𝑢 𝑗

With our two choices of 𝑨 integration by parts is permitted and (19.4.26) is equivalent
to
𝑛 ∫
𝑗 𝜕𝑄
∑︁
(−1) 𝑗 𝑃 [𝜑 (𝑡, 𝑥)] 𝜕𝑥 [𝜑 (𝑡, 𝑥)] d𝑥 = 0. (19.4.27)
𝑗=0 X 𝜕𝑢 𝑗

The trouble is that we do not really have access to every solution of (19.4.23).
In practice, the only way to prove that (19.4.26) and (19.4.27) hold is by showing
that the integrands are “divergences”, which is to say, that there exist polynomials
Φ, Ψ ∈ 𝔓 such that
𝑛
∑︁ 𝜕𝑄
D 𝑗 (𝑃 [ℎ]) [ℎ] = D (Φ [ℎ]) (19.4.28)
𝑗=0
𝜕ℎ 𝑗

or, after “integration by parts”,

𝑛
∑︁ 𝜕𝑄
𝑃 [ℎ] (−1) 𝑗 D 𝑗 [ℎ] = D (Ψ [ℎ]) ; (19.4.29)
𝑗=0
𝜕ℎ 𝑗

and not just for all solutions of (19.4.23) but actually for all ℎ ∈ 𝑨.
816 19 Classical Analytic Formalism

To attain maximum generality (i.e., independence from the choice of the function
algebra 𝑨) we avail ourselves of the chain-rule derivation 𝔡 and the definition (19.4.4)
to rewrite (19.4.28) as
𝜗𝑃 𝑄 = 𝔡Φ (19.4.30)
and (19.4.29) as
𝛿𝑄
𝑃 = 𝔡Ψ. (19.4.31)
𝛿𝑢
(Φ and Ψ are often called fluxes.) This completes the elimination of the analytic
operation of definite integral over 𝑥-space; 𝑥 does not figure any more in the equations.
Integration by parts is simply replaced by the Leibniz rule. The equivalence of
(19.4.30) and (19.4.31) (given Φ for each choice of Ψ, and vice versa) is a direct
consequence of (19.4.13).
We may now introduce the mathematically correct
Definition 19.4.21 A polynomial 𝑄 ∈ 𝔓 shall be called a conserved polynomial of
the evolution equation (19.4.23) if 𝜗𝑃 𝑄 or, equivalently, 𝑃 𝛿𝑄
𝛿𝑢 belong to 𝔡𝔓.
If 𝑄 ∈ 𝔓 is a conserved polynomial of the evolution equation (19.4.23) the
functions 𝑄 [ℎ], with ℎ a solution of (19.4.23) in a suitable function algebra 𝑨, are
often referred to as constants of motions or first integrals of (19.4.23); these names
may also be applied directly to 𝑄. By Proposition 19.4.5 the properties 𝑃 𝛿𝑄 𝛿𝑢 ∈ 𝔡𝔓

𝛿 𝛿𝑄
and 𝛿𝑢 𝑃 𝛿𝑢 = 0 are equivalent.
Proposition 19.4.22 Every 𝑄 ∈ 𝔡𝔓 is a conserved polynomial of the evolution
equation (19.4.23) whatever 𝑃 ∈ 𝔓.
𝛿
Proof Indeed, 𝜗𝑃 𝔡𝑅 = 𝔡𝜗𝑃 𝑅 or, equivalently, 𝑃 𝛿𝑢 𝔡𝑅 = 0 (Proposition 19.4.5). □
Proposition 19.4.22 tells us that conserved polynomials of the form 𝔡Φ may be
disregarded.
Proposition 19.4.23 Each 𝑄 ∈ 𝔓 is a conserved polynomial of the evolution equa-
tion (19.4.23) in which 𝑃 = 𝔡 𝛿𝑄
𝛿𝑢 .

2
1 𝛿𝑄
Proof Indeed, 𝛿𝑄 𝛿𝑢 𝔡 𝛿𝑄
𝛿𝑢 = 𝔡 2 𝛿𝑢 . □

In the remainder of this section we shall describe a special sequence of polyno-

mials of the form 𝔡 𝛿𝑄
𝛿𝑢 .

19.4.7 The stabilizer of the Sturm–Liouville symbol

We go back to the group GAform (𝔓). By the Sturm–Liouville symbol we mean

𝐿 = 𝜆2 − 𝑢 0 ∈ GAform (𝔓); in standard notation 𝜉 2 − 𝑢 0 is the symbol of the
b
differential operator 𝐿 = −𝜕𝑥2 − 𝑢 0 ; when 𝑢 0 is a given function of the single variable
𝑥 (possibly complex) −𝐿 𝑓 = 𝑔 is the simplest version of the classical (in the literal
sense) Sturm–Liouville differential equation.
19.4 Symbolic Calculus and the KdV Hierarchy 817

We will now characterize the stabilizer of b

𝐿, by which we mean the set of
Í∞ 𝑚− 𝑗 (𝑚)
𝔞 (𝜆) = 𝜆 + 𝑗=1 𝜆
𝑚 𝑃 𝑗 ∈ GAform (𝔓) (𝑚 ∈ Z) such that 𝔞#b𝐿#𝔞 #−1 = b
𝐿, i.e.,
h i
𝐿 = 0. We will make use of the following
𝔞, b
#

(2) (1)
Lemma 19.4.24 To every 𝔞 ∈ GAform (𝔓) there is a 𝔟 ∈ GAform (𝔓) such that
𝔞 = 𝔟#2 = 𝔟#𝔟.

Proof Let 𝔞 (𝑧, 𝜆) = 𝜆2 + ∞

Í 2− 𝑗 𝑃 ∈ GA (2) (𝔓). To determine 𝔟 we begin
𝑗=1 𝜆 𝑗 form
by reasoning formally, meaning that we seek a sequence of polynomials 𝑄 𝑗 ∈ 𝔓,
𝑄 𝑗 [0] = 0 ( 𝑗 = 1, 2, ...), satisfying

∑︁ ∞ ∑︁ ∞ ∑︁ ∞
­𝜆 + 𝜆1− 𝑗 𝑄 𝑗 ® # ­𝜆 + 𝜆1− 𝑗 𝑄 𝑗 ® = 𝜆2 + 𝜆2− 𝑗 𝑃 𝑗
© ª © ª

« 𝑗=1 ¬ « 𝑗=1 ¬ 𝑗=1

where the left-hand side is equal to

∑︁∞ ∑︁∞
−𝑗
­1 + 𝜆 𝑄 𝑗 ® 𝜆# ­1 + 𝜆− 𝑗 𝑄 𝑗 ® 𝜆
© ª © ª

« 𝑗=1 ¬ « 𝑗=1 ¬
∑︁∞ ∞
∑︁ ∞
∑︁
= ­1 + 𝜆− 𝑗 𝑄 𝑗 ® # ­1 + 𝜆− 𝑗 𝑄 𝑗 ® 𝜆2 + 𝜆1− 𝑗 𝔡𝑄 𝑗
© ª © ª

« 𝑗=1 ¬ « 𝑗=1 ¬ 𝑗=1

∑︁∞ ∑︁∞ ∞
∑︁
= ­1 + 𝜆− 𝑗 𝑄 𝑗 ® ­1 + 𝜆− 𝑗 𝑄 𝑗 ® 𝜆2 + 𝜆1− 𝑗 𝔡𝑄 𝑗
© ª© ª

« 𝑗=1 ¬« 𝑗=1 ¬ 𝑗=1

∞ ∞ ∞
∑︁ 1 ©∑︁ ℓ − 𝑗 ª ©∑︁ − 𝑗 ℓ ª
+ ­ 𝜕𝜆 𝜆 𝑄 𝑗 ® ­ 𝜆 𝔡 𝑄 𝑗 ®
ℓ=1
ℓ! 𝑗=1
« ¬ « 𝑗=1 ¬
[cf. (19.4.17)]. It follows that we must have
∞
∑︁ ∞
∑︁ ∞
∑︁
2 𝜆− 𝑗 𝑄 𝑗 + 𝜆− 𝑗−𝑘 𝑄 𝑗 𝑄 𝑘 + 𝜆− 𝑗 𝔡𝑄 𝑗−1
𝑗=1 𝑗,𝑘=1 𝑗=2
∞ ∞
∑︁ ( 𝑗 + ℓ − 1)! − 𝑗−ℓ−𝑘 ∑︁
+ (−1) ℓ 𝜆 𝑄 𝑗 𝔡ℓ 𝑄 𝑘 = 𝜆− 𝑗 𝑃 𝑗 .
𝑗,ℓ,𝑘=1
ℓ! ( 𝑗 − 1)! 𝑗=1

We can solve these equations recursively with respect to the 𝑄 𝑗 :

818 19 Classical Analytic Formalism
∑︁
2𝑄 𝑗 = 𝑃 𝑗 − 𝔡𝑄 𝑗−1 − 𝜆−𝛼−𝛽 𝑄 𝛼 𝑄 𝛽 (19.4.32)
𝛼 ≥1,𝛽 ≥1
𝛼+𝛽= 𝑗
∑︁ (𝛼 + 𝛾 − 1)! −𝛼−𝛾−𝛽−𝛾
− (−1) 𝛾 𝜆 𝑄 𝛼𝔡 𝛾 𝑄 𝛽 .
𝛼 ≥1,𝛽 ≥1,𝛾 ≥1
(𝛼 − 1)!𝛾!
𝛼+𝛽+𝛾= 𝑗

Assume that the coefficients 𝑃 𝑗 satisfy (FA𝔓). Induction on 𝑗 = 0, 1... and the
equations (19.4.32) make it easy to prove that the functions 𝑄 𝑗 also satisfy (FA𝔓).
The fact that 𝑃 𝑗 [0] = 0 and 𝔡𝔓◦ ⊂ 𝔓◦ implies that 𝑄 𝑗 [0] = 0 for all 𝑗 = 1, 2, .... □
Remark 19.4.25 An obvious modification of the argument in the preceding proof
would allow us to prove the following two statements:
(𝑚)
(1) given any integer 𝑚 ≥ 1, to every 𝔞 ∈ GAform (𝔓) there is a unique 𝔟 ∈
(1) #𝑚
GAform (𝔓) such that 𝔞 = 𝔟 ;
(0) (0)
(2) to every 𝔞 ∈ GAform (𝔓) there is a unique 𝔟 ∈ GAform (𝔓) such that 𝔞 = 𝔟#2 .
1
If 𝔞 = 𝔟#𝔟 we shall use the notation 𝔟 = 𝔞 # 2 ; then, whatever 𝑚 ∈ Z we have
𝔟#𝑚 = 𝔞 #𝑚/2 . The series 𝔞 #𝑚/2 , 𝑚 ∈ Z, form an Abelian subgroup of GAform (𝔓);
every one of its elements belongs to the centralizer of 𝔞.
Theorem 19.4.26 The centralizer of b 𝐿 in the group GAform (𝔓) is the Abelian
1
subgroup of GAform (𝔓) consisting of the powers b 𝐿 # 2 𝑚 , 𝑚 ∈ Z.

Proof Given 𝔞 (𝑧, 𝜆) = 𝜆 𝑚 + ∞

Í 𝑚− 𝑗 𝑃 ∈ A (0) (𝔓) we have
𝑗=1 𝜆 𝑗 form

𝐿#𝔞 = 𝜆2 − 𝑢 0 𝔞 + 2𝜆𝔡𝔞 + 𝔡2𝔞,
b (19.4.33)
∞
∑︁ 1 𝑘
𝐿 = 𝜆2 − 𝑢 0 𝔞 −
𝔞#b 𝜕 𝔞 𝑢𝑘 .
𝑘=1
𝑘! 𝜆
h i
If b𝐿, 𝔞 = 0 we derive
#

∞
∑︁ 1 𝑘
𝔡2𝔞 + 2𝜆𝔡𝔞 + 𝜕 𝔞 𝑢 𝑘 = 0. (19.4.34)
𝑘=1
𝑘! 𝜆

Í∞ −𝑗𝑃 (0)
Assuming that 𝔞 = 1+ 𝑗=1 𝜆 𝑗 ∈ GAform (𝔓) and taking into account that 𝑃 𝑗 (0)
if 𝑗 ≥ 1, we get
∞ ∞ ∞ ∑︁∞
∑︁ ∑︁ ∑︁ 1 𝑘 −𝑗
2 𝜆1− 𝑗 𝔡𝑃 𝑗 + 𝜆− 𝑗 𝔡2 𝑃 𝑗 = − 𝜕 𝜆 𝑃 𝑗 𝑢𝑘 .
𝑗=1 𝑗=1 𝑗=1 𝑘=1
𝑘! 𝜆

We equate to zero the coefficient of 𝜆− 𝑗 for each 𝑗 ∈ Z+ ; we get directly 𝑃1 = 0 and,

as a consequence, 𝑃 𝑗 [𝑢] ≡ 0 for every 𝑗 = 2, 3.... In other words, 𝔞 = 1.
19.4 Symbolic Calculus and the KdV Hierarchy 819

If 𝔞 = 𝜆 𝑚 + ∞
Í 𝑚− 𝑗 𝑃 ∈ GA (𝑚) (𝔓) (0 ≠ 𝑚 ∈ Z) belongs to the centralizer
𝑗=1 𝜆 𝑗 form
(0) (0)
of b
𝐿 in GAform (𝔓) so does b 𝐿 #(−𝑚/2) #𝔞 ∈ GAform 𝐿 #𝑚/2 .
(𝔓), implying 𝔞 = b □
We deduce from (19.4.33)

Res b 𝐿#𝔞 = 𝑃𝑚+3 + 2𝔡𝑃𝑚+2 + 𝔡2 − 𝑢 0 𝑃𝑚+1 . (19.4.35)

We introduce the version acting on 𝔓 of what is called the Lenard operator:

𝔏 = 𝔡3 − 4𝑢 0𝔡−2𝑢 1 . (19.4.36)
h i
Proposition 19.4.27 Let 𝔞 = ∞
Í 𝑚− 𝑗 𝑃 ∈ A (𝑚) (𝔓) be such that b 𝔞 = 0;
𝑗=0 𝜆 𝑗 form
𝐿,
#
then
4𝔡 Res b 𝐿#𝔞 = 𝔏 Res 𝔞. (19.4.37)
h i
Proof If b 𝐿, 𝔞 = 0 we derive from (19.4.34):
#

𝔡2 𝑃𝑚+1 + 2𝔡𝑃𝑚+2 = 0,
𝔡2 𝑃𝑚+2 + 2𝔡𝑃𝑚+3 − 𝑢 1 𝑃𝑚+1 = 0,

implying
1 1
𝔡𝑃𝑚+3 = 𝑢 1 𝑃𝑚+1 + 𝔡3 𝑃𝑚+1 .
2 2
We derive from (19.4.35):

𝔡 Res b 𝐿#𝔞 = 𝔡𝑃𝑚+3 + 2𝔡2 𝑃𝑚+2 + 𝔡3 𝑃𝑚+1 − 𝑢 0𝔡𝑃𝑚+1 − 𝑢 1 𝑃𝑚+1

1 1 3
= 𝑢 1 𝑃𝑚+1 + 𝔡 𝑃𝑚+1 − 𝑢 0𝔡𝑃𝑚+1 − 𝑢 1 𝑃𝑚+1
2 2
1 1
= 𝔡3 𝑃𝑚+1 − 𝑢 0𝔡2 𝑃𝑚+1 − 𝑢 1 𝑃𝑚+1 ,
4 2
whence (19.4.37). □
# ( 21 𝑚+1)
Of course, Res 𝜆2 − 𝑢 0 = 0 if 𝑚 ∈ 2Z+ .

19.4.8 The KdV hierarchy

We define, for 𝑚 = 0, 1...,

# ( 𝑚+ 21 )
𝑆 𝑚 = −22𝑚+1 Res 𝜆2 − 𝑢 0 ; 𝑅𝑚 = 𝔡𝑆 𝑚 . (19.4.38)

Thus 𝑆0 = 𝑢 0 , 𝑅0 = 𝑢 1 , and if 𝑚 ∈ Z+ , (19.4.37) implies

820 19 Classical Analytic Formalism

𝑆 𝑚+1 = 𝔡−1𝔏𝑆 𝑚 , 𝑅𝑚+1 = 𝔏𝑆 𝑚 = 𝔏𝔡−1 𝑅𝑚 . (19.4.39)

We list the expressions of 𝑆 𝑚 and 𝑅𝑚 for 𝑚 = 1, 2, 3:

𝑆1 = 𝑢 2 − 3𝑢 20 ; (19.4.40)
𝑆2 = 𝑢 4 − 10𝑢 0 𝑢 2 − 5𝑢 21 + 10𝑢 30 ;
𝑆3 = 𝑢 6 − 14𝑢 0 𝑢 4 − 28𝑢 1 𝑢 3 − 35𝑢 40 − 21𝑢 22 + 70𝑢 0 𝑢 21 + 70𝑢 20 𝑢 2

𝑅1 = 𝑢 3 − 6𝑢 0 𝑢 1 ; (19.4.41)
𝑅2 = 𝑢 5 − 10𝑢 0 𝑢 3 − 20𝑢 1 𝑢 2 + 30𝑢 20 𝑢 1 ;
𝑅3 = 𝑢 7 − 14𝑢 0 𝑢 5 − 42𝑢 1 𝑢 4 − 70𝑢 2 𝑢 3
+ 70𝑢 20 𝑢 3 + 280𝑢 0 𝑢 1 𝑢 2 + 70𝑢 31 − 140𝑢 30 𝑢 1 .

The KdV equation reads

𝑢 𝑡 = 𝜕𝑥3 𝑢 − 6𝑢𝜕𝑥 𝑢. (19.4.42)
Other texts use different coefficients of 𝑢𝜕𝑥 𝑢; there is latitude to do so, by rescaling
𝑡, 𝑥, 𝑢. We shall refer to the evolution equation

𝑢 𝑡 = 𝑅𝑚 [𝑢]

as the 𝑚th KdV equation (𝑚 = 1, 2, ...). These equations form the KdV hierarchy.
We are now going to prove that the polynomials 𝑆 𝑚 are hamiltonians (see Sub-
section 19.4.3). We need the following

Lemma 19.4.28 If 𝐹 ∈ 𝔓 and if 𝑃𝐹 ∈ 𝔡𝔓 for every 𝑃 ∈ 𝔓 then necessarily 𝐹 = 0.

Proof Putting 𝑃 = 1 the hypothesis demands 𝐹 ∈ 𝔡𝔓. Then 𝐹 = 𝐴0 (𝑢 0 , ..., 𝑢 𝜈 ) +

𝐴1 (𝑢 0 , ..., 𝑢 𝜈 ) 𝑢 𝜈+1 for some 𝜈 ∈ Z+ and 𝐴1 ≠ 0. But 𝑢 𝜈+1 𝐹 cannot belong to 𝔡𝔓.□

Theorem 19.4.29 For each 𝑚 ∈ Z+ we have

1 𝛿𝑆 𝑚+1
𝑆𝑚 = − . (19.4.43)
2 (2𝑚 + 3) 𝛿𝑢
Proof If 𝑃 ∈ 𝔓 we have
3 3
𝜗𝑃 Res Λ# ( 𝑚+ 2 ) = Res 𝜗𝑃 Λ# ( 𝑚+ 2 )
∑︁ 1 1
1

= Res Λ# 2 𝑝 # 𝜗𝑃 Λ# 2 #Λ# 2 𝑞 .
𝑝+𝑞=2(𝑚+1)
19.4 Symbolic Calculus and the KdV Hierarchy 821

Let stand for “congruent mod 𝔡𝔓”. Thanks to Propositions 19.4.12 and 19.4.13
we have
3
1

𝜗𝑃 Res Λ# ( 𝑚+ 2 ) (2𝑚 + 3) Res 𝜗𝑃 Λ# 2 #Λ#(𝑚+1)
1 1
1

(2𝑚 + 3) Res Λ# 2 # 𝜗𝑃 Λ# 2 #Λ# ( 𝑚+ 2 )

3 1

𝑚+ Res 𝜗𝑃 Λ#Λ# ( 𝑚+ 2 ) ,
2

whence −1
3 3 1
𝑚+ 𝜗𝑃 Res Λ# ( 𝑚+ 2 ) −𝑃 Res Λ# ( 𝑚+ 2 ) .
2
But
3 𝛿 3
𝜗𝑃 Res Λ# ( 𝑚+ 2 ) 𝑃 Res Λ# ( 𝑚+ 2 )
𝛿𝑢
by (19.4.13), whence
−1
3 𝛿 3 1
𝑚+ Res Λ# ( 𝑚+ 2 ) = − Res Λ# ( 𝑚+ 2 )
2 𝛿𝑢

by Lemma 19.4.28. The claim then follows from (19.4.38). □

Corollary 19.4.30 For every 𝑚 ∈ Z+ , 𝑆 𝑚 is a hamiltonian.

Every KdV polynomial 𝑅𝑚 = 𝔡𝑆 𝑚 is Hamiltonian.

Corollary 19.4.31 For each 𝑚 ∈ Z+ we have

1 𝜕𝑆 𝑚+1
𝑆𝑚 = − . (19.4.44)
2 (2𝑚 + 3) 𝜕𝑢 0
Proof Combine Proposition 19.4.6 with Theorem 19.4.29. □
The coefficient in (19.4.43) has been chosen so that, for each 𝑚 ∈ Z+ , 𝑆 𝑚 = 𝑢 2𝑚 +
terms of order < 2𝑚; this is a direct consequence of (19.4.39); the KdV differential
operator 𝑅𝑚 is of the form 𝜕𝑥2𝑚+1 + lower order terms. We define 𝑄 𝑚+2 as the
1
polynomial congruent to (−1) 𝑚−1 2(2𝑚+3) 𝑆 𝑚+1 mod 𝔡𝔓 whose term of highest order
1 2
is 2 𝑢 𝑚 (𝑚 ∈ Z+ ). From (19.4.43) we derive directly

𝛿𝑄 𝑚+2
𝑆 𝑚 = (−1) 𝑚 , 𝑚 ∈ Z+ . (19.4.45)
𝛿𝑢
822 19 Classical Analytic Formalism

19.4.9 First integrals of the KdV equations

# ( 𝑚+ 21 )
In the following statements b 𝐿+ (𝑚 ∈ Z+ ) stands for the polynomial part of
1 1 1 # ( 𝑚+ 12 )
𝐿 # ( 𝑚+ 2 ) , meaning b
b 𝐿 # ( 𝑚+ 2 ) ∈ 𝔓 [𝜆] and b 𝐿 # ( 𝑚+ 2 ) − b
𝐿+ ∈ 𝜆−1 A (0) (𝔓). form

Lemma 19.4.32 We have

# ( 𝑚+ 21 )
𝑅𝑚 = 22𝑚 b𝐿+ 𝐿 .
,b (19.4.46)
#
Í∞ −𝑗𝑃 (0)
Proof Indeed, given arbitrary 𝔞 = 𝑗=1 𝜆 𝑗 ∈ 𝜆−1 Aform (𝔓), (19.4.33) implies
∞
h i ∑︁ 1 𝜕 𝑘𝔞
𝐿, 𝔞
b = 2𝜆𝔡𝔞 + 𝔡2𝔞 + 𝑢𝑘
#
𝑘=1
𝑘! 𝜕𝜆 𝑘
h i
and therefore the polynomial part of b𝐿, 𝔞 is equal to 2𝔡𝑃1 . We have
#

1 1
# ( 𝑚+ 2 ) # ( 𝑚+ 2 )
𝐿+
b ,b
𝐿 =− b
𝐿−b
𝐿+ ,b
𝐿
# #
1
𝐿 # ( 𝑚+ 2 ) = 2−2𝑚𝔡𝑆 𝑚 ,
= −2𝔡 Res b

the last equation a consequence of (19.4.38). □

Remark 19.4.33 The formula (19.4.46) and the fact that 𝑅𝑚 = 𝔡𝑆 𝑚 is Hamiltonian
# ( 𝑚+ 21 )
is often expressed by the statement that b
𝐿+ and b
𝐿 form a Lax pair, after the
work [Lax, 1968] (see [Dickey, 2003], p. 12).

Lemma 19.4.34 For all 𝑚 ∈ Z+ , 𝑛 ∈ Z,

𝑛 # ( 𝑚+ 12 ) 𝑛
𝐿 # 2 = −22𝑚 b
𝜗𝑅𝑚 b 𝐿+ 𝐿# 2 .
,b (19.4.47)
#

# ( 𝑚+ 12 )
Proof Since both 𝜗𝑅𝑚 and Ad b 𝐿+ are derivations in the algebra Aform (𝔓)
(with respect to the composition law #, Proposition 19.4.8) it suffices to prove the
claim for 𝑛 = 1. The Leibniz rule implies

# ( 𝑚+ 21 ) # 12 # 12 # 12 # ( 𝑚+ 12 ) # 12 # ( 𝑚+ 12 )
𝐿+
b ,𝐿
b #𝐿 + 𝐿 # 𝐿+
b b b ,𝐿
b = 𝐿+
b , 𝐿 = 2−2𝑚 𝑅𝑚 ,
# #
1 1 1 1
#
# # # # 2
𝜗𝑅𝑚 b 𝐿 2 #b𝐿 2 +b 𝐿 2 #𝜗𝑅𝑚 b 𝐿 2 = 𝜗𝑅𝑚 𝜆 − 𝑢 0 = −𝑅𝑚 ,

whence
19.4 Symbolic Calculus and the KdV Hierarchy 823

# ( 𝑚+ 21 ) 1 1 1
22𝑚 b𝐿+ 𝐿 # 2 + 𝜗𝑅𝑚 b
,b 𝐿 # 2 #b
𝐿# 2 (19.4.48)
#

1
1
#2 2𝑚 b # ( 𝑚+ 2 ) b# 2
1
# 12
+𝐿 # 2
b 𝐿+ ,𝐿 + 𝜗𝑅𝑚 𝐿b = 0.
#

If 𝔞 ∈ Aform (𝔓) satisfies

1 1
𝐿# 2 + b
𝔞#b 𝐿 # 2 #𝔞 = 0 (19.4.49)
then 𝔞 = 0. Indeed, if 𝔞 = 𝜆 𝑁 𝑃0 +𝑂 𝜆 𝑁 −1 with 𝑁 ∈ Z and 𝑃0 ≠ 0 then the left-hand

side in (19.4.49) has the form 2𝑃0 𝜆 𝑁 +1 + 𝑂 𝜆 𝑁 , a contradiction. Applying this to
(19.4.48) proves (19.4.47). □

Theorem 19.4.35 For all pairs of integers 𝑚, 𝑛 ∈ Z+ , 𝜗𝑅𝑚 𝑆 𝑛 ∈ 𝔡𝔓.

1
𝐿 # ( 𝑛+ 2 ) and
Proof Indeed, 𝑆 𝑚 = −22𝑛+1 Res b
1 1
𝐿 # ( 𝑛+ 2 ) = Res 𝜗𝑅𝑚 b
𝜗𝑅𝑚 Res b 𝐿 # ( 𝑛+ 2 )

# ( 𝑚+ 21 ) # ( 𝑛+ 12 )
= Res 𝐿 +
b ,𝐿b
#

belongs to 𝔡𝔓 by Proposition 19.4.13. □

In other words, every polynomial 𝑆 𝑛 and therefore also every polynomial congru-
ent to 𝑆 𝑛 mod 𝔡𝔓 is a conserved polynomial of 𝑢 𝑡 = 𝑅𝑚 [𝑢].

Theorem 19.4.36 For all pairs of integers 𝑚, 𝑛 ∈ Z+ , 𝑄 𝑛+2 is a conserved polynomial

of the 𝑚 th KdV equation 𝑢 𝑡 = 𝑅𝑚 [𝑢].

Proof It follows from (19.4.44) and Theorem 19.4.35 that 𝛿𝑄 𝑛+2 /𝛿𝑢 is a conserved
polynomial of 𝑅𝑚 ; then (19.4.13) implies 𝜗𝑅𝑚 𝑄 𝑛+2 ∈ 𝔡𝔓. □
To summarize, all the KdV polynomials share the same conserved polynomials
𝑄 𝑛+2 and order 𝑄 𝑚+2 = 𝑚. In a kind of converse, the latter determine the KdV
hierarchy through the map 𝑄 𝑚+2 ↦→ 𝑅𝑚 = 𝔡 𝛿𝑄𝛿𝑢𝑚+2 . Here are the first polynomials
𝑄 𝑚+1 (𝑚 ≤ 5):
1 2
𝑄2 = 𝑢 ,
2 0
1 2
𝑄3 = 𝑢 + 𝑢 30 ,
2 1
1 2 5
𝑄4 = 𝑢 2 + 5𝑢 0 𝑢 21 + 𝑢 40 ,
2 2
1 2
𝑄5 = 𝑢 + 7𝑢 0 𝑢 2 + 35𝑢 20 𝑢 21 + 7𝑢 50 ,
2
2 3
1 2 35
𝑄6 = 𝑢 + 9𝑢 0 𝑢 23 − 10𝑢 32 + 63𝑢 20 𝑢 22 − 𝑢 41 + 210𝑢 30 𝑢 21 + 21𝑢 60 .
2 4 2
It is traditional to add 𝑄 1 = 𝑢 0 to the list.
824 19 Classical Analytic Formalism

Remark 19.4.37 The construction of the KdV Hierarchy can be duplicated starting
𝐿 = 𝜆2 − 𝑢 0 :
from more general polynomials than b

𝐿 𝑚 = 𝜆 𝑚 + 𝑢 𝑚−2 𝜆 𝑚−2 + · · · + 𝑢 0 , 𝑚 ≥ 3.
b

One can then form the root b𝐿 1/𝑚 ∈A e(1) (𝔓) (see Remark 19.4.25) and associate
𝑚 form
to b
𝐿 𝑚 a hierarchy of equations generalizing the KdV hierarchy, called a Gelfand–
Dickey or GD hierarchy. The evolution equations in these hierarchies have soliton
solutions (see [Dickey, 2003], Ch. 1).
Chapter 20
Germ Fourier Integral Operators in Complex
Space

In this chapter and the two following ones our aim is to describe the action (mostly
microlocal) on a distribution (or a hyperfunction) of an operator originating from an
integral ∫
𝑨𝔠 𝑓 (𝜃, 𝜆) = e𝑖𝜆𝜑 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃, 𝜆) 𝑓 (𝑧) d𝑧 (20.1)
𝔠
with a complex phase-function 𝜑 ∈ O (Ω × Θ), a complex-analytic symbol 𝑎 (𝑧, 𝜃, 𝜆)
in Ω×Θ (Ω, Θ: domains in Euclidean complex spaces of possibly different dimensions
𝑛, 𝑁; see Definition 20.1.1 below) and 1 ≤ 𝜆 ↗ +∞; or, alternatively, from an
integral of the following kind:
∫
𝑨ℭ 𝑓 (𝑧, 𝜆) = e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) 𝑓 (𝑤) d𝑤d𝜃 (20.2)
ℭ

with 𝜑 ∈ O (Ω1 × Ω2 × Θ) and 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) a complex-analytic symbol in Ω1 ×Ω2 ×

Θ (Ω1 , Ω2 : domains in C𝑛1 , C𝑛2 ). The contours of integration 𝔠, ℭ will be totally real
submanifolds of Ω and Ω2 × Θ respectively, of real dimension 𝑛 and 𝑛2 + 𝑁. For our
purposes, we shall be taking 𝔠 (resp., ℭ) to be a relatively compact open subset of a
totally real submanifold of C𝑛 (resp., C𝑛2 +𝑁 ), X ⊂ Ω (resp., X ⊂ Ω2 × Θ), of class
C ∞ or, more often than not, C 𝜔 , of maximum dimension 𝑛 (resp., 𝑛2 + 𝑁). Note,
indeed, that in the familiar cases such as the Fourier transform (say, of compactly
supported distributions) or the FBI transform of the same (in both X = R𝑛 ), the
resulting 𝑨𝔠 𝑓 (𝜃, 𝜆) is a holomorphic (and even an entire) function of 𝜃.
The contrasting significance of the integrations with respect to 𝑧 and with respect
to 𝜃 made us refer to the integral operators defined by (20.1) and the linear maps
between spaces of germs of (holomorphic or generalized) functions constructed from
these, as transforms. In contrast, the role of 𝑤 in (20.2) is akin to that of 𝑧 and we
most often assume 𝑛1 = 𝑛2 ; generally speaking, 𝑨ℭ 𝑓 will be an object in the same
class as 𝑓 (think of pseudodifferential or of Fourier integral operators). In view of
this we shall refer to the integral operators defined by (20.2) and to the derived sheaf
homomorphisms as operators.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 825
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_20
826 20 Germ Fourier Integral Operators in Complex Space

Note that inverting the transform (20.1) compels us to handle functions that
depend also on 𝜆; it therefore makes sense to let 𝑓 depend on 𝜆 from the start. This
leads to consider, in (20.1), holomorphic functions 𝑓 (𝑧, 𝜆) in Ω that satisfy growth
conditions of the kind

∀𝜀 > 0, ∃𝐶 𝜀 , max | 𝑓 (𝑧, 𝜆)| ≤ 𝐶 𝜀 exp (𝜆Φ (𝑧) + 𝜀𝜆) (20.3)

𝑧 ∈𝐾 ,𝜆>1

where Φ is real-valued, smooth and often, in the applications, plurisubharmonic

(see Section 11.2); 𝐾 is an arbitrary compact subset of Ω. The set of these functions
𝑓 (𝑧, 𝜆) were introduced in [Sjöstrand, 1982]; we denote them by O (Φ) (Ω) and
refer to Φ as an exponent-weight. Bounds on (20.1) will depend on Ψ (𝑧, 𝜃) =
Φ (𝑧) − Im 𝜑 (𝑧, 𝜃) and the critical points of Ψ in 𝑧-space, 𝑧 (𝜃). We refer to (𝜑, Φ)
as a Sjöstrand pair at (𝑧◦ , 𝜃 ◦ ) if 𝑧 ◦ ∈ 𝔠 is a critical point as well as a saddle
point of Ψ (𝑧, 𝜃 ◦ ). Under the assumption that 𝑛 = 𝑁 the critical value function
Φ∗ (𝜃) = Ψ (𝑧 (𝜃) , 𝜃) can be regarded as an exponent-weight in 𝜃-space; 𝜃 ◦ is
a saddle point of Ψ∗ (𝜃, 𝑧◦ ) = Φ∗ (𝜃) + Im 𝜑 (𝑧 ◦ , 𝜃); we refer to (−𝜑, Φ∗ ) as the
Sjöstrand pair at (𝜃 ◦ , 𝑧◦ ) dual of (𝜑, Φ). It is easily checked that the dual pair
of (−𝜑, Φ∗ ) at (𝑧 ◦ , 𝜃 ◦ ) is (𝜑, Φ). The integral (20.1) defines a continuous linear
∗
operator 𝑨𝔠 : O (Φ) (Ω) −→ O (Φ ) (Θ). Similar considerations in the last section of
this chapter apply to (20.2) and lead naturally to Sjöstrand triads. These results have
applications in the following chapters.
Often (but not always!) the symbol 𝑎 (𝑧, 𝜃, 𝜆) in (20.1) and the amplitude
𝑎 (𝑧, 𝑤, 𝜃, 𝜆) in (20.2) will be classical (Definition 19.1.14), in which case we shall
fully avail ourselves of the results in Ch. 19.
As for the germ aspects this will come about by contraction (and deformation,
whence the frequent use of Stokes’ Theorem) of the contours of integration around
central points 𝑧 ◦ ∈ 𝔠, 𝑤 ◦ ∈ Ω♭ or (𝑧 ◦ , 𝜃 ◦ ) ∈ ℭ; the use of “germ” as an adjective
will be a tiresome recurrence, there to remind us that we are really talking about
sheaf maps while focusing on single stalks. Sheaf theory is by now the unavoidable
approach to functions and more general objects in the complex-analytic category.
Aside from “germification” the essential novelty in this chapter, when compared
to Chapters 17 and 18, is that the phase-functions in both (20.1) and (20.2) are
complex. Of course, this was already the case for the FBI transform (introduced in
Section 3.4) but transforms of the FBI type were not covered in Ch. 18.
The material in this chapter and the two following ones has been extracted, almost
entirely, from [Sjöstrand, 1982] (whose contents go well beyond the relatively small
portion gleaned here.) The basic concept in the first section of this chapter originates
with [Boutet-Krée, 1967].
20.1 Analytic Symbols 827

20.1 Analytic Symbols

20.1.1 Analytic symbols. Definitions

As we have seen in previous chapters the description of a pseudodifferential calculus

begins with the selection of amplitudes and symbols (Definitions 16.1.1; 17.1.4,
17.2.3; 18.1.1; also Definition 19.1.7). We shall now introduce a relatively wide
class of complex-analytic symbols which, later, will allow various specializations
(including the notion of analytic amplitude). Thus, in this subsection, 𝑛 and 𝑁 will
be arbitrary positive integers; Ω (resp., Θ) will be open subsets of C𝑛 (resp., C 𝑁 )
where the variable is 𝑧 = (𝑧 1 , ..., 𝑧 𝑛 ) [resp., 𝜃 = (𝜃 1 , ..., 𝜃 𝑁 )]; “analytic” will always
mean “complex-analytic”; we shall use “real-analytic” or C 𝜔 in the real set-up.
Definition 20.1.1 By a complex-analytic (or simply analytic) symbol in Ω × Θ we
shall mean a measurable function of 𝜆 ∈ [0, +∞) valued in O (Ω × Θ), 𝑎 (𝑧, 𝜃, 𝜆),
satisfying the following condition:
(AS) To every compact subset 𝐾 of Ω × Θ and every 𝜀 > 0 there is a 𝐶𝐾 , 𝜀 > 0 such
that
∀ (𝑧, 𝜃) ∈ 𝐾, ∀𝜆 ≥ 1, |𝑎 (𝑧, 𝜃, 𝜆)| ≤ 𝐶𝐾 , 𝜀 e 𝜀𝜆 . (20.1.1)
We shall denote by 𝑆a (Ω × Θ) the ring (with respect to ordinary addition and
multiplication) of analytic symbols in Ω × Θ.
Analytic amplitudes are simply analytic symbols where 𝑧 is replaced by (𝑧, 𝑤) ∈
Ω1 × Ω2 , Ω 𝑗 an open subset of C𝑛 , and 𝑛 by 2𝑛.
In Definition 20.1.1 Θ = ∅ is not precluded, in which case we would be dealing
with elements of the ring 𝑆a (Ω).
Remark 20.1.2 Suppose 𝑁 = 𝑛 and let 𝑎 ∈ 𝑆a (Ω × Θ). If we put 𝜆 = |𝜁 |, 𝜃 = 𝜁/|𝜁 | ∈
Θ ∩ S2𝑛−1 , then 𝑎 (𝑧, 𝜃, 𝜆) = 𝑎 (𝑧, 𝜁) becomes a function of infra-exponential type
with respect to 𝜁 [cf. Definition 8.1.1, and (8.1.13)] not in the whole of Ω × C𝑛 but
just in the complex “wedge” Ω × Γ ⊂ C2𝑛 where Γ = {𝜁 ∈ C𝑛 ; 𝜁/|𝜁 | ∈ Θ}.
The ring 𝑆a (Ω × Θ) becomes a Fréchet–Montel algebra when equipped with the
topology defined by the seminorms

𝑎 ↦→ sup e−𝜀𝜆 𝑎 (𝑧, 𝜃, 𝜆) (20.1.2)

(𝑧, 𝜃) ∈𝐾 ,𝜆≥1

with 𝐾 a compact subset of Ω × Θ and 𝜀 > 0.

Proposition 20.1.3 Every hyperdifferential operator in Ω (Definition 8.1.14) defines
a linear endomorphism of 𝑆a (Ω × Θ).
Proof Consider a hyperdifferential operator in Ω,
∑︁
𝐽 (𝑧, D𝑧 ) = 𝑐 𝛼 (𝑧) D𝑧𝛼 .
𝛼∈Z+𝑛
828 20 Germ Fourier Integral Operators in Complex Space

Let the compact subset 𝐾 of Ω and 𝜀 > 0 be arbitrary; there is a 𝐵 𝐾 , 𝜀 > 0 such that
(8.1.15) holds:
𝜀|𝛼|
∀𝛼 ∈ Z+𝑛 , max |𝑐 𝛼 (𝑧) | ≤ 𝐵 𝐾 , 𝜀 .
𝑧 ∈𝐾 𝛼!
Let 𝑎 ∈ 𝑆a (Ω × Θ) and 𝐾 ′ be a compact subset of Θ; to every 𝜀 ′ > 0 there is a
𝐶𝐾×𝐾 ′ , 𝜀′ > 0 such that
| 𝛼 |+1 ′
∀ (𝑧, 𝜃) ∈ 𝐾 × 𝐾 ′, ∀𝜆 ≥ 1, D𝑧𝛼 𝑎 (𝑧, 𝜃, 𝜆) ≤ 𝐶𝐾×𝐾 ′ , 𝜀 ′ 𝛼!e
𝜀𝜆
.

We derive, for (𝑧, 𝜃) ∈ 𝐾 × 𝐾 ′,

∑︁
| 𝛼 |+1 | 𝛼 | ′| 𝛼 |
|𝐽 (𝑧, D𝑧 ) 𝑎 (𝑧, 𝜃, 𝜆)| ≤ 𝐵 𝐾 , 𝜀 e 𝜀𝜆 𝐶𝐾×𝐾 ′ , 𝜀′ 𝜀 𝜀 .
𝛼∈Z+𝑛

If we take 𝜀 ′ < 1 and 𝜀 < 𝐶𝐾×𝐾

−1
′ , 𝜀 ′ we get

1
|𝐽 (𝑧, D𝑧 ) 𝑎 (𝑧, 𝜃, 𝜆)| ≤ 𝐶𝐾×𝐾 ′ , 𝜀′ 𝐵 𝐾 , 𝜀 e 𝜀𝜆 . □
1 − 𝐶𝐾×𝐾 ′ , 𝜀′ 𝜀𝜀 ′

Replacing Ω by Ω × Θ in Definition 19.1.10 leads to the following

Definition 20.1.4 By an exponentially decaying symbol in Ω × Θ we shall mean

an analytic symbol in Ω × Θ satisfying the following condition:
(EDCS) To every compact subset 𝐾 of Ω × Θ there are positive constants 𝐶, 𝜅, such
that
∀𝜆 ≥ 1, max |𝑎 (𝑧, 𝜃, 𝜆)| ≤ 𝐶e−𝜅𝜆 .
𝐾

We shall denote by 𝑆a−𝜔 (Ω × Θ) the ring of exponential decaying symbols in

Ω × Θ.

Obviously, 𝑆a−𝜔 (Ω × Θ) is an ideal in the ring 𝑆a (Ω × Θ) and 𝑆a−𝜔 (Ω) =

O (−𝜔) (Ω).
The next definition introduces an important subclass of analytic symbols (cf.
Definition 17.1.4).

Definition 20.1.5 An analytic symbol 𝑎 (𝑧, 𝜃, 𝜆) is said to be of order 𝑚(∈ R) if the

following is valid:
(FAS) To every compact subset 𝐾 of Ω × Θ there is a 𝐶𝐾 > 0 such that, for every
𝜆 ≥ 1, max |𝑎 (𝑧, 𝜃, 𝜆)| ≤ 𝐶𝐾 𝜆 𝑚 .
(𝑧, 𝜃) ∈𝐾

We shall denote by 𝑆a𝑚 (Ω × Θ) the linear space of symbols of order 𝑚.

The elements of the ring (with respect to ordinary addition and multiplication)
Ø
𝑆a𝑚 (Ω × Θ) (20.1.3)
𝑚∈R
20.1 Analytic Symbols 829

are the analytic symbols of finite order. The elements of the ring 𝑆a0 (Ω × Θ) =
Ob (Ω × Θ) (cf. Definition 19.1.11) are the bounded symbols in Ω × Θ. In the sequel
Θ ⊂ R 𝑁 will be often replaced by a domain Ξ in C𝑛 .
Given two analytic symbols in Ω × Ξ, 𝑎 (𝑧, 𝜁, 𝜆), 𝑏 (𝑧, 𝜁, 𝜆), we can define the
composition law # [cf. (16.2.16), (19.1.12)]:
∑︁ 1
𝜆− |𝛾 | D 𝜁 𝑎 𝜕𝑧 𝑏
𝛾 𝛾
𝑎#𝑏 = (20.1.4)
𝛾 ∈Z 𝑛 𝛾!
+

√
where, as usual, D = − −1𝜕.

Proposition 20.1.6 If 𝑎 (𝑧, 𝜁, 𝜆), 𝑏 (𝑧, 𝜁, 𝜆) are two analytic symbols in Ω × Ξ then
to every compact subset 𝐾 of Ω × Ξ and every 𝜀 > 0 there is a 𝐶𝐾 , 𝜀 > 0 such that,
for all 𝛾 ∈ Z+𝑛 and 𝜆 ≥ 1,

1
𝛾

𝛾
max D 𝜁 𝑎 𝜕𝑧 𝑏 ≤ 𝐶𝐾 , 𝜀 𝛾!e 𝜀𝜆 . (20.1.5)
𝛾! 𝐾
Proof Follows directly from Definition 20.1.1 and the Cauchy inequalities. □
Proposition 20.1.6 means that we can regard (20.1.4) as a formal analytic series
with coefficients in 𝑆a (Ω × Ξ), i.e., an element of Í Aform 𝑆a (Ω × Θ) (cf.ÍDefinition
19.1.18). It is an easy exercise to prove that if 𝑎 = ∞ 𝑗=0 𝑎 𝑗 𝜆
− 𝑗 and 𝑏 = ∞
𝑗=0 𝑏 𝑗 𝜆
−𝑗

are themselves formal analytic series belonging to Aform 𝑆a (Ω × Θ) then the same
is true of
∞
∑︁ ∑︁ 1 − |𝛾 |− 𝑗−𝑘 𝛾 𝛾
𝑎#𝑏 = 𝜆 D 𝜁 𝑎 𝑗 𝜕𝑧 𝑏 𝑘 . (20.1.6)
𝛾 ∈Z𝑛 𝑗,𝑘=0
𝛾!
+

Thus Aform 𝑆a (Ω × Θ) becomes a (topological, associative, noncommutative) alge-

bra with respect to the composition law #.

Remark 20.1.7 The Í algebra Aform 𝑆a (Ω × Θ) should not be confused with

Aform (Ω × Θ): if ∞ 𝑗=0 𝑎 𝑗 𝜆 −𝑗 ∈ A
form 𝑆 a (Ω × Θ) the coefficients 𝑎 𝑗 are functions
of 𝜆 satisfying Condition (AS) in Definition 20.1.1 with constants 𝐶 𝜀 depending on
𝑗 as required by Definition 19.1.18 with 𝑬 the Fréchet–Montel algebra 𝑆a (Ω × Θ).

20.1.2 Analytic symbols of finite order and pseudoanalytic symbols

It is natural to ask what is the relation between pseudoanalytic (Definition 17.1.4)

and analytic symbols. In this subsection we answer this question for symbols of finite
order, when 𝑁 = 𝑛; Ω and Θ are open subsets of C𝑛 intersecting R𝑛 . The next
statement can be regarded as a variant of Lemma 17.2.27.
830 20 Germ Fourier Integral Operators in Complex Space

Theorem 20.1.8 Suppose 𝑎 (𝑧, 𝜃,n 𝜆) ∈ 𝑆a𝑚 (Ω × Θ), 𝑚 ∈ oR. Let 𝜃 ◦ ∈ Θ ∩ R𝑛 \ {0}
◦
and let 𝑟 > 0 be such that Γ𝑟 = 𝜉 ∈ R𝑛 ; | 𝜉𝜉 | − |𝜃𝜃 | < 𝑟 is contained in the cone
generated by Θ ∩ R𝑛 . Suppose that the function R+ ∋ 𝜆 ↦→ 𝑎 (𝑧, 𝜃, 𝜆) is C ∞ and
satisfies the following condition:

(FAS’) To every compact subset 𝐾 of Ω there are positive numbers 𝐶𝐾 , 𝜌 𝐾 such

that, for all 𝑝 ∈ Z+ , 𝜆 ≥ 𝜌 𝐾 max (1, 𝑝) and 𝜃 ∈ Θ,

max 𝜕𝜆𝑝 𝑎 (𝑧, 𝜃, 𝜆) ≤ 𝐶𝐾𝑝+1 𝑝!𝜆 𝑚− 𝑝 .

𝑧 ∈𝐾

Under this hypothesis and possibly after decreasing 𝑟 > 0, there is a pseudo-
analytic symbol 𝑎♭ (𝑥, 𝜉) of order 𝑚 in (Ω ∩ R𝑛 ) × Γ𝑟 equal to 𝑎 𝑥, | 𝜉𝜉 | , |𝜉 | if
|𝜉 | > 1.

Proof We define 𝑎♭ (𝑧, 𝜉) = 𝜒 (|𝜉 |) 𝑎 𝑧, | 𝜉𝜉 | , |𝜉 | with 𝜒 (𝑡) ∈ C ∞ (R), 𝜒 (𝑡) = 0
if 𝑡 < 1/2 and 𝜒 (𝑡) = 1 if 𝑡 > 1. When 𝑛 = 1 the result is immediate since
𝑎 𝑧, | 𝜉𝜉 | , |𝜉 | = 𝑎 (𝑧, ±1, |𝜉 |). Suppose 𝑛 ≥ 2. We introduce the following vector
fields in Θ × (0, +∞),
𝜕
𝐿 𝑗 = 𝜆−1𝑇 𝑗 + 𝜃 𝑗 ,
𝜕𝜆
𝑛
𝜕 ∑︁ 𝜕
𝑇𝑗 = − 𝜃𝑗 𝜃𝑘
𝜕𝜃 𝑗 𝑘=1
𝜕𝜃 𝑘

𝑗 = 1, ..., 𝑛. The 𝐿 𝑗 commute pairwise; 𝐿 𝑗 = 𝜕𝜕𝜉 𝑗 when 𝜃 = | 𝜉𝜉 | and 𝜆 = |𝜉 |, 𝜉 ∈ R𝑛 .

Let 𝐾 and 𝑧 ∈ 𝐾 be as in (FAS’); we take 𝑟 < 1 and define

𝑉𝑟 = {𝜉 ∈ Γ𝑟 ; 1 − 𝑟 < |𝜉 | < 1 + 𝑟 } .

We will apply the Ovsyannikov technique (cf. Ch. 5). Suppose we have proved, for
some 𝑞 ∈ Z+ , 𝐶 > 0 and all 𝛾 ∈ Z+𝑛 , |𝛾| ≤ 𝑞, all 𝑝 ∈ Z+ , 𝜆 ≥ 𝜌 𝐾 max (1, 𝑝 + 𝑞) and
𝑠, 0 < 𝑠 < 𝑟,
( 𝑝 + 𝑞)! 𝑚− 𝑝−𝑞
∀𝜃 ∈ 𝑉𝑠 ,sup 𝜕𝜆𝑝 𝐿 𝛾 𝑎 (𝑧, 𝜃, 𝜆) ≤ 𝐶 𝑞+ 𝑝+1 𝜆 (20.1.7)
(𝑟 − 𝑠) 𝑞
𝛾 𝛾
(𝐿 𝛾 = 𝐿 1 1 · · · 𝐿 𝑛𝑛 ). We reason by induction on 𝑞; by (FAS’) this is true when 𝑞 = 0.
We have, for |𝛾| = 𝑞,

𝜕𝜆𝑝 𝐿 𝑗 𝐿 𝛾 𝑎 = 𝜕𝜆𝑝 𝜆−1𝑇 𝑗 + 𝜃 𝑗 𝜕𝜆 𝐿 𝛾 𝑎

= 𝑇 𝑗 𝜕𝜆𝑝 𝜆−1 𝐿 𝛾 𝑎 + 𝜃 𝑗 𝜕𝜆𝑝+1 𝐿 𝛾 𝑎

= 𝜆−1𝑇 𝑗 𝜕𝜆𝑝 𝐿 𝛾 𝑎 − 𝑝𝜆−2𝑇 𝑗 𝜕𝜆𝑝−1 (𝐿 𝛾 𝑎) + 𝜃 𝑗 𝜕𝜆𝑝+1 𝐿 𝛾 𝑎.

20.1 Analytic Symbols 831

Since 𝜃 ∈ 𝑉𝑟 =⇒ |𝜃| < 2 we have, by (20.1.7) provided 𝜆 ≥ 𝜌 𝐾 ( 𝑝 + 𝑞 + 1),

( 𝑝 + 𝑞 + 1)! 𝑚− 𝑝−𝑞−1
𝜃 𝑗 𝜕𝜆𝑝+1 𝐿 𝛾 𝑎 (𝑧, 𝜃, 𝜆) ≤ 2𝐶 𝑞+ 𝑝+2 𝜆 . (20.1.8)
(𝑟 − 𝑠) 𝑞

If 0 < 𝑠 ′ < 𝑠 and ℎ ∈ O (𝑉) the Cauchy inequalities imply

5
sup 𝑇 𝑗 𝑓 (𝜃) ≤ sup | 𝑓 (𝜃)| .
𝜃 ∈𝑉𝑠′ 𝑠 − 𝑠 ′ 𝜃 ∈𝑉𝑠

Combining this with (20.1.7) yields, for 𝜆 ≥ 𝜌 𝐾 max (1, 𝑝 + 𝑞), 𝑧 ∈ 𝐾 and 𝜃 ∈ 𝑉𝑠′ ,
and
( 𝑝 + 𝑞)!
𝜆−1 𝑇 𝑗 𝜕𝜆𝑝 𝐿 𝛾 𝑎 (𝑧, 𝜃, 𝜆) ≤ 5𝐶 𝑞+ 𝑝+1 𝜆 𝑚− 𝑝−𝑞−1 ,
(𝑟 − 𝑠) 𝑞 (𝑠 − 𝑠 ′)
( 𝑝 + 𝑞)!
𝑝𝜆−2 𝑇 𝑗 𝜕𝜆𝑝−1 (𝐿 𝛾 𝑎) ≤ 𝐶 𝑞+ 𝑝 𝜆 𝑚− 𝑝−𝑞−1 .
(𝑟 − 𝑠) 𝑞 (𝑠 − 𝑠 ′)

We select 𝑠 − 𝑠 ′ = 𝑟−𝑠
𝑝+𝑞+1 ; these last inequalities become

( 𝑝 + 𝑞 + 1)!
𝜆−1 𝑇 𝑗 𝜕𝜆𝑝 𝐿 𝛾 𝑎 (𝑧, 𝜃, 𝜆) ≤ 5𝐶 𝑞+ 𝑝+1 𝜆 𝑚− 𝑝−𝑞−1 , (20.1.9)
(𝑟 − 𝑠) 𝑞+1
( 𝑝 + 𝑞 + 1)! 𝑚− 𝑝−𝑞−1
𝑝𝜆−2 𝑇 𝑗 𝜕𝜆𝑝−1 (𝐿 𝛾 𝑎) ≤ 𝐶 𝑞+ 𝑝 𝜆 .
(𝑟 − 𝑠) 𝑞+1
We derive from (20.1.8) and (20.1.9):
( 𝑝 + 𝑞 + 1)! 𝑚− 𝑝−𝑞−1
𝐿 𝑗 𝜕𝜆𝑝 𝐿 𝛾 𝑎 (𝑧, 𝜃, 𝜆) ≤ 5𝐶 −1 + 𝐶 −2 + 2𝑟 𝐶 𝑞+ 𝑝+2 𝜆 .
(𝑟 − 𝑠) 𝑞+1
(20.1.10)
By taking 𝐶 sufficiently large and 𝑟 sufficiently small so that 5𝐶 −1 + 𝐶 −2 + 2𝑟 ≤ 1 we
have proved that (20.1.7) is valid for 𝑞 + 1 replacing 𝑞 provided 𝜆 ≥ 𝜌 𝐾 ( 𝑝 + 𝑞 + 1).
We select 𝑠 = 21 𝑟; putting 𝑝 = 0 in (20.1.10) we derive, for all 𝛾 ∈ Z+𝑛 and 𝜉 ∈ R𝑛 ,
|𝜉 | > 𝜌 𝐾 max (1, |𝛾|),
|𝛾 |+1
max 𝜕 𝜉 𝑎♭ (𝑧, 𝜉) ≤ 2𝑟 −1 𝐶 |𝜉 | −𝑚−|𝛾 | .
𝛾
𝑧 ∈𝐾

This proves the claim. □

832 20 Germ Fourier Integral Operators in Complex Space

20.1.3 Analytic symbols and nonclassical differential operators of

infinite order

Let Ω and Ξ be open subsets of C𝑛 and (𝑧 ◦ , 𝜁 ◦ ) ∈ Ω × Ξ; let 𝑝 (𝑧, 𝜁, 𝜆) be an analytic

symbol in Ω × Ξ (Definition 20.1.1). For fixed 𝜆 ≥ 1 the Taylor expansion
∑︁ 1
𝑝 (𝑧, 𝜁, 𝜆) = (𝜁 − 𝜁 ◦ ) 𝛼 𝜕𝜁𝛼 𝑝 𝑗 (𝑧, 𝜁 ◦ , 𝜆) (20.1.11)
𝛼∈Z𝑛
𝛼!
+

converges absolutely uniformly in 𝐾 × 𝑉, 𝐾 a compact subset of Ω and 𝑉 a suffi-

ciently small neighborhood of 𝜁 ◦ in Ξ (generally depending on 𝐾). We introduce the
differential operator of infinite order
∑︁ 1
𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) = 𝜆− | 𝛼 | 𝜕𝜁𝛼 𝑝 (𝑧, 𝜁 ◦ , 𝜆) D𝑧𝛼 . (20.1.12)
𝛼∈Z 𝑛 𝛼!
+

Property (AS), (20.1.1) and the Cauchy inequalities imply that to every compact
subset 𝐾 of Ω and to every 𝜀 > 0 there is a 𝐶𝐾 , 𝜀 > 0 such that

1
∀𝜆 ≥ 1, max 𝜕 𝛼 𝑝 (𝑧, 𝜁 ◦ , 𝜆) ≤ 𝐶𝐾| 𝛼, |+1 𝜀𝜆
𝜀 e .
𝛼! 𝑧 ∈𝐾 𝜁
This suggests the introduction of a wide class of differential operators of infinite
order with coefficients in 𝑆a (Ω).

Definition 20.1.9 Let Ω be an open subset of C𝑛 . By an analytic differential oper-

ator of infinite order in Ω we shall mean a formal series
∑︁
𝐿 (𝑧, D𝑧 , 𝜆) = 𝑐 𝛼 (𝑧, 𝜆) 𝜆−| 𝛼 | D𝑧𝛼 (20.1.13)
𝛼∈Z+𝑛

where the coefficients 𝑐 𝛼 (𝑧, 𝜆) are functions of (𝑧, 𝜆) ∈ Ω × [1, +∞) holomorphic
with respect to 𝑧 and satisfy the following condition:

(AS”) To every compact subset 𝐾 of Ω there are positive constants 𝑀𝐾 and, to every
𝜀 > 0, 𝐶𝐾 , 𝜀 such that

∀𝛼 ∈ Z+𝑛 , ∀𝜆 ≥ 1, max |𝑐 𝛼 (𝑧, 𝜆)| ≤ 𝑀𝐾 𝐶𝐾| 𝛼, |𝜀 e 𝜀𝜆 . (20.1.14)

𝑧 ∈𝐾

If we compare Definition 20.1.9 to Definition 19.2.1 we see that the modification

follows the rules of transition from formal analytic series to analytic symbols. A
difference is that an analytic differential operator of infinite order 𝐿 (𝑧, D𝑧 , 𝜆) in Ω
may not map 𝑆a (Ω) into itself.
20.1 Analytic Symbols 833

Example 20.1.10 Suppose 𝑛 = 1 and let 𝐿 (D𝑧 , 𝜆) = ∞ −𝑘 𝑘 1

Í
𝑘=0 𝜆 𝜕𝑧 . If 1−𝑧 is identified
with an analytic symbol in the open unit disk, we have, for |𝑧| < 1,
1 ∑︁ 𝑘!
𝐿 (D𝑧 , 𝜆) = 𝜆−𝑘 .
1 − 𝑧 𝑘 ∈Z (1 − 𝑧) 𝑘+1
+

Things become more interesting if we apply 𝐿 (𝑧, D𝑧 , 𝜆) to series 𝑎 (𝑧, 𝜆) =

Í∞ −𝑗 ∈ A
𝑗=0 𝑎 𝑗 (𝑧, 𝜆) 𝜆 form 𝑆 a (Ω) (cf. Definition 19.1.18). Such series satisfy the
appropriate (GFA) condition: to every compact subset 𝐾 of Ω and every 𝜀 > 0 there
are positive constants 𝑀𝐾′ , 𝐶𝐾′ , 𝜀 such that

′𝑗
∀ 𝑗 ∈ Z+ ,max 𝑎 𝑗 (𝑧, 𝜆) ≤ 𝑀𝐾′ 𝐶𝐾 . 𝜀 𝑗!e 𝜀𝜆 . (20.1.15)
𝑧 ∈𝐾

Proposition 20.1.11 An analytic differential operator of infinite order 𝐿 (𝑧, D𝑧 , 𝜆)

in Ω defines a linear endomorphism of Aform 𝑆a (Ω).

Proof The Cauchy inequalities and (20.1.15) entail, for some 𝐵 𝐾 > 0,
′𝑗
∀ 𝑗 ∈ Z+ , ∀𝛼 ∈ Z+𝑛 ,max 𝜕𝑧𝛼 𝑎 𝑗 (𝑧, 𝜆) ≤ 𝑀𝐾′ 𝐵 𝐾| 𝛼 | 𝐶𝐾 . 𝜀 𝑗!𝛼!e 𝜀𝜆 . (20.1.16)
𝑧 ∈𝐾

We have
∞
∑︁ ∑︁
𝐿 (𝑧, D𝑧 , 𝜆) 𝑎 (𝑧, 𝜆) = 𝜆−𝑚 𝑐 𝛼 (𝑧, 𝜆) D𝑧𝛼 𝑎 𝑚−| 𝛼 | (𝑧, 𝜆) .
𝑚=0 | 𝛼 | ≤𝑚

Combining (20.1.14) with (20.1.15 ) yields, for a sufficiently large 𝐶

e𝐾 , 𝜀 > 0 and all
𝑧 ∈ 𝐾, 𝜆 ≥ 1,

∑︁ ∑︁
𝑐 𝛼 (𝑧, 𝜆) D𝑧𝛼 𝑎 𝑚−| 𝛼 | (𝑧, 𝜆) ≤ 𝑀𝐾′ e 𝜀𝜆 𝐶𝐾| 𝛼, |+1 | 𝛼 | ′𝑚−| 𝛼 |
𝜀 𝐵 𝐾 𝐶𝐾 . 𝜀 𝑗! (𝑚 − |𝛼|)!
| 𝛼 | ≤𝑚 | 𝛼 | ≤𝑚

≤ 𝑀𝐾′ 𝐶𝐾𝑚, 𝜀 𝐶
e𝑚 𝑚!e 𝜀𝜆 .
𝐾,𝜀 □
834 20 Germ Fourier Integral Operators in Complex Space

20.2 Contours and Function Spaces

20.2.1 Contours of integration

We begin by specifying the nature of the contour of integration 𝔠 ⊂ Ω in integrals

such as (20.1). Here Ω and Θ will be open subsets of C𝑛 and C 𝑁 respectively.
Definition 20.2.1 By a contour of integration in Ω (resp., Θ) we shall mean an
open and connected subset of a totally real C ∞ submanifold X of Ω (resp., Θ) of real
dimension 𝑛 (resp., 𝑁) whose closure is contained in X and whose boundary in X is
Lipschitz continuous.

√ We recall the definition: an R-linear subspace 𝑬 of ∞

C𝑛 is totally real if 𝑬 ∩
−1𝑬 = {0}. The tangent space to the totally real C submanifold X at a point
𝑧 ∈ X is a totally real vector subspace 𝑇𝑧 X of 𝑇𝑧 C𝑛 (the latter identified with C𝑛
through its basis 𝜕𝑧𝜕1 , ..., 𝜕𝑧𝜕𝑛 ) and (in Definition 20.2.1) dimR 𝑇𝑧 X = 𝑛.
Much use will be made of the concepts in the following
Definition 20.2.2 Let 𝑓 ∈ C (Ω; R). We shall say that a contour of integration 𝔠 ⊂ Ω
(Definition 20.2.1) is well-shaped at 𝑧 ◦ ∈ 𝔠 for 𝑓 if 𝑓 (𝑧 ) − 𝑓 (𝑧◦ ) ≤ 0 for every
𝑧 ∈ 𝔠 and that 𝔠 is strongly well-shaped at 𝑧◦ for 𝑓 if there is a 𝛾◦ > 0 such that

∀𝑧 ∈ 𝔠, 𝑓 (𝑧 ) − 𝑓 (𝑧◦ ) ≤ −𝛾◦ |𝑧 − 𝑧 ◦ | 2 .

We shall say that 𝔠 is locally well-shaped (resp., locally strongly well-shaped) at

𝑧◦ for 𝑓 if there is a neighborhood 𝑈 of 𝑧◦ in Ω such that 𝔠 ∩ 𝑈 is a contour of
integration well-shaped (resp., strongly well-shaped) at 𝑧◦ for 𝑓 .
We can relate the notion of local strong well-shapedness for 𝑓 ∈ C 2 (Ω; R) at
𝑧◦∈ 𝔠 to the Hessian of 𝑓 at 𝑧◦ . The Hessian of 𝑓 , Hess 𝑓 , is the (2𝑛) × (2𝑛)
2𝑛 𝜕2 𝑓
matrix computed in the canonical coordinates 𝑠1 , ..., 𝑠2𝑛 in R , 𝜕𝑠 𝑗 𝜕𝑠𝑘 .
1≤ 𝑗,𝑘 ≤2𝑛
For 𝔠 to be locally strongly well-shaped at 𝑧 ◦ for 𝑓 it is necessary and sufficient that
d 𝑓 (𝑧 ◦ ) = 0 and the restriction of (Hess 𝑓 ) (𝑧◦ ) to 𝑇𝑧 ◦ 𝔠 be negative definite. We point
out a few related facts:
(1) The totally real linear subspaces of C𝑛 of a given dimension 𝑑 form an open
subset 𝔄𝑑 of the Grassmannian of all real linear subspaces of C𝑛 of dimension
𝑑; the latter is a compact real-analytic manifold and thus 𝔄𝑑 can be regarded as
a C 𝜔 manifold.
(2) An 𝑛-dimensional R-linear subspace 𝑬 of C𝑛 is totally real if and only if there
is a C-linear automorphism of C𝑛 that maps 𝑬 onto R𝑛 ⊂ C𝑛 . This defines
a C 𝜔 diffeomorphism of 𝔄𝑛 onto the quotient GL (𝑛, C) /GL (𝑛, R), itself a
connected C 𝜔 manifold; thus the C 𝜔 manifold 𝔄𝑛 is connected.
(3) Let 𝑄 be a real quadratic form on C𝑛 R2𝑛 . Denote by 𝔄𝑛(𝑄) the subset of 𝔄𝑛
consisting of the vector subspaces 𝑬 ∈ 𝔄𝑛 such that 𝑄| 𝑬 is negative definite; it
is obvious that 𝔄𝑛(𝑄) is an open subset of 𝔄𝑛 .
20.2 Contours and Function Spaces 835

(4) If 𝑄 1 and 𝑄 2 are two real quadratic forms on C𝑛 whose restrictions to 𝑬 are
negative definite then the same is true of 𝜆𝑄 1 for all 𝜆 > 0 and of 𝑡𝑄 1 +(1 − 𝑡) 𝑄 2 ,
0 ≤ 𝑡 ≤ 1: these quadratic forms make up a convex open cone Γ (𝑬).
Let us regard 𝑬 ∈ 𝔄𝑛 as a subspace of C𝑛 𝑇0 C𝑛 . If Σ is a C ∞ submanifold
of a neighborhood 𝑈 of the origin in R2𝑛 such that 𝑇0 Σ = 𝑬 then there is a smooth
homotopy that transforms Σ ∩ 𝑉 (𝑉: a suitably small neighborhood of 0) into 𝑬 ∩ 𝑉,
with 𝑬 now regarded as a subspace of R2𝑛 . This is evident when 𝑬 = R𝑛 : the
homotopy is Σ ∩ 𝑉 ∋ 𝑧 = 𝑥 + 𝑖𝑦 ↦→ 𝑥 + 𝑖𝑡𝑦, 0 ≤ 𝑡 ≤ 1. Transforming R𝑛 into 𝑬 by a
C-linear automorphism of C𝑛 proves the claim in the general case.
Putting all this together allows us to state
Proposition 20.2.3 Let 𝔠◦ and 𝔠1 be two contours of integration in Ω containing 𝑧 ◦
and strongly well-shaped at 𝑧◦ for the function 𝑓 ∈ C ∞ (Ω). There are contours
𝔠 (𝑡) ⊂⊂ Ω strongly well-shaped at 𝑧◦ for the function 𝑓 and depending smoothly
on 𝑡 ∈ [0, 1] and a neighborhood of 𝑧◦ , 𝑈 ⊂ Ω, such that 𝔠 (0) ∩ 𝑈 = 𝔠◦ ∩ 𝑈,
𝔠 (1) ∩ 𝑈 = 𝔠1 ∩ 𝑈.
In important situations the central point 𝑧 ◦ will be a saddle point of 𝑓 .
Definition 20.2.4 Let 𝑓 ∈ C ∞ (Ω; R). We shall say that ℘◦ ∈ Ω is a saddle point of
𝑓 if d 𝑓 (℘◦ ) = 0, det Hess 𝑓 (℘◦ ) ≠ 0 and sign Hess 𝑓 (℘◦ ) = 0.
The signature of the Hessian of 𝑓 at ℘◦ , sign Hess 𝑓 , is 𝜈+ − 𝜈− , 𝜈+ = number
of eigenvalues 𝜒 > 0 of Hess 𝑓 , 𝜈− = number of eigenvalues 𝜒 < 0. In Definition
20.2.4 the requirement is that 𝜈+ = 𝜈− = 𝑛 at ℘◦ , and therefore in a full neighborhood
of ℘◦ . It is an easy exercise to check that this requirement is coordinate free. This
fact is also a consequence of the following application of the Morse lemma (without
parameters, cf. Lemma 18.A.3).
Proposition 20.2.5 Let Ω be an open subset of R2𝑛 and the function 𝑓 ∈ C ∞ (Ω)
be real-valued. If ℘◦ ∈ Ω is a saddle point of 𝑓 then there are C ∞ coordinates
𝑥1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛 in a neighborhood 𝑈 ⊂ Ω of ℘◦ such that 𝑥 𝑗 (℘◦ ) = 𝑦 𝑗 (℘◦ ) = 0
for all 𝑗 = 1, ..., 𝑛, and
1 2
𝑓 (𝑥, 𝑦) = 𝑓 (℘◦ ) + |𝑥| − |𝑦| 2
2
in 𝑈.

Example 20.2.6 Take Ω = C𝑛 and 𝑓 (𝑧) = 1
2 𝛼 |𝑥| 2 − 𝛽 |𝑦| 2 , 𝛼 > 0, 𝛽 > 0; the
origin is a saddle point of 𝑓 . If 𝜅 2 > 𝛼/𝛽 the contour

𝔠 = {𝑧 ∈ C𝑛 ; |𝑥| < 1, 𝑦 = 𝜅𝑥}

is strongly well-shaped at 0 for 𝑓 .

The preceding observations also apply to integrals such as (20.2): it suffices to
substitute (𝑤, 𝜃) for 𝑧. Thus ℭ will be a contour of integration in (𝑤, 𝜃)-space Ω2 × Θ
in the sense of Definition 20.2.1, well-shaped or strongly well-shaped with respect
to a relevant function 𝑓 (𝑤, 𝜃) ∈ C (Ω; R).
836 20 Germ Fourier Integral Operators in Complex Space

20.2.2 Holomorphic function spaces defined by an exponent-weight

Let Ω ⊂ C𝑛 be an open set and Φ a continuous function Ω −→ R.

Definition 20.2.7 We shall denote by O (Φ) (Ω) the space of measurable functions
ℎ (𝑧, 𝜆) : Ω× [1, +∞) −→ C, holomorphic with respect to 𝑧 and having the following
property: to every compact set 𝐾 ⊂ Ω and every 𝜀 > 0 there is a positive constant
𝐶𝐾 , 𝜀 such that

∀𝑧 ∈ 𝐾, ∀𝜆 ≥ 1, |ℎ (𝑧, 𝜆)| ≤ 𝐶𝐾 , 𝜀 e𝜆(Φ(𝑧)+𝜀) . (20.2.1)

We shall refer to Φ as an exponent-weight.

A basic motivation for introducing Definition 20.2.7 is to delimit more precisely
the range of a transform such as (20.1) and compose it with transformations within
an array of holomorphic function spaces.
The space O (Φ) (Ω) equipped with its natural locally convex topology, defined
by the seminorms

O (Φ) (Ω) ∋ ℎ ↦→ sup e−𝜆(Φ(𝑧)+𝜀) |ℎ (𝑧, 𝜆)| ,

𝑧 ∈𝐾,𝜆≥1

(𝐾 ⊂ Ω compact, 𝜀 > 0) is a Fréchet–Montel space. The linear spaces O𝑧(Φ) ◦ of

germs at 𝑧◦ of functions ℎ ∈ O (Φ) (Ω) are the stalks of a sheaf O (Φ) over Ω.
Proposition 20.2.8 If Φ1 and Φ2 are real-valued continuous functions in Ω such
that Φ1 ≤ Φ2 then O (Φ1 ) (Ω) ⊂ O (Φ2 ) (Ω) and O𝑧(Φ◦ 1 ) ⊂ O𝑧(Φ◦ 2 ) for every 𝑧 ◦ ∈ Ω; the
embeddings are continuous.
This is self-evident.
Remark 20.2.9 It is sometimes convenient to use seminorms based on norms other
than the maximum norm, for instance the 𝐿 2 norm. One may want to use the space
O𝐿(Φ)
2 ,loc (Ω) of measurable functions ℎ (𝑧, 𝜆) : Ω × [1, +∞) −→ C, holomorphic with
respect to 𝑧 and having the following property: to every compact set 𝐾 ⊂ Ω and
every 𝜀 > 0 there is a positive constants 𝐶𝐾 , 𝜀 such that
∫
∀𝜆 ≥ 1, e−2𝜆Φ(𝑧) |ℎ (𝑧, 𝜆)| 2 d𝑥d𝑦 ≤ 𝐶𝐾 , 𝜀 e 𝜀𝜆 .
𝐾

The space O𝐿(Φ)

2 ,loc (Ω) is equipped with the topology defined by the seminorms

∫ 21
−𝜀𝜆 −2𝜆Φ(𝑧) 2
ℎ ↦→ sup e e |ℎ (𝑧, 𝜆)| d𝑥d𝑦 .
𝜆≥1 𝐾

The stability of the spaces O (Φ) (Ω) under analytic differential operators is an
important property.
20.2 Contours and Function Spaces 837

Proposition 20.2.10 Let Φ : Ω −→ R be a continuous function and ℎ ∈ O (Φ) (Ω).

To each compact subset 𝐾 of Ω and each 𝜀 > 0 there is a constant 𝑀𝐾 , 𝜀 > 0 such
that, for every 𝛼 ∈ Z+𝑛 ,

∀𝑧 ∈ 𝐾, e−𝜆Φ(𝑧) 𝜕𝑧𝛼 ℎ (𝑧, 𝜆) ≤ 𝑀𝐾| 𝛼, 𝜀|+1 𝛼!e 𝜀𝜆 . (20.2.2)

Thus ℎ ∈ O (Φ) (Ω) implies 𝜕𝑧𝛼 ℎ ∈ O (Φ) (Ω).

Proof Let 𝐾 ⊂⊂ Ω and 𝜀 > 0 be arbitrary. We use the notation

n o
Δ𝑟𝑛 (𝑧∗ ) = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 − 𝑧∗𝑗 ≤ 𝑟, 𝑗 = 1, ..., 𝑛 , 𝑧∗ ∈ C𝑛 .
Ø
We select 𝑟 ◦ > 0 such that 𝐾𝑟◦ = Δ𝑟◦ (𝑧 ∗ ) is a compact subset of Ω. The Cauchy
𝑧 ∗ ∈𝐾
inequalities imply, for arbitrary 𝑟 ∈ (0, 𝑟 ◦ ) and 𝑧∗ ∈ 𝐾,

−𝜆Φ(𝑧 ∗ ) ∗ −| 𝛼 | 𝜆(Φ(𝑧)−Φ(𝑧 ∗ ))
e 𝛼
𝜕𝑧 ℎ (𝑧 , 𝜆) ≤ 𝛼!𝑟 max ∗ e max e−𝜆Φ(𝑧) |ℎ (𝑧, 𝜆)|
𝑧 ∈𝜕Δ𝑟 (𝑧 ) 𝑧 ∈𝜕Δ𝑟 (𝑧 ∗ )
1 ∗ ))
≤ 𝛼!𝑟 −| 𝛼 | 𝐶𝐾 , 𝜀/2 e 2 𝜀𝜆 max ∗ e𝜆(Φ(𝑧)−Φ(𝑧
𝑧 ∈𝜕Δ𝑟 (𝑧 )

where [cf. (20.2.1)]

−𝜆Φ− 21 𝜀𝜆
𝐶𝐾 , 𝜀/2 = sup maxe |ℎ| .
𝜆≥1 𝐾𝑟◦

By the Borel–Lebesgue Lemma we can select 𝑟 = 𝑟 (𝜀) > 0 so small that

1
∀𝑧∗ ∈ 𝐾, max (Φ (𝑧) − Φ (𝑧∗ )) ≤ 𝜀,
𝑧 ∈𝜕Δ𝑟 (𝑧 ∗ ) 2

whence
−𝜆Φ(𝑧 ∗ )
max
∗
e 𝜕𝑧
𝛼
ℎ (𝑧 ∗
, 𝜆) ≤ 𝛼!𝑟 (𝜀) − | 𝛼 | 𝐶𝐾 , 𝜀/2 e 𝜀𝜆 .
𝑧 ∈𝐾

Taking 𝑀𝐾 , 𝜀 = max 𝑟 (𝜀) −1 , 𝐶𝐾 , 𝜀/2 proves (20.2.2). □

Remark 20.2.11 In what precedes the case Φ ≡ 0 is not precluded. When Φ (𝑧) < 0
for some 𝑧 ∈ Ω we have O (0) (Ω) ⊄ O (Φ) (Ω); but, obviously, O (Φ) (Ω) is a
O (0) (Ω)-module (with respect to ordinary addition and multiplication). Note that
if Θ is an open subset of C 𝑁 then O (0) (Ω × Θ) = Sa (Ω × Θ) (Definition 20.1.1).
When Φ ≤ 0, O (Φ) (Ω) is an ideal in the commutative ring O (0) (Ω).

Corollary 20.2.12 Let Φ : Ω −→ R be a continuous function. A hyperdifferential

operator in Ω (Definition 8.1.14),
∑︁
𝐽 (𝑧, D𝑧 ) = 𝑎 𝛼 (𝑧) D𝑧𝛼 ,
𝛼∈Z+𝑛
838 20 Germ Fourier Integral Operators in Complex Space

defines an endomorphism of O (Φ) (Ω).

Proof Combine (20.2.2) with (8.1.15). □

We shall make frequent use of the rings O (−𝜔) (Ω) and O𝑧(−𝜔)
◦ (Definition
19.1.10).
Proposition 20.2.13 Let 𝑧◦ ∈ Ω ⊂ C𝑛 be arbitrary and Φ ∈ C (Ω; R). If Φ (𝑧◦ ) ≥ 0
then O𝑧(−𝜔)
◦ ⊂ O𝑧(Φ)
◦ .

Proof Let 𝑈 ⊂ Ω be a neighborhood of 𝑧◦ and ℎ ∈ O (𝑈) be such that sup |ℎ| ≤

𝑈
𝐶e−𝜅𝜆 for some positive constants 𝐶, 𝜅. After contracting 𝑈 about 𝑧◦ we may assume
inf Φ ≥ −𝜅, whence sup e−𝜆Φ |ℎ| < +∞. □
𝑈 𝑈

Generally speaking, one cannot say that O𝑧(−𝜔)

◦ is an ideal in O𝑧(Φ)
◦ since O𝑧(Φ)
◦ is
not a ring with respect to ordinary multiplication unless Φ ≤ 0 in a neighborhood
of 𝑧 ◦ .

20.2.3 Representations of the identity operator in O𝒛(𝚽)

◦

Let Ω be a domain in C𝑛 and 𝑧◦ ∈ Ω. In this subsection Φ ∈ C ∞ (Ω; R) shall be

such that Φ (𝑧◦ ) = 0, dΦ (𝑧◦ ) = 0. As an (important) example of (20.2) we look at
the integral
𝑛 ∫
𝜆 ◦
𝑰 ℭ (𝑧) ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) · (𝜁 −𝜁 ) ℎ (𝑤, 𝜆) 𝐷 (𝑤, 𝜁) d𝑤d𝜁 (20.2.3)
2𝜋 ℭ (𝑧)

for an appropriate choice of the contour of integration ℭ (𝑧) ⊂ Ω × C𝑛 , ℭ (𝑧) ∋

(𝑧◦ , 𝜁 ◦ ), and a concomitant choice of 𝐷 (𝑤, 𝜁); throughout, ℎ ∈ O (Φ) (Ω). All
the contours of integration shall be totally real smooth submanifolds of C2𝑛 of
(real) dimension 2𝑛, not all bounded. But we shall always assume that (𝑤, 𝜁) ∈ ℭ (𝑧)
entails 𝑤 ∈ 𝐾, a compact subset of Ω.
There is no loss of generality in assuming 𝑧◦ = 0. We select the contour ℭ (𝑧) to
be the set of points (𝑢 + 𝑖𝑣, 𝜉 + 𝑖𝜂) such that

|𝑢| < 𝑟, 𝑣 = 0, 𝜉 = 𝜉 ◦ + 𝜃, (20.2.4)

1
𝜃 ∈ R𝑛 , |𝜃| < 𝜌, 𝜂 = 𝜂◦ + |𝜃| (𝑧 − 𝑢) ,
2
where 𝑟 > 0 is small enough that |𝑧| ≤ 𝑟 entails 𝑧 ∈ 𝐾; 𝜌 > 0 is arbitrary. We get

𝑰 ℭ (𝑧) ℎ (𝑧, 𝜆) (20.2.5)

𝑛 ∫ ∫
𝜆 1 2
= e𝑖𝜆𝜃 · (𝑧−𝑢)− 2 𝜆 | 𝜃 | ⟨𝑧−𝑢⟩ ℎ (𝑢, 𝜆) Δ 1 (𝑧 − 𝑢, 𝜃) d𝑢d𝜃,
2𝜋 | 𝜃 |<𝜌 |𝑢 |<𝑟 2
20.2 Contours and Function Spaces 839

where ⟨𝑧⟩ 2 = 𝑧 · 𝑧 and [cf. (3.4.6)]

1 𝜃
Δ 1 (𝑧, 𝜃) = 1 + 𝑖 𝑧 · .
2 2 |𝜃|

We refer the reader to Remark 3.4.4: if we define the FBI transform of the function
of 𝑢 equal to ℎ (𝑢, 𝜆) if |𝑢| < 𝑟 and to zero if |𝑢| ≥ 𝑟, by the formula
∫
1 2
𭟋 1 ℎ(𝑧, 𝜃, 𝜆) = e𝑖 𝜃 · (𝑧−𝑢)− 2 | 𝜃 | ⟨𝑧−𝑢⟩ ℎ (𝑢, 𝜆) Δ 1 (𝑧 − 𝑢, 𝜃) d𝑢
2 2
|𝑢 |<𝑟

then [cf. (3.4.10)], for |𝑢| < 𝑟,

𝑛 ∫
𝜆
ℎ (𝑢, 𝜆) = 𭟋 1 ℎ(𝑢, 𝜆𝜃, 𝜆)d𝜃.
2𝜋 R𝑛 2

By holomorphic extension we have, for 𝑧 ∈ C𝑛 , |𝑧| < 𝑟,

𝑛 ∫
𝜆
ℎ (𝑧, 𝜆) = 𭟋 1 ℎ(𝑧, 𝜃, 𝜆)d𝜃.
2𝜋 R𝑛 2

We derive, from this and from (20.2.4), (20.2.5),

−𝑛
𝜆
ℎ (𝑧, 𝜆) − 𝑰 ℭ (𝑧) ℎ (𝑧, 𝜆) (20.2.6)
2𝜋
∫ ∫
1 2
= e𝑖𝜆(𝑧−𝑢) · 𝜃− 2 𝜆 | 𝜃 | ⟨𝑧−𝑢⟩ ℎ (𝑢, 𝜆) Δ 1 (𝑧 − 𝑢, 𝜃) d𝑢d𝜃.
2
| 𝜃 |>𝜌 |𝑢 |<𝑟

We underline the fact that the right-hand side is an entire function of 𝑧 for each fixed
𝜆 ≥ 1.
By Definition 20.2.7 to every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that
(Φ)
∥ℎ∥ 𝐾 = max e−𝜆Φ(𝑤) |ℎ (𝑤, 𝜆)| ≤ 𝐶 𝜀 e 𝜀𝜆 . (20.2.7)
𝑤 ∈𝐾

Proposition 20.2.14 Let Ω′ be a domain in C𝑛 , 𝑧◦ ∈ Ω′ ⊂⊂ Ω. If 𝑟 > 0 and diam Ω′

are sufficiently small and 𝜌 is sufficiently large there are positive constants 𝐶, 𝜅,
such that
(Φ)
sup ℎ (𝑧, 𝜆) − 𝑰 ℭ (𝑧) ℎ (𝑧, 𝜆) ≤ 𝐶e−𝜅𝜆 ∥ℎ∥ 𝐾 ,
𝑧 ∈Ω′

where ℎ ∈ O (Φ) (Ω) is arbitrary and 𝐾 = Ω′.

Proof In the right-hand side in n(20.2.6) we deform the contour

o of 𝑢-integration from
𝜃
the ball |𝑢| < 𝑟 to the domain 𝑤 ∈ Ω; |𝑢| < 𝑟, 𝑣 = −𝛿 | 𝜃 | (𝛿 > 0 small). By the
Cauchy Integral Theorem we find
840 20 Germ Fourier Integral Operators in Complex Space
∫
𭟋 1 ℎ(𝑧, 𝜆𝜃, 𝜆) = e𝑖𝜆𝐸1 (𝑧,𝑤, 𝜃) ℎ (𝑤, 𝜆) Δ 1 (𝑧 − 𝑤, 𝜃) d𝑢 (20.2.8)
2
|𝑢 |<𝑟 2 𝑤=𝑢−𝑖 𝛿 | 𝜃𝜃 |
∫ 1∫
+ e𝑖𝜆𝐸𝑡 (𝑧,𝑤, 𝜃) ℎ (𝑤, 𝜆) Δ 1 (𝑧 − 𝑤, 𝜃) d𝑢d𝑡,
0 |𝑢 |=𝑟 2 𝑤=𝑢−𝑖𝑡 𝛿 | 𝜃𝜃 |

where
2
𝜃 𝜃 1 𝜃
𝐸𝑡 𝑧, 𝑢 − 𝑖𝑡𝛿 , 𝜃 = 𝑧 − 𝑢 + 𝑖𝑡𝛿 · 𝜃 + 𝑖 |𝜃| 𝑧 − 𝑢 + 𝑖𝑡𝛿 ,
|𝜃| |𝜃| 2 |𝜃|

0 ≤ 𝑡 ≤ 1. Since Φ (𝑧 ◦ ) = 0, dΦ (𝑧◦ ) = 0, we have

𝜃
Φ 𝑢 − 𝑖𝑡𝛿 ≤ 𝐶◦ |𝑢| 2 + 𝛿2 . (20.2.9)
|𝜃|

We derive

𝜃 1
Im 𝐸 1 𝑧, 𝑢 − 𝑖𝑡𝛿 , 𝜃 = 𝑦 · 𝜃 + 𝛿 |𝜃| + |𝜃| |𝑥 − 𝑢| 2 − (|𝑦| + 𝛿) 2 ,
|𝜃| 2

whence
1
(Φ (𝑤) − Im 𝐸 1 (𝑧, 𝑤, 𝜃))| 𝑤=𝑢−𝑖 𝛿 𝜃 ≤ − (𝛿 − |𝑦|) |𝜃|+𝐶◦ 𝑟 2 + 𝛿2 + (|𝑦| + 𝛿) 2 |𝜃| .
|𝜃| 2
By requiring 𝛿, 𝑟 and |𝑦| /𝛿 to be sufficiently small we get

𝜃 𝜃 1 1
Φ 𝑢 − 𝑖𝑡𝛿 − Im 𝐸 1 𝑧, 𝑢 − 𝑖𝑡𝛿 , 𝜃 ≤ − 𝛿 |𝜃| ≤ − 𝛿𝜌.
|𝜃| |𝜃| 2 2

We reach the conclusion that

∫ ∫
e𝑖𝜆𝐸1 (𝑧,𝑤, 𝜃) ℎ (𝑤, 𝜆) Δ 1 (𝑧 − 𝑤, 𝜃) d𝑢d𝜃 (20.2.10)
| 𝜃 |>𝜌 |𝑢 |<𝑟 2 𝑤=𝑢−𝑖 𝛿 | 𝜃𝜃 |
∫
1 (Φ) 1 (Φ)
≲ e− 4 𝜆 𝛿 | 𝜃 | d𝜃 ∥ℎ∥ 𝐾 ≲ 𝛿−𝑛 e− 2 𝛿𝜌𝜆 ∥ℎ∥ 𝐾 .
| 𝜃 |>𝜌

Next we look at the second integral on the right-hand side of (20.2.8). We have

𝜃 2

𝜃 1 1
Im 𝐸 𝑡 𝑧, 𝑢 − 𝑖𝑡𝛿 , 𝜃 = 𝑦 · 𝜃 − 𝑡𝛿 |𝜃| + |𝜃| |𝑥 − 𝑢| 2 − |𝜃| 𝑦 + 𝑡𝛿 .
|𝜃| 2 2 |𝜃|

We obtain, by (20.2.9),

(Φ (𝑤) − Im 𝐸 1 (𝑧, 𝑤, 𝜃))| 𝑤=𝑢−𝑖𝑡 𝛿 𝜃

|𝜃|

1 1 1
≤ − |𝜃| |𝑥 − 𝑢| 2 + |𝜃| |𝑦| + |𝑦| 2 + 𝛿2 + 𝐶◦ 𝑟 2 + 𝛿2 .
2 2 2
20.3 Sjöstrand Pairs 841

We require 𝑧 ∈ 𝐾, |𝑢| = 𝑟, to imply |𝑥 − 𝑢| 2 ≥ 𝑟 2 /2, |𝑦| + 1

2 |𝑦| 2 + 12 𝛿2 ≤ 𝛿. By
requiring 𝛿 + 𝑟 ≪ 𝜌 we obtain
∫ ∫ 1∫
e𝑖𝜆𝐸𝑡 (𝑧,𝑤, 𝜃) ℎ (𝑤, 𝜆) Δ 1 (𝑧 − 𝑤, 𝜃) d𝑢d𝑡d𝜃
| 𝜃 |>𝜌 0 |𝑢 |=𝑟 2 𝑤=𝑢−𝑖𝑡 𝛿 | 𝜃𝜃 |
−𝑐𝜌𝜆 (Φ)
≲ e ∥ℎ∥ 𝐾

for some 𝑐 > 0. Combining this with (20.2.10) completes the proof of Proposition
20.2.14. □

Corollary 20.2.15 Assume the same hypotheses as in Proposition 20.2.14. There

is a neighborhood Ω′ ⊂ Ω of 𝑧◦ such that ℎ − 𝑰 ℭ ℎ ∈ O (−𝜔) (Ω′) [cf. Definition
19.1.10, also (20.2.3)] whatever ℎ ∈ O (Φ) (Ω).

Proof Formula (20.2.5) shows that 𝑰 ℭ (𝑧) ℎ (𝑧, 𝜆) is a holomorphic function of 𝑧 in a

neighborhood of 𝑧 ◦ ; then Proposition 20.2.14 implies the claim. □
It follows from Corollary 20.2.15 that if ℭ1 (𝑧) and ℭ2 (𝑧) are two contours
(20.2.4) defined with different (but sufficiently small) values of 𝑟, 𝑟 1 and 𝑟 2 , then

𝑰 ℭ1 ℎ − 𝑰 ℭ2 ℎ ∈ O (−𝜔) (Ω′) .

Inspection of the proof of Proposition 20.2.14 also shows that the following is
true:

Proposition 20.2.16 Assume the same hypotheses as in Proposition 20.2.14. If 𝑟 is

sufficiently small then 𝑰 ℭ ℎ ∈ O (−𝜔) (Ω′) for every ℎ ∈ O (−𝜔) (Ω).

This and the hypothesis that Φ (𝑧◦ ) = 0 (cf. Proposition 20.2.13) allow us to
contract Ω′ and Ω about 𝑧 ◦ as much as needed and use the integrals (20.2.3) to define
one and the same germ operator I𝑧 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ ←↪. Corollary 20.2.15 then leads
to the following conclusion:

Theorem 20.2.17 If Φ ∈ C ∞ (Ω; R) and Φ (𝑧 ◦ ) = 0, dΦ (𝑧◦ ) = 0, then I𝑧 ◦ , defined

by the integrals (20.2.3), is the identity operator of O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ .

20.3 Sjöstrand Pairs

20.3.1 Sjöstrand pairs. Definition

Throughout this subsection the phase-function 𝜑 shall be defined and holomorphic

in the open subset Ω × Θ of C𝑛 × C 𝑁 where the variable point is (𝑧, 𝜃). The analysis
will take place in neighborhoods of a point (𝑧◦ , 𝜃 ◦ ) ∈ Ω × Θ such that 𝜑 (𝑧◦ , 𝜃 ◦ ) = 0.
842 20 Germ Fourier Integral Operators in Complex Space

We introduce an exponent-weight Φ ∈ C ∞ (Ω; R) (Definition 20.2.7); keep in

mind that Φ (𝑧◦ ) = 0, ensuring that O𝑧(−𝜔)
◦ ⊂ O𝑧(Φ)
◦ (Proposition 20.2.13). Our
further requirements on 𝜑 and Φ will involve the function, defined and C ∞ in Ω × Θ,

Ψ (𝑧, 𝜃) = Φ (𝑧) − Im 𝜑 (𝑧, 𝜃) . (20.3.1)

Remark 20.3.1 If Φ is plurisubharmonic in Ω (Definition 11.2.6) then Ψ is plurisub-

harmonic in Ω × Θ since Im 𝜑 (𝑧, 𝜃) is pluriharmonic in Ω × Θ.
The critical points of 𝑧 ↦→ Ψ (𝑧, 𝜃), denoted by 𝑧 (𝜃), are the solutions of the
equations

𝜕𝑥 Φ (𝑧) = 𝜕𝑥 Im 𝜑 (𝑧, 𝜃) , (20.3.2)

𝜕𝑦 Φ (𝑧) = 𝜕𝑦 Im 𝜑 (𝑧, 𝜃) = 𝜕𝑥 Re 𝜑 (𝑧, 𝜃) ,

i.e.,
1
𝜕𝑧 Φ (𝑧) = 𝜕𝑧 Im 𝜑 (𝑧, 𝜃) = 𝜕𝑧 𝜑 (𝑧, 𝜃) . (20.3.3)
2𝑖
In the sequel we assume that 𝑧 ◦ is a critical point of Ψ (𝑧, 𝜃 ◦ ); this implies
1
𝜕𝑧 Φ (𝑧 ◦ ) = 𝜕𝑧 𝜑 (𝑧 ◦ , 𝜃 ◦ ) . (20.3.4)
2𝑖
Lemma 20.3.2 Suppose that 𝑧◦ is a saddle point (Definition 20.2.4) of the function
Ψ (𝑧, 𝜃 ◦ ). After contracting Ω about 𝑧◦ and Θ about 𝜃 ◦ the equation (20.3.3) has
a unique C ∞ solution 𝑧 (𝜃) such that 𝑧 (𝜃 ◦ ) = 𝑧 ◦ . For each 𝜃 ∈ Θ the following
properties hold:
(1) 𝑧 (𝜃) is an isolated critical point of 𝑧 ↦→ Ψ (𝑧, 𝜃);
(2) 𝑧 (𝜃) is a saddle point of the function 𝑧 ↦→ Ψ (𝑧, 𝜃).
Proof The fact that 𝑧◦ is a saddle point of Ψ (𝑧, 𝜃 ◦ ) requires that the Hessian of
Ψ (𝑧, 𝜃) with respect to (𝑥, 𝑦) be nonsingular at 𝑧◦ when 𝜃 = 𝜃 ◦ ; this remains true
when 𝜃 varies in Θ suitably contracted about 𝜃 ◦ . From this we deduce, by applying
the Implicit Function Theorem to the system of 2𝑛 equations (20.3.2), the existence
and uniqueness of the C ∞ solution 𝑧 (𝜃) of (20.3.3) such that 𝑧 (𝜃 ◦ ) = 𝑧 ◦ , which
implies Claim (1). Claim (2) follows directly from the stability of the signature of
the Hessian. □
When Property (1) in Lemma 20.3.2 holds

Φ∗ (𝜃) = Ψ (𝑧 (𝜃) , 𝜃) (20.3.5)

is the critical value of 𝑧 ↦→ Ψ (𝑧, 𝜃).

Lemma 20.3.3 Suppose that 𝑧◦ is a saddle point of the function Ψ (𝑧, 𝜃 ◦ ). The real
rank at 𝜃 ◦ of the map 𝜃 ↦→ 𝑧 (𝜃) in Lemma 20.3.2 is equal to twice the complex rank
of 𝜕𝑧 𝜕𝜃 𝜑 (𝑧◦ , 𝜃 ◦ ).
20.3 Sjöstrand Pairs 843

Proof We differentiate the equations

𝜕𝑥 Ψ (𝑧 (𝜃) , 𝜃) = 0, 𝜕𝑦 Ψ (𝑧 (𝜃) , 𝜃) = 0

with respect to 𝜎 = Re 𝜃, 𝜏 = Im 𝜃; by (20.3.3) we get, at (𝑧◦ , 𝜃 ◦ ),

2
𝜕 𝑥 Ψ 𝜕 𝑥 𝜕𝑦 Ψ 𝜕 𝜎 𝑥 𝜕 𝜏 𝑥 𝜕𝜎 𝜕𝑥 Im 𝜑 𝜕𝜏 𝜕𝑥 Im 𝜑
= .
𝜕𝑥 𝜕𝑦 Ψ 𝜕𝑦2 Ψ 𝜕 𝜎 𝑦 𝜕𝜏 𝑦 𝜕𝜎 𝜕𝑦 Im 𝜑 𝜕𝜏 𝜕𝑦 Im 𝜑

If 𝑧 ◦ is a saddle point of the function Ψ (𝑧, 𝜃 ◦ ) then the matrix [conjugate to

𝜕𝑥 𝜕𝑦 Φ (𝑧◦ , 𝜃 ◦ )]
𝜕𝑧 Ψ (𝑧◦ , 𝜃 ◦ ) 𝜕𝑧 𝜕𝑧¯ Ψ (𝑧 ◦ , 𝜃 ◦ )
2

𝜕𝑧 𝜕𝑧¯ Ψ (𝑧 ◦ , 𝜃 ◦ ) 𝜕𝑧2¯ Ψ (𝑧◦ , 𝜃 ◦ )

is invertible. Still at (𝑧 ◦ , 𝜃 ◦ ) we have, whatever (𝑢, 𝑣) ∈ R2𝑛 ,

𝜕𝜎 𝜕𝑥 Im 𝜑 𝜕𝜏 𝜕𝑥 Im 𝜑 𝑢 𝜕 𝜕 Im 𝜑
= Re (𝑢 + 𝑖𝑣) 𝜃 𝑥
𝜕𝜎 𝜕𝑦 Im 𝜑 𝜕𝜏 𝜕𝑦 Im 𝜑 𝑣 𝜕𝜃 𝜕𝑦 Im 𝜑

1 1 𝜕 𝜕 𝜑
= Re (𝑢 + 𝑖𝑣) 𝑥 𝜃
2 𝑖 𝜕𝑦 𝜕 𝜃 𝜑

1 −𝑖 (𝜕𝑧 𝜕𝜃 𝜑) (𝑢 + 𝑖𝑣)
= Re
2 (𝜕𝑧 𝜕𝜃 𝜑) (𝑢 + 𝑖𝑣)

1 Im ((𝜕𝑧 𝜕𝜃 𝜑) (𝑢 + 𝑖𝑣))
= .
2 Re ((𝜕𝑧 𝜕𝜃 𝜑) (𝑢 + 𝑖𝑣))

𝜕𝜎 𝑥 (𝜃 ◦ ) 𝜕𝜏 𝑥 (𝜃 ◦ )

This implies that the (real) ranks of and (𝜕𝜃 𝜕𝑧 𝜑) (𝑧◦ , 𝜃 ◦ ) are
𝜕𝜎 𝑦 (𝜃 ◦ ) 𝜕𝜏 𝑦 (𝜃 ◦ )
equal. □
In the sequel we shall reason under the following hypothesis:
2
𝜕 𝜑
rank (𝑧◦ , 𝜃 ◦ ) = min (𝑛, 𝑁) . (20.3.6)
𝜕𝑧 𝑗 𝜕𝜃 𝑘 1≤ 𝑗 ≤𝑛
1≤𝑘 ≤ 𝑁

Corollary 20.3.4 Suppose that (20.3.6) holds and 𝑁 ≤ 𝑛. If Θ is sufficiently small

then, as 𝜃 ranges over Θ, the critical points of the map 𝑧 ↦→ Ψ (𝑧, 𝜃) describe a
2𝑁-dimensional C ∞ submanifold of Ω.

Corollary 20.3.5 Suppose (20.3.6) holds and 𝑁 ≥ 𝑛. After appropriate contractions

of Ω about 𝑧 ◦ and of Θ about 𝜃 ◦ the map 𝜃 ↦→ 𝑧 (𝜃) in Lemma 20.3.2 is a C ∞
surjection Θ −→ Ω of constant rank 2𝑛.

Thus, under the hypotheses of Corollary 20.3.5, we can represent

Θ as a fiber bun-
dle over Ω with fibers the (𝑁 − 𝑛)-dimensional C ∞ submanifolds 𝜃 ∈ Θ; 𝑧 (𝜃) = 𝑧♭ ,
𝑧♭ ∈ Ω.
844 20 Germ Fourier Integral Operators in Complex Space

Definition 20.3.6 We shall say that the phase-function 𝜑 ∈ O (Ω × Θ) and the

exponent-weight Φ ∈ C ∞ (Ω; R) form a Sjöstrand pair at (𝑧 ◦ , 𝜃 ◦ ) ∈ Ω × Θ if

Im 𝜑 (𝑧◦ , 𝜃 ◦ ) = Φ (𝑧 ◦ ) = 0 (20.3.7)

and if 𝑧 ◦ is a saddle point of the function Ψ (𝑧, 𝜃 ◦ ).

Note that (20.3.5) and (20.3.7) imply

Φ∗ (𝜃 ◦ ) = 0. (20.3.8)

Proposition 20.3.7 Let (𝑧 ◦ , 𝜃 ◦ ) ∈ Ω × Θ. To every 𝜑 ∈ O (Ω × Θ) such that

Im 𝜑 (𝑧◦ , 𝜃 ◦ ) = 0 there is a Φ ∈ C ∞ (Ω; R) such that (𝜑, Φ) is a Sjöstrand pair
at (𝑧◦ , 𝜃 ◦ ). The function Φ can be taken to be strictly plurisubharmonic (Definition
11.2.11).
Proof The claim is a statement about the Taylor expansion of order 2 about 𝑧◦ of
the function Ψ (𝑧, 𝜃 ◦ ) = Φ (𝑧) − Im 𝜑 (𝑧, 𝜃 ◦ ). Taking 𝑧 ◦ = 0 we assume that

𝜑 (𝑧, 𝜃 ◦ ) = 𝜑 (𝑧◦ , 𝜃 ◦ ) + (𝑏 ◦ + 𝑖𝑎 ◦ ) · 𝑧 + (𝐵◦ + 𝑖 𝐴◦ ) 𝑧 · 𝑧 + 𝑂 |𝑧| 3 ,

where 𝑎 ◦ , 𝑏 ◦ ∈ R𝑛 and 𝐴◦ , 𝐵◦ real symmetric 𝑛 × 𝑛 matrices. We take

Φ (𝑧) = 𝑎 · 𝑥 + 𝑏 · 𝑦 + 𝐴𝑥 · 𝑥 + 2𝐵◦ 𝑥 · 𝑦 + 𝐶 𝑦 · 𝑦.

Thus we are dealing with the real polynomial of degree 2,

𝐹 (𝑥, 𝑦) = 𝑎 · 𝑥 + 𝑏 · 𝑦 + 𝐴𝑥 · 𝑥 + 2𝐵𝑥 · 𝑦 + 𝐶 𝑦 · 𝑦 − 𝑃◦ (𝑥, 𝑦) ,

where 𝑎, 𝑏 ∈ R𝑛 , 𝐴, 𝐵, 𝐶 are real symmetric 𝑛 × 𝑛 matrices, and

𝑃◦ (𝑥, 𝑦) = 𝑐 ◦ + 𝑎 ◦ · 𝑥 + 𝑏 ◦ · 𝑦 + 𝐴◦ 𝑥 · 𝑥 + 2𝐵◦ 𝑥 · 𝑦 − 𝐴◦ 𝑦 · 𝑦

with 𝑐 ◦ = − Im 𝜑 (0, 𝜃 ◦ ).
We need 0 ∈ R2𝑛 to be a saddle point, in particular a critical point, of 𝐹 (cf.
Definition 20.2.4); thus we must have 𝑎 ◦ = 𝑎, 𝑏 ◦ = 𝑏. Moreover, we need the 2𝑛 × 2𝑛
real symmetric matrix
𝐴 − 𝐴◦ 𝐵 − 𝐵 ◦
𝐵 − 𝐵 ◦ 𝐶 + 𝐴◦
to be nonsingular and have signature equal to zero. It suffices to select 𝐴 and 𝐶 such
that 𝐴◦ − 𝐴 and 𝐴◦ + 𝐶 are positive definite and, say, 𝐵 = 𝐵◦ . We can moreover take
𝐶 − 𝐴 positive definite, to ensure that Φ is strictly plurisubharmonic. □
The next statement provides a kind of normal form for the function Ψ. It can
be viewed as a complex variant (with greater precision) of Proposition 20.2.5. By
Lemma 20.3.2 𝑧 (𝜃) is a critical point of 𝑧 ↦→ Ψ (𝑧, 𝜃), therefore
1
Ψ (𝑧, 𝜃) = Φ∗ (𝜃) + 𝑄 (𝜃; 𝑧 − 𝑧 (𝜃)) + 𝑂 |𝑧 − 𝑧 (𝜃)| 3 , (20.3.9)
2
20.3 Sjöstrand Pairs 845

where 𝑧 ↦→ 𝑄 (𝜃; 𝑧) is a (real) quadratic form in R2𝑛 with coefficients in C ∞ (Θ).

Proposition 20.3.8 There is a 𝑇 ∈ GL (𝑛, C) such that, for all (𝑧, 𝜃) ∈ Ω × Θ

[possibly contracted about (𝑧◦ , 𝜃 ◦ )]
𝑛
∑︁ 2
𝑄 (𝜃; 𝑇 (𝑧 − 𝑧 (𝜃))) = − 𝜒 𝑗 𝑥 𝑗 − 𝑥 𝑗 (𝜃) (20.3.10)
𝑗=1
𝑛
∑︁
2
+ 𝜒 ′𝑗 𝑦 𝑗 − 𝑦 𝑗 (𝜃) + 𝑂 |𝑧 − 𝑧◦ | 3 + |𝜃 − 𝜃 ◦ | 3 ,
𝑗=1

with 𝜒 𝑗 > 0, 𝜒 ′𝑗 > 0, 𝑗 = 1, ..., 𝑛.

Proof By hypothesis 𝑧 ◦ is a saddle point of the function

1
Ψ (𝑧, 𝜃 ◦ ) = Ψ (𝑧◦ , 𝜃 ◦ ) + 𝑄 (𝜃 ◦ ; 𝑧 − 𝑧 ◦ ) + 𝑂 |𝑧 − 𝑧 ◦ | 3 .
2
Lemma 11.2.28, implies that there is a 𝑇 ∈ GL (𝑛, C) such that
𝑛
∑︁ 2
𝑄 (𝜃 ◦ ; 𝑇 (𝑧 − 𝑧 (𝜃))) = − 𝜒 𝑗 𝑥 𝑗 − 𝑥 𝑗 (𝜃)
𝑗=1
𝑛
∑︁ 2
+ 𝜒 ′𝑗 𝑦 𝑗 − 𝑦 𝑗 (𝜃)
𝑗=1

with 𝜒 𝑗 > 0, 𝜒 ′𝑗 > 0, 𝑗 = 1, ..., 𝑛. We have

|𝑄 (𝜃; 𝑇 (𝑧 − 𝑧 (𝜃))) − 𝑄 (𝜃 ◦ ; 𝑇 (𝑧 − 𝑧 (𝜃)))| ≲ |𝑧 − 𝑧 (𝜃)| 2 |𝜃 − 𝜃 ◦ |

for all 𝑧 ∈ C𝑛 , 𝜃 ∈ Θ′; also,

|𝑧 − 𝑧 (𝜃)| 2 − 2 |𝑧 − 𝑧◦ | 2 ≤ 2 |𝑧◦ − 𝑧 (𝜃)| 2 ≲ |𝜃 − 𝜃 ◦ | 2 ,

whence (20.3.10). □

The numbers −𝜒1 , ..., −𝜒𝑛 , 𝜒1′ , ..., 𝜒𝑛′ are the eigenvalues of the Hessian of Ψ with
respect to (𝑥, 𝑦), Hess 𝑥,𝑦 Ψ (𝑧, 𝜃), at (𝑧◦ , 𝜃 ◦ ). We can state (cf. Lemma 11.2.28)

Corollary 20.3.9 For 𝑧 ↦→ Ψ (𝑧, 𝜃) to be strictly plurisubharmonic in a neighbor-

hood of (𝑧◦ , 𝜃 ◦ ) it is necessary and sufficient that 0 < 𝜒 𝑗 < 𝜒 ′𝑗 , 𝑗 = 1, ..., 𝑛.
846 20 Germ Fourier Integral Operators in Complex Space

20.3.2 Dual Sjöstrand pairs

20.3.2.1 The case 𝑵 = 𝒏.

In this subsubsection we assume 𝑁 = 𝑛; so as not to lose sight of this fact we

denote by 𝜁 the “fiber variable” instead of 𝜃; 𝜁 varies in a neighborhood Ξ of 𝜁 ◦ in
C𝑛 . Until specified otherwise we require the phase-function 𝜑 ∈ O (Ω × Ξ) to be
nondegenerate in the sense that
2
𝜕 𝜑
∀ (𝑧, 𝜁) ∈ Ω × Ξ, det (𝑧, 𝜁) ≠ 0. (20.3.11)
𝜕𝑧 𝑗 𝜕𝜁 𝑘 1≤ 𝑗,𝑘 ≤𝑛

When referring to (𝜑, Φ) [Φ ∈ C ∞ (Ω; R)] as a Sjöstrand pair it will be tacitly

assumed that these conditions are satisfied.

Proposition 20.3.10 Let (𝜑, Φ) be a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) in Ω × Ξ. There is a

neighborhood Ω′ ⊂ Ω of 𝑧 ◦ and a unique C ∞ map Ω′ ∋ 𝑧 ↦→ 𝜁 (𝑧) ∈ Ξ such that
𝜁 (𝑧◦ ) = 𝜁 ◦ and that 𝜁 (𝑧) satisfies (20.3.3), i.e.,

∀𝑧 ∈ Ω′, 𝜕𝑧 𝜑 (𝑧, 𝜁 (𝑧)) = 2𝑖𝜕𝑧 Φ (𝑧) . (20.3.12)

This map is the inverse of the map Ξ′ ∋ 𝜁 ↦→ 𝑧 (𝜁) ∈ Ω in Lemma 20.3.2.

Proof Indeed, under the present hypotheses, 𝑁 = 𝑛 and (20.3.11), the map in Lemma
20.3.2, is a C ∞ diffeomorphism of Ξ′ onto an open subset Ω′ of Ω mapping 𝜁 ◦ onto
𝑧◦ . □
We introduce the analogue of Ψ for −𝜑 and Φ∗ [see (20.3.5)], namely the function

Ψ∗ (𝑧, 𝜁) = Φ∗ (𝜁) + Im 𝜑 (𝑧, 𝜁) (20.3.13)

= Ψ (𝑧 (𝜁) , 𝜁) + Im 𝜑 (𝑧, 𝜁) ,

where (𝑧, 𝜁) ∈ Ω × Ξ.

Proposition 20.3.11 Let (𝜑, Φ) be a Sjöstrand pair at (𝑧 ◦ , 𝜁 ◦ ) in Ω × Ξ. The critical

points of the function 𝜁 ↦→ Ψ∗ (𝑧, 𝜁) in Ξ are the values of the map 𝜁 (𝑧) in Proposition
20.3.10; moreover,
∀𝑧 ∈ Ω, Φ (𝑧) = Ψ∗ (𝑧, 𝜁 (𝑧)) . (20.3.14)

Proof The inverse of the map 𝜁 ↦→ 𝑧 (𝜁) being the map 𝑧 ↦→ 𝜁 (𝑧) (Proposition
20.3.10) we derive from (20.3.13):

Ψ∗ (𝑧, 𝜁 (𝑧)) = Ψ (𝑧, 𝜁 (𝑧)) + Im 𝜑 (𝑧, 𝜁 (𝑧)) = Φ (𝑧) ,

the last equality a consequence of the definition (20.3.1). □

The identity (20.3.14) is the analogue of (20.3.5): Φ (𝑧) is the critical value of
the function 𝜁 ↦→ Ψ∗ (𝑧, 𝜁).
20.3 Sjöstrand Pairs 847

Corollary 20.3.12 If (𝜑, Φ) is a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) then we have, for all
(𝑧, 𝜁) ∈ Ω × Ξ,
Ψ (𝑧, 𝜁) = Ψ∗ (𝑧, 𝜁 (𝑧)) − Im 𝜑 (𝑧, 𝜁) . (20.3.15)

Proof Combine (20.3.14) with (20.3.1). □

2
Example 20.3.13 Consider the phase-function 𝜑 (𝑧, 𝜁) = 21 𝑖 𝑗=1 𝑧 𝑗 − 𝜁 𝑗 in C𝑛 ×
Í𝑛
C𝑛 ; we have 2
𝜕 𝜑
(𝑧, 𝜁) = −𝑖𝐼𝑛
𝜕𝑧 𝑗 𝜕𝜁 𝑘 𝑗,𝑘=1,...,𝑛

(𝐼𝑛 : the 𝑛 × 𝑛 identity matrix). If Φ (𝑧) = 21 𝛼 |𝑥| 2 − 𝛽 |𝑦| 2 , 𝛼 > 0, 𝛽 > 0, then

1
Ψ (𝑧, 𝜁) = 𝛼 |𝑥| 2 − 𝛽 |𝑦| 2 − |𝑥 − 𝜉 | 2 + |𝑦 − 𝜂| 2
2
(𝑧 = 𝑥 +𝑖𝑦, 𝜁 = 𝜉 +𝑖𝜂). Provided 𝛼 ≠ 1, 𝛽 ≠ 1, the critical points 𝑧 (𝜁) of the function
𝑧 ↦→ Ψ (𝑧, 𝜁) are given by
𝜉 𝜂
𝑥 (𝜁) = , 𝑦 (𝜁) = .
1−𝛼 1−𝛽

Thus (𝜉, 𝜂) ↦→ (𝑥 (𝜁) , 𝑦 (𝜁)) is a linear automorphism of R2𝑛 (cf. Proposition

20.3.8). The origin is the critical point, as well as a saddle point of

∗ 1 𝛼 2 𝛽 2
Φ (𝜁) = Ψ (𝑧 (𝜁) , 𝜁) = − |𝜉 | − |𝜂| .
2 𝛼−1 𝛽−1

For Φ to be plurisubharmonic it is necessary and sufficient that 𝛼 ≥ 𝛽 > 0. Provided

either 𝛽 ≤ 𝛼 < 1 or 1 < 𝛽 ≤ 𝛼, we see directly that Φ∗ (𝜁) is plurisubharmonic
𝑡
since the function 𝑡 ↦→ 𝑡−1 is decreasing in both intervals (0, 1) and (1, +∞). We
have

∗ 1 𝛼 2 𝛽 2 1
Ψ (𝑧, 𝜁) = − |𝜉 | − |𝜂| + |𝑥 − 𝜉 | 2 − |𝑦 − 𝜂| 2 .
2 𝛼−1 𝛽−1 2

The main result of this subsection, Theorem 20.3.16 below, will be a direct
consequence of the following derivation of a normal form of Ψ∗ from that of Ψ
provided in Proposition 20.3.8.

Proposition 20.3.14 Assume the same hypotheses as in Proposition 20.3.8. If

(20.3.10) holds then there is an 𝑆 ∈ GL (𝑛, C) such that
1
Ψ∗ (𝑧, 𝜁) = Φ (𝑧) + 𝑄 ∗ (𝑧; 𝜁 − 𝜁 (𝑧)) + 𝑂 |𝜁 − 𝜁 (𝑧)| 3
2
with
848 20 Germ Fourier Integral Operators in Complex Space

𝑄 ∗ (𝑧; 𝑆 (𝜁 − 𝜁 (𝑧))) (20.3.16)

𝑛 𝑛
∑︁ 1 2 ∑︁ 1 2
=− ′ 𝜉 𝑗 − 𝜉 𝑗 (𝑧) + 𝜂 𝑗 − 𝜂 𝑗 (𝑧)
𝜒
𝑗=1 𝑗
𝜒
𝑗=1 𝑗

+𝑂 |𝑧 − 𝑧◦ | 3 + |𝜁 − 𝜁 ◦ | 3 .

Proof The definitions (20.3.1) and (20.3.5) and Proposition 20.3.8 imply

Ψ (𝑧, 𝜁) = Φ (𝑧) − Im 𝜑 (𝑧, 𝜁)

1
= Φ∗ (𝜁) + 𝑄 (𝜁; 𝑧 − 𝑧 (𝜁)) + 𝑂 |𝑧 − 𝑧 (𝜁)| 3 ,
2
whence
1
Φ (𝑧) − Φ∗ (𝜁) = Im 𝜑 (𝑧, 𝜁) + 𝑄 (𝜁; 𝑧 − 𝑧 (𝜁)) (20.3.17)
2
3
+ 𝑂 |𝑧 − 𝑧 (𝜁)| .

Combining (20.3.5) and (20.3.17) yields

1
Ψ∗ (𝑧, 𝜁) = Φ (𝑧) − 𝑄 (𝜁; 𝑧 − 𝑧 (𝜁)) + 𝑂 |𝑧 − 𝑧 (𝜁)| 3
2
and therefore
1
Ψ∗ (𝑧 ◦ , 𝜁) = − 𝑄 (𝜁; 𝑧 ◦ − 𝑧 (𝜁)) + 𝑂 |𝜁 ◦ − 𝜁 | 3 (20.3.18)
2
since Φ (𝑧◦ ) = 0 and |𝑧 ◦ − 𝑧 (𝜁)| ≤ 𝐶 |𝜁 − 𝜁 ◦ |.
We assume that the transformation 𝑇 in Proposition 20.3.8 has been carried out:
Formula (20.3.10) reads
𝑛
∑︁ 2
𝑄 (𝜁; 𝑧 − 𝑧 (𝜁)) = 𝜒 ′𝑗 𝑦 𝑗 − 𝑦 𝑗 (𝜁) (20.3.19)
𝑗=1
𝑛
∑︁
2
− 𝜒 𝑗 𝑥 𝑗 − 𝑥 𝑗 (𝜁) + 𝑂 |𝜁 − 𝜁 ◦ | 3 .
𝑗=1

By (20.3.17) we have

1 𝜕𝑄 𝜕Φ 𝜕
(𝜁; 𝑧 − 𝑧 (𝜁)) = −𝜒 𝑗 𝑥 𝑗 − 𝑥 𝑗 (𝜁) (𝑧) − Im 𝜑 (𝑧, 𝜁) ,
2 𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑗
1 𝜕𝑄 𝜕Φ 𝜕
(𝜁; 𝑧 − 𝑧 (𝜁)) = 2𝜒 ′𝑗 𝑦 𝑗 − 𝑦 𝑗 (𝜁)

(𝑧) − Im 𝜑 (𝑧, 𝜁) ,
2 𝜕𝑦 𝑗 𝜕𝑦 𝑗 𝜕𝑦 𝑗

where stands for congruent mod 𝑂 (|𝑧 − 𝑧 (𝜁)| |𝜁 − 𝜁 ◦ |). We obtain

20.3 Sjöstrand Pairs 849

1 𝜕
𝜕𝜁 𝑥 𝑗 (𝜁) − 𝜕𝜁 Im 𝜑 (𝑧, 𝜁) ,
𝜒 𝑗 𝜕𝑥 𝑗
1 𝜕
𝜕𝜁 𝑦 𝑗 (𝜁) ′ 𝜕𝜁 Im 𝜑 (𝑧, 𝜁) .
𝜒 𝑗 𝜕𝑦 𝑗

Since 2𝑖𝜕𝑤 Im 𝜑 = 𝜕𝑤 𝜑 we get

1 1 𝜕 1 1 𝜕
𝜕𝜁 𝑥 𝑗 = − 𝜕𝜁 𝜑 = − 𝜕𝜁 𝜑,
2𝑖 𝜒 𝑗 𝜕𝑥 𝑗 2𝑖 𝜒 𝑗 𝜕𝑧 𝑗
1 1 𝜕 1 1 𝜕
𝜕𝜁 𝑦 𝑗 = ′ 𝜕𝜁 𝜑 = 𝜕𝜁 𝜑,
2𝑖 𝜒 𝑗 𝜕𝑦 𝑗 2 𝜒 ′𝑗 𝜕𝑧 𝑗

at (𝑧◦ , 𝜁 ◦ ). We derive

1 𝜕2
𝑄 (𝜁; 𝑧 ◦ − 𝑧 (𝜁))
2 𝜕𝜁 𝑝 𝜕𝜁 𝑞
𝜁 =𝜁 ◦
𝑛
!
∑︁ 𝜕𝑦 𝑗 𝜕𝑦 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑗
= 𝜒 ′𝑗 − 𝜒𝑗 (𝑧 ◦ , 𝜁 ◦ )
𝑗=1
𝜕𝜁 𝑝 𝜕𝜁 𝑞 𝜕𝜁 𝑝 𝜕𝜁 𝑞

𝑛
!
1 ∑︁ 1 1 𝜕2 𝜑 𝜕2 𝜑
= ′ − (𝑧◦ , 𝜁 ◦ ) (𝑧◦ , 𝜁 ◦ ) .
4 𝑗=1 𝜒 𝑗 𝜒 𝑗 𝜕𝜁 𝑝 𝜕𝑧 𝑗 𝜕𝜁𝑞 𝜕𝑧 𝑗

Using the notation, for arbitrary 𝑤 ∈ C𝑛 ,

𝑛
∑︁ 𝜕𝑓
𝑤 · 𝜕𝜁 𝑓 = 𝑤𝑝 , 𝑓 ∈ C ∞ (Ξ) ,
𝑝=1
𝜕𝜁 𝑝

we derive from (20.3.18)

𝜕 2 Ψ∗
𝑛
∑︁
𝑤 𝑝 𝑤¯ 𝑞 (𝑧 ◦ , 𝜁 ◦ )
𝑝,𝑞=1 𝜕𝜁 𝑝 𝜕𝜁 𝑞
𝑛
1 ∑︁ 𝜕2
=− 𝑤 𝑝 𝑤¯ 𝑞𝑞 𝑄 (𝜁 ◦ ; 𝑧◦ − 𝑧 (𝜁))
2 𝑝,𝑞=1 𝜕𝜁 𝑝 𝜕𝜁 𝑞 𝜁 =𝜁 ◦
!
𝜕𝜑 ◦ ◦ 2
𝑛
1 ∑︁ 1 1
= − ′ 𝑤 · 𝜕𝜁 (𝑧 , 𝜁 ) .
4 𝑗=1 𝜒 𝑗 𝜒𝑗 𝜕𝑧 𝑗
850 20 Germ Fourier Integral Operators in Complex Space

Likewise,

1 𝜕2
𝑄 (𝜁; 𝑧 ◦ − 𝑧 (𝜁))
2 𝜕𝜁 𝑝 𝜕𝜁𝑞 𝜁 =𝜁 ◦
𝑛
∑︁ 𝜕𝑦 𝑗 𝜕𝑦 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑗
= 𝜒 ′𝑗 − 𝜒𝑗 (𝑧 ◦ , 𝜁 ◦ )
𝑗=1
𝜕𝜁 𝑝 𝜕𝜁 𝑞 𝜕𝜁 𝑝 𝜕𝜁 𝑞

𝑛
!
1 ∑︁ 1 1 𝜕2 𝜑 𝜕2 𝜑
= ′ + (𝑧 ◦ , 𝜁 ◦ ) (𝑧◦ , 𝜁 ◦ ) ,
4 𝑗=1 𝜒 𝑗 𝜒 𝑗 𝜕𝜁 𝑝 𝜕𝑧 𝑗 𝜕𝜁𝑞 𝜕𝑧 𝑗

whence

𝜕 2 Ψ∗
𝑛
∑︁
𝑤 𝑝 𝑤𝑞 (𝑧 ◦ , 𝜁 ◦ )
𝑝,𝑞=1
𝜕𝜁 𝑝 𝜕𝜁𝑞
𝑛
1 ∑︁ 𝜕2
=− 𝑤 𝑝 𝑤𝑞 𝑄 (𝜁 ◦ ; 𝑧◦ − 𝑧 (𝜁))
2
𝑝,𝑞=1
𝜕𝜁 𝑝 𝜕𝜁 𝑞 𝜁 =𝜁 ◦
𝑛
! 2
1 ∑︁ 1 1 𝜕𝜑
=− + 𝑤 · 𝜕𝜁 (𝑧◦ , 𝜁 ◦ ) .
4 𝑗=1 𝜒 𝑗 𝜒 ′𝑗 𝜕𝑥 𝑗

We reach the conclusion that

𝜕 2 Ψ∗ 𝜕 2 Ψ∗
𝑛
∑︁ 𝑛
∑︁
𝑤 𝑝 𝑤¯ 𝑞 (𝑧◦ , 𝜁 ◦ ) + Re 𝑤 𝑝 𝑤𝑞 (𝑧 ◦ , 𝜁 ◦ )
𝑝,𝑞=1 𝜕𝜁 𝑝 𝜕𝜁 𝑞 𝑝,𝑞=1
𝜕𝜁 𝑝 𝜕𝜁 𝑞
𝑛 2
1 ∑︁ 1 𝜕𝜑 ◦ ◦
=− Re 𝑤 · 𝜕 (𝑧 , 𝜁 )
2 𝑗=1 𝜒 ′𝑗
𝜁
𝜕𝑥 𝑗
𝑛 2
1 ∑︁ 1 𝜕𝜑 ◦ ◦
+ Im 𝑤 · 𝜕𝜁 (𝑧 , 𝜁 ) .
2 𝑗=1 𝜒 𝑗 𝜕𝑥 𝑗

The nondegeneracy of 𝜑, (20.3.11), implies that the linear map

𝜕𝜑 ◦ ◦ 𝜕𝜑 ◦ ◦
𝑤 ↦→ 𝑆𝑤 = 𝑤 · 𝜕𝜁 (𝑧 , 𝜁 ) , ..., 𝑤 · 𝜕𝜁 (𝑧 , 𝜁 ) (20.3.20)
𝜕𝑧 1 𝜕𝑧 𝑛

is a linear automorphism of C𝑛 . Putting 𝑤 = 𝜁 − 𝜁 (𝑧) proves

𝑄 ∗ (𝑧◦ ; 𝑆 (𝜁 − 𝜁 (𝑧))) (20.3.21)

𝑛 𝑛
∑︁ 1 2 ∑︁ 1 2
= 𝜂 𝑗 − 𝜂 𝑗 (𝑧) − ′ 𝜉 𝑗 − 𝜉 𝑗 (𝑧)
𝜒
𝑗=1 𝑗
𝜒
𝑗=1 𝑗

+ 𝑂 |𝜁 − 𝜁 ◦ | 3 .
20.3 Sjöstrand Pairs 851

From this point on the proof is the same as that of Proposition 20.3.8 in which 𝑧 and
𝜁 are exchanged. □
Exercise 20.3.15 Conjecture what ought to be the relation between the linear trans-
formation 𝑇 in Proposition 20.3.8 and the transformation 𝑆 in Proposition 20.3.14,
then prove the conjecture.
Theorem 20.3.16 If (𝜑, Φ) is a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) in Ω × Ξ then (−𝜑, Φ∗ )
with Φ∗ given by (20.3.5) is a Sjöstrand pair at (𝜁 ◦ , 𝑧◦ ) in Ξ′ × Ω′ (Ω′ ⊂ Ω a
neighborhood of 𝑧◦ , Ξ′ ⊂ Ξ a neighborhood of 𝜁 ◦ , both appropriately small).
Proof We must prove that 𝜁 ◦ is a saddle point of the function Ψ∗ (𝑧◦ , 𝜁); it suffices to
prove that this is true of the function 𝜁 ↦→ 𝑄 ∗ (𝑧 ◦ ; 𝜁 − 𝜁 (𝑧)), which is an immediate
consequence of (20.3.21) since 𝜒 𝑗 > 0, 𝜒 ′𝑗 > 0, 𝑗 = 1, ..., 𝑛. □
Theorem 20.3.16 can be supplemented as follows:
Theorem 20.3.17 If (𝜑, Φ) is a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) in Ω × Ξ and if Φ is strictly
plurisubharmonic at 𝑧◦ then Φ∗ is a strictly plurisubharmonic function in Ξ at 𝜁 ◦ .
Proof Here also it suffices to prove the claim with 𝑄 ∗ (𝑧◦ ; 𝜁 − 𝜁 (𝑧)) replacing
Ψ∗ (𝑧◦ , 𝜁). It is then an immediate consequence of (20.3.21) since the hypothesis is
that 0 < 𝜒 𝑗 < 𝜒 ′𝑗 , 𝑗 = 1, ..., 𝑛 (cf. Corollary 20.3.9). □
Definition 20.3.18 Let (𝜑, Φ) be a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) in Ω × Ξ. We shall refer
to (−𝜑, Φ∗ ) as the dual of the Sjöstrand pair (𝜑, Φ).
Because of (20.3.14) the map (𝜑, Φ) ↦→ (−𝜑, Φ∗ ) is an involution.
Remark 20.3.19 Let Hess 𝜉 , 𝜂 Ψ∗ (𝑧, 𝜁) denote the Hessian of the function 𝜁 = 𝜉 +
𝑖𝜂 ↦→ Ψ∗ (𝑧, 𝜁). Formula (20.3.16) implies that the eigenvalues of Hess 𝜉 , 𝜂 Ψ (𝑧◦ , 𝜁 ◦ )
are −𝜒1′−1 , ..., −𝜒𝑛′−1 , 𝜒1−1 , ..., 𝜒𝑛−1 . In particular, 𝑄 ∗ (𝑧 ◦ ; 𝜁) is pluriharmonic if and
only if the same is true of 𝑄 (𝜁 ◦ ; 𝑧). Referring the reader to Remark 20.3.1 and
introducing the (2𝑛) × (2𝑛) fundamental symplectic matrix (Definition 13.1.17)

0 −𝐼𝑛
𭟋2𝑛 = ,
𝐼𝑛 0

we see that

Hess 𝜉 , 𝜂 Ψ∗ (𝑧 ◦ , 𝜁 ◦ ) = 𝑇2 𭟋2𝑛𝑇1−1 Hess 𝑥,𝑦 Ψ (𝑧 ◦ , 𝜁 ◦ ) 𝑇1 𭟋2𝑛𝑇2−1 , (20.3.22)

where 𝑇1 , 𝑇2 ∈ O (2𝑛, R).

20.3.2.2 The cases 𝑵 > 𝒏

When 𝑁 > 𝑛 we can attach a dual Sjöstrand pair to (𝜑, Φ) to each 𝑛-dimensional
affine complex subspace of C 𝑁 , 𝑬 ∋ 𝜃 ◦ , such that the restriction of 𝜕𝑧 𝜕𝜃 𝜑 (𝑧 ◦ , 𝜃 ◦ )
to 𝑬 is invertible. By Lemma 20.3.3 the surjection 𝜃 ↦→ 𝑧 (𝜃) in Corollary 20.3.5
induces a diffeomorphism of a neighborhood of 𝜃 ◦ in 𝑬 onto a neighborhood of 𝑧◦ .
Replacing Θ by 𝑬 ∩ Θ brings us back to the case 𝑁 = 𝑛.
852 20 Germ Fourier Integral Operators in Complex Space

20.3.2.3 The cases 𝑵 < 𝒏

When 𝑁 < 𝑛 we can augment the number of variables 𝜃 𝑗 by substituting

e (𝑧, 𝜃, 𝜃 ′′) = 𝜑 (𝑧, 𝜃) + (𝑧 ′′ − 𝑧 ◦′′) · 𝜃 ′′ for 𝜑 (𝑧, 𝜃) where 𝑧 ′′ = (𝑧 𝑁 +1 , ..., 𝑧 𝑛 )
𝜑
and 𝜃 ′′ ∈ C𝑛−𝑁 . This allows us to define a dual Sjöstrand pair to ( 𝜑 e, Φ) at
(𝑧◦ , (𝜃 ◦ , 0)) ∈ Ω × Θ × C𝑛−𝑁 .

20.3.3 The case of the canonical phase-function

In this subsection we take the phase-function to be the canonical phase-function

at (𝑧 ◦ , 𝜁 ◦ ), by which we mean the phase-function − (𝑧 − 𝑧 ◦ ) · (𝜁 − 𝜁 ◦ ), not to be
confused with the phase-function of the identity (Example 18.2.14). Here Φ (𝑧) will
be a real-valued C ∞ function in Ω ⊂ C𝑛 such that Φ (𝑧◦ ) = 0 and [cf. (20.3.1)]

Ψ (𝑧, 𝜁) = Φ (𝑧) + Im ((𝑧 − 𝑧 ◦ ) · (𝜁 − 𝜁 ◦ )) . (20.3.23)

The critical points of the function 𝑧 ↦→ Ψ (𝑧, 𝜁) are the solutions of the equation

𝜁 = 𝜁 ◦ − 2𝑖𝜕𝑧 Φ (𝑧) (20.3.24)

[cf. (20.3.3)]. An immediate consequence of (20.3.23) and (20.3.24) is the following

Proposition 20.3.20 Let 𝑧◦ ∈ Ω and Φ ∈ C ∞ (Ω; R) be such that Φ (𝑧◦ ) = 0. For

𝜑 (𝑧, 𝜁) = − (𝑧 − 𝑧 ◦ ) · (𝜁 − 𝜁 ◦ ) and Φ to form a Sjöstrand pair at (𝑧 ◦ , 𝜁 ◦ ) (Definition
20.3.6) it is necessary and sufficient that 𝑧 ◦ be a saddle point of Φ (𝑧).

Example 20.3.21 If 𝜁 ◦ ∈ C𝑛 , 𝑎 > 0, 𝑏 > 0 then 𝜑 (𝑧, 𝜁) = −𝑧 · (𝜁 − 𝜁 ◦ ) and

Φ (𝑧) = 21 𝑎 |𝑥| 2 − 12 𝑏 |𝑦| 2 form a Sjöstrand pair at (0, 𝜉 ◦ ). Indeed,

1 1
Ψ (𝑧, 𝜁) = 𝑎 |𝑥| 2 − 𝑏 |𝑦| 2 + Im (𝑧 · (𝜁 − 𝜁 ◦ ))
2 2
1 1
= |𝜉 − 𝜉 | − |𝜂 − 𝜂◦ | 2
◦ 2
𝑏 𝑎
1 2 1
+ 𝑎 𝑥 + 𝑎 (𝜂 − 𝜂◦ ) − 𝑏 𝑦 + 𝑏 −1 (𝜉 − 𝜉 ◦ ) 2 .
−1
2 2
The following direct consequence of Corollary 11.2.27 is noteworthy.

Proposition 20.3.22 Suppose 𝜑 (𝑧, 𝜁) = − (𝑧 − 𝑧◦ ) · (𝜁 − 𝜁 ◦ ) and Φ ∈ C ∞ (Ω; R)

such that Φ (0) = 0 form a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) and Φ is plurisubharmonic in
Ω. If Φ1 ∈ C ∞ (Ω; R) is plurisubharmonic, Φ1 ≤ Φ and Φ1 (𝑧 ◦ ) = 0 then 𝑧◦ is also
a saddle point of Φ1 .

In the remainder of this subsection we assume that Φ (𝑧◦ ) = 0 and that 𝑧◦ is a

saddle point of Φ (𝑧); in particular, 𝑧◦ is an isolated critical point of Φ. Possibly after
a contraction of Ω about 𝑧◦ we can assume (Proposition 20.3.8) that Ω ∋ 𝑧 ↦→ 𝜁 =
20.3 Sjöstrand Pairs 853

𝜁 ◦ −2𝑖𝜕𝑧 Φ (𝑧) is a diffeomorphism; we define Ξ as its range; 𝜁 ◦ ∈ Ξ. The graph ΛΦ of

this map is a C ∞ submanifold of Ω × Ξ containing (𝑧◦ , 𝜁 ◦ ); the coordinate projection
(𝑧, 𝜁) ↦→ 𝜁 is a bijection of ΛΦ onto Ξ. Since det 𝜕𝑧2 Φ (𝑧) ≠ 0 in a neighborhood 𝑈
of 𝑧◦ there is a unique C ∞ solution 𝑧 (𝜁) in 𝑈 of (20.3.24) such that 𝑧 (𝜁 ◦ ) = 𝑧◦ . We
have [cf. (20.3.5)]

Φ∗ (𝜁) = Ψ (𝑧 (𝜁) , 𝜁) = Φ (𝑧 (𝜁)) + Im (𝑧 (𝜁) · (𝜁 − 𝜁 ◦ )) . (20.3.25)

We deduce

𝜕𝜁 Φ∗ (𝜁) = (𝜕𝑧 Φ) (𝑧 (𝜁)) · 𝜕𝜁 𝑧 (𝜁) + (𝜕𝑧¯ Φ) (𝑧 (𝜁)) · 𝜕𝜁 𝑧¯ (𝜁)

1 1 1
+ 𝜕𝜁 𝑧 (𝜁) · (𝜁 − 𝜁 ◦ ) − 𝜕𝜁 𝑧 (𝜁) · 𝜁 − 𝜁¯◦ + 𝑧 (𝜁) .
2𝑖 2𝑖 2𝑖
From (20.3.24) we derive
1 1
(𝜕𝑧 Φ) (𝑧 (𝜁)) + (𝜁 − 𝜁 ◦ ) = (𝜕𝑧¯ Φ) (𝑧 (𝜁)) − 𝜁 − 𝜁¯◦ = 0,
2𝑖 2𝑖
whence
𝑧 (𝜁) = 2𝑖𝜕𝜁 Φ∗ (𝜁) . (20.3.26)
Thus ΛΦ can also be regarded as the graph of the map 𝜁 ↦→ 2𝑖𝜕𝜁 Φ∗ (𝜁). Here
(20.3.14) reads
Φ (𝑧) = Φ∗ (𝜁 (𝑧)) − Im (𝑧 · (𝜁 (𝑧) − 𝜁 ◦ )) . (20.3.27)
Recall that 𝜁 (𝑧) = 𝜁 ◦ − 2𝑖𝜕𝑧 Φ (𝑧).

Theorem 20.3.23 If − (𝑧 − 𝑧 ◦ ) · (𝜁 − 𝜁 ◦ ) and Φ (𝑧) form a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ )

then (𝑧 − 𝑧◦ ) · (𝜁 − 𝜁 ◦ ) and Φ∗ (𝜁) form a Sjöstrand pair at (𝜁 ◦ , 𝑧◦ ).

Theorem 20.3.23 is a special case of Theorem 20.3.16. We shall nevertheless

give a proof of it as the linear algebra argument makes even the general case more
transparent.
Proof To lighten the notation we take 𝑧 ◦ = 0. If
1
𝑄 (𝑧) + 𝑂 |𝑧| 3 ,
Ψ (𝑧, 𝜁 (𝑧)) = Φ (𝑧) − 2 Re (𝑧 · 𝜕𝑧 Φ (𝑧)) =
2
1 ∗
∗ ∗
Ψ (𝑧 (𝜁) , 𝜁) = Φ (𝜁) − Im (𝑧 (𝜁) · 𝜁) = 𝑄 (𝜁) + 𝑂 |𝜁 | 3 ,
2
with 𝑄 and 𝑄 ∗ real quadratic forms in C𝑛 and if (20.3.10) holds we can write
−1
Ψ 𝑇 𝑧, 𝑇 ⊤ 𝜁 = Φ (𝑇 𝑧) + Im (𝑧 · (𝜁 − 𝜁 ◦ ))
𝑛 𝑛
1 ∑︁ 1 ∑︁ ′ 2
=− 𝜒 𝑗 𝑥 2𝑗 + 𝜒 𝑗 𝑦 𝑗 + Im (𝑧 · (𝜁 − 𝜁 ◦ )) + 𝑂 |𝑧| 3 .
2 𝑗=1 2 𝑗=1
854 20 Germ Fourier Integral Operators in Complex Space

Then (20.3.24) reads

𝜁 𝑗 − 𝜁 ◦𝑗 = −𝜒 ′𝑗 𝑦 𝑗 + 𝑖 𝜒 𝑗 𝑥 𝑗 + 𝑂 |𝑧| 2 , 𝑗 = 1, ..., 𝑛,

and therefore the critical points 𝑧 (𝜁) of 𝑧 ↦→ Ψ 𝑇 𝑧, (𝑇 ⊤ ) −1 𝜁 are given by

1
𝑥𝑗 = 𝜂 𝑗 − 𝜂◦𝑗 + 𝑂 |𝜁 − 𝜁 ◦ | 2 ,
𝜒𝑗
1
𝑦 𝑗 = − ′ 𝜉 𝑗 − 𝜉 ◦𝑗 + 𝑂 |𝜁 − 𝜁 ◦ | 2 .
𝜒𝑗

A simple calculation shows that

𝑛
!
1 ∑︁ 1 2 1 2
⊤ −1 ◦ ◦
Ψ 𝑇 𝑧 (𝜁) , 𝑇 𝜁 =− 𝜉𝑗 − 𝜉𝑗 − 𝜂𝑗 − 𝜂𝑗 (20.3.28)
2 𝑗=1 𝜒 ′𝑗 𝜒𝑗

+ 𝑂 |𝜁 − 𝜁 ◦ | 3 ,

in agreement with Proposition 20.3.14 where we can take 𝑆 = (𝑇 ⊤ ) −1 , the contra-

gredient of 𝑇. Formula (20.3.22) reads here:

Hess 𝜉 , 𝜂 Ψ∗ (𝑧◦ , 𝜁 ◦ ) = 𝑇𭟋2𝑛𝑇 ⊤ Hess 𝑥,𝑦 Ψ (𝑧 ◦ , 𝜁 ◦ ) 𝑇𭟋2𝑛𝑇 ⊤ . □

Proposition 20.3.24 Let 𝑧◦ = 0. Possibly after a C-linear change of the coordinates

𝑧 𝑗 = 𝑥 𝑗 + 𝑖𝑦 𝑗 , 𝑗 = 1, ..., 𝑛, there are sufficiently small numbers 𝑟 > 0 and 𝑟 ′ > 0 such
that
Φ∗ (𝜁) = min ′ max Ψ (𝑧, 𝜁) . (20.3.29)
|𝑦 | ≤𝑟 | 𝑥 | ≤𝑟

It is tacitly assumed that 𝔅𝑟(𝑛) × 𝔅𝑟(𝑛) ′ = {𝑥 + 𝑖𝑦 ∈ C𝑛 ; |𝑥| ≤ 𝑟, |𝑦| ≤ 𝑟 ′ } ⊂⊂ Ω;

Ψ is given by (20.3.23).

Proof We may suppose that Φ (𝑧) − 2 Re (𝑧 · 𝜕𝑧 Φ (𝑧)) = 21 𝑄 (𝑧) + 𝑂 |𝑧| 3 with
Í
𝑄 (𝑧) = 𝑛𝑗=1 𝜒 ′𝑗 𝑦 2𝑗 − 𝜒 𝑗 𝑥 2𝑗 . Then, if 𝑟, 𝑟 ′ and the neighborhood Ξ′ ⊂ Ξ of 𝜁 ◦ are
sufficiently small, for any given (𝑦, 𝜁) ∈ 𝔅𝑟(𝑛)
′ × Ξ′ the critical point 𝑥 (𝑦, 𝜁) of the
function
𝑛
1 ∑︁ ′ 2
𝔅𝑟(𝑛) ∋ 𝑥 → 𝜒 𝑗 𝑦 𝑗 − 𝜒 𝑗 𝑥 2𝑗
2 𝑗=1
𝑛
∑︁ 1
+ 𝑥𝑗 𝜂𝑗 − 𝜂◦𝑗 + 𝑦𝑗 𝜉𝑗 − 𝜉 ◦𝑗 + Φ − 𝑄 (𝑧)
𝑗=1
2
20.4 Germ Fourier-like Transforms 855

is its maximum; 𝑥 (𝑦, 𝜁) ∈ 𝔅𝑟(𝑛) is the (smooth) solution in 𝔅𝑟(𝑛) ′

′ × Ξ of a system of
equations
1
𝑥𝑗 = 𝜂 𝑗 − 𝜂◦𝑗 + 𝑂 𝑧2 , 𝑗 = 1, ..., 𝑛,
𝜒𝑗
such that 𝑥 (0, 𝜁 ◦ ) = 0. The critical point of the function
𝑛
1 ∑︁ ′ 2
𝑦→ 𝜒 𝑗 𝑦 𝑗 − 𝜒 𝑗 𝑥 2𝑗 (𝑦, 𝜁) +
2 𝑗=1
𝑛
∑︁
𝑥 𝑗 (𝑦, 𝜁) 𝜂 𝑗 − 𝜂◦𝑗 + 𝑦 𝑗 𝜉 𝑗 − 𝜉 ◦𝑗 + (Φ − 𝑄) (𝑥 (𝑦, 𝜁) + 𝑖𝑦)
𝑗=1

is its minimum; it is the (smooth) solution 𝑦 (𝜁) ∈ 𝔅𝑟(𝑛) ′

′ in Ξ of a system of equations

1
𝑦𝑗 = ′ 𝜉 𝑗 − 𝜉 ◦𝑗
𝜒𝑗
𝑛
1 ∑︁ 𝜕𝑥
𝜒𝑘 𝑥 𝑘 (𝑦, 𝜁) − 𝜂 𝑗 − 𝜂◦𝑗 (𝑦, 𝜁) + 𝑂 |𝑦| 2 + |𝜁 − 𝜁 ◦ | 2
𝑘
− ′
𝜒 𝑗 𝑘=1 𝜕𝑦 𝑗
1
= ′ 𝜉 𝑗 − 𝜉 ◦𝑗 + 𝑂 |𝑦| 2 + |𝜁 − 𝜁 ◦ | 2
𝜒𝑗

such that 𝑦 (𝜁 ◦ ) = 0. □
The equation (20.3.29) shows the kinship between the transform Φ∗ and the Leg-
endre transform of a convex function (see, e.g., [Arnold, 1989]). Actually, analogous
results are valid for a general Sjöstrand pair (𝜑, Φ) when 𝑁 ≠ 𝑛.

20.4 Germ Fourier-like Transforms

20.4.1 Deforming the contours of integration

Let Ω and Θ be domains in C𝑛 and C 𝑁 respectively; the analysis will take place in
a neighborhood of (𝑧◦ , 𝜃 ◦ ) ∈ Ω × Θ. The purpose of this section is to define a germ
transform starting from the integral (20.1) in which:
(1) 𝜑 ∈ O (Ω × Θ) such that Im 𝜑 (𝑧 ◦ , 𝜃 ◦ ) = 0,
(2) 𝑎 (𝑧, 𝜃, 𝜆) ∈ 𝑆a (Ω × Θ) (Definition 20.1.1),
(3) ℎ ∈ O (Φ) (Ω),
(4) the contour 𝔠 ⊂⊂ Ω is strongly well-shaped (Definition 20.2.2) at 𝑧◦ for the
function 𝑧 ↦→ Ψ (𝑧, 𝜃 ◦ ).
856 20 Germ Fourier Integral Operators in Complex Space

In (4) Ψ (𝑧, 𝜃) = Φ (𝑧) − Im 𝜑 (𝑧, 𝜃) [cf. (20.3.1)] with Φ ∈ C ∞ (Ω; R) an

exponent-weight such that (𝜑, Φ) is a Sjöstrand pair at (𝑧◦ , 𝜃 ◦ ) ∈ Ω × Θ (Defi-
nition 20.3.6).
The following lemma will play an important role in the forthcoming construction.

Lemma 20.4.1 If the domains Ω and Θ are sufficiently small then, to each 𝜃 ∈ Θ
there exist a unique saddle point 𝑧 (𝜃) of the function Ω ∋ 𝑧 ↦→ Ψ (𝑧, 𝜃) and a
contour of integration 𝔠 (𝜃) ⊂⊂ Ω with the following properties:
(1) 𝔠 (𝜃) depends smoothly on 𝜃 and 𝔠 (𝜃 ◦ ) = 𝔠;
(2) the saddle point 𝑧 (𝜃) of the function Ω ∋ 𝑧 ↦→ Ψ (𝑧, 𝜃) belongs to 𝔠 (𝜃);
(3) 𝔠 (𝜃) is strongly well-shaped for the function 𝑧 ↦→ Ψ (𝑧, 𝜃) at 𝑧 (𝜃);
(4) the orientation of 𝔠 (𝜃) is independent of 𝜃.

Proof We apply Proposition 20.3.8: after a linear automorphism of C𝑛 we can

assume that
1
Ψ (𝑧, 𝜃) = Ψ (𝑧 (𝜃) , 𝜃) + 𝑄 (𝜃; 𝑧 − 𝑧 (𝜃)) + 𝑂 |𝑧 − 𝑧◦ | 3 + |𝜃 − 𝜃 ◦ | 3
2
where
𝑛
∑︁ 𝑛
∑︁
2 2
𝑄 (𝜃; 𝑧 − 𝑧 (𝜃)) = − 𝜒 𝑗 𝑥 𝑗 − 𝑥 𝑗 (𝜃) + 𝜒 ′𝑗 𝑦 𝑗 − 𝑦 𝑗 (𝜃)
𝑗=1 𝑗=1

with 0 < 𝜒 𝑗 ≤ 𝜒 ′𝑗 , 𝑗 = 1, ..., 𝑛. In conformity with Definition 20.2.2 and after

contraction of Ω about 𝑧◦ the new coordinates 𝑧1 , ..., 𝑧 𝑛 can be chosen so that 𝔠 is
defined by an equation 𝑦 − 𝑦 ◦ = 𝑓 (𝑥 −√𝑥 ◦ ) with 𝑓 ∈ C ∞ (𝑈; R𝑛 ), 𝑈 a neighborhood
of 0 such that 𝑈 + 𝑧◦ ⊂⊂ Ω, | 𝑓 (𝑥)| ≤ 𝛿 |𝑥| for some 𝛿 ≥ 0, 𝛿 < min ( 𝜒 𝑗 /𝜒 ′𝑗 ), and
1≤ 𝑗 ≤𝑛
all 𝑥 ∈ 𝑈. If 𝑈 and Θ are small enough the contours 𝔠 (𝜃) defined by the equations
𝑦 − 𝑦 (𝜃) = 𝑓 (𝑥 − 𝑥 (𝜃)) satisfy the requirements of the lemma. □
We assume that, as 𝜃 varies in Θ, the contours 𝔠 (𝜃) in Lemma 20.4.1 are all
contained in an open set Ω′ ⊂⊂ Ω. We shall also assume that the boundaries of 𝔠 (𝜃)
are smooth and depend smoothly on 𝜃. The function
∫
𝑨𝔠 ( 𝜃) ℎ (𝜃, 𝜆) = e𝑖 𝜑 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃, 𝜆) ℎ (𝑧, 𝜆) d𝑧 (20.4.1)
𝔠 ( 𝜃)

is of class C∞ with respect to 𝜃 ∈ Θ. Keep in mind that 𝔠 (𝜃 ◦ ) = 𝔠.

Proposition 20.4.2 Let Φ∗ be as in (20.3.5). If Θ is sufficiently small then to every

𝜀 > 0 and every ℎ ∈ O (Φ) (Ω) there is a 𝐶 𝜀 > 0 such that
∗ ( 𝜃)+𝜀)
∀𝜃 ∈ Θ, 𝜆 ≥ 1, 𝑨𝔠 ( 𝜃) ℎ (𝜃, 𝜆) ≤ 𝐶 𝜀 e𝜆(Φ . (20.4.2)

Proof By (AS), Definition 20.1.1, combined with (20.3.1) and (20.3.10) we know
that to every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that
20.4 Germ Fourier-like Transforms 857
∫
∗ ( 𝜃)
e−𝜆Φ 𝑨𝔠 ( 𝜃) ℎ (𝜃, 𝜆) ≤ 𝐶 𝜀 e 𝜀𝜆 e𝜆(Ψ(𝑧, 𝜃)−Ψ(𝑧 ( 𝜃), 𝜃)) e−𝜆Φ(𝑧) |ℎ (𝑧, 𝜆) d𝑧| .
𝔠 ( 𝜃)

By applying Lemma 20.4.1, Property (3), we obtain directly

∫
∗
e−𝜆Φ ( 𝜃) 𝑨𝔠 ( 𝜃) ℎ (𝜃, 𝜆) ≤ 𝐶 𝜀 e 𝜀𝜆 e−𝜆Φ(𝑧) |ℎ (𝑧, 𝜆) d𝑧|
𝔠 ( 𝜃)

for all 𝜃 ∈ Θ. □
Proposition 20.4.3 If diam Θ and 𝜅 > 0 are sufficiently small then

∀𝜃 ∈ Θ, 𝜆 ≥ 1, 𝑨𝔠 ( 𝜃) ℎ (𝜃, 𝜆) − 𝑨𝔠 ℎ (𝜃, 𝜆) ≲ e−𝜅𝜆 (20.4.3)

for each ℎ ∈ O (Φ) (Ω).

Proof Let [0, 1] ∋ ◦ 1
1 1
𝑠 → 𝜃 (𝑠) ∈ Θ be a smooth curve joining 𝜃 = 𝜃 (0) to 𝜃 = 𝜃 (1)
and set 𝔠 = 𝔠 𝜃 . By Stokes’ Theorem we have
∫ 1∫
𝑨𝔠1 ℎ (𝜃, 𝜆) − 𝑨𝔠 ℎ (𝜃, 𝜆) = e𝑖𝜆𝜑 (𝑧, 𝜃 (𝑠)) ℎ (𝑧, 𝜆) d𝜇 𝑠 (𝑧) d𝑠,
0 𝜕𝔠 ( 𝜃 (𝑠))

where d𝜇 𝑠 (𝑧) is the appropriate measure on the boundary 𝜕𝔠 (𝜃 (𝑠)). Suppose

∀𝑧 ∈ Ω, ∀𝜆 ≥ 1, |ℎ (𝑧, 𝜆)| ≤ 𝐶 𝜀 e𝜆(Φ(𝑧)+𝜀) .

We get, thanks to Property (3), Lemma 20.4.1,

𝑨𝔠1 ℎ (𝜃, 𝜆) − 𝑨𝔠 ℎ (𝜃, 𝜆)

∫ 1∫
∗
≤ 𝐶𝜀e 𝜀𝜆
e𝜆Φ ( 𝜃 (𝑠)) e𝜆(Ψ(𝑧, 𝜃)−Ψ(𝑧 ( 𝜃 (𝑠)) , 𝜃 (𝑠))) |d𝜇 𝑠 (𝑧)| d𝑠
0 𝜕𝔠 ( 𝜃 (𝑠))
∫ 1∫ 2
∗ ( 𝜃 (𝑠))
≤ 𝐶 𝜀 e 𝜀𝜆 e𝜆Φ e−𝜅◦ 𝜆 |𝑧−𝑧 ( 𝜃 (𝑠)) | |d𝜇 𝑠 (𝑧)| d𝑠.
0 𝜕𝔠 ( 𝜃 (𝑠))

After contracting Θ about 𝜃 ◦ we may assume that

min dist (𝑧 (𝜃) , 𝜕𝔠 (𝜃)) = 𝛿 > 0,

𝜃 ∈Θ

whence
∗
e−𝜆Φ ( 𝜃) 𝑨𝔠1 ℎ (𝜃, 𝜆) − 𝑨𝔠 ℎ (𝜃, 𝜆)
∫ 1∫
∗ 2
≤ 𝐶 𝜀 e 𝜀𝜆 e𝜆Φ ( 𝜃 (𝑠))−𝜅◦ 𝜆 𝛿 |d𝜇 𝑠 (𝑧)| d𝑠.
0 𝜕𝔠 ( 𝜃 (𝑠))

Further contracting Θ about 𝜃 ◦ allows us to take 𝜀 > 0 so small that

1
∀𝜃 ∈ Θ, |Φ∗ (𝜃)| + 𝜀 < 𝜅 ◦ 𝜆𝛿2 ,
2
858 20 Germ Fourier Integral Operators in Complex Space

whence (20.4.3). □
∗
Corollary 20.4.4 If 𝔠 and Θ are sufficiently small then 𝑨𝔠 ℎ ∈ O (Φ ) (Θ) whatever
ℎ ∈ O (Φ) (Ω).
Proof Combine (20.4.2) with (20.4.3). □
Thanks to Corollary 20.4.4 we may regard 𝑨𝔠 as a bona fide linear operator
∗
O (Φ) (Ω) −→ O (Φ ) (Θ) (obviously bounded with respect to the natural Fréchet–
Montel topologies).
The next statement is self-evident.
Proposition 20.4.5 If 𝔠 ⊂ Ω and Θ are sufficiently small then 𝑨𝔠 ℎ ∈ O (−𝜔) (Θ)
(Definition 19.1.10) if ℎ ∈ O (−𝜔) (Ω).

20.4.2 Germ Fourier-like transforms. Definition

As before (𝜑, Φ) is a Sjöstrand pair at (𝑧 ◦ , 𝜃 ◦ ) ∈ Ω × Θ; the integral 𝑨𝔠 ℎ is given by

(20.1). The definition we are seeking is based on the following statement.
Theorem 20.4.6 If the neighborhood Ω′ of 𝑧◦ in Ω and the neighborhood Θ′ of 𝜃 ◦
in Θ are sufficiently small and 𝔠 ⊂⊂ Ω′ is strongly well-shaped at 𝑧◦ for the function
𝑧 ↦→ Ψ (𝑧, 𝜃 ◦ ) then, whatever ℎ ∈ O (Φ) (Ω′), the equivalence class mod O (−𝜔)
𝜃◦ of
∗
the germ at 𝜃 ◦ of the function 𝑨𝔠 ℎ ∈ O (Φ ) (Θ′) (Corollary 20.4.4) only depends on
·
the equivalence class h𝑧 ◦ ∈ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ of the germ of ℎ at 𝑧 ◦ .

Recall that Φ (𝑧◦ ) = 0 implies O𝑧(−𝜔)

◦ ⊂ O𝑧(Φ)
◦ (Proposition 20.2.13); likewise,
∗
O (−𝜔)
𝜃◦ ⊂ O (Φ ) ′ ◦
𝜃 ◦ by (20.3.9). By contracting Ω (and therefore also 𝔠) about 𝑧 and,
′ ◦
correspondingly, Θ about 𝜃 , Theorem 20.4.6 and Corollary 20.4.4 allow us to
define a germ linear operator
∗
A𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ . (20.4.4)

Definition 20.4.7 Let 𝜑 ∈ O (Ω × Θ) and Φ ∈ C ∞ (Ω; R) be such that (𝜑, Φ) is a

Sjöstrand pair at (𝑧◦ , 𝜃 ◦ ) ∈ Ω × Θ. Let 𝑎 (𝑧, 𝜃, 𝜆) ∈ 𝑆a (Ω × Θ). We shall refer to
(20.4.4) as a germ Fourier-like transform at (𝑧◦ , 𝜃 ◦ ). We shall say that (20.4.4) is
classical if 𝑎 (𝑧, 𝜃, 𝜆) can be taken to be a finite realization of a classical analytic
symbol (Definitions 19.1.7, 19.1.14).1
In the sequel we will often say that the germ operator (20.4.4) originates from
an integral (20.1) [in which ℎ = ℎ (𝑧, 𝜆) ∈ O (Φ) (Ω)] or that in defining (20.4.4) we
start from an integral (20.1).
1 In the remainder of this chapter and in the following two chapters, we often omit the adjective
analytic. It is tacitly understood that all phase-functions, amplitudes and symbols are analytic as
specified in Section 20.1.
20.4 Germ Fourier-like Transforms 859

We now proceed with the proof of Theorem 20.4.6. It consists of a couple of steps
which we formalize as propositions.
For 𝑈 ⊂ Ω′ open, 𝑧 ◦ ∈ 𝔠 ∩ 𝑈, we introduce the integral
∫
𝑨𝔠,𝑈 ℎ (𝜃, 𝜆) = e𝑖𝜆𝜑 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃, 𝜆) ℎ (𝑧, 𝜆) d𝑧. (20.4.5)
𝔠∩𝑈

Proposition 20.4.8 Let 𝑈, 𝑉 be open subsets of Ω′ such that 𝑧 ◦ ∈ 𝑈 ∩ 𝑉. If the

neighborhood Θ′ ⊂ Θ of 𝜃 ◦ and 𝜅 > 0 are sufficiently small then

∀𝜃 ∈ Θ′, ∀𝜆 ≥ 1, 𝑨𝔠,𝑈 ℎ (𝜃, 𝜆) − 𝑨𝔠,𝑉 ℎ (𝜃, 𝜆) ≲ e−𝜅𝜆 (20.4.6)

for each ℎ ∈ O (Φ) (Ω′).

Proof The set 𝐾 = 𝔠\ (𝔠 ∩ 𝑈 ∩ 𝑉) ⊂ Ω\ {𝑧 ◦ } is compact; since 𝔠 is well-shaped for

the function 𝑧 ↦→ Ψ (𝑧, 𝜃 ◦ ) at 𝑧 ◦ we have maxΨ (𝑧, 𝜃 ◦ ) ≤ Ψ (𝑧◦ , 𝜃 ◦ ) − 𝛿 for some
𝑧 ∈𝐾
𝛿 > 0. Since Ψ (𝑧 ◦ , 𝜃 ◦ ) = 0 and dΨ (𝑧◦ , 𝜃 ◦ ) = 0 we also have Ψ (𝑧, 𝜃) ≤ 𝐶◦ |𝜃 − 𝜃 ◦ | 2
for some 𝐶◦ > 0 and all 𝜃 in a neighborhood Θ′ ⊂ Θ of 𝜃 ◦ ; we deduce that if Θ′ is
sufficiently small then
1
Ψ (𝑧, 𝜃) ≤ − 𝛿 < 0 (20.4.7)
2
for all 𝑧 ∈ 𝐾, 𝜃 ∈ Θ′. We get, for suitable positive constants 𝐶 𝜀 , 𝐶 𝜀′ and all 𝜆 ≥ 1
and all 𝜃 ∈ Θ′,
∫
𝑨𝔠,𝑈 ℎ (𝜃, 𝜆) − 𝑨𝔠,𝑉 ℎ (𝜃, 𝜆) ≤ e𝜆Ψ(𝑧, 𝜃) e−𝜆Φ(𝑧) |ℎ (𝑧, 𝜆) d𝑧|
𝐾
∫
≤ 𝐶 𝜀 e 𝜀𝜆 e𝜆Ψ(𝑧, 𝜃) |d𝑧|
𝐾
′ − ( 21 𝛿−𝜀 ) 𝜆
≤ 𝐶𝜀e ,

the last inequality a consequence of (20.4.7); (20.4.6) ensues by selecting 𝜀 < 21 𝛿.□

Corollary 20.4.4 and Proposition 20.4.8 allow us to define the map

∗
𝑨𝔠,𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ (20.4.8)
·
that transforms the coset h𝑧 ◦ into the equivalence class mod O (−𝜔)
𝜃◦ of the germ at 𝜃 ◦
(Φ ∗) ′
of the function 𝑨𝔠 ℎ ∈ O (Θ ).

Proposition 20.4.9 The linear operator (20.4.8) is independent of the contour of

integration 𝔠 (assumed to have the “correct” orientation).

Proof If 𝔠 and 𝔠 1 are contours of integration centered at 𝑧◦ both with the “correct”
orientation [and both strongly well-shaped for Ψ (𝑧, 𝜃 ◦ ) at 𝑧◦ ] inspection of the
proof of Lemma 20.4.1 shows that we can find a smooth curve of contours 𝔠 (𝑠),
𝑠 ∈ [0, 1] ∋ 𝑠 ↦→ 𝔠 (𝑠) ⊂⊂ Ω, joining 𝔠 = 𝔠 (0) to 𝔠 (1) = 𝔠 1 ; we can even select 𝔠 (𝑠)
860 20 Germ Fourier Integral Operators in Complex Space

to be centered at 𝑧 ◦ whatever 𝑠 ∈ [0, 1]. Inspection of the proof of Proposition 20.4.3

shows that, if Θ′ and |𝑠 − 𝑠 ′ | are sufficiently small then the analogue of (20.4.3)
holds:

∀𝜃 ∈ Θ′, ∀𝜆 ≥ 1, 𝑨𝔠 (𝑠) ℎ (𝜃, 𝜆) − 𝑨𝔠 (𝑠′ ) ℎ (𝜃, 𝜆) ∈ O (−𝜔) (Θ′) .

Then the Borel–Lebesgue Lemma enables us to conclude that the same is true when
𝑠 = 0, 𝑠 ′ = 1. □
Removing 𝔠 from (20.4.8) completes the proof of Theorem 20.4.6.

Remark 20.4.10 The leeway in the choice of contours 𝔠 in the integral (20.1) and
the fact that the concept of strongly well-shapedness is coordinate-free provides us
with some especially convenient choices. By Proposition 20.3.8 we can assume that
𝑛 2
1 ∑︁
Ψ (𝑧, 𝜃 ◦ ) = − 𝜒 𝑗 𝑥 𝑗 − 𝑥 ◦𝑗
2 𝑗=1
𝑛 2
1 ∑︁ ′
+ 𝜒 𝑗 𝑦 𝑗 − 𝑦 ◦𝑗 + 𝑂 |𝑧 − 𝑧 ◦ | 3
2 𝑗=1

with 𝜒 𝑗 > 0, 𝜒 ′𝑗 > 0, 𝑗 = 1, ..., 𝑛. We may then take 𝔠 to be the open ball in R𝑛 + 𝑖𝑦 ◦ ,

𝔅𝑟(𝑛) (𝑥 ◦ ) + 𝑖𝑦 ◦ = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C𝑛 ; |𝑥 − 𝑥 ◦ | < 𝑟, 𝑦 = 𝑦 ◦ } , 𝑟 > 0,

in which case we have, for all 𝑧 = 𝑥 ∈ 𝔠,

𝑛 2
1 ∑︁
Ψ (𝑧, 𝜃 ◦ ) = − 𝜒 𝑗 𝑥 𝑗 − 𝑥 ◦𝑗 + 𝑂 |𝑥 − 𝑥 ◦ | 3 .
2 𝑗=1

Proposition 20.4.11 If the analytic symbol 𝑎 (𝑧, 𝜃, 𝜆) in the integral (20.1) belongs
′
to O (−𝜔) (Ω × Θ) and if 𝔠 and the open set Θ′ ⊂ Θ (Θ ∋ 𝜃 ◦ ) are sufficiently small
then 𝑨𝔠 ℎ (𝜃, 𝜆) ∈ O (−𝜔) (Θ′) whatever ℎ (𝑧, 𝜆) ∈ O (Φ) (Ω).

Proof There are positive constants 𝜅, 𝐶 such that

∀𝜆 ≥ 1, sup |𝑎 (𝑧, 𝜃, 𝜆)| ≤ 𝐶e−𝜅𝜆

(𝑧, 𝜃) ∈𝔠×Θ′

whence ∫
| 𝑨𝔠 ℎ (𝜃, 𝜆)| ≤ 𝐶e−𝜅𝜆 e𝜆Ψ(𝑧, 𝜃) e−𝜆Φ(𝑧) |ℎ (𝑧, 𝜆) d𝑧|
𝔠
for all 𝜃 ∈ Θ′, 𝜆 ≥ 1, with Ψ as in (20.3.1). We apply (20.2.1) with 𝐾 the closure of
𝔠: to every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that
∫
−(𝜅−𝜀)𝜆
| 𝑨𝔠 ℎ (𝜃, 𝜆)| ≤ 𝐶 𝜀 𝐶e e𝜆Ψ(𝑧, 𝜃) |d𝑧| .
𝔠
20.4 Germ Fourier-like Transforms 861

Since Ψ (𝑧◦ , 𝜃 ◦ ) = 0 we can take 𝜀 and 𝔠 × Θ′ so small that Ψ (𝑧, 𝜃) + 𝜀 ≤ 21 𝜅, whence

∫
1
| 𝑨𝔠 ℎ (𝜃, 𝜆)| ≤ 𝐶 𝜀 𝐶e− 2 𝜅𝜆 |d𝑧| . □
𝔠

Corollary 20.4.12 If 𝑎 (𝑧, 𝜃, 𝜆) in (20.1) belongs to O (−𝜔) (Ω × Θ) then 𝑨 𝑧 ◦ , 𝜃 ◦ = 0.

Remark 20.4.13 When dealing with a classical analytic symbol
∞
∑︁
𝑎 (𝑧, 𝜃, 𝜆) = 𝜆 𝑚− 𝑗 𝑎 𝑗 (𝑧, 𝜃)
𝑗=0

(Definition 19.1.14) the amplitude in (20.1) must be taken to be a finite realization

𝑎 (𝜈) (𝑧, 𝜃, 𝜆) of 𝑎. This can be kept unchanged throughout the proof of Theorem
20.4.6; Lemma 19.1.9 applied to 𝑎, and Corollary 20.4.12 imply that a change of
the finite realization of 𝑎 has no effect on the definition of the germ operator 𝑨 𝑧 ◦ , 𝜃 ◦ ,
which will then be called classical.
Remark 20.4.14 It should be emphasized that, until now in this section (throughout
the construction of the germ operator 𝑨 𝑧 ◦ , 𝜃 ◦ ), there was no need to assume 𝑁 = 𝑛.
This is not so in the remainder of this subsection.
We are going to need germ Fourier-like transforms that “go the other way”:
∗
B 𝜃 ◦ ,𝑧 ◦ : O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ , (20.4.9)

defined by integrals
𝑛 ∫
𝜆
𝑩𝔠∗ 𝑔 (𝑧, 𝜆) = e−𝑖𝜆𝜑 (𝑧, 𝜃) 𝑏 (𝑧, 𝜃, 𝜆) 𝑔 (𝜃, 𝜆) d𝜃. (20.4.10)
2𝜋 𝔠∗
𝑛
[The factor 2𝜆𝜋 serves to remind us of the analogy with the inverse Fourier
transform.] Here we are taking full advantage of Theorem 20.3.16 (and therefore
we must assume 𝑁 = 𝑛): (−𝜑, Φ∗ ) is a Sjöstrand pair at (𝜃 ◦ , 𝑧◦ ) in Θ × Ω (possibly
contracted). The definition of (20.4.9) is the same as that of 𝑨 𝑧 ◦ , 𝜃 ◦ after exchange
of 𝑧 and 𝜃.
The composite operator

B 𝜃 ◦ ,𝑧 ◦ 𝑨 𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (20.4.11)

is well-defined. Actually, (20.4.11) can be defined directly, starting from an integral

(20.2) where the phase-function and the contour of integration have a special form:
𝑛 ∫ ∫
𝜆
e𝑖𝜆( 𝜑 (𝑧, 𝜃)−𝜑 (𝑤, 𝜃)) 𝑏 (𝑧, 𝜃, 𝜆) 𝑎 (𝑤, 𝜃, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜃. (20.4.12)
2𝜋 𝔠∗ 𝔠

The following section studies integrals generalizing (20.4.12).

862 20 Germ Fourier Integral Operators in Complex Space

20.4.3 Germ Fourier-like transforms of hyperfunctions

Hyperfunctions are well defined on totally real C 𝜔 submanifolds of C𝑛 since a

local biholomorphism can deform a contour of integration into an open subset
𝑈 ⊂ Ω ∩ R𝑛 ⊂ C𝑛 ; as before Ω (resp., Θ) is a domain in C𝑛 (resp., C 𝑁 ). Consider
a phase-function 𝜑 ∈ O (Ω × Θ) and a symbol 𝑎 (𝑧, 𝜃, 𝜆) ∈ 𝑆a (Ω × Θ); we assume
that 𝜑 (𝑥 ◦ , 𝜃 ◦ ) = 0 where 𝑥 ◦ ∈ Ω ∩ R𝑛 , 𝜃 ◦ ∈ Θ. Here we shall briefly discuss
the Fourier-like transform at (𝑥 ◦ , 𝜃 ◦ ) of the germ at 𝑥 ◦ of a hyperfunction 𝑓 in 𝑈,
f 𝑥 ◦ ∈ B 𝑥 ◦ , defined by 𝜑 and 𝑎.
Let an analytic functional 𝜇 ∈ O ′ (𝐾) represent 𝑓 in a neighborhood 𝑉 of 𝑥 ◦
in R𝑛 ; we assume that 𝐾 = 𝑉 and supp 𝜇 ⊂ 𝐾. This means that 𝑓 | 𝑉 is the coset
of 𝜇 in O ′ (𝐾) /O ′ (𝜕𝑉) (cf. Definition 7.1.1). We are interested in the Fourier-like
transform of 𝜇 defined by the duality bracket (taken in 𝑧-space)
D E
𝑨𝜇 (𝜃, 𝜆) = 𝜇 𝑧 ,e𝑖𝜆𝜑 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃, 𝜆) ; (20.4.13)

𝑨𝜇 is a holomorphic function of 𝜃 ∈ Θ. If 𝐾 𝜀 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝐾) ≤ 𝜀}, 𝜀 > 0

such that 𝐾 𝜀 ⊂ Ω, there is a 𝐶 𝜀 > 0 such that

| 𝑨𝜇 (𝜃, 𝜆)| ≤ 𝐶 𝜀 e 𝜀𝜆 max e−𝜆 Im 𝜑 (𝑧, 𝜃)
𝑧 ∈𝐾 𝜀

≤ 𝐶 𝜀 e 𝜀𝜆 exp 𝜆 max (− Im 𝜑 (𝑧, 𝜃)) + 𝜀 .
𝑧 ∈𝐾 𝜀

We have
max (− Im 𝜑 (𝑧, 𝜃)) ≤ max (− Im 𝜑 (𝑥, 𝜃)) + 𝐶 ′ 𝜀.
𝑧 ∈𝐾 𝜀 𝑥 ∈𝐾

Since 𝜃 ↦→ − Im 𝜑 (𝑧, 𝜃) is pluriharmonic and Θ ∋ 𝜃 ↦→ max (− Im 𝜑 (𝑥, 𝜃)) is con-

𝑥 ∈𝐾
tinuous Proposition 11.2.8 implies that Φ∗𝐾 (𝜃) = max (− Im 𝜑 (𝑥, 𝜃)) is a plurisub-
𝑥 ∈𝐾
harmonic function in Θ. After a redefinition of 𝜀 we obtain
∗
| 𝑨𝜇 (𝜃, 𝜆)| ≤ 𝐶 𝜀 e𝜆 ( Φ𝐾 ( 𝜃)+𝜀 ) , (20.4.14)
∗
i.e., 𝑨𝜇 ∈ O ( Φ𝐾 ) (Θ).
We assume that 𝑈 is well-shaped at 𝑥 ◦ for the function 𝑧 ↦→ − Im 𝜑 (𝑧, 𝜃) (Defi-
nition 20.2.2) uniformly with respect to 𝜃 on compact subsets of Θ, implying that to
every compact set 𝐾 ∗ ⊂ Θ there is a 𝛿 > 0 (depending on 𝐾 and 𝐾 ∗ ) such that

∀𝜃 ∈ 𝐾 ∗ , sup (− Im 𝜑 (𝑥, 𝜃)) ≤ − Im 𝜑 (𝑥 ◦ , 𝜃) − 𝛿.

𝑥 ∈𝜕𝑉

If supp 𝜇 ⊂ 𝜕𝑉 we have, for all 𝜆 ≥ 1,

∀𝜃 ∈ 𝐾 ∗ , | 𝑨𝜇 (𝜃, 𝜆)| ≤ 𝐶 𝜀 exp 𝜆 (− Im 𝜑 (𝑥 ◦ , 𝜃) − 𝛿 + 𝜀) .

20.5 Sjöstrand Triads and Germ Fourier Integral Operators 863

We can contract Θ′ about 𝜃 ◦ so that sup Im 𝜑 (𝑥 ◦ , 𝜃) < 𝛿/4 and select 𝜀 < 𝛿/4 to
𝜃 ∈Θ′
ensure that
∀𝜆 ≥ 1, sup | 𝑨𝜇 (𝜃, 𝜆)| ≤ 𝐶 𝜀 exp (−𝛿𝜆/2) .
𝜃 ∈Θ′

We see that if supp 𝜇 ⊂ 𝜕𝑉 and Θ′ ∋ 𝜃 ◦ is sufficiently small then 𝑨𝜇 ∈ O (−𝜔) (Θ′).

The same argument shows that if 𝜇 𝑗 ∈ O ′ (𝐾), 𝑗 = 1, 2, represent the same
hyperfunction in 𝑉 then 𝑨𝜇1 − 𝑨𝜇2 ∈ O (−𝜔) (Θ′). Putting all this together
we see that (20.4.13) associates to the hyperfunction 𝑓 ∈ B (𝑉) an element of
∗
O ( Φ𝐾 ) (Θ′) /O (−𝜔) (Θ′).
If 𝑥 ◦ ∈ 𝑉 ′ ⊂ 𝑉 and 𝐾 ′ = 𝑉 ′ ⊂ 𝐾 then

∀𝜃 ∈ Θ, Φ∗𝐾 ′ (𝜃) = max′ (− Im 𝜑 (𝑥, 𝜃)) ≤ Φ∗𝐾 (𝜃) .

𝑥 ∈𝐾

As we contract 𝐾 to {𝑥 ◦ } the functions Φ∗𝐾 (𝜃) decrease to Φ∗𝑥 ◦ (𝜃) = − Im 𝜑 (𝑥 ◦ , 𝜃),

which is pluriharmonic in Θ. This shows that (20.4.13) associates to f 𝑥 ◦ the germ at
𝜃 ◦ , h 𝑥 ◦ , 𝜃 ◦ , of a measurable function of (𝜃, 𝜆), holomorphic with respect to 𝜃, that has
the following property: to every 𝜀 > 0 there is a representative of the germ h 𝑥 ◦ , 𝜃 ◦ in
a neighborhood Θ′ ⊂ Θ of 𝜃 ◦ , of the type 𝑨𝜇 (𝜃, 𝜆) with 𝜇 ∈ O ′ (𝐾) representing 𝑓
in 𝑉, such that
∗
∀𝜆 ≥ 1, ∀𝜃 ∈ Θ′, | 𝑨𝜇 (𝜃, 𝜆)| ≤ 𝐶 𝜀 e𝜆 ( Φ 𝑥 ◦ ( 𝜃)+𝜀 ) . (20.4.15)

We must point out that this does not mean that there is a representative ℎ (𝜃, 𝜆) ∈
∗
O ( Φ 𝑥 ◦ ) (Θ′) of h 𝑥 ◦ , 𝜃 ◦ : the latter would mean that we can choose 𝜇 and Θ′ such that
(20.4.15) holds for all 𝜀 > 0, which has not been proved.
Note also that h 𝑥 ◦ , 𝜃 ◦ is a microlocal object and ought to be associated to the
microfunction at (𝑥 ◦ , 𝜃 ◦ ) defined by 𝑓 . But the Fourier-like transform as defined
here does not allow us to mod off the hyperfunctions that are microanalytic at
(𝑥 ◦ , 𝜃 ◦ ).

20.5 Sjöstrand Triads and Germ Fourier Integral Operators

20.5.1 Sjöstrand triads

In this section we turn our attention to germ operators defined by integrals of the type
(20.2). We describe the framework: Ω1 , Ω2 are domains in C𝑛1 and C𝑛2 , Θ is a domain
in C 𝑁 , containing “central” points 𝑧◦ , 𝑤 ◦ and 𝜃 ◦ respectively. About the phase-
function 𝜑 ∈ O (Ω1 × Ω2 × Θ) we shall always assume ◦ ◦ ◦
that Im 𝜑 (𝑧 , 𝑤 , 𝜃 ) = 0.
We introduce two exponent-weights Φ 𝑗 ∈ C Ω 𝑗 ; R ( 𝑗 = 1, 2) such that Φ1 (𝑧◦ ) =
∞

Φ2 (𝑤 ◦ ) = 0. The following C ∞ functions in Ω1 × Ω2 × Θ will play a crucial role in

the forthcoming analysis:
864 20 Germ Fourier Integral Operators in Complex Space

Ψ1 (𝑧, 𝑤, 𝜃) = Φ1 (𝑧) − Im 𝜑 (𝑧, 𝑤, 𝜃) , (20.5.1)

Ψ2 (𝑧, 𝑤, 𝜃) = Φ2 (𝑤) + Im 𝜑 (𝑧, 𝑤, 𝜃) . (20.5.2)

Definition 20.5.1 We shall say that (𝜑, Φ1 , Φ2 ) is a Sjöstrand triad at (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) ∈

Σ 𝜑 in Ω1 × Ω2 × Θ if the following conditions are satisfied:
(1) (𝑧 ◦ , 𝜃 ◦ ) is a saddle point of Ψ1 (𝑧, 𝑤 ◦ , 𝜃);
(2) (𝑤 ◦ , 𝜃 ◦ ) is a saddle point of Ψ2 (𝑧◦ , 𝑤, 𝜃);
(3) the critical value of (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤, 𝜃) is equal to Φ2 (𝑤) in a neighborhood
of (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ).

Property (1) in Definition 20.5.1 means that (𝜑 (𝑧, 𝑤, 𝜃) , Φ1 (𝑧)) is a Sjöstrand

pair at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) in (Ω1 × Θ) ×Ω2 (cf. Definition 20.3.6); here (𝑧, 𝜃) is the variable
in the base, it plays the role of 𝑧 in Definition 20.3.6; the role of 𝜃 is played here by
𝑤. Note that when 𝑛1 = 𝑛2 = 𝑁 the number of base variables 𝑧 𝑗 , 𝜃 𝑘 is the double of
that of fiber variables, 𝑤 ℓ . This is also valid for Property (2) in Definition 20.5.1 if
we exchange 𝑧 and 𝑤 as well as 𝜑 and −𝜑.

Remark 20.5.2 Definition 20.5.1, (1), does not imply that (𝜑 (𝑧, 𝑤 ◦ , 𝜃) , Φ1 (𝑧)) is a
Sjöstrand pair at (𝑧◦ , 𝜃 ◦ ). As an example take Ω1 = Ω2 = Ω ∋ 𝑧◦ = 𝑤 ◦ , 𝜑 (𝑧, 𝑤, 𝜃) =
(𝑧 − 𝑤) · (𝜃 − 𝜃 ◦ ), Φ1 ≡ 0. Clearly, (𝑧 ◦ , 𝜃 ◦ ) is a critical
point of Ψ1 (𝑧, 𝑧◦ , 𝜃) =
0 𝐼
𝜑 (𝑧, 𝑧◦ , 𝜃) and the Hessian of 𝜑 (𝑧, 𝑧◦ , 𝜃) is 𝑛
, whose eigenvalues are ±1,
𝐼𝑛 0
◦ ◦
each with multiplicity 𝑛. But Ψ1 (𝑧, 𝑧 , 𝜃 ) ≡ 0 (cf. Definition 20.3.6).

The critical points of (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤, 𝜃) are the points (𝑧, 𝜃) ∈ Ω1 ×Θ satisfying
the system of equations

𝜕𝑧 Im 𝜑 (𝑧, 𝑤, 𝜃) = 𝜕𝑧 Φ1 (𝑧) , (20.5.3)

𝜕𝜃 Im 𝜑 (𝑧, 𝑤, 𝜃) = 0.

We apply Lemma 20.3.2: the system of equations (20.5.3) has a unique C ∞ solution
(𝑧 (𝑤) , 𝜃 (𝑤)) such that 𝑧 (𝑤 ◦ ) = 𝑧 ◦ , 𝜃 (𝑤 ◦ ) = 𝜃 ◦ . Condition (3) in Definition 20.5.1
states that
Ψ1 (𝑧 (𝑤) , 𝑤, 𝜃 (𝑤)) = Φ2 (𝑤) . (20.5.4)
It is convenient to refer to a point (𝑧, 𝑤, 𝜃) satisfying (20.5.3) as a critical point
of (20.5.1) although it is a point in Ω1 × Ω2 × Θ. The same terminology will be used
in dealing with (20.5.2). With this terminology (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) is a critical point of both
(20.5.1) and (20.5.2).

Proposition 20.5.3 Let (𝜑, Φ1 , Φ2 ) be a Sjöstrand triad at (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) in Ω1 ×Ω2 ×Θ.

If Ω1 , Ω2 , Θ are sufficiently small then (20.5.1) and (20.5.2) have the same critical
points (in Ω1 × Ω2 × Θ) and the critical value of (𝑧, 𝜃) ↦→ Ψ2 (𝑧, 𝑤, 𝜃) is equal to
Φ1 (𝑧).
20.5 Sjöstrand Triads and Germ Fourier Integral Operators 865

Proof Lemma 20.3.2 implies that the critical points of (20.5.2) in Ω1 × Ω2 × Θ are
the solutions (𝑧, 𝑤 (𝑧) , 𝜃 (𝑧)) of the system of equations

𝜕𝑤 Im 𝜑 (𝑧, 𝑤, 𝜃) = −𝜕𝑤 Φ2 (𝑤) , (20.5.5)

𝜕𝜃 Im 𝜑 (𝑧, 𝑤, 𝜃) = 0,

such that 𝑤 (𝑧◦ ) = 𝑤 ◦ , 𝜃 (𝑧◦ ) = 𝜃 ◦ ; as 𝑧 varies in Ω1 (suitably adjusted) they describe

a 2𝑛-dimensional C ∞ submanifold L passing through (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ).
We differentiate both sides in (20.5.4) with respect to 𝑤; (20.5.1) and (20.5.3)
entail

𝜕𝑤 Φ2 (𝑤) = (𝜕𝑤 Ψ1 ) (𝑧 (𝑤) , 𝑤, 𝜃 (𝑤))

= − (𝜕𝑤 Im 𝜑) (𝑧 (𝑤) , 𝑤, 𝜃 (𝑤)) .

We conclude that (𝑧 (𝑤) , 𝑤, 𝜃 (𝑤)) satisfies (20.5.5). As 𝑤 varies in Ω2 (suitably

adjusted) the point (𝑧 (𝑤) , 𝑤, 𝜃 (𝑤)) describes a neighborhood of (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) in L.□

Corollary 20.5.4 If (𝜑, Φ1 , Φ2 ) is a Sjöstrand triad at (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) in Ω1 × Ω2 × Θ

then (−𝜑, Φ2 , Φ1 ) is a Sjöstrand triad at (𝑤 ◦ , 𝑧◦ , 𝜃 ◦ ) in Ω2 × Ω1 × Θ.

As an example we discuss an important particular case of a Sjöstrand triad.

Proposition 20.5.5 Let 𝑓 : Ω2 −→ Ω1 be a biholomorphism such that 𝑓 (𝑤 ◦ ) = 𝑧 ◦

and let (𝜑, Φ) be a Sjöstrand pair at (𝑧 ◦ , 𝜃 ◦ ) in Ω1 × Θ. Under these hypotheses and
provided Ω1 × Ω2 × Θ is sufficiently small, if 𝜓 (𝑧, 𝑤, 𝜃) = 𝜑 ( 𝑓 (𝑤) , 𝜃) − 𝜑 (𝑧, 𝜃)
[∈ O (Ω1 × Ω2 × Θ) ] then (𝜓, Φ (𝑧) , Φ ( 𝑓 (𝑤))) is a Sjöstrand triad at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ )
in Ω1 × Ω2 × Θ.

Proof We have

Ψ1 (𝑧, 𝑤, 𝜃) = Φ1 (𝑧) − Im 𝜑 (𝑧, 𝜃) + Im 𝜑 ( 𝑓 (𝑤) , 𝜃) ,

= Ψ (𝑧, 𝜃) + Im 𝜑 ( 𝑓 (𝑤) , 𝜃) .

On the one hand, by (20.3.9) and Proposition 20.3.8 we have

1
Ψ (𝑧, 𝜃) = Φ∗ (𝜃) − 𝑄 (𝜃; 𝑧 − 𝑧 (𝜃)) + 𝑂 |𝑧 − 𝑧◦ | 3 + |𝜃 − 𝜃 ◦ | 3
2
with Φ∗ defined in (20.3.5) and the (real) quadratic form 𝑧 ↦→ 𝑄 (𝜃; 𝑧) nondegenerate
and having signature zero. On the other hand, by (20.3.13),

Φ∗ (𝜃) + Im 𝜑 ( 𝑓 (𝑤) , 𝜃) = Ψ∗ ( 𝑓 (𝑤) , 𝜃) .

By Propositions 20.3.11, 20.3.14, we have

1
Ψ∗ ( 𝑓 (𝑤) , 𝜃) = Φ ( 𝑓 (𝑤)) + 𝑄 ∗ ( 𝑓 (𝑤) ; 𝜃 − 𝜁 ( 𝑓 (𝑤))) + 𝑂 |𝜃 − 𝜁 ( 𝑓 (𝑤))| 3 ,
2
866 20 Germ Fourier Integral Operators in Complex Space

where 𝜁 (𝑧) has the same meaning as in Proposition 20.3.11 [but 𝜁 (𝑧◦ ) = 𝜃 ◦ in our
present notation] and the (real) quadratic form 𝜃 ↦→ 𝑄 ∗ (𝑧; 𝜃) is nondegenerate and
has signature zero. Putting all this together yields
1 1
Ψ1 (𝑧, 𝑤, 𝜃) = Φ ( 𝑓 (𝑤)) + 𝑄 (𝜃; 𝑧 − 𝑧 (𝜃)) + 𝑄 ∗ ( 𝑓 (𝑤) ; 𝜃 − 𝜁 ( 𝑓 (𝑤)))
2 2
+ 𝑂 |𝑧 − 𝑧 (𝜃)| 3 + |𝜃 − 𝜁 ( 𝑓 (𝑤))| 3 .

First of all, this proves that Φ ( 𝑓 (𝑤)) is the critical value of (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤, 𝜃).
Next, putting 𝑤 = 𝑤 ◦ and therefore 𝑓 (𝑤) = 𝑧 ◦ we get
1 1
Ψ1 (𝑧, 𝑤 ◦ , 𝜃) = 𝑄 (𝜃; 𝑧 − 𝑧 (𝜃)) + 𝑄 ∗ (𝑧 ◦ ; 𝜃 − 𝜃 ◦ )
2 2
+ 𝑂 |𝑧 − 𝑧 (𝜃)| 3 + |𝜃 − 𝜃 ◦ | 3

since 𝜁 (𝑧◦ ) = 𝜃 ◦ and Φ (𝑧◦ ) = 0. Since (𝑧 − 𝑧 (𝜃) , 𝜃) ↦→ (𝑧, 𝜃) is a diffeomorphism

of a neighborhood of (0, 𝜃 ◦ ) onto one of (𝑧 ◦ , 𝜃 ◦ ) we see that (𝜓, Φ, Φ ◦ 𝑓 ) has the
properties (1) and (3) in Definition 20.5.1. The same argument after exchanging 𝑧
and 𝑤 proves Property (2). □
Note that for 𝜓 as in Proposition 20.5.5 the equation 𝜕𝜃 𝜓 (𝑧, 𝑤, 𝜃) = 0 defines the
set [cf. (20.5.12) below]

Σ 𝜓 = {(𝑧, 𝑤, 𝜃) ∈ Ω × Ω × Θ; 𝑧 = 𝑓 (𝑤)} . (20.5.6)

20.5.2 Germ FIOs. Action on holomorphic functions

Through the remainder of this section we shall assume that (𝜑, Φ1 , Φ2 ) is a Sjöstrand
triad at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 in Ω1 × Ω2 × Θ. Under this hypothesis we study integrals
of the type (20.2),
𝑛 ∫
𝜆
𝑨ℭ1 ℎ1 (𝑤, 𝜆) = e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ1 (𝑧, 𝜆) d𝑧d𝜃, (20.5.7)
2𝜋 ℭ1

or
𝑛 ∫
𝜆
𝑩ℭ2 ℎ2 (𝑧, 𝜆) = e−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃, (20.5.8)
2𝜋 ℭ2

where 𝑎, 𝑏 ∈ 𝑆a (Ω1 × Ω2 × Θ) (Definition 20.1.1), ℭ 𝑗 ⊂⊂ Ω 𝑗 × Θ is a contour

of integration and ℎ 𝑗 ∈ O ( Φ 𝑗 ) Ω 𝑗 ( 𝑗 = 1, 2); ℭ1 (resp., ℭ2 ) is strongly well-

shaped at (𝑧◦ , 𝜃 ◦ ) [resp., (𝑤 ◦ , 𝜃 ◦ )] for the function (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤 ◦ , 𝜃) [resp.,
20.5 Sjöstrand Triads and Germ Fourier Integral Operators 867

(𝑤, 𝜃) ↦→ Ψ2 (𝑧 ◦ , 𝑤, 𝜃); see (20.5.1), (20.5.2)]. It should be underlined that we are

now regarding Φ 𝑗 as an exponent-weight in Ω 𝑗 × Θ and ℎ 𝑗 as a function belonging
to O ( Φ 𝑗 ) Ω 𝑗 × Θ , both independent of 𝜃.

We can apply directly the results of Section 20.3, mainly Corollary 20.4.4, Propo-
sition 20.4.5 and Theorem 20.4.6:

Theorem 20.5.6 Let (𝜑, Φ1 , Φ2 ) be a Sjöstrand triad at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) in Ω1 × Ω2 × Θ.

If the neighborhood Ω1′ × Θ′ of (𝑧◦ , 𝜃 ◦ ) in Ω1 × Θ and the neighborhood Ω2′ of 𝑤 ◦
in Ω2 are sufficiently small and ℭ1 ⊂⊂ Ω′ × Θ′ then 𝑨ℭ1 ℎ ∈ O (Φ2 ) Ω2′ whatever

(−𝜔)
ℎ ∈ O (Φ1 ) (Ω1 ) and the equivalence class mod O𝑤 ◦ of the germ at 𝑤 ◦ of the
·
function 𝑨ℭ1 ℎ only depends on the equivalence class h𝑧 ◦ ∈ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ of the germ
◦
of ℎ at 𝑧 .

This defines a germ operator starting from the integral (20.5.7),

A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) : O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦
(Φ2 )
−→ O𝑤 (−𝜔)
◦ /O 𝑤 ◦ . (20.5.9)

Exchanging 𝑧 and 𝑤 as well as 𝜑 and −𝜑 yields a germ operator

(Φ2 ) (−𝜔)
B𝑤 ◦ ,𝑧 ◦ , 𝜃 ◦ : O𝑤 ◦ /O 𝑤 ◦ −→ O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ (20.5.10)

starting from the integral (20.5.8). Both germ operators (20.5.9), (20.5.10), depend
holomorphically on 𝜃 ◦ ∈ Θ, in the same sense that (19.3.19) was said to depend
holomorphically on 𝜃. As 𝑧◦ and 𝑤 ◦ range over Ω1 and Ω2 respectively, they define
homomorphisms of one of the sheaves O (Φ1 ) /O (−𝜔) , O (Φ2 ) /O (−𝜔) onto the other.

Definition 20.5.7 We shall refer to (20.5.9) as a germ Fourier integral operator (in
short, germ FIO) at the point (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) with phase-function 𝜑. We shall say that
A𝑧 ◦ ,𝑤 ◦ ,𝜁 ◦ is classical if 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) in (20.5.7) can be taken to be a finite realization
of a classical amplitude (Definitions 19.1.7, 19.1.14).

20.5.3 Deforming the contours of integration

We return to the integral (20.5.7) and explore possible deformations of the contour
of integration ℭ1 that might be helpful in the coming analysis. Let Ψ1 (𝑧, 𝑤, 𝜃) be the
function defined in (20.5.1). Because of our hypothesis that (𝑧◦ , 𝜃 ◦ ) is a saddle point
of Ψ1 (𝑧, 𝑤 ◦ , 𝜃) or, in other words, that (𝜑 (𝑧, 𝑤, 𝜃) ,Ψ1 (𝑧, 𝑤, 𝜃)) is a Sjöstrand pair
at ((𝑧◦ , 𝜃 ◦ ) , 𝑤 ◦ ) in (Ω1 × Θ) × Ω2 , we will need to, and can, avail ourselves of all the
results in the construction of germ Fourier-like transforms (Section 20.3), in addition
to Corollary 20.4.4, Proposition 20.4.5 and Theorem 20.4.6. In this subsection we
content ourselves with listing those results.
First of all, since the contour ℭ ⊂⊂ Ω × Ξ is strongly well-shaped at (𝑧◦ , 𝜃 ◦ ) for
the function Ψ1 (𝑧, 𝑤 ◦ , 𝜃) we can avail ourselves of Lemma 20.4.1:
868 20 Germ Fourier Integral Operators in Complex Space

If Ω′𝑗 ⊂ Ω ( 𝑗 = 1, 2) and Θ′ ⊂ Θ are sufficiently small (with 𝑧◦ ∈ Ω1′ , 𝑤 ◦ ∈ Ω2′ ,

𝜃 ◦ ∈ Θ′) then, to each 𝑤 ∈ Ω2′ there exist a unique saddle point (𝑧 (𝑤) , 𝜃 (𝑤)) of the
function Ω1′ × Θ′ ∋ (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤, 𝜃) and a contour of integration ℭ (𝑤) ⊂⊂ Ω′
with the following properties:

(1) ℭ (𝑤) depends smoothly on 𝑤 ∈ Ω′ and ℭ (𝑤 ◦ ) = ℭ1 ;

(2) the saddle point (𝑧 (𝑤) , 𝜃 (𝑤)) belongs to ℭ (𝑤);
(3) ℭ (𝑤) is strongly well-shaped for 𝑧 ↦→ Ψ1 (𝑧, 𝑤, 𝜃) at (𝑧 (𝑤) , 𝜃 (𝑤));
(4) the orientation of ℭ (𝑤) is independent of 𝑤.
We also assume that the boundaries of the contours ℭ (𝑤) are smooth and depend
smoothly on 𝑤. The functions
𝑛 ∫
𝜆
𝑨ℭ (𝑤) ℎ (𝑤, 𝜆) = e𝑖 𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ (𝑧, 𝜆) d𝑧d𝜃 (20.5.11)
2𝜋 ℭ (𝑤)

are of class C ∞ with respect to 𝑤 ∈ Ω2′ . The analogues of Propositions 20.4.2 and
20.4.3 are valid:

Proposition 20.5.8 If diam Ω2′ is sufficiently small then to every 𝜀 > 0 and ℎ ∈
O (Φ1 ) (Ω) there is a 𝐶 𝜀 > 0 such that

∀𝑤 ∈ Ω2′ , 𝜆 ≥ 1, 𝑨ℭ (𝑤) ℎ (𝑤, 𝜆) ≤ 𝐶 𝜀 e𝜆(Φ2 (𝑤)+𝜀) .

Proposition 20.5.9 If diam Ω2′ and 𝜅 > 0 are sufficiently small then

∀𝑤 ∈ Ω2′ , 𝜆 ≥ 1, 𝑨ℭ (𝑤) ℎ (𝑤, 𝜆) − 𝑨ℭ1 ℎ (𝑤, 𝜆) ≲ e−𝜅𝜆

for each ℎ ∈ O (Φ1 ) (Ω1 ).

20.5.4 Symplectic considerations

In the preceding two subsections 𝜑 was simply a holomorphic function in Ω1 ×Ω2 ×Θ

real at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ). In the applications one usually puts more stringent conditions on
𝜑. In this subsection we shall assume that 𝑛1 = 𝑛2 = 𝑛.
First of all we observe that the critical sets of Ψ 𝑗 [cf. (20.5.1), (20.5.2)] lie in the
critical set of 𝜑,

Σ 𝜑 = {(𝑧, 𝑤, 𝜃) ∈ Ω1 × Ω2 × Θ; 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) = 0} (20.5.12)

(cf. Definition 18.1.3). We are extending the terminology of Chapter 18 to complex

phase-space; more precisely, we adopt the following
20.5 Sjöstrand Triads and Germ Fourier Integral Operators 869

Definition 20.5.10 A function 𝜑 ∈ O (Ω1 × Ω2 × Θ) shall be referred to as a phase-

function in Ω1 ×Ω2 ×Θ if 𝜑 satisfies the following conditions (cf. Definition 18.1.4):
(1) The critical set Σ 𝜑 is a closed complex-analytic submanifold of Ω1 × Ω2 × Θ of
dimension 2𝑛.
(2) If 𝑝 𝜑 (𝑧, 𝑤, 𝜃) = (𝑧, 𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃)), 𝑞 𝜑 (𝑧, 𝑤, 𝜃) = (𝑤, −𝜕𝑤 𝜑 (𝑧, 𝑤, 𝜃)), the dia-
gram [cf. (18.1.19)]
Σ𝜑
𝑝𝜑 ↙ ↘ 𝑞𝜑 (20.5.13)
𝜒𝜑
𝑇 (1,0) Ω1 \0 −→ 𝑇 (1,0) Ω2 \0
is commutative and each map 𝑝 𝜑 and 𝑞 𝜑 is a local biholomorphism.

For the meaning of 𝑇 (1,0) Ω 𝑗 \0 see (9.4.24). Note that Property (2) would not
make sense if we had allowed 𝑛1 ≠ 𝑛2 .

Proposition 20.5.11 Let 𝜑 be a phase-function in Ω1 ×Ω2 ×Θ. The map 𝜒 𝜑 is a local

biholomorphism and a local symplectomorphism. The graph of 𝜒 𝜑 , Λ
𝜑 , is an im-

mersed Lagrangian complex-analytic submanifold of 𝑇 (1,0) Ω1 \0 × 𝑇 (1,0) Ω2 \0 .

Here the (complex) symplectic structure of 𝑇 (1,0) Ω1 \0 × 𝑇 (1,0) Ω2 \0 is defined
by the “holomorphic” twisted 2-form d𝜁 ∧ d𝑧 − d𝜗 ∧ d𝑤.
Proof Let U1 be a neighborhood of (𝑧∗ , 𝜁 ∗ ) in 𝑇 (1,0) Ω1 \0 and let V be an open
subset of Σ 𝜑 such that 𝑝 𝜑 V is a biholomorphism of V onto U1 ; if U1 and V are
sufficiently small 𝑞 𝜑 V will be a biholomorphism of V onto a neighborhood U2 of
−1
𝜒 𝜑 (𝑧 ∗ , 𝜁 ∗ ) in 𝑇 (1,0) Ω2 \0. It follows that 𝜒 𝜑 U1 = 𝑞 𝜑 V ◦ 𝑝š
𝜑 V is a biholomorphism
of U1 onto U2 .
The pullback to Σ 𝜑 of the (1, 0)-form in Ω1 × Ω2 × Θ,

d𝜑 = (𝜕𝑧 𝜑) d𝑧 + (𝜕𝑤 𝜑) d𝑤 + (𝜕𝜃 𝜑) d𝜃,

reduces to (𝜕𝑧 𝜑) d𝑧 + (𝜕𝑤 𝜑) d𝑤. Thus the pullback under the map

𝑝 𝜑 , 𝑞 𝜑 : Σ 𝜑 −→ 𝑇 (1,0) Ω1 \0 × 𝑇 (1,0) Ω2 \0

of the 1-form 𝜎 = 𝜁 · d𝑧 − 𝜗 · d𝑤 in 𝑇 (1,0) Ω1 \0 × 𝑇 (1,0) Ω2 \0 is equal to d𝜑,
implying d𝜎 = 0, i.e., d𝜁 ∧ d𝑧 = d𝜗 ∧ d𝑤 on Λ 𝜑 . □

Definition 20.5.12 The phase-function

𝜑 is said to be nondegenerate at (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ )
if the differentials d 𝜕𝜃 𝑗 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ), 𝑗 = 1, ..., 𝑁, are linearly independent (cf.
𝜕𝜑

Definition 18.1.6).
870 20 Germ Fourier Integral Operators in Complex Space

Remark 20.5.13 When 𝑛1 = 𝑛2 = 𝑁 = 𝑛 the condition

𝜕 𝜕 𝜑 𝜕𝑧 𝜕𝜃 𝜑
det 𝑧 𝑤 ≠0 (20.5.14)
𝜕𝑤 𝜕𝜃 𝜑 𝜕𝜃2 𝜑

ensures that 𝜑 is a nondegenerate phase-function in Ω1 × Ω2 × Θ. Property (20.5.14)

and the Implicit Function Theorem imply that the equations (defining 𝑝 𝜑 )

𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃) = 𝜁, 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) = 0 (20.5.15)

have a unique solution 𝑤 = 𝑤 (𝑧, 𝜁), 𝜃 = 𝜃 (𝑧, 𝜁), such that

𝑤 (𝑧◦ , 𝜁 ◦ ) = 𝑤 ◦ , 𝜃 (𝑧 ◦ , 𝜁 ◦ ) = 𝜃 ◦ , (20.5.16)

where 𝜁 ◦ = 𝜕𝑧 𝜑 (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ). This confirms that we can use (𝑧, 𝜁) as local coordinates

in Σ 𝜑 . The analogous argument applies to 𝑞 𝜑 and 𝑤, 𝜗 = −𝜕𝑤 𝜑 (𝑧, 𝑤, 𝜃).

Remark 20.5.14 Let 𝑓 : Ω2 −→ Ω1 and the Sjöstrand pair (𝜑, Φ) in Ω1 × Θ be as in

Proposition 20.5.5. We have 𝜓 (𝑧, 𝑤, 𝜃) = 𝜑 ( 𝑓 (𝑤) , 𝜃)−𝜑 (𝑧, 𝜃) ∈ O (Ω1 × Ω2 × Θ).
If we assume that det 𝜕𝑧 𝜕𝜃 𝜑 ≠ 0 at every point of Ω1 × Θ then the equations
𝜕𝜃 𝜓 (𝑧, 𝑤, 𝜃) = 0 and 𝑧 = 𝑓 (𝑤) are equivalent and the matrix

𝜕𝑧 𝜕𝑤 𝜓 (𝑧, 𝑤, 𝜃) 𝜕𝑧 𝜕𝜃 𝜓 (𝑧, 𝑤, 𝜃) 0 −𝜕𝑧 𝜕𝜃 𝜑 (𝑧, 𝜃)
=
𝜕𝑤 𝜕𝜃 𝜓 (𝑧, 𝑤, 𝜃) 𝜕𝜃2 𝜓 (𝑧, 𝑤, 𝜃) 𝜕𝑧 𝜕𝜃 𝜑 ( 𝑓 (𝑤) , 𝜃) 𝜕𝜃2 𝜓 (𝑧, 𝑤, 𝜃)

𝑤, (𝜕𝑤 𝑓 ) ⊤ (𝑤) 𝜁

is nonsingular. It is readily checked that 𝜒 𝜓 (𝑧, 𝜁) = [cf.
−1
(20.5.13)] with 𝑤 = 𝑓 (𝑧). The 𝑛 × 𝑛 matrix (𝜕𝑤 𝑓 ) ⊤ is the contragredient of
−1
𝜕𝑧 𝑓 .
Chapter 21
Germ Pseudodifferential Operators in Complex
Space

This chapter revisits, in greater detail, the procedures and results of Ch. 20 for a
special but important subclass of germ Fourier Integral Operators, which it makes
sense to call germ pseudodifferential operators: those originating from integral op-
erators in which the phase-function is (𝑧 − 𝑤) · 𝜁 (or is reducible to this form), with
special attention to those with classical amplitudes or symbols, and to their relations
to differential operators of infinite order. This subclass comprises the composition
of two germ FIOs with opposite phase-functions; it is stable under conjugation
A ↦→ FAF−1 , where F is an elliptic germ FIO usually associated to a microlocal
(biholomorphic) symplectomorphism, with the effect seen in the germ version of
the important Egorov Theorem. The action of germ pseudodifferential operators on
germs of distributions is easy to describe (Section 21.3). Their actions on germs of
hyperfunctions is defined (in Section 21.4) under a more stringent requirement on
their (complex) phase-function; these phase-functions (here called effective) gener-
alize the FBI phase-functions introduced in the following chapter, and might be of
interest in applications where FBI phase-functions could not be used. They allow
us to define the action of the corresponding pseudodifferential operators on analytic
functionals, thence on hyperfunctions, and, importantly, on hyperfunction boundary
values of holomorphic functions in wedges, which allows us to prove that these pseu-
dodifferential operators decrease the analytic wave-front set, allowing us to extend
their action on singularity hyperfunctions and on microfunctions.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 871
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_21
872 21 Germ Pseudodifferential Operators in Complex Space

21.1 Germ Pseudodifferential Operators

21.1.1 Germ pseudodifferential operators. Definition

A special case of germ FIOs are germ pseudodifferential operators. Their definition
is based on the following variant of Proposition 20.5.5 (in which 𝑓 = Identity).

Proposition 21.1.1 Let Ω and Ξ be domains in C𝑛 , 𝑧◦ ∈ Ω, 𝜁 ◦ ∈ Ξ. Let 𝜑◦ (𝑧, 𝑤, 𝜁) =

(𝑧 − 𝑤) · 𝜁 and Φ ∈ C ∞ (Ω; R), Φ (𝑧 ◦ ) = 0. For (𝜑◦ (𝑧, 𝑤, 𝜁) , Φ (𝑤) , Φ (𝑧)) to be
a Sjöstrand triad at (𝑧◦ , 𝑧◦ , 𝜁 ◦ ) in Ω × Ω × Ξ (assumed to be suitably small) it is
necessary and sufficient that 𝜕𝑧 Φ (𝑧◦ ) = 21 𝑖𝜁 ◦ .

We have exchanged the roles of 𝑧 and 𝑤 compared to the preceding chapter [cf.
Definition 20.5.1, also compare (21.1.2) below to (20.5.7)].
Proof Define [cf. (20.5.1)–(20.5.2)]

Ψ1 (𝑧, 𝑤, 𝜁) = Φ (𝑤) − Im ((𝑧 − 𝑤) · 𝜁) , (21.1.1)

Ψ2 (𝑧, 𝑤, 𝜁) = Φ (𝑧) + Im ((𝑧 − 𝑤) · 𝜁) .

The critical points of (𝑤, 𝜁) ↦→ Ψ1 (𝑧, 𝑤, 𝜁) near (𝑧 ◦ , 𝑧◦ , 𝜁 ◦ ) are the solutions of

1
𝜕𝑤 Φ (𝑤) = 𝜕𝑤 Im 𝜑◦ (𝑧, 𝑤, 𝜁) = 𝑖𝜁,
2
𝜕𝜁 Im 𝜑◦ (𝑧, 𝑤, 𝜁) = 𝑧 − 𝑤 = 0

[cf. (20.5.3)]; they are given by 𝑤 = 𝑧, 𝜁 = −2𝑖𝜕𝑤 Φ (𝑤). The critical value of
(𝑤, 𝜁) ↦→ Ψ1 (𝑧, 𝑤, 𝜁) is Φ (𝑧). For (𝑧 ◦ , 𝜁 ◦ ) to be a critical point of Ψ1 (𝑧◦ , 𝑤, 𝜁) it
is necessary and sufficient that 𝜕𝑧 Φ (𝑧 ◦ ) = 21 𝑖𝜁 ◦ .
It remains to show that (𝑧◦ , 𝜁 ◦ ) is a saddle point of Ψ1 (𝑧◦ , 𝑤, 𝜁). The same
argument applies to Ψ2 (𝑧, 𝑤 ◦ , 𝜁). Let 𝐻 denote the Hessian of Φ (𝑤) with respect
to (Re 𝑤, Im 𝑤) at 𝑧 ◦ and let 𝜒1 , ..., 𝜒2𝑛 be the eigenvalues of 𝐻. The Hessian
determinant of Ψ1 (𝑧 ◦ , 𝑤, 𝜁) with respect to (Re 𝑤, Im 𝑤, 𝜉, 𝜂) at (𝑧◦ , 𝜁 ◦ ) is equal to
that of the 4𝑛 × 4𝑛 matrix
𝐻 𝐼2𝑛
,
𝐼2𝑛 0
√︂ 2
whose eigenvalues are 21 𝜒 𝑗 ± 1+ 1
2 𝜒𝑗 , 𝑗 = 1, ..., 2𝑛. □

We combine Proposition 21.1.1 and Theorem 20.5.6:

Theorem 21.1.2 Let Φ ∈ C ∞ (Ω; R), Φ (𝑧 ◦ ) = 0, 𝜕𝑧 Φ (𝑧◦ ) = 12 𝑖𝜁 ◦ , 𝑝 (𝑧, 𝑤, 𝜁, 𝜆) ∈

𝑆a (Ω × Ω × Ξ) (cf. Definition 20.1.1). If the neighborhood Ω′ ×Ξ′ of (𝑧◦ , 𝜁 ◦ ) in Ω×Ξ
is sufficiently small and if the contour of integration in (𝑤, 𝜁)-space, ℭ ⊂⊂ Ω′ × Ξ′,
is strongly well-shaped at (𝑧◦ , 𝜁 ◦ ) for the function Ψ1 (𝑧◦ , 𝑤, 𝜁) [cf. (21.1.1)] then
21.1 Germ Pseudodifferential Operators 873
𝑛 ∫
𝜆
𝑷ℭ ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝 (𝑧, 𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁, (21.1.2)
2𝜋 ℭ

belongs to O (Φ) (Ω′) whatever ℎ ∈ O (Φ) (Ω) and the equivalence class mod O𝑧(−𝜔)
◦
·
of the germ at 𝑧◦ of the function 𝑨ℭ ℎ only depends on the equivalence class h𝑧 ◦ ∈
O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ of the germ of ℎ at 𝑧◦ .

Starting from the integral (21.1.2) we define a germ operator

P𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (21.1.3)

to which we shall refer as a germ pseudodifferential operator at (𝑧◦ , 𝜁 ◦ ). However,

to facilitate the transition to complex manifolds there is an advantage in using phase-
functions less coordinate-dependent than (𝑧 − 𝑤) · 𝜁 . This is made possible by the
following

Proposition 21.1.3 Suppose that both Ω and Θ are domains in C𝑛 . Let (𝑧 ◦ , 𝜃 ◦ ) ∈

Ω × Θ and let 𝜑 ∈ O (Ω × Ω × Θ) have the following properties

∀ (𝑧, 𝜃) ∈ Ω × Θ, 𝜑 (𝑧, 𝑧, 𝜃) = 0, (21.1.4)

𝜕2 𝜑

det (𝑧 ◦ , 𝑧◦ , 𝜃 ◦ ) ≠ 0. (21.1.5)
𝜕𝑤 𝑗 𝜕𝜃 𝑘 1≤ 𝑗,𝑘 ≤𝑛

Let Φ ∈ C ∞ (Ω; R) be such that Φ (𝑧 ◦ ) = 0, 𝜕𝑧 Φ (𝑧 ◦ ) = − 21 𝑖𝜕𝑤 𝜑 (𝑧◦ , 𝑧◦ , 𝜃 ◦ ).

Under these hypotheses, if Ω′ ⊂ Ω and Θ′ ⊂ Θ are sufficiently small neighbor-
hoods of 𝑧◦ and 𝜃 ◦ respectively, then (𝜑 (𝑧, 𝑤, 𝜃) , Φ (𝑤) , Φ (𝑧)) is a Sjöstrand triad
in Ω′ × Ω′ × Θ′ at (𝑧◦ , 𝑧◦ , 𝜃 ◦ ).

In the remainder of this subsection we use the notation 𝜁 ◦ = −𝜕𝑤 𝜑 (𝑧 ◦ , 𝑧◦ , 𝜃 ◦ ).

Proof By (21.1.4) we have 𝜑 (𝑧, 𝑤, 𝜃) = (𝑧 − 𝑤) · 𝜙1 (𝑧, 𝑤, 𝜃) with

𝜙1 ∈ O (Ω × Ω × Θ; C𝑛 ) and
𝜙1 (𝑧, 𝑧, 𝜃) = − (𝜕𝑤 𝜑) (𝑧, 𝑧, 𝜃) .

We have D𝜙 1 ◦ ◦ ◦ ′ ′ ′
D𝜃 (𝑧 , 𝑧 , 𝜃 ) ≠ 0 by (21.1.5). If Ω × Ω × Θ is sufficiently small the Im-
plicit Function Theorem implies that there is a solution 𝜃 (𝑧, 𝑤, 𝜁) ∈ O (Ω′ × Ω′ × Ξ)
of the equation 𝜙1 (𝑧, 𝑤, 𝜃) = 𝜁 such that 𝜃 (𝑧 ◦ , 𝑧◦ , 𝜁 ◦ ) = 𝜃 ◦ . The local biholomor-
phism 𝜃 → 𝜁 transforms 𝜑 (𝑧, 𝑤, 𝜃) into (𝑧 − 𝑤) · 𝜁. Then Proposition 21.1.1 implies
the desired conclusion. □
It follows from (21.1.4) that 𝜕𝑧 𝜑 (𝑧, 𝑧, 𝜃) = −𝜕𝑤 𝜑 (𝑧, 𝑧, 𝜃) for all (𝑧, 𝜃) ∈ Ω × Θ;
this implies that (21.1.5) is equivalent to
2
𝜕 𝜑 ◦ ◦ ◦
det (𝑧 , 𝑧 , 𝜃 ) ≠ 0. (21.1.6)
𝜕𝑧 𝑗 𝜕𝜃 𝑘
874 21 Germ Pseudodifferential Operators in Complex Space

We can carry out the change of variables 𝜃 ⇝ 𝜁 = 𝜙1 (𝑧, 𝑤, 𝜃) (𝜙1 as in the proof
of Proposition 21.1.3) in the integral
𝑛 ∫
𝜆
𝑷ℭ ℎ (𝑧, 𝜆) = e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑝 (𝑧, 𝑤, 𝜃, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜃, (21.1.7)
2𝜋 ℭ

transforming it into an integral of type (21.1.2),

𝑛 ∫
𝜆
𝑷ℭ̃ ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝˜ (𝑧, 𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁 (21.1.8)
2𝜋 ℭ̃ (𝑧)

where
D𝜃
𝑝˜ (𝑧, 𝑤, 𝜁, 𝜆) = 𝑝 (𝑧, 𝑤, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆) (𝑧, 𝑤, 𝜁)
D𝜁
and

ℭ̃ (𝑧) = {(𝑤, 𝜁) ∈ Ω′ × Ξ; ∃𝜃 ∈ Θ′, (𝑤, 𝜃) ∈ ℭ, 𝜁 = 𝜁 (𝑧, 𝑤, 𝜃)} .

Starting from (21.1.8) and, if needed, using the contour deformations described in
Subsection 20.4.1, we can define a germ pseudodifferential operator of the type
(21.1.3).

Definition 21.1.4 Let the phase-function 𝜑 ∈ O (Ω × Ω × Θ) satisfy the conditions

(21.1.4), (21.1.5) and let Φ ∈ C ∞ (Ω; R) be such that Φ (𝑧◦ ) = 0, 𝜕𝑧 Φ (𝑧◦ ) = 12 𝑖𝜁 ◦ .
We shall refer to the germ operator (21.1.3) defined starting from the integral (21.1.7)
as a germ pseudodifferential operator at (𝑧◦ , 𝜃 ◦ ).

21.1.2 Reducing the integrals

We return to the integral (21.1.2). We apply Lemma 20.4.1, in the present context, and
carry out contour deformations 𝑧 ↦→ ℭ (𝑧) with ℭ (𝑧◦ ) = ℭ and ℭ (𝑧) ⊂⊂ Ω′ × Ξ′
strongly well-shaped at (𝑤 (𝑧) , 𝜁 (𝑧)) for the function (𝑤, 𝜁) ↦→ Ψ1 (𝑧, 𝑤, 𝜁) [cf.
(21.1.1)]. Now
𝑛 ∫
𝜆
𝑷ℭ (𝑧) ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝 (𝑧, 𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁
2𝜋 ℭ (𝑧)

with 𝑝 (𝑧, 𝑤, 𝜁, 𝜆) ∈ 𝑆a (Ω × Ω × Ξ), ℎ ∈ O (Φ) (Ω).

We avail ourselves of Propositions 20.5.8 and 20.5.9:
(1) If Ω′ is sufficiently small then to every 𝜀 > 0 and every ℎ ∈ O (Φ) (Ω) there is a
𝐶 𝜀 > 0 such that

∀𝑧 ∈ Ω′, 𝜆 ≥ 1, 𝑷ℭ (𝑧) ℎ (𝑧, 𝜆) ≤ 𝐶 𝜀 e𝜆(Φ(𝑧)+𝜀) . (21.1.9)

21.1 Germ Pseudodifferential Operators 875

(2) If Ω′ and 𝜅 > 0 are sufficiently small then to every ℎ ∈ O (Φ) (Ω) there is a
𝐶 > 0 such that

∀𝑧 ∈ Ω′, 𝜆 ≥ 1, 𝑷ℭ (𝑧) ℎ (𝑧, 𝜆) − 𝑷ℭ ℎ (𝑧, 𝜆) ≤ 𝐶e−𝜅𝜆 (21.1.10)

[𝑷ℭ ℎ is defined in (21.1.2)].

The Taylor expansion of 𝑝 (𝑧, 𝑤, 𝜁, 𝜆) in the 𝑤 variable, about 𝑤 = 𝑧, is normally
convergent in Ω [uniformly with respect to (𝑧, 𝜁) on compact subsets of Ω × Ξ]; we
derive
−𝑛
𝜆
𝑷ℭ (𝑧) ℎ (𝑧, 𝜆)
2𝜋
∑︁ 1 ∫
= e𝑖𝜆(𝑧−𝑤) ·𝜁 (𝑤 − 𝑧) 𝛼 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁
𝛼∈Z+𝑛 𝛼! ℭ (𝑧)

∑︁ (−𝜆) −| 𝛼 | ∫
= D 𝜁𝛼 e𝑖𝜆(𝑧−𝑤) ·𝜁 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁.
𝛼∈Z𝑛
𝛼! ℭ (𝑧)
+

For fixed 𝛼 ∈ Z+𝑛 the Cauchy Integral Theorem implies that the difference between
∫
(−1) | 𝛼 | D 𝜁𝛼 e𝑖𝜆(𝑧−𝑤) ·𝜁 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁 (21.1.11)
ℭ (𝑧)

and ∫
e𝑖𝜆(𝑧−𝑤) ·𝜁 D 𝜁𝛼 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁 (21.1.12)
ℭ (𝑧)

is a finite linear combination of integrals of the type

∫
𝐽 ( 𝛼,𝛽) (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 D 𝜁 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝜇 𝛽 (𝑧)
𝛽
𝜕ℭ (𝑧)

where 𝛽 𝑗 < 𝛼 𝑗 for each 𝑗 = 1, ..., 𝑛, and d𝜇 𝛽 (𝑧) is the appropriate measure on
the boundary 𝜕ℭ (𝑧) (assumed to be a smooth submanifold of Ω′ × Ξ′). From the
definition of Ψ1 we derive

𝐽 ( 𝛼,𝛽) (𝑧, 𝜆)
∫
e𝜆Ψ1 (𝑧,𝑤,𝜁 ) D 𝜁 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) e−𝜆Φ(𝑤) |ℎ (𝑤, 𝜆)| d𝜇 𝛽 (𝑧) .
𝛽
≤
𝜕ℭ (𝑧)

On 𝜕ℭ (𝑧) we have, by Condition (3) in Definition 20.5.1 and the fact that ℭ (𝑧) is
strongly well-shaped at (𝑤 (𝑧) , 𝜁 (𝑧)) for Ψ1 (𝑧, 𝑤, 𝜁),

Ψ1 (𝑤 (𝑧) , 𝑧, 𝜁 (𝑧)) ≤ Φ (𝑧) − 𝑐 |𝑤 − 𝑤 (𝑧)| 2 + |𝜁 − 𝜁 (𝑧)| 2 (21.1.13)

(𝑐 > 0), whence

876 21 Germ Pseudodifferential Operators in Complex Space

e−𝜆Φ(𝑧) 𝐽 ( 𝛼,𝛽) (𝑧, 𝜆)

∫
2
−𝑐𝜆|𝜁 −𝜁 (𝑧) | 2
e−𝑐𝜆|𝑤−𝑤 (𝑧) | D 𝜁 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) e−𝜆Φ(𝑤) |ℎ (𝑤, 𝜆)| d𝜇 𝛽 (𝑧) .
𝛽
≤
𝜕ℭ (𝑧)

With (𝑤, 𝜁) ∈ 𝜕ℭ (𝑧), 𝑧 ∈ Ω′ and 𝜆 ≥ 1 we can make use of the following properties:
(1) to each 𝜀 > 0 there is a constant 𝐶 𝜀 > 0 independent of (𝑤, 𝜆), such that
e−𝜆Φ(𝑤) |ℎ (𝑤, 𝜆)| ≤ 𝐶 𝜀 e 𝜀𝜆 ;
(2) to each 𝜀 > 0 there is a 𝐶 𝜀′ > 0 such that, for all 𝛼, 𝛽 ∈ Z+𝑛 , 𝜆 ≥ 1,
| 𝛼+𝛽 |+1
D 𝜁 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) ≤ 𝐶 𝜀′
𝛽
sup 𝛼!𝛽!e 𝜀𝜆 . (21.1.14)
(𝑧,𝜁 ) ∈Ω′ ×Ξ′

If 𝜅 and diam Ω′ are sufficiently small, and therefore |𝑤 (𝑧) − 𝑧◦ | and |𝜁 (𝑧) − 𝜁 ◦ |
are also small, we have

𝑐 |𝑤 − 𝑤 (𝑧)| 2 + 𝑐 |𝜁 − 𝜁 (𝑧)| 2 ≥ 𝜅

for all (𝑤, 𝜁) ∈ 𝜕ℭ (𝑧). We end up with an estimate

1 −𝜆Φ(𝑧) ( 𝛼,𝛽) | 𝛼+𝛽 |+1
e 𝐽 (𝑧, 𝜆) ≤ 𝐶 𝜀′′ 𝛽!e−(𝜅−2𝜀)𝜆 .
𝛼!
It suffices to require 𝜀 < 41 𝜅 and Φ (𝑧) < 14 𝜅 if 𝑧 ∈ Ω′ to conclude that the absolute

value of the difference between (21.1.11) and (21.1.12) is of the order of exp − 41 𝜅𝜆 .
Of course, summing over all 𝛼 ∈ Z+𝑛 introduces infinite series, but we can, thanks
to the inequalities (21.1.14), replace them by the finite realizations (cf. Definition
19.1.7)
∑︁ 𝜆− | 𝛼 |
𝑝 (𝑅) (𝑧, 𝜁, 𝜆) = D 𝛼 𝜕 𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) (21.1.15)
𝛼! 𝜁 𝑤
| 𝛼 | ≤𝜆/𝑅

(𝑅 ≫ 1) and
𝑛 ∫
𝜆
𝑷ℭ(𝑅)
(𝑧)
ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝 (𝑅) (𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁.
2𝜋 ℭ (𝑧)

By using the inequalities (21.1.14) and mimicking the proof of Lemma 19.1.9 we
derive from (21.1.10) the following conclusion, under the assumptions that Ω′ ⊂ Ω
is a sufficiently small neighborhood of 𝑧◦ [with ℭ (𝑧) ⊂⊂ Ω′], that 𝑅 is sufficiently
large and 𝜅 ′ > 0 is sufficiently small:
′
∀𝑧 ∈ Ω′, 𝜆 ≥ 1, 𝑷ℭ(𝑅)
(𝑧)
ℎ (𝑧, 𝜆) − 𝑷ℭ ℎ (𝑧, 𝜆) ≲ e−𝜅 𝜆

[𝑷ℭ ℎ defined in (21.1.2)]; and if 𝑅 ′ > 𝑅,

′ ′
∀𝑧 ∈ Ω′, 𝜆 ≥ 1, 𝑷ℭ(𝑅(𝑧)) ℎ (𝑧, 𝜆) − 𝑷ℭ(𝑅)
(𝑧)
ℎ (𝑧, 𝜆) ≲ e−𝜅 𝜆 .
21.1 Germ Pseudodifferential Operators 877

The problem that 𝑷ℭ(𝑅)(𝑧)

ℎ (𝑧, 𝜆) is a smooth, not necessarily holomorphic, function
of 𝑧 is remedied by applying once again Lemma 20.4.1, to deform ℭ (𝑧) back to a
contour ℭ independent of all variables. We end up with an estimate (21.1.10) with
𝑝 (𝑧, 𝑤, 𝜁, 𝜆) replaced by 𝑝 (𝑅) (𝑧, 𝜁, 𝜆) [see (21.1.15)], leading to the conclusion that
𝑛 ∫
(𝑅) 𝜆
𝑷ℭ ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝 (𝑅) (𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁 (21.1.16)
2𝜋 ℭ

belongs to the same coset in O (Φ) (Ω′) /O (−𝜔) (Ω′) as 𝑷ℭ ℎ. We can state

Proposition 21.1.5 If the neighborhood of 𝑧◦ , Ω′ ⊂ Ω, and the contour ℭ ⊂⊂ Ω′ ×Ξ′

are sufficiently small and 𝑅 is sufficiently large then 𝑷ℭ and 𝑷ℭ(𝑅) (followed
by restriction to Ω′) define the same linear map of O (Φ) (Ω) /O (−𝜔) (Ω) into
O (Φ) (Ω′) /O (−𝜔) (Ω′).

As a direct consequence we obtain

Theorem 21.1.6 Let P𝑧 ◦ ,𝜁 ◦ denote the germ pseudodifferential operator (21.1.3)

defined by integrals (21.1.2). Then P𝑧 ◦ ,𝜁 ◦ can also be defined by the analogous
integrals where 𝑝 is replaced by some finite realization of the formal series

−1 D
∑︁ 𝜆− | 𝛼 |
e𝜆 𝜁 𝜕𝑤
𝑝 (𝑧, 𝑤, 𝜁,𝜆) = D 𝜁𝛼 𝜕𝑤𝛼 𝑝 (𝑧, 𝑧, 𝜁, 𝜆) . (21.1.17)
𝑤=𝑧
𝛼∈Z 𝑛 𝛼!
+

Dealing with integrals of the type (21.1.16) allows us to make full use of the
symbolic calculus (cf. Section 16.2), the composition law # [cf. (16.2.16)], the
formula for the transpose, etc.
Instead of a Taylor expansion in the variable 𝑤 about 𝑧 we could have used, in the
integral (21.1.2), a Taylor expansion in the variable 𝑧 about 𝑤:
−𝑛
𝜆
𝑷ℭ ℎ (𝑧, 𝜆)
2𝜋
∑︁ 1 ∫
= e𝑖𝜆(𝑧−𝑤) ·𝜁 (𝑧 − 𝑤) 𝛼 𝜕𝑧𝛼 𝑝 (𝑤, 𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝜁d𝑤
𝛼∈Z+𝑛
𝛼! ℭ

∑︁ 𝜆− | 𝛼 | ∫
= D 𝜁𝛼 e𝑖𝜆(𝑧−𝑤) ·𝜁 𝜕𝑧𝛼 𝑝 (𝑤, 𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝜁d𝑤.
𝛼∈Z𝑛
𝛼! ℭ
+

This would have led us to the conclusion that the germ pseudodifferential operator
P𝑧 ◦ ,𝜁 ◦ can also be defined by the integrals
𝑛 ∫
𝜆
e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝˜ (𝑅) (𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝜁d𝑤 (21.1.18)
2𝜋 ℭ

where
878 21 Germ Pseudodifferential Operators in Complex Space

∑︁ (−1) | 𝛼 | − | 𝛼 | 𝛼 𝛼
𝑝˜ (𝑅) (𝑤, 𝜁, 𝜆) = 𝜆 D 𝜁 𝜕𝑧 𝑝 (𝑤, 𝑤, 𝜁, 𝜆) . (21.1.19)
𝛼!
| 𝛼 | ≤𝜆/𝑅

21.1.3 Germ Fourier transform of holomorphic functions and its

inversion

As before let Φ ∈ C ∞ (Ω; R), Φ (𝑧 ◦ ) = 0, 𝜕𝑧 Φ (𝑧 ◦ ) = 21 𝑖𝜁 ◦ . Now we take 𝜑 (𝑧, 𝜁) =

− (𝑧 − 𝑧◦ ) · 𝜁; the critical points of

𝑧 ↦→ Ψ (𝑧, 𝜁) = Φ (𝑧) + Im ((𝑧 − 𝑧◦ ) · 𝜁)

are defined by the equation 𝜕𝑧 Φ (𝑧) = 21 𝑖𝜁; by hypothesis, 𝑧◦ is a critical point of

Ψ (𝑧, 𝜁 ◦ ). We assume that (𝜑, Φ) is a Sjöstrand pair at (𝑧◦ , 𝜁 ◦ ) ∈ Ω × (C𝑛 \ {0});
this means (cf. Proposition 20.3.20) that 𝑧◦ is a saddle point of Φ (𝑧) which in
turn implies that, in the vicinity of 𝑧◦ , there is a unique C ∞ solution 𝑧 (𝜁) of
𝜕𝑧 Φ (𝑧) = 21 𝑖𝜁 such that 𝑧 (𝜁 ◦ ) = 𝑧◦ . By (20.3.5) Φ∗ (𝜁) = Ψ (𝑧 (𝜁) , 𝜁); it follows
that Φ∗ (𝜁 ◦ ) = Φ (𝑧 ◦ ) = 0. The germ map
∗
F𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O 𝜁(Φ◦ ) /O 𝜁(−𝜔)
◦ (21.1.20)

arrived at, through application of Theorem 20.4.6, starting from the integral
∫
◦
𝑭𝔠 ℎ (𝜁, 𝜆) = e−𝑖𝜆(𝑧−𝑧 ) ·𝜁 ℎ (𝑧, 𝜆) d𝑧, (21.1.21)
𝔠
can be regarded as a germ Fourier transform at (𝑧 ◦ , 𝜁 ◦ ) ∈ Ω × Ξ. For 𝜆 fixed C𝑛 ∋
𝜁 ↦→ 𝑭𝔠 ℎ (𝜁, 𝜆) is an entire function of exponential type; it can be interpreted as the
Borel–Laplace transform of an analytic functional in C𝑛 (Definition 6.2.1) carried by
𝔠¯. The contour 𝔠 is strongly well-shaped at 𝑧◦ for Ψ (𝑧, 𝜁 ◦ ) = Φ (𝑧)+Im ((𝑧 − 𝑧◦ ) · 𝜁 ◦ )
and ℎ ∈ O (Φ) (Ω).
By (20.3.25) and Theorem 20.3.23, we know that ((𝑧 − 𝑧◦ ) · 𝜁, Φ∗ (𝜁)) is a
Sjöstrand pair at (𝜁 ◦ , 𝑧◦ ); we have Φ∗ (𝜁 ◦ ) = 0 and 𝜁 ◦ is a saddle point of
Ψ∗ (𝑧◦ , 𝜁) = Φ∗ (𝜁). Consider the germ operator
∗
F∗𝜁 ◦ ,𝑧 ◦ : O 𝜁(Φ◦ ) /O 𝜁(−𝜔)
◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ , (21.1.22)

arrived at starting from the integral

𝑛 ∫
𝜆 ◦
𝑭𝔠∗∗ 𝑓 (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑧 ) ·𝜁 𝑓 (𝜁, 𝜆) d𝜁 (21.1.23)
2𝜋 𝔠∗

where the contour of integration 𝔠 ∗ (such that 𝜁 ◦ ∈ 𝔠 ∗ ⊂⊂ Ξ) is strongly well-shaped

∗
at 𝜁 ◦ for the function Φ∗ (𝜁) ; 𝑓 ∈ O (Φ ) (Ξ).
21.1 Germ Pseudodifferential Operators 879

The composite F∗𝜁 ◦ ,𝑧 ◦ F𝑧 ◦ ,𝜁 ◦ is defined starting from the double integral

𝑛 ∫ ∫
𝜆
𝑭𝔠∗∗ 𝑭𝔠 ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 ℎ (𝑤, 𝜆) d𝑤d𝜁. (21.1.24)
2𝜋 𝔠∗ 𝔠

We can avail ourselves of Remark 20.4.10: after an affine transformation 𝑧 ↦→

𝑧◦ + 𝑇 (𝑧 − 𝑧 ◦ ), with 𝑇 a linear automorphism of C𝑛 we can assume that
𝑛
1 ∑︁ 2
Φ (𝑧) = Φ∗ (𝜁) − 𝜒 𝑗 𝑥 𝑗 − 𝑥 𝑗 (𝜁) (21.1.25)
2 𝑗=1
𝑛
1 ∑︁ ′ 2
+ 𝜒 𝑗 𝑦 𝑗 − 𝑦 𝑗 (𝜁) + 𝑂 |𝑧 − 𝑧 (𝜁)| 3
2 𝑗=1

and take 𝔠 to be a ball |𝑥 − 𝑥 ◦ | < 𝑟 in the affine subspace of C𝑛 defined by 𝑦 = 𝑦 ◦ ;

𝑧 (𝜁) is the C ∞ solution of the equation 𝜕𝑧 Φ (𝑧) = 21 𝑖𝜁 such that 𝑧 (𝜁 ◦ ) = 𝑧◦ . We can
also avail ourselves of the proof of Theorem 20.3.23, and perform the transformation
𝜁 ↦→ 𝜁 ◦ + 𝑆𝜁 where 𝑆 = (𝑇 ⊤ ) −1 is the contragredient of 𝑇; this done we have
𝑛 𝑛
2 ∑︁ 2
∑︁ 1 1
Φ∗ (𝜁) = 𝜂 𝑗 − 𝜂◦𝑗 − ′ 𝜉 𝑗 − 𝜉 ◦𝑗 + 𝑂 |𝜁 − 𝜁 ◦ | 3 . (21.1.26)
𝜒
𝑗=1 𝑗
𝜒
𝑗=1 𝑗

We take 𝔠 ∗ to be a ball |𝜉 − 𝜉 ◦ | < 𝜌 in the affine subspace of C𝑛 defined by 𝜂 = 𝜂◦ .

There is no loss of generality in assuming that 𝑧 ◦ = 𝜉 ◦ = 0; thus,
𝑛 ∫ ∫
𝜆
𝑭𝔠∗∗ 𝑭𝔠 ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑢) · 𝜉 ℎ (𝑢, 𝜆) d𝑢d𝜉.
2𝜋 | 𝜉 |<𝜌 |𝑢 |<𝑟

We deform the domain of 𝜉-integration from the ball |𝜉 | < 𝜌 to the image of this
ball under the map 𝜉 ↦→ 𝜉 + 𝑖𝜅 | 𝜉𝜉 | (𝑧 − 𝑢), 𝜅 > 0. By the Cauchy Integral Theorem
we obtain
𝑛 ∫ ∫
𝜆 2
∗
𝑭𝔠∗ 𝑭𝔠 ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑢) · 𝜉 −𝜅 ⟨𝑧−𝑢⟩ ℎ (𝑢, 𝜆) d𝑢d𝜉
2𝜋 | 𝜉 |<𝜌 |𝑢 |<𝑟
𝑛 ∫ 1∫ ∫
𝜆 2
+ 𝜌 𝑛−1 e𝑖𝜆(𝑧−𝑢) · 𝜉 −𝜅𝑡 ⟨𝑧−𝑢⟩ ℎ (𝑢, 𝜆) d𝑢d𝜇 (𝜉) d𝑡,
2𝜋 0 | 𝜉 |=𝜌 |𝑢 |<𝑟

where d𝜇 (𝜉) is the appropriate measure on the sphere S𝑛−1 .

We use the notation (20.2.7) with 𝐾 a compact subset of Ω whose interior contains
the set defined by |𝑥| < 𝑟, 𝑦 = 0. Thus, (21.1.25)–(21.1.26) imply
880 21 Germ Pseudodifferential Operators in Complex Space
∫
2
+𝜅𝜆𝜆𝑡 |𝑦 | 2
e𝑖𝜆(𝑧−𝑢) · 𝜉 −𝜅𝜆𝜆𝑡 |𝑥−𝑢 | ℎ (𝑢, 𝜆) d𝑢
|𝑢 |<𝑟
∫
(Φ) −𝜆𝑦· 𝜉 +𝜅𝜆𝑡 |𝑦 | 2
≤ ∥ℎ∥ 𝐾 e e𝜆Φ(𝑢) d𝑢
|𝑢 |<𝑟
∫
(Φ) −𝜆𝑦· 𝜉 +𝜅𝜆 |𝑦 | 2 +𝜆Φ∗ ( 𝜉 ) 2
≤ ∥ℎ∥ 𝐾 e e−𝑐𝜆 |𝑢−𝑥 ( 𝜉 ) | d𝑢
|𝑢 |<𝑟
(Φ) −𝜆𝑦· 𝜉 +𝜅𝜆 |𝑦 | 2 +𝜆Φ∗ ( 𝜉 ) (Φ) −𝜆𝑦· 𝜉 +𝜅𝜆 |𝑦 | 2
−𝑐𝜆 | 𝜉 | 2
≲ ∥ℎ∥ 𝐾 e ≤ ∥ℎ∥ 𝐾 e

(𝑐 > 0 suitably small). We derive

∫ 1∫ ∫
2
e𝑖𝜆(𝑧−𝑢) · 𝜉 −𝜅𝑡 ⟨𝑧−𝑢⟩ ℎ (𝑢, 𝜆) d𝑢d𝜇 (𝜉) d𝑡
0 | 𝜉 |=𝜌 |𝑢 |<𝑟
(Φ) 𝜆 ( 𝜅 |𝑦 | 2 +𝜌 |𝑦 |−𝑐𝜌2 ) 2
+𝜌 |𝑦 |−𝑐𝜌2 )
≲ ∥ℎ∥ 𝐾 e ≤ 𝐶 𝜀 e𝜆 ( 𝜀+𝜅 |𝑦 | ,

the last inequality a consequence of (20.2.7). It suffices to require |𝑦| and 𝜀 > 0 to
be so small that 𝜀 + 𝜅 |𝑦| 2 + 𝜌 |𝑦| < 21 𝑐𝜌 and then let 𝜅 ↘ 0, in order to conclude, in
the notation of (20.2.3), that

𝑭𝔠∗∗ 𝑭𝔠 ℎ Ω′ − 𝑰 ℭ ℎ| Ω′ ∈ O (−𝜔) (Ω′) (21.1.27)

with Ω′ ⊂ Ω a sufficiently small neighborhood of 𝑧◦ (= 0). Combining this with

Corollary 20.2.15, enables us to state
Theorem 21.1.7 Let Φ ∈ C ∞ (Ω; R) be such that Φ (𝑧◦ ) = 0, 𝜕𝑧 Φ (𝑧◦ ) = 12 𝑖𝜁 ◦ ,
and 𝑧◦ ∈ Ω be a saddle point of Ψ (𝑧, 𝜁 ◦ ) = Φ (𝑧) + Im ((𝑧 − 𝑧 ◦ ) · 𝜁 ◦ ). If the
domain Ω′ ⊂ Ω containing 𝑧◦ is sufficiently small then (21.1.27) holds whatever
ℎ ∈ O (Φ) (Ω).
Corollary 21.1.8 Under the hypotheses of Theorem 21.1.7, F∗𝜁 ◦ ,𝑧 ◦ F𝑧 ◦ ,𝜁 ◦ = I𝑧(Φ)
◦ , the

identity operator of O𝑧(Φ) (−𝜔)

◦ /O 𝑧 ◦ .
Thus Theorem 21.1.7 provides an integral representation of the identity different
from that presented in Subsection 20.2.3. In connection with this the following
remark is pertinent.
Remark 21.1.9 The contour of integration defined by (20.2.4) is not strongly well-
shaped at (𝑧 ◦ , 𝜁 ◦ ) for the function Ψ1 (𝑧◦ , 𝑤, 𝜁) [cf. (21.1.1)] whereas the contour 𝔠 ×
𝔠 ∗ in (21.1.24) is, as a direct consequence of (21.1.25) and (21.1.26) [cf. Proposition
21.1.11 below]. By applying Lemma 20.4.1, we can replace the contour 𝔠 by contours
𝔠 (𝜁) strongly well-shaped for 𝑤 ↦→ Ψ1 (𝑧◦ , 𝑤, 𝜁) at 𝑧 (𝜁), the point in (21.1.25). Then
the contour
ℭ = {(𝑤, 𝜁) ∈ Ω × Ξ; 𝑤 ∈ 𝔠 (𝜁) , 𝜁 ∈ 𝔠 ∗ }
is strongly well-shaped at (𝑧◦ , 𝜁 ◦ ) for Ψ1 (𝑧◦ , 𝑤, 𝜁). By applying Proposition 20.2.3,
given any other contour ℭ1 strongly well-shaped at (𝑧 ◦ , 𝜁 ◦ ) for Ψ1 (𝑧◦ , 𝑤, 𝜁), and
selecting sufficiently small neighborhoods Ω′ ⊂ Ω and Ξ′ ⊂ Ξ of 𝑧 ◦ and 𝜁 ◦ we can
deform ℭ ∩ (Ω′ × Ξ′) into ℭ1 ∩ (Ω′ × Ξ′).
21.1 Germ Pseudodifferential Operators 881

Remark 21.1.10 Using F∗𝜁 ◦ ,𝑧 ◦ and taking advantage of Theorem 21.1.6 one can
give another definition of a germ pseudodifferential operator: such an operator is
∗
the composite F∗𝜁 ◦ ,𝑧 ◦ A𝑧 ◦ ,𝜁 ◦ , where A𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O 𝜁(Φ◦ ) /O 𝜁(−𝜔)
◦ is a
Fourier-like transform (Definition 20.4.7) with phase-function −𝑧 · 𝜁 [cf. (21.1.22)–
(21.1.23)].

21.1.4 Composition of germ Fourier-like transforms with opposite

phase-functions

An important special case to consider in Proposition 21.1.3 is that of the phase-

function 𝜑 (𝑧, 𝜃) − 𝜑 (𝑤, 𝜃), with 𝜑 ∈ O (Ω × Θ) satisfying the following conditions
[cf. (20.3.11); here dim Ω = dim Θ = 𝑛]
2
𝜕 𝜑
∀ (𝑧, 𝜃) ∈ Ω × Θ, det (𝑧, 𝜃) ≠ 0. (21.1.28)
𝜕𝑧 𝑗 𝜕𝜃 𝑘 1≤ 𝑗,𝑘 ≤𝑛

As before, let Φ ∈ C ∞ (Ω; R), Φ (𝑧 ◦ ) = 0, 𝜕𝑧 Φ (𝑧◦ ) = 21 𝑖𝜁 ◦ . We assume that

(𝜑, Φ) is a Sjöstrand pair at (𝑧◦ , 𝜃 ◦ ); then (−𝜑, Φ∗ ) is a Sjöstrand pair at (𝜃 ◦ , 𝑧◦ ).
Since 𝑧◦ is a critical point of Ψ (𝑤, 𝜃 ◦ ) = Φ (𝑤) − Im 𝜑 (𝑤, 𝜃 ◦ ) we must have
𝜕𝑧 𝜑 (𝑧◦ , 𝜃 ◦ ) = 2𝑖𝜕𝑧 Φ (𝑧◦ ) = −𝜁 ◦ .
We start from the two integrals (20.1) and (20.4.10). If we replace 𝑔 by 𝑨𝔠 ℎ in
the latter we get the double integral (20.4.12), here
𝑛 ∫ ∫
𝜆
𝑻 𝔠,𝔠∗ ℎ (𝑧, 𝜆) = e𝑖𝜆( 𝜑 (𝑤, 𝜃)−𝜑 (𝑧, 𝜃)) 𝑏 (𝑧, 𝜃, 𝜆) (21.1.29)
2𝜋 𝔠 ∗ 𝔠
× 𝑎 (𝑤, 𝜃, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜃,

where ℎ ∈ O (Φ) (Ω). The contour 𝔠 ⊂⊂ Ω is strongly well-shaped at 𝑧◦ ∈ 𝔠 for

Ψ (𝑤, 𝜃 ◦ ) = Φ (𝑤) − Im 𝜑 (𝑤, 𝜃 ◦ ); 𝔠 ∗ ⊂⊂ Θ is strongly well-shaped at 𝜃 ◦ ∈ 𝔠 ∗ for
Ψ∗ (𝑧◦ , 𝜃) = Φ∗ (𝜃) + Im 𝜑 (𝑧◦ , 𝜃) [Φ∗ (𝜃) is the critical value of 𝑤 ↦→ Ψ (𝑤, 𝜃)].
The integral (20.1) leads to the definition of the germ Fourier-like transform
(20.4.4), which we reproduce here,
∗
A𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ , (21.1.30)

whereas (20.4.10) leads to the definition of the germ operator (20.4.9),

∗
B 𝜃 ◦ ,𝑧 ◦ : O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ . (21.1.31)

Naturally, (21.1.29) leads to the composite (20.4.11), B 𝜃 ◦ ,𝑧 ◦ A𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)

◦ /O 𝑧 ◦ ←↪.
A priori, (21.1.29) does not have the canonical form (20.5.7). We show how to
transition from the former to the latter without affecting the germ operators.
882 21 Germ Pseudodifferential Operators in Complex Space

By hypothesis 𝑧 ◦ is a saddle-point of 𝑤 ↦→ Ψ (𝑤, 𝜃 ◦ ) = Φ (𝑤) − Im 𝜑 (𝑤, 𝜃 ◦ )

while 𝜃 ◦ is a saddle-point of 𝜃 ↦→ Ψ∗ (𝑧◦ , 𝜃) = Φ∗ (𝜃) + Im 𝜑 (𝑧◦ , 𝜃). We apply
Lemma 20.4.1: to each 𝜃 ∈ Θ (resp., 𝑧 ∈ Ω) we attach a contour of integration
𝔠 (𝜃) ⊂⊂ Ω [resp., 𝔠 ∗ (𝑧) ⊂⊂ Θ] depending smoothly on 𝜃 ∈ Θ (resp., 𝑧 ∈ Ω),
such that 𝔠 (𝜃 ◦ ) = 𝔠 [resp., 𝔠 ∗ (𝑧 ◦ ) = 𝔠 ∗ ] and, moreover, such that, for some positive
constants 𝛾◦ , 𝛾1 and all 𝑧 ∈ Ω, 𝜃 ∈ 𝔠 ∗ (𝑧), 𝑤 ∈ 𝔠 (𝜃), we have

Ψ (𝑤, 𝜃) ≤ Ψ (𝑤 (𝜃) , 𝜃) − 𝛾◦ |𝑤 − 𝑤 (𝜃)| 2 ,

Ψ∗ (𝑧, 𝜃) ≤ Ψ∗ (𝑧, 𝜃 (𝑧)) − 𝛾1 |𝜃 − 𝜃 (𝑧)| 2 ,

where 𝑤 (𝜃) is the saddle point of the function Ω ∋ 𝑤 ↦→ Ψ (𝑤, 𝜃) and 𝜃 (𝑧) is the
saddle point of the function Θ ∋ 𝜃 ↦→ Ψ∗ (𝑧, 𝜃). Recalling (20.3.5) and (20.3.14) we
get

Ψ (𝑤, 𝜃) ≤ Φ∗ (𝜃) − 𝛾◦ |𝑤 − 𝑤 (𝜃)| 2 , (21.1.32)

∗ 2
Ψ (𝑧, 𝜃) ≤ Φ (𝑧) − 𝛾1 |𝜃 − 𝜃 (𝑧)| . (21.1.33)

We define the following “composite” contour of integration in (𝑤, 𝜃)-space:

ℭ (𝑧) = {(𝑤, 𝜃) ∈ Ω × Θ; 𝑤 ∈ 𝔠 (𝜃) , 𝜃 ∈ 𝔠 ∗ (𝑧)} . (21.1.34)

Proposition 21.1.11 The contour of integration ℭ (𝑧) is strongly well-shaped at

(𝑤 (𝑧) , 𝜃 (𝑧)) ∈ Ω × Θ for the function

(𝑤, 𝜃) ↦→ Ψ (𝑤, 𝜃) − Ψ (𝑧, 𝜃) .

Proof By subtracting Im 𝜑 (𝑧, 𝜃) from both sides of (21.1.33) we obtain

Φ∗ (𝜃) ≤ Ψ (𝑧, 𝜃) − 𝛾1 |𝜃 − 𝜃 (𝑧)| 2 ;

putting this into (21.1.32) yields

Ψ (𝑤, 𝜃) − Ψ (𝑧, 𝜃) ≤ −𝛾◦ |𝑤 − 𝑤 (𝜃)| 2 − 𝛾1 |𝜃 − 𝜃 (𝑧)| 2 . □

The argument ends with repeated applications of Lemma 20.4.1:

(1) In the integral (21.1.29) we deform the contour 𝔠 ∗ to the contour 𝔠 ∗ (𝑧).
(2) This done, we deform the contour 𝔠 to the contour 𝔠 (𝜃) for each 𝜃 ∈ 𝔠 ∗ (𝑧).
After Steps (1) and (2) the integration is carried out over the “composite” contour
(21.1.34); we have
∫
e−𝜆Φ(𝑧) 𝑻 𝔠,𝔠∗ ℎ (𝑧, 𝜆) ≤ e𝜆(Ψ(𝑤, 𝜃)−Ψ(𝑧, 𝜃)) 𝑏 (𝑧, 𝜃, 𝜆) 𝑎 (𝑤, 𝜃, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜃.
ℭ (𝑧)

Next we deform the contour ℭ (𝑧) to ℭ (𝑧◦ ) = ℭ. Hypothesis (21.1.28) combined with
the hypothesis 𝜕𝑧 Φ (𝑧◦ ) = 21 𝑖𝜁 ◦ allows us to apply Proposition 21.1.3 to 𝜑 (𝑤, 𝜃) −
𝜑 (𝑧, 𝜃). We can state:
21.1 Germ Pseudodifferential Operators 883

Theorem 21.1.12 Assume that 𝜑 ∈ O (Ω × Θ) satisfies (21.1.28) and 𝜕𝑧 Φ (𝑧 ◦ ) =

1 ◦
2 𝑖𝜁 . Under these hypotheses the composite B 𝜃 ,𝑧 A 𝑧 , 𝜃 of (21.1.30) and (21.1.31)
◦ ◦ ◦ ◦
(Φ) (−𝜔)
is a germ pseudodifferential operator O𝑧 ◦ /O𝑧 ◦ ←↪.

Let us return to (21.1.29). We put 𝜑 (𝑤, 𝜃) − 𝜑 (𝑧, 𝜃) = (𝑧 − 𝑤) · 𝜙1 (𝑤, 𝑧, 𝜃); keep

in mind that 𝜙1 (𝑧, 𝑧, 𝜃) = −𝜕𝑧 𝜑 (𝑧, 𝜃). The hypothesis (21.1.28) ensure that the map
𝜃 ↦→ 𝜁 = 𝜁 ◦ + 𝜙1 (𝑤, 𝑧, 𝜃) is a local biholomorphism from Θ onto a neighborhood
of 𝜁 ◦ in C𝑛 whose inverse we denote by 𝜁 ↦→ 𝜃 (𝑧, 𝑤, 𝜁). It transforms (21.1.29) into
𝑛 ∫ ∫
𝜆
e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑐 (𝑧, 𝑤, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁, (21.1.35)
2𝜋 ∗
𝔠 𝔠 (𝑤,𝑧)

where 𝔠 ∗ (𝑤, 𝑧) is the image of 𝔠 ∗ under the map 𝜃 ↦→ 𝜁 and

D𝜃
𝑐 (𝑧, 𝑤, 𝜁, 𝜆) = 𝑎 (𝑤, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆) 𝑏 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆) (𝑧, 𝑤, 𝜁) . (21.1.36)
D𝜁
We can use the standard deformations of the contours (cf. Lemma 20.4.1) and move
the integration with respect to 𝜁 to a contour independent of (𝑤, 𝑧) at the price of an
error that belongs to O (−𝜔) (Ω′) (Ω′ ⊂ Ω a small open set, 𝑧 ◦ ∈ Ω′). We can also
carry out the reduction of the integrals described in the precedent subsection and
replace 𝑐 (𝑤, 𝑧, 𝜁) by

−1 D
∑︁ 𝜆−| 𝛼 |
e𝜆 𝜁 𝜕𝑤
𝑐 (𝑧, 𝑤, 𝜁,𝜆) = D 𝜁𝛼 𝜕𝑤𝛼 𝑐 (𝑧, 𝑧, 𝜁, 𝜆) . (21.1.37)
𝑤=𝑧
𝛼∈Z 𝑛 𝛼!
+

We remark that, since 𝜕𝜃 𝜁 | 𝑤=𝑧 = −𝜕𝑧 𝜕𝜃 𝜑 (𝑧, 𝜃), by (21.1.28) a nonsingular 2𝑛 × 2𝑛

complex matrix, we have
2 −1
D𝜃 𝜕 𝜑
(𝑧, 𝑧, 𝜁) = det (𝑧, 𝜃 (𝑧, 𝑧, 𝜁)) . (21.1.38)
D𝜁 𝜕𝑧 𝑗 𝜕𝜃 𝑘 1≤ 𝑗,𝑘 ≤𝑛

The principal symbol of B 𝜃 ◦ ,𝑧 ◦ A𝑧 ◦ , 𝜃 ◦ is easy to derive by the analogue of Formula

(21.2.17) from the phase-function and from the principal symbols of A𝑧 ◦ , 𝜃 ◦ and
B 𝜃 ◦ ,𝑧 ◦ . We shall give the formula in the more general case of a classical elliptic germ
FIO instead of a germ Fourier-like transform.

21.1.5 Composition of germ FIOs with opposite phase-functions

We prove the analogue of Theorem 21.1.12 for composite germ FIOs with “opposite”
phase-functions (cf. Subsection 18.5.2). Let (𝜑, Φ1 , Φ2 ) be a Sjöstrand triad at
(𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) in (Ω1 , Ω2 , Θ); recall that Φ1 (𝑧◦ ) = Φ2 (𝑤 ◦ ) = 0. We shall reason
under the hypotheses that dim Ω1 = dim Ω2 = dim Θ = 𝑛. Given a Sjöstrand triad
(𝜑, Φ1 , Φ2 ) at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 in Ω1 × Ω2 × Θ (Definition 20.5.1) we start from
884 21 Germ Pseudodifferential Operators in Complex Space

integrals
𝑛 ∫
𝜆
𝑨ℭ2 ℎ2 (𝑧, 𝜆) = e−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃, (21.1.39)
2𝜋 ℭ2
𝑛 ∫
𝜆
𝑩ℭ1 ℎ1 (𝑤, 𝜆) = e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ1 (𝑧, 𝜆) d𝑧d𝜃, (21.1.40)
1𝜋 ℭ1

where ℎ 𝑗 ∈ O ( Φ 𝑗 ) Ω 𝑗 and the contour ℭ 𝑗 ⊂⊂ Ω 𝑗 ×Θ is strongly well-shaped for the

function Φ 𝑗 ( 𝑗 = 1, 2) at the saddle point which is the projection of (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ). Note
that 𝜕𝑧 Φ1 (𝑧◦ ) = 21 𝑖𝜕𝑧 𝜑 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ); likewise, 𝜕𝑤 Φ2 (𝑤 ◦ ) = − 21 𝑖𝜕𝑤 𝜑 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ).
Out of (21.1.39)–(21.1.40) we construct germ FIOs [cf. (20.5.9) and (20.5.10)]
(Φ2 ) (−𝜔)
A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) : O𝑤 ◦ /O 𝑤 ◦ −→ O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ , (21.1.41)
B𝑤 ◦ ,𝑧 ◦ (𝜃 ◦ ) : O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ −→ (Φ2 )
O𝑤 (−𝜔)
◦ /O 𝑤 ◦ . (21.1.42)

Theorem 21.1.13 Assume that the phase-function 𝜑 ∈ O (Ω1 , Ω2 , Θ) satisfies Con-

dition (20.5.14) and let (𝜑, Φ1 , Φ2 ) be a Sjöstrand triad at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 .
The composite of the germ operators (21.1.41) and (21.1.42) is a germ pseu-
dodifferential operator A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) B𝑤 ◦ ,𝑧 ◦ (𝜃 ◦ ) : O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ ←↪ at (𝑧◦ , 𝜁 ◦ ),
𝜁 ◦ = 𝜕𝑧 𝜑 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ).

Proof We take advantage of the fact that (𝜑 (𝑧, 𝑤, 𝜃) , Φ2 (𝑤)) is a Sjöstrand pair
at (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) in (Ω2 × Θ) × Ω1 [with (𝑤, 𝜃) the “base variable” and 𝑧 the “fiber
variable”]: we apply Lemma 20.4.1 to the integral (21.1.39) and deform ℭ2 to a
smooth curve of contours ℭ2 (𝑧) ⊂⊂ Ω2 × Θ strongly well-shaped for the function
(𝑤, 𝜃) ↦→ Ψ2 (𝑧, 𝑤, 𝜃) at its saddle point (𝑤 (𝑧) , 𝜃 (𝑧)):
𝑛 ∫
𝜆
𝑨ℭ2 ℎ2 (𝑧, 𝜆) e−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃
2𝜋 ℭ2 (𝑧)

with meaning congruent mod O (−𝜔) (Ω1 ). Likewise, we deform ℭ1 in the integral
(21.1.40) to a curve of contours ℭ1 (𝑤) ⊂⊂ Ω1 × Θ strongly well-shaped for the
function (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤, 𝜃) at its saddle point (𝑧 (𝑤) , 𝜃 (𝑤)):
𝑛 ∫
𝜆
𝑩ℭ1 ℎ1 (𝑤, 𝜆) e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ1 (𝑧, 𝜆) d𝑧d𝜃,
2𝜋 ℭ1 (𝑤)

where ℎ1 ∈ O (Φ1 ) (Ω1 ) [cf. (20.5.8)] and means congruent mod O (−𝜔) Ω2′ ,

Ω2′ ⊂ Ω2 a suitably small neighborhood of 𝑧 ◦ . The composite germ operator
A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) B𝑤 ◦ ,𝑧 ◦ (𝜃 ◦ ) originates with the composite integral operator

𝑨ℭ2 𝑩ℭ1 ℎ2 (𝑧, 𝜆)

2𝑛 ∫ ∫
𝜆 ˜
e𝑖𝜆 ( 𝜑 ( 𝑧˜ ,𝑤, 𝜃 ) −𝜑 (𝑧,𝑤, 𝜃) ) 𝑐 𝑧, 𝑧˜, 𝑤, 𝜃, 𝜃,
˜ 𝜆 ℎ2 ( 𝑧˜, 𝜆) d𝑤d𝜃d𝑧˜d𝜃,
˜

2𝜋 ℭ2 (𝑧) ℭ1 (𝑤)
21.1 Germ Pseudodifferential Operators 885

where
˜ 𝜆 = 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) 𝑏 𝑧˜, 𝑤, 𝜃,
˜ 𝜆

𝑐 𝑧, 𝑧˜, 𝑤, 𝜃, 𝜃, (21.1.43)
is a bona fide analytic symbol in 𝑧, 𝑧˜, 𝑤, 𝜃, 𝜃˜ -space. We derive

−2𝑛
𝜆
e−Φ1 (𝑧) 𝑩ℭ2 𝑨ℭ1 ℎ1 (𝑧, 𝜆)
2𝜋
∫ ∫
˜
≤ e𝜆 ( Ψ1 ( 𝑧˜ ,𝑤, 𝜃 ) −Ψ1 (𝑧,𝑤, 𝜃) )
ℭ2 (𝑧) ℭ1 (𝑤)

× e−𝜆Φ1 ( 𝑧˜ ) 𝑐 𝑧, 𝑧˜, 𝑤, 𝜃, 𝜃,
˜ 𝜆 ℎ1 ( 𝑧˜, 𝜆) d𝑧˜d𝜃d𝑤d𝜃
˜

for all 𝑧 ∈ Ω1′ . By Definition 20.5.1, the critical value of (𝑧, 𝜃) ↦→ Ψ1 (𝑧, 𝑤, 𝜃) is
equal to Φ2 (𝑤) in a neighborhood of (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ); it follows that we have, in ℭ1 (𝑤),

2
Ψ1 𝑧˜, 𝑤, 𝜃˜ ≤ Φ2 (𝑤) − 𝛾◦ | 𝑧˜ − 𝑧˜ (𝑤)| 2 + 𝜃˜ − 𝜃˜ (𝑤) .

(21.1.44)

By Proposition 20.5.3, the critical value of (𝑧, 𝜃) ↦→ Ψ2 (𝑧, 𝑤, 𝜃) is equal to Φ1 (𝑧);

it follows that we have, in ℭ2 (𝑧),

Ψ2 (𝑧, 𝑤, 𝜃) ≤ Φ1 (𝑧) − 𝛾1 |𝑤 − 𝑤 (𝑧)| 2 + |𝜃 − 𝜃 (𝑧)| 2 . (21.1.45)

We form the composite contour

˜ 𝑤, 𝜃 ∈ C4𝑛 ; 𝑧˜, 𝜃˜ ∈ ℭ1 (𝑤) , (𝑤, 𝜃) ∈ ℭ2 (𝑧) .

ℭ (𝑧) = 𝑧˜, 𝜃,

We generalize Proposition 21.1.11: the contour of integration ℭ (𝑧) is strongly well-

shaped at

𝑧˜ (𝑤 (𝑧)) , 𝑤 (𝑧) , 𝜃˜ (𝑤 (𝑧)) , 𝜃 (𝑧) ∈ Ω1 × Ω2 × Θ × Θ

for the function

˜ 𝑤, 𝜃 →↦ Ψ1 𝑧˜, 𝑤, 𝜃˜ − Ψ1 (𝑧, 𝑤, 𝜃) .

𝑧˜, 𝜃,
Proof of claim: By adding Im 𝜑 (𝑧, 𝑤, 𝜃) to both sides of (21.1.45) we obtain

Φ2 (𝑤) ≤ Ψ1 (𝑧, 𝑤, 𝜃) − 𝛾1 |𝑤 − 𝑤 (𝑧)| 2 + |𝜃 − 𝜃 (𝑧)| 2 .

Putting this into (21.1.44) yields, for some 𝐶 > 0 independent of 𝛾◦ , 𝛾1 ,

2
Ψ1 𝑧˜, 𝑤, 𝜃˜ ≤ Ψ1 (𝑧, 𝑤, 𝜃) − 𝛾◦ | 𝑧˜ − 𝑧˜ (𝑤)| 2 + 𝜃˜ − 𝜃˜ (𝑤)


− 𝛾1 |𝑤 − 𝑤 (𝑧)| 2 + |𝜃 − 𝜃 (𝑧)| 2
886 21 Germ Pseudodifferential Operators in Complex Space

2
≤ Ψ1 (𝑧, 𝑤, 𝜃) − 𝛾◦ | 𝑧˜ − 𝑧˜ (𝑤 (𝑧))| 2 + 𝜃˜ − 𝜃˜ (𝑤 (𝑧))

− (𝛾1 − 𝐶𝛾◦ ) |𝑤 − 𝑤 (𝑧)| 2 + |𝜃 − 𝜃 (𝑧)| 2 .

We can take 𝛾◦ so small that 𝐶𝛾◦ < 21 𝛾1 . □

◦
After deforming the contour ℭ (𝑧) to ℭ = ℭ (𝑧 ) we end up with the integral

( 𝑨𝑩) ℭ ℎ (𝑧, 𝜆) (21.1.46)

2𝑛 ∫
𝜆 ˜
e𝑖𝜆 ( 𝜑 ( 𝑧˜ ,𝑤, 𝜃 ) −𝜑 (𝑧,𝑤, 𝜃) ) 𝑐 𝑧, 𝑧˜, 𝑤, 𝜃, 𝜃,
˜ 𝜆 ℎ ( 𝑧˜, 𝜆) d𝑧˜d𝜃d𝑤d𝜃
˜

=
2𝜋 ℭ

[ℎ ∈ O (Φ1 ) (Ω1 )] that leads to the definition of the composite A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) B𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ )

and is simpler to handle if we want to show that A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) B𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) is a germ
pseudodifferential operator O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ ←↪.
Taylor expansion yields

𝜑 𝑧˜, 𝑤, 𝜃˜ − 𝜑 (𝑧, 𝑤, 𝜃) = 𝜑 𝑧 (𝑧, 𝑤, 𝜃) · ( 𝑧˜ − 𝑧) + 𝜑 𝜃 (𝑧, 𝑤, 𝜃) · 𝜃˜ − 𝜃

1 1 ˜
+ ( 𝑧˜ − 𝑧) · 𝜕𝑧2 𝜑 (𝑧, 𝑤, 𝜃) ( 𝑧˜ − 𝑧) + 𝜃 − 𝜃 · 𝜕𝜃2 𝜑 (𝑧, 𝑤, 𝜃) 𝜃˜ − 𝜃

2 2
3
+ 𝜃˜ − 𝜃 · 𝜕𝑧 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) ( 𝑧˜ − 𝑧) + 𝑂 | 𝑧˜ − 𝑧| 3 + 𝜃˜ − 𝜃 ,

whence
𝜑 𝑧˜, 𝑤, 𝜃˜ − 𝜑 (𝑧, 𝑤, 𝜃) = 𝜁 · ( 𝑧˜ − 𝑧) + 𝑡 · 𝜏

where 𝜏 = 𝜃˜ − 𝜃 and
1
𝜁 = 𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃) + 𝜕𝑧2 𝜑 (𝑧, 𝑤, 𝜃) ( 𝑧˜ − 𝑧) (21.1.47)
2
1
+ 𝜕𝑧 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) 𝜏 + 𝑂 | 𝑧˜ − 𝑧| 2 + |𝜏| 2 ,
2
1
𝑡 = 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) + 𝜕𝜃2 𝜑 (𝑧, 𝑤, 𝜃) 𝜏
2
1
+ 𝜕𝑧 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) ( 𝑧˜ − 𝑧) + 𝑂 | 𝑧˜ − 𝑧| 2 + |𝜏| 2 .
2
If we regard 𝑧, 𝑧˜, 𝑤, 𝜃, 𝜏 as independent variables we get, when 𝑧˜ = 𝑧 and 𝜏 = 0,

𝜕𝑤 𝜁 = 𝜕𝑧 𝜕𝑤 𝜑 (𝑧, 𝑤, 𝜃) , 𝜕𝜃 𝜁 = 𝜕𝑧 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) ,
𝜕𝑤 𝑡 = 𝜕𝑤 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) , 𝜕𝜃 𝑡 = 𝜕𝜃2 𝜑 (𝑧, 𝑤, 𝜃) ,

and therefore, by (20.5.14) we have, at the point (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ),

𝐷 (𝜁, 𝑡)
= Δ𝜑 ≠ 0 (21.1.48)
𝐷 (𝑤, 𝜃)
21.1 Germ Pseudodifferential Operators 887

where
𝜕𝑧 𝜕𝑤 𝜑 𝜕𝑧 𝜕𝜃 𝜑
Δ 𝜑 = det ,
𝜕𝑤 𝜕𝜃 𝜑 𝜕𝜃2 𝜑
a notation we shall continue to use in the sequel. As a consequence of (21.1.48)
and the Implicit Function Theorem, the system of equations (21.1.47) admits unique
holomorphic solutions 𝑤 = 𝑓 (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏), 𝜃 = 𝑔 (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏) in a neighborhood of
(𝑧◦ , 𝑧◦ , 0, 𝜁 ◦ , 0) [with 𝜁 ◦ = 𝜕𝑧 𝜑 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ )] such that

𝑓 (𝑧◦ , 𝑧◦ , 0, 𝜁 ◦ , 0) = 𝑤 ◦ , 𝑔 (𝑧 ◦ , 𝑧◦ , 0, 𝜁 ◦ , 0) = 0.

In (21.1.46) we carry out the changes of variables, first 𝜃˜ ⇝ 𝜏 + 𝜃 and next

( 𝑧˜, 𝑤, 𝜃, 𝜏) ⇝ ( 𝑧˜, 𝑡, 𝜁, 𝜏), and we deform the contour of integration (cf. Subsec-
tion 20.4.3) to a suitably small contour ℭ̃ strongly well-shaped at (𝑧◦ , 0, 𝜁 ◦ , 0) in
( 𝑧˜, 𝑡, 𝜁, 𝜏)-space for the appropriate function (see the proof of Theorem 21.1.13 and
preceding subsection). We see that, if Ω1′ ∋ 𝑧◦ is sufficiently small, the restriction of
( 𝑨𝑩) ℭ ℎ (𝑧, 𝜆) to Ω1′ is congruent mod O (−𝜔) Ω1′ to

2𝑛 ∫
𝜆
e−𝑖𝜆𝜁 · (𝑧−𝑧˜ )+𝑖𝜆𝑡 · 𝜏 𝑐♭ (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏, 𝜆) (21.1.49)
2𝜋 ℭ̃
×ℎ ( 𝑧˜, 𝜆) d𝑧˜d𝜁d𝑡d𝜏,

where

𝑐♭ (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏, 𝜆) (21.1.50)

𝐷 ( 𝑓 , 𝑔)
= 𝑐 (𝑧, 𝑧˜, 𝑤, 𝜃, 𝜏 + 𝜃, 𝜆) (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏)
𝐷 (𝑡, 𝜁)

with 𝑤 = 𝑓 (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏), 𝜃 = 𝑔 (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏). It follows from (21.1.50) and Def-
inition 21.1.4 that the integral operator ( 𝑨𝑩) ℭ leads to the definition of a germ
pseudodifferential operator O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ ←↪. □
We can go one step further beyond (21.1.49)–(21.1.50), first by deforming the
contour ℭ̃ to a product 𝔠 ×𝔠 ′ with 𝔠 a contour in ( 𝑧˜, 𝜁)-space and 𝔠 ′ one in (𝑡, 𝜏)-space,
then define
𝑛 ∫
𝜆
♮
𝑐 (𝑧, 𝑧˜, 𝜁, 𝜆) = e−𝑖𝜆𝑡 · 𝜏 𝑐♭ (𝑧, 𝑧˜, 𝑡, 𝜁, 𝜏, 𝜆) d𝑡d𝜏.
2𝜋 𝔠′

If we apply Theorem 21.1.7 and the observation that precedes its statement we see
that
𝑐 ♮ (𝑧, 𝑧˜, 𝜁, 𝜆) − 𝑐♭ (𝑧, 𝑧˜, 0, 𝜁, 0, 𝜆) ∈ O (−𝜔) Ω1′ × Ω1′ × Ξ

provided Ω1′ and Ξ are sufficiently small neighborhoods of 𝑧 ◦ and 𝜁 ◦ respectively.

Lastly, we can apply Theorem 21.1.6 and replace 𝑐 ♮ (𝑧, 𝑧˜, 𝜁, 𝜆) by
−1 D
𝑐 ♯ (𝑧, 𝜁, 𝜆) = e𝜆 𝜁 𝜕𝑧˜
𝑐♭ (𝑧, 𝑧˜, 0, 𝜁, 0, 𝜆) .
𝑧˜ =𝑧
888 21 Germ Pseudodifferential Operators in Complex Space

Taking (21.1.43) and (21.1.50) into account this reads

−1 D 𝐷 (𝑤, 𝜃)
𝑐 ♯ (𝑧, 𝜁, 𝜆) = e𝜆 𝜁 𝜕𝑧˜
𝑎 ( 𝑧˜, 𝑤, 𝜃, 𝜆) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) (𝑧, 𝑧˜, 0, 𝜁, 0)
𝐷 (𝑡, 𝜁) 𝑧˜ =𝑧
(21.1.51)
where 𝑤 = 𝑓 (𝑧, 𝑧˜, 0, 𝜁, 0), 𝜃 = 𝑔 (𝑧, 𝑧˜, 0, 𝜁, 0). It remains to prove that (21.1.51) is
an analytic symbol, but this is a routine matter. We end up with the congruence
𝑛 ∫
𝜆
( 𝑨𝑩) ℭ̃ ℎ (𝑧, 𝜆) e𝑖𝜆𝜁 · (𝑧−𝑧˜ ) 𝑐 ♯ (𝑧, 𝜁, 𝜆) ℎ ( 𝑧˜, 𝜆) d𝑧˜d𝜁. (21.1.52)
2𝜋 𝔠

For application in the next section we shall take a closer look at the relation
between the amplitudes 𝑎 and 𝑏 in (20.5.7) and (20.5.8) and the formal series
amplitude (21.1.51). Actually we limit our attention to 𝑐 ♮ (𝑧, 𝑧, 𝜁, 𝜆) (pushing the
analysis beyond this is too complicated). Backtracking to (21.1.50) we get

𝑐♭ (𝑧, 𝑧, 0, 𝜁, 0, 𝜆)
= 𝑐 (𝑧, 𝑧, 𝑓 (𝑧, 𝑧, 0, 𝜁, 0) , 𝑔 (𝑧, 𝑧, 0, 𝜁, 0) , 𝑔 (𝑧, 𝑧, 0, 𝜁, 0) , 𝜆)
𝐷 ( 𝑓 , 𝑔)
× (𝑧, 𝑧, 0, 𝜁, 0) .
𝐷 (𝑡, 𝜁)
Let us use the notation

𝐹 (𝑧, 𝑡, 𝜁) = 𝑓 (𝑧, 𝑧, 𝑡, 𝜁, 0) ,
𝐺 (𝑧, 𝑡, 𝜁) = 𝑔 (𝑧, 𝑧, 𝑡, 𝜁, 0) ,

whence
𝐷 ( 𝑓 , 𝑔) 𝐷 (𝐹, 𝐺)
(𝑧, 𝑧, 𝑡, 𝜁, 0) = (𝑧, 𝑡, 𝜁)
𝐷 (𝑡, 𝜁) 𝐷 (𝑡, 𝜁)
and

𝑐♭ (𝑧, 𝑧, 0, 𝜁, 0, 𝜆) (21.1.53)
𝐷 (𝐹, 𝐺)
= 𝑐 (𝑧, 𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁) , 𝜆) (𝑧, 0, 𝜁) .
𝐷 (𝑡, 𝜁)

Putting 𝑧˜ = 𝑧, 𝜏 = 0, 𝜃˜ = 𝐺 (𝑧, 𝑡, 𝜁) = 𝜃 in (21.1.47) reduces these equations to

𝜕𝑧 𝜑 (𝑧, 𝐹 (𝑧, 𝑡, 𝜁) , 𝐺 (𝑧, 𝑡, 𝜁)) = 𝜁,

𝜕𝜃 𝜑 (𝑧, 𝐹 (𝑧, 𝑡, 𝜁) , 𝐺 (𝑧, 𝑡, 𝜁)) = 𝑡.

Regarding (𝑧, 𝑡, 𝜁) as independent variables we derive

−1
𝜕𝜁 𝐹 𝜕𝑡 𝐹 𝜕𝑧 𝜕𝑤 𝜑 𝜕𝑧 𝜕𝜃 𝜑
= ,
𝜕𝜁 𝐺 𝜕𝑡 𝐺 𝜕𝑤 𝜕𝜃 𝜑 𝜕𝜃2 𝜑
21.1 Germ Pseudodifferential Operators 889

whence D(𝐹,𝐺) −1
D(𝜁 ,𝑡) = Δ 𝜑 [cf. (21.1.48)].
In terms of the diagram (20.5.13) and the definition (20.5.12) the submanifold
defined by 𝑡 = 0 is Σ 𝜑 . Thus (𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁)) ∈ Σ 𝜑 ; 𝑝 𝜑 is the map

(𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁)) ↦→ (𝑧, 𝜁)

and 𝑞 𝜑 is the map

(𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁)) ↦→ (𝐹 (𝑧, 0, 𝜁) , −𝜕𝑤 𝜑 (𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁)))

while 𝜒 𝜑 is the map

(𝑧, 𝜁) ↦→ (𝐹 (𝑧, 0, 𝜁) , −𝜕𝑤 𝜑 (𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁))) .

Recalling that 𝑐 (𝑧, 𝑧, 𝑤, 𝜃, 𝜃, 𝜆) = 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) [cf. (21.1.43)] we see

that

𝑐♭ (𝑧, 𝑧, 0, 𝜁, 0, 𝜆)
𝐷 (𝐹, 𝐺) 𝑎𝑏
= (𝑎𝑏) (𝑧, 𝐹 (𝑧, 0, 𝜁) , 𝐺 (𝑧, 0, 𝜁) , 𝜆) (𝑧, 0, 𝜁) = .
𝐷 (𝑡, 𝜁) Δ𝜑 Σ𝜑

In (21.1.52) we have
𝑎𝑏
𝑐 ♯ (𝑧, 𝜁, 𝜆) = + 𝑂 𝜆−1 (21.1.54)
Δ𝜑 Σ𝜑

where we use the 𝑧 𝑗 and 𝜁 𝑘 as coordinates on Σ 𝜑 .

21.1.6 Pseudodifferential operators and differential operators of

infinite order

Let Ω (resp., Ξ) be an open subset of C𝑛 (esp., C𝑛 \ {0}) and (𝑧◦ , 𝜁 ◦ ) ∈ Ω × Ξ; let

𝑝 (𝑧, 𝜁, 𝜆) be an analytic symbol in Ω × Ξ (Definition 20.1.1). Let Φ ∈ C ∞ (Ω; R)
be such that Φ (𝑧◦ ) = 0, 𝜕𝑧 Φ (𝑧 ◦ ) = 21 𝑖𝜁 ◦ . We define

Ψ (𝑤, 𝜁) = Φ (𝑤) − Im (𝑧 ◦ − 𝑤) · 𝜁 = Ψ1 (𝑧◦ , 𝑤, 𝜁)

[cf. (20.5.1)]. In this subsection we shall assume that 𝑧◦ is a saddle point of Ψ (𝑤, 𝜁 ◦ );
since 𝑧◦ is a critical point of Ψ (𝑤, 𝜁 ◦ ) our hypothesis is that the Hessian matrix of
Φ (𝑤) at 𝑧◦ is nonsingular and its signature is equal to zero. Under this hypothesis,
((𝑧 − 𝑤) · 𝜁, Φ (𝑤)) is a Sjöstrand pair at all points 𝑧 near 𝑧 ◦ . The critical points
of 𝑤 ↦→ Ψ (𝑧◦ , 𝑤, 𝜁) near 𝑤 = 𝑧◦ are the solutions (uniquely defined) 𝑤 (𝜁) of
890 21 Germ Pseudodifferential Operators in Complex Space

the equation 𝜁 = −2𝑖𝜕𝑤 Φ (𝑤). We define Φ∗ (𝜁) = Ψ (𝑧 ◦ , 𝑤 (𝜁) , 𝜁) whence [cf.

(20.3.13)] Ψ∗ (𝑤, 𝜁) = Φ∗ (𝜁) + Im (𝑧 ◦ − 𝑤) · 𝜁, implying that the critical point of
𝜁 ↦→ Ψ∗ (𝑧◦ , 𝜁) is the same as that of Φ∗ (𝜁), namely 𝜁 ◦ .
Consider then an integral of the type
𝑛 ∫ ∫
𝜆
𝑷𝔠,𝔠∗ ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝 (𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁, (21.1.55)
2𝜋 𝔠 𝔠 ∗

where ℎ ∈ O (Φ) (Ω) and the contour 𝔠 (resp., 𝔠 ∗ ) is strongly well-shaped at 𝑧◦ (resp.,
𝜁 ◦ ) for the function Ψ (𝑤, 𝜁 ◦ ) [resp., 𝜁 ↦→ Φ∗ (𝜁)].
Since our viewpoint is microlocal we can assume, without loss of generality, that
the Taylor expansion (20.1.11), which we reproduce here for convenience,
∑︁ 1
𝑝 (𝑧, 𝜁, 𝜆) = (𝜁 − 𝜁 ◦ ) 𝛼 𝜕𝜁𝛼 𝑝 𝑗 (𝑧, 𝜁 ◦ , 𝜆) (21.1.56)
𝛼∈Z𝑛
𝛼!
+

converges absolutely and uniformly in Ω × Ξ. We introduce the differential operator

of infinite order (20.1.12), 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆), which allows us to write
𝑛 ∫ ∫
◦ 𝜆
𝑷𝔠,𝔠 ℎ (𝑧, 𝜆) = 𝐿 𝜁 (𝑧, D𝑧 − 𝜆𝜁 , 𝜆)
∗ ◦ e𝑖𝜆(𝑧−𝑤) ·𝜁 ℎ (𝑤, 𝜆) d𝑤d𝜁.
2𝜋 𝔠 𝔠∗

Introducing the integral

𝑛 ∫ ∫
𝜆
𝑰𝔠,𝔠∗ ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) ·𝜁 ℎ (𝑤) d𝑤d𝜁.
2𝜋 𝔠 𝔠∗

we apply Theorem 21.1.7:

𝑰𝔠,𝔠∗ ℎ (𝑧, 𝜆) Ω′ − ℎ (𝑧, 𝜆)| Ω′ ∈ O (−𝜔) (Ω′)

if Ω′ ⊂ Ω is a sufficiently small neighborhood of 𝑧 ◦ . As a consequence, we have, in

Ω′,
𝑷𝔠,𝔠∗ ℎ (𝑧, 𝜆) 𝐿 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ , 𝜆) ℎ (𝑧, 𝜆) , (21.1.57)
where means congruent mod O (−𝜔) (Ω′). This can be rewritten in the following
◦ ◦ ◦
manner. Set ℎ˜ (𝑧, 𝜆) = e−𝑖 (𝑧−𝑧 ) ·𝜁 ℎ (𝑧, 𝜆) and 𝑷 𝜁 ◦ ,𝔠,𝔠∗ ℎ˜ = e−𝑖𝑧·𝜁 𝑷𝔠,𝔠∗ ℎ, i.e.,
𝑛 ∫ ∫
𝜆 ◦
˜
𝑷 𝜁 ◦ ,𝔠,𝔠∗ ℎ (𝑧, 𝜆) = e𝑖𝜆(𝑧−𝑤) · (𝜁 −𝜁 ) 𝑝 (𝑧, 𝜁, 𝜆) ℎ˜ (𝑤, 𝜆) d𝑤d𝜁,
2𝜋 𝔠 𝔠∗
(21.1.58)
then
𝑷 𝜁 ◦ ,𝔠,𝔠∗ ℎ˜ (𝑧, 𝜆) 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) ℎ˜ (𝑧, 𝜆) . (21.1.59)

We have ℎ˜ ∈ O ( Φ̃) (Ω), Φ̃ (𝑧) = Φ (𝑧) +Im ((𝑧 − 𝑧◦ ) · 𝜁 ◦ ), Φ̃ (𝑧 ◦ ) = 0, 𝜕𝑧 Φ̃ (𝑧◦ ) = 0;

note that Ψ (𝑤, 𝜁) = Φ̃ (𝑤) +Im ((𝑤 − 𝑧 ◦ ) · (𝜁 − 𝜁 ◦ )). This shows that we could have
started from the integral (21.1.58) instead of (21.1.55).
21.2 Classical Germ Pseudodifferential Operators 891

The following statements about finite realizations 𝐿 𝜁(𝑅)

◦ (𝑧, D 𝑧 , 𝜆) are applications
of Proposition 20.1.11.
Proposition 21.1.14 Let Φ ∈ C ∞ (Ω; R) be such that Φ (𝑧◦ ) = 0, 𝜕𝑧 Φ (𝑧 ◦ ) = 21 𝑖𝜁 ◦ .
To every open set Ω′ ⊂⊂ Ω there is an 𝑅◦ > 0 such that if 𝑅 > 𝑅◦ then the range of
the map O (Φ) (Ω) ∋ ℎ ↦→ 𝐿 𝜁(𝑅) ◦
◦ (𝑧, D 𝑧 − 𝜆𝜁 , 𝜆) ℎ
′
is contained in O (Φ) (Ω′).
Ω

Proposition 21.1.15 With Φ and Ω′ as in Proposition 21.1.14 there is an 𝑅1 > 0

such that if 𝑅 ′ > 𝑅 > 𝑅1 then the range of the map
′
O (Φ) (Ω) ∋ ℎ ↦→ 𝐿 𝜁(𝑅) ◦
◦ (𝑧, D 𝑧 − 𝜆𝜁 , 𝜆) ℎ − 𝐿 𝜁(𝑅◦ ) (𝑧, D𝑧 − 𝜆𝜁 ◦ , 𝜆) ℎ
Ω′ Ω′

is contained in O (−𝜔) (Ω′).

These propositions allow us to define a sheaf endomorphism

L 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ , 𝜆) : O (Φ) (Ω) /O (−𝜔) (Ω) ←↪ (21.1.60)

which the considerations above show to be equal to the germ pseudodifferential

operator (21.1.3).

21.2 Classical Germ Pseudodifferential Operators

21.2.1 Classical germ pseudodifferential operators

As before Ω, Ξ are domains in C𝑛 , 𝑧 ◦ ∈ Ω, 𝜁 ◦ ∈ Ξ; Φ ∈ C ∞ (Ω; R), Φ (𝑧 ◦ ) = 0,

𝜕𝑧 Φ (𝑧◦ ) = 21 𝑖𝜁 ◦ ; by Proposition 21.1.1 𝜑 (𝑧, 𝑤, 𝜁) = (𝑧 − 𝑤) · (𝜁 − 𝜁 ◦ ), Φ (𝑧) and
Φ (𝑤) form a Sjöstrand triad at (𝑧◦ , 𝑧◦ , 𝜁 ◦ ). In (21.1.2) and availing ourselves of
Theorem 21.1.6 we replace 𝑝 (𝑧, 𝑤, 𝜁, 𝜆) by a classical analytic symbol (Definition
19.1.14) of degree zero (for simplicity),
∞
∑︁
𝑝 (𝑧, 𝜁, 𝜆) = 𝜆− 𝑗 𝑝 𝑗 (𝑧, 𝜁) , (21.2.1)
𝑗=0

and its finite realizations (21.1.15). Keep in mind that 𝑝 (𝑧, 𝜁, 𝜆) satisfies Condition
(FA) (Definition 19.1.1) which reads here
(FA) To every compact subset 𝐾 of Ω × Ξ there is a 𝐶𝐾 > 0 such that
𝑗+1
∀ 𝑗 ∈ Z+ , max 𝑝 𝑗 (𝑧, 𝜁) ≤ 𝐶𝐾 𝑗!. (21.2.2)
(𝑧,𝜁 ) ∈𝐾

Theorem 21.1.2 allows us to define a germ pseudodifferential operator

P𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ ←↪ (21.2.3)
892 21 Germ Pseudodifferential Operators in Complex Space

starting from integrals (21.1.16) in which ℎ ∈ O (Φ) (Ω) and the contour of integration
ℭ ⊂⊂ Ω × Ξ is strongly well-shaped at (𝑧 ◦ , 𝜁 ◦ ) ∈ ℭ for Ψ (𝑧◦ , 𝑤, 𝜁) = Φ (𝑤) +
Im (𝑤 − 𝑧◦ ) · 𝜁. That (21.2.3) is independent of suitably large 𝑅 [in (21.1.16)] is a
consequence of the following

Proposition 21.2.1 Let Ω′ ⊂ Ω and Ξ′ ⊂ Ξ be suitably small neighborhoods of 𝑧◦

and 𝜁 ◦ respectively, such that ℭ ⊂⊂ Ω′ × Ξ′. If 𝑅 is sufficiently large and 𝑅 ′ > 𝑅
′
then 𝑷ℭ(𝑅) ℎ − 𝑷ℭ(𝑅 ) ℎ ∈ O (−𝜔) (Ω′) whatever ℎ ∈ O (Φ) (Ω).

Proof We apply Proposition 19.1.13: if 𝑅 is sufficiently large and 𝑅 ′ > 𝑅 we have

′
𝑝 (𝑅) (𝑧, 𝜁, 𝜆) − 𝑝 (𝑅 ) (𝑧, 𝜁, 𝜆) ∈ O (−𝜔) (Ω′ × Ξ′) . (21.2.4)

Then the claim follows from Proposition 20.4.11. □

Definition 21.2.2 We shall refer to P𝑧 ◦ ,𝜁 ◦ as a classical germ pseudodifferential

operator at (𝑧 ◦ , 𝜁 ◦ ).

The germ operator (21.2.3) is said to be elliptic if the principal symbol 𝑝 0 (𝑧, 𝜁)
of P𝑧 ◦ ,𝜁 ◦ does not vanish at (𝑧◦ , 𝜁 ◦ ).

Theorem 21.2.3 If the classical germ pseudodifferential operator (21.2.3) is elliptic

then there is an elliptic classical germ pseudodifferential operator
(Φ) (−𝜔)
P−1
𝑧 ◦ ,𝜁 ◦ : O 𝑧 ◦ /O 𝑧 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (21.2.5)

such that
(Φ)
P𝑧 ◦ ,𝜁 ◦ P−1 −1
𝑧 ◦ ,𝜁 ◦ = P 𝑧 ◦ ,𝜁 ◦ P 𝑧 ◦ ,𝜁 ◦ = I 𝑧 ◦ (21.2.6)

(I𝑧(Φ) (Φ) (−𝜔)

◦ : the identity operator on O 𝑧 ◦ /O 𝑧 ◦ ).

Proof (𝑧, 𝜁) does not vanish in Ω × Ξ the total symbol 𝑝 (𝑧, 𝜁, 𝜆) =

Í∞ −If𝑗 𝑝 0
𝑗=0 𝜆 𝑝 𝑗 (𝑧, 𝜁) has an inverse with respect to the composition law # (Theorem
19.1.17) which defines (21.2.5); (21.2.6) follows directly from Theorem 21.1.13
applied to finite realizations of the total symbols. □

21.2.2 Classical germ pseudodifferential operators and differential

operators of infinite order

Everything said in Subsection 21.1.6 can be repeated in the classical case, with the
added fact that the differential operator of infinite order associated to a classical
pseudodifferential operator defined by an integral (21.1.2), more precisely to its
classical symbol 𝑝 (𝑧, 𝜁, 𝜆) = ∞ −𝑗
Í
𝑗=0 𝑗 (𝑧, 𝜁) 𝜆 , can be taken itself to be classical.
𝑝
All this has been spelled out in detail in Subsection 19.2.5. We recall it rapidly here.
Given open subsets Ω and Ξ of C𝑛 such that (𝑧◦ , 𝜁 ◦ ) ∈ Ω × Ξ, the differential
operator is [(19.2.16), Proposition 19.2.23]
21.2 Classical Germ Pseudodifferential Operators 893
∞ ∑︁
∑︁ 1 − 𝑗−| 𝛼 | 𝛼
𝐿 𝜁 ◦ (𝑧, D, 𝜆) = 𝜆 𝜕𝜁 𝑝 𝑗 (𝑧, 𝜁 ◦ ) D 𝛼 . (21.2.7)
𝑗=0 𝛼∈Z 𝑛 𝛼!
+

The finite realizations (21.1.15) of 𝑝 (𝑧, 𝜁, 𝜆), 𝑝 (𝑅) (𝑧, 𝜁, 𝜆), are polynomials with
respect to 𝜆−1 with coefficients in O (Ω × Ξ). Going back to the integral (21.1.16)
we can form the Taylor expansion with respect to 𝜁 about 𝜁 ◦ .

𝑷ℭ(𝑅) ℎ (𝑧, 𝜆)
𝑛 ∑︁ ∫
𝜆 1 𝛼 (𝑅)
= 𝜕𝜁 𝑝 (𝑧, 𝜁 ◦ , 𝜆) e𝑖𝜆(𝑧−𝑤) ·𝜁 (𝜁 − 𝜁 ◦ ) 𝛼 ℎ (𝑤, 𝜆) d𝑤d𝜁,
2𝜋 𝛼∈Z 𝑛 𝛼! ℭ
+

normally convergent provided |𝜁 − 𝜁 ◦ | is sufficiently small. We introduce the corre-

sponding differential operators of infinite order,
∑︁ 1
𝐿 𝜁(𝑅)
◦ (𝑧, D 𝑧 , 𝜆) = 𝜆−| 𝛼 | 𝜕𝜁𝛼 𝑝 (𝑅) (𝑧, 𝜁 ◦ , 𝜆) D𝑧𝛼 ; (21.2.8)
𝛼∈Z 𝑛 𝛼!
+

we see that
◦
e−𝑖𝜆𝑧·𝜁 𝑷ℭ(𝑅) ℎ (𝑧, 𝜆) (21.2.9)
𝑛 ∫
𝜆 ◦
= 𝐿 𝜁(𝑅)
◦ (𝑧, D𝑧 , 𝜆) e−𝑖𝜆𝑧·𝜁 e𝑖𝜆(𝑧−𝑤) ·𝜁 ℎ (𝑤, 𝜆) d𝑤d𝜁 .
2𝜋 ℭ

In all this 𝑅 > 0 can be as large as needed. The analogues of Propositions 21.1.14,
21.1.15 are valid. They enable us to state

Proposition 21.2.4 If the neighborhood Ω′ ⊂ Ω of 𝑧◦ and the contour ℭ ⊂⊂

Ω′ × Ξ′ are sufficiently small, and if 𝑅 is sufficiently large, then the operators
𝑷ℭ(𝑅) and 𝐿 𝜁(𝑅) ◦
◦ (𝑧, D 𝑧 − 𝜆𝜁 , 𝜆) define the same linear map of O
(Φ) (Ω) /O (−𝜔) (Ω)

into O (Φ) ′
(Ω ) /O (−𝜔) ′
(Ω ).

The equivalence classes mod O𝑧(−𝜔)

◦ of the germs at 𝑧◦ of 𝐿 𝜁(𝑅) ◦
◦ (𝑧, D 𝑧 − 𝜆𝜁 , 𝜆) ℎ

and 𝑷ℭ(𝑅) ℎ are equal (𝑅 > 0 suitably large). The resulting equivalence class is
· ·
independent of 𝑅 and equal to P𝑧 ◦ ,𝜁 ◦ h𝑧 ◦ (h𝑧 ◦ : the equivalence class mod O𝑧(−𝜔)
◦ of
h𝑧 ◦ , the germ of ℎ at 𝑧 ◦ ). The finite realizations 𝐿 𝜁(𝑅)
◦ (𝑧, D 𝑧 , 𝜆) can be used to define
a germ linear operator defined by integrals (21.1.10) and a sheaf endomorphism

L 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ , 𝜆) : O (Φ) (Ω) /O (−𝜔) (Ω) ←↪ (21.2.10)

equal to the germ operator P𝑧 ◦ ,𝜁 ◦ [cf. (21.1.3)].

894 21 Germ Pseudodifferential Operators in Complex Space

Remark 21.2.5 Reasoning as in Section 19.2, we can use (21.2.10) to define a germ
operator Ob,𝑧 ◦ /O𝑧(−𝜔)
◦ ←↪ [cf. (19.2.6)] but we cannot state that (21.2.10) extends the
latter (or vice-versa) since the fact that 𝑧◦ is a saddle point of Φ implies Ob,𝑧 ◦ ⊄ O𝑧(Φ)
◦

and O𝑧(Φ)
◦ ⊄ Ob,𝑧 ◦ .

21.2.3 Action on classical Fourier-like transforms

We assume that (𝜑, Φ) is a Sjöstrand pair at (𝑧 ◦ , 𝜃 ◦ ) in Ω × Θ; (−𝜑, Φ∗ ) is the dual

Sjöstrand pair at (𝜃 ◦ , 𝑧◦ ) (Definitions 20.3.6, 20.3.18). We take the symbol in the
integral (20.4.10) to be classical: 𝑏 (𝑧, 𝜃, 𝜆) = +∞ −𝑗
Í
𝑗=0 𝜆 𝑏 𝑗 (𝑧, 𝜃) in Ω × Θ. We return
to the integral operators (21.1.16) in which we put
𝑛 ∫
𝜆
ℎ (𝑤, 𝜆) = 𝑩𝔠∗ 𝑔 (𝑤, 𝜆) = e−𝑖𝜆𝜑 (𝑤, 𝜃) 𝑏 (𝑤, 𝜃, 𝜆) 𝑔 (𝜃, 𝜆) d𝜃 (21.2.11)
2𝜋 𝔠∗

where the contour 𝔠 ∗ ⊂⊂ Θ is strongly well-shaped at 𝜃 ◦ for Ψ∗ (𝑧◦ , 𝜃) = Φ∗ (𝜃) +

∗
Im 𝜑 (𝑧◦ , 𝜃) and 𝑔 ∈ O (Φ ) (Θ). We get, by (21.2.9),
−𝑛
𝜆
𝑷ℭ(𝑅) ℎ (𝑧, 𝜆)
2𝜋
𝑛 ∫ ∫
𝜆
= e𝑖𝜆(𝑧−𝑤) ·𝜁 −𝑖𝜆𝜑 (𝑤, 𝜃) 𝑝 (𝑅) (𝑧, 𝜁, 𝜆) 𝑏 (𝑤, 𝜃, 𝜆) 𝑔 (𝜃, 𝜆) d𝜃d𝑤d𝜁
2𝜋 ℭ 𝔠∗
∫
𝑖𝜆𝑧·𝜁 ◦ (𝑅) ◦
e 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e−𝑖𝜆( 𝜑 (𝑧, 𝜃)+𝑧·𝜁 ) 𝑏 (𝑧, 𝜃, 𝜆) 𝑔 (𝜃, 𝜆) d𝜃
𝔠∗

where means congruent mod O (−𝜔) (Ω) (Ω contracted about 𝑧◦ ). We require

𝜕𝑧 𝜑 (𝑧 ◦ , 𝜃 ◦ ) = −𝜁 ◦ , which allows us to apply Corollary 19.2.20: indeed, if we set
𝜓 (𝑧, 𝜃) = 𝜑 (𝑧, 𝜃) + 𝑧 · 𝜁 ◦ then 𝜕𝑧 𝜓 (𝑧 ◦ , 𝜃 ◦ ) = 0 and we have, formally,

( 𝜓)
e𝑖𝜆𝜓 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e−𝑖𝜆𝜓 𝑏 (𝑧, 𝜃, 𝜆) = 𝐿 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ , 𝜆) 𝑏 (𝑧, 𝜃, 𝜆) , (21.2.12)

where 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) is the operator (21.2.7) and

∑︁
( 𝜓)
𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) = 𝑐♭𝛼 (𝑧, 𝜆) 𝜆− | 𝛼 | D𝑧𝛼
𝛼∈Z+𝑛

is a differential operator of infinite order in Ω (contracted about 𝑧◦ ). It is directly

checked that
( 𝜓)
𝑏♭ (𝑧, 𝜃, 𝜆) = 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) 𝑏 (𝑧, 𝜃, 𝜆)
is a classical symbol. We define
21.2 Classical Germ Pseudodifferential Operators 895
𝑛 ∫
𝜆
𝑩𝔠♭∗ 𝑔 (𝑧, 𝜆) = e−𝑖𝜆𝜑 (𝑧, 𝜃) 𝑏♭ (𝑧, 𝜃, 𝜆) 𝑔 (𝜃, 𝜆) d𝜃 (21.2.13)
2𝜋 𝔠∗

and derive 𝑷ℭ 𝑩𝔠∗ 𝑔 − 𝑩𝔠♭∗ 𝑔 ∈ O (−𝜔) (Ω′). (Of course, in order to have a true integral
operator we must replace 𝑏♭ by one of its finite realizations.) Going, without further
ado, to germ operators we get

Theorem 21.2.6 Suppose that 𝜕𝑧 𝜑 (𝑧 ◦ .𝜃 ◦ ) = −𝜁 ◦ . Let P𝑧 ◦ ,𝜁 ◦ be the classical germ

pseudodifferential operator (21.2.3) and let
∗
B 𝜃 ◦ ,𝑧 ◦ : O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (21.2.14)

be the classical germ Fourier-like transform defined by the integral operator

(21.2.11). The composite P𝑧 ◦ ,𝜁 ◦ B 𝜃 ◦ ,𝑧 ◦ is the classical germ Fourier-like transform
∗
B♭𝜃 ◦ ,𝑧 ◦ : O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (21.2.15)

defined by the integral operator (21.2.13).

Perhaps the simplest way to compute 𝑏♭ from the symbol 𝑏 (and the phase 𝜑) is
to use the equation (21.2.12) rewritten as

𝑏♭ (𝑧, 𝜃, 𝜆) = 𝐿 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ − 𝜆𝜕𝑧 𝜑 (𝑧, 𝜃) , 𝜆) 𝑏 (𝑧, 𝜃, 𝜆) (21.2.16)

∞
∑︁ ∑︁ 1 𝛼
= 𝜆− 𝑗 𝜕𝜁𝛼 𝑝 𝑗 (𝑧, 𝜁 ◦ ) 𝜆−1 D𝑧 − 𝜆𝜁 ◦ − 𝜕𝑧 𝜑 (𝑧, 𝜃) 𝑏 𝑗 (𝑧, 𝜃, 𝜆) .
𝑗=0 𝛼∈Z𝑛
𝛼!
+

We get straightaway the relation between principal symbols:

𝑏♭0 (𝑧, 𝜃) = 𝑝 0 (𝑧, −𝜕𝑧 𝜑 (𝑧, 𝜃)) 𝑏 0 (𝑧, 𝜃) . (21.2.17)

In particular, 𝑏♭0 (𝑧 ◦ , 𝜃 ◦ ) = 𝑝 0 (𝑧 ◦ , 𝜁 ◦ ) 𝑏 0 (𝑧 ◦ , 𝜃 ◦ ).
Rather than looking at the left action of P𝑧 ◦ ,𝜁 ◦ on a classical germ Fourier-like
transform B 𝜃 ◦ ,𝑧 ◦ we could have looked at the right action P𝑧 ◦ ,𝜁 ◦ on a classical germ
Fourier-like transform (20.4.4),
∗
A𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ ,

defined by an integral (20.1) (cf. Theorem 20.4.6) with a classical symbol 𝑎 (𝑧, 𝜃, 𝜆).
It would have resulted in a classical germ Fourier-like transform
∗
A𝑧 ◦ , 𝜃 ◦ P𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ −→ O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ (21.2.18)

with same phase-function 𝜑 as (20.1) and principal symbol

𝑎♭0 (𝑧, 𝜃) = 𝑝 0 (𝑧, 𝜕𝑧 𝜑 (𝑧, 𝜃)) 𝑎 0 (𝑧, 𝜃) . (21.2.19)

896 21 Germ Pseudodifferential Operators in Complex Space

Assume that A𝑧 ◦ , 𝜃 ◦ is classical elliptic and let A−1𝑧 ◦ , 𝜃 ◦ be its inverse. By Theorem
−1
21.2.6 P𝑧 ◦ ,𝜁 ◦ A𝑧 ◦ , 𝜃 ◦ is a germ Fourier-like transform with phase-function −𝜑 (𝑧, 𝜃).
By applying Theorem 21.1.12 we deduce that
(Φ) (−𝜔)
A𝑧 ◦ , 𝜃 ◦ P𝑧 ◦ ,𝜁 ◦ A−1
𝑧 ◦ , 𝜃 ◦ : O 𝑧 ◦ /O 𝑧 ◦ ←↪

is a classical germ pseudodifferential operator whose principal symbol is easy to

derive from (21.2.17). We shall give the formula in the more general case of a
classical elliptic germ FIO instead of a germ Fourier-like transform.

21.2.4 Action on classical germ FIOs

We continue to deal with the classical germ pseudodifferential operator P𝑧 ◦ ,𝜁 ◦ :

O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ ←↪ defined by the integrals (21.1.2). Keep in mind that Φ (𝑧◦ ) = 0,
◦ 1 ◦
𝜕𝑧 Φ (𝑧 ) = 2 𝑖𝜁 (Proposition 21.1.1). In this subsection we essentially duplicate for
classical germ FIOs the results about classical germ Fourier-like transforms in the
preceding section; the transform (21.2.14) can be replaced by a classical germ FIO,
(Φ2 ) (−𝜔)
B𝑤 ◦ ,𝑧 ◦ (𝜃 ◦ ) : O𝑤 ◦ /O 𝑤 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (21.2.20)

defined starting from an integral (20.5.8),

𝑛 ∫
𝜆
𝑩ℭ2 ℎ2 (𝑧, 𝜆) = e−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃 (21.2.21)
2𝜋 ℭ2

where 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) = +∞ −𝑗
Í
𝑗=0 𝜆 𝑏 𝑗 (𝑧, 𝑤, 𝜃) is classical of order zero and ℎ ∈
O (Φ 2 ) (Ω2 ), Ω2 ⊂ C open. This assumes that (−𝜑 (𝑧, 𝑤, 𝜃) , Φ2 (𝑤) , Φ (𝑧)) or,
𝑛

equivalently (Corollary 20.5.4), (𝜑 (𝑧, 𝑤, 𝜃) , Φ (𝑧) , Φ2 (𝑤)) is a Sjöstrand triad at

(𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) ∈ Ω × Ω2 × Θ with 𝜑 ∈ O (Ω × Ω2 × Θ), Φ2 ∈ C ∞ (Ω2 ) such that
Φ2 (𝑤 ◦ ) = Im 𝜑 (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) = 0. Definition 20.5.1, requires that
(1) (𝑧◦ , 𝜃 ◦ ) be a saddle point of Ψ (𝑧, 𝑤 ◦ , 𝜃) = Φ (𝑧) − Im 𝜑 (𝑧, 𝑤 ◦ , 𝜃),
(2) (𝑤 ◦ , 𝜃 ◦ ) be a saddle point of Ψ2 (𝑧◦ , 𝑤, 𝜃) = Φ2 (𝑤) + Im 𝜑 (𝑧◦ , 𝑤, 𝜃),
(3) the critical value of (𝑧, 𝜃) ↦→ Ψ (𝑧, 𝑤, 𝜃) be equal to Φ2 (𝑤) in a neighborhood
of (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ).
Condition (1) combined with 𝜕𝑧 Φ (𝑧◦ ) = 21 𝑖𝜁 ◦ entails

𝜕𝑧 𝜑 (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) = −𝜁 ◦ . (21.2.22)

We have, for 𝑧 ∈ Ω′ ⊂ Ω (Ω′: a suitably small neighborhood of 𝑧◦ ),

21.2 Classical Germ Pseudodifferential Operators 897
−𝑛
𝜆
𝑷ℭ 𝑩ℭ2 ℎ2 (𝑧, 𝜆)
2𝜋
𝑛 ∫
𝜆
= e𝑖𝜆(𝑧−𝑧˜ ) ·𝜁 −𝑖𝜆𝜑 ( 𝑧˜ ,𝑤, 𝜃) 𝑝 (𝑧, 𝜁, 𝜆) 𝑏 ( 𝑧˜, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃d𝑧˜d𝜁
2𝜋 ℭ×ℭ 2
∫
𝑖𝜆𝑧·𝜁 ◦ ◦
e 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e𝑖𝜆(𝑧−𝑧˜ ) ·𝜁 −𝑖𝜆𝜑 ( 𝑧˜ ,𝑤, 𝜃)−𝑖𝜆𝑧·𝜁
ℭ×ℭ 2
×𝑏 ( 𝑧˜, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃d𝑧˜d𝜁

where means congruent mod O (−𝜔) (Ω′); (21.2.9) implies

𝑛 ∫ ∫
𝜆
e𝑖𝜆(𝑧−𝑧˜ ) ·𝜁 −𝑖𝜆𝜑 ( 𝑧˜ ,𝑤, 𝜃) 𝑏 ( 𝑧˜, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃𝑑 𝑧˜d𝜁
2𝜋 ℭ ℭ2
∫
e−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃,
ℭ2

whence
−𝑛
𝜆
𝑷ℭ 𝑩ℭ2 ℎ2 (𝑧, 𝜆)
2𝜋
∫ ∫
◦ ◦
e𝑖𝜆𝑧·𝜁 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e−𝑖𝜆( 𝜑 (𝑧,𝑤, 𝜃)+𝑧·𝜁 ) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃𝑑 𝑧˜d𝜁 .
ℭ ℭ2

Thanks to (21.2.22), here too we can apply Corollary 19.2.20: there is a classical
amplitude 𝑏♭ (𝑧, 𝑤, 𝜃, 𝜆) = +∞ −𝑗 ♭
Í
𝑗=0 𝜆 𝑏 𝑗 (𝑧, 𝑤, 𝜃) such that
∫
◦ ◦
e𝑖𝜆𝑧·𝜁 𝐿 𝜁 ◦ (𝑧, D𝑧 , 𝜆) e−𝑖𝜆( 𝜑 (𝑧,𝑤, 𝜃)+𝑧·𝜁 ) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃
ℭ2
∫
−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) ♭
e 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃.
ℭ2

If we define
𝑛 ∫
𝜆
𝑩♭ℭ2 ℎ2 (𝑤, 𝜆) = e−𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑏♭ (𝑧, 𝑤, 𝜃, 𝜆) ℎ2 (𝑤, 𝜆) d𝑤d𝜃 (21.2.23)
2𝜋 ℭ2

then, provided ℭ2 ⊂⊂ Ω2 (𝑤 ◦ ∈ ℭ2 ) is sufficiently small,

∀ℎ2 ∈ O (Φ2 ) (Ω2 ) , 𝑷ℭ 𝑩ℭ2 ℎ2 − 𝑩♭ℭ2 ℎ2 ∈ O (−𝜔) (Ω′) .

Since 𝑷ℭ( 𝑁 ) 𝑩ℭ2 ℎ2 and 𝑩♭ℭ2 ℎ2 both belong to O (−𝜔) (Ω′) whenever ℎ2 ∈ O (−𝜔) (Ω2 )
(Proposition 20.4.5) we reach the following conclusion:

Proposition 21.2.7 Provided the domains Ω′ ⊂ Ω, Ω2 and Θ are sufficiently small

the integral operator 𝑷ℭ 𝑩ℭ2 and 𝑩♭ℭ2 define one and the same linear mapping of
O (Φ2 ) (Ω2 ) /O (−𝜔) (Ω2 ) into O (Φ) (Ω′) /O (−𝜔) (Ω′).
898 21 Germ Pseudodifferential Operators in Complex Space

Going to germ operators we may state

Theorem 21.2.8 Suppose 𝜕𝑧 Φ (𝑧◦ ) = 21 𝑖𝜁 ◦ . Let P𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ) (−𝜔)

◦ /O 𝑧 ◦ ←↪ be the
classical germ pseudodifferential operator defined by the integral operator (21.1.2)
and let B𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) be the classical germ FIO (21.2.20) defined by the integral
(21.2.21). The composite P𝑧 ◦ ,𝜁 ◦ B𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) is the classical germ FIO
(Φ2 ) (−𝜔)
B♭𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) : O𝑤 ◦ /O 𝑤 ◦ −→ O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ (21.2.24)

defined by the integral operator (21.2.23) [and therefore with the same phase-function
as B𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ )].

Using the relation

𝑏♭ (𝑧, 𝑤, 𝜃, 𝜆) = 𝐿 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ − 𝜆𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃) , 𝜆) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) (21.2.25)

∞ ∑︁
∑︁ 1 −𝑗 𝛼 𝛼
= 𝜆 𝜕𝜁 𝑝 𝑗 (𝑧, 𝜁 ◦ ) 𝜆−1 D𝑧 − 𝜁 ◦ − 𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆)
𝑗=0 𝛼∈Z𝑛
𝛼!
+

we derive the expected relation between principal symbols:

𝑏♭0 (𝑧, 𝑤, 𝜃) = 𝑝 0 (𝑧, −𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃)) 𝑏 0 (𝑧, 𝑤, 𝜃) . (21.2.26)

By (21.2.22) we have

𝑏♭0 (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ) = 𝑝 0 (𝑧 ◦ , 𝜁 ◦ ) 𝑏 0 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) . (21.2.27)

21.2.5 Inversion of classical elliptic germ Fourier-like transforms

We go back to the germ operators with opposite phase-functions, (21.1.30) and

(21.1.31). We are going to apply Theorem 21.1.12 in which the hypotheses include
(21.1.28) and 𝜕𝑧 Φ (𝑧 ◦ ) = 21 𝑖𝜁 ◦ .
Now we assume that the amplitudes in the integrals (20.1) and (20.4.10) [defining
(21.1.30) and (21.1.31) respectively] are classical:
∞
∑︁
𝑎 (𝑧, 𝜃, 𝜆) = 𝜆− 𝑗 𝑎 𝑗 (𝑧, 𝜃) , (21.2.28)
𝑗=0
∞
∑︁
𝑏 (𝑧, 𝜃, 𝜆) = 𝜆− 𝑗 𝑏 𝑗 (𝑧, 𝜃) . (21.2.29)
𝑗=0
21.2 Classical Germ Pseudodifferential Operators 899

We refer the reader to Formula (21.1.37) for the symbol of the composite
B 𝜃 ◦ ,𝑧 ◦ A𝑧 ◦ , 𝜃 ◦ , here yielding the classical symbol

𝑐˜ (𝑤, 𝜁, 𝜆) (21.2.30)
∑︁ 𝜆−| 𝛼 |
𝑎 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆) 𝑏 (𝑤, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆)

= D 𝜁𝛼 𝜕𝑧𝛼
𝛼∈Z+𝑛
𝛼! det 𝜕𝜃 𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃) 𝑧=𝑤
∞
∑︁ ∑︁ 𝜆− 𝑗− | 𝛼 | ∑︁
𝑎 𝑘 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁)) 𝑏 ℓ (𝑤, 𝜃 (𝑧, 𝑤, 𝜁))
= D 𝜁𝛼 𝜕𝑧𝛼 ,
𝑗=0 𝛼∈Z𝑛
𝛼! 𝑘+ℓ= 𝑗 det 𝜕𝜃 𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃) 𝑧=𝑤
+

where 𝜃 (𝑧, 𝑤, 𝜁) is the (unique) solution of the equation 𝜙1 (𝑧, 𝑤, 𝜃) = 𝜁 in

a neighborhood of (𝑧 ◦ , 𝑧◦ , 𝜃) such that 𝜃 (𝑧 ◦ , 𝑤 ◦ , 𝜁 ◦ ) = 𝜃 ◦ . Keep in mind that
𝜙1 (𝑧◦ , 𝑧◦ , 𝜃) = −𝜕𝑧 𝜑 (𝑧◦ , 𝜃 ◦ ) = 𝜁 ◦ [cf. (21.2.22)]. We can state (cf. Theorem
21.1.12)

Theorem 21.2.9 Assume that the phase-function 𝜑 ∈ O (Ω × Θ) satisfies (21.1.28)

and that 𝜕𝑧 Φ (𝑧 ◦ ) = 21 𝑖𝜁 ◦ . If the germ operators A𝑧 ◦ , 𝜃 ◦ and B 𝜃 ◦ ,𝑧 ◦ [i.e., (21.1.30) and
(21.1.31)] are classical so is their composite, the germ pseudodifferential operator

B 𝜃 ◦ ,𝑧 ◦ A𝑧 ◦ , 𝜃 ◦ : O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ ←↪ . (21.2.31)

It is not difficult to see that if 𝑎 0 (𝑧,Í𝜃) does not vanish at any point of Ω × Θ then,
given arbitrarily a classical symbol ∞ −𝑗
𝑗=0 𝜆 𝑐 𝑗 (𝑧, 𝜁) in Ω × Ξ, the holomorphic
functions 𝑏 ℓ (𝑤, 𝜃) can be determined recursively to solve

𝜆−1 D 𝜁 𝜕𝑤 D𝜃
e 𝑎 (𝑤, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆) 𝑏 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁) , 𝜆) (𝑤, 𝑧, 𝜁) =1
D𝜁 𝑤=𝑧
(21.2.32)
[cf. (21.1.36), (21.1.37)]. The solution (21.2.29) is the symbol of the inverse A−1 𝑧 ◦ , 𝜃 ◦ of
A𝑧 ◦ , 𝜃 ◦ and (21.2.31) is the identity map of O𝑧(Φ) (−𝜔)
◦ /O 𝑧 ◦ . To get explicit formulas for
the 𝑏 ℓ by solving the recurrence equations derived from (21.2.32) is not practicable
beyond low values of ℓ. However, determining the principal symbol is immediate,
since (21.2.32) implies

𝑎 0 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁)) 𝑏 0 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁)) = Δ (𝑧, 𝜁) (21.2.33)

where Δ (𝑧, 𝜁) = det 𝜕𝜃 𝜕𝑧 𝜑 (𝑧, 𝜃 (𝑧, 𝜁)) with 𝜃 (𝑧, 𝜁) the solution of the equation
𝜕𝑧 𝜑 (𝑧, 𝜃) = −𝜁 such that 𝜃 (𝑧◦ , 𝜁 ◦ ) = 𝜃 ◦ . We can state:

Theorem 21.2.10 Assume that the phase-function 𝜑 ∈ O (Ω × Θ) satisfies (21.1.28)

and that
1 1
𝜕𝑧 Φ (𝑧◦ ) = − 𝑖𝜕𝑧 𝜑 (𝑧◦ , 𝜃 ◦ ) = 𝑖𝜁 ◦ .
2 2
If the classical germ Fourier-like transform A𝑧 ◦ , 𝜃 ◦ [i.e., (21.1.30)] is elliptic there
is a classical “inverse” germ Fourier-like transform A−1 −1
𝑧 ◦ , 𝜃 ◦ such that A 𝑧 ◦ , 𝜃 ◦ A 𝑧 , 𝜃
◦ ◦

is the identity operator on O𝑧(Φ) (−𝜔)

◦ /O 𝑧 ◦ and A𝑧 ◦ , 𝜃 ◦ A−1
𝑧 ◦ , 𝜃 ◦ is the identity operator on
900 21 Germ Pseudodifferential Operators in Complex Space
∗
O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ . If 𝑎 0 (𝑧, 𝜃) is the principal symbol of A 𝑧 , 𝜃 the principal symbol of
◦ ◦
−1
A𝑧 ◦ , 𝜃 ◦ is
𝑎 0 (𝑧, 𝜃 (𝑧, 𝑤, 𝜁)) −1 Δ (𝑧, 𝜁) . (21.2.34)

21.2.6 Inversion of classical elliptic germ FIOs. Egorov’s theorem

The results about germ Fourier-like transforms in the preceding section can be
essentially duplicated for germ FIOs. We assume throughout that (𝜑, Φ1 , Φ2 ) is a
Sjöstrand triad at (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) ∈ Σ 𝜑 in Ω1 × Ω2 × Θ. In the integrals (20.5.7) and
(20.5.8) we take the amplitudes to be classical:
∞
∑︁
𝑎 (𝑧, 𝑤, 𝜃, 𝜆) = 𝜆− 𝑗 𝑎 𝑗 (𝑧, 𝑤, 𝜃) , (21.2.35)
𝑗=0
∞
∑︁
𝑏 (𝑧, 𝑤, 𝜃, 𝜆) = 𝜆− 𝑗 𝑏 𝑗 (𝑧, 𝑤, 𝜃) . (21.2.36)
𝑗=0

Theorem 21.1.13 and Formulas (21.1.48), (21.1.54) have the following direct con-
sequence:

Theorem 21.2.11 Assume that 𝜑 satisfy Condition (20.5.14) and dΦ1 (𝑧◦ ) = 21 𝑖𝜁 ◦ .
Let A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) and B𝑤 ◦ ,𝑧 ◦ (𝜃 ◦ ) be the classical germ FIOs (20.5.9) and (20.5.10)
with phase-functions 𝜑 and −𝜑 and amplitudes (21.2.35) and (21.2.36) respectively.
Under these hypotheses the composite B𝑤 ◦ ,𝑧 ◦ (𝜃 ◦ ) A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) : O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ ←↪
is a classical germ pseudodifferential operator at (𝑧◦ , 𝜁 ◦ ), 𝜁 ◦ = −𝜕𝑧 𝜑 (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ),
with principal symbol 𝑎Δ0 𝑏𝜑 0 .
Σ𝜑

The only difference with the results of the preceding section is that the amplitudes
and, consequently, operators are classical.
We remind the reader that

𝜕 𝜕 𝜑 𝜕𝑧 𝜕𝜃 𝜑
Δ 𝜑 = det 𝑧 𝑤 , (21.2.37)
𝜕𝑤 𝜕𝜃 𝜑 𝜕𝜃2 𝜑

and Σ 𝜑 is the submanifold of Ω1 × Ω2 × Θ defined by 𝜑 𝜃 (𝑧, 𝑤, 𝜃) = 0. If A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ )

is elliptic, i.e., 𝑎 0 ≠ 0 everywhere in a neighborhood of (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ ), we can solve
the equation [cf. (21.1.48)]

−1 D 𝐷 (𝑤, 𝜃)
e𝜆 𝜁 𝜕𝑧˜
𝑎 ( 𝑧˜, 𝑤, 𝜃, 𝜆) 𝑏 (𝑧, 𝑤, 𝜃, 𝜆) (𝑧, 𝑧˜, 0, 𝜁, 0) =1 (21.2.38)
𝐷 (𝑡, 𝜁) 𝑧˜ =𝑧
21.2 Classical Germ Pseudodifferential Operators 901

where 𝑤 = 𝑓 (𝑧, 𝑧˜, 0, 𝜁, 0), 𝜃 = 𝑔 (𝑧, 𝑧˜, 0, 𝜁, 0), regarding the coefficients 𝑎 𝑗 in
(21.2.35) as given and the coefficients 𝑏 𝑗 in (21.2.36) as unknowns. The result-
ing classical symbol (21.2.36) is the total symbol of the inverse A−1 ◦
𝑧 ◦ ,𝑤 ◦ (𝜃 ) of
A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ). We can state:

Theorem 21.2.12 Assume the same hypotheses as in Theorem 21.2.11. If the germ
operator
A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) : O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦
(Φ2 )
−→ O𝑤 (−𝜔)
◦ /O 𝑤 ◦ (21.2.39)
(Φ2 ) (−𝜔)
is elliptic there is an elliptic classical germ FIO O𝑤 ◦ /O 𝑤 ◦ −→ O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦
−1 ◦ −1 ◦ ◦ ◦
◦
with phase-function −𝜑, A𝑧 ◦ ,𝑤 ◦ (𝜃 ), such that A𝑧 ◦ ,𝑤 ◦ (𝜃 ) A𝑧 ,𝑤 (𝜃 ) is the identity
operator of O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ and A𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) A−1 ◦
𝑧 ◦ ,𝑤 ◦ (𝜃 ) is the identity operator of
(Φ2 ) (−𝜔) Δ𝜑
O𝑤 ◦ /O 𝑤 ◦ . The principal symbol of A−1 ◦
𝑧 ◦ ,𝑤 ◦ (𝜃 ) is equal to 𝑎0 Σ .
𝜑

Next, we apply Theorem 21.2.8 and (21.2.26) with B𝑧 ◦ ,𝑤 ◦ (𝜃 ◦ ) = A−1 ◦

𝑧 ◦ ,𝑤 ◦ (𝜃 ):
let P𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ ←↪ be a germ pseudodifferential operator with prin-
cipal symbol 𝑝 0 (𝑧, 𝜁). The composite P𝑧 ◦ , 𝜔 ◦ A−1 ◦
𝑧 ◦ ,𝑧 ◦ (𝜃 ) is a classical germ FIO
(Φ2 ) (−𝜔)
O𝑤 ◦ /O 𝑤 ◦ −→ O𝑧(Φ◦ 1 ) /O𝑧(−𝜔)
◦ with phase-function −𝜑 and principal symbol

𝑝 0 (𝑧, −𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃)) 𝑎 0 (𝑧, 𝑤, 𝜃) .

At this stage we can apply Theorem 21.1.13 and combine (21.1.54) with the last
claim in Theorem 21.2.12 to state

Theorem 21.2.13 Assume the same hypotheses as in Theorem 21.2.11. Let (21.2.39)
be an elliptic classical germ FIO with phase-function 𝜑 and P𝑧 ◦ ,𝜁 ◦ : O𝑧(Φ◦ 2 ) /O𝑧(−𝜔)
◦ ←↪
be a germ pseudodifferential operator with principal symbol 𝑝 0 (𝑧, 𝜁). Under these
hypotheses
A−1 ◦ ◦
𝑧 ◦ ,𝑤 ◦ (𝜃 ) P 𝑧 ◦ ,𝜁 ◦ A 𝑧 ◦ ,𝑤 ◦ (𝜃 )
(Φ2 ) (−𝜔)
is a germ pseudodifferential operator O𝑤 ◦ /O 𝑤 ◦ ←↪ with principal symbol
𝑝 0 (𝑤, −𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃))| Σ 𝜑 [equal to 𝑝 0 (𝑤 , 𝜁 ) at (𝑧 ◦ , 𝑤 ◦ , 𝜃 ◦ )].
◦ ◦

Theorem 21.2.13 is the version of the Egorov theorem (cf. Theorem 18.5.31) for
germ pseudodifferential operators and FIOs. The associated symplectomorphism is

Ω1 × Ξ1 ∋ (𝑧, −𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜃)) ↦→ (𝑤, 𝜕𝑤 𝜑 (𝑧, 𝑤, 𝜃)) ∈ Ω2 × Ξ2

under the assumptions that 𝜕𝜃 𝜑 (𝑧, 𝑤, 𝜃) = 0 and 𝜕𝑧 𝜑 (𝑧◦ , 𝑤 ◦ , 𝜃 ◦ ) = −𝜁 ◦ .

Remark 21.2.14 By positing dim Θ = 0, i.e., Θ = {𝜃 ◦ }, we deduce from Theorem

21.2.13 the version of the Egorov theorem for classical germ pseudodifferential
operators and classical germ Fourier-like transforms (cf. Remark 20.5.2).
902 21 Germ Pseudodifferential Operators in Complex Space

21.3 Action on distributions

It is reasonable to ask how to get from integrals such as

𝑛 ∫
𝜆
e𝑖𝜆(𝑧−𝑤) ·𝜁 𝑝 (𝑧, 𝜁, 𝜆) ℎ (𝑤, 𝜆) d𝑤d𝜁 (21.3.1)
2𝜋 ℭ

to linear operators acting on germs (at a point 𝑥 ◦ ∈ R𝑛 ) of distributions or hyper-

functions. The answer is fairly simple for distributions when the symbol 𝑝 (𝑧, 𝜁, 𝜆)
is classical as in (21.2.1). Let us take ℭ = 𝑈 × 𝑉 where 𝑈 ∋ 𝑥 ◦ is a domain in R𝑛
and 𝑉 ∋𝜉 ◦ a bounded domain in R𝑛 , 0 ∉ 𝑉.
We use once again the representation of an arbitrary distribution 𝑢 ∈ E ′ (𝑈) as a
finite sum of derivatives of functions 𝑓 ∈ 𝐿 c1 (𝑈). Let 𝛼 ∈ Z+𝑛 ; we can regard
∫ ∞∫ D E
𝐼 𝑘 (𝑥) = e𝑖𝜆( 𝑥−𝑠) · 𝜉 , D 𝛼 𝑓 (𝑠) 𝑝 𝑘 (𝑥, 𝜉) 𝜆 𝑛−𝑘 d𝜉d𝜆
1 𝑉

as an oscillatory integral, as follows. First, we note that

D E
e𝑖𝜆( 𝑥−𝑠) · 𝜉 , D 𝛼 𝑓 (𝑠)
D E
= (−1) | 𝛼 | D𝑠𝛼 e𝑖𝜆( 𝑥−𝑠) · 𝜉 . 𝑓 (𝑠)
∫
= D 𝑥𝛼 e𝑖𝜆( 𝑥−𝑠) · 𝜉 𝑓 (𝑠) d𝑠 = D 𝑥𝛼 e𝑖𝜆𝑥· 𝜉 b𝑓 (𝜆𝜉) .
𝑈

The Fourier transform b 𝑓 is a C 𝜔 function in R𝑛 vanishing at infinity.

Next, we apply the transpose Leibniz rule (1.1.5),

| 𝛼−𝛽 | 𝛼
∑︁
𝛼 𝑖𝜆𝑥· 𝜉 𝛼−𝛽 𝛽
D𝑥 e 𝑝 𝑘 (𝑥, 𝜉) = (−1) D𝑥 e𝑖𝜆𝑥· 𝜉 D 𝑥 𝑝 𝑘 (𝑥, 𝜉) .
𝛽⪯𝛼
𝛽

We see that 𝐼 𝑘 (𝑥) is a linear combination of derivatives of functions

∫ ∞∫
𝛽
𝐼 𝑘,𝛽 (𝑥) = e𝑖𝜆𝑥· 𝜉 D 𝑥 𝑝 𝑘 (𝑥, 𝜉) b
𝑓 (𝜆𝜉) 𝜆 𝑛−𝑘 d𝜉d𝜆.
1 𝔠

Since
𝑓 (𝜆𝜉) = (−1) 𝑛+1 𝜆−2𝑛+2 |𝜉 | −2𝑛−2 Δ
b › 𝑛+1 𝑓 (𝜆𝜉)

and |𝜉 | > 𝛿 > 0 in 𝑉, we obtain

∫ ∞∫
e𝑖𝜆𝑥· 𝜉 |𝜉 | −2𝑛−2 D 𝑥 𝑝 𝑘 (𝑥, 𝜉) Δ 𝑛+1 𝑓 (𝜆𝜉) 𝜆 −𝑘−𝑛−2 d𝜉d𝜆,
𝛽
𝐼 𝑘,𝛽 (𝑥) = (−1) 𝑛+1 ›
1 𝑉

which is of the same kind as 𝐼 𝑘 except that 𝑘 has been replaced by 𝑘 + 𝑛 + 2 and
𝑝 𝑘 (𝑥, 𝜉) by |𝜉 | −2𝑛−2 D 𝑥 𝑝 𝑘 (𝑥, 𝜉). The same argument as that applied to 𝐼 𝑘 shows
𝛽

that 𝐼 𝑘,𝛽 is a linear combination of derivatives of functions

21.4 Action on Hyperfunctions and Microfunctions 903
∫ ∞∫
𝑓 (𝜆𝜉) 𝜆−2 d𝜉d𝜆 (𝛾 ∈ Z+𝑛 );
𝛾
𝐼 𝑘,𝛾 (𝑥) = e𝑖𝜆𝑥· 𝜉 D 𝑥 𝑝 𝑘 (𝑥, 𝜉) b
1 𝑉

such functions are continuous in a neighborhood of 𝑥 ◦ . This proves that if 𝑢 ∈ E ′ (𝑈)

then ∫ ∞∫
(2𝜋) −𝑛 e𝑖𝜆𝑥· 𝜉 𝑝 𝑘 (𝑥, 𝜉) b
𝑢 (𝜉) 𝜆 𝑛−𝑘 d𝜉d𝜆
1 𝑉
is a distribution in a neighborhood of 𝑥 ◦ in R𝑛 . The same is then true if we replace
𝑝 𝑘 by an arbitrary finite realization (21.1.15) of 𝑝.
′
Now consider two finite realizations of 𝑝, 𝑝 (𝑅) and 𝑝 (𝑅 ) , with 𝑅 ′ > 𝑅 > 𝑅◦ and
𝑅◦ sufficiently large that (21.2.4) applies: here we are dealing with an “integral”
∫ ∞∫
e𝑖𝜆𝑥· 𝜉 𝑎 (𝑥, 𝜉, 𝜆) D
š 𝛼 𝑓 (𝜆𝜉) 𝜆 𝑛−𝑘 d𝜉d𝜆, (21.3.2)
1 𝑉
′
where 𝑎 (𝑥, 𝜉, 𝜆) = 𝑝 (𝑅) (𝑥, 𝜉, 𝜆) − 𝑝 (𝑅 ) (𝑥, 𝜉, 𝜆) is extendible as a holomorphic
function of (𝑧, 𝜁) in a complex neighborhood of (𝑥 ◦ , 𝜉 ◦ ) and satisfies there an
inequality |𝑎 (𝑧, 𝜁, 𝜆)| ≲ e−𝜅𝜆 for some 𝜅 > 0. Keeping in mind that D š𝛼 𝑓 (𝜆𝜉) has

polynomial growth as 𝜆 → +∞ we derive right-away that (21.3.2) can be extended

as a holomorphic function in a suitably small complex neighborhood of 𝑥 ◦ . We
reach the conclusion that (21.3.1) defines a linear map 𝑷 𝑥 ◦ of the quotient vector
space D′𝑥 ◦ /C 𝑥𝜔◦ into itself; D′𝑥 ◦ /C 𝑥𝜔◦ can be viewed as the space of germs at 𝑥 ◦ of
analytic-singularity distributions (cf. Definition 11.3.4).
To derive from (21.3.1) an operator acting on singularity hyperfunctions and
thence on microfunctions, one approach is to make use of duality: put ℎ ∈ O (C𝑛 )
into (21.3.1) and let 𝜇 ∈ O ′ (R𝑛 ) act on the resulting integrals. For real phase-
functions the subject has already been dealt with in the last two sections of Ch.
18. At this juncture we shall carry out this part of the discussion after restricting
our choice of phase-functions 𝜑 (𝑥, 𝑦, 𝜃) such that Im 𝜑 (𝑥, 𝑦, 𝜃) > 0 if 𝑥 ≠ 𝑦. This
strengthening of the hypotheses enables us to shift the emphasis from the well-
shapedness of the contours (Definition 20.2.2) to the behavior of Im 𝜑 on contours
that are open subsets of real space.

21.4 Action on Hyperfunctions and Microfunctions

21.4.1 Effective phase-functions

In this section we will need to maintain a sharp distinction between real and complex
domains. Thus Ω shall be a domain in R𝑛 and ΩC one in C𝑛 such that Ω ⊂ ΩC ∩R𝑛 ; Θ
will also be a domain in C𝑛 . Our “central point” will be (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) ∈ Ω × Ω × Θ. We
are going to look at the action of germs FIOs on analytic functionals, hyperfunctions
and microfunctions, starting from integrals
904 21 Germ Pseudodifferential Operators in Complex Space
∫
e𝑖𝜆𝜑 ( 𝑥,𝑠, 𝜃) 𝑎 (𝑥, 𝑠, 𝜃, 𝜆) 𝑓 (𝑠) d𝑠 (21.4.1)
𝑈

where 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) isan analytic amplitude in ΩC × ΩC × Θ and the phase-function

𝜑 ∈ O ΩC × ΩC × Θ satisfies the following conditions [cf. (21.1.4), (21.1.5)]

𝜑 (𝑧, 𝑧, 𝜃) ≡ 0, (21.4.2)
𝜕2 𝜑

det (𝑧, 𝑧, 𝜃) ≠ 0, (21.4.3)
𝜕𝑧 𝑗 𝜕𝜃 𝑘 1≤ 𝑗,𝑘 ≤𝑛
𝜕2 𝜑

det (𝑧, 𝑧, 𝜃) ≠ 0, (21.4.4)
𝜕𝑤 𝑗 𝜕𝜃 𝑘 1≤ 𝑗,𝑘 ≤𝑛

for all (𝑧, 𝑤, 𝜃) ∈ ΩC × ΩC × Θ. [Evidently, (21.4.2)–(21.4.3) implies (21.4.4).]

We stress the fact that the contour of integration in (21.4.1) is an open subset of
R𝑛 , 𝑈 ⊂⊂ Ω. Recall Proposition 21.1.3: if 𝜑 satisfies (21.4.2)–(21.4.3)–(21.4.4)
and 𝑈 is well-shaped for 𝜑 at (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) (Definition 20.2.2) the germ FIO defined
starting from (21.4.1) will necessarily be a germ pseudodifferential operator. The
well-shapedness of 𝑈 for 𝜑 at (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) will be a feature of the restricted class of
phase-functions we now introduce.

Definition 21.4.1 We shall say that a phase-function 𝜑 ∈ O ΩC × ΩC × Θ is
effective if 𝜑 satisfies (21.4.2)–(21.4.3)–(21.4.4) and − Im 𝜑 (𝑥, 𝑦, 𝜃) < 0 for all
(𝑥, 𝑦, 𝜃) ∈ Ω × Ω × Θ such that 𝑥 ≠ 𝑦.
Í 2𝑘
Example 21.4.2 If Ω = R𝑛 let 𝜑 (𝑧, 𝑤, 𝜃) = (𝑧 − 𝑤) · 𝜃 + 𝑖 ⟨𝜃⟩ 𝑛𝑗=1 𝑧 𝑗 − 𝑤 𝑗 with
⟨𝜃⟩ the main branch square-root of 𝜃 · 𝜃 and 1 ≤ 𝑘 ∈ Z+ . Assuming that |Im 𝜃| /|Re 𝜃|
is sufficiently small, 𝜑 is effective.

Proposition 21.4.3 Suppose 𝜑 is effective. If 𝐾1 , 𝐾2 are compact subsets of Ω such

that 𝐾1 ∩ 𝐾2 = ∅ and 𝐾 ∗ is a compact subset of Θ then there is an 𝜀 > 0 such that
− Im 𝜑 (𝑧, 𝑤, 𝜃) < 0 for all (𝑧, 𝑤, 𝜃) ∈ ΩC × ΩC × 𝐾 ∗ such that dist (𝑧, 𝐾1 ) ≤ 𝜀,
dist (𝑤, 𝐾2 ) ≤ 𝜀.

Proof We have Im 𝜑 (𝑥, 𝑦, 𝜃) > 0 in 𝐾1 × 𝐾2 × 𝐾 ∗ , hence there are numbers 𝜀 > 0

and 𝑐 𝜀 > 0 such that Im 𝜑 (𝑧, 𝑤, 𝜃) > 𝑐 𝜀 if dist (𝑧, 𝐾1 ) ≤ 𝜀, dist (𝑤, 𝐾2 ) ≤ 𝜀,
𝜃 ∈ 𝐾 ∗. □
If Ω is convex we must have 𝜀 < 21 dist (𝐾1 , 𝐾2 ) in Proposition 21.4.3, otherwise
there would be an 𝑥 ∈ Ω such that dist (𝑥, 𝐾1 ) ≤ 𝜀, dist (𝑥, 𝐾2 ) ≤ 𝜀, an impossibility
since 𝜑 (𝑥, 𝑥, 𝜃) ≡ 0.

Remark 21.4.4 Effectiveness of the phase-function as defined above is a global

property in Ω × Ω × Θ (as the points 𝑥, 𝑦 are allowed to be arbitrarily distant from one
another) which is unreasonable to expect to hold in a general set-up. But here our
viewpoint is very local, even microlocal (in the limitations of Θ), and it is convenient
to deal with sets ΩC and Θ that are, from the start, as small as needed.
21.4 Action on Hyperfunctions and Microfunctions 905

Proposition 21.4.5 If 𝜑 is effective then 𝜕𝜃 𝜑 (𝑥, 𝑦, 𝜃) ≠ 0 for all (𝑥, 𝑦, 𝜃) ∈ Ω×Ω×Θ,

𝑥 ≠ 𝑦.

Proof If 𝜕𝜃 𝜑 (𝑥, 𝑦, 𝜃) = 0 the Euler homogeneity identity implies 𝜑 (𝑥, 𝑦, 𝜃) = 0.

The latter is impossible unless 𝑥 = 𝑦. □
Thus the (real) critical set Σ 𝜑 of 𝜑 [cf. (20.5.12)] lies above the diagonal of Ω × Ω.
Property (21.4.4) allows us to apply the Implicit Function Theorem and solve the
system of equations
𝜕𝜑
(𝑧, 𝑤, 𝜃) = 𝜁 𝑗 , 𝑗 = 1, ..., 𝑛, (21.4.5)
𝜕𝑧 𝑗
with respect to 𝜃; the solution 𝜃 (𝑧, 𝑤, 𝜁) will be holomorphic in ΩC × ΩC × ΞC , with
ΩC possibly contracted about 𝑥 ◦ and ΞC a domain in C𝑛 ; if 𝜉 ◦ = 𝜕𝑧 𝜑 (𝑥 ◦ , 𝑥 ◦ , 𝜃 ◦ ) we
will have 𝜃 ◦ = 𝜃 (𝑥 ◦ , 𝑥 ◦ , 𝜉 ◦ ). If 𝜑 is effective so is the phase-function 𝜑˜ (𝑧, 𝑤, 𝜁) =
𝜑 (𝑧, 𝑤, 𝜃 (𝑧, 𝑤, 𝜁)); if 𝜉 ∈ Ξ = ΞC ∩ R𝑛 we have 𝜕𝑧 𝜑 (𝑧, 𝑤, 𝜉) = 𝜉.

21.4.2 Action on holomorphic functions in wedges

Throughout the sequel we reason under the hypothesis that 𝜑 is effective. Recall the
notion of a wedge W𝛿 (𝑈, Γ) = {𝑧 ∈ C𝑛 ; 𝑥 ∈ 𝑈, 𝑦 ∈ Γ, |𝑦| < 𝛿}, 𝑈 an open subset
of R𝑛 , Γ a convex open cone in R𝑛 with vertex at 0, and 𝛿 > 0; we assume that
W𝛿 (𝑈, Γ) ⊂⊂ ΩC (in particular, 𝑈 ⊂⊂ Ω). Given ℎ ∈ O (W𝛿 (𝑈, Γ)) we define
∫
e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ (𝑧) d𝑧,
♮
𝑨𝑈,𝑦 ◦ ℎ (𝑤, 𝜃, 𝜆, 𝜏) = (21.4.6)
𝑈+𝑖 𝜏 𝑦 ◦

where 𝑦 ◦ ∈ Γ ∩ S𝑛−1 , 𝑈 + 𝑖 [0, 𝛿] 𝑦 ◦ ⊂ ΩC , 0 < 𝜏 < 𝛿, 𝜆 > 0; for each (𝜆, 𝜏) the
♮
function (𝑤, 𝜃) ↦→ 𝑨𝑈,𝑦 ◦ ℎ (𝑤, 𝜃, 𝜆, 𝜏) is holomorphic in ΩC × Θ.

Lemma 21.4.6 Suppose that 𝜕𝑈 is smooth and that there is a 𝑐 ◦ > 0 such that

∀𝑦 ∈ Γ, 𝑦 · 𝜉 ◦ ≥ 𝑐 ◦ |𝑦| . (21.4.7)

Let 𝑉 ⊂⊂ Ω be an open set in R𝑛 such that dist (𝑉, 𝜕𝑈) > 0 and 𝐾 ∗ a compact
subset of Θ. Provided 𝑈 ∪ 𝑉 and 𝐾 ∗ are sufficiently small there are positive numbers
𝜏 ◦ ∈ (0, 𝛿), 𝜀◦ , 𝜅, such that

𝑨𝑈,𝑦 ◦ ℎ (𝑤, 𝜃, 𝜆, 𝜏) ≤ 𝐶e−𝜅𝜆

♮
max |ℎ| (21.4.8)
𝑈+𝑖 [𝜏, 𝜏 ◦ ] 𝑦 ◦

for all ℎ ∈ O (W𝛿 (𝑈, Γ)), 𝑤 = 𝑠 + 𝑖𝑡, 𝑠 ∈ 𝑉, |𝑡| < 𝜀◦ , 𝜃 ∈ Θ, 𝜆 > 0, 0 < 𝜏 ≤ 𝜏 ◦ .

Keep in mind that, generally speaking, none of the two sides in (21.4.8) is bounded
as 𝜏 ↘ 0. It must be stressed, however, that 𝜅 and 𝐶 are independent of 𝜏 ≤ 𝜏 ◦ .
906 21 Germ Pseudodifferential Operators in Complex Space

Proof We deform the contour of integration in (21.4.6) from 𝑈 + 𝑖𝜏𝑦 ◦ to 𝑈 + 𝑖𝜏 ◦ 𝑦 ◦ .

The Cauchy Integral Theorem yields
∫
e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ (𝑧) d𝑧
♮
𝑨𝑈,𝑦 ◦ ℎ (𝑤, 𝜃, 𝜆, 𝜏) = (21.4.9)
𝑧 ∈𝑈+𝑖 𝜏 ◦ 𝑦 ◦
∫ 𝜏◦ ∫
− e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) ℎ (𝑧) d𝜎𝑈 (𝑥) d𝜏 ′,
𝜏 𝑧 ∈𝜕𝑈+𝑖 𝜏 ′ 𝑦 ◦

where d𝜎𝑈 (𝑥) is the appropriate volume element on 𝜕𝑈. Proposition 21.4.3 implies
that − Im 𝜑 (𝑧, 𝑤, 𝜃) ≤ −2𝜅 for some 𝜅 > 0, all 𝑧 ∈ 𝜕𝑈 + 𝑖𝜏𝑦 ◦ , 0 ≤ 𝜏 ≤ 𝜏 ◦ , and all
𝑤 = 𝑠 + 𝑖𝑡, 𝑠 ∈ 𝑉, |𝑡| < 𝜀◦ (for some 𝜀◦ > 0). We take into account the fact that to
every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that

|𝑎 (𝑧, 𝑠 + 𝑖𝑡, 𝜃, 𝜆)| ≤ 𝐶 𝜀 e 𝜀𝜆

for all 𝑧 ∈ 𝑈 + 𝑖𝜏𝑦 ◦ , 0 ≤ 𝜏 ≤ 𝜏 ◦ , 𝑠 ∈ 𝑉, |𝑡| < 𝜀◦ , 𝜃 ∈ Θ. Selecting 𝜀 < 𝜅 implies

directly an upper bound of the type (21.4.8) for the second integral in the right-hand
side of (21.4.9). To deal with the first integral we observe that

𝜑 (𝑧, 𝑤, 𝜃) = 𝜉 ◦ · (𝑥 − 𝑥 ◦ + 𝑖𝑦) + 𝑂 |𝑧 − 𝑥 ◦ | 2 + |𝑤 − 𝑥 ◦ | + |𝜃 − 𝜃 ◦ | (21.4.10)

whence, by (21.4.7),

− Im 𝜑 (𝑥 + 𝑖𝜏 ◦ 𝑦 ◦ , 𝑤, 𝜃) ≤ −𝑐 ◦ 𝜏 ◦ − 𝐶 |𝑥 − 𝑥 ◦ | 2 + |𝑤 − 𝑥 ◦ | + |𝜃 − 𝜃 ◦ | .

We select 𝑈, 𝑉, Θ sufficiently small that − Im 𝜑 (𝑥 + 𝑖𝜏 ◦ 𝑦 ◦ , 𝑤, 𝜃) ≤ − 21 𝑐 ◦ 𝜏 ◦ ≤ −𝜅

for all (𝑥, 𝑤, 𝜃) ∈ 𝑈 × 𝑉 × Θ. □

Corollary 21.4.7 Assume the same hypotheses as in Lemma 21.4.6. Provided

diam 𝑈, diam 𝑉 and diam Γ ∩ S𝑛−1 are sufficiently small and the number 𝜅 is as in
♮
(21.4.8) then, for arbitrary ℎ ∈ O ΩC , the integral 𝑨𝑈,𝑦 ◦ ℎ (𝑤, 𝜃, 𝜆, 𝜏) converges
♮
uniformly to a function 𝑨𝑈 ℎ (𝑤, 𝜃, 𝜆) as 𝜏 ↘ 0. We have

𝑨𝑈 ℎ (𝑤, 𝜃, 𝜆) ≲ e−𝜅𝜆
♮
max |ℎ| (21.4.11)
𝑈+𝑖 [0, 𝜏 ◦ ] 𝜉 ◦

for all 𝑤 = 𝑠 + 𝑖𝑡, 𝑠 ∈ 𝑉, |𝑡| < 𝜀◦ , 𝜃 ∈ Θ, 𝜆 > 0.

Proof If ℎ ∈ O ΩC we can let 𝜏 ↘ 0 in (21.4.8). □
21.4 Action on Hyperfunctions and Microfunctions 907

21.4.3 Action on analytic functionals

We continue to assume that 𝜑 is effective. We associate to each compact subset 𝐾 of

Ω the exponent-weight

Φ𝐾 (𝑧, 𝜃) = sup (− Im 𝜑 (𝑧, 𝑤, 𝜃)) . (21.4.12)

𝑤 ∈𝐾

It follows from Proposition 11.2.8 that Φ𝐾 (𝑧, 𝜃) is plurisubharmonic. We define, for

𝜇 ∈ O ′ (Ω),
D E
𝑨𝜇 (𝑧, 𝜃, 𝜆) = 𝜇 𝑤 , e𝑖𝜆𝜑 (𝑧,𝑤, 𝜃) 𝑎 (𝑧, 𝑤, 𝜃, 𝜆) , 𝜆 > 0. (21.4.13)

If 𝜇 ∈ O ′ (𝐾) then to every 𝜀 > 0 there is a 𝐶 𝜀 > 0 such that

| 𝑨𝜇 (𝑧, 𝜃, 𝜆)| ≤ 𝐶 𝜀 exp (𝜆Φ𝐾 (𝑧, 𝜃) + 𝜀𝜆) (21.4.14)

for all (𝑧, 𝜃) ∈ ΩC × Θ, 𝜆 > 0. Thus 𝑨 defines a continuous linear map

O ′ (𝐾) −→ O (Φ𝐾 ) ΩC × Θ , the continuity following easily from the definitions of
the topologies of O ′ (Ω) and O (Φ𝐾 ) ΩC × Θ .

Proposition 21.4.8 Let 𝑈 ⊂⊂ Ω be an open set and 𝐾 be a compact subset of

Ω. If 𝐾 ∩ 𝑈 = ∅ there is an open subset 𝑈 C of C𝑛 , 𝑈 C ∩ R𝑛 ⊃ 𝑈, such that
𝑨𝜇 (𝑧, 𝜃, 𝜆) ∈ O (−𝜔) 𝑈 C × Θ for all 𝜇 ∈ O ′ (𝐾).

Proof Let 𝑈 ′ ⊂⊂ 𝑈 be open; then dist ′

(𝑈 , 𝐾) > 0; Proposition 21.4.3 implies
′
Φ𝐾 (𝑧, 𝜃) ≤ −𝑐 ◦ for some 𝑐 ◦ > 0 if dist 𝑧, 𝑈 is sufficiently small. Taking 𝜀 < 𝑐 ◦ /2
in (21.4.14) implies | 𝑨𝜇 (𝑧, 𝜃, 𝜆)| ≤ 𝐶 𝜀 exp (−𝜆𝑐 ◦ /2). □

Corollary 21.4.9 There is an open subset 𝑈 C of C𝑛 , 𝑈 C ∩ R𝑛 = 𝑈, such that

𝑨𝜇 (𝑧, 𝜃, 𝜆) ∈ O (−𝜔) 𝑈 C for all 𝜇 ∈ O ′ (𝜕𝑈).

Let now 𝜇 ∈ O ′ (Ω) and ℎ ∈ O ΩC , both arbitrary; let 𝑈 ⊂⊂ Ω be an open set.

We define
∫
⟨𝑨𝑈 𝜇 (𝜃, 𝜆) , ℎ⟩ = 𝑨𝜇 (𝑥, 𝜃, 𝜆) ℎ (𝑥) d𝑥 (21.4.15)
𝑈
∫
= 𝜇𝑤 , e𝑖𝜆𝜑 ( 𝑥,𝑤, 𝜃) 𝑎 (𝑥, 𝑤, 𝜃, 𝜆) ℎ (𝑥) d𝑥 ;
𝑈

in other words (cf. Corollary 21.4.7)

D E
♮
⟨𝑨𝑈 𝜇 (𝜃, 𝜆) , ℎ⟩ = 𝜇 𝑤 , 𝑨𝑈 ℎ (𝑤, 𝜃, 𝜆) . (21.4.16)

It is readily seen that 𝑨𝑈 𝜇 (𝜃, 𝜆) is an analytic functional in R𝑛 with supp 𝑨𝑈 𝜇 ⊂ 𝑈,

depending holomorphically on 𝜃.
908 21 Germ Pseudodifferential Operators in Complex Space

Proposition 21.4.10 Let 𝑈, 𝐾, be as in Proposition 21.4.8 and 𝐾 ∗ be a compact

subset of Ξ ∩ S𝑛−1 . If 𝜇 ∈ O ′ (𝐾) then the analytic functional 𝑨𝑈 𝜇 (𝜁, 𝜆) in C𝑛 [see
(21.4.16)] is carried by 𝑈 and to every 𝜀 > 0 there are positive numbers 𝜅, 𝐶, such
that
|⟨𝑨𝑈 𝜇 (𝜉, 𝜆) , ℎ⟩| ≤ 𝐶e−𝜅𝜆 sup |ℎ (𝑧)| (21.4.17)
dist(𝑧,𝑈) < 𝜀

for all ℎ ∈ O (C𝑛 ), 𝜉∈ 𝐾 ∗, 𝜆 ≥ 1.

Proof Indeed, there are positive constants 𝜀, 𝐶 ′ such that, for all ℎ ∈ O (C𝑛 ), 𝜉 ∈ 𝐾 ∗ ,
𝜆 ≥ 1, D E
𝜇 (𝜉, 𝜆) , 𝑨𝑈 ℎ ≤ 𝐶 ′ sup 𝑨𝑈 ℎ (𝑤, 𝜉, 𝜆) ,
♮ ♮

𝑤 ∈𝐾 𝜀C

where 𝐾 C𝜀 = {𝑤 ∈ C𝑛 ; Re 𝑤 ∈ 𝐾, |Im 𝑤| < 𝜀} ⊂⊂ ΩC . The claim then follows

from (21.4.15) and the definition (21.4.16). □
Note that (21.4.17) is the same as
∫
𝑨𝜇 (𝑥, 𝜉, 𝜆) ℎ (𝑥) d𝑥 ≤ 𝐶e−𝜅𝜆 sup |ℎ| . (21.4.18)
𝑈 dist(𝑧,𝑈) < 𝜀

21.4.4 Integrating 𝑨𝑼 𝝁

It is now convenient to assume that the effective phase-function 𝜑 (𝑧, 𝑤, 𝜁) is such

that 𝜕𝑧 𝜑 (𝑧, 𝑧, 𝜁) ≡ 𝜁 in ΩC × ΩC × ΞC [cf. (21.4.5)]. It is also convenient to assume
that ΩC is a Rungedomain, meaning that the restrictions of entire functions in C𝑛
are dense in O ΩC .
Proposition 21.4.11 Let 𝑈, 𝑈 C , 𝐾 be as in Proposition 21.4.8, 𝐾 ∗ a compact
subset of Ξ = ΞC∫ ∩ R𝑛 and 𝑁 ∈ Z+ . If 𝜇 ∈ O ′ (𝐾) and if 𝜀 > 0 is suffi-
+∞
ciently small then 1 𝑨𝜇 (𝑧, 𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 is a holomorphic function of (𝑧, 𝜁) in
𝑈 C × {𝜁 ∈ C𝑛 ; dist (𝜁, 𝐾 ∗ ) < 𝜀}.
This follows directly from Proposition 21.4.8.
Proposition 21.4.12 Let 𝑈, 𝐾, be as in Proposition 21.4.8 and let Ξ′ a domain in
R𝑛 , Ξ′ ⊂⊂ Ξ. If 𝜇 ∈ O ′ (𝐾) and 𝑁 ∈ Z+ the integral
∫ +∞
𝑨𝑈 𝜇 (𝜉, 𝜆) 𝜆 𝑁 −1 d𝜆 (21.4.19)
1

is a C 𝜔 function of 𝜉 ∈ Ξ′ ∩ S𝑛−1 valued in O ′ 𝑈 .

Proof Let 𝐾 ∗ be the closure of Ξ′. We derive from (21.4.17) that (21.4.19) is a
holomorphic function of 𝜁 ∈ ΞC , dist (𝜁, 𝐾 ∗ ) sufficiently small, and that to every
𝜀 > 0 there is a 𝜅 > 0 such that
21.4 Action on Hyperfunctions and Microfunctions 909
∫ +∞
|⟨𝑨𝑈 𝜇 (𝜁, 𝜆) , ℎ⟩| 𝜆 𝑁 −1 d𝜆 ≲ 𝑁!𝜅 −𝑁 sup |ℎ (𝑧)| (21.4.20)
1 dist(𝑧,𝑈) < 𝜀

for these same 𝜁. Then the claim follows

from Definition 6.1.3 and the definition of
′
the Fréchet space topology on O 𝑈 . □

Thus, under the hypotheses ∫of Proposition 21.4.12, whatever 𝜇 ∈ O ′ (𝐾) and
+∞
𝑁 ∈ Z+ , 𝑁 ≥ 1, the integral 1 𝑨𝑈 𝜇 (𝜉, 𝜆) 𝜆 𝑁 −1 d𝜆 is a C 𝜔 function of 𝜉 ∈

Ξ ∩ S𝑛−1 valued in O ′ 𝑈 . Let d𝑆 (𝜉) denote the “volume” element in S𝑛−1 such
that 𝜆 𝑛−1 d𝜆d𝑆 (𝜉) = d𝜉 if 𝜆 = |𝜉 |.

Corollary 21.4.13 Let 𝑈 and 𝐾 be as in Proposition 21.4.8. If Γ is an open cone in

R𝑛 \ {0} such that Γ ∩ S𝑛−1 ⊂⊂ Ξ then, for every 𝜇 ∈ O ′ (𝐾), the integral
∫ +∞ ∫ ∫
𝑨𝑈 𝜇 (𝜉, 𝜆) 𝜆 𝑛−1 d𝜆d𝑆 (𝜉) = 𝑨𝑈 𝜇 (𝜉/|𝜉 | , |𝜉 |) d𝜉 (21.4.21)
1 𝜉 ∈Γ∩S𝑛−1 Γ

represents an analytic functional in C𝑛 carried by 𝑈.

Obviously, in Corollary 21.4.13 the density d𝑆 (𝜉) can be replaced by any distri-
bution in Γ ∩ S𝑛−1 .
If 𝜑 (𝑥, 𝑤, 𝜉) stands for |𝜉 | 𝜑 (𝑥, 𝑤, 𝜉/|𝜉 |) and 𝑎 (𝑧, 𝑤, 𝜉) for 𝑎 (𝑥, 𝑤, 𝜉/|𝜉 | , |𝜉 |)
we see that
∫ ∫ ∫ ∫
𝑖 𝜑 ( 𝑥,𝑤, 𝜉 )
𝑨𝜇 (𝑥, 𝜉/|𝜉 | , |𝜉 |) ℎ (𝑥) d𝑥d𝜉 = 𝜇 𝑤 , e 𝑎 (𝑥, 𝑤, 𝜉) ℎ (𝑥) d𝑥d𝜉
𝑈 Γ 𝑈 Γ
(21.4.22)
whatever ℎ ∈ O (C𝑛 ).

Remark 21.4.14 If 𝜇 is a compactly supported distribution the same is true of

(21.4.19).

21.4.5 Action on hyperfunctions and mapping into singularity

hyperfunctions

We continue to deal with an effective phase-function 𝜑 (𝑧, 𝑤, 𝜁) such that 𝜕𝑧 𝜑 (𝑧, 𝑧, 𝜁)

= 𝜁; if 𝐾 is a compact subset of Ω the exponent-weight Φ𝐾 is defined in (21.4.12).

Proposition 21.4.15 Let 𝑈, 𝐾 and 𝑈 C be as in Proposition 21.4.8; suppose, more-

over, that there is a bounded domain 𝑈 ′ in R𝑛 such that 𝑈 ⊂⊂ 𝑈 ′ and 𝜕𝑈 ′ ⊂ 𝐾. The
operator 𝑨 acting on analytic functionals supported in 𝑈 [cf. (21.4.12)] induces a
linear map

B (𝑈) −→ O (Φ𝐾 ) 𝑈 C × Ξ /O (−𝜔) 𝑈 C × Ξ . (21.4.23)
910 21 Germ Pseudodifferential Operators in Complex Space

Proof Each 𝑓 ∈ B (𝑈) can be defined by an analytic functional 𝜇 ∈ O ′ (𝑈 ′ ∪ 𝐾).

Let 𝜇 𝑗 ∈ O ′ (𝑈 ′ ∪ 𝐾), 𝑗 = 1, 2, define the same hyperfunction 𝑓 in 𝑈 ′. This implies
supp (𝜇1 − 𝜇2 ) ⊂ 𝐾 and, as a consequence of Proposition 21.4.8,

𝑨 (𝜇1 − 𝜇2 ) (𝑧, 𝜁, 𝜆) ∈ O (−𝜔) 𝑈 C × Ξ . (21.4.24)

□
The same argument allows us to exploit the restriction mappings from 𝑈 to a
smaller domain in R𝑛 and thence define germ maps

A 𝑥 ◦ (𝜁) : B 𝑥 ◦ −→ O 𝑥(Φ◦ 𝐾 ) /O 𝑥(−𝜔)

◦ (21.4.25)

where 𝑥 ◦ ∈ Ω, 𝜁 ∈ Ξ.

Proposition 21.4.16 Let 𝑈, 𝐾, 𝛿 and 𝑈 C be asin Proposition 21.4.8, 𝐾 ∗ be a compact

∫ +∞
subset of Ξ and 𝑁 ∈ Z+ . The operator O ′ 𝑈 ∋ 𝜇 ↦→ 1 𝑨𝑈 𝜇 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆
(1 ≤ 𝑁 ∈ Z+ ) induces a linear map B (𝑈) −→ B (𝑈) /C 𝜔 (𝑈) depending on
𝜁 ∈ ΞC , dist (𝜁, 𝐾 ∗ ) sufficiently small.
∫ +∞
Proof It follows directly from Proposition 21.4.12 that 𝜇 ↦→ 1 𝑨𝑈 𝜇 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆

induces a continuous linear endomorphism of O ′ 𝑈 depending holomorphically

on 𝜁 ∈ ΞC , dist (𝜁, 𝐾 ∗ ) sufficiently small. If 𝜇 𝑗 ∈ O ′ 𝑈 , 𝑗 = 1, 2, define the same
hyperfunction in 𝑈 we have supp (𝜇1 − 𝜇2 ) ⊂ 𝜕𝑈; then Proposition 21.4.8 implies
(21.4.24). It follows that
∫ +∞
( 𝑨𝑈 𝜇2 (𝜁, 𝜆) − 𝑨𝑈 𝜇1 (𝜁, 𝜆)) 𝜆 𝑁 −1 d𝜆 ∈ O 𝑈 C × ΞC ,
1
∫ +∞
implying that 1 𝑨𝑈 𝜇 𝑗 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆, 𝑗 = 1, 2, define the same singularity hyper-
function in 𝑈. □

Definition 21.4.17 Given a hyperfunction 𝑓 in 𝑈 defined by 𝜇 ∈ O ′ 𝑈 we shall
∫ +∞
denote by 1 𝑨𝑈 𝑓 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 the singularity hyperfunction in 𝑈 defined by the
analytic functional (21.4.19).
∫ +∞
Let 1 [Af (𝜁, 𝜆)] 𝑥 ◦ 𝜆 𝑁 −1 d𝜆 denote the germ at 𝑥 ◦ of the singularity hyperfunc-
∫ +∞
tion 1 𝑨𝑈 𝑓 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆. Inspection of the proof of Proposition 21.4.16 shows
that contracting 𝑈 about 𝑥 ◦ leads to a linear map (depending on 𝜁 in a neighborhood
of Ξ ∩ S𝑛−1 in C𝑛 ),
∫ +∞
sing
B 𝑥 ◦ ∋ f 𝑥 ◦ ↦→ [Af (𝜁, 𝜆)] 𝑥 ◦ 𝜆 𝑁 −1 d𝜆 ∈ B 𝑥 ◦ /C 𝑥𝜔◦ = B 𝑥 ◦ . (21.4.26)
1
21.4 Action on Hyperfunctions and Microfunctions 911

When 𝑁 = 𝑛 and Γ is an open cone in R𝑛 \ {0} such that Γ ∩ S𝑛−1 ⊂⊂ Ξ, (21.4.22)

leads to the map
∫
sing
B 𝑥 ◦ ∋ f 𝑥 ◦ ↦→ [Af (𝜉/|𝜉 | , |𝜉 |)] 𝑥 ◦ d𝜉 ∈ B 𝑥 ◦ . (21.4.27)
Γ

21.4.6 Effect on the analytic wave-front set. Action on singularity

hyperfunctions and microfunctions

We have not shown that the kernels of the germ maps (21.4.26), (21.4.27) contain
C 𝑥𝜔◦ . We now prove the microlocal (and therefore stronger) version of this result.
We continue to deal with an effective phase-function 𝜑 such that 𝜕𝑧 𝜑 (𝑤, 𝑤, 𝜁) =
𝜁. We go back to (and use the notation in) the proof of Lemma 21.4.6 where, however,
we exchange the roles of 𝑧 and 𝑤 and replace 𝜉 ◦ = −𝜕𝑤 𝜑 (𝑥 ◦ , 𝑥 ◦ , 𝜉 ◦ ) by −𝜉 ◦ [cf.
(21.4.2) with 𝜁 substituted for 𝜃]. Here we assume that W𝛿 (𝑈, −Γ) ⊂⊂ ΩC . Given
ℎ ∈ O (W𝛿 (𝑈, −Γ)) and the fact that 𝑣 ◦ ∈ Γ =⇒ (−𝑣 ◦ ) · (−𝜉 ◦ ) > 0 we can replace
everywhere (in the proof of Lemma 21.4.6) 𝑣 ◦ by −𝑣 ◦ ; we change notation [compared
to (21.4.6)]: if 0 < 𝜏 < 𝜏 ◦ < 𝛿,
∫
𝑨𝑈,𝑣 ℎ (𝑧, 𝜁, 𝜆, 𝜏) =
◦ e𝑖𝜆𝜑 (𝑧,𝑤,𝜁 ) 𝑎 (𝑧, 𝑤, 𝜁, 𝜆) ℎ (𝑤) d𝑤 (21.4.28)
𝑈−𝑖 𝜏𝑣 ◦

is a holomorphic function of (𝑧, 𝜁) in ΩC × ΞC and (21.4.8) now reads

𝑨𝑈,𝑣 ◦ ℎ (𝑧, 𝜁, 𝜆, 𝜏) ≲ e−𝜅𝜆 max |ℎ| (21.4.29)

𝑈−𝑖 [𝜏, 𝜏 ◦ ]𝑣 ◦

for all 𝑧 ∈ ΩC , 𝑥 ∈ 𝐾, |𝑦| < 𝜀, 𝜁 ∈ ΞC , |𝜁 − 𝜉 ◦ | < 𝜀 ≪ 1, 𝜆 ≥ 1, 0 < 𝜏 ≤ 𝜏 ◦ < 𝛿.

′ ◦ defined as follows
We introduce the analytic functionals 𝜇𝑈 ℎ, 𝜏 ∈ O 𝑈 − 𝑖𝜏𝑣

[cf. (7.2.1)]: if ℎ1 ∈ O ΩC ,
D E ∫
𝜇𝑈
ℎ, 𝜏 1 =
, ℎ ℎ (𝑤) ℎ1 (𝑤) d𝑤 (21.4.30)
𝑈−𝑖 𝜏𝑣 ◦
∫
= ℎ (𝑠 − 𝑖𝜏𝑣 ◦ ) ℎ1 (𝑠 − 𝑖𝜏𝑣 ◦ ) d𝑠.
𝑈

As 𝜏 ↘ 0 the analytic functionals 𝜇𝑈 ℎ, 𝜏 converge (in the sense of Property ♣, Lemma

′ ◦
ℎ ∈ O 𝑈 independent of the choice of 𝑣 ∈ Γ. By
7.2.2) to an analytic functional 𝜇𝑈
Definition 7.2.3, 𝜇𝑈ℎ defines the hyperfunction 𝑏𝑈 ℎ in 𝑈 ⊂ R𝑠 ⊂ C𝑤 ; by Definition
𝑛 𝑛
◦ ◦ ◦
7.4.7, 𝑏𝑈 ℎ is microanalytic at (𝑥 , 𝜉 ) (since −𝑦 ∈ Γ =⇒ 𝑦 · 𝜉 < 0). By Lemma
7.2.2, to every 𝜀 > 0 there is a 𝜏 (𝜀) ∈ (0, 𝜏 ◦ ) and 𝐶 𝜀 > 0 such that
D E
𝜇𝑈 𝑈
ℎ − 𝜇 ℎ, 𝜏 ( 𝜀) , ℎ1 ≤ 𝐶 𝜀 max |ℎ1 (𝑤)| (21.4.31)
dist(𝑤,𝜕𝑈) ≤ 𝜀
912 21 Germ Pseudodifferential Operators in Complex Space

for all ℎ1 ∈ O ΩC . In what follows 𝑨𝜇𝑈
ℎ
(𝑧, 𝜁, 𝜆) is defined by (21.4.12) where
𝜇 = 𝜇 ℎ (and 𝜁 replaces 𝜃).
𝑈

Lemma 21.4.18 Let 𝑈, 𝑉 be domains in R𝑛 , 𝑉 ⊂⊂ 𝑈 ⊂⊂ Ω, satisfying the hypothe-

ses in Lemma 21.4.6; suppose |𝜉 ◦ | = 1. If the positive constants 𝑐 ◦ , 𝜀 are sufficiently
small and ℎ ∈ O (W𝛿 (𝑈, −Γ)) then 𝑨𝜇𝑈 ℎ
(𝑧, 𝜁, 𝜆) is a holomorphic function of
(𝑧, 𝜁) in Ω × Ξ such that
C C

−𝑐◦ 𝜆
∀𝜆 > 0, 𝑨𝜇𝑈
ℎ (𝑧, 𝜁, 𝜆) ≤ 𝐶◦ e (21.4.32)

for some 𝐶◦ > 0 and all 𝑧 = 𝑥 + 𝑖𝑦, 𝑥 ∈ 𝑉, |𝑦| < 𝜀, 𝜁 ∈ ΞC , |𝜁 − 𝜉 ◦ | < 𝜀.

Proof We put ℎ1 = 𝑎 (𝑧, 𝑤, 𝜁, 𝜆) exp (𝑖𝜆𝜑 (𝑧, 𝑤, 𝜁)) in (21.4.30), in which case [cf.
(21.4.6)] D E
𝜇𝑈
ℎ, 𝜏 , ℎ 1 = 𝑨𝑈,𝑣 ◦ ℎ (𝑧, 𝜁, 𝜆, 𝜏) (21.4.33)

is a holomorphic function of (𝑧, 𝜁) ∈ ΩC × ΞC (0 < 𝜏 < 𝜏 ◦ ). Let 𝑉 ⊂⊂ 𝑈 be a

suitably small neighborhood of 𝑥 ◦ in R𝑛 ; it follows from the estimates in the proof
of Lemma 21.4.6 that there are positive constants 𝐶, 𝑐 1 such that

∀𝜆 ≥ 1, max |ℎ1 (𝑧, 𝑤, 𝜁, 𝜆)| ≤ 𝐶e−𝑐1 𝜆 (21.4.34)

dist(𝑤,𝜕𝑈) ≤ 𝜀

for all 𝜀 > 0 sufficiently small and all 𝑧 = 𝑥 +𝑖𝑦, 𝑥 ∈ 𝑉, |𝑦| < 𝜀, 𝜁 ∈ ΞC , |𝜁 − 𝜉 ◦ | < 𝜀,
𝜆 ≥ 1. Combining this with (21.4.29) and (21.4.33) yields
D E
−𝜅𝜆
∀𝜆 ≥ 1, 𝜇𝑈 ℎ, 𝜏 1 ≲ e
, ℎ max |ℎ| .
𝑈−𝑖 [𝜏, 𝜏 ◦ ]𝑣 ◦

ℎ , ℎ1 = 𝑨𝜇 ℎ
𝑈 (𝑧,
Formula (21.4.6) implies 𝜇𝑈 𝜁, 𝜆), whence, by (21.4.31) and for
the same (𝑧, 𝜁, 𝜆) as in (21.4.34),
D E
∀𝜆 ≥ 1, 𝑨𝜇𝑈 ℎ − 𝜇𝑈
ℎ, 𝜏 , ℎ 1 ≤ 𝐶e−𝑐◦ 𝜆 ,

whence the claim. □

Corollary 21.4.19 Let 𝑈, 𝑉 and 𝜉 ◦ be as in Lemma∫ +∞ 21.4.18. If 𝜀 > 0 is sufficiently
small and ℎ ∈ O (W𝛿 (𝑈, −Γ)) then the integral 1 𝑨𝜇𝑈 ℎ
(𝑧, 𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 (𝑁 > 0)
is a holomorphic function of (𝑧, 𝜁), 𝑥 ∈ 𝑉, |𝑦| < 𝜀, 𝜁 ∈ Ξ , |𝜁 − 𝜉 ◦ | < 𝜀.
C

Proof We know that 𝑨𝜇𝑈 (𝑧, 𝜁, 𝜆) is holomorphic with respect to (𝑧, 𝜁) ∈ ΩC × ΞC ;

∫ ℎ+∞
(21.4.32) implies that 1 𝑨𝜇𝑈 ℎ
(𝑧, 𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 is holomorphic with respect to
(𝑧, 𝜁). □
If we define 𝑨𝑏𝑈 ℎ (𝑥, 𝜆) to be the hyperfunction defined in 𝑉 by 𝑨𝜇𝑈 ℎ
(𝑧, 𝜁, 𝜆)
we obtain the linear map
∫ +∞
O (W𝛿 (𝑈, Γ)) ∋ ℎ ↦→ 𝑨𝑏𝑈 ℎ (𝑥, 𝜆) 𝜆 𝑁 −1 d𝜆 ∈ C 𝜔 (𝑉) .
1
21.4 Action on Hyperfunctions and Microfunctions 913

◦ ◦
Suppose a hyperfunction 𝑓 in R is microanalytic at (𝑥 , 𝜉 ) ∈
Proposition 21.4.20 𝑛

ΩC × ΞC ∩ S𝑛−1 . If 𝑈 ⊂ ΩC∫ is a sufficiently small neighborhood of 𝑥 ◦ in R𝑛 then

+∞
the singularity hyperfunction 1 𝑨𝑈 𝑓 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 (Definition 21.4.17) vanishes
identically in a neighborhood 𝑉 ⊂⊂ 𝑈 of 𝑥 ◦ in R𝑛 (𝑁 > 0, 𝜁 ∈ ΞC , |𝜁 − 𝜉 ◦ | < 𝜀 ≪
1).

Proof Let ℎ 𝑗 ∈ O W2 𝛿 𝑈, Γ 𝑗 ( 𝑗 = 1, ..., 𝜈) Í be such that every cone Γ 𝑗 is
· ◦ < 0 and 𝑓 = 𝜈
contained
Í𝜈 in the half-space 𝑦 𝜉 𝑗=1 𝑏𝑈 ℎ 𝑗 . We define 𝑨𝑈 𝑓 =
𝑗=1 𝑨𝑏𝑈 ℎ 𝑗 . By Corollary 21.4.19, if 𝑈, 𝑉 ⊂⊂ 𝑈, 𝜀 > 0, are sufficiently small then
∫ +∞
𝑨𝑏𝑈 ℎ 𝑗 (𝑥, 𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 ∈ C 𝜔 (𝑉) , 𝑗 = 1, ..., 𝜈,
1

for all 𝜁 ∈ ΞC , |𝜁 − 𝜉 ◦ | < 𝜀, whence the claim. □

In accordance with Definition 7.4.7, we can state

Corollary 21.4.21 Let 𝑓 be a hyperfunction in ΩC . To every open subset 𝑈 of R𝑛 ,

𝑈 ⊂⊂ Ω, and every 𝑁 > 0 there is an 𝜀 > 0 such that
∫ +∞
𝑊 𝐹a 𝑨𝑈 𝑓 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 ⊂ 𝑊 𝐹a ( 𝑓 ) (21.4.35)
1

for all 𝜁 ∈ ΞC , |𝜁 − 𝜉 ◦ | < 𝜀.

We avail ourselves of the fact that 𝜉 ◦ can be any point in Ξ ∩ S𝑛−1 [by (21.4.2)–
(21.4.3)]. By letting 𝜉 ◦ range over S𝑛−1 we get immediately

Corollary 21.4.22 Suppose S𝑛−1 ⊂ Ξ. If the hyperfunction 𝑛

∫ +∞𝑓 in R is real-analytic
◦
in a neighborhood of 𝑥 in R the same is true of 𝑥 ↦→ 1 𝑨𝑈 𝑓 (𝜁, 𝜆) 𝜆 𝑁 −1 d𝜆 as
𝑛

𝜁 ranges over a sufficiently thin neighborhood of S𝑛−1 in ΞC .

sing
Corollary 21.4.23 The map (21.4.26) induces a linear endomorphism B 𝑥 ◦ ←↪
depending on 𝜉 ∈ Ξ ∩ S𝑛−1 ,
∫ +∞
[f] 𝑥 ◦ ↦→ [Af (𝜉, 𝜆)] 𝑥 ◦ 𝜆 𝑁 −1 d𝜆. (21.4.36)
1

The map (21.4.36) depends “analytically” on 𝜉 in the sense that there is a neigh-
borhood 𝑈 of 𝑥 ◦ in R𝑛 and an analytic functional 𝜇 representing [f] 𝑥 ◦ in 𝑈 such
that
∫ +∞
𝑨𝑈 𝜇 (𝜉, 𝜆) 𝜆 𝑁 −1 d𝜆 is a C function of 𝜉 ∈ Ξ ∩ S
𝜔 C 2𝑛−1 valued in O 𝑈 .′
1

Corollary 21.4.24 Let Γ be an open cone in R𝑛 \ {0} such that Γ ∩ S𝑛−1 ⊂ Ξ; the
sing
map (21.4.27) induces a linear endomorphism B 𝑥 ◦ ←↪,
∫
[f] 𝑥 ◦ ↦→ AΓ [f] 𝑥 ◦ = [Af (𝜉/|𝜉 | , |𝜉 |)] 𝑥 ◦ d𝜉. (21.4.37)
Γ
914 21 Germ Pseudodifferential Operators in Complex Space

If 𝜇 ∈ O ′ (Ω) the integral

∫ ∫ +∞
𝑨𝑈,Γ 𝜇 = 𝑨𝑈 𝜇 (𝜉, 𝜆) 𝜆 𝑛−1 d𝜆𝑑𝑆 (𝜉) (21.4.38)
1
∫Γ∩S
𝑛−1

= 𝑨𝑈 𝜇 (𝜉, |𝜉 |) d𝜉
Γ

represents an analytic functional in Ω with supp 𝑨𝑈,Γ 𝜇 ⊂ 𝑈. Let 𝑓 be the hyperfunc-

tion in 𝑈 represented by 𝜇; 𝑨𝑈,Γ 𝜇 defines a hyperfunction in 𝑈 which we denote by
𝑨𝑈,Γ 𝑓 ; the singularity hyperfunction corresponding to 𝑨𝑈,Γ 𝜇 does not depend on
the representative 𝜇. We derive from (21.4.37)

Theorem 21.4.25 If 𝑓 is a hyperfunction (resp., a singularity hyperfunction) in Ω

then 𝑊 𝐹a 𝑨𝑈,Γ 𝑓 ⊂ 𝑊 𝐹a ( 𝑓 ).

The action of 𝑨𝑈,Γ on microfunctions is a direct consequence of Theorem 21.4.25:

Let [f] ( 𝑥 ◦ , 𝜉 ◦ ) be a germ at (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑈 × Γ of a microfunction, i.e., the value at
(𝑥 ◦ , 𝜉 ◦ ) of a continuous section [f] of the sheaf Bmicro over R𝑛 (remember that the
sheaf Bmicro is flabby). If 𝑓1 and 𝑓2 are two singularity hyperfunctions representing
[f] ( 𝑥 ◦ , 𝜉 ◦ ) then 𝑓1 − 𝑓2 is microanalytic at every point (𝑥, 𝜉) ∈ 𝑈 ×Σ with Σ ⊂ Γ∩S𝑛−1
a neighborhood of 𝜉 ◦ . Theorem 21.4.25 tells us that 𝑨𝑈,Γ ( 𝑓1 − 𝑓2 ) is microanalytic
in 𝑈 ′ × Σ ′ with 𝑥 ◦ ∈ 𝑈 ′ ⊂ 𝑈 and 𝜉 ◦ ∈ Σ ′ ⊂ Σ = Γ ∩ S𝑛−1 . It follows that 𝑨𝑈,Γ 𝑓1
and 𝑨𝑈,Γ 𝑓2 define the same microfunction in 𝑈 ′ × Σ ′, whose germ at (𝑥 ◦ , 𝜉 ◦ ) is the
sought image of [f] ( 𝑥 ◦ , 𝜉 ◦ ) . The same concept is arrived at by contracting 𝑈 about
𝑥 ◦ and Σ about 𝜉 ◦ and we may therefore denote said image by A [f] ( 𝑥 ◦ , 𝜉 ◦ ) . This is
the definition of the microlocal analogues of (21.4.37):

Bmicro micro
( 𝑥 ◦ , 𝜉 ◦ ) ∋ [f] ( 𝑥 ◦ , 𝜉 ◦ ) ↦→ A [f] ( 𝑥 ◦ , 𝜉 ◦ ) ∈ B ( 𝑥 ◦ , 𝜉 ◦ ) . (21.4.39)
Chapter 22
Germ FBI Transforms

This chapter is devoted to a special class of the integral transforms introduced in Ch.
21, generalizing the FBI transform and designed for the specific purpose of analyzing
the wave-front set of a distribution. We follow and strengthen the strategy initiated
in the last section of the preceding chapter: shifting the burden of the strong well-
shapedness of the contours of integration from the geometry and the exponent-weight
to the imaginary part of the phase-function 𝜑 (𝑧, 𝜃) itself. We abide by Sjöstrand’s
terminology and call 𝜑 (𝑧, 𝜃) an FBI phase-function. For these phase-functions it is
convenient, and not limitative, to assume from the start that the contour of integration
𝔠 is a smoothly bounded open subset 𝑈 of R𝑛 (⊂ C𝑛 ). The main raison d’être of
the chapter lies in Sjöstrand’s Equivalence Theorem, proved in the last section and
needed in Ch. 24.

22.1 Germ FBI Transforms

22.1.1 FBI phase-functions. Definition

Let ΩC , Θ be domains in C𝑛 , 0 ∉ Θ; the variable in ΩC × Θ is (𝑧, 𝜃), 𝑧 = 𝑥 + 𝑖𝑦 =

(𝑧1 , ..., 𝑧 𝑛 ), 𝜃 = (𝜃 1 , ..., 𝜃 𝑛 ). Throughout this chapter 𝑥 ◦ ∈ Ω = ΩC ∩ R𝑛 , 𝜃 ◦ ∈ Θ.

Definition 22.1.1 By an FBI phase-function in ΩC × Θ at a point (𝑥 ◦ , 𝜃 ◦ ) we shall

mean a holomorphic function 𝜑 (𝑧, 𝜃) in ΩC × Θ having the following properties:

(1) 𝜑(𝑥 ◦ , 𝜃 ◦ ) = 0, 𝜕𝜑 ◦ ◦ ◦
𝜕𝑧 (𝑥 , 𝜃 ) = −𝜉 ∈ R \ {0};
𝑛
2
(2) det 𝜕𝜃𝜕𝑗 𝜕𝑧 𝜑
(𝑥 ◦ , 𝜃 ◦ ) ≠ 0, i.e., (21.1.28) holds;
1≤ 𝑗,𝑘 ≤𝑛
𝑘
2
(3) Im 𝜕𝑧𝜕𝑗 𝜕𝑧 𝜑
(𝑥 ◦ , 𝜃 ◦ ) is positive definite.
𝑘 1≤ 𝑗,𝑘 ≤𝑛

Note that if 𝜑 is an FBI phase-function this not true of −𝜑.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 915
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_22
916 22 Germ FBI Transforms

Example 22.1.2 The prototype of an FBI phase-function in R𝑛 × C𝑛 at (𝑥 ◦ , 𝜃 ◦ ) is

− (𝑧 − 𝑥 ◦ ) · 𝜃 + 21 𝑖 ⟨𝑧 − 𝑥 ◦ ⟩ 2 [⟨𝑧⟩ 2 = 𝑧 · 𝑧]. Condition (1) in Definition 22.1.1 demands
𝜃 ◦ = 𝜉 ◦ ∈ R𝑛 \ {0}.
Proposition 22.1.3 If 𝜑 ∈ O ΩC × Θ is an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ) then the

following properties hold at (𝑥 ◦ , 𝜃 ◦ ):
(1) 𝜕𝑥 Im 𝜑 = 0, 𝜕𝑦 Im 𝜑 = −𝜉 ◦ ∈ R𝑛 \ {0};
(2) the 𝑛 × 𝑛 complex matrices 𝜕Re 𝜃 𝜕𝑧 Im 𝜑, 𝜕Im 𝜃 𝜕𝑧 Im 𝜑, 𝜕𝜃 𝜕𝑥 Im 𝜑, 𝜕𝜃 𝜕𝑦 Im 𝜑 are
invertible;
(3) the 𝑛 × 𝑛 real symmetric matrix 𝜕𝑥2 Im 𝜑 = −𝜕𝑦2 Im 𝜑 is positive definite.
Proof The Cauchy–Riemann equations imply directly

𝜕𝑧 𝜑 = 2𝑖𝜕𝑧 Im 𝜑 = 𝑖𝜕𝑥 + 𝜕𝑦 Im 𝜑. (22.1.1)

This shows that Property (1) in Definition 22.1.1 implies (1) in Proposition 22.1.3.
All the matrices in (2), Proposition 22.1.3, are equal to 𝜕𝜃 𝜕𝑧 𝜑 up to a constant factor
±2, ±2𝑖, and thus are nonsingular at (𝑥 ◦ , 𝜃 ◦ ) by Property (2) in Definition 22.1.1.
Since Im 𝜕𝑧2 𝜑 = −𝜕𝑦2 Im 𝜑 Property (3) in Definition 22.1.1 entails (3) in Proposition
22.1.3. □
Remark 22.1.4 Property (1) in Proposition 22.1.3 precludes Im 𝜑 (𝑧, 𝜃 ◦ ) from hav-
ing a critical point in the neighborhood of 𝑥 ◦ .
Remark 22.1.5 Property (2) in Proposition 22.1.3 is equivalent to the invertibility
at (𝑥 ◦ , 𝜃 ◦ ) of the 2𝑛 × 2𝑛 matrix

𝜕Re 𝜃 𝜕𝑥 Im 𝜑 𝜕Re 𝜃 𝜕𝑦 Im 𝜑
𝜕Re 𝜃 ,Im 𝜃 𝜕𝑥,𝑦 Im 𝜑 =
𝜕Im 𝜃 𝜕𝑥 Im 𝜑 𝜕Im 𝜃 𝜕𝑦 Im 𝜑

𝜕Re 𝜃 𝜕𝑥 Im 𝜑 𝜕Im 𝜃 𝜕𝑥 Im 𝜑
= .
𝜕Im 𝜃 𝜕𝑥 Im 𝜑 −𝜕Re 𝜃 𝜕𝑥 Im 𝜑

u
Indeed, if 𝐴, 𝐵 are real 𝑛 × 𝑛 matrices, a vector ∈ R2𝑛 belongs to the null space
v

𝐴 𝐵
of the real 2𝑛 × 2𝑛 matrix if and only if ( 𝐴 − 𝑖𝐵) (u + 𝑖v) = 0.
𝐵 −𝐴
Proposition 22.1.6 Let 𝜑 (𝑧, 𝜃) ∈ O ΩC × Θ be an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ )

and let 𝑓 (resp., 𝑔) be a biholomorphism of an open subset ΩC1 (resp., Θ1 ) of C𝑛
onto ΩC (resp., Θ) mapping ΩC1 ∩ R𝑛 onto Ω. If 𝑥 ◦ = 𝑓 (𝑥 ∗ ) and 𝜃 ◦ = 𝑔 (𝜃 ∗ ) then
𝜑 ( 𝑓 (𝑧) , 𝑔 (𝜃)) is an FBI phase-function at (𝑥 ∗ , 𝜃 ∗ ).
𝜕𝑓 ∗ 𝜕𝑓
(𝜃 ∗ ) are
Proof The hypotheses entail that the Jacobian matrices 𝜕𝑧 (𝑥 ) and 𝜕𝜃
𝜕𝑓 𝜕𝑓
nonsingular and that 𝜕𝑧 (𝑥) = 𝜕𝑥 (𝑥) is real for 𝑥 ∈ Ω1 ∩ R𝑛 .
C
In turn this en-
tails directly that 𝜑 ( 𝑓 (𝑧) , 𝑔 (𝜃)) satisfies Conditions (1), (2), in Definition 22.1.1;
𝜑 ( 𝑓 (𝑧) , 𝑔 (𝜃)) also satisfies Conditions (3) since, given arbitrary 𝑛 × 𝑛 matrices
𝐴 complex and 𝐵 real, the matrix Im (𝐵⊤ 𝐴𝐵) = 𝐵⊤ (Im 𝐴) 𝐵 is positive definite if
Im 𝐴 is positive definite and 𝐵 is nonsingular. □
22.1 Germ FBI Transforms 917

22.1.2 Sjöstrand pairs associated to an arbitrary FBI phase-function

If 𝜑 ∈ O ΩC × Θ is an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ), exponent-weights Φ ∈

C ∞ ΩC ; R such that (𝜑 (𝑧, 𝜃) , Φ (𝑧)) is a Sjöstrand pair at (𝑥 ◦ , 𝜃 ◦ ) (Definition
20.3.6) are readily available. We shall now show thatthis is true when Φ (𝑧) = −𝜉 ◦ · 𝑦
whatever the FBI phase-function 𝜑 ∈ O ΩC × Θ such that 𝜉 ◦ = −𝜕𝑧 𝜑(𝑥 ◦ , 𝜃 ◦ )
(assumed to be the case throughout the sequel); here [cf. (20.3.1)]

Ψ (𝑧, 𝜃) = −𝜉 ◦ · 𝑦 − Im 𝜑 (𝑧, 𝜃)

is pluriharmonic.
Consider the Taylor expansion of 𝑧 ↦→ − Im 𝜑 (𝑧, 𝜃) about 𝑥 ◦ up to third order:

− Im 𝜑 (𝑧, 𝜃) = − Im 𝜑 (𝑥 ◦ , 𝜃) + 𝑝 (𝜃) · (𝑥 − 𝑥 ◦ ) − 𝑞 (𝜃) · 𝑦 (22.1.2)

1 1
− (𝑥 − 𝑥 ◦ ) · 𝑀◦ (𝑥 − 𝑥 ◦ ) + 𝑦 · 𝑀◦ 𝑦 − (𝑥 − 𝑥 ◦ ) · 𝑁◦ 𝑦
2 2
+ 𝑂 |𝑧 − 𝑥 ◦ | 3 + |𝑧 − 𝑥 ◦ | 2 |𝜃 − 𝜃 ◦ | ,

where

𝑝 (𝜃) = −𝜕𝑥 Im 𝜑 (𝑥 ◦ , 𝜃) , 𝑞 (𝜃) = 𝜕𝑦 Im 𝜑 (𝑥 ◦ , 𝜃) , (22.1.3)

𝑀◦ = 𝜕𝑥2 Im 𝜑 (𝑥 ◦ , 𝜃 ◦ ) , 𝑁◦ = 𝜕𝑥 𝜕𝑦 Im 𝜑 (𝑥 ◦ , 𝜃 ◦ ) .

Proposition 22.1.3 states that

𝑝 (𝜃 ◦ ) = 0, 𝑞 (𝜃 ◦ ) = −𝜉 ◦ , det 𝜕𝜃 𝑞 (𝜃 ◦ ) ≠ 0 (22.1.4)

and that 𝑀◦ is positive definite; 𝑁◦ is symmetric.

The critical points of 𝑧 ↦→ Ψ (𝑧, 𝜃) are the solutions (𝑥, 𝑦) of the following
congruences:

𝑀◦ (𝑥 − 𝑥 ◦ ) + 𝑁◦ 𝑦 𝑝 (𝜃) , (22.1.5)
𝑁◦ (𝑥 − 𝑥 ◦ ) − 𝑀◦ 𝑦 −𝑞 (𝜃) − 𝜉 ◦ ,

where means mod terms 𝑂 |𝑧 − 𝑥 ◦ | 2 + |𝜃 − 𝜃 ◦ | |𝑧 − 𝑥 ◦ | . Solving (22.1.5) yields

𝑥 − 𝑥 ◦ 𝑀1−1 𝑝 (𝜃) − 𝑁◦ 𝑀◦−1 (𝑞 (𝜃) + 𝜉 ◦ ) (22.1.6)

𝑦 𝑀1−1 𝑁◦ 𝑀◦−1 𝑝 (𝜃) + 𝑞 (𝜃) + 𝜉 ◦ ,

where 𝑀1 = 𝑀◦ + 𝑁◦ 𝑀◦−1 𝑁◦ , an obviously positive definite, real 𝑛 × 𝑛 matrix;

(22.1.4) implies 𝑥 = 𝑥 ◦ , 𝑦 = 0 when 𝜃 = 𝜃 ◦ . The Hessian of Ψ (𝑧, 𝜃) with respect to
(𝑥, 𝑦) at the critical points is the (2𝑛) × (2𝑛) matrix
918 22 Germ FBI Transforms

−𝑀◦ 𝑁◦
.
𝑁◦ 𝑀◦

That these critical points are saddle points is a consequence of the following ele-
mentary result in linear algebra:
Lemma 22.1.7 Let 𝐴1 , 𝐴2 be two real symmetric 𝑛×𝑛 matrices, both positive definite,
and let 𝐵 be an arbitrary real 𝑛 × 𝑛 matrix. The real symmetric (2𝑛) × (2𝑛) matrix

−𝐴1 𝐵
𝑇=
𝐵 ⊤ 𝐴2

is nonsingular and its signature is equal to zero.

u
Proof If 𝑇 = 0 then 𝐵v = 𝐴1 u, 𝐵⊤ u = −𝐴2 v, implying
v

u · 𝐴1 u = u · 𝐵v = v · 𝐵⊤ u = −v · 𝐴2 v

hence u = v = 0 by the hypothesis on 𝐴1 , 𝐴2 . For every 𝑡 ∈ [0, 1] the real symmetric

(2𝑛) × (2𝑛) matrix
−𝐴1 𝑡𝐵
𝑇=
𝑡𝐵⊤ 𝐴2
remains invertible, implying that its signature is independent of 𝑡; the signature of
𝑇 (0) being equal to zero the same is true of 𝑇 (1) = 𝑇. □
We can state
Proposition 22.1.8 If 𝜑 ∈ O ΩC × Θ is an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ) and

Φ (𝑧) = −𝜉 ◦ · 𝑦 then (𝜑, Φ) is a Sjöstrand pair at (𝑥 ◦ , 𝜃 ◦ ).
The critical value Φ∗ (𝜃) of 𝑧 ↦→ Ψ (𝑧, 𝜃) = −𝜉 ◦ · 𝑦 − Im 𝜑 (𝑧, 𝜃) at the points
(22.1.3) is fairly complicated to write down and we shall not make use of its precise
expression. We have Φ∗ (𝜃 ◦ ) = 0 [cf. (20.3.8)] and if Ψ∗ is defined as in (20.3.13)
then 𝜃 ◦ is a saddle point of Ψ∗ (𝑥 ◦ , 𝜃) (Theorem 20.3.16).
Remark 22.1.9 It follows from Proposition 22.1.8 and Theorem 20.3.16 that
(−𝜑, Φ∗ ) is a Sjöstrand pair at (𝜃 ◦ , 𝑥 ◦ ).
Example 22.1.10 If 𝜑 (𝑧, 𝜃) = (𝑧 − 𝑥 ◦ ) · 𝜃 + 12 𝑖 ⟨𝑧 − 𝑥 ◦ ⟩ 2 (cf. Example 22.1.2) and
Φ (𝑧) = 𝑦 · 𝜃 ◦ [𝜃 ◦ = −𝜉 ◦ ∈ R𝑛 \ {0}, cf. Definition 22.1.1, (1)] then
1 1
Ψ (𝑧, 𝜃) = (𝑥 − 𝑥 ◦ ) · Im 𝜃 + 𝑦 · (Re 𝜃 − 𝜃 ◦ ) − |𝑥 − 𝑥 ◦ | 2 + |𝑦| 2
2 2
and the critical points of 𝑧 ↦→ Ψ (𝑧, 𝜃) are given by 𝑥 − 𝑥 ◦ = Im 𝜃, 𝑦 = − (Re 𝜃 − 𝜃 ◦ ),
with the critical value
1 1
Φ∗ (𝜃) = |Im 𝜃| 2 − |Re 𝜃 − 𝜃 ◦ | 2 . (22.1.7)
2 2
22.1 Germ FBI Transforms 919

The function
1
𝜃 ↦→ Ψ∗ (𝑥, 𝜃) = |Im 𝜃 + 𝑥 − 𝑥 ◦ | 2 − |Re 𝜃 − 𝜃 ◦ | 2
2
has a saddle point at 𝜃 = 𝜃 ◦ − 𝑖 (𝑥 − 𝑥 ◦ ) with zero critical value [in agreement with
the fact that Φ (𝑥) ≡ 0].

We go back to the general case.

Proposition 22.1.11 Let Ψ (𝑧, 𝜃) = −𝜉 ◦ ·𝑦−Im 𝜑 (𝑧, 𝜃). If Ω′C ⊂⊂ ΩC is a sufficiently

small neighborhood of 𝑥 ◦ then every neighborhood 𝑈 ⊂⊂ Ω′C ∩ R𝑛 of 𝑥 ◦ with a
smooth boundary, regarded as a contour of integration, is strongly well-shaped at
𝑥 ◦ (Definition 20.2.2) for the function 𝑧 ↦→ Ψ(𝑧, 𝜃 ◦ ).

Proof From (22.1.2), (22.1.4) and the fact that Im 𝜑 (𝑥 ◦ , 𝜃 ◦ ) = 0 we derive

2
−1

Ψ (𝑥, 𝜃 ◦ ) = − 𝑀◦ 2 (𝑥 − 𝑥 ◦ ) + 𝑂 |𝑥 − 𝑥 ◦ | 3 . □

Remark 22.1.12 By adapting the proofs of Propositions 22.1.8 and 22.1.11 we see
that they remain valid for the more general exponent-weights
1
Φ (𝑧) = −𝜉 ◦ · 𝑦 − (𝑥 − 𝑥 ◦ ) · 𝑆1 (𝑥 − 𝑥 ◦ ) (22.1.8)
2
1
+ 𝑦 · 𝑆2 𝑦 + (𝑥 − 𝑥 ◦ ) · 𝑆 ′ 𝑦 + 𝑂 |𝑧 − 𝑥 ◦ | 3
2
where 𝑆1 , 𝑆2 and 𝑆 ′ are real symmetric 𝑛 × 𝑛 matrices with the following properties:
(1) 𝑆2 − 𝑆1 is positive semidefinite;
(2) 𝑀◦ + 𝑆 𝑗 [ 𝑗 = 1, 2, 𝑀◦ as in (22.1.2)] is positive definite.
The function (22.1.8) is plurisubharmonic in a neighborhood of 𝑥 ◦ in C𝑛 .

22.1.3 Germ FBI-like transforms of holomorphic functions. Classical

FBI-like transforms

Let 𝜑 ∈ O ΩC × Θ be an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ) and 𝜉 ◦ = −𝜕𝑧 𝜑(𝑥 ◦ , 𝜃 ◦ ) ∈

Ξ∩R𝑛 and 𝑎 (𝑧, 𝜃, 𝜆) be an analytic symbol in ΩC ×Θ (Definition 20.1.1). Propositions
22.1.8 and 22.1.11 allow us to exploit the construction in Subsection 20.3.2, and
define a germ operator of the type (20.4.4) starting from an integral (20.1) in which
we take the contour 𝔠 to be an open subset 𝑈 of R𝑛 , 𝑥 ◦ ∈ 𝑈 ⊂⊂ ΩC :
∫
𝑨𝑈 ℎ (𝜃, 𝜆) = e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) ℎ (𝑥, 𝜆) d𝑥 (22.1.9)
𝑈
920 22 Germ FBI Transforms
∗
with ℎ ∈ O (Φ) ΩC , Φ (𝑧) = −𝜉 ◦ ·𝑦. From Theorem 20.4.6, we get 𝑨𝑈 ℎ ∈ O (Φ ) (Θ)

with Φ∗ (𝜃) the critical value (22.1.7). By contracting ΩC and 𝑈 about 𝑥 ◦ and Θ
about 𝜃 ◦ and taking into account Corollary 20.4.4, the integrals (22.1.9) define a
germ Fourier-like transform (Definition 20.4.7),
∗
𝑨 𝑥 ◦ , 𝜃 ◦ : O 𝑥(Φ) (−𝜔)
◦ /O 𝑥 ◦ −→ O (Φ ) (−𝜔)
𝜃 ◦ /O 𝜃 ◦ . (22.1.10)

Definition 22.1.13 We shall refer to the operator (22.1.10) as a germ FBI-like

transform at (𝑥 ◦ , 𝜃 ◦ ) associated to the FBI phase-function 𝜑. We shall say that
𝑨 𝑥 ◦ , 𝜃 ◦ is classical if the symbol 𝑎 (𝑥, 𝜃, 𝜆) in (22.1.9) can be taken to be (a finite
realization of) a classical symbol (Definitions 19.1.7, 19.1.14).

When 𝑎 (𝑧, 𝜃, 𝜆) is classical and elliptic 𝑨 𝑥 ◦ , 𝜃 ◦ has an inverse

(Φ ) ∗
(−𝜔)
𝑨−1
𝜃 ◦ , 𝑥 ◦ : O 𝜃 ◦ /O 𝜃 ◦ −→ O 𝑥(Φ) (−𝜔)
◦ /O 𝑥 ◦ (22.1.11)

(Theorem 21.2.10; about O 𝑥(−𝜔)

◦ see Definition 19.1.10) defined by integrals
𝑛 ∫
𝜆
𝑩𝔠∗ 𝑔 (𝑧, 𝜆) = e−𝑖𝜆𝜑 (𝑧, 𝜃) 𝑏 (𝑧, 𝜃, 𝜆) 𝑔 (𝜃, 𝜆) d𝜃 (22.1.12)
2𝜋 𝔠∗

where the contour of integration 𝔠 ∗ ⊂⊂ Ξ is strongly well-shaped at 𝜃 ◦ for the function

∗
Ψ∗ (𝑥 ◦ , 𝜃) = Φ∗ (𝜃) + Im 𝜑 (𝑥 ◦ , 𝜃), 𝑔 ∈ O (Φ ) (Ξ) and 𝑏 is a classical, elliptic symbol
in Ω × Ξ. Note, however, that −𝜑 (𝑧, 𝜃) is not an FBI phase-function.
C

22.2 Germ FBI Transforms of Distributions

22.2.1 FBI phase-functions at real points

We continue to deal with the FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ), 𝜑 ∈ O ΩC × Θ . In this

subsection we focus on the behavior of − Im 𝜑 at the germ of real space R𝑛 in C𝑛 .

Proposition 22.2.1 The domains ΩC and Θ in C𝑛 can be selected so that, for each
𝜃 ∈ Θ, the function 𝑥 ↦→ − Im 𝜑 (𝑥, 𝜃) admits a unique maximum in Ω = ΩC ∩ R𝑛 ,
𝑥 max (𝜃), where its Hessian matrix is negative definite; we have 𝑥max (𝜃 ◦ ) = 𝑥 ◦ ; the
map Θ ∋ 𝜃 ↦→ 𝑥 max (𝜃) ∈ R𝑛 is of class C 𝜔 .

Proof The Implicit Function Theorem and Properties (1) and (3) in Proposition
22.1.6 imply that the system of (real) equations

𝜕 Im 𝜑
(𝑥, 𝜃) = 0, 𝑗 = 1, ..., 𝑛, (22.2.1)
𝜕𝑥 𝑗
22.2 Germ FBI Transforms of Distributions 921

has a unique solution 𝑥max (𝜃) ∈ C 𝜔 (Θ; R) such that 𝑥max (𝜃 ◦ ) = 𝑥 ◦ (provided Θ
is sufficiently small). Property (3) in Proposition 22.1.6 implies the claim about the
Hessian matrix. □

Proposition 22.2.2 The following properties are equivalent: 1) 𝜕𝑧 𝜑 (𝑥, 𝜃) ∈ R𝑛 ; 2)

𝑥 = 𝑥max (𝜃).

Proof When (22.2.1) holds we have 𝜕𝑧 𝜑 (𝑥, 𝜃) = 𝜕𝑦 Im 𝜑 (𝑥, 𝜃) [cf. (22.1.1)] and
conversely, this equation implies (22.2.1). □

Lemma 22.2.3 If Θ is sufficiently small and if 𝑥 ∗ ∈ Ω is sufficiently close to 𝑥 ◦

then the function 𝜃 ↦→ |𝑥max (𝜃) − 𝑥 ∗ | 2 is strictly plurisubharmonic in Θ (Definition
11.2.6).

Proof We have
𝑛 𝜕𝑥max, 𝑗
𝜕 ∑︁
|𝑥max (𝜃) − 𝑥 ∗ | 2 = 2 𝑥max, 𝑗 (𝜃) − 𝑥 ∗𝑗 (𝜃) ,
𝜕𝜃 𝛼 𝑗=1
𝜕𝜃 𝛼

𝜕2
𝑛
∑︁ 𝜕 2 𝑥 max, 𝑗
|𝑥max (𝜃) − 𝑥 ∗ | 2 = 2 𝑥max, 𝑗 (𝜃) − 𝑥 ∗𝑗 (𝜃)
𝜕𝜃 𝛼 𝜕𝜃 𝛽 𝑗=1 𝜕𝜃 𝛼 𝜕𝜃 𝛽
𝑛
∑︁ 𝜕𝑥 max, 𝑗 𝜕𝑥max, 𝑗
+2 (𝜃) (𝜃) ,
𝑗=1
𝜕𝜃 𝛼 𝜕𝜃 𝛽

whence, for all 𝜃 ∈ Θ, 𝑤 ∈ C𝑛 ,

𝑛 𝑛
1 ∑︁ ∑︁ 𝜕2
𝑤 𝛼𝑤𝛽 |𝑥max (𝜃) − 𝑥 ∗ | 2 (22.2.2)
2 𝛼=1 𝛽=1 𝜕𝜃 𝛼 𝜕𝜃 𝛽
𝑛 ∑︁
𝑛 2
∑︁ 𝜕𝑥 max, 𝑗
= 𝑤𝛼 (𝜃) + 𝑂 |𝑥max (𝜃) − 𝑥 ∗ | |𝑤| 2 .
𝑗=1 𝛼=1
𝜕𝜃 𝛼

From (22.2.1) we deduce

𝑛
1 𝜕 2 Im 𝜑 ∑︁ 𝜕 2 Im 𝜑 𝜕𝑥 max,𝑘
(𝑥 max (𝜃) , 𝜃) = (𝑥 max (𝜃) , 𝜃) (𝜃) .
2𝑖 𝜕𝜃 𝛼 𝜕𝑥 𝑗 𝑘=1
𝜕𝑥 𝑗 𝜕𝑥 𝑘 𝜕𝜃 𝛼

This identity
together
with Properties (2) and (3) in Proposition 22.1.6 imply that the
is nonsingular in a neighborhood of 𝜃 ◦ . This fact and
𝜕𝑥max,𝑘
matrix 𝜕𝜃 𝛼 (𝜃)
1≤ 𝛼,𝑘 ≤𝑛
(22.2.2) imply
𝑛 ∑︁
𝑛
∑︁ 𝜕2
𝑤 𝛼 𝑤¯ 𝛽 |𝑥 max (𝜃) − 𝑥 ∗ | 2 ≥ 𝑐 |𝑤| 2
𝛼=1 𝛽=1 𝜕𝜃 𝛼 𝜕𝜃 𝛽

for some 𝑐 > 0, all 𝜃 sufficiently close to 𝜃 ◦ and all 𝑤 ∈ C𝑛 . □

922 22 Germ FBI Transforms

In the sequel we make use of the following notation [cf. (22.2.1)]:

Φ♭ (𝜃) = − Im 𝜑 (𝑥 max (𝜃) , 𝜃) = sup (− Im 𝜑 (𝑥, 𝜃)) . (22.2.3)

𝑥 ∈Ω

We have Φ♭ (𝜃 ◦ ) = 0.

Proposition 22.2.4 Let the function Φ♭ be given by (22.2.3). If ΩC and Θ are suffi-
ciently small then there is a 𝑐 > 0 such that, for every (𝑥, 𝜃) ∈ (Ω) × Θ,

Φ♭ (𝜃) + Im 𝜑 (𝑥, 𝜃) ≥ 𝑐 |𝑥 − 𝑥max (𝜃)| 2 . (22.2.4)

The functions Φ♭ and 𝜃 ↦→ Φ♭ (𝜃) + Im 𝜑 (𝑥, 𝜃) are strictly plurisubharmonic in Θ.

Proof We have, for (𝑥, 𝜃) ∈ (Ω) × Θ,

Φ♭ (𝜃) = − Im 𝜑 (𝑥max (𝜃) , 𝜃) = − Im 𝜑 (𝑥, 𝜃)

𝑛
1 ∑︁ 𝜕 2 Im 𝜑
+ (𝑥max (𝜃) , 𝜃) 𝑥 𝑗 − 𝑥max, 𝑗 (𝜃) 𝑥 𝑘 − 𝑥max,𝑘 (𝜃)
2 𝑗,𝑘=1 𝜕𝑥 𝑗 𝜕𝑥 𝑘

+ 𝑂 |𝑥max (𝜃) − 𝑥| 3 .

Proposition 22.1.6, (3), implies that the Hessian matrix 𝜕𝑥2 Im 𝜑 (𝑧, 𝜃) is positive
definite in ΩC ×Θ [possibly contracted about (𝑥 ◦ , 𝜃 ◦ )], whence (22.2.4). The function
𝜃 ↦→ − Im 𝜑 (𝑥, 𝜃) being pluriharmonic for every 𝑥 ∈ Ω we derive from Proposition
11.2.8 that Φ♭ is plurisubharmonic. If we fix 𝑥 the two sides in (22.2.4) are C 𝜔
plurisubharmonic functions of 𝜃 ∈ Θ and both vanish when 𝑥 = 𝑥max (𝜃) by (22.2.3);
by Lemma 22.2.3 the right-hand side is strictly plurisubharmonic in a neighborhood
of 𝜃 ◦ . It follows immediately that this must also be true of the left-hand side. □
The following statement reveals an important difference between Φ♭ and Φ∗ [cf.
(22.1.7) and Proposition 22.1.8].

Corollary 22.2.5 The pair of functions −𝜑, Φ♭ is not a Sjöstrand pair at (𝜃 ◦ , 𝑥 ◦ ).

Proof If 𝜃 ◦ were a saddle point of Ψ♭ (𝑥 ◦ , 𝜃) = Φ♭ (𝜃) + Im 𝜑 (𝑥 ◦ , 𝜃) we would have

Ψ♭ (𝑥 ◦ , 𝜃) < 0 at some points 𝜃 arbitrarily close to 𝜃 ◦ , contradicting (22.2.4). □

Example 22.2.6 When 𝜑 (𝑧, 𝜃) = 𝜑◦ (𝑧 − 𝑥 ◦ , 𝜃) is the phase-function in Example

22.1.2 we have
1 1
Im 𝜑 (𝑧, 𝜃) = − (𝑥 − 𝑥 ◦ ) · Im 𝜃 − 𝑦 · Re 𝜃 + |𝑥 − 𝑥 ◦ | 2 − |𝑦| 2 ,
2 2
whence
𝜕 Im 𝜑
(𝑥, 𝜃) = 0 ⇐⇒ 𝑥 − 𝑥 ◦ = Im 𝜃
𝜕𝑥
and Φ♭ (𝜃) = 1
2 |Im 𝜃| 2 . In this case,
22.2 Germ FBI Transforms of Distributions 923

Ψ♭ (𝑧, 𝜃) = Φ♭ (𝜃) + Im 𝜑 (𝑧, 𝜃)

1 1 1
= |Im 𝜃| 2 − (𝑥 − 𝑥 ◦ ) · Im 𝜃 − 𝑦 · Re 𝜃 + |𝑥 − 𝑥 ◦ | 2 − |𝑦| 2
2 2 2
1 1
= |𝑥 − 𝑥 ◦ − Im 𝜃| 2 − 𝑦 · Re 𝜃 − |𝑦| 2
2 2
and 𝜃 ↦→ Ψ♭ (𝑧, 𝜃) has no saddle point, whatever 𝑧 ∈ C𝑛 . Compare Φ♭ and Ψ♭ to Φ∗
and Ψ∗ (Example 22.1.10).

22.2.2 Germ FBI-like transforms of integrable functions and of

distributions

As before ΩC and
Θ are open subsets of C containing
𝑛 𝑥 ◦ and 𝜃 ◦ respectively and
◦ ◦
𝜑 ∈ O Ω × Θ is an FBI phase-function at (𝑥 , 𝜃 ). We return to the integral
C

operator 𝑨𝑈 of (22.1.9) where we no longer require the “test function” ℎ to be

holomorphic.
Proposition 22.2.7 Let Θ ⊂ C𝑛 \ {0} be as required in Proposition 22.2.1. Let 𝐾 ⊂ Ω
be a compact set and [cf. (22.2.3)]

Φ♭𝐾 (𝜃) = sup (− Im 𝜑 (𝑥, 𝜃)) , 𝜃 ∈ Θ. (22.2.5)

𝑥 ∈𝐾

1 (Ω) continuously into

If 𝑈 ⊂ 𝐾 the linear operator 𝑨𝑈 [see (22.1.9)] maps 𝐿 loc
O ( Φ𝐾 ) (Θ) (Definition 20.2.7).
♭

Proof We take Definition 20.1.1 into account. By (22.1.9) and (22.2.5), to every
𝜀 > 0 there is a 𝐶 𝜀 > 0 such that
∫
−𝜆Φ♭𝐾 ( 𝜃) 𝜀𝜆
e | 𝑨𝑈 𝑓 (𝜃, 𝜆)| ≤ 𝐶 𝜀 e | 𝑓 (𝑥)| d𝑥. (22.2.6)
𝐾

□
Proposition 22.2.8 Let 𝑈 ⊂⊂ Ω be a neighborhood of 𝑥 ◦ in R𝑛 . If 𝑓 ∈ 𝐿 loc1 (Ω) the
(−𝜔)
equivalence class mod O 𝜃 ◦ (cf. Definition 19.1.10) of the germ at 𝜃 ◦ of 𝑨𝑈 𝑓 (𝜃, 𝜆)
depends solely on the germ of 𝑓 at 𝑥 ◦ .
Proof Suppose

𝔅𝑅 (𝑥 ◦ ) = {𝑥 ∈ R𝑛 ; |𝑥 − 𝑥 ◦ | < 𝑅} ⊂⊂ 𝑈 ⊂ 𝐾

and supp 𝑓 ⊂ 𝑈\𝔅𝑅 (𝑥 ◦ ). We have

∫
𝑨𝑈 𝑓 (𝜃, 𝜆) = e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) 𝑓 (𝑥) d𝑥.
| 𝑥−𝑥 ◦ |>𝑅
924 22 Germ FBI Transforms

By (22.1.2) we have

− Im 𝜑 (𝑥, 𝜃) = − Im 𝜑 (𝑥 ◦ , 𝜃) + 𝑝 (𝜃) · (𝑥 − 𝑥 ◦ )
1
− (𝑥 − 𝑥 ◦ ) · 𝑀◦ (𝑥 − 𝑥 ◦ ) + 𝑂 |𝑥 − 𝑥 ◦ | 3 + |𝜃 − 𝜃 ◦ | |𝑥 − 𝑥 ◦ | 2 .
2
Recalling that Im 𝜑 (𝑥 ◦ , 𝜃 ◦ ) = 0 and 𝑝 (𝜃 ◦ ) = 0 we see that there are constants 𝜅 𝑗 > 0
( 𝑗 = 1, 2) such that

− Im 𝜑 (𝑥, 𝜃) ≤ −𝜅 1 |𝑥 − 𝑥 ◦ | 2 + 𝜅2 |𝜃 − 𝜃 ◦ | ,

whence, for all 𝑥 ∈ 𝑈\𝔅𝑅 (𝑥 ◦ ),

− Im 𝜑 (𝑥, 𝜃) ≤ −𝜅 1 𝑅 2 + 𝜅 2 |𝜃 − 𝜃 ◦ | .

If we take |𝜃 − 𝜃 ◦ | ≤ (𝜅 1 /2𝜅 2 ) 𝑅 2 we get, for every 𝜀 > 0,

1 2𝜆
| 𝑨𝑈 𝑓 (𝜃, 𝜆)| ≤ 𝐶 𝜀 e 𝜀𝜆− 2 𝜅1 𝑅 ∥ 𝑓 ∥ 𝐿1 ,

which proves that 𝑨𝑈 𝑓 (𝜃, 𝜆) ∈ O (−𝜔) (Θ′) provided the neighborhood Θ′ of 𝜃 ◦ is

sufficiently small (independently of 𝜆). This implies the claim. □

Let us denote by L1𝑥 ◦ the linear space of germs at 𝑥 ◦ of functions 𝑓 ∈ 𝐿 1 (R𝑛 ).

Propositions 22.2.7, 22.2.8 (and Corollary 20.4.4) allow us to state:

Proposition 22.2.9 The integral (22.1.9) defines a continuous linear map

( Φ♭ )
A 𝑥 ◦ , 𝜃 ◦ : L1𝑥 ◦ −→ O 𝜃 ◦ /O (−𝜔)
𝜃◦ . (22.2.7)

We shall also refer to operators such as (22.2.7) as germ FBI-like transforms

(cf. Definition 22.1.13).
Next we extend the action of integral operators such as (22.1.9) to distributions
𝑢 ∈ E ′ (Ω); thus we study the duality bracket
∫
𝑨𝑢 (𝜃, 𝜆) = e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) 𝑢 (𝑥) d𝑥. (22.2.8)
R𝑛

Of course, (22.2.8) must be understood as an oscillatory integral: to every 𝜀 > 0

there are functions 𝑓 𝛼 ∈ 𝐿 c1 (Ω), 𝛼 ∈ Z𝑛 , |𝛼| ≤ 𝑚, such that
∑︁
𝑢= D𝛼 𝑓𝛼 (22.2.9)
| 𝛼 | ≤𝑚

and dist (supp 𝑓 𝛼 , supp 𝑢) < 𝜀. If 𝑢 is given by (22.2.9) then

∑︁ ∫
𝑨𝑢 (𝜃, 𝜆) = e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑏 𝛼 (𝑥, 𝜃, 𝜆) 𝑓 𝛼 (𝑥) d𝑥, (22.2.10)
| 𝛼 | ≤𝑚 R𝑛
22.3 The Equivalence Theorem for Distributions 925

where

𝑏 𝛼 (𝑧, 𝜃, 𝜆) = e−𝑖𝜆𝜑 (𝑧, 𝜃) (−D𝑧 ) 𝛼 e𝑖𝜆𝜑 (𝑧, 𝜃) 𝑎 (𝑧, 𝜃, 𝜆) (22.2.11)

is easily seen to be an analytic symbol in ΩC × Θ, classical if 𝑎 is classical, with

principal symbol 𝜆 | 𝛼 | (−𝜕𝑧 𝜑) | 𝛼 | 𝑎 0 (𝑎 0 : principal symbol of 𝑎).
In the next statements the linear operator 𝑨 : E ′ (Ω) −→ O (Θ) is defined by
(22.2.8).

Proposition 22.2.10 Let 𝜑 ∈ O ΩC × Θ be an FBI phase-function at (𝑥 ◦ , 𝜉 ◦ ). Let

𝐾 ⊂ Ω be a compact set and Φ♭𝐾 (𝜃) as in (22.2.5). The linear operator 𝑨 maps
E ′ (𝐾) continuously into O ( Φ𝐾 ) (Θ).
♭

Proof Let 𝜀 > 0 be arbitrary; it suffices to deal with 𝑢 = 𝑓 𝛼 ∈ 𝐿 c1 (Ω) [cf. (22.2.9)]
with dist (supp 𝑓 𝛼 , R𝑛 \𝐾) so small that Φ♭𝐾 (𝜃) + Im 𝜑 (𝑥, 𝜃) ≥ −𝜀 if 𝑥 ∈ supp 𝑓 𝛼 .
Since 𝑏 𝛼 (𝑧, 𝜃, 𝜆) satisfies (AS) (Definition 20.1.1) we derive
∫
e−𝜆Φ𝐾 ( 𝜃) | 𝑨𝑢 (𝜃, 𝜆)| ≤ 𝐶 𝜀 e2𝜀𝜆
♭
| 𝑓 𝛼 (𝑥)| d𝑥. □
R𝑛

Proposition 22.2.11 Let 𝑈 ⊂⊂ ΩC be a neighborhood of 𝑥 ◦ and 𝑢 ∈ E ′ (𝑈). The

equivalence class mod O (−𝜔)
𝜃◦ of the germ at 𝜃 ◦ of 𝑨𝑢 (𝜃, 𝜆) depends solely on the
◦
germ at 𝑥 of 𝑢.

Proof In view of (22.2.10) it suffices to deal with 𝑢 = 𝑓 belonging to a bounded

subset of 𝐿 c1 (Ω), in which case the result is stated in Proposition 22.2.8. □
We deduce from Propositions 22.2.10 and 22.2.11 that (22.2.8) defines a contin-
uous linear map
( Φ♭ )
A 𝑥 ◦ , 𝜃 ◦ : D′𝑥 ◦ −→ O 𝜃 ◦ /O (−𝜔)
𝜃◦ . (22.2.12)
We shall extend the terminology of Definitions 22.1.13 to the operators (22.2.12).

22.3 The Equivalence Theorem for Distributions

22.3.1 Difference of two FBI phase-functions

◦ ◦
We are going to compare the (generic) FBI phase-function at (𝑥 , 𝜃 ) ∈ Ω ×Θ
(Ω = Ω ∩ R ) of the preceding section, 𝜑 (𝑧, 𝜃) ∈ O Ω × Θ , to 𝜑◦ (𝑧 − 𝑥 ◦ , 𝜁) =
C 𝑛 C

− (𝑧 − 𝑥 ◦ ) · 𝜁 + 21 𝑖 ⟨𝑧 − 𝑥 ◦ ⟩ 2 , regarded as an FBI phase-function at (𝑥 ◦ , 𝜉 ◦ ) in C𝑛 ×

(C𝑛 \ {0}) (cf. Example 22.1.2). Condition (1) in Definition 22.1.1 implies

𝜕𝑧 𝜑 (𝑥 ◦ , 𝜃 ◦ ) = 𝜕𝑧 𝜑◦ (𝑧 − 𝑥 ◦ , 𝜉 ◦ )| 𝑧=𝑥 ◦ = −𝜉 ◦ ∈ R𝑛 \ {0} . (22.3.1)

926 22 Germ FBI Transforms

As a general rule we shall assume that the domains ΩC and Θ are sufficiently small
(and geometrically simple) that the assertions based on their sizes and shapes are
valid. We associate with 𝜑◦ the exponent-weight (cf. Example 22.2.6)
1
Φ♭◦ (𝜁) = sup (− Im 𝜑◦ (𝑥, 𝜁)) = |Im 𝜁 | 2 . (22.3.2)
𝑥 ∈Ω 2

We introduce the phase-function

𝜓 (𝑧, 𝑤, 𝜁, 𝜃) = 𝜑 (𝑧, 𝜃) − 𝜑◦ (𝑧 − 𝑤, 𝜁) (22.3.3)

1
= 𝜑 (𝑧, 𝜃) + (𝑧 − 𝑤) · 𝜁 − 𝑖 ⟨𝑧 − 𝑤⟩ 2
2
in (𝑧, 𝑤, 𝜁, 𝜃)-space ΩC ×C𝑛 ×(C𝑛 \ {0})×Θ; (22.3.1) entails 𝜕𝑧 𝜓 (𝑥 ◦ , 𝑥 ◦ , 𝜉 ◦ , 𝜃 ◦ ) = 0.
We define, in ΩC × C𝑛 × (C𝑛 \ {0}) × Θ,

Ψ (𝑧, 𝑤, 𝜁, 𝜃) = Φ♭◦ (𝜁) − Im 𝜓 (𝑧, 𝑤, 𝜁, 𝜃) (22.3.4)

1
= |𝑥 − 𝑠 − 𝜂| 2 − (𝑦 − 𝑡) · 𝜉
2
1
− |𝑦 − 𝑡| 2 − Im 𝜑 (𝑧, 𝜃) ,
2
where 𝑠 = Re 𝑤, 𝑡 = Im 𝑤, 𝜉 = Re 𝜁, 𝜂 = Im 𝜁.

Proposition 22.3.1 There are neighborhoods Ω′C ⊂ ΩC of 𝑥 ◦ , Θ′ ⊂ Θ of 𝜃 ◦ , in C𝑛 ,

𝑈 of 𝑥 ◦ in R𝑛 , such that, for each (𝑠, 𝜃) ∈ 𝑈 × Θ′, the function (𝑧, 𝜁) ↦→ Ψ (𝑧, 𝑠, 𝜁, 𝜃)
has a unique critical point (𝑧 (𝑠, 𝜃) , 𝜁 (𝑠, 𝜃)) in Ω′C × Θ, with

𝑧 (𝑠, 𝜃) = 𝑥max (𝜃) , (22.3.5)

𝜁 (𝑠, 𝜃) = −𝜕𝑦 Im 𝜑 (𝑥max (𝜃) , 𝜃) + 𝑖 (𝑥max (𝜃) − 𝑠) , (22.3.6)

where 𝑥max (𝜃) is the unique critical point of the function 𝑥 ↦→ − Im 𝜑 (𝑥, 𝜃) in
Ω′ = Ω′C ∩ R𝑛 (cf. Proposition 22.2.1). The critical value of (22.3.4) is Φ♭ (𝜃) =
− Im 𝜑 (𝑥max (𝜃) , 𝜃) [cf. (22.2.3)] and (𝑧 (𝑠, 𝜃) , 𝜁 (𝑠, 𝜃)) is a saddle point of the
function (𝑧, 𝜁) ↦→ Ψ (𝑧, 𝑠, 𝜁, 𝜃).

Proof On the one hand, 𝜕𝑧 Ψ (𝑧, 𝑠, 𝜁, 𝜃) = 0 entails

𝜉 = −𝑦 − 𝜕𝑦 Im 𝜑 (𝑧, 𝜃) ,
𝜂 = 𝑥 − 𝑠 − 𝜕𝑥 Im 𝜑 (𝑧, 𝜃) .

On the other hand, 𝜕𝜁 Ψ (𝑧, 𝑠, 𝜁, 𝜃) = 0 entails 𝜂 = 𝑥 − 𝑠, 𝑦 = 0, and therefore

𝜉 = −𝜕𝑦 Im 𝜑 (𝑥, 𝜃) , 𝜕𝑥 Im 𝜑 (𝑥, 𝜃) = 0.

22.3 The Equivalence Theorem for Distributions 927

Then Proposition 22.2.2 entails 𝑥 = 𝑥 max (𝜃), whence (22.3.5)–(22.3.6). Substitution

of these values in Ψ (𝑧, 𝑠, 𝜁, 𝜃) yields Φ♭ (𝜃). Putting 𝑠 = 𝑥 ◦ , 𝜃 = 𝜃 ◦ in (22.3.5)–
(22.3.6) yields 𝑧 = 𝑥 ◦ , 𝜁 = 𝜉 ◦ ; because of stability, it suffices to prove that (𝑥 ◦ , 𝜉 ◦ )
is a saddle point of Ψ (𝑧, 𝑥 ◦ , 𝜁, 𝜃 ◦ ); this follows directly from the next proposition.□
Proposition 22.3.2 If diam Ω′C , diam Ξ and 𝜏 > 0 are sufficiently small then

ℭ +𝜏 = (𝑧, 𝜁) ∈ Ω′C × Ξ; 𝑥 − 𝑥 ◦ = 𝜂, 𝑦 = 𝜏 (𝜉 − 𝜉 ◦ )

is a strongly well-shaped contour for (𝑧, 𝜁) ↦→ Ψ (𝑧, 𝑥 ◦ , 𝜁, 𝜃 ◦ ) at (𝑥 ◦ , 𝜉 ◦ ), whereas

ℭ − = (𝑧, 𝜁) ∈ Ω′C × Ξ; 𝑥 = 𝑥 ◦ , 𝜉 − 𝜉 ◦ = 𝑦

is strongly well-shaped for (𝑧, 𝜁) ↦→ −Ψ (𝑧, 𝑥 ◦ , 𝜁, 𝜃 ◦ ) at (𝑥 ◦ , 𝜉 ◦ ).

Proof To simplify the notation we take 𝑥 ◦ = 0. Taylor expansion of Im 𝜑 (𝑧, 𝜃 ◦ )
about 0 and (22.3.4) yield
1 1
Ψ (𝑧, 0, 𝜁, 𝜃 ◦ ) = |𝑥 − 𝜂| 2 − 𝑦 · (𝜉 − 𝜉 ◦ ) − |𝑦| 2 (22.3.7)
2 ! 2
2 2
1 1 1
− 𝑀◦2 𝑥 − 𝑀◦2 𝑦
2

+ 𝑦 · 𝑁◦ 𝑥 + 𝑂 |𝑧| 3 ,

where 𝑀◦ = 𝜕𝑥2 Im 𝜑 (0, 𝜃 ◦ ) is positive definite, 𝑁◦ = −𝜕𝑥 𝜕𝑦 Im 𝜑 (0, 𝜃 ◦ ). We have,

in ℭ +𝜏 ,
2
1 1 1
Ψ (𝑧, 0, 𝜁, 𝜃 ) ≤ − 𝑀◦ 𝑥 − 𝜏 1 + 𝜏 (1 − ∥𝑀◦ ∥) |𝜉 − 𝜉 ◦ | 2
◦ 2
2 2
1
+ 𝜏 ∥𝑁◦ ∥ |𝑥| |𝜉 − 𝜉 ◦ | + 𝑂 |𝑥| 3 + 𝜏 3 |𝜉 − 𝜉 ◦ | 3 ,
2
which proves the first claim in the statement. In ℭ − we have

1 2
◦ 1/2
Ψ (𝑧, 0, 𝜁, 𝜃 ) ≥ |𝜂| + |𝑦| + 𝑀◦ 𝑦 + 𝑂 |𝑦| 3 ,
2 2
2

which proves the second claim. □

End of the proof of Proposition 22.3.1. Each one of the sets ℭ +𝜏 and ℭ − is an open
subset of a 2𝑛-dimensional real affine subspace of C2𝑛 through (𝑥 ◦ , 𝜉 ◦ ). The two
affine subspaces are obviously totally real and transversal to one another, which
implies that (𝑥 ◦ , 𝜉 ◦ ) is a saddle point of Ψ (𝑧, 𝑥 ◦ , 𝜁, 𝜃 ◦ ). □
Remark 22.3.3 The fact that (𝑥 ◦ , 𝜉 ◦ ) is a saddle point of Ψ (𝑧, 𝜃 ◦ , 𝑥 ◦ , 𝜁) can also be
proved directly, i.e., without relying on Proposition 22.3.2. Indeed, a simple linear
algebra argument shows that the Hessian of Ψ (𝑧, 𝑥 ◦ , 𝜁, 𝜃 ◦ ) at (𝑥 ◦ , 𝜉 ◦ ),
928 22 Germ FBI Transforms

𝐼𝑛 − 𝑀◦ 𝑁◦ 0 −𝐼𝑛
𝑁◦ 𝑀◦ − 𝐼𝑛 −𝐼𝑛 0 ®
© ª
®,
­
0 −𝐼𝑛 0 0 ®
­
­
« −𝐼𝑛 0 0 𝐼𝑛 ¬

[cf. (22.3.8; 𝐼𝑛 : identity 𝑛 × 𝑛 matrix)] is nonsingular and its signature is equal to

zero.

In the terminology of Definition 20.3.6 and regarding (𝑧, 𝜁) ∈ ΩC × (C𝑛 \ {0}) as

the “base variable” and (𝑤, 𝜃) ∈ C𝑛 × Θ as the “fiber variable”, Proposition 22.3.1
leads to the following

Proposition 22.3.4 If 𝜓 (𝑧, 𝜃, 𝑤, 𝜁) = 𝜑 (𝑧, 𝜃) − 𝜑◦ (𝑧 − 𝑤, 𝜁) then 𝜓, Φ♭◦ [cf.
(22.3.2)] is a Sjöstrand pair at ((𝑥 ◦ , 𝜉 ◦ ) , (𝑥 ◦ , 𝜃 ◦ )) ∈ Ω × (R𝑛 \ {0}) × Ω × Θ.

We continue to look at the variables (𝑧, 𝜁) and (𝑤, 𝜃) as playing the role of 𝑧 and
𝜃 respectively in Ch. 20; the next statement establishes an extension of (21.1.28).

Proposition 22.3.5 In a neighborhood of ((𝑥 ◦ , 𝜉 ◦ ) , (𝑥 ◦ , 𝜃 ◦ )) ∈ ΩC × (C𝑛 \ {0}) ×

C𝑛 × Θ we have 2 2 !
𝜕 𝜓 𝜕 𝜓
det 𝜕𝑧𝜕𝑤 𝜕𝑧𝜕𝜃
𝜕2 𝜓 𝜕2 𝜓 ≠ 0.
𝜕𝑤𝜕𝜁 𝜕𝜁 𝜕𝜃

Proof We derive from (22.3.3):

𝜕2𝜓
(𝑧, 𝜃, 𝑤, 𝜁) = 𝑖𝐼𝑛 ,
𝜕𝑧𝜕𝑤
𝜕2𝜓 𝜕2 𝜑
(𝑧, 𝜃, 𝑤, 𝜁) = (𝑧, 𝜃) ,
𝜕𝑧𝜕𝜃 𝜕𝑧𝜕𝜃
𝜕2𝜓
(𝑧, 𝜃, 𝑤, 𝜁) = −𝐼𝑛 ,
𝜕𝑤𝜕𝜁
𝜕2𝜓
(𝑧, 𝜃, 𝑤, 𝜁) = 0.
𝜕𝜁 𝜕𝜃
We have
𝑖𝐼𝑛 𝐴
det = det 𝐴.
−𝐼𝑛 0
The claim then follows from (21.1.28). □

Corollary 22.3.6 If Φ♭ (𝜃) is the function (22.2.3) then −𝜓, Φ♭ is the dual Sjös-

trand pair of 𝜓, Φ♭◦ at ((𝑥 ◦ , 𝜃 ◦ ) , (𝑥 ◦ , 𝜉 ◦ )) (Definition 20.3.18).

Proof Combine the preceding results with Theorem 20.3.16. □

22.3 The Equivalence Theorem for Distributions 929

22.3.2 Transfer operators

We continue to use the notation 𝜓 (𝑧, 𝑤, 𝜁, 𝜃) = 𝜑 (𝑧, 𝜃) − 𝜑◦ (𝑧 − 𝑤, 𝜁). Propositions

22.3.2 and 22.3.4 enable us to make use of the construction of the operator (20.4.4)
( Φ♭ ) ( Φ♭ )
and define a germ operator O ( 𝑥 ◦◦, 𝜉 ◦ ) −→ O ( 𝑥 ◦ , 𝜃 ◦ ) starting from integrals
𝑛 ∫
𝜆
e𝑖𝜆𝜓 (𝑧,𝑤,𝜁 , 𝜃) 𝑎 (𝑧, 𝑤, 𝜁, 𝜃, 𝜆) 𝑔(𝑧, 𝜁, 𝜆)d𝑧d𝜁 (22.3.8)
2𝜋 ℭ +𝜏,𝑟

with 𝑎 (𝑧, 𝑤, 𝜁, 𝜃, 𝜆) ∈ 𝑆class ΩC × C𝑛 × Ξ × Θ , 𝑔 ∈ O ( Φ◦ ) ΩC × Ξ , Ξ a domain

♭
in C𝑛 , 𝜉 ◦ ∈ Ξ. The contour of integration is the set ℭ +𝜏,𝑟 of points (𝑧, 𝜁) ∈ ΩC × Ξ
such that
𝑥 ∈ 𝑈, 𝑦 = 𝜏 (𝜉 − 𝜉 ◦ ) , |𝜉 − 𝜉 ◦ | < 𝑟, 𝜂 = 𝑥 − 𝑥 ◦ ,
with 𝑈 a smoothly bounded neighborhood of 𝑥 ◦ in Ω′ and sufficiently small diam 𝑈,
𝜏 > 0, 𝑟 > 0; Proposition 22.3.2 implies that ℭ +𝜏,𝑟 is a strongly well-shaped contour
for (𝑧, 𝜁) ↦→ Ψ (𝑧, 𝑥 ◦ , 𝜁, 𝜃 ◦ ) at (𝑥 ◦ , 𝜉 ◦ ). Actually, for our purpose we can simplify
the contour of integration.

Lemma 22.3.7 Let 𝑉 C ⊂ ΩC be a neighborhood of 𝑥 ◦ in C𝑛 , Θ′ ⊂ Θ a domain in

C𝑛 , Θ′ ∋ 𝜃 ◦ . If 𝑟 > 0, 𝜏 ◦ > 0, diam 𝑉 C , diam Θ′, are sufficiently small then the
restriction to 𝑉 C × Θ′ of the integral (22.3.8) is congruent mod O (−𝜔) 𝑉 C × Θ′ to
the integral

𝑨𝑔 (𝑤, 𝜃, 𝜆)
e (22.3.9)
𝑛 ∫ ∫
𝜆
= e𝑖𝜆𝜓 (𝑧,𝑤,𝜁 , 𝜃) 𝑎 (𝑥, 𝑤, 𝜁, 𝜃, 𝜆) 𝑔 (𝑥, 𝜁, 𝜆) d𝑥d𝜉,
2𝜋 | 𝜉 − 𝜉 ◦ |<𝑟 𝑈

where 𝑔 ∈ O ( Φ◦ ) ΩC × Ξ is arbitrary, 𝜁 = 𝜉 + 𝑖 (𝑥 − 𝑥 ◦ ).
♭

Proof Once again we assume that 𝑥 ◦ = 0; (22.3.7) yields

2
1 1
Ψ (𝑥 + 𝑖𝜏 (𝜉 − 𝜉 ◦ ) , 𝑤, 𝜉 + 𝑖𝑥, 𝜃) = − 𝑀◦2 𝑥 − 𝜏 (1 + 𝜏/2) |𝜉 − 𝜉 ◦ | 2
2
2
1 1
+ 𝜏 2 𝑀◦2 (𝜉 − 𝜉 ◦ ) + 𝜏 |𝜉 − 𝜉 ◦ | · 𝑁◦ 𝑥 + 𝑂 |𝑥| 3 + 𝜏 3 |𝜉 − 𝜉 ◦ | 3 ,
2

implying that if 𝑥 ∈ 𝜕𝑈, and if 𝑟 and 𝜏 ◦ are sufficiently small, then

Ψ (𝑥 + 𝑖, 𝑤, 𝜉 + 𝑖𝑥, 𝜃) ≤ − 𝑐 ◦ 𝑑 2 − 𝜏 |𝜉 − 𝜉 ◦ | 2 − 𝐶 (|𝑤| + |𝜃 − 𝜃 ◦ |)

for some 𝑐 ◦ , all (𝑤, 𝜃) ∈ 𝑉 C × Θ′, 𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | < 𝑟, 0 ≤ 𝜏 ≤ 𝜏 ◦ . As a

consequence of this we have
930 22 Germ FBI Transforms
∫ ∫
e𝑖𝜆𝜓 (𝑧,𝑤,𝜁 , 𝜃) 𝑎 (𝑧, 𝑤, 𝜁, 𝜃, 𝜆) 𝑔 (𝑥 + 𝑖𝜏 |𝜉 − 𝜉 ◦ | , 𝜁) d𝑥d𝜉
| 𝜉 − 𝜉 ◦ |<𝑟 𝜕𝑈

≤ 𝐶e−𝜅𝜆 e−𝜆Φ◦ ( 𝜉 +𝑖 𝑥) |𝑔 (𝑥 + 𝑖𝜏 |𝜉 − 𝜉 ◦ | , 𝜁)| .

♭
max
𝑥 ∈𝜕𝑈, | 𝜉 − 𝜉 ◦ |<𝑟

where 𝜁 = 𝜉 + 𝑖𝑥 and 𝐶 > 0, 𝜅 > 0 are suitably large and small respectively. Then
the claim is a consequence of Stokes’ Theorem and of Definition 20.2.7. □
Since it is obvious that (22.3.8) and (22.3.9) map O (−𝜔) ΩC × Ξ into

O (−𝜔) 𝑉 C × Θ′ , we reach the conclusion that these integrals define a linear map

e ( 𝑥 ◦ , 𝜉 ◦ ), ( 𝑥 ◦ , 𝜃 ◦ ) : O ( Φ◦◦ ) ◦ /O (−𝜔) ( Φ♭ )
♭
A (𝑥 , 𝜉 ) ( 𝑥◦ , 𝜉 ◦)
−→ O ( 𝑥 ◦ , 𝜃 ◦ ) /O (−𝜔)
( 𝑥◦ , 𝜉 ◦)
. (22.3.10)

Before proceeding we deduce the following application of Theorem 21.2.10, and

Corollary 22.3.6.
Proposition 22.3.8 If the classical symbol 𝑎 is elliptic then (22.3.10) is invertible.
Proof The definition (22.3.3) implies

𝜕𝑧 𝜓 (𝑥 ◦ , 𝑥 ◦ , 𝜉 ◦ , 𝜃 ◦ ) = 𝜕𝜁 𝜓 (𝑥 ◦ , 𝑥 ◦ , 𝜉 ◦ , 𝜃 ◦ ) = 0

while (22.3.2) shows that 𝜕𝜁 Φ♭◦ (𝜉 ◦ ) = 0. Together with Proposition 22.3.5 this
proves that the hypotheses in Theorem 21.2.10, are satisfied. □
With the classical symbol 𝑎 elliptic we obtain the germ operator inverse of
(22.3.10),
e−1◦ ◦ ◦ ◦ : O ( Φ◦ ) ◦ −→ O ( Φ◦◦ ) ◦ .
♭ ♭
A ( 𝑥 , 𝜃 ), ( 𝑥 , 𝜉 ) (𝑥 ,𝜃 ) (𝑥 , 𝜉 )
(22.3.11)
The term transfer operators in the heading of this subsection refers to (elliptic)
operators such as (22.3.10) or (22.3.11).

22.3.3 Exploiting the transfer operators

Let 𝑈 be the neighborhood of 𝑥 ◦ introduced in the preceding subsection and 𝑓 ∈

E ′ (𝑈); recalling that 𝜑◦ (𝑧 − 𝑥 ◦ , 𝜁) = − (𝑧 − 𝑥 ◦ ) · 𝜁 + 21 𝑖 ⟨𝑧 − 𝑥 ◦ ⟩ 2 we introduce [cf.
(7.3.4)] ∫
◦ ,𝜁 )
𝑭 𝑓 (𝑥 ◦ , 𝜆𝜁) = e𝑖𝜆𝜑◦ (𝑠−𝑥 𝑓 (𝑠) d𝑠 (22.3.12)

with the integral standing for the duality bracket. Proposition 22.2.7 and Ex-
ample 22.2.6 imply that the linear operator 𝑭 maps E ′ (Ω) continuously into
O ( Φ◦ ) ΩC × Θ , where Φ♭◦ (𝜁) = 21 |Im 𝜁 | 2 .
♭

Let 𝜑 be an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ) in ΩC × Θ satisfying the conditions

𝑨𝑭 𝑓 (𝑥 ◦ , 𝜁) [see (22.3.9)] to 𝑨 𝑓 (𝜃, 𝜆) as defined
(22.3.1); we are going to relate e
in (22.2.5). By Proposition 22.2.7 we have 𝑨 𝑓 (𝜃, 𝜆) ∈ O ( Φ ) (Θ′), where Φ♭ is the
♭
22.3 The Equivalence Theorem for Distributions 931

exponent-weight (22.2.3) and Θ′ ⊂ Θ open, 𝜃 ◦ ∈ Θ′. We put 𝑔 = 𝑭 𝑓 in (22.3.9)

where we take 𝑎 (𝑧, 𝑤, 𝜁, 𝜃, 𝜆) = 𝑎 (𝑧, 𝜃, 𝜆):
−𝑛
𝜆 e𝑨𝑭 𝑓 (𝑥 ◦ , 𝜃, 𝜆) (22.3.13)
2𝜋
∫ ∫ ∫
= 𝑎 (𝑧, 𝜃, 𝜆) e𝑖𝜆𝜒 (𝑧,𝑠,𝜁 , 𝜃) 𝑓 (𝑠) d𝑠 d𝑧d𝜁,
| 𝜉 − 𝜉 ◦ |<𝑟 𝑈

where

𝜒 (𝑧, 𝑤, 𝜁, 𝜃) = 𝜑 (𝑧, 𝜃) − 𝜑◦ (𝑧 − 𝑥 ◦ , 𝜁) + 𝜑◦ (𝑤 − 𝑥 ◦ , 𝜁) (22.3.14)

1
= 𝜑 (𝑧, 𝜃) + (𝜁 + 𝑖𝑥 ◦ ) · (𝑧 − 𝑤) − 𝑖 ⟨𝑧⟩ 2 − ⟨𝑤⟩ 2 .
2
We now carry out a reduction of the integrals (22.3.9) and (22.3.13) or, rather, of
their equivalence classes mod O (−𝜔) (Θ). We begin by looking at Im 𝜒 (𝑧, 𝑠, 𝜁, 𝜃)
on ℭ +𝜏,𝑟 .

Lemma 22.3.9 Let 𝑈 ⊂ Ω be a neighborhood of 𝑥 ◦ in R𝑛 . Provided diam 𝑈 and the

positive constants 𝑟, 𝜏 ◦ , 𝑐 ◦ are sufficiently small we have

Im 𝜒 (𝑥 + 𝑖𝜏 (𝜉 − 𝜉 ◦ ) , 𝑠, 𝜉 + 𝑖 (𝑥 − 𝑥 ◦ ) , 𝜃 ◦ ) (22.3.15)
1
≥ 𝑐 ◦ |𝑥 − 𝑥 ◦ | 2 + 𝑐 ◦ 𝜏 |𝜉 − 𝜉 ◦ | 2 + |𝑥 − 𝑠| 2
2
for all 𝑥 ∈ 𝑈, 𝑠 ∈ R𝑛 , 0 ≠ 𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | < 𝑟, 𝜏 ∈ [0, 𝜏 ◦ ].

Proof To lighten the notation we take 𝑥 ◦ = 0; we have

1
Im (𝜁 · (𝑧 − 𝑠)) − Re ⟨𝑧⟩ 2 − ⟨𝑠⟩ 2 = 𝜉 · 𝑦 + 𝜂 · (𝑥 − 𝑠)
2
1 2
+ |𝑥| − |𝑦| 2 − |𝑠| 2 + |𝑡| 2 .
2
Since 𝑖𝜕𝑥 + 𝜕𝑦 Im 𝜑 (𝑥 ◦ , 𝜃 ◦ ) = −𝜉 ◦ (Definition 22.1.1, (1)), Taylor expansion of

Im 𝜑 (𝑧, 𝜃 ◦ ) about 𝑧 = 0 reads
!
2 2
1 1 1
Im 𝜑 (𝑧, 𝜃 ◦ ) = −𝑦 · (𝜉 ◦ − 𝑁◦ 𝑥) + 𝑀◦2 𝑥 − 𝑀◦2 𝑦 + 𝑂 |𝑧| 3
2

whence, by (22.3.14),
1 2
Im 𝜒 (𝑧, 𝑠, 𝜁, 𝜃 ◦ ) = 𝑦 · (𝜉 − 𝜉 ◦ ) + (𝑥 − 𝑠) · 𝜂 − |𝑥| − |𝑦| 2 − |𝑠| 2 (22.3.16)
! 2
2 2
1 1 1
+ 𝑀◦2 𝑥 − 𝑀◦2 𝑦 + 𝑦 · 𝑁◦ 𝑥 + 𝑂 |𝑧| 3 .
2
932 22 Germ FBI Transforms

We put 𝑦 = 𝜏 (𝜉 − 𝜉 ◦ ), 𝜂 = 𝑥:
2
1 1 1
Im 𝜒 (𝑧, 𝑠, 𝜁, 𝜃 ◦ )| ℭ+𝜏,𝑟 = 𝑀◦2 𝑥 + |𝑥 − 𝑠| 2
2 2
2
1 1 1
+ 𝜏 1 + 𝜏 |𝜉 − 𝜉 ◦ | 2 − 𝜏 2 𝑀◦2 (𝜉 − 𝜉 ◦ )
2 2

+ 𝜏 (𝜉 − 𝜉 ◦ ) · 𝑁◦ 𝑥 + 𝑂 |𝑥| 3 + 𝜏 3 |𝜉 − 𝜉 ◦ | 3 .

Since 𝑀◦ is positive definite the claim ensues easily. □

In order to prove the result we are seeking, we need a variant of Lemma 22.3.10
(in which |𝜉 − 𝜉 ◦ | < 𝑟) for the case |𝜉 − 𝜉 ◦ | > 𝑟. Here also we look at the imaginary
part of 𝜒 (𝑧, 𝑠, 𝜁, 𝜃).

Lemma 22.3.10 To every 𝑟 > 0 there are sufficiently small positive constants 𝜏 ◦ , 𝑐,
such that
𝜉 − 𝜉◦

1
Im 𝜒 𝑥 + 𝑖𝜏 ◦
, 𝑠, 𝜉 + 𝑖𝑥, 𝜃 ≥ 𝑐 |𝑥 − 𝑥 ◦ | 2 + 𝑐𝜏 |𝜉 − 𝜉 ◦ | + |𝑥 − 𝑠| 2
◦
|𝜉 − 𝜉 | 2

for all 𝑥 ∈ 𝑈, 𝑠 ∈ R𝑛 , 0 ≠ 𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | ≥ 𝑟, 𝜏 ∈ [0, 𝜏 ◦ ].

Proof Again we assume 𝑥 ◦ = 0. First of all, we require 𝜏 ◦ to be small enough that

𝑥 + 𝑖𝜏v ∈ ΩC for all 0 ≤ 𝜏 ≤ 𝜏 ◦ , v ∈ R𝑛 , |v| ≤ 1, and 𝑥 ∈ 𝑈. This allows us to put
◦
𝑦 = 𝜏 | 𝜉𝜉 −− 𝜉𝜉 ◦ | , Im 𝜉 = 𝑥 in (22.3.16), to get

𝜉 − 𝜉◦

◦ ◦ 1 2 2

Im 𝜒 𝑥 + 𝑖𝜏 , 𝑠, 𝜉 + 𝑖𝑥, 𝜃 = 𝜏 |𝜉 − 𝜉 | + |𝑥 − 𝑠| + 𝜏
|𝜉 − 𝜉 ◦ | 2
2 1 𝜉 − 𝜉◦
2
1 1 1
+ 𝑀◦2 𝑥 − 𝜏 2 𝑀◦2
2 2 |𝜉 − 𝜉 ◦ |
𝜉 − 𝜉◦
3 3

+𝜏 · 𝑁 ◦ 𝑥 + 𝑂 |𝑥| + 𝜏 .
|𝜉 − 𝜉 ◦ |
We use the fact that
𝜉 − 𝜉◦
𝜏 · 𝑁◦ 𝑥 ≥ −𝜏 4/3 − 𝜏 2/3 |𝑁◦ 𝑥| 2
|𝜉 − 𝜉 ◦ |
to obtain
𝜉 − 𝜉◦

◦ 1
Im 𝜒 𝑥 + 𝑖𝜏 ◦
, 𝑠, 𝜉 + 𝑖𝑥, 𝜃 ≥ 𝜏 |𝜉 − 𝜉 ◦ | + |𝑥 − 𝑠| 2
|𝜉 − 𝜉 | 2
2
1 1
+ 𝑀◦2 𝑥 − 𝐶 |𝑥| 3 + 𝜏 4/3
2

for a suitably large 𝐶 > 0; the claim ensues easily. □

22.3 The Equivalence Theorem for Distributions 933

Proposition 22.3.11 If 𝑈 and Θ are sufficiently small and if the boundary 𝜕𝑈 of 𝑈

is smooth then
𝑨𝑭 𝑓 (𝑥 ◦ , 𝜃, 𝜆) ∈ O (−𝜔) (Θ)
𝑨 𝑓 (𝜃, 𝜆) − e (22.3.17)
for every 𝑓 ∈ E ′ (𝑈).

Proof We continue to assume 𝑥 ◦ = 0. Until

specified otherwise we assume 𝑓 ∈
𝐿 c (𝑈). Lemma 22.3.7 implies that 𝑨𝑭 𝑓 (𝑠, 𝜃, 𝜆) is congruent mod O (−𝜔) (Θ) to
1 e

𝑛 ∫ ∫ ∫
𝜆
e𝑖𝜆𝜒 ( 𝑥,𝑠, 𝜉 +𝑖 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) 𝑓 (𝑠) d𝑠d𝑥d𝜉 (22.3.18)
2𝜋 | 𝜉 − 𝜉 ◦ |<𝑟 𝑥 ∈𝑈 𝑠 ∈R𝑛

provided sup |𝜃 − 𝜃 ◦ | is sufficiently small. We then look at the integral

𝜃 ∈Θ
∫ ∫
𝑰𝑈, 𝜏 (𝑠, 𝜃, 𝜆) = e𝑖𝜆𝜒 ( 𝑥,𝑠, 𝜉 +𝑖 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) 𝑓 (𝑠) d𝑠d𝑥d𝜉
ℭ′𝜏,𝑟 𝑠 ∈R𝑛

where
𝜉 − 𝜉◦

′
ℭ 𝜏,𝑟 = (𝑧, 𝜁) ∈ ΩC × Ξ; 𝑥 ∈ 𝑈, 𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | > 𝑟, 𝜂 = 𝑥, 𝑦 = 𝜏 .
|𝜉 − 𝜉 ◦ |
◦
We take diam 𝑈 and 𝜏 ◦ > 0 sufficiently small that 𝑥 + 𝑖𝜏 | 𝜉𝜉 −− 𝜉𝜉 ◦ | stays in a compact
subset of Ω′C whatever 𝑥 ∈ 𝑈, 𝜉 ∈ R𝑛 \ {𝜉 ◦ }, 0 ≤ 𝜏 ≤ 𝜏 ◦ , and that Lemma
22.3.10 applies with 𝑟 = inf◦ |𝜉 − 𝜉 ◦ |. We get for some 𝐶 > 0 and all 𝑥 ∈ 𝑈,
| 𝜉 − 𝜉 |>𝑟
𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | > 𝑟,

𝜉 − 𝜉◦

2
Im 𝜒 𝑥 + 𝑖𝜏 , 𝑠, 𝜉 + 𝑖𝑥, 𝜃 ≥ 𝑐 |𝑥| + 𝜏𝑟 − 𝐶 |𝜃 − 𝜃 ◦ | . (22.3.19)
|𝜉 − 𝜉 ◦ |

Putting 𝜏 = 𝜏 ◦ in (22.3.19) shows that 𝑰𝑈, 𝜏 ◦ (𝑠, 𝜃, 𝜆) ∈ O −𝜔 (Θ) if 𝐶 |𝜃 − 𝜃 ◦ | <

′
𝑐𝜏 ◦ 𝑟/2.
We apply Stokes’ Theorem and deform ℭ 𝜏 ◦ ,𝑟 to the contour of integration
′
ℭ𝑟 = {(𝑥, 𝜁) ∈ 𝑈 × Ξ; 𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | > 𝑟, 𝜂 = 𝑥} .

This is possible because min |𝑥| ≥ 𝑑 > 0 and therefore, by (22.3.19), the contribution
𝜕𝑈
from the “cylindrical” part,

𝜉 − 𝜉◦

(𝑧, 𝜁) ∈ C2𝑛 ; 𝑥 ∈ 𝜕𝑈, 𝜉 ∈ R𝑛 , |𝜉 − 𝜉 ◦ | > 𝑟, 𝜂 = 𝑥, 𝑦 = 𝜏 , 0 ≤ 𝜏 ≤ 𝜏 ◦
,
|𝜉 − 𝜉 ◦ |

belongs to O (−𝜔) (Θ) provided 𝐶 |𝜃 − 𝜃 ◦ | < 𝑐𝑑 2 /2. Let Θ′ denote the subset of
Θ defined by 𝐶 |𝜃 − 𝜃 ◦ | < 21 𝑐 min 𝑑 2 , 𝜏 ◦ 𝑟 . We have proved that 𝑰𝑈,0 (𝑠, 𝜃, 𝜆) −
𝑰𝑈, 𝜏 ◦ (𝑠, 𝜃, 𝜆) ∈ O −𝜔 (Θ′), implying that
934 22 Germ FBI Transforms
∫ ∫
𝑰𝑈,0 (𝑠, 𝜃, 𝜆) = e𝑖𝜆𝜑 (𝑧, 𝜃)−𝑖𝜆𝜑◦ (𝑧,𝜁 ) 𝑎 (𝑥, 𝜃, 𝜆) 𝑭 𝑓 (𝑠, 𝜉 + 𝑖𝑥) d𝑥d𝜉
| 𝜉 − 𝜉 ◦ |>𝑟 𝑈

(−𝜔) (Θ ′ ). By adding 𝑰
to O
belongs
𝑈,0 (𝑠, 𝜃, 𝜆) to (22.3.18) we conclude that
−𝜔
𝑨𝑭 𝑓 (𝑠, 𝜃, 𝜆) is congruent mod O (Θ) to
e

𝑛 ∫ ∫ ∫
𝜆
e𝑖𝜆𝜒 ( 𝑥,𝑠, 𝜉 +𝑖 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) 𝑓 (𝑠) d𝑠d𝑥d𝜉. (22.3.20)
2𝜋 R𝑛 𝑥 ∈𝑈 𝑠 ∈R𝑛

By (22.3.14) (where 𝑥 ◦ = 0) we have

𝜒 (𝑥, 𝑠, 𝜉 + 𝑖𝑥, 𝜃) = 𝜑 (𝑥, 𝜃) − 𝜑◦ (𝑥, 𝜉 + 𝑖𝑥) + 𝜑◦ (𝑠, 𝜉 + 𝑖𝑥)

1
= 𝜑 (𝑥, 𝜃) + (𝑥 − 𝑠) · 𝜉 + 𝑖 |𝑥 − 𝑠| 2 .
2

We derive that e 𝑨𝑭 𝑓 (𝑠, 𝜃, 𝜆) is congruent mod O −𝜔 (Θ) to
∫
e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎 (𝑥, 𝜃, 𝜆) 𝑔 (𝑥) d𝑥, (22.3.21)
𝑈

where 𝑛 ∫
𝜆 1 2
𝑔 (𝑥) = e𝑖𝜆( 𝑥−𝑠) · 𝜉 − 2 𝜆 | 𝑥−𝑠 | 𝑓 (𝑠) d𝑠d𝜉.
2𝜋 R2𝑛

We can perform the change of variables 𝜆𝜉 ⇝ 𝜉. [Recall that 𝑎 (𝑥, 𝜃, 𝜆) ≡ 0 if

𝜆 < 1.] We get
∫
1 2
−𝑛
𝑔 (𝑥) = (2𝜋) e𝑖 ( 𝑥−𝑠) · 𝜉 − 2 𝜆 | 𝑥−𝑠 | 𝑓 (𝑠) d𝑠d𝜉.
R2𝑛

The Fourier inversion formula entails 𝑔 = 𝑓 and (22.3.21) is seen to be equal to

𝑨 𝑓 (𝜃, 𝜆) [cf. (22.2.5)].
Every compactly supported distribution in Ω is the sum of a finite number of
derivatives of functions 𝑓 ∈ 𝐿 c1 (𝑈). Since the commutators of the integral operators
𝑨, e𝑨, 𝑭 with partial derivatives are operators of the same kind with modified am-
plitudes, the claim in Proposition 22.3.11 extends straightforwardly to 𝑓 ∈ E ′ (𝑈).□

22.3.4 Equivalence of classical elliptic germ FBI transforms of

distributions with different phase-functions

We return to our generic “central” point 𝑥 ◦ ∈ Ω. We can “germify” Proposition

22.3.11 using the germ operators (22.2.8), A 𝑥 ◦ , 𝜃 ◦ , (22.3.10), A
e ( 𝑥 ◦ , 𝜉 ◦ ), ( 𝑥 ◦ , 𝜃 ◦ ) , and
F 𝑥 ◦ , 𝜃 ◦ , the latter defined by the “integral” (22.3.12).
22.3 The Equivalence Theorem for Distributions 935

Proposition 22.3.12 We have

e ( 𝑥 ◦ , 𝜉 ◦ ), ( 𝑥 ◦ , 𝜃 ◦ ) F 𝑥 ◦ , 𝜉 ◦ = A 𝑥 ◦ , 𝜃 ◦ : D′ ◦ −→ O ( Φ◦ ) /O (−𝜔)
♭
A 𝑥 𝜃 𝜃◦ .

We now state and prove the result we were aiming for. We regard O 𝑥(−𝜔)
◦ as a
linear subspace of L1𝑥 ◦ .

Theorem 22.3.13 Let 𝜉 ◦ ∈ R𝑛 \ {0} and, for 𝑗 = 1, 2, let 𝜑 𝑗 be an FBI phase-function

(𝜑𝑗)

𝜕𝜑
in Ω×Θ at 𝑥 ◦ , 𝜃 ( 𝑗) (0 ≠ 𝜃 ( 𝑗) ∈ C𝑛 ) such that 𝜕𝑧𝑗 (𝑥 ◦ , 𝜃 ( 𝑗) ) = −𝜉 ◦ . If A 𝑥 ◦ , 𝜃 ( 𝑗) is the
analogue of (22.2.8) with 𝜑 𝑗 substituted for 𝜑, the elliptic symbol 𝑎 𝑗 ∈ 𝑆class (Ω × Θ)
( 𝑗)
substituted for 𝑎 and 𝜃 ( 𝑗) for 𝜃 ◦ , then the germ operators A 𝑥 ◦ , 𝜃 ( 𝑗) : L1𝑥 ◦ /O (−𝜔)
𝜃◦ −→

Φ♭𝑗
O 𝜃 ( 𝑗) /O (−𝜔)
𝜃 ( 𝑗)
, 𝑗 = 1, 2, are equivalent, in the sense that there is an invertible linear
e (1,2) ( Φ♭1 ) ( Φ♭2 ) (2) e (1,2) (1)
operator A 𝑥 ◦ , 𝜉 ◦ : O 𝜃 (1) −→ O 𝜃 (2) such that A 𝑥 ◦ , 𝜃 (2) = A 𝑥 ◦ , 𝜉 ◦ A 𝑥 ◦ , 𝜃 (1) .

We have used the notation Φ♭𝑗 (𝜃) = sup − Im 𝜑 𝑗 (𝑥, 𝜃) (assumed to be
𝑥 ∈Ω∩R𝑛
< +∞).
Proof First consider the case 𝜃 (1) = 𝜃 (2) = 𝜃 ◦ . Let A e ( 𝑗)◦ ◦ ◦ ◦ denote the
( 𝑥 , 𝜉 ), ( 𝑥 , 𝜃 )
analogue of (22.3.10) for 𝜑 = 𝜑 𝑗 , 𝑎 = 𝑎 𝑗 ; since 𝑎 𝑗 is elliptic the map A e ( 𝑗)◦ ◦ ◦ ◦
( 𝑥 , 𝜉 ), ( 𝑥 , 𝜃 )
is bijective (Proposition 22.3.8). Proposition 22.3.12 implies
−1 −1
e (2)◦ ◦ ◦ ◦ (𝜑 ) e (1)◦ ◦ ◦ ◦ (𝜑 )
F𝑥◦ , 𝜉 ◦ = A ( 𝑥 , 𝜉 ), ( 𝑥 , 𝜃 )
A 𝑥 ◦ ,2𝜃 ◦ = A ( 𝑥 , 𝜉 ), ( 𝑥 , 𝜃 )
A 𝑥 ◦ ,1𝜃 ◦ .

The claim is proved in this case by taking

−1
e (1,2)
A = A
e (2)
A
e (1)
.
𝑥◦ , 𝜉 ◦ ( 𝑥 ◦ , 𝜉 ◦ ), ( 𝑥 ◦ , 𝜃 ◦ ) ( 𝑥 ◦ , 𝜉 ◦ ), ( 𝑥 ◦ , 𝜃 ◦ )

Now consider the case 𝜃 (1) ≠ 𝜃 (2) and define 𝜑˜ 2 (𝑧, 𝜃) = 𝜑 𝑗 𝑧, 𝜃 − 𝜃 (1) + 𝜃 (2) , 𝑗 =

𝜕 𝜑˜
1, 2; then 𝜑˜ 𝑗 0, 𝜃 (1) = 0, 𝜕𝑧𝑗 (𝑥 ◦ , 𝜃 (1) ) = −𝜉 ◦ , and the preceding argument applies
with 𝜑˜ 2 in place of 𝜑2 . The translation
↦ 𝜃 − 𝜃 (1) + 𝜃 (2) induces isomorphisms
𝜃 →
of each relevant germ space at 𝑥 ◦ , 𝜃 (1) for 𝜑˜ 2 onto its analogue at 𝑥 ◦ , 𝜃 (2) for
(𝜑 ) ( 𝜑˜ )
𝜑2 and the corresponding pullback of the germ operator A 𝑥 ◦ ,2𝜃 (2) to A 𝑥 ◦ ,2𝜃 (1) (in the
latter, the defining amplitude 𝑎˜ 2 is the translation of the amplitude 𝑎 2 in the former);
the germ operator A e (1,2)
𝑥 ◦ , 𝜉 ◦ (defined with the phase-function 𝜑2 ) is then the conjugate
of the analogue defined with 𝜑˜ 2 via the appropriate pullback and pushforward. □
Theorem 22.3.13 points to the most important class of FBI-like transforms of
distributions. For these no reference to the phase-function is needed.
936 22 Germ FBI Transforms

Definition 22.3.14 Let 𝜑 be an FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ) ∈ (Ω ∩ R𝑛 ) × Θ such

that 𝜕𝜑 ◦ ◦ ◦
𝜕𝑧 (𝑥 , 𝜃 ) = −𝜉 ∈ R \ {0}. If the symbol 𝑎 (𝑧, 𝜃, 𝜆) ∈ 𝑆 (Ω × Θ) is classical
𝑛

and elliptic we shall refer to (22.2.8) as a germ FBI transform at (𝑥 ◦ , 𝜉 ◦ ).

Theorem 22.3.15 below explains the microlocal significance of Definition 22.3.14.

Let the distribution 𝑢 ∈ D ′ (Ω ∩ R𝑛 ) be arbitrary and 𝑊 𝐹a (𝑢) denote its analytic
wave-front set (Definition 3.5.1). The conic set 𝑊 𝐹a (𝑢) ∩ 𝑇𝑥∗◦ Ω is defined solely
by the germ of 𝑢 at 𝑥 ◦ , u 𝑥 ◦ ; we shall therefore denote it by 𝑊 𝐹a (u 𝑥 ◦ ). To say that
𝑊 𝐹a (u 𝑥 ◦ ) = ∅ is the same as saying that 𝑢 is a C 𝜔 function in a neighborhood of
𝑥 ◦ in R𝑛 (Proposition 3.5.2).

Theorem 22.3.15 Let A 𝑥 ◦ , 𝜃 ◦ be a germ FBI transform at (𝑥 ◦ , 𝜉 ◦ ) and f 𝑥 ◦ ∈ D′𝑥 ◦ be

arbitrary. The following two properties are equivalent: 1) (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (f 𝑥 ◦ ); 2)
A 𝑥 ◦ , 𝜃 ◦ f 𝑥 ◦ = 0.

Proof By Theorem 22.3.13 it suffices to prove the claim when A 𝑥 ◦ , 𝜃 ◦ = F 𝑥 ◦ , 𝜃 ◦

defined by the “integrals” (22.3.12). To say that F 𝑥 ◦ , 𝜃 ◦ f 𝑥 ◦ = 0 is equivalent to saying
that 𝑭 𝑓 (𝜃, 𝜆) ∈ O (−𝜔) (Θ) with 𝑓 ∈ E ′ (𝑈) representing f 𝑥 ◦ , provided 𝑈 and Θ are
sufficiently small. In (22.3.12) replace 𝜃 by | 𝜃𝜃 | and put 𝜆 = |𝜃|: the claim accords
with Definition 3.5.1. □
Part VII
Analytic Pseudodifferential Operators of
Principal Type
Chapter 23
Analytic PDEs of Principal Type. Local
Solvability

The results proved in this chapter rely on many of the concepts and methods in-
troduced so far in the book but they are focused on a special question: determining
whether or not a linear PDE with complex C 𝜔 coefficients and simple real character-
istics admits local distribution solutions. The simple real characteristics requirement
has the effect that all the results of this and the following chapter are solely based on
the properties of the principal symbol. As soon as this requirement is relinquished
the lower order terms make their presence felt; the problems become greatly more
difficult and the possibility of a comprehensive theory is hard to imagine. One of the
aims of the present chapter is to provide guidance for the extension to the PDEs of
principal type with C ∞ coefficients. The latter case is considerably harder to settle
and requires much finer microlocal analysis (see [Beals-Fefferman, 1973], [Beals-
Fefferman, 1974], [Hörmander, 1983, IV], Ch. XXVI). Much of the distaste toward
analysis in the C 𝜔 category comes from the need to prove the convergence of power
series expansions, and the handling of all those factorials. This need will not be felt
here, because previous chapters have provided tools that help to bypass those hurdles.
And it is not unfair to say that analyticity allows for a simplicity and elegance of the
theory somewhat less visible in the smooth case.
Foremost among these tools are analytic pseudodifferential operators (Ch. 17)
and analytic Fourier integral operators (Ch. 18). Actually, the results of this and
the following chapter apply not only to linear differential operators but also to
pseudodifferential operators (with simple real characteristics) that are antipodal,
meaning that 𝑃𝑚 (𝑥, −𝜉) = (−1) 𝑚 𝑃𝑚 (𝑥, 𝜉) if 𝑃𝑚 is the principal symbol of the
operator. The results of the present chapter are purely local and apply only to
distribution solutions. In fact they remain valid for hyperfunction solutions. This will
be established in the following chapter as a byproduct of the study of the propagation
of analytic singularities; this study will enable us to extend the determination of
solvability in hyperfunctions from the local to the semiglobal, meaning to domains
with compact closure in the base C 𝜔 manifold, under the standard hypothesis of the
absence of trapped null bicharacteristics (Definition 23.1.15 and in Ch. 24, Definition
24.8.8).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 939
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_23
940 23 Analytic PDEs of Principal Type. Local Solvability

23.1 Pseudodifferential Operators of Principal Type

23.1.1 Pseudodifferential operators of principal type. Definition

In this chapter we shall deal solely with classical pseudodifferential operators (Defi-
nitions 16.3.6, 17.2.28). In most of the statements and proofs it will suffice to consider
a pseudodifferential operator 𝑃 of order 𝑚 ∈ Z in an open subset Ω of R𝑛 , with total
symbol
∞
∑︁
𝑃 (𝑥, 𝜉) = 𝑃𝑚− 𝑗 (𝑥, 𝜉) , (23.1.1)
𝑗=0

𝑃𝑚− 𝑗 (𝑥, 𝜉) ∈ C ∞ (Ω × (R𝑛 \ {0})) homogeneous of order 𝑚 − 𝑗, meaning that

∀𝜆 > 0, 𝑃𝑚− 𝑗 (𝑥, 𝜆𝜉) = 𝜆 𝑚− 𝑗 𝑃𝑚− 𝑗 (𝑥, 𝜉)

whatever (𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}).

We recall some standard notation and terminology:
A subset of 𝑇 ∗ Ω\0 Ω × (R𝑛 \ {0}) is conic if it is invariant under the fiber
dilations (𝑥, 𝜉) ↦→ (𝑥, 𝜆𝜉), 𝜆 > 0;
𝑃𝑚 (𝑥, 𝜉) is the principal symbol of 𝑃; 𝑃𝑚 is a C ∞ function in 𝑇 ∗ Ω\0 (Subsection
17.2.5);
the null set of 𝑃𝑚 in 𝑇 ∗ Ω\0, Char 𝑃, is the characteristic set of 𝑃 (ibid.);
the operator 𝑃 (𝑥, D) is elliptic if Char 𝑃 = ∅;
the tautological one-form is the one-form 𝜎 = 𝜉 · d𝑥 = 𝑛𝑗=1Í𝜉 𝑗 d𝑥 𝑗 on 𝑇 ∗ Ω
Í
(Subsection 19.3.3); d𝜎 is the fundamental symplectic two-form 𝑛𝑗=1 d𝜉 𝑗 ∧ d𝑥 𝑗
(Definition 13.3.1);
a C ∞ change of variables 𝑦 = 𝑦 (𝑥, 𝜉), 𝜂 = 𝜂 (𝑥, 𝜉) is symplectic if it preserves
Í Í
the fundamental two-form, i.e., 𝑛𝑗=1 d𝑦 𝑗 ∧ d𝜂 𝑗 = 𝑛𝑗=1 d𝑥 𝑗 ∧ d𝜉 𝑗 . We shall say that
it is homogeneous if

∀𝜆 > 0, 𝑦 (𝑥, 𝜆𝜉) = 𝑦 (𝑥, 𝜉) , 𝜂 (𝑥, 𝜆𝜉) = 𝜆𝜂 (𝑥, 𝜉) . (23.1.2)

The generalization of the above and of most forthcoming results to a generic C ∞

manifold M is straightforward, through the use of local coordinates. The following
statement will be useful.

Lemma 23.1.1 For a homogeneous, C ∞ change of variables in a conic open subset

U of 𝑇 ∗ M\0 to be symplectic it is necessary and sufficient that it preserve the
tautological form 𝜎.

Proof The statement is microlocal and we can use local coordinates 𝑥 𝑖 , 𝜉 𝑗 . If 𝜂 · d𝑦 =

𝜉 · d𝑥 then necessarily the change of variables preserves the fundamental two-form.
Now suppose the latter is true; we have d (𝜉 · d𝑥 − 𝜂 · d𝑦) = 0. We derive that there
exists a smooth function 𝜒 (𝑥, 𝜉) such that 𝜉 · d𝑥 − 𝜂 · d𝑦 = d𝜒, which implies
23.1 Pseudodifferential Operators of Principal Type 941

∑︁ 𝜕𝑦 𝑗 𝑛
𝜕𝜒
= 𝜉𝑖 − 𝜂𝑗 , (23.1.3)
𝜕𝑥𝑖 𝑗=1
𝜕𝑥𝑖
𝑛
∑︁ 𝜕𝑦 𝑗
𝜕𝜒
=− 𝜂𝑗 . (23.1.4)
𝜕𝜉𝑖 𝑗=1
𝜕𝜉𝑖

Since 𝑦 (𝑥, 𝜉) is homogeneous of degree zero the Euler homogeneity identities imply
Í𝑛 𝜕𝑦 𝑗 Í𝑛 𝜕𝜒
𝑖=1 𝜉𝑖 𝜕 𝜉𝑖 = 0, hence 𝑖=1 𝜉𝑖 𝜕 𝜉𝑖 = 0 by (23.1.4): 𝜒 is also homogeneous of degree
zero. But then the left-hand side in (23.1.3) [resp., (23.1.4)] is homogeneous of degree
0 (resp., −1) whereas the right-hand side is homogeneous of degree +1 (resp., 0) –
unless both sides vanish identically. We conclude that d𝜒 = 0. □
In dealing with the single function 𝑃𝑚 we introduce a definition more general
than Definition 13.4.1; the connection will soon be made clear (Proposition 23.1.8
below).
Definition 23.1.2 We say that a classical pseudodifferential operator 𝑃 in the man-
ifold M (as well as its principal symbol 𝑃𝑚 = 𝐴 + 𝑖𝐵 in 𝑇 ∗ M\0) is said to be of
principal type at a point ℘ ∈ Char 𝑃 if d𝑃𝑚 ∧ 𝜎 ≠ 0 at ℘. We shall say that 𝑃 or
𝐴 + 𝑖𝐵 is of principal type if it is of principal type at every point of Char 𝑃.
Definition 23.1.2 can be reformulated in terms of the Hamiltonian vector field
(Definition 13.3.4) of the principal symbol of the operator 𝑃, in local coordinates
𝑛
∑︁ 𝜕𝑃𝑚 𝜕 𝜕𝑃𝑚 𝜕
𝐻 𝑃𝑚 = − , (23.1.5)
𝑗=1
𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝜉 𝑗

and of the “radial” vector field in the fibers of 𝑇 ∗ M,

𝑛
∑︁ 𝜕
𝜌= 𝜉𝑗 ; (23.1.6)
𝑗=1
𝜕𝜉 𝑗

𝜌, like the tautological form 𝜎, is invariant under C ∞ changes of coordinates in the

base.
Proposition 23.1.3 For 𝑃 to be of principal type at ℘ ∈ Char 𝑃 it is necessary and
sufficient that 𝜌 ∧ 𝐻 𝑃𝑚 ≠ 0 at ℘.
Proof Let 𝜛e℘ : 𝑇℘ M −→ 𝑇℘∗ M be the canonical isomorphism defined by the

symplectic form d𝜉 ∧ d𝑥 [see (13.1.1)–(13.1.2)]. Recall that 𝜛
e℘ 𝐻 𝑓 = −d 𝑓 if
𝑓 ∈ C ∞ (M); it follows that 𝜛e℘ (𝜌) = 𝜎 whence the equivalence of 𝜎 ∧ d𝑃𝑚 ≠ 0
and 𝜌 ∧ 𝐻 𝑃𝑚 ≠ 0 (at a point). □
For 𝑃 to be of principal type at ℘ ∈ Char 𝑃 it is necessary and sufficient that at
least one of the two real pseudodifferential operators 𝐴 = Re 𝑃 or 𝐵 = Im 𝑃 be of
principal type at ℘, simply because 𝜎 is a real one-form. If 𝑃 is elliptic then 𝑃 is
of principal type, since Char 𝑃 = ∅. When 𝑛 = 1 we have 𝑃𝑚 (𝑥, 𝜉) = 𝑎 (𝑥, 𝜉) 𝜉 𝑚
942 23 Analytic PDEs of Principal Type. Local Solvability

𝑎 (𝑥,
with 𝜉) = 𝑎 (𝑥, ±1) if 𝜉 ≠ 0. If 𝜉 ≠ 0 we have 𝜕 𝜉 𝑃𝑚 (𝑥, 𝜉) ≠ 0 if and only if
𝜉
𝑎 𝑥, | 𝜉 | ≠ 0: 𝑃 is of principal type at (𝑥, 𝜉) if and only if it is elliptic at that point.
In the sequel we shall assume 𝑛 ≥ 2.
The most basic examples of differential operators of principal type are the (com-
plex, smooth) vector fields that do not vanish at any point. Elliptic pseudodifferential
operators are of principal type; so are hyperbolic differential operators.

Example 23.1.4 The Tricomi differential operator in R𝑛 ,

𝑛−1
∑︁
𝑃 (𝑥, D) = D2𝑛 − 𝑥 𝑛 D2𝑗 ,
𝑗=1

is of principal type. It is evident when 𝑥 𝑛 ≠ 0 since then d 𝜉 𝑃 ≠ 0. When 𝑥 𝑛 = 𝜉 𝑛 = 0

(and |𝜉 | ≠ 0) then d 𝜉 𝑃 = 0 but 𝜎 ∧ d𝑃 = |𝜉 | 2 𝑛−1
Í
𝑗=1 𝜉 𝑗 d𝑥 𝑗 ∧ d𝑥 𝑛 ≠ 0.

The principal symbol 𝑃𝑚 might vanish identically in the whole cotangent space
at a point and 𝑃 (𝑥, D) still be of principal type, as exemplified by the rotation vector
field in the plane:

Example 23.1.5 Suppose 𝑛 = 2 and 𝑃𝑚 (𝑥, 𝜉) = 𝑥2 𝜉1 − 𝑥 1 𝜉2 (thus 𝑚 = 1); we have

d 𝑥 𝑃𝑚 = −𝜉2 d𝑥1 + 𝜉1 d𝑥2 and

d𝑃𝑚 ∧ 𝜎| 𝑥=0 = − |𝜉 | 2 d𝑥1 ∧ d𝑥2 .

23.1.2 Microlocal normal forms

Let 𝑃 (𝑥, D) be a classical pseudodifferential operator of order 𝑚 (and of class

C ∞ ) in a domain Ω in R𝑛 ; we are going to reason microlocally, here meaning that
(𝑥, 𝜉) ∈ R2𝑛 will be confined to a conic neighborhood U = 𝑈 × Γ of a point
(𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 in 𝑇 ∗ Ω\0, with 𝑈 ⊂⊂ Ω an open set containing the “central point”
𝑥 ◦ and Γ ⊂ R𝑛 \ {0} a convex open cone containing 𝜉 ◦ .

Proposition 23.1.6 If d 𝜉 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) ≠ 0 then 𝑃 (𝑥, D) is of principal type at (𝑥 ◦ , 𝜉 ◦ ).

Proof Indeed, 𝜉 ≠ 0 =⇒ 𝜎 ∧ d𝜉 𝑗 ≠ 0. □
To say that d 𝜉 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) ≠ 0 is equivalent to saying that the base projection
(𝑥, 𝜉) ↦→ 𝑥 of the Hamiltonian vector field of 𝑃𝑚 does not vanish at (𝑥 ◦ , 𝜉 ◦ ).

Remark 23.1.7 When 𝑚 > 0 the Euler homogeneity identities imply 𝑃𝑚 (𝑥, 𝜉) =
1 Í𝑛 𝜕𝑃𝑚 ◦ ◦ ◦ ◦
𝑚 𝑗=1 𝑗 𝜕 𝜉 𝑗 (𝑥, 𝜉) and thus d 𝜉 𝑃 𝑚 (𝑥 , 𝜉 ) = 0 implies (𝑥 , 𝜉 ) ∈ Char 𝑃.
𝜉
23.1 Pseudodifferential Operators of Principal Type 943

Proposition 23.1.8 If 𝑃 (𝑥, D) is of principal type at ℘ ∈ Char 𝑃 and if d 𝜉 𝑃𝑚 = 0

at ℘ there is a homogeneous, symplectic, C ∞ change of coordinates

(𝑥1 , ..., 𝑥 𝑛 , 𝜉1 , ..., 𝜉 𝑛 ) ⇝ (𝑦 1 , ..., 𝑦 𝑛 , 𝜂1 , ..., 𝜂 𝑛 )

in a conic neighborhood U of ℘ such that d 𝜂 𝑃𝑚 ≠ 0 at ℘.

A rephrasing of Proposition 23.1.8 is that if 𝑃 is of principal type at ℘ ∈ Char 𝑃

there is a homogeneous, symplectic, smooth change of the local coordinates in a conic
neighborhood of ℘ in the cotangent bundle ensuring that, in the new coordinates, the
Hamiltonian tangent vector 𝐻 𝑃𝑚 ℘ is not tangent to 𝑇℘∗ Ω, i.e., the base projection of
𝐻 𝑃𝑚 ℘ is not equal to zero.
Proof An affine change of coordinates in R𝑛 allows us to take ℘ = (0, 𝜉 ◦ ), 𝜉 ◦ =
(1, 0, ..., 0); we can also assume that 𝑚 = 0 after substituting |𝜉 | −𝑚 𝑃𝑚 (𝑥, 𝜉) for
𝑃𝑚 (𝑥, 𝜉). The hypothesis is that 𝑃𝑚 (0, 𝜉 ◦ ) = 0, d 𝜉 𝑃𝑚 (0, 𝜉 ◦ ) = 0 and 𝜎| (0, 𝜉 ◦ ) ∧
d 𝑥 𝑃𝑚 (0, 𝜉 ◦ ) ≠ 0. Since 𝜎| (0, 𝜉 ◦ ) = d𝑥1 the latter means that 𝜕𝑃 ◦
𝜕𝑥 𝑗 (0, 𝜉 ) ≠ 0
𝑚

for some 𝑗 ≥ 2. After a linear change of the coordinates 𝑥 𝑗 , 𝑗 = 2, ..., 𝑛, and

multiplication of 𝑃𝑚 by a complex number 𝜁 ≠ 0 we can assume that the linear part
of Re 𝑃𝑚 (𝑥, 𝜉 ◦ ) at 0 is equal to 𝑥2 . We seek a symplectic, homogeneous change of
the coordinates (𝑥1 , 𝑥2 , 𝜉1 ,𝜉2 ) ↦→ (𝑦 1 , 𝑦 2 , 𝜂1 ,𝜂2 ) such that d 𝜂 𝑥 2 ≠ 0. In what follows
we assume 𝑥 𝑗 = 𝑦 𝑗 , 𝜉 𝑗 = 𝜂 𝑗 for all 𝑗 = 3, ..., 𝑛. If 𝑓 (𝜂1 , 𝜂2 ) is homogeneous of
degree 1 the transformation in R2 × R2 \ {0} ,

𝜕𝑓 𝜕𝑓
𝑥1 = 𝑦 2 + (𝜂1 , 𝜂2 ) , 𝑥2 = 𝑦 1 + (𝜂1 , 𝜂2 ) ,
𝜕𝜂2 𝜕𝜂1
𝜉1 = 𝜂2 , 𝜉2 = 𝜂1 ,

is homogeneous; it is symplectic because, by the Euler homogeneity identities,

𝜕𝑓 𝜕𝑓 𝜕𝑓 𝜕𝑓
𝜂1 d + 𝜂2 d = d 𝜂1 + 𝜂2 −d𝑓 = 0
𝜕𝜂1 𝜕𝜂2 𝜕𝜂1 𝜕𝜂2

and therefore 𝜉 · d𝑥 = 𝜂 · d𝑦. The point 𝑥 = 0, 𝜉 ◦ = (1, 0, ..., 0), is transformed into
the point 𝑦 = 0, 𝜂◦ = (0, 1, 0, ..., 0) , provided

𝜕𝑓 𝜕𝑓
(0, 1) = (0, 1) = 0.
𝜕𝜂1 𝜕𝜂2

One can take, for instance, 𝑓 (𝜂) = |𝜂| − 𝜂2 . □

Example 23.1.9 Suppose 𝑛 = 2, 𝑃1 (𝑥, 𝜉) = 𝑥2 𝜉1 − 𝑥1 𝜉2 and 𝑥 ◦ = 0, 𝜉 ◦ = (0, 1). If

we carry out the symplectic change of variables in R2 × R2 \ {0} ,
𝜂2 𝜂1
𝑥1 = 𝑦 2 + − 1, 𝑥2 = 𝑦 1 + , 𝜉1 = 𝜂2 , 𝜉2 = 𝜂1 ,
|𝜂| |𝜂|
944 23 Analytic PDEs of Principal Type. Local Solvability

we obtain
𝑃1 (𝑥, 𝜉) = 𝜂1 + 𝑦 1 𝜂2 − 𝜂1 𝑦 2 ,
whence d 𝜂 𝑃1 (𝑥 (𝑦, 𝜂) , 𝜂2 , 𝜂1 ) 𝑦=0
= d𝜂1 .

In the remainder of this subsection we require 𝑃 (𝑥, D) to be a classical ana-

lytic pseudodifferential operator in Ω (Definition 17.2.31); in particular, 𝑃𝑚 (𝑥, 𝜉) ∈
C 𝜔 (𝑇 ∗ Ω\0). In the sequel we will frequently deal with classical analytic pseudod-
ifferential operators defined in a conic open subset U of 𝑇 ∗ Ω\0 (or of 𝑇 ∗ M\0 with
M a C 𝜔 manifold). The precise definitions and properties have been given in Ch.
17, solely for symbols in 𝑇 ∗ Ω\0; but as far as the symbols are concerned all remains
valid if we substitute U for 𝑇 ∗ Ω\0; to define the corresponding operators in U one
can use the cutoffs 𝑔 𝑅 (D) (Subsection 17.3.1).
Let 𝑇◦ : (𝑥, 𝜉) ↦→ (𝑦, 𝜂) denote a symplectomorphism of U onto a conic open
set V ⊂R𝑛 × (R𝑛 \ {0}). We can associate to the graph of 𝑇◦ | U in U × V a unitary
Fourier integral operator T◦ (Definition 18.5.27). Egorov’s Theorem 18.5.31(more
accurately, its analytic version; cf. Theorem 21.2.12), implies that T◦ 𝑃T◦∗ is a clas-
sical analytic pseudodifferential operator of order 𝑚 in V.
Proposition 23.1.10 Let 𝑃𝑚 (𝑥, 𝜉) ∈ C 𝜔 (𝑇 ∗ Ω\0) be such that d 𝜉 𝑃𝑚 ≠ 0 at a point
(𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑇 ∗ Ω\0. Possibly after a linear change of the coordinates in R𝑛 there are
two analytic functions in a conic neighborhood U of (𝑥 ◦ , 𝜉 ◦ ), 𝐸 ◦ (𝑥, 𝜉) homogeneous
of degree 𝑚 − 1 and nowhere vanishing in U, 𝑞 (𝑥, 𝜉 ′) homogeneous of degree 1
[where 𝜉 ′ = (𝜉1 , ..., 𝜉 𝑛−1 )], such that 𝑃𝑚 (𝑥, 𝜉) = 𝐸 ◦ (𝑥, 𝜉) (𝜉 𝑛 + 𝑞 (𝑥, 𝜉 ′)) in U.
Proof After a linear change of variables we may assume that 𝜕𝑃 ◦ ◦
𝜕 𝜉𝑛 (𝑥 , 𝜉 ) ≠ 0; then
𝑚

the claim follows directly from the complex-analytic Implicit Function Theorem (or
the Weierstrass Preparation Theorem 14.3.4). □
Next we apply the construction ∫in𝑥Subsection 18.3.3, and
Remark 18.3.9: we form
the unitary operator T = exp −𝑖 0 𝑛 Re 𝑞 (𝑥 ′, 𝑡, D 𝑥′ ) d𝑡 where 𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛−1 )
and D 𝑥′ = D 𝑥1 , ..., D 𝑥𝑛−1 , D 𝑥 𝑗 = √1 𝜕𝑥𝜕 𝑗 ; we have

−1

T ∗ D 𝑥𝑛 + Re 𝑞 (𝑥, D 𝑥′ ) T = D 𝑥𝑛

(T ∗ is the adjoint, as well as the inverse, of T ). We apply again the analytic Egorov
Theorem: T ∗ Im 𝑞 (𝑥 ′, 𝑡, D 𝑥′ ) T is a classical analytic pseudodifferential operator of
order 1 in the conic neighborhood U of (𝑥 ◦ , 𝜉 ◦ ) with principal symbol 𝐵 (𝑥, 𝜉 ′) =
Im 𝑞 (𝑥, 𝜉 ′). Thus, in U, the analysis has been reduced to the case 𝑚 = 1 and to the
normal form
𝑃1 (𝑥, 𝜉) = 𝜉 𝑛 + 𝑖𝐵 (𝑥, 𝜉 ′) . (23.1.7)
Remark 23.1.11 Proposition 23.1.10 remains valid in the C ∞ category as seen by
applying the Weierstrass–Malgrange Preparation Theorem (for a proof see [Niren-
berg, 1971]).
It is often convenient to factorize microlocally (under the hypothesis in Proposition
23.1.10) the total symbol of 𝑃, not solely its principal symbol.
23.1 Pseudodifferential Operators of Principal Type 945

Proposition 23.1.12 Let 𝑃 (𝑥, D) be a classical analytic pseudodifferential operator

of order 𝑚 in Ω such that d 𝜉 𝑃𝑚 ≠ 0 at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑇 ∗ Ω\0. Possibly after a
linear change of the coordinates in R𝑛 there exists a unique pair of classical analytic
symbols in a conic neighborhood U of (𝑥 ◦ , 𝜉 ◦ ), 𝐸 (𝑥, 𝜉) elliptic of order 𝑚 − 1 and
𝑄 (𝑥, 𝜉 ′) independent of 𝜉 𝑛 of order 1, such that 𝑃 (𝑥, 𝜉) = 𝐸 (𝑥, 𝜉) # (𝜉 𝑛 + 𝑄 (𝑥, 𝜉 ′))
in U.

By # we mean the composition law (16.2.16).

Proof We extend the analysis to the complex domain: ΩC is a domain in C𝑛 such
that ΩC ∩ R𝑛 = Ω, ΓC is a convex open cone in C𝑛 \ {0} such that 𝜉 ◦ ∈ ΓC , and each
homogeneous term 𝑃 𝑗 in the total symbol 𝑃 = ∞
Í
𝑗=0 𝑃 𝑚− 𝑗 extends as a holomorphic
function of (𝑧, 𝜁) to Ω ∩ Γ homogeneous of degree 𝑚 − 𝑗 (with respect to 𝜁) such
C C

that, given any compact subset K of ΩC × ΓC ∩ S2𝑛−1 there is a 𝐶 K > 0 such that
𝑗+1
∀ 𝑗 ∈ Z+ , max 𝑃 𝑗 (𝑧, 𝜁) ≤ 𝐶 K 𝑗!. (23.1.8)
(𝑧,𝜁 ) ∈K

The factorization in Proposition 23.1.10 extends holomorphically to a conic

neighborhood U C of (𝑥 ◦ , 𝜉 ◦ ) in ΩC ∩ ΓC : 𝑃𝑚 (𝑧, 𝜁) = 𝐸 ◦ (𝑧, 𝜁) (𝜁 𝑛 + 𝑞 (𝑧, 𝜁 ′))
with 𝐸 ◦ (𝑧, 𝜁), 𝑞 (𝑧, 𝜁 ′) ∈ O U C . Let us write
∞
∑︁ ∞
∑︁
𝐸= 𝐸 𝑚−1− 𝑗 , 𝑄 = 𝑞 1− 𝑗 ,
𝑗=0 𝑗=0

with 𝐸 𝑚−1− 𝑗 (𝑧, 𝜁) ∈ O U C homogeneous of degree 𝑚 − 1 − 𝑗 and 𝑞 1− 𝑗 =
𝑞 1− 𝑗 (𝑧, 𝜁 ′) ∈ O U C homogeneous of degree 1 − 𝑗 and independent of 𝜁 𝑛 . We
have
∞
∑︁ ∞ ∑︁
∑︁ ∑︁|−1 ℓ−∑︁
ℓ−|𝛾 𝑗− |𝛾 |
1 𝛾
𝐸 𝑚−1−ℓ (𝜁 𝑛 + 𝑞 (𝑧, 𝜁 ′))+
𝛾
𝐸#𝑄 = D 𝜁 𝐸 𝑚−1− 𝑗 𝜕𝑧 𝑞 1−𝑘 .
ℓ=0 ℓ=0 𝛾 ∈Z+𝑛 𝑗=0 𝑘=0
𝛾!
|𝛾 | ≤ℓ

We seek to determine recursively the homogeneous functions 𝐸 𝑚−1− 𝑗 and 𝑞 1−𝑘 as

solutions of the equations
|𝛾 |−1
ℓ−∑︁
∑︁ 1 𝛾
𝛾
𝐸 𝑚−1−ℓ (𝜁 𝑛 + 𝑞 1 ) + D 𝜁 𝐸 𝑚−1− 𝑗 𝜕𝑧 𝑞 1+ 𝑗+|𝛾 |−ℓ = 𝑃𝑚−ℓ
𝛾 ∈Z+𝑛 𝑗=0
𝛾!
|𝛾 | ≤ℓ

(ℓ ∈ Z+ ). Obviously, we choose 𝐸 𝑚−1 = 𝐸 ◦ and 𝑞 1 = 𝑞 (𝑧, 𝜁 ′). For ℓ ≥ 1 we assume

that the function
946 23 Analytic PDEs of Principal Type. Local Solvability

ℓ−1 ∑︁
∑︁ 1 𝛾
𝛾
𝑅𝑚−ℓ = 𝑃𝑚−ℓ − D 𝜁 𝐸 𝑚−1− 𝑗 𝜕𝑧 𝑞 1+ 𝑗+|𝛾 |−ℓ
𝑗=0 |𝛾 | ≤ℓ− 𝑗
𝛾!
∑︁ 1
𝛾 𝛾
− D 𝜁 𝐸 ◦ 𝜕𝑧 𝑞 1+|𝛾 |−ℓ
𝛾!
0< |𝛾 | ≤ℓ

is known (since the intervening 𝐸 𝑚−1− 𝑗 and 𝑞 1−𝑘 have been determined by the in-
duction hypothesis). The Weierstrass Division Theorem
14.3.5 implies the existence
and uniqueness of the functions 𝐸 𝑚−1−ℓ ∈ O U C and 𝑞 1−ℓ ∈ O U C such that

𝐸 ◦−1 𝑅𝑚−ℓ = 𝐸 𝑚−1−ℓ (𝜁 𝑛 + 𝑞) + 𝑞 1−ℓ

in a neighborhood of (𝑧◦ , 𝜁 ◦ ) in Ω × S𝑛−1 . To complete the proof one must prove

that 𝐸 𝑚−1−ℓ and 𝑞 1−ℓ satisfy inequalities of the kind (23.1.8). This follows from the
estimates deriving from explicit formulas in the division by 𝜁 𝑛 + 𝑞 (see the proof of
Theorem 14.3.5) and the estimates of 𝐸 ◦−1 𝑅𝑚−ℓ provided by the induction hypothesis.
We leave the details as an exercise. □

23.1.3 Null bicharacteristic curves, surfaces

We return to the general case of a C 𝜔 manifold M and a conic open subset U of

𝑇 ∗ M\0. We deal with two real-valued functions 𝐴, 𝐵 ∈ C 𝜔 (U), both homogeneous
of degree 𝑚, and we make use of the Nagano foliation in U defined by the Hamil-
tonian vector fields 𝐻 𝐴 and 𝐻 𝐵 (Subsection 12.2.2); these generate a Lie algebra of
vector fields in U with respect to the commutation bracket, henceforth denoted by
𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ). The Nagano Theorem 12.4.2 allows us to state: There is a foliation
of U consisting of integral manifolds (Definition 12.2.1) of 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ), denoted
in the sequel by 𝔑𝔞𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) and whose elements shall be referred to as the
Nagano leaves of (𝐻 𝐴, 𝐻 𝐵 ) or of 𝐴 + 𝑖𝐵 in U.
Recall that a smooth function 𝑓 is constant on each integral curve of its Hamilto-
nian field 𝐻 𝑓 .

Proposition 23.1.13 If a one-dimensional Nagano leaf 𝔠 of (𝐻 𝐴, 𝐻 𝐵 ) in U intersects

Char ( 𝐴 + 𝑖𝐵) then 𝔠 ⊂ Char ( 𝐴 + 𝑖𝐵).

Proof Suppose ℘◦ ∈ 𝔠 ∩ Char ( 𝐴 + 𝑖𝐵). We exploit the fact that 𝔠 is smooth and
connected (analyticity is not required although it would slightly simplify the argu-
ment). Let 𝛾 be an open subset of 𝔠 in which 𝐻 𝐴 ≡ 0, which is equivalent to d𝐴 ≡ 0;
𝐴 must be locally constant in 𝛾. If 𝐻 𝐴 never vanishes in 𝛾, the fact that 𝐻 𝐴 𝐴 ≡ 0
everywhere implies that 𝐴 is locally constant in 𝛾. We conclude that 𝐴 is locally
constant in an open and dense subset of 𝔠. Since 𝐴 is smooth 𝐴 must be constant in
𝔠; then 𝐴 (℘◦ ) = 0 implies 𝐴 ≡ 0 in 𝔠. The same reasoning applies to 𝐵. □
23.1 Pseudodifferential Operators of Principal Type 947

Remark 23.1.14 Proposition 23.1.13 does not require 𝐴 + 𝑖𝐵 to be of principal type

in a neighborhood of the curve 𝔠 but this will be automatically true if the radial vector
field (23.1.6) is nowhere tangent to 𝔠.

Definition 23.1.15 Let 𝐴, 𝐵 ∈ C 𝜔 (U) be real-valued and homogenous of degree

𝑚. By a null bicharacteristic leaf of 𝐴 + 𝑖𝐵 (or of any classical pseudodifferential
operator 𝑃 in M with principal symbol 𝑃𝑚 = 𝐴 + 𝑖𝐵) we shall mean a Nagano
leaf L of (𝐻 𝐴, 𝐻 𝐵 ) in U contained in Char ( 𝐴 + 𝑖𝐵). We shall say that L is a null
bicharacteristic curve (resp., surface) if dim L = 1 (resp., 2).

Keep in mind that two distinct Nagano leaves do not intersect.

Example 23.1.16 The null bicharacteristic surfaces in 𝑇 ∗ R2 \0 of the vector field

D2 + 𝑖𝑥1 D1 (with symbol 𝜉2 + 𝑖𝑥1 𝜉1 ) are the half-planes 𝑥1 = 𝜉2 = 0, 𝜉1 ≷ 0.

The invariance of Definition 23.1.15 under multiplication by an “elliptic” symbol

as stated next is important in the forthcoming applications. Here we limit ourselves
to the case of one- or two-dimensional leaves.

Proposition 23.1.17 Let 𝐴 + 𝑖𝐵 be of principal type in U and L be a Nagano leaf

of (𝐻 𝐴, 𝐻 𝐵 ) in U. If L ⊂ Char ( 𝐴 + 𝑖𝐵) and dim L ≤ 2 then L is also a null
bicharacteristic leaf of ℎ ( 𝐴 + 𝑖𝐵) in U, with ℎ = 𝑓 + 𝑖𝑔 ∈ C 𝜔 (U) homogeneous
and nowhere vanishing.

Proof It suffices to reason microlocally, in the neighborhood of (𝑥 ◦ , 𝜉 ◦ ) ∈ L in

U = 𝑈 × Γ, where 𝑈 is a domain in R𝑛 and Γ ⊂ R𝑛 \ {0} a convex open cone. We
assume 𝐴 = 𝜉 𝑛 , 𝐵 = 𝐵 (𝑥, 𝜉 ′) homogeneous of degree 1 (cf. Proposition 23.1.10).
Let L ∈ 𝔑𝔞𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) and L ⊂ Char ( 𝐴 + 𝑖𝐵). We have

𝑋 𝑘 = (Ad 𝐻 𝐴) 𝑘 𝐻 𝐵 = 𝐻𝜕𝑘 𝐵/𝜕𝑥𝑛𝑘

𝑛−1 𝑛
∑︁ 𝜕 𝑘+1 𝐵 𝜕 ∑︁ 𝜕 𝑘+1 𝐵 𝜕
= − .
𝑗=1
𝜕𝜉 𝑗 𝜕𝑥 𝑛𝑘 𝜕𝑥 𝑗 𝑗=1 𝜕𝑥 𝑗 𝜕𝑥 𝑛𝑘 𝜕𝜉 𝑗

An important property, used below, is that the vector fields 𝐻 𝐴 and 𝑋 𝑘 are tangent
to L, implying
∀𝑘 ∈ Z+ , 𝐻 𝐴𝑘 𝐵 ≡ 0, 𝑋 𝑘 𝐴 ≡ 𝑋 𝑘 𝐵 ≡ 0 in L. (23.1.9)
Let (𝑥 ◦ , 𝜉 ◦ ) ∈ L be arbitrary. First consider the case where 𝑋 𝑘 = 0 at (𝑥 ◦ , 𝜉 ◦ ) for
all 𝑘 ∈ Z+ . This means that 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) is spanned at (𝑥 ◦ , 𝜉 ◦ ) by 𝐻 𝐴, implying
dim L = 1 and that 𝐻 𝐴 spans 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) in the whole of L. Then the invariance
under multiplication by ℎ is obvious.
Now suppose there is a least integer 𝑘 ∈ Z+ such that 𝐻 𝐴 = 𝜕𝑥𝜕𝑛 and 𝑋 𝑘 span
𝑇( 𝑥 ◦ , 𝜉 ◦ ) L; let us call it 𝑘 ◦ . Note, for use below that 𝑋 𝑘 | ( 𝑥 ◦ , 𝜉 ◦ ) = 0 for all 𝑘 < 𝑘 ◦ .
Let V ⊂ U be a conic neighborhood of (𝑥 ◦ , 𝜉 ◦ ) so that 𝑋 𝑘◦ ≠ 0 at every point
of L ∩ V; this requires dim L = 2. Since 𝐵 is independent of 𝜉 𝑛 the Hamiltonian
fields 𝑋 𝑘 must be collinear to 𝑋 𝑘◦ for all 𝑘 ∈ Z+ : in V we must have 𝑋 𝑘 = 𝛾 𝑘 𝑋 𝑘◦ ,
𝛾 𝑘 ∈ C 𝜔 (L ∩ V; R). We have
948 23 Analytic PDEs of Principal Type. Local Solvability

[𝐻 𝐴, 𝑋 𝑘 ] = 𝑋 𝑘+1 = 𝛾 𝑘+1 𝑋 𝑘◦ (23.1.10)

𝑋 𝑘◦ , 𝑋 𝑘 = 𝑋 𝑘◦ 𝛾 𝑘 𝑋 𝑘◦ .

We see that 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) is spanned by 𝐻 𝐴 and 𝑋 𝑘◦ at every point of L ∩ V.

Let now 𝑓 + 𝑖𝑔 be as in the statement; obviously, Char (ℎ ( 𝐴 + 𝑖𝐵)) = Char ( 𝐴 + 𝑖𝐵)
since ℎ never vanishes in U. We have

𝐻 𝑓 𝐴−𝑔𝐵 = 𝑓 𝐻 𝐴 − 𝑔𝐻 𝐵 + 𝜉 𝑛 𝐻 𝑓 − 𝐵𝐻𝑔
= 𝑓 𝐻 𝐴 − 𝑔𝛾0 𝑋 𝑘◦ + 𝜉 𝑛 𝐻 𝑓 − 𝐵𝐻𝑔 ,
𝐻𝑔 𝐴+ 𝑓 𝐵 = 𝑔𝐻 𝐴 + 𝑓 𝐻 𝐵 + 𝜉 𝑛 𝐻𝑔 + 𝐵𝐻 𝑓 .

It is an easy exercise to derive from (23.1.9) and (23.1.10) that

𝔤 U, 𝐻 𝑓 𝐴−𝑔𝐵 , 𝐻𝑔 𝐴+ 𝑓 𝐵 ⊂ 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 )

at every point of L ∩ V. If we assume that | 𝑓 + 𝑖𝑔| does not vanish we conclude

that the converse inclusion is true given that the linear transformation ( 𝐴, 𝐵) ↦→
( 𝑓 𝐴 − 𝑔𝐵, 𝑔 𝐴 + 𝑓 𝐵) is invertible. □
The following statement is self-evident.

Proposition 23.1.18 Let U be a conic open subset of 𝑇 ∗ M\0 and 𝐴 + 𝑖𝐵 ∈ C 𝜔 (U)

be homogeneous. If 𝐴 + 𝑖𝐵 is of principal type in U then the Hamiltonian vector
fields 𝐻 𝐴 and 𝐻 𝐵 define an orientation on each null bicharacteristic curve (Definition
23.1.15) of 𝐴 + 𝑖𝐵 in U.

It is possible for 𝐴+𝑖𝐵 not to have any null bicharacteristic leaf, as in the following

Example 23.1.19 Consider the Mizohata vector fields in the plane, 𝐿 𝑘 = 𝜕𝑥𝜕 2 +
𝑖𝑥 2𝑘 𝜕𝑥𝜕 1 (1 ≤ 𝑘 ∈ Z+ ; see [Mizohata, 1962]); if we denote by 𝐴 + 𝑖𝐵 the symbol of
−𝑖𝐿 𝑘 then 𝐴 = 𝜉2 , 𝐵 = 𝑥 2𝑘 𝜉1 and

𝜕 𝜕 𝜕
𝐻𝐴 = , 𝐻 𝐵 = 𝑥2𝑘 − 𝑘𝜉1 𝑥2𝑘−1 .
𝜕𝑥2 𝜕𝑥1 𝜕𝜉2
We have
Char 𝐿 𝑘 = (𝑥, 𝜉) ∈ R4 ; 𝑥 2 = 𝜉2 = 0, 𝜉1 ≠ 0

and the Lie algebra generated (with respect to the commutation bracket) by 𝐻 𝐴, 𝐻 𝐵
is
𝜕 𝜕 𝜕
𝔤 (𝐻 𝐴, 𝐻 𝐵 ) = span , , 𝜉1 .
𝜕𝑥1 𝜕𝑥2 𝜕𝜉2
The Nagano leaves of (𝐻 𝐴, 𝐻 𝐵 ) in 𝑇 ∗ R2 \ {0} are the affine hyperplanes

(𝑥, 𝜉) ∈ R4 ; 𝜉1 = 𝜉1◦ ≠ 0

and the affine planes

23.1 Pseudodifferential Operators of Principal Type 949

(𝑥, 𝜉) ∈ R4 ; 𝜉1 = 0, 𝜉2 = 𝜉2◦ ≠ 0 .

The former intersect Char 𝐿 𝑘 along the 𝑥1 -lines

(𝑥, 𝜉) ∈ R4 ; 𝑥2 = 𝜉2 = 0, 𝜉1 = 𝜉1◦ ≠ 0 ;

the latter do not intersect Char 𝐿 𝑘 .

There are differential operators with null bicharacteristic curves whose base pro-
jections are single points.
Example 23.1.20 Consider the rotation operator 𝐿 = 𝑥2 𝜕𝑥𝜕 1 − 𝑥1 𝜕𝑥𝜕 2 in R2 (Example
√
23.1.5); up to a factor −1 its Hamiltonian vector field is

𝜕 𝜕 𝜕 𝜕
𝑥2 − 𝑥1 + 𝜉2 − 𝜉1 .
𝜕𝑥1 𝜕𝑥2 𝜕𝜉1 𝜕𝜉2

The circles 𝜉12 + 𝜉22 = 𝜌 2 (𝜌 > 0) in 𝑇0∗ R2 are null bicharacteristic curves of

𝐿. The other null bicharacteristic curves of 𝐿 are defined, in R2 × R2 \ {0} , by the

equations
𝑥12 + 𝑥22 = 𝑟 2 , 𝜉 = 𝑐𝑥 (𝑟 > 0, 𝑐 ≠ 0).
We wish to look at Definition 23.1.15 from a slightly different viewpoint. As
before suppose 𝑃𝑚 = 𝐴 + 𝑖𝐵.
Definition 23.1.21 We shall say that a point ℘ ∈ Char 𝑃 is of finite (resp., infinite)
type if it is of finite (resp., infinite) type for the pair of functions ( 𝐴, 𝐵) (Definition
13.3.28).
Thus ℘ is of finite type for 𝐴 + 𝑖𝐵 if (and only if) at least one of the multibrackets
(13.3.23) in which each 𝑓 𝜄𝑘 = 𝐴 or 𝐵 does not vanish at ℘.
Proposition 23.1.22 For ℘ ∈ U ∩ Char 𝑃 to be of infinite type it is necessary and
sufficient that the Nagano leaf L of (𝐻 𝐴, 𝐻 𝐵 ) in U through ℘ be contained in
Char 𝑃.
Proof The vector fields in 𝔤 (𝐻 𝐴, 𝐻 𝐵 ) span the whole tangent bundle 𝑇 L. To say
that all the multibrackets (13.3.23) in which 𝑓 𝜄𝑘 = 𝐴 or 𝐵 for every 𝜄 𝑘 vanish at ℘ is
the same as saying that the restrictions to L of 𝐴 and 𝐵 vanish to infinite order at ℘,
which is equivalent to L ⊂ Char 𝑃. □
Proposition 23.1.22 yields a partition:

Char 𝑃 = Char∞ 𝑃 ∪ Charfin 𝑃, (23.1.11)

where Charfin 𝑃 and Char∞ 𝑃 are the sets of all points in Char 𝑃 of finite and infinite
type; Char∞ 𝑃 is foliated by the Nagano leaves of (𝐻 𝐴, 𝐻 𝐵 ) contained in Char 𝑃. The
Mizohata vector fields 𝐿 𝑘 (Example 23.1.19) have the property that Char∞ 𝐿 𝑘 = ∅.
If Char 𝑃 is foliated by the null bicharacteristic curves of 𝑃 (as when 𝑃 is real) then
Charfin 𝑃 = ∅.
950 23 Analytic PDEs of Principal Type. Local Solvability

23.1.4 Condition (P) for vector fields

In this subsection we limit our attention to C 𝜔 vector fields in a C 𝜔 manifold M.

First consider a C 𝜔 vector field in an open subset Ω of R𝑛 , 𝐿 = 𝑋 + 𝑖𝑌 , 𝑋 = Re 𝐿,
𝑌 = Im 𝐿:
𝑛 𝑛
∑︁ 𝜕 ∑︁ 𝜕
𝑋= 𝑎 𝑗 (𝑥) ,𝑌 = 𝑏 𝑗 (𝑥) (23.1.12)
𝑗=1
𝜕𝑥 𝑗 𝑗=1
𝜕𝑥 𝑗

with 𝑎 𝑗 , 𝑏 𝑗 ∈ C 𝜔 (Ω; R). It is convenient to assume that 0 ∈ Ω. If 𝐿 ≠ 0 at

every point of Ω then 𝐿 is a differential operator of principal type in Ω. (There are
vector fields of principal type that have critical points, as in√Example 23.1.5.) After
contracting Ω about 0 and, if necessary, multiplying 𝐿 by −1 we can assume that
𝑋 does not vanish at any point of Ω and√ we can label the coordinates in such a way
that 𝑎 𝑛 (𝑥) ≠ 0. Division by 𝑎 𝑛 (𝑥) + −1𝑏 𝑛 (𝑥) and a redefinition of the coefficients
allows us to assume that

𝜕
𝑛−1
∑︁ √ 𝜕
𝐿= + 𝑎 𝑗 (𝑥) + −1𝑏 𝑗 (𝑥) . (23.1.13)
𝜕𝑥 𝑛 𝑗=1 𝜕𝑥 𝑗

We wish to “straighten” the integral curves of 𝑋; to do this we solve the initial value
problem
d𝑥 𝑗
= 𝑎 𝑗 (𝑥 ′, 𝑡) , 𝑥 𝑗 𝑡=0 = 𝑦 𝑗 , 𝑗 = 1, ..., 𝑛 − 1, (23.1.14)
d𝑡
[where 𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛−1 )] assuming 𝑦 𝑗 < 𝜀, 𝑗 = 1, ..., 𝑛 − 1, 𝜀 > 0 suitably
small. Let 𝑥 𝑗 (𝑦, 𝑡), 𝑗 = 1, ..., 𝑛 − 1, be the solutions of (23.1.14); the Jacobian
( 𝑥1 ,..., 𝑥𝑛−1 )
determinant 𝐷 𝐷 ( 𝑦1 ,...,𝑦𝑛−1 ) is equal to 1 when 𝑡 = 0. We can therefore solve with
respect to 𝑦 ′ the equations 𝑥 𝑗 (𝑦 ′, 𝑡) = 𝑥 𝑗 , 𝑗 = 1, ..., 𝑛 − 1; this yields functions
𝑦 𝑗 (𝑥 ′, 𝑡), 𝑗 = 1, ..., 𝑛 − 1; to these we adjoin 𝑦 𝑛 (𝑥 ′, 𝑡) = 𝑡. If we carry out the change
of variable 𝑥 ↦→ 𝑦(𝑥 ′, 𝑥 𝑛 ) we get, by (23.1.14),
𝑛−1
𝜕 𝜕 ∑︁ 𝜕𝑥 𝑗 𝜕
= + = 𝑋.
𝜕𝑦 𝑛 𝜕𝑥 𝑛 𝑗=1 𝜕𝑦 𝑛 𝜕𝑥 𝑗

We also have
𝑛−1
𝜕 ∑︁ 𝜕𝑦 𝑘 𝜕
= (𝑥 ′ (𝑦 ′, 𝑦 𝑛 ) , 𝑦 𝑛 ) , 𝑗 = 1, ..., 𝑛 − 1,
𝜕𝑥 𝑗 𝑘=1 𝜕𝑥 𝑗 𝜕𝑦 𝑘

whence, by (23.1.13),
𝑛−1
∑︁ 𝜕𝑦 𝑘 ′ ′ 𝜕
𝑌= 𝑏 𝑗 (𝑥 ′ (𝑦 ′, 𝑦 𝑛 ) , 𝑦 𝑛 ) (𝑥 (𝑦 , 𝑦 𝑛 ) , 𝑦 𝑛 ) .
𝑗,𝑘=1
𝜕𝑥 𝑗 𝜕𝑦 𝑘
23.1 Pseudodifferential Operators of Principal Type 951

In other words, a change of coordinates in the base brings us into the situation where
the symbol of 𝑋 is 𝑖𝜉 𝑛 and that of 𝑌 is independent of 𝜉 𝑛 . Back to the notation 𝑥 𝑗 for
the coordinates we assume that
𝑛−1
𝜕 ∑︁ 𝜕
𝐿= +𝑖 𝑏 𝑗 (𝑥) (23.1.15)
𝜕𝑥 𝑛 𝑗=1
𝜕𝑥 𝑗

in the whole domain Ω. In the sequel we write b (𝑥) = (𝑏 1 (𝑥) , ..., 𝑏 𝑛−1 (𝑥)), 𝑥 ′ =
(𝑥1 , ..., 𝑥 𝑛−1 ), 𝜉 ′ = (𝜉1 , ..., 𝜉 𝑛−1 ). At this point it is convenient to take Ω = Ω′ ×
(−𝑇, 𝑇) with Ω′ a neighborhood of the origin in R𝑛−1 .
Definition 23.1.23 We say that the vector field (23.1.15) has Property (P), or satisfies
Condition (P), in Ω if, given 𝑥 ′◦ ∈ Ω′ and 𝜉 ′ ∈ R𝑛−1 \ {0} arbitrarily, the linear form
⟨𝜉 ′, b (𝑥)⟩ does not change sign in the interval

𝑰 ◦ (𝑥 ′◦ , 𝑇) = {𝑥 ∈ Ω; 𝑥 ′ = 𝑥 ′◦ , |𝑥 𝑛 | < 𝑇 } . (23.1.16)

Example 23.1.24 The Mizohata vector fields in the plane, 𝐿 𝑘 = 𝜕𝑥𝜕 2 + 𝑖𝑥 2𝑘 𝜕𝑥𝜕 1 (see
Example 23.1.9) have Property (P) in any rectangle 𝑎 < 𝑥1 < 𝑏 (𝑎, 𝑏 ∈ R), |𝑥2 | < 𝑇,
if and only if 𝑘 ∈ 2Z+ .
For (23.1.15) Property (P) means that b (𝑥) does not change direction nor orien-
tation in 𝑰 ◦ (𝑥 ′◦ , 𝑇): either b ≡ 0 in 𝑰 ◦ (𝑥 ′◦ , 𝑇) or else there is a unit vector v (𝑥 ′)
such that
∀𝑥 ∈ 𝑰 ◦ (𝑥 ′◦ , 𝑇) , b (𝑥) = |b (𝑥)| v (𝑥 ′) . (23.1.17)
Indeed, in the latter case, the zeros of 𝑥 𝑛 ↦→ |b (𝑥 ′◦ , 𝑥 𝑛 )| in 𝑰 ◦ (𝑥 ′◦ , 𝑇) are isolated
b( 𝑥)
and v (𝑥 ′) = |b( ′
𝑥) | when b (𝑥) ≠ 0; v (𝑥 ) is a C
𝜔 map of a neighborhood of

𝑥 ′◦ in R𝑛−1 into S𝑛−1 . The geometric interpretation of (23.1.17) is clearest when

Ω = Ω′ × (−𝑇, 𝑇) with Ω′ ⊂ R𝑛−1 open. Let 𝒁 be the analytic subvariety of Ω′
consisting of the points 𝑥 ′ such that b (𝑥 ′, 𝑥 𝑛 ) = 0 for all 𝑥 𝑛 ∈ (−𝑇, 𝑇). Assuming
that 𝒁 ≠ Ω′, i.e., excluding the possibility that 𝐿 = 𝜕𝑥𝜕𝑛 (i.e., that 𝐿 is real in the
original coordinates, after division by a nonvanishing scalar factor), Ω is foliated by
the “vertical lines” {𝑥 ′◦ } × (−𝑇, 𝑇), 𝑥 ′◦ ∈ 𝒁, and the “vertical surfaces” 𝔠 × (−𝑇, 𝑇)
with 𝔠 an integral curve of v in Ω′\𝒁. The sign of 2𝑖1 𝐿 ∧ 𝐿 is well-defined on each
surface 𝔠 × (−𝑇, 𝑇) and turns it into a Riemann surface.
These properties can easily be “globalized” in an analytic manifold M. Let
𝐿 = 𝑋 + 𝑖𝑌 be a complex vector field of class C 𝜔 in M.
Definition 23.1.25 We say that 𝐿 satisfies Condition (P) in the neighborhood of a
point 𝑥 ◦ ∈ M if there is a local coordinates chart (Ω, 𝑥1 , ..., 𝑥 𝑛 ) centered at 𝑥 ◦ , in
which 𝑞 (𝑥) 𝐿, 𝑞 ∈ C 𝜔 (Ω), has an expression (23.1.15) and satisfies Condition (P)
(Definition 23.1.23). We say that 𝐿 satisfies Condition (P) in M if it does in the
neighborhood of every point of M.
It is a trivial consequence of (23.1.15) that a real vector field in M without critical
points (i.e., whose coefficients do not vanish at any point of M) satisfies Condition
(P) in M.
952 23 Analytic PDEs of Principal Type. Local Solvability

It is convenient to introduce the Lie algebra of vector fields (with respect to the
commutation bracket) generated by 𝑋, 𝑌 , 𝔤 (M, 𝑋, 𝑌 ). It defines a Nagano foliation
𝔑𝔞𝔤 (M, 𝑋, 𝑌 ) in M (Theorem 12.4.2). By the integral curves (resp., surfaces) of
𝐿 in M we shall mean the leaves in 𝔑𝔞𝔤 (M, 𝑋, 𝑌 ) of dimension 1 (resp., 2). The
zero-dimensional Nagano leaves are the critical points of 𝐿.

Example 23.1.26 Let 𝐿 = 𝜕𝑥𝜕𝑛 + 𝑖𝑥 𝑛2 𝜕𝑥𝜕 1 in R𝑛 ; there are no integral curves of 𝐿.

If 𝑛 = 2 the whole space R2 is an integral surface of 𝐿; we have 𝐿 ∧ 𝐿 = 0 when
𝑥 𝑛 = 0. When 𝑛 ≥ 3 the integral surfaces of 𝐿 are the affine planes parallel to the
(𝑥1 , 𝑥 𝑛 )-coordinates plane.

The next statement implies the invariance of Property (P) under multiplication of
𝐿 by a nowhere vanishing C 𝜔 function.

Proposition 23.1.27 A C 𝜔 vector field 𝐿 without critical points in M has Property

(P) everywhere in M if and only if the following two properties hold:
(1) M is foliated by the integral surfaces and the integral curves of 𝐿;
(2) the sign of 2𝑖1 𝐿 ∧ 𝐿 is constant in every integral surface of 𝐿.

Proof It suffices to reason in a local chart (Ω, 𝑥1 , ..., 𝑥 𝑛 ) where 𝐿 is given by

(23.1.15); this precludes that Ω contain critical points of 𝐿. It helps to assume that
Ω = Ω′ × (−𝑇, 𝑇) with Ω′ a domain in 𝑥 ′-space R𝑛−1 . Note that, regardless of whether
(P) is valid or not, the integral curves of (23.1.15) in Ω are the “vertical” intervals
{𝑥 ′◦ } × (−𝑇, 𝑇) with 𝑥 ′◦ ∈ 𝒁, meaning that b (𝑥 ′◦ , 𝑥 𝑛 ) = 0 for all 𝑥 𝑛 ∈ (−𝑇, 𝑇).
Indeed, no point 𝑥 ◦ ∈ (Ω′\𝒁) × (−𝑇, 𝑇) can belong to an integral curve of 𝐿 since
the zeros of b (𝑥 ′◦ , 𝑥 𝑛 ) are isolated in (−𝑇, 𝑇). Definition 23.1.23 and (23.1.17) show
immediately that (1) and (2) are direct consequences of (P). Now suppose (1) and
(2) hold. By (2) there is an integral surface S of 𝐿 through 𝑥 ◦ ∈ (Ω′\𝒁) × (−𝑇, 𝑇).
Since 𝜕𝑥𝜕𝑛 is tangent to S at every point of S there is a neighborhood 𝑈 of 𝑥 ◦ in Ω
such that S ∩𝑈 = 𝔠 × 𝑥 𝑛◦ − 𝜀, 𝑥 𝑛◦ + 𝜀 with 𝔠 an analytic, one-dimensional, connected

submanifold of Ω′. Let v (𝑥 ′) be an analytic vector field in a neighborhood 𝑈 ′ of
𝑥 ◦′ tangent to 𝔠 and such that |v (𝑥 ′)| = 1 at every point of 𝔠. This, combined
with (23.1.15), implies that, in S ∩ 𝑈, we have Im 𝐿 = 𝛽 (𝑥) v (𝑥 ′) with 𝛽 ∈
C 𝜔 (S ∩ 𝑈; R), and therefore
𝑛−1
𝜕 𝜕 ∑︁ 𝜕
∧ Im 𝐿 = 𝛽 (𝑥) ∧ 𝑣 𝑗 (𝑥 ′) .
𝜕𝑥 𝑛 𝜕𝑥 𝑛 𝑗=1 𝜕𝑥 𝑗

Property (2) requires that 𝛽 (𝑥) not change sign in S ∩ 𝑈; we conclude that
𝑥 𝑛◦ − 𝜀, 𝑥 𝑛◦ + 𝜀 ∋ 𝑥 𝑛 ↦→ 𝛽 (𝑥 ′, 𝑥 𝑛 ) does not change sign whatever 𝑥 ′ ∈ 𝔠.

□

Corollary 23.1.28 If 𝐿 does not have critical points and satisfies (P) in M then the
integral curves and the integral surfaces of 𝐿 are the leaves in 𝔑𝔞𝔤 (𝑋, 𝑌 ).

Proof Follows from Proposition 23.1.27 and the general fact that if 𝑋 and 𝑌 are
tangent to a smooth submanifold the same is true of their Lie brackets of all orders.□
23.1 Pseudodifferential Operators of Principal Type 953

Example 23.1.29 The whole plane is an integral surface of the Mizohata vector
fields 𝐿 𝑘 = 𝜕𝑥𝜕 2 + 𝑖𝑥 2𝑘 𝜕𝑥𝜕 1 (𝑘 ∈ Z+ ). When 𝑘 is odd the signs of 2𝑖1 𝐿 𝑘 ∧ 𝐿 𝑘 in the
half-planes 𝑥 2 > 0 and 𝑥2 < 0 are different.

Remark 23.1.30 There are vector fields 𝐿 = 𝑋 + 𝑖𝑌 in R𝑛 (𝑛 ≥ 3) that do not satisfy

(P) and yet are such that the dimension of every leaf in 𝔑𝔞𝔤 (𝑋, 𝑌 ) does not exceed
2. Example: 𝐿 = 𝜕𝑥𝜕 1 + 𝑖𝑥1 𝜕𝑥𝜕 2 in R3 .

Remark 23.1.31 Let 𝑛 ≥ 3 and 2 ≤ 𝑑 ≤ 𝑛. There are vector fields 𝐿 = 𝑋 + 𝑖𝑌

in R𝑛 such that every leaf in 𝔑𝔞𝔤 (𝑋, 𝑌 ) is 𝑑-dimensional. Example: 𝐿 = 𝜕𝑥𝜕 1 +
Í
𝑖 𝑑𝑘=2 𝑥1𝑘−2 𝜕𝑥𝜕𝑘 .

Next, we take a look at the geometry “above” the base M, in the cotangent bundle
𝑇 ∗ M.

Proposition 23.1.32 Let 𝐿 be an analytic vector field without critical points in M

[but not required to satisfy (P)]. The following two properties of an immersed one-
dimensional analytic submanifold 𝔠 of M are equivalent: 1) 𝔠 is an integral curve
of 𝐿; 2) 𝔠 is the base projection of a null bicharacteristic curve of 𝐿 in 𝑇 ∗ M\0
(Definition 23.1.15).

Proof It suffices to prove the result locally and therefore to reason in the Euclidean
set-up. Suppose 𝑥 ◦ ∈ 𝔠; a change of variables brings us to a vector field (23.1.15)
in which case there is an arc of 𝔠 equal to 𝑰 ◦ (𝑥 ′◦ , 𝑇) in the local coordinates, with
b (𝑥 ′◦ , 𝑥 𝑛 ) = 0 for all 𝑥 𝑛 ∈ 𝑥 𝑛◦ − 𝑇, 𝑥 𝑛◦ + 𝑇 . Every straight-line segment

{(𝑥, 𝜉) ∈; 𝑥 ∈ 𝑰 ◦ (𝑥 ′◦ , 𝑇) , 𝜉 𝑛 = 0, 𝜉 ′ = 𝜉 ′◦ }

(𝜉 ′◦
√ ≠ 0) is contained in Char 𝐿 and 𝐿 is (always) equal to the base projection of
−1𝐻 𝐿 . □
Í𝑛
With 𝑋, 𝑌 Ías in (23.1.12) we use the notation 𝐴 = ⟨𝜎, 𝑋⟩ = 𝑘=1 𝑎 𝑘 (𝑥) 𝜉 𝑘 ,
𝐵 = ⟨𝜎, 𝑌 ⟩ = 𝑛𝑘=1 𝑏 𝑘 (𝑥) 𝜉 𝑘 .

Proposition 23.1.33 Let 𝐿 = 𝑋 +𝑖𝑌 be an analytic vector field in an open subset Ω of

R𝑛 , 𝑛 ≥ 3. If 𝐿 satisfies (P) then every integral surface of 𝐿 is the base projection of a
null bicharacteristic surface of 𝐴 +𝑖𝐵 (Definition 23.1.21) on which the Hamiltonian
vector fields 𝐻 𝐴 and 𝐻 𝐵 define a Riemann surface structure.

Proof Let 𝐿 be given by (23.1.15) and assume that (23.1.17) holds; Char 𝐿 is
defined by the equations 𝜉 𝑛 = ⟨𝜉 ′, 𝒃 (𝑥)⟩ = 0. The Hamiltonian fields of 𝐴 = 𝜉 𝑛 ,
𝐵 = ⟨𝜉 ′, 𝒃 (𝑥)⟩ are
𝑛−1 𝑛−1 ∑︁
𝑛
𝜕 ∑︁ 𝜕 ∑︁ 𝜕𝑏 𝑗 𝜕
𝐻𝐴 = , 𝐻𝐵 = 𝑏 𝑗 (𝑥) − 𝜉𝑗 .
𝜕𝑥 𝑛 𝑗=1
𝜕𝑥 𝑗 𝑗=1 𝑘=1
𝜕𝑥 𝑘 𝜕𝜉 𝑘
954 23 Analytic PDEs of Principal Type. Local Solvability

Recall that
𝒁 = {𝑥 ′ ∈ Ω′; ∀𝑥 𝑛 ∈ (−𝑇, 𝑇) , 𝒃 (𝑥 ′, 𝑥 𝑛 ) = 0} .
𝒃 ( 𝑥)
Here we restrict the variation of 𝑥 ′ to Ω′\𝒁; we have 𝒗 = ∈ C 𝜔 Ω′\𝒁; S𝑛−2 .

|𝒃 ( 𝑥) |
Note that the function

(Ω′\𝒁) × (−𝑇, 𝑇) ∋ 𝑥 ↦→ |𝒃 (𝑥)| ∈ [0, +∞)

is analytic, since
𝜕 𝜕𝒃
|𝒃 (𝑥)| = 𝒗 · , 𝑗 = 1, ..., 𝑛.
𝜕𝑥 𝑗 𝜕𝑥 𝑗
We have

𝜕𝒗 (𝑥 ′) 𝜕
𝑛−1
∑︁ 𝜕
𝐻 𝐵 = |𝒃 (𝑥)| 𝑣 𝑗 (𝑥 ′) − 𝜉 ′,
𝑗=1
𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝜉 𝑗
𝑛
∑︁ 𝜕 |𝒃 (𝑥)| 𝜕
− ⟨𝜉 ′, 𝒗 (𝑥 ′)⟩ .
𝑗=1
𝜕𝑥 𝑗 𝜕𝜉 𝑗

Since 𝒃 (𝑥) = 0 implies d |𝒃 (𝑥)| = 0 the restriction of 𝐻 𝐵 to Char 𝐿 is equal to

|𝒃 (𝑥)| 𝐻𝒗 with
𝑛−1 ′

𝜕 ′ 𝜕𝒗 (𝑥 ) 𝜕
∑︁
′
𝐻𝒗 = 𝑣 𝑗 (𝑥 ) − 𝜉, .
𝑗=1
𝜕𝑥 𝑗 𝜕𝑥 𝑗 𝜕𝜉 𝑗

Let 𝑥 ′◦ ∈ Ω′\𝒁 be arbitrary and let R ∋ 𝑡 ↦→ 𝐸 (𝑥 ′◦ , 𝑡) = (exp 𝑡 𝑋) 𝑥 ′◦

be the canonical C 𝜔 immersion of R into Ω′\𝒁 as the integral curve 𝔠 of
Í
𝑋 = 𝑛−1 ′ 𝜕 ′◦
𝑗=1 𝑣 𝑗 (𝑥 ) 𝜕𝑥 𝑗 through 𝑥 . We now regard R as a submanifold of R
𝑛−1

(defined by the equations 𝑥 𝑗 = 0, 𝑗 = 1, ..., 𝑛 − 2) and we use the map 𝐸 (𝑥 ′◦ , 𝑡) −1

to push forward the tangent line to 𝔠 into the tangent line 𝑇R ⊂ 𝑇R𝑛−1 . We can pull
back via 𝐸 (𝑥 ′◦ , 𝑡) −1 the conormal bundle of R in R𝑛−1 , 𝑁 ∗ R R × R𝑛−2 , onto the
conormal bundle 𝑁 ∗𝔠 of 𝔠 in Ω′. Let 𝒆 = (𝑒 1 , ..., 𝑒 𝑛−2 ) ∈ R𝑛−2 \ {0} be arbitrary; if
we view 𝒆 as a (constant) section of 𝑁 ∗ R its pullback via 𝐸 (𝑥 ′◦ , 𝑡) −1 is a C 𝜔 section
𝜍 (𝒆, 𝑡) of 𝑁 ∗𝔠, 𝜍 (𝒆, 𝑡) ≠ 0 for all 𝑡 ∈ R; we have ⟨𝜍 (𝒆, 𝑡) , 𝒗 (𝐸 (𝑥 ′◦ , 𝑡))⟩ = 0 for all
𝑡 ∈ R.
To better visualize what we are doing we make an analytic change of variables in a
neighborhood 𝜔 (𝑥 ′◦ ) of 𝑥 ′◦ that transforms 𝑥 ′◦ into 0, ..., 0, 𝑥 𝑛−1
◦ and the connected
component of 𝔠 ∩ 𝜔 (𝑥 ′◦ ) that contains 𝑥 ′◦ into an interval

𝑥 ∈ R𝑛−1 ; 𝑥 ′′ = 0, 𝑥 𝑛−1 − 𝑥 𝑛−1

◦

<𝜀 ,

where 𝑥 ′′ = ( 𝑥1 , ..., 𝑥 𝑛−2 ). [Since we are dealing with immersions it may happen
that there are infinitely many connected component of 𝔠 ∩ 𝜔 (𝑥 ′◦ ).] Provided 𝜀 > 0
is sufficiently small there is a neighborhood 𝜔 ′′ of 0 in R𝑛−2 such that
23.1 Pseudodifferential Operators of Principal Type 955

𝜔 ′′ × (−𝜀, 𝜀) ∋ (𝑥 ′′, 𝑡) ↦→ 𝐸 (𝑥 ′′, 𝑡) = (exp 𝑡 𝑋) 𝑥 ′′, 𝑥 𝑛−1

◦

is a C 𝜔 diffeomorphism onto a neighborhood 𝜔 ′ of 𝑥 ′◦ in Ω′\𝒁. If we identify

𝒆 with the differential d (𝒆 · 𝑥 ′′) in R𝑛−2 we can pullback the function 𝑥 ′ ↦→ 𝒆 ·
𝑥 ′′ in 𝜔 ′′ × (−𝜀, 𝜀) via the map 𝐸 (𝑥 ′′, 𝑡) ↦→ (𝑥 ′′, 𝑡) as a C 𝜔 function 𝜑 in 𝜔 ′
that vanishes identically on 𝔠 ∩ 𝜔 ′ and is such that d𝜑| 𝐸 (0,𝑡) = 𝜍 (𝑡). The latter
implies ⟨d𝜑, 𝒗⟩ ≡ 0 in 𝔠 ∩ 𝜔 ′; if we regard (𝐸 (0, 𝑡) , 𝑥 𝑛 ) as the variable point in
(𝔠 ∩ 𝜔 ′) × (−𝑇, 𝑇) and regard 𝜉 (𝑡) = (𝜍 (𝑡) , 0) as a section of 𝑇 ∗ Ω over 𝔠 ∩ 𝜔 ′ we
see that ((𝐸 (0, 𝑡) , 𝑥 𝑛 ) , 𝜉 (𝑡)) describes a surface Σ (𝑥 ′◦ ) ⊂ Char 𝐿. It is obvious
that 𝐻 𝐴 = 𝜕𝑥𝜕𝑛 is tangent to Σ (𝑥 ′◦ ). The equation defining Σ (𝑥 ′◦ ) in 𝑇 ∗ Ω can be
taken to be
𝜕𝜑 ′
𝑥 ′ = 𝐸 (0, 𝑡) , 𝜉 𝑗 = (𝑥 ) , 𝜉 𝑛 = 0.
𝜕𝑥 𝑗
We see immediately that, when 𝑥 ′ ∈ 𝔠 ∩ 𝜔 ′, 𝜉 ′ = d𝜑 (𝑥 ′),

𝜕𝜑 ′ 𝜕𝒗 𝑘 (𝑥 ′)
𝑛−1
𝜕2 𝜑

𝜕𝜑 ′ ∑︁
𝐻𝒗 𝜉 𝑗 − (𝑥 ) = − 𝑣 𝑘 (𝑥 ′) (𝑥 ′) + (𝑥 ) ,
𝜕𝑥 𝑗 𝑘=1
𝜕𝑥 𝑗 𝜕𝑥 𝑘 𝜕𝑥 𝑘 𝜕𝑥 𝑗
𝑛−1
𝜕 ∑︁ 𝜕𝜑 ′
=− 𝑣 𝑘 (𝑥 ′) (𝑥 )
𝜕𝑥 𝑗 𝑘=1 𝜕𝑥 𝑘

vanishes identically. We also have 𝐻𝒗 𝑥 𝑗 − 𝐸 𝑗 (0, 𝑡) = 0 if |𝑡| < 𝜀 since the
equations 𝑥 𝑗 − 𝐸 𝑗 (0, 𝑡) = 0 ( 𝑗 = 1, ..., 𝑛 − 1) describe the arc 𝔠 ∩ 𝜔 ′. This proves
that 𝐻 𝐵 is also tangent to Σ (𝑥 ′◦ ) and implies that the same is true of the restriction
to Σ (𝑥 ′◦ ) of every vector field in 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) [U: a neighborhood of Σ (𝑥 ′◦ ) in
𝑇 ∗ Ω\0]. As (𝑥 ′◦ , 𝑥 𝑛 ) ranges over 𝔠 × (−𝑇, 𝑇) we can patch together the surfaces
Σ (𝑥 ′◦ ) to get an immersed two-dimensional analytic submanifold Σ (𝔠, 𝒆) of 𝑇 ∗ Ω\0
with the following properties:
(1) Σ (𝔠, 𝒆) ⊂ Char 𝐿;
(2) Σ (𝔠, 𝒆) is a leaf belonging to 𝔑𝔞𝔤 (𝑇 ∗ Ω\0, 𝐻 𝐴, 𝐻 𝐵 );
(3) the base projection of Σ (𝔠, 𝒆) is the integral surface 𝔠 × (−𝑇, 𝑇) of 𝐿;
(4) Σ (𝔠, 𝒆) does not contain any null bicharacteristic curve of 𝐿 [the base projection
of the latter being intervals {𝑥 ′◦ } × (−𝑇, 𝑇), 𝑥 ′◦ ∈ 𝒁].
Noting that 𝐻𝒗 does not vanish at any point of Σ (𝔠, 𝒆) and that the zeros of
(−𝑇, 𝑇) ∋ 𝑥 𝑛 ↦→ |𝒃 (𝑥 ′◦ , 𝑥 𝑛 )| are isolated we see that 𝐻 𝐴 ∧ 𝐻 𝐵 = |𝒃| 𝐻 𝐴 ∧ 𝐻𝒗
defines an orientation, hence a Riemann surface structure on Σ (𝔠, 𝒆). □
The claim in Proposition 23.1.33 is patently false when 𝑛 = 2, as shown by the
operator 𝜕𝑥𝜕 1 − 𝑖 𝜕𝑥𝜕 2 in the plane. By contrast, we have the following

Example 23.1.34 Consider 𝜕𝑥𝜕 1 − 𝑖 𝜕𝑥𝜕 2 in R3 ; the Nagano leaves of (𝑋, 𝑌 ) are the
affine planes 𝑥3 = 𝑥3◦ . Here 𝐴 = 𝜉2 , 𝐵 = −𝜉1 ,

Char 𝐿 = (𝑥, 𝜉) ∈ 𝑇 ∗ R3 ; 𝜉1 = 𝜉2 = 0, 𝜉3 ≠ 0 .

956 23 Analytic PDEs of Principal Type. Local Solvability

The Nagano leaves of (𝐻 𝐴, 𝐻 𝐵 ) are the affine planes

(𝑥, 𝜉) ∈ 𝑇 ∗ R3 ; 𝜉1 = 𝜉2 = 0, 𝑥3 = 𝑥 3◦ , 𝜉3 = 𝜉3◦ ≠ 0 .

Remark 23.1.35 Even when 𝑛 ≥ 3 there might be analytic submanifolds of 𝑇 ∗ M\0

whose base projection is not a leaf of (𝑋, 𝑌 ), yet are contained in the intersection of
Char 𝐿 with a Nagano leaf of (𝐻 𝐴, 𝐻 𝐵 ). It is evident in the case of 𝐿 = 𝜕𝑥𝜕 1 + 𝑖𝑥12 𝜕𝑥𝜕 2
in R2 . As shown below it remains true if 𝐿 is regarded as a vector field in R3 [that
satisfies (P), obviously].

Example 23.1.36 If 𝐿 = 𝜕
𝜕𝑥1 + 𝑖𝑥12 𝜕𝑥𝜕 2 in R3 we have

Char 𝐿 = (𝑥, 𝜉) ∈ 𝑇 ∗ R3 ; 𝑥1 𝜉2 = 𝜉1 = 0, 𝜉22 + 𝜉32 ≠ 0 .

In the notation of the proof of Proposition 23.1.33 we have 𝒁 = ∅; 𝔤 R3 , 𝑋, 𝑌 is

spanned by 𝜕𝑥𝜕 1 , 𝜕𝑥𝜕 2 ; the Nagano leaves of (𝑋, 𝑌 ) are the affine planes 𝑥3 = 𝑥3◦ . If
we write 𝐴 = 𝜉1 , 𝐵 = 𝑥12 𝜉2 , we have

𝜕 𝜕 𝜕
𝐻𝐴 = , 𝐻 𝐵 = 𝑥12 − 2𝑥1 𝜉2 ,
𝜕𝑥 1 𝜕𝑥 2 𝜕𝜉1
1 𝜕 𝜕
[𝐻 𝐴, 𝐻 𝐵 ] = 𝑥 1 − 𝜉2 ,
2 𝜕𝑥2 𝜕𝜉1
1 𝜕
[𝐻 𝐴, [𝐻 𝐴, 𝐻 𝐵 ]] = .
2 𝜕𝑥2
The 3D Nagano leaves of (𝐻 𝐴, 𝐻 𝐵 ) are the affine spaces parallel to the (𝑥1 , 𝑥2 , 𝜉1 )-
coordinates space,

Π± 𝑥 3◦ , 𝜉2◦ , 𝜉3◦ = (𝑥, 𝜉) ∈ 𝑇 ∗ R3 ; 𝑥3 = 𝑥3◦ , 𝜉2 = 𝜉2◦ ≠ 0, 𝜉3 = 𝜉3◦ ;

they intersect Char 𝐿 along straight lines parallel to the 𝑥2 -axis.

23.1.5 Condition (P) for analytic pseudodifferential operators

Before restricting our attention to the analytic framework consider a pair of real-
valued C ∞ functions 𝐴, 𝐵 in 𝑇 ∗ M\0, both homogeneous of degree 𝑚; M is a C ∞
manifold; U is a conic open subset of 𝑇 ∗ M\0. We shall reason, provisionally, under
an asymmetrical hypothesis, in that it gives a preponderant role to the real part 𝐴 of
the symbol 𝑃𝑚 = 𝐴 + 𝑖𝐵:
(prtA) ∀℘ ∈ U, if 𝐴 (℘) = 0 then d𝐴 ∧ 𝜎 ≠ 0 at ℘.
23.1 Pseudodifferential Operators of Principal Type 957

Let 𝐻 𝐴 denote the Hamiltonian vector field of 𝐴; (prtA) is equivalent to saying

that 𝐻 𝐴 is linearly independent from the radial vector field (23.1.6) at every point
℘ ∈ U such that 𝐴 (℘) = 0 (Proposition 23.1.3). Obviously, (prtA) implies that 𝐴
and therefore 𝑃𝑚 is of principal type in U.

Definition 23.1.37 Let 𝑃 be a classical pseudodifferential operator in M with prin-

cipal symbol 𝑃𝑚 = 𝐴 + 𝑖𝐵. We shall say that 𝐴 + 𝑖𝐵 (as well as 𝑃) has Property
(P) in U if Condition (prtA) holds and if the restriction of 𝐵 to an arbitrary null
bicharacteristic curve of 𝐴 in U (Definition 23.1.15) does not change sign.

Property (P) is invariant under homogeneous symplectic transformations [see

(23.1.2)] but not, obviously, under multiplication of 𝐴 + 𝑖𝐵 by 𝑖. This will now be
remedied in the C 𝜔 set-up (see Theorem 23.1.46 and Remark 23.1.44 below) to
which we now return: in the remainder of this section pseudodifferential operators
and symbols shall be defined and analytic in analytic submanifolds of a C 𝜔 manifold
M or of 𝑇 ∗ M\0.
Suppose that the base projection of the conic open set U is contained in the
domain of analytic local coordinates 𝑥1 , ..., 𝑥 𝑛 . This allows us to reason in an open
subset Ω of R𝑛 and take U = 𝑈 × Γ, 𝑈 ⊂⊂ Ω open, Γ ⊂ R𝑛 \ {0} an open cone.
For simplicity we assume that 0 ∈ 𝑈 and center the analysis at a point (0, 𝜉 ◦ ),
𝜉 ◦ ∈ Γ. In the coming reasoning we let 𝑈 contract about 0 and Γ about the ray
(0, 𝜆𝜉 ◦ ), 𝜆 > 0, as needed. The results we are seeking will be invariant under
homogeneous symplectic transformations in what concerns symbols; they will be
invariant under conjugation with unitary FIOs in what concerns pseudodifferential
operators. We can therefore avail ourselves of Proposition 23.1.8 and take 𝐴 = 𝜉 𝑛 ,
𝐵 = 𝐵 (𝑥, 𝜉 ′) ∈ C 𝜔 (U) real-valued, homogeneous of degree 1 [cf. (23.1.7)].
We are going to exploit the following elementary statement.

Lemma 23.1.38 For a nonconstant, monic polynomial 𝑝 (𝑡) ∈ C [𝑡] to be real (i.e.,
to have real coefficients) it is necessary and sufficient that this be true of 𝑝 (𝑡) 𝑚 for
some positive integer 𝑚.

Proof If 𝑝 (𝑡) is real every (integer) power of 𝑝 (𝑡) is real. Conversely, suppose
𝑝 𝑚 (𝑡) ∈ R [𝑡] for some positive integer 𝑚; this implies that the rational function
deg
∑︁𝑝 1
d 𝑚 𝑝 ′ (𝑡)
𝑝 (𝑡) −𝑚 𝑝 (𝑡) = 𝑚 =𝑚
d𝑡 𝑘=1
𝑡 − 𝜌𝑘 𝑝 (𝑡)

deg
Ö𝑝
is real for all 𝑡 ∈ R such that 𝑝 (𝑡) = (𝑡 − 𝜌 𝑘 ) ≠ 0. This demands that if a root
𝑘=1
𝜌 𝑘 is not real then there be 𝑘 ′ ≠ 𝑘 such that 𝜌 𝑘′ = 𝜌 𝑘 , whence 𝑝 (𝑡) ∈ R [𝑡]. □
958 23 Analytic PDEs of Principal Type. Local Solvability

Proposition 23.1.39 For 𝜉 𝑛 + 𝑖𝐵 (𝑥, 𝜉 ′) [cf. (23.1.7)] to have Property (P) in U =

𝑈 × Γ, suitably contracted about (0, 𝜉 ◦ ), it is necessary and sufficient that there be
two real-valued analytic functions in U, 𝐺 (𝑥, 𝜉 ′) ≥ 0 and 𝐹 (𝑥 ′, 𝜉 ′), homogeneous
of degree 0 and 1 respectively, such that

∀ (𝑥, 𝜉 ′, 0) ∈ U, 𝐵 (𝑥, 𝜉 ′) = 𝐺 (𝑥, 𝜉 ′) 𝐹 (𝑥 ′, 𝜉 ′) . (23.1.18)

Proof The sufficiency of the condition is evident since the null bicharacteristics of
𝐴 = 𝜉 𝑛 in U are intervals in the 𝑥 𝑛 -lines 𝑥 ′ = const ., 𝜉 ′ = const ., 𝜉 𝑛 = 0. We prove
the necessity of the condition. We take 𝑈 = 𝑈 ′ × (−𝑇, 𝑇), 𝑈 ′ a neighborhood of the
origin in 𝑥 ′-space R𝑛−1 and 𝑇 > 0. We shall assume that 𝐵 changes sign in every
neighborhood of (0, 𝜉 ◦ ); if this were not true we could take 𝐺 = |𝐵| and 𝐹 = ±1.
The function ∫ 𝑇
𝑀 (𝑥 ′, 𝜉 ′) = 𝐵 (𝑥 ′, 𝑥 𝑛 , 𝜉 ′) d𝑥 𝑛
−𝑇
is analytic in 𝑈 ′ × Γ. If 𝐵 (𝑥 ′, 𝑥 𝑛 , 𝜉 ′) > 0 (resp., < 0) for a given (𝑥 ′, 𝑥 𝑛 , 𝜉 ′) ∈ U,
𝑥 𝑛 ∈ (−𝑇, 𝑇), then 𝑀 (𝑥 ′, 𝜉 ′) > 0 (resp., < 0). If 𝑀 (𝑥 ′, 𝜉 ′) = 0 then 𝐵 (𝑥 ′, 𝑥 𝑛 , 𝜉 ′) =
0 for all 𝑥 𝑛 ∈ (−𝑇, 𝑇).
Because of the homogeneity we may limit the variation of 𝜉 ′ to a neighborhood of
𝜉 in R𝑛−1 . It is convenient to regard the 𝑥 𝑗 and 𝜉 𝑘 − 𝜉 𝑘◦ (1 ≤ 𝑗, 𝑘 ≤ 𝑛 − 1) as forming
′◦

a single set of 2 (𝑛 − 1) real variable, which we will denote by 𝜃 = (𝜃 1 , ..., 𝜃 2𝑛−2 ),

varying in an open polycube
n o
𝔔 𝜀(2𝑛−2) = 𝜃 ∈ R2(𝑛−1) ; |𝜃 𝑘 | < 𝜀, 𝑘 = 1, ..., 2𝑛 − 2 .

We write 𝐵 (𝜃, 𝑥 𝑛 ) and 𝑀 (𝜃) in the place of 𝐵 (𝑥 ′, 𝑥 𝑛 , 𝜉 ′) and 𝑀 (𝑥 ′, 𝜉 ′). We ap-

ply the Weierstrass Preparation Theorem (Theorem 14.3.4) and, possibly after a
relabeling of the variables 𝜃 𝑘 , write

𝑀 (𝜃) = 𝐾 (𝜃) 𝑀1𝑚1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀 𝑁

𝑚𝑁
(𝜃 ′′; 𝜃 1 )

where 𝐾 (𝜃) ≠ 0 for all 𝜃 ∈ 𝔔 𝜀(2𝑛−2) and the 𝑀 𝛼𝑚 𝛼 (𝜃 ′′; 𝜃 1 ) are irreducible Weier-
strass polynomials (Definition 14.3.3); 𝜃 ′′ = (𝜃 2 , ..., 𝜃 2𝑛−2 ) and the 𝑚 𝛼 are positive
integers. We label the polynomials 𝑀 𝛼 (𝜃 ′′; 𝜃 1 ) so that

𝑀1𝑚1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀𝜈𝑚𝜈 (𝜃 ′′; 𝜃 1 )

changes sign in 𝔔 𝜀(2𝑛−2) whereas 𝑀𝜈+1 𝑚𝜈+1 ′′ 𝑚𝑁

(𝜃 ; 𝜃 1 ) · · · 𝑀 𝑁 (𝜃 ′′; 𝜃 1 ) does not; more-
over, we take 𝜈 to be the smallest positive integer such that this is valid. We claim that
if 𝛼 ≤ 𝜈 then 𝑀 𝛼𝑚 𝛼 (𝜃 ′′; 𝜃 1 ) is real and changes sign in 𝔔 𝜀(2𝑛−2) ; by Lemma 23.1.38
this is equivalent to saying that 𝑀 𝛼 (𝜃 ′′; 𝜃 1 ) is real and changes sign in 𝔔 𝜀(2𝑛−2) . If
𝑚
this were not the case there would be 𝛽 ≠ 𝛼 such that 𝑀𝛽 𝛽 (𝜃 ′′; 𝜃 1 ) = 𝑀 𝛼𝑚 𝛼 (𝜃 ′′; 𝜃 1 );
but this would imply that either

𝑀 𝛼𝑚 𝛼 (𝜃 ′′; 𝜃 1 ) 𝑀𝛽 𝛽 (𝜃 ′′; 𝜃 1 ) 𝑀𝜈+1

𝑚𝜈+1 ′′
(𝜃 ′′; 𝜃 1 ) if 𝛽 ≤ 𝜈,
𝑚 𝑚𝑁
(𝜃 ; 𝜃 1 ) · · · 𝑀 𝑁
23.1 Pseudodifferential Operators of Principal Type 959

or
𝑀 𝛼𝑚 𝛼 (𝜃 ′′; 𝜃 1 ) 𝑀𝛽 𝛽 (𝜃 ′′; 𝜃 1 )𝑀𝜈+1
𝑚𝜈+1 ′′
(𝜃 ′′; 𝜃 1 ) if 𝛽 > 𝜈,
𝑚 𝑚𝑁
(𝜃 ; 𝜃 1 ) · · · 𝑀 𝑁

does not change sign in 𝔔 𝜀(2𝑛−2) , contradicting the minimality of 𝜈. The powers 𝑚 𝛼
must be odd integers if 𝛼 ≤ 𝜈.
From the properties of 𝑀 relative to 𝐵 we know that 𝑀 𝛼 (𝜃 ′′; 𝜃 1 ) = 0 entails
𝐵 (𝜃, 𝑥 𝑛 ) = 0 for all 𝑥 𝑛 ∈ (−𝑇, 𝑇). With possibly a smaller 𝜀 > 0, 𝑀 𝛼 can be
extended as a holomorphic function in the polydisk

Δ (2𝑛−2) = 𝜃˜ ∈ C2𝑛−2 ; 𝜃˜ 𝑗 < 𝜀, 𝑗 = 1, ..., 2𝑛 − 2

𝜀

˜ 𝑥 𝑛 in Δ (2𝑛−2)

while 𝐵 can be extended as an analytic function of 𝜃, 𝜀 × (−𝑇, 𝑇) holo-
morphic with respect to 𝜃.˜ The equation 𝑀 𝛼 𝜃˜ = 0 defines an irreducible complex-
analytic subvariety 𝑽 C𝛼 of Δ (2𝑛−2)
𝜀 ; setting 𝑽 𝛼 = 𝑽 C𝛼 ∩ R𝑛 we have dimR 𝑽 𝛼 = 2𝑛 − 3,
otherwise 𝑀 𝛼 would not change sign in 𝔔 𝜀(2𝑛−2) ; as a consequence, dimC 𝑽 𝛼 = 2𝑛−3.
C

Since 𝑽 C𝛼 is irreducible the regular part ℜ 𝑽 C𝛼 is connected (Proposition 14.3.22).

Since 𝐵 (𝜃, 𝑥 𝑛 ) ≡ 0 in ℜ 𝑽 C𝛼 ∩ 𝔔 𝜀(2𝑛−2) for every 𝑥 𝑛 ∈ (−𝑇, 𝑇) the same is true

in an open subset of ℜ 𝑽 C𝛼 and therefore in the whole of ℜ 𝑽 C𝛼 . It follows that
𝐵 (𝜃, 𝑥 𝑛 ) is divisible by 𝑀 𝛼 (𝜃) for each 𝛼 = 1, ..., 𝜈:

𝐵 (𝜃, 𝑥 𝑛 ) = 𝐺 (𝜃, 𝑥 𝑛 ) 𝑀1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀𝜈 (𝜃 ′′; 𝜃 1 ).

We derive from this that, for 𝜃 ∈ 𝔔 𝜀(2𝑛−2) fixed such that 𝑀1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀𝜈 (𝜃 ′′; 𝜃 1 ) ≠
0, 𝑥 𝑛 ↦→ 𝐺 (𝜃, 𝑥 𝑛 ) does not change sign in (−𝑇, 𝑇). We now have

𝑀 (𝜃) = 𝐾1 (𝜃) 𝑀1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀𝜈 (𝜃 ′′; 𝜃 1 )

where, owing to the fact that the powers 𝑚 𝛼 − 1 are even,

𝜈
Ö 𝑁
Ö
𝑀 𝛼𝑚 𝛼 −1 (𝜃 ′′; 𝜃 1 ) 𝑀𝛽 𝛽 (𝜃 ′′; 𝜃 1 )
𝑚
𝐾1 (𝜃) = 𝐾 (𝜃)
𝛼=1 𝛽=𝜈+1

does not change sign in 𝔔 𝜀(2𝑛−2) . Since

∫ 𝑇
𝐺 (𝜃, 𝑥 𝑛 ) d𝑥 𝑛 = 𝐾1 (𝜃)
−𝑇

we conclude that either 𝐺 (𝜃, 𝑥 𝑛 ) ≥ 0 or 𝐺 (𝜃, 𝑥 𝑛 ) ≤ 0 in the whole region

n o
(𝜃, 𝑥 𝑛 ) ∈ 𝔔 𝜀(2𝑛−2) × (−𝑇, 𝑇) ; 𝑀1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀𝜈 (𝜃 ′′; 𝜃 1 ) ≠ 0 ,

which is dense in 𝔔 𝜀(2𝑛−2) × (−𝑇, 𝑇). We conclude that (23.1.12) holds with
𝐺 (𝑥, 𝜉 ′) = 𝐺 (𝜃, 𝑥 𝑛 ) and 𝐹 (𝑥 ′, 𝜉 ′) = 𝜍 𝑀1 (𝜃 ′′; 𝜃 1 ) · · · 𝑀𝜈 (𝜃 ′′; 𝜃 1 ), 𝜍 = +1 or −1. □
960 23 Analytic PDEs of Principal Type. Local Solvability

Corollary 23.1.40 Suppose the pair 𝜉 𝑛 + 𝑖𝐵 (𝑥, 𝜉 ′) has Property (P) in U = 𝑈 ′ ×

(−𝑇, 𝑇) × Γ. Then, given arbitrarily (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑈 ′ × (−𝑇, 𝑇) × Γ such that 𝜉 𝑛◦ = 0,
𝐵 (𝑥 ◦ , 𝜉 ◦′) = 0, at least one of the following two conditions is satisfied:

∀𝑥 𝑛 ∈ (−𝑇, 𝑇) , 𝐵 (𝑥 ◦′, 𝑥 𝑛 , 𝜉 ◦′) = 0, (23.1.19)

or d 𝜉 ′ 𝐵 (𝑥 ◦ , 𝜉 ◦′) = 0.

Proof Back to (23.1.18) we note that 𝐺 (𝑥 ◦ , 𝜉 ◦ ) = 0 entails d𝐺 (𝑥 ◦ , 𝜉 ◦ ) = 0 and

therefore also d 𝜉 ′ 𝐵 (𝑥 ◦ , 𝜉 ◦′) = 0. If 𝐺 (𝑥 ◦ , 𝜉 ◦ ) ≠ 0 then 𝐵 (𝑥 ◦ , 𝜉 ◦′) = 0 entails
𝐹 (𝑥 ◦ , 𝜉 ◦′) = 0 and therefore also (23.1.19). □

Remark 23.1.41 The factorization (23.1.18) is possible because the symbols are
analytic; it is not possible in the general C ∞ case. The reader might try to prove that
it is impossible in the following special case [where 𝑛 = 3, 𝜉 ′ = (𝜉1 , 𝜉2 ) ]:

𝐵 (𝑥, 𝜉 ′) = |𝜉 ′ | −1 (𝜉2 − 𝑥 3 𝜉1 exp (−1/𝑥2 )) 2 if 𝑥 2 > 0,

𝐵 (𝑥, 𝜉 ′) = |𝜉 ′ | −1 𝜉22 if 𝑥2 = 0,
𝐵 (𝑥, 𝜉 ′) = |𝜉 ′ | −1 𝜉2 (𝜉2 − 𝜉1 exp (1/𝑥 2 )) if 𝑥2 < 0.

Remark 23.1.42 When dealing with the vector field (23.1.15) we have 𝐵 (𝑥, 𝜉 ′) =
Í𝑛−1 ′
𝑗=1 𝑏 𝑗 (𝑥) 𝜉 𝑗 ; Property (P) requires that 𝑥 𝑛 ↦→ 𝐵 (𝑥, 𝜉 ) not change sign on any
∗
null bicharacteristic curve of 𝑋 in 𝑇 Ω, i.e., on any straight-line interval (23.1.16).
Formula (23.1.17) provides the vector field version of (23.1.18): 𝐺 (𝑥, 𝜉 ′) = |𝒃 (𝑥)|,
𝐹 (𝑥 ′, 𝜉 ′) = ⟨𝜉 ′, 𝒗 (𝑥 ′)⟩.

Proposition 23.1.43 Suppose that 𝑃𝑚 = 𝐴 + 𝑖𝐵 is of principal type in a conic open

subset U of 𝑇 ∗ M\0 and that 𝑃𝑚 has Property (P) in U. Let L be a Nagano
leaf of (𝐻 𝐴, 𝐻 𝐵 ) in U (see Subsection 23.1.3). If there are homogeneous functions
𝑓 , 𝑔 ∈ C 𝜔 (U; R) such that L ∩ Char 𝑃 contains an integral curve 𝔠 of 𝐻 𝑓 𝐴−𝑔𝐵 in
U then L ⊂ Char 𝑃 and dim L ≤ 2. When dim L = 2 the pair of vectors (𝐻 𝐴, 𝐻 𝐵 )
define an orientation on L.

Proof It suffices to prove the claims microlocally: as above we can take U = 𝑈 × Γ,

𝑈 ⊂ R𝑛 open, Γ ⊂ R𝑛 \ {0} an open cone; let (𝑥 ◦ , 𝜉 ◦ ) ∈ (𝑈 × Γ) ∩Char ( 𝐴 + 𝑖𝐵). We
are allowed to contract 𝑈 about 𝑥 ◦ and Γ about the ray (𝑥◦ , 𝜌𝜉 ◦ ), 𝜌 > 0, as much as
needed. It is convenient to take 𝑈 = 𝑈 ′ × 𝑥 𝑛◦ − 𝑇, 𝑥 𝑛◦ + 𝑇 , 𝑈 ′ a neighborhood of 𝑥 ◦′
in R𝑛−1 , 𝑇 > 0. All the properties in the statement are symplectically invariant and
therefore there is no loss of generality in assuming that (23.1.7) holds with 𝐴 = 𝜉 𝑛 ,
𝐵 ∈ C 𝜔 (U; R); thus 𝜉 𝑛◦ = 0. The hypothesis (P) enables us to apply Proposition
23.1.39: 𝐵 (𝑥, 𝜉) = 𝐺 (𝑥, 𝜉 ′) 𝐹 (𝑥 ′, 𝜉 ′), 𝐺 ≥ 0, and

U ∩ Char 𝑃 = {(𝑥, 𝜉) ∈ 𝑈 × Γ; 𝜉 𝑛 = 𝐺 (𝑥, 𝜉 ′) 𝐹 (𝑥 ′, 𝜉 ′) = 0} .

We have
𝐻 𝐴 = 𝜕𝑥𝑛 , 𝐻 𝐵 = 𝐹 (𝑥 ′, 𝜉 ′) 𝐻𝐺 + 𝐺 (𝑥, 𝜉 ′) 𝐻𝐹 . (23.1.20)
23.1 Pseudodifferential Operators of Principal Type 961

On 𝔠 ⊂ L ∩ Char 𝑃 we have 𝐴 = 𝐵 = 0 and therefore

𝐻 𝑓 𝐴−𝑔𝐵 = 𝑓 𝐻 𝐴 − 𝑔𝐻 𝐵 .

On 𝔠, if 𝐹 . 0 then necessarily 𝐺 ≡ 0, which requires d𝐺 ≡ 0 in 𝔠 since 𝐺 ≥ 0

in U, implying 𝐻 𝐵 ≡ 0 and 𝐻 𝑓 𝐴−𝑔𝐵 = 𝑓 𝐻 𝐴 in 𝔠. We conclude that 𝑓 ≠ 0 and 𝜕𝑥𝑛
spans the tangent line at every point of 𝔠, meaning that 𝔠 is the interval

𝑰 (𝑥 ◦′, 𝜉 ◦′) = (𝑥, 𝜉) ∈ 𝑈 × Γ; 𝑥 ′ = 𝑥 ◦′, 𝑥 𝑛 − 𝑥 𝑛◦ < 𝑇, 𝜉 ′ = 𝜉 ◦′, 𝜉 𝑛 = 0 .

(23.1.21)

We have Ad 𝑘 𝐻 𝐴 𝐻 𝐵 = 𝜕𝑥𝑘𝑛 𝐺 𝐻𝐹 ≡ 0 in 𝔠, whatever 𝑘 ∈ Z+ . This means
that 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) is spanned along 𝔠 by 𝐻 𝐴 alone and, therefore, that 𝔠 is a null
bicharacteristic curve of 𝐴 + 𝑖𝐵 in U (Definition 23.1.15) and L = 𝔠.
Now suppose that 𝐹 (𝑥 ◦′, 𝜉 ◦′) = 0 and 𝐺 . 0 on 𝔠. In this case 𝐻 𝐵 = 𝐺 (𝑥, 𝜉 ′) 𝐻𝐹
and
𝜕
𝐻 𝑓 𝐴−𝑔𝐵 = 𝑓 + 𝑔𝐺 (𝑥, 𝜉 ′) 𝐻𝐹
𝜕𝑥 𝑛
on 𝔠. Since 𝐻 𝐴 𝐹 ≡ 0, [𝐻 𝐴, 𝐻𝐹 ] ≡ 0, we derive that 𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ) is spanned along
𝔠 by 𝐻 𝐴 and 𝐻𝐹 . If 𝐻𝐹 ≡ 0 in 𝔠 we have dim L = 1 and L = 𝔠; otherwise dim L = 2.
All integral curves of 𝐻 𝐴, i.e., intervals 𝑰 (𝑥 ◦′, 𝜉 ◦′), that intersect L are contained
in L. Suppose dim L = 2. On 𝑰 (𝑥 ◦′, 𝜉 ◦′) the vector field 𝐻𝐹 is constant, tangent to
L and 𝐻𝐹 ≠ 0; 𝐻 𝐵 = 𝐺𝐻𝐹 is transversal to 𝑰 (𝑥 ◦′, 𝜉 ◦′) at every point where 𝐺 ≠ 0;
since 𝐵 ≡ 0 on every integral curve of 𝐻 𝐵 that intersects 𝑰 (𝑥 ◦′, 𝜉 ◦′) we derive that
𝐵 ≡ 0 in an open subset of L and therefore in the whole of L. The 2-vector 𝐻 𝐴 ∧ 𝐻𝐹
never vanishes and defines an orientation on L. □

Remark 23.1.44 Example 23.1.36 shows that Hypothesis (P) does not preclude the
existence of Nagano leaves of (𝐻 𝐴, 𝐻 𝐵 ) that intersect Char 𝑃 and have dimension
≥ 3.

The properties of 𝑃𝑚 = 𝐴+𝑖𝐵 in Proposition 23.1.43 only describe the component

Char∞ 𝑃 in the partition (23.1.11) and cannot by themselves ensure the validity of
(P) as Char∞ 𝑃 might be empty and (P) may or may not be valid (Example 23.1.19).
We now complete the geometric characterization of Property (P), first by proving

Proposition 23.1.45 Suppose that 𝑃𝑚 = 𝐴 + 𝑖𝐵 is of principal type in a conic open

subset U of 𝑇 ∗ M\0. If 𝑃𝑚 has Property (P) in U then the pair of functions ( 𝐴, 𝐵)
is of finite even type (Definition 13.3.28) at every point of U ∩ Charfin 𝑃𝑚 .

Proof Once again it suffices to prove the claim microlocally; we reason in the same
set-up as in the proof of Proposition 23.1.43. To simplify the notation we carry
out a translation in 𝑥-space so as to have 𝑥 ◦ = 0. We take (0, 𝜉 ◦ ) ∈ (𝑈 × Γ) ∩
Charfin ( 𝐴 + 𝑖𝐵) with 𝑈 = 𝑈 ′ × (−𝑇, 𝑇), 𝑇 > 0, 𝑈 ′ a neighborhood of 0 in R𝑛−1 . Let
L be the Nagano leaf of (𝐻 𝐴, 𝐻 𝐵 ) through (0, 𝜉 ◦ ) and let the null bicharacteristic of
𝐴, 𝑰 (0, 𝜉 ◦′) [see (23.1.21)], be contained in L; 𝑰 (0, 𝜉 ◦′) ∩Char ( 𝐴 + 𝑖𝐵) is a discrete
set; by decreasing 𝑇 > 0 we can ensure that (0, 𝜉 ◦ ) is the sole zero of 𝐵 in 𝑰 (0, 𝜉 ◦′),
962 23 Analytic PDEs of Principal Type. Local Solvability

necessarily of even order 2𝜈 ≥ 2, meaning that 𝐻 2𝜈 ◦ ◦ 2ℓ ◦ ◦

𝐴 𝐵 (𝑥 , 𝜉 ) ≠ 0, 𝐻 𝐴 𝐵 (𝑥 , 𝜉 ) = 0
if ℓ < 2𝜈. After contracting 𝑈 ′ about 0 and Γ about the ray (0, 𝜌𝜉 ◦ ), 𝜌 > 0, and
decreasing 𝑇 > 0, we can apply the Weierstrass Preparation Theorem and write, in
𝑈 × Γ,
2𝜈−1
∑︁
𝐵 (𝑥, 𝜉 ′) = 𝐸 (𝑥, 𝜉 ′) ­𝑥 𝑛2𝜈 + 𝑏 𝑗 (𝑥 ′, 𝜉 ′) 𝑥 𝑛 ® ,
© 𝑗ª

« 𝑗=0 ¬
where the real-valued, analytic functions 𝐸 and 𝑏 𝑗 are homogeneous of degree 1 and
0 respectively, with 𝐸 nowhere zero and 𝑏 𝑗 (0, 𝜉 ◦′) = 0 for all 𝑗. Now suppose we
had
d 𝑥′ , 𝜉 ′ 𝐵 (0, 𝜉 ◦′) = d 𝑥′ , 𝜉 ′ 𝑏 0 (0, 𝜉 ◦′) ≠ 0.
There would be points (𝑥, 𝜉) arbitrarily close to (0, 𝜉 ◦ ) where 𝑏 0 (𝑥 ′, 𝜉 ′) < 0 and
where the function 𝑥 𝑛 ↦→ 𝐵 (𝑥 ′, 𝑥 𝑛 , 𝜉 ′) would change sign, contradicting (P). This
implies that 𝐻 𝐵 = 0 at (0, 𝜉 ◦ ) and therefore 𝐻 𝐵 𝐻 ℓ𝐴 𝐵 (𝑥 ◦ , 𝜉 ◦ ) = 0 whatever ℓ ∈ Z+ .
This means that the pair of functions ( 𝐴, 𝐵) is of type 2𝜈 at (0, 𝜉 ◦ ). □
Inspection of the proofs of Propositions 23.1.43 and 23.1.43 shows how to prove
the following

Theorem 23.1.46 Suppose 𝑃𝑚 = 𝐴 + 𝑖𝐵 is of principal type in a conic open subset

U of 𝑇 ∗ M\0. For 𝑃𝑚 to have Property (P) in U it is necessary and sufficient that
the following conditions be satisfied:
(i) if L is a null bicharacteristic leaf of 𝐴 + 𝑖𝐵 in U (Definition 23.1.15) then
dim L ≤ 2;
(ii) 𝐻 𝐴 ∧ 𝐻 𝐵 does not change sign in any (connected) null bicharacteristic surface
of 𝐴 + 𝑖𝐵 in U;
(iii) the system of functions ( 𝐴, 𝐵) is of finite even type at every point of
Charfin ( 𝐴 + 𝑖𝐵).

Corollary 23.1.47 Property (P) of 𝑃𝑚 in U is invariant under homogeneous sym-

plectomorphisms and multiplication of 𝑃𝑚 by a nowhere vanishing, homogeneous
function ℎ ∈ C 𝜔 (U).

Proof The invariance of (i), (ii), (iii) under homogeneous symplectic transformations
is true since they involve solely the Hamiltonian vector fields of 𝐴 and 𝐵. They are
also invariant under multiplication of 𝐴 + 𝑖𝐵 by a nowhere vanishing, homogeneous,
complex symbol ℎ = 𝑓 +𝑖𝑔 ∈ C 𝜔 (U). For (i) this follows from Proposition 23.1.17.
For (ii) it suffices to note that

𝐻 ℎ𝑃𝑚 ∧ 𝐻 ℎ𝑃𝑚 = |ℎ| 2 𝐻 𝑃𝑚 ∧ 𝐻 𝑃𝑚

at every point of U where 𝑃𝑚 = 0. For (iii) it follows from Proposition 13.3.31. □

Remark 23.1.48 If 𝐻 𝐴 ∧ 𝐻 𝐵 does not change sign in a two-dimensional Nagano

leaf L ∈ 𝔑𝔞𝔤 (U, 𝐴, 𝐵) it defines an orientation, i.e., a Riemann surface structure,
on L.
23.2 Local Solvability of Analytic PDEs of Principal Type 963

23.2 Local Solvability of Analytic PDEs of Principal Type

23.2.1 Basic definition. A functional-analytic condition

Throughout this section, unless specified otherwise 𝑃 will be a differential operator

of order 𝑚, with C 𝜔 complex coefficients, in an open subset Ω of R𝑛 or, more
generally, in a C 𝜔 manifold M (countable at infinity). The definitions and results in
this subsection are classical and analyticity is not needed, C ∞ suffices.
Definition 23.2.1 We shall say that 𝑃 is locally solvable in M if there is a covering
of M by open sets 𝑈 𝜄 in which the following is true:
(LS) Given 𝑓 ∈ Cc∞ (𝑈 𝜄 ) there is a 𝑢 ∈ D ′ (𝑈 𝜄 ) such that 𝑃𝑢 = 𝑓 in 𝑈 𝜄 .
We shall say that 𝑃 is locally solvable at a point ℘ ∈ M if 𝑃 is locally solvable
in a neighborhood of ℘.
The following statement is self-evident but important.
Proposition 23.2.2 If 𝑃 is locally solvable at a point ℘ ∈ M there is a neighborhood
𝑈 of ℘ such that 𝑃 is locally solvable at every point of 𝑈.
We are going to use the Sobolev spaces 𝐻c𝑠 (M) and 𝐻loc 𝑠 (M).
The fastest way
to define these spaces is to equip M with a Riemannian metric d𝑠2 ; let then Δ denote
the Laplace–Beltrami operator and d𝑉 the volume element. For every 𝑠 ∈ R we
can define a standard pseudodifferential operator (1 − Δ) 𝑠/2 of order 𝑠. We define
𝐻c0 (M) = 𝐿 c2 (M) and 𝐻c𝑠 (M) = 𝑇 𝑠 𝐿 c2 (M) where 𝑇 𝑠 is a properly supported
(Definition 2.3.6) representative of (1 − Δ) 𝑠/2 ; 𝐻loc
𝑠 (M)
is the space of distributions
∞
𝑢 in M such that 𝜒𝑢 ∈ 𝐻c (M) whatever 𝜒 ∈ Cc (M). These definitions are easily
𝑠

seen to be independent of the choice of the Riemannian metric (e.g., by using

partitions of unity and local coordinates charts) and of the representative 𝑇 𝑠 (two of
the latter differ by a smoothing operator). The spaces 𝐻c𝑠 (M) and 𝐻loc 𝑠 (M)
shall
carry their natural locally convex topologies; 𝐻loc 𝑠 (M) −𝑠
and 𝐻c (M) are naturally
isomorphic to the dual of each other.
If 𝐾 is a compact subset of M we denote by Cc∞ (𝐾) [resp., 𝐻c𝑠 (𝐾)] the subspace
of Cc∞ (M) [resp., 𝐻c𝑠 (M)] consisting of the functions (resp., distributions) with
support in 𝐾; E ′ (𝐾) is the spaces of all distributions in M with support in 𝐾; we
have
Ù∞ ∞
Ø
Cc∞ (𝐾) = 𝐻c𝑘 (𝐾) , E ′ (𝐾) = 𝐻c−𝑘 (𝐾) . (23.2.1)
𝑘=0 𝑘=0

We denote by ∥·∥ 𝑠 the norm in (𝐾) (it depends on the choice of the metric d𝑠2 );
𝐻c𝑠
the norms ∥·∥ 𝑠 define a Fréchet space structure on Cc∞ (𝐾).
The following functional-analytic condition, necessary and sufficient for solvabil-
ity, is a classical result of L. Hörmander (see, e.g., [Hörmander, 1963]). Let 𝑃 be a
differential operator with complex C ∞ coefficients in a C ∞ manifold M; we denote
by 𝑃⊤ the transpose of 𝑃 relative to the density d𝑉:
964 23 Analytic PDEs of Principal Type. Local Solvability
∫ ∫
𝑓 𝑃⊤ 𝑔d𝑉 = 𝑔𝑃 𝑓 d𝑉 (23.2.2)

for 𝑓 , 𝑔 ∈ Cc∞ (M).

Lemma 23.2.3 Let 𝐾 be a compact subset of M. Suppose that to every 𝑓 ∈ Cc∞ (𝐾)
there is an open set 𝑈 ⊂ M, 𝑈 ⊃ 𝐾, and a distribution 𝑢 ∈ D ′ (𝑈) such that
𝑃𝑢 = 𝑓 . Under this hypothesis there exist integers 𝑘 ◦ , 𝑘 ∈ Z+ and 𝐶 > 0 such that
∫
∀ 𝑓 , 𝑔 ∈ Cc∞ (𝐾) , 𝑓 𝑔𝑑𝑉 ≤ 𝐶 ∥ 𝑓 ∥ 𝑘◦ 𝑃⊤ 𝑔 𝑘 . (23.2.3)

Conversely, if (23.2.3) holds then there is a bounded linear operator 𝑮 : 𝐻 𝑘◦ (𝐾) −→

−𝑘 (M) such that 𝑃𝑮 𝑓 = 𝑓 in the interior Int 𝐾 for every 𝑓 ∈ 𝐻 𝑘◦ (𝐾).
𝐻loc

Proof The hypothesis implies that the map 𝑔 ↦→ 𝑃⊤ 𝑔 of Cc∞ (𝐾) into itself is
∫injective. Indeed, it implies that there is a 𝑢 ∈ D ′ (𝑈) such that 𝑃𝑢 = 𝑔¯ whence
|𝑔| 2 d𝑉 = ⟨𝑢, 𝑃⊤ 𝑔⟩. Let 𝑬 (𝑃, 𝐾) denote the space Cc∞ (𝐾) equipped with the
topology defined by the norms ∥𝑃⊤ 𝑔∥ 𝑘 (𝑘 ∈ Z+ ); it is metrizable (but possibly not
complete). The bilinear functional
∫
Cc∞ (𝐾) × 𝑬 (𝑃, 𝐾) ∋ ( 𝑓 , 𝑔) ↦→ 𝑓 𝑔d𝑉 ∈ C (23.2.4)
∫
is separately continuous: the continuity of 𝑓 ↦→ 𝑓 𝑔d𝑉 for 𝑔 fixed is a triviality. For
fixed 𝑓 we select 𝑢 ∈ D ′ (𝑈) such that 𝑃𝑢 = 𝑓 in ′
𝑈. Let 𝑈 ⊂⊂ 𝑈 be a neighborhood
of 𝐾; by (23.2.1) there is 𝑘 ∈ Z+ and 𝑢 ′ ∈ 𝐻c−𝑘 𝑈 ′ such that 𝑢 ′ = 𝑢 in 𝑈 ′, implying
∫
𝑓 𝑔𝑑𝑉 = |⟨𝑃𝑢 ′, 𝑔⟩| = 𝑢 ′, 𝑃⊤ 𝑔 ≤ 𝐶 𝑃⊤ 𝑔 𝑘
.

The validity of (23.2.3), i.e., the continuity of (23.2.4), then follows from the Banach–
Steinhaus theorem (see e.g., [Treves, 1967], Theorem 34.1).
Now suppose that (23.2.3) holds for a given 𝑓 ∈ Cc∞ (𝐾). This means ∫ that, on
the vector subspace 𝑃⊤ Cc∞ (𝐾), the linear functional 𝐺 𝑓 : 𝑃⊤ 𝑔 ↦→ 𝑓 𝑔d𝑉 is
continuous with respect to the norm ∥·∥ 𝑘 ; this is therefore true when said functional
is extended to the whole space 𝐻c𝑘 (M) by the Hahn–Banach Theorem. By letting
𝑓 vary this defines a linear map 𝑮 : Cc∞ (𝐾) −→ 𝐻loc −𝑘 (M). It follows from (23.2.3)

that 𝑮 is continuous with respect to the norm ∥·∥ 𝑘◦ and therefore extends to 𝐻 𝑘◦ (𝐾)
as a continuous linear map 𝑮 : 𝐻 𝑘◦ (𝐾) −→ 𝐻loc −𝑘 (M). By continuity we have

⟨𝑮 𝑓 , 𝑃 𝑔⟩ = ⟨ 𝑓 , 𝑔⟩ whatever 𝑓 ∈ 𝐻 (𝐾), 𝑔 ∈ Cc∞ (Int 𝐾), whence 𝑃𝑮 𝑓 = 𝑓 in

⊤ 𝑘 ◦

Int 𝐾. □

Remark 23.2.4 For a much more elaborated version of Lemma 23.2.3 we refer the
reader to [Hörmander, 1983, IV], Lemma 26.4.5.
23.2 Local Solvability of Analytic PDEs of Principal Type 965

To prove the necessity of Condition (P) for the local solvability of the equation
𝑃𝑢 = 𝑓 we shall determine a sequence of approximate solutions of the homogeneous
equation 𝑃⊤ 𝑔 = 0, more precisely C ∞ functions 𝑔 (𝑥, 𝜆) ∈ Cc∞ (𝐾) depending on
a parameter 𝜆 > 0 such that 𝑃⊤ 𝑔 → 0 as 𝜆 ↗ +∞ whereas the left-hand side in
(23.2.3) remains bounded away from zero for a suitable choice of 𝑓 . To do this we
shall use a special complex version of the classical KWB method; we devote the next
subsection to its description.

23.2.2 Formal series solutions and approximate solutions of a linear

PDE of principal type

Unless specified otherwise 𝑃 = 𝑃 (𝑥, D) shall denote a (linear) differential operator

with C 𝜔 coefficients, of order 𝑚 ∈ Z+ , 𝑚 ≥ 1, in a domain Ω in R𝑛 ; 𝑃𝑚 (𝑥, 𝜉)
shall denote its principal symbol. We assume that the coefficients of 𝑃 (𝑥, D) extend
holomorphically to a domain ΩC in C𝑛 such that ΩC ∩ R𝑛 = Ω. Here we shall assume
that 𝑃 is of principal type in the strong sense, meaning that 𝜕 𝜉 𝑃𝑚 (𝑥, 𝜉) ≠ 0 for all
(𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}). The analysis will take place in a neighborhood 𝑈 of a point
𝑥 ◦ ∈ Ω. We select a domain 𝑈 C ⊂ ΩC in C𝑛 such that 𝑈 C ∩ R𝑛 = 𝑈; throughout the
reasoning we will be allowed to contract 𝑈 C about 0 as needed.
We shall reason under the hypothesis that the eikonal equation

𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) = 0 (23.2.5)

such that 𝜕𝑧 𝜑 (0) = 𝜉 ◦ ∈ R𝑛 \ {0}.

admits a solution 𝜑 ∈ O 𝑈 C

Remark 23.2.5 The solution 𝜑 ∈ O 𝑈 C is not unique because the condition
𝜕𝑧 𝜑 (0) = 𝜉 ◦ does not completely determine 𝜑. In the applications, the values
of 𝜑 on a complex hypersurface in 𝑈 C noncharacteristic with respect to 𝑃 at 𝑥 ◦
(Definition 1.3.3) will be specified, implying its uniqueness.

Under the hypotheses above we seek a formal analytic series (Definition 19.1.1)
∞
∑︁
𝑎 (𝑧, 𝜆) = 𝜆− 𝑗 𝑎 𝑗 (𝑧) ∈ Aform 𝑈 C ;
𝑗=0

for the convenience of the reader we recall that 𝑎 𝑗 ∈ O 𝑈 C for all 𝑗 and the
condition

(FA) To every compact subset 𝐾 of 𝑈 C there is a 𝐶𝐾 > 0 such that

𝑗+1
∀ 𝑗 ∈ Z+ , max 𝑎 𝑗 ≤ 𝐶𝐾 𝑗!. (23.2.6)
𝐾
966 23 Analytic PDEs of Principal Type. Local Solvability

We require that 𝑎 (𝑧, 𝜆) solves the (formal) equation

𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 (𝑧) 𝑎 (𝑧, 𝜆) (23.2.7)
∞
∑︁
= 𝜆− 𝑗 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 (𝑧) 𝑎 𝑗 (𝑧) = 0
𝑗=0

under the Cauchy conditions

𝑎 0 | Σ = 1, 𝑎 𝑗 Σ
= 0 if 𝑗 ≥ 1, (23.2.8)

where Σ is a complex-analytic hypersurface in 𝑈 C noncharacteristic with respect to

𝑃 at 𝑥 ◦ .
The generalized Leibniz rule (1.1.4) implies, for every 𝑓 ∈ C ∞ (Ω),
∑︁ 1
e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑓 = e−𝑖𝜆𝜑 𝜕 𝜉𝛼 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 D𝑧𝛼 𝑓 (23.2.9)
𝛼∈Z𝑛
𝛼!
+
𝑚−1
∑︁ ∑︁ 1 𝑘
𝛽

= 𝜆 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑓 ,
𝑘=0 | 𝛼 | ≤𝑚−𝑘
𝛼!

where 𝑝 𝑘, 𝛼 (𝑧, 𝜁1 , ..., 𝜁ℓ ) is a homogeneous polynomial of degree 𝑘 with respect to

𝛽
𝜁1 , ..., 𝜁ℓ , whose coefficients are C-linear combinations of those of 𝑃; 𝜕𝑧 𝜑 stands
𝛽1 𝛽ℓ
for varying sets of ℓ partial derivatives 𝜕𝑧 𝜑, ..., 𝜕𝑧 𝜑, |𝛽1 | + · · · + |𝛽ℓ | ≤ 𝑚 − |𝛼|. A
straightforward calculation shows that

𝛽
𝑝 𝑚,0 𝑧, 𝜕𝑧 𝜑 = 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) , (23.2.10)
𝑛
∑︁
𝛽
1 𝜕𝑃𝑚 ∑︁ 𝜕
𝑝 𝑚−1, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 = √ (𝑧, 𝜕𝑧 𝜑) , (23.2.11)
| 𝛼 |=1 −1 𝑗=1 𝜕𝜉 𝑗 𝜕𝑧 𝑗
𝑛

𝛽
1 ∑︁ 𝜕 2 𝑃𝑚 𝜕2 𝜑
𝑝 𝑚−1,0 𝑧, 𝜕𝑧 𝜑 = 𝑃𝑚−1 (𝑧, 𝜕𝑧 𝜑) + √ (𝑧, 𝜕𝑧 𝜑) .
2 −1 𝑗,𝑘=1 𝜕𝜉 𝑗 𝜕𝜉 𝑘 𝜕𝑧 𝑗 𝜕𝑧 𝑘
(23.2.12)

Thus (23.2.5) and (23.2.7) imply

∞ 𝑚−1
∑︁ ∑︁ ∑︁ 1 𝑘− 𝑗
𝛽

𝜆 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗 = 0. (23.2.13)
𝑗=0 𝑘=0 | 𝛼 | ≤𝑚−𝑘
𝛼!

By equating to zero the coefficients of 𝜆− 𝑗 in the left-hand side we get the standard
transport equations:
23.2 Local Solvability of Analytic PDEs of Principal Type 967
∑︁
𝛽
𝜕 𝜉𝛼 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) D𝑧𝛼 𝑎 0 + 𝑝 𝑚−1,0 𝑧, 𝜕𝑧 𝜑 𝑎 0 = 0 (23.2.14)
| 𝛼 |=1

and for 𝑗 ≥ 1,
∑︁
𝛽
𝜕 𝜉𝛼 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) D𝑧𝛼 𝑎 𝑗 + 𝑝 𝑚−1,0 𝑧, 𝜕𝑧 𝜑 𝑎 𝑗 (23.2.15)
| 𝛼 |=1
𝑗 𝑚−1
∑︁ ∑︁ ∑︁ 1
𝛽

=− 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗−ℓ .
ℓ=1 𝑘=0 | 𝛼 |=ℓ+𝑘−𝑚+1
𝛼!

Let 𝐾 be an arbitrary compact subset of a domain ΩC in C𝑛 and let 𝑟 𝑗 ∈ (0, 1)

( 𝑗 = 1, ..., 𝑛) be such that
n o
𝐾𝑟 = 𝑧 ∈ C𝑛 ; ∃𝑧∗ ∈ 𝐾, 𝑧 𝑗 − 𝑧 ∗𝑗 ≤ 𝑟 𝑗 , 𝑗 = 1, ..., 𝑛

is a compact subset of ΩC . We will avail ourselves of the following lemma, where

we use the notation

𝑁 𝑠 ( 𝑓 ) = max | 𝑓 | , 𝑓 ∈ O ΩC , 0 ≤ 𝑠 ≤ 1. (23.2.16)
𝐾𝑠𝑟

Lemma 23.2.6 Let 𝑓 ∈ O ΩC and suppose there are 𝑀 > 0, 𝑘 ∈ Z+ \ {0} such
that e 𝑘
𝑁𝑠 ( 𝑓 ) ≤ 𝑀 (23.2.17)
1−𝑠
for every 𝑠, 0 ≤ 𝑠 < 1. Then, if 𝛼 ∈ Z+𝑛 we have
𝑀 e 𝑘+| 𝛼 |
𝑁 𝑠 𝜕𝑧𝑎 𝑓 ≤ 𝛼 (𝑘 + 1) · · · (𝑘 + |𝛼|) . (23.2.18)
𝑟 1−𝑠

In (23.2.18) 𝑟 𝛼 = 𝑟 1𝛼1 · · · 𝑟 𝑛𝛼𝑛 .

Proof If 0 < 𝜀 < 1 the Cauchy inequalities imply

𝜕𝑓 1
𝑁𝑠 ≤ 𝑁 𝑠+𝜀 ( 𝑓 ) , 𝑗 = 1, ..., 𝑛.
𝜕𝑧 𝑗 𝜀𝑟 𝑗

We derive from this estimate and from (23.2.17), for 𝜀 < 1 − 𝑠,

𝜕𝑓 𝑀 e 𝑘
𝑁𝑠 ≤ .
𝜕𝑧 𝑗 𝜀𝑟 𝑗 1 − 𝑠 − 𝜀
968 23 Analytic PDEs of Principal Type. Local Solvability

1−𝑠
If we select 𝜀 = 1+𝑘 we get
𝑘
𝜕𝑓 𝑀 1 e 𝑘+1
𝑁𝑠 ≤ (𝑘 + 1) 1 +
𝜕𝑧 𝑗 e𝑟 𝑗 𝑘 1−𝑠
𝑀 e 𝑘+1
≤ (𝑘 + 1) .
𝑟𝑗 1−𝑠

Iterating this result or using induction on |𝛼| yields directly (23.2.18). □

23.2.7 Assume that to every 𝑔 ∈ O 𝑈 there is a unique solution
Proposition C

𝑓 ∈ O 𝑈 of the Cauchy problem

C

∑︁
𝛽
𝜕 𝜉𝛼 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) D𝑧𝛼 𝑓 + 𝑝 𝑚−1,0 𝑧, 𝜕𝑧 𝜑 𝑓 = 𝑔, 𝑓 | Σ = 0 or 1.
| 𝛼 |=1

Then there is a unique sequence 𝑎 𝑗 𝑗 ∈Z+
, 𝑎 𝑗 ∈ O 𝑈 C , solution of (23.2.14)–
(23.2.15) satisfying (23.2.8). Moreover, possibly after a contraction of 𝑈 C about 0,
the functions 𝑎 𝑗 satisfy (FA) (Definition 1.3.3).

Proof The existence and uniqueness of the solutions 𝑎 𝑗 ∈ O 𝑈 C is a direct conse-
quence of the hypothesis, induction on 𝑗 ∈ Z+ and the finiteness of the sums in the
right-hand side of (23.2.15).
We take 𝑈 C sufficiently small that 𝑎 0 (𝑧) ≠ 0 and 𝜕𝑃 𝜕 𝜉ℓ (𝑧, 𝜕𝑧 𝜑) ≠ 0 for some ℓ
𝑚

and all 𝑧 ∈ 𝑈 C . [Recall that 𝜕𝑧 𝜑 (𝑧) = 𝜉 ◦ + 𝑂 (|𝑧 − 𝑥 ◦ |) by (24.3.10).] We are going

to apply Lemma 23.2.6 to a geometrically simple compact subset 𝐾 of 𝑈 C whose
interior contains the origin, and prove that, for 𝐶𝐾 > 0 sufficiently large,
e 𝑗
𝑗+1
∀ 𝑗 ∈ Z+ , ∀𝑠 ∈ (0, 1) , 𝑁 𝑠 𝑎 𝑗 ≤ 𝐶𝐾 𝑗! . (23.2.19)
1−𝑠
This will prove, by fixing 𝑠 and redefining 𝐶𝐾 , that the 𝑎 𝑗 satisfy (FA).
First of all, we require 𝐶𝐾 ≥ 𝑁1 (𝑎 0 ). Putting 𝑎 𝑗 = 𝑎 0 𝑏 𝑗 in (23.2.15) yields
∑︁
𝜕 𝜉𝛼 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) D𝑧𝛼 𝑎 𝑗
| 𝛼 |=1
𝑗 𝑚−1
∑︁ ∑︁ ∑︁ ∑︁ 1
−𝑎 −1
𝛽
= 0 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 0 𝑎 𝑗−ℓ
ℓ=1 𝑘=0 | 𝛼 |=ℓ+𝑘−𝑚+1
𝛼!
𝑗 𝑚−1
∑︁ ∑︁ ∑︁ ∑︁ 1
= −𝑎 −1 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼− 𝛼˜ 𝑎 0 D𝑧𝛼˜ 𝑎 𝑗−ℓ .
𝛽
0
ℓ=1 𝑘=0 | 𝛼 |=ℓ+𝑘−𝑚+1 𝛼˜ ⪯ 𝛼
˜ (𝛼 − 𝛼)!
𝛼! ˜

𝛽
Since the 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 are finitely many there is a constant 𝑀◦ > 0 such that

e 𝑗−ℓ
𝑗−ℓ+1
𝑁 𝑠 𝑏 𝑗−ℓ ≤ 𝐶𝐾 ( 𝑗 − ℓ)! (23.2.20)
1−𝑠
23.2 Local Solvability of Analytic PDEs of Principal Type 969

for all relevant 𝑘, 𝛼, whence

© ∑︁ 𝛼
𝜕 𝜉 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) 𝜕𝑧𝛼 𝑎 𝑗 ® (23.2.21)
ª
𝑁𝑠 ­
« | 𝛼 |=1 ¬
𝑗 𝑚−1
∑︁ ∑︁ ∑︁ ∑︁ 1
≤ 𝑀◦ 𝑁1 𝑎 −1
0 𝑁 𝑠 𝜕𝑧
𝛼− 𝛼˜
𝑎 0 𝑁 𝑠 𝜕𝑧
𝛼˜
𝑎 𝑗−ℓ .
𝛼˜ ⪯ 𝛼
˜ (𝛼 − 𝛼)!
𝛼! ˜
ℓ=1 𝑘=0 | 𝛼 |=ℓ+𝑘−𝑚+1

Induction on 𝑗 ≥ 1 and (23.2.19) yields, for 1 ≤ ℓ ≤ 𝑗,

e 𝑗−ℓ
𝑗−ℓ+1
𝑁 𝑠 𝑎 𝑗−ℓ ≤ 𝐶𝐾 ( 𝑗 − ℓ)!
1−𝑠
whence, by Lemma 23.2.6, for every 𝛼 ∈ Z+𝑛 ,
e 𝑗−ℓ+| 𝛼 |
𝑗−ℓ+1
𝑁 𝑠 𝜕𝑧𝑎 𝑎 𝑗−ℓ ≤ 𝐶𝐾 ( 𝑗 − ℓ + |𝛼|)! .
1−𝑠
Putting this into (23.2.21) yields

© ∑︁ 𝛼
𝜕 𝜉 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) 𝜕𝑧𝛼 𝑎 𝑗 ® ≤
ª
𝑁𝑠 ­
« | 𝛼 |=1 ¬
𝑗 𝑚−1
∑︁ ∑︁ ∑︁ ∑︁ |𝛼 − 𝛼|! ˜ e 𝑗−ℓ+| 𝛼 |
˜ ( 𝑗 − ℓ + | 𝛼|)!
2
𝑁1 𝑎 −1
𝑗−ℓ
𝑀◦ 𝐶𝐾 0 𝐶𝐾
𝛼˜ ⪯ 𝛼
˜ (𝛼 − 𝛼)!
𝛼! ˜ 1−𝑠
ℓ=1 𝑘=0 | 𝛼 |=ℓ+𝑘−𝑚+1
𝑗 ∑︁ ∑︁ | 𝛼|!
∑︁ ˜ ( 𝑗 − | 𝛼|)!
˜ e 𝑗
≤ 𝑚𝑀◦ 𝐶𝐾 𝑁1 𝑎 −1 1−ℓ
𝑗+1
0 𝐶𝐾 .
𝛼˜ ⪯ 𝛼
˜ (𝛼 − 𝛼)!
𝛼! ˜ 1−𝑠
ℓ=1 | 𝛼 |=ℓ

Assuming 𝐶𝐾 ≥ 1 we have
𝑗
1 ∑︁ 1−ℓ ∑︁ ∑︁ | 𝛼|!
˜ ( 𝑗 − | 𝛼|)!
˜
𝐶
𝑗! ℓ=1 𝐾 𝛼˜ ⪯ 𝛼
˜
𝛼! (𝛼 − ˜
𝛼)!
| 𝛼 |=ℓ
𝑗
∑︁
1−ℓ
∑︁ ∑︁ 1 ( 𝑗 − | 𝛼|)!
˜ | 𝛼|!
˜
= 𝐶𝐾
(𝛼 − 𝛼)!
˜ 𝛼!˜ 𝑗!
ℓ=1 | 𝛼 |=ℓ 𝛼˜ ⪯ 𝛼
𝑗 ∑︁ 1 ∑︁
∑︁
1−ℓ 𝛼!
≤ 𝐶𝐾
𝛼! 𝛼˜ ⪯ 𝛼 (𝛼 − 𝛼)!
˜ 𝛼!˜
ℓ=1 | 𝛼 |=ℓ
𝑗
∑︁ ∑︁ 1
1−ℓ
≤ 𝐶𝐾 2ℓ ≤ e2𝑛 .
𝛼!
ℓ=1 | 𝛼 |=ℓ
970 23 Analytic PDEs of Principal Type. Local Solvability

There is a holomorphic Í change of the variables 𝑧1 , ..., 𝑧 𝑛 that transforms the

holomorphic vector field | 𝛼 |=1 𝜕 𝜉𝛼 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) 𝜕𝑧𝛼 (possibly after contraction of 𝑈 C
about 0) into 𝜕𝑧𝜕1 . We can take 𝑈 C to be a polydisk

Δ𝜌(𝑛) = 𝑧 ∈ C𝑛 ; 𝑧 𝑗 < 𝜌 𝑗 , 𝑗 = 1, ..., 𝑛

and 𝐾 the closure of a similar polydisk; then the same is true of 𝐾1 . In the new
coordinates,
e 𝑗
𝜕𝑎 𝑗
≤ (𝑚 − 1) 𝑀◦ e2𝑛 𝑁1 𝑎 −1
𝑗+1
𝑁𝑠 0 𝐶 𝑗! .
𝜕𝑧1 𝐾 1−𝑠

We derive
e 𝑗
𝑁 𝑠 𝑎 𝑗 ≤ 𝜌1 (𝑚 − 1) 𝑀◦ e2𝑛 𝑁1 𝑎 −1
𝑗+1
0 𝐶𝐾 𝑗! .
1−𝑠
Requiring 𝜌1 (𝑚 − 1) 𝑀◦ e2𝑛 𝑁1 𝑎 −1

0 ≤ 1 proves (23.2.19). □

Once the 𝑎 𝑗 are determined we form a series

∞
∑︁
𝑎♭ (𝑧, 𝜆) = 𝑎 0 (𝑧) + 𝜆− 𝑗 𝜒 (𝜆/𝑅 𝑗) 𝑎 𝑗 (𝑧) (23.2.22)
𝑗=1

where 𝜒 ∈ C ∞ (R), 𝜒 (𝜆) = 0 for 𝜆 < 1, 𝜒 (𝜆) = 1 for 𝜆 > 2, and 𝑅 > 0.

Proposition 23.2.8 Let 𝑈 C be as in Proposition 23.2.7. To every compact subset

𝐾 of 𝑈 C there are positive numbers 𝑅, 𝐶 such that the series (23.2.22) converges
absolutely uniformly on 𝐾 and

∀𝜆 ≥ 1, max 𝑎♭ (𝑧, 𝜆) ≤ 𝐶. (23.2.23)

𝑧 ∈𝐾

Moreover, given an arbitrary pair of positive integers 𝑁1 , 𝑁2 , the constants 𝑅 and

𝐶 can be chosen sufficiently large that
∑︁
max e−𝑖𝜆𝜑 (𝑧) 𝜕𝑧𝛼 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 (𝑧) 𝑎♭ (𝑧, 𝜆) ≤ 𝐶𝜆−𝑁2 . (23.2.24)
𝐾
| 𝛼 | ≤ 𝑁1

Proof From (23.2.6) and the Stirling inequality we derive

𝜆− 𝑗 𝜒 (𝜆/𝑅 𝑗) max 𝑎 𝑗 ≤ 𝐶𝐾 (𝐶𝐾 /𝑅 𝑗) 𝑗 𝑗!

𝐾
≤ 𝐶◦ 𝐶𝐾 (𝐶𝐾 /𝑅) 𝑗

where 𝐶◦ is a universal constant. By selecting 𝑅 ≥ 𝐶𝐾 /𝜀 (𝜀 > 0) we get

𝜆− 𝑗 𝜒 (𝜆/𝑅 𝑗) max 𝑎 𝑗 ≤ 𝐶◦ 𝐶𝐾 𝜀 𝑗 .
𝐾
23.2 Local Solvability of Analytic PDEs of Principal Type 971

It follows that the series (23.2.22) converges uniformly in 𝐾 and that (23.2.23) holds.
It is a simple exercise to derive from (23.2.14)–(23.2.15) and the Cauchy inequalities
that (23.2.24) holds provided 𝜀 is sufficiently small. □

23.2.3 Local solvability of analytic differential operators of principal

type. Necessity of Condition (P)

As before let 𝑃 be a differential operator with C 𝜔 coefficients, of order 𝑚 ∈ Z+ ,

𝑚 ≥ 1, in the C 𝜔 manifold M; 𝑃𝑚 will be its principal symbol; we assume that
d 𝜉 𝑃𝑚 ≠ 0 at every point of 𝑇 ∗ M\0. We shall reason under the (NonP) hypothesis:
Property (P) (Definition 23.1.37) does not hold at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃.
We need to take a look at the microlocal geometry resulting from (NonP) in a
suitably “acute” conic open subset U of 𝑇 ∗ M\0 containing (𝑥 ◦ , 𝜉 ◦ ). We carry out
a local symplectic transformation bringing us to the situation in which d 𝜉 𝑃𝑚 ≠ 0 at
every point of U. Using local coordinates in a neighborhood of 𝑥 ◦ we replace M
by an open set Ω in R𝑛 . Possibly after a further symplectic transformation we can
assume that, in U,
√
𝑃𝑚 (𝑥, 𝜉) = 𝐸 ◦ (𝑥, 𝜉) 𝜉 𝑛 − −1𝐵 (𝑥, 𝜉 ′) (23.2.25)

with 𝐸 ◦ ∈ C 𝜔 (U; C) nowhere vanishing, homogeneous of degree 𝑚 − 1, and

𝐵 ∈ C 𝜔 (U; R) homogeneous of degree 1 [see Proposition 23.1.10 and (23.1.7);
𝜉 ′ = (𝜉1 , ..., 𝜉 𝑛−1 )]; thus 𝜉 𝑛◦ = 0. We shall take U = 𝑈 × Γ, 𝑈 a neighborhood of
𝑥 ◦ in R𝑛 and Γ = {𝜉 ∈ R𝑛 \ {0} ; 𝜉 ′ ∈ Γ ′, |𝜉 𝑛 | < 𝜀 |𝜉 ′ |}, Γ ′ a convex open cone in
R𝑛−1 \ {0}, 𝜀 > 0. We will use the notation 𝑥 ′ = (𝑥 1 , ..., 𝑥 𝑛−1 ). Our hypothesis will
be
(NonP) ( 𝑥 ◦ , 𝜉 ◦ ) The function 𝑥 𝑛 ↦→ 𝐵 (𝑥 ◦′, 𝑥 𝑛 , 𝜉 ◦′) changes sign at 𝑥 𝑛◦ .
To simplify the notation we shall assume that 𝑥 ◦ = 0 and 𝑈 = 𝑈 ′ × (−𝜏, 𝜏), 𝑈 ′ ⊂
R𝑛−1 open, 0 ∈ 𝑈 ′, 𝜏 > 0. By the Weierstrass Preparation Theorem (NonP) ( 𝑥 ◦ , 𝜉 ◦ )
implies that there is an integer 𝜈 ≥ 1 such that

∑︁ 2𝜈
′ ′
𝐵 (𝑥, 𝜉 ) = 𝐸 (𝑥, 𝜉 ) ­2𝜈𝑥 𝑛2𝜈−1 + 𝑎 𝑗 (𝑥 ′, 𝜉 ′) 𝑥 𝑛 ®
© 𝑗ª

« 𝑗=0 ¬
in 𝑈 × Γ, with 𝑎 𝑗 ∈ C 𝜔 (𝑈 ′ × Γ ′) homogeneous of degree 0, 𝑎 𝑗 (0, 𝜉 ◦′) = 0 for
all 𝑗 and 𝐸 (𝑥, 𝜉 ′) ∈ C 𝜔 (𝑈 × Γ ′; R) homogeneous of degree 1 and nowhere zero.
Given that the regular part of Char 𝑃 is open and dense in Char 𝑃 (Proposition
14.2.8) there are points (𝑥 ′, 𝜉 ′) arbitrarily close to (0, 𝜉 ◦′) where 𝑥 𝑛 ↦→ 𝐵 (𝑥 ′, 𝑥 𝑛 , 𝜉 ′)
change sign across an analytic hypersurface Σ on which 𝐵 (𝑥, 𝜉 ′) vanishes to some
odd order 2𝑞 − 1 (1 ≤ 𝑞 ≤ 𝐵) uniformly; Σ can be defined locally by an equation
𝑥 𝑛 = 𝑓 (𝑥 ′, 𝜉 ′) with 𝑓 ∈ C 𝜔 (𝑈 ′; R). Noting that {𝜉 𝑛 , 𝑥 𝑛 − 𝑓 (𝑥 ′, 𝜉 ′)} = 1 we apply
972 23 Analytic PDEs of Principal Type. Local Solvability

the Darboux Theorem 13.3.20 to select Darboux coordinates 𝑥 𝑗 , 𝜉 𝑗 (1 ≤ 𝑗 ≤ 𝑛)

with 𝜉 𝑛 unchanged, 𝑥 𝑛 replaced by 𝑥 𝑛 + 𝑓 (𝑥 ′, 𝜉 ′) so that Σ is defined, in the new
coordinates, by 𝑥 𝑛 = 0. Possibly after a linear transformation in 𝑥 ′-space we assume
that 𝜉 ◦ = (1, 0, ..., 0). We move our “center” to a point of Σ or rather, to simplify
matters further, we start from the assumption that said point is (0, 𝜉 ◦ ) and

𝐵 (𝑥, 𝜉 ′) = 2𝜈𝐸 (𝑥, 𝜉 ′) 𝑥 𝑛2𝜈−1 (23.2.27)

in 𝑈 × Γ. (That it suffices to contradict local solvability after such a transfer of the

central point follows from Proposition 23.2.2.)
By (23.2.25) and the displacement of the central point carried out above, the
eikonal equation (23.2.5) is equivalent to

𝜕𝑥𝑛 𝜑 − 2𝜈𝑥 𝑛2𝜈−1 𝐸 (𝑥, D 𝑥′ 𝜑) = 0. (23.2.28)

We require 𝜑| 𝑥𝑛 =0 = 𝑥1 + 𝑖 |𝑥 ′ | 2 . By the Cauchy–Kovalevskaya Theorem the solution

𝜑 is unique and

𝜑 (𝑥) = 𝑥1 + 𝑖 |𝑥 ′ | 2 + 𝑖𝐸 0 (𝑥 ′) 𝑥 𝑛2𝜈 + 𝑂 |𝑥 ′ | 3 + |𝑥 𝑛 | 2𝜈+1 , (23.2.29)

where 𝐸 0 (𝑥 ′) = 𝐸 (𝑥 ′, 0, 𝜉 ◦ ), recalling that 𝜉 ◦ = (1, 0, ..., 0); we derive, for 𝑥 ′ ∈ 𝑈 ′,

Im 𝜑 (𝑥) = |𝑥 ′ | 2 + 𝐸 0 (𝑥 ′) 𝑥 𝑛2𝜈 + 𝑂 |𝑥 ′ | 3 + |𝑥 𝑛 | 2𝜈+1 . (23.2.30)

At this point we make our hypothesis more precise: we assume that 𝐸 0 (𝑥 ′) > 0 for
𝑥 ′ ∈ 𝑈 ′; this is equivalent to the following:
(Ψ∗ ) (0, 𝜉 ◦ ) The function 𝑥 𝑛 ↦→ 𝐵 (0, 𝑥 𝑛 , 𝜉 ′◦ ) changes sign at 0 from − to +.

Remark 23.2.9 If the function 𝑥 𝑛 ↦→ 𝐵 (0, 𝑥 𝑛 , 𝜉 ′◦ ) changes sign at 0 from + to − then

(Ψ∗ ) (0,− 𝜉 ◦ ) holds true and we replace (1, 0, ..., 0) by (−1, 0, ..., 0) in the ongoing
argument.

We are going to use the formal series 𝑎 (𝑥, 𝜆) = ∞ −𝑗

Í
𝑗=0 𝜆 𝑎 𝑗 (𝑥) in Proposition
23.2.7, the solution of the transport equations (23.2.14)–(23.2.15), but with two
modifications:
(1) we replace 𝑃 (𝑥, D 𝑥 ) by its transpose 𝑃 (𝑥, D 𝑥 ) ⊤ ;
(2) the condition (23.2.8) is replaced by the following Cauchy conditions:

𝑎 0 | 𝑥𝑛 =0 = 1, 𝑎 𝑗 𝑥𝑛 =0
= 0. (23.2.31)

Back to Lemma 23.2.3 we select 𝑔 (𝑥, 𝜆) = 𝜒 (𝑥) e𝑖𝜆𝜑 ( 𝑥) 𝑎♭ (𝑥, 𝜆) with 𝑎♭ as given
∞

in (23.2.22) and 𝜒 ∈ Cc 𝔅𝜌 where 𝔅𝜌 = {𝑥 ∈ R𝑛 ; |𝑥| ≤ 𝜌} ⊂⊂ 𝑈 and 𝜒 ≡ 1 in
𝔅𝜌′ , 𝜌 ′ < 𝜌. We have 𝑃⊤ 𝑔 = 𝜒𝑃⊤ e𝑖𝜆𝜑 𝑎 + ℎe𝑖𝜆𝜑 , ℎ ∈ Cc∞ 𝔅𝜌 , ℎ ≡ 0 in 𝔅𝜌′ .
From (23.2.24) we derive
23.2 Local Solvability of Analytic PDEs of Principal Type 973
∑︁
𝜒 (𝑥) 𝜕𝑥𝛼 𝑃 (𝑥, D 𝑥 ) ⊤ e𝑖𝜆𝜑 ( 𝑥) 𝑎♭ (𝑥, 𝜆) ≤ 𝐶𝜆−𝑁2 e−𝜆 Im 𝜑 ( 𝑥) .
| 𝛼 | ≤ 𝑁1

We also have
∑︁
𝜕𝑥𝛼 ℎ (𝑥) e𝑖𝜆𝜑 ( 𝑥) ≤ 𝐶 ′𝜆 𝑁1 ′ max e−𝜆 Im 𝜑 ( 𝑥) .
𝜌 ≤ | 𝑥 | ≤𝜌
| 𝛼 | ≤ 𝑁1

It follows from (23.2.30) that Im 𝜑 (𝑥) ≥ 0 for all 𝑥 ∈ 𝔅𝜌 (assuming that 𝜌 is

sufficiently small) and ′ min Im 𝜑 (𝑥) ≥ 𝑐 > 0. Putting all this together we obtain
𝜌 ≤ |𝑥 | ≤𝜌
∑︁
𝜕𝑥𝛼 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑔 (𝑥, 𝜆) ≤ 𝐶𝜆−𝑁2 + 𝐶 ′𝜆 𝑁1 e−𝑐𝜆 . (23.2.32)
| 𝛼 | ≤ 𝑁1

Lastly we select 𝑓 (𝑥) = 𝑓1 (𝜆𝑥) with 𝑓1 ∈ Cc∞ 𝔅𝜌 ; note that supp 𝑓 ⊂ 𝔅𝜌/𝜆

and 𝑛
∥ 𝑓 ∥ 𝑘 ◦ ≤ 𝜆 𝑘 ◦ − 2 ∥ 𝑓1 ∥ 𝑘 ◦ . (23.2.33)
Now we have
∫ ∫
−1 𝑥
𝑓 𝑔d𝑥 = 𝜆−𝑛 e𝑖𝜆𝜑 ( 𝜆 ) 𝑓1 (𝑥) 𝑎♭ 𝜆−1 𝑥, 𝜆 𝜒 𝜆−1 𝑥 d𝑥.

On the one hand, (23.2.30), (23.2.31) and the fact that 𝜒 (0) = 1 imply that, as
𝜆 → +∞,
∫ ∫
−1
e𝑖𝜆𝜑 ( 𝜆 𝑥 ) 𝑓1 (𝑥) 𝑎♭ 𝜆−1 𝑥, 𝜆 𝜒 𝜆−1 𝑥 d𝑥 ↦→ e𝑖 𝑥1 𝑓1 (𝑥) d𝑥 = b
𝑓1 (−1, 0, ..., 0) .

On the other hand, (23.2.32) and (23.2.33) imply, if 𝑁1 ≫ 𝑘,

𝑛

∥ 𝑓 ∥ 𝑘◦ 𝑃⊤ 𝑔 𝑘 ≤ ∥ 𝑓1 ∥ 𝑘◦ 𝜆 𝑘◦ − 2 𝐶𝜆−𝑁2 + 𝐶 ′𝜆 𝑁1 e−𝑐𝜆 .

By selecting 𝑓1 so that its Fourier transform b 𝑓1 does not vanish at (−1, 0, ..., 0) we
have reached the sought negation of (23.2.3). This proves the necessity of Condition
(P) for the local solvability of the analytic differential operator 𝑃 (𝑥, D) at 𝑥 ◦ .
Actually, the reasoning in this subsection and the preceding one can be modified
so that its conclusions apply to an analytic classical pseudodifferential operator in
Ω. The modifications are twofold:
(1) We must take into account the fact that pseudodifferential operators are pseu-
dolocal, not local (unless they are differential operators), and therefore must
use properly supported representatives of the operators or, which amounts to
the same, insert well-placed cut-off functions in the computations.
974 23 Analytic PDEs of Principal Type. Local Solvability

(2) The fact that, generally speaking, the total symbol of a classical pseudodif-
ferential operator has infinitely many terms (of homogeneity degree ↘ −∞)
complicates the system of transport equations (23.2.15) – without really altering
it.
These twin hurdles are manageable [about (2) cf. Subsection 18.3.3] and the
reasoning above, properly modified, leads to the following result:
Theorem 23.2.10 Let 𝑃 be a classical analytic pseudodifferential operator of order
𝑚 in the 𝑛-dimensional C 𝜔 manifold M. If 𝑃 is locally solvable at a point 𝑥 ◦ ∈ M
then each point (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 has the following property:
(𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) ) There is a conic neighborhood U of (𝑥 ◦ , 𝜉 ◦ ) in 𝑇 ∗ M\0 such that, given
any homogenous function ℎ ∈ C 𝜔 (U) such that the Hamiltonian Í vector
field of Re (ℎ𝑃𝑚 ) is not collinear to the radial vector field 𝑛𝑗=1 𝜉 𝑗 𝜕𝜕𝜉 𝑗 at
any point in U, then Im (ℎ𝑃𝑚 ) does not change sign from − to + on any
null bicharacteristic curve of Re (ℎ𝑃𝑚 ) in U.
In (𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) ) the null bicharacteristic curves of Re (ℎ𝑃𝑚 ) are oriented by the
(nowhere vanishing) vector field 𝐻Re(ℎ𝑃𝑚 ) .
Corollary 23.2.11 Let 𝑃 be a classical analytic pseudodifferential operator of prin-
cipal type and order 𝑚 in the 𝑛-dimensional C 𝜔 manifold M. Suppose that the
principal symbol of 𝑃 is antipodal, meaning that 𝑃𝑚 (𝑥, −𝜉) = (−1) 𝑚 𝑃𝑚 (𝑥, 𝜉) ev-
erywhere in 𝑇 ∗ M\0. If 𝑃 is locally solvable at a point 𝑥 ◦ ∈ M then there is an open
neighborhood 𝑈 of 𝑥 ◦ such that 𝑃 has Property (P) in 𝑇 ∗ M\0|𝑈 .
Proof Suppose there is a conic open subset U of 𝑇 ∗ M\0|𝑈 and a homogenous
function ℎ ∈ C 𝜔 (U) such that theÍHamiltonian vector field of Re (ℎ𝑃𝑚 ) is not
collinear to the radial vector field 𝑛𝑗=1 𝜉 𝑗 𝜕𝜕𝜉 𝑗 at any point of U and Im (ℎ𝑃𝑚 )
changes sign from + to − on a null bicharacteristic curve of Re (ℎ𝑃𝑚 ) in U, at some
point (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃. Then Im (ℎ (𝑥, −𝜉) 𝑃𝑚 ) changes sign from − to + on the
null bicharacteristic curve of Re (ℎ (𝑥, −𝜉) 𝑃𝑚 ) through (𝑥 ◦ , −𝜉 ◦ ) and thus violates
(𝚿 ( 𝑥 ◦ ,− 𝜉 ◦ ) ). □
We shall return to the condition (𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) ) in a later chapter.
Remark 23.2.12 If one wishes to remove the limitation that the differential operator
𝑃 (𝑥, D) is of principal type in the strong sense, meaning d 𝜉 𝑃𝑚 ≠ 0 everywhere
in 𝑇 ∗ M\0, and extend the necessity of Condition (P) to the general principal type
case (Definition 23.1.2; cf. also Example 23.1.4) then Corollary 23.2.11 is needed.
Indeed, the symplectic transformation that reduces the analysis to the case d 𝜉 𝑃𝑚 ≠ 0
transforms (microlocally) the differential operator 𝑃 (𝑥, D) into a classical analytic
pseudodifferential, generally not differential, operator.
Example 23.2.13 Consider the operator D 𝑥2 + 𝑖𝑥 2 D 𝑥1 acting on 𝑢 ∈ Cc∞ R2 ,

∫
D 𝑥2 𝑢 + 𝑖𝑥 2 D 𝑥1 𝑢 = (2𝜋) −𝑛 e𝑖 𝑥· 𝜉 (𝜉2 + 𝑖𝑥2 |𝜉1 |) b
𝑢 (𝜉) d𝜉.
23.2 Local Solvability of Analytic PDEs of Principal Type 975

The null bicharacteristic curves of 𝐴 = 𝜉2 are the straight-lines 𝑥1 = 𝑥1◦ , 𝜉1 = 𝜉1◦ ≠ 0,

𝜉2 = 0. Along these lines (oriented by 𝜕/𝜕𝑥2 ) 𝐵 = 𝑥2 |𝜉1 | changes sign from − to +.
By Theorem 23.2.10 the equation D 𝑥2 𝑢 + 𝑖𝑥2 D 𝑥1 𝑢 = 𝑓 is not locally solvable at
points of the 𝑥1 -axis. On the other hand, note that D 𝑥2 𝑢 − 𝑖𝑥 2 D 𝑥1 𝑢 = 𝑓 ∈ Cc∞ R2
admits the solution
∫ +∞
−1
𝑢 (𝑥) = (2𝜋) e𝑖 𝑥1 𝜉1 b
𝑢 (𝜉1 , 𝑥2 ) d𝜉1
−∞

where ∫ 𝑥2
1 2 2
𝑢 (𝜉1 , 𝑥2 ) = 𝑖
b e− 2 ( 𝑥2 −𝑡 ) | 𝜉1 | b
𝑓 (𝜉1 , 𝑡) d𝑡.
0

23.2.4 Sufficiency of Condition (P). 1. A functional-analytic a priori

estimate

We will need a result about an evolution equation of a special type in Hilbert space.
Actually, we shall be dealing with two complex Hilbert spaces H0 and H1 ⊂ H0 ,
H1 dense in H0 ; the Hermitian product and the norm in H0 will be denoted by (·, ·) 0
and ∥·∥ 0 , respectively. The equation will be

d𝑢
𝑳𝑢 = − 𝑮 (𝑡) 𝚽𝑢 − 𝑮 1 (𝑡) 𝑢 = 𝑓 , (23.2.34)
d𝑡
where 𝑮 (𝑡) and 𝑮 1 (𝑡) are bounded linear operators H0 −→ H0 depending con-
tinuously on 𝑡 ∈ [−𝑇, 𝑇] (𝑇 > 0); 𝚽 is a bounded linear operator H1 −→ H0 ;
𝑢 ∈ Cc1 ((−𝑇, 𝑇) ; H1 ).
We shall reason under the following hypotheses:

(h1) 𝚽 is self-adjoint, in the sense that (𝚽𝑢, 𝑣) 0 = (𝑢, 𝚽𝑣) 0 for all 𝑢, 𝑣 ∈ H1 ;
(h2) [𝑮 (𝑡) , 𝚽] = 𝑮 (𝑡) 𝚽 − 𝚽𝑮 (𝑡) and [[𝑮 (𝑡) , 𝚽] , 𝚽] (a priori defined on H1 )
extend as bounded linear operators H0 −→ H0 depending continuously on
𝑡 ∈ [−𝑇, 𝑇];
(h3) 𝑮 (𝑡) is self-adjoint and (𝑮 (𝑡) 𝑢, 𝑢) 0 ≥ 0 whatever 𝑢 ∈ H0 .

We begin with a reminder about unbounded self-adjoint operators in Hilbert

space (here H0 ). First of all, we observe that the set D (𝚽) of elements 𝑣 ∈ H0
such that the linear functional H1 ∋ 𝑢 ↦→ (𝚽𝑢, 𝑣) 0 extends as a continuous linear
functional 𝜆 𝚽 (𝑣) in H0 form a linear subspace of H0 . By (h1) H1 ⊂ D (𝚽) and
therefore D (𝚽) is dense in H0 ; if 𝑣 ∈ D (𝚽) \H1 we can define 𝑭𝑣 as the element
in H0 such that 𝜆𝚽 (𝑣) (𝑢) = (𝑢, 𝚽𝑣) 0 . When equipped with the “graph norm”
21
𝑣 ↦→ ∥𝑣∥ 20 + ∥𝚽𝑣∥ 20 the space D (𝚽) is complete; it is therefore a Hilbert space.
We shall simply assume that D (𝚽) = H1 .
976 23 Analytic PDEs of Principal Type. Local Solvability

We recall that the spectrum of 𝚽, Spect 𝚽, is the set of 𝑧 ∈ C such that the
operator 𝑧𝑰 − 𝚽 : H1 −→ H0 is not invertible. Since, for arbitrary 𝑢 ∈ H1 and
𝑧 = 𝑥 + 𝑖𝑦 ∈ C,

∥𝑧𝑢 − 𝚽𝑢∥ 20 = 𝑦 2 ∥𝑢∥ 20 + ∥𝑥𝑢 − 𝚽𝑢∥ 20 ≥ 𝑦 2 ∥𝑢∥ 20 ,

we see that Spect 𝚽 ⊂ R and also that, if 𝑦 ≠ 0,

(𝑧𝑰 − 𝚽) −1 ≤ |𝑦| −1 . (23.2.35)

(∥·∥: norm of bounded linear operators H0 −→ H0 .)

Let ker 𝚽 denote the orthogonal of 𝚽H1 in H0 ; we denote by 𝜋◦ the orthogonal
projection H0 −→ ker 𝚽 and define H0♭ = (ker 𝚽) ⊤ ; we define H1♭ = H1 ∩ H0♭ .
The restriction of 𝚽 to the Hilbert space H1♭ is an injective self-adjoint operator. We
shall denote by 𝝅 + the spectral projection of 𝚽 corresponding to (0, +∞) ∩ Spect 𝚽
and 𝜋− that corresponding to (−∞, 0) ∩ Spect 𝚽; we have 𝜋+ = 𝜋+∗ = 𝜋+2 , likewise
for 𝜋− and therefore 𝜋+ 𝜋− = 0; 𝜋+ + 𝜋− is the identity map of H0♭ , 𝜋+ + 𝜋− + 𝜋◦ = 𝑰,
the identity operator in H0 ; thus we have the orthogonal decomposition H0 =
𝜋+ H0 ⊕ 𝜋− H0 ⊕ ker 𝚽. Also, [𝜋± , 𝚽] = 0 and if we write 𝚽± = ±𝜋± 𝚽 then
𝚽 = 𝚽+ − 𝚽− and

∀𝑢 ∈ H1 , (𝚽+ 𝑢, 𝑢) 0 ≥ 0, (𝚽− 𝑢, 𝑢) 0 ≥ 0.

There are several approaches to the definition of the positive square root of the
operators 𝚽± . Here we start with the absolute value of 𝚽, |𝚽| = 𝚽+ + 𝚽− : H1 −→
H0 , and the operator-valued integral
∫ √
1
2
1 −1 𝑧d𝑧
𝚽+ (𝜀) = ( 𝑰 + 𝜀 |𝚽|) (𝑧𝑰 − 𝚽)
2𝜋𝑖 𝔠 1 + 𝜀𝑧
√
where 𝑧 is the main branch of the square-root function and the contour of integration
𝔠 = 𝔠− ∪ 𝔠◦ ∪ 𝔠+ is the union (oriented clockwise starting from ∞ in the lower right
quadrant) of the following three contours

𝔠◦ = {𝑧 = 𝑖𝑦 ∈ C; |𝑦| ≤ 1} ,
𝔠± = {𝑧 = 𝑥 + 𝑖𝑦 ∈ C; 0 < 𝑥 < +∞, 𝑦 = ± (1 + 𝑥)} .

Since ∥𝑢∥ 0 ≤ (( 𝑰 + 𝜀 |𝚽|) 𝑢, 𝑢) 0 for all 𝑢 ∈ H1 the map ( 𝑰 + 𝜀 |𝚽|) H1 ∋

( 𝑰 + 𝜀 |𝚽|) 𝑢 ↦→ 𝑢 can be extended as a bounded linear operator H0 −→ H0 ,
( 𝑰 + 𝜀 |𝚽|) −1 , of operator norm ≤ 1 and converging to 𝑰 as 𝜀 ↘ 0. We leave it as an
1
2
exercise to deduce that 𝚽 + (𝜀) converges
to a bounded linear operator H1 −→ H0 ,
1 1 1
precisely, 𝚽+2 . We have 𝚽+2 𝑢, 𝑢 ≥ 0 for all 𝑢 ∈ H1 . We define 𝚽−2 similarly. We
2 0
1 1 1 1
have 𝚽± = 𝚽± , 𝚽+ , 𝚽 = 𝚽+ , 𝜋± = 0 and likewise for 𝚽−2 .
2 2 2
23.2 Local Solvability of Analytic PDEs of Principal Type 977

Note that, by (23.2.35), the integral

∫ √
1 −1 𝑧d𝑧
(𝑧𝑰 − 𝚽)
2𝜋𝑖 𝔠◦ 1 + 𝜀𝑧

converges to a bounded linear operator H0 −→ H0 as 𝜀 ↘ 0.

Lemma 23.2.14 Under the hypotheses (h1) and (h2),

1 1 1
2 2 2
𝚽± , 𝑮 (𝑡) and 𝚽± , 𝚽± , 𝑮 (𝑡)

extend as bounded linear operators H0 −→ H0 depending continuously on 𝑡 ∈

[−𝑇, 𝑇].

Proof From (23.2.35) we first derive, for 0 < |𝑦| ≤ 1,

𝑮 (𝑡) , (𝑖𝑦𝑰 − 𝚽) −1 ≤ 2 ∥𝑮 (𝑡) ∥ |𝑦| −1

and, as a consequence, for all 𝜀 ≥ 0,

∫ √
−1 𝑧d𝑧
𝑮 (𝑡) , (𝑧𝑰 − 𝚽) ≲ ∥𝑮 (𝑡) ∥ , (23.2.36)
𝔠◦ 1 + 𝜀𝑧

where ≲ means up to a factor that is a universal positive constant. The formula

𝑮 (𝑡) , (𝑧𝑰 − 𝚽) −1 = − (𝑧𝑰 − 𝚽) −1 [𝑮 (𝑡) , 𝚽] (𝑧𝑰 − 𝚽) −1

is evident if we multiply both sides from the left and the right by 𝑧𝑰 − 𝚽. From this
and (23.2.35) we derive

∀𝑧 ∈ 𝔠± , 𝑮 (𝑡) , (𝑧𝑰 − 𝚽) −1 ≲ (1 + 𝑥) −2 ∥ [𝑮 (𝑡) , 𝚽] ∥ ,

whence ∫ √
−1
𝑧d𝑧
𝑮 (𝑡) , (𝑧𝑰 − 𝚽) ≲ ∥ [𝑮 (𝑡) , 𝚽] ∥ .
𝔠± 1 + 𝜀𝑧
Combining this with (23.2.36) yields

1
∀𝜀 > 0, 𝑮 (𝑡) , 𝚽+2 (𝜀) ≲ ∥ [𝑮 (𝑡) , 𝚽] ∥ + ∥𝑮 (𝑡) ∥ ,

whence, as 𝜀 ↘ 0,

1
𝑮 (𝑡) , 𝚽+2 ≲ ∥ [𝑮 (𝑡) , 𝚽] ∥ + ∥𝑮 (𝑡) ∥ .
978 23 Analytic PDEs of Principal Type. Local Solvability

Since [[𝑮 (𝑡) , 𝚽] , 𝚽] is bounded we may replace 𝑮 (𝑡) by

[𝑮 (𝑡) , 𝚽] in what
1
precedes and conclude that [−𝑇, 𝑇] ∋ 𝑡 ↦→ [𝑮 (𝑡) , 𝚽] , 𝚽+2 is a continuous map

1
into the algebra of bounded linear operators on H0 . Since 𝚽+2 , 𝚽 = 0 we have

1 1
2 2
𝑮 (𝑡) , 𝚽+ , 𝚽 = [𝑮 (𝑡) , 𝚽] , 𝚽+ .

1
2
It follows that we may apply the first part of the proof with 𝑮 (𝑡) , 𝚽+ in the

1 1
2 2
place of 𝑮 (𝑡) and therefore conclude that [−𝑇, 𝑇] ∋ 𝑡 ↦→ 𝑮 (𝑡) , 𝚽+ , 𝚽+ is
a continuous map into the algebra of bounded linear operators on H0 . The same
1 1
argument applies with 𝚽−2 replacing 𝚽+2 . □

Lemma 23.2.15 Under the hypotheses (h1), (h2), (h3), there is a constant 𝑐 ◦ > 0
such that 0 < 𝜏 < 𝑐 ◦ implies
∫ 𝜏 21 ∫ +𝜏 21
𝑐◦ ∥𝑢 (𝑡)∥ 20 d𝑡 ≤𝜏 ∥ 𝑳𝑢 (𝑠) ∥ 20 d𝑠
−𝜏 −𝜏

for all 𝑢 ∈ Cc1 ((−𝜏, 𝜏) ; H1 ).

Proof We will use the fact that [𝜕𝑡 , 𝝅 ◦ ] = 0 and [𝜕𝑡 , 𝝅 ± ] = 0. We get first, for
𝑢 ∈ Cc1 ((−𝜏, 𝜏) ; H1 ),

2 Re ( 𝑳𝑢, 𝝅 ◦ 𝑢) 0 = 𝜕𝑡 ∥𝝅 ◦ 𝑢∥ 20 − 2 Re (𝑮 1 (𝑡) 𝑢, 𝝅 ◦ 𝑢) 0

[𝑮 1 as in (23.2.34)], whence

𝜕𝑡 ∥𝝅 ◦ 𝑢∥ 20 ≤ 2 Re ( 𝑳𝑢, 𝝅 ◦ 𝑢) 0 + 2 max ∥𝑮 1 (𝑡) ∥ ∥𝑢∥ 20 . (23.2.37)
|𝑡 | ≤𝜏

Next,

2 Re ( 𝑳𝑢, 𝝅 + 𝑢) 0 = 𝜕𝑡 ∥𝝅 + 𝑢∥ 20 − 2 Re (𝑮 (𝑡) 𝑢, 𝚽+ 𝑢) 0 − 2 Re (𝑮 1 (𝑡) 𝑢, 𝝅 + 𝑢) 0

1 1
= 𝜕𝑡 ∥𝝅 + 𝑢∥ 20 − 2 𝑮 (𝑡) 𝚽+2 𝑢, 𝚽+2 𝑢
0

1 1 1 1
2 2 2 2
+ 𝑮 (𝑡) , 𝚽+ 𝑢, 𝚽+ 𝑢 + 𝚽+ 𝑢, 𝑮 (𝑡) , 𝚽+ 𝑢
0 0
− 2 Re (𝑮 1 (𝑡) 𝑢, 𝝅 + 𝑢) 0 .
∗
1 1
2 2
We exploit the fact that 𝑮 (𝑡) , 𝚽+ = − 𝑮 (𝑡) , 𝚽+ :
23.2 Local Solvability of Analytic PDEs of Principal Type 979

1 1
2 Re ( 𝑳𝑢, 𝝅 + 𝑢) 0 = 𝜕𝑡 ∥𝝅 + 𝑢∥ 20 − 2 𝑮 (𝑡) 𝚽+2 𝑢, 𝚽+2 𝑢
0

1 1
− 𝑮 (𝑡) , 𝚽+2 , 𝚽+2 𝑢, 𝑢 − 2 Re (𝑮 1 (𝑡) 𝑢, 𝝅 + 𝑢) 0 .
0

From this and from Lemma 23.2.14 and (h3) we derive

− 𝜕𝑡 ∥𝝅 + 𝑢∥ 20 ≤ −2 Re ( 𝑳𝑢, 𝝅 + 𝑢) 0 + 𝐶1 ∥𝑢∥ 20 . (23.2.38)

(In the remainder of the proof 𝐶1 , 𝐶2 , etc., are positive constants independent of 𝑢
2
1
and of 𝜏.) Recalling that 𝚽− = 𝜋− 𝚽 = − 𝚽−2 we see that

1 1
2 Re ( 𝑳𝑢, 𝝅 − 𝑢) 0 = 𝜕𝑡 ∥𝝅 − 𝑢∥ 20 + 2 Re 𝑮 (𝑡) 𝚽−2 𝑢, 𝚽−2 𝑢
0

1 1
+ 𝑮 (𝑡) , 𝚽−2 , 𝚽−2 𝑢, 𝑢 − 2 Re (𝑮 1 (𝑡) 𝑢, 𝝅 − 𝑢) 0 ,
0

whence, again by Lemma 23.2.14 and (h3),

𝜕𝑡 ∥𝝅 − 𝑢∥ 20 ≤ 2 Re ( 𝑳𝑢, 𝝅 − 𝑢) 0 + 𝐶2 ∥𝑢∥ 20 . (23.2.39)

We apply the Cauchy–Schwarz inequality and integrate (23.2.37) and (23.2.39)

from −𝜏 to 𝑡:
∫ 𝑡 ∫ 𝑡
1 2
∥𝝅 ◦ 𝑢 (𝑡) ∥ 0 ≤ ∥ 𝑳𝑢 (𝑠)∥ 0 ∥𝝅 ◦ 𝑢 (𝑠)∥ 0 d𝑠 + 𝐶3 ∥𝑢 (𝑠) ∥ 20 d𝑠,
2 −𝜏 −𝜏
∫ 𝑡 ∫ 𝑡
1 2
∥𝝅 − 𝑢 (𝑡) ∥ 0 ≤ ∥ 𝑳𝑢 (𝑠)∥ 0 ∥𝝅 − 𝑢 (𝑠)∥ 0 d𝑠 + 𝐶3 ∥𝑢 (𝑠) ∥ 20 d𝑠.
2 −𝜏 −𝜏

We integrate (23.2.38) from 𝑡 to 𝜏:

∫ 𝜏 ∫ 𝜏
1
∥𝝅 + 𝑢 (𝑡) ∥ 20 ≤ ∥ 𝑳𝑢 (𝑠)∥ 0 ∥𝝅 + 𝑢 (𝑠)∥ 0 d𝑠 + 𝐶4 ∥𝑢 (𝑠) ∥ 20 d𝑠.
2 𝑡 𝑡

Adding up these last three inequalities we get

∫ +𝜏 ∫ 𝜏
1 2
∥𝑢 (𝑡) ∥ 0 ≤ ∥ 𝑳𝑢 (𝑠)∥ 0 ∥𝑢 (𝑠)∥ 0 d𝑠 + 𝐶5 ∥𝑢 (𝑠) ∥ 20 d𝑠.
2 −𝜏 −𝜏

If we further integrate from −𝜏 to 𝜏 and once again apply the Cauchy–Schwarz

inequality we get

∫ 𝜏 21 ∫ +𝜏 21 ∫ 𝜏 21
1
∥𝑢 (𝑡) ∥ 20 d𝑡 ≤ ∥ 𝑳𝑢 (𝑡) ∥ 20 d𝑡 + 𝐶5 ∥𝑢 (𝑡) ∥ 20 d𝑡 .
4𝜏 −𝜏 −𝜏 −𝜏
980 23 Analytic PDEs of Principal Type. Local Solvability

Requiring 8𝐶𝜏 ≤ 1 yields the desired result with 𝑐 ◦ = (8𝐶) −1 . □

23.2.5 Sufficiency of Condition (P). 2. Patching up the microlocal

estimates

We return to our standard Euclidean data: the open set Ω in R𝑛 ; the analytic dif-
ferential operator 𝑃 (𝑥, D) of order 𝑚 ≥ 1 in Ω. Here we posit the stronger ver-
sion of principal type: d 𝜉 𝑃𝑚 does not vanish at any point of 𝑇 ∗ Ω\0. Note that
d 𝜉 𝑃𝑚 (𝑥, 𝜉) = 0 =⇒ 𝑃𝑚 (𝑥, 𝜉) = 0 by the Euler homogeneity identities. The analy-
sis takes place in a neighborhood 𝑈 of a point 𝑥 ◦ of Ω; we let 𝜉 ◦ ∈ 𝑇𝑥∗◦ Ω\ {0} belong
to a convex open cone Γ in R𝑛 \ {0}. We avail ourselves of Proposition 23.1.10 and
the considerations that follow it and start from Formula (23.1.7); more precisely, we
start from the hypothesis that we have (in 𝑈 × Γ)

(−1) 𝑚 𝑖𝐸 ◦ (𝑥, D 𝑥 ) −1 𝑃𝑚 (𝑥, D 𝑥 ) = T 𝑳T ∗ (23.2.40)

where 𝐸 ◦ (𝑥, D 𝑥 ) is elliptic of order 𝑚 − 1 and

𝑳 = 𝜕𝑥𝑛 − 𝐵 (𝑥, D 𝑥′ ) (23.2.41)

with 𝐵 (𝑥, 𝜉 ′) homogeneous of degree 1 and real-valued. Note that

Char 𝑳 = {(𝑥, 𝜉) ∈ 𝑈 × Γ; 𝜉 𝑛 = 0, 𝐵 (𝑥, 𝜉 ′) = 0} .

This allows us to take

Γ = {𝜉 ∈ R𝑛 \ {0} ; 𝜉 ′ ∈ Γ ′, |𝜉 𝑛 | < 𝜀 |𝜉 ′ |} (23.2.42)

with Γ ′ a convex open cone Γ in 𝜉 ′-space R𝑛−1 \ {0} containing 𝜉 ◦′ ≠ 0 and 𝜀 > 0.
There is no loss of generality in assuming 𝑥 ◦ = 0 and 𝑈 = 𝑈 ′ × (−𝜏, 𝜏) with 𝑈 ′
a neighborhood of 0 in R𝑛−1 and 𝜏 > 0. We introduce cut-off functions 𝜒1 (𝑥 ′) ∈
Cc∞ (𝑈 ′), 𝜒1 ≡ 1 in a neighborhood 𝜔 ′ ⊂ 𝑈 ′ of 0, 0 ≤ 𝜒1 ≤ 1 everywhere, and
′ ∞ ′ , 𝜒2 ≡ 1 in a neighborhood of 𝜉 ◦′, 0 ≤ 𝜒2 ≤ 1 everywhere

𝜒2 (𝜉 ) ∈ Cc Γ ∩ S 𝑛−2

in Γ ∩ S ; we extend 𝜒2 as a C ∞ function, homogeneous of degree zero, in the

′ 𝑛−2

whole R𝑛−1 \ {0}, 𝜒2 ≡ 0 in R𝑛−1 \Γ ′. We point out that 𝐵 (𝑥, D 𝑥′ ) 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) is

a well-defined (C ∞ , not C 𝜔 ) pseudodifferential operator in 𝑈 ′ (depending smoothly
on 𝑥 𝑛 ). At this point we avail ourselves of Condition (P) through Proposition 23.1.39:
in (23.2.41) we can write

𝐵 (𝑥, D 𝑥′ ) = 𝐺 (𝑥, D 𝑥′ ) 𝐹 (𝑥 ′, D 𝑥′ ) in 𝑈 ′ × (−𝜏, 𝜏) × Γ ′.

Here 𝐹 (𝑥 ′, 𝜉 ′) is homogeneous of degree 1, 𝐺 (𝑥, 𝜉 ′) ≥ 0 is homogeneous of degree

0.
23.2 Local Solvability of Analytic PDEs of Principal Type 981

Lemma 23.2.16 There is a 𝑐 ◦ > 0 such that, for every 𝜏 ∈ (0, 𝑐 ◦ ) and 𝑢 ∈
Cc∞ (𝜔 ′ × (−𝜏, 𝜏)),

𝑐 ◦ ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐿 2 ≤ 𝜏 ∥ 𝑳 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐿 2 . (23.2.43)

Proof The operator 𝚽 = 21 𝐹 (𝑥 ′, D 𝑥′ ) + 𝐹 (𝑥 ′, D 𝑥′ ) ∗ is self-adjoint, whereas

1 ′ ′ ∗
is of order zero since 𝐹 (𝑥 ′, 𝜉 ′) is real. The Sharp

2 𝐹 (𝑥 , D 𝑥 ) − 𝐹 (𝑥 , D 𝑥 )
′ ′

Gårding Inequality (Theorem 16.1.21) implies the existence of a classical pseudo-

differential operator 𝑮 (𝑥 𝑛 ) in 𝑈 ′, depending smoothly on 𝑥 𝑛 ∈ (−𝜏, 𝜏), with princi-
pal symbol 𝐺 (𝑥, 𝜉 ′) (hence, of order zero), which is positive semidefinite in 𝐿 2 (𝑈 ′).
We see that
1
𝐺 (𝑥, D 𝑥′ ) 𝐹 (𝑥 ′, D 𝑥′ ) − 𝐹 (𝑥 ′, D 𝑥′ ) ∗

𝑮 1 (𝑥 𝑛 ) =
2
+ (𝐺 (𝑥, D 𝑥′ ) − 𝑮 (𝑥 𝑛 )) 𝚽

is of order zero. We have

𝑳 = 𝜕𝑥𝑛 − 𝑮 (𝑥 𝑛 ) 𝚽 − 𝑮 1 (𝑥 𝑛 ) .

The pseudodifferential operators [𝚽, 𝐺 (𝑥, D 𝑥′ )], [𝚽, [𝚽, 𝐺 (𝑥, D 𝑥′ )]] are of order
zero, by (16.2.20). This shows that the hypotheses (h1), (h2) and (h3) are satisfied:
Lemma 23.2.16 is an application of Lemma 23.2.15. □
The commutator [𝑳, 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ )] is of order zero and thus

∥ 𝑳 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐿 2 ≤ ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑳𝑢∥ 𝐿 2 + 𝐶1 ∥𝑢∥ 𝐿 2 ,

where 𝐶1 > 0 depends on the diameters of supp 𝜒1 and supp 𝜒2 but not on 𝜏. We
can therefore choose 𝜏 so that 𝑐 ◦ 𝐶1 𝜏 ≤ 12 ; we derive from (23.2.43):

2𝑐 ◦ ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐿 2 ≤ 𝜏 ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑳𝑢∥ 𝐿 2
≤ 𝜏 T 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) T ∗ 𝐸 ◦ (𝑥, D 𝑥 ) −1 𝑃𝑚 (𝑥, D 𝑥 ) T 𝑢 𝐿2
.
∞ ′
Provided 𝜏 stays sufficiently small, when acting
′ ∗
on Cc (𝜔 × (−𝜏,−1𝜏)), 𝐾 (𝑥, D) =
T 𝜒1 (𝑥 ) 𝜒2 (D 𝑥′ ) T and the commutator 𝐾 (𝑥, D) , 𝐸 ◦ (𝑥, D 𝑥 ) are pseudod-
ifferential operators of order zero and −𝑚 respectively. From this and from the
ellipticity of order 𝑚 − 1 of 𝐸 ◦ (𝑥, D 𝑥 ) we deduce, after replacing 𝑢 by T ∗ 𝑢,

∥𝐾 (𝑥, D) 𝑢∥ 𝐿 2 ≤ 𝐶2 𝜏 ∥𝑃𝑚 (𝑥, D 𝑥 ) 𝑢∥ 𝐻 1−𝑚 + 𝐶2 𝜏 ∥𝑢∥ 𝐿 2 ,

where 𝐶2 > 0 is independent of 𝜏 [also, needless to say, of 𝑢 ∈ Cc∞ (𝜔 ′ × (−𝜏, 𝜏))].

Here we note that 𝑃 (𝑥, D 𝑥 ) ⊤ − (−1) 𝑚 𝑃𝑚 (𝑥, D 𝑥 ) is a differential operator of order
≤ 𝑚 − 1; the preceding estimate yields, with an increased constant 𝐶2 ,

∥𝐾 (𝑥, D) 𝑢∥ 𝐿 2 ≤ 𝐶2 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐻 1−𝑚

+ 𝐶2 𝜏 ∥𝑢∥ 𝐿 2 . (23.2.44)
982 23 Analytic PDEs of Principal Type. Local Solvability

We let Γ range over a finite covering {Γ1 , ..., Γ𝑟 } of R𝑛 \ {0} by open cones. To each
cone Γ 𝜄 we can associate a neighborhood 𝑈 𝜄 of 𝑥 ◦ such that the previous construction
applies when 𝜉 ◦ is replaced by a point in Γ 𝜄 ; note that this implies that the role of
𝜉 𝑛 might be played by 𝜉 𝑗 for some 𝑗 < 𝑛 depending on 𝜄. To each 𝜄 = 1, ..., 𝑟, there
corresponds an operator 𝐾 𝜄 (𝑥, D); we derive from the analogues of (23.2.44), for 𝑢
supported in a suitably small neighborhood of 𝑥 ◦ , 𝑈 (𝜏) ⊂ 𝑈1 ∩ · · · ∩ 𝑈𝑟 ,
𝑟
∑︁ 2
𝐾 𝜄 (𝑥, D) ∗ 𝐾 𝜄 (𝑥, D) 𝑢, 𝑢 ≤ 𝐶3 𝜏 2 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 + 𝐶3 𝜏 2 ∥𝑢∥ 2𝐿 2 ,

𝐿2 𝐻 1−𝑚
𝜄=1

with 𝐶3 > 0 independent of 𝜏. If the neighborhood 𝑉 of 𝑥 ◦ , 𝑈 (𝜏) ⊂ 𝑉 ⊂ 𝑈1 ∩· · ·∩𝑈𝑟 ,

is
Í𝑟sufficiently small (but is independent of 𝜏) then the pseudodifferential operator
∗
𝐾
𝜄=1 𝜄 (𝑥, D) 𝐾 𝜄 (𝑥, D) is elliptic of order zero when acting on Cc∞ (𝑉). We obtain,
for some 𝐶4 ≥ 𝐶3 independent of 𝜏, and all 𝑢 ∈ Cc∞ (𝑈 (𝜏)),

∥𝑢∥ 𝐿 2 ≤ 𝐶4 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐻 1−𝑚

+ 𝐶4 𝜏 ∥𝑢∥ 𝐿 2 ,

whence, if we take 𝜏 ≤ (2𝐶4 ) −1 ,

∥𝑢∥ 𝐿 2 ≤ 2𝐶4 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐻 1−𝑚

. (23.2.45)

This enables us to prove the following statement, in which 𝐻0𝑚−1 (𝑈◦ ) (𝑈◦ : a domain
in R𝑛 ) is the closure of Cc∞ (𝑈◦ ) in 𝐻 𝑚−1 (𝑈◦ ); 𝐻0𝑚−1 (𝑈◦ ) and 𝐻 1−𝑚 (𝑈◦ ) can be
identified with the dual of each other when viewed as spaces of distributions in 𝑈◦ .
Theorem 23.2.17 Let 𝑃 (𝑥, D) be an analytic differential operator of order 𝑚 in Ω
such that d 𝜉 𝑃𝑚 ≠ 0 at every point of Ω × (R𝑛 \ {0}) (regarded as a conic open subset
of 𝑇 ∗ Ω\0). Suppose there is a neighborhood 𝑈 of 𝑥 ◦ ∈ Ω such that 𝑃𝑚 has Property
(P) in 𝑈 × (R𝑛 \ {0}). Then there is a neighborhood 𝑈◦ ⊂ 𝑈 of 𝑥 ◦ and a bounded
linear operator 𝑮 : 𝐿 2 (𝑈◦ ) −→ 𝐻0𝑚−1 (𝑈◦ ) such that 𝑃 (𝑥, D) 𝑮 𝑓 = 𝑓 for every
𝑓 ∈ 𝐿 2 (𝑈◦ ).
Proof In the notation of the argument leading to (23.2.45) we take 𝑈◦ = 𝑈 (𝜏). We
regard 𝑃 (𝑥, D 𝑥 ) ⊤ Cc∞ (𝑈◦ ) as a linear subspace of the Sobolev space 𝐻 1−𝑚 (𝑈◦ );
(23.2.45) implies that the map 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 ↦→ 𝑢 into 𝐿 2 (𝑈◦ ) is continuous
with respect to the norm ∥·∥ 𝐻 1−𝑚 ; therefore, it can be extended to the clo-
sure of 𝑃 (𝑥, D 𝑥 ) ⊤ Cc∞ (𝑈◦ ) in 𝐻 1−𝑚 (𝑈◦ ) and thence to the entire Hilbert space
𝐻 1−𝑚 (𝑈◦ ) by setting it equal to 0 in the orthogonal of 𝑃 (𝑥, D 𝑥 ) ⊤ Cc∞ (𝑈◦ ); call
𝑮 ′ : 𝐻 1−𝑚 (𝑈◦ ) −→ 𝐿 2 (𝑈◦ ) the operator thus extended. We take the operator 𝑮 to
be the transpose of 𝑮 ′; then we have, for all 𝜑 ∈ Cc∞ (𝑈◦ ) and 𝑓 ∈ 𝐿 2 (𝑈◦ ),

∫
⟨ 𝑓 , 𝜑⟩ = 𝑓 (𝑥) 𝑮 ′ 𝑃 (𝑥, D 𝑥 ) ⊤ 𝜑 (𝑥) d𝑥

= ⟨𝑃 (𝑥, D 𝑥 ) 𝑮 𝑓 , 𝜑⟩ ,

whence the claim. □

23.2 Local Solvability of Analytic PDEs of Principal Type 983

Remark 23.2.18 In proving the crucial estimate (23.2.45) we have only exploited
the properties of the principal symbol 𝑃𝑚 (𝑥, 𝜉); (23.2.45) would remain valid if
𝑃 (𝑥, D 𝑥 ) ⊤ were replaced by an operator of the form ±𝑃𝑚 (𝑥, D 𝑥 ) + 𝑩, with 𝑩 a
bounded linear operator 𝐿 2 (𝑈◦ ) −→ 𝐻 1−𝑚 (𝑈◦ ). Moreover, we have never really
made use of the fact that 𝑃𝑚 (𝑥, 𝜉) is a polynomial with respect to 𝜉. We have only
used the fact, resulting from the analyticity of 𝑃𝑚 and Property (P), that, in a conic
neighborhood of (𝑥 ◦ , 𝜉 ◦ ) and possibly after a symplectic transformation, we had a
factorization
𝑃𝑚 (𝑥, 𝜉) = 𝐸 (𝑥, 𝜉) (𝜉 𝑛 + 𝑖𝐺 (𝑥, 𝜉 ′) 𝐹 (𝑥 ′, 𝜉 ′))
with 𝐸 elliptic, 𝐺 ≥ 0, etc. It follows that Theorem 23.2.17 remains true for any
classical pseudodifferential operator 𝑃 (𝑥, D) of principal type in Ω whose principal
symbol is analytic and satisfies (P) in 𝑈 × (R𝑛 \ {0}).
Remark 23.2.19 The hypothesis that d 𝜉 𝑃𝑚 never vanishes is essential, as shown by
Example 23.1.5: the equation 𝑃𝑢 = 𝜕𝜃 𝑢 = 𝑓 in an arbitrary neighborhood of 0 can
∫ 2𝜋
be solved if and only if 0 𝑓 (𝑟, 𝜃) d𝜃 = 0 for all 𝑟 > 0.
By combining Corollary 23.2.11 and Theorem 23.2.17 we can state
Theorem 23.2.20 Let 𝑃 (𝑥, D) be an analytic differential operator of order 𝑚 in Ω
such that d 𝜉 𝑃𝑚 ≠ 0 at every point of Ω × (R𝑛 \ {0}). For 𝑃 (𝑥, D) to be locally solv-
able in Ω it is necessary and sufficient that 𝑃 (𝑥, D) satisfy Condition (P) everywhere
in 𝑇 ∗ Ω\0.
As stated in Corollary 23.2.11 and Remark 23.2.18 we are allowed to replace
“analytic differential operator” by “classical analytic pseudodifferential operator” of
order 𝑚 in Ω, 𝑃 (𝑥, D), such that 𝑃𝑚 (𝑥, −𝜉) = ±𝑃𝑚 (𝑥, 𝜉).
The inequality (23.2.45) can be extended to the Sobolev spaces 𝐻 𝑠 , 𝑠 ∈ R (in the
place of 𝐻 0 = 𝐿 2 ) but at a cost: the small constant 𝜏 must be appropriately decreased.
As an example we prove

∥𝑢∥ 𝐻 𝑚−1 ≤ 2𝐶4 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐿2

(23.2.46)

for all 𝑢 ∈ Cc∞ (𝑈 (𝜏)). In (23.2.45) we replace 𝑢 by D 𝛼 𝑢, |𝛼| ≤ 𝑚 − 1. We get, for

a suitably large constant 𝐶5 > 0,
∑︁ ∑︁
∥D 𝛼 𝑢∥ 𝐿 2 ≤ 2𝐶4 𝜏 D 𝛼 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐻 1−𝑚
| 𝛼 | ≤𝑚−1 | 𝛼 | ≤𝑚−1
∑︁
𝑃 (𝑥, D 𝑥 ) ⊤ , D 𝛼 𝑢

+ 2𝐶4 𝜏 𝐻 1−𝑚
| 𝛼 | ≤𝑚−1

≤ 𝐶5 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐿2
+ 𝐶5 𝜏 ∥𝑢∥ 𝐻 𝑚−1

since 𝑃 (𝑥, D 𝑥 ) ⊤ , DÍ𝛼 has order 𝑚 + |𝛼| − 1 ≤ 2 (𝑚 − 1). Since the norms 𝑢 ↦→

∥𝑢∥ 𝐻 𝑚−1 and 𝑢 ↦→ | 𝛼 | ≤𝑚−1 ∥D 𝛼 𝑢∥ 𝐿 2 are equivalent we get, for some constant
𝐶6 ≥ 𝐶5 ,
∥𝑢∥ 𝐻 𝑚−1 ≤ 𝐶6 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐿 2 + 𝐶6 𝜏 ∥𝑢∥ 𝐻 𝑚−1 .
984 23 Analytic PDEs of Principal Type. Local Solvability

Taking 2𝐶6 𝜏 ≤ 1 yields (23.2.46).

Remark 23.2.21 Inequalities such as (23.2.45) or (23.2.46) are often said to reflect
the “loss of one derivative” when dealing with general differential operators of
principal type. If 𝑃 (𝑥, D) were elliptic (of order 𝑚) there would not be any loss of
derivatives: the norm in the left-hand side could then be ∥𝑢∥ 𝐻 𝑚 .

Remark 23.2.22 We underline the fact that the constant 𝜏 in (23.2.46) can be made to
decrease to zero with diam (supp 𝑢). A consequence of this is that (23.2.46) cannot be
valid unless 𝑃 (𝑥, D) is of principal type, as can be shown by inserting sequences of
◦
test-functions 𝑢 (𝑥, 𝜆) = e𝑖𝜆𝑥· 𝜉 𝑣 (𝑥, 𝜆) with 𝜉 ◦ ∈ S𝑛−1 such that d 𝜉 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) = 0,
𝑣 ∈ Cc∞ (Ω), 𝑣 ≡ 1 in a neighborhood of 𝑥 ◦ and diam (supp 𝑣) ↘ 0 as 𝜆 ↗ +∞. If
we do not require 𝜏 to go to zero with diam (supp 𝑢) then (23.2.46) is satisfied by a
differential operator 𝑃 (𝑥, D) (say, with smooth coefficients) elliptic of order 𝑚 − 1.

Remark 23.2.23 That the constant 𝜏 must be decreased as we increase |𝑠|, in esti-
mates
∥𝑢∥ 𝐻 𝑠+𝑚−1 ≲ 𝜏 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 𝐻 𝑠 , 𝑢 ∈ Cc∞ (𝑈 (𝜏)) , (23.2.47)
makes it impossible to deduce solely from the results of this chapter that, if 𝑃
is locally solvable at a point 𝑥 ◦ , then there is a neighborhood 𝑈 of 𝑥 ◦ such that
𝑃 (𝑥, D 𝑥 ) C ∞ (𝑈) ⊃ Cc∞ (𝑈). This is proved in [Hörmander, 1983, IV] using more
sophisticated analytic techniques.
Chapter 24
Analytic PDEs of Principal Type. Regularity of
the Solutions

In this chapter, explicitly or not, we take it as a given that the partial differential
operator under study, 𝑃(𝑥, D), is analytic, of principal type and locally solvable,
i.e., has Property (P) (Definition 23.1.37). The first part of this chapter centers on
the statement and proof of Theorem 17.4.10, which gathers the results originally
established in [Treves, 1971], [Treves, 1970, 2], [Treves, 1971, 2]. Our first purpose,
here, is to characterize those, among the operators of principal type, that are hypoel-
liptic (Definition 2.4.5) or analytic hypoelliptic (Definition 3.1.2). Corollary 2.4.7
tells us that the hypoellipticity of 𝑃(𝑥, D) implies the local solvability of its trans-
pose, 𝑃 (𝑥, D) ⊤ ; the latter must satisfy (P); but the principal symbols of 𝑃(𝑥, D) and
𝑃 (𝑥, D) ⊤ are the same up to sign (this is also true of antipodal pseudodifferential
operators) and therefore also 𝑃(𝑥, D) has Property (P). It follows that the question of
C ∞ hypoellipticity cannot be raised in dealing with differential operators that are not
locally solvable. As will be shown, for differential operators (with C 𝜔 coefficients)
of principal type, C ∞ and C 𝜔 hypoellipticity are equivalent. As a matter of fact, they
are equivalent to a much more precise property which we introduce at the start, subel-
lipticity. This concept has been in use since the 1960s in a much ampler field than
differential operators of principal type: it emerged in the study of the so-called Kohn
Laplacian on the smooth boundary of a pseudoconvex domain ([Kohn-Nirenberg,
1965]) and was the subject of intense investigations in SVC theory and in the sum
of squares theory.
Restricting our attention to analytic singularities, we gather (and prove) the results
in [Hanges, 1981], [Hanges-Sjöstrand, 1982] and [Sjöstrand, 1982], implying that
the analytic singularities of the solutions of an equation 𝑃(𝑥, D)𝑢 = 𝑓 ∈ C 𝜔 (Ω),
where Ω is a domain in R𝑛 , propagate along null bicharacteristic leaves of dimension
≤ 2. This last limitation happens to be a requirement of Property (P) (Proposition
23.1.43). The chapter concludes with the exploitation of the propagation theorem,
to show how it implies the semiglobal solvability of our equation in hyperfunctions,
under the hypothesis that null bicharacteristic leaves trapped over Ω (Definition
24.8.8) do not exist.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 985
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_24
986 24 Analytic PDEs of Principal Type. Regularity of the Solutions

24.1 A New Concept: Subellipticity

24.1.1 Subellipticity. Definition and basic properties

In this section 𝑃 = 𝑃 (𝑥, D) shall be a differential operator of order 𝑚 ≥ 1 in Ω with

C ∞ complex coefficients and principal symbol 𝑃𝑚 (𝑥, 𝜉).
We now call into play a new concept, originally introduced (in a different context)
in [Kohn-Nirenberg, 1965]. We denote most often by ∥·∥ 𝑠 the norm in the Sobolev
space 𝐻 𝑠 (R𝑛 ) (𝑠 ∈ R); when it is important to stress the number of variables we
denote that same norm by ∥·∥ 𝐻 𝑠 (R𝑛 ) .

Definition 24.1.1 The differential operator 𝑃 is called subelliptic if to every compact

subset 𝐾 of Ω there are constants 𝑠◦ ∈ (0, 1], 𝑐 𝐾 > 0, such that

𝑐 𝐾 ∥𝑢∥ 𝑚−1+𝑠◦ ≤ ∥𝑃𝑢∥ 𝐿 2 + ∥𝑢∥ 𝑚−1 (24.1.1)

for every 𝑢 ∈ Cc∞ (𝐾).

Proposition 24.1.2 For 𝑃 to be subelliptic it is necessary and sufficient that every

point 𝑥 ◦ ∈ Ω has a neighborhood 𝑈 ⊂ Ω such that

∃𝑠◦ ∈ (0, 1], ∀𝑢 ∈ Cc∞ (𝑈) , ∥𝑢∥ 𝑚−1+𝑠◦ ≲ ∥𝑃𝑚 (𝑥, D) 𝑢∥ 𝐿 2 . (24.1.2)

Proof The condition is sufficient. Let 𝐾 be a compact subset of Ω; there is a

covering of 𝐾 by finitely many open sets 𝑈𝑖 (𝑖 = 1, Í ..., 𝑟) in each of which (24.1.2)
holds. Let 𝜑𝑖 ∈ Cc∞ (𝑈𝑖 ), 𝑖 = 1, ..., 𝑟, be such that 𝑟𝑖=1 𝜑𝑖 ≡ 1 in a neighborhood of
𝐾. If 𝑢 ∈ Cc∞ (𝐾) then
𝑟
∑︁
∥𝑢∥ 𝑚−1+𝑠◦ ≤ 𝐶 ′ ∥𝜑𝑖 𝑢∥ 𝑚−1+𝑠◦
𝑖=1
∑︁𝑟
≤ 𝐶 ′′ ∥𝑃 (𝜑𝑖 𝑢)∥ 𝐿 2 + ∥(𝑃 − 𝑃𝑚 ) (𝜑𝑖 𝑢) ∥ 𝐿 2
𝑖=1
𝑟
∑︁
≤ 𝐶 ′′ ∥𝑃𝑢∥ 𝐿 2 + 𝐶 ′′ ∥ [𝑃, 𝜑𝑖 ] 𝑢∥ 𝐿 2 + ∥ (𝑃 − 𝑃𝑚 ) (𝜑𝑖 𝑢) ∥ 𝐿 2 ,
𝑖=1

whence (24.1.1).
The condition is necessary: (24.1.1) implies for some 𝑐 > 0 small, 𝐶 > 0 large
and all 𝑢 ∈ Cc∞ (𝑈),

𝑐 ∥𝑢∥ 𝑚−1+𝑠◦ ≤ ∥𝑃𝑚 𝑢∥ 𝐿 2 + ∥(𝑃 − 𝑃𝑚 ) (𝜑𝑖 𝑢) ∥ 𝐿 2 + ∥𝑢∥ 𝑚−1

≤ ∥𝑃𝑚 𝑢∥ 𝐿 2 + 𝐶 ∥𝑢∥ 𝑚−1 .

We apply Proposition 2.2.5: by contracting 𝑈 about 𝑥 ◦ we can ensure that 𝐶 ∥𝑢∥ 𝑚−1 ≤
1 ∞
2 𝑐 ∥𝑢∥ 𝑚−1+𝑠◦ for all 𝑢 ∈ Cc (𝑈), whence (24.1.2). □
24.1 A New Concept: Subellipticity 987

A slightly different formulation of Proposition 24.1.2 will be helpful.

Proposition 24.1.3 For 𝑃 to be subelliptic it is necessary and sufficient that every

point 𝑥 ◦ ∈ Ω has a neighborhood 𝑈 ⊂ Ω such that, for some 𝑠◦ ∈ (0, 1] and 𝐶 > 0
and every 𝑢 ∈ Cc∞ (𝑈),

∥𝑢∥ 𝑠◦ ≤ 𝐶 ∥𝑃𝑚 (𝑥, D 𝑥 ) 𝑢∥ 1−𝑚 . (24.1.3)

1−𝑚
Proof We follow standard procedure: replacing 𝑢 ∈ Cc∞ (𝑈) by 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 𝑢
with 𝜓 ∈ Cc∞ (𝑈), 𝜓 ≡ 1 in a neighborhood of supp 𝑢, enables us to rewrite (24.1.2)
as
𝑚−1
1−𝑚

(1 − Δ 𝑥 ) 2 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 𝑢
𝑠◦
1−𝑚

≤ 𝐶𝑈 𝑃𝑚 (𝑥, D 𝑥 ) 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 𝑢 2 .
𝐿

We have, for some 𝐶 ′ > 0 depending only on 𝜓,

𝑚−1
1−𝑚

(1 − Δ 𝑥 ) 2 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 𝑢 ≥ ∥𝑢∥ 𝑠◦ − 𝐶 ′ ∥𝑢∥ 𝑠◦ −1 .
𝑠◦
h 1−𝑚
i
Since the pseudodifferential operator 𝑃𝑚 (𝑥, D 𝑥 ) , 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 is of order
zero we also have
1−𝑚
𝑃𝑚 (𝑥, D 𝑥 ) 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 𝑢 ≤ ∥𝑃𝑚 (𝑥, D 𝑥 ) 𝑢∥ 1−𝑚
𝐿2
h 1−𝑚
i
+ 𝑃𝑚 (𝑥, D 𝑥 ) , 𝜓 (𝑥) (1 − Δ 𝑥 ) 2 𝑢
𝐿2
≤ ∥𝑃𝑚 (𝑥, D 𝑥 ) 𝑢∥ 1−𝑚 + 𝐶 ′′ ∥𝑢∥ 𝐿 2 ,

with 𝐶 ′′ > 0 depending only on 𝜓. We apply once again Proposition 2.2.5: since
0 < 𝑠◦ ≤ 1 the claim ensues by decreasing diam 𝑈. □

24.1.2 Subellipticity implies hypoellipticity

Lemma 24.1.4 Let 𝜌 ∈ Cc∞ (R𝑛 ) and define 𝜌 𝜀 (𝑥) = 𝜀 −𝑛 𝜌 𝑥𝜀 (𝜀 > 0). Given

arbitrarily 𝑎 ∈ C ∞ (R𝑛 ) the linear operators 𝑢 ↦→ 𝑎 (𝜌 𝜀 ∗ 𝑢) − 𝜌 𝜀 ∗ (𝑎𝑢) (0 < 𝜀 < 1)
form a bounded set of standard pseudodifferential operators of order −1 (Definition
16.1.14).

Proof Let 𝑢 ∈ 𝐻c𝑠 (R𝑛 ). Since the sets supp (𝜌 𝜀 ∗ 𝑢) converge to supp 𝑢 it suffices
to prove the claim when supp 𝑎 is compact. The Fourier inversion formula implies
988 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫ ∫ ∫
(2𝜋) 𝑛 (𝑎 (𝜌 𝜀 ∗ 𝑢) − 𝜌 𝜀 ∗ (𝑎𝑢)) = 𝑎 (𝑥) e𝑖 𝜉 · ( 𝑥−𝑧) 𝜌 𝜀 (𝑧 − 𝑦) 𝑢 (𝑦) d𝑦d𝑧d𝜉
∫ ∫ ∫
− e𝑖 𝜉 · (𝑧−𝑦) 𝜌 𝜀 (𝑥 − 𝑧) 𝑎 (𝑧) 𝑢 (𝑦) d𝑦d𝑧d𝜉

∫ ∫ ∫
= 𝑎 (𝑥) e𝑖 𝜉 · ( 𝑥−𝑦−𝑤) 𝑢 (𝑦) d𝑦d𝑤d𝜉
∫ ∫ ∫
− e𝑖 𝜉 · ( 𝑥−𝑦−𝑤) 𝜌 𝜀 (𝑤) 𝑎 (𝑥 − 𝑤) 𝑢 (𝑦) d𝑦d𝑤d𝜉
∫ ∫
= e𝑖 𝜉 · ( 𝑥−𝑦) 𝑏 𝜀 (𝑥, 𝜉) 𝑢 (𝑦) d𝑦d𝜉,

where
∫
𝑏 𝜀 (𝑥, 𝜉) = e−𝑖 𝜉 ·𝑤 (𝑎 (𝑥) − 𝑎 (𝑥 − 𝑤)) 𝜌 𝜀 (𝑤) d𝑤
𝑛 ∫
∑︁ 𝜕
=𝑖 e−𝑖 𝜉 ·𝑤 𝑎 ′𝑗 (𝑥, 𝑤) 𝜌 𝜀 (𝑤) d𝑤
𝑗=1
𝜕𝜉 𝑗

with 𝑎 ′ = 𝑎 1′ , ..., 𝑎 𝑛′ ∈ C𝑛 such that 𝑎 (𝑥) − 𝑎 (𝑥 − 𝑤) = 𝑛𝑗=1 𝑤 𝑗 𝑎 ′𝑗 (𝑥, 𝑤). We

Í
define ∫
𝑐 𝜀, 𝑗 (𝑥, 𝜉) = e−𝑖 𝜉 ·𝑤 𝑎 ′𝑗 (𝑥, 𝑤) 𝜌 𝜀 (𝑤) d𝑤.

The function R𝑛 ∋ 𝑥 ↦→ 𝑎 ′𝑗 (𝑥, 𝑤) ∈ C𝑛 is C ∞ and its support is contained in a

compact set independent of 𝑤 ∈ 𝐾 = supp 𝜌. It follows that

D 𝑥𝛼 D𝑤 𝑎 ′𝑗 (𝑥, 𝑤) < +∞.

𝛽
∀𝛼, 𝛽 ∈ Z+𝑛 , max (24.1.4)
( 𝑥,𝑤) ∈R𝑛 ×𝐾

For 𝛽, 𝛾 ∈ Z+𝑛 , 𝛾 ⪯ 𝛽,
∫
|𝛽−𝛾 |
D𝑤 e−𝑖 𝜉 ·𝑤 𝑤 𝛽 D 𝑥𝛼 𝑎 ′𝑗 (𝑥, 𝑤) 𝜌 𝜀 (𝑤) d𝑤
𝛾 𝛽 𝛾
𝜉 D 𝜉 D 𝑥𝛼 𝑐 𝜀, 𝑗 (𝑥, 𝜉) = (−1)
∫
e−𝑖 𝜉 ·𝑤 D𝑤 𝑤 𝛽 D 𝑥𝛼 𝑎 ′𝑗 (𝑥, 𝑤) 𝜌 𝜀 (𝑤) d𝑤.
𝛾
=

The Leibniz rule yields

D𝑤 𝑤 𝛽 D 𝑥𝛼 𝑎 ′𝑗 (𝑥, 𝑤) 𝜌 𝜀 (𝑤)
𝛾

𝛾−𝛾′ 𝛼 ′ 𝛾′
∑︁ 𝛾!
𝛽
= D 𝑤 D 𝑎 (𝑥, 𝑤) D 𝑤 𝑤 𝜌 𝜀 (𝑤) .
𝛾′ ⪯𝛾
𝛾 ′! (𝛾 − 𝛾 ′)! 𝑥 𝑗
24.1 A New Concept: Subellipticity 989

𝛾′
Since 𝛾 ′ ⪯ 𝛾 ⪯ 𝛽 we see that 𝜀 𝑛 D𝑤 𝑤 𝛽 𝜌 𝜀 (𝑤) has an upper bound independent

of 𝜀. If we combine this with (24.1.4) we get

∑︁
𝛽
𝜉 𝛾 D 𝜉 D 𝑥𝛼 𝑐 𝜀, 𝑗 (𝑥, 𝜉) ≤ 𝐶 𝛼,𝛽 .
𝛾 ⪯𝛽

This proves that the 𝑐 𝜀, 𝑗 (𝑥, 𝜉) form a bounded set of standard symbols of order zero
Í 𝜕𝑐
(Definition 16.1.1). Since 𝑏 𝜀 = 𝑖 𝑛𝑗=1 𝜕 𝜉𝜀,𝑗 𝑗 the claim is thus proved. □

The following corollary is the classical Friedrichs Lemma.

∫
Corollary 24.1.5 Let 𝜌 ∈ Cc∞ (R𝑛 ) be such that 𝜌 (𝑥) d𝑥 = 1 and let 𝑎 ∈ C ∞ (R𝑛 )
be arbitrary. Whatever 𝑠 ∈ R and 𝑢 ∈ 𝐻c𝑠 (R𝑛 ), as 𝜀 ↘ 0 the C ∞ functions
𝑎 (𝜌 𝜀 ∗ 𝑢) − 𝜌 𝜀 ∗ (𝑎𝑢) converge to zero in 𝐻c𝑠+1 (R𝑛 ).

Proof On the one hand, by Lemma 24.1.4 the commutators [𝑎, 𝜌 𝜀 ∗] form an
equicontinuous set of continuous linear maps 𝐻c𝑠 (R𝑛 ) −→ 𝐻c𝑠+1 (R𝑛 ) (cf. Re-
mark 16.1.20). On the other hand, the hypothesis on 𝜌 implies that [𝑎, 𝜌 𝜀 ∗] 𝑢 → 0
in Cc∞ (R𝑛 ) whatever 𝑢 ∈ Cc∞ (R𝑛 ). The fact that Cc∞ (R𝑛 ) is dense in 𝐻c𝑠 (R𝑛 )
[precisely because 𝜌 𝜀 ∗ 𝑢 ↦→ 𝑢 in 𝐻c𝑠 (R𝑛 )] and a classical theorem of Functional
Analysis (see, e.g., [Treves, 1967], Proposition 32.5) entails then that [𝑎, 𝜌 𝜀 ∗] 𝑢 → 0
in 𝐻c𝑠+1 (R𝑛 ) whatever 𝑢 ∈ 𝐻c𝑠 (R𝑛 ). □

Corollary 24.1.6 Let 𝜌 ∈ Cc∞ (R𝑛 ) and let 𝑃 (𝑥, D 𝑥 ) be a linear partial differential
operator of order 𝑚 with C ∞ coefficients in Ω. Let Ω′ be an open subset of R𝑛 whose
closure is contained in Ω and let 𝜀◦ > 0 be sufficiently small that Ω′ + supp 𝜌 𝜀 ⊂⊂ Ω
for all 𝜀 ∈ (0, 𝜀◦ ]. The commutators [𝑃 (𝑥, D 𝑥 ) , 𝜌 𝜀 ∗] (0 < 𝜀 ≤ 𝜀◦ ) form a bounded
set of pseudodifferential operators of order 𝑚 − 1 in Ω′.

Proof It suffices to prove the statement when 𝑃 (𝑥, D 𝑥 ) = 𝑎 (𝑥) D𝑥𝛼 , 𝑎 ∈ C∞ (Ω),
𝛼 ∈ Z+𝑛 , in which case it follows directly from Lemma 24.1.4 since D 𝑥𝛼 , 𝜌 𝜀 ∗ = 0.□

Theorem 24.1.7 If 𝑃 (𝑥, D 𝑥 ) is a subelliptic linear partial differential operator of

order 𝑚 with C ∞ coefficients in Ω then, for every 𝑣 ∈ D ′ ( Ω), 𝑃 (𝑥, D 𝑥 ) 𝑣 ∈ C ∞ (Ω)
implies 𝑣 ∈ C ∞ (Ω).

Proof Let 𝑣 ∈ D ′ (Ω) be such that 𝑃 (𝑥, D 𝑥 ) 𝑣 ∈ C ∞ (𝑈) and let 𝑉 ⊂⊂ Ω be

a neighborhood of 𝑥 ◦ in which (24.4.75) is valid for all 𝑢 ∈ Cc∞ (𝑉). We select
a sequence of neighborhoods of 𝑥 ◦ , 𝑊 𝑗 ( 𝑗 ∈ Z+ ), such that 𝑊 𝑗+1 ⊂⊂ 𝑊 𝑗 ⊂⊂
𝑊 of 𝑥 ◦ . We can then select a
Ñ
𝑊0 ⊂⊂ 𝑉 and 𝑗 𝑊 𝑗 contains a neighborhood
∞

sequence of functions 𝑔 𝑗 ∈ Cc 𝑊 𝑗−1 , 𝑗 = 1, 2, ..., such that 𝑔 𝑗 (𝑥) ≡ 1 for every
𝑠+𝑚−1 (𝑉)
𝑥 ∈ 𝑊 𝑗 . Let 𝑠 be such that 𝑣 ∈ 𝐻loc and suppose we have proved that
𝑔𝑗𝑣 ∈ 𝐻 𝑠+𝑚−1+( 𝑗−1) 𝜎◦ (R ).
𝑛
∫
Let 𝜌 ∈ Cc∞ (R𝑛 ) be such that 𝜌 (𝑥) ≠ 0 =⇒ |𝑥| < 1 and 𝜌 (𝑥) d𝑥 = 1, and select
𝜀◦ > 0 such that dist(𝑥, 𝑊0 ) ≤ 𝜀◦ =⇒ 𝑥 ∈ 𝑉. Then 𝑢 𝑗, 𝜀 = 𝜌 𝜀 ∗ 𝑔 𝑗 𝑣 ∈ Cc∞ (𝑉) for

all 𝑗 = 1, 2, ..., and 𝜀 ∈ (0, 𝜀◦ ]; we have
990 24 Analytic PDEs of Principal Type. Regularity of the Solutions
2 2 2
𝑢 𝑗+1, 𝜀 𝑠+𝑚−1+ 𝑗 𝜎◦
≤ 𝐶 𝑗 𝑃 (𝑥, D 𝑥 ) 𝑢 𝑗+1, 𝜀 𝑠+( 𝑗−1) 𝜎◦
+ 𝐶 𝑗 𝑢 𝑗+1, 𝜀 𝑠+𝑚−1+( 𝑗−1) 𝜎◦
(24.1.5)
and

𝑃 (𝑥, D 𝑥 ) 𝑢 𝑗+1, 𝜀 = 𝜌 𝜀 ∗ 𝑃 (𝑥, D 𝑥 ) 𝑔 𝑗+1 𝑣 + [𝑃 (𝑥, D 𝑥 ) , 𝜌 𝜀 ∗]

= 𝜌 𝜀 ∗ 𝑔 𝑗+1 𝑃 (𝑥, D 𝑥 ) 𝑣 + 𝜌 𝜀 ∗ 𝑃 (𝑥, D 𝑥 ) , 𝑔 𝑗+1 𝑣
+ [𝑃 (𝑥, D 𝑥 ) , 𝜌 𝜀 ∗] 𝑔 𝑗+1 𝑣.

Since 𝑔 𝑗+1 𝑔 𝑗 ≡ 𝑔 𝑗+1 we have 𝑃 (𝑥, D 𝑥 ) , 𝑔 𝑗+1 𝑣 = 𝑃 (𝑥, D 𝑥 ) , 𝑔 𝑗+1 𝑔 𝑗 𝑣 and
2 2
≤ 𝐶 ′𝑗 𝑔 𝑗 𝑣

𝑃 (𝑥, D 𝑥 ) , 𝑔 𝑗+1 𝑣 𝑠+( 𝑗−1) 𝜎◦ 𝑠+𝑚−1+( 𝑗−1) 𝜎◦
.

By Lemma 24.1.4,
2
[𝑃 (𝑥, D 𝑥 ) , 𝜌 𝜀 ∗] 𝑔 𝑗+1 𝑣 𝑠+( 𝑗−1) 𝜎◦
≤ 𝐶 ′′𝑗 𝑔 𝑗+1 𝑣 𝑠+𝑚−1+( 𝑗−1) 𝜎◦
2
≤ 𝐶 ′′′
𝑗 𝑔𝑗𝑣 𝑠+𝑚−1+( 𝑗−1) 𝜎◦
.

Taking these two estimates into account we derive from (24.1.5):

2 2
𝜌 𝜀 ∗ 𝑔 𝑗+1 𝑣 𝑠+𝑚−1+ 𝑗 𝜎◦
≤ 𝐶 𝑗+1 𝑔 𝑗+1 𝑃 (𝑥, D 𝑥 ) 𝑣 𝑠+( 𝑗−1) 𝜎◦
2
+ 𝐶 𝑗+1 𝑔 𝑗 𝑣 𝑠+𝑚−1+( 𝑗−1) 𝜎◦
.

By letting 𝜀 ↘ 0 we derive that 𝑔 𝑗+1 𝑣 ∈ 𝐻 𝑠+𝑚−1+ 𝑗 𝜎◦ (R𝑛 ). By letting 𝑗 ↗ +∞ we

conclude that 𝑣| 𝑊 ∈ C ∞ (𝑊). □

24.1.3 From microlocal to local subellipticity

First suppose that (24.1.3) holds. Let 𝑥 ◦ ∈ Ω, 𝜉 ◦ ∈ S𝑛−1 , be such that (𝑥 ◦ , 𝜉 ◦ ) ∈

Char 𝑃, otherwise arbitrary. We select a suitably small neighborhood of 𝑥 ◦ , 𝑈 ⊂
Ω, and a convex open cone in R𝑛 \ {0}, Γ ∋ 𝜉 ◦ . It is convenient (but not really
necessary) to avail ourselves of Proposition 23.1.10 and Remark 23.1.11: we assume
that d 𝜉 𝑃𝑚 (𝑥, 𝜉) ≠ 0 in 𝑈 × Γ and that the unitary transformation T in (23.2.40) has
√
been carried out: 𝑃𝑚 = 𝐸 ◦ 𝐿 in 𝑈 ×Γ with 𝐿 (𝑥, 𝜉) = −1𝜉 𝑛 − 𝐵 (𝑥, 𝜉 ′), 𝐸 ◦ (𝑥, 𝜉) ≠ 0
whatever (𝑥, 𝜉) ∈ 𝑈 × Γ, 𝐿 and 𝐸 ◦ homogeneous; thus

(𝑈 × Γ) ∩ Char 𝑃 = (𝑈 × Γ) ∩ Char 𝐿
= {(𝑥, 𝜉) ∈ 𝑈 × Γ; 𝜉 𝑛 = 0, 𝐵 (𝑥, 𝜉 ′) = 0} .

It follows that 𝜉 𝑛◦ = 0, |𝜉 ◦′ | = 1, 𝐵 (𝑥 ◦ , 𝜉 ◦′) = 0. To simplify the notation we assume

that 𝑥 ◦ = 0. We then take
24.1 A New Concept: Subellipticity 991

𝑈 = 𝑈 ′ × (−𝜏, 𝜏) , Γ = {𝜉 ∈ R𝑛 ; 𝜉 ′ ∈ Γ ′, |𝜉 𝑛 | < 𝑐 |𝜉 ′ |} ,

with 𝑈 ′ a neighborhood of 0 in R𝑛−1 , Γ ′ a convex open cone in R𝑛 \ {0}, 𝜉 ◦′ ∈ Γ ′, 𝜏 >

0, 𝑐 > 0 suitably small. We make use of cut-off functions 𝜒1 ∈ Cc∞ (𝑈 ′), 0 ≤ 𝜒1 ≤ 1
everywhere, 𝜒1 ≡ 1 in a neighborhood 𝑈◦′ of 𝑥 ◦′, and 𝜒2 ∈ Cc∞ Γ ′ ∩ S𝑛−2 extended
to R𝑛−1 \ {0} as a homogeneous function of degree zero, 0 ≤ 𝜒2 ≤ 1 everywhere,
𝜒2 ≡ 1 in a convex open cone Γ◦′ ∋ 𝜉 ◦′. In what follows 𝐶1 , 𝐶2 , etc. will be positive
constants independent of 𝑢 ∈ Cc∞ (𝑈).
Let us first assume that (24.1.3) is valid. We replace 𝑢 by 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢 ∈
∞
Cc (𝑈); since 𝐸 ◦ (𝑥, D 𝑥 ) is an elliptic pseudodifferential operator of order 𝑚 − 1 in
𝑈 × Γ we derive

∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝑠◦ ≤ 𝐶1 ∥𝑃𝑚 (𝑥, D 𝑥 ) 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 1−𝑚

≤ 𝐶2 ∥𝐿 (𝑥, D 𝑥 ) 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 0 + 𝐶3 ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 0
≤ 𝐶2 ∥ 𝜒1 (𝑥 ′) 𝐿 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝑢∥ 0 + 𝐶4 ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 0 .

We apply Proposition 2.2.5: since 𝑠◦ > 0, by decreasing diam 𝑈 we achieve that

1
𝐶4 ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 0 ≤ ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝑠◦ ,
2
whence

∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝑠◦ ≤ 2𝐶2 ∥ 𝜒1 (𝑥 ′) 𝐿 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐿 2 . (24.1.6)

Since 𝐿 = 𝐸 ◦−1 𝑃𝑚 the same type of argument entails

∥ 𝜒1 (𝑥 ′) 𝐿 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐿 2
≤ 𝐶5 ∥𝑃𝑚 (𝑥, D 𝑥 ) ( 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢) ∥ 1−𝑚 + 𝐶5 ∥𝑢∥ 𝐿 2 .

Combining this with (24.1.6) yields

∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝑠◦ ≤ 𝐶6 ∥𝑃 (𝑥, D 𝑥 ) ( 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢) ∥ 1−𝑚 + 𝐶7 ∥𝑢∥ 𝐿 2 .

(24.1.7)
If (𝑥 ◦ , 𝜉 ◦ ) ∉ Char 𝑃 we can select 𝑈 and Γ such that 𝑈 ∩ Γ ∩ Char 𝑃 = ∅; then
𝑃 (𝑥, D 𝑥 ) 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) is elliptic in 𝑈◦′ × (−𝜏, 𝜏) × Γ◦ , where

Γ◦ = 𝜉 ∈ R𝑛 ; 𝜉 ′ ∈ Γ◦′ , |𝜉 𝑛 | < 𝑐 |𝜉 ′ | .

This yields optimal estimates, meaning (24.1.7) with 𝑠◦ = 1.

Conversely, we can patch together the microlocal estimates (24.1.7) in the same
fashion that we have derived (23.2.45) from (23.2.44), thereby proving

∥𝑢∥ 𝑠◦ ≤ 𝐶8 ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 1−𝑚 + 𝐶8 ∥𝑢∥ 𝐿 2 (24.1.8)

or, equivalently,
∥𝑢∥ 𝑚−1+𝑠◦ ≲ ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 𝐿 2 + ∥𝑢∥ 𝑚−1 (24.1.9)
992 24 Analytic PDEs of Principal Type. Regularity of the Solutions

for every 𝑢 ∈ Cc∞ (𝑈◦ ) with 𝑈◦ a suitably small neighborhood of 𝑥 ◦ ∈ Ω.

24.2 Statement of the Main Theorem

We now state the main result of this part of the present chapter.
Theorem 24.2.1 Let 𝑃 = 𝑃 (𝑥, D) be a differential operator of principal type with
C 𝜔 coefficients in the domain Ω in R𝑛 . The following properties are equivalent:
(a) 𝑃 is hypoelliptic,
(b) 𝑃 is analytic hypoelliptic,
(c) 𝑃 is subelliptic,
and they imply that 𝑃 satisfies Condition (P) (Definition 23.1.37). Moreover, they
are equivalent to the following property:
(Q) there are no null bicharacteristic leaves of 𝑃 of positive dimension in 𝑇 ∗ Ω\0
(Definition 23.1.15).
The entailment (c)=⇒(a) follows from Theorem 24.1.7.
The claim that each one of the properties (a), (b), (c) implies (P) is a direct
consequence of their equivalence and of the following
Proposition 24.2.2 A hypoelliptic differential operator of principal type 𝑃 in Ω
satisfies Condition (P) at every point of 𝑇 ∗ Ω\0 and therefore 𝑃 is locally solvable at
every point of Ω.
Proof By Corollary 2.4.7, if 𝑃 is hypoelliptic then its transpose 𝑃⊤ is locally solvable
at every point of Ω. By the necessity part in Theorem 23.2.20, if 𝑃⊤ is locally solvable
in Ω the principal symbol of 𝑃⊤ , 𝑃𝑚 (𝑥, −𝜉) = (−1) 𝑚 𝑃𝑚 (𝑥, 𝜉), must have Property
(P) at every point of 𝑇 ∗ Ω\0 Ω × (R𝑛 \ {0}). Then the latter must also be true of
𝑃𝑚 (𝑥, 𝜉) which proves the claim, by the sufficiency part in loc. cit. □
It is easy to give examples of differential operators that have Property (P) but are
not hypoelliptic.
Example 24.2.3 The vector field 𝐿 = 𝜕𝑥𝜕 2 + 𝑖𝑥 1 𝜕𝑥𝜕 1 satisfies Condition (P) (Def-
inition 23.1.23) but 𝐿 is not hypoelliptic since the signum function 𝑥1 /|𝑥1 | is
a distribution solution of the equation 𝐿ℎ = 0. Each of the two half-planes
(𝑥, 𝜉) ∈ R4 ; 𝑥1 = 𝜉2 = 0, 𝜉1 ≷ 0 is a null bicharacteristic surface of 𝐿.
From now on we shall assume that 𝑃 (𝑥, D) satisfies Condition (P).
In view of Proposition 23.1.43 Property (Q) states that there are no null bichar-
acteristic curves or surfaces of 𝑃 in 𝑇 ∗ Ω\0. The following should be kept in mind.
Proposition 24.2.4 Let 𝑃 be an analytic pseudodifferential operator of order 𝑚 in Ω,
of principal type and satisfying Condition (P). Suppose that no null bicharacteristic
leaf of 𝑃 of positive dimension contains (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃. If d Re 𝑃𝑚 does not vanish
at (𝑥 ◦ , 𝜉 ◦ ) then the following properties hold:
24.3 Hypoellipticity Implies (Q) 993

(1) the zeros of Im 𝑃𝑚 along the null bicharacteristic curve 𝔠 of Re 𝑃𝑚 through

(𝑥 ◦ , 𝜉 ◦ ) must be isolated (and of even order);
(2) there is a conic neighborhood of (𝑥 ◦ , 𝜉 ◦ ) in 𝑇 ∗ Ω\0 which does not intersect any
null bicharacteristic leaf of 𝑃 of positive dimension.

Proof Let us write 𝑃𝑚 = 𝐴 + 𝑖𝐵 and take 𝑥 ◦ = 0. By the results in Section 23.1

(in particular Proposition 23.1.39), there is no loss of generality in assuming that
𝐴 = 𝜉 𝑛 , 𝐵 = 𝐺 (𝑥, 𝜉 ′) 𝐹 (𝑥 ′, 𝜉 ′), 𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛−1 ), 𝜉 ′ = (𝜉1 , ..., 𝜉 𝑛−1 ), 𝐺 ≥ 0
everywhere. Then 𝐻 𝐴 = 𝜕𝑥𝜕𝑛 and

𝔠 = {(𝑥, 𝜉) ∈ 𝑇 ∗ Ω; 𝑥 ′ = 0, − 𝑇1 < 𝑥 𝑛 < 𝑇2 , 𝜉 ′ = 𝜉 ′◦ ≠ 0, 𝜉 𝑛 = 0}

with 𝑇 𝑗 > 0, 𝑗 = 1, 2. On 𝔠 we have

𝐻 𝐵 = 𝐹 (0, 𝜉 ′◦ ) 𝐻𝐺 + 𝐺 (0, 𝑥 𝑛 , 𝜉 ′◦ ) 𝐻𝐹 .

We reason by contradiction. Suppose 𝐵 ≡ 0 on 𝔠, meaning that either 𝐹 (0, 𝜉 ′◦ ) = 0

or 𝐺 (0, 𝑥 𝑛 , 𝜉 ′◦ ) = 0 for all 𝑥 𝑛 ∈ (−𝑇1 , 𝑇2 ). If 𝐹 (0, 𝜉 ′◦ ) ≠ 0 then necessarily the
latter would be true, implying 𝐻𝐺 = 0 at every point of (0, 𝑥 𝑛 , 𝜉 ′◦ ) (since 𝐺 ≥ 0) and
therefore 𝐻 𝐵 ≡ 0 on 𝔠. This means that 𝔠 is a null bicharacteristic leaf of 𝑃 contrary
to our hypothesis. Thus we must assume that 𝐹 (0, 𝜉 ′◦ ) = 0, 𝐻𝐹 ≠ 0 on 𝔠 (𝐻𝐹 is
constant on 𝔠) and the zeros of 𝑥 𝑛 ↦→ 𝐺 (0, 𝑥 𝑛 , 𝜉 ′◦ ) are isolated. This implies that the
Lie algebra 𝔤 (𝐻 𝐴, 𝐻 𝐵 ) is spanned at every point of 𝔠 by the linearly independent,
commuting vector fields 𝐻 𝐴 and 𝐻𝐹 . If 𝛾 is an integral curve of 𝐻𝐹 intersecting 𝔠 we
must have 𝐴 ≡ 0 and 𝐹 ≡ 0 on 𝛾; it follows that 𝛾 ⊂ Char 𝑃. We conclude that the
Nagano leaf of 𝔤 (𝐻 𝐴, 𝐻 𝐵 ) is a surface Σ such that 𝔠 ⊂ Σ ⊂ Char 𝑃, contradicting
our hypothesis. This completes the proof of (1). Claim (2) is a consequence of (1),
as the latter implies that 𝐵 cannot vanish identically on any null bicharacteristic
curve of 𝐴 passing through points of a sufficiently small neighborhood of (𝑥 ◦ , 𝜉 ◦ ) in
𝑇 ∗ Ω\0. □
Taking Proposition 24.2.4 into account, inspection of the forthcoming proof of
Theorem 24.2.1 will show that it also proves the following microlocal result:

Theorem 24.2.5 Let 𝑃 = 𝑃 (𝑥, D) be a differential operator of principal type with

C 𝜔 coefficients in Ω that satisfy Condition (P) at (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃. If no null
bicharacteristic leaf of 𝑃 of positive dimension contains (𝑥 ◦ , 𝜉 ◦ ) then there is a conic
neighborhood U of (𝑥 ◦ , 𝜉 ◦ ) in 𝑇 ∗ Ω\0 such that U ∩ 𝑊 𝐹a (𝑢) = U ∩ 𝑊 𝐹a (𝑃𝑢) for
all 𝑢 ∈ D ′ (Ω).

24.3 Hypoellipticity Implies (Q)

Our starting point is exactly the same as in the proof of the sufficiency of Condition
(P) for local solvability (Subsections 23.2.4, 23.2.5): Let 𝑥 ◦ ∈ Ω, 𝜉 ◦ ∈ S𝑛−1 be such
that (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃, otherwise arbitrary. We select a suitably small neighborhood
994 24 Analytic PDEs of Principal Type. Regularity of the Solutions

of 𝑥 ◦ , 𝑈 ⊂ Ω, and a convex open cone in R𝑛 \ {0}, Γ ∋ 𝜉 ◦ . We assume that

d 𝜉 𝑃𝑚 (𝑥, 𝜉) ≠ 0 in 𝑈 × Γ and that the unitary transformation T in (23.2.40) has
√
been carried out: 𝑃𝑚 = 𝐸 ◦ 𝐿 in 𝑈 ×Γ with 𝐿 (𝑥, 𝜉) = −1𝜉 𝑛 − 𝐵 (𝑥, 𝜉 ′), 𝐸 ◦ (𝑥, 𝜉) ≠ 0
whatever (𝑥, 𝜉) ∈ 𝑈 × Γ, 𝐿 and 𝐸 homogeneous; thus

(𝑈 × Γ) ∩ Char 𝑃 = (𝑈 × Γ) ∩ Char 𝐿 (24.3.1)

= {(𝑥, 𝜉) ∈ 𝑈 × Γ; 𝜉 𝑛 = 0, 𝐵 (𝑥, 𝜉 ′) = 0} .

As in loc. cit., 𝑈 = 𝑈 ′ × (−𝜏, 𝜏) with 𝑈 ′ a neighborhood of 𝑥 ◦′ in R𝑛−1 and 𝜏 > 0; Γ

is defined in (23.2.42) with 𝜀 > 0 and Γ ′ a convex open cone in 𝜉 ′-space R𝑛−1 \ {0}
containing 𝜉 ◦′ ∈ S𝑛−2 ; thus 𝜉 ◦ = (𝜉 ◦′, 0), |𝜉 ◦′ | = 1.

Proposition 24.3.1 For the Hamiltonian vector fields 𝐻Re 𝑃𝑚 and 𝐻Im 𝑃𝑚 to be
collinear at (𝑥, (𝜉 ′, 0)) ∈ (𝑈 × Γ) ∩ Char 𝑃 it is necessary and sufficient that
d𝐵 (𝑥, 𝜉 ′) = 0.

Proof At every point of (𝑈 × Γ) ∩ Char 𝑃 the span of 𝐻Re 𝑃𝑚 and 𝐻Im 𝑃𝑚 is the same
as that of 𝐻 𝜉𝑛 = 𝜕𝑥𝜕𝑛 and 𝐻 𝐵 . To say that the latter are not collinear at (𝑥, (𝜉 ′, 0)) is
the same as saying that 𝐻 𝐵 ≠ 0, i.e., d𝐵 ≠ 0 at (𝑥, (𝜉 ′, 0)). □

Corollary 24.3.2 Along every null bicharacteristic curve of 𝑃𝑚 in 𝑈 × Γ we have

d𝐵 ≡ 0.

In this and later sections, unless otherwise specified the coefficients of 𝑃𝑚 (𝑥, D 𝑥 )
shall be of class C 𝜔 . This and Property (P) allow us to apply Proposition 23.1.39:

∀ (𝑥, 𝜉 ′) ∈ 𝑈 × Γ ′, 𝐵 (𝑥, 𝜉 ′) = 𝐺 (𝑥, 𝜉 ′) 𝐹 (𝑥 ′, 𝜉 ′) (24.3.2)

with C 𝜔 functions 𝐺 (𝑥, 𝜉 ′) ≥ 0 and 𝐹 (𝑥 ′, 𝜉 ′) homogeneous of degree 0 and 1

respectively. We have

𝐺 (𝑥, 𝜉 ′) = 0 =⇒ d𝐺 (𝑥, 𝜉 ′) = 0 =⇒ d𝐵 (𝑥, 𝜉 ′) = 0.

Thus, if (𝑥, 𝜉) ∈ (𝑈 × Γ) ∩ Char 𝑃 then d𝐵 (𝑥, 𝜉 ′) ≠ 0 is equivalent to

𝐺 (𝑥, 𝜉 ′) ≠ 0, 𝐹 (𝑥 ′, 𝜉 ′) = 0, d𝐹 (𝑥 ′, 𝜉 ′) ≠ 0. (24.3.3)

Proposition 24.3.3 Suppose (24.3.2) holds. If (𝑥 ◦ , 𝜉 ◦ ) does not lie on a null bichar-
acteristic curve of 𝑃 then 𝐵 (𝑥, 𝜉 ′) does not change sign in a neighborhood of (𝑥 ◦ , 𝜉 ◦′)
in 𝑈 × Γ ′.

Proof If (𝑥 ◦ , 𝜉 ◦ ) does not lie on a null bicharacteristic curve of 𝑃 then 𝐹 (𝑥 ′, 𝜉 ◦′) ≠ 0

and therefore 𝐹 (𝑥 ′, 𝜉 ′) ≠ 0 in a neighborhood 𝜔 ′ of (𝑥 ◦′, 𝜉 ′◦ ) in 𝑈 ′ × Γ ′ and thus 𝐹
does not change sign in 𝜔 ′, whence the claim by (24.3.2). □
In connection with Proposition 24.3.3 the following result will be needed later.
24.3 Hypoellipticity Implies (Q) 995

Proposition 24.3.4 Suppose that 𝐵 (𝑥, 𝜉 ′) does not change sign in 𝑈 × Γ ′. Then,
to every compact subset 𝐾 of 𝑈 and every open cone Γ◦′ in R𝑛−1 \ {0} such that
Γ◦′ ∩ S𝑛−2 ⊂⊂ Γ ′ there is a constant 𝐶𝐾 ,Γ◦′ > 0 such that
2
|𝜉 ′ | −1 |𝜕𝑥 𝐵 (𝑥, 𝜉 ′)| 2 + |𝜉 ′ | 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′) ≤ 𝐶𝐾 ,Γ◦′ |𝐵 (𝑥, 𝜉 ′)| (24.3.4)

for every (𝑥, 𝜉 ′) ∈ 𝐾 × Γ◦′ .

Proof We may assume 𝐵 ≥ 0 in 𝑈×Γ ′. Since both sides in (24.3.4) are homogeneous
of degree 1 it suffices to prove the claim when |𝜉 | = 1, in which case it reduces to
|d𝐵 (𝑥, 𝜉 ′)| 2 ≤ 𝐶𝐾 ,Γ◦′ 𝐵 (𝑥, 𝜉 ′). There is a 𝛿 > 0 such ′ ′
that (𝑥′ + 𝑦, 𝜉 + 𝜂 ) stays in a
′
compact subset of 𝑈 × Γ if (𝑥, 𝜉 ) ∈ 𝐾 × Γ◦ ∩ S ′ 𝑛−2 , |𝑦| + |𝜂 | ≤ 𝛿; Taylor expansion
yields, for some constant 𝑀 ≥ 0 and all such points,

0 ≤ 𝐵 (𝑥 + 𝑦, 𝜉 ′ + 𝜂 ′) ≤ 𝐵 (𝑥, 𝜉 ′) + 𝑦 · 𝜕𝑥 𝐵 (𝑥, 𝜉 ′) +𝜂 ′ · 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′) + 𝑀 |𝑥| 2 + |𝜉 ′ | 2 .

By selecting 𝑦 = −𝜀𝜕𝑥 𝐵 (𝑥, 𝜉 ′), 𝜂 ′ = −𝜀𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′), 𝜀 > 0 sufficiently small that
|𝑦| + |𝜂 ′ | ≤ 𝛿, we get

2
𝜀 (1 − 𝑀𝜀) |𝜕𝑥 𝐵 (𝑥, 𝜉 ′)| 2 + 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′) ≤ 𝐵 (𝑥, 𝜉 ′) ,

whence (24.3.4) if 𝜀 ≤ 21 𝑀 −1 , 𝐶𝐾 ,Γ◦′ ≥ 2𝜀 −1 . □

At this stage, taking 𝑥◦
= 0 simplifies the notation considerably: our “central”
point will be (0, (𝜉 ◦′, 0)) ∈ (𝑈 × Γ) ∩ Char 𝑃. We assume that the following holds:
(A1) 𝐵 (𝑥, 𝜉 ′) extends
holomorphically to an open subset 𝑈 C × Γ ′C of C𝑛 ×
C \ {0} , with 𝑈 C ∩ R𝑛 = 𝑈 and
𝑛−1

Γ ′C = 𝜁 ′ = 𝜉 ′ + 𝑖𝜂 ′ ∈ C𝑛−1 ; 𝜉 ′ ∈ Γ ′; |𝜂 ′ | < 𝜀◦′ |𝜉 ′ | (𝜀◦′ > 0).

We shall reason by contradiction and hypothesize that (a) holds (and thus the nec-
essary condition for hypoellipticity in Lemma 2.4.6 is satisfied) and, simultaneously,
that the curve 𝔠◦ defined by

𝜉 𝑛 = 0, 𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) = 0, |𝑥 𝑛 | < 𝜏, (24.3.5)

is a null bicharacteristic curve of 𝐿 (𝑥, D 𝑥 ) = 𝜕𝑥𝜕𝑛 − 𝐵 (𝑥, D 𝑥′ ) (hence of 𝑃) in 𝑈 × Γ.

By Proposition 24.3.1 the latter is equivalent to saying that d𝐵 ≡ 0 on 𝔠◦ .
Under this hypothesis we seek functions 𝑢 (𝑥, 𝜆) = e𝑖𝜆𝜑 ( 𝑥) 𝑎 (𝑥, 𝜆) ∈ Cc∞ (𝑈)
which violate the inequalities (2.4.6) as 𝜆 ↗ +∞. The (complex) phase-function
𝜑 ∈ C 𝜔 (𝑈) must solve the initial value problem

𝜕𝜑
𝑖 = 𝐵 (𝑥, 𝜕𝑥′ 𝜑) , 𝜑| 𝑥𝑛 =0 = 𝑥 ′ · 𝜉 ◦′ + 𝑖 |𝑥 ′ | 2 . (24.3.6)
𝜕𝑥 𝑛

We assume that 𝑈 is sufficiently small that 𝜕𝑥 𝜑 stays inside Γ ′C as 𝑥 ranges over 𝑈.

This has the important consequence that
996 24 Analytic PDEs of Principal Type. Regularity of the Solutions

∀𝑥 ∈ 𝑈, 𝑃𝑚 (𝑥, 𝜕𝑥 𝜑) = 0. (24.3.7)

Let us write 𝜑 (𝑥) = 𝑥 · 𝜉 ◦ + 𝑖𝜓 (𝑥) = 𝑥 ′ · 𝜉 ◦′ + 𝑖𝜓 (𝑥); 𝜓 must solve the initial

value problem

𝜕𝜓
+ 𝐵 (𝑥, 𝜉 ◦′ + 𝑖𝜕𝑥′ 𝜓 (𝑥)) = 0, 𝜓| 𝑥𝑛 =0 = |𝑥 ′ | 2 . (24.3.8)
𝜕𝑥 𝑛
Since 𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) ≡ 0 and d𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) ≡ 0 we see that

𝐵 (𝑥, 𝜉 ◦′ + 𝜕𝑥′ 𝜓 (𝑥)) = 𝐵 (𝑥, 𝜉 ◦′) + 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ◦′) · 𝜕𝑥′ 𝜓 (𝑥) + 𝑂 |𝜕𝑥′ 𝜓 (𝑥)| 2
1 ′ 2
𝑥 · 𝜕𝑥′ 𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) 𝑥 ′ + 𝜕 𝜉 ′ 𝜕𝑥′ 𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) 𝑥 ′ · 𝜕𝑥′ 𝜓 (𝑥)

=
2
+ 𝑂 |𝜕𝑥′ 𝜓 (𝑥)| 2 + |𝑥 ′ | 3 ,

whence, for some 𝐶1 > 0 independent of 𝑥 ∈ 𝑈,

|𝐵 (𝑥, 𝜉 ◦′ + 𝑖𝜕𝑥′ 𝜓 (𝑥))| ≤ 𝐶1 |𝑥 ′ | 2 + |𝜕𝑥′ 𝜓 (𝑥)| 2 . (24.3.9)

We derive from (24.3.8), for 𝑗 = 1, ..., 𝑛 − 1,

𝑛−1
𝜕2𝜓
∑︁
𝜕 𝜕𝜓 𝜕𝐵
+ (𝑥, 𝜉 ◦′ + 𝜕𝑥′ 𝜓 (𝑥))
𝜕𝑥 𝑛 𝜕𝑥 𝑗 𝑘=1
𝜕𝜉 𝑘 𝜕𝑥 𝑗 𝜕𝑥 𝑘
𝜕𝐵
+ (𝑥, 𝜉 ◦′ + 𝜕𝑥′ 𝜓 (𝑥)) = 0.
𝜕𝑥 𝑗

Putting 𝑥 ′ = 0 in this equation we get, for 𝑗 = 1, ..., 𝑛 − 1,

𝑛−1
𝜕2𝜓
∑︁
𝜕 𝜕𝜓 𝜕𝐵
(0, 𝑥 𝑛 ) + (0, 𝑥 𝑛 ) (0, 𝑥 𝑛 , 𝜉 ◦′ + 𝜕𝑥′ 𝜓 (0, 𝑥 𝑛 ))
𝜕𝑥 𝑛 𝜕𝑥 𝑗 𝑘=1
𝜕𝑥 𝑗 𝜕𝑥 𝑘 𝜕𝜉 𝑘

𝜕𝐵
+ (0, 𝑥 𝑛 , 𝜉 ◦′ + 𝜕𝑥′ 𝜓 (0, 𝑥 𝑛 )) = 0.
𝜕𝑥 𝑗
𝜕𝜓
We can regard this as a system of ODEs in the unknowns 𝜕𝑥 𝑗 (0, 𝑥 𝑛 ), 𝑗 = 1, ..., 𝑛 − 1.
Since 𝜕𝜓
𝜕𝑥 𝑗 (0) = 0 by (24.3.8) and since d𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) = 0 for all 𝑥 𝑛 ∈ (−𝜏, 𝜏) the
𝜕𝜓
uniqueness theorem for systems of ODEs implies 𝜕𝑥 𝑗 (0, 𝑥 𝑛 ) ≡ 0. Combining this
𝜕𝜓
with (24.3.5) and (24.3.6) shows that 𝜕𝑥𝑛 (0, 𝑥 𝑛 ) ≡ 0. Since 𝜓 (0) = 0 we obtain

∀𝑥 𝑛 ∈ (−𝜏, 𝜏) , 𝜓 (0, 𝑥 𝑛 ) = 0, d𝜓 (0, 𝑥 𝑛 ) = 0,

whence, by (24.3.8),
𝜓 (𝑥) = |𝑥 ′ | 2 (1 + 𝑂 (|𝑥 𝑛 − 𝑡|)) . (24.3.10)
24.4 Property (Q) Implies Subellipticity 997

If 𝜏 is sufficiently small (24.3.10) entails the following inequalities, for all 𝑥 ∈ 𝑈,

1 ′2
|𝑥 | ≤ Im 𝜑 (𝑥) = Re 𝜓 (𝑥) ≤ 2 |𝑥 ′ | 2 , (24.3.11)
2
|𝜕𝑥′ 𝜓 (𝑥)| 2 ≤ 𝐶2 Re 𝜓 (𝑥) , (24.3.12)

with 𝐶2 > 0 independent of 𝑥.

We make use of the formal power series 𝑎 (𝑧, 𝜆) = ∞ −𝑗
Í
𝑗=0 𝜆 𝑎 𝑗 (𝑧), the solution of
(23.2.14)–(23.2.15) under the Cauchy conditions (23.2.31), or rather of the approx-
imate solution e𝑖𝜆𝜑 (𝑧) 𝑎♭ (𝑧, 𝜆) of 𝑃 (𝑥, D 𝑥 ) ℎ = 0 [cf. (23.2.22)]. Proposition 23.2.8
implies, for all 𝜆 ≥ 1 and 𝑥 ∈ 𝐾1 ,

e𝑖𝜆𝜑 ( 𝑥) 𝑎♭ (𝑥, 𝜆) ≤ 𝐶e−𝜆 Im 𝜑 ( 𝑥) , (24.3.13)

∑︁
𝜕𝑥𝛼 𝑃 (𝑥, D 𝑥 ) e𝑖𝜆𝜑 ( 𝑥) 𝑎♭ (𝑥, 𝜆) ≤ 𝐶𝜆−𝑁2 e−𝜆 Im 𝜑 ( 𝑥) . (24.3.14)
| 𝛼 | ≤𝑁1

Putting this into (2.4.6) and taking (24.3.11) into account yields
∑︁
max D 𝑥𝛼 e𝑖𝜆𝜑 𝑎♭ ≤ 𝐶 1 + 𝜆−𝑁2 .
𝐾
| 𝛼 | ≤𝑁

From (24.3.6) and (24.3.10) we derive

∑︁ ∑︁
max D 𝛼 e𝑖𝜆𝜑 𝑎♭ (𝑥, 𝜆) ≥ 𝜆2 |D 𝛼 𝜑 (0)| − 𝐶 ′𝜆
𝐾
| 𝛼 |=2 | 𝛼 |=2

≥ 𝜆 − 𝐶 ′′𝜆.
2

If we take 𝑁 ≥ 2 and let 𝜆 ↗ +∞ the last two inequalities lead to a contradiction.

This completes the proof that non-(Q) entails non-(a).

24.4 Property (Q) Implies Subellipticity

24.4.1 Construction of a parametrix I. Complex phase-functions under

a no-change of sign hypothesis

We continue to reason within the framework of the preceding subsection and to

use the same notation. In particular, we assume that (A1) holds. We hypothesize,
moreover, that the following consequence of Property (Q), Theorem 24.2.1, holds
(cf. Proposition 24.3.3):

(A2) 𝐵 ≤ 0 everywhere in 𝑈 × Γ ′.
998 24 Analytic PDEs of Principal Type. Regularity of the Solutions

We are going to carry out a construction akin to, but with a different objective
from, that of the functions e𝑖𝜆𝜑 𝑎 in Subsection 23.2.2. Besides the hypothesis (A2)
a difference between the present situation and the situation in Ch. 23 is that more
variables will now be involved: 𝑡 ∈ (−𝜏, 𝜏), 𝜉 ′ ∈ Γ ′ with |𝜉 ′ | playing the role of 𝜆.
The more important difference is that, here, we have to deal with pseudodifferential,
in general not differential, operators. Here, when dealing with an analytic pseudo-
differential operator 𝑃 (assumed to be properly supported, for convenience), by a
parametrix at a point ℘ ∈ 𝑈 × Γ ′ we mean a linear operator 𝐺 : E ′ (𝑈) −→ D ′ (𝑈)
such that ℘ ∉ 𝑊 𝐹a (𝑃𝐺𝑢 − 𝑢) whatever 𝑢 ∈ E ′ (𝑈). The analogous terminology is
used in the C ∞ class, in which case ℘ ∉ 𝑊 𝐹 (𝑃𝐺𝑢 − 𝑢).
In this subsection we focus on the phase-function 𝜑. We will use the same
notation as in the preceding subsection for analogous objects: so 𝜑 (𝑡, 𝑥, 𝜉) ∈
C 𝜔 ((−𝜏, 𝜏) × 𝑈 × Γ) will be the unique analytic solution in (−𝜏, 𝜏) × 𝑈 × Γ of
the initial value problem [cf. (23.1.7), (24.3.6)]

𝜕𝜑
𝑖 = 𝐵 (𝑥, 𝜕𝑥′ 𝜑) , 𝜑| 𝑥𝑛 =𝑡 = 𝑥 ′ · 𝜉 ′. (24.4.1)
𝜕𝑥 𝑛
Obviously, 𝜑 is independent of 𝜉 𝑛 ; we write 𝜑 (𝑡, 𝑥, 𝜉) = 𝜑 (𝑡, 𝑥, 𝜉 ′) = 𝑥 ′ · 𝜉 ′ +
𝑖𝜓 (𝑡, 𝑥, 𝜉 ′); we have 𝜓 (𝑡, 𝑥, 𝜆𝜉 ′) = 𝜆𝜓 (𝑡, 𝑥, 𝜉 ′) (𝜆 > 0) and
𝜕𝜓
+ 𝐵 (𝑥, 𝜉 ′ + 𝑖𝜕𝑥′ 𝜓) = 0, 𝜓| 𝑥𝑛 =𝑡 = 0. (24.4.2)
𝜕𝑥 𝑛
Keep in mind that Re 𝜓 = Im 𝜑. Throughout the sequel we shall tacitly assume that
(𝑥 ′, 𝜉 ′) ∈ 𝑈 ′ × Γ ′ and (𝑥 𝑛 , 𝑡) ∈ (−𝜏, 𝜏) 2 entail 𝜕𝑥′ 𝜑 = 𝜉 ′ + 𝑖𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉 ′) ∈ Γ ′C [see
(A1)].

Proposition 24.4.1 If (A1) and (A2) hold and 𝜏 > 0 is sufficiently small then there
exist a neighborhood 𝑈◦′ ⊂⊂ 𝑈 ′ of 0, an open cone Γ◦′ ⊂ R𝑛−1 \ {0}, 𝜉 ◦′ ∈ Γ◦′ ∩
S𝑛−2 ⊂⊂ Γ ′ ∩ S𝑛−2 , and a constant 𝐶 > 0 such that
2
|𝜉 ′ | −1 |𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉 ′)| 2 + |𝜉 ′ | 𝜕 𝜉 ′ 𝜓 (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶 (𝑥 𝑛 − 𝑡) Re 𝜓 (𝑡, 𝑥, 𝜉 ′) , (24.4.3)
|Im 𝜓 (𝑡, 𝑥, 𝜉 ′)| 2 ≤ 𝐶 |𝜉 ′ | Re 𝜓 (𝑡, 𝑥, 𝜉 ′) , (24.4.4)

for all (𝑥 ′, 𝜉 ′) ∈ 𝑈◦′ × Γ◦′ and all (𝑥 𝑛 , 𝑡) ∈ (−𝜏, 𝜏) 2 such that −𝜏 < 𝑡 ≤ 𝑥 𝑛 < 𝜏.

Proof We derive from (24.4.2), for 𝑗 = 1, ..., 𝑛 − 1,

𝜕2𝜓 𝜕𝐵
+ (𝑥, 𝜉 ′ + 𝑖𝜕𝑥′ 𝜓)
𝜕𝑥 𝑛 𝜕𝑥 𝑗 𝜕𝑥 𝑗
𝑛−1
∑︁ 𝜕 2 𝜓 𝜕𝐵 𝜕𝜓
+𝑖 (𝑥, 𝜉 ′ + 𝑖𝜕𝑥′ 𝜓) = 0, = 0,
𝑘=1
𝜕𝑥 𝑗 𝜕𝑥 𝑘 𝜕𝜉 𝑘 𝜕𝑥 𝑗 𝑥𝑛 =𝑡
𝑛−1
𝜕2𝜓 𝜕2𝜓

∑︁ 𝜕𝐵 𝜕𝜓
+ 1+𝑖 (𝑥, 𝜉 ′ + 𝑖𝜕𝑥′ 𝜓) = 0, = 0.
𝜕𝑥 𝑛 𝜕𝜉 𝑗 𝑘=1 𝜕𝑥 𝑘 𝜕𝜉 𝑗 𝜕𝜉 𝑘 𝜕𝜉 𝑗 𝑥𝑛 =𝑡
24.4 Property (Q) Implies Subellipticity 999

This can be viewed as an initial value problem for a system of ODEs in the unknowns
𝜕𝜓 𝜕𝜓
𝜕𝑥 𝑗 , 𝜕 𝜉 𝑗 , leading to estimates of the following kind:
∫ 𝑥𝑛 ∫ 𝑥𝑛
|𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉)| ≤ 𝐶1 |𝜕𝑥′ 𝐵 (𝑥 ′, 𝑠, 𝜉 ′)| d𝑠 + 𝐶1 |𝜉 ′ | 𝜕 𝜉 ′ 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠,
𝑡 𝑡
∫ 𝑥𝑛
𝜕 𝜉 ′ 𝜓 (𝑡, 𝑥, 𝜉) ≤ 𝐶1 𝜕 𝜉 ′ 𝐵 (𝑥 , 𝑠, 𝜉 ′) d𝑠,
′
𝑡

and by the Cauchy–Schwarz inequalities,

∫ 𝑥𝑛
|𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉)| 2 ≤ 𝐶 (𝑥 𝑛 − 𝑡) |𝜕𝑥′ 𝐵 (𝑥 ′, 𝑠, 𝜉 ′)| 2 d𝑠
𝑡
∫ 𝑥𝑛
′ 2
+ 𝐶 |𝜉 | (𝑥 𝑛 − 𝑡) 𝜕 𝜉 ′ 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠,
𝑡
∫ 𝑥𝑛
2 2
𝜕 𝜉 ′ 𝜓 (𝑡, 𝑥, 𝜉) ≤ 𝐶 (𝑥 𝑛 − 𝑡) 𝜕 𝜉 ′ 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠.
𝑡

Applying (24.3.4) yields

∫ 𝑥𝑛
2
|𝜉 ′ | −1 |𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉)| 2 + |𝜉 ′ | 𝜕 𝜉 ′ 𝜓 (𝑡, 𝑥, 𝜉) ≤ −𝐶2 (𝑥 𝑛 − 𝑡) 𝐵 (𝑥, 𝜉 ′) d𝑥.
𝑡
(24.4.5)
We return to (24.4.2) and derive
𝜕𝜓
(𝑡, 𝑥, 𝜉 ′) + 𝐵 (𝑥, 𝜉 ′) − 𝑖 (𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉 ′)) · 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉)
𝜕𝑥 𝑛
∑︁ 1
− 𝑐 𝛼 (𝑡, 𝑥, 𝜉 ′) (𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉 ′)) 𝛼 = 0
𝛼!
| 𝛼 |=2

where 𝑐 𝛼 (𝑡, 𝑥, 𝜉 ′) is holomorphic and homogeneous of degree −1. Combined with

(24.4.5) this last equation implies
∫ 𝑥𝑛
Re 𝜓 (𝑡, 𝑥, 𝜉) + 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠 ≥ − (𝜕𝑥′ Im 𝜓 (𝑡, 𝑥, 𝜉)) · 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉) (24.4.6)
𝑡
∫ 𝑥𝑛
− 𝐶3 (𝑥 𝑛 − 𝑡) 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠 .
𝑡

We also deduce, from (24.4.2) and (24.4.5),

∫ 𝑥𝑛
2 ′
(𝜕 Im 𝜓 (𝑡, 𝑥, 𝜉)) · 𝜕 𝐵 (𝑥, 𝜉)
𝑥′ 𝜉′ ≤ 𝐶4 (𝑥 𝑛 − 𝑡) 𝐵 (𝑥, 𝜉 ) 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠.
𝑡

If 𝜏 > 0 is sufficiently small and −𝜏 < 𝑥 𝑛 ≤ 𝑡 < 𝜏 then the right-hand side in
(24.4.6) will be bounded from below by
1000 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫ 𝑥𝑛 ∫ 𝑥𝑛
1 ′ 1 ′
− 𝐵 (𝑥 , 𝑠, 𝜉 ) d𝑠 = 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠,
2 𝑡 2 𝑡

whence ∫ 𝑥𝑛
− 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠 ≤ 2 Re 𝜓 (𝑡, 𝑥, 𝜉 ′) . (24.4.7)
𝑡
Putting this last estimate into (24.4.5) yields (24.4.3). Since
∫ 𝑥𝑛
′
𝜓 (𝑡, 𝑥, 𝜉 ) = − 𝐵 (𝑥 ′, 𝑠, 𝜉 ′) d𝑠
𝑡
∫ 𝑥𝑛
− (𝜕𝑥′ 𝜓 (𝑠, 𝑥, 𝜉)) · 𝜕 𝜉 ′ 𝐵 (𝑥 ′, 𝑠, 𝜉) d𝑠 + 𝑂 |𝜕𝑥′ 𝜓| 2
𝑡

we derive from (24.4.7) and (24.4.3)

|Im 𝜓 (𝑡, 𝑥, 𝜉 ′)| ≤ 2 Re 𝜓 (𝑡, 𝑥, 𝜉 ′) + 𝐶5 |𝜕𝑥′ 𝜓 (𝑡, 𝑥, 𝜉)|

√︁ √︁ √︁
≤ 2 Re 𝜓 (𝑡, 𝑥, 𝜉 ′) + 𝐶6 |𝜉 ′ | Re 𝜓 (𝑡, 𝑥, 𝜉 ′),

whence (24.4.4). □
Proposition 24.4.2 Assume the same hypotheses as in Proposition 24.4.1 and the
same choice of 𝑈◦′ , Γ◦′ and 𝜏. To every pair (𝛼, 𝛽) ∈ Z+𝑛−1 × Z+𝑛−1 there is a 𝐶 𝛼,𝛽 > 0
such that
′
∑︁ 1 ′
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) ≤ 𝐶 𝛼,𝛽 ((𝑥 𝑛 − 𝑡) |𝜉 ′ |) ℓ |𝜉 ′ | − |𝛽 | e− 2 Re 𝜓 (𝑡 , 𝑥, 𝜉 )
𝛽

ℓ ≤ 21 | 𝛼+𝛽 |
(24.4.8)
for every 𝑥 ′ ∈ 𝑈◦′ , 𝜉 ′ ∈ Γ◦′ , |𝜉 ′ | ≥ 1, and all (𝑥 𝑛 , 𝑡) ∈ (−𝜏, 𝜏) 2 such that −𝜏 < 𝑡 ≤
𝑥 𝑛 < 𝜏.
Proof We claim that
′)
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉
𝛽
(24.4.9)
∑︁ ∑︁ ∑︁ 𝛽˜ ′)
= 𝑞 𝛼, 𝛼,𝛽,
˜ 𝛽,𝑘
′
˜ (𝑡, 𝑥, 𝜉 ) (𝜕𝑥 ′ 𝜓)
𝛼˜
𝜕𝜉 ′ 𝜓 e−𝜓 (𝑡 , 𝑥, 𝜉
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑘 ≤ 1 | 𝛼− 𝛼+𝛽−
˜ 𝛽˜ |
2

where the 𝑞 𝛼, 𝛼,𝛽, ˜ are linear combinations (with coefficients in Z) of “monomials”

˜ 𝛽,𝑘
in the partial derivatives of 𝜓 with respect to (𝑥 ′, 𝜉 ′) of the following type:
𝑘
𝛽 [ℓ ]
Ö [ℓ ]
𝔐 (𝑡, 𝑥, 𝜉 ′) = 𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝜓, 𝛼 [ℓ ] , 𝛽 [ℓ ] ∈ Z+𝑛−1 , (24.4.10)
ℓ=1
𝑘
∑︁ 𝑘
∑︁
𝛼 [ℓ ] = 𝛼 − 𝛼,
˜ 𝛽 [ℓ ] = 𝛽 − 𝛽,
˜ 𝛼 [ℓ ] + 𝛽 [ℓ ] ≥ 2;
ℓ=1 ℓ=1

𝔐 (𝑡, 𝑥, 𝜉 ′) is homogeneous (with respect to 𝜉 ′) of degree 𝑘 − 𝛽 − 𝛽˜ .

24.4 Property (Q) Implies Subellipticity 1001

To prove the claim we reason by induction on |𝛼 + 𝛽|, the statement being trivial
when |𝛼 + 𝛽| = 0. We have, for 1 ≤ 𝑗 ≤ 𝑛 − 1,
𝛽˜ ˜ 𝑗⟩ 𝛽˜
e 𝜓 𝜕𝑥 𝑗 e−𝜓 (𝜕𝑥′ 𝜓) 𝛼˜ 𝜕 𝜉 ′ 𝜓 𝔐 = − (𝜕𝑥′ 𝜓) 𝛼+⟨ 𝜕𝜉 ′ 𝜓 𝔐 (24.4.11)
˜
+ (𝜕𝑥′ 𝜓) 𝛼˜ 𝜕 𝜉 ′ 𝜓 𝜕𝑥 𝑗 𝔐
𝛽
∑︁
˜ ⟨ 𝑗 ⟩− ⟨ 𝑗 ′ ⟩ 𝛽˜
+ (𝜕𝑥′ 𝜓) 𝛼− 𝜕𝜉 ′ 𝜓 𝜕𝑥 𝑗 𝜕𝑥 𝑗 ′ 𝜓 𝔐
⟨ 𝑗 ′ ⟩ ⪯ 𝛼−
˜ ⟨𝑗⟩
˜ 𝑗′ ⟩
∑︁
˜ ⟨𝑗⟩ 𝛽−⟨
+ (𝜕𝑥′ 𝜓) 𝛼− 𝜕𝜉 ′ 𝜓 𝜕𝑥 𝑗 𝜕 𝜉 𝑗 ′ 𝜓 𝔐
⟨ 𝑗 ′ ⟩ ⪯ 𝛽˜

(⟨ 𝑗⟩ : the multi-index whose entries are all 0 except the 𝑗 th one, equal to 1). The
last two sums in the right-hand side of (24.4.11) are absent if 𝛼 𝑗 = 0. In (24.4.10)
𝛼 [ℓ ] = 𝛼 + ⟨ 𝑗⟩ − ( 𝛼˜ + ⟨ 𝑗⟩) and therefore the first term at the right in
Í𝑘
we have ℓ=1
𝛽˜
(24.4.11) is of the same type as (𝜕𝑥′ 𝜓) 𝛼˜ 𝜕 𝜉 ′ 𝜓 𝔐. About the second term we have
𝑘 Ö
𝑘
𝛼 [ℓ ] +⟨ 𝑗 ⟩ 𝛽 [ℓ ]
∑︁
𝜕𝑥 𝑗 𝔐 = 𝜕𝑥 ′ 𝜕𝜉 ′ 𝜓
𝑗 ′ =1 ℓ=1

𝛼 [ℓ ] + ⟨ 𝑗⟩ = 𝛼 + ⟨ 𝑗⟩ − 𝛼.
Í𝑘
and ℓ=1 ˜ Supposing 𝛼 𝑗 ≥ 1 we see that the last two sums in
the right-hand side of (24.4.11) are of the same type as 𝔐 except
that 𝑘 is replaced
by 𝑘 + 1 and 𝛼 [𝑘+1] ′
= ⟨ 𝑗⟩ + ⟨ 𝑗 ⟩, 𝛽 [𝑘+1] = 0 in 𝜕𝑥 𝑗 𝜕𝑥 𝑗′ 𝜓 𝔐 and 𝛼 [𝑘+1] = ⟨ 𝑗⟩,

𝛽 [𝑘+1] = ⟨ 𝑗 ′⟩ in 𝜕𝑥 𝑗 𝜕 𝜉 𝑗′ 𝜓 𝔐. We have

𝑘+1
∑︁ 𝑘
∑︁ 𝑘+1
∑︁
𝛼 [ℓ ] = 𝛼 [ℓ ] + ⟨ 𝑗⟩ + ⟨ 𝑗 ′⟩ = 𝛼 + ⟨ 𝑗⟩ − ( 𝛼˜ − ⟨ 𝑗 ′⟩) , 𝛽 [ℓ ] = 𝛽 − 𝛽,
˜
ℓ=1 ℓ=1 ℓ=1
𝑘+1
∑︁ 𝑘
∑︁ 𝑘+1
∑︁ 𝑘
∑︁
𝛼 [ℓ ] = 𝛼 [ℓ ] + ⟨ 𝑗⟩ = 𝛼 + ⟨ 𝑗⟩ − 𝛼, 𝛽 [ℓ ] + ⟨ 𝑗 ′⟩ = = 𝛽 − 𝛽˜ − ⟨ 𝑗 ′⟩ ,

˜
ℓ=1 ℓ=1 ℓ=1 ℓ=1

in 𝜕𝑥 𝑗 𝜕𝑥 𝑗′ 𝜓 𝔐 and 𝜕𝑥 𝑗 𝜕 𝜉 𝑗′ 𝜓 𝔐 respectively. This proves the claim for
(𝛼 + ⟨ 𝑗⟩ , 𝛽). The proof for (𝛼, 𝛽 + ⟨ 𝑗⟩) is the same except for the exchange of
𝑥 ′ and 𝜉 ′. We reach the following conclusion, for some 𝐶 𝛼,𝛽
′ > 0 independent of
′
𝑡, 𝑥, 𝜉 and of the monomial 𝔐,
˜
|𝔐 (𝑡, 𝑥, 𝜉 ′)| ≤ 𝐶 𝛼,𝛽
′
(𝑥 𝑛 − 𝑡) 𝑘 |𝜉 ′ | 𝑘− | 𝛽−𝛽 | .

From (24.4.3) we derive

𝛽˜ 2 ˜
˜ 𝛽 | ′ | 𝛼˜ |− | 𝛽 |
˜ ˜
(𝜕𝑥′ 𝜓) 𝛼˜ 𝜕 𝜉 ′ 𝜓 ≤ (𝐶 (𝑥 𝑛 − 𝑡)) | 𝛼+ |𝜉 | (Re 𝜓) | 𝛼+
˜ 𝛽|
.
1002 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Putting these last two estimates into (24.4.9) yields

𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓
𝛽

1
∑︁ ∑︁ ˜
′′
≤ 𝐶 𝛼,𝛽 ((𝑥 𝑛 − 𝑡) |𝜉 ′ |) ℓ |𝜉 ′ | − |𝛽 | (Re 𝜓) 2 | 𝛼+
˜ 𝛽 | − Re 𝜓
e ,
𝛼˜ ⪯ 𝛼 1
˜ 𝛽˜ | ≤ℓ ≤ 12 | 𝛼+𝛽 |
| 𝛼+
2
𝛽˜ ⪯𝛽

1

˜
whence (24.4.8) since (Re 𝜓) 2 | 𝛼+
˜ 𝛽 | − 2 Re 𝜓 1 1 ˜
e ≤ 2 2 | 𝛼+
˜ 𝛽|
Γ 1
2 𝛼˜ + 𝛽˜ + 1 . □

Corollary 24.4.3 Assume the same hypotheses as in Proposition 24.4.1. To every

pair (𝛼, 𝛽) ∈ Z+𝑛−1 × Z+𝑛−1 there is a 𝐶 𝛼,𝛽 > 0 such that
′) 1 1
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | 2 | 𝛼 |− 2 |𝛽 |
𝛽
(24.4.12)

for every 𝑥 ′ ∈ 𝑈◦′ , 𝜉 ′ ∈ Γ◦′ , |𝜉 ′ | ≥ 1, and all (𝑥 𝑛 , 𝑡) ∈ (−𝜏, 𝜏) 2 such that −𝜏 < 𝑡 ≤
𝑥 𝑛 < 𝜏.
1
Proof In (24.4.8) we have ((𝑥 𝑛 − 𝑡) |𝜉 ′ |) ℓ ≤ |𝜉 ′ | 2 | 𝛼+𝛽 | since |𝑥 𝑛 − 𝑡| ≤ 𝜏 < 1 and
|𝜉 ′ | ≥ 1. □

24.4.2 Application to the action of 𝑩 (𝒙, D 𝒙′ )

We are going to exploit Corollary 24.4.3 in connection with integrals

′ ′
′

(2𝜋) 𝑛−1 e−𝑖 𝑥 · 𝜉 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑢 (𝑥 ′)
∫
′ ′ ′ ′ ′
= e𝑖 ( 𝑥 −𝑦 ) · ( 𝜃 − 𝜉 )−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝐵 (𝑥, 𝜃 ′) 𝜒2 (𝜃 ′) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′
R2𝑛−2

in which 𝜒1 and 𝜒2 are the cut-off functions introduced in Subsection 24.1.3 and
𝑢 ∈ C ∞ (𝑈 ′). If we change variables from 𝜃 ′ to 𝜆 (𝜃 ′ + 𝜉 ′/𝜆), 𝜆 = |𝜉 ′ |, and take into
account the homogeneity of 𝐵 (𝑥, 𝜃 ′) 𝜒2 (𝜃 ′) we get
′ ′
′

(2𝜋) 𝑛−1 e−𝑖 𝑥 · 𝜉 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜉 /𝜆) 𝜒1 (𝑥 ′) 𝑢 (𝑥 ′)
∫
′ ′ ′ ′
= 𝜆𝑛 e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 −𝜆𝜓 (𝑡 ,𝑦, 𝜉 /𝜆) 𝐵♭ (𝑥, 𝜃 ′) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′,
R2𝑛−2

where
𝐵♭ (𝑥, 𝜃 ′) = 𝐵 (𝑥, 𝜃 ′ + 𝜉 ′/𝜆) 𝜒2 (𝜃 ′ + 𝜉 ′/𝜆) . (24.4.13)
To lighten the notation in the coming estimates it is convenient to temporarily dis-
regard the variables 𝑡, 𝜉 ′/𝜆, and let 𝜆𝜑 (𝑥) stand for 𝜑 (𝑡, 𝑥, 𝜉 ′), 𝜆𝜓 (𝑥) for 𝜓 (𝑡, 𝑥, 𝜉 ′).
Thus we are dealing with an integral
24.4 Property (Q) Implies Subellipticity 1003
′ ′

𝜆−1 e−𝑖 𝑥 · 𝜉 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖𝜆𝜑 ( 𝑥) 𝜒1 (𝑥 ′) 𝑢 (𝑥 ′) (24.4.14)
𝑛−1 ∫
𝜆 ′ ′ ′ ′
= e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 −𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝐵♭ (𝑥, 𝜃 ′) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′.
2𝜋 R2𝑛−2

Let 𝜌 > 0 be small; we decompose the integral on the right in (24.4.14):

𝑛−1 ∫
𝜆 ′ ′ ′ ′
e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 −𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝐵♭ (𝑥, 𝜃 ′) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′
2𝜋 R2𝑛−2
= 𝑩˜ 𝜌 𝑢 (𝑥, 𝜆) + 𝑹˜ 𝜌 𝑢 (𝑥, 𝜆)

where

𝑩˜ 𝜌 𝑢 (𝑥, 𝜆)
𝑛−1 ∫ ∫
𝜆 ′ ′ ′ ′
= e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 −𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝐵♭ (𝑥, 𝜃 ′) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′,
2𝜋 R𝑛−1 | 𝜃 ′ |<𝜌

𝑹˜ 𝜌 𝑢 (𝑥, 𝜆)
𝑛−1 ∫ ∫
𝜆 ′ ′ ′ ′
= e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 −𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝐵♭ (𝑥, 𝜃 ′) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′.
2𝜋 R𝑛−1 | 𝜃 ′ |>𝜌

Both 𝑩˜ 𝜌 𝑢 and 𝑹˜ 𝜌 𝑢 depend on 𝑡 as well as on 𝜉 ′. We begin with an estimate of 𝑹˜ 𝜌 𝑢.

Lemma 24.4.4 Let 𝑈◦′ , Γ◦′ and 𝜏 be as in Proposition 24.4.1. To every 𝑢 ∈ Cc∞ 𝑈◦′

and 𝑁 ∈ Z+ there is a 𝐶 𝑁 (𝑢, 𝜌) > 0 such that

∀𝜆 ∈ [1, +∞), 𝑹˜ 𝜌 𝑢(𝑥, 𝜆) ≤ 𝐶 𝑁 (𝑢, 𝜌) 𝜆−𝑁 −1

for all 𝑥 ′ ∈ 𝑈◦′ , 𝜉 ′/𝜆 ∈ Γ◦′ and (𝑡, 𝑥 𝑛 ) such that −𝜏 ≤ 𝑡 < 𝑥 𝑛 ≤ 𝜏.

Proof We see that (2𝜋) 𝑛−1 𝜆2𝑁 +𝑛+1 𝑹˜ 𝜌 𝑢(𝑥, 𝜆) is equal to

∫ ∫
′ ′ ′
e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) Δ 𝑦𝑁′ +𝑛 e−𝑖𝜆𝑦 · 𝜃 𝑔 𝑁 (𝑥, 𝜃 ′)d𝑦 ′d𝜃 ′
R𝑛−1 | 𝜃 ′ |>𝜌
∫ ∫
′ ′ ′
= e−𝑖𝜆𝑦 · 𝜃 Δ 𝑦𝑁′ +𝑛 e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) 𝑔 𝑁 (𝑥, 𝜃 ′)d𝑦 ′d𝜃 ′,
R𝑛−1 | 𝜃 ′ |>𝜌

where 𝑔 𝑁 (𝑥, 𝜃 ′) = 𝐵♭ (𝑥, 𝜃 ′) |𝜃 ′ | −2𝑁 −2𝑛 . Corollary 24.4.3 and a straightforward

application of the Leibniz rule leads to an estimate
∫
′
Δ 𝑦𝑁′ +𝑛 e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′ ≤ 𝐶 𝑁
′
(𝑢) 𝜆 𝑁 +𝑛 (24.4.15)
R𝑛−1
′ (𝑢) > 0 independent of all variables, 𝐶 ′ (𝑠𝑢) = 𝑠𝐶 ′ (𝑢) for all 𝑠 ≥ 0], whence
[𝐶 𝑁 𝑁 𝑁
1004 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫
′
𝑹˜ 𝜌 𝑢(𝑥, 𝜆) ≤ 𝐶 𝑁 (𝑢) 𝜆−𝑁 −1 𝐵♭ (𝑥, 𝜃 ′) |𝜃 ′ | −2𝑁 −2𝑛 d𝜃 ′.
| 𝜃 ′ |>𝜌

Since 𝐵♭ (𝑥, 𝜃 ′) ≤ const. (1 + |𝜃 ′ |) and 𝑛 ≥ 2 the claim is proved. □

We now turn our attention to 𝑩˜ 𝜌 𝑢. If 𝜌 and 𝛿 are sufficiently small |𝜉 ′/𝜆 − 𝜉 ◦′ | ≤ 𝛿,

|𝜃 ′ |< 𝜌 imply 𝜒2 (𝜃 ′ + 𝜉 ′/𝜆) = 1, and by Hypothesis (A1), that the Taylor expansion
∑︁ 1 ′𝛼 𝛼
𝐵 (𝑥, 𝜃 ′ + 𝜉 ′/𝜆) = 𝜃 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′/𝜆)
𝛼!
𝛼∈Z+𝑛−1

converges absolutely uniformly. By (24.4.13) we have

∑︁ 1 𝑛−| 𝛼 | 𝛼
𝑩˜ 𝜌 𝑢 (𝑥, 𝜆) = 𝜆 𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′/𝜆) 𝑲 𝜌[ 𝛼] 𝑢 (𝑥, 𝜆) ,
𝛼!
𝛼∈Z+𝑛−1

where

𝑲 𝜌[ 𝛼] 𝑢 (𝑥, 𝜆)
𝑛−1 ∫ ∫
−|𝛼| 𝜆 ′ ′ ′ ′
= (−1) e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) D 𝑦𝛼′ e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 d𝑦 ′d𝜃 ′
2𝜋 R 𝑛−1 ′
| 𝜃 |<𝜌
𝑛−1 ∫ ∫
𝜆 ′ ′ ′ ′
= e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 D 𝑦𝛼′ e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′
2𝜋 R𝑛−1 | 𝜃 ′ |<𝜌
𝑛−1 ∫
𝜆 ′ ′ ′ ′
= e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 D 𝑦𝛼′ e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′
2𝜋 R2𝑛−2
𝑛−1 ∫ ∫
𝜆 ′ ′ ′ ′
− e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 D 𝑦𝛼′ e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′.
2𝜋 R𝑛−1 | 𝜃 ′ |>𝜌

If we apply the Fourier inversion formula in R𝑛−1 and restrict the variation of 𝑥 ′ to
a neighborhood of 0, 𝑉 ′ ⊂⊂ 𝑈◦′ , where 𝜒1 ≡ 1, we get

D 𝑥𝛼′ e−𝜆𝜓 ( 𝑥) 𝑢 (𝑥 ′) (24.4.16)
𝑛−1 ∫
𝜆 ′ ′ ′ ′
= e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 D 𝑦𝛼′ e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′.
2𝜋 R2𝑛−2

Lemma 24.4.5 Let 𝑈◦′ , Γ◦′ and 𝜏 be as in Proposition 24.4.1. To every 𝑢 ∈ Cc∞ (𝑉 ′)
and 𝑁 ∈ Z+ there is a 𝐶 𝛼, 𝑁 (𝑢, 𝜌) > 0 such that
∫ ∫
′ ′ ′ ′
e𝑖𝜆( 𝑥 −𝑦 ) · 𝜃 D 𝑦𝛼′ e−𝜆𝜓 ( 𝑦 , 𝑥𝑛 ) 𝜒1 (𝑦 ′) 𝑢 (𝑦 ′) d𝑦 ′d𝜃 ′
R𝑛−1 | 𝜃 ′ |>𝜌
≤ 𝐶 𝛼, 𝑁 (𝑢, 𝜌) 𝜆−𝑛−𝑁

for all 𝜆 ≥ 1, 𝑥 ′ ∈ 𝑈◦′ , 𝜉 ′/𝜆 ∈ Γ◦′ and (𝑡, 𝑥 𝑛 ) such that −𝜏 ≤ 𝑡 < 𝑥 𝑛 ≤ 𝜏.
24.4 Property (Q) Implies Subellipticity 1005

The proof of Lemma 24.4.5 is the same as that of Lemma 24.4.4 with the added
simplification that 𝐵♭ is replaced by 1.
We shall now use to mean equivalent mod formal series powers ∞ −𝑗
Í
𝑗=0 𝜆 𝑐 𝑗 (𝑥, 𝜆)
whose coefficients 𝑐 𝑗 (𝑥, 𝜆) = 𝑐 𝑗 (𝑡, 𝑥, 𝜉 ′/𝜆, 𝜆) decay as 𝜆 ↗ +∞ faster than any
power of 𝜆−1 uniformly with respect to (𝑥 ′, 𝜉 ′/𝜆) ∈ 𝑈◦′ × Γ◦′ and to (𝑡, 𝑥 𝑛 ) such that
−𝜏 ≤ 𝑡 < 𝑥 𝑛 ≤ 𝜏. Combining Lemmas 24.4.4, 24.4.5 and (24.4.16) yields

𝑩˜ 𝜌 𝑢 (𝑥, 𝜆) + 𝑹˜ 𝜌 𝑢 (𝑥, 𝜆)
∑︁ 1
𝜆 𝑛−| 𝛼 | 𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′/𝜆) D 𝑥𝛼′ e−𝜆𝜓 ( 𝑥) 𝑢 (𝑥 ′) .
𝑛−1
𝛼!
𝛼∈Z+

If we combine this with (24.4.14) and (24.4.16) and take into account the fact that
𝜒2 ≡ 1 in Γ◦′ we obtain
′ ′

e−𝑖 𝑥 · 𝜉 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖𝜆𝜑 ( 𝑥) 𝜒1 (𝑥 ′) 𝑢 (𝑥 ′)
∑︁ 1
𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) D 𝑥𝛼′ e−𝜆𝜓 ( 𝑥) 𝑢 (𝑥 ′)
𝑛−1
𝛼!
𝛼∈Z+

and, finally, back to the complete set of variables 𝑡, 𝑥, 𝜉 ′,

′

𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑢 (𝑥 ′) (24.4.17)
∑︁ 1 𝛼 𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ′ )
𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) D 𝑥′ − 𝜉 ′ e 𝑢 (𝑥 ′) ,
𝑛−1
𝛼!
𝛼∈Z+

where 𝑥 ′ ∈ 𝑉 ′, 𝜉 ′ ∈ Γ◦′ , −𝜏 ≤ 𝑡 < 𝑥 𝑛 ≤ 𝜏.

A last remark will be useful: what precedes is valid (and simpler) if we replace
𝐵 (𝑥, D 𝑥′ ) by 𝜕𝑥𝜕𝑛 and therefore we can use the congruence

𝜕 𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ′ ) 𝜕 𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ′ )
𝜒2 (D 𝑥′ ) e 𝜒1 (𝑥 ′) 𝑢 (𝑥 ′) e 𝑢 (𝑥 ′) (24.4.18)
𝜕𝑥 𝑛 𝜕𝑥 𝑛
valid for the same 𝑥, 𝑡, 𝜉 ′. In the next subsection these congruences will be applied
to functions like 𝑢 but depending also on 𝑡, 𝑥 𝑛 , 𝜉 ′.

24.4.3 Construction of a parametrix II. Transport equations and

formal series amplitudes

We now determine a classical amplitude (Definition 17.2.28)

∞
∑︁
𝑎 (𝑡, 𝑥, 𝜉 ′) = 𝑎 𝑗 (𝑡, 𝑥, 𝜉 ′)
𝑗=0
1006 24 Analytic PDEs of Principal Type. Regularity of the Solutions

with 𝑎 𝑗 ∈ C 𝜔 ((−𝜏, 𝜏) × 𝑈 × Γ ′) homogeneous of degree − 𝑗, satisfying the (formal

series) equation

𝜕 ′
− 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑎 (𝑡, 𝑥, 𝜉 ′) (24.4.19)
𝜕𝑥 𝑛
∞
∑︁ 𝜕 ′
= − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑎 𝑗 (𝑡, 𝑥, 𝜉 ′) = 0
𝑗=0
𝜕𝑥 𝑛

under the Cauchy conditions

𝑎 0 | 𝑥𝑛 =𝑡 = 1, 𝑎 𝑗 𝑥𝑛 =𝑡
= 0 if 𝑗 ≥ 1. (24.4.20)

By expansion in formal series in the powers of 𝜆−1 , 𝜆 = |𝜉 ′ |, (24.4.19) implies

∞ ∞
© 𝜕𝜑 ∑︁ 1− 𝑗 ∑︁ 𝜕𝑎 𝑗 ª
𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) ­𝑖 𝜆 𝑎𝑗 + 𝜆− 𝑗 ®
𝜕𝑥 𝑛 𝑗=0 𝑗=0
𝜕𝑥 𝑛
« ¬
∞
∑︁
− e−𝑖𝜆𝜑 𝜆− 𝑗 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) e𝑖𝜆𝜑 𝜒1 (𝑥 ′) 𝑎 𝑗 = 0.
𝑗=0

We reason under the hypothesis that 𝑥 ′ ∈ 𝑉 ′, −𝜏 ≤ 𝑡 < 𝑥 𝑛 ≤ 𝜏, 𝜉 ′ ∈ Γ◦′ (recall

that 𝜒1 ≡ 1 in 𝑉 ′ and 𝜒2 ≡ 1 in Γ◦′ ). We replace 𝑢 (𝑥 ′) by 𝑎 (𝑡, 𝑥, 𝜉 ′) in (24.4.17)–
(24.4.18). This yields the following congruence mod formal series in the powers of
𝜆−1 whose coefficients depend on (𝑡, 𝑥, 𝜉 ′) and decay as 𝜆 = |𝜉 ′ | ↗ +∞ faster than
any power of 𝜆−1 uniformly with respect to (𝑡, 𝑥, 𝜉 ′/𝜆),

𝜕𝜑 𝜕𝑎 ∑︁ 1 𝛼
𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′) D 𝑥′ + 𝑖𝜕𝑥′ 𝜓 𝑎 0.
𝛼
𝑖 𝑎+ −
𝜕𝑥 𝑛 𝜕𝑥 𝑛 𝛼!
𝛼∈Z+𝑛−1

Since 𝜑 = 𝑥 ′ · 𝜉 ′ + 𝑖𝜓 is homogeneous of degree 1 and 𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) is homogeneous

of degree 1 − |𝛼| the eikonal equation can be rewritten as

𝜕𝜑 ∑︁ 1
𝑖 = 𝜕 𝛼′ 𝐵 (𝑥, 𝜉 ′) (𝑖𝜕𝑥′ 𝜓) 𝛼 .
𝜕𝑥 𝑛 𝑛−1
𝛼! 𝜉
𝛼∈Z+

The above congruence is seen to be equivalent to

𝜕𝑎 ∑︁
𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) 𝑄 [ 𝛼] 𝑎 (24.4.21)
𝜕𝑥 𝑛 𝑛−1
𝛼∈Z+

where
24.4 Property (Q) Implies Subellipticity 1007

1
𝑄 [ 𝛼] 𝑎 =
𝛼
D 𝑥′ + 𝑖𝜕𝑥′ 𝜓 − (𝑖𝜕𝑥′ 𝜓) 𝛼 𝑎
𝛼!
1 1
= e 𝜓 D 𝑥𝛼′ e−𝜓 𝑎 −

(𝑖𝜕𝑥′ 𝜓) 𝛼 𝑎.
𝛼! 𝛼!
At this juncture it is convenient to reintroduce the abbreviated notation of the
preceding subsection: ignoring the parameters 𝑡 , 𝜉 ′/𝜆, and writing (𝑥)
Í 𝜆𝜑− 𝑗 for
𝜑 (𝑡, 𝑥, 𝜉 ′), 𝜆𝜓 (𝑥) for 𝜓 (𝑡, 𝑥, 𝜉 ′), 𝑎 𝑗 (𝑥) for 𝑎 𝑗 (𝑡, 𝑥, 𝜉 ′/𝜆); thus 𝑎 = ∞
𝑗=0 𝜆 𝑎 𝑗 (𝑥)
and (24.4.21) reads
∞ ∞
∑︁ 𝜕𝑎 𝑗 ∑︁ ∑︁
𝜆− 𝑗 = 𝜆− 𝑗 𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) 𝑄 [ 𝛼] 𝑎 𝑗 . (24.4.22)
𝑗=0
𝜕𝑥 𝑛 𝑗=0 𝑛−1 𝛼∈Z+

We must take a closer look at 𝑄 [ 𝛼] 𝑎 𝑗 . By the Leibniz rule,

𝑄 [ 𝛼] 𝑎 𝑗 (𝑥, 𝜆) (24.4.23)
∑︁ 1
e𝜆𝜓 ( 𝑥) D 𝑥′ e−𝜆𝜓 ( 𝑥) D 𝑥′ 𝑎 𝑗 (𝑥) − (𝑖𝜆𝜕𝑥′ 𝜓 (𝑥)) 𝛼 𝑎 𝑗 (𝑥) .
𝛽 𝛼−𝛽
=
𝛽⪯𝛼
𝛽! (𝛼 − 𝛽)!

We are going to need the following

Lemma 24.4.6 We have, for all 𝛼 ∈ Z+𝑛−1 and 𝜆 ≥ 1,

e𝜆𝜓 D 𝑥𝛼′ e−𝜆𝜓 = (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼 (24.4.24)
1 ∑︁
[ 𝛼] 𝜕2𝜓
+ 𝑖𝜆 𝐶 𝑘,ℓ (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼− ⟨𝑘 ⟩− ⟨ℓ ⟩ + 𝑂 𝜆 | 𝛼 |−2 ,
2 𝜕𝑥 𝑘 𝜕𝑥ℓ
⟨𝑘 ⟩+⟨ℓ ⟩ ⪯ 𝛼

where
[ 𝛼] 𝛼!
𝐶 𝑘,ℓ = .
(𝛼 − ⟨𝑘⟩ − ⟨ℓ⟩)!
Proof We use induction on |𝛼| ≥ 1, the claim being obvious when 𝛼 = 0 and
|𝛼| = 1. Suppose it has been proved for a given 𝛼 ∈ Z+𝑛−1 , |𝛼| ≥ 1. If 1 ≤ 𝑚 ≤ 𝑛 − 1
and ⟨𝑘⟩ + ⟨ℓ⟩ ⪯ 𝛼 then
𝜕𝜓 𝜆𝜓 𝛼 −𝜆𝜓
e𝜆𝜓 D 𝑥𝛼+⟨𝑚⟩
′ e−𝜆𝜓 = 𝑖𝜆 e D 𝑥′ e + D 𝑥𝑚 e𝜆𝜓 D 𝑥𝛼′ e−𝜆𝜓 (24.4.25)
𝜕𝑥 𝑚
= (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼+⟨𝑚⟩
1 ∑︁
[ 𝛼] 𝜕2𝜓
𝑖𝜆 𝐶 𝑘,ℓ (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼+⟨𝑚⟩− ⟨𝑘 ⟩−⟨ℓ ⟩
2 𝜕𝑥 𝑘 𝜕𝑥ℓ
⟨𝑘 ⟩+⟨ℓ ⟩ ⪯ 𝛼
∑︁ 𝜕2𝜓
+ 𝑖𝜆 𝛼 𝑘 (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼+⟨𝑚⟩− ⟨𝑘 ⟩− ⟨𝑚⟩ + 𝑂 𝜆 | 𝛼 |−1 .
𝜕𝑥 𝑘 𝜕𝑥 𝑚
⟨𝑘 ⟩+⟨𝑚⟩ ⪯ 𝛼+⟨𝑚⟩
1008 24 Analytic PDEs of Principal Type. Regularity of the Solutions

If 𝛼𝑚 = 0 then

[ 𝛼+⟨𝑚⟩ ] (𝛼 + ⟨𝑚⟩)! (𝛼 + ⟨𝑚⟩)!

𝐶 𝑘,𝑚 = 𝛼𝑘 = = .
(𝛼 − ⟨𝑘⟩)! (𝛼 + ⟨𝑚⟩ − ⟨𝑘⟩ − ⟨𝑚⟩)!
Suppose 𝛼𝑚 ≥ 1; in (24.4.25) there are terms

1 ∑︁
[ 𝛼] 𝜕2𝜓
𝐶𝑚,ℓ (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼+⟨𝑚⟩− ⟨𝑚⟩−⟨ℓ ⟩
2 𝜕𝑥 𝑚 𝜕𝑥ℓ
⟨𝑚⟩+⟨ℓ ⟩ ⪯ 𝛼
ℓ≠𝑚
1 ∑︁
[ 𝛼] 𝜕2𝜓
+ 𝐶 𝑘,𝑚 (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼+⟨𝑚⟩− ⟨𝑘 ⟩− ⟨𝑚⟩
2 𝜕𝑥 𝑘 𝜕𝑥 𝑚
⟨𝑘 ⟩+⟨𝑚⟩ ⪯ 𝛼
𝑘≠𝑚
∑︁
[ 𝛼] 𝜕2𝜓
= 𝐶 𝑘,𝑚 (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼+⟨𝑚⟩− ⟨𝑘 ⟩−⟨𝑚⟩
𝜕𝑥 𝑘 𝜕𝑥 𝑚
⟨𝑘 ⟩ ⪯ 𝛼−⟨𝑚⟩

and we get, assuming that ⟨𝑘⟩ ⪯ 𝛼 − ⟨𝑚⟩, 𝑘 ≠ 𝑚,

[ 𝛼+⟨𝑚⟩ ] [ 𝛼]
𝐶 𝑘,𝑚 = 𝐶 𝑘,𝑚 + 𝛼𝑘

1 1
= 𝛼! +
(𝛼 − ⟨𝑘⟩ − ⟨𝑚⟩)! (𝛼 − ⟨𝑘⟩)!

𝛼! (𝛼 − ⟨𝑘⟩)!
= 1+
(𝛼 − ⟨𝑘⟩)! (𝛼 − ⟨𝑘⟩ − ⟨𝑚⟩)!
𝛼! (𝛼𝑚 + 1) (𝛼 + ⟨𝑚⟩)!
= = .
(𝛼 − ⟨𝑘⟩)! (𝛼 + ⟨𝑚⟩ − ⟨𝑘⟩ − ⟨𝑚⟩)!
If 𝑘 = 𝑚, then 𝛼𝑚 ≥ 2 and
[ 𝛼+⟨𝑚⟩ ] [ 𝛼]
𝐶𝑚,𝑚 = 𝐶𝑚,𝑚 + 2𝛼𝑚 ,
[ 𝛼] 𝛼!
𝐶𝑚,𝑚 = + 2𝛼𝑚
(𝛼 − 2 ⟨𝑚⟩)!
= 𝛼𝑚 (𝛼𝑚 − 1) + 2𝛼𝑚 .

When 𝑘 ≠ 𝑚, ℓ ≠ 𝑚,

[ 𝛼] 𝛼! (𝛼 + ⟨𝑚⟩)! [ 𝛼+⟨𝑚⟩ ]
𝐶 𝑘,ℓ = = = 𝐶 𝑘,ℓ . □
(𝛼 − ⟨𝑘⟩ − ⟨ℓ⟩)! (𝛼 + ⟨𝑚⟩ − ⟨𝑘⟩ − ⟨ℓ⟩)!

Corollary 24.4.7 For each 𝛼 ∈ Z+𝑛−1 ,

∑︁ 1
𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) e𝜆𝜓 D 𝑥′ e−𝜆𝜓 D 𝑥′ 𝑎 𝑗 = 𝑂 𝜆−1 .
𝛼−𝛽 𝛽
(𝛼 − 𝛽)!𝛽!
|𝛽 | ≥2
𝛽⪯𝛼
24.4 Property (Q) Implies Subellipticity 1009

Proof Indeed,

𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′) = 𝜆1−| 𝛼 | 𝜕 𝜉𝛼′ 𝐵 (𝑥, 𝜉 ′/𝜆) and

e−𝜆𝜓 = (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼−𝛽 + 𝑂 𝜆 | 𝛼−𝛽 |−1
𝛼−𝛽
e𝜆𝜓 D 𝑥′

by Lemma 24.4.6. □
The transport equations [cf. (23.2.14)–(23.2.15)] obtain by equating the coeffi-
cients of each power of 𝜆−1 in both sides of (24.4.22). Taking (24.4.24) into account
and summing the Taylor expansions yields
∑︁ ∑︁ 1
𝜕 𝜉𝛼 𝐵 (𝑥, 𝜉 ′) e𝜆𝜓 D 𝑥′ e−𝜆𝜓 D 𝑥′ 𝑎 𝑗
𝛼−𝛽 𝛽
(𝛼 − 𝛽)!
𝛼∈Z+𝑛−1 |𝛽 |=1
𝛽⪯𝛼
∑︁ ∑︁ 1 − | 𝛼 | 𝛼+𝛽
𝜕 𝜉 𝐵 (𝑥, 𝜉 ′/𝜆) e𝜆𝜓 D 𝑥𝛼′ e−𝜆𝜓 D 𝑥′ 𝑎 𝑗
𝛽
= 𝜆
𝛼!
|𝛽 |=1 𝛼∈Z+𝑛−1
∑︁ ∑︁ 1 𝛼+𝛽 ∑︁
𝜕 𝜉 𝐵 (𝑥, 𝜉 ′/𝜆) (𝑖𝜕𝑥′ 𝜓) 𝛼 D 𝑥′ 𝑎 𝑗 + 𝑂 𝜆−1 D 𝑥′ 𝑎 𝑗
𝛽 𝛽
=
𝛼!
|𝛽 |=1 𝛼∈Z+𝑛−1 |𝛽 |=1
∑︁ ∑︁
𝜕 𝜉 𝐵 (𝑥, 𝜉 ′/𝜆 + 𝑖𝜕𝑥′ 𝜓) D 𝑥′ 𝑎 𝑗 + 𝑂 𝜆−1 D 𝑥′ 𝑎 𝑗
𝛽 𝛽 𝛽
=
|𝛽 |=1 |𝛽 |=1

as well as
∑︁ 1 𝛼
𝜕 𝜉 ′ 𝐵 (𝑥, 𝜉 ′) e𝜆𝜓 D 𝑥𝛼′ e−𝜆𝜓 − (𝑖𝜆𝜕𝑥′ 𝜓) 𝛼 𝑎 𝑗
𝛼!
𝛼∈Z+𝑛−1
𝑛−1
1 ∑︁ 𝜕 2 𝐵 𝜕2𝜓
=− (𝑥, 𝜉 ′/𝜆 + 𝑖𝜕𝑥′ 𝜓) 𝑎 𝑗 + 𝑂 𝜆−1 𝑎 𝑗 .
2𝑖 𝑘,ℓ=1 𝜕𝜉 𝑘 𝜕𝜉ℓ 𝜕𝑥 𝑘 𝜕𝑥ℓ

We introduce the vector field

𝑛−1
∑︁ 𝜕𝐵
𝔏 (𝑥, D 𝑥 ) = D 𝑥𝑛 + 𝑖 (𝑥, 𝜕𝑥′ 𝜑 (𝑥)) D 𝑥𝑘 (24.4.26)
𝑘=1
𝜕𝜉 𝑘

and the coefficient

𝑛−1
1 ∑︁ 𝜕 2 𝐵 𝜕2 𝜑
𝑐 0 (𝑥) = (𝑥, 𝜕𝑥′ 𝜑 (𝑥)) (𝑥) . (24.4.27)
2𝑖 𝑘,ℓ=1 𝜕𝜉 𝑘 𝜕𝜉ℓ 𝜕𝑥 𝑘 𝜕𝑥ℓ

Keep in mind that here 𝜑 (𝑥) = 𝑥 · 𝜉 ′/𝜆 + 𝑖𝜓 (𝑡, 𝑥, 𝜉 ′/𝜆). Equating the coefficients of
the zero power of 𝜆−1 in (24.4.22) gives us the first transport equation

𝔏 (𝑥, D 𝑥 ) 𝑎 0 + 𝑐 0 𝑎 0 = 0. (24.4.28)
1010 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Next, we derive from (24.4.22), for arbitrary 𝑗 ≥ 1,

𝑗
∑︁ ∑︁ 1
𝔏 (𝑥, D 𝑥 ) 𝑎 𝑗 + 𝑐 0 𝑎 𝑗 = 𝑐 𝑗,𝑘, 𝛼 D 𝑥𝛼′ 𝑎 𝑗−𝑘 , (24.4.29)
𝑘=1 | 𝛼 | ≤𝑘+1
𝛼!

where the coefficients 𝑐 𝑗,𝑘, 𝛼 are C 𝜔 functions of (𝑡, 𝑥, 𝜉 ′/𝜆) in the appropriate
domains (cf. above). The Cauchy conditions for the system of equations (24.4.28)–
(24.4.29) are (24.4.20).

24.4.4 Microlocal parametrices under Hypotheses (A1) and (A2)

We shall now deal with finite sums

𝑁
∑︁
𝑎 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) = 𝑎 𝑗 (𝑡, 𝑥, 𝜉 ′)
𝑗=0

with 𝑎 𝑗 (𝑡, 𝑥, 𝜉 ′/𝜆) solving (24.4.28)–(24.4.29) and satisfying (24.4.20). The con-
gruence (24.4.21) leads to the following statement:

Proposition 24.4.8 If 𝑈◦′ , Γ◦′ ∩S𝑛−2 and 𝜏◦ are sufficiently small then to every integer
𝑁 ′ ≥ 1 there is another integer 𝑁 ≥ 𝑁 ′ such that

𝜕 ′
− 𝐵 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑎 [𝑁 ] (𝑡, 𝑥, 𝜉 ′)
′ (24.4.30)
𝜕𝑥 𝑛
′
= e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′)

where 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) ∈ C ∞ (−𝜏◦ , 𝜏◦ ) × 𝑈◦′ × (−𝜏◦ , 𝜏◦ ) × Γ◦′ satisfies estimates

′
∀𝑘 ∈ Z+ , ∀𝛼 ∈ Z+𝑛 , 𝛽 ∈ Z+𝑛−1 , 𝜕𝑡𝑘 𝜕𝑥𝛼 𝜕 𝜉 ′ 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶 𝑘, 𝛼,𝛽 |𝜉 ′ | −𝑁 − |𝛽 |
𝛽

(24.4.31)
for all (𝑡, 𝑥, 𝜉 ′) ∈ (−𝜏◦ , 𝜏◦ ) × 𝑈◦′ × (−𝜏◦ , 𝜏◦ ) × Γ◦′ , |𝜉 ′ | ≥ 1.

Proof Recalling the meaning of in (24.4.21) one uses the transport equations
(24.4.28)–(24.4.29) and standard formulas for the error term in the finite Taylor
expansion of 𝐵 (𝑥, 𝜉 ′) in a conic neighborhood of (0, 𝜉 ◦′). We leave the details as an
exercise. □
Let Υ (𝑥 𝑛 ) denote the Heaviside function; (24.4.1), (24.4.20) and (24.4.30) imply

𝜕 ′
− 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) Υ (𝑥 𝑛 − 𝑡) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑎 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) (24.4.32)
𝜕𝑥 𝑛
′ ′ ′
= 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) e𝑖 𝑥 · 𝜉 ⊗ 𝛿 (𝑥 𝑛 − 𝑡) + Υ (𝑥 𝑛 − 𝑡) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) .
24.4 Property (Q) Implies Subellipticity 1011

We introduce one more cut-off function: 𝜒˜ 2 ∈ Cc∞ Γ◦′ ∩ S𝑛−2 extended to R𝑛−1 \ {0}

as a homogeneous function of degree zero, 0 ≤ 𝜒˜ 2 ≤ 1 everywhere, 𝜒˜ 2 ≡ 1 in a
neighborhood of supp 𝜒2 . We shall make use of the following integrals, in which
∞ 𝑈 ′ × (−𝜏 , 𝜏 ) with 𝜏 as in

𝑥 𝑛 ∈ (−𝜏, 𝜏) with 𝜏 as in Proposition 24.4.1, 𝑢 ∈ C c ◦ ◦ ◦ ◦
∫ ′ ′
𝑢 (𝜉 ′, 𝑡) = R𝑛−1 e−𝑖𝑦 · 𝜉 𝑢 (𝑦 ′, 𝑡) d𝑦 ′:
Proposition 24.4.8, b

𝑨 𝜑[𝑁 ] 𝑢 (𝑥) (24.4.33)

∫ ∫ 𝑥𝑛
𝑖 𝑥 ′ · 𝜉 ′ −𝜓 (𝑡 , 𝑥, 𝜉 ′ )
= (2𝜋) 1−𝑛 e 𝑎 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) 𝜒˜ 2 (𝜉 ′) b
𝑢 (𝜉 ′, 𝑡) d𝜉 ′d𝑡,
| 𝜉 ′ |>1 −∞

𝑹 𝜑[𝑁 ] 𝑢 (𝑥) (24.4.34)

∫ ∫ 𝑥𝑛
′ ′ −𝜓 (𝑡 , 𝑥, 𝜉 ′ )
= (2𝜋) 1−𝑛 e𝑖 𝑥 · 𝜉 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) 𝜒˜ 2 (𝜉 ′) b
𝑢 (𝜉 ′, 𝑡) d𝜉 ′d𝑡.
| 𝜉 ′ |>1 −∞

The operators 𝑨 𝜑[𝑁 ] and 𝑹 𝜑[𝑁 ] are not standard pseudodifferential operators in the
sense of Definition 16.1.14: while the symbols 𝑎 ( 𝑁 ) and 𝑅 [𝑁 ] are standard (Defi-
nitions 16.1.1, 16.2.1) this is not so of e−𝜓 𝑎 ( 𝑁 ) and e−𝜓 𝑅 [ 𝑁 ] , as shown in Corol-
lary 24.4.3; due to the inequalities (24.4.8) they belong to a class often denoted
by 𝑆 1 , 1 or, here, more precisely, 𝑆 01 1 (𝑈◦′ ). They define pseudodifferential oper-
2 2 2,2
ators in 𝑈◦′ × (−𝜏◦ , 𝜏◦ ) which share most (but not all) of the basic properties of
the standard pseudodifferential operators and, first of all, the property of defin-
operators Cc∞ 𝑈◦′ × (−𝜏 ∞ 𝑈 ′ × (−𝜏 , 𝜏 ) and

ing continuous linear ◦ , 𝜏◦ ) −→ C ◦ ◦ ◦
E ′ 𝑈◦′ × (−𝜏◦ , 𝜏◦ ) −→ D ′ 𝑈◦′ × (−𝜏◦ , 𝜏◦ ) . We shall now get rough estimates of

the actions of 𝑨 𝜑[𝑁 ] and 𝑹 𝜑[𝑁 ] on Sobolev spaces (cf. Theorem 16.1.19). We need a
preliminary result, about the Fourier transform
∫
(𝑁) ′ ′ ′
𝑎
b ′ ′
(𝑡, 𝑥 𝑛 , 𝜃 , 𝜉 ) = e−𝑖 𝑥 · 𝜃 −𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑎 ( 𝑁 ) (𝑡, 𝑥, 𝜉 ′) d𝑥 ′. (24.4.35)

Lemma 24.4.9 If diam 𝑈◦′ , diam Γ◦′ ∩ S𝑛−2 and 𝜏◦ > 0 are sufficiently small then

to every 𝜈 ∈ R+ there is a 𝐶 (𝑁, 𝜈) > 0 such that
∫ 𝜈
𝑎 ( 𝑁 ) (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′) 1 + |𝜃 ′ | 2 d𝜃 ′ ≤ 𝐶 (𝑁, 𝜈) |𝜉 ′ | 𝜈+𝑛
b (24.4.36)

for all (𝑡, 𝑥 𝑛 ) such that −𝜏◦ < 𝑡 < 𝑥 𝑛 < 𝜏◦ , 𝜉 ′ ∈ Γ ′, |𝜉 ′ | ≥ 1.

Proof Since Re 𝜓 ≥ 0 if −𝜏◦ < 𝑡 < 𝑥 𝑛 < 𝜏◦ (Proposition 24.4.1) we have

1

𝑎 ( 𝑁 ) 𝑡, 𝑥 𝑛 , |𝜉 ′ | 2 𝜃 ′, 𝜉 ′ ≤ 𝐶0 ,
b (24.4.37)

where 𝐶0 > 0 is independent of 𝑡, 𝑥, 𝜉 ′, 𝜃 ′, like 𝐶1 (𝜈), 𝐶2 (𝜈) ... in the sequel. First
consider the case 𝜈 ∈ Z+ ; we have
1012 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫
′ −𝑖 𝑥 ′ · 𝜃 ′
𝑎 (𝜈) (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′) |𝜃 ′ | 2(𝜈+𝑛) =
b e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑎 ( 𝑁 ) (𝑡, 𝑥, 𝜉 ′) Δ𝜈+𝑛 𝑥′ e d𝑥 ′
∫
′ ′ −𝜓 (𝑡 , 𝑥, 𝜉 ′ ) (𝑁)
= e−𝑖 𝑥 · 𝜃 Δ𝜈+𝑛
𝑥 ′ e 𝜒 1 (𝑥 ′
) 𝑎 (𝑡, 𝑥, 𝜉 ′
) d𝑥 ′.

The inequalities (24.4.8) and the fact that 𝑎 ( 𝑁 ) (𝑡, 𝑥, 𝜉 ′) is a standard symbol imply
directly [cf. (24.4.15)]

−𝜓 (𝑡 , 𝑥, 𝜉 ′ ) ′ (𝑁) ′
Δ𝜈+𝑛
𝑥 ′ e 𝜒1 (𝑥 ) 𝑎 (𝑡, 𝑥, 𝜉 ) ≤ 𝐶1 (𝜈) |𝜉 ′ | 𝜈+𝑛 .

We conclude that

𝑎 ( 𝑁 ) (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′) |𝜃 ′ | 2(𝜈+𝑛) ≤ 𝐶2 (𝜈) |𝜉 ′ | 𝜈+𝑛 .

b

1
After the change of variables 𝜃 ′ ⇝ |𝜉 ′ | − 2 𝜃 ′ the last inequality reads
1

𝑎 ( 𝑁 ) 𝑡, 𝑥 𝑛 , |𝜉 ′ | 2 𝜃 ′, 𝜉 ′ |𝜃 ′ | 2(𝜈+𝑛) ≤ 𝐶2 (𝜈) .
b (24.4.38)

Now let 𝑠 ∈ (0, 2); from (24.4.37) we derive, if |𝜃 ′ | ≤ 1,

1

𝑎 ( 𝑁 ) 𝑡, 𝑥 𝑛 , |𝜉 ′ | 2 𝜃 ′, 𝜉 ′ |𝜃 ′ | 2(𝜈+𝑛)+𝑠 ≤ 𝐶0 ;
b

from (24.4.38) we derive, if |𝜃 ′ | > 1,

1

𝑎 ( 𝑁 ) 𝑡, 𝑥 𝑛 , |𝜉 ′ | 2 𝜃 ′, 𝜉 ′ |𝜃 ′ | 2(𝜈+𝑛)+𝑠 ≤ 𝐶2 (𝜈 + 1) .
b

This proves that to every 𝜈 ∈ R+ ,

𝜈+𝑛
𝑎 ( 𝑁 ) (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′) 1 + |𝜃 ′ | 2
b ≤ 𝐶3 (𝜈) |𝜉 ′ | 𝜈+𝑛 .

We derive
d𝜃 ′
∫ 𝜈 ∫
𝑎 ( 𝑁 ) (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′) 1 + |𝜃 ′ | 2 d𝜃 ′ ≤ 𝐶3 (𝜈) |𝜉 ′ | 𝜈+𝑛
b 𝑛 . □
1 + |𝜃 ′ | 2

Proposition 24.4.10 If 𝑠 ∈ R, 𝑠 ≤ 𝑛/2, then the operator 𝑨 𝜑[𝑁 ] induces a continuous

linear map 𝐻c𝑠 𝑈◦′ −→ 𝐻loc 𝑈◦′ whose norm is bounded independently of
𝑠−𝑛/2
𝑥 𝑛 ∈ (−𝜏◦ , 𝜏◦ ).

Proof We derive from (24.4.33) and (24.4.35):

24.4 Property (Q) Implies Subellipticity 1013
∫
′ ′
(2𝜋) 2𝑛−2 e−𝑖 𝑥 · 𝜃 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] 𝑢 (𝑥) d𝑥 ′
∫ 𝑥𝑛 ∫
= 𝑎 ( 𝑁 ) (𝑡, 𝑥 𝑛 , 𝜃 ′ − 𝜉 ′, 𝜉 ′) 𝜒˜ 2 (𝜉 ′) b
b 𝑢 (𝜉 ′, 𝑡) d𝜉 ′d𝑡.
−∞ | 𝜉 ′ |>1

We now use the inequality

𝜅 |𝜅 | 𝜅
1 + |𝜃 ′ | 2 ≤ 2 |𝜅 |/2 1 + |𝜃 ′ − 𝜉 ′ | 2 1 + |𝜉 ′ | 2

[cf. (16.1.19), 𝜅 ∈ R]. It implies

∫ 𝜅
′ ′
(2𝜋) 𝑛−1 − |𝑠 |/2
2 e−𝑖 𝑥 · 𝜃 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] 𝑢 (𝜅) d𝑥 ′ 1 + |𝜃 ′ | 2
∫ |𝜅 | 𝜅
≤ 𝑎 ( 𝑁 ) (𝑡, 𝜅 𝑛 , 𝜃 ′ − 𝜉 ′, 𝜉 ′) 1 + |𝜃 ′ − 𝜉 ′ | 2
b 𝑢 (𝜉 ′, 𝑡)| 1 + |𝜉 ′ | 2 d𝜉 ′.
| 𝜒˜ 2 (𝜉 ′) b

We apply (24.4.36) with 𝜈 = |𝜅| and the classical inequality

∥ 𝑓 ∗ 𝑔∥ 𝐿 2 ≤ ∥ 𝑓 ∥ 𝐿 1 ∥𝑔∥ 𝐿 2 , 𝑓 ∈ 𝐿 1 R𝑛−1 , 𝑔 ∈ 𝐿 2 R𝑛−1 ,

to get
∫ ∫ 2 𝜅
′ ′
e−𝑖 𝑥 · 𝜃 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] 𝑢 (𝑥) d𝑥 ′ 1 + |𝜃 ′ | 2 d𝜃 ′
∫ 𝜅+ 21 |𝜅 |+ 12 𝑛
≲ 𝑢 (𝜉 ′, 𝑡)| 2 1 + |𝜉 ′ | 2
|b d𝜉 ′.

1
If 𝑠 ≤ 𝑛/2 we can take 𝜅 = 2𝑠 − 𝑛 equivalent to 𝜅 + 2 |𝜅| + 12 𝑛 = 𝑠, thereby proving
the claim. □

Proposition 24.4.11 To every integer 𝑘 ≥ 1 there is an integer 𝑁 such that, whatever

𝑠 ∈ R, the operator 𝑹 𝜑[𝑁 ] induces a continuous linear map 𝐻c𝑠 𝑈◦′ −→ 𝐻loc 𝑈◦′
𝑠+𝑘

whose norm is bounded independently of 𝑥 𝑛 ∈ (−𝜏◦ , 𝜏◦ ).

Proof We derive from (24.4.34):

∫
′ ′
(2𝜋) 2𝑛−2 e−𝑖 𝑥 · 𝜃 𝜒1 (𝑥 ′) 𝑅 [𝑁 ] 𝑢 (𝑥) d𝑥 ′
∫ 𝑥𝑛 ∫
= 𝑅ˆ [𝑁 ] (𝑡, 𝑥 𝑛 , 𝜃 ′ − 𝜉 ′, 𝜉 ′) 𝜒˜ 2 (𝜉 ′) b
𝑢 (𝜉 ′, 𝑡) d𝜉 ′d𝑡
−∞ | 𝜉 ′ |>1

where
∫
′ ′ ′
𝑅ˆ [𝑁 ] (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′) = e−𝑖 𝑥 · 𝜃 −𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) d𝑥 ′.
1014 24 Analytic PDEs of Principal Type. Regularity of the Solutions

We have, for 𝜅 ∈ R,
𝜅
1 + |𝜃 ′ | 2 𝑅ˆ [𝑁 ] (𝑡, 𝑥 𝑛 , 𝜃 ′, 𝜉 ′)
∫ 𝜅
′ ′ ′
= e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) 1 + |Δ 𝑥′ | 2 e−𝑖 𝑥 · 𝜃 d𝑥 ′
∫ 𝜅
′ ′ ′
= e−𝑖 𝑥 · 𝜃 1 + |Δ 𝑥′ | 2 e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) d𝑥 ′.

Combining Corollary 24.4.3 and (24.4.30) yields, for some constant 𝐶 > 0 and all
(𝑡, 𝑥 𝑛 ) such that −𝜏◦ < 𝑡 < 𝑥 𝑛 < 𝜏◦ , 𝜉 ′ ∈ Γ ′, |𝜉 ′ | ≥ 1, 𝜃 ∈ R𝑛−1 ,
𝜅 ′
′
1 + |Δ 𝑥′ | 2 e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜒1 (𝑥 ′) 𝑅 [𝑁 ] (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶 |𝜉 ′ | 3𝜅−𝑁 .

1
Selecting 𝜅 ≥ 2 (|𝑠| + 𝑘) + 𝑛 and 𝑁 ′ ≥ 3𝜅 yields
1 ( |𝑠 |+𝑘) d𝜃 ′
∫ ∫
ˆ [𝑁 ] ′ ′ ′ 2 2 ′
𝑅 (𝑡, 𝑥 𝑛 , 𝜃 , 𝜉 ) 1 + |𝜃 | d𝜃 ≤ 𝐶 𝑛 .
1 + |𝜃 ′ | 2

The proof is completed in the same manner as that of Proposition 24.4.10. □

Multiplying both sides in (24.4.32) by 𝜒˜ 2 (𝜉 ′) yields

𝜕
− 𝐵 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] 𝑢 (𝑥) = 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) 𝜒˜ 2 (D 𝑥′ ) 𝑢 (𝑥)
′
𝜕𝑥 𝑛
+ 𝑹 𝜑[𝑁 ] 𝑢 (𝑥) .

Since 𝜒1 ≡ 1 in the neighborhood 𝑉 ′ ⊂ 𝑈◦′ of 0 we see that

𝜒2 (D 𝑥′ ) [ 𝜒1 (𝑥 ′) , 𝜒˜ 2 (D 𝑥′ )]

is a smoothing operator 𝑺 in 𝑉 ′; since 𝜒2 (D 𝑥′ ) 𝜒˜ 2 (D 𝑥′ ) = 𝜒2 (D 𝑥′ ) we reach the

conclusion that

𝜕
− 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] 𝑢 (𝑥) (24.4.39)
𝜕𝑥 𝑛
= 𝜒2 (D 𝑥′ ) 𝑢 (𝑥) + 𝑹 𝜑[𝑁 ] 𝑢 (𝑥) + 𝑺𝑢 (𝑥)

for every 𝑢 ∈ Cc∞ (−𝜏◦ , 𝜏◦ ) ; 𝐻c𝑠 (𝑉 ′) . Thus, according to Proposition 24.4.10 the

hypotheses
(A1), (A2) provide us with an approximate solution of the equation
𝜕
𝜕𝑥𝑛 − 𝐵 (𝑥, D 𝑥 ′ ) 𝜒2 (D 𝑥 ′ ) 𝑣 = 𝑢, namely 𝑨 𝜑[𝑁 ] 𝑢, with degraded regularity (“loss”
of 21 𝑛 derivatives) with respect to the 𝑥 ′ variables. The next step is to show that
Hypothesis (Q) in Theorem 24.2.1 greatly improves the regularity of 𝑨 𝜑[𝑁 ] , to a
fractional gain of derivatives with respect to the 𝑥 ′ variables (and of one derivative
with respect to 𝑥 𝑛 ).
24.4 Property (Q) Implies Subellipticity 1015

24.4.5 Consequences of Hypothesis (Q)

We assume that Condition (Q) is satisfied; this adds the following hypothesis to
(A1), (A2):
(A3) The function (−𝜏, 𝜏) ∋ 𝑥 𝑛 ↦→ 𝐵 (0, 𝑥 𝑛 , 𝜉 ◦′) vanishes to order 2𝑘 at 𝑥 𝑛 = 0
(𝑘 ∈ Z+ , 1 ≤ 𝑘 < +∞).

Remark 24.4.12 The case 𝑘 = 0 is covered by the ongoing argument but corresponds
to the ellipticity of 𝜕𝑥𝜕𝑛 − 𝐵 (𝑥, D 𝑥′ ) at (𝑥 ◦ , 𝜉 ◦ ), a case in which the properties sought
here have been already established.

We go back to (24.4.1) recalling that 𝜑 = 𝜑 (𝑡, 𝑥, 𝜉 ′) and

𝐵 (𝑥, 𝜕𝑥′ 𝜑 (𝑡, 𝑥, 𝜉 ◦′))| 𝑥𝑛 =𝑡 = 𝐵 (𝑥 ′, 𝑡, 𝜉 ◦′) .

Hypotheses (A1) and (A3) allows us to apply the Weierstrass Preparation Theorem
14.3.4: in a conic neighborhood of ((0, 𝑡) , 𝜉 ◦′), which we take to be 𝑈 × Γ ′ (as long
as −𝜏 < 𝑡 < 𝜏 and 𝜏 is sufficiently small),

𝐵 (𝑥, 𝜕𝑥′ 𝜑 (𝑡, 𝑥, 𝜉 ′)) = −𝐸 1 (𝑡, 𝑥, 𝜉 ′) 𝑓 (𝑡, 𝑥 ′, 𝜉 ′; 𝑥 𝑛 ) , (24.4.40)

′ ′ 2𝑘 ′ ′ 2𝑘−1
𝑓 (𝑡, 𝑥 , 𝜉 ; 𝑥 𝑛 ) = (𝑥 𝑛 − 𝑡) + 𝑐 1 (𝑡, 𝑥 , 𝜉 ) (𝑥 𝑛 − 𝑡) + · · · + 𝑐 2𝑘 (𝑡, 𝑥 , 𝜉 ′) ,
′

where 𝑐 𝑗 (𝑡, 𝑥 ′, 𝜉 ′) ∈ C 𝜔 ((−𝜏, 𝜏) × 𝑈 ′ × Γ ′; R) are homogeneous of degree zero,

𝑐 ℓ (𝑡, 0, 𝜉 ◦′) = 0 for every ℓ = 1, ..., 2𝑘, 𝑡 ∈ (−𝜏, 𝜏) [thus 𝜉 ′ ↦→ 𝑓 (𝑡, 𝑥 ′, 𝜉 ′; 𝑠) is
homogeneous of degree zero]; 𝐸 1 (𝑡, 𝑥, 𝜉 ′) ∈ C 𝜔 ((−𝜏, 𝜏) × 𝑈 × Γ ′; R) is homo-
geneous of degree 1; 𝐸 1 (𝑡, 𝑥, 𝜉 ′) ≥ 𝑐 ◦ |𝜉 ′ | for some 𝑐 ◦ > 0 and all (𝑡, 𝑥, 𝜉 ′) ∈
(−𝜏, 𝜏) × 𝑈 × Γ ′ [the latter takes also (A2) into account]. Then, in the notation of
(24.4.2) we have
∫ 𝑥𝑛
′
𝜓 (𝑡, 𝑥, 𝜉 ) = 𝐸 1 (𝑡, 𝑥 ′, 𝑡 𝑛 , 𝜉 ′) 𝑓 (𝑡, 𝑥 ′, 𝜉 ′; 𝑡 𝑛 ) d𝑡 𝑛 . (24.4.41)
𝑡

We introduce the following numbers

𝑠◦ = (2𝑘 + 1) −1 , 𝛿 = 𝑘 𝑠◦ , 𝜌 = (𝑘 + 1) 𝑠◦ ; (24.4.42)

note that 𝛿 = 𝜌 − 𝑠◦ = 1 − 𝜌 and 0 < 𝛿 < 21 < 𝜌 < 1. In all the statements that follow
it is assumed that diam 𝑈 ′, 𝜏 > 0 and diam Γ ′ ∩ S𝑛−2 are as small as needed.

Lemma 24.4.13 Let 𝑎 (𝑡, 𝑥, 𝜉 ′) be a standard symbol of order zero (Definition

16.2.1) in 𝑈 ′ depending smoothly on (𝑡, 𝑥 𝑛 ) ∈ [−𝜏, 𝜏] 2 . To every pair (𝛼, 𝛽) ∈
Z+𝑛−1 × Z+𝑛−1 there is a 𝐶 𝛼,𝛽 > 0 such that, for all (𝑥 ′, 𝜉 ′) ∈ 𝑈 ′ × Γ ′, |𝜉 ′ | ≥ 1,
1016 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫ 𝜏
′
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) d𝑥 𝑛 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | −𝑠◦ + 𝛿 | 𝛼 |−𝜌 |𝛽 | ,
𝛽
sup
−𝜏 ≤𝑡 ≤𝜏 𝑡
(24.4.43)
∫ 𝑥𝑛
′
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) d𝑡 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | −𝑠◦ + 𝛿 | 𝛼 |−𝜌 |𝛽 | .
𝛽
sup
−𝜏 ≤𝑥𝑛 ≤𝜏 −𝜏
(24.4.44)

Proof We apply the Leibniz rule

′

𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′)
𝛽

∑︁ ∑︁
𝛼− 𝛼˜ 𝛽− 𝛽˜ 𝛽˜ ′
= 𝛼 𝛽
𝛼˜ 𝛽˜ 𝜕 𝑥 ′ 𝜕 𝜉 ′ 𝑎 (𝑡, 𝑥, 𝜉 ′
) 𝜕𝑥𝛼˜′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) .
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽

On the one hand, we apply Proposition 24.4.2:

𝛽˜ ˜
((𝑥 𝑛 − 𝑡) |𝜉 ′ |) ℓ |𝜉 ′ | − | 𝛽 | e− 2 Re 𝜓 (𝑡 , 𝑥, 𝜉 ) ;
′
∑︁ 1 ′
𝜕𝑥𝛼˜′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) ≤ 𝐶 𝛼,
˜ 𝛽˜
ℓ ≤ 21 | 𝛼+
˜ 𝛽˜ |
(24.4.45)
on the other hand, by Definition 16.2.1, we have
𝛽− 𝛽˜ ˜
𝜕𝑥𝛼−
′
𝛼˜ ′[ 𝑁 ]
𝜕 𝜉 ′ 𝑎 (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶 𝛼− ˜
𝛼,𝛽− 𝛽˜
|𝜉 ′ | − | 𝛽−𝛽 | . (24.4.46)

′[𝑁 ]
(Needless to say, 𝐶 𝛼,˜ 𝛽˜ > 0 and 𝐶 𝛼− 𝛼,𝛽− ˜ 𝛽˜
> 0 are independent of 𝑡, 𝑥, 𝜉 ′, |𝜉 ′ | ≥ 1,
like all constants in the remainder of this subsection.) We get
∫ 𝑥𝑛
′
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) d𝑡
𝛽
−𝜏
∑︁ ∫ 𝑥𝑛
1 ′
′′
≤ 𝐶 𝛼,𝛽 |𝜉 ′ | ℓ− |𝛽 | (𝑥 𝑛 − 𝑡) ℓ e− 2 Re 𝜓 (𝑡 , 𝑥, 𝜉 ) d𝑡.
−𝜏
ℓ ≤ 21 | 𝛼+𝛽 |

Formula (24.4.41) allows us to apply Corollary 24.A.5 in which we put 𝜆 = |𝜉 ′ |,

𝜈 = 0, 𝜃 = 𝜃 ′ = 0, 𝑚 = 2𝑘, ℎ2 ≡ 1, as well as
1
𝑏 (𝑡 𝑛 ) = Re (𝐸 1 (𝑡, 𝑥 ′, 𝑡 𝑛 , 𝜉 ′) 𝑓 (𝑡, 𝑥 ′, 𝜉 ′; 𝑡 𝑛 )) ;
2
we get directly
∫ 𝑥𝑛
′
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) d𝑡 ≤ 𝐶 𝛼,𝛽
′′′
|𝜉 ′ | ℓ− |𝛽 |−𝑠◦ (ℓ+1) .
𝛽
max
−𝜏 ℓ ≤ 21 | 𝛼+𝛽 |

By (24.4.42)
1
ℓ= |𝛼 + 𝛽| =⇒ ℓ − |𝛽| − 𝑠◦ (ℓ + 1) = 𝑠◦ (𝑘 |𝛼| − (𝑘 + 1) |𝛽| − 1) .
2
24.4 Property (Q) Implies Subellipticity 1017

The same argument, after exchange of 𝑥 𝑛 and 𝑡 < 𝑥 𝑛 , yields (24.4.44). □

Lemma 24.4.14 Let 𝑎 (𝑡, 𝑥, 𝜉 ′) be as in Lemma 24.4.13. To every pair (𝛼, 𝛽) ∈

Z+𝑛−1 × Z+𝑛−1 there is a 𝐶 𝛼,𝛽 > 0 such that, for all (𝑥 ′, 𝜉 ′) ∈ 𝑈 ′ × Γ ′, |𝜉 ′ | ≥ 1,
∫ 𝜏
𝛽 𝜕 ′
sup 𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) d𝑥 𝑛 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | 𝛿 | 𝛼 |−𝜌 |𝛽 | ,
−𝜏 ≤𝑡 ≤𝜏 𝑡 𝜕𝑥 𝑛
(24.4.47)
∫ 𝑥𝑛
𝛽 𝜕 ′
sup 𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) d𝑡 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | 𝛿 | 𝛼 |−𝜌 |𝛽 | .
−𝜏 ≤𝑥𝑛 ≤𝜏 −𝜏 𝜕𝑥 𝑛
(24.4.48)

Proof We content ourselves with proving (24.4.48); (24.4.47) is proved in a similar

𝜕𝑎
fashion. Since 𝜕𝑥 𝑛
have estimates of the same type as 𝑎, (24.4.46), we have the
analogues of (24.4.43)–(24.4.44), in particular,
∫ 𝑥𝑛
′ 𝜕𝑎
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) (𝑡, 𝑥, 𝜉 ′) d𝑡 ≤ 𝐶 𝑁 , 𝛼,𝛽 |𝜉 ′ | −𝑠◦ − 𝛿 | 𝛼 |−𝜌 |𝛽 | .
𝛽
−𝜏 𝜕𝑥 𝑛

We must therefore prove estimates of the following kind:

∫ 𝑥𝑛
−𝜓 (𝑡 , 𝑥, 𝜉 ′ ) ′ 𝜕𝜓
sup 𝛼 𝛽
𝜕𝑥 ′ 𝜕 𝜉 ′ e 𝑎 (𝑡, 𝑥, 𝜉 ) (𝑡, 𝑥, 𝜉 ) d𝑡 ≤ 𝐶 |𝜉 ′ | 𝛿 | 𝛼 |−𝜌 |𝛽 | .
′
−𝜏 ≤𝑥𝑛 ≤𝜏 −𝜏 𝜕𝑡
(24.4.49)
[All constants in the proof will depend on 𝑁, 𝛼, 𝛽 but not on (𝑡, 𝑥, 𝜉 ′).]
We apply the Leibniz rule and exploit (24.4.46):

𝜕𝜓 ˜ 𝛽˜ 𝜕𝜓 −𝜓
|𝜉 ′ | − | 𝛽−𝛽 | 𝜕𝑥𝛼˜′ 𝜕 𝜉 ′
∑︁ ∑︁
𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 𝑎 ≤ 𝐶′
𝛽
e . (24.4.50)
𝜕𝑥 𝑛 𝛼˜ ⪯ 𝛼 ˜
𝜕𝑥 𝑛
𝛽 ⪯𝛽

Also,

𝜕𝜓 −𝜓 𝛽˜ 𝜕𝜓 𝛼˜ 𝛽˜ −𝜓
𝜕𝑥𝛼˜′ 𝜕 𝜉 ′
e ≤ 𝜕 ′𝜕 ′e
𝜕𝑥 𝑛 𝜕𝑥 𝑛 𝑥 𝜉
∑︁ ∑︁ ˜ ˜ ♭
𝛼˜ 𝛼♭ 𝛽 ♭ 𝜕𝜓
𝜕 𝜉 ′ e−𝜓
𝛽 ˜ ♭ 𝛽−𝛽
+ 𝑎♭ 𝛽 ♭ 𝜕𝑥 ′ 𝜕 𝜉 ′ 𝜕𝑥 𝜕𝑥𝛼−𝛼
′
♭ ˜ 𝑛
♭ 𝑎 ⪯ 𝛼˜ 𝛽 ⪯ 𝛽
| 𝛼♭ +𝛽♭ | =1
𝛼˜ 𝛽˜ 𝜕𝜓 ˜ ♭
∑︁ ∑︁
𝛽♭
𝜕 𝜉 ′ e−𝜓 .
♭ ˜ ♭ 𝛽−𝛽
+ 𝑎♭ 𝛽 ♭ 𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝜕 𝛼−𝛼
′
𝜕𝑥 𝑛 𝑥
𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜
| 𝛼♭ +𝛽♭ | ≥2

We combine the last two estimates and apply (24.4.45):

1018 24 Analytic PDEs of Principal Type. Regularity of the Solutions

1 𝜕𝜓 ∑︁ ∑︁ ∑︁ 𝜕𝜓
Re 𝜓 −𝜓
≤ 𝐶 ′′ (𝑥 𝑛 − 𝑡) ℓ |𝜉 ′ | ℓ− |𝛽 |
𝛽
e 2 𝜕𝑥𝛼′ 𝜕 𝜉 ′ e 𝑎
𝜕𝑥 𝑛 𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 ℓ ≤ 1 | 𝛼+
𝜕𝑥 𝑛
2 ˜ 𝛽˜ |
(24.4.51)
♭ 𝛽 ♭ 𝜕𝜓
(𝑥 𝑛 − 𝑡) ℓ |𝜉 ′ | ℓ− | 𝛽−𝛽 | 𝜕𝑥𝛼′ 𝜕 𝜉 ′
∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭
+𝐶 ′′
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
𝜕𝑥 𝑛
2 ˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽

| 𝛼♭ +𝛽♭ | =1
♭ 𝛽 ♭ 𝜕𝜓
(𝑥 𝑛 − 𝑡) ℓ |𝜉 ′ | ℓ− | 𝛽−𝛽 | 𝜕𝑥𝛼′ 𝜕 𝜉 ′
∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭
+𝐶 ′′ .
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
𝜕𝑥 𝑛
2 ˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽

| 𝛼♭ +𝛽♭ | ≥2

To deal with the first two sums in the right-hand side of (24.4.51) we use (24.4.41).
Firstly, we have
∫ 𝜏
∑︁ ∑︁ ∑︁ 𝜕𝜓 − 1 Re 𝜓
|𝜉 ′ | ℓ−|𝛽 | (𝑥 𝑛 − 𝑡) ℓ e 2 d𝑡
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 ℓ ≤ 1 | 𝛼+ 𝑡 𝜕𝑥 𝑛
˜
2 ˜ 𝛽|
∑︁ ∑︁ ∑︁ ∫ 𝑥𝑛
1
≤ 𝑀◦ |𝜉 ′ | 1+ℓ− |𝛽 | (𝑥 𝑛 − 𝑡) ℓ e− 2 Re 𝜓 𝑓 d𝑡,
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 ℓ ≤ 1 | 𝛼+ ˜ −𝜏
2 ˜ 𝛽|

where 𝑀◦ is the maximum of 𝐸 1 in the appropriate domain. We apply Corollary

24.A.5 with 𝜆 = |𝜉 ′ |, 𝜈 = 0, 𝜃 = 0, 𝜃 ′ = 1, 𝑚 = 2𝑘:
∑︁ ∑︁ ∑︁ ∫ 𝑥𝑛
1 𝑘 𝑘+1
|𝜉 ′ | 1+ℓ− |𝛽 | (𝑥 𝑛 − 𝑡) ℓ e− 2 Re 𝜓 𝑓 d𝑡 ≤ 𝐶 ′′′ |𝜉 ′ | 2𝑘+1 | 𝛼 |− 2𝑘+1 |𝛽 | .
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 ℓ ≤ 1 | 𝛼+ ˜ −𝜏
2 ˜ 𝛽|

Secondly,

♭ 𝛽 ♭ 𝜕𝜓
(𝑥 𝑛 − 𝑡) ℓ |𝜉 ′ | ℓ− | 𝛽−𝛽 | 𝜕𝑥𝛼′ 𝜕 𝜉 ′
∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭

𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
𝜕𝑥 𝑛
2 ˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽

| 𝛼♭ +𝛽♭ | =1
(𝑥 𝑛 − 𝑡) ℓ |𝜉 ′ | ℓ− | 𝛽−𝛽 | 𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝐸 1 𝑓
∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭ ♭ 𝛽♭
≤
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽
2
| 𝛼 |
♭ +𝛽 ♭ =1

(𝑥 𝑛 − 𝑡) ℓ |𝜉 ′ | 1+ℓ− | 𝛽−𝛽 | 𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝑓 .

∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭ ♭ 𝛽♭
+𝑀◦
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽
2
| 𝛼♭ +𝛽♭ | =1

The first term in the right-hand side is dealt with like the first sum in the right-hand
side of (24.4.51). Since 𝛼♭ + 𝛽♭ = 1 Proposition 24.3.4 implies

𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝑓 ≤ 𝐶◦ |𝜉 ′ | − | 𝛽 | 𝑓 ,
♭ 𝛽♭ ♭ √︁
24.4 Property (Q) Implies Subellipticity 1019

whence

|𝜉 ′ | 1+ℓ− | 𝛽−𝛽 | (𝑥 𝑛 − 𝑡) ℓ 𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝑓

∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭ ♭ 𝛽♭

𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽
2
| 𝛼♭ +𝛽♭ | =1
∑︁ ∑︁ ∑︁ 1
≤ 𝐶◦ |𝜉 ′ | 1+ℓ− |𝛽 | (𝑥 𝑛 − 𝑡) ℓ 𝑓 2 .
𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 ℓ ≤ 1 ( | 𝛼˜ |+ | 𝛽˜ | −1)
2

We apply Corollary 24.A.5 with 𝜆 = |𝜉 ′ |, 𝜈 = 0, 𝜃 = 0, 𝜃 ′ = 21 , 𝑚 = 2𝑘:

∑︁ ∫ 𝑥𝑛
♭ 𝛽♭ 1
′ 1+ℓ−|𝛽 |
|𝜉 | (𝑥 𝑛 − 𝑡) ℓ 𝜕𝑥𝛼′ 𝜕 𝜉 ′ 𝑓 e− 2 Re 𝜓 d𝑡
−𝜏
ℓ ≤ 12 ( | 𝛼˜ |+ | 𝛽˜ | −1)
𝑘 𝑘+1
≤ 𝐶 (iv) |𝜉 ′ | 2𝑘+1 | 𝛼 |− 2𝑘+1 |𝛽 | ,

which is what we wanted.

Lastly, we deal with the remaining terms in the left-hand side of (24.4.51); using
the fact that
♭ 𝛽 ♭ 𝜕𝜓
≤ const. |𝜉 ′ | 1− | 𝛽 |
♭
𝜕𝑥𝛼′ 𝜕 𝜉 ′
𝜕𝑥 𝑛
we get

|𝜉 ′ | ℓ− | 𝛽−𝛽 |
∑︁ ∑︁ ∑︁ ∑︁ ∑︁ ♭

𝛼˜ ⪯ 𝛼 𝛽˜ ⪯𝛽 𝑎♭ ⪯ 𝛼˜ 𝛽 ♭ ⪯ 𝛽˜ ℓ ≤ 1 (| 𝛼−𝛼
˜ | | ˜ ♭ |)
♭ + 𝛽−𝛽
2
| 𝛼 |
♭ +𝛽 ♭ ≥2
∫ 𝑥𝑛
♭ 𝛽 ♭ 𝜕𝜓 1
× (𝑥 𝑛 − 𝑡) ℓ 𝜕𝑥𝛼′ 𝜕 𝜉 ′ e− 2 Re 𝜓 d𝑡
−𝜏 𝜕𝑥 𝑛
∑︁ ∫ 𝑥𝑛
1
≤ 𝐶 (v) |𝜉 ′ | 1+ℓ−|𝛽 | (𝑥 𝑛 − 𝑡) ℓ e− 2 Re 𝜓 d𝑡
−𝜏
ℓ ≤ 21 ( | 𝛼 |+|𝛽 |)−1
∑︁ ∫
1
≤ 𝐶 (vi) |𝜉 ′ | ℓ− |𝛽 | (𝑥 𝑛 − 𝑡) ℓ e− 2 Re 𝜓 d𝑡.
ℓ ≤ 21 | 𝛼+𝛽 |

We apply Corollary 24.A.5 with 𝜆 = |𝜉 ′ |, 𝜈 = 0, 𝜃 = 0, 𝜃 ′ = 0, 𝑚 = 2𝑘:

∑︁ ∫ 𝑥𝑛
1
′ 1+ℓ− |𝛽 |
|𝜉 | (𝑥 𝑛 − 𝑡) ℓ e− 2 Re 𝜓 d𝑡
−𝜏
ℓ ≤ 21 ( | 𝛼 |+|𝛽 |)−1
2𝑘
≤ 𝐶 (vii) max |𝜉 ′ | 2(𝑘+1) (1+ℓ)− |𝛽 |
ℓ ≤ 21 ( | 𝛼 |+|𝛽 |)−1
𝑘
= 𝐶 (vii) |𝜉 ′ | 2𝑘+1 ( | 𝛼 |+|𝛽 |)− |𝛽 | . □
1020 24 Analytic PDEs of Principal Type. Regularity of the Solutions

24.4.6 A class of operator-valued pseudodifferential operators

Lemmas 24.4.13 and 24.4.14 suggest that we take a look at the class of nonstandard
symbols satisfying estimates of the type (24.4.43) or (24.4.47). For the sake of
generality we define these symbols in a domain Ω of R𝑛 . (In the applications of
interest to us Ω will be replaced by the open subset 𝑈 of R𝑛−1 .) Moreover, we need
nonscalar symbols, symbols valued in the Banach algebra L (H) of bounded linear
operators on a (complex) Hilbert space H. [In the applications of interest to us, here,
H = 𝐿 2 (−𝜏, 𝜏).] Thus we consider a C ∞ function 𝒂 (𝑥, 𝜉) in U valued in L (H) that
satisfies the following inequalities, for some 𝑚 ∈ R and all (𝛼, 𝛽) ∈ Z+𝑛 × Z+𝑛 ,

∀ (𝑥, 𝜉) ∈ 𝑇 ∗ Ω\0, 𝜕𝑥𝛼 𝜕 𝜉 𝒂 (𝑥, 𝜉) ≤ 𝐶 𝛼,𝛽 (1 + |𝜉 |) 𝑚+ 𝛿 | 𝛼 |−𝜌 | 𝛼 |

𝛽
(24.4.52)

(∥·∥: the operator norm; the norm in H will be denoted by ∥·∥ H ). When H = C the
set of such symbols is often denoted by 𝑆 𝜌, 𝑚 (Ω); then 𝑆 𝑚 (Ω) denotes the space
𝛿 1,0
of standard symbols of order 𝑚 in Ω (Definition 16.2.1). We shall denote the set
of symbols 𝒂 satisfying (24.4.52) by 𝑆 𝜌, 𝑚 (Ω; L (H)). We introduce the following
𝛿
operator, at first acting on test-functions 𝒖 ∈ Cc∞ (Ω; H),
∫
−𝑛
𝑨𝒖 (𝑥) = (2𝜋) e𝑖 ( 𝑥−𝑦) · 𝜉 𝒂 (𝑥, 𝜉) 𝒖 (𝑦) d𝑦d𝜉. (24.4.53)
R2𝑛

The Fourier transform of 𝒖 is well-defined:

∫
𝒖 (𝜉) =
b e−𝑖𝑦· 𝜉 𝒖 (𝑦) d𝑦; (24.4.54)
R𝑛

it extends to C𝑛 as an entire function of exponential type valued in H; we can

therefore write ∫
−𝑛
𝑨𝒖 (𝑥) = (2𝜋) e𝑖 𝑥· 𝜉 𝒂 (𝑥, 𝜉) b
𝒖 (𝜉) d𝜉. (24.4.55)
R𝑛
We are going to need the Sobolev spaces of distributions in R𝑛 valued in H:
In (24.4.55) 𝒖 can be taken to be a tempered distribution in R𝑛 valued in H, i.e.,
𝒖 ∈ S ′ (R𝑛 ; H). By 𝐻 𝑠 (R𝑛 ; H) (𝑠 ∈ R) we shall mean the space of distributions
𝒖 ∈ S ′ (R𝑛 ; H) whose Fourier transform is a measurable function in R𝑛 valued in
H, such that
∫ 𝑠
∥𝒖∥ 2𝑠 = (2𝜋) −𝑛 ∥b𝒖 (𝜉)∥ 2H 1 + |𝜉 | 2 d𝜉 < +∞; (24.4.56)
R𝑛

𝐻 𝑠 (R𝑛 ; H) is a Hilbert space with inner product

∫ 𝑠
(𝒖, 𝒗) 𝑠 = (2𝜋) −𝑛 (b 𝒗 (𝜉)) H 1 + |𝜉 | 2 d𝜉
𝒖 (𝜉) , b (24.4.57)
R𝑛
24.4 Property (Q) Implies Subellipticity 1021

[(·, ·) H : inner product in H]. We denote by 𝐻c𝑠 ( Ω; H) the space of distributions

𝒖 ∈ 𝐻 𝑠 (R𝑛 ; H) such that supp 𝒖 ⊂⊂ Ω and by 𝐻loc𝑠 (
Ω; H) the space of distributions
𝒖 ∈ D ′ (Ω; H) such that 𝜒𝒖 ∈ 𝐻 𝑠 (R𝑛 ; H) whatever 𝜒 ∈ Cc∞ (Ω).
The following result is important for the proof developed in this section.

Proposition 24.4.15 Assume that 0 < 𝛿 < 𝜌 ≤ 1 − 𝛿 and let 𝑠,𝑚 ∈ R be arbitrary.
𝑚 (Ω; L (H)) then 𝑨 [defined in (24.4.53)] induces a continuous linear
If 𝒂 ∈ 𝑆 𝜌, 𝛿
operator 𝐻c𝑠 ( Ω; H) −→ 𝐻loc
𝑠−𝑚 (
Ω; H).

Notice that the hypotheses in this statement imply 𝛿 < 1/2.

Proof We start by proving the claim when 𝑚 = 𝑠 = 0. Suppose
−ℓ (𝜌− 𝛿)
𝒂 ∈ 𝑆 𝜌, 𝛿 (Ω; L (H)) , ℓ ∈ Z+ large.

Let 𝜒 ∈ Cc∞ (Ω), 𝑢 ∈ Cc∞ (Ω; H) be arbitrary; we have

∫
𝜒 𝑨𝒖 (𝜉) = (2𝜋) −𝑛
š c𝒂 (𝜉 − 𝜂, 𝜂) b
𝜒 𝒖 (𝜂) d𝜂,
R𝑛

where ∫
c𝒂 (𝜉 − 𝜂, 𝜂) =
𝜒 e−𝑖 𝑥· ( 𝜉 −𝜂) 𝜒 (𝑥) 𝒂 (𝑥, 𝜂) d𝑥.
R𝑛
We have
𝑁 ∫
1 + |𝜉 − 𝜂| 2 𝜒 c𝒂 (𝜉 − 𝜂, 𝜂) = e−𝑖 𝑥· ( 𝜉 −𝜂) (1 − Δ 𝑥 ) 𝑁 ( 𝜒 (𝑥) 𝒂 (𝑥, 𝜂)) d𝑥,
R𝑛

whence
−𝑁 − 21 ℓ (𝜌− 𝛿)+ 𝛿 𝑁
c𝒂 (𝜉 − 𝜂, 𝜂)∥ ≤ 𝐶 1 + |𝜉 − 𝜂| 2
∥𝜒 1 + |𝜂| 2 .

If 2𝑁 > 𝑛 and ℓ (𝜌 − 𝛿) > 𝑛 + 𝛿𝑁 the classical inequality for convolution implies

∫
d𝜉
∥ 𝜒 𝑨𝒖∥ 𝐿 2 ≤ ∥𝒖∥ 𝐿 2 sup ∥𝜒
c𝑎 (𝜉, 𝜂) ∥ 𝑁 .
𝜂 ∈R𝑛
1 + |𝜉 | 2

We reason by descending induction on ℓ =, ..., 1, 0; we select 𝑀 ∈ C ∞ (Ω),

𝑀 > 0, such that the operator on H,
−ℓ (𝜌− 𝛿)
𝑀 (𝑥) 1 + |𝜉 | 2 𝐼H − 𝒂 (𝑥, 𝜉) ∗ 𝒂 (𝑥, 𝜉) (24.4.58)

is positive definite for all (𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}) [𝒂 (𝑥, 𝜉) ∗ : adjoint of the bounded
linear operator on H, 𝒂 (𝑥, 𝜉); 𝐼H : identity operator on H]. It can be checked di-
−ℓ (𝜌− 𝛿)
rectly that the square-root of (24.4.58), 𝒃 (𝑥, 𝜉), belongs to 𝑆 𝜌, 𝛿 (Ω; L (H)). Let
𝑩 denote the pseudodifferential operator valued in L (H) defined by substituting
1022 24 Analytic PDEs of Principal Type. Regularity of the Solutions

𝒃 (𝑥, 𝜉) for 𝒂 (𝑥, 𝜉) in (24.4.53). The symbol of the composite 𝑨∗ 𝜒 (𝑥) 𝑨 differs
−(ℓ+1) (𝜌+ 𝛿)
from 𝜒 (𝑥) 𝒂 (𝑥, 𝜉) ∗ 𝒂 (𝑥, 𝜉) by a symbol belonging to 𝑆 𝜌, 𝛿 (Ω; L (H)). The
operator
1 1
𝑹 = (1 − Δ) − 2 ℓ (𝜌− 𝛿) 𝜒 (𝑥) 𝑀 (𝑥) (1 − Δ) − 2 ℓ (𝜌− 𝛿) − 𝑨∗ 𝜒 (𝑥) 𝑨 − 𝑩∗ 𝜒 (𝑥) 𝑩
−(ℓ+1) (𝜌+ 𝛿)
can then be defined by a symbol belonging to 𝑆 𝜌, 𝛿 (Ω; L (H)) We derive:
∫ 2
∫
1
𝜒 (𝑥) 𝑀 (𝑥) (1 − Δ) − 2 ℓ (𝜌− 𝛿) 𝒖 (𝑥) d𝑥 ≥ 𝜒 (𝑥) ∥ 𝑨𝒖 (𝑥) ∥ 2H d𝑥
H
∫ ∫
+ 𝜒 (𝑥) ∥ 𝑩𝒖 (𝑥)∥ 2H d𝑥 − (𝒖 (𝑥) , 𝑹𝒖 (𝑥)) H d𝑥.

The induction on ℓ implies

∫
(𝒖 (𝑥) , 𝑹𝒖 (𝑥)) H d𝑥 ≤ 𝐶 ∥𝒖∥ 2𝐿 2

−ℓ (𝜌− 𝛿)
for some 𝐶 > 0 independent of 𝑢 ∈ Cc∞ (Ω; H). Thus, if 𝒂 ∈ 𝑆 𝜌, 𝛿 (Ω; L (H))
then ∫ ∫
𝜒 (𝑥) ∥ 𝑨𝒖 (𝑥)∥ 2H d𝑥 ≤ 𝐶 ′ ∥𝒖 (𝑥) ∥ 2H d𝑥.

This proves that if 𝒂 ∈ 𝑆 0𝜌, 𝛿 (Ω; L (H)) then 𝑨 maps 𝐿 c2 (Ω; H) into 𝐿 loc
2 (Ω; H).

The case when 𝑚 and 𝑠 are arbitrary real numbers ensues from the eas-
ily proved fact that if 𝒂 ∈ 𝑆 𝜌, 𝑚 (Ω; L (H)) then (1 − Δ) −(𝑚−𝑠)/2 𝑨 (1 − Δ) −𝑠/2
𝛿
is defined [cf. (24.4.53)] by a symbol belonging to 𝑆 0𝜌, 𝛿 (Ω; L (H)), implying
that (1 − Δ) 𝑠/2 𝑨𝜒 (𝑥) (1 − Δ) −𝑠/2 maps 𝐿 c2 (Ω; H) into 𝐻loc
−𝑚 (Ω; H) and 𝑨 maps

𝐻c (Ω; H) into 𝐻loc (Ω; H).

𝑠 𝑠−𝑚
□

Remark 24.4.16 Under the hypotheses of Proposition 24.4.15 a calculus of pseu-

dodifferential operators of type (𝜌, 𝛿) can be developed by duplicating, practically
verbatim, the definitions and proofs for standard pseudodifferential operators (see
Ch. 16). For an elementary presentation we refer the reader to [Treves, 1980], p. 223
et seq. For a much more general and detailed theory see [Hörmander, 1983, III], Ch.
XVIII.

To apply Proposition 24.4.15 with the choice H = 𝐿 2 (−𝜏, 𝜏) we will use the
following.

Lemma 24.4.17 Let ( 𝑿, d𝑥) and (𝒀, d𝑦) be two measure spaces. Let 𝐹 (𝑥, 𝑦) be a
complex-valued function on 𝑋 × 𝑌 , measurable for the measure d𝑥d𝑦. If
∫ ∫
′
𝒀 ∋ 𝑦 ↦→ |𝐹 (𝑥, 𝑦 ′) 𝐹 (𝑥, 𝑦)| d𝑥d𝑦 (24.4.59)
𝒀 𝑿
24.4 Property (Q) Implies Subellipticity 1023
∫
belongs to 𝐿 ∞ (𝒀, d𝑦) then ℎ (𝑦) ↦→ 𝒀 𝐹 (𝑥, 𝑦) ℎ (𝑦) d𝑦 is a bounded linear map
𝐿 2 (𝒀, d𝑦) −→ 𝐿 2 ( 𝑿, d𝑥) whose norm does not exceed the square-root of twice the
𝐿 ∞ norm of (24.4.59).
Proof We have
∫ ∫ 2
𝐹 (𝑥, 𝑦) ℎ (𝑦) d𝑦 d𝑥
∫𝑿 ∫ 𝒀
≤ |𝐹 (𝑥, 𝑦) 𝐹 (𝑥, 𝑦 ′)| |ℎ (𝑦)| |ℎ (𝑦 ′)| d𝑥d𝑦d𝑦 ′
𝑿 𝒀 ×𝒀 , |ℎ( 𝑦) | ≤ |ℎ( 𝑦 ′ ) |
∫ ∫
+ 𝐹 (𝑥, 𝑦) 𝐹 (𝑥, 𝑦 ′) |ℎ (𝑦)| |ℎ (𝑦 ′)| d𝑥d𝑦d𝑦 ′
𝑿 𝒀 ×𝒀 , |ℎ( 𝑦′ ) | ≤ |ℎ( 𝑦) |
∫ ∫
≤ |𝐹 (𝑥, 𝑦 ) 𝐹 (𝑥, 𝑦)| d𝑥d𝑦 |ℎ (𝑦 ′)| 2 d𝑦 ′
′
𝒀 𝑿 ×𝒀
∫ ∫
+ ′ ′
|𝐹 (𝑥, 𝑦 ) 𝐹 (𝑥, 𝑦)| d𝑥d𝑦 |ℎ (𝑦)| 2 d𝑦,
𝒀 𝑿 ×𝒀

whence the claim. □

Corollary 24.4.18 If
∫ ∫
∞
|𝐹 (𝑥, 𝑦)| d𝑥 ∈ 𝐿 (𝒀, d𝑦) , |𝐹 (𝑥, 𝑦)| d𝑦 ∈ 𝐿 ∞ ( 𝑿, d𝑥) ,
𝑿 𝒀

𝐿∞
and if the ∫ norm of each one of these two functions does not exceed 𝐶 > 0 then
ℎ (𝑦) ↦→ 𝒀 𝐹 (𝑥, 𝑦) ℎ (𝑦) d𝑦 is a bounded linear map 𝐿 2 (𝒀, d𝑦) −→ 𝐿 2 ( 𝑿, d𝑥)
√
with norm ≤ 2𝐶.
Proof The 𝐿 ∞ norm of (24.4.59) does not exceed the 𝐿 ∞ norm of
∫ ∫
𝒀 ∋ 𝑦 ′ ↦→ |𝐹 (𝑥, 𝑦)| d𝑦 |𝐹 (𝑥, 𝑦 ′)| d𝑥,
𝒀 𝐿 ∞ ( 𝑿 ,d𝑥) 𝑿

which is ≤ 𝐶 2 . □

24.4.7 End of the proof that (Q) implies subellipticity

24.4.7.1 Microlocal parametrix in the C ∞ class under Hypothesis (Q)

We combine the results of the preceding subsections. Let 𝑎 (𝑡, 𝑥, 𝜉 ′) be a standard

symbol of order zero in 𝑈 ′ depending smoothly on (𝑡, 𝑥 𝑛 ) ∈ [−𝜏, 𝜏] 2 and let 𝛼, 𝛽 ∈
Z+𝑛−1 . Regarding (𝑥 ′, 𝜉 ′) as parameters we denote by 𝐹𝛼,𝛽 (𝑥 𝑛 , 𝑡) the function
′

(𝑥 𝑛 , 𝑡) ↦→ Υ (𝑥 𝑛 − 𝑡) 𝜕𝑥𝛼′ 𝜕 𝜉 ′ e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) .
𝛽
(24.4.60)
1024 24 Analytic PDEs of Principal Type. Regularity of the Solutions

(Υ: the Heaviside function). Lemma 24.4.13 states that

∫ 𝜏
𝐹𝛼,𝛽 (𝑥 𝑛 , 𝑡) d𝑥 𝑛 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | −𝑠◦ + 𝛿 | 𝛼 |−𝜌 |𝛽 | ,
𝑡
∫ 𝑥𝑛
𝐹𝛼,𝛽 (𝑥 𝑛 , 𝑡) d𝑥 𝑛 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | −𝑠◦ + 𝛿 | 𝛼 |−𝜌 |𝛽 | ,
−𝜏

with 𝑠◦ , 𝜌, 𝛿 given by (24.4.42). By Corollary 24.4.18 we see that (24.4.60) defines

a bounded linear operator on 𝐿 2 (−𝜏, 𝜏),
∫ 𝜏
ℎ ↦→ 𝐹𝛼,𝛽 (𝑥 𝑛 , 𝑡) ℎ (𝑡) d𝑡,
−𝜏
√
with operator-norm ≤ 2𝐶 𝛼,𝛽 |𝜉 ′ | −𝑠◦ + 𝛿 | 𝛼 |−𝜌 |𝛽 | . In the terminology of the preceding
′
subsection, this means that Υ (𝑥 𝑛 − 𝑡) e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) ∈ 𝑆 −𝑠 ◦ ′
𝜌, 𝛿 (𝑈 ; L (H)), H =
2
𝐿 (−𝜏, 𝜏). The oscillatory integral, in which 𝑢 ∈ Cc (𝑈 ), ∞ ′

∫
′ ′ ′
(2𝜋) 1−𝑛 Υ (𝑥 𝑛 − 𝑡) e𝑖 𝑥 · 𝜉 −𝜓 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) b
𝑢 (𝜉 ′) d𝜉 ′
∫
′
= (2𝜋) 1−𝑛 Υ (𝑥 𝑛 − 𝑡) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) b
𝑢 (𝜉 ′) d𝜉 ′,

defines a pseudodifferential operator of type (𝜌, 𝛿) in 𝑈 ′ and order −𝑠◦ valued in

the algebra of bounded linear operators on 𝐿 2 (−𝜏, 𝜏). In this integral we may let 𝑢
depend also on 𝑥 𝑛 : 𝑢 ∈ Cc∞ (𝑈 ′ × (−𝜏, 𝜏)). Proposition 24.4.15 allows us to conclude
that the integral
∫ 𝑥𝑛 ∫
′
(2𝜋) 1−𝑛
e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) b
𝑢 (𝜉 ′, 𝑡) d𝜉 ′d𝑡 (24.4.61)
−𝜏 R𝑛−1

defines a bounded linear operator

𝑨 : 𝐿 2 −𝜏, 𝜏; 𝐻c𝑠 (𝑈 ′) −→ 𝐿 2 −𝜏, 𝜏; 𝐻loc (𝑈 ′)
𝑠+𝑠◦
(24.4.62)

whatever 𝑠 ∈ R.
We can duplicate almost verbatim the preceding reasoning when applying Lemma
24.4.14, which states that
∫ 𝜏
𝜕𝐹𝛼,𝛽
(𝑥 𝑛 , 𝑡) d𝑥 𝑛 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | 𝛿 | 𝛼 |−𝜌 |𝛽 | ,
−𝜏 𝜕𝑥 𝑛
∫ 𝜏
𝜕𝐹𝛼,𝛽
(𝑥 𝑛 , 𝑡) d𝑥 𝑛 ≤ 𝐶 𝛼,𝛽 |𝜉 ′ | 𝛿 | 𝛼 |−𝜌 |𝛽 | .
−𝜏 𝜕𝑥 𝑛

Through Proposition 24.4.15 it then implies that the integral

∫ 𝑥𝑛 ∫
1−𝑛 𝜕 𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ′ )
(2𝜋) e 𝑎 (𝑡, 𝑥, 𝜉 ′) b𝑢 (𝜉 ′, 𝑡) d𝜉 ′d𝑡 (24.4.63)
−𝜏 R𝑛−1 𝜕𝑥 𝑛
24.4 Property (Q) Implies Subellipticity 1025

defines a continuous linear operator

𝑨1 : 𝐿 2 −𝜏, 𝜏; 𝐻c𝑠 (𝑈 ′) −→ 𝐿 2 −𝜏, 𝜏; 𝐻loc (𝑈 ′)

𝑠

whatever 𝑠 ∈ R; (24.4.1) implies, for every 𝑢 ∈ Cc∞ (𝑈 ′ × (−𝜏, 𝜏)),

∫
𝜕 1 ′ ′
𝑨𝑢 − 𝑨1 𝑢 = e𝑖 𝑥 · 𝜉 𝑎 (𝑥 𝑛 , 𝑥, 𝜉 ′) b
𝑢 (𝜉 ′, 𝑥 𝑛 ) d𝜉 ′. (24.4.64)
𝜕𝑥 𝑛 (2𝜋) 𝑛−1 R𝑛−1

This implies that 𝜕𝑥𝜕𝑛 𝑨 defines a continuous linear operator 𝐿 2 −𝜏, 𝜏; 𝐻c𝑠 (𝑈 ′) −→


𝑠 (𝑈 ′ )
𝐿 2 −𝜏, 𝜏; 𝐻loc .
′

We apply these results to 𝑎 = 𝑎 [𝑁 ] = 𝑁𝑗=0 𝑎 𝑗 with 𝑎 𝑗 𝑡, 𝑥, | 𝜉𝜉 ′ | solving
Í

(24.4.28)–(24.4.29) and satisfying (24.4.20). This leads us to identify the integral

(24.4.61) with 𝑨 𝜑[𝑁 ] 𝑢 given by (24.4.33).

Remark 24.4.19 The integrals with respect to 𝜉 ′ such as (24.4.61) are carried out
over the whole of R𝑛−1 whereas those defining 𝑨 𝜑[𝑁 ] and 𝑹 𝜑[𝑁 ] , (24.4.33) and
(24.4.34) respectively, are carried out over the domain |𝜉 ′ | > 1. The difference
between the two integrations produces operators that are smoothing in 𝑥 ′-space, and
thus map 𝐿 2 −𝜏, 𝜏; 𝐻c𝑠 (𝑈 ′) into 𝐿 2 (−𝜏, 𝜏; C ∞ (𝑈 ′)) whatever 𝑠 ∈ R, and as a
consequence, do not have any effect on the results in the present subsection.

We avail ourselves of Proposition 24.4.8 and end up with the equation (24.4.39)
where 𝑢 ∈ Cc∞ (−𝜏◦ , 𝜏◦ ) ; 𝐻c𝑠 (𝑉 ′) (𝑠 ∈ R arbitrary): regarding the operators as
acting on 𝐿 2 ((−𝜏◦ , 𝜏◦ ) ; E ′ (𝑉 ′)) we have the following relation between operators
acting on Cc∞ (−𝜏◦ , 𝜏◦ ) ; 𝐻c𝑠 (𝑉 ′) :

𝜕
− 𝐵 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] = 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ )+𝑹 𝜑[ 𝑁 ] +𝑺. (24.4.65)
′
𝜕𝑥 𝑛

The new feature is that, under Hypothesis (Q), 𝑨 𝜑[𝑁 ] is a continuous linear operator

𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻c𝑠 (𝑉 ′) −→ 𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻loc (𝑉 ′)
𝑠+𝑠◦
(24.4.66)

while 𝑨 𝜑[𝑁 ] 𝜕𝑥𝜕𝑛 and 𝜕

𝜕𝑥𝑛 𝑨 𝜑[𝑁 ] are continuous linear operators

𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻c𝑠 (𝑉 ′) −→ 𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻loc (𝑉 ′) .

𝑠
(24.4.67)

By Proposition 24.4.11, to every 𝑘 ∈ R there is an 𝑁 > 0 such that the operator

𝑹 𝜑[𝑁 ] is a continuous linear operator

𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻c𝑠 (𝑉 ′) −→ 𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻loc (𝑉 ′) .
𝑠+𝑘
(24.4.68)

In (24.4.66), (24.4.67) and (24.4.68) 𝑠 ∈ R is arbitrary.

1026 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Actually, we exploit the construction of 𝑨 𝜑[𝑁 ] and 𝑹 𝜑[𝑁 ] for the transpose of
𝑃 = 𝑃 (𝑥, D 𝑥 ), 𝑃 (𝑥, D 𝑥 ) ⊤ , instead of 𝑃. Since the principal symbol of 𝑃⊤ differs
from 𝑃𝑚 (𝑥, D 𝑥 ) only by a factor (−1) 𝑚 , there is nothing much to modify in said
construction; we may replace the cone Γ ′ by its opposite −Γ ′ and 𝜉 ◦ by −𝜉 ◦ but,
anyway, the unit covector 𝜉 ◦ was arbitrary. We note that
⊤
𝜕 𝜕
𝜒2 (−D 𝑥′ ) − 𝐵 (𝑥, D 𝑥′ ) + − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (−D 𝑥′ )
𝜕𝑥 𝑛 𝜕𝑥 𝑛
= − (𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ )) ⊤ − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (−D 𝑥′ )

is a standard pseudodifferential operator 𝑻 in 𝑉 ′ of order zero depending smoothly

on 𝑥 𝑛 ∈ (−𝜏, 𝜏) [cf. (16.2.13)]. With minor modifications of 𝑨 𝜑[𝑁 ] , 𝑹 𝜑[𝑁 ] and 𝑺 we
obtain the analogue of (24.4.65):
⊤
𝜕
− 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (−D 𝑥′ ) 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] = 𝜒1 (𝑥 ′) 𝜒2 (−D 𝑥′ )
𝜕𝑥 𝑛
+ 𝑻 𝜒1 (𝑥 ′) 𝑨 𝜑[𝑁 ] + 𝑹 𝜑[𝑁 ] + 𝑺.
2 ′

𝑠 ∈ R is arbitrary
Since and 𝐿 −𝜏◦ , 𝜏◦ ; 𝐻c (𝑉 ) can be identified with the dual of
𝑠

−𝑠 (𝑉 ′ )
𝐿 2 −𝜏◦ , 𝜏◦ ; 𝐻loc and vice versa, the transpose 𝑨 𝜑[𝑁 ]⊤ of 𝑨 𝜑[𝑁 ] is a continuous
linear operator (24.4.66) and the transpose 𝑹 𝜑[𝑁 ]⊤ of 𝑹 𝜑[𝑁 ] is a continuous linear
operator (24.4.68). By transposition we derive from the preceding formula

[𝑁 ]⊤ ′ 𝜕
𝑨 𝜑 𝜒1 (𝑥 ) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) = 𝜒2 (D 𝑥′ ) 𝜒1 (𝑥 ′)
𝜕𝑥 𝑛
+ 𝑨 𝜑[𝑁 ]⊤ 𝜒1 (𝑥 ′) 𝑻 ⊤ + 𝑹 𝜑[𝑁 ]⊤ + 𝑺⊤ .

Since 𝜒1 ≡ 1 in 𝑉 ′ the operator [ 𝜒2 (D 𝑥′ ) , 𝜒1 (𝑥 ′)] is smoothing in 𝑉 ′; it maps

𝑠′ (𝑉 ′ )
𝐻c𝑠 (𝑉 ′) into 𝐻𝑙𝑜c whatever 𝑠, 𝑠 ′ ∈ R. We have the following equation between
operators on 𝐿 2 (−𝜏◦ , 𝜏◦ ) ; 𝐻c𝑠 (𝑉 ′) ,

[𝑁 ]⊤ ′ 𝜕
𝑨 𝜑 𝜒1 (𝑥 ) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) = 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) (24.4.69)
𝜕𝑥 𝑛
+ 𝑨 𝜑[𝑁 ]⊤ 𝜒1 (𝑥 ′) 𝑻 ⊤ + 𝑹 𝜑[𝑁 ]⊤ + 𝑺 ′

with 𝑺 ′ smoothing in 𝑉 ′.
24.4 Property (Q) Implies Subellipticity 1027

24.4.7.2 Proof that (Q) implies (c)

From (24.4.69) we derive, for every 𝑢 ∈ Cc∞ (𝑉 ′ × (−𝜏◦ , 𝜏◦ )),

∫ 𝜏∫
𝑨 𝜑[𝑁 ] 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢 (𝑥) 𝜒1 (𝑥 ′)
−𝜏 𝑈 ′

𝜕
× − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 (𝑥) d𝑥
𝜕𝑥 𝑛
= ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐿 2 (R𝑛 )
∫ 𝜏∫
+ 𝑨 𝜑[𝑁 ] 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢 (𝑥) 𝑻 𝜒1 (𝑥 ′) 𝑢 (𝑥) d𝑥
′
∫−𝜏𝜏 ∫𝑈
+ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢 (𝑥) 𝑹 𝜑[𝑁 ]⊤ 𝑢 (𝑥) d𝑥
′
∫−𝜏𝜏 ∫𝑈
+ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢 (𝑥) 𝑺 ′𝑢 (𝑥) d𝑥.
−𝜏 𝑈′

By the Cauchy–Schwarz inequality and the fact that 𝑨 𝜑[𝑁 ] is a continuous linear
operator (24.4.66) we get

∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐿 2 (R𝑛 )
∫ 𝜏◦ 2
𝜕
≤𝐶 𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛
−𝜏◦ 𝜕𝑥 𝑛 𝐻 −𝑠◦ ( R𝑛−1 )
∫ 𝜏◦
+𝐶 ∥𝑢∥ 2𝐻 −𝑠◦ R𝑛−1 d𝑥 𝑛 .
−𝜏◦
( )
1
Inserting the usual operators (1 − Δ 𝑥′ ) 2 (𝑠+𝑠◦ ) and keeping in mind that commutators
of standard pseudodifferential operators of order 𝑑 and 𝑑 ′ respectively have order
𝑑 + 𝑑 ′ − 1, we see that to every 𝑠 ∈ R there are constants 𝐶 (𝑠) > 0 such that
∫ 𝜏◦
∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐻 𝑠+𝑠◦ ( R𝑛−1 ) d𝑥 𝑛 (24.4.70)
−𝜏◦
∫ 𝜏◦ 2
𝜕
≤ 𝐶 (𝑠) 𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛
−𝜏◦ 𝜕𝑥 𝑛 𝐻 𝑠 ( R𝑛−1 )
∫ 𝜏◦
+ 𝐶 (𝑠) ∥𝑢∥ 2𝐻 𝑠 d𝑥 𝑛
−𝜏◦
( R𝑛−1 )

for every 𝑢 ∈ Cc∞ (𝑉 ′ × (−𝜏◦ , 𝜏◦ )). For simplicity, let us take 𝑠 = 0, in which case
1028 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫ 𝜏◦ 2
𝜕
𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛 (24.4.71)
−𝜏◦ 𝜕𝑥 𝑛 𝐻 0 ( R𝑛−1 )
2
𝜕
= 𝜒1 (𝑥 ′) 𝐸 ◦ (𝑥, D 𝑥 ) −1 𝐸 ◦ (𝑥, D 𝑥 ) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢
𝜕𝑥 𝑛 𝐿 2 (R𝑛 )

≥ ∥ 𝜒1 (𝑥 ′) 𝑃𝑚 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐻 1−𝑚 (R𝑛 )

− 𝐶1 ∥ 𝜒1 (𝑥 ′) 𝑃𝑚 (𝑥, D 𝑥 ) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐻 −𝑚 (R𝑛 )
≥ ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑃 (𝑥, D 𝑥 ) 𝑢∥ 2𝐻 1−𝑚 (R𝑛 ) − 𝐶2 ∥𝑢∥ 2𝐿 2 (R𝑛 )

since the commutator 𝜒1 (𝑥 ′) , 𝐸 ◦ (𝑥, D 𝑥 ) −1 has order −𝑚 while

[𝑃𝑚 (𝑥, D 𝑥 ) , 𝜒2 (D 𝑥′ )] and 𝑃 (𝑥, D 𝑥 ) − 𝑃𝑚 (𝑥, D 𝑥 )

have order 𝑚 − 1. Thanks to these lower bounds we derive from (24.4.70):

∫ 𝜏◦
∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐻 𝑠◦ ( R𝑛−1 ) d𝑥 𝑛 (24.4.72)
−𝜏◦

≤ 𝐶3 ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑃 (𝑥, D 𝑥 ) 𝑢∥ 2𝐻 1−𝑚 (R𝑛 ) + 𝐶3 ∥𝑢∥ 2𝐿 2 (R𝑛 ) .

We also exploit the analogue of (24.4.64): 𝜕𝑥𝜕𝑛 𝑨 [𝑁 ]⊤ defines a continuous lin-

ear operator (24.4.67) and so, therefore, does its transpose 𝑨 [𝑁 ] 𝜕𝑥𝜕𝑛 . We have the
analogue of (24.4.69):

𝜕 𝜕 𝜕𝑢
𝑨 𝜑[𝑁 ]⊤ 𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 = 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ )
𝜕𝑥 𝑛 𝜕𝑥 𝑛 𝜕𝑥 𝑛
𝜕 𝜕 [𝑁 ]⊤ 𝜕𝑢
+𝑨 𝜑[𝑁 ]⊤ 𝜒1 (𝑥 ′) 𝑻 ⊤ 𝑢 + 𝑹 𝜑 𝑢 + 𝑺 ′′
𝜕𝑥 𝑛 𝜕𝑥 𝑛 𝜕𝑥 𝑛
with 𝑺 ′′ smoothing in 𝑉 ′ and depending smoothly on 𝑥 𝑛 and 𝑢 ∈ Cc∞ (𝑉 ′ × (−𝜏◦ , 𝜏◦ )).
In this connection it is helpful to require that the positive square-roots of 𝜒1
and 𝜒2 be C ∞ functions with the same properties (a requirement easy to satisfy
by replacing 𝜒1 and 𝜒2 , from the start, by 𝜒12 and 𝜒22 ). If we take then the in-
ner product in 𝐿 2 −𝜏, 𝜏; 𝐻 𝑠 R𝑛−1 of both sides of the preceding equation with

𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝜕𝑥
𝜕𝑢
𝑛
we obtain, for 𝑠 ′ < 𝑠 arbitrary, suitably large positive constants
′′
𝐶𝑠 , 𝐶𝑠,𝑠′ ,
24.4 Property (Q) Implies Subellipticity 1029

𝜕𝑢 2
∫ 𝜏◦
𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ )
d𝑥 𝑛
−𝜏◦ 𝜕𝑥 𝑛 𝐻 𝑠 ( R𝑛−1 )
∫ 𝜏◦ 2
𝜕
≤ 𝐶𝑠′′ 𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛
−𝜏◦ 𝜕𝑥 𝑛 𝐻 𝑠 ( R𝑛−1 )
∫ 𝜏◦ ∫ 𝜏◦ 2
′′ 2 √ 𝜕𝑢 √
+𝐶𝑠 ∥𝑢∥ 𝐻 𝑠 R𝑛−1 d𝑥 𝑛 + 𝐶𝑠,𝑠′ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛 .
−𝜏◦
( ) −𝜏◦ 𝜕𝑥 𝑛 ′
𝐻 𝑠 ( R𝑛−1 )

In turn, taking 𝑠 ′ ≤ 𝑠 − 1 this leads to

𝜕𝑢 2
∫ 𝜏◦ ∫ 𝜏◦
′ ′′
𝜒1 (𝑥 ) 𝜒2 (D 𝑥′ ) d𝑥 𝑛 ≤ 𝐶𝑠 ∥𝑢∥ 2𝐻 𝑠 R𝑛−1 d𝑥 𝑛
−𝜏◦ 𝜕𝑥 𝑛 𝐻 𝑠 ( R𝑛−1 ) −𝜏◦
( )
∫ 𝜏◦ 2
𝜕
+𝐶𝑠′′′ 𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛
−𝜏◦ 𝜕𝑥 𝑛 𝐻 𝑠 ( R𝑛−1 )
∫ 𝜏◦ 2
√ 𝜕 √
+𝐶𝑠,𝑠′ 𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛 .
𝜕𝑥 𝑛 ′
−𝜏◦ 𝐻 𝑠 ( R𝑛−1 )

Here too we take 𝑠 = 0 and 𝑠 ′ ≤ −1; we avail ourselves of (24.4.71) and also the
inequality
∫ 𝜏◦ 2
√ 𝜕 √
𝜒1 (𝑥 ′) − 𝐵 (𝑥, D 𝑥′ ) 𝜒2 (D 𝑥′ ) 𝑢 d𝑥 𝑛
𝜕𝑥 𝑛 ′
−𝜏◦ 𝐻 𝑠 ( R𝑛−1 )
√ √ 2
≤ 𝐶3 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑃 (𝑥, D 𝑥 ) 𝑢 𝐻 𝑠′ (R𝑛 ) + 𝐶3 ∥𝑢∥ 2𝐿 2 (R𝑛 )

to conclude
2
𝜕𝑢
𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) ≤ 𝐶0′′ ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 2𝐻 1−𝑚 (R𝑛 ) + 𝐶4 ∥𝑢∥ 2𝐿 2 (R𝑛 ) .
𝜕𝑥 𝑛 𝐿 2 (R𝑛 )
𝑠◦ 𝑠◦ (24.4.73)
2
Since 1 + |𝜉 | ≤ 1+ |𝜉 ′ | 2 2
+ |𝜉 𝑛 | we have
∫ 𝜏◦
′
∥ 𝜒1 (𝑥 ) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐻 𝑠◦ (R𝑛 ) ≤ ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 2𝐻 𝑠◦ ( R𝑛−1 ) d𝑥 𝑛
−𝜏◦
2
𝜕𝑢
+ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) .
𝜕𝑥 𝑛 𝐿 2 (R𝑛 )

By combining this estimate with (24.4.72)–(24.4.73) we obtain

𝜕𝑢
𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) + ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐻 𝑠◦ (R𝑛 )
𝜕𝑥 𝑛 𝐿 2 (R𝑛 )
≤ 𝐶5 ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 𝐻 1−𝑚 (R𝑛 ) + 𝐶5 ∥𝑢∥ 𝐿 2 (R𝑛 ) .
1030 24 Analytic PDEs of Principal Type. Regularity of the Solutions
1
By replacing 𝑢 by 𝜒 (𝑥) (1 − Δ 𝑥 ) 2 (𝑚−1) 𝑢 with 𝜒 ∈ Cc∞ (Ω), 𝜒 ≡ 1 in an open set
containing
h supp 𝑢, and by taking into
i account the fact that the pseudodifferential
1
operator 𝑃 (𝑥, D 𝑥 ) , (1 − Δ 𝑥 ) 2 (𝑚−1) has order 2 (𝑚 − 1), it is not difficult to derive
the inequality

𝜕𝑢
𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) + ∥ 𝜒1 (𝑥 ′) 𝜒2 (D 𝑥′ ) 𝑢∥ 𝐻 𝑚−1+𝑠◦ (R𝑛 ) (24.4.74)
𝜕𝑥 𝑛 𝐻 𝑚−1 (R𝑛 )
≤ 𝐶5 ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 𝐿 2 (R𝑛 ) + 𝐶6 ∥𝑢∥ 𝐻 𝑚−1 (R𝑛 ) .

This completes the proof that (Q)=⇒(c) in Theorem 24.2.1, at least microlocally;
the local result then ensues as explained in Subsection 24.1.3. We can state:
If Property (Q) in Theorem 24.2.1 holds and provided the neighborhood 𝑈 ⊂ Ω
of 𝑥 ◦ is sufficiently small there are numbers 𝜎◦ ∈ (0, 1], 𝐶 > 0 such that

∥𝑢∥ 𝐻 𝑚−1+𝜎◦ (R𝑛 ) ≤ 𝐶 ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 𝐿 2 (R𝑛 ) + 𝐶 ∥𝑢∥ 𝐻 𝑚−1 (R𝑛 ) (24.4.75)

for every 𝑢 ∈ Cc∞ (𝑈).

Replacing 𝑢 (𝑥) by 𝜒 (𝑥) (1 − Δ 𝑥 ) 𝑠/2 𝑢 (𝑥), 𝑠 ∈ R, 𝜒 ∈ Cc∞ (𝑈), 𝜒 ≡ 1 in a
neighborhood 𝑉 of 𝑥 ◦ , we derive from (24.4.76):

∥𝑢∥ 𝐻 𝑠+𝜎◦ (R𝑛 ) ≤ 𝐶 (𝑠) ∥𝑃 (𝑥, D 𝑥 ) 𝑢∥ 𝐻 𝑠−𝑚+1 (R𝑛 ) + 𝐶 (𝑠) ∥𝑢∥ 𝐻 𝑠 (R𝑛 ) (24.4.76)

for some 𝐶 (𝑠) > 0 and every 𝑢 ∈ Cc∞ (𝑉).

24.5 Analytic Hypoellipticity Implies (Q)

In this context also we shall reason by contradiction and hypothesize that (b) holds
and, simultaneously, that the curve 𝔠◦ defined by (24.3.5) is a null bicharacteristic
curve of 𝑃 (𝑥, D 𝑥 ) in 𝑈 × Γ. We make use of the holomorphic extensions of the
phase-function 𝜑 and the formal
amplitude 𝑎 of the previous subsection. Under
Hypothesis (A1) 𝜑 ∈ O 𝑈 C is the solution of the complex Cauchy problem

𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) = 0, 𝜑| 𝑧𝑛 =0 = 𝑧 ′ · 𝜉 ◦′ + 𝑖𝑧 ′ · 𝑧 ′ (24.5.1)

[cf. (23.2.5) and Remark 23.2.2].

As before, the formal power series 𝑎 (𝑧, 𝜆) = ∞ −𝑗
Í
𝑗=0 𝜆 𝑎 𝑗 (𝑧) is the solution
of (23.2.7) that satisfies (23.2.31). We seek approximate solutions of the equation
𝑃 (𝑥, D 𝑥 ) ℎ = 0 of the type ℎ [𝑁 ] = e𝑖𝜆𝜑 𝑎 [𝑁 ] where
∑︁
𝑎 [𝑁 ] (𝑧, 𝜆) = 𝜆− 𝑗 𝑎 𝑗 (𝑧, 𝜆) (24.5.2)
𝑗 ≤𝜆/𝑁
24.5 Analytic Hypoellipticity Implies (Q) 1031

with 𝑁 > 0 large. In passing, notice that [1, +∞) ∋ 𝜆 ↦→ 𝑎 [𝑁 ] (𝑧, 𝜆) is not
a continuous function: if 𝑚𝑁 < 𝜆 < (𝑚 + 1) 𝑁 (𝑚 ∈ Z+ ) , 𝑎 [𝑁 ] (𝑧, 𝜆) =
e𝑖𝜆𝜑 (𝑧) 𝑚 −𝑗
Í
𝑗=0 𝑗 (𝑧) 𝜆 . We shall fully exploit Proposition 23.2.7: 𝑎 𝑗 ∈ O 𝑈
𝑎 C

and the estimates (23.2.6) hold.

Lemma 24.5.1 Given an arbitrary compact subset 𝐾 of 𝑈 C let 𝐶𝐾 > 0 be the

constant in (23.2.6). If 𝑁 ≥ 𝐶𝐾 then

∀𝜆 ≥ 1,max 𝑎 [𝑁 ] (𝑧, 𝜆) ≤ 𝐶◦ 𝐶𝐾 (24.5.3)

𝑧 ∈𝐾

𝑗!
with 𝐶◦ = 2 sup √ .
𝑗 ∈Z+ 𝑗+1( 𝑗/𝑒) 𝑗

Proof By (23.2.6) and Stirling’s inequality we have

𝑗
1 √︁ 𝐶𝐾
𝜆− 𝑗 max 𝑎 𝑗 (𝑧) ≤ 𝐶◦ 𝐶𝐾 𝑗 + 1 ,
𝑧 ∈𝐾 2 e𝑁

whence
[𝜆/𝑁
∑︁ ] √︁ 𝑗
[𝑁 ] 1 𝐶𝐾
max 𝑎 (𝑧, 𝜆) ≤ 𝐶◦ 𝐶𝐾 𝑗 +1 .
𝑧 ∈𝐾 2 𝑗=0
e𝑁
Í∞ √︁
If 𝑁 ≥ 𝐶𝐾 then (24.5.3) follows from the fact that 𝑗=0 𝑗 + 1e− 𝑗 < 2. □
We return to (23.2.7) or, equivalently, to the system of equations (23.2.14)–
(23.2.15). We derive
∑︁
𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 (𝑧) 𝑎 [𝑁 ] (𝑧, 𝜆) = 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 (𝑧) 𝑎 𝑗 (𝑧, 𝜆) .
𝑗>𝜆/𝑁

Recalling that the order of 𝑃 is equal to 𝑚 and applying the generalized Leibniz rule
[cf. (23.2.9)] we get

∞ 𝑚−1
∑︁ ∑︁ ∑︁
e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 𝑗 =
𝛽
𝜆 𝑘 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗 ,
𝑗=0 𝑘=0 | 𝛼 | ≤𝑚−𝑘

whence
∑︁ 𝑚−1
∑︁ ∑︁
e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 [𝑁 ] =
𝛽
𝜆 𝑘− 𝑗 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗 .
𝑗>𝜆/𝑁 𝑘=0 | 𝛼 | ≤𝑚−𝑘
(24.5.4)

Proposition 24.5.2 To every compact subset 𝐾 of 𝑈 C and large 𝑁 > 0 there is a

𝐶˜𝐾 , 𝑁 > 0 such that

max e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶˜𝐾 , 𝑁 e−𝜆/𝑁 . (24.5.5)
𝐾
1032 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Proof Select 𝜀 > 0 such that 𝐾 𝜀 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝐾) ≤ 𝜀} ⊂ 𝑈 C . By the Cauchy

inequalities the estimates (23.2.6) imply
∑︁
′
𝐶𝐾 𝜀 𝜀 −𝑚 𝑚! 𝑗!
𝑗+1
max D𝑧𝛼 𝑎 𝑗 ≤ 𝐶𝑚
𝐾
| 𝛼 | ≤𝑚

′ =
where 𝐶𝑚
Í 𝛼!
| 𝛼 | ≤𝑚 𝑚! . We derive from (24.5.4) and Stirling’s inequality
∑︁
e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶𝑚
′′ −𝑚 𝑚
𝐶𝐾 𝜀 𝑗!𝜆− 𝑗
𝑗+1
𝜀 𝜆
𝑗>𝜆/𝑁
∑︁
′′
𝐶𝐾 𝜀 𝜀 −𝑚 𝜆 𝑚 𝐶𝐾 𝜀 𝑗! ( 𝑗 𝑁) − 𝑗
𝑗+1
≤ 𝐶𝑚
𝑗>𝜆/𝑁
𝑗
′′′
∑︁ √︁ 𝐶𝐾
≤ 𝐶𝑚 𝐶𝐾 𝜀 𝜀 −𝑚 𝜆 𝑚 𝑗 +1
e𝑁
𝑗>𝜆/𝑁
∑︁ √︁
′′′
≤ 𝐶𝑚 𝐶𝐾 𝜀 𝜀 −𝑚 𝜆 𝑚 e−𝜆/𝑁 𝑗 + 1 (𝐶𝐾 /𝑁) 𝑗 .
𝑗>𝜆/𝑁

Taking 𝑁 ≥ 2𝐶𝐾 yields

e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶𝑚
𝑖𝑣
𝐶𝐾 𝜀 𝜀 −𝑚 𝜆 𝑚+1 2−𝜆/𝑁 e−𝜆/𝑁 .

whence (24.5.5). □

Since we are assuming that (b) holds we can avail ourselves of Lemma 3.1.6: to
every compact subset 𝐾 C of 𝑈 C there exist a compact subset 𝐾1C of 𝑈 C , a compact
subset 𝐾 ′ of 𝑈 and a constant 𝐶 > 0 such that

max e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶max ′
e 𝑖𝜆𝜑 [𝑁 ]
𝑎 + 𝐶max 𝑃 (𝑧, D 𝑧 ) e 𝑖𝜆𝜑 [ 𝑁 ]
𝑎 . (24.5.6)
𝐾C 𝐾 𝐾1C

First, we exploit (24.3.11):

max
′
e𝑖𝜆𝜑 ≤ max
′
e−𝜆 Im 𝜓 ≤ 1
𝐾 𝐾

provided 𝑈 is sufficiently small, whence, by (24.5.3),

∀𝜆 ≥ 1, max
′
e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶◦ 𝐶𝐾 ′ . (24.5.7)
𝐾

From Proposition 24.5.2 we get

max 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 [ 𝑁 ] ≤ 𝐶˜𝐾 C , 𝑁 e𝜆 Re 𝜑−𝜆/𝑁 .
𝐾1C 1
24.6 Property (Q) Implies Analytic Hypoellipticity 1033

Here we adjust our choice of 𝑈 C . We have Re 𝜑 (𝑥 + 𝑖𝑦) = Im 𝜓 (𝑥) + 𝑂 (|𝑦|);

by (24.5.3) this implies |Re 𝜑 (𝑥 + 𝑖𝑦)| ≤ 𝐶 |𝑦| and we may therefore require that
|Re 𝜑 (𝑧)| ≤ 1/𝑁 whatever 𝑧 ∈ 𝑈 C , ensuring that

∀𝜆 ≥ 1, max 𝑃 (𝑧, D𝑧 ) e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶1 < +∞ (24.5.8)
𝐾1C

whatever the compact subset 𝐾1C of 𝑈 C . Last, we look at the left-hand side of (24.5.6)
where we select 𝐾 C = {(−𝑖𝑠𝜉 ◦′, 0, ..., 0)} with 𝑠 > 0 sufficiently small that 𝐾 C ⊂ 𝑈 C .
From (24.5.1) we deduce that Im 𝜑 (−𝑖𝑠𝜉 ◦′, 0, ..., 0) = −𝑠 (1 + 𝑠) and from (23.2.31)
that 𝑎 [ 𝑁 ] (−𝑖𝑠𝜉 ◦′, 0, ..., 0) = 1 + 𝑂 (𝑠), whence

∀𝜆 ≥ 1, max e𝑖𝜆𝜑 𝑎 [𝑁 ] = (1 + 𝑂 (𝑠)) e𝑠 (𝑠+1)𝜆 . (24.5.9)

𝐾C

Putting (24.5.7), (24.5.8) and (24.5.9) into (24.5.6) leads to a contradiction for 𝑠
sufficiently small and 𝜆 sufficiently large. This completes the proof that non-(Q)
entails non-(b).

24.6 Property (Q) Implies Analytic Hypoellipticity

24.6.1 Construction of an analytic parametrix I. Complex

phase-function and amplitudes

We shall once again reason in the complex domain ΩC ; we shall assume that every
coefficient of 𝑃 (𝑥, D) extends holomorphically to ΩC . We can, and shall, assume that
𝜕𝑚 𝑃𝑚 ◦ C ◦ ◦′ ◦′
𝜕 𝜉𝑛𝑚 (𝑧, 𝜉 ) ≠ 0 whatever 𝑧 ∈ Ω , 𝜉 = (𝜉 , 0) ∈ Char 𝑃, |𝜉 | = 1. [We continue to
′ ′
use the notation 𝜁 = (𝜁1 , ..., 𝜁 𝑛−1 ), 𝜉 = (𝜉1 , ..., 𝜉 𝑛−1 ).] As before, it is convenient to
assume that the central point 𝑥 ◦ ∈ Ω = ΩC ∩R𝑛 is the origin; as before, 𝑈 ⊂ Ω = ΩC ∩
R𝑛 is a neighborhood of 0 in R𝑛 ; Γ = {𝜉 ∈ R𝑛 ; 𝜉 ′ ∈ Γ ′, 𝜉 𝑛 < 𝜀◦ |𝜉 ′ |} (𝜀◦ > 0) with
Γ ′ ∋ 𝜉 ◦′ a convex open cone in R𝑛−1 \ {0}. We select an open subset 𝑈 C × ΓC of C𝑛 ×
(C𝑛 \ {0}) such that 𝑈 C ∩ R𝑛 = 𝑈 and ΓC = {𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛 ; 𝜉 ∈ Γ, |𝜂| < 𝜅 |𝜉 |}
(0 < 𝜅 ≪ 1) with the following property: the eikonal equation 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) = 0
has a unique solution 𝜑 ∈ O 𝑈 C × ΓC such that 𝜑| 𝑧𝑛 =𝑡 = 𝑧 ′ · 𝜁 ′ [cf. (24.4.1)];
𝜑 = 𝜑 (𝑡, 𝑧, 𝜁 ′) depends holomorphically on 𝑡 ∈ C, |𝑡| < 𝜏 (i.e., 𝑡 ∈ Δ 𝜏 ).
We avail ourselves of Proposition 23.1.12:

𝑃 (𝑧, 𝜁) = 𝐸 (𝑧, 𝜁) (𝑖𝜁 𝑛 − 𝐵 (𝑧, 𝜁 ′) + 𝐶 (𝑧, 𝜁 ′)) , (24.6.1)

where 𝐸 (𝑧, 𝜁) and 𝐶 (𝑧, 𝜁 ′) are classical analytic symbols in 𝑈 C × ΓC (cf. Definition

′
zero respectively; 𝐸 (𝑧, 𝜁) ≠ 0 for all (𝑧, ◦′
17.2.28) of order 𝑚 − 1 and 𝜁) ∈ 𝑈 C × ΓC ;
𝐵 (𝑧, 𝜁 ) ∈ O 𝑈 × Γ is homogeneous of degree 1 and 𝐵 (0, 𝜉 ) = 0. Thus
C C

(A1) holds. Moreover, provided 𝑈 C and ΓC ∩ S2𝑛−1 are sufficiently small, we have
1034 24 Analytic PDEs of Principal Type. Regularity of the Solutions

𝐸 ◦ (𝑧, 𝜕𝑧 𝜑) ≠ 0 and
𝜕𝜑
= 𝐵 (𝑧, 𝜕𝑧′ 𝜑) (24.6.2)
𝜕𝑧 𝑛
for all (𝑡, 𝑧, 𝜁) ∈ Δ 𝜏 × 𝑈 C × ΓC.
𝜉′
We revert briefly to omitting 𝑡, in the notation and setting |𝜉 ′ | = 𝜆. We revisit
| 𝜉′ |
the formal power series 𝑎 (𝑧, 𝜆) = ∞ −𝑗
Í
𝑗=0 𝜆 𝑎 𝑗 (𝑧) with coefficients 𝑎 𝑗 ∈ O 𝑈
C

that satisfy (23.2.7), meaning that they solve the transport equations, and satisfy
the “initial” conditions (24.4.20). We can apply Proposition 23.2.7: the 𝑎 𝑗 satisfy
Condition (FA) [cf. (23.2.6)]. Since

𝜕 𝜉𝛼′ 𝑃𝑚 (𝑧, 𝜕𝑧 𝜑) = 𝐸 ◦ (𝑧, 𝜕𝑧 𝜑) 𝜕𝜁𝛼 (𝑖𝜁 𝑛 − 𝐵 (𝑧, 𝜁 ′)) if |𝛼| = 1,

𝜁 =𝜕𝑧 𝜑

we see that (23.2.14)–(23.2.15) are equivalent to

𝔏 (𝑧, D𝑧 ) 𝑎 0 + 𝑐 0 𝑎 0 = 0, (24.6.3)
𝑗 ∑︁
∑︁ 1
𝔏 (𝑧, D𝑧 ) 𝑎 𝑗 + 𝑐 0 𝑎 𝑗 = 𝑐 𝑘, 𝛼 D𝑧𝛼 𝑎 𝑗−𝑘 , 𝑗 ≥ 1, (24.6.4)
𝛼!
𝑘=1 | 𝛼 | ≤𝑘

where 𝔏 (𝑧, D𝑧 ) is given by (24.4.26) and

𝑐 0 (𝑧) = 𝐸 ◦ (𝑧, 𝜕𝑧 𝜑) −1 𝑝 𝑚−1,0 𝑧, 𝜕𝑧 𝜑 ,
𝛽

𝑐 𝑘, 𝛼 (𝑧) = 𝐸 ◦ (𝑧, 𝜕𝑧 𝜑) −1 𝑝 𝑚−1−𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 .
𝛽

𝛽
The functionals 𝑝 𝑚−1−𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 are defined in (23.2.9): note that necessarily
|𝛼| ≤ 𝑚, 𝑘 ≤ 𝑚 − 1. Restricting to real-space and formally speaking, we now have,
by (23.2.7) and (24.6.1),

𝐸 (𝑥, D 𝑥 ) −1 𝑃 (𝑥, D 𝑥 ) e𝑖𝜆𝜑 ( 𝑥) 𝑎 (𝑥, 𝜆) (24.6.5)

𝜕
= − 𝐵 (𝑥, D 𝑥′ ) + 𝐶 (𝑥, D 𝑥′ ) e𝑖𝜆𝜑 ( 𝑥) 𝑎 (𝑥, 𝜆) = 0.
𝜕𝑥 𝑛

Remark 24.6.1 Strictly speaking, the action of 𝐵 (𝑥, D 𝑥′ ) and 𝐶 (𝑥, D 𝑥′ ) on

e𝑖𝜆𝜑 ( 𝑥) 𝑎 (𝑥, 𝜆) is not well-defined unless we interpret such action as that of properly
supported representatives of these pseudodifferential operators or, equivalently, if
we insert a cutoff 𝑔 (𝑥 ′) ∈ Cc∞ (𝑈 ′), 𝑔 ≡ 1 in a neighborhood of 0. But if we do
this the equality signs must be replaced by congruences mod symbols that decay
exponentially as 𝜆 ↗ +∞ in a neighborhood of (0, 𝜉 ′◦ ). The presence of 𝑔 has no
effect on the final result and we shall not burden the notation further by inserting it.
24.6 Property (Q) Implies Analytic Hypoellipticity 1035
′

It is now convenient to go back to the notation 𝑎 𝑗 (𝑡, 𝑧, 𝜉 ′) = |𝜉 ′ | − 𝑗 𝑎 𝑗 𝑡, 𝑧, | 𝜉𝜉 ′ | ;
thus 𝑎 (𝑡, 𝑧, 𝜉 ′) = +∞ ′
Í
𝑗=0 𝑎 𝑗 (𝑡, 𝑧, 𝜉 ). We exploit (23.2.6) and its consequence: if the
neighborhoods 𝑈 of 0 ∈ R and Γ ∩ S𝑛−2 of 𝜉 ′◦′ are sufficiently small then there is
C 𝑛

a constant 𝐶 > 0 such that, for all 𝑗 ∈ Z, 𝛼 ∈ Z+𝑛 , 𝛽 ∈ Z+𝑛 ,

∀𝑡 ∈ Δ 𝜏 , ∀𝑧 ∈ 𝑈 C , ∀𝜉 ′ ∈ Γ, D𝑧𝛼 D 𝜉 ′ 𝑎 𝑗 (𝑡, 𝑧, 𝜉 ′) ≤ 𝐶 𝑗+| 𝛼 |+|𝛽 |+1 𝑗!𝛼!𝛽! |𝜉 ′ | − 𝑗−|𝛽 | .

𝛽

(24.6.6)
We apply Lemma 17.2.27: there is a sequence of Ehrenpreis cut-off functions 𝜒 𝑗 (𝑠)
( 𝑗 = 1, 2, ...) relative to the pair of half-lines {𝑠 ∈ R; 𝑠 ≥ 1}, {𝑠 ∈ R; 𝑠 ≥ 2} and a
number 𝜌 > 0 such that
∞
∑︁
𝑎♭ (𝑡, 𝑥, 𝜉 ′) = 𝑎 0 (𝑡, 𝑥, 𝜉 ′) + 𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) 𝑎 𝑗 (𝑡, 𝑥, 𝜉 ′) (24.6.7)
𝑗=1

is a bona fide pseudoanalytic symbol (Definition 17.1.4) in 𝑈 × Γ depending analyt-

ically on 𝑡 ∈ (−𝜏, 𝜏). It is convenient to assume 0 ≤ 𝜒 𝑗 ≤ 1 for all 𝑗. Keep in mind
that the coefficients 𝑎 𝑗 satisfy the Cauchy conditions (24.4.20); as a consequence,

𝑎♭ (𝑡, 𝑥, 𝜉 ′) ≡ 1. (24.6.8)
𝑥𝑛 =𝑡

24.6.2 Formal and approximate solution of an operator equation

We introduce a cut-off function in 𝜉-space, 𝑔 𝑅 (𝑅 > 0 large), of the type described

in Lemma 17.3.2, with the choice of convex cones Γ◦ , Γ♭ in R𝑛 \ {0}, with

(𝜉 ◦′, 0) ∈ Γ◦ ∩ S𝑛−1 ⊂⊂ Γ♭ ∩ S𝑛−1 ⊂⊂ Γ.

It might be helpful to recall the properties of 𝑔 𝑅 : first of all, 0 ≤ 𝑔 𝑅 (𝜉) ≤ 1 for all
𝜉 ∈ R𝑛 \ {0}; and for every 𝛼 ∈ Z+𝑛 , |𝜉 | ≥ 2𝑅 |𝛼| =⇒√|D 𝛼 𝑔 𝑅 (𝜉)| ≤ 2 (𝐶1 /𝑅) | 𝛼 | .
Moreover, 𝑔 𝑅 extends as a C ∞ function 𝑔 𝑅 (𝜁) in C𝑛 \ −1R𝑛 having the following
properties:
(1) 𝜉 ∈ Γ◦ =⇒ 𝑔 𝑅 (𝜁) = 1;
(2) 𝜉 ∉ Γ♭ =⇒ 𝑔 𝑅 (𝜁) = 0;
(3) |𝑔 𝑅 (𝜁)| ≤ exp (𝐶1 |𝜂| /𝑅);
(4) 𝜕 𝜁 𝑔 𝑅 (𝜁) ≤ 𝐶2 exp ((𝐶3 |𝜂| − |𝜉 |) /𝑅);

𝐶1 , 𝐶2 and 𝐶3 are positive constants independent of 𝑅 and, of course, of 𝜁 = 𝜉 +𝑖𝜂,

𝜉 ≠ 0.
We seek a solution 𝑨 to the following linear operators equation:

𝑃 (𝑥, D 𝑥 ) 𝑨 = 𝐸 (𝑥, D 𝑥 ) 𝑔 𝑅 (D 𝑥 ) (24.6.9)

1036 24 Analytic PDEs of Principal Type. Regularity of the Solutions

where 𝐸 (𝑥, D 𝑥 ) is the elliptic pseudodifferential operator in 𝑈 with symbol 𝐸 (𝑥, 𝜉)

[cf. (24.6.1)]. Actually we shall content ourselves with an approximate solution 𝑨♭ ,
in the sense that 𝑃 (𝑥, D 𝑥 ) 𝑨♭ − 𝐸 (𝑥, D 𝑥 ) 𝑔 𝑅 (D 𝑥 ) will be analytic regularizing in a
conic neighborhood of (0, 𝜉 ◦ ). We seek 𝑨♭ of the form
∫ 𝑥𝑛 ∫
′
𝑨♭ 𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 )+𝑖 𝜉𝑛 𝑡 𝑎♭ (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) b
𝑢 (𝜉) d𝜉d𝑡 (24.6.10)
−𝜏 R𝑛

[𝑢 ∈ Cc∞ (𝑈) and b𝑢 (𝜉) = e−𝑖 𝑥· 𝜉 𝑢 (𝑥) d𝑥 ]. We derive from (24.6.5)

∫
R𝑛

♭ 𝜕
𝑃 (𝑥, D 𝑥 ) 𝑨 𝑢 (𝑥) = 𝐸 (𝑥, D 𝑥 ) − 𝐵 (𝑥, D 𝑥′ ) + 𝐶 (𝑥, D 𝑥′ ) 𝑨♭ 𝑢 (𝑥) .
𝜕𝑥 𝑛

The conditions 𝜑| 𝑥𝑛 =𝑡 = 𝑥 ′ · 𝜉 ′ and (24.6.8) imply

(2𝜋) 𝑛 𝐸 (𝑥, D 𝑥 ) −1 𝑃 (𝑥, D 𝑥 ) 𝑨♭ 𝑢 (𝑥)

∫
= e𝑖 𝑥· 𝜉 𝑎♭ (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) b𝑢 (𝜉) d𝜉d𝑡
R𝑛
∫ 𝑥𝑛 ∫
𝑖 𝜉𝑛 𝑡 𝜕
+ e − 𝐵 (𝑥, D 𝑥′ ) + 𝐶 (𝑥, D 𝑥′ )
−𝜏 R𝑛 𝜕𝑥 𝑛
′

× e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑎♭ (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) b
𝑢 (𝜉) d𝜉d𝑡.

Formally, we have

𝜕 ′
− 𝐵 (𝑥, D 𝑥′ ) + 𝐶 (𝑥, D 𝑥′ ) e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝑎 (𝑡, 𝑥, 𝜉 ′) = 0, (24.6.11)
𝜕𝑥 𝑛

implying that we would get (24.6.10) if we substituted the formal series 𝑎 (𝑡, 𝑥, 𝜉 ′)
for the pseudoanalytic amplitude 𝑎♭ (𝑡, 𝑥, 𝜉 ′) (and 𝑨 for 𝑨♭ ). We begin by estimating
the error ensuing from the substitution of 𝑎♭ for 𝑎.

24.6.3 Estimate of the error resulting from the insertion of the cut-offs

Recall that the transport equations express the vanishing of the homogeneous terms
in the formal series equation [cf. (23.2.13)]
∞ 𝑚−1
∑︁ ∑︁ ∑︁ 1
|𝜉 ′ | 𝑘− 𝑗 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗 = 0,
𝛽
𝛼!
𝑗=0 𝑘=0 | 𝛼 | ≤𝑚

[𝑚: order of 𝑃 (𝑥, D 𝑥 )] which implies

24.6 Property (Q) Implies Analytic Hypoellipticity 1037

∞ 𝑚−1
∑︁ ∑︁ ∑︁ 1
𝜒 𝑗 (|𝜉 ′ |) |𝜉 ′ | 𝑘− 𝑗 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗
𝛽
𝛼!
𝑗=1 𝑘=0 | 𝛼 | ≤𝑚
∞ 𝑚−1
∑︁ ∑︁ ∑︁ 1
𝜒 𝑗 (|𝜉 ′ |) − 1 |𝜉 ′ | 𝑘− 𝑗 𝑝 𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗 .
𝛽
=
𝛼!
𝑗=1 𝑘=0 | 𝛼 | ≤𝑚

We look at the error terms of the same homogeneity degree (as |𝜉 ′ | −→ ∞) in the
right-hand side; by (24.6.3) they are

𝔢 𝑗 = 𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) 𝔏 (𝑧, D𝑧 ) 𝑎 𝑗 + 𝑐 0 𝑎 𝑗

min(∑︁
𝑗,𝑚)−1 ∑︁ 1
− 𝑐 𝑘, 𝛼 (𝑡, 𝑧, 𝜉 ′) 𝜒 𝑗−𝑘 (|𝜉 ′ | /( 𝑗 − 𝑘) 𝜌) D𝑧𝛼 𝑎 𝑗−𝑘
𝛼!
𝑘=1 | 𝛼 | ≤𝑚

( 𝑗 = 1, 2, ...) with

𝑐 𝑘, 𝛼 (𝑡, 𝑧, 𝜉 ′) = 𝐸 ◦ (𝑧, 𝜕𝑧 𝜑) −1 𝑝 𝑚−1−𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 .
𝛽

We return to (23.2.14)–(23.2.15) and to the “source equation” (23.2.7); we reach the

following conclusion, valid for all 𝑡 ∈ Δ 𝜏 , 𝑧 ∈ 𝑈 C , 𝜉 ′ ∈ Γ ′,
′
′
𝑃 (𝑧, D𝑧 ) e𝑖 𝜑 (𝑡 ,𝑧, 𝜉 ) 𝑎♭ (𝑡, 𝑧, 𝜉 ′) = e𝑖 𝜑 (𝑡 ,𝑧, 𝜉 ) 𝔢 (𝑡, 𝑧, 𝜉 ′) , (24.6.12)

where 𝑎♭ is defined in (24.6.7) and 𝔢 (𝑡, 𝑧, 𝜉 ′) = ∞ ′

Í
𝑗=1 𝔢 𝑗 (𝑡, 𝑧, 𝜉 ). The Cauchy con-
ditions (24.4.20) imply
𝔢 (𝑡, 𝑧, 𝜉 ′)| 𝑧𝑛 =𝑡 ≡ 0. (24.6.13)
Lemma 24.6.2 We have for suitably large 𝐶𝐾 > 0 and all 𝛼 ∈ Z+𝑛 , 𝑡 ∈ (−𝜏, 𝜏),
𝜉 ′ ∈ Γ ′, |𝜉 ′ | ≥ max (1, |𝛼|),
′
′ ′ 4
max D 𝑥𝛼 e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ) 𝔢 (𝑡, 𝑥, 𝜉 ′) ≤ 2𝑛 𝐶𝐾| 𝛼 |+1 𝜌 | 𝛼 | 𝛼!e− | 𝜉 |/2 𝜌 . (24.6.14)
𝑥 ∈𝐾

Proof By (23.2.20) and the Cauchy inequalities we derive that, for some constant
𝐶◦ > 0 and all 𝑘 < 𝑚, |𝛼| ≤ 𝑚,
𝛽′ |𝛽′ |+1 ′ ′
∀𝑡 ∈ Δ 𝜏 , ∀𝑧 ∈ 𝑈 C , ∀𝜉 ′ ∈ Γ, D 𝜉 ′ 𝑐 𝑘, 𝛼 (𝑡, 𝑧, 𝜉 ′) ≤ 𝐶◦ 𝛽 ! |𝜉 ′ | −𝑘− |𝛽 | . (24.6.15)

We derive from (24.6.4)

min(∑︁
𝑗,𝑚)−1 ∑︁ 1
𝑐 𝑘, 𝛼 𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) − 𝜒 𝑗−𝑘 (|𝜉 ′ | /( 𝑗 − 𝑘) 𝜌) D𝑧𝛼 𝑎 𝑗−𝑘 ,

𝔢𝑗 =
𝛼!
𝑘=1 | 𝛼 | ≤𝑚

and from (24.6.6)

∀𝑡 ∈ Δ 𝜏 , ∀𝑧 ∈ 𝑈 C , ∀𝜉 ′ ∈ Γ ′, D𝑧𝛼 𝑎 𝑗 (𝑡, 𝑧, 𝜉 ′) ≤ 𝐶 𝑗+| 𝛼 |+1 𝑗!𝛼! |𝜉 ′ | − 𝑗 .

1038 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Combining this with (24.6.15) yields

𝔢 𝑗 (𝑡, 𝑧, 𝜉 ′)
min(∑︁
𝑗,𝑚)−1
( 𝑗 − 𝑘)! 𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) − 𝜒 𝑗−𝑘 (|𝜉 ′ | /( 𝑗 − 𝑘) 𝜌) |𝜉 ′ | − 𝑗+𝑘
𝑚+1 𝑗−𝑘
≤ 𝐶1 𝐶 𝐶◦
𝑘=1

for some 𝐶1 > 0 depending only on 𝑚, and for all 𝑡 ∈ Δ 𝜏 , 𝑧 ∈ 𝑈 C , 𝜉 ′ ∈ Γ ′ . We know

that

𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) − 𝜒 𝑗−𝑘 (|𝜉 ′ | /( 𝑗 − 𝑘) 𝜌) ≠ 0 =⇒ ( 𝑗 − 𝑘) 𝜌 < |𝜉 ′ | < 2 𝑗 𝜌,

whence

𝔢 𝑗 (𝑡, 𝑧, 𝜉 ′)
min(∑︁
𝑗,𝑚)−1
≤ 𝐶1 𝐶 𝑚+1 (𝐶◦ /𝜌) 𝑗−𝑘 ( 𝑗 − 𝑘)! ( 𝑗 − 𝑘) − 𝑗+𝑘
𝑘=1
× 𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) − 𝜒 𝑗−𝑘 (|𝜉 ′ | /( 𝑗 − 𝑘) 𝜌) .

If we take 𝜌 > 2𝐶◦ the Sterling inequality implies

𝔢 𝑗 (𝑡, 𝑧, 𝜉 ′)
min(∑︁
𝑗,𝑚)−1
√︁
≤ 𝐶1 𝐶 𝑚+1
𝑗 − 𝑘2− 𝑗+𝑘 e 𝑘 e− 𝑗 𝜒 𝑗 (|𝜉 ′ | / 𝑗 𝜌) − 𝜒 𝑗−𝑘 (|𝜉 ′ | /( 𝑗 − 𝑘) 𝜌)
𝑘=1
min(∑︁
𝑗,𝑚)−1
!
√︁ ′ |/4𝜌 ′ |/4𝜌
≤ 𝐶1 𝐶 𝑚+1 𝑗 − 𝑘2− 𝑗+𝑘 e 𝑘 e− 𝑗/2 e−| 𝜉 ≤ 𝐶2 e− 𝑗/2 e− | 𝜉
𝑘=1

with 𝐶2 > 0 sufficiently large. It follows directly that

∞
∑︁ ′ |/4𝜌
C
∀𝑡 ∈ Δ 𝜏 , ∀𝑧 ∈ 𝑈 , ∀𝜉 ∈ Γ , ′ ′
𝔢 𝑗 (𝑡, 𝑧, 𝜉 ′) ≤ 2𝐶2 e− | 𝜉 , (24.6.16)
𝑗=1

and then, from the Cauchy inequalities, that if 𝐾 is a compact subset of 𝑈 C there is
a 𝐶𝐾 > 0 such that
′ |/4𝜌
∀𝛼 ∈ Z+𝑛 , ∀𝑡 ∈ Δ 𝜏 , ∀𝜉 ′ ∈ Γ ′, max 𝜕𝑧𝛼𝔢 (𝑡, 𝑧, 𝜉 ′) ≤ 𝐶𝐾| 𝛼 |+1 𝛼!e− | 𝜉 . (24.6.17)
𝑧 ∈𝐾

Proposition 24.4.1 enables us to apply Corollary 24.B.2 (Appendix B to the present

chapter) with 𝜓 (𝑥, 𝜉 ′) = 𝑖𝑥 ′ · 𝜉 ′ −𝑖𝜑 (𝑡, 𝑥, 𝜉 ′) in the place of 𝜙 and 𝜆 = |𝜉 ′ | (regarding
𝑡 and 𝜆−1 𝜉 ′ as parameters) and thus obtain the inequality
′) √ 1 ′
max 𝜕𝑥𝛼 e−𝜓 (𝑡 , 𝑥, 𝜉 ≤ 𝐶𝐾| 𝛼 | 𝛼! |𝜉 ′ | | 𝛼 |/2 e− 2 Re 𝜓 (𝑡 , 𝑥, 𝜉 ) if |𝜉 ′ | ≥ |𝛼| , (24.6.18)
𝑥 ∈𝐾
24.6 Property (Q) Implies Analytic Hypoellipticity 1039

where now 𝐾 ⊂ 𝑈. We apply the Leibniz rule taking (24.6.17) and (24.6.18) into
account; we get, for a suitably increased 𝐶𝐾 and all 𝛼 ∈ Z+𝑛 , 𝑡 ∈ (−𝜏, 𝜏), 𝜉 ′ ∈ Γ ′,
|𝜉 ′ | ≥ max (1, |𝛼|),
√ 1 ′ ′
max 𝜕𝑥𝛼 e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝔢 (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶𝐾| 𝛼 |+1 𝛼! |𝜉 ′ | 2 | 𝛼 | e− | 𝜉 |/4𝜌
𝑥 ∈𝐾

implying, for a further increased 𝐶𝐾 and the same 𝛼, 𝑡, 𝜉 ′,

′ ′
max 𝜕𝑥𝛼 e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝔢 (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶𝐾| 𝛼 |+1 𝜌 | 𝛼 |/2 𝛼!e− | 𝜉 |/8𝜌 . (24.6.19)
𝑥 ∈𝐾

We apply again the Leibniz rule, to derive from (24.6.19):

1 𝛼 𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ′ ) ∑︁ 1
′ |𝛽′ | 𝛼−(𝛽′ ,0)

−𝜓 (𝑡 , 𝑥, 𝜉 )

𝜕𝑥 e 𝔢 (𝑡, 𝑥, 𝜉 ′) ≤ |𝜉 | 𝜕 𝑥 e 𝔢 (𝑡, 𝑥, 𝜉 ′
)
𝛼! 𝛽⪯𝛼
𝛽 ′!
∑︁ 1 ′ | 𝛼′ −𝛽′ |+𝛼𝑛 +1 | 𝛼′ −𝛽′ |/2 − | 𝜉 ′ |/23 𝜌
≤ ′ ′ ′
|𝜉 ′ | |𝛽 | 𝐶𝐾 𝜌 e
𝛽⪯𝛼
𝛽 ! (𝛼 − 𝛽 )!𝛼𝑛 !
′
∑︁ |𝛽′ | ′ 4
≤ 𝐶𝐾| 𝛼 |+1 𝜌 | 𝛼 | 24 /𝐶𝐾 e−| 𝜉 |/2 𝜌 .
𝛽⪯𝛼

Requiring 𝐶𝐾 ≥ 25 leads to the desired conclusion. □

We define a pseudodifferential operator by the integral
∫ 𝑥𝑛 ∫
′
𝑹𝑢 (𝑥) = (2𝜋) −𝑛 e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 )+𝑖 𝜉𝑛 𝑡 𝔢 (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) b
𝑢 (𝜉) d𝜉d𝑡 (24.6.20)
−𝜏 R𝑛

[𝑢 ∈ Cc∞ (𝑈) and b𝑢 (𝜉) = e−𝑖 𝑥· 𝜉 𝑢 (𝑥) d𝑥 ].

∫
R𝑛

Proposition 24.6.3 Let 𝑈◦ ⊂⊂ 𝑈 be an open set; the restriction of 𝑹𝑢 to 𝑈◦ is

analytic, whatever 𝑢 ∈ Cc∞ (𝑈).

Proof Because of (24.6.13) we have

∫ 𝑥𝑛 ∫
𝜕 𝜕 𝑖 𝜑 (𝑡 , 𝑥, 𝜉 ′ )+𝑖 𝜉𝑛 𝑡
𝑹𝑢 (𝑥) = (2𝜋) −𝑛 e 𝔢 (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) b
𝑢 (𝜉) d𝜉d𝑡,
𝜕𝑥 𝑛 −𝜏 R𝑛 𝜕𝑥 𝑛

whence
∫ 𝑥𝑛 ∫
′
D 𝑥𝛼 𝑹𝑢 (𝑥) = (2𝜋) −𝑛 D 𝑥𝛼 e𝑖 𝜑 (𝑡 , 𝑥, 𝜉 )+𝑖 𝜉𝑛 𝑡 𝔢 (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) b
𝑢 (𝜉) d𝜉d𝑡.
−𝜏 R𝑛

′
We derive directly from (24.6.14), taking 𝐾 = 𝑈◦ and 𝐶 | 𝛼 |+1 ≥ 2𝑛+1 𝐶𝐾| 𝛼 |+1 𝜌 | 𝛼 | 𝜏,
∫
| 𝛼 |+1 ′
𝛼
D 𝑥 𝑹𝑢 (𝑥) ≤ 𝐶 𝛼! e− | 𝜉 |/16𝜌 𝑔 𝑅 (𝜉) |b
𝑢 (𝜉)| d𝜉.
R𝑛
1040 24 Analytic PDEs of Principal Type. Regularity of the Solutions

𝑢 ∈ 𝐿 1 (R𝑛 ) we obtain
Since b

max D 𝑥𝛼 𝑹𝑢 (𝑥) ≤ 𝐶 | 𝛼 |+1 𝛼! ∥b

𝑢 ∥ 𝐿1 . □
𝐾

Remark 24.6.4 In the statement and proof of Proposition 24.6.3 the cut-off 𝑔 𝑅 did
not need to satisfy all the conditions of Lemma 17.3.2; these will be needed below.

24.6.4 Analyticity off the diagonal of the distribution kernel associated

to the microlocal parametrix

We return to (24.6.10), more precisely to the distribution kernel

∫ ∫ 𝑥𝑛
′
𝐴♭ (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜓 (𝑡 , 𝑥, 𝜉 )+𝑖 𝜉𝑛 (𝑡−𝑥𝑛 ) 𝑎♭ (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) d𝜉d𝑡,
R𝑛 −𝜏
(24.6.21)
where we deform the contour of 𝜉-integration in (24.6.10)
from R 𝑛 to the image

Σ 𝜀 (𝑥, 𝑦) of R𝑛 under the map 𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜀 |𝜉 ′ | 𝑥 ′ − 𝑦 ′, (𝑥 𝑛 − 𝑦 𝑛 ) 𝑁 with
𝑥, 𝑦 ∈ 𝑈. We shall use the notation

Φ (𝑡, 𝑥, 𝑦, 𝜁) = (𝑥 ′ − 𝑦 ′) · 𝜁 + (𝑡 − 𝑦 𝑛 ) 𝜁 𝑛 + 𝑖𝜓 (𝑡, 𝑥, 𝜁 ′) .

We define the following extensions to the complex domain [cf. (24.6.7)]:

∞
∑︁
𝑎♭ (𝑡, 𝑥, 𝜁 ′) = 𝜒 𝑗 (|Re 𝜁 | / 𝑗 𝜌) 𝑎 𝑗 (𝑡, 𝑥, 𝜁 ′) . (24.6.22)
𝑗=1

It is always tacitly assumed that |𝑥 − 𝑦| and 𝜀 > 0 are sufficiently small that all the
extensions make sense. We apply Stokes’ Theorem:
∫ 𝑥𝑛 ∫
e𝑖Φ(𝑡 , 𝑥,𝑦, 𝜉 ) 𝑎♭ (𝑡, 𝑥, 𝜉 ′) 𝑔 𝑅 (𝜉) d𝜉d𝑡
−𝜏 R𝑛
∫ 𝑥𝑛 ∫
− e𝑖Φ(𝑡 , 𝑥,𝑦,𝜁 ) 𝑎♭ (𝑡, 𝑥, 𝜁 ′) 𝑔 𝑅 (𝜁) d𝜁d𝑡
−𝜏 Σ 𝜀 ( 𝑥,𝑦)
∫ 𝑥𝑛 ∫
= e𝑖Φ(𝑡 , 𝑥,𝑦,𝜁 ) 𝜕 𝜁 𝑔 𝑅 (𝜁) 𝑎♭ (𝑡, 𝑥, 𝜁 ′) ∧ d𝜁d𝑡,
−𝜏 Λ 𝜀 ( 𝑥,𝑦)
Ð
where Λ 𝜀 (𝑥, 𝑦) = 0<𝑐< 𝜀 Σ𝑐 (𝑥, 𝑦).
We need more detailed information on the phase-function Φ (𝑡, 𝑥, 𝑦, 𝜁) in the
submanifolds Σ𝑐 (𝑥, 𝑦), 0 ≤ 𝑐 ≤ 𝜀. We have

Im Φ (𝑡, 𝑥, 𝑦, 𝜁) = 𝑐 |𝜉 ′ | |𝑥 ′ − 𝑦 ′ | 2 + 𝑐 |𝜉 ′ | (𝑥 𝑛 − 𝑦 𝑛 ) 𝑁 (𝑡 − 𝑦 𝑛 ) + Re 𝜓 (𝑡, 𝑥, 𝜁 ′)
24.6 Property (Q) Implies Analytic Hypoellipticity 1041

and

𝜓 (𝑡, 𝑥, 𝜁 ′) = 𝜓 (𝑡, 𝑥, 𝜉 ′) + 𝑖𝑐 |𝜉 ′ | (𝑥 ′ − 𝑦 ′) · 𝜕 𝜉 ′ 𝜓 (𝑡, 𝑥, 𝜉 ′) + |𝜉 ′ | 𝑂 𝑐2 |𝑥 − 𝑦| 2 .

Once again we avail ourselves of Proposition 24.4.1:

Re 𝜓 (𝑡, 𝑥, 𝜁 ′) ≥ Re 𝜓 (𝑡, 𝑥, 𝜉 ′) − 𝑀1 𝑐2 |𝜉 ′ | |𝑥 ′ − 𝑦 ′ | 2
1 1 √︁
− 𝑀2 |𝑥 𝑛 − 𝑡| 2 𝑐 |𝜉 ′ | 2 |𝑥 ′ − 𝑦 ′ | Re 𝜓 (𝑡, 𝑥, 𝜉 ′),

whence
1
Re 𝜓 (𝑡, 𝑥, 𝜁 ′) ≥Re 𝜓 (𝑡, 𝑥, 𝜉 ′) − 𝑀3 𝑐2 |𝜉 ′ | |𝑥 ′ − 𝑦 ′ | 2 ,
2
where 𝑀1 , 𝑀2 , 𝑀3 are positive constants independent of all the variables and of 𝑐.
If 𝜀 > 0 is sufficiently small and 0 ≤ 𝑐 ≤ 𝜀, we get, in Σ𝑐 (𝑥, 𝑦),
1 1
Im Φ (𝑡, 𝑥, 𝑦, 𝜁) ≥ Re 𝜓 (𝑡, 𝑥, 𝜉 ′) + 𝑐 |𝜉 ′ | |𝑥 ′ − 𝑦 ′ | 2 + 𝑐 |𝜉 ′ | (𝑥 𝑛 − 𝑦 𝑛 ) 𝑁 (𝑡 − 𝑦 𝑛 ) .
2 2
We go back to (24.4.40), (24.4.41), and apply Corollary 24.A.2 to get, for some
𝑏 > 0, in Σ𝑐 (𝑥, 𝑦),
1
Re 𝜓 (𝑡, 𝑥, 𝜉 ′) ≥ 𝑏 |𝜉 ′ | (𝑥 𝑛 − 𝑡) 2𝑘+1 .
2
We derive from this and the preceding inequality
1
|𝜉 ′ | −1 Im Φ (𝑡, 𝑥, 𝑦, 𝜁) ≥ 𝑏 (𝑥 𝑛 − 𝑡) 2𝑘+1 + 𝑐 |𝑥 ′ − 𝑦 ′ | 2
2
− 𝑐 (𝑥 𝑛 − 𝑦 𝑛 ) 2𝑘+1 (𝑥 𝑛 − 𝑡) + 𝑐 (𝑥 𝑛 − 𝑦 𝑛 ) 2𝑘+2 .

Since
1
1
𝑐 (𝑥 𝑛 − 𝑦 𝑛 ) 2𝑘+1 (𝑥 𝑛 − 𝑡) ≤ 2𝑐 2𝑘+1 /𝑏 𝑐 (𝑥 𝑛 − 𝑦 𝑛 ) 2𝑘+2 + 𝑏 (𝑥 𝑛 − 𝑡) 2𝑘+2
2

it suffices to require 𝑐 ≤ (𝑏/4) 2𝑘+1 to obtain

2 |𝜉 ′ | −1 Im Φ (𝑡, 𝑥, 𝑦, 𝜁) ≥ 𝑏 (𝑥 𝑛 − 𝑡) 2𝑘+1 + 𝑐 |𝑥 ′ − 𝑦 ′ | 2 + 𝑐 (𝑥 𝑛 − 𝑦 𝑛 ) 2𝑘+2 .

Here (and in the sequel) we take advantage of the fact that in Γ ⊃ supp 𝑔 𝑅 we
have 𝜉 𝑛 < 𝜀◦ |𝜉 ′ | hence |𝜉 | ≤ (1 + 𝜀◦ ) |𝜉 ′ |, whence

4 |𝜉 | −1 Im Φ (𝑡, 𝑥, 𝑦, 𝜁) ≥ 𝑏 (𝑥 𝑛 − 𝑡) 2𝑘+1 + 𝑐 |𝑥 ′ − 𝑦 ′ | 2 + 𝑐 (𝑥 𝑛 − 𝑦 𝑛 ) 2𝑘+2 (24.6.23)

for all (𝑥, 𝑦) ∈ 𝑈 × 𝑈, −𝜏 ≤ 𝑡 ≤ 𝑥 𝑛 , 𝜁 ∈ Σ𝑐 (𝑥, 𝑦), 𝜉 = Re 𝜁 ∈ Γ.

1042 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Next we take a closer look at the amplitudes. By (24.6.22)

∞
1 ∑︁
𝜕 𝜁 𝑎♭ (𝑡, 𝑥, 𝜁 ′) = 𝑎 𝑗 (𝑡, 𝑥, 𝜁 ′) 𝜕 𝜉 𝜒 𝑗 (|𝜉 | / 𝑗 𝜌) , 𝜉 = Re 𝜁.

2 𝑗=1

Recall that 𝜒 𝑗 (𝑠) ( 𝑗 = 1, 2, ...) is a sequence of Ehrenpreis cut-off functions

relative to the pair of half-lines {𝑠 ∈ R; 𝑠 ≥ 1}, {𝑠 ∈ R; 𝑠 ≥ 2}; in particular,
𝜕 𝜉 𝜒 𝑗 (|𝜉 | / 𝑗 𝜌) = 0 if |𝜉 | ≤ 𝑗 𝜌 or if |𝜉 | ≥ 2 𝑗 𝜌. A simpler version of the ar-
gument leading to (24.6.19) applies here and enables us to prove, for all 𝜉 ∈ Γ,
′
4
max e−𝜓 (𝑡 , 𝑥, 𝜉 ) 𝜕 𝜁 𝑎♭ (𝑡, 𝑥, 𝜉 ′) ≤ 𝐶𝐾 e− | 𝜉 |/2 𝜌 (24.6.24)
𝑥 ∈𝐾

with 𝐾 a compact subset of 𝑈 as in (24.6.17) and (24.6.19).

With this established, if 𝜁 = 𝜉 + 𝑖𝜀 |𝜉 | (𝑥 − 𝑦) with 𝜀 |𝑥 − 𝑦| sufficiently small it
is not difficult to deduce from (24.6.23) and (24.6.24) the following estimate, for an
increased 𝐶𝐾 and all 𝜁 ∈ Σ𝑐 (𝑥, 𝑦), 0 ≤ 𝑐 ≤ 𝜀, such that 𝜉 = Re 𝜁 ∈ Γ:
′
5
e𝑖 ( 𝑥−𝑦) ·𝜁 −𝜓 (𝑡 , 𝑥,𝜁 ) 𝜕 𝜁 𝑎♭ (𝑡, 𝑥, 𝜁 ′) ≤ 𝐶𝐾 e−| 𝜉 |/2 𝜌 . (24.6.25)

We also have, by (17.3.3),

|𝑔 𝑅 (𝜁)| ≤ exp (𝜀𝐶1 |𝑥 − 𝑦| |𝜉 | /𝑅) . (24.6.26)

Combining (24.6.25) and (24.6.26) yields the following result: if 𝜀 and 𝜀/𝑅 are
sufficiently small the integral
∫ 𝑥𝑛 ∫
𝐽1 (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖Φ(𝑡 , 𝑥,𝑦,𝜁 ) 𝑔 𝑅 (𝜁) 𝜕 𝜁 𝑎♭ (𝑡, 𝑥, 𝜁 ′)∧d𝜁d𝑡 (24.6.27)
−𝜏 Λ 𝜀 ( 𝑥,𝑦)

can be extended as a holomorphic function of (𝑥, 𝑦) in an open ball centered at

(0, 0) in C2𝑛 . In this extension the integral with respect to 𝑡 can be taken to be the
straight-line segment joining −𝜏 to 𝑥 𝑛 in the complex plane; the same applies to the
forthcoming holomorphic extensions.
Next we look at the integral
∫ 𝑥𝑛 ∫
𝐽2 (𝑥, 𝑦) = (2𝜋) −𝑛 e𝑖Φ(𝑡 , 𝑥,𝑦,𝜁 ) 𝑎♭ (𝑡, 𝑥, 𝜁 ′) b
𝑢 (𝜁, 𝑡) 𝜕 𝜁 𝑔 𝑅 (𝜁) ∧ d𝜁d𝑡,
−𝜏 Λ 𝜀 ( 𝑥,𝑦)
(24.6.28)
where 𝜁 = 𝜉 + 𝑖𝑐 |𝜉 ′ | (𝑥 − 𝑦). Here we avail ourselves of (17.3.4):

𝜕 𝜁 𝑔 𝑅 (𝜉 + 𝑖𝑐 |𝜉 ′ | (𝑥 − 𝑦)) ≤ 𝐶2 exp (− (1 − 𝐶3 𝜀 |𝑥 − 𝑦|) |𝜉 ′ | /𝑅) .

If we select 𝜀 > 0 sufficiently small we get

𝜕 𝜁 𝑔 𝑅 (𝜉 + 𝑖𝑐 |𝜉 | (𝑥 − 𝑦)) ≤ 𝐶2 exp (− |𝜉 | /2𝑅) (24.6.29)

24.6 Property (Q) Implies Analytic Hypoellipticity 1043

for all 𝑥, 𝑦 ∈ 𝑈, 𝜉 ∈ Γ. We conclude easily that if 𝜀 is sufficiently small the integral

𝐽2 (𝑥, 𝑦) can be extended as a holomorphic function of (𝑥, 𝑦) in an open ball centered
at (0, 0) in C2𝑛 .
We have established that mod functions of (𝑥, 𝑦) holomorphically extendible to
a neighborhood of (0, 0) in C2𝑛 the distribution kernel (24.6.21) is congruent to
∫ 𝑥𝑛 ∫
D𝜁
(2𝜋) −𝑛 e𝑖Φ(𝑡 , 𝑥,𝑦,𝜁 ) 𝑎♭ (𝑡, 𝑥, 𝜁 ′) 𝑔 𝑅 (𝜁) d𝜉d𝑡, (24.6.30)
−𝜏 R𝑛 D𝜉

where 𝜁 = 𝜉 + 𝑖𝜀 |𝜉 | (𝑥 − 𝑦). The inequality (24.6.23) implies that if (𝑥 ∗ , 𝑦 ∗ ) ∈ 𝑈 ×𝑈

is sufficiently close to the origin but 𝑥 ∗ ≠ 𝑦 ∗ the function (24.6.30) can be extended
as a holomorphic function of (𝑥, 𝑦) in a neighborhood of (𝑥 ∗ , 𝑦 ∗ ) in C2𝑛 . The same
is therefore true for (24.6.21).

24.6.5 End of the proof of the main theorem: (Q) implies (b)

We have proved that if 𝑈 is a suitably small neighborhood of 𝑥 ◦ (= 0 in the proof)

then, for every 𝑢 ∈ Cc∞ (𝑈),
∫ ∫
𝑃 (𝑥, D 𝑥 ) 𝐴♭ (𝑥, 𝑦) 𝑢 (𝑦) d𝑦 = 𝐸 (𝑥, D 𝑥 ) 𝑔 𝑅 (D 𝑥 ) 𝑢 (𝑥) + 𝐹 (𝑥, 𝑦) 𝑢 (𝑦) d𝑦
(24.6.31)
and 𝐴♭ (𝑥, 𝑦) is C 𝜔 in (𝑈 × 𝑈) \ diag (𝑈 × 𝑈) whereas 𝐹 ∈ C 𝜔 (𝑈 × 𝑈). We point
out that 𝐴♭ (𝑥, 𝑦) is semiregular (Definition 2.3.1) as readily proved, or as a conse-
quence of the fact that 𝑃 (𝑥, D 𝑥 ) and its transpose 𝑃 (𝑥, D 𝑥 ) ⊤ are hypoelliptic [since
(Q)=⇒(c)=⇒(a) has been proved].
At this point there are two ways of completing the proof, one local and one
microlocal.
Local approach.
We select a finite set of functions of the type 𝑔 𝑅 , 𝑔 𝑅[𝑖 ] (D 𝑥 ), 𝑖 = 1, ..., 𝑟, such
that 𝑟𝑖=1 𝑔 𝑅[𝑖 ] (𝜉) ≥ 1 if 𝜉 ∈ S𝑛−1 and that the analogue of (24.6.31) can be proved,
Í

naturally with appropriate distribution kernels 𝐴♭[𝑖 ] , 𝐹 [𝑖 ] and operators 𝐸 [𝑖 ] , in the

intersection of the domains 𝑈 [𝑖 ] where the desired properties are valid. We have seen
that this is possible in a conic neighborhood of an arbitrary point (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃;
it is even simpler to do if (𝑥 ◦ , 𝜉 ◦ ) ∉ Char 𝑃, i.e., in the “elliptic” region. Summing
the formulas (24.6.31) for 𝑖 = 1, ..., 𝑟 leads to a local formula
∫ ∫
♭
𝑃 (𝑥, D 𝑥 ) 𝐴 (𝑥, 𝑦) 𝑢 (𝑦) d𝑦 = 𝐸 (𝑥, D 𝑥 ) 𝑢 (𝑥) + 𝐹 (𝑥, 𝑦) 𝑢 (𝑦) d𝑦 (24.6.32)

with 𝐸 (𝑥, D 𝑥 ) elliptic in 𝑈. Replacing 𝑢 (𝑥) by 𝜒 (𝑥) 𝐸 (𝑥, D 𝑥 ) −1 𝑢 (𝑥), 𝜒 ∈ Cc∞ (𝑈),
𝜒 ≡ 1 in a neighborhood 𝑈◦ of supp 𝑢, yields a “true” parametrix in 𝑈◦ ,
1044 24 Analytic PDEs of Principal Type. Regularity of the Solutions
∫
−1
𝑮𝑢 (𝑥) = 𝐴♭ (𝑥, 𝑦) 𝜒 (𝑦) 𝐸 𝑦, D 𝑦 𝑢 (𝑦) d𝑦. (24.6.33)

Theorem 3.1.4, then allows us to conclude that 𝑃 (𝑥, D 𝑥 ) ⊤ is analytic hypoelliptic.

Since (Q) is invariant under transposition this completes the proof of Theorem 24.2.1
Microlocal approach.
Here we reason in the conic neighborhood 𝑈◦ × Γ◦ of (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃; recall
that 𝑔 𝑅 (𝜁) ≡ 1 if 𝜉 = Re 𝜁 ∈ Γ◦ . We regard (24.6.31) as defining a (microlocal)
parametrix 𝑮 ◦ of 𝑃 (𝑥, D 𝑥 ) 𝐸 (𝑥, D 𝑥 ) −1 in 𝑈◦ × Γ◦ such that

(𝑈◦ × Γ◦ ) ∩ 𝑊 𝐹a 𝑮 ⊤

◦ 𝑢 ⊂ (𝑈◦ × Γ◦ ) ∩ 𝑊 𝐹a 𝑢. (24.6.34)

By transposition we get that

𝑢 − 𝑃 (𝑥, D 𝑥 ) 𝐸 (𝑥, D 𝑥 ) −1 𝑮 ⊤
◦𝑢

is microanalytic (cf. Definition 3.5.1) in 𝑈◦ × Γ◦ . By transposition we get

⊤
⊤ −1 ⊤
(𝑈◦ × (−Γ◦ )) ∩ 𝑊 𝐹a 𝑮 ◦ 𝐸 (𝑥, D 𝑥 ) 𝑃 (𝑥, D 𝑥 ) 𝑢 − 𝑢 = ∅.

Combining this with (24.6.34) and taking the ellipticity of 𝐸 (𝑥, D 𝑥 ) ⊤ in 𝑈◦ × (−Γ◦ )
into account yields

𝑊 𝐹a 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑢 ∩ (𝑈◦ × (−Γ◦ )) = 𝑊 𝐹a 𝑢 ∩ (𝑈◦ × (−Γ◦ )) ,

(24.6.35)

which expresses the microlocal analytic-hypoellipticity of 𝑃 (𝑥, D 𝑥 ) ⊤ at (𝑥 ◦ , −𝜉 ◦ ).

Exchange of 𝑃 (𝑥, D 𝑥 ) and 𝑃 (𝑥, D 𝑥 ) ⊤ proves the microlocal analytic-hypoellipticity
of 𝑃 (𝑥, D 𝑥 ) at (𝑥 ◦ , 𝜉 ◦ ), i.e., the microlocal version of the entailment (Q)=⇒(b).

24.7 The C ∞ Situation

The equivalence hypoellipticity ≈ subellipticity is not true when the coefficients of

𝑃 (𝑥, D 𝑥 ) are smooth but not analytic, as we show now.
∫𝑡
Proposition 24.7.1 Suppose 𝑏 ∈ C ∞ (R) and 𝑡 2 𝑏 (𝑡) d𝑡 > 0 for every pair of real
1
numbers 𝑡1 < 𝑡2 . The operator in R2 ,
𝜕 𝜕
𝐿= − 𝑖𝑏 (𝑡) , (24.7.1)
𝜕𝑡 𝜕𝑥
is hypoelliptic.

Proof Let 𝑢 ∈ E ′ R2 be such that 𝐿𝑢 = 𝑓 ∈ Cc∞ R2 and, after Fourier transform

with respect to 𝑥,
24.7 The C ∞ Situation 1045

𝜕b𝑢
− 𝑖𝑏 (𝑡) 𝜉b 𝑢= b 𝑓.
𝜕𝑡
A solution of 𝐿𝑢 = 𝑓 has the expression
∫ 𝑡 ∫ 𝑦=+∞ ∫ +∞
1 ∫𝑡
𝑢 (𝑥, 𝑡) = e𝑖 𝜉 ( 𝑥−𝑦)− 𝜉 𝑡 ′ 𝑏 (𝑠)d𝑠 𝑓 (𝑦, 𝑡 ′) d𝜉d𝑦d𝑡 ′
2𝜋 −∞ 𝑦=−∞ 𝜉 =0
∫ +∞ ∫ 𝑦=+∞ ∫ +∞ ∫ 𝑡′
1
− e−𝑖 𝜉 ( 𝑥−𝑦)− 𝜉 𝑡 𝑏 (𝑠)d𝑠 𝑓 (𝑦, 𝑡 ′) d𝜉d𝑦d𝑡 ′.
2𝜋 𝑡 𝑦=−∞ 𝜉 =0

This means that

∫ +∞
1
∫𝑡
𝑖 𝜉 ( 𝑥−𝑦)− 𝜉 𝑏 (𝑠)d𝑠
𝐺 (𝑥, 𝑡, 𝑦, 𝑡 ′) = e 𝑡′
d𝜉 (24.7.2)
2𝜋 0

is a fundamental solution (Definition 2.4.1) of 𝐿. Note that

𝜕2

𝐺 (𝑥, 𝑡, 𝑦, 𝑡 ) = 1 − 2 𝐺 ◦ (𝑥, 𝑡, 𝑦, 𝑡 ′) ,
′
𝜕𝑥
∫ +∞
1 𝑖 𝜉 ( 𝑥−𝑦)− 𝜉 𝑡 ′ 𝑏 (𝑠)d𝑠 d𝜉
∫𝑡
𝐺 ◦ (𝑥, 𝑡, 𝑦, 𝑡 ′) = e ,
2𝜋 0 1 + 𝜉2

are well-defined semiregular distribution kernels in R2 × R2 (Definition 2.3.1). In

the oscillatory integral (24.7.2) we deform the domain of integration from R+ to the
image of R+ under the map

𝜉 ↦→ 𝜁 = 𝜉 + 𝑖𝜉 (𝑥 − 𝑦) .

The Cauchy Integral Theorem implies

∫ +∞
1
∫𝑡 ∫𝑡
′ 𝑖 𝜉 𝑥−𝑦− 𝑡 ′ 𝑏 (𝑠)d𝑠 − 𝜉 | 𝑥−𝑦 | 2 + 𝑡 ′ 𝑏 (𝑠)d𝑠
𝐺 (𝑥, 𝑡, 𝑦, 𝑡 ) = e d𝜉.
2𝜋 0

Since (𝑥, 𝑡) ≠ (𝑦, 𝑡 ′) entails

∫ 𝑡
|𝑥 − 𝑦| 2 + 𝑏 (𝑠) d𝑠 ≠ 0
𝑡′

we derive immediately 𝐺 (𝑥, 𝑡, 𝑦, 𝑡 ′) is a C ∞ function in the complement of the

diagonal in R2 × R2 . □
A minor modification of the argument developed in this chapter to prove that
(a)⇐⇒(Q) (Theorem 24.2.1) yields a proof of the following result:
1046 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Theorem 24.7.2 Let 𝑃 (𝑥, D) be a differential operator of principal type with C ∞

coefficients in Ω. The following two properties are equivalent:
(a) 𝑃 (𝑥, D) is hypoelliptic;
(PQ) 𝑃 (𝑥, D) satisfies Condition (P) and there are no null bicharacteristic curves
of 𝑃 in 𝑇 ∗ Ω\0.
We emphasize the fact that, in Condition (PQ), by “null bicharacteristic curves”
we mean immersed one-dimensional C ∞ submanifolds of 𝑇 ∗ Ω\0 contained in Char 𝑃
to which the Hamiltonian fields 𝐻Re 𝑃 and 𝐻Im 𝑃 are tangent at every point (cf.
Definition 23.1.15).
Back to Example (24.7.1) we have the following statement:
Proposition 24.7.3 Suppose 𝑏 ∈ C ∞ (R) and 𝑏 (𝑡) > 0 if 𝑡 ≠ 𝑡 ◦ ∈ R. If 𝑏 (𝑡)
vanishes to infinite order at 𝑡 ◦ then the operator (24.7.1) in R2 is not subelliptic in
any rectangle (𝑥 ◦ − 𝑟 1 , 𝑥 ◦ + 𝑟 2 ) × (𝑡 ◦ − 𝑇1 , 𝑡 ◦ + 𝑇2 ) (𝑟 𝑗 , 𝑇 𝑗 positive numbers).
∫𝑡 ∫
Proof We take 𝑥 ◦ = 𝑡 ◦ = 0. Let 𝐵 (𝑡) = 0 𝑏 (𝑠) d𝑠 and b 𝑢 (𝜉) = e−𝑖 𝑥 𝜉 𝑢 (𝑥) d𝑥,
𝑢 ∈ Cc∞ (R), supp 𝑢 ⊂ (−1, 1). Then, for every 𝜆 ≥ 1,
2 3
𝑢 (𝜉/𝜆) e 𝐵(𝑡) 𝜉 − 𝜉 /𝜆
ℎˆ (𝜉, 𝑡, 𝜆) = b
𝜕 ℎˆ
is a smooth solution of + 𝑏 (𝑡) 𝜉 ℎˆ = 0 in R2 rapidly decaying at infinity. We write
𝜕𝑡
∫
1
ℎ (𝑥, 𝑡, 𝜆) = e−𝑖 𝑥 𝜉 ℎˆ (𝜉, 𝑡, 𝜆) d𝜉.
2𝜋
We introduce two cut-off functions 𝜒1 ∈ Cc∞ (−𝑟 1 , 𝑟 2 ), 0 ≤ 𝜒1 ≤ 1, 𝜒2 ∈
Cc∞ (−𝑇1 , 𝑇2 ). Suppose we had, for some 𝑠 > 0,

∥ 𝜒1 (𝑥) 𝜒2 (𝑡) ℎ∥ 2𝑠 ≤ 𝐶 ∥𝐿 ( 𝜒1 (𝑥) 𝜒2 (𝑡) ℎ)∥ 20 (24.7.3)

2
≤ 𝐶 ′ ∥ 𝜒1 (𝑥) 𝐿 ( 𝜒2 (𝑡) ℎ)∥ 20 + 𝐶 ′ 𝜒1′ (𝑥) 𝜒2 (𝑡) 𝑏 (𝑡) ℎ 0
,

where ∥·∥ 𝑠 is the norm in the Sobolev space 𝐻 𝑠 R2 and the constants are indepen-

dent of ℎ, 𝜒1 , 𝜒2 . We avail ourselves of the fact that 𝐿 is invariant under translations
in 𝑥; this means that we can select 𝜒1 and a sequence of its translates 𝜒1 𝑥 − 𝑥 𝑗
such that, for some 𝑀 > 0 and every 𝑥 ∈ R,
∞
∑︁ ∞
∑︁ 2
𝜒1′ 𝑥 − 𝑥 𝑗

𝜒1 𝑥 − 𝑥 𝑗 = 1, ≤ 𝑀.
𝑗=0 𝑗=0

We derive from (24.7.3):

2
∥ 𝜒2 (𝑡) ℎ∥ 2𝑠 ≤ 𝐶 ′′ 𝜒2′ (𝑡) ℎ 0
+ 𝐶 ′ 𝑀 ∥ 𝜒2 (𝑡) 𝑏 (𝑡) ℎ∥ 20 . (24.7.4)

Actually we replace 𝜒2 (𝑡) by 𝜒2 (𝜆 𝜅 𝑡) with 𝜅 > 0 to be selected below; the support

of 𝑡 ↦→ 𝜒2 (𝜆 𝜅 𝑡) is contained in (−𝜆−𝜅 𝑇1 , 𝜆−𝜅 𝑇2 ) ⊂⊂ (−𝑇1 , 𝑇2 ); then (24.7.4) entails
24.7 The C ∞ Situation 1047
∫ ∫ ∫ ∫
2 2
𝜉 2𝑠 ℎˆ (𝜉, 𝑡) 𝜒2 (𝜆 𝜅 𝑡) d𝜉d𝑡 ≤ 𝐶 ′′𝜆2𝜅 ℎˆ (𝜉, 𝑡) 𝜒 ′2 (𝜆 𝜅 𝑡) d𝜉d𝑡
∫ ∫
′ 2
+𝐶 𝑀 ℎˆ (𝜉, 𝑡) 𝜒2 (𝜆 𝜅 𝑡) d𝜉d𝑡,

i.e.,
∫ ∫
2 3
𝑢 (𝜉/𝜆)| 2 e2𝐵(𝑡) 𝜉 −2 𝜉 /𝜆 𝜒2 (𝜆 𝜅 𝑡) d𝜉d𝑡
𝜉 2𝑠 |b
∫ ∫
−𝜅 2
= 𝜆1+2𝑠−𝜅 𝑢 (𝜉)| 2 e2𝜆𝐵(𝜆 𝑡) 𝜉 −2 𝜉 /𝜆 𝜒2 (𝑡) d𝜉d𝑡
𝜉 2𝑠 |b
∫ ∫
−𝜅 2
≤𝐶 𝜆′′ 1+𝜅
𝑢 (𝜉)| 2 e2𝜆𝐵(𝜆 𝑡) 𝜉 −2 𝜉 /𝜆 𝜒 ′2 (𝑡) d𝜉𝑑𝑡d𝜉
|b
∫ ∫
−𝜅 2
+ 𝐶 ′ 𝑀𝜆1−𝜅 𝑢 (𝜉)| 2 e2𝜆𝐵(𝜆 𝑡) 𝜉 −2 𝜉 /𝜆 𝜒2 (𝑡) d𝜉d𝑡.
|b

We select 𝜅 > 0 such that 𝑠 > 2𝜅, whence

∫ ∫ 𝑇2
−𝜅 2
𝜆 𝑠
𝑢 (𝜉)| 2 e2𝜆𝐵(𝜆 𝑡) 𝜉 − 𝜉 /𝜆 𝜒2 (𝑡) d𝜉d𝑡
𝜉 2𝑠 |b
−𝑇1
∫ ∫ 𝑇2
−𝜅 2
≤ 𝐶 ′′ 𝑢 (𝜉)| 2 e2𝜆𝐵(𝜆 𝑡) 𝜉 −2 𝜉 /𝜆 𝜒 ′2 (𝑡) d𝜉𝑑𝑡d𝜉
|b
−𝑇1
∫ ∫
−𝜅 2
+ 𝐶 ′ 𝑀𝜆−2𝜅 𝑢 (𝜉)| 2 e2𝜆𝐵(𝜆 𝑡) 𝜉 −2 𝜉 /𝜆 𝜒2 (𝑡) d𝜉d𝑡.
|b

Since to every 𝜅 > 0 there is a 𝐶 𝜅 > 0 such that 𝜆𝐵 (𝜆−𝜅 𝑡) ≤ 𝐶 𝜅 𝜆−1 for every 𝑡 ∈
[−𝑇1 , 𝑇2 ] the left-hand side tends to +∞ as 𝜆 ↗ +∞ whereas the right-hand side
converges to
∫ ∫ 𝑇2
2
𝐶 𝑢 (𝜉)| 2 𝜒 ′2 (𝑡) d𝜉d𝑡d𝜉 = 𝐶 ∥𝑢∥ 2𝐿 2 𝜒 ′2
|b 𝐿2
,
−𝑇1

which proves the impossibility of (24.7.4). □

About subellipticity in the C ∞ set-up the following can be stated:

Theorem 24.7.4 For a differential operator 𝑃 (𝑥, D) of principal type with C ∞ co-
efficients in Ω to be subelliptic it is necessary and sufficient that 𝑃 (𝑥, D) satisfies
Condition (P) and that the pair of functions (Re 𝑃𝑚 , Im 𝑃𝑚 ) is of finite (necessarily
even) type (Definition 13.3.28) at every point of Char 𝑃.

The proof of Theorem 24.7.4 (see [Treves, 1970, 2]) is essentially the same as that
of the equivalence of (b)⇐⇒(Q) in Theorem 24.2.1; the only important difference
is that the use of the Weierstrass Preparation Theorem [cf. Proposition 23.1.10
and (24.4.40)] must be replaced by that of the Weierstrass–Malgrange Preparation
Theorem (for a simple proof of the latter see [Nirenberg, 1971]).
1048 24 Analytic PDEs of Principal Type. Regularity of the Solutions

24.8 Propagation of Analytic Singularities

24.8.1 Preliminaries

We continue to study a differential operator 𝑃 with complex C 𝜔 coefficients in a

C 𝜔 manifold M. The main question addressed in this section is the following: Let
𝑢 ∈ D ′ (M) be such that 𝑃𝑢 ∈ C 𝜔 (M) and let L be a connected C 𝜔 submanifold
of 𝑇 ∗ M\0. Suppose that (𝑥 ◦ , 𝜉 ◦ ) ∈ L ∩ 𝑊 𝐹a (𝑢). For which submanifolds L does
it follow that L ⊂ 𝑊 𝐹a (𝑢)? By the Sato Theorem 7.4.19 we know that 𝑊 𝐹a (𝑢) ⊂
Char 𝑃. We shall always reason under the hypothesis that 𝑃 is of principal type. By
Theorem 24.2.1 we know that (𝑥 ◦ , 𝜉 ◦ ) must lie in a null bicharacteristic leaf of 𝑃
of positive dimension in 𝑇 ∗ M\0. Under Condition (P) the dimension of such a leaf
cannot exceed 2 (Proposition 23.1.43).
In the matter of propagation of analytic singularities local implies global:

Lemma 24.8.1 Let 𝑢 ∈ D ′ (M) be such that 𝑃𝑢 ∈ C 𝜔 (M) and let L be a connected
C 𝜔 submanifold of 𝑇 ∗ M\0. The following two properties are equivalent:
(i) L ⊂ 𝑊 𝐹a (𝑢);
(ii) Every point (𝑥 ◦ , 𝜃 ◦ ) ∈ L ∩ 𝑊 𝐹a (𝑢) has a neighborhood U in 𝑇 ∗ M such that
U ∩ L ⊂ 𝑊 𝐹a (𝑢).

Proof That (i)=⇒(ii) is trivial. If (ii) holds then L ∩ 𝑊 𝐹a (𝑢) is open in L; since
𝑊 𝐹a (𝑢) is closed in 𝑇 ∗ M\0 we see that L ∩ 𝑊 𝐹a (𝑢) is closed in L. Since L is
connected necessarily L ∩ 𝑊 𝐹a (𝑢) = L. □
We limit ourselves to study a differential operator 𝑃 = 𝑃 (𝑥, D) of order 𝑚 ≥ 1
and principal type in an open subset Ω ∋ 𝑥 ◦ of R𝑛 , with coefficients in C 𝜔 (Ω).
Suppose 𝑃𝑢 ∈ C 𝜔 (Ω). Using the Cauchy–Kovalevskaya Theorem 5.2.4, we can find
a solution ℎ ∈ C 𝜔 (𝑈) of the equation 𝑃ℎ = 𝑃𝑢 in a suitably small neighborhood 𝑈
of 𝑥 ◦ ; thus, replacing 𝑢 by 𝑢 − ℎ reduces the question to distribution solutions of the
homogeneous equation 𝑃𝑢 = 0.
In order to prove the results in the next two sections we are going to apply Theorem
22.3.15; for this we shall make use of an FBI phase-function 𝜑 (Definition 22.1.1). We
also introduce a classical symbol (Definition 19.1.14) 𝑎 (𝑥, 𝜃, 𝜆) = +∞ −𝑗
Í
𝑗=0 𝜆 𝑎 𝑗 (𝑥, 𝜃)
with 𝑎 𝑗 (𝑥, 𝜃) ∈ C (Ω × Θ) homogeneous of degree 𝑚− 𝑗; Θ is a domain in C𝑛 \ {0}.
𝜔

We assume that 𝑎 is elliptic, i.e., 𝑎 0 nowhere vanishes in Ω × Θ (cf. Proposition

17.2.30). To make the objects we are handling more concrete we introduce (cf.
Lemma 17.2.27) a sequence of Ehrenpreis cut-off functions 𝜒 𝑗 (𝑡) ( 𝑗 = 1, 2, ...,
0 ≤ 𝜒 𝑗 ≤ 1 for all 𝑗) relative to the pair of half-lines {𝑡 ∈ R; 𝑡 ≥ 1}, {𝑡 ∈ R; 𝑡 ≥ 2}
and a number 𝜌 > 0 such that
∞
∑︁
𝑎♭ (𝑥, 𝜃, 𝜆) = 𝑎 0 (𝑥, 𝜃) + 𝜒 𝑗 (𝜆/ 𝑗 𝜌) 𝜆− 𝑗 𝑎 𝑗 (𝑥, 𝜃) (24.8.1)
𝑗=1
24.8 Propagation of Analytic Singularities 1049

is a bona fide pseudoanalytic symbol (Definition 17.1.4) in Ω × Γ (Θ) if Γ (Θ) is the

cone spanned by Θ and 𝜆 = |𝜃|. We then define the FBI transform [cf. (22.2.8) and
Definition 22.3.14]
∫
𝑨𝑢 (𝜃, 𝜆) = e𝑖𝜆𝜑 ( 𝑥, 𝜃) 𝑎♭ (𝑥, 𝜃, 𝜆) 𝑢 (𝑥) d𝑥, (24.8.2)
R𝑛

where 𝑢 ∈ E ′ (𝑈) (thus the integral stands for a duality bracket). Theorem 22.3.15,
implies

Theorem 24.8.2 Let 𝑢 ∈ E ′ (𝑈). Under the preceding hypotheses the property that
(𝑥 ◦ , 𝜃 ◦ ) ∉ 𝑊 𝐹a (𝑢) is equivalent to the following:
There exist neighborhoods 𝑈1 ⊂ 𝑈 of 𝑥 ◦ in R𝑛 , Θ1 ⊂ Θ of 𝜃 ◦ in C𝑛 and a constant
𝑐 ◦ > 0 such that | 𝑨𝑢 (𝜃, 𝜆)| ≲ exp (−𝑐 ◦ 𝜆) for all (𝑥, 𝜃) ∈ 𝑈1 × Θ1 , 𝜆 > 0.

24.8.2 Propagation of analytic singularities along a null

bicharacteristic curve

Throughout this subsection and the next we deal with a distribution 𝑢 ∈ E ′ (𝑈)
such that 𝑃𝑢 = 𝑃 (𝑥, D 𝑥 ) 𝑢 = 0 in a neighborhood 𝑈 of 𝑥 ◦ ∈ Ω and let (𝑥 ◦ , 𝜉 ◦ ) ∈
𝑊 𝐹a (𝑢) ∩ Char 𝑃 (see the beginning of the preceding subsection). Let 𝑃𝑚 (𝑥, 𝜉)
denote the principal symbol of 𝑃 (𝑥, D 𝑥 ); 𝑃 is of principal type (Definition 23.1.2).
Let 𝔠 be a null bicharacteristic curve of 𝑃 (𝑥, D 𝑥 ) in 𝑇 ∗ Ω\0 Ω × (R𝑛 \ {0}) passing
through (𝑥 ◦ , 𝜉 ◦ ); thus (Definition 23.1.15) 𝔠 is an integral curve in Ω × (R𝑛 \ {0}) of
the Lie algebra 𝔤 𝐻Re 𝑃𝑚 , 𝐻Im 𝑃𝑚 generated by the Hamiltonian vector fields 𝐻Re 𝑃𝑚 ,
𝐻Im 𝑃𝑚 (𝔠 is connected!). This concept is symplectically invariant and therefore (cf.
Proposition 23.1.8) we can suppose that d 𝜉 𝑃𝑚 (𝑥, 𝜉) ≠ 0 in U = 𝑈 × ℭ, where ℭ
is an open cone in R𝑛 \ {0}, ℭ ∋𝜉 ◦ . By Proposition 23.1.8, it is also invariant under
division of 𝑃𝑚 by an elliptic (i.e., nowhere vanishing) factor. We can therefore assume
that 𝑃𝑚 has the microlocal normal form (23.1.7) in U, 𝑃1 (𝑥, 𝜉) = 𝜉 𝑛 + 𝑖𝐵 (𝑥, 𝜉 ′),
𝜉 ′ = (𝜉1 , ..., 𝜉 𝑛−1 ); our hypothesis that (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 implies 𝜉 𝑛◦ = 𝐵 (𝑥, 𝜉 ′◦ ) = 0.
The vector field 𝐻 𝐵 must be collinear to 𝐻Re 𝑃𝑚 = 𝜕/𝜕𝑥 𝑛 , which requires 𝜕𝑥 𝐵 ≡ 0,
𝜕 𝜉 ′ 𝐵 ≡ 0, along 𝔠. Since 𝔠 is an integral line of 𝜕/𝜕𝑥 𝑛 and 𝐻 𝐵 ≡ 0 in 𝔠 we deduce
that 𝔠 is an interval in the straight line 𝑥 ′ = 𝑥 ′◦ , 𝜉 ◦ = (𝜉 ′◦ , 0).
We propose to exploit Theorem 24.8.2. We begin by defining the FBI phase-
function 𝜑. This is best done in the complex-analytic framework, more precisely in
a neighborhood 𝑈 C × Θ of (𝑥 ◦ , 𝜃 ◦ ) in C𝑛 such that 𝑈 = 𝑈 C ∩ R𝑛 ; 𝜃 ◦ ∈ C𝑛 satisfies
𝜕𝑧 𝜑 (0, 𝑥 ◦ , 𝜃 ◦ ) = −𝜉 ◦ . Then 𝜑 will be the (unique) holomorphic solution 𝜑 (𝑡, 𝑧, 𝜃)
in 𝑈 C × Θ of the following initial value problem:

𝜕𝜑 𝜕𝜑
− + 𝑖𝐵 (𝑧, −𝜕𝑧′ 𝜑) = 0, (24.8.3)
𝜕𝑡 𝜕𝑧 𝑛
𝜑 (0, 𝑧, 𝜃) = (𝑧 − 𝑥 ◦ ) · 𝜃 + 𝑖 |𝜃| ⟨𝑧 − 𝑥 ◦ ⟩ 2 ;
1050 24 Analytic PDEs of Principal Type. Regularity of the Solutions

𝜑 depends, holomorphically, on the 𝑡 ∈ C, |𝑡| < 𝛿. Putting 𝑧 = 𝑥 ◦ in (24.8.3) yields

𝜕𝜑 𝜕𝜑
(𝑡, 𝑥 ◦ , 𝜃) − (𝑡, 𝑥 ◦ , 𝜃) + 𝑖𝐵 (𝑥 ◦ , −𝜕𝑧′ 𝜑 (𝑡, 𝑥 ◦ , 𝜃)) = 0, (24.8.4)
𝜕𝑡 𝜕𝑧 𝑛
𝜑 (0, 𝑥 ◦ , 𝜃) = 0, 𝜕𝑧 𝜑 (0, 𝑥 ◦ , 𝜃) = 𝜃.

Proposition 24.8.3 We have 𝜃 ◦ = −𝜉 ◦ and 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = 0, 𝜕𝑧 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = −𝜉 ◦ ,

𝜕𝜃 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = 0, for all 𝑡 ∈ C, |𝑡| < 𝛿.
Proof Putting 𝜃 = 𝜃 ◦ in (24.8.4) yields 𝜕𝑧 𝜑 (0, 𝑥 ◦ , 𝜃 ◦ ) = 𝜃 ◦ = −𝜉 ◦ . Since

𝜕𝜑
𝐵 (𝑥 ◦ , −𝜕𝑧′ 𝜑 (0, 𝑥 ◦ , 𝜃 ◦ )) = 0, (0, 𝑥 ◦ , 𝜃 ◦ ) = −𝜉 𝑛◦ = 0,
𝜕𝑧 𝑛

uniqueness in the Cauchy problem implies 𝜕𝜑 ◦ ◦ ◦ ◦

𝜕𝑡 (𝑡, 𝑥 , 𝜃 ) ≡ 0, i.e. 𝜑 (𝑡, 𝑥 , 𝜃 ) = 0,
for all 𝑡’s. Differentiating (24.8.3) with respect to 𝑧 𝑗 ( 𝑗 = 1, ..., 𝑛) and then putting
𝑧 = 𝑥 ◦ , 𝜃 = 𝜃 ◦ , yields

𝜕2 𝜑 𝜕2 𝜑 𝜕𝐵 ◦
(𝑡, 𝑥 ◦ , 𝜃 ◦ ) − (𝑡, 𝑥 ◦ , 𝜃 ◦ ) − 𝑖 (𝑥 , −𝜕𝑧′ 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ ))
𝜕𝑡𝜕𝑧 𝑗 𝜕𝑧 𝑗 𝜕𝑧 𝑛 𝜕𝑧 𝑗
𝑛−1
∑︁ 𝜕𝐵 ◦ 𝜕2 𝜑
+𝑖 (𝑥 , −𝜕𝑧′ 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ )) (𝑡, 𝑥 ◦ , −𝜕𝑧′ 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ )) = 0,
𝑘=1
𝜕𝜉 𝑘 𝜕𝑧 𝑗 𝜕𝑧 𝑘

𝜕𝜑
(0, 𝑥 ◦ , 𝜃 ◦ ) = 𝜃 ◦𝑗 .
𝜕𝑧 𝑗

Since 𝐻 𝐵 ≡ 0 in 𝔠 uniqueness in the Cauchy problem implies that 𝜕𝜑

𝜕𝑧 𝑗 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = 𝜃 ◦𝑗

is the solution of this last Cauchy problem. Differentiating (24.8.3) with respect to
𝜃 𝑗 ( 𝑗 = 1, ..., 𝑛) and then putting 𝑧 = 𝑥 ◦ yields

𝜕2 𝜑 𝜕2 𝜑
(𝑡, 𝑥 ◦ , 𝜃) − (𝑡, 𝑥 ◦ , 𝜃)
𝜕𝑡𝜕𝜃 𝑗 𝜕𝑧 𝑛 𝜕𝜃 𝑗
𝑛−1
∑︁ 𝜕𝐵 ◦ 𝜕2 𝐵
+𝑖 (𝑥 , −𝜕𝑧′ 𝜑 (𝑡, 𝑥 ◦ , 𝜃)) (𝑥 ◦ , −𝜕𝑧′ 𝜑 (𝑡, 𝑥 ◦ , 𝜃)) = 0,
𝑘=1
𝜕𝜉 𝑘 𝜕𝑧 𝑘 𝜕𝜃 𝑗

𝜕𝜑
(0, 𝑥 ◦ , 𝜃) = 0.
𝜕𝜃 𝑗

Here 𝐻 𝐵 ≡ 0 in 𝔠 and uniqueness in the Cauchy problem imply 𝜕𝜑

𝜕𝜃 𝑗 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = 0

for all 𝑡 ∈ C, |𝑡| < 𝛿.

Corollary 24.8.4 If diam U C and 𝛿 > 0 are sufficiently small then 𝜑 (𝑡, 𝑧, 𝜃) is an
FBI phase-function at (𝑥 ◦ , 𝜃 ◦ ) in U C (Definition 22.1.1) for every 𝑡 ∈ C, |𝑡| < 𝛿.
24.8 Propagation of Analytic Singularities 1051

Proof Indeed, it follows from the Cauchy

conditions
in (24.8.3) that,if diamU C and
𝜕2 𝜑 2𝜑
𝛿 > 0 are sufficiently small, then det 𝜕𝑧 𝑗 𝜕𝜃𝑘 ≠ 0 and Im 𝜕𝑧𝜕𝑗 𝜕𝑧
1≤ 𝑗,𝑘 ≤𝑛 𝑘 1≤ 𝑗,𝑘 ≤𝑛
will be positive definite at every point of U C for every 𝑡 ∈ C, |𝑡| < 𝛿. □

We have the analogue of Proposition 24.4.1.

Proposition 24.8.5 Let 𝜓 (𝑡, 𝑥, 𝜃) = 𝜑 (𝑡, 𝑥, 𝜃) − (𝑥 − 𝑥 ◦ ) · 𝜃. If diam 𝑈 × Θ and 𝛿 > 0

are sufficiently small there is a 𝐶 > 0 such that
1
Im 𝜓 (𝑡, 𝑥, 𝜃) ≥ |𝜃| |𝑥 − 𝑥 ◦ | 2
2
|𝜕𝑥 𝜓 (𝑡, 𝑥, 𝜃)| 2 + |𝜕𝜃 𝜓 (𝑡, 𝑥, 𝜃)| 2 ≤ 𝐶 Im 𝜓 (𝑡, 𝑥, 𝜃) ,

for all (𝑥, 𝜃) ∈ 𝑈 × Θ, 𝑡 ∈ (−𝛿, 𝛿).

Proof By (24.8.3) we can select diam 𝑈 × Θ and 𝛿 > 0 sufficiently small

Im 𝜓 (𝑡, 𝑥, 𝜃) ≥ 21 |𝜃| |𝑥 − 𝑥 ◦ | 2 . Since 𝜕𝑧 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = 𝜃 ◦ = −𝜉 ◦ , 𝜕𝜃 𝜑 (𝑡, 𝑥 ◦ , 𝜃 ◦ ) = 0,
for all 𝑡 ∈ (−𝛿, 𝛿) we will have

|𝜕𝑥 𝜓 (𝑡, 𝑥, 𝜃 ◦ )| = |𝜕𝑥 𝜑 (𝑡, 𝑥, 𝜃 ◦ ) − 𝜃 ◦ | ≲ |𝑥 − 𝑥 ◦ | ,

|𝜕𝜃 𝜓 (𝑡, 𝑥, 𝜃 ◦ )| = |𝜕𝜃 𝜑 (𝑡, 𝑥, 𝜃 ◦ ) − (𝑥 − 𝑥 ◦ )| ≲ |𝑥 − 𝑥 ◦ | . □

In what follows 𝜑 is the solution of (24.8.3). Í

Next we seek a classical symbol 𝑎 (𝑡, 𝑥, 𝜃) = ∞ −𝑗
𝑗=0 𝜆 𝑎 𝑗 (𝑡, 𝑥, 𝜃); we wish to
⊤
relate it to the transpose 𝑃 (𝑥, D 𝑥 ) of 𝑃 (𝑥, D 𝑥 ); recall that the principal symbol of
𝑃 (𝑥, D 𝑥 ) ⊤ is 𝑃1 (𝑥, −D 𝑥 ) = −D 𝑥𝑛 + 𝑖𝐵 (𝑥, 𝜉 ′). At first reasoning formally (and in
the complex set-up), we take 𝑎 to be the formal solution of the initial value problem

D𝑡 + 𝑃 (𝑧, D𝑧 ) ⊤ e𝑖𝜆𝜑 (𝑡 ,𝑧, 𝜃) 𝑎 (𝑡, 𝑧, 𝜃, 𝜆) = 0, (24.8.7)
𝑎 (0, 𝑧, 𝜃, 𝜆) = 1 + 𝑖𝜃 · (𝑧 − 𝑥 ◦ ) (24.8.8)

[cf. (23.2.7)–(23.2.8)], obtained by the geometric optics method (we hope, by now
familiar to the reader); (24.8.8) ensures that the classical symbol will be elliptic
provided |𝑧 − 𝑥 ◦ | is sufficiently small. The transport equations are the analogues of
(23.2.14)–(23.2.15):
𝑛−1
∑︁
′ 𝛽
D𝑡 𝑎 0 − D 𝑧 𝑛 𝑎 𝑗 − 𝑖 𝜕 𝜉𝛼𝑘 𝐵 (𝑧, −𝜕𝑧′ 𝜑) D𝑧𝑘 𝑎 0 + 𝑝 0,0 𝑧, 𝜕𝑧 𝜑 𝑎 0 = 0 (24.8.9)
𝑘=1

and for 𝑗 ≥ 1,
1052 24 Analytic PDEs of Principal Type. Regularity of the Solutions

𝑛−1
∑︁
′ 𝛽
D𝑡 𝑎 𝑗 − D 𝑧 𝑛 𝑎 𝑗 + 𝑖 𝜕 𝜉𝛼𝑘 𝐵 (𝑧, −𝜕𝑧′ 𝜑) D𝑧𝑘 𝑎 𝑗 + 𝑝 0,0 𝑧, 𝜕𝑧 𝜑 𝑎 𝑗 (24.8.10)
𝑘=1
𝑗
∑︁ ∑︁ 1
𝑝 ′𝑘, 𝛼 𝑧, 𝜕𝑧 𝜑 D𝑧𝛼 𝑎 𝑗−ℓ = 0.
𝛽
+
𝛼!
ℓ=1 | 𝛼 |=ℓ

In passing note that −𝜕𝑧 𝜑 (𝑡, 𝑧, 𝜃) remains close to 𝜉 ◦ provided (𝑡, 𝑧, 𝜃) stays close
to (0, 𝑥 ◦ , 𝜃 ◦ ). It is easy to adapt the proof of Proposition 23.2.7 to show that its
conclusion remains valid, namely that if diam U C and 𝛿 > 0 are sufficiently small
and if 𝐶 > 0 is sufficiently large, then

max 𝑎 𝑗 (𝑡, 𝑧, 𝜃) ≤ 𝐶 𝑗+1 𝑗!. (24.8.11)

(𝑧, 𝜃) ∈U C , |𝑡 |< 𝛿

The next step is either to form a finite realization the formal symbol 𝑎 (Definition
19.1.7) or else use Ehrenpreis’ cutoffs exactly as in (24.8.1). The only novelty is the
presence of the variable 𝑡:
∞
∑︁
𝑎♭ (𝑡, 𝑥, 𝜃, 𝜆) = 𝑎 0 (𝑡, 𝑥, 𝜃) + 𝜒 𝑗 (𝜆/ 𝑗 𝜌) 𝜆− 𝑗 𝑎 𝑗 (𝑡, 𝑥, 𝜃) ;
𝑗=1

𝑎♭ (𝑡, 𝑥, 𝜃, |𝜃|) is a bona fide pseudoanalytic symbol in 𝑈 × (Θ ∩ R𝑛 ) depending

analytically on 𝑡 ∈ (−𝛿, 𝛿). Substituting 𝑎♭ for 𝑎 only provides an approximate
solution of (24.8.7) but the error can be estimated by exploiting the variant of Lemma
24.6.2, proved by applying Corollary 24.B.2, Appendix B to the present chapter; the
latter is made possible by Proposition 24.8.3 (playing the role of Proposition 24.4.1
in the proof of Lemma 24.6.2). To summarize, we have a solution 𝑤 (𝑡, 𝑥, 𝜃, 𝜆) =
e𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) 𝑎♭ (𝑡, 𝑥, 𝜃, 𝜆) of the equation

D𝑡 𝑤 (𝑡, 𝑥, 𝜃, 𝜆) + 𝑃 (𝑥, D 𝑥 ) ⊤ 𝑤 (𝑡, 𝑥, 𝜃, 𝜆) = e𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) 𝑟 (𝑡, 𝑥, 𝜃, 𝜆) (24.8.12)

satisfying the initial condition

◦ ) · 𝜃−𝜆| 𝑥−𝑥 ◦ | 2
𝑤 (0, 𝑥, 𝜃) = e𝑖𝜆( 𝑥−𝑥 (1 + 𝑖𝜃 · (𝑧 − 𝑥 ◦ )) (24.8.13)

[cf. (24.8.3), (24.8.8)]. The right-hand side in (24.8.12) satisfies the following es-
timates: for suitably large positive constants 𝑅, 𝐶 and all 𝛼 ∈ Z+𝑛 , (𝑥, 𝜃) ∈ 𝑈 × Θ,
|𝜃| ≥ max (1, |𝛼|), all 𝑡 ∈ (−𝛿, 𝛿) and 𝑥 ∈ 𝑈, |𝑥 − 𝑥 ◦ | < 𝜀,

D 𝑥𝛼 e𝑖 𝜑 (𝑡 , 𝑥, 𝜃 ,𝜆) 𝑟 (𝑡, 𝑥, 𝜃, 𝜆) ≤ 𝐶𝛼!e− | 𝜃 |/𝑅 . (24.8.14)

We are now in a position to prove the following result of [Hanges, 1981].

Theorem 24.8.6 Let 𝔠 be a null bicharacteristic curve of 𝑃 in 𝑇 ∗ Ω. If 𝑢 ∈ D ′ (Ω) is

such that 𝑃𝑢 ∈ C 𝜔 (Ω) then either 𝔠 ∩ 𝑊 𝐹a (𝑢) = ∅ or 𝔠 ⊂ 𝑊 𝐹a (𝑢).
24.8 Propagation of Analytic Singularities 1053

Proof We shall prove that the set 𝑇 ∗ Ω\𝑊 𝐹a (𝑢), obviously open in 𝔠, is also closed
in 𝔠 and therefore equal to 𝔠. To do this we are going to show that the hypothesis that
(𝑥 ◦ , 𝜉 ◦ ) belongs to the boundary in 𝔠 of 𝔠 ∩ (𝑇 ∗ Ω\𝑊 𝐹a (𝑢)) leads to a contradiction.
In the vicinity of (𝑥 ◦ , 𝜉 ◦ ) we can assume that 𝑥 ◦ = (𝑥 ′◦ , 0) and that 𝔠 is the image of
the C 𝜔 map (−𝛿, 𝛿) ∋ 𝑡 ↦→ ((𝑥 ′◦ , 𝑡) , 𝜉 ◦ ). By our hypothesis, at least one half of the
arc 𝑡 ↦→ ((𝑥 ′◦ , 𝑡) , 𝜉 ◦ ) is contained in 𝑇 ∗ Ω\𝑊 𝐹a (𝑢); let it be the arc in which 𝑡 > 0.
We will use the fact that 𝔠 is also a null bicharacteristic curve of the transpose 𝑃⊤
of 𝑃. There is no loss of generality in assuming supp 𝑢 ⊂⊂ Ω. [This precludes that
𝑢 is C 𝜔 in Ω but we are now reasoning microlocally, near (𝑥 ◦ , 𝜉 ◦ )]. In what follows,
𝜆 is a large positive number; we have, by (24.8.12),
√ 𝜕 D 𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) ♭ E
−1 𝑢, e 𝑎 (𝑡, 𝑥, 𝜃, 𝜆)
D 𝜕𝑡 E D 𝑥
E
⊤ 𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) ♭ 𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃)
= 𝑢, 𝑃 (𝑥, D 𝑥 ) e 𝑎 (𝑡, 𝑥, 𝜃, 𝜆) + 𝑢, e 𝑟 (𝑡, 𝑥, 𝜃, 𝜆) ,
𝑥 𝑥

where ⟨·, ·⟩ 𝑥 stands for the duality bracket between E ′ and C ∞ in 𝑥-space. We derive,
by (24.8.14),

𝜕 D 𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) ♭ E D E
𝑢, e 𝑎 (𝑡, 𝑥, 𝜃, 𝜆) − 𝑃 (𝑥, D 𝑥 ) 𝑢, e𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) 𝑎♭ (𝑡, 𝑥, 𝜃, 𝜆)
𝜕𝑡 𝑥 𝑥

≲ e−𝜆/𝑅 .

On the one hand, since ((𝑥 ◦′, 𝑡) , 𝜃 ◦ ) ∉ 𝑊 𝐹a (𝑃𝑢) whatever 𝑡 ∈ (−𝛿, 𝛿) and since 𝑎♭
is elliptic, Theorem 24.8.2 allows us to select neighborhoods 𝑈1 ⊂ 𝑈 of 𝑥 ◦ , Θ1 ⊂ Θ
of 𝜃 ◦ , and the positive constants 𝛿, 𝑐 ◦ such that, if supp 𝑢 ⊂ 𝑈1 , then
D E
𝑃 (𝑥, D 𝑥 ) 𝑢, e𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) 𝑎♭ (𝑡, 𝑥, 𝜃, 𝜆) ≲ e−𝑐◦ 𝜆
𝑥

for all 𝜃 ∈ Θ1 , 𝑡 ∈ (−𝛿, 𝛿). By integrating from 0 to 𝑡 > 0 (fixed) we derive from the
preceding two inequalities and from (24.8.13)
D E
𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) ♭ 𝑖𝜆( 𝑥−𝑥 ◦ ) · 𝜃−𝜆| 𝑥−𝑥 ◦ | 2 𝜃 ◦
𝑢, e 𝑎 (𝑡, 𝑥, 𝜃, 𝜆) − 𝑢, e 1+𝑖 (𝑥 − 𝑥 )
𝑥 |𝜃| 𝑥
≲ |𝑡| e−𝑐◦ 𝜆 .

By Theorem 24.8.2 our hypothesis that ((𝑥 ◦′, 𝑡) , 𝜃 ◦ ) ∉ 𝑊 𝐹a (𝑢) whatever 𝑡 ∈ (0, 𝛿)
implies
D E
∀ (𝑥, 𝜃) ∈ 𝑈1 × Θ1 , 𝑢, e𝑖𝜆𝜑 (𝑡 , 𝑥, 𝜃) 𝑎♭ (𝑡, 𝑥, 𝜃, 𝜆) ≲ e−𝑐◦ 𝜆 ,
𝑥

possibly with decreased 𝑐 ◦ > 0 and diam 𝑈1 × Θ1 . Combining the preceding two
inequalities yields

◦ ◦ 2 𝜃
𝑢, e𝑖𝜆( 𝑥−𝑥 ) · 𝜃−𝜆| 𝑥−𝑥 | 1 + 𝑖 (𝑥 − 𝑥 ◦ ) ≲ e−𝑐◦ 𝜆 .
|𝜃| 𝑥
1054 24 Analytic PDEs of Principal Type. Regularity of the Solutions

By Theorem 24.8.2 (or, simply, by Definition 3.5.1) this proves that (𝑥 ◦ , 𝜉 ◦ ) ∉

𝑊 𝐹a (𝑢). □

24.8.3 Propagation of analytic singularities along a null

bicharacteristic surface

In this subsection we prove the following consequence of Theorem 9.2, p. 62,

in [Sjöstrand, 1982]. As before 𝑃 = 𝑃 (𝑥, D 𝑥 ) is a differential operator with C 𝜔
complex coefficients in the open subset Ω of R𝑛 ; the order of 𝑃 is 𝑚 ≥ 1.
Theorem 24.8.7 Let 𝑃 be of principal type and satisfy Condition (P). Let Σ be a null
bicharacteristic surface of 𝑃 in Ω (Definition 23.1.15). If 𝑢 ∈ D ′ (Ω) is such that
𝑃𝑢 ∈ C 𝜔 (Ω) then either Σ ∩ 𝑊 𝐹a (𝑢) = ∅ or Σ ⊂ 𝑊 𝐹a (𝑢).
Proof It follows from Proposition 23.1.43, that an arbitrary null bicharacteristic
surface of 𝑃 is orientable and carries the natural structure of a Riemann surface. Let
(𝑥 ◦ , 𝜉 ◦ ) ∈ Σ, (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑢). We adapt the argument in the proof of Theorem
◦ ◦
24.8.6, now interpreting 𝑡 as a complex coordinate in Σ, 𝑡 = 0 at (𝑥 , 𝜃 ), and
𝜕 1 𝜕 𝜕
𝜕𝑡 = 2 𝜕 Re 𝑡 − 𝑖 𝜕 Im 𝑡 ; we identify Δ 𝛿 = {𝑡 ∈ C; |𝑡| < 𝛿}, 𝛿 > 0 suitably small,
with an open subset of Σ. Then 𝜑 (𝑡, 𝑧, 𝜃) is the holomorphic solution of (24.8.3),
𝑎 (𝑡, 𝑧, 𝜃, 𝜆) is the holomorphic solution of (24.8.7)–(24.8.8), all this in a suitable
conic neighborhood of (0, 𝑥 ◦ , 𝜃 ◦ ) in Δ 𝛿 × ΩC × Θ. From there on the reasoning
proceeds exactly as in the proof of Theorem 24.8.6 replacing everywhere [−𝛿, 𝛿] by
Δ𝛿. □

24.8.4 Nonconfinement of the null bicharacteristic leaves and some

consequences

We continue to deal with a differential operator 𝑃 = 𝑃 (𝑥, D 𝑥 ) with C 𝜔 complex

coefficients in the open subset Ω of R𝑛 , of order 𝑚 ≥ 1.
Definition 24.8.8 Let 𝑃 be of principal type. We shall say that a null bicharacteristic
leaf of 𝑃, L ⊂ 𝑇 ∗ Ω\0, is trapped or confined if the base projection of L is contained
in a compact subset of Ω.
𝜕 𝜕 𝜕
Example 24.8.9 The characteristic set of the rotation vector field 𝜕𝜃 = 𝑥 𝜕𝑦 − 𝑦 𝜕𝑥
in R2 \ {0} is defined by the equation 𝐵 = 𝑥𝜂 − 𝑦𝜉 = 0. It is a C 𝜔 hypersurface Σ in
2
R2 \ {0} . The Hamiltonian vector field 𝐻 𝐵 = 𝜕𝜃 𝜕 𝜕
+ 𝜉 𝜕𝜂 − 𝜂 𝜕𝜕𝜉 is different from
zero and tangent to Σ at every point. As a consequence, the null bicharacteristic
𝜕
leaves of 𝜕𝜃 are the integral curves of 𝐻 𝐵 in Σ. These are defined by the parametric
equations
𝑥 = 𝑅 cos 𝜃, 𝑦 = 𝑅 sin 𝜃, 𝜉 = 𝜌 cos 𝜃, 𝜂 = 𝜌 sin 𝜃,
24.8 Propagation of Analytic Singularities 1055

with 𝑅 > 0, 𝜌 > 0 arbitrary. Every one of these null bicharacteristic leaves is trapped.
Theorem 24.8.10 Let 𝑃 be an analytic differential operator of principal type and
satisfy Condition (P). If there are no trapped null bicharacteristic leaves of 𝑃 of
positive dimension in 𝑇 ∗ Ω\0 then the homogeneous equation 𝑃𝑢 = 0 does not have
any solution 𝑢 ∈ E ′ (Ω) other than zero.
Proof We reason by contradiction. Let 𝑢 ∈ E ′ (Ω) satisfy 𝑃𝑢 = 0 in Ω. We select an
open set Ω′ ⊂⊂ Ω with smooth boundary 𝜕Ω′ and such that supp 𝑢 ⊂ Ω′. Let 𝑥 ◦ ∈
supp 𝑢, dist (𝑥 ◦ , 𝜕Ω′) = dist (supp 𝑢, 𝜕Ω′). There is a neighborhood 𝑈 ⊂ Ω of 𝑥 ◦ and
a function 𝑓 ∈ C 𝜔 (𝑈) such that d 𝑓 ≠ 0 everywhere in 𝑈, 𝑓 (𝑥 ◦ ) = 0 and 𝑓 (𝑥) ≤
0 =⇒ dist (𝑥, 𝜕Ω′) ≤ dist (supp 𝑢, 𝜕Ω′). This allows us to apply Theorem 3.5.12,
and conclude that if 𝜉 ◦ · d𝑥 = d 𝑓 (𝑥 ◦ ) then (𝑥 ◦ , ±𝜉 ◦ ) ∈ 𝑊 𝐹a (𝑢). The Holmgren
Theorem 5.3.12, demands 𝑃𝑚 (𝑥 ◦ , ±𝜉 ◦ ) = 0. Suppose no null bicharacteristic leaf of
𝑃 of positive dimension contains (𝑥 ◦ , 𝜉 ◦ ). We apply Theorem 24.2.5: there is a conic
neighborhood U of (𝑥 ◦ , 𝜉 ◦ ) in 𝑇 ∗ Ω\0 such that U ∩ 𝑊 𝐹a (𝑢) = U ∩ 𝑊 𝐹a (𝑃𝑢) = ∅
contradicting (𝑥 ◦ , 𝜉 ◦ ) ∈ 𝑊 𝐹a (𝑢). Lastly, suppose that (𝑥 ◦ , 𝜉 ◦ ) belongs to a null
bicharacteristic leaf L of 𝑃 of dimension 1 or 2. By Theorems 24.8.6 and 24.8.7 we
have L ⊂ 𝑊 𝐹a (𝑢), implying that the base projection of L is contained in supp 𝑢
and thereby contradicting the nonconfinement hypothesis. □
Corollary 24.8.11 Let 𝑃 be an analytic differential operator of principal type and
satisfy Condition (P). If there are no trapped null bicharacteristic leaves of 𝑃 of
positive dimension in Ω then 𝑃C ∞ (Ω) is dense in C ∞ (Ω) and 𝑃D ′ (Ω) is dense in
D ′ (Ω).
Proof Follows from Theorem 24.8.10 and the Hahn–Banach Theorem. □

24.8.5 Solvability in hyperfunctions of an analytic PDE of principal

type

Let 𝑃 = 𝑃 (𝑥, D) be an analytic differential operator in the domain Ω ⊂ R𝑛 as in the

preceding subsection.
Theorem 24.8.12 Let 𝑃 be of principal type and satisfy Condition (P). Let 𝑈 ⊂⊂ Ω
be an open set. If there are no trapped null bicharacteristic leaves of 𝑃 of positive
dimension above 𝑈 then 𝑃 (𝑥, D) B (𝑈) = B (𝑈).
About the space B (𝑈) of hyperfunctions in 𝑈 see Ch. 7. A conclusion such as
that in Theorem 24.8.12 is often referred to as a result on semiglobal solvability, in
contrast to Corollary 24.8.13 and Theorem 24.8.14 below, results in local solvability.
Proof We apply Theorem 7.1.22: let 𝑢 ∈ C ∞ (Ω) be such that 𝑢| 𝑉 ∈ C 𝜔 (𝑉),
𝑉 ⊂ Ω open, 𝜕𝑈 ⊂ 𝑉. By hypothesis every null bicharacteristic leaf of 𝑃 of positive
dimension in 𝑇 ∗ Ω\0|𝑈 intersects 𝑇 ∗ Ω\0| 𝑉 whereas 𝑊 𝐹a (𝑢) does not. It ensues that
𝑊 𝐹a (𝑢) = ∅, by Theorems 24.8.6 and 24.8.7. This proves that Condition (A) in
Theorem 7.1.22, is satisfied, whence the claim. □
1056 24 Analytic PDEs of Principal Type. Regularity of the Solutions

Corollary 24.8.13 Let 𝑃 be of principal type and satisfy Condition (P). An arbitrary
point 𝑥 ◦ ∈ Ω has a neighborhood 𝑈 ⊂ Ω such that 𝑃 (𝑥, D) B (𝑈) = B (𝑈).

Proof It is an easy exercise to prove that, under the hypotheses of Theorem 24.8.12,
𝑥 ◦ ∈ Ω has a neighborhood 𝑈 ⊂ Ω such that there are no trapped null bicharacteristic
leaves of 𝑃 of positive dimension in 𝑇 ∗ Ω\0|𝑈 . □
Corollary 24.8.13 leads naturally to the question of whether Condition (P) is also
necessary for the local solvability of 𝑃 (𝑥, D) in hyperfunctions. The latter property
is equivalent to the germ equation P 𝑥 ◦ B 𝑥 ◦ = B 𝑥 ◦ by Proposition 7.1.20 [B 𝑥 ◦ is the
stalk at 𝑥 ◦ of the sheaf of hyperfunctions in R𝑛 and P 𝑥 ◦ is the linear endomorphism
of B 𝑥 ◦ induced by 𝑃 (𝑥, D 𝑥 )]. We shall now show that the answer to our question is
positive.

Theorem 24.8.14 Let 𝑃 (𝑥, D 𝑥 ) be a differential operator of principal type with C 𝜔

complex coefficients in the open subset Ω of R𝑛 and let 𝑥 ◦ ∈ Ω be arbitrary. For P 𝑥 ◦
to be surjective it is necessary and sufficient that 𝑃 (𝑥, D) satisfies Condition (P).

Proof To prove the necessity of (P) for P 𝑥 ◦ B 𝑥 ◦ = B 𝑥 ◦ to hold we shall reason by

contradiction. We shall reason under the hypothesis that Property (P) does not hold
at (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃; there is no loss of generality in taking 𝜉 ◦ = (0, ..., 0, 1). We
follow to a large extent the reasoning in Subsection 23.2.3. Let 𝑈 C be a domain in
C𝑛 such that 𝑈 C ∩ R𝑛 = 𝑈 ∋ 𝑥 ◦ . We seek approximate solutions of the equation
𝑃 (𝑧, D𝑧 ) ⊤ ℎ = 0 [𝑃 (𝑧, D𝑧 ) ⊤ : transpose of 𝑃 (𝑧, D𝑧 )] of the type ℎ [𝑁 ] = e𝑖𝜆𝜑 𝑎 [𝑁 ]
where

(1) the phase-function 𝜑 ∈ O 𝑈 C is the solution of the eikonal equation (23.2.28)
[equivalent to (23.2.5)] satisfying the initial condition 𝜑| 𝑧𝑛 =0 = 𝑧 1 + 𝑖𝑧 ′ · 𝑧 ′,
whence the finite expansion (23.2.29);
(2) the amplitude ∑︁
𝑎 [𝑁 ] (𝑧, 𝜆) = 𝜆− 𝑗 𝑎 𝑗 (𝑧, 𝜆) (𝑁 > 0)
𝑗 ≤𝜆/𝑁

originates from the solutions 𝑎 𝑗 of the transport equations (23.2.14)–(23.2.15) in

which 𝑃 (𝑧, D𝑧 ) is replaced by 𝑃 (𝑧, D𝑧 ) ⊤ and the initial conditions are 𝑎 0 | 𝑧𝑛 =0 =
1, 𝑎 𝑗 𝑧𝑛 =0 = 0.

We apply Proposition 23.2.7, and Lemma 24.5.1: 𝑎 𝑗 ∈ O 𝑈 C and the estimates
(23.2.6) hold; possibly after a contraction of 𝑈 C about 𝑥 ◦ there are numbers 𝐶1 > 0
and 𝑁 sufficiently large that

∀𝜆 ≥ 1, sup 𝑎 [ 𝑁 ] (𝑧, 𝜆) ≤ 𝐶◦ . (24.8.15)

𝑧 ∈𝑈 C

We also apply Proposition 24.5.2: provided 𝑁 and 𝐶2 > 0 are sufficiently large,

sup e−𝑖𝜆𝜑 𝑃 (𝑧, D𝑧 ) ⊤ e𝑖𝜆𝜑 𝑎 [𝑁 ] ≤ 𝐶2 e−𝜆/𝑁 . (24.8.16)
𝑈C
24.A Appendix: Properties of Real Polynomials in a Single Variable 1057

At this stage we avail ourselves of Lemma 7.1.21. If 𝑃B (𝑈) = B (𝑈) then to

every 𝜀 > 0 there are 𝛿 > 0, 𝐶3 > 0 such that, for every 𝑓 ∈ O (C𝑛 ),

∥ 𝑓 ∥ 𝐿 ∞ (𝑈 𝛿 ) ≤ 𝐶3 𝑃 (𝑧, D𝑧 ) ⊤ 𝑓 𝐿 ∞ (𝑈 𝜀 ) + ∥ 𝑓 ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 ) . (24.8.17)

We recall that 𝑆 𝜀 = {𝑧 ∈ C𝑛 ; dist (𝑧, 𝑆) < 𝜀}, 𝑆 ⊂ C𝑛 , 𝜀 > 0. Since 𝑈 is Runge

the estimate (24.8.17) remains valid for all 𝑓 ∈ O 𝑈 𝜀◦ , with 𝜀◦ > 0 such that
𝑈 𝜀◦ ⊂⊂ ΩC , a domain in C𝑛 such that Ω ⊂ ΩC ∩R𝑛 ; we can assume 0 < 𝛿 < 𝜀 < 𝜀◦ .
This allows us to put 𝑓 = e𝑖𝜆𝜑 𝑎 [𝑁 ] in (24.8.17). From (24.8.16) we derive, for all
𝑧 ∈ 𝑈C,

𝑃 (𝑧, D𝑧 ) ⊤ e𝑖𝜆𝜑 (𝑧) 𝑎 [𝑁 ] (𝑧, 𝜆) ≤ 𝐶2 e−𝜆 Im 𝜑 (𝑧)−𝜆/𝑁 . (24.8.18)

From (23.2.30) we derive, for suitable positive constants 𝑐, 𝐶, and all 𝑧 = 𝑥+𝑖𝑦 ∈ 𝑈 𝜀◦ ,

Im 𝜑 (𝑧) ≥ 𝑐 |𝑥 ′ | 2 + 𝑥 𝑛2𝜈 − 𝐶 |𝑦 ′ | 2 + |𝑦 𝑛 | 2𝜈+1 . (24.8.19)

We take n o
𝑈 = 𝑥 ∈ R𝑛 ; |𝑥 ′ | 2 + 𝑥 𝑛2𝜈 ≤ 𝑟 2

with 𝑟 > 0 sufficiently small that 𝑈 ⊂⊂ Ω. In (24.8.17) we can take 𝜀 ≪

min 𝑟, 𝑁 −1 . From (24.8.18) and (24.8.19) we derive
2
𝑃 (𝑧, D𝑧 ) ⊤ 𝑓 𝐿 ∞ (𝑈 𝜀 )
≤ 𝐶2 e−𝜆/𝑁 +𝐶 𝜀 𝜆 ,
2 𝜆+𝐶 𝜀 2 𝜆
∥ 𝑓 ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 ) ≤ 𝐶◦ e−𝑐𝑟 .

1
If we select 𝜀 2 ≤ min 𝑁 −1 , 𝑐𝑟 2 we get

2𝐶

1 2
𝑃 (𝑧, D𝑧 ) ⊤ 𝑓 𝐿 ∞ (𝑈 𝜀 )
+ ∥ 𝑓 ∥ 𝐿 ∞ ( (𝜕𝑈) 𝜀 ) ≤ 𝐶2 e−𝜆/2𝑁 + 𝐶◦ e− 2 𝑐𝑟 𝜆 .

We also have
1 = | 𝑓 (0)| ≤ ∥ 𝑓 ∥ 𝐿 ∞ (𝑈 𝛿 ) .
Putting this into (24.8.17) and letting 𝜆 ↗ +∞ leads directly to a contradiction. □

24.A Appendix: Properties of Real Polynomials in a Single

Variable

Let P 𝑚 denote the (complex) vector subspace of C [𝑡] consisting of the polynomials
of degree ≤ 𝑚 (∈ Z+ ); dim P 𝑚 = 𝑚 + 1.

Lemma 24.A.1 There is a constant 𝐶𝑚 ≥ 1 such that, for all 𝑔 ∈ P 𝑚 and all pairs
of real numbers 𝑡 ≤ 𝑡 ′,
1058 24 Analytic PDEs of Principal Type. Regularity of the Solutions
𝑚 ∫ 𝑡′
∑︁ d𝑗𝑔 𝐶𝑚
(𝑡 ′ − 𝑡) 𝑗 (𝑡) ≤ ′ |𝑔 (𝑠)| d𝑠. (24.A.1)
𝑗=0
d𝑡 𝑗 𝑡 −𝑡 𝑡

Proof Suppose 𝑡 < 𝑡 ′ and define ℎ (𝑥) = 𝑔 (𝑡 ′ − (𝑡 ′ − 𝑡) 𝑥). We have

𝑚 𝑚
∑︁ d𝑗𝑔 ∑︁ d 𝑗 ℎ
(𝑡 ′ − 𝑡) 𝑗 (𝑡) = (1) (24.A.2)
𝑗=0
d𝑡 𝑗
𝑗=0
d𝑡 𝑗

𝑡 ′ −𝑠
and the change of variable 𝑥 = 𝑡 ′ −𝑡 yields
∫ 𝑡′ ∫ 1
1
|𝑔 (𝑠)| d𝑠 = |ℎ (𝑥)| d𝑥. (24.A.3)
𝑡 − 𝑡′ 𝑡 0

The claim follows from the fact that the right-hand sides in both (24.A.2) and
(24.A.3) are norms on P 𝑚 , per force equivalent. The case 𝑡 = 𝑡 ′ follows by going to
the limit as 𝑡 ↗ 𝑡 ′. □

Corollary 24.A.2 If 𝑓 ∈ P 𝑚 is monic of degree 𝑚 and 𝑡 ≤ 𝑡 ′ then

∫ 𝑡′
′ 𝑚+1 𝐶𝑚
(𝑡 − 𝑡) ≤ | 𝑓 (𝑠)| d𝑠. (24.A.4)
𝑚! 𝑡

In the remainder of this appendix 𝑓 ∈ R [𝑡] shall be monic of degree 𝑚 and satisfy
the following condition, in which 𝑇0 and 𝑇1 are real numbers, 𝑇0 < 𝑇1 :

∀𝑡 ∈ [𝑇0 , 𝑇1 ] , 𝑓 (𝑡) ≥ 0. (24.A.5)

Let ℓ, 𝜈 ∈ [0, +∞), 𝜃 ∈ [0, 1] and consider the following integral, in which 𝑇0 < 𝑡 <
𝑇1 and 𝜆 ≥ 1,
∫ 𝑇1 ∫ 𝑡 ′ 𝜈 ∫ 𝑡′
𝐼ℓ,𝜈, 𝜃 (𝑡, 𝜆) = (𝑡 − 𝑡 ′) ℓ 𝑓 (𝑠) d𝑠 e−𝜆 𝑡 𝑓 (𝑠)d𝑠 𝑓 (𝑡 ′) 𝜃 d𝑡 ′. (24.A.6)
𝑡 𝑡

Lemma 24.A.3 If (24.A.5) holds then there is a constant 𝐶𝑚,ℓ,𝜈 > 0 independent of
𝑓 and 𝑇0 , 𝑇1 such that
ℓ+1 𝑚 ′
𝑓 (𝑡) 𝜃 𝐼ℓ,𝜈, 𝜃 ′ (𝑡, 𝜆) ≤ 𝐶𝑚,ℓ,𝜈 𝜆−𝜈− 𝑚+1 − 𝑚+1 ( 𝜃+𝜃 ) (24.A.7)

for all 𝜆 ≥ 1, 𝑡 ∈ [𝑇0 , 𝑇1 ], 𝜃 ≥ 0, 𝜃 ′ ≥ 0, 𝜃 + 𝜃 ′ ≤ 1.

Proof Applying (24.A.4) we get

ℓ
∫ 𝑇1 ∫ 𝑡′ 𝜈+ 𝑚+1 ∫ 𝑡′ ′
𝐶𝑚 𝑓 (𝑠)d𝑠
𝐼ℓ,𝜈, 𝜃 ′ (𝑡, 𝜆) ≤ ℓ
𝑓 (𝑠) d𝑠 e−𝜆 𝑡 𝑓 (𝑡 ′) 𝜃 d𝑡 ′.
(𝑚!) 𝑚+1 𝑡 𝑡

We also have
24.A Appendix: Properties of Real Polynomials in a Single Variable 1059
ℓ
∫ 𝑡′ 𝜈+ 𝑚+1 ∫ 𝑡′
1 ℓ
𝑓 (𝑠)d𝑠
𝑓 (𝑠) d𝑠 e− 2 𝜆 𝑡
′
≤ 𝐶𝑑,ℓ,𝜈 𝜆−𝜈− 𝑚+1
𝑡

with
′ ℓ
𝜈+ 𝑚+1 ℓ
𝐶𝑚,ℓ,𝜈 ≈2 Γ 𝜈+ +1 ,
𝑚+1
1
∫ 𝑡′
𝑓 (𝑠)d𝑠
whence, writing 𝑢 (𝑡, 𝑡 ′, 𝜆) = e− 2 𝜆 𝑡 ,
∫ 𝑇1
ℓ ′
′′
𝑓 (𝑡) 𝜃 𝐼ℓ,𝜈, 𝜃 (𝑡, 𝜆) ≤ 𝐶𝑑,ℓ,𝜈 𝜆−𝜈− 𝑚+1 𝑢 (𝑡, 𝑡 ′, 𝜆) 𝑓 (𝑡) 𝜃 𝑓 (𝑡 ′) 𝜃 d𝑡 ′. (24.A.8)
𝑡
′ ′ ′
Since 𝑓 (𝑡) 𝜃 𝑓 (𝑡 ′) 𝜃 ≤ 𝑓 (𝑡) 𝜃+𝜃 + 𝑓 (𝑡 ′) 𝜃+𝜃 it suffices to prove the result when either
𝜃 or 𝜃 ′ or both are equal to zero. We are also going to apply the Hölder inequalities:
∫ 𝑇1 ∫ 𝑇1
′ 𝜃′ ′ ′
′
𝑢 (𝑡, 𝑡 , 𝜆) 𝑓 (𝑡 ) d𝑡 = ′
𝑢 (𝑡, 𝑡 ′, 𝜆) 1−𝜃 (𝑢 (𝑡, 𝑡 ′, 𝜆) 𝑓 (𝑡 ′)) 𝜃 d𝑡 ′ (24.A.9)
𝑡 𝑡
∫ 𝑇1 1−𝜃 ′ ∫ 𝑇1 𝜃′
′ ′ ′ ′ ′
≤ 𝑢 (𝑡, 𝑡 , 𝜆) d𝑡 𝑢 (𝑡, 𝑡 , 𝜆) 𝑓 (𝑡 ) d𝜃
𝑡 𝑡

and
∫ 𝑇1 ∫ 𝑇1
′ 𝜃
𝑢 (𝑡, 𝑡 , 𝜆) 𝑓 (𝑡) d𝑡 ≤ ′
𝑢 (𝑡, 𝑡 ′, 𝜆) 1−𝜃 (𝑢 (𝑡, 𝑡 ′, 𝜆) 𝑓 (𝑡)) 𝜃 d𝑡 ′ (24.A.10)
𝑡 𝑡
∫ 𝑇1 1−𝜃 ∫ 𝑇1 𝜃
≤ 𝑢 (𝑡, 𝑡 ′, 𝜆) d𝑡 ′ 𝑓 (𝑡) 𝑢 (𝑡, 𝑡 ′, 𝜆) d𝑡 ′ .
𝑡 𝑡
1060 24 Analytic PDEs of Principal Type. Regularity of the Solutions

This shows that it suffices to deal with the following three cases:
Case 1: 𝜃 = 𝜃 ′ = 0. We apply again (24.A.4) setting 21 𝐶𝑑−1 𝑚! = 𝑐 𝑚 :
∫ 𝑇1 ∫ 𝑇1
′ ′ ′ 𝑚+1
𝑢 (𝑡, 𝑡 , 𝜆) d𝑡 ≤ e−𝑐𝑚 𝜆(𝑡 −𝑡) d𝑡 ′,
𝑡 𝑡

whence
∫ 𝑇1 ∫ +∞
1
𝑢 (𝑡, 𝑡 ′, 𝜆) d𝑡 ′ ≤ e−𝑐𝑚 𝜆𝑠 d𝑠 = 𝐶𝑚 𝜆− 𝑚+1 .
𝑚+1
(24.A.11)
𝑡 0

Putting this in (24.A.8) yields (24.A.7) in Case 1.

Case 2: 𝜃 = 0, 𝜃 ′ = 1. We have
∫ 𝑇1 ∫ 𝑇1
1
∫ 𝑡′
𝑓 (𝑠)d𝑠
𝑢 (𝑡, 𝑡 ′, 𝜆) 𝑓 (𝑡 ′) d𝑡 ′ = e− 2 𝜆 𝑡 𝑓 (𝑡 ′) d𝑡 ′
𝑡 𝑡

whence
∫ 𝑇1 ∫ +∞
1
′ ′ ′
𝑢 (𝑡, 𝑡 , 𝜆) 𝑓 (𝑡 ) d𝑡 ≤ e− 2 𝜆𝑠 d𝑠 = 2𝜆−1 . (24.A.12)
𝑡 0

Putting this in (24.A.8) yields (24.A.7) in Case 2.

Case 3: 𝜃 = 1, 𝜃 ′ = 0. We derive from (24.A.1)
∫ 𝑡′
(𝑡 ′ − 𝑡) 𝑓 (𝑡) ≤ 𝐶𝑚 𝑓 (𝑠) d𝑠,
𝑡

whence ∫ ∫
𝑇1 𝑇1
1 ′
𝑓 (𝑡) ′
𝑢 (𝑡, 𝑡 , 𝜆) d𝑡 ≤ 𝑓 (𝑡) ′
e− 2 𝐶𝑚 𝜆(𝑡 −𝑡) 𝑓 (𝑡) d𝑡 ′
𝑡 𝑡
and therefore
∫ 𝑇1 ∫ +∞
𝑓 (𝑡) 𝑢 (𝑡, 𝑡 ′, 𝜆) d𝑡 ′ ≤ 𝑓 (𝑡) e−𝑐𝑚 𝜆 𝑓 (𝑡)𝑠 d𝑠 = 𝐶𝑚 (2𝜆) −1 . (24.A.13)
𝑡 0

Putting this in (24.A.8) yields (24.A.7) in Case 3.

By combining (24.A.9), (24.A.11) and (24.A.12) we obtain (24.A.7) when 𝜃 = 0,
0 < 𝜃 ′ < 1. By combining (24.A.10), (24.A.11) and (24.A.13) we obtain (24.A.7)
when 0 < 𝜃 < 1, 𝜃 ′ = 0. □
Remark 24.A.4 The constant 𝐶𝑚,ℓ,𝜈 in (24.A.7) depends neither on the segment
[𝑇0 , 𝑇1 ] nor on 𝜃, 𝜃 ′.
Now, with ℓ, 𝜈, 𝜃, 𝜃 ′ ∈ R+ , 𝜃 ≤ 1, 𝑡 ∈ [𝑇0 , 𝑇1 ] as before, consider the integral
∫ 𝑇1 ∫ 𝑡′ 𝜈 ∫ 𝑡′ ′
𝑏 (𝑠)d𝑠
𝐽ℓ,𝜈, 𝜃 (𝑡, 𝜆) = (𝑡 − 𝑡 ′) ℓ ℎ1 (𝑠) 𝑏 (𝑠) d𝑠 e−𝜆 𝑡 𝑏 (𝑡 ′) 𝜃 ℎ2 (𝑡 ′) d𝑡 ′.
𝑡 𝑡
24.B Appendix: Analytic Estimates of Exponential Amplitudes 1061

Corollary 24.A.5 Let ℎ 𝑗 (𝑡) ∈ C ([𝑇0 , 𝑇1 ]), 𝑗 = 1, 2, 0 ≤ 𝜃 ′ ≤ 1 − 𝜃. If 𝑏 ∈

C ( [𝑇0 , 𝑇1 ]) is such that

∃𝑐 ◦ > 0, ∀𝑡 ∈ [𝑇0 , 𝑇1 ] , 𝑐 ◦ 𝑓 (𝑡) ≤ 𝑏 (𝑡) ≤ 𝑐−1

◦ 𝑓 (𝑡) (24.A.14)

then there is a 𝐶 > 0 such that

ℓ+1 𝑚 ′
𝑏 (𝑡) 𝜃 𝐽ℓ,𝜈, 𝜃 ′ (𝑡, 𝜆) ≤ 𝐶𝜆−𝜈− 𝑚+1 − 𝑚+1 ( 𝜃+𝜃 ) (24.A.15)

for all 𝜆 ≥ 1, 𝑡 ∈ [𝑇0 , 𝑇1 ].

Proof Since (24.A.14) and 𝑡 < 𝑡 ′ imply
∫ 𝑡′ ∫ 𝑡′ ∫ 𝑡′
𝑐◦ 𝑓 (𝑠) d𝑠 ≤ 𝑏 (𝑠) d𝑠 ≤ 𝑐−1
◦ 𝑓 (𝑠) d𝑠
𝑡 𝑡 𝑡

we have

′ −𝜈
𝜃
𝑏 (𝑡) 𝐽ℓ,𝜈, 𝜃 (𝑡, 𝜆) ≤ 𝑐−𝜃−𝜃
◦ max |ℎ1 | 𝜈
max |ℎ2 | 𝑓 (𝑡) 𝜃 𝐼ℓ,𝜈, 𝜃 ′ (𝑡, 𝑐 ◦ 𝜆) .
[𝑇0 ,𝑇1 ] [𝑇0 ,𝑇1 ]

By (24.A.7) we have
ℓ+1 𝑚 ′
𝑓 (𝑡) 𝜃 𝐼ℓ,𝜈, 𝜃 ′ (𝑡, 𝑐 ◦ 𝜆) ≤ 𝐶𝑚,ℓ,𝜈 (𝑐 ◦ 𝜆) −𝜈− 𝑚+1 − 𝑚+1 ( 𝜃+𝜃 ),

whence (24.A.15). □

24.B Appendix: Analytic Estimates of Exponential Amplitudes

Let Ω be an open subset of R 𝑁 and 𝜙 ∈ C 𝜔 (Ω; C).

Lemma 24.B.1 Suppose there is a 𝐶∗ > 0 such that

∀𝑥 ∈ Ω, |∇𝜙 (𝑥)| 2 + |𝜙 (𝑥)| 2 ≤ 𝐶∗ Re 𝜙 (𝑥) . (24.B.1)

Then, to each compact set 𝐾 ⊂ Ω there are positive numbers 𝜀 < 1 and 𝐶 such that,
for all 𝜆 ≥ 1 and 𝛼 ∈ Z+𝑁 ,

1 √ Ö 𝑁
𝛼 /2
max 𝜕𝑥𝛼 e−𝜆𝜙 ( 𝑥) ≤ 𝐶 | 𝛼 | 𝜀 − 2 | 𝛼 | 𝛼! max 𝜆, 𝛼 𝑗 𝑗 e−(1−𝜀)𝜆 Re 𝜙 ( 𝑥) .
𝑥 ∈𝐾
𝑗=1
(24.B.2)
Proof Assume 𝜙 extends holomorphically to the domain ΩC ⊂ C 𝑁 , Ω ⊂ ΩC . Let
𝑟 = (𝑟 1 , ..., 𝑟 𝑁 ), 𝑟 𝑗 > 0, be such that
n o
∀𝑥 ◦ ∈ 𝐾, Δ𝑟( 𝑁 ) (𝑥 ◦ ) = 𝑧 ∈ C 𝑁 ; 𝑧 𝑗 − 𝑥 ◦𝑗 < 𝑟 𝑗 , 𝑗 = 1, ..., 𝑁 ⊂⊂ ΩC .
1062 24 Analytic PDEs of Principal Type. Regularity of the Solutions

We may assume 𝐶∗ = 1; 𝐶1 , 𝐶2 ... will be suitably large positive constants,

independent of 𝛼 ∈ Z+𝑁 , 𝜆 ≥ 1 and 𝑥 ◦ ∈ 𝐾. The claim being trivial for 𝛼 = 0 we may
assume |𝛼| ≥ 1. As a matter of fact, we have the right to ignore every variable 𝑥 𝑗
such that 𝛼 𝑗 = 0. In other words, there is no loss of generality in assuming 𝛼 𝑗 ≥ 1 for
every 𝑗 = 1, ..., 𝑁. Set 𝜇 𝑗 = 𝛼 𝑗 /max 𝜆, 𝛼 𝑗 ; we have 0 < 𝜇 𝑗 ≤ 1 and 𝜆𝜇 𝑗 ≤ 𝛼 𝑗 . If
√
𝑧 𝑗 − 𝑥 ◦𝑗 ≤ 𝜀𝜇 𝑗 , 𝑗 = 1, ..., 𝑁, 0 < 𝜀 < min𝑟 2𝑗 , then the hypothesis (24.B.1) implies
𝑗

𝑁
∑︁
𝜆 |𝜙 (𝑧) − 𝜙 (𝑥 ◦ )| ≤ 𝜆 |(𝑧 − 𝑥 ◦ ) · ∇𝜙 (𝑥 ◦ )| + 𝐶1 𝜆 𝜇𝑗
𝑗=1

≤ 𝜀𝜆 Re 𝜙 (𝑥 ◦ ) + 𝐶2 𝜀 |𝛼| ,

whence
◦
max e−𝜆 Re 𝜙 (𝑧) ≤ 𝐶3| 𝛼 | e−(1−𝜀)𝜆 Re 𝜙 ( 𝑥 ) ,
( )
𝑧 ∈Δ√𝑁𝜀 𝜇 ( 𝑥 ◦ )

√ √ √
where 𝜀𝜇 = 𝜀𝜇1 , ..., 𝜀𝜇 𝑁 . The Cauchy inequalities imply, for all 𝑥 ◦ ∈ 𝐾,

𝑁
1 𝜕 𝛼 −𝜆𝜙 ( 𝑥 ◦ ) Ö 𝛼 /2 ◦
𝛼
e ≤ 𝐶4| 𝛼 |+1 𝜀 −𝛼/2 |𝛼| − | 𝛼 |/2 max 𝜆, 𝛼 𝑗 𝑗 e−(1−𝜀)𝜆 Re 𝜙 ( 𝑥 ) ,
𝛼! 𝜕𝑥 𝑗=1

whence the claim by the Stirling inequality. □

Corollary 24.B.2 If (24.B.1) holds then to each compact set 𝐾 ⊂ Ω there is a 𝐶 > 0
such that, for all 𝛼 ∈ Z+𝑁 and 𝜆 ≥ |𝛼|,
1 √ 1 1
max 𝜕𝑥𝛼 e−𝜆𝜙 ( 𝑥) ≤ 𝐶 | 𝛼 | 𝜀 − 2 | 𝛼 | 𝛼!𝜆 2 | 𝛼 | e− 2 𝜆 Re 𝜙 ( 𝑥) . (24.B.3)
𝑥 ∈𝐾

Let ℓ ∈ Z, ℓ ≥ 1. We have, for arbitrary 𝛼 ∈ Z+𝑁 ,

ℓ−1
1 𝛼 ℓ ∑︁ ∑︁
𝜕 𝜙 = 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽 𝑄 𝛼,ℓ
𝑝,𝛽 (𝜙) (24.B.4)
𝛼! 𝑥 𝛽⪯𝛼
𝑝=max(0,ℓ− | 𝛼 |)
|𝛽 | ≤ℓ− 𝑝

where 𝛽1 𝛽𝑛
𝜕𝜙 𝛽 𝜕𝜙
(𝜕𝑥 𝜙) = ···
𝜕𝑥 1 𝜕𝑥 𝑛
and 𝑄 𝛼,ℓ
𝑝,𝛽 (𝜙) is a polynomial with rational coefficients, of degree 𝜈 ≥ 0, in the partial
derivatives of 𝜙 of order ≥ 2. Necessarily 𝜈 ≤ 21 |𝛼 − 𝛽| and, since ℓ = 𝑝 + 𝑞 + 𝜈, we
get
1
ℓ − 𝑝 ≤ |𝛼 + 𝛽| . (24.B.5)
2
24.B Appendix: Analytic Estimates of Exponential Amplitudes 1063

For use below we point out the following consequence of (24.B.4):

𝛼,ℓ
𝑄 ℓ− |𝛽 |,𝛽
(𝜙) = 0 if |𝛽| < |𝛼| . (24.B.6)

𝛼,ℓ
Indeed, 𝑄 ℓ−|𝛽 |,𝛽
(𝜙) must be a constant but if |𝛽| < |𝛼| the order of differentiation
in the right-hand side of (24.B.4) is strictly less than that of the left-hand side.

Lemma 24.B.3 Let 𝛼 ∈ Z+𝑁 , ℓ ∈ Z+ be such that ℓ𝛼 ≠ 0, and 𝜙 ∈ C 𝜔 (Ω; C). To

every compact set 𝐾 ⊂ Ω there is a 𝐶𝐾 > 0 such that

| 𝛼 |+ℓ ( 𝑝 + |𝛽|)!
max 𝑄 𝛼,ℓ
𝑝,𝛽 (𝜙) ≤ 𝐶𝐾 (24.B.7)
𝐾 𝑝!𝛽!

for all pairs ( 𝑝, 𝛽) ∈ Z+ × Z+𝑁 such that 𝛽 ⪯ 𝛼, 𝑝 + |𝛽| ≤ ℓ.

𝛼,ℓ 1 𝛼
Proof When ℓ = 1, |𝛼| ≥ 2, we have 𝑄 0,0 (𝜙) ≡ 𝛼! 𝜕𝑥 𝜙; by the analyticity of 𝜙
there is a constant 𝑀𝐾 > 0 such that
1
max 𝜕𝑥𝛼 𝜙 ≤ 𝑀𝐾| 𝛼 |+1 (24.B.8)
𝛼! 𝐾
implying (24.B.7) directly. When ℓ ≥ 1, |𝛼| = 1, we have 𝑝 = ℓ − 1, 𝛽 = 𝛼 and
𝛼,ℓ
𝑄 ℓ−1, 𝛼
(𝜙) = ℓ, which proves (24.B.7) in this case. In the remainder of the proof we
assume ℓ ≥ 2, |𝛼| ≥ 2.
In what follows we denote by ⟨ 𝑗⟩ the 𝑁-tuple with all entries equal to zero except
the 𝑗 th one, equal to 1. The Leibniz rule implies
1 𝛼 ℓ ∑︁ 1 𝛼−𝛽 1 𝛽
𝜕 𝜙 = 𝜕𝑥 𝜙 𝜕𝑥 𝜙ℓ−1
𝛼! 𝑥 𝛽⪯𝛼
(𝛼 − 𝛽)! 𝛽!
1 ∑︁ 𝜕𝜙 1 𝛼−⟨ 𝑗 ⟩ ℓ−1
= 𝜙𝜕𝑥𝛼 𝜙ℓ−1 + 𝜕 𝜙
𝛼! 𝜕𝑥 𝑗 (𝛼 − ⟨ 𝑗⟩)! 𝑥
⟨ 𝑗⟩≺𝛼
∑︁ 1 𝛼−𝛽 1 𝛽
+ 𝜕𝑥 𝜙 𝜕𝑥 𝜙ℓ−1 .
𝛽⪯𝛼
(𝛼 − 𝛽)! 𝛽!
| 𝛼−𝛽 | ≥2

It then follows from (24.B.4) that

1064 24 Analytic PDEs of Principal Type. Regularity of the Solutions

ℓ−1
∑︁ ∑︁
𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽 𝑄 𝛼,ℓ
𝑝,𝛽 (𝜙)
𝑝=max(0,ℓ−| 𝛼 |) 𝛽⪯𝛼
|𝛽 | ≤ℓ− 𝑝
ℓ−2
∑︁ ∑︁
= 𝜙 𝑝+1 (𝜕𝑥 𝜙) 𝛽 𝑄 𝛼,ℓ−1
𝑝,𝛽 (𝜙)
𝑝=max(0,ℓ−1−| 𝛼 |) 𝛽⪯𝛼
|𝛽 | ≤ℓ− 𝑝−1
ℓ−2
∑︁ ∑︁ ∑︁
𝛼−⟨ 𝑗 ⟩,ℓ−1
+ 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽+⟨ 𝑗 ⟩ 𝑄 𝑝,𝛽 (𝜙)
𝑝=max(0,ℓ−1−| 𝛼 |) 𝛽⪯𝛼 ⟨ 𝑗⟩⪯𝛼
|𝛽 | ≤ℓ− 𝑝−1
ℓ−2
∑︁ 1
𝛼−𝛽
∑︁ ∑︁
𝛽,ℓ−1
+ 𝜕𝑥 𝜙 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛾 𝑄 𝑝,𝛾 (𝜙) .
𝛽⪯𝛼
(𝛼 − 𝛽)! 𝛾 ⪯𝛽
𝑝=max(0,ℓ−1−|𝛽 |)
| 𝛼−𝛽 | ≥2 |𝛾 | ≤ℓ−1− 𝑝

We make the natural changes of summation indices. In dealing with the first term in
the right-hand side we use the fact that max (0, ℓ − 1 − 𝑘) + 1 = max (0, ℓ − 𝑘) = 1
if ℓ = 𝑘 + 1. We have
ℓ−1
∑︁ ∑︁
𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽 𝑄 𝛼,ℓ
𝑝,𝛽 (𝜙) (24.B.9)
𝑝=max(0,ℓ−| 𝛼 |) 𝛽⪯𝛼
|𝛽 | ≤ℓ− 𝑝
ℓ−2
∑︁ ∑︁
= 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽 𝑄 𝛼,ℓ−1
𝑝−1,𝛽
(𝜙)
𝑝=max(0,ℓ− | 𝛼 |) 𝛽⪯𝛼
|𝛽 | ≤ℓ− 𝑝
ℓ−2
∑︁ ∑︁ ∑︁
𝛼−⟨ 𝑗 ⟩,ℓ−1
+ 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽 𝑄 𝑝,𝛽− ⟨ 𝑗 ⟩ (𝜙)
𝑝=max(0,ℓ−1−| 𝛼 |) 𝛽⪯𝛼 ⟨ 𝑗 ⟩ ⪯𝛽
|𝛽 | ≤ℓ− 𝑝
ℓ−2
∑︁ ∑︁ ∑︁ 1 𝛾 𝛼−𝛾,ℓ−1
+ 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛽 𝜕 𝜙 𝑄 𝑝,𝛽 (𝜙) .
𝛽⪯𝛼 𝛾 ⪯ 𝛼−𝛽
𝛾! 𝑥
𝑝=max(0,ℓ−| 𝛼 |+1)
|𝛽 | ≤ℓ−1− 𝑝 2≤ |𝛾 | ≤ | 𝛼 |+ 𝑝−ℓ+1

First, we look at the terms in which 𝑝 + |𝛼| = ℓ. It follows from (24.B.5) that this
is equivalent to 𝛽 = 𝛼; the latter implies that the third sum in the right-hand side of
(24.B.9) vanishes identically:
∑︁
𝛼,ℓ 𝛼,ℓ−1 𝛼− ⟨ 𝑗 ⟩,ℓ−1
𝑄 ℓ− | 𝛼 |, 𝛼
(𝜙) = 𝑄 ℓ− | 𝛼 |−1, 𝛼
(𝜙) + 𝑄 ℓ− | 𝛼 |, 𝛼− ⟨ 𝑗 ⟩ (𝜙) .
⟨ 𝑗⟩⪯𝛼
24.B Appendix: Analytic Estimates of Exponential Amplitudes 1065

Induction on ℓ ≥ 2 (recall that |𝛼| ≥ 2) proves the claim in this case, more
precisely,
𝛼,ℓ ℓ!
𝑄 ℓ−| 𝛼 |, 𝛼
(𝜙) = .
𝛼! (ℓ − |𝛼|)!
We look at the terms in (24.B.9) in which 𝑝 + |𝛼| ≤ ℓ − 1, implying ℓ ≥ 3,
𝑝 ≤ ℓ − 3. We have
∑︁
𝛼− ⟨ 𝑗 ⟩,ℓ−1
𝑄 𝛼,ℓ 𝛼,ℓ−1
𝑝,𝛽 (𝜙) = 𝑄 𝑝−1,𝛽 (𝜙) + 𝑄 𝑝,𝛽− ⟨ 𝑗 ⟩ (𝜙)
⟨ 𝑗 ⟩ ⪯𝛽
∑︁ 1 𝛾 𝛼−𝛾,ℓ−1
+ 𝜕 𝜙 𝑄 𝑝,𝛽 (𝜙) ,
𝛾 ⪯ 𝛼−𝛽
𝛾! 𝑥
2≤ |𝛾 | ≤ | 𝛼 |+ 𝑝−ℓ+1

Again we reason by induction on ℓ; (24.B.8) implies

| 𝛼 |+ℓ−1 ( 𝑝 − 1 + |𝛽|)!
max 𝑄 𝛼,ℓ
𝑝,𝛽 (𝜙) ≤ 𝐶𝐾
𝐾 ( 𝑝 − 1)!𝛽!
∑︁ ( 𝑝 − 1 + |𝛽|)! ∑︁ ( 𝑝 + |𝛽|)!
|𝛾 |+1 | 𝛼−𝛾 |+ℓ−1
+𝐶𝐾| 𝛼 |+ℓ−2 + 𝑀𝐾 𝐶𝐾
𝑝! (𝛽 − ⟨ 𝑗⟩)! 𝛾 ⪯ 𝛼−𝛽
𝑝!𝛽!
⟨ 𝑗 ⟩ ⪯𝛽
2≤ |𝛾 | ≤ | 𝛼 |+ 𝑝−ℓ+1

( 𝑝 + |𝛽|)! © 𝑝 −1 |𝛽|
∑︁
≤ 𝐶𝐾| 𝛼 |+ℓ−1 + 𝐶𝐾 + 𝑀𝐾 (𝑀𝐾 /𝐶𝐾 ) |𝛾 | ® .
ª
𝑝 + |𝛽| 𝑝 + |𝛽|
­
𝑝!𝛽!
« |𝛾 | ≥2 ¬
It is clear that the last inequality implies (24.B.7) if 𝐶𝐾 ≫ max (1, 𝑀𝐾 ). □

Lemma 24.B.4 Suppose Condition (24.B.1) is satisfied. Then, to each compact set
𝐾 ⊂ Ω there are positive numbers 𝜀 < 1 and 𝐶 such that, for all 𝛼 ∈ Z+𝑁 , ℓ ∈ Z+
and 𝜆 > |𝛼|,
1 1 √
max 𝜕𝑥𝛼 𝜕𝜆ℓ e−𝜆𝜙 ( 𝑥) ≤ 𝐶 | 𝛼 |+ℓ 𝜀 − 2 ( | 𝛼 |+ℓ) 𝜆 2 ( | 𝛼 |−ℓ) ℓ!𝛼!e−(1−𝜀)𝜆 Re 𝜙 . (24.B.10)
𝑥 ∈𝐾

Proof We apply the Leibniz rule and (24.B.4):

1 𝛼 ℓ −𝜆𝜙 1
𝜕𝑥 𝜕𝜆 e = 𝜕𝑥𝛼 𝜙ℓ e−𝜆𝜙
𝛼! 𝛼!
∑︁ 1 1
𝜕𝑥 e−𝜆𝜙
𝛽 ℓ 𝛼−𝛽
= 𝜕𝑥 𝜙
𝛽⪯𝛼
𝛽! (𝛼 − 𝛽)!
ℓ−1
∑︁ ∑︁ ∑︁
𝛽,ℓ 1 𝛼−𝛽 −𝜆𝜙
= 𝜙 𝑝 (𝜕𝑥 𝜙) 𝛾 𝑄 𝑝,𝛾 (𝜙) 𝜕 e .
𝛽 ⪯ 𝛼 𝑝=max(0,ℓ− |𝛽 |) 𝛾 ⪯𝛽
(𝛼 − 𝛽)! 𝑥
|𝛾 | ≤ℓ− 𝑝

From (24.B.2) and (24.B.7) we derive, for 𝜆 ≥ |𝛼|,

1066 24 Analytic PDEs of Principal Type. Regularity of the Solutions

1 𝛼 ℓ −𝜆𝜙
𝜕 𝜕 e
𝛼! 𝑥 𝜆
ℓ−1
∑︁ ∑︁ ∑︁ 1 1 ( 𝑝 + |𝛾|)!
≤ 𝐶 | 𝛼 |+ℓ 𝜀 − 2 | 𝛼−𝛽 | √︁
𝛽 ⪯ 𝛼 𝑝=max(0,ℓ− |𝛽 |) 𝛾 ⪯𝛽 (𝛼 − 𝛽)! 𝑝!𝛾!
|𝛾 | ≤ℓ− 𝑝
1
×𝜆 2 | 𝛼−𝛽 | |𝜙| 𝑝 |(𝜕𝑥 𝜙) 𝛾 | e−(1−𝜀)𝜆 Re 𝜙 ( 𝑥) .

Once again we avail ourselves of (24.B.1):

1
|𝜙| 𝑝 |(𝜕𝑥 𝜙) 𝛾 | e−𝜀𝜆 Re 𝜙 ( 𝑥) ≤ (Re 𝜙) 2 ( 𝑝+|𝛾 |) e−𝜀𝜆 Re 𝜙
√︁ 1
≤ ( 𝑝 + |𝛾|)! (𝜀𝜆) − 2 ( 𝑝+|𝛾 |) .

If 𝑝 ≥ ℓ − |𝛽| then |𝛼 − 𝛽| − ( 𝑝 + |𝛾|) ≥ |𝛼| − ℓ − |𝛾| ≥ |𝛼 − 𝛽| − ℓ (since 𝛾 ⪯ 𝛽)

whence

𝐶 −| 𝛼 |−ℓ e (1−2𝜀)𝜆 Re 𝜙 𝜕𝑥𝛼 𝜕𝜆ℓ e−𝜆𝜙

√︄
ℓ−1
∑︁ ∑︁ ∑︁
− 21 | 𝛼−𝛽+𝛾 |− 12 𝑝 1
( | 𝛼−𝛽 |−ℓ) ( 𝑝 + |𝛾|)! ( 𝑝 + |𝛾|)!
≤ 𝜀 𝜆 2 𝛼!
𝛽 ⪯ 𝛼 𝑝=max(0,ℓ− |𝛽 |) 𝛾 ⪯𝛽
(𝛼 − 𝛽)! 𝑝!𝛾!
|𝛾 | ≤ℓ− 𝑝
√︄ !
1 √ 𝛼! 1
𝐶1| 𝛼 |+ℓ 𝜀 − 2 ( | 𝛼 |+ℓ)
√︁
≤ ℓ!𝛼! 𝜆− 2 |𝛽 | 𝛽!.
(𝛼 − 𝛽)!𝛽!

If 𝜆 ≥ |𝛼| ≥ |𝛽| the Stirling inequality and a redefinition of 𝜀 imply (24.B.10). □

Chapter 25
Solvability of Constant Vector Fields of Type
(1,0)

This chapter may seem a departure from the road to our understanding of the
solvability of analytic pseudodifferential equations of principal type so far followed
in this Part of the book, but in fact it is a détour. The main reason for its insertion
at this late juncture is that its main result (cf. Theorems 25.3.17, 25.3.18) is the
foundation of the original proof of the Trépreau theorem (see [Trépreau, 1984],
and [Ancona, 1994], [Trépreau, 1992], for the auxiliary results here found in the
Appendix). This proof will be outlined in the next chapter. Here, however, we deal
solely with holomorphic functions and vector fields of type (1,0), mainly in domains
in C𝑛 . Thus solvability, be it local or global, is always to be understood in the
holomorphic category. In C𝑛 , vector fields of type (1,0) are linear combinations of
𝜕/𝜕𝑧1 , ..., 𝜕/𝜕𝑧 𝑛 with holomorphic coefficients. In a complex-analytic manifold M
vector fields of type (1,0) have such an expression in every complex-analytic local
chart (see Subsection 9.4.5).
With the exception of the last section (on differential complexes) the contents of
this chapter are almost entirely extracted from the monograph [Hörmander, 1994]
(which contains much more material than presented or alluded to here). We have
departed in one item of terminology: what is called 𝑍-convexity in [Hörmander,
1994] is called 𝑍-pseudoconvexity here, 𝑍 denoting a complex vector field of
type (1,0) with constant coefficients. There are two reasons for this change: Firstly,
the invariant definition of 𝑍-convexity of the boundary of a domain Ω in C𝑛 in
[Hörmander, 1994] (Definition 7.1.7, p. 363; Definition 25.2.19 in this text) is based
on the use of a generalized Levi form, not unlike pseudoconvexity in the usual sense
(except for its link to the vector field 𝑍, of course). Secondly, there is a useful
definition of 𝑍-convexity which adapts to 𝑍 the concept of C-convexity (Definition
25.1.1 in this text) and is relevant in one of the formulations of the main result alluded
to above.
The characterization of solvability of a single vector field at a point 𝑧◦ of the
boundary (Definition 25.2.1) of a strictly pseudoconvex domain Ω leads to the
question of the local exactness at 𝑧◦ (in the analogous sense: local exactness on
the Ω-side of the boundary, in the complex-analytic category) of the differential
complex defined by an involutive system of vector fields of type (1,0). Theorem

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1067
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_25
1068 25 Solvability of Constant Vector Fields of Type (1,0)

25.4.9 characterizes the local exactness in every degree (a feature of the De Rham and
the Dolbeault complexes) but not in a given single degree while possibly not valid in a
different degree (as happens in CR complexes). This is an important open question,
since its answer could lead to a characterization of the microlocal exactness of
differential complexes defined by an involutive system of analytic pseudodifferential
operators of principal type and, thence, to at least a conjecture on the local exactness
of all such systems, even in the C ∞ class.

25.1 C-Convexity and Global Solvability

25.1.1 C-convex domains in C𝒏

Throughout this section Ω will denote an open subset of C𝑛 , Ω its closure, 𝜕Ω = Ω\Ω
its boundary. We recall that an affine subspace of a vector space is the translate of a
vector subspace and an affine transformation is a linear transformation followed by
a translation. By an affine complex line 𝐿 in C𝑛 we mean a subset 𝐿 ◦ + 𝑧◦ of C𝑛 with
𝐿 ◦ a complex linear subspace of C𝑛 , dimC 𝐿 ◦ = 1.

Definition 25.1.1 The open subset Ω of C𝑛 is said to be C-convex if, given arbitrarily
an affine complex line 𝐿 in C𝑛 , Ω ∩ 𝐿 is either connected and simply connected in
𝐿 or else empty.

Proposition 25.1.2 A C-convex open subset Ω of C𝑛 is connected and simply con-

nected.

Proof That Ω is connected is evident: if 𝑧1 , 𝑧2 ∈ Ω let 𝐿 be the affine complex line

containing 𝑧 1 and 𝑧2 ; 𝑧1 and 𝑧 2 belong to the connected set Ω ∩ 𝐿 (open in 𝐿). We
must prove that Ω is simply connected.
Let 𝔠 denote a loop in Ω, by which we mean the range of a continuous map
[0, 1] ∋ 𝑡 ↦→ 𝔠 (𝑡) ∈ Ω such that 𝔠 (𝑡) ≠ 𝔠 (𝑡 ′) if 0 ≤ 𝑡 < 𝑡 ′ < 1 and 𝔠 (1) = 𝔠 (0).
If 𝑧, 𝑧 ′ ∈ C𝑛 , 𝑧 ≠ 𝑧 ′, we write [𝑧, 𝑧 ′] = {𝜃𝑧 + (1 − 𝜃) 𝑧 ′; 0 ≤ 𝜃 ≤ 1}, the segment
joining 𝑧 and 𝑧 ′. There is an 𝜀 > 0 such that if |𝑡 − 𝑡 ′ | < 𝜀, 0 ≤ 𝑡 < 𝑡 ′ ≤ 1, then the
following is true:
(1) [𝔠 (𝑡) , 𝔠 (𝑡 ′)] ⊂ Ω;
(2) there is a continuous deformation in Ω from [𝔠 (𝑡) , 𝔠 (𝑡 ′)] to the arc in 𝔠 joining
𝔠 (𝑡) to 𝔠 (𝑡 ′), i.e., the range of the map [𝑡, 𝑡 ′] ∋ 𝑠 ↦→ 𝔠 (𝑠).
We select a finite set of points 0 < 𝑡1 < · · · < 𝑡 𝑗 < 𝑡 𝑗+1 < · · · < 𝑡 𝑁 < 1
such that 𝑡1 < 𝜀, 𝑡 𝑗+1 < 𝑡 𝑗 + 𝜀, 1 − 𝜀 < 𝑡 𝑁 (𝑁 ≥ 2). For each 𝑗 = 0, 1, ..., 𝑁
we can continuously deform within Ω the range of the map 𝑡 𝑗 , 𝑡 𝑗+1 ∋ 𝑠 ↦→ 𝔠 (𝑠)

to the segment 𝔠 𝑡 𝑗 , 𝔠 𝑡 𝑗+1 (with 𝑡0 = 0, 𝑡 𝑁 +1 = 1). Thus 𝔠 is homotopically
equivalent in Ω to the piecewise linear loop 𝔠♭ with vertices 𝔠 (0), 𝔠 (𝑡1 )..., 𝔠 (𝑡 𝑁 ).
If 𝛿 > 0 is sufficiently small then, given any (℘0 , ℘1 , ..., ℘ 𝑁 ) ∈ Ω 𝑁 +1 such that
25.1 C-Convexity and Global Solvability 1069

max ℘𝑗 − 𝔠 𝑡𝑗 < 𝛿, the following is true: ℘ 𝑗 ≠ ℘ 𝑘 if 𝑗 ≠ 𝑘 and the piecewise
𝑗=0,1,..., 𝑁
linear closed curve 𝔠 ♮ with successive vertices ℘ 𝑗 is contained in Ω and has no self-
intersections (thus it is a loop). Provided 𝛿 is sufficiently small, 𝔠 ♮ is homotopically
equivalent to 𝔠♭ and therefore to 𝔠 in Ω.
Now let Ω∗𝑁 +1 be the subset of Ω 𝑁 +1 consisting of the elements (℘0 , ℘1 , ..., ℘ 𝑁 )
such that ℘ 𝑗 ≠ ℘ 𝑘 if 𝑗 ≠ 𝑘. The complement Ω 𝑁 +1 \Ω∗𝑁 +1 is a complex-analytic
𝑁
Ö −1
2
subvariety of Ω 𝑁 +1 defined by the equation ℘ 𝑗 − ℘ 𝑗+1 = 0. By the Riemann
𝑗=0
Extension Theorem 14.3.7, it follows that Ω∗𝑁 +1 is open, connected and dense in
Ω 𝑁 +1 . Let then Ω∗∗ 𝑁 +1 be the (open) subset of Ω 𝑁 +1 consisting of the elements
∗
(℘0 , ℘1 , ..., ℘ 𝑁 ) such that e 𝑗 = ℘ 𝑗+1 − ℘ 𝑗 and e 𝑘 = ℘ 𝑘+1 − ℘ 𝑘 (with ℘ 𝑁 +1 = ℘0 ) are
C-linearly independent if 𝑗 ≠ 𝑘. The complement Ω∗𝑁 +1 \Ω∗∗ 𝑁 +1 is a complex-analytic
+1 1
subvariety of Ω∗ defined by the 2 𝑁 (𝑁 + 1) equations e 𝑗 ∧e 𝑘 = 0, 0 ≤ 𝑗 < 𝑘 ≤ 𝑁.
𝑁

Applying once again Theorem 14.3.7, we derive that Ω∗∗ 𝑁 +1 is open, connected and
+1
dense in Ω∗ and therefore in Ω
𝑁 𝑁 +1 . Let (℘0 , ℘1 , ..., ℘ 𝑁 ) ∈ Ω∗∗𝑁 +1 be arbitrary and

𝐿 𝑗 ( 𝑗 = 0, 1, ..., 𝑁) denote the complex affine line containing the segment ℘ 𝑗 , ℘ 𝑗+1 .
There are continuous curves without self-intersections 𝛾 𝑗 in the simply connected
open subset Ω ∩ 𝐿 𝑗 of 𝐿 𝑗 joining ℘ 𝑗 to℘ 𝑗+1 ; since 𝐿 𝑗 ≠ 𝐿 𝑘 if 𝑗 < 𝑘 the union of
these curves 𝛾0 , 𝛾1 ., ..., 𝛾 𝑁 is also without self-intersections, i.e., it is a loop in Ω.
It defines a homotopy class and it is not difficult to show that this class is constant
in a full neighborhood of (℘0 , ℘1 , ..., ℘ 𝑁 ) in Ω∗∗ 𝑁 +1 ; since Ω 𝑁 +1 is connected all its
∗∗
elements define a single homotopy class. By density this is also the class of the loops
𝔠♭ and 𝔠 of the beginning of the proof. Since there is a (℘0 , ℘1 , ..., ℘ 𝑁 ) ∈ Ω∗∗ 𝑁 +1

whose components all belong to an open ball contained in Ω, this homotopy class is
0. □

Proposition 25.1.3 The preimage of a C-convex domain Ω under an affine complex

map 𝜑 : C 𝑝 −→ C𝑛 and the image of Ω under a surjective affine complex map
𝜓 : C𝑛 −→ C𝑞 are C-convex.
−1
Proof If Λ ⊂ C 𝑝 is an affine complex line intersecting 𝜑 (Ω) then either 𝜑 (Λ) is
−1
a single point 𝑧◦ ∈ Ω, in which case Λ ⊂ 𝜑 (Ω), or else 𝜑| Λ is a bijection onto
−1
the affine complex line 𝜑 (Λ) ⊂ C𝑛 , in which case Λ ∩ 𝜑 (Ω) is homeomorphic to
𝜑 (Λ) ∩ Ω, which is connected and simply connected.
If 𝜓 is surjective 𝜓 (Ω) is open and connected. Let Λ be an affine complex line
intersecting 𝜓 (Ω); we claim that Λ ∩ 𝜓 (Ω) is connected and simply connected. To
−1 −1
prove this we may as well replace Ω by 𝜓 (Λ) ∩ Ω. Indeed, if 𝐿 ⊂ 𝜓 (Λ) is an affine
complex line in C𝑛 then 𝐿 ∩ Ω is connected and simply connected; in other words,
−1 −1
𝜓 (Λ) ∩ Ω is C-convex in the affine complex space 𝜓 (Λ) (of dimension 𝑛 − 𝑞 + 1).
In other words, we can and shall assume 𝑞 = 1.
1070 25 Solvability of Constant Vector Fields of Type (1,0)

Consider aloop [0, 1] ∋ 𝑡 ↦→ 𝔠 (𝑡) in 𝜓 (Ω) ⊂ C. We construct a loop b 𝔠 ⊂ C𝑛

such that 𝜓 b 𝔠 = 𝔠. Let 𝑁 ∈ Z+ be sufficiently large that, for each 𝑗 = 0, 1, ..., 𝑁,
h i −1
the arc of curve 𝔠 𝑗 , the range of 𝑁 𝑗+1 , 𝑁𝑗+1
+1 ∋ 𝑡 ↦→ 𝔠 (𝑡), can be lifted under 𝜓

𝔠 𝑗 ⊂ Ω with end points ℘ 𝑗 , ℘′𝑗 such that 𝜓 ℘ 𝑗 = 𝔠 𝑁 𝑗+1 ,

as a continuous curve b

𝜓 ℘′𝑗 = 𝔠 𝑁𝑗+1

+1 = 𝜓 ℘ 𝑗+1 (𝜓 induces a homeomorphism of 𝔠 𝑗 onto 𝔠 𝑗 ). If
b
℘ 𝑗 ≠ ℘ 𝑗+1 we insert a continuous curve [0, 1] ∋ 𝑡 ↦→ 𝛾 𝑗 (𝑡) ∈ Ω joining ℘′𝑗
′

to ℘ 𝑗+1 in the affine complex line 𝐿 𝑗 passing through ℘′𝑗 and ℘ 𝑗+1 . Note that
n o
𝜓 𝐿 𝑗 = 𝔠 𝑁𝑗+1 +1 , implying 𝐿 𝑗 ≠ 𝐿 𝑘 if 𝑗 ≠ 𝑘 since 𝔠 does not have any self-
intersection. By inserting in ordered fashion the curves 𝛾 𝑗 between the arcs b 𝔠 𝑗 we
−1
𝔠 from ℘0 back to ℘0 in 𝜓 (Λ) that does not have self-
obtain a closed curve b
intersections and 𝜓 b
𝔠 = 𝔠. By Proposition 25.1.2 we can contract b
𝔠 to a point in
−1
𝜓 (Λ), implying that 𝔠 can be contracted to a point in Λ. This proves the claim. □
We now relate C-convexity to other notions of convexity.

Definition 25.1.4 The domain Ω ⊂ C𝑛 is said to be:

linearly convex if every point in C𝑛 \Ω lies in an affine complex hyperplane
contained in C𝑛 \Ω;
weakly linearly convex if every point in 𝜕Ω lies in an affine complex hyperplane
contained in C𝑛 \Ω.

Obviously, linear convexity entails weak linear convexity. An open subset 𝑈 of a

real vector space 𝑽 is convex if every point in 𝑽\𝑈 lies in a real affine hyperplane
𝑯 ⊂ 𝑽\𝑈. Since every real affine hyperplane 𝑯 ⊂ C𝑛 \Ω contains an affine complex
hyperplane, every convex domain in C𝑛 is linearly convex. Intersections of linearly
convex (resp., weakly linearly convex) sets are linearly convex (resp., weakly linearly
convex).

Proposition 25.1.5 A weakly linearly convex domain Ω ⊂ C𝑛 is pseudoconvex (Def-

inition 11.2.15).

Proof We show that an arbitrary compact subset 𝐾 of Ω is contained in a plurisub-

harmonic compact subset (Definition 11.2.14) of Ω. If Π is a bounded convex domain
containing 𝐾 then every connected component of Ω ∩ Π is weakly linearly convex;
there is no loss of generality in assuming that 𝐾 is contained in one of these com-
ponents or, more simply, that Ω is bounded. If 𝑧 ◦ ∈ 𝜕Ω there is an affine complex
hyperplane 𝑯 such that 𝑧 ◦ ∈ 𝑯 ⊂ C𝑛 \Ω. Let 𝜆 be a complex linear functional on
C𝑛 such that 𝑯 = {𝑧 ∈ C𝑛 ; 𝜆 (𝑧) = 𝜆 (𝑧 ◦ )}; then 𝑓 (𝑧) = − log |𝜆 (𝑧) − 𝜆 (𝑧◦ )| is
plurisubharmonic and continuous in Ω. Since 𝑓 (𝑧) → +∞ as 𝑧 ∈ Ω tends to 𝑧◦ there
is a neighborhood 𝑈𝑧 ◦ of 𝑧 ◦ such that 𝑓 (𝑧) > sup 𝑓 whatever 𝑧 ∈ 𝑈𝑧 ◦ . Since 𝜕Ω is
𝐾
compact there are finitely many continuous plurisubharmonic functions 𝑓1 , ..., 𝑓𝑟 in
Ω such that
25.1 C-Convexity and Global Solvability 1071

𝑧 ∈ Ω; 𝑓 𝑗 (𝑧) ≤ sup 𝑓 𝑗 , 𝑗 = 1, ..., 𝑟
𝐾

is a compact subset of Ω, obviously plurisubharmonic convex in Ω. □

Below we make use of the one-dimensional complex projective space CP1 . For
arbitrary 𝑛 = 1, 2, ..., the 𝑛-dimensional complex projective space CP𝑛 is the quotient
of C𝑛+1 \ {0} under the action 𝑧 ↦→ 𝜆𝑧, 0 ≠ 𝜆 ∈ C or, equivalently, the quotient of
S2𝑛+1 under the equivalence relation 𝑧 ≈ 𝑧 ′, meaning that there is a 𝜃 ∈ [0, 2𝜋]
such that 𝑧 ′ = e𝑖 𝜃 𝑧; CP𝑛 inherits from C𝑛+1 \ {0} the structure of an 𝑛-dimensional,
compact complex-analytic manifold (thus CP𝑛 is orientable, in contrast to the real
projective spaces). Let 𝜋 (𝑛) : C𝑛+1 −→ CP𝑛 be the natural projection and (𝑧1 , ..., 𝑧 𝑛+1 )
be the coordinates in C𝑛+1 ; the 𝑧 𝑗 are often called homogeneous coordinates in CP𝑛 .
If 𝐿 is a complex line in C𝑛+1 on which 𝑧 𝑗 ≠ 0 off the origin for some 𝑗, then

𝑧 1 /𝑧 𝑗 , ..., 𝑧 𝑗−1 /𝑧 𝑗 , 𝑧 𝑗+1 /𝑧 𝑗 , ..., 𝑧 𝑛 /𝑧 𝑗 (25.1.1)

are true local complex-analytic coordinates in a neighborhood of 𝜋 (𝑛) (𝐿) in CP𝑛 .

With ℘ ∈ CP𝑛 arbitrary we can identify C𝑛 with the open subset CP𝑛 \ {℘} of CP𝑛
(℘ can then be regarded as the point at infinity). Using coordinates 𝑧1 , ..., 𝑧 𝑛+1 in
C𝑛+1 we can take ∞ to be the image of the 𝑧 𝑛+1 -axis under 𝜋 (𝑛) and identify C𝑛 with
the points (𝑧1 , ..., 𝑧 𝑛 , 1) ∈ CP𝑛 .
The set of complex lines through 0 in C𝑛+1 can be regarded as a complex line
bundle over CP𝑛 , the canonical bundle of CP𝑛 denoted here by Can CP𝑛 : the
◦ ◦ ◦

fiber at 𝑧 = 𝑧1 , ..., 𝑧 𝑛+1 (in the homogeneous coordinates) is the complex line
through 𝑧◦1 , ..., 𝑧 ◦𝑛+1 . We denote by 𝑝 : Can CP𝑛 −→ CP𝑛 the base projection.
Let 𝑧◦ ∈ C𝑛+1 , 𝑧◦ ≠ 0, be arbitrary and 𝐿 𝑧 ◦ the complex ◦
line spanned by 𝑧 . If
𝐿 ⊥𝑧 ◦ is the orthogonal hyperplane in C𝑛+1 and 𝔅 𝛿 (𝑧 ◦ ) = 𝑧 ∈ C𝑛+1 ; |𝑧 − 𝑧◦ | < 𝛿 ,
𝛿 ≤ |𝑧 ◦ |, then 𝑝 induces a biholomorphism of 𝔅 𝛿 (𝑧◦ ) ∩ 𝐿 ⊥𝑧 ◦ onto a neighborhood U
of 𝑝 (𝐿 𝑧 ◦ ); the inverse mapping is a nowhere vanishing complex-analytic section s
of Can CP𝑛 over U. We get local coordinates in Can CP𝑛 | U by taking coordinates
of the kind (25.1.1) in the base U and the coordinate 𝑐s (𝜁) −→ 𝑐 ∈ C (𝜁 ∈ U) in
the fiber.

Proposition 25.1.6 If the domain Ω is C-convex then Ω is linearly convex.

Proof Let 𝑧◦ be an arbitrary point of C𝑛 \Ω; we must show that there is an affine
complex hyperplane containing 𝑧◦ and not intersecting Ω; we may as well take 𝑧◦ = 0,
in which case we must show that there is a complex hyperplane (∋ 0) that does not
intersect Ω. Suppose the claim proved for 𝑛 = 2. Let 𝑬 be a complex linear subspace
of C𝑛 , dimC 𝑬 = 2, such that 𝑬 ∩ Ω ≠ ∅; then 𝑬 ∩ Ω is C-convex in 𝑬. It follows that
there is a complex line (through 0) 𝐿 ⊂ 𝑬 such that 𝐿 ∩ Ω = ∅. If 𝜋 : C𝑛 −→ 𝐿 ⊥
is the orthogonal projection then the image 𝜋 (Ω) is C-convex in 𝐿 ⊥ (by Proposition
25.1.3). Induction on 𝑛 allows us to select a complex hyperplane 𝑯 ′ in 𝐿 ⊤ such that
𝑯 ′ ∩ 𝜋 (Ω) = ∅. The orthogonal sum 𝑯 ′ ⊕ 𝐿 is a complex hyperplane in C𝑛 which
does not intersect Ω.
1072 25 Solvability of Constant Vector Fields of Type (1,0)

It remains to prove the claim when 𝑛 = 2. We start from the hypothesis that
𝑝 (Ω) = CP1 and show that this leads to a contradiction. To do this we define a
smooth, real-valued argument function 𝐴 as follows. An arbitrary point 𝜁 ◦ ∈ CP1 has
a neighborhood U with the following property: there is a complex-analytic section
𝜁 ↦→ 𝑧 (𝜁) of Can CP1 over U such that 𝑧 (𝜁) ∈ Ω for every 𝜁 ∈ U. Set 𝐿 𝜁 = C𝑧 (𝜁),
the complex line spanned by 𝑧 (𝜁); 𝐿 𝜁 has the orientation inherited from C2 . Since
𝐿 𝜁 ∩ Ω is connected and simply connected we can define the argument function 𝐴
in 𝐿 𝜁 ∩ Ω by setting it equal to zero at 𝑧 (𝜁); 𝐴 maps 𝐿 𝜁 ∩ Ω onto an interval (𝑎, 𝑏),
−𝜋 < 𝑎 < 0 < 𝑏 < 𝜋. The latter fact allows us to use a C ∞ partition of unity in CP1 , to
patch together these local definitions, thereby defining 𝐴 ∈ C ∞ (Ω; R). It is important
to note that this patching may modify the value of 𝐴 (𝑧 (𝜁)) if 𝜁 ∈ U and 𝑧 (𝜁) is the
section of Can CP1 at the beginning of this paragraph. Nonetheless,
to each 𝜁 ∈ CP
1

there is a unique ray 𝜆𝑧♭ (𝜁), 𝑧♭ (𝜁) ∈ Ω, 𝜆 > 0, such that 𝐴 𝑧♭ (𝜁) = 0. Applying
the Implicit Function Theorem to this equation shows that 𝑧♭ (𝜁) is a C ∞ function
in CP1 , nowhere zero since 0 ∉ Ω. It follows that 𝜁 ↦→ 𝑧♭ (𝜁) / 𝑧♭ (𝜁) is a smooth
section of Can CP1 over CP1 , contradicting Lemma 25.1.9 below. □
Corollary 25.1.7 A C-convex domain is weakly linearly convex.
Corollary 25.1.8 A C-convex domain is pseudoconvex.
Proof Combine Corollary 25.1.7 and Proposition 25.1.5 □
A partial converse of Corollary 25.1.7 is true. We shall now focus our attention
on domains Ω in C𝑛 whose boundary is of class C 𝑚 for 𝑚 = 1, 2, by which we mean
that an arbitrary point 𝑧 ◦ ∈ 𝜕Ω has a neighborhood 𝑈 in C𝑛 such that there is a
function 𝜌 ∈ C 𝑚 (𝑈) defining Ω in 𝑈, i.e., Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} and d𝜌 ≠ 0
at every point of 𝜕Ω ∩ 𝑈 (or of 𝑈). We need the following
Lemma 25.1.9 Suppose that the domain Ω is bounded and that its boundary is of
class C 1 . If Ω is locally weakly linearly convex then every affine complex line 𝐿 such
that 𝐿 ∩ Ω ≠ ∅ intersects 𝜕Ω transversally at every point of 𝐿 ∩ 𝜕Ω.
That Ω is locally weakly linearly convex means that every point 𝑧◦ ∈ 𝜕Ω has
a neighborhood 𝑈 in C𝑛 such that every point of 𝑈 ∩ 𝜕Ω lies in an affine complex
hyperplane contained in C𝑛 \ (𝑈 ∩ Ω). That 𝐿 intersects 𝜕Ω at 𝑧◦ transversally
means that there is a real (affine) line through 𝑧◦ contained in 𝐿 and transversal to
𝜕Ω.
Proof Suppose 0 ∈ 𝜕Ω and let 𝑯 R
√ denote the real hyperplane tangent to 𝜕Ω at 0.
Our hypothesis is that 𝑯 = 𝑯 ∩ −1𝑯 R is the unique complex hyperplane with the
R

following property: there is a neighborhood 𝑈 of 0 in C𝑛 such that 𝑯∩Ω∩𝑈 = ∅. Let

𝐿 be a complex line intersecting 𝜕Ω at 0; either 𝐿 ⊂ 𝑯 and therefore 𝐿 ∩ Ω ∩ 𝑈 = ∅
or else 𝐿 intersects 𝜕Ω transversally at 0. □
Proposition 25.1.10 Suppose that the domain Ω is bounded and that its boundary
is of class C 1 . If Ω is locally weakly linearly convex (Definition 25.1.4) then Ω is
C-convex.
25.1 C-Convexity and Global Solvability 1073

Proof Let 𝐿 ∋ 0 such that 𝐿 ∩Ω ≠ ∅. Since Ω is connected we can join two arbitrary
points 𝑧◦ , 𝑧♭ ∈ 𝐿 ∩ Ω by a continuous curve [0, 1] ∋ 𝑡 ↦→ 𝑧 (𝑡) ∈ Ω, 𝑧 (0) = 𝑧 ◦ ,
𝑧 (1) = 𝑧♭ . The affine complex line through 𝑧◦ and 𝑧 (𝑡) is
n o
𝐿 𝑡 = 𝑧 ∈ C𝑛 ; ∃𝜁 ∈ C, 𝑧 = 𝑧◦ + 𝜁 𝑧 (𝑡) − 𝑧♭ .

Lemma 25.1.9 implies that 𝐿 𝑡 ∩ Ω has finitely many connected components and that
each one is bounded and has a C 1 boundary in 𝐿 𝑡 . The set of all 𝑡 such that 𝑧 ◦ and
𝑧 (𝑡) are in the same component of 𝐿 𝑡 ∩ Ω is open and so is the set of all 𝑡 such that
𝑧◦ and 𝑧 (𝑡) are in different components of 𝐿 𝑡 ∩ Ω; one of these two sets must be
empty. Since for small 𝑡 > 0, 𝑧◦ and 𝑧 (𝑡) are in the same component of 𝐿 𝑡 ∩ Ω the
same is true for all 𝑡. This proves that 𝐿 ∩ Ω ≠ ∅ is connected.
The set of affine complex lines in C𝑛 parallel to a given complex line 𝐿 ◦ ∋ 0 is in
one-to-one correspondence with the points in the complex hyperplane 𝐿 ◦⊥ orthogonal
to 𝐿, therefore with the points in C𝑛−1 ; 𝐿 ◦ itself can be identified with a point in
PC𝑛−1 . Thus the set of affine complex lines in C𝑛 can be identified to C𝑛−1 × PC𝑛−1 ,
a complex-analytic manifold M of dimension 2𝑛 − 2. It is clear that the subset MΩ
of M whose elements are the affine complex lines 𝐿 such that 𝐿 ∩ Ω ≠ ∅ is open and
connected. As 𝐿 varies in MΩ the number of connected components of 𝐿 ∩ C𝑛 \Ω
is locally constant, due to the transversality discussed at the beginning; it is therefore
constant throughout MΩ . Let then 𝔅 be an open ball containing Ω and such that
there is a 𝑧 ◦ ∈ 𝔅 ∩ 𝜕Ω. It is easily seen (after suitably choosing the coordinates in
C𝑛 in a neighborhood of 𝑧◦ ) that there is an affine complex line 𝐿 in C𝑛 such that
𝐿 ∩ Ω is simply connected, therefore the same is true of all 𝐿 ∈ MΩ . This completes
the proof that Ω is C-convex. □
We can now state

Theorem 25.1.11 For a bounded domain Ω in C𝑛 with a C 1 boundary, C-convexity,

linear convexity, weak linear convexity, and local weak linear convexity are equiva-
lent properties. If they hold Ω is pseudoconvex.

Proof The equivalence of the four kinds of convexity when 𝜕Ω is of class C 1 and the
ensuing pseudoconvexity follow directly from Propositions 25.1.5, 25.1.6, 25.1.10
and Corollary 25.1.7. □

Corollary 25.1.12 Suppose that the boundary 𝜕Ω of the bounded domain Ω is

of class C 2 . If the equivalent convexity properties in Theorem 25.1.11 hold then
there is a neighborhood 𝑈 of 𝜕Ω in C𝑛 and a function 𝜌 ∈ C 2 (𝑈; R) such that
𝜕Ω = {𝑧 ∈ 𝑈; 𝜌 (𝑧) = 0}, d𝜌 ≠ 0 and the Levi form of 𝜌 is positive semidefinite at
every point of 𝜕Ω.

Proof Follows from the pseudoconvexity of Ω and Proposition 11.2.19. □

1074 25 Solvability of Constant Vector Fields of Type (1,0)

25.1.2 A pseudoconvex domain 𝛀 that is not C-convex

Without the C 1 boundary hypothesis the claim in Proposition 25.1.10 is not valid. The
classes of C-convex, weakly linearly convex, locally weakly linearly convex domains
are all different. It follows that there are pseudoconvex domains that do not have the
convexity properties in Theorem 25.1.11. We describe the example in [Hörmander,
1994], pp. 297–298, of a weakly linearly convex (hence pseudoconvex, Corollary
25.1.8) domain that is not linearly convex (hence not C-convex, Proposition 25.1.6).
We seek a bounded domain Ω in C2 with the following properties:

(1) the boundary 𝜕Ω is Lipschitz continuous but not C 1 ;

(2) every complex line 𝐿 intersecting 𝜕Ω is contained in C2 \Ω;
(3) there is an open subset 𝑈 of C2 \Ω such that every affine complex line in C2
intersecting 𝑈 also intersects Ω.
The construction of Ω is based on that of two bounded disjoint compact subsets
𝐾, 𝐾 ′, of a real hyperplane 𝐻 in C2 . Our first (crucial) requirement is that every
(affine) real line 𝐿 in 𝐻 intersecting 𝐾 ′ must intersect 𝐾. Our second important
requirement is that 𝐾 be the closure of Ω ∩ 𝐻 and Ω ∩ 𝐻 be a domain in 𝐻 whose
boundary in 𝐻 is Lipschitz continuous but not C 1 (whence (1) above).
Note that the complex lines Π in C2 that do not intersect 𝐾 are of one of three
types: Π ∩ 𝐻 = ∅ (i.e., Π is parallel to 𝐻); Π ∩ 𝐻 is a single real line not intersecting
𝐾; Π is the union of real lines in 𝐻 that do not intersect 𝐾. Suppose 𝐾 = Ω ∩ 𝐻
and 𝐾 ′ are defined. Let 𝐹 be the union of all the affine complex lines in C2 that
do not intersect Ω; 𝐹 is closed in C2 . Provided dist (𝐾, 𝜕Ω) is sufficiently small, Ω
is a connected component of C2 \𝐹 and C2 \𝐹 has a connected component Ω′ ≠ Ω
(Ω′ intersects 𝐾 ′). On the one hand, since 𝜕Ω ⊂ 𝐹 every complex line Π such that
Π ∩ 𝜕Ω ≠ ∅ is contained in C2 \Ω: Ω is weakly linearly convex. On the other hand,
every complex line intersecting Ω′ intersects Ω: Ω is not linearly convex.
We proceed with the construction of 𝐾 and 𝐾 ′. It is convenient to identify C2 with
4
R where the canonical coordinates are 𝜉 𝑗 ; we take 𝐻 to be the affine hyperplane in
R4 containing 𝒗 1 = (1, 0, 0, 0), 𝒗 2 = (0, 1, 0, 0), 𝒗 3 = (0, 0, 1, 0), 𝒗 4 = (0, 0, 0, 1),
the vertices of the standard simplex


 4
∑︁ 


𝑆 = 𝜉 ∈ R4 ; 𝜉 𝑗 = 1, ∀ 𝑗, 𝜉 𝑗 ≥ 0 ;
 
 𝑗=1 

𝑆 has 4 two-dimensional faces Σ 𝑗 = 𝜉 ∈ 𝑆; 𝜉 𝑗 = 0 ; Σ 𝑗 is a triangle in the hyper-
plane 𝜉 𝑗 = 0 with vertices 𝒗 𝑘 , 𝑘 ≠ 𝑗. For 31 < 𝜀 < 21 the set

𝑆 (𝜀) = 𝜉 ∈ 𝑆; max𝜉 𝑗 ≥ 𝜀
𝑗
25.1 C-Convexity and Global Solvability 1075

is the closure of an open and connected subset of 𝐻. We denote by 𝑇 (𝜀) the union
of all the segments 𝐿 ∩ 𝑆 with 𝐿 a straight line in 𝐻 such that 𝐿 ∩ 𝑆 (𝜀) = ∅; 𝑇 (𝜀)
is open in 𝑆. A further adjustment of 𝜀 will ensure that 𝔔 (𝜀) = 𝑆\ (𝑆 (𝜀) ∪ 𝑇 (𝜀))
is closed and nonempty.
To define 𝑇 (𝜀) we introduce the planar set


𝜔 𝑗 (𝜀) = Σ 𝑗 \ Σ 𝑗 ∩ 𝑆 (𝜀) = 𝜉 ∈ 𝑆; 𝜉 𝑗 = 0, max𝜉 𝑘 < 𝜀 ;
𝑘≠ 𝑗

𝜔 𝑗 (𝜀) is a nonempty triangle, open in Σ 𝑗 since 𝜉 ∈ 𝜔 𝑗 (𝜀) demands 0 < 𝜉 𝑘 < 𝜀 for
every 𝑘 ≠ 𝑗. This has two consequences:
(1) 𝜔 𝑗 (𝜀) ∩ 𝜔 𝑘 (𝜀) = ∅;
(2) the interior of Σ 𝑗 ∩ 𝑆 (𝜀) in Σ 𝑗 separates 𝜔 𝑗 (𝜀) from
Ø
𝜕Σ 𝑗 = [𝒗 𝑘 , 𝒗 ℓ ] .
1≤𝑘<ℓ ≤4,𝑘≠ 𝑗,ℓ≠ 𝑗

It follows that to each straight line 𝐿 in 𝐻 such that 𝐿 ∩ 𝑆 ≠ ∅, 𝐿 ∩ 𝑆 (𝜀) = ∅,

there is a pair of indices 𝑗, 𝑘, 1 ≤ 𝑗 < 𝑘 ≤ 4, such that 𝐿 intersects 𝜔 𝑗 (𝜀) at a single
point and exits 𝑆 at a point of 𝜔 𝑘 (𝜀). The convex hull of 𝜔 𝑗 (𝜀) ∪ 𝜔 𝑘 (𝜀), 𝑗 ≠ 𝑘, is
the following open subset of 𝑆:

𝔠 𝑗,𝑘 (𝜀) = 𝜉 ∈ 𝑆; 0 < 𝜉 𝑗 + 𝜉 𝑘 < 𝜀, 0 < 𝜉ℓ < 𝜀 if ℓ ≠ 𝑗, ℓ ≠ 𝑘 . (25.1.2)

If a straight line 𝐿 in 𝐻 is such that 𝐿 ∩ 𝑆 (𝜀) = ∅ then either 𝐿 ∩ 𝑆 = ∅ or else

Ø
𝐿 ∩ 𝑆 ⊂ 𝑇 (𝜀) = 𝔠 𝑗,𝑘 (𝜀) .
1≤ 𝑗<𝑘 ≤4

It follows that every straight line in 𝐻 that does not intersect 𝑆 (𝜀) does not intersect
𝔔 (𝜀), i.e., 𝔔 (𝜀) ∩ 𝑇 (𝜀) = ∅.
We come to the final adjustment of 𝜀 ∈ 31 , 12 . It follows from

4
∑︁
𝜉 ∈ 𝑆\𝑆 (𝜀) ⇐⇒ 𝜉 𝑗 = 1, max𝜉 𝑗 < 𝜀,
𝑗
𝑗=1

and from (25.1.2), that for 𝜉 ∈ 𝑆\𝑆 (𝜀) not to belong to 𝑇 (𝜀) it is necessary and
sufficient that 𝜉 𝑗 + 𝜉 𝑘 ≥ 𝜀 for all ( 𝑗, 𝑘), 1 ≤ 𝑗 < 𝑘 ≤ 4. Adding any three of these
lower bounds implies
3
𝜉 𝑗 + 𝜉 𝑘 + 𝜉ℓ ≥ 𝜀 if 1 ≤ 𝑗 < 𝑘 < ℓ ≤ 4.
2
1076 25 Solvability of Constant Vector Fields of Type (1,0)

We derive 𝜉 𝑗 ≤ 1 − 23 𝜀, 𝑗 = 1, 2, 3, 4. If 52 < 𝜀 < 21 we have 𝜉 ∉ 𝑆 (𝜀), implying

dist (𝔔 (𝜀) , 𝑆 (𝜀)) > 0. Since 𝑇 (𝜀)
is the complement of 𝔔 (𝜀) in 𝑆\𝑆 (𝜀), 𝔔 (𝜀)
2 1
must be closed. With 𝜀 ∈ 5 , 2 chosen we see that we can take 𝐾 = 𝑆 (𝜀),
𝐾 ′ = 𝔔 (𝜀).

25.1.3 Global solvability of (1,0)-vector fields with constant coefficients

The heading refers to vector fields

𝑛
∑︁ 𝜕
𝑍= 𝑎𝑗 , 𝑎 = (𝑎 1 , ..., 𝑎 𝑛 ) ∈ C𝑛 , |𝑎| ≠ 0. (25.1.3)
𝑗=1
𝜕𝑧 𝑗

We regard them as continuous linear operators on O (Ω).

The next lemma is a result in Several Complex Variables theory, crucial to our
present purpose but whose proof requires methods well beyond those presented in
this book. A proof based on Cartan’s Theorem B can be found in [Gunning and Rossi,
1965], p. 245, also in [Hörmander, 1966], p. 192; a proof based on estimates in 𝐿 2 -
spaces with weight exp (−Φ), Φ plurisubharmonic, can be found in [Hörmander,
1994], p. 262.
Lemma 25.1.13 Let Ω be a pseudoconvex domain in C𝑛 , 𝑓1 , ..., 𝑓𝑟 (1 ≤ 𝑟 < +∞)
be holomorphic functions in Ω and Ω′ ⊂ Ω be an open subset of C𝑛 such that
𝑽 = {𝑧 ∈ Ω′; 𝑓1 (𝑧) = · · · = 𝑓𝑟 (𝑧) = 0} is closed in Ω. For every ℎ ∈ O (Ω′) there
is an ℎ˜ ∈ O (Ω) such that ℎ˜ ≡ ℎ in 𝑽.
A first consequence of Lemma 25.1.13 is the following
Proposition 25.1.14 Let Ω be a pseudoconvex domain in C𝑛 and 𝐿 an affine complex
line in C𝑛 to which the vector field (25.1.3) is tangent. If 𝑍 O (Ω) = O (Ω) then every
connected component of 𝐿 ∩ Ω is simply connected.
Proof Suppose there is a point 𝑧◦ belonging to a relatively compact open subset
𝜔 of 𝐿 such that 1) 𝜔 ∩ Ω = ∅; 2) there is a loop in Ω ∩ 𝐿 going around 𝑧◦ .
Property 1) implies that there is an open subset Ω′ of Ω, Ω ∩ 𝐿 ⊂ Ω′, such that
(𝑧 − 𝑧 ◦ ) −1 ∈ O (Ω′). By Lemma 25.1.13 (in which 𝑽 = 𝐿 ∩Ω′) there is an 𝑓 ∈ O (Ω)
whose restriction to Ω ∩ 𝐿 is equal to (𝑧 − 𝑧◦ ) −1 . Property 2) implies that there is
no holomorphic solution 𝑢 of the equation 𝑍𝑢 = 𝑓 in Ω ∩ 𝐿, contradicting the
hypothesis. □
The following two lemmas are further applications of Lemma 25.1.13 needed
below.
Lemma 25.1.15 Let Ω be a pseudoconvex domain in C𝑛 , 𝐿 an affine complex line
in C𝑛 and 𝑧 ◦ ∈ 𝐿 ∩ Ω. Let 𝜁 (1) , ..., 𝜁 (𝑛−1) ∈ C𝑛 be linearly independent and such
that the equations 𝜁 ( 𝑗) · (𝑧 − 𝑧 ◦ ) = 0, 𝑗 = 1, ..., 𝑛 − 1, define 𝐿. If the restriction of
25.1 C-Convexity and Global Solvability 1077

ℎ ∈ O (Ω) to Ω ∩ 𝐿 vanishes identically there are 𝑔 𝑗 ∈ O (Ω), 𝑗 = 1, ..., 𝑛 − 1, such

that
𝑛−1
∑︁
∀𝑧 ∈ Ω, ℎ (𝑧) = 𝑔 𝑗 (𝑧) 𝜁 ( 𝑗) · (𝑧 − 𝑧◦ ) . (25.1.4)
𝑗=1

Proof To simplify the notation we take 𝑧◦ = 0. When 𝑛 = 1 we have 𝐿 = {0}; Ω

is an arbitrary open set in the plane such that 0 ∈ Ω; if ℎ ∈ O (Ω), ℎ (0) = 0, then
𝑧−1 ℎ (𝑧) ∈ O (Ω). For 𝑛 > 1 we take 𝜁 ( 𝑗) ·𝑧 = 𝑧 𝑗+1 , 𝑗 = 1, ...,
𝑛−1; then 𝐿 is the com-
plex 𝑧1 -line. Using induction on 𝑛 and the fact that Ω′ = 𝑧 ′ ∈ C𝑛−1 ; (𝑧 ′, 0) ∈ Ω
is pseudoconvex (Proposition 11.2.17), there are functions ℎ 𝑗 ∈ O (Ω′) such that
𝑛−1
∑︁
∀𝑧 ′ ∈ Ω′, ℎ (𝑧 ′, 0) = 𝑧 𝑗 ℎ 𝑗 (𝑧 ′) ,
𝑗=2

where 𝑧 ′ = (𝑧 1 , ..., 𝑧 𝑛−1 ). Now we apply Lemma 25.1.13 to 𝑽 = Ω′ × {0}: for each
𝑗 there is a function 𝑔 𝑗 ∈ O (Ω) whose restriction to Ω′ × {0} is equal to ℎ 𝑗 ; if
have 𝑔 (𝑧 ′, 0) = ℎ (𝑧 ′, 0), whence 𝑔𝑛 (𝑧) = 𝑧−1
Í
𝑔 = 𝑛−1𝑗=2 𝑧 𝑗 ℎ 𝑗 we Í 𝑛 (ℎ (𝑧) − 𝑔 (𝑧)) ∈
O (Ω) and ℎ (𝑧) = 𝑛𝑗=2 𝑧 𝑗 𝑔 𝑗 . □

Lemma 25.1.16 Assume the same hypotheses as in Proposition 25.1.14. If 𝑧◦ , 𝑧♭

belong to different connected components of 𝐿 ∩ Ω then there is an 𝜀 > 0 such that
𝑧◦ + 𝜀𝒗, 𝑧♭ + 𝜀𝒗 belong to different connected components of (𝐿 + 𝜀𝒗) ∩ Ω whatever
𝒗 ∈ C𝑛 , |𝒗| < 1.

Proof Let Γ be the connected component of 𝐿 ∩ Ω such that 𝑧◦ ∈ Γ; by Lemma

25.1.13 there is an ℎ ∈ O (Ω) whose restriction to 𝐿 ∩ Ω is equal to 1 in Γ and to
0 in (𝐿 ∩ Ω) \Γ; we have 𝑍 ℎ = 0 in 𝐿 ∩ Ω. We apply Lemma 25.1.15: there are
𝑔 𝑗 ∈ O (Ω), 𝑗 = 1, ..., 𝑛 − 1, such that
𝑛−1
∑︁
∀𝑧 ∈ Ω, 𝑍 ℎ (𝑧) = 𝑔 𝑗 (𝑧) 𝜁 ( 𝑗) · (𝑧 − 𝑧◦ ) .
𝑗=1

Since 𝑍 O (Ω) = O (Ω) there is a function

𝑓 𝑗 ∈ O (Ω)
such that 𝑍 𝑓 𝑗 = 𝑔 𝑗 for each
𝑗. Since 𝑍 is tangent to 𝐿 we have 𝑍 𝜁 ( 𝑗) · (𝑧 − 𝑧 ◦ ) = 0 for all 𝑗. It follows that
𝑍 (ℎ − 𝑓 ) = 0 if we define
𝑛−1
∑︁
𝑓 = 𝑓 𝑗 (𝑧) 𝜁 ( 𝑗) · (𝑧 − 𝑧 ◦ ) .
𝑗=1

This means that ℎ − 𝑓 is constant in each connected component of (𝐿 + 𝜀𝒗) ∩ Ω,

whatever 𝒗 ∈ C𝑛 . Since 𝑓 (𝑧◦ ) = 0 we see that if 𝜀 is sufficiently
|𝒗| < 1
small and
then |ℎ (𝑧 ◦ + 𝜀𝒗) − 𝑓 (𝑧 ◦ + 𝜀𝒗)| > 1/2 whereas ℎ 𝑧♭ + 𝜀𝒗 − 𝑓 𝑧♭ + 𝜀𝒗 < 1/2.
This would not be possible if these two numbers were equal. □
1078 25 Solvability of Constant Vector Fields of Type (1,0)

We are now in a position to prove the global solvability result we were aiming at.

Theorem 25.1.17 Let Ω be a C-convex domain in C𝑛 . Then 𝑍 O (Ω) = O (Ω) for

every vector field (25.1.3).

Proof It suffices to deal with 𝑍 = 𝜕𝑧𝜕𝑛 ; we write 𝑧 ′ = (𝑧1 , ..., 𝑧 𝑛−1 ); let 𝜋 : 𝑧 ↦→ 𝑧 ′
be the coordinate projection. By Proposition 25.1.3 𝜋 (Ω) is C-convex. For each
𝑧 ′ ∈ 𝜋 (Ω) we denote by 𝐿 𝑧′ the affine line in C𝑛 passing through (𝑧 ′, 0). For 𝑧 𝑛 ∈ C
let 𝑊 ′ (𝑧 𝑛 ) denote the set of points 𝑧 ′ ∈ 𝜋 (Ω) such that (𝑧 ′, 𝑧 𝑛 ) ∈ Ω [then, of course,
(𝑧 ′, 𝑧 𝑛 ) ∈ Ω ∩ 𝐿 𝑧′ ]; as 𝑧 𝑛 ranges over C the 𝑊 ′ (𝑧 𝑛 ) that are not empty form an open
covering 𝔉 of 𝜋 (Ω). We can find a locally finite open covering {U 𝜄 } 𝜄 ∈𝐼 of 𝜋 (Ω)
finer than 𝔉 and a C ∞ partition of unity {𝜑 𝜄 } 𝜄 ∈𝐼 subordinate to the covering {U 𝜄 } 𝜄 ∈𝐼 .
For each 𝜄 ∈ 𝐼 we select 𝜁 𝜄 ∈ C such that U 𝜄 ⊂ 𝑊 ′ (𝜁 𝜄 ); then supp 𝜑 𝜄 ⊂ 𝑊 ′ (𝜁 𝜄 ).
Now let 𝑓 ∈ O (Ω) be arbitrary. Since Ω ∩ 𝐿 𝑧′ is connected, to each 𝑧 ′ ∈ 𝑊 ′ (𝜁 𝜄 )
there is a smooth simple curve 𝛾 ⊂ Ω ∩ 𝐿 𝑧′ joining ∫ (𝑧 ′, 𝜁 𝜄 ) to (𝑧 ′, 𝑧 𝑛 ) ∈ Ω ∩ 𝐿 𝑧′ .
Since Ω ∩ 𝐿 𝑧′ is simply connected the integral 𝛾 𝑓 (𝑧 ′, 𝜁) d𝜁 is independent of the
∫𝑧
choice of 𝛾; we denote it by 𝑢 𝜄 (𝑧) = 𝜁 𝑛 𝑓 (𝑧 ′, 𝜁) d𝜁. We must prove that there is a
𝜄
𝜕𝑢 𝜕𝑢 𝜄 −1
𝑢 ∈ O (Ω) such that 𝜕𝑧 𝑛
= 𝜕𝑧𝑛 in Ω ∩ 𝜋 (U 𝜄 ) for every index 𝜄. If U 𝜄 ∩ U𝜅 ≠ ∅ we
introduce the integral
∫ 𝜁𝜅
′
𝑣 𝜄,𝜅 (𝑧 ) = 𝑢 𝜄 (𝑧) − 𝑢 𝑘 (𝑧) = 𝑓 (𝑧 ′, 𝜁) d𝜁 ∈ O (U 𝜄 ∩ U𝜅 ) (25.1.5)
𝜁𝜄

𝜑𝜆 𝑣 𝜄,𝜆 ∈ C ∞ (𝜋 (Ω)). We have, in U 𝜄 ∩ U𝜅 ,

Í
and we define 𝑣 𝜄 = 𝜆∈𝐼
∑︁ ∑︁
𝑣 𝜄 − 𝑣𝜅 = 𝜑𝜆 𝑣 𝜄,𝜆 − 𝑣 𝑘,𝜆 = 𝜑𝜆 𝑣 𝜄,𝜅 = 𝑣 𝜄,𝜅 . (25.1.6)
𝜆∈𝐼 𝜆∈𝐼

Let 𝜕¯ ′ = 𝜕𝜕𝑧¯1 , ..., 𝜕𝑧¯𝜕𝑛−1 ; we derive 𝜕¯ ′ 𝑣 𝜄 = 𝜕¯ ′ 𝑣 𝜅 in U 𝜄 ∩ U𝜅 , implying that there
is a C ∞ 𝜕¯ ′-closed (0, 1)-form in 𝜋 (Ω) whose restriction to U 𝜄 is equal to 𝜕¯ ′ 𝑣 𝜄
whatever 𝜄. Since 𝜋 (Ω) is pseudoconvex Theorem 11.2.21 enables us to conclude
that there is a 𝑣 ∈ C ∞ (𝜋 (Ω)) such that 𝜕¯ ′ 𝑣 = 𝜕¯ ′ 𝑣 𝜄 in U 𝜄 whatever 𝜄. We have
𝑤 𝜄 = 𝑣 𝜄 − 𝑣 ∈ O (U 𝜄 ) and 𝑣 𝜄,𝜅 = 𝑤 𝜄 − 𝑤 𝜅 in U 𝜄 ∩ U𝜅 whence, by (25.1.5)–(25.1.6),
𝑤 𝜄 − 𝑣 𝜄 = 𝑤 𝜅 − 𝑣 𝜅 and thus 𝑢 𝜄 − 𝑤 𝜄 ◦ 𝜋 = 𝑢 𝜅 − 𝑤 𝜅 ◦ 𝜋. We conclude that there is
−1
a 𝑢 ∈ O (Ω) whose restriction to Ω ∩ 𝜋 (U 𝜄 ) is equal to 𝑢 𝜄 − 𝑤 𝜄 ◦ 𝜋 for every 𝜄,
𝜕𝑢
thereby satisfying 𝜕𝑧 𝑛
= 𝑓 since 𝜕𝑧𝜕𝑛 (𝑤 𝜄 ◦ 𝜋) = 0. □

Theorem 25.1.18 If the pseudoconvex domain Ω in C𝑛 is bounded and if 𝑍 O (Ω) =

O (Ω) for every vector field (25.1.3) then Ω is C-convex.

Proof By Proposition 25.1.14 the hypotheses imply that if 𝐿 is an affine complex line
in C𝑛 intersecting Ω then every connected component of 𝐿 ∩ Ω is simply connected;
we must show that 𝐿 ∩ Ω is connected. Let 𝔄 be the subset of Ω × Ω consisting of
the pairs (𝑧, 𝑤) such that either 𝑧 = 𝑤 or 𝑧 and 𝑤 ≠ 𝑧 belong to the same connected
component of 𝐿 𝑧,𝑤 ∩ Ω, 𝐿 𝑧,𝑤 denoting the affine complex line that contains 𝑧 and
25.2 Local Solvability at the Boundary. First Steps 1079

𝑤. If (𝑧◦ , 𝑤 ◦ ) and (𝑧, 𝑤) are sufficiently close, an arbitrary continuous path joining
𝑧◦ to 𝑤 ◦ in 𝐿 𝑧 ◦ ,𝑤 ◦ ∩ Ω can be mapped onto a continuous path joining 𝑧 to 𝑤 in
𝐿 𝑧,𝑤 ∩ Ω. This means that 𝔄 is open in Ω × Ω; we claim that 𝔄 is closed in Ω × Ω.
The hypothesis that the closure of 𝐿 𝑧,𝑤 ∩ Ω in C𝑛 is compact ensures that if 𝜀 > 0
is sufficiently small and if (𝑧, 𝑤) is sufficiently close to (𝑧◦ , 𝑤 ◦ ) then
Ø
𝐿 𝑧,𝑤 ∩ Ω ⊂ 𝐿 𝑧 ◦ ,𝑤 ◦ + 𝜀𝒗 ,
|𝒗 |<1

and secondly, by Lemma 25.1.16, (𝑧, 𝑤) ∉ 𝔄. This proves that 𝔄 = Ω × Ω. □

Corollary 25.1.19 If the pseudoconvex domain Ω in C𝑛 is bounded then the following

two properties are equivalent: 1) 𝑍 O (Ω) = O (Ω) for every vector field (25.1.3); 2)
Ω is C-convex.

Corollary 25.1.20 Let the pseudoconvex domain Ω in C𝑛 be bounded and have a

C 1 boundary. If every vector field (25.1.3) defines a surjection of O (Ω) onto itself
then Ω is linearly convex (hence weakly and locally weakly linearly convex).

Proof Combine Theorem 25.1.18 with Theorem 25.1.11. □

25.2 Local Solvability at the Boundary. First Steps

25.2.1 Local holomorphic solvability at the boundary. Basics

Throughout this section Ω shall be a domain in C𝑛 with boundary 𝜕Ω = Ω∩ (C𝑛 \Ω),

Ω: closure of Ω. There will be regularity requirements put upon 𝜕Ω but they are not
needed in the following

Definition 25.2.1 Let 𝑧 ◦ ∈ 𝜕Ω and 𝑍 be a (1, 0)-vector field with holomorphic

coefficients and no critical point in a neighborhood 𝜔 of 𝑧◦ in C𝑛 ,
𝑛 𝑛
∑︁ 𝜕 ∑︁
𝑍= 𝑎 𝑗 (𝑧) , 𝑎 𝑗 ∈ O (𝜔) , 𝑎 𝑗 ≠ 0. (25.2.1)
𝑗=1
𝜕𝑧 𝑗 𝑗=1

We say that the equation 𝑍𝑢 = 𝑓 is locally holomorphically solvable (or, simply,

locally solvable) in Ω at 𝑧◦ if to every neighborhood 𝑈 ⊂ 𝜔 of 𝑧◦ in C𝑛 there is
a neighborhood 𝑉 ⊂ 𝑈 of 𝑧◦ in C𝑛 such that the equation 𝑍𝑢 = 𝑓 has a solution
𝑢 ∈ O (Ω ∩ 𝑉) whatever 𝑓 ∈ O (Ω ∩ 𝑈).

Remark 25.2.2 Because O (Ω ∩ 𝑈) and O (Ω ∩ 𝑉) are Fréchet–Montel spaces a

standard Baire category argument shows that the neighborhood 𝑉 ⊂ 𝑈 of 𝑧◦ in
Definition 25.2.1 can be chosen independently of 𝑓 ∈ O (Ω ∩ 𝑈) and thus to depend
solely on 𝑈 (besides Ω and 𝑍, of course).
1080 25 Solvability of Constant Vector Fields of Type (1,0)

Two observations are in order. Firstly and evidently, Definition 25.2.1 can be
restated in terms of germs. If 𝑧◦ ∈ Ω we denote by O𝑧 ◦ (Ω) the ring of germs of
holomorphic functions in the germ of set (Ω, 𝑧◦ ); if 𝑧 ◦ ∈ Ω we have O𝑧 ◦ (Ω) =
O𝑧 ◦ . Here we are interested in the case 𝑧 ◦ ∈ 𝜕Ω. Explicitly, a representative of a
germ f𝑧 ◦ ∈ O𝑧 ◦ (Ω) is a holomorphic function in an open set 𝑉 ∩ Ω where 𝑉 is a
neighborhood of 𝑧◦ in C𝑛 . Two functions 𝑓 𝑗 ∈ O Ω ∩ 𝑉 𝑗 , 𝑗 = 1, 2, belong to the
same coset f𝑧 ◦ if there is a neighborhood 𝑉3 ⊂ 𝑉1 ∩ 𝑉2 of 𝑧◦ in C𝑛 such that 𝑓1 = 𝑓2
in 𝑉3 ∩ Ω. As 𝑧◦ ranges over 𝜕Ω the rings O𝑧 ◦ (Ω) are the stalks of a sheaf over
𝜕Ω, which we shall denote by O𝜕Ω (Ω). The vector field (25.2.1) defines a sheaf
endomorphism of O𝜕Ω (Ω)| 𝜔∩𝜕Ω into itself, which we denote by boldface Z. The
restriction of Z to a stalk O𝑧 ◦ (Ω) is a derivation Z𝑧 ◦ of the ring O𝑧 ◦ (Ω) to which
we refer as the derivation of O𝑧 ◦ (Ω) induced by 𝑍; the equation 𝑍𝑢 = 𝑓 is locally
solvable in Ω at 𝑧◦ ∈ 𝜕Ω if this derivation is surjective.
The second observation is that (25.2.1) can always be transformed (up to a
nonvanishing factor) into a vector field (25.1.3) by a local biholomorphism preserving
𝑧◦ . That is why, in the sequel, we deal only with vector fields (25.1.3) or even, and
most often, with 𝑍 = 𝜕𝑧𝜕1 or 𝑍 = 𝜕𝑧𝜕𝑛 . In passing note that this facilitation was not
available in the preceding section when dealing with global solvability.
Let 𝑍 be as in (25.1.3); by a 𝑍-line we shall mean an affine complex line 𝐿 to
which the vector field 𝑍 is tangent at every point. If we were dealing with a vector
field (25.2.1) instead of 𝑍-lines we would be talking of integral manifolds of 𝑍.
In connection with our first observation we point out that in the following statement
there are no regularity conditions on the boundary 𝜕Ω.

Proposition 25.2.3 Let the domain Ω be pseudoconvex (Definition 11.2.15), 𝑧◦ ∈ 𝜕Ω

and 𝑍 be a vector field (25.1.3). If the derivation Z𝑧 ◦ of the ring O𝑧 ◦ (Ω) induced
by 𝑍 is surjective and if 𝑈 is a sufficiently small neighborhood of 𝑧 ◦ in C𝑛 then the
derivation Z𝑧 of O𝑧 (Ω) induced by 𝑍 is also surjective, whatever 𝑧 ∈ 𝑈 ∩ 𝜕Ω.

Proof Assume that there is a sequence of points 𝑧 ( 𝑗) ∈ 𝜕Ω ( 𝑗 = 1, 2, ...) converging

to 𝑧◦ and a corresponding sequence of functions 𝑓 𝑗 ∈ O Ω ∩ 𝑈 𝑗 , 𝑈 𝑗 a neighborhood
of 𝑧 ( 𝑗) in C𝑛 , such that the equation 𝑍𝑢 = 𝑓 𝑗 has no solution 𝑢 ∈ O Ω ∩ 𝑉 𝑗 whatever

the neighborhood 𝑉 𝑗 ⊂ 𝑈𝑗 of 𝑧 ( 𝑗) in C𝑛 . We take the neighborhoods 𝑈 𝑗 to be disjoint
and select 𝜑 𝑗 ∈ Cc∞ 𝑈 𝑗 , 𝜑 𝑗 (𝑧) = 1 if 𝑧 ∈ 𝑈 ′𝑗 ⊂ 𝑈 𝑗 , 𝑈 ′𝑗 a neighborhood of 𝑧 ( 𝑗)
Ø∞
in C𝑛 . Let 𝑔 = ∞ ∞ Ω; 𝑇 (0,1) ; we have 𝑔 (𝑧) = 0 if 𝑧 ∈ 𝑈 ′𝑗
Í
𝑓
𝑗=1 𝑗 𝜕𝜑 𝑗 ∈ C
𝑗=1
and 𝜕𝑔 ≡ 0 in Ω. Since Ω is pseudoconvex, Theorem Í 11.2.21 applies: there is a
𝑣 ∈ C ∞ (Ω) such that 𝜕𝑣 = 𝑔. It follows that ℎ = ∞ 𝑗=1 𝑓 𝑗 𝜑 𝑗 − 𝑣 ∈ O (Ω) and

that ℎ − 𝑓 𝑗 𝑈′ = − 𝑣|𝑈′𝑗 ∈ O 𝑈 ′𝑗 . Our definition of 𝑓 𝑗 implies that the equation

𝑗
𝑍𝑤 = ℎ has no solution in O 𝑈 ′𝑗 . Since every neighborhood of 𝑧 ◦ contains 𝑈 ′𝑗 if 𝑗
is sufficiently large, the derivation Z𝑧 ◦ : O𝑧 ◦ (Ω) ←↪ is not surjective. □
25.2 Local Solvability at the Boundary. First Steps 1081

We shall mostly deal with domains Ω in C𝑛 whose boundary 𝜕Ω is at least of class

C 1 . We assume that there are a neighborhood 𝑈 in C𝑛 of a point 𝑧◦ ∈ 𝜕Ω (often taken
to be the origin) and a function 𝜌 ∈ C 1 (𝑈; R) such that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0}
and d𝜌 (𝑧) ≠ 0 whatever 𝑧 ∈ 𝜕Ω ∩ 𝑈 (or even 𝑧 ∈ 𝑈). We always refer to 𝜌 as a
defining function of 𝜕Ω (or of Ω) in 𝑈.

Theorem 25.2.4 Let Ω be a domain in C𝑛 with a C 1 boundary 𝜕Ω. If a vector field

(25.1.3) is not tangent (i.e., is transverse) to 𝜕Ω at 𝑧◦ ∈ 𝜕Ω then the derivation Z𝑧 ◦
of the ring O𝑧 ◦ (Ω) induced by 𝑍 is surjective.

Proof We can take 𝑧 ◦ = 0 and 𝑍 = 𝜕 1

𝜕𝑧𝑛 . Assume 𝜌 is a C defining function of
𝜕Ω (or of Ω) in the neighborhood 𝑈 of 𝑧◦ ; by hypothesis, 𝜕𝑧
𝜕𝜌
𝑛
(𝑧◦ ) ≠ 0. Suppose
𝜕𝜌
𝜕𝑦𝑛 (𝑧◦ ) ≠ 0; the Implicit Function Theorem allows us to replace 𝜌 by −𝑦 𝑛 +
(𝑛−1)
𝑔 (𝑧 ′, 𝑥 1 ′
𝑛 ), 𝑔 ∈ C (𝑈), 𝑧 = (𝑧 1 , ..., 𝑧 𝑛−1 ). We can take 𝑈 = 𝔅 𝜀 × Δ 𝛿 , where
𝔅 𝜀(𝑛−1) = 𝑧 ′ ∈ C𝑛−1 ; |𝑧 ′ | < 𝛿 , Δ 𝛿 = {𝑧 𝑛 ∈ C; |𝑧 𝑛 | < 𝛿}. Let 𝐿 𝑧′ denote the 𝑍-line

through the point (𝑧 ′, 0). If 𝜀 and 𝛿 are sufficiently small then, whatever 𝑧 ′ ∈ Δ (𝑛−1)
𝜀 ,
the intersection 𝐿 𝑧′ ∩ Ω ∩ 𝑈 is a connected and simply connected open subset of 𝐿 𝑧′
containing
∫ 𝑧𝑛 𝑧 ′, 𝑧◦𝑛 with 𝑧 ◦𝑛 ∈ Δ 𝛿 independent of 𝑧 ′. If 𝑓 ∈ O (Ω ∩ 𝑈) the integral
𝑓 (𝑧 , 𝑡) d𝑡 along any simple smooth curve in Ω ∩ 𝑈 ∩ 𝐿 𝑧′ joining 𝑧 ′, 𝑧◦𝑛 to 𝑧
′

𝑧◦
𝑛
𝜕𝑢
defines a function 𝑢 ∈ O (Ω ∩ 𝑈) satisfying 𝜕𝑧𝑛 = 𝑓. □
In the next subsection we restrict our attention to a strictly pseudoconvex domain
Ω in C𝑛 (𝑛 ≥ 2) with a C 2 boundary (Definition 11.2.20); we “construct” defining
functions of Ω at a point of 𝜕Ω that have expressions suitable for our purposes.

Remark 25.2.5 If we assume that 𝜕Ω is nondegenerate (Definition 13.6.12) but not

pseudoconvex the problem of local solvability for a vector field (25.2.1) at a point
𝑧◦ ∈ 𝜕Ω is trivial. Indeed, suppose that Ω is defined in a neighborhood ◦
2 𝑈 of𝑧 in C
𝑛

by an inequality 𝜌 (𝑧) < 0, 𝜌 ∈ C ∞ (𝑈). If the (Levi) matrix 𝜕𝑧 𝑗 𝜕𝑧¯ 𝑘 (𝑧◦ )

𝜕 𝜌
1≤ 𝑗,𝑘<𝑛
has an eigenvalue 𝜒 < 0 then, by a classical theorem (see [Lewy, 1956]), an ar-
bitrary holomorphic function 𝑓 in Ω ∩ 𝑈 extends as a holomorphic function in a
neighborhood 𝑉 ⊂ 𝑈 of 𝑧 ◦ in C𝑛 where the equation 𝑍𝑢 = 𝑓 has a holomorphic
solution.

25.2.2 The positive definite case

We begin with a simple but significant special case.

Proposition 25.2.6 Let Ω be a domain in C𝑛 with a C 2 boundary 𝜕Ω and let 0 ∈ 𝜕Ω.

Suppose there is a neighborhood 𝑈 of 0 in C𝑛 such that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0}
with 𝜌 given by
𝜌 (𝑧) = −2 Im 𝑧 𝑛 + 𝑄 (𝑧) + 𝑜 |𝑧| 2 , (25.2.2)
1082 25 Solvability of Constant Vector Fields of Type (1,0)

where
𝑛−1
∑︁ 𝑛−1
∑︁
2
𝑄 (𝑧) = 𝑧 𝑗 + Re 𝑐 𝑗 𝑧 2𝑗 .
𝑗=1 𝑗=1

If the quadratic form 𝑄 (𝑧) is positive definite then the derivation of O0 (Ω) induced
by an arbitrary vector field 𝑍 given by (25.1.3), tangent to 𝜕Ω at 0, is surjective.

Proof We replace 𝑧 𝑗 by 𝑧 𝑗 exp − 21 𝑖𝜔 𝑗 , arg 𝑐 𝑗 = 𝜔 𝑗 , or, equivalently, we just
assume 𝑐 𝑗 ∈ R+ ( 𝑗 = 1, ..., 𝑛 − 1). In this case, we have
𝑛−1
∑︁
𝑄 (𝑧) = 𝑎 𝑗 𝑥 2𝑗 + 𝑏 𝑗 𝑦 2𝑗 , 𝑎 𝑗 , 𝑏 𝑗 ∈ R.
𝑗=1

Suppose that 𝑎 𝑗 > 0, 𝑏 𝑗 > 0 for all 𝑗; this implies 𝑦 𝑛 > 0 in Ω ∩ 𝑈 if 𝑈 is

sufficiently small. There is no loss of generality in assuming 𝑍 = 𝜕𝑧𝜕1 . We use the
notation 𝑧 ′′ = (𝑧 2 , ..., 𝑧 𝑛 ) and define Σ+𝜀 = 𝑧 ′′ ∈ C𝑛−1 ; |𝑧 ′′ | < 𝜀, 𝑦 𝑛 > 0 , 𝜀 > 0.

If 𝜀 is sufficiently small there is a neighborhood 𝑉 𝜀 ⊂ 𝑈 of 0 in C𝑛 such that

𝑧 ∈ Ω ∩ 𝑈, 𝑧 ′′ ∈ Σ+𝜀 , implies 𝑧 ∈ 𝑉 𝜀 . The expression (25.2.2) ensures that 𝑉 𝜀
can be chosen so that if 𝐿 𝑧′′ = C × {𝑧 ′′ } then Ω ∩ 𝑉 𝜀 ∩ 𝐿 𝑧′′ is close to an ellipse
centered at (0, 𝑧 ′′) and 2
∫ 𝑧 has a C boundary in 𝐿 𝑧′′ . Given arbitrarily 𝑓 ∈ O (Ω ∩ 𝑉 𝜀 )
we define 𝑢 (𝑧) = 0 1 𝑓 (𝜁, 𝑧 ′′) d𝜁 with integration on a simple smooth curve in
𝜕𝑢
Ω ∩ 𝑉 𝜀 ∩ 𝐿 𝑧′′ ; clearly 𝑢 is holomorphic and satisfies 𝜕𝑧 1
= 𝑓 in Ω ∩ 𝑊 𝜀 with
𝑊 𝜀 ⊂ 𝑉 𝜀 a neighborhood of 0 in C . 𝑛 □

25.2.3 Adapted defining functions

A linear transformation of C𝑛 allows us to select the tangent real hyperplane to 𝜕Ω

at the central point 𝑧 ◦ : we take it to be defined by 𝑦 𝑛 = 0 (we also take 𝑧◦ = 0, to
lighten the notation); this will not change in the sequel. The linear transformation
also provides a reasonably simple normal form of a defining function to start from.

Lemma 25.2.7 Let Ω be a strictly pseudoconvex domain in C𝑛 with a C 2 boundary

𝜕Ω and let 0 ∈ 𝜕Ω. After a linear change of coordinates in C𝑛 there is a defining
function of Ω in a neighborhood of 0 in C𝑛 of the form
𝑛
∑︁
𝜌 (𝑧) = −2𝑦 𝑛 + |𝑧 ′ | 2 + Re 𝑐 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + 𝑜 |𝑧| 2 (25.2.3)
𝑗,𝑘=1

where 𝑧 ′ = (𝑧1 , ..., 𝑧 𝑛−1 ) and 𝑐 𝑗,𝑘 = 𝑐 𝑘, 𝑗 ∈ C.

Proof A C-linear change of coordinates in C𝑛 transforms the real tangent space

to 𝜕Ω at 0 into the hyperplane 𝑦 𝑛 = 0. Let 𝑈 be a neighborhood of 0 in C𝑛 and
𝜌 ∈ C 2 (𝑈), d𝜌 ≠ 0 at every point of 𝑈, such that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0}.
25.2 Local Solvability at the Boundary. First Steps 1083

Division by a nonvanishing factor and Taylor expansion about 0 to order 2 imply

that, if 𝑈 is sufficiently small, then
𝑛
∑︁ 𝑛
∑︁
𝜌 (𝑧) = −2𝑦 𝑛 + 𝑎 𝑗,𝑘 𝑧 𝑗 𝑧¯ 𝑘 + Re 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + 𝑜 |𝑧| 2
𝑗,𝑘=1 𝑗,𝑘=1

with the self-adjoint matrix 𝑎 𝑗,𝑘 1≤ 𝑗,𝑘<𝑛 positive definite and 𝑏 𝑗,𝑘 1≤ 𝑗,𝑘<𝑛
complex symmetric. A C-linear change of the variables 𝑧1 , ..., 𝑧 𝑛−1 transforms
𝑧 𝑗 𝑧¯ 𝑘 into |𝑧 ′ | 2 ; in the new coordinates,
Í𝑛−1
𝑗,𝑘=1 𝑎 𝑗,𝑘

𝑛
∑︁ 𝑛
∑︁
𝜌 (𝑧) = −2𝑦 𝑛 + |𝑧 ′ | 2 + Re 𝑎 𝑗,𝑛 𝑧 𝑗 𝑧¯𝑛 + Re 𝑏 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + 𝑜 |𝑧| 2 . (25.2.4)
𝑗=1 𝑗,𝑘=1

Since
𝑛
∑︁ 𝑛 𝑛
©∑︁ ∑︁
Re 𝑎 𝑗,𝑛 𝑧 𝑗 𝑧¯𝑛 = Re ­ 𝑎 𝑗,𝑛 𝑧 𝑗 𝑧 𝑛 ® + 2𝑦 𝑛 Im
ª
𝑎 𝑗,𝑛 𝑧 𝑗
𝑗=1 « 𝑗=1 ¬ 𝑗=1

(25.2.4) is equivalent to

𝑛
∑︁ 𝑛
© ∑︁
𝜌 (𝑧) = −2 ­1 − Im 𝑎 𝑗,𝑛 𝑧 𝑗 ® 𝑦 𝑛 + |𝑧 ′ | 2 + Re ­ 𝑐 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 ® + 𝑜 |𝑧| 2 .
© ª ª

« 𝑗=1 ¬ « 𝑗,𝑘=1 ¬
Í𝑛
Division by 1 − Im 𝑗=1 𝑎 𝑗,𝑛 𝑧 𝑗 yields (25.2.3). □
In view of Theorem 25.2.4 we propose to study the local solvability at 0 ∈ 𝜕Ω
of a vector field (25.1.3) tangent to 𝜕Ω at 0; there is no loss of generality in taking
it to be 𝜕/𝜕𝑧 1 . In the sequel we are allowed to carry out changes of holomorphic
coordinates in C𝑛 provided they preserve the origin, the hyperplane tangent to 𝜕Ω at
0 and, more importantly, the vector field 𝜕/𝜕𝑧 1 up to a nonvanishing factor, thereby
preserving the foliation of C𝑛 by the 𝜕/𝜕𝑧 1 -lines. We shall refer to such a change of
variables as adapted to 𝜕/𝜕𝑧 1 .

Lemma 25.2.8 After a change of variables adapted to 𝜕/𝜕𝑧1 there is a defining

function of Ω in a suitably small neighborhood 𝑈 of 0 in C𝑛 , 𝜌 ∈ C 2 (𝑈), of the
following kind. If 𝑛 = 2,

𝜌 (𝑧) = −2𝑦 2 + 1 + 𝑐 1,1 𝑥12 + 1 − 𝑐 1,1 𝑦 21

(25.2.5)

2 2
+ 2𝑥 2 Re 𝑐 1,2 𝑧1 + 𝑜 |𝑧1 | + 𝑥2 ,

with 𝑐 1,1 ≥ 0, whereas if 𝑛 ≥ 3,

1084 25 Solvability of Constant Vector Fields of Type (1,0)

𝑛−1
∑︁ 2
𝑐 1,1 𝑥12 𝑐 1,1 𝑦 21

𝜌 (𝑧) = −2𝑦 𝑛 + 1 + + 1− + 𝑧𝑗 (25.2.6)
𝑗=2

+ 2𝑐 1,2 (𝑥1 𝑥2 − 𝑦 1 𝑦 2 ) + 2𝑥 𝑛 Re 𝑐 1,𝑛 𝑧1 + 𝑜 |𝑧 ′ | 2 + 𝑥 𝑛2

with 𝑐 1, 𝑗 ≥ 0, 𝑗 = 1, ..., 𝑛 − 1.

Proof The change of coordinate

𝑛
1 ∑︁
𝑧𝑛 − 𝑖 𝑐 𝑗,𝑛 𝑧 𝑗 𝑧 𝑘 ⇝ 𝑧 𝑛
2 𝑗,𝑘=2

is adapted to 𝜕/𝜕𝑧1 . After the due modification of 𝑐 1, 𝑗 (2 ≤ 𝑗 ≤ 𝑛) it simplifies

(25.2.3) to

𝑛
1 ∑︁
𝜌 (𝑧) = 2 Re ­𝑖𝑧 𝑛 + 𝑐 1,1 𝑧21 + 𝑧 1 𝑐 1, 𝑗 𝑧 𝑗 ® + |𝑧 ′ | 2 + 𝑜 |𝑧| 2 . (25.2.7)
© ª
2 𝑗=2
« ¬
Let 𝛼 𝑗 = arg 𝑐 1, 𝑗 ∈ [0, 2𝜋) for 1 ≤ 𝑗 ≤ 𝑛 − 1; we can replace 𝑧 1 by e−𝑖 𝛼1 /2 𝑧1 ; this
done, we replace 𝑧 𝑗 by e−𝑖 𝛼 𝑗 +𝑖 𝛼1 /2 𝑧 𝑗 for 1 < 𝑗 < 𝑛; this brings us to the case where
𝑐 1, 𝑗 ∈ R+ for all 𝑗 = 1, ..., 𝑛 − 1.
When 𝑛 = 2 this reduces (25.2.7) to

𝜌 (𝑧) = −2𝑦 2 + Re 𝑐 1,1 𝑧21 + 2𝑐 1,2 𝑧 1 𝑧 2 + |𝑧 1 | 2 + 𝑜 |𝑧| 2 ,

whence

𝜌 (𝑧) = −2𝑦 2 1 + Im 𝑐 1,2 𝑧1 + |𝑧1 | 2 + 𝑐 1,1 Re 𝑧21

(25.2.8)

+ 2𝑥 2 Re 𝑐 1,2 𝑧 1 + 𝑜 |𝑧| 2 .

When 𝑛 ≥ 4 a real orthogonal transformation of 𝑧 2 , ..., 𝑧 𝑛−1 brings us to the case

where 𝑐 1,𝑘 = 0 for every 𝑘 = 3, ..., 𝑛 − 1, reducing (25.2.7) to

𝜌 (𝑧) = −2𝑦 𝑛 1 + Im 𝑐 1,𝑛 𝑧1 + |𝑧 ′ | 2

(25.2.9)

+ Re 𝑐 1,1 𝑧 21 + 2𝑐 1,2 𝑧1 𝑧 2 + 2𝑥 𝑛 Re 𝑐 1,𝑛 𝑧1 + 𝑜 |𝑧| 2 .

When 𝑛 = 3 (25.2.9) is a direct consequence of (25.2.7).

Provided 𝑈 is sufficiently small division by 1 + Im 𝑐 1,𝑛 𝑧1 brings us to the
expression 𝜌 (𝑧) = −2𝑦 𝑛 + 𝜑 (𝑧 ′, 𝑥 𝑛 ) as the defining function of Ω in 𝑈 (without any
change of variables). It is readily derived from (25.2.8) [resp., (25.2.9)] that it must
have the expression (25.2.5) [resp., (25.2.6)] when 𝑛 = 2 (resp., 𝑛 ≥ 3). □
25.2 Local Solvability at the Boundary. First Steps 1085

25.2.4 Further decomposition of the defining function

We assume 𝜌 given by (25.2.5) when 𝑛 = 2 and by (25.2.6) when 𝑛 ≥ 3. It is

convenient to take

𝑈 = 𝑧 ∈ C𝑛 ; |𝑥1 | < 𝑟 1 , |𝑦 1 | < 𝑟 1′ , |𝑧 ′′ | < 𝑟 ′′

(25.2.10)

with 𝑧 ′′ = (𝑧 2 , ..., 𝑧 𝑛 ) and positive numbers 𝑟 1 , 𝑟 1′ , 𝑟 ′′ as small as needed. When

𝑛 ≥ 3 we are going to use systematically the notation 𝜃 = (𝑧 2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ). Thus we
rewrite (25.2.6) as 𝜌 (𝑧) = −2𝑦 𝑛 + 𝜙 (𝑧1 , 𝜃) with

𝜙 (𝑧 1 , 𝜃) = 1 + 𝑐 1,1 𝑥12 + 1 − 𝑐 1,1 𝑦 21 + 2𝑐 1,2 (𝑥1 𝑥2 − 𝑦 1 𝑦 2 )

(25.2.11)
𝑛−1
∑︁
2
+ 2 (𝑎𝑥1 − 𝑏𝑦 1 ) 𝑥 𝑛 + 𝑧 𝑗 + 𝑜 |𝑧 1 | 2 + |𝜃| 2 ,
𝑗=2

where 𝑎 = Re 𝑐 1,𝑛 , 𝑏 = Im 𝑐 1,𝑛 . When 𝑛 = 2 we replace 𝜃 by 𝑥 2 : 𝜌 (𝑧) = −2𝑦 𝑛 +

𝜙 (𝑧1 , 𝑥2 ) and (25.2.5) yields

𝜙 (𝑧1 , 𝑥2 ) = 1 + 𝑐 1,1 𝑥12 + 1 − 𝑐 1,1 𝑦 21

(25.2.12)

+ 2 (𝑎𝑥1 − 𝑏𝑦 1 ) 𝑥2 + 𝑜 |𝑧1 | 2 + 𝑥22 .

Lemma 25.2.9 Suppose 𝑛 ≥ 3 and 𝜌 (𝑧) = −2𝑦 𝑛 + 𝜙 (𝑧1 , 𝜃) ∈ C 2 (𝑈) with 𝜙 given
by (25.2.11). Let
𝜙min (𝑦 1 , 𝜃) = min 𝜙 (𝑧1 , 𝜃) .
|𝑥1 |<𝑟1

There is a function
√︁ 𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛
𝜓 (𝑧1 , 𝜃) = 1 + 𝑐 1,1 𝑥 1 + √︁ + 𝑜 (|𝑧1 | + |𝜃|) ∈ C 1 (𝑈) (25.2.13)
1 + 𝑐 1,1

such that 𝜙 (𝑧 1 , 𝜃) = 𝜙min (𝑦 1 , 𝜃) + 𝜓 2 (𝑧1 , 𝜃).

When 𝑛 = 2 the same formulas are valid if we put 𝑐 1,2 = 0, 𝜃 = 𝑥 2 .
If 𝜙 ∈ C ∞ (𝑈) [resp., 𝜙 ∈ C 𝜔 (𝑈)] the same is true of 𝜓 (𝑧 1 , 𝜃) and 𝜙min (𝑦 1 , 𝜃).

Proof We deal with the case 𝑛 ≥ 3; the case 𝑛 = 2 is simpler. We derive from
(25.2.11):
1 𝜕𝜙
(𝑧 1 , 𝜃) = 1 + 𝑐 1,1 𝑥 1 + 𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛 + 𝑜 (|𝑧1 | + |𝜃|) .
2 𝜕𝑥 1

The equation 𝜕𝜙
𝜕𝑥1 = 0 has a unique C 1 solution 𝑥1 = 𝜒 (𝑦 1 , 𝜃) with

𝑐 1,2 𝑥 2 + 𝑎𝑥 𝑛
𝜒 (𝑦 1 , 𝜃) = − + 𝑜 (|𝑧 1 | + |𝜃|) . (25.2.14)
1 + 𝑐 1,1
1086 25 Solvability of Constant Vector Fields of Type (1,0)

Provided 𝑟 1′ , 𝑟 ′′ are sufficiently small the critical value is a minimum of [−𝑟 1 , 𝑟 1 ] ∋

𝑥1 ↦→ 𝜙 (𝑧 1 , 𝜃),
2
𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛 𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛
𝜙min (𝑦 1 , 𝜃) = − 2 𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛 (25.2.15)
1 + 𝑐 1,1 1 + 𝑐 1,1
𝑛−1
∑︁
2
+ 1 − 𝑐 1,1 𝑦 21 − 2 𝑐 1,2 𝑦 2 + 𝑏𝑥 𝑛 𝑦 1 + 𝑧 𝑗 + 𝑜 𝑦 21 + |𝜃| 2 .

𝑗=2

If we define
1
𝜕2 𝜙
∫
1
𝑘 (𝑧1 , 𝜃) = ((1 − 𝑡) 𝜒 (𝑦 1 , 𝜃) + 𝑡𝑥1 ) (1 − 𝑡) d𝑡,
2 0 𝜕𝑥12

we can write

𝜙 (𝑧 1 , 𝜃) − 𝜙min (𝑦 1 , 𝜃) = (𝑥1 − 𝜒 (𝑦 1 , 𝜃)) 2 𝑘 (𝑧1 , 𝜃) . (25.2.16)

2
1𝜕 𝜙
Since 2 𝜕𝑥 2 = 1 + 𝑐 1,1 + 𝑜 (1) we have 𝑘 (𝑧1 , 𝜃) = 1 + 𝑐 1,1 + 𝑜 (1) ≥ 1 + 𝑜 (1); we
1
define √︁
𝜓 (𝑧1 , 𝜃) = (𝑥 1 − 𝜒 (𝑦 1 , 𝜃)) 𝑘 (𝑧 1 , 𝜃). (25.2.17)
We have 𝜙min (0) = 𝜓 (0) = 0; 𝑘 and 𝜓 are continuous functions in a neighborhood
of 0 (which we take to be 𝑈); the function

𝑘 1 (𝑧1 , 𝜃) = (𝑥 1 − 𝜒 (𝑦 1 , 𝜃)) 𝑘 (𝑧1 , 𝜃)

1 1
∫
𝜕 𝜕𝜙
= (1 − 𝑡) ((1 − 𝑡) 𝜒 (𝑦 1 , 𝜃) + 𝑡𝑥1 ) d𝑡
2 0 𝜕𝑡 𝜕𝑥 1
∫ 1
1 𝜕𝜙
= ((1 − 𝑡) 𝜒 (𝑦 1 , 𝜃) + 𝑡𝑥1 ) d𝑡
2 0 𝜕𝑥1
√︁
is of class C 1 . Thus 𝜓 = (𝑥 1 − 𝜒) 𝑘 1 is of class C 1 in the subset of 𝑈 where
1 1
𝑥1 − 𝜒 (𝑦 1 , 𝜃) ≠ 0; but the fact that d𝜓 = 𝑘 2 (d𝑥1 − d𝜒) + 12 𝑘 − 2 d𝑘 1 proves that
the partial derivatives of 𝜓 extend continuously to the whole of 𝑈 and therefore
𝜓 ∈ C 1 (𝑈). From (25.2.14) and (25.2.17) we derive (25.2.13). From (25.2.16) and
(25.2.17) we derive 𝜙 = 𝜙min + 𝜓 2 . The C ∞ (resp., C 𝜔 ) regularity of 𝜓 is a direct
consequence of (25.2.14) and (25.2.17); the regularity of 𝜙min = 𝜙 − 𝜓 2 ensues. □
Our new defining function in 𝑈 is

𝜌 (𝑧) = −2𝑦 𝑛 + 𝜙min (𝑦 1 , 𝜃) + 𝜓 2 (𝑧1 , 𝜃) (25.2.18)

with 𝜙min given by (25.2.15) and 𝜓 by (25.2.13). The following statements are direct
consequence of these formulas.
25.2 Local Solvability at the Boundary. First Steps 1087

Proposition 25.2.10 The preimages of points under the coordinate projection 𝜋1 :

Ω ∩ 𝑈 ∋ 𝑧 ↦→ (𝑦 1 , 𝑧 ′′) ∈ R × C𝑛−1 are open intervals ′′ } ⊂ C𝑛

in R 𝑥1 + 𝑖 {𝑦 1 } × {𝑧
𝑐1,2 𝑥2 +𝑎𝑥𝑛
centered approximately at − 1+𝑐1,1 + 𝑖𝑦 1 , 𝑧 ′′ .

Proof A direct consequence of (25.2.13) and (25.2.18) is that Ω ∩ 𝑈 is defined by

an inequality
2
2 𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛
𝜓 (𝑧1 , 𝜃) = 1 + 𝑐 1,1 𝑥1 + + 𝑜 (|𝑧1 | + |𝜃|) (25.2.19)
1 + 𝑐 1,1
< 2𝑦 𝑛 − 𝜙min (𝑦 1 , 𝜃) .

The claim follows directly from (25.2.19). □

Proposition 25.2.11 We have (𝑦 1 , 𝑧 ′′) ∈ 𝜋1 (Ω ∩ 𝑈) if and only if

𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 . (25.2.20)

Proof Since 𝜙min is a minimum of the function [−𝑟 1 , 𝑟 1 ] ∋ 𝑥1 ↦→ 𝜙1 (𝑧1 , 𝜃) (25.2.20)

is equivalent to the existence of points 𝑥1 ∈ [−𝑟 1 , 𝑟 1 ] such that 𝜙 (𝑧1 , 𝜃) < 2𝑦 𝑛 . □
Let us denote by 𝜋1′ the coordinate projection R × C𝑛−1 ∋ (𝑦 1 , 𝑧 ′′) ↦→ 𝑧 ′′ ∈ C𝑛−1 ;
the composite 𝜋1′ 𝜋1 is equal to the coordinate projection 𝜋 ′′ : Ω ∩ 𝑈 ∋ 𝑧 ↦→ 𝑧 ′′ ∈
C𝑛−1 .

Proposition 25.2.12 We have 𝑧 ′′ ∈ 𝜋 ′′ (Ω ∩ 𝑈) if and only if there is a 𝑦 1 ∈ −𝑟 1′ , 𝑟 1′

satisfying (25.2.20).

This is self-evident.

25.2.5 Necessary conditions for local holomorphic solvability at the

boundary

In this subsection we do not need the full force of Lemma 25.2.9; Lemma 25.2.8
will suffice.

Proposition 25.2.13 Let Ω be a strictly pseudoconvex domain in C𝑛 with a C 2

boundary 𝜕Ω and let 0 ∈ 𝜕Ω. Suppose there is a neighborhood 𝑈 of 0 in C𝑛 such
that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} with 𝜌 given by (25.2.5) when 𝑛 = 2 or by (25.2.6)
Í 2
when 𝑛 ≥ 3. If 𝜕/𝜕𝑧1 induces a surjective derivation of O0 (Ω) then 𝑛−1
𝑗=1 𝑐 1, 𝑗 ≤ 1.

Proof Suppose 𝑛 = 2. If the neighborhood 𝑉 ⊂ 𝑈 of 0 in C2 is sufficiently small

the function ℎ (𝑧) = 2𝑖𝑧 2 + 𝑐 1,1 𝑧 21 does not vanish in Ω ∩ 𝑉: indeed, if |𝑧1 | ≪ 1,

ℎ (𝑧) = 0 entails |𝑧 2 | ≲ |𝑧1 | 2 and 𝜌 (𝑧) = |𝑧1 | 2 + 𝑜 |𝑧 1 | 2 > 0 by (25.2.5). If 𝑐 1,1 > 1
and 𝜀 > 0 is sufficiently small then 𝜌 (𝑖𝑡, 𝑖𝑦 2 ) = 1 − 𝑐 1,1 𝑡 2 − 2𝑦 2 + 𝑜 𝑡 2 + 𝑦 22 < 0

1088 25 Solvability of Constant Vector Fields of Type (1,0)

for all (𝑡, 𝑦 2 ) ∈ [−𝜀, 𝜀] × [0, 𝜀], 𝑡 2 + 𝑦 2 > 0. If 𝜕𝑧𝜕𝑢

1
= ℎ1 , 𝑢 ∈ O (Ω ∩ 𝑉), then, for
0 < 𝑦 2 ≤ 𝜀, ∫ +𝜀
d𝑡
𝑢 (𝑖𝜀, 𝑖𝑦 2 ) − 𝑢 (−𝑖𝜀, 𝑖𝑦 2 ) = −𝑖 2
−𝜀 2𝑦 2 + 𝑐 1,1 𝑡

and the right-hand side ↗ +∞ as 𝑦 2 ↘ 0 while the left-hand side converges to

𝑢 (𝑖𝜀, 0) − 𝑢 (−𝑖𝜀, 0).
When 𝑛 ≥ 3 we duplicate the preceding reasoning with ℎ (𝑧) = 2𝑖𝑧 𝑛 + 𝑐 1,1 𝑧21 +
𝑐 1,2 𝑧1 𝑧2 : ℎ (𝑧) = 0 entails |𝑧 𝑛 | ≲ |𝑧 1 | 2 + |𝑧 2 | 2 and also, by (25.2.6),
𝑛−1
∑︁ 2
𝜌 (𝑧) = (1 + 𝑜 (1)) 𝑧𝑗 ≥0
𝑗=1

if |𝑧1 | 2 +|𝑧2 | 2 ≪ 1. If 𝑐 1,1 > 1 and 𝜀 is sufficiently small then 𝜌 (𝑖𝑡, 0, ..., 0, 𝑖𝑦 𝑛 ) =
1 − 𝑐 1,1 𝑡 2 − 2𝑦 𝑛 + 𝑜 𝑡 2 + 𝑦 2𝑛 < 0 for all (𝑡, 𝑦 𝑛 ) ∈ [−𝜀, 𝜀] × [0, 𝜀], 𝑡 2 + 𝑦 𝑛 > 0.
Using the same argument as in the case 𝑛 = 2 (after replacing 𝑦 2 by 𝑦 𝑛 ) we conclude
Í 2
that 𝑐 1,1 ≤ 1. This proves the claim in the case 𝑛−1 𝑗=2 𝑐 1, 𝑗 = 0.
In the remainder of the proof we assume 𝑐 1,𝑘 ≠ 0 for some 𝑘 ∈ [2, 𝑛 − 1]. We
Í
make a unitary change of the variables 𝑧 2 , ..., 𝑧 𝑛−1 transforming 𝑛−1 𝑘=2 𝑐 1,𝑘 𝑧 𝑘 into
𝜇𝑧 𝑛−1 (𝜇 > 0), followed by a change of variables 𝑧 𝑛 + 21 𝑖𝛾𝑧2𝑛−1 ⇝ 𝑧 𝑛 , 𝛾 > 0 to be
chosen below; (25.2.7) becomes

𝜌 (𝑧) = −2𝑦 𝑛 + |𝑧 ′ | 2 + 𝑐 1,1 Re 𝑧 21 + 2𝜇 Re (𝑧1 𝑧 𝑛−1 )

+ 𝛾 Re 𝑧2𝑛−1 + 𝑐 1,𝑛 Re (𝑧1 𝑧 𝑛 ) + 𝑜 |𝑧| 2 .

We will reason by contradiction and show that if 𝑐21,1 + 𝜇2 > 1 then there is a
𝜕𝑢
polynomial 𝑃 (𝑧) of degree 2 nowhere zero in Ω ∩ 𝑈 such that the equation 𝜕𝑧1
=
1 𝜕𝑃
𝑃 𝜕𝑧1 has no holomorphic solution 𝑢 in 𝑉 ∩ Ω whatever the neighborhood 𝑉 ⊂ 𝑈
of 0 in C𝑛 . We take
𝑃 (𝑧) = 2𝑖𝑧 𝑛 + (𝛼𝑧1 + 𝛽𝑧 𝑛−1 ) 2
where 𝛼, 𝛽 ∈ R; 𝑃 (𝑧) = 0 implies |𝑧 𝑛 | ≲ |𝑧 ′ | 2 + |𝑧 ′′ | 2 as well as 2𝑦 𝑛 =
Re (𝛼𝑧 1 + 𝛽𝑧 𝑛−1 ) 2 , whence

𝜌 (𝑧) = |𝑧 ′ | 2 + 𝑐 1,1 − 𝛼2 Re 𝑧21 − 2 (𝜇 − 𝛼𝛽) Re (𝑧1 𝑧 𝑛−1 ) (25.2.21)

+ 𝛾 − 𝛽2 Re 𝑧 2𝑛−1 + 𝑜 |𝑧| 2 .

We select 𝛼 > 0 such that 𝑐 1,1 − 𝛼2 < 1, then 𝛽 = 𝜇/𝛼, lastly 𝛾 = 𝛽2 . It is clear
that the right-hand side in (25.2.21) is > 0 if 0 < |𝑧| ≪ 1. In other words, 𝑃(𝑧) ≠ 0
for every 𝑧 ∈ Ω ∩ 𝑈 provided 𝑈 is sufficiently small.
Suppose that 𝑢 ∈ O (Ω ∩ 𝑈) satisfies 𝑃𝜕𝑢/𝜕𝑧 1 = 𝜕𝑃/𝜕𝑧1 in Ω ∩ 𝑈; we have
25.2 Local Solvability at the Boundary. First Steps 1089

𝜕𝑢 𝜕𝑃
𝑃 = = 2𝛼 (𝛼𝑧1 + 𝛽𝑧 𝑛−1 ) ,
𝜕𝑧1 𝜕𝑧1
𝜕𝑃
= 2𝛽 (𝛼𝑧1 + 𝛽𝑧 𝑛−1 ) .
𝜕𝑧 𝑛−1
We derive
𝜕2𝑢
𝑃 = 2𝛼2 − 4𝛼2 (𝛼𝑧 1 + 𝛽𝑧 𝑛−1 ) 2 /𝑃,
𝜕𝑧21
𝜕2𝑢
𝑃 = 2𝛼𝛽 − 4𝛼𝛽 (𝛼𝑧1 + 𝛽𝑧 𝑛−1 ) 2 /𝑃,
𝜕𝑧1 𝜕𝑧 𝑛−1
whence
𝜕2𝑢 𝜕2𝑢
𝛼 − 𝛽 2 = 0. (25.2.22)
𝜕𝑧1 𝜕𝑧 𝑛−1 𝜕𝑧1
Let Π1 denote the complex linear subspace of C𝑛 defined by the equations 𝑧 𝑗 = 0,
𝑗 ∈ [2, 𝑛 − 2]; Π1 (𝑧1 , 𝑧 𝑛−1 , 𝑧 𝑛 )-space C3 . We have, in Ω ∩ 𝑈 ∩ Π1 ,

𝜌 = −2𝑦 𝑛 + |𝑧1 | 2 + |𝑧 𝑛−1 | 2 + 𝑐 1,1 Re 𝑧21 (25.2.23)

𝛾 Re 𝑧2𝑛−1

+ 2𝜇 Re (𝑧1 𝑧 𝑛−1 ) + + Re 𝑐 1,𝑛 𝑧1 𝑧 𝑛

+ 𝑜 |𝑧 1 | 2 + |𝑧 𝑛−1 | 2 + |𝑧 𝑛 | 2 .

We define, in Ω ∩ 𝑈 ∩ Π1 ,
𝜕𝑢 𝜕𝑢
𝑣 (𝑧1 , 𝑧 𝑛−1 , 𝑧 𝑛 ) = 𝛼 −𝛽 ; (25.2.24)
𝜕𝑧 𝑛−1 𝜕𝑧 1
𝜕𝑣
then (25.2.22) entails 𝜕𝑧 1
≡ 0, i.e., 𝑣 (𝑧1 , 𝑧 𝑛−1 , 𝑧 𝑛 ) = 𝑣 (𝑧 𝑛−1 , 𝑧 𝑛 ).
For 𝜅, 𝜆 ∈ R let 𝑬 𝜅 ,𝜆 denote the real vector subspace of Π1 defined by the
equation 𝑧1 = 𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 . By (25.2.21) and neglecting terms 𝑜 |𝑧| 2 we derive:
if (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 , 𝑧 𝑛−1 , 𝑧 𝑛 ) ∈ Ω ∩ 𝑈 ∩ Π1 then, in Ω ∩ 𝑈 ∩ 𝑬 𝜅 ,𝜆 ,

𝜌 = 2 Re (𝑖𝑧 𝑛 ) + |𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 | 2 + |𝑧 𝑛−1 | 2

+ Re 𝑐 1,1 (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 ) 2 + 2𝜇 (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 ) 𝑧 𝑛−1 + 𝛾𝑧2𝑛−1

+ Re 𝑐 1,𝑛 (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 ) 𝑧 𝑛
= 2 Re 𝑖 1 − 𝑖𝑐 1,𝑛 (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 ) 𝑧 𝑛 + 𝐴 |𝑧 𝑛−1 | 2 + 𝐵 Re 𝑧2𝑛−1 ,

where

𝐴 = 1 + 𝜅 2 + 𝜆2 + 2𝜅𝜆𝑐 1,1 + 2𝜆𝜇,

𝐵 = 2𝜅𝜆 + 𝜅 2 + 𝜆2 𝑐 1,1 + 2𝜅𝜇 + 𝛾.
1090 25 Solvability of Constant Vector Fields of Type (1,0)

The hypersurface 𝜌 (0, ..., 0, 𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 , 0, ..., 0, 𝑧 𝑛−1 , 𝑧 𝑛 ) = 0 in (𝑧 𝑛−1 , 𝑧 𝑛 )-

space C2 has the Levi form 𝐴 |𝑧 𝑛−1 | 2 . The minimum of 𝐴 with respect to 𝜅 occurs at
𝜅 = −𝜆𝑐 1,1 and the minimum value of 𝐴 is 𝐴min = 1 + 2𝜆𝜇 + 𝜆 1 − 𝑐21,1 . This is
2

where we use the hypothesis 𝜇2 + 𝑐21,1 > 1: it implies 𝐴 < 0 for a suitable choice of
(𝜅, 𝜆). We apply Lemma 25.2.14 below: since 𝑣 is defined and holomorphic in the
submanifold Ω ∩ 𝑈 ∩ 𝑬 𝜅 ,𝜆 of 𝑈 defined by the conditions

𝑧1 = · · · = 𝑧 𝑛−2 = 0, 𝜌 (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 , 𝑧 𝑛−1 , 𝑧 𝑛 ) < 0,

𝑣 extends as a holomorphic function in a neighborhood 𝑉 ′ of the origin in C2 .

We can solve the equation 𝜕𝑧𝜕𝜓 𝑛−1
(𝑧 𝑛−1 , 𝑧 𝑛 ) = 𝛼−1 𝑣 (𝑧 𝑛−1 , 𝑧 𝑛 ) in 𝑉 ′ (keep in mind
that 𝛼 > 0, 𝛽 > 0). If 𝑉 is the preimage of 𝑉 ′ under the coordinate projection
𝜋 : Π1 ∋ (𝑧1 , 𝑧 𝑛−1 , 𝑧 𝑛 ) ↦→ (𝑧 𝑛−1 , 𝑧 𝑛 ) the function 𝑢♭ = 𝑢 − 𝜋 ∗ 𝜓 in Ω ∩𝑈 ∩𝑉 satisfies

𝜕𝑢♭ 1 𝜕𝑃 𝜕𝑢♭ 𝜕𝑢♭

= , = 𝛼−1 𝛽 ,
𝜕𝑧1 𝑃 𝜕𝑧1 𝜕𝑧 𝑛−1 𝜕𝑧1

the second equation a consequence of (25.2.24) and of 𝜕𝑧𝜕𝜓 𝑛−1

= 𝛼−1 𝑣. Let
𝑢˜ (𝑧 𝑛−1 , 𝑧 𝑛 ) = 𝑢♭ (𝜅𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 , 𝑧 𝑛−1 , 𝑧 𝑛 ); a priori 𝑢˜ (𝑧 𝑛−1 , 𝑧 𝑛 ) is not holomorphic
in the submanifold Ω ∩ 𝑈 ∩ 𝑬 𝜅 ,𝜆 but, as a consequence of (25.2.22),

𝜕 𝑢˜ 𝜅 + 𝛽/𝛼 𝜕𝑃
(𝑧 𝑛−1 , 𝑧 𝑛 ) = .
𝜕𝑧 𝑛−1 𝑃 𝜕𝑧 1 Ω∩𝑈∩𝑬 𝜅,𝜆

Recall that 𝑃 does not vanish in Ω ∩ 𝑈 ∩ 𝑬 𝜅 ,𝜆 , where we also have

𝜕 𝑢˜ 𝜕𝑢♭ 𝜆 𝜕𝑃
(𝑧 𝑛−1 , 𝑧 𝑛 ) = 𝜆 = , (25.2.25)
𝜕 𝑧¯𝑛−1 𝜕 𝑧¯1 𝑃 𝜕𝑧1
implying
𝜕 𝑢˜ 𝜕 𝑢˜
𝜆 = (𝜅 + 𝛽/𝛼)
𝜕𝑧 𝑛−1 𝜕 𝑧¯𝑛−1
and therefore
𝜕 𝑢˜
(𝜆 − (𝜅 + 𝛽/𝛼)) = 0,
𝜕𝑥 𝑛−1
𝜕 𝑢˜
(𝜆 + 𝜅 + 𝛽/𝛼) = 0.
𝜕𝑦 𝑛−1

We can wiggle by an arbitrarily small amount 𝜅+ 𝛽/𝛼 to ensure that 𝜆2 + (𝜅 + 𝛽/𝛼) 2 ≠

0; we derive that 𝑢˜ is independent of (𝑥 𝑛−1 , 𝑦 𝑛−1 ) in Ω ∩ 𝑈 ∩ 𝑬 𝜅 ,𝜆 , contradicting
(25.2.25) where the right-hand side is seen to be equal to

(𝜅 + 𝛽/𝛼) 𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1

𝜆 . □
𝑖𝛼−2 𝑧 𝑛 + 1
2 ((𝜅 + 𝛽/𝛼) 𝑧 𝑛−1 + 𝜆 𝑧¯𝑛−1 ) 2
25.2 Local Solvability at the Boundary. First Steps 1091

Lemma 25.2.14 Let 𝑈 be a neighborhood of 0 in C2 [where the variables are 𝑧 2 , 𝑧3 ]

and 𝜌 ∈ C 2 (𝑈) be such that

𝜌 (𝑧 2 , 𝑧3 ) = 2 Re (𝑖𝑧 3 + 𝑄 (𝑧2 , 𝑧3 )) + 𝐴 |𝑧2 | 2 + 𝑜 |𝑧2 | 2 + |𝑧3 | 2 . (25.2.26)

If 𝐴 < 0 every holomorphic function 𝑢 in the region {𝑧 ∈ 𝑈; 𝜌 (𝑧 2 , 𝑧3 ) < 0} extends

holomorphically to a neighborhood of 0 independent of 𝑢.
√
Proof The change of variables −𝐴𝑧 1 ⇝ 𝑧1 , 𝜁 = 𝑧 2 − 𝑖𝑄 (𝑧1 , 𝑧2 ) transforms
(25.2.26) into
𝜌 (𝜁, 𝑧 1 ) = −2 Im 𝜁 − |𝑧1 | 2 + 𝑜 |𝜁 | 2 + |𝑧1 | 2 .
𝑧2
We carry out the change of a single complex variable 𝜁 = 1−𝑖𝑧2 /2 ; in the new
coordinates

𝜌 (𝑧 1 , 𝑧2 ) = 1 − |𝑧1 | 2 − |𝑧 2 + 𝑖| 2 + 𝑜 |𝑧1 | 2 + |𝑧2 | 2 . (25.2.27)

Set 𝜔𝑡 = {𝑧 ∈ 𝑉; 𝜌 (𝑧 1 , 𝑧2 ) < −𝑡}, 0 < 𝑡 ≪ 1; 𝑢 ∈ O (𝜔0 ) has a smooth restric-

tion to 𝜕𝜔𝑡 if 𝑡 > 0. By the Approximation Theorem in [Baouendi-Treves, 1981]
(see also [Stein-Shakarchi, 2011], p. 299 et seq.) there is a neighborhood 𝑉1 ⊂ 𝑉 of 0
in C2 such that 𝑢 is the uniform limit in Σ𝑡 = 𝜕𝜔𝑡 ∩ 𝑉1 of a sequence of polynomials
𝑃 𝜈 (𝑧1 , 𝑧2 ); it follows that 𝑃 𝜈 converges uniformly in the polynomially convex hull
of Σ𝑡 , bΣ𝑡 . When 𝑡 is sufficiently small (25.2.27) shows directly that bΣ𝑡 contains a
neighborhood of 0 in C2 . □
Remark 25.2.15 A consequence of Theorem 25.1.17 and Proposition 25.2.13 is that,
given a strictly pseudoconvex domain Ω with a C 2 boundary and a point 𝑧◦ ∈ 𝜕Ω,
it is not true, in general, that every neighborhood 𝑈 of 𝑧 ◦ in C𝑛 contains another
neighborhood 𝑉 of 𝑧 ◦ in C𝑛 such that Ω ∩ 𝑉 is C-convex.

25.2.6 A sufficient condition for local holomorphic solvability at the

boundary

Proposition 25.2.16 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a

C 2 boundary 𝜕Ω and let 0 ∈ 𝜕Ω. Suppose there is a neighborhood 𝑈 of 0 in C𝑛 such
that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} with 𝜌 given by (25.2.5) when 𝑛 = 2, or by (25.2.6)
Í 2
where 𝑛−1𝑗=1 𝑐 1, 𝑗 ≤ 1, when 𝑛 ≥ 3. If 𝑐 1,1 < 1 then 𝜕/𝜕𝑧 1 induces a surjective
derivation of O0 (Ω).
Proof Case 𝑛 = 2. When 𝑐 1,2 = 0, (25.2.5) reads

𝜌 (𝑧) = −2𝑦 2 + 1 + 𝑐 1,1 𝑥 12 + 1 − 𝑐 1,1 𝑦 21 + 𝑜 |𝑧| 2

and the claim is a particular case of Proposition 25.2.6. Suppose 𝑐 1,2 ≠ 0; for |𝑧1 |
small (25.2.5) can be rewritten as
1092 25 Solvability of Constant Vector Fields of Type (1,0)

𝜌 (𝑧) = −2𝑦 2 + |𝑧 1 | 2 + Re 𝑐 1,1 𝑧21 + 2𝑐 1,2 𝑧 1 𝑧 2

+ 𝑜 |𝑧 1 | 2 + 𝑥22 + 𝑂 (|𝑧1 𝑦 2 |) .

The change of variable 𝑧1 ⇝ 𝑧1 + 𝜏𝑧 2 (𝜏 ∈ C to be chosen) transforms the quadratic

form at the right into

𝑄 (𝑧) = |𝑧1 + 𝜏𝑧 2 | 2 + 𝑐 1,1 Re (𝑧1 + 𝜏𝑧2 ) 2 + 2 Re 𝑐 1,2 𝑧2 (𝑧 1 + 𝜏𝑧 2 )

= |𝑧1 | 2 + 𝑐 1,1 Re 𝑧21 + |𝜏𝑧 2 | 2 + 𝑐 1,1 Re (𝜏𝑧 2 ) 2 + 2 Re 𝜏¯ + 𝜏𝑐 1,1 + 𝑐 1,2 𝑧1 𝑧2

+4 Im ( 𝜏𝑧
¯ 1 𝑦2) .

Selecting 𝜏 ∈ C such that 𝜏¯ + 𝜏𝑐 1,1 + 𝑐 1,2 = 0 (implying 𝜏 ≠ 0) yields

𝑄 (𝑧) = |𝑧1 | 2 + 𝑐 1,1 Re 𝑧21 + |𝜏𝑧2 | 2 + 𝑐 1,1 Re (𝜏𝑧 2 ) 2 + 𝑂 (|𝑧1 𝑦 2 |) .

It follows that, if 𝑐 1,1 < 1, then

(1 + 𝑂 (|𝑧|)) −1 𝜌 (𝑧 1 + 𝜏𝑧 2 , 𝑧2 ) = |𝑧1 | 2 + 𝑐 1,1 Re 𝑧21 + |𝜏𝑧 2 | 2 + 𝑐 1,1 Re (𝜏𝑧 2 ) 2

is positive definite and we can apply Proposition 25.2.6.

Case 𝑛 ≥ 3. Let 𝜌 = 𝜓 (𝑧1 , 𝜃) − 𝑦 𝑛 , 𝜃 = (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ); we can assume 𝜓 is
given by (25.2.12); then (25.2.15) implies
2
𝑐 1,2 𝑦 2 + 𝑏𝑥 𝑛 2

𝑐 1,2 𝑥 2 + 𝑎𝑥 𝑛
𝜓min (𝑦 1 , 𝜃) = 1 − 𝑐 1,1 𝑦 1 − − (25.2.28)
1 − 𝑐 1,1 1 + 𝑐 1,1
2 ∑︁ 𝑛−1
𝑐 1,2 𝑦 2 + 𝑏𝑥 𝑛 2

− + 𝑧 𝑗 + 𝑜 𝑦 21 + |𝜃| 2 .
1 − 𝑐 1,1 𝑗=2

Let 𝑈 be as in (25.2.10) and 𝜋 ′′ : C𝑛𝑧 −→ C𝑛−1

𝑧 ′′ denote the coordinate projec-
tion. Combining (25.2.28) with (25.2.19) shows that if 𝑟 ′′ is sufficiently small
−1
and 𝑧 ′′ ∈ 𝜋′′ (Ω ∩ 𝑈) then the preimage 𝜋 ′′ (𝑧 ′′) is approximately a rectangle
𝑐1,2 𝑥2 +𝑎𝑥𝑛 𝑐1,2 𝑦2 +𝑏𝑥𝑛
contained in (−𝑟 1 , 𝑟 1 ) × −𝑟 1′ , 𝑟 1′ ; this planar

centered at − 1+𝑐1,1 , 1−𝑐1,1
domain 𝔇 (𝑧 ′′) is connected and simply connected. It is also evident that every
point ′′◦ ∈ 𝜋 ′′ (Ω ∩ 𝑈) has a neighborhood N (𝑧 ′′◦ ) ⊂ 𝜋 ′′ (Ω ∩ 𝑈) such that
Ù𝑧
𝔇 (𝑧 ′′) ≠ ∅.
𝑧 ′′ ∈N (𝑧 ′′◦ )
We derive from (25.2.28)

𝑐21,2
! 𝑛−1
1 1 1 ∑︁ 2
𝜓min (0, 𝜃) − 𝑦 𝑛 = 1− 𝑥22 + 𝑦 22 + 𝑧𝑗
2 1 + 𝑐 1,1 2 2 𝑗=3
1
− 𝑦𝑛 − 𝑎 2 𝑥 𝑛2 + 2𝑎𝑐 1,2 𝑥2 𝑥 𝑛 + 𝑜 |𝜃| 2 .
2 1 + 𝑐 1,1
25.2 Local Solvability at the Boundary. First Steps 1093

Since 𝑐21,2 ≤ 1 − 𝑐21,1 we see that

1 2 1 2
𝑐21,2
Δ 𝑥2 ,𝑦2 𝑐 1,1 𝑥2 + 𝑦 2 = 2 − ≥ 1 + 𝑐 1,1 ,
2 2 1 + 𝑐 1,1

proving that the domain in 𝑧 ′′-space defined by 𝜓min (0, 𝜃) < 𝑦 𝑛 is strictly pseudocon-
vex. On the one hand this is also true of the domains Ω′′ 𝑦1 defined by 𝜓min (𝑦 1 , 𝜃) < 𝑦 𝑛
if |𝑦 1 | ≤ 𝑟 1′ provided 𝑟 1′ is sufficiently small. Since 𝑐 1,1 < 1 (25.2.28) shows
that 𝑦 1 ↦→ 𝜓min (𝑦 1 , 𝜃) will have a minimum in −𝑟 1′ , 𝑟 1′ at some point 𝑦 ◦1 . If

′ ′
−𝑟 1 , 𝑟 1 ∋ 𝑦 1 ≠ 𝑦 ◦1 then Ω′′ ′′ ′′ ′′
𝑦1 ⊂ Ω 𝑦 ◦ and therefore Ω 𝑦 ◦ is equal to 𝜋 (Ω ∩ 𝑈).
1 1
We reach the conclusion that if 𝑟 1′ is sufficiently small then 𝜋 ′′ (Ω ∩ 𝑈) is strictly
pseudoconvex.
From the earlier statement about the domains 𝔇 (𝑧 ′′) we deduce that there is
a locally finite covering {U 𝜄 } 𝜄 ∈𝐼 of Ω ∩ 𝑈 by open sets such that to each index
𝜄 ∈ 𝐼 there is a 𝑧1( 𝜄) ∈ C such that 𝑧1( 𝜄) , 𝑧 ′′ ∈ Ω ∩ 𝑈 for every 𝑧 ′′ ∈ U 𝜄 . Let now
∫𝑧
𝑓 ∈ O (Ω ∩ 𝑈) be arbitrary. For each 𝜄 ∈ 𝐼 we define 𝑢 𝜄 (𝑧) = 𝑧 ( 1𝜄) 𝑓 (𝑡, 𝑧 ′′) d𝑡 where
1
−1
the integration is performed over a simple smooth curve contained in 𝜋 ′′ (𝑧 ′′)∩Ω∩𝑈.
Since 𝜋 ′′ (Ω ∩ 𝑈) is pseudoconvex we can complete the proof exactly as in that of
Theorem 25.1.17. □

25.2.7 𝒁 -pseudoconvexity and strict 𝒁 -pseudoconvexity

So far all the results of this section have been stated and proved for special choices
of the defining function of Ω and of the coordinates in C𝑛 . It is convenient to seek
a modicum of invariance and, incidentally, to name the conditions in Proposition
25.2.13 and Theorem 25.2.23. We shall say that Ω (or 𝜕Ω) is 𝜕/𝜕𝑧1 -pseudoconvex
Í 2
at 0 ∈ 𝜕Ω if either 𝜕/𝜕𝑧1 is transverse to 𝜕Ω at 0 or else 𝑛−1
𝑗=1 𝑐 1, 𝑗 ≤ 1 in (25.2.3).
We can reformulate this condition so that it applies to any vector field (14.5.2). If
we restrict our attention to defining functions (25.2.3) we can content ourselves with
the following
Í
Definition 25.2.17 Let 𝑍 = 𝑎 · 𝜕𝑧 = 𝑛𝑗=1 𝑎 𝑗 𝜕𝑧𝜕 𝑗 , 𝑎 ∈ C𝑛 \0. We shall say that Ω (or
𝜕Ω) is 𝑍-pseudoconvex
Í at 0 if either
𝑍 is transverse to 𝜕Ω at 0 or else the linear
𝑛−1
functional 𝑧 ↦→ 𝑍 𝑗,𝑘=1 𝑗,𝑘 𝑗 𝑘 has norm ≤ |𝑎|.
𝑐 𝑧 𝑧

The following claim is self-evident.

Proposition 25.2.18 Let Ω be a strictly pseudoconvex domain with C 2 boundary
𝜕Ω ∋ 0. Suppose that 𝜕/𝜕𝑧 1 is tangent to 𝜕Ω at 0 and that there is a neighborhood
𝑈 of 0 in C𝑛 such that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} with 𝜌 given by (25.2.5) when
𝑛 = 2 or by (25.2.7) when 𝑛 ≥ 3. The domain Ω is 𝜕/𝜕𝑧1 -pseudoconvex at 0 ∈ 𝜕Ω
Í 2
if and only if 𝑛−1
𝑗=1 𝑐 1, 𝑗 ≤ 1.
1094 25 Solvability of Constant Vector Fields of Type (1,0)

Greater invariance (from biholomorphisms preserving 𝑍 up to a nonvanishing

factor) and independence from the choice of 𝜌 requires further generalization. In
the nontransverse case, we introduce a generic holomorphic vector field 𝑤 · 𝜕𝑧
(0 ≠ 𝑤 ∈ C𝑛 ) tangent to 𝜕Ω at 𝑧, meaning 𝑤 · 𝜕𝑧 𝜌 (𝑧) = 0, and we form a new
vector field Z𝑤 = 𝑤 · 𝜕𝑧 + 𝑤 0 𝑍 (𝑤 0 ∈ C) and an extended Levi form (cf. Definition
11.2.18):

Z𝑤 Z 𝑤 𝜌 = |𝑤 0 | 2 𝑍 𝑍 𝜌 + (𝑤 · 𝜕𝑧 ) (𝑤 · 𝜕𝑧¯ ) 𝜌 + 2 Re (𝑤 0 (𝑤 · 𝜕𝑧 ) 𝑍 𝜌) . (25.2.29)

We compute Z𝑤 Z 𝑤 𝜌 at 𝑧 = 0 when the defining function is given by (25.2.3) and

𝑛−1
∑︁ 𝜕
𝑍= 𝑎𝑗 , 𝑎 = (𝑎 1 , ..., 𝑎 𝑛−1 ) ∈ C𝑛−1 , |𝑎| ≠ 0. (25.2.30)
𝑗=1
𝜕𝑧 𝑗

The hypothesis that 𝑤 · 𝜕𝑧 is tangent to 𝜕Ω at 0 means that 𝑤 𝑛 = 0; we have

(𝑤 · 𝜕𝑧 ) (𝑤 · 𝜕𝑧¯ ) 𝜌 (𝑧) = (1 + 𝑜 (1)) |𝑤| 2 ,

𝑍 𝑍 𝜌 (𝑧) = (1 + 𝑜 (1)) |𝑎| 2 ,
𝑛−1
∑︁
(𝑤 · 𝜕𝑧 ) 𝑍 𝜌 (0) = 𝑐 𝑗,𝑘 𝑎 𝑗 𝑤 𝑘 + 𝑜 (1) |𝑎| |𝑤| ,
𝑗,𝑘=1

whence
𝑛−1
∑︁
Z𝑤 Z 𝑤 𝜌 (0) = |𝑤 0 𝑎| 2 + |𝑤| 2 + 2 Re ­𝑤 0 𝑐 𝑗,𝑘 𝑎 𝑗 𝑤 𝑘 ® . (25.2.31)
© ª

« 𝑗,𝑘=1 ¬
Consider the special case where 𝑎 1 = 1, 𝑎 𝑗 = 0 if 2 ≤ 𝑗 ≤ 𝑛 − 1:

𝑛−1
© ∑︁
Z𝑤 Z 𝑤 𝜌 (0) = |𝑤 0 | 2 + |𝑤| 2 + 2 Re ­𝑤 0
ª
𝑐 1, 𝑗 𝑤 𝑗 ®
« 𝑗=1 ¬
1
𝑛−1 2
©∑︁ 2ª
≥ |𝑤 0 | 2 + |𝑤| 2 − 2 ­ 𝑐 1, 𝑗 ® |𝑤 0 | |𝑤| ,
« 𝑗=1 ¬
the last lower bound a consequence of the Cauchy–Schwarz inequalities. If we
require Z𝑤 Z 𝑤 𝜌 (0) ≥ 0 for all (𝑤 0 , 𝑤) ∈ C𝑛+1 such that 𝑤 𝑛 = 0 we must have
Í𝑛−1 2
𝑗=1 𝑐 1, 𝑗 ≤ 1.
All this justifies the introduction of definitions independent of the choices of the
defining function 𝜌 and of the complex coordinates:
25.2 Local Solvability at the Boundary. First Steps 1095

Definition 25.2.19 Let 𝑍 = 𝑎 · 𝜕𝑧 = 𝑛𝑗=1 𝑎 𝑗 𝜕𝑧𝜕 𝑗 , 𝑎 ∈ C𝑛 \ {0}, and 𝜌 be a C 2

Í

function defining Ω in a neighborhood of 𝑧◦ ∈ 𝜕Ω. We shall say that Ω (or 𝜕Ω) is

𝑍-pseudoconvex at 𝑧◦ if either 𝑍 is transverse to 𝜕Ω at 𝑧 ◦ or else Z𝑤 Z 𝑤 𝜌 (𝑧) ≥ 0
for all 𝑧 ∈ 𝜕Ω in a neighborhood of 𝑧 ◦ and all (𝑤 0 , 𝑤 1 , ..., 𝑤 𝑛 ) ∈ C𝑛+1 \ {0} such
𝜕𝜌
(𝑧 ◦ ) = 0.
Í
that 𝑛𝑗=1 𝑤 𝑗 𝜕𝑧 𝑗

Definition 25.2.20 Let 𝑍 , Ω, 𝜌 be as in Definition 25.2.19. We shall say that Ω (or

𝜕Ω) is strictly 𝑍-pseudoconvex at 𝑧◦ ∈ 𝜕Ω if Ω is 𝑍-pseudoconvex at 𝑧◦ and, when
𝑍 is tangent to 𝜕Ω at 𝑧 ◦ , we have Z𝑤 Z 𝑤 𝜌 (𝑧◦ ) > 0 if 𝑤 = 𝜆𝑎, 0 ≠ 𝜆 ∈ C, 𝑤 0 ∈ C
arbitrary.

We can state:

Proposition 25.2.21 Let Ω be a strictly pseudoconvex domain with C 2 boundary

𝜕Ω ∋ 0. Suppose that 𝜕/𝜕𝑧 1 is tangent to 𝜕Ω at 0 and that there is a neighborhood
𝑈 of 0 in C𝑛 such that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} with 𝜌 given by (25.2.5) when
𝑛 = 2 or by (25.2.7) when 𝑛 ≥ 3. The domain Ω is strictly 𝜕/𝜕𝑧1 -pseudoconvex at
Í 2
0 ∈ 𝜕Ω if and only if 𝑐 1,1 < 1 and 𝑛−1
𝑗=1 𝑐 1, 𝑗 ≤ 1.

Proof We put 𝑤 0 = −𝑐 1,1 𝑤 1 (0 ≤ 𝑐 1,1 ≤ 1), 𝑤 𝑗 = 0 for 𝑗 > 1 in the right-hand side
of (25.2.31), thus equating it to 1 − 𝑐 1,1 |𝑤 1 | 2 .

□

Remark 25.2.22 It is evident that we should think of (𝑤 0 , 𝑤 1 , ..., 𝑤 𝑛 ) as representing

a point in complex 𝑛-dimensional projective space.

Propositions 25.2.13 and 25.2.16 can be restated as follows:

Theorem 25.2.23 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a C 2

boundary 𝜕Ω and let 𝑧◦ ∈ 𝜕Ω. For 𝑍 to induce a surjective derivation of O𝑧 ◦ (Ω)
it is necessary that Ω be 𝑍-pseudoconvex at 𝑧 ◦ and sufficient that Ω be strictly
𝑍-pseudoconvex at 𝑧◦ .

The next section is devoted to filling the obvious gap between the necessary
condition and the sufficient condition in Theorem 25.2.23.

25.2.8 Symplectic interpretation of 𝒁 -pseudoconvexity

In preparation for the results in the next section we close this section with a symplectic
interpretation of 𝑍-pseudoconvexity. To this end we introduce the outer conormal
∗ 𝜕Ω of 𝜕Ω and related concepts, defined in Definition 13.6.16, where
bundle 𝑁out
the domain Ω was assumed to be strictly pseudoconvex and 𝜕Ω to be smooth (It
could as well have been C 2 .) But rather than revising Ch. 13 we shall assume, in this
subsection, that 𝜕Ω is C ∞ . (In the final result of this chapter 𝜕Ω is assumed to be
∗ 𝜕Ω is a half-line subbundle of 𝑇 (1,0) Ω
C 𝜔 .) We recall that 𝑁out , I-symplectic and
𝜕Ω
R-Lagrangian at every one of its points (Proposition 13.6.17.)
1096 25 Solvability of Constant Vector Fields of Type (1,0)

We continue to reason in a neighborhood 𝑈 of 𝑧◦ ∈ 𝜕Ω such that Ω ∩ 𝑈 =

{𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0}, 𝜌 ∈ C ∞ (𝑈); then
n o
∗
𝑁out 𝜕Ω 𝑈 = (𝑧, 𝜁) ∈ 𝑇 (1,0) Ω ; 𝜌 (𝑧) = 0, 𝜁 = 𝜆𝜕𝑧 𝜌 (𝑧) , 𝜆 > 0 , (25.2.32)
𝑈

𝜕𝜌 𝜕𝜌
where 𝜕𝑧 𝜌 = 𝜕𝑧1 , ..., 𝜕𝑧𝑛 . Actually, to simplify the notation, we write M =
∗ 𝜕Ω ;
𝑁out we take 𝑧 ◦ = 0 and 𝜌 given by (25.2.3):
𝑈

𝑛−1
∑︁ 𝑛
2 © ∑︁
𝜌 (𝑧) = −2𝑦 𝑛 + 𝑧 𝑗 + Re ­ 𝑐 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 ® + 𝑜 |𝑧| 2 . (25.2.33)
ª
𝑗=1 « 𝑗,𝑘=1 ¬
It follows from (25.2.32) that the pullback to M of the (1, 0)-form 𝜁 ·d𝑧 in 𝑇 (1,0) Ω
is equal to 𝜆𝜕 𝜌; the pullback to M of the real one-form d𝜌 = 𝜕 𝜌 + 𝜕 𝜌 vanishes
identically and therefore the pullback to M of 𝜕 𝜌 is purely imaginary (this is not the
same as saying that ∇𝑧 𝜌 is purely imaginary!); the same is true of the sum on the
right in
𝑛
∑︁ 𝜕2 𝜌
d (𝜆𝜕 𝜌) = d𝜆 ∧ 𝜕 𝜌 + 𝜆 d𝑧¯ 𝑘 ∧ d𝑧 𝑗 . (25.2.34)
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

We can take the two-form −𝑖d (𝜆𝜕 𝜌) as the fundamental symplectic two-form 𝜛M on
the (I-)symplectic C ∞ manifold M [(25.2.34) shows also that M is R-Lagrangian].
We can use 𝑥 𝑗 ( 𝑗 = 1, ..., 𝑛), 𝑦 𝑘 (𝑘 = 1, ..., 𝑛 − 1) and 𝜆 as coordinates (but not as
Darboux coordinates; cf. Definition 13.3.19) in M. The restriction of the covector
d𝑦 𝑛 | 0 = − 21 d𝜌| 0 to 𝑇(0,𝜁 ) M vanishes; therefore, by (25.2.33) the pullback of 𝜕 𝜌| 0
is equal to
𝜕𝜌 1 𝜕𝜌
d𝑧 𝑛 = (0) d𝑥 𝑛 = 𝑖d𝑥 𝑛 .
𝜕𝑧 𝑛 0 2𝑖 𝜕𝑦 𝑛
From this and from (25.2.33) we deduce
𝑛−1
∑︁
−𝑖d (𝜆𝜕 𝜌)| 0 = d𝜆 ∧ d𝑥 𝑛 − 𝑖𝜆 d𝑧¯ 𝑗 ∧ d𝑧 𝑗 ,
𝑗=1

i.e.,
𝑛−1
∑︁
−𝑖d (𝜆𝜕 𝜌)| 0 = d𝜆 ∧ d𝑥 𝑛 + 2𝜆 d𝑥 𝑗 ∧ d𝑦 𝑗 . (25.2.35)
𝑗=1

This implies that there are Darboux coordinates (Definition 13.3.19) on M, 𝑥˜ 𝑗 , 𝜉e𝑘
( 𝑗, 𝑘 = 1, ..., 𝑛) such that

𝑥˜ 𝑗 − 𝑥 𝑗 ( 𝑗 = 1, ..., 𝑛), 𝜉e𝑘 − 2𝑦 𝑘 (𝑘 = 1, ..., 𝑛 − 1), 𝜉e𝑛 + 𝜆

25.2 Local Solvability at the Boundary. First Steps 1097

vanish to second order at 𝑧 = 0. We can select the Poisson bracket in M of a pair of

functions 𝑓 , 𝑔 ∈ C 1 (M) to be
𝑛−1
1 ∑︁ 𝜕 𝑓 𝜕𝑔 𝜕 𝑓 𝜕𝑔 𝜕 𝑓 𝜕𝑔 𝜕 𝑓 𝜕𝑔
{ 𝑓 , 𝑔} M = − + − . (25.2.36)
2 𝑗=1 𝜕𝑥 𝑗 𝜕𝑦 𝑗 𝜕𝑦 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑛 𝜕𝜆 𝜕𝜆 𝜕𝑥 𝑛

Theorem 25.2.24 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a C ∞

boundary 𝜕Ω and let 𝑧◦ ∈ 𝜕Ω; let 𝑍 be a vector field (25.1.3) with symbol 𝜎 (𝑍).
The following two properties are equivalent:
(a) Ω is 𝑍-pseudoconvex at 𝑧◦ ;
(b) ∀ (𝑧◦ , 𝜁) ∈ M =𝑁out
∗ 𝜕Ω, 𝜎 (𝑍) (𝑧 ◦ , 𝜁) = 0 implies

{Re 𝜎 (𝑍) , Im 𝜎 (𝑍)} M (𝑧◦ , 𝜁) ≤ 0.

Proof The intersection of 𝑁out ∗ 𝜕Ω with 𝑇 ∗ C𝑛 is a half-line bundle over 𝜕Ω;

𝑁out 𝜕Ω 𝑧 ◦ ∩ 𝑇𝑧 ◦ C is the ray 𝑅 𝑧 ◦ of points (0, 𝜆𝜕𝑧 𝜌 (𝑧◦ )), 𝜆 > 0. It follows that
∗ ∗ 𝑛

either 𝜎 (𝑍) = −𝑖 ⟨𝜁 · d𝑧, 𝑍⟩ vanishes everywhere or nowhere on 𝑅 𝑧 ◦ . To say that

𝑍 is tangent to 𝜕Ω at 𝑧◦ is equivalent to saying that 𝜎 (𝑍) ≡ 0 on 𝑅 𝑧 ◦ . We assume
this to be the case. As before we take 𝑧 ◦ = 0. It suffices to prove the statement for
𝑍 = 𝜕/𝜕𝑧1 . The symbol of 𝜕/𝜕𝑧 1 as a function in 𝑇 ∗ C𝑛 is equal to 𝑖𝜁1 ; by (25.2.32)
𝜕𝜌
its restriction to M is equal to 𝑖𝜆 𝜕𝑧 1
; by (25.2.3) we have
𝑛
𝜕𝜌 ∑︁
= 𝑧¯1 + 𝑐 1,𝑘 𝑧 𝑘 + 𝑂 |𝑧| 2 .
𝜕𝑧 1 𝑘=1

𝜕𝜌
We apply (25.2.36) with 𝑓 = 𝑔¯ = 𝑖𝜆 𝜕𝑧 1
; by ( 25.2.33) 𝜕𝜕𝜆𝑓 = 𝑖 𝜕𝑧𝜕𝜌
1
vanishes at
𝑧 = 0. A direct calculation shows that if 𝑧 = 0 then
𝑛−1 𝑛
! 𝑛
!
𝜕𝜌 𝜕𝜌 1 ∑︁ 𝜕 ∑︁ 𝜕 ∑︁
, = 𝑖 𝑥1 + 𝑐 1,𝑘 𝑥 𝑘 𝑦1 − 𝑐¯1,𝑘 𝑦 𝑘
𝜕𝑧1 𝜕 𝑧¯1 M 2 𝑗=1 𝜕𝑥 𝑗 𝑘=1
𝜕𝑦 𝑗 𝑘=1
𝑛−1 𝑛
! 𝑛
!
1 ∑︁ 𝜕 ∑︁ 𝜕 ∑︁
− 𝑖 𝑥1 + 𝑐¯1,𝑘 𝑥 𝑘 𝑦𝑗 + 𝑐 1,𝑘 𝑦 𝑘
2 𝑗=1 𝜕𝑥 𝑗 𝑘=1
𝜕𝑦 𝑗 𝑘=1

∑︁𝑛
2ª
= 𝑖 ­1 − 𝑐 1, 𝑗 ® .
©

« 𝑗=1 ¬
n o
2
By Definition 25.2.17 (a) means that 𝑛𝑗=1 𝑐 1, 𝑗 ≤ 1, hence 𝑖 −1 𝜕𝑧
𝜕𝜌 𝜕𝜌
Í
1
, 𝜕 𝑧¯1 M ≥ 0 at
n o
𝜕𝜌
𝑧 = 0, which is the same as Re 𝜕𝑧 1
, Im 𝜕𝜕𝜌
𝑧¯1 ≤ 0 at 𝑧 = 0. □
M
1098 25 Solvability of Constant Vector Fields of Type (1,0)

Theorem 25.2.25 Let Ω and 𝑍 be as in Theorem 25.2.24. The following two prop-
erties are equivalent:
(a) Ω is strictly 𝑍-pseudoconvex at 𝑧◦ (Definition 25.2.20);
(b) ∀ (𝑧 ◦ , 𝜁) ∈ M =𝑁out
∗ 𝜕Ω, 𝜎 (𝑧 ◦ , 𝜁) = 0 implies

{Re 𝜎 (𝑍) , Im 𝜎 (𝑍)} M (𝑧 ◦ , 𝜁) < 0.

Proof An affine automorphism of C𝑛 brings us to the case where 𝑧 ◦ = 0 and

𝑍 = 𝜕/𝜕𝑧1 . We limit ourselves to the case 𝑛 ≥ 3; the case 𝑛 = 2 is simpler. As
pointed out above the restriction to M =𝑁out ∗ 𝜕Ω of the symbol of 𝜕/𝜕𝑧 (which is
1
𝑖𝜁1 as a function on 𝑇 ∗ C𝑛 ) is equal to 𝑖𝜆 𝜕𝑧
𝜕𝜌
1
. Lemma 25.2.9 allows us to assume that

𝜌 (𝑧) = −2𝑦 𝑛 + 𝜓min (𝑦 1 , 𝜃) + ℎ2 (𝑧1 , 𝜃) ,

where 𝜃 = (𝑧2 ., ..., 𝑧 𝑛−1 , 𝑥 𝑛 ). We derive

𝜕𝜌 1 𝜕𝜓min 𝜕ℎ
= −𝑖 (𝑦 1 , 𝜃) + ℎ (𝑧1 , 𝜃) (𝑧 1 , 𝜃)
𝜕𝑧1 2 𝜕𝑦 1 𝜕𝑦 1
𝜕ℎ
+ ℎ (𝑧 1 , 𝜃) (𝑧1 , 𝜃) .
𝜕𝑥1
Thus, if we define

𝜕𝜌 𝜕𝜌 1 𝜕𝜌
𝑓 = Im 𝑖𝜆 = 𝜆 Re = 𝜆 ,
𝜕𝑧1 𝜕𝑧1 2 𝜕𝑥1

𝜕𝜌 𝜕𝜌 1 𝜕𝜌
𝑔 = Re 𝑖𝜆 = −𝜆 Im = 𝜆 ,
𝜕𝑧1 𝜕𝑧1 2 𝜕𝑦 1

then
𝜕ℎ
𝜆−1 𝑓 (𝑧1 , 𝜃) = ℎ (𝑧1 , 𝜃) (𝑧1 , 𝜃) ,
𝜕𝑥1
1 𝜕𝜓min 𝜕ℎ
𝜆−1 𝑔 (𝑧1 , 𝜃) = (𝑦 1 , 𝜃) + ℎ (𝑧1 , 𝜃) (𝑧 1 , 𝜃) .
2 𝜕𝑦 1 𝜕𝑦 1

We reason in a neighborhood 𝑈 ′ of 0 in 𝑧1 , 𝜃 -space C𝑛−1 × R assumed to be as small

as needed. From (25.2.13) we derive
𝜕ℎ √︁
(𝑧1 , 𝜃) = 1 + 𝑐 1,1 + 𝑂 (|𝑧1 | + |𝜃|) .
𝜕𝑥1
We see that 𝑓 (𝑧1 , 𝜃) = 0 ⇐⇒ ℎ (𝑧1 , 𝜃) = 0 in 𝑈 ′. With the Poisson brackets in M
defined by (25.2.36) we obtain

−1 𝜕ℎ 𝜕ℎ
𝜆 { 𝑓 , 𝑔} M = {ℎ, 𝑔} M + ℎ ,𝑔 ,
𝜕𝑥1 𝜕𝑥1 M
25.2 Local Solvability at the Boundary. First Steps 1099

and therefore 𝑓 = 0 implies

𝜕ℎ 1 𝜕ℎ 𝜕𝜓min
𝜆−2 { 𝑓 , 𝑔} M = 𝜆−1 {ℎ, 𝑔} M = ℎ, . (25.2.37)
𝜕𝑥1 2 𝜕𝑥1 𝜕𝑦 1 M

Formula (25.2.13) and the Implicit Function Theorem yield a factorization ℎ =

𝐸 𝜒 with
𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛
𝜒 (𝑧1 ,𝜃) = 𝑥1 + + 𝑂 |𝑦 1 | 2 + |𝜃| 2
1 + 𝑐 1,1
and 𝐸 ∈ C 𝜔 (𝑈; R), 𝐸 > 0 everywhere in 𝑈 ′. The change of variable
𝑐 1,2 𝑥2 + 𝑎𝑥 𝑛
𝑧1 + ⇝ 𝑧1 (25.2.38)
1 + 𝑐 1,1

preserves the vector field 𝜕/𝜕𝑧 1 as well as the hyperplane 𝑦 𝑛 = 0. In the new
coordinates 𝜒 (𝑧1 , 𝜃) = 𝑥1 + 𝑂 |𝑦 1 | 2 + |𝜃| 2 and therefore 𝑓 = 0 entails

1 𝜕 2 𝜓min

𝜕𝜓min
𝜒, = + 𝑂 (|𝑦 1 | + |𝜃|) . (25.2.39)
𝜕𝑦 1 M 2 𝜕𝑦 21

The change of variable (25.2.38) transforms 𝜓min (𝑦 1 , 𝜃) into

♭ 𝑐 1,2 𝑦 2 + 𝑎𝑦 𝑛
𝜓min (𝑦 1 , 𝜃) = 𝜓min 𝑦 1 − ,𝜃 .
1 + 𝑐 1,1

From (25.2.15) we derive that 𝑓 = 0 entails

𝜕 2 𝜓min
♭

(𝑦 1 , 𝜃) = 2 1 − 𝑐 1,1 𝑂 (|𝑦 1 | + |𝜃|) .
𝜕𝑦 21

From this equation and (25.2.39) we derive

( )
♭
2
𝜕𝜓min
|( 𝑓 + 𝑖𝑔) (0)| = 0 =⇒ 𝜒, (0) = 1 − 𝑐 1,1 > 0.
𝜕𝑦 1
M

Combining this entailment with (25.2.37) and the fact that {Re 𝜎 (𝑍) , Im 𝜎 (𝑍)} M =
− { 𝑓 , 𝑔} M proves the claim. □
1100 25 Solvability of Constant Vector Fields of Type (1,0)

25.3 Local Solvability at the Boundary. Final Characterization

25.3.1 A new condition: quasiconvexity

In view of Proposition 25.2.16, Theorem 25.2.23 and in the language of Theorems

25.2.24, 25.2.25, it remains to settle the case in which Ω is 𝑍-pseudoconvex but
not strictly pseudoconvex at 𝑧◦ . In terms of 𝑍 = 𝜕/𝜕𝑧 1 (and 𝑧 ◦ = 0) this is the
case where 𝑐 1,1 = 1 and 𝑐 1, 𝑗 = 0 if 1 < 𝑗 < 𝑛. We return to the “streamlined”
defining function constructed in Subsection 25.2.4 and we systematically use the
notation of that subsection, in particular the defining function (25.2.11) when 𝑛 ≥ 3,
𝜌 (𝑧) = −2𝑦 𝑛 + 𝜙 (𝑧1 , 𝜃) with 𝜃 = (𝑧 2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) and
𝑛−1
∑︁
2
𝜙 (𝑧 1 , 𝜃) = 2𝑥 12 + 2 (𝑎𝑥1 − 𝑏𝑦 1 ) 𝑥 𝑛 + 𝑧 𝑗 + 𝑜 |𝑧 1 | 2 + |𝜃| 2 , (25.3.1)
𝑗=2

where 𝑎 = Re 𝑐 1,𝑛 , 𝑏 = Im 𝑐 1,𝑛 . When 𝑛 = 2 we replace 𝜃 by 𝑥 2 :

𝜙 (𝑧1 , 𝑥2 ) = 2𝑥 12 + 2 (𝑎𝑥1 − 𝑏𝑦 1 ) 𝑥 2 + 𝑜 |𝑧 1 | 2 + 𝑥22 . (25.3.2)

We get, when 𝑛 ≥ 3 [cf. (25.2.15)],

𝑛−1
1 ∑︁ 2

𝜙min (𝑦 1 , 𝜃) = − 𝑎 2 𝑥 𝑛2 − 2𝑏𝑦 1 𝑥 𝑛 + 𝑧 𝑗 + 𝑜 𝑦 21 + |𝜃| 2 , (25.3.3)
2 𝑗=2

or, when 𝑛 = 2,
1
𝜙min (𝑦 1 , 𝑥2 ) = − 𝑎 2 𝑥22 − 2𝑏𝑦 1 𝑥2 + 𝑜 𝑦 21 + 𝑥22 . (25.3.4)
2
Also [cf. (25.2.13)],
√ 1
𝜓 (𝑧1 , 𝜃) = 𝜙 (𝑧1 , 𝜃) − 𝜙min (𝑦 1 , 𝜃) = 2𝑥 1 + √ 𝑎𝑥 𝑛 + 𝑜 (|𝑧 1 | + |𝜃|) (25.3.5)
2
with 𝜃 replaced by 𝑥2 when 𝑛 = 2.
It follows from these formulas that the analogue of Proposition 25.2.10 remains
valid:

Proposition 25.3.1 The preimages of points under the coordinate projection 𝜋1 :

Ω ∩ 𝑈 ∋ 𝑧 ↦→ (𝑦 1 , 𝑧2 , ..., 𝑧 𝑛 ) ∈ R × C𝑛−1 are open intervals centered approximately
at − 21 𝑎𝑥 𝑛 .

Before tackling the general case we discuss two special cases.

25.3 Local Solvability at the Boundary. Final Characterization 1101

Example 25.3.2 The model case: 𝑛 ≥ 3 (when 𝑛 = 2 things are simpler) and there
′ 2 2
are no terms 𝑜 |𝑧 | + 𝑥 𝑛 . Thus,

𝑛−1
1 1 ∑︁ 2
𝜙min (𝑦 1 , 𝜃) = − 𝑎 2 𝑥 𝑛2 − 𝑏𝑦 1 𝑥 𝑛 + 𝑧𝑗 . (25.3.6)
4 2 𝑗=2

The function 𝑦 1 ↦→ 𝜙min (𝑦 1 , 𝜃) is either constant (when 𝑏𝑥 𝑛 = 0) or monotone. In

both cases the preimages of points under 𝜋 ′′ are connected and simply connected,
and 𝜋 ′′ (Ω ∩ 𝑈) is strictly pseudoconvex (cf. the proof of Theorem 25.2.23). We
conclude that 𝜕/𝜕𝑧 1 defines a map of O0 onto itself.

Example 25.3.3 Suppose that there is a sequence of points 𝜃 = (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) ∈
C𝑛−2 × R converging to 0, such that the function 𝑦 1 ↦→ 𝜙min (𝑦 1 , 𝜃) has an isolated
maximum at 𝑦 1 = 0. After suitably decreasing 𝑟 1′ and 𝑟 ′′ we can assume that for
some 𝜃 ∈ C𝑛−2 × R arbitrarily close to 0, to every 𝑦 −1 ∈ −𝑟 1′ , 0 there is a 𝑦 +1 ∈ 0, 𝑟 1′

such that
𝜙min 𝑦 −1 , 𝜃 = 𝜙min 𝑦 +1 , 𝜃 < 2𝑦 𝑛 ,

1
where 𝑦 𝑛 = (0, 𝜃). Then 𝑧 ′′ = (𝑧2 , ..., 𝑧 𝑛 ) ∈ 𝜋 ′′ (Ω ∩ 𝑈) and the set
2 𝜙min
−1
𝜋1′ (Ω ∩ 𝑈) contains 𝑦 −1 , 𝑧 ′′ and 𝑦 +1 , 𝑧 ′′ but not (0, 𝑧 ′′); thus 𝜋 ′′ (𝑧 ′′) ∩ Ω ∩ 𝑈

is not connected.

These examples show that the key condition finalizing the characterization of
local holomorphic solvability of 𝜕/𝜕𝑧 1 at the boundary
must be sought in the totality
of the defining function 𝜌, including in the terms 𝑜 |𝑧 ′ | 2 + 𝑥 𝑛2 in (25.3.1)–(25.3.2),
unlike the conditions in the results proved so far. For this we will introduce a new
concept; but first we recall an old one.
A real-valued function 𝑓 in an open interval I ⊂ R is said to be convex if,
whatever 𝑎, 𝑏 ∈ I,

∀𝜆 ∈ (0, 1) , 𝑓 ((1 − 𝜆) 𝑎 + 𝜆𝑏) ≤ (1 − 𝜆) 𝑓 (𝑎) + 𝜆 𝑓 (𝑏) . (25.3.7)

We shall say that 𝑓 is strictly convex if (25.3.7) holds with < replacing ≤. We recall
classical properties of convex functions.

Lemma 25.3.4 For 𝑓 : 𝐼 ↦→ R to be convex it is necessary and sufficient that

𝑓 (𝑎+ 𝛿)− 𝑓 (𝑎)
𝛿 be an increasing function of 𝛿 ≠ 0, whatever 𝑎 ∈ I such that 𝑎 + 𝛿 ∈ I.

Proof Putting 𝑏 − 𝑎 = 𝛿 > 0 in (25.3.7) we get

𝑓 (𝑎 + 𝜆𝛿) − 𝑓 (𝑎) 𝑓 (𝑎 + 𝛿) − 𝑓 (𝑎)

∀𝜆 ∈ (0, 1) , ≤ . (25.3.8)
𝜆𝛿 𝛿
Conversely, putting 𝑏 = 𝑎+𝛿 in (25.3.7) yields (25.3.8). We also get, after exchanging
𝜆 and 1 − 𝜆,
1102 25 Solvability of Constant Vector Fields of Type (1,0)

𝑓 (𝑏) − 𝑓 (𝑏 − 𝛿) 𝑓 (𝑏) − 𝑓 (𝑏 − 𝜆𝛿)

∀𝜆 ∈ (0, 1) , ≤ . □
𝛿 𝛿𝜆

Corollary 25.3.5 Every convex function in I is locally Lipschitz continuous, mean-

ing that its distribution derivative is locally 𝐿 ∞ in I.

Corollary 25.3.6 For 𝑓 ∈ C 1 (I; R) to be convex it is necessary and sufficient that

𝑓 ′ be monotone increasing. For 𝑓 ∈ C 2 (I; R) to be convex it is necessary and
sufficient that 𝑓 ′′ ≥ 0.

The property of functions of interest to us is a weakening of convexity.

Definition 25.3.7 A continuous real-valued function 𝑓 in an open subset A of R

is said to be quasiconvex if given any compact segment [𝑡1 , 𝑡2 ] ⊂ A we have
𝑓 (𝑡) ≤ max ( 𝑓 (𝑡 1 ) , 𝑓 (𝑡 2 )) for every 𝑡 ∈ [𝑡1 , 𝑡2 ]. A continuous real-valued function
𝑓 in a topological space L homeomorphic to A shall be called quasiconvex if there
is a homeomorphism 𝜒 : A −→ L such that 𝑓 ◦ 𝜒 is quasiconvex in A.

In other words, 𝑓 is quasiconvex if 𝑓 does not have a local maximum. Among

continuous functions, convex functions and monotone functions are quasiconvex;
strictly concave functions ( 𝑓 is strictly concave if − 𝑓 is strictly convex) are not.
For arbitrary 𝑐 ∈ R the set Σ𝑐 (A, 𝑓 ) = {𝑡 ∈ A; 𝑓 (𝑡) < 𝑐} is referred to as a
sublevel of 𝑓 in A.

Proposition 25.3.8 Let I ⊂ R be an open interval and the function 𝑓 : 𝐼 ↦→ R be

continuous. For 𝑓 to be quasiconvex it is necessary and sufficient that every sublevel
set of 𝑓 in I be connected.

Proof The condition is necessary. Suppose Σ𝑐 (I, 𝑓 ) is not connected; there are
points 𝑡◦ , 𝑡 ′, 𝑡 ′′ ∈ I, 𝑡 ′ < 𝑡◦ < 𝑡 ′′ such that 𝑓 (𝑡◦ ) ≥ 𝑐, and 𝑓 (𝑡 ′) < 𝑐, 𝑓 (𝑡 ′′) < 𝑐,
implying that I is not quasiconvex. The condition is sufficient. Let [𝑡 1 , 𝑡2 ] ⊂ I,
−∞ < 𝑡1 < 𝑡2 < +∞. Suppose there were 𝑡◦ ∈ [𝑡1 , 𝑡2 ] such that 𝑓 (𝑡 ◦ ) > 𝑀 =
max ( 𝑓 (𝑡1 ) , 𝑓 (𝑡2 )). If 𝑀 < 𝑐 < 𝑓 (𝑡◦ ) the set Σ𝑐 (I, 𝑓 ) would not be connected. □

Proposition 25.3.9 Let A ⊂ R be an open set and the function 𝑓 : A −→ R be

continuous. The following two properties are equivalent:
(a) 𝑓 is not quasiconvex in A;
(b) A contains two compact segments [𝑎 0 , 𝑏 0 ], [𝑎 1 , 𝑏 1 ] such that [𝑎 0 , 𝑏 0 ] ⊂ (𝑎 1 , 𝑏 1 )
and 𝑓 ≡ 𝑀 in [𝑎 0 , 𝑏 0 ], 𝑓 < 𝑀 in [𝑎 1 , 𝑎 0 ) ∪ (𝑏 0 , 𝑏 1 ].

Proof If 𝑓 is not quasiconvex we can find [𝑎, 𝑏] ⊂ A such that max ( 𝑓 (𝑎) , 𝑓 (𝑏)) <
max 𝑓 = 𝑀. Let 𝑎 1 > 𝑎, 𝑏 1 < 𝑏 0 be such that 𝑓 ≡ 𝑀 in [𝑎 1 , 𝑏 1 ] and 𝑏 1 − 𝑎 1
[𝑎,𝑏]
is maximal (although possibly zero) for these properties. By continuity there are
numbers 𝑎 0 ∈ (𝑎, 𝑎 1 ), 𝑏 0 ∈ (𝑏 1 , 𝑏) sufficiently close to 𝑎 1 and 𝑏 1 , respectively, that
the properties in the statement hold. That (b)=⇒(a) follows directly from Definition
25.3.7. □
25.3 Local Solvability at the Boundary. Final Characterization 1103

The next statement has important connections with the results in the next chapter.
Proposition 25.3.10 Let I ⊂ R be an open interval. For 𝑓 ∈ C 1 (I; R) to be
quasiconvex it is necessary and sufficient that, for every pair of points 𝑠 < 𝑡 of I,
𝑓 ′ (𝑠) > 0 implies 𝑓 ′ (𝑡) ≥ 0.
Proof If there are points 𝑠 < 𝑡 of I such that 𝑓 ′ (𝑠) > 0 and 𝑓 ′ (𝑡) < 0 there is
necessarily 𝑠 ′ > 𝑠, 𝑡 ′ ∈ (𝑠 ′, 𝑡), such that 𝑓 (𝑠 ′) > 𝑓 (𝑠), 𝑓 (𝑡 ′) > 𝑓 (𝑡) and therefore 𝑓
is not quasiconvex in I. Conversely, if 𝑓 is not quasiconvex in I we apply Proposition
25.3.9: if 𝑎 0 < 𝑎 1 ≤ 𝑏 1 < 𝑏 0 are as in (b) there are points 𝑡 1 ∈ (𝑎 0 , 𝑎 1 ), 𝑡2 ∈ (𝑏 1 , 𝑏 0 ]
where 𝑓 ′ (𝑡1 ) > 0 and 𝑓 ′ (𝑡 2 ) < 0. □

25.3.2 Local holomorphic solvability at the boundary when 𝒄 1,1 = 1.

Necessity of quasiconvexity

We continue to deal with a domain Ω in C𝑛 defined in a neighborhood 𝑈 of 0 ∈

𝜕Ω by the inequality 𝜌 (𝑧) = 𝜙 (𝑧 ′, 𝑥 𝑛 ) − 2𝑦 𝑛 < 0, and to use the notation 𝜃 =
(𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) ∈ C𝑛−2 × R with the understanding that 𝜃 = 𝑥2 when 𝑛 = 2. In this
subsection we reason under the hypothesis that 𝜌 ∈ C 2 (𝑈) and that 𝑐 1,1 = 1, 𝑐 1, 𝑗 = 0
if 1 < 𝑗 < 𝑛 [cf. (25.3.1) when 𝑛 ≥ 3, (25.3.2) when 𝑛 = 2]. Thus 𝜙 = 𝜙min + 𝜓 2 with
𝜙min (𝑦 1 , 𝜃) given by (25.3.3) [(25.3.4) when 𝑛 = 2] and 𝜓 (𝑧1 , 𝜃) given by (25.3.5).
It is convenient to take the neighborhood 𝑈 of the origin of the form

𝑧 ∈ C𝑛 ; |𝑥 1 | < 𝑟 1 , |𝑦 1 | < 𝑟 1′ , |𝜃| < 𝑟, |𝑦 𝑛 | < 𝑟 𝑛 .

(25.3.9)

The condition alluded to in the heading is the following:

(QCX) If 𝑟 1′ > 0 and 𝑟 > 0 are sufficiently small and |𝜃| < 𝑟 then −𝑟 1′ , 𝑟 1′ ∋ 𝑦 1 ↦→

𝜙min (𝑦 1 , 𝜃) is quasiconvex.
In this subsection we prove that (QCX) holds when 𝜕/𝜕𝑧 1 is locally holomorphi-
cally solvable in Ω at 0 ∈ 𝜕Ω under the hypotheses about 𝜌 stated above. We are
going to make use of the following
Lemma 25.3.11 Suppose 𝑐 1,1 = 1, 𝑐 1, 𝑗 = 0 if 1 < 𝑗 < 𝑛 in (25.3.1) when 𝑛 ≥ 3,
(25.3.2) when 𝑛 = 2. Then there is a neighborhood 𝑈 of 0 in C𝑛 such that, for every
𝑤 ∈ 𝑈 ∩ 𝜕Ω, the polynomial
1 𝜕𝜌
𝑃 (𝑤; 𝑧) = (𝑧 − 𝑤) · 𝜕𝑧 𝜌 (𝑤) − 𝑖 (𝑧 1 − 𝑤 1 ) 2 (𝑤) (25.3.10)
2 𝜕𝑧 𝑛
does not vanish at any point 𝑧 ∈ 𝑈 ∩ Ω.
Proof By (25.3.1) we have
𝑛−1
∑︁
2
𝜌 (𝑧) = −2𝑦 𝑛 + 2𝑥 12 + 2 (𝑎𝑥1 − 𝑏𝑦 1 ) 𝑥 𝑛 + 𝑧 𝑗 + 𝑜 |𝑧 ′ | 2 + 𝑥 𝑛2 ,
𝑗=2
1104 25 Solvability of Constant Vector Fields of Type (1,0)

where 𝑧 ′ = (𝑧1 , ..., 𝑧 𝑛−1 ); we derive

𝜕𝜌 𝜕𝜌
(𝑤) = 𝑂 (|𝑤|) , 𝑗 = 1, ..., 𝑛 − 1, (𝑤) = 𝑖 + 𝑂 (|𝑤|) , (25.3.11)
𝜕𝑧 𝑗 𝜕𝑧 𝑛
𝜕2 𝜌 𝜕2 𝜌
2
(𝑤) = 1 + 𝑜 (1) , (𝑤) = 𝑜 (1) , 1 ≤ 𝑗, 𝑘 ≤ 𝑛.
𝜕𝑧1 𝜕𝑧 𝑗 𝜕𝑧 𝑘
𝜕2 𝜌
(𝑤) = 𝛿 𝑗,𝑘 + 𝑜 (1) , 1 ≤ 𝑗, 𝑘 ≤ 𝑛 − 1.
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘

Since 𝜌 (𝑤) = 0, Taylor expansion about 𝑤 yields

𝜌 (𝑧) = 2 Re ((𝑧 − 𝑤) · 𝜕𝑧 𝜌 (𝑤))

1
+ Re (𝑧1 − 𝑤 1 ) 2 + |𝑧 ′ − 𝑤 ′ | 2 + 𝑜 |𝑧 ′ − 𝑤 ′ | 2 .
2
It follows from (25.3.11) that 𝑃 (𝑤; 𝑧) = 0 entails

2 Re ((𝑧 − 𝑤) · 𝜕𝑧 𝜌 (𝑤)) = − (1 + 𝑂 (|𝑤|)) Re (𝑧1 − 𝑤 1 ) 2 ,

whence
1 ′
𝜌 (𝑧) = |𝑧 − 𝑤 ′ | 2 + 𝑂 (|𝑤|) (𝑧1 − 𝑤 1 ) 2 + 𝑜 |𝑧 ′ − 𝑤 ′ | 2 ,
2
implying 𝜌 (𝑧) > 0 provided |𝑧 ′ − 𝑤 ′ | and |𝑤| are sufficiently small. □

Theorem 25.3.12 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a C 2

boundary 𝜕Ω and let 0 ∈ 𝜕Ω. Suppose there is a neighborhood 𝑈 of 0 in C𝑛 such
that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} where 𝜌 (𝑧) = 𝜙 (𝑧1 , 𝜃) − 2𝑦 𝑛 ∈ C 2 (𝑈), 𝜙 given
by (25.3.1) when 𝑛 ≥ 3, with 𝑐 1,1 = 1, 𝑐 1, 𝑗 = 0 if 1 < 𝑗 < 𝑛, by (25.3.2) when 𝑛 = 2,
with 𝑐 1,1 = 1. If 𝜕/𝜕𝑧1 induces a surjective derivation of O0 (Ω) then Condition
(QCX) holds in a set (25.3.9).

Proof We assume 𝑈 given by (25.3.9). We shall reason under the hypothesis that
(QCX) does not hold:
(∼ QCX) To every pair of positive numbers 𝛿1 < 𝑟 1′ , 𝛿 < 𝑟, there is a 𝜃 =
(𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) ∈ C𝑛−2 × R, |𝜃| < 𝛿, such that (−𝛿1 , 𝛿1 ) ∋ 𝑦 1 ↦→
𝜙min (𝑦 1 , 𝜃) is not quasiconvex,
and deduce from it that the equation 𝜕𝑧 𝜕𝑢
1
= 𝑃 (𝑤; 𝑧) −1 , with 𝑃 (𝑤; 𝑧) given by
(25.3.10) and 𝑤 ∈ 𝜕Ω ∩ 𝑈 suitably chosen, has no holomorphic solution in Ω ∩ 𝑉
whatever the neighborhood 𝑉 ⊂ 𝑈 of 0. Henceforth we assume that 𝜃 satisfies the
requirements in (∼ QCX) for arbitrary small 𝛿1 , 𝛿.
We apply Lemma 25.2.9: 𝜓 (𝑧 1 , 𝜃) = 0 determines 𝑥1 unambiguously in terms of
(𝑦 1 , 𝜃) and (25.2.13) implies that
25.3 Local Solvability at the Boundary. Final Characterization 1105

Λ (𝜃) = {𝑧 ∈ 𝑈; 𝜓 (𝑧1 , 𝜃) = 0, (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) = 𝜃}

is a two-dimensional C 1 submanifold of 𝑈, in which we can use the coordinates 𝑦 1 ,

𝑦 𝑛 . The intersection Λ (𝜃) ∩Ω∩𝑈 [resp., Λ (𝜃) ∩ 𝜕Ω∩𝑈] is defined by the inequality
𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 [resp., the equation 𝜙min (𝑦 1 , 𝜃) = 2𝑦 𝑛 ]; thus Λ (𝜃) ∩ 𝜕Ω ∩ 𝑈 is
a one-dimensional (embedded) C 1 submanifold of 𝑈 parametrized by 𝑦 1 . We take
𝑤 = 𝑠 + 𝑖𝑡 to belong to a suitably selected arc of curve 𝔠 (𝜃) ⊂ Λ (𝜃) ∩ 𝜕Ω ∩ 𝑈. As
just pointed out this implies (𝑤 2 , ..., 𝑤 𝑛−1 , 𝑠 𝑛 ) = 𝜃, 𝜓 (𝑤 1 , 𝜃) = 0, 𝜙min (𝑡1 , 𝜃) = 2𝑡 𝑛 .
To determine 𝔠 (𝜃) we avail ourselves of Proposition 25.3.9: we select two seg-
ments [𝑎 0 , 𝑏 0 ] ⊂ (𝑎 1 , 𝑏 1 ) ⊂⊂ (−𝛿1 , 𝛿1 ) satisfying the following two conditions:

(1) If 𝑦 1 ∈ [𝑎 1 , 𝑏 1 ] then 𝑧 = 𝑥1 + 𝑖𝑦 1 , 𝜃 + 21 𝑖𝜙min (𝑦 1 , 𝜃) 𝒆 𝑛 ∈ 𝜕Ω ∩ 𝑈, 𝒆 𝑛 =
(0, ..., 0, 1) ∈ C𝑛−1 ;
(2) There is an 𝑀 (𝜃) ∈ R such that 𝜙min (𝑦 1 , 𝜃) = 𝑀 (𝜃) if 𝑦 1 ∈ [𝑎 0 , 𝑏 0 ],
𝜙min (𝑦 1 , 𝜃) < 𝑀 (𝜃) if 𝑦 1 ∈ [𝑎 1 , 𝑏 1 ] \ [𝑎 0 , 𝑏 0 ].
We define 𝔠 (𝜃) ⊂ Λ (𝜃) ∩ 𝜕Ω ∩ 𝑈 as the curve

1
[𝑎 0 , 𝑏 0 ] ∋ 𝑡 1 → 𝑤 1 , 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛
2

with 𝑠1 satisfying 𝜓 (𝑠1 + 𝑖𝑡1 , 𝜃) = 0. On 𝔠 (𝜃) we have 𝜙min (𝑡 1 , 𝜃) = 𝑀 (𝜃); 𝔠 (𝜃)

can be a single point (necessarily so if 𝜙min is analytic). To help visualize 𝔠 (𝜃) in
𝑧-space C𝑛 we point out that 𝔠 (𝜃) is contained in the intersection of Λ (𝜃) ∩ 𝜕Ω ∩ 𝑈
with the affine complex line (parallel to the 𝑧1 -plane)

1
𝐿 (𝜃) = 𝑧 ∈ C𝑛 ; (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) = 𝜃, 𝑦 𝑛 = 𝑀 (𝜃) .
2

For use below we note that, on 𝔠 (𝜃),

𝜕𝜌 1 𝜕𝜙min 𝜕𝜓
(𝑤 1 , 𝜃) = − 𝑖 (𝑡1 , 𝜃) + 𝜓 (𝑤 1 , 𝜃) (𝑤 1 , 𝜃) = 0, (25.3.12)
𝜕𝑧1 2 𝜕𝑦 1 𝜕𝑧1
𝜕𝜌 1 𝜕𝜙min 𝜕𝜓
(𝑤 1 , 𝜃) = 𝑖 + (𝑡 1 , 𝜃) + 𝜓 (𝑤 1 , 𝜃) (𝑤 1 , 𝜃)
𝜕𝑧 𝑛 2 𝜕𝑥 𝑛 𝜕𝑥 𝑛
1 𝜕𝜙min
=𝑖+ (𝑡 1 , 𝜃) .
2 𝜕𝑥 𝑛
Next, for fixed 𝜃 and arbitrarily small 𝜀 > 0 we introduce the complex line,
parallel to 𝐿 (𝜃),

1
𝐿 𝜀 (𝜃) = 𝑧 ∈ C𝑛 ; (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) = 𝜃, 𝑦 𝑛 = 𝑀 (𝜃) + 𝜀
2

in which we use the coordinate 𝑧 1 . We see that Ω 𝜀 (𝜃) = 𝐿 𝜀 (𝜃) ∩ Ω ∩ 𝑈 is defined

in 𝐿 𝜀 (𝜃) by the inequality
1106 25 Solvability of Constant Vector Fields of Type (1,0)

𝜙 (𝑧1 , 𝜃) = 𝜙min (𝑦 1 , 𝜃) + 𝜓 2 (𝑧 1 , 𝜃) < 𝑀 (𝜃) + 2𝜀. (25.3.13)

The image of 𝔠 (𝜃) under the orthogonal projection C𝑛 −→ 𝐿 𝜀 (𝜃),

𝔠 𝜀 (𝜃) = {𝑧 1 ∈ C; 𝜓 (𝑧1 , 𝜃) = 0, 𝑎 0 ≤ 𝑦 1 ≤ 𝑏 0 } ,

is a compact subset of Ω 𝜀 (𝜃). We select 𝑎 ∈ (𝑎 1 , 𝑎 0 ), 𝑏 ∈ (𝑏 0 , 𝑏 1 ). If 𝜀, 𝑎 0 − 𝑎,

𝑏 0 − 𝑏 are small enough then the closed set

𝔠 𝜀 (𝜃) = {𝑧 1 ∈ C; 𝜓 (𝑧1 , 𝜃) = 0, 𝑎 ≤ 𝑦 1 ≤ 𝑏}
b

is also a C 1 curve in Ω 𝜀 (𝜃) since

𝔠 𝜀 (𝜃) , 𝑦 1 ∈ [𝑎, 𝑏]\ [𝑎 0 , 𝑏 0 ] =⇒ 𝜙 (𝑧 1 , 𝜃) < 𝑀 (𝜃) .

∀𝑧1 ∈ b (25.3.14)

We shall denote by 𝑧 1 (𝑎, 𝜃), 𝑧1 (𝑏, 𝜃) the end points of b

𝔠 𝜀 (𝜃) [viewed as a curve in
an open subset of C].
Let 𝑤 ∈ 𝔠 (𝜃) be given and suppose that the equation 𝜕𝑧 𝜕𝑢
1
= 𝑃 (𝑤; 𝑧) −1 has a
holomorphic solution in ΩØ ∩ N where N ⊂ 𝑈 is a neighborhood of 𝔠 (𝜃) in C𝑛𝑧
containing the closure of Ω 𝜀 (𝜃). The difference
0< 𝜀 ≪1

1 1
𝑢 𝑧1 (𝑏, 𝜃) , 𝜃 + 𝑖 𝑀 (𝜃) + 𝜀 𝒆 𝑛 − 𝑢 𝑧 1 (𝑎, 𝜃) , 𝜃 + 𝑖 𝑀 (𝜃) + 𝜀 𝒆 𝑛
2 2
(25.3.15)
is equal to the integral (oriented by 𝑎 ≤ 𝑦 1 ↗ 𝑏)
∫
d𝑧1
. (25.3.16)
𝔠 𝜀 𝑃 𝑤; 𝑧 , 𝜃 + 𝑖 1 𝑀 (𝜃) + 𝜀 𝒆
1
b
2 𝑛

We know, by Lemma 25.3.11, that 𝑃 𝑤; 𝑧 1 , 𝜃 + 𝑖 21 𝑀 (𝜃) + 𝜀 𝒆 𝑛 ≠ 0 in Ω 𝜀 (𝜃).
On the one hand, if 𝜀 ↘ 0 (25.3.14)
implies that (25.3.15) converges in C since
both 𝑧1 (𝑎, 𝜃) , 𝜃 + 21 𝑖𝑀 (𝜃) 𝒆 𝑛 , 𝑧1 (𝑏, 𝜃) , 𝜃 + 21 𝑖𝑀 (𝜃) 𝒆 𝑛 belong to Ω ∩ N . On
the other hand, 𝑤 = 𝑠 + 𝑖𝑡 ∈ 𝔠 (𝜃) implies

1 𝜕𝜙min
(𝑤 2 , ..., 𝑤 𝑛−1 , 𝑠 𝑛 ) = (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) = 𝜃, 𝑡 𝑛 = 𝑀 (𝜃) , (𝑡1 , 𝜃) = 0.
2 𝜕𝑦 1
By (25.3.10) and (25.3.12) we obtain

1 1
𝑃 𝑤; 𝑧1 , 𝜃 + 𝑖 𝑀 (𝜃) + 𝜀 𝒆 𝑛 = 𝜅 (𝑡1 , 𝑤) (𝑧1 − 𝑤 1 ) 2 − 𝜀 ,
2 2

with 𝜅 (𝑡1 , 𝑤) = 1 − 12 𝑖 𝜕𝜙 min

𝜕𝑥𝑛 (𝑡 1 , 𝜃). If 𝑦 1 = 𝑡 1 ∈ [𝑎 0 , 𝑏 0 ] then 𝑥 1 = 𝑠1 since both
sides satisfy 𝜓 (𝜉 + 𝑖𝑡1 , 𝜃) = 0 and therefore 𝑧 1 = 𝑤 1 = 𝑠1 + 𝑖𝑡1 . It follows that the
integral (25.3.16) tends to ∞ as 𝜀 ↘ 0. We reach a contradiction.
25.3 Local Solvability at the Boundary. Final Characterization 1107

Hypothesis (∼ QCX) allows us to select a sequence of curves 𝔠 𝜃 (𝜈) ⊂
𝜈=1,2,...
𝜕Ω∩𝑈 analogous
to 𝔠 (𝜃) with 𝜃 (𝜈)
→ 0, such that there is no neighborhood N𝜈 ⊂ 𝑈
of 𝔠 𝜃 (𝜈) 𝜕𝑢
in C where the equation 𝜕𝑧
𝑛
1
= 𝑃 (𝑤; 𝑧) −1 is holomorphically solvable.
Since Ω is pseudoconvex in the neighborhood of 0 the same argument as in the proof
of Proposition 25.2.3 applies and we reach the conclusion that if (∼ QCX) holds the
derivation of O0 (Ω) induced by 𝜕/𝜕𝑧1 is not surjective. □

25.3.3 Local holomorphic solvability at an analytic boundary when

𝒄 1,1 = 1. Sufficiency of Condition (QCX)

In this subsection we prove the converse of Theorem 25.3.12 under the hypothesis of
analyticity of the boundary 𝜕Ω. We continue to use the notation 𝜃 = (𝑧 2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 )
if 𝑛 ≥ 3, 𝜃 = 𝑥 𝑛 if 𝑛 = 2.
Theorem 25.3.13 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a C 𝜔
boundary 𝜕Ω and let 0 ∈ 𝜕Ω. Suppose there is a neighborhood 𝑈 of 0 in C𝑛 such
that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} where 𝜌 (𝑧) = 𝜙 (𝑧 1 , 𝜃) − 2𝑦 𝑛 ∈ C 𝜔 (𝑈), 𝜙 given
by (25.3.1) when 𝑛 ≥ 3, with 𝑐 1,1 = 1, 𝑐 1, 𝑗 = 0 if 1 < 𝑗 < 𝑛, by 25.3.2) when 𝑛 = 2,
with 𝑐 1,1 = 1. If (QCX) holds then 𝜕/𝜕𝑧 1 induces a surjective derivation of O0 (Ω).
Proof Before engaging in the proof we go back to the definitions of 𝜙, (25.2.11)
when 𝑛 ≥ 3, (25.2.12) when 𝑛 = 2, and of 𝜓, (25.2.13): by Lemma 25.2.9, if
𝜌 ∈ C 𝜔 (𝑈) then 𝜓 (𝑧 1 , 𝜃) and 𝜙min (𝑦 1 , 𝜃) = 𝜌 (𝑧) − 𝜓 2 (𝑧 1 , 𝜃) + 2𝑦 𝑛 belong to
C 𝜔 (𝑈). This said, we take 𝑈 as in (25.3.9). If 𝜋1 : 𝑧 ↦→ (𝑦 1 , 𝑧 ′′) [𝑧 ′′ = (𝑧2 , ..., 𝑧 𝑛 )]
−1
is the coordinate projection then Ω ∩ 𝑈 ∩ 𝜋1 (𝑦 1 , 𝑧 ′′) is defined by the inequality

𝜓 2 (𝑧1 , 𝜃) < 2𝑦 𝑛 − 𝜙min (𝑦 1 , 𝜃) . (25.3.17)

−1
A direct consequence of (25.3.5) and (25.3.17) is that Ω ∩ 𝑈 ∩ 𝜋1 (𝑦 1 , 𝑧 ′′) can
be identified with an interval (𝛼 (𝑦 1 , 𝑧 ′′) , 𝛽 (𝑦 1 , 𝑧 ′′)) in the 𝑥 1 -axis, empty unless
𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 , otherwise containing a point 𝑥1 such that 𝜓 (𝑥1 + 𝑖𝑦 1 , 𝜃) = 0. The
latter fact implies that 𝜋1 (Ω ∩ 𝑈) is defined by the inequality 𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 . We
avail ourselves of (QCX) and apply Proposition 25.3.8; if 𝜋 ′′ is the coordinate projec-
tion 𝑧 ↦→ 𝑧 ′′ and 𝑧 ′′ ∈ 𝜋 ′′ (Ω ∩ 𝑈) then the sublevel set {𝑦 1 ∈ R; 𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 }
−1
is connected. We reach the conclusion that 𝜋 ′′ (𝑧 ′′) is a connected and simply con-
nected subset of the 𝑧 1 -plane. It is readily checked that every 𝑧 ′′◦ ∈ 𝜋 ′′ (Ω ∩ 𝑈) has a
Ù −1
neighborhood N (𝑧 ′′◦ ) ⊂ 𝜋 ′′ (Ω ∩ 𝑈) such that 𝜋 ′′ (𝑧 ′′) ≠ ∅. It remains to
𝑧 ′′ ∈N (𝑧 ′′◦ )
show that the domain 𝜋 ′′ (Ω ∩ 𝑈) ⊂ C𝑛−1 is pseudoconvex, enabling us to complete
the proof of the claim by the standard argument, used in the last part of the proof
of Theorem 25.1.17. Since every open subset of the complex plane is pseudoconvex
we shall assume 𝑛 ≥ 3 in the remainder of the proof.
1108 25 Solvability of Constant Vector Fields of Type (1,0)

For |𝑦 1 | ≤ 𝑟 1′ we define

Ω′′ (𝑦 1 ) = 𝑧 ′′ ∈ C𝑛−1 ; |𝜃| < 𝛿, |𝑦 𝑛 | < 𝑟 𝑛 , 𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 .

From (25.3.17) we derive that 𝜋 ′′ (Ω ∩ 𝑈) is the set of points 𝑧 ′′ = 𝜃 +𝑖𝑦 𝑛 𝒆 𝑛 ∈ C𝑛−1 ,

|𝜃| < 𝛿, |𝑦 𝑛 | < 𝑟 𝑛 , such that 𝜙min (𝑦 1 , 𝜃) < 2𝑦 𝑛 for some 𝑦 1 ∈ −𝑟 1′ , 𝑟 1′ ; indeed, if
such a number 𝑦 1 exists then (𝑥1 + 𝑖𝑦 1 , 𝑧 ′′) ∈ Ω ∩ 𝑈 for 𝑥 1 ∈ (𝛼 (𝑦 1 , 𝑧 ′′) , 𝛽 (𝑦 1 , 𝑧 ′′))
satisfying 𝜓 (𝑧1 , 𝜃) = 0. In other words,
Ø
𝜋 ′′ (Ω ∩ 𝑈) = Ω′′ (𝑦 1 ) .
|𝑦1 |<𝑟1′

If 𝜙min ( 𝑦˜ 1 , 𝜃) − 2𝑦 𝑛 < 𝜙min (𝑦 1 , 𝜃) − 2𝑦 𝑛 , max (|𝑦 1 | , | 𝑦˜ 1 |) ≤ 𝑟 1′ , then Ω′′ (𝑦 1 ) ⊂

Ω′′ ( 𝑦˜ 1 ); the case Ω′′ (𝑦 1 ) = ∅ is not precluded. The function

𝑚 (𝜃) = inf ′ 𝜙min (𝑦 1 , 𝜃)

|𝑦1 |<𝑟1

is locally Lipschitz continuous (Appendix, Proposition 25.A.3). We reach the fol-

lowing conclusion:
(1) 𝜋 ′′ (Ω ∩ 𝑈) is the set of points 𝑧 ′′ = 𝜃 + 𝑖𝑦 𝑛 𝒆 𝑛 ∈ C𝑛−1 , |𝑧 ′′ | < 𝑟 ′′, such that
2𝑦 𝑛 > 𝑚 (𝜃);
(2) 𝑈 ∩ 𝜕𝜋 ′′ (Ω ∩ 𝑈) is the set of points 𝑧 ′′ = 𝜃 + 𝑖𝑦 𝑛 𝒆 𝑛 ∈ C𝑛−1 , |𝑧 ′′ | ≤ 𝑟 ′′, such
that 2𝑦 𝑛 = 𝑚 (𝜃).
It follows directly from (25.3.3) that Ω′′ (𝑦 1 ), when not empty, is a domain in
C𝑛−1 strictly pseudoconvex at points 𝑧 ′′ = (𝜃, 𝑦 𝑛 ) sufficiently close to the origin.
Thanks to the hypotheses in Theorem 25.3.13 we can apply Theorem 25.A.11 in the
Appendix, with 𝑡 = 𝑦 1 , 𝐹 (𝑧, 𝑡) = 𝜙min (𝑦 1 , 𝜃) − 2𝑦 𝑛 (thus, here, 𝑧 is replaced by 𝑧 ′′)
and conclude that 𝜋 ′′ (Ω ∩ 𝑈) ⊂ C𝑛−1 is pseudoconvex. □

25.3.4 Summary and symplectic interpretation

Throughout this subsection Ω shall be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2)

with a C 𝜔 boundary 𝜕Ω ∋ 0. By combining 25.2.13, 25.2.16, Theorem 25.3.12,
25.3.13, we can state

Theorem 25.3.14 If 𝜕/𝜕𝑧 1 is transverse to 𝜕Ω at 0 then 𝜕/𝜕𝑧1 induces a surjective

derivation of O0 (Ω). Suppose that 𝜕/𝜕𝑧1 is tangent to 𝜕Ω at 0 and that there is a
neighborhood 𝑈 of 0 in C𝑛 such that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} with 𝜌 given by
(25.2.5) when 𝑛 = 2 or by (25.2.6) when 𝑛 ≥ 3. For 𝜕/𝜕𝑧 1 to induce a surjective
derivation of O0 (Ω) it is necessary and sufficient that one of the following, mutually
exclusive, conditions be satisfied:
25.3 Local Solvability at the Boundary. Final Characterization 1109
√︃
(i) 0 ≤ 𝑐 1,1 < 1, and 0 ≤ 𝑐 1,2 ≤ 1 − 𝑐21,1 if 𝑛 ≥ 3;
(ii) 𝑐 1,1 = 1, (QCX) holds, and if 𝑛 ≥ 3, 𝑐 1,2 = 0.

Based on Definitions 25.2.19, 25.2.20, and Propositions 25.2.18, 25.2.21, the

following is a restatement of Theorem 25.3.14.

Theorem 25.3.15 For 𝜕/𝜕𝑧 1 to induce a surjective derivation of O0 (Ω) it is nec-

essary and sufficient that Ω be 𝜕/𝜕𝑧1 -pseudoconvex at 0 (Definition 25.2.19) and
that either Ω be strictly 𝜕/𝜕𝑧 1 -pseudoconvex at 0 (Definition 25.2.20) or else (QCX)
hold.

Needless to say, there is no special meaning attached to the origin and 0 can be
replaced by any point 𝑧◦ ∈ 𝜕Ω.
We return to the interpretation of the conditions in Theorem 25.3.15 in symplectic
terms, focusing on the case in which Ω is 𝜕/𝜕𝑧1 -pseudoconvex but not strictly 𝜕/𝜕𝑧1 -
pseudoconvex at 0. If we assume that the defining function of Ω in a neighborhood
of 0 is given by (25.2.5) when 𝑛 = 2 or by (25.2.6) when 𝑛 ≥ 3, this corresponds to
the case 𝑐 1,1 = 1 and, if 𝑛 ≥ 3, 𝑐 1,2 = 0. We shall content ourselves with dealing with
the cases 𝑛 ≥ 3. In a neighborhood 𝑈 of 0 ∈ 𝜕Ω we define 𝜙 (𝑧 ′, 𝑥 𝑛 ) = 𝜌 (𝑧) + 2𝑦 𝑛 ;
by (25.2.7) we have
𝑛−1
∑︁
2
𝜙 (𝑧 ′, 𝑥 𝑛 ) = 2𝑥12 + 𝑧 𝑗 + 2𝑥 𝑛 (𝑎𝑥1 − 𝑏𝑦 1 ) + 𝑂 |𝑧 ′ | 3 + |𝑥 𝑛 | 3 . (25.3.18)
𝑗=2

We write 𝜙 = 𝜙min + 𝜓 2 with 𝜙min given by (25.3.3) and 𝜓 given by (25.3.5). For the
convenience of the reader we recall these formulas here

𝜙min (𝑦 1 , 𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) (25.3.19)

𝑛−1 𝑛−1
1 ∑︁ 2
∑︁ 3
= − 𝑎 2 𝑥 𝑛2 − 2𝑏𝑦 1 𝑥 𝑛 + 𝑧 𝑗 + 𝑂 ­ |𝑦 1 | 3 + 𝑧 𝑗 + |𝑥 𝑛 | 3 ® ,
© ª
2 𝑗=2 𝑗=2
« ¬
√ 1
𝜓 (𝑧 ′, 𝑥 𝑛 ) = 2𝑥 1 + √ 𝑎𝑥 𝑛 + 𝑂 |𝑧 ′ | 2 + 𝑥 𝑛2 . (25.3.20)
2
We take the neighborhood 𝑈 of 0 in the form
 𝑛−1
∑︁ 

 2 

𝑈 = 𝑧 ∈ C𝑛 ; |𝑥1 | < 𝑟 1 , |𝑦 1 | < 𝑟 1′ , 𝑧 𝑗 + 𝑥 𝑛2 < 𝑟 2 , |𝑦 𝑛 | < 𝑟 𝑛 . (25.3.21)
 
 𝑗=2 
In this set-up Condition (QCX) reads
2
(QCX) If 𝑟 1′ , 𝑟, are sufficiently small and 𝑛−1
Í 2 2
′ ′
𝑗=2 𝑧 𝑗 + 𝑥 𝑛 < 𝑟 , then −𝑟 1 , 𝑟 1 ∋
𝑦 1 ↦→ 𝜙min (𝑦 1 , 𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) is quasiconvex.
Proposition 25.3.10 provides an equivalent statement of (QCX):
1110 25 Solvability of Constant Vector Fields of Type (1,0)
2
(Ψ𝑈) If 𝑟 1′ , 𝑟, are sufficiently small and 𝑧 𝑗 + 𝑥 𝑛2 < 𝑟 2 , then −𝑟 1′ , 𝑟 1′ ∋ 𝑦 1 ↦→
Í𝑛−1
𝑗=2
𝜕𝜙min
𝜕𝑦1 (𝑦 1 , 𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ) does not change sign from + to − at any point.

We go back to the set-up in Subsection 25.2.8; thus we transfer the analysis to the
outer conormal bundle 𝑁out∗ 𝜕Ω of 𝜕Ω (Definition 13.6.16). Given a sufficiently small

neighborhood 𝑈 of 0 in C𝑛 we define M = 𝑁out ∗ 𝜕Ω ; as pointed out in Subsection

𝑈
25.2.8 we may use 𝑥1 , ..., 𝑥 𝑛 , 𝑦 1 , ..., 𝑦 𝑛−1 , 𝜆 as coordinates in M. We regard M as a
real C 𝜔 manifold equipped with the (real) symplectic structure in which the Poisson
bracket is given by (25.2.36), implying that the Hamiltonian vector field of a C 1
function 𝑓 in M is
𝑛−1
1 ∑︁ 𝜕 𝑓 𝜕 𝜕𝑓 𝜕 𝜕𝑓 𝜕 𝜕𝑓 𝜕
𝐻𝑓 = − + − .
2 𝑗=1 𝜕𝑥 𝑗 𝜕𝑦 𝑗 𝜕𝑦 𝑗 𝜕𝑥 𝑗 𝜕𝑥 𝑛 𝜕𝜆 𝜕𝜆 𝜕𝑥 𝑛

Proposition 25.3.16 Property (Ψ𝑈) holds if and only if the symbol 𝜁1 = 𝜉1 + 𝑖𝜂1 of
−𝑖𝜕/𝜕𝑧1 restricted to M has the following property:
(ΨM) In M, 𝜂1 does not change sign from − to + along any null bicharacteristic of
𝜉1 .

Proof Recall that (𝑧, 𝜁) ∈ M if 𝑧 ∈ 𝑈 ∩ 𝜕Ω and 𝜁 = 𝜆𝜕𝑧 𝜌 (𝑧) , 𝜆 > 0. Keeping in

mind that 𝑐 1,1 = 1, 𝑐 1,2 = 0, we obtain

𝜕𝜌 𝜕𝜙min 𝜕𝜓
𝜆−1 𝜁1 = = + 2𝜓
𝜕𝑧 1 𝜕𝑧1 𝜕𝑧 1
𝜕𝜓
= 2𝑥 1 + (𝑎 + 𝑖𝑏) 𝑥 𝑛 + 2𝜓 ,
𝜕𝑧1
and therefore, since 𝜙min is independent of 𝑥 1 ,

𝜕𝜓
𝜆−1 𝜉1 = 𝜓 ,
𝜕𝑥1
1 𝜕𝜙min 𝜕𝜓
𝜆−1 𝜂1 = − −𝜓 .
2 𝜕𝑦 1 𝜕𝑦 1
𝜕𝜓
We note that 𝜉1 = 0 entails 𝜓 𝜕𝑥 1
= 0; by (25.3.20) we have, on an arbitrary null
bicharacteristic 𝔠 of 𝜉1 in M,
𝜕𝜓 √
= 2 + 𝑂 (|𝑧 ′ | + |𝑥 𝑛 |) ,
𝜕𝑥1

implying that 𝜓 ≡ 0 and 𝜂1 = − 21 𝜆 𝜕𝜙 min

𝜕𝑦1 on 𝔠 (𝜕/𝜕𝑥 1 is not tangent to 𝔠). Since
2
𝜕 𝜉1 𝜕𝜓
𝜕𝑥1 = 𝜕𝑥1 + 𝑂 (|𝑧 ′ | + |𝑥 𝑛 |) we can take 𝑦 1 as the coordinate on 𝔠 whence the
claim. □
25.3 Local Solvability at the Boundary. Final Characterization 1111

We note that, on the one hand, if (ΨM) holds then 𝜁1 = 0 entails {𝜉1 , 𝜂1 } ≤ 0;
on the other hand, if (𝑧◦ ,𝜁 ◦ ) ∈ M, 𝜁1◦ = 0 and {𝜉1 , 𝜂1 }| (𝑧 ◦ ,𝜁 ◦ ) < 0 then (ΨU) holds
in a sufficiently small neighborhood U of (𝑧◦ ,𝜁 ◦ ) in M. If 𝜕/𝜕𝑧1 is not tangent to
𝜕Ω at 𝑧◦ then 𝜁1 ≠ 0 and (ΨM) is trivial. This allows us to summarize in symplectic
terms all that has been proved about local solvability of 𝜕/𝜕𝑧 1 at a point of 𝜕Ω.
Theorem 25.3.17 Let 𝑧◦ ∈ 𝜕Ω. For 𝜕/𝜕𝑧1 to induce a surjective derivation of
O𝑧 ◦ (Ω) it is necessary and sufficient that 𝜂1 not change sign from − to + along any
null bicharacteristic of 𝜉1 in a conic neighborhood of (𝑧 ◦ , 𝜕𝑧 𝜌 (𝑧 ◦ )) in 𝑁out
∗ 𝜕Ω.

The extension of this result to a generic vector field (25.2.1), 𝑍, ensues directly
by performing a microlocal complex-analytic symplectomorphism that transforms
the symbol 𝜎 (𝑍) into 𝜁1 .
Theorem 25.3.18 Let ℎ be a holomorphic function in a neighborhood 𝑈 of 𝑧◦ ∈ 𝜕Ω
and suppose that Im (ℎ𝑍) ≠ 0 at 𝑧◦ . For 𝑍 to induce a surjective derivation of
O𝑧 ◦ (Ω) it is necessary and sufficient that Im (ℎ𝜎 (𝑍)) not change sign from −
to + along any null bicharacteristic of Re (ℎ𝜎 (𝑍)) in a conic neighborhood of
(𝑧◦ , 𝜕𝑧 𝜌 (𝑧◦ )) in 𝑁out
∗ 𝜕Ω.

25.3.5 𝒁 -convexity at a boundary point when 𝒁 has constant

coefficients

Let Ω be a domain in C𝑛 (𝑛 ≥ 2) with a C 1 boundary 𝜕Ω; throughout this subsection

𝑍 shall be a vector field (25.1.3), meaning a vector field (25.2.1) with constant
coefficients. The following definition is a localization of Definition 25.1.1) at a
boundary point of Ω . By a 𝑍-line we shall mean an affine complex line 𝐿 to which
𝑍 is tangent at every point.
Definition 25.3.19 We shall say that Ω (or 𝜕Ω) is 𝑍-convex at 𝑧 ◦ ∈ 𝜕Ω if there exists
a basis of neighborhoods {𝑈} 𝜈=1,2,... of 𝑧 ◦ in C𝑛 that have the following property:
for every 𝜈 = 1, 2, ..., and every 𝑍-line 𝐿, either 𝐿 ∩ Ω ∩ 𝑈𝜈 = ∅ or 𝐿 ∩ Ω ∩ 𝑈𝜈 is a
connected and simply connected open subset of 𝐿.
The property of 𝑍-convexity is invariant under local biholomorphisms that pre-
serve 𝑧 ◦ and the vector field 𝑍 up to a nonvanishing factor.
The next statement is simply a reformulation of Theorem 25.2.4 or, rather, of its
proof.
Proposition 25.3.20 If 𝑍 is transverse to 𝜕Ω at 𝑧◦ ∈ 𝜕Ω then Ω is 𝑍-convex at 𝑧◦ .
Remark 25.3.21 Obviously, in Proposition 25.3.20, the analogous claim could be
made about the complement of Ω. The same will not be true in the forthcoming
statements.
1112 25 Solvability of Constant Vector Fields of Type (1,0)

Inspection of the proof of Proposition 25.2.5 shows that the following was proved.
Proposition 25.3.22 Let Ω be a strictly pseudoconvex domain in C𝑛 with a C 2
boundary 𝜕Ω and let 0 ∈ 𝜕Ω. Suppose there is a neighborhood 𝑈 of 0 in C𝑛 such
that Ω ∩ 𝑈 = {𝑧 ∈ 𝑈; 𝜌 (𝑧) < 0} with 𝜌 given by (25.2.5) when 𝑛 = 2 or by (25.2.6)
when 𝑛 ≥ 3. If 𝑎 1,1 > 1 then Ω is not 𝑍-convex at 0.
The proof of Proposition 25.2.4 consists in showing that if the quadratic form
𝑛−1
∑︁ 2 𝑛−1
∑︁ 2
𝑄 (𝑧) = 𝑧 𝑗 − 𝑧 ◦𝑗 + Re 𝑐 𝑗 𝑧 𝑗 − 𝑧◦𝑗
𝑗=1 𝑗=1

is positive definite then Ω is 𝑍-convex at 𝑧◦ whatever the vector field 𝑍 [given by

(25.1.3)] tangent to 𝜕Ω at 0. Taking this fact into account inspection of the proof of
Proposition 25.2.16 shows that the following statement was proved:
Proposition 25.3.23 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a
C 𝜔 boundary 𝜕Ω and let 𝑧◦ ∈ 𝜕Ω. If Ω is strictly 𝑍-pseudoconvex at 𝑧 ◦ then Ω is
𝑍-convex at 𝑧◦ .
Keep in mind that strict 𝑍-pseudoconvexity subsumes 𝑍-pseudoconvexity.
Proposition 25.3.24 Let Ω be as in Proposition 25.3.23 and let 𝜌 be a C 𝜔 defining
function of Ω in a neighborhood 𝑈 of 0 ∈ 𝜕Ω. Let the vector field 𝜕/𝜕𝑧1 be tangent
to 𝜕Ω at 0. If Ω is 𝜕/𝜕𝑧1 -pseudoconvex at 0 then the 𝜕/𝜕𝑧 1 -convexity of Ω at 0 is
equivalent to the property that either Ω is strictly 𝜕/𝜕𝑧1 -pseudoconvex at 0 or else
(QCX) holds in a set (25.3.9).
Proof Proposition 25.3.23 states that if Ω is strictly 𝜕/𝜕𝑧 1 -pseudoconvex at 0 then
Ω is 𝜕/𝜕𝑧1 -convex at 0. The first step in the proof of Theorem 25.3.13 was to prove
−1
that every fiber 𝜋 ′′ (𝑧 ′′) [𝑧 ′′ = (𝑧2 , ..., 𝑧 𝑛 ) ∈ 𝜋 ′′ (Ω ∩ 𝑈)] is a connected and simply
connected subset of the 𝑧1 -plane. Thus if Ω is 𝜕/𝜕𝑧1 -pseudoconvex at 0 and if (QCX)
holds then Ω is 𝜕/𝜕𝑧1 -convex at 0.
In the notation of (25.2.11)–(25.2.12) we are left with proving that if Ω is 𝜕/𝜕𝑧 1 -
convex at 0 then (QCX) holds, in the case 𝑐 1,1 = 1 when 𝑛 = 2, 𝑐 1,1 = 1, 𝑐 1,2 = 0
when 𝑛 ≥ 3 (since 𝜕/𝜕𝑧 1 -pseudoconvexity at 0 entails 𝑐21,1 + 𝑐21,2 ≤ 1). We will
prove that (∼ QCX) implies that Ω is not 𝜕/𝜕𝑧1 -convex at 0. We limit ourselves to
dealing with the cases 𝑛 ≥ 3; the case 𝑛 = 2 is simpler and dealt with in a similar
fashion.
We now extract a proof of our claim from the proof of Theorem 25.3.12 under the
hypothesis that 𝜕Ω is of class C 𝜔 . Assuming that 𝑈 is given by (25.3.9) we make
use of the defining function

𝜌 (𝑧) = −2𝑦 𝑛 + 𝜙min (𝑦 1 , 𝜃) + 𝜓 2 (𝑧1 , 𝜃) ∈ C 𝜔 (𝑈)

where 𝜃 = (𝑧2 , ..., 𝑧 𝑛−1 , 𝑥 𝑛 ), 𝜙min is given by (25.3.3) and 𝜓 by (25.3.5). Provided
𝑟 1′ and 𝑟 are sufficiently small the equation 𝜓 (𝑧1 , 𝜃) = 0 has a unique C 𝜔 solution
𝑥1 = 𝜑 (𝑦 1 , 𝜃) ∈ (−𝑟 1 , 𝑟 1 ) such that 𝜑 (0, 0) = 0 in the domain
25.3 Local Solvability at the Boundary. Final Characterization 1113

(𝑦 1 , 𝜃) ∈ R × C𝑛−1 × R; |𝑦 1 | < 𝑟 1′ , |𝜃| < 𝑟 .

By Proposition 25.3.1 the preimages of points under the coordinate projection 𝜋1 :

Ω ∩ 𝑈 ∋ 𝑧 ↦→ (𝑦 1 , 𝑧2 , ..., 𝑧 𝑛 ) ∈ R × C𝑛−1 are closed intervals defined by 𝜓 2 (𝑧1 , 𝜃) ≤
2𝑦 𝑛 − 𝜙min (𝑦 1 , 𝜃) centered approximately at − 21 𝑎𝑥 𝑛 [cf. (25.2.19)]. The hypothesis
(∼ QCX) implies that there exist a segment [𝑎 1 , 𝑏 1 ] ⊂ (−𝛿1 , 𝛿1 ) and a point 𝑦 1 (𝜃) ∈
(𝑎 1 , 𝑏 1 ) such that 𝜙min (𝑦 1 , 𝜃) < 𝜙min (𝑦 1 (𝜃) , 𝜃) = 𝑀 (𝜃) if 𝑦 1 ∈ [𝑎 1 , 𝑏 1 ], 𝑦 1 ≠
𝑦 1 (𝜃). If 𝛿 is sufficiently small we have 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛 ∈ 𝜋 ′′ Ω ∩ 𝑈 [as before, 𝜋 ′′
is the coordinate projection 𝑧 ↦→ 𝑧 ′′ = (𝑧 2 , ..., 𝑧 𝑛 )] and if 𝑦 1 ∈ [𝑎 1 , 𝑏 1 ],

𝑧 (𝑦 1 , 𝜃) = (𝜑 (𝑦 1 , 𝜃) + 𝑖𝑦 1 , 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛 ) ∈ Ω.

We have 𝑧 (𝑦 1 (𝜃) , 𝜃) ∈ 𝜕Ω since 𝑦 𝑛 = 𝜙min (𝑦 1 (𝜃) , 𝜃) [thus 𝜓 (𝑧 (𝑦 1 (𝜃) , 𝜃) , 𝜃) =

0] whereas 𝑧 (𝑦 1 , 𝜃) ∈ Ω if 𝑦 1 ∈ [𝑎 1 , 𝑏 1 ], 𝑦 1 ≠ 𝑦 1 (𝜃), since this implies
𝜙min (𝑦 1 , 𝜃) < Im 𝑧 𝑛 (𝑦 1 , 𝜃). To summarize: if (𝑧1 , 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛 ) ∈ Ω then either

𝜓 2 (𝑧1 , 𝜃) ≤ 2𝑀 (𝜃) − 𝜙min (𝑦 1 , 𝜃) , 𝑦 1 ∈ [𝑎 1 , 𝑏 1 ] , 𝑦 1 ≠ 𝑦 1 (𝜃) ,

and therefore (𝑧1 , 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛 ) ∈ Ω or else

(𝑧1 , 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛 ) = 𝑧 (𝑦 1 (𝜃) , 𝜃) ∈ 𝜕Ω.

We reach the conclusion that the set of points (𝑧1 , 𝜃 + 𝑖𝑀 (𝜃) 𝒆 𝑛 ) ∈ Ω is not con-
nected and simply connected. □
Proposition 25.3.24 allows us to restate Theorem 25.3.17 in a more general and
invariant manner.

Theorem 25.3.25 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a C 𝜔

boundary 𝜕Ω and let 𝑧◦ ∈ 𝜕Ω. For a vector field (25.1.3), 𝑍, to induce a surjective
derivation of O𝑧 ◦ (Ω) it is necessary and sufficient that Ω be 𝑍-pseudoconvex and
𝑍-convex at 𝑧◦ .

Remark 25.3.26 Since Ω can be 𝜕/𝜕𝑧 1 -pseudoconvex at 0 without being strictly

𝜕/𝜕𝑧 1 -pseudoconvex at 0 nor (QCX) holding in a set (25.3.9) it follows that 𝜕/𝜕𝑧1 -
pseudoconvexity does not imply 𝜕/𝜕𝑧 1 -convexity. Proposition 25.3.22 shows that
the converse implication is true when 𝑛 = 2, whence the next statement.

Theorem 25.3.27 Let Ω be a strictly pseudoconvex domain in C2 with a C 𝜔 bound-

ary 𝜕Ω and let 𝑧 ◦ ∈ 𝜕Ω. For a vector field (25.1.3), 𝑍, to induce a surjective
derivation of O𝑧 ◦ (Ω) it is necessary and sufficient that Ω be 𝑍-convex at 𝑧◦ .

The fact that 𝜕/𝜕𝑧 1 -convexity does not imply 𝜕/𝜕𝑧1 -pseudoconvexity if 𝑛 > 2 is
shown in the following example.
1114 25 Solvability of Constant Vector Fields of Type (1,0)

Example 25.3.28 The domain in C3

Ω = 𝑧 ∈ C3 ; |𝑧 1 | 2 + |𝑧2 | 2 + 4 Re (𝑧1 𝑧2 ) < 2𝑦 3

is not 𝜕/𝜕𝑧1 -pseudoconvex at 0 since 𝑐 1,2 = 2 in the notation of (25.2.6). But if 𝐿 is

an arbitrary 𝜕/𝜕𝑧1 -line then 𝐿 ∩ Ω is either empty, when 2𝑦 3 + 3 |𝑧 2 | 2 ≤ 0, or it is the
21
disk in 𝐿 (where the coordinate is 𝑧1 ) centered at 2𝑧¯2 with radius 2𝑦 3 + 3 |𝑧 2 | 2
(> 0).
For a generic vector field (25.1.3), 𝑍, we denote by 𝐿 ◦ the 𝑍-line through 0 and
by 𝜋 𝑍 : C𝑛 −→ C𝑛 /𝐿 ◦ the quotient map. We introduce a new definition.
Definition 25.3.29 Let Ω be a domain in C𝑛 (𝑛 ≥ 2) with a C 2 boundary 𝜕Ω and
let 𝑧◦ ∈ 𝜕Ω. Let 𝑍 be a vector field (25.1.3). We shall say that Ω is transversally
𝑍-pseudoconvex at 𝑧◦ if there exists a basis of neighborhoods {𝑈𝜈 } 𝜈=1,2,... of 𝑧 ◦ in
C𝑛 such that for every 𝜈 = 1, 2, ..., 𝜋 𝑍 (Ω ∩ 𝑈𝜈 ) is a pseudoconvex domain in C𝑛 /𝐿 ◦
( C𝑛−1 ).
When 𝑛 = 2 every domain in C𝑛 /𝐿 ◦ is pseudoconvex and Definition 25.3.29 is a
triviality. Inspection of the proofs in this section shows that the following is valid.
Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with a C 𝜔 boundary 𝜕Ω and
let 0 ∈ 𝜕Ω. Assuming that Ω is 𝜕/𝜕𝑧 1 -convex at 0 then Ω is 𝜕/𝜕𝑧1 -pseudoconvex at
0 if and only if Ω is transversally 𝜕/𝜕𝑧1 -pseudoconvex at 𝑧◦ , the invariant form of
this claim being
Proposition 25.3.30 Let 𝑍 be a vector field (25.1.3). Assuming that Ω is 𝑍-convex at
𝑧◦ then Ω is 𝑍-pseudoconvex at 𝑧◦ if and only if Ω is transversally 𝑍-pseudoconvex
at 𝑧 ◦ .
Theorem 25.3.25 can be restated as follows.
Theorem 25.3.31 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 2) with
a C 𝜔 boundary 𝜕Ω and let 𝑧◦ ∈ 𝜕Ω. For a vector field (25.1.3), 𝑍, to induce a
surjective derivation of O𝑧 ◦ (Ω) it is necessary and sufficient that Ω be 𝑍-convex
and transversally 𝑍-pseudoconvex at 𝑧◦ .

25.4 The Differential Complex. Generalities

25.4.1 The global differential complex defined by a system of constant

vector fields of type (1,0)

Let Ω be a bounded domain in C𝑛 and 𝑧◦ ∈ Ω. We now consider a system 𝒁

= (𝑍1 , ..., 𝑍 𝑚 ) of vector fields (25.1.3), linearly independent (over C; thus 𝑚 ≤ 𝑛).
All the results in this section have been proved in the preceding sections when 𝑚 = 1.
We shall therefore assume 𝑚 ≥ 2, throughout.
25.4 The Differential Complex. Generalities 1115

There are affine complex functionals 𝜆 ( 𝑗) (𝑧), 𝑗 = 1, ..., 𝑚, such that

𝑍 𝑗 𝜆 (𝑘) (𝑧) = 𝛿 𝑗,𝑘 , 𝜆 (𝑘) (𝑧 ◦ ) = 0, 𝑗, 𝑘 = 1, ..., 𝑚. (25.4.1)

Since the 𝑍 𝑗 commute they define a differential complex over Ω (Definition 9.4.2).
We introduce the ( 𝑝, 0)-forms
∑︁
𝑓 = 𝑓 𝐼 (𝑧) d𝜆 ( 𝜄1 ) ∧ · · · ∧ d𝜆 ( 𝜄 𝑝 ) (25.4.2)
|𝐼 |= 𝑝

where 𝐼 = 𝜄1 , ..., 𝜄 𝑝 is the multi-index such that 1 ≤ 𝜄1 < · · · < 𝜄 𝑝 ≤ 𝑚, 𝑓 𝐼 ∈ O (Ω),
and the differential operator
𝑚
∑︁ ∑︁
𝜕𝒁 𝑓 = 𝑍 𝑗 𝑓 𝐼 (𝑧) d𝜆 ( 𝑗) ∧ d𝜆 ( 𝜄1 ) ∧ · · · ∧ d𝜆 ( 𝜄 𝑝 ) . (25.4.3)
|𝐼 |= 𝑝 𝑗=1

Since 𝑍 𝑗 , 𝑍 𝑘 = 0 we have 𝜕𝒁2 = 0. If we denote by O𝑇 ( 𝑝,0) (Ω, 𝒁) the complex

vector space [or O (Ω)-module] consisting of the ( 𝑝, 0)-forms (25.4.2) then the
differential complex under consideration is the sequence of differential operators
𝜕𝒁 𝜕𝒁 𝜕𝒁
0 −→ O (Ω) −→ · · · −→ O𝑇 ( 𝑝,0) (Ω) −→ (25.4.4)
𝜕𝒁 𝜕𝒁
O𝑇 (0, 𝑝+1) (Ω) −→ · · · −→ O𝑇 (𝑚,0) (Ω) → 0.

Things become clearer if we carry out a C-linear change of the variables 𝑧 𝑗

transforming 𝑍 𝑗 into 𝜕𝑧𝜕 𝑗 for each 𝑗, in which case 𝜆 (𝑘) (𝑧) = 𝑧 𝑘 . Then the differential
complex we are interested in becomes the sequence of differential operators
(𝑚) (𝑚) (𝑚)
𝜕𝑧 𝜕𝑧 𝜕𝑧
0 −→ O (Ω) −→ · · · −→ O𝑇 ( 𝑝,0) (Ω) −→ (25.4.5)
(𝑚) (𝑚)
𝜕𝑧 𝜕𝑧
O𝑇 (0, 𝑝+1) (Ω) −→ · · · −→ O𝑇 (𝑚,0) (Ω) → 0

where O𝑇 ( 𝑝,0) (Ω) denotes the complex vector space [or O (Ω)-module] consisting
of the ( 𝑝, 0)-forms ∑︁
𝑓 = 𝑓 𝐼 (𝑧) d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 (25.4.6)
|𝐼 |= 𝑝

[where 𝐼 = 𝜄1 , ..., 𝜄 𝑝 , 1 ≤ 𝜄1 < · · · < 𝑖 𝑝 ≤ 𝑚 ] and
𝑚
∑︁ ∑︁ 𝜕 𝑓𝐼
𝜕𝑧(𝑚) 𝑓 = (𝑧) d𝑧 𝑗 ∧ d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 . (25.4.7)
𝜕𝑧 𝑗
|𝐼 |= 𝑝
𝑗=1

Obviously 𝜕𝑧(𝑚) ◦ 𝜕𝑧(𝑚) = 0.

Based on results in the preceding sections we can prove the global exactness of
the differential complex (25.4.3) in a C-convex domain in C𝑛 .
1116 25 Solvability of Constant Vector Fields of Type (1,0)

Theorem 25.4.1 Suppose the domain Ω in C𝑛 is C-convex and the pair (𝑚, 𝑝) ∈ Z2+
is such that 1 ≤ 𝑝 ≤ 𝑚 ≤ 𝑛. To every 𝜕𝑧(𝑚) -closed form 𝑓 ∈ O𝑇 ( 𝑝,0) (Ω) there is a
𝑢 ∈ O𝑇 ( 𝑝−1,0) (Ω) such that 𝜕𝑧(𝑚) 𝑢 = 𝑓 .

Proof We reason by induction on 𝑚. When 𝑚 = 1 we have 𝑓 = 𝑓 (1) d𝑧1 , 𝑓 (1) ∈ O (Ω)

𝜕𝑢
(with no other condition). The existence of 𝑢 ∈ O (Ω) such that 𝜕𝑧 1
= 𝑓 (1) is stated
in Theorem 25.1.17. When 𝑚 > 1 we write
∑︁ ∑︁
𝑓 = 𝑓 𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 + 𝑓 𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 ∧ d𝑧 𝑚 .
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚 |𝐼 |= 𝑝,𝑖 𝑝 =𝑚

By the induction hypothesis there is a

∑︁
𝑣= 𝑣 𝐼 ′ d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 , 𝑣 𝐼 ′ ∈ O (Ω) ,
|𝐼 ′ |= 𝑝−1,𝑖 𝑝−1 <𝑚

[where 𝐼 ′ = 𝜄1 , ..., 𝜄 𝑝−1 ] such that

∑︁
𝜕𝑧(𝑚−1) 𝑣 = 𝑓 𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 ,
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚

whence

∑︁ 𝜕𝑣 𝐼 ′
𝑓 − 𝜕𝑧(𝑚) 𝑣 = 𝑓 (𝐼 ′ .𝑚) − d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 ∧ d𝑧 𝑚 .
𝜕𝑧 𝑚
|𝐼 ′ |= 𝑝−1

Since the left-hand side is 𝜕𝑧(𝑚) -closed the ( 𝑝 − 1)-form

∑︁ 𝜕𝑣 𝐼 ′
𝑔= 𝑓 (𝐼 ′ .𝑚) − d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1
𝜕𝑧 𝑚
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚

is 𝜕𝑧(𝑚−1) -closed. By the induction hypothesis (on 𝑚) there is a

∑︁
𝑤= 𝑤 𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 , 𝑤 𝐼 ∈ O (Ω) ,
|𝐼 |= 𝑝−1,𝑖 𝑝−1 <𝑚

such that 𝜕𝑧(𝑚−1) 𝑤 = 𝑔 whence 𝑓 = 𝜕𝑧(𝑚) (𝑣 + 𝑤 ∧ d𝑧 𝑚 ). □

We denote by O𝐻 𝑝 (Ω, 𝒁), 𝑝 = 0, 1..., the cohomology groups of the differential
complex (25.4.4); reverting from 𝜕𝑧𝜕 𝑗 to 𝑍 𝑗 ( 𝑗 = 1, ..., 𝑚) allows us to restate Theorem
25.4.1 as follows:

Theorem 25.4.2 Suppose the domain Ω in C𝑛 is C-convex and the differential com-
plex (25.4.4) is defined by a system 𝒁 = (𝑍1 , ..., 𝑍 𝑚 ) of 𝑚 linearly independent
vector fields (25.1.3); then O𝐻 𝑝 (Ω, 𝒁) = 0 for every 𝑝 = 1, ..., 𝑚.
25.4 The Differential Complex. Generalities 1117

25.4.2 The differential complex in germs at boundary points. Basics

Let 𝒁 = (𝑍1 , ..., 𝑍 𝑚 ) be a system of vector fields (25.1.3), linearly independent over
C and let Ω be a domain in C𝑛 ; until specified otherwise we put no condition on
its boundary, 𝜕Ω. We are interested in the differential complex (25.4.4) localized at
a point 𝑧◦ ∈ 𝜕Ω; it is best described at the level of germs. We introduce the affine
complex functionals 𝜆 ( 𝑗) (𝑧) ( 𝑗 = 1, ..., 𝑚 ≤ 𝑛) satisfying (25.4.1), and the germs at
the germ of set (Ω, 𝑧◦ ) of the differentials forms (25.4.2) which we denote by
∑︁
f𝑧 ◦ = f𝐼 d𝜆 ( 𝜄1 ) ∧ · · · ∧ d𝜆 ( 𝜄 𝑝 ) , (25.4.8)
|𝐼 |= 𝑝

where f𝐼 ∈ O𝑧 ◦ (Ω) with O𝑧 ◦ (Ω) the stalk at 𝑧◦ of the sheaf O𝝏Ω (Ω) defined at the
(
beginning of Section 25.2. The germs f make up the stalk OT𝑧 ◦𝑝,0) (Ω, 𝒁) at 𝑧◦ of a
( 𝑝,0)
sheaf OT𝝏Ω (Ω, 𝒁). Since every 𝑍 𝑗 defines a derivation Z 𝑗 of O𝑧 ◦ (Ω) the system
( (
𝒁 defines a linear map 𝝏 Z : OT𝑧 ◦𝑝,0) (Ω, 𝒁) −→ OT𝑧 ◦𝑝+1,0) (Ω, 𝒁):
∑︁
𝝏 Z f𝑧 ◦ = Z 𝑗 f𝐼 d𝜆 ( 𝑗) ∧ d𝜆 ( 𝜄1 ) ∧ · · · ∧ d𝜆 ( 𝜄 𝑝 ) . (25.4.9)
|𝐼 |= 𝑝

We have 𝝏 2Z = 0. We propose to analyze the exactness of the differential complex

𝝏Z( 𝝏Z 𝝏Z
0 −→ O𝑧 ◦ (Ω) −→ · · · −→ OT𝑧 ◦𝑝,0) (Ω, 𝒁) −→ (25.4.10)
( 𝝏Z 𝝏Z
OT𝑧 ◦𝑝+1,0) (Ω, 𝒁) −→ · · · −→ OT𝑧(𝑚,0)
◦ (Ω, 𝒁) → 0.

An observation similar to Remark 25.2.2 is pertinent here.

Remark 25.4.3 A priori the exactness of (25.4.10) in degree 𝑝 ≥ 1 means that if

∑︁
𝑓 = 𝑓 𝐼 (𝑧) d𝜆 ( 𝜄1 ) ∧ · · · ∧ d𝜆 ( 𝜄 𝑝 ) ∈ O𝑇 ( 𝑝,0) (Ω ∩ 𝑈, 𝒁) (25.4.11)
|𝐼 |= 𝑝

(𝑈: a neighborhood of 𝑧 ◦ in C𝑛 ) satisfies

𝑚
∑︁ ∑︁
𝝏𝒁 𝑓 = 𝑍 𝑗 𝑓 𝐼 (𝑧) d𝜆 ( 𝑗) ∧ d𝜆 ( 𝜄1 ) ∧ · · · ∧ d𝜆 ( 𝜄 𝑝 ) = 0 (25.4.12)
|𝐼 |= 𝑝 𝑗=1

there is a 𝑢 ∈ O𝑇 ( 𝑝−1,0) (Ω ∩ 𝑉, 𝒁) (𝑉: a neighborhood of 𝑧◦ in C𝑛 , 𝑉 ⊂ 𝑈) such

that 𝝏 𝒁 𝑢 = 𝑓 in 𝑉. Since all the topological vector spaces O𝑇 ( 𝑝,0) (Ω ∩ 𝑈, 𝒁) are
Fréchet–Montel spaces a standard Baire category argument shows that 𝑉 can be
chosen independently of 𝑓 .

Of course, the exactness of (25.4.10) in degree 𝑝 can be stated in terms of the 𝑝 th

cohomological group: O𝑧 ◦ 𝐻 𝑝 (Ω, 𝒁) = 0.
1118 25 Solvability of Constant Vector Fields of Type (1,0)

As in the preceding subsection we carry out a C-affine

change
of variables trans-
(𝑚) 𝝏 𝝏
forming the system 𝒁 into the system 𝝏 𝑧 = 𝝏𝑧1 , ..., 𝝏𝑧𝑚 and to assume that
𝑧◦ = 0, leading to the germ analogue of (25.4.5),
(𝑚) (𝑚) 𝝏𝑧(𝑚)
𝝏𝑧 𝝏𝑧 (

0 −→ O0 (Ω) −→ · · · −→ OT0 𝑝,0) Ω, 𝝏 𝑧(𝑚) −→ (25.4.13)
𝝏𝑧(𝑚) (𝑚)
𝝏𝑧
(
OT0 𝑝+1,0) Ω, 𝝏 𝑧(𝑚) −→ · · · −→ OT0(𝑛−1,0) Ω, 𝝏 𝑧(𝑚) → 0.

We shall often refer to (25.4.13) as the differential complex 𝝏 𝑧(𝑚) . To streamline the
next statement it is helpful to revise the meaning of 𝝏 𝑧(𝑚−1) : rather than regarding
it as a differential complex over the germ of Ω at 0 it is convenient to define it as
a differential complex over the germ of 𝜋 𝑚 (Ω) at 0 where 𝜋 𝑚 is the coordinate
projection C𝑛 ∋ 𝑧 ↦→ (𝑧 1 , ..., 𝑧 𝑚−1 , 𝑧 𝑚+1 , ..., 𝑧 𝑛 ) ∈ C𝑛−1 . Note that 0 might belong to
the domain 𝜋 𝑚 (Ω) and not to its boundary 𝜕𝜋 𝑚 (Ω), in which case Property (ii) in
the following statement is trivial.

Theorem 25.4.4 Let Ω be a domain in C𝑛 such that 0 ∈ 𝜕Ω. If 𝜕/𝜕𝑧 𝑚 induces a

surjective derivation of O0 (Ω) then 𝝏 𝑧(𝑚) is exact in degree 𝑚 and the following
properties are equivalent, for every 𝑝 = 1, ..., 𝑚 − 1:

(i) The differential complex 𝝏 𝑧(𝑚) [i.e., (25.4.13)] is exact in degree 𝑝.

(ii) The differential complex 𝝏 𝑧(𝑚−1) is exact in degree 𝑝.

Proof Let 𝑝 ∈ [1, 𝑚] and

∑︁
(
f= f𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 ∈ OT0 𝑝,0) Ω, 𝝏 𝑧(𝑚)
|𝐼 |= 𝑝

be 𝝏 𝑧(𝑚) -closed; we have f = f◦ ∧ d𝑧 𝑚 + g, where

∑︁
f◦ = f𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 ,
|𝐼 |= 𝑝−1,𝑖 𝑝−1 <𝑚
∑︁
g= g𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 .
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚

If 𝜕/𝜕𝑧 𝑚 induces a surjective derivation of O𝑧 ◦ (Ω) then, to every f𝐼 there is a

v𝐼 ∈ O0 (Ω) such that f𝐼 =𝝏v𝐼 /𝝏𝑧 𝑚 and thus f◦ ∧ d𝑧 𝑚 =𝝏 𝑧(𝑚) v − h, where
∑︁
v = (−1) 𝑝−1 v𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 ,
|𝐼 |= 𝑝−1,𝑖 𝑝−1 <𝑚
𝑚−1
∑︁ ∑︁ 𝜕v𝐼
h= d𝑧 𝑗 ∧ d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 .
𝜕𝑧 𝑗
|𝐼 |= 𝑝−1,𝑖 𝑝−1 <𝑚 𝑗=1
25.4 The Differential Complex. Generalities 1119

If 𝑝 = 𝑚 necessarily h = 0 since |𝐼 | = 𝑚 − 1 =⇒ 𝑖 𝛼 = 𝛼 = 1, ..., 𝑚 − 1. This

proves that 𝝏 𝑧(𝑚) is exact in degree 𝑚.
In the remainder of the proof 𝑝 ≤ 𝑚 − 1. We have f =𝝏 𝑧(𝑚) v + g♭ with
∑︁
g♭ = g − h = g♭𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚

and 𝝏 𝑧(𝑚) g♭ = 0, implying

∑︁ 𝝏g♭𝐼
d𝑧 𝜄 ∧ · · · ∧ d𝑧 𝜄 𝑝 = 0, (25.4.14)
𝝏𝑧 𝑚 1
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚
𝑚−1
∑︁ ∑︁ 𝝏g♭𝐼
d𝑧 𝑗 ∧ d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝 = 0. (25.4.15)
𝝏𝑧 𝑗
|𝐼 |= 𝑝,𝑖 𝑝 <𝑚 𝑗=1

From (25.4.14) we derive that each coefficient g♭𝐼 is independent of 𝑧 𝑚 and from

(
(25.4.15) that the 𝑝-form g♭ is 𝝏 𝑧(𝑚−1) -closed. If there is a u♭ ∈ OT0 𝑝−1,0) Ω, 𝝏 𝑧(𝑚−1)
such that 𝝏 𝑧(𝑚−1) u♭ = g♭ then

𝝏 𝑧(𝑚−1) u♭ = 𝝏 𝑧(𝑚) u♭ = g♭ .
𝑧𝑚 =0 𝑧𝑚 =0

We reach the conclusion that f =𝝏 𝑧(𝑚) v+ u♭ , proving that (ii)=⇒(i).
𝑧𝑚 =0

( 𝑝,0)
Now assume 𝑝 ≤ 𝑚 − 1 and let g ∈ OT0 (Ω) , 𝝏 𝑧(𝑚−1) be 𝝏 𝑧(𝑚−1) -closed;
𝜋𝑚
∗ g is a ( 𝑝, 0)-form in Ω whose coefficients are independent of 𝑧 and
the pullback 𝜋 𝑚 𝑚
which is therefore 𝝏 𝑧(𝑚) -closed. Let
∑︁
(
v= v𝐼 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 ∈ OT0 𝑝−1,0) Ω, 𝝏 𝑧(𝑚)
|𝐼 |= 𝑝−1,𝑖 𝑝−1 ≤𝑚

be such that 𝝏 𝑧(𝑚) v =𝜋 𝑚

∗ g; then 𝝏 (𝑚−1) v =𝜋 ∗ g, where
𝑧 ◦ 𝑚
∑︁
v◦ = v𝐼 | 𝑧𝑚 =0 d𝑧 𝜄1 ∧ · · · ∧ d𝑧 𝜄 𝑝−1 .
|𝐼 |= 𝑝−1,𝑖 𝑝−1 <𝑚

(
We can view v◦ as an element of OT0 𝑝−1,0) 𝜋 𝑚 (Ω) , 𝝏 𝑧(𝑚−1) satisfying 𝝏 𝑧(𝑚−1) v◦ = g.
This proves that (i)=⇒(ii). □

Corollary 25.4.5 Suppose the boundary 𝜕Ω is of class C 1 . Properties (1) and (2)
hold if 𝜕/𝜕𝑧 𝑚 is transversal to 𝜕Ω.

Proof Follows from Theorem 25.4.4 and Theorem 25.2.4. □

1120 25 Solvability of Constant Vector Fields of Type (1,0)

Remark 25.4.6 The restatement of Theorem 25.4.4 in terms of an arbitrary system

𝒁 = (𝑍1 , ..., 𝑍 𝑚 ) of vector fields (25.1.3), linearly independent over C, is evident: 𝑍 𝑚
replaces 𝜕/𝜕𝑧 𝑚 and 𝜋 𝑚 must be replaced by the projection of C𝑛 onto the complex
hyperplane of C𝑛 orthogonal to 𝑍 𝑚 .

In the next two statements Ω is a domain in C𝑛 (𝑛 ≥ 3) and 𝑧◦ ∈ 𝜕Ω; there are no

conditions on 𝜕Ω.

Corollary 25.4.7 Assume that 𝑚 = 2 and the differential complex (25.4.10) is exact
in degree 1. If 𝑍2 induces a surjective derivation of O𝑧 ◦ (Ω) then the same is true of
𝑍1 .

Corollary 25.4.8 The following properties are equivalent:

(a) Each vector field 𝑍 𝑗 ( 𝑗 = 1, ..., 𝑚) induces a surjective derivation of O𝑧 ◦ (Ω).
(b) For every 𝜈, 1 ≤ 𝜈 ≤ 𝑚, and every subset ( 𝑗 1 , ..., 𝑗 𝜈 ) of (1, ..., 𝑚) the analogous

differential complex where 𝒁 = (𝑍1 , ..., 𝑍 𝑚 ) is replaced by 𝑍 𝑗1 , ..., 𝑍 𝑗𝜈 is exact
at every degree.

Proof Obviously (b)=⇒(a). To prove (a)=⇒(b) we reason by induction on 𝑚, the

claim being trivial when 𝑚 = 1. Assume that 𝑚 ≥ 2 and that (b) is valid for the
analogue of (25.4.10) with 𝑚 − 1 substituted for 𝑚. Then the validity of (b) for
(25.4.10) is a direct consequence of (a) and Theorem 25.4.4. □
Many differential complexes (such as the De Rham and the Dolbeault complexes)
satisfy their version of Property (b).

Theorem 25.4.9 Let Ω be a strictly pseudoconvex domain in C𝑛 (𝑛 ≥ 3) with a

C ∞ boundary 𝜕Ω ∋ 𝑧 ◦ . For the differential complex (25.4.10) to have Property
(b), Corollary 25.4.8, it is necessary and sufficient that Ω be 𝑍 𝑗 -pseudoconvex
(Definition 25.2.20) and 𝑍 𝑗 -convex at 𝑧◦ (Definition 25.3.19) for every 𝑗 = 1, ..., 𝑚.

Proof Combine Theorem 25.3.27 and Theorem 25.4.4. □

25.A Appendix: Minima of Families of Plurisubharmonic

Functions

25.A.1 Minima of real-valued continuous functions in real space

In the present subsection we limit our attention to real-valued functions in real space.
Let Ω be a domain in R𝑛 , I = [𝑡 ◦ , 𝑡1 ] a compact segment in R and Ω × I ∋ (𝑥, 𝑡) ↦→
𝐹 (𝑥, 𝑡) ∈ R a continuous function. We define

𝑓 (𝑥) = min𝐹 (𝑥, 𝑡) , (25.A.1)

𝑡 ∈I
I𝐹 (𝑥) = {𝑡 ∈ I; 𝐹 (𝑥, 𝑡) = 𝑓 (𝑥)} . (25.A.2)
25.A Appendix: Minima of Families of Plurisubharmonic Functions 1121

Proposition 25.A.1 The function (25.A.1) is continuous in Ω.

Proof Let 𝐾 ⊂ Ω be compact. To every 𝜀 > 0 there is a 𝛿 > 0 such that the
following is true: if 𝑥, 𝑦 ∈ 𝐾, 𝑠 ∈ I𝐹 (𝑥), 𝑡 ∈ I𝐹 (𝑦), then |𝑥 − 𝑦| < 𝛿 implies
𝑓 (𝑥) − 𝑓 (𝑦) ≤ 𝐹 (𝑥, 𝑡) − 𝐹 (𝑦, 𝑡) ≤ 𝜀, 𝑓 (𝑦) − 𝑓 (𝑥) ≤ 𝐹 (𝑦, 𝑠) − 𝐹 (𝑥, 𝑠) ≤ 𝜀. □
Generally speaking, I𝐹 (𝑥) is a closed subset of I; it is a compact subset of the
interior (𝑡 ◦ , 𝑡1 ) of I if and only if 𝑡 ◦ , 𝑡1 ∉ I𝐹 (𝑥). In the trivial example I = [−1, 1],
𝐹 (𝑥, 𝑡) = Φ (𝑥) − 𝑡 2 , Φ ∈ C (Ω) arbitrary, I𝐹 (𝑥) = 𝜕𝐼 is not connected. However,
in a special case of crucial importance to us, the latter cannot occur:

Proposition 25.A.2 If the function I ∋𝑡 ↦→ 𝐹 (𝑥, 𝑡) is quasiconvex then I𝐹 (𝑥) is

connected.

Proof By Definition 25.3.7, if 𝐹 (𝑥, 𝑎) = 𝐹 (𝑥, 𝑏) = 𝑓 (𝑥), [𝑎, 𝑏] ⊂ (𝑡◦ , 𝑡1 ), then

𝐹 (𝑥, 𝑡) ≤ 𝑓 (𝑥) for every 𝑡 ∈ (𝑎, 𝑏). □
Next we take a look at the regularity of 𝑓 .

Proposition 25.A.3 If the function Ω ∋ 𝑥 ↦→ 𝐹 (𝑥, 𝑡) ∈ R is of class C 1 then 𝑓 is

locally Lipschitz continuous in Ω.

Proof Let 𝔅 ⊂ Ω be a compact ball and 𝑥, 𝑥 ′ ∈ 𝔅; let 𝑡 ∈ I𝐹 (𝑥), 𝑡 ′ ∈ I𝐹 (𝑥 ′).

Taylor expansion about 𝑥 ′ yields

𝑓 (𝑥) ≤ 𝐹 (𝑥, 𝑡 ′) = 𝐹 (𝑥 ′, 𝑡 ′) + ∇ 𝑥 𝐹 (𝑥 ′, 𝑡 ′) · (𝑥 − 𝑥 ′) + 𝑜 (|𝑥 − 𝑥 ′ |) ,

whence 𝑓 (𝑥) − 𝑓 (𝑥 ′) ≲ |𝑥 − 𝑥 ′ |. Exchanging 𝑥 and 𝑥 ′ proves the claim. □

Proposition 25.A.4 If the function Ω ∋ 𝑥 ↦→ 𝐹 (𝑥, 𝑡) ∈ R is of class C 1 then

𝑓 ∈ C 1 (Ω) if and only if

I𝐹 (𝑥) ∋ 𝑡 ↦→ ∇ 𝑥 𝐹 (𝑥, 𝑡) ∈ R𝑛

is constant and equal to ∇ 𝑓 (𝑥) for every 𝑥 ∈ Ω.

Proof If 𝑓 ∈ C 1 (Ω) then the function 𝐹 (𝑥, 𝑡) − 𝑓 (𝑥) ≥ 0 in Ω × I is C 1 with

respect to 𝑥 and ∇ 𝑥 𝐹 (𝑥, 𝑡) = ∇ 𝑥 𝑓 (𝑥) for every 𝑡 ∈ I𝐹 (𝑥). Conversely, suppose
∇ 𝑥 𝐹 (𝑥, 𝑡) = 𝒈 (𝑥) for all 𝑥 ∈ Ω, 𝑡 ∈ I𝐹 (𝑥); we have 𝒈 (𝑥) ∈ C (Ω; R𝑛 ). Let 𝜉 ∈ R𝑛 ,
𝜉 ≠ 0, be suitably close to 0; we apply Taylor expansion twice: On the one hand, we
have, trivially,

𝑓 (𝑥 + 𝜉) ≤ min 𝐹 (𝑥 + 𝜉, 𝑡) ≤ 𝑓 (𝑥) + 𝜉 · 𝒈 (𝑥) + 𝑜 (|𝜉 |) ,

𝑡 ∈I𝐹 ( 𝑥)

whence
𝑓 (𝑥 + 𝜉) − 𝑓 (𝑥) 𝜉
− · 𝒈 (𝑥) ≤ |𝜉 | −1 𝑜 (|𝜉 |) .
|𝜉 | |𝜉 |
1122 25 Solvability of Constant Vector Fields of Type (1,0)

On the other hand, to every 𝛿 > 0 we can select a neighborhood U 𝛿 of I𝐹 (𝑥) in I

such that |∇ 𝑥 𝐹 (𝑥, 𝑡) − 𝒈 (𝑥)| < 𝛿 for all 𝑡 ∈ U 𝛿 (and all 𝑥 in a neighborhood of some
𝑥 ◦ ∈ Ω). Taylor expansion yields, for |𝜉 | sufficiently small that I𝐹 (𝑥 + 𝜉) ∩ U 𝛿 ≠ ∅,

𝑓 (𝑥 + 𝜉) = inf 𝐹 (𝑥 + 𝜉, 𝑡) = inf (𝐹 (𝑥, 𝑡) + 𝜉 · ∇ 𝑥 𝐹 (𝑥, 𝑡)) + 𝑜 (|𝜉 |) ,

𝑡 ∈U 𝛿 𝑡 ∈U 𝛿

whence
𝑓 (𝑥 + 𝜉) − 𝑓 (𝑥) 𝜉
− · 𝒈 (𝑥) ≥ −𝛿 + |𝜉 | −1 𝑜 (|𝜉 |) .
|𝜉 | |𝜉 |
We conclude that ∇ 𝑥 𝑓 = 𝒈. □
We recall that C 1,1 (Ω) is the subring of C 1 (Ω) consisting of the functions whose
gradient is Lipschitz continuous in Ω.
Proposition 25.A.5 Suppose 𝐹 ∈ C 2 (Ω × I). For 𝑓 (𝑥) = min𝐹 (𝑥, 𝑡) to be of class
𝑡 ∈I
C 1,1 in Ω it is necessary and sufficient that the following two conditions be satisfied:
(i) The function 𝑡 ↦→ ∇ 𝑥 𝐹 (𝑥, 𝑡) is constant in I𝐹 (𝑥) = {𝑡 ∈ I; 𝐹 (𝑥, 𝑡) = 𝑓 (𝑥)}
for every 𝑥 ∈ Ω.
(ii𝑛 ) To every compact set 𝐾 ⊂ Ω there is an 𝑀 > 0 such that
2
𝜕 𝜕2 𝐹
∇ 𝑥 𝐹 (𝑥, 𝑡) ≤ (𝑥, 𝑡) (Δ 𝑥 𝐹 (𝑥, 𝑡) + 𝑀) (25.A.3)
𝜕𝑡 𝜕𝑡 2
𝜕𝐹
for all 𝑥 ∈ 𝐾, 𝑡 ∈ I𝐹 (𝑥) satisfying 𝜕𝑡 (𝑥, 𝑡) = 0.
In (25.A.3) Δ 𝑥 is the Laplacian in 𝑥-space R𝑛 .
2
Remark 25.A.6 Suppose 𝐹 ∈ C 2 (Ω × I). If 𝑡 ∈ I𝐹 (𝑥) then necessarily 𝜕𝜕𝑡𝐹2 (𝑥, 𝑡) ≥
0; if 𝑡 ∈ I𝐹 (𝑥) ∩ (𝑡◦ , 𝑡1 ) then 𝜕𝐹
𝜕𝑡 (𝑥, 𝑡) = 0.

Proof By Proposition 25.A.3 we know that (i) is necessary and sufficient for 𝑓 to
be of class C 1 . Assuming that (i) holds, in order to prove that (ii𝑛 ) is necessary and
sufficient it suffices to prove the claim when 𝑛 = 1. In the remainder of the proof we
assume that Ω is an open interval in R and we prove that 𝑓 ∈ C 1,1 (Ω) if and only if
(ii1 ) holds true.
Necessity of (ii1 ). Suppose 𝑓 ∈ C 1,1 (Ω) and let 𝑀 > 0 be such that
| 𝑓 (𝑥) − 𝑓 ′ (𝑦)| ≤ 𝑀 |𝑥 − 𝑦| for all 𝑥, 𝑦 ∈ 𝑈 ⊂⊂ Ω, 𝑈 an open interval such
′

that 𝐾 ⊂ 𝑈. It follows that 𝑓 ′ (𝑥) + 𝑀𝑥 is a monotone increasing function in 𝑈 and

𝑓 𝑀 (𝑥) = 𝑓 (𝑥) + 21 𝑀𝑥 2 is convex in 𝑈. We apply (25.3.8) to 𝑓 𝑀 where 𝑎 = 𝑥 ∈ 𝑈,
𝑥 + ℎ ∈ 𝑈 and we let 𝜆 → 0; we get 𝑓 𝑀 (𝑥 + ℎ) ≥ 𝑓 𝑀 (𝑥) + ℎ 𝑓 𝑀 ′ (𝑥), whence

1
𝑓 (𝑥 + ℎ) ≥ 𝑓 (𝑥) + ℎ 𝑓 ′ (𝑥) − 𝑀 ℎ2 .
2
If 𝑡 ∈ I𝐹 (𝑥), implying 𝜕𝐹
𝜕𝑥 (𝑥, 𝑡) = 𝑓 ′ (𝑥) by Proposition 25.A.4, and if 𝑠 + 𝑡 ∈ I we
get
25.A Appendix: Minima of Families of Plurisubharmonic Functions 1123

𝜕𝐹 1
𝐹 (𝑥 + ℎ, 𝑠 + 𝑡) ≥ 𝑓 (𝑥 + ℎ) ≥ 𝐹 (𝑥, 𝑡) + ℎ (𝑥, 𝑡) − 𝑀 ℎ2 .
𝜕𝑥 2
Taylor expansion with respect to 𝑥 yields

1 2 𝜕2 𝐹

𝜕𝐹
𝐹 (𝑥, 𝑠 + 𝑡) + ℎ (𝑥, 𝑠 + 𝑡) + ℎ (𝑥, 𝑠 + 𝑡) + 𝑀
𝜕𝑥 2 𝜕𝑥 2
𝜕𝐹
+𝑜 ℎ2 ≥ 𝐹 (𝑥, 𝑡) + ℎ (𝑥, 𝑡) .
𝜕𝑥
We apply Taylor expansion with respect to 𝑡 to the left-hand side of this last inequality:

𝜕𝐹 1 𝜕2 𝐹
𝐹 (𝑥, 𝑠 + 𝑡) = 𝐹 (𝑥, 𝑡) + 𝑠 (𝑥, 𝑡) + 𝑠2 2 (𝑥, 𝑡) + 𝑜 𝑠2 ,
𝜕𝑡 2 𝜕𝑡
𝜕𝐹 𝜕𝐹 𝜕2 𝐹
(𝑥, 𝑠 + 𝑡) = (𝑥, 𝑡) + 𝑠 (𝑥, 𝑡) + 𝑜 (𝑠) ,
𝜕𝑥 𝜕𝑥 𝜕𝑥𝜕𝑡
𝜕𝐹
whence, since 𝜕𝑡 (𝑥, 𝑡) = 0,
2𝐹 𝜕2 𝐹
2
2𝜕 2 𝜕 𝐹
𝑠 (𝑥, 𝑡) + 2ℎ𝑠 (𝑥, 𝑡) + ℎ (𝑥, 𝑡) + 𝑀
𝜕𝑡 2 𝜕𝑥𝜕𝑡 𝜕𝑥 2

+𝑜 𝑠2 + ℎ𝑜 (𝑠) + 𝑜 ℎ2 ≥ 0.

Putting ℎ = 𝜆𝑠, dividing by 𝑠2 and letting 𝑠 → 0 shows that

𝜕2 𝐹 𝜕2 𝐹
2
𝜕 𝐹
∀𝜆 ∈ R, 2 (𝑥, 𝑡) + 2𝜆 (𝑥, 𝑡) + 𝜆2 (𝑥, 𝑡) + 𝑀 ≥ 0, (25.A.4)
𝜕𝑡 𝜕𝑥𝜕𝑡 𝜕𝑥 2

which is equivalent to (25.A.3).

Sufficiency of (ii1 ). Let 𝑥, 𝑦 ∈ 𝐾 and let 𝑠 ∈ I𝐹 (𝑥), 𝑡 ∈ I𝐹 (𝑦) such that
𝜕𝐹 𝜕𝐹
𝜕𝑡 (𝑥, 𝑠) = 𝜕𝑡 (𝑦, 𝑡) = 0. Taylor expansions yield

1
𝑓 (𝑥) − 𝑓 (𝑦) ≤ 𝐹 (𝑥, 𝑡) − 𝐹 (𝑦, 𝑡) ≤ (𝑥 − 𝑦) 𝑓 ′ (𝑦) + 𝐶𝐾 (𝑦 − 𝑥) 2 ,
2
1
′
𝑓 (𝑦) − 𝑓 (𝑥) ≤ 𝐹 (𝑦, 𝑠) − 𝑓 (𝑥) ≤ (𝑦 − 𝑥) 𝑓 (𝑥) + 𝐶𝐾 (𝑥 − 𝑦) 2
2
and adding the two equations

(𝑥 − 𝑦) ( 𝑓 ′ (𝑥) − 𝑓 ′ (𝑦)) ≤ 𝐶𝐾 (𝑥 − 𝑦) 2 ,

which implies
′ ′
(𝑦) ≤ (𝐶𝐾 + 𝑀) (𝑥 − 𝑦) 2

(𝑥 − 𝑦) 𝑓 𝑀 (𝑥) − 𝑓 𝑀
′ (𝑥) = 𝑓 ′ (𝑥) + 𝑀𝑥 =
again in the notation 𝑓 𝑀 𝜕𝐹
(𝑥, 𝑡) + 𝑀𝑥. We have
𝜕𝑥
1124 25 Solvability of Constant Vector Fields of Type (1,0)
2
𝜕𝐹 𝜕𝐹 𝜕 𝐹
(𝑥, 𝑡) + 𝑀𝑥 = (𝑦, 𝑡) + 𝑀 𝑦 + (𝑥 − 𝑦) (𝑦, 𝑡) + 𝑀 + 𝑜 (|𝑥 − 𝑦|) .
𝜕𝑥 𝜕𝑥 𝜕𝑥 2
2 2
′
Since 𝜕𝜕𝑥𝐹2 (𝑦, 𝑡) ≥ 0 (25.A.3) implies 𝜕𝜕𝑥𝐹2 (𝑦, 𝑡) + 𝑀 > 0. We conclude that 𝑓 𝑀
is monotone increasing, whence 0 ≤ 𝑓 𝑀 ′ (𝑥) − 𝑓 ′ (𝑦) ≤ (𝐶 + 𝑀) (𝑥 − 𝑦) and
𝑀 𝐾
therefore
−𝑀 (𝑥 − 𝑦) ≤ 𝑓 ′ (𝑥) − 𝑓 ′ (𝑦) ≤ 𝐶𝐾 (𝑥 − 𝑦) . □

25.A.2 Minima of functions plurisubharmonic in complex space,

quasiconvex in time

In this subsection we look at functions in a domain Ω in C𝑛 depending on a real

parameter (the “time” variable) varying in a compact segment of the real line. The
key results are proved in the case 𝑛 = 1, the extension to higher complex dimensions
being straightforward. We need a couple of elementary facts.
Lemma 25.A.7 For all 𝛼, 𝛽 ∈ C,
∫ 2𝜋
min Re 𝛼e𝑖 𝜃 , Re 𝛽e𝑖 𝜃 d𝜃 = −2 |𝛼 − 𝛽| . (25.A.5)
0

Proof Suppose 𝛼 ≠ 𝛽, otherwise the claim is trivial. After a rotation in the plane we
can assume that 𝛼 − 𝛽 = − |𝛼 − 𝛽|, in which case the integral on the left in (25.A.5)
is equal to
∫ 𝜋/2 ∫ 𝜋/2
Re (𝛼 − 𝛽) e𝑖 𝜃 d𝜃 = − |𝛼 − 𝛽| cos 𝜃d𝜃. □
− 𝜋/2 − 𝜋/2

Lemma 25.A.8 Let Ω be a domain in C and 𝑢 ∈ C 1,1 (Ω); then we have, in the sense
of distributions,
∫ 2𝜋
2 𝑢 𝑧 + 𝑟e𝑖 𝜃 − 𝑢 (𝑧)
Δ𝑢 (𝑧) = lim d𝜃. (25.A.6)
𝜋 𝑟↘0 0 𝑟2
Proof It suffices to prove the claim when 𝑢 is real-valued. We select a sequence of
functions 𝑢 𝑘 ∈ C 2 (Ω), 𝑘 = 1, 2, ..., converging to 𝑢 uniformly on compact subsets
of Ω; Taylor expansion with respect to (𝑧, 𝑧¯) yields, for 𝑟 > 0,
∫ 2𝜋
1
𝑢 𝑘 𝑧 + 𝑟e𝑖 𝜃 − 𝑢 𝑘 (𝑧) d𝜃 = Δ𝑢 𝑘 (𝑧) + 𝑂 (𝑟) , (25.A.7)
2𝜋𝑟 2 0

whence (25.A.6) with 𝑢 𝑘 in the place of 𝑢. Since 𝑢 ∈ C 1,1 (Ω) we have Δ𝑢 ∈

∞ (Ω); integrating both sides of (25.A.7) with respect to a measure 𝜑 (𝑥, 𝑦) d𝑥d𝑦,
𝐿 loc
𝜑 ∈ Cc∞ (Ω), and letting 𝑟 go to zero directly proves the claim. □
25.A Appendix: Minima of Families of Plurisubharmonic Functions 1125

Proposition 25.A.9 Let Ω be a domain in C, I = [𝑡 ◦ , 𝑡1 ] ⊂⊂ R and suppose

𝐹 ∈ C 2 (Ω × I; R). If 𝑓 (𝑧) = min𝐹 (𝑧, 𝑡) is subharmonic in Ω (Definition 11.2.1)
𝑡 ∈I
then the following three conditions are satisfied:
(a) The function 𝑡 ↦→ 𝜕𝐹𝜕𝑧 (𝑧, 𝑡) is constant in I𝐹 (𝑧) = {𝑡 ∈ I; 𝐹 (𝑧, 𝑡) = 𝑓 (𝑧)} for
every 𝑧 ∈ Ω.
(b) For all 𝑧 ∈ Ω, 𝑡 ∈ I𝐹 (𝑧) implies Δ 𝑥 𝐹 (𝑥, 𝑡) ≥ 0.
(c) For all 𝑧 ∈ Ω, 𝑡 ∈ I𝐹 (𝑧) ∩ (𝑡 ◦ , 𝑡1 ), we have
2
𝜕2 𝐹 𝜕2 𝐹 𝜕2 𝐹
(𝑧, 𝑡) ≤ (𝑧, 𝑡) 2 (𝑧, 𝑡) . (25.A.8)
𝜕𝑧𝜕𝑡 𝜕𝑧𝜕 𝑧¯ 𝜕𝑡

If these conditions are satisfied then 𝑓 ∈ C 1,1 (Ω).

Proof Since 𝐹 is real-valued d 𝑥,𝑦 𝐹 = 2 Re 𝜕𝑧 𝐹d𝑧 and therefore (a) is equivalent to
(i), Proposition 25.A.3. From Proposition 25.A.1 we know that 𝑓 is locally Lipschitz
continuous. We assume that 𝑓 is subharmonic. Then
∫ 2𝜋
𝑓 𝑧 + 𝑟e𝑖 𝜃 − 𝑓 (𝑧) d𝜃 ≥ 0. (25.A.9)
0

If 𝑠, 𝑡 ∈ I𝐹 (𝑧) we have

𝑓 𝑧 + 𝑟e𝑖 𝜃 − 𝑓 (𝑧) ≤ min 𝐹 𝑧 + 𝑟e𝑖 𝜃 , 𝑠 − 𝐹 (𝑧, 𝑠) , 𝐹 𝑧 + 𝑟e𝑖 𝜃 , 𝑡 − 𝐹 (𝑧, 𝑡)

𝑖 𝜃 𝜕𝐹 𝑖 𝜃 𝜕𝐹
≤ 2𝑟 min Re e (𝑧, 𝑠) , Re e (𝑧, 𝑡) + 𝐶𝑟 2 ,
𝜕𝑧 𝜕𝑧

whence, after division of the integral in (25.A.9) by 2𝑟 and letting 𝑟 ↘ 0,

∫ 2𝜋
𝑖 𝜃 𝜕𝐹 𝑖 𝜃 𝜕𝐹
0≤ min Re e (𝑧, 𝑠) , Re e (𝑧, 𝑡) d𝜃
0 𝜕𝑧 𝜕𝑧
𝜕𝐹 𝜕𝐹
= −2 (𝑧, 𝑠) − (𝑧, 𝑡) ,
𝜕𝑧 𝜕𝑧

the last equality a consequence of Lemma 25.A.9. This proves (a).

We apply Lemma 25.A.8 with 𝐹 (𝑧, 𝑡) in the place of 𝑢 (𝑧). If 𝑡 ∈ I𝐹 (𝑧),
∫ 2𝜋
1 𝐹 𝑧 + 𝑟e𝑖 𝜃 , 𝑡 − 𝐹 (𝑧, 𝑡) d𝜃
Δ 𝑥,𝑦 𝐹 (𝑧, 𝑡) = lim (25.A.10)
4 𝑟↘0 0 𝑟2 2𝜋
∫ 2𝜋
𝑓 𝑧 + 𝑟e𝑖 𝜃 − 𝑓 (𝑧) d𝜃
≥ lim ,
𝑟↘0 0 𝑟2 2𝜋

whence (b) by (25.A.9).

Now let 𝑡 ∈ I𝐹 (𝑧) ∩ (𝑡◦ , 𝑡1 ) and 𝜆, 𝜇 ∈ R be such that

|𝜆 + 𝑖𝜇| ≤ 𝑟 −1 min (|𝑡 − 𝑡◦ | , |𝑡 − 𝑡1 |) ,

1126 25 Solvability of Constant Vector Fields of Type (1,0)

implying 𝑡 + 𝑟 (𝜆 cos 𝜃 + 𝜇 sin 𝜃) ∈ I; we have

2𝜋

𝑓 𝑧 + 𝑟e𝑖 𝜃 − 𝑓 (𝑧) d𝜃
∫
0 ≤ lim
𝑟↘0 0 𝑟2 2𝜋
∫ 2𝜋
2 𝐹 𝑧 + 𝑟e , 𝑡 + 𝑟 (𝜆 cos 𝜃 + 𝜇 sin 𝜃) − 𝐹 (𝑧, 𝑡)
𝑖 𝜃
≤ lim d𝜃.
𝑟↘0 𝜋 0 𝑟2
Taylor expansion yields

𝐹 𝑧 + 𝑟e𝑖 𝜃 , 𝑡 + 𝑟 (𝜆 cos 𝜃 + 𝜇 sin 𝜃) − 𝐹 (𝑧, 𝑡)
𝑟2
2

−1 𝑖 𝜃 𝜕𝐹 −1 𝜕𝐹 2𝑖 𝜃 𝜕 𝐹
= 2𝑟 Re e (𝑧, 𝑡) + 𝑟 (𝜆 cos 𝜃 + 𝜇 sin 𝜃) (𝑧, 𝑡) + Re e (𝑧, 𝑡)
𝜕𝑧 𝜕𝑡 𝜕𝑧 2
𝜕2 𝐹 𝜕2 𝐹

+2 (𝑧, 𝑡) + 2 (𝜆 cos 𝜃 + 𝜇 sin 𝜃) Re e𝑖 𝜃 (𝑧, 𝑡)
𝜕𝑧𝜕 𝑧¯ 𝜕𝑧𝜕𝑡
1 𝜕2 𝐹
+ (𝜆 cos 𝜃 + 𝜇 sin 𝜃) 2 2 (𝑧, 𝑡) + 𝑟 −2 𝑜 𝑟 2 1 + 𝜆2 + 𝜇2 ,
2 𝜕𝑡
whence
2𝜋

𝐹 𝑧 + 𝑟e𝑖 𝜃 , 𝑡 + 𝑟𝜆 cos 𝜃 + 𝑟 𝜇 sin 𝜃 − 𝐹 (𝑧, 𝑡)
∫
2
d𝜃
𝜋 0 𝑟2
𝜕2 𝐹 𝜕2 𝐹 𝜕2 𝐹
= Δ𝐹 (𝑧, 𝑡) + 2𝜆 (𝑧, 𝑡) + 2𝜇 (𝑧, 𝑡) + 𝜆2 + 𝜇2 (𝑧, 𝑡)
𝜕𝑥𝜕𝑡 𝜕𝑦𝜕𝑡 𝜕𝑡 2

+𝑟 −2 𝑜 𝑟 2 1 + 𝜆2 + 𝜇2 .

If we let 𝑟 go to zero we reach the conclusion that

𝜕2 𝐹 𝜕2 𝐹 𝜕2 𝐹
Δ𝐹 (𝑧, 𝑠) + 2𝜆 (𝑧, 𝑠) + 2𝜇 (𝑧, 𝑠) + 𝜆2 + 𝜇2 (𝑧, 𝑠) ≥ 0 (25.A.11)
𝜕𝑥𝜕𝑡 𝜕𝑦𝜕𝑡 𝜕𝑡 2

for all (𝜆, 𝜇) ∈ R2 ; this property is equivalent to (25.A.8), whence (c). □

In the last two results of this Appendix we strengthen the hypotheses on the
function 𝐹. The strengthened conditions are satisfied in the applications in this book
of Theorem 25.A.11 below and they considerably simplify the proofs, but they can
be dispensed with, as shown in [Hörmander, 1994], Ch. IV.

Proposition 25.A.10 Let Ω be a domain in C, I = [𝑡 ◦ , 𝑡1 ] ⊂⊂ R and 𝐹 ∈

C 𝜔 (Ω × I; R). If the function I ∋ 𝑡 ↦→ 𝐹 (𝑧, 𝑡) is quasiconvex for every 𝑥 ∈ Ω,
Conditions (a), (b), (c) in Proposition 25.A.9 imply that 𝑓 (𝑧) = min𝐹 (𝑧, 𝑡) is
𝑡 ∈I
subharmonic in Ω.
25.A Appendix: Minima of Families of Plurisubharmonic Functions 1127

Proof Since I ∋ 𝑡 ↦→ 𝐹 (𝑧, 𝑡) is quasiconvex and C 𝜔 then either I𝐹 (𝑧) = I or

I𝐹 (𝑧) consists of a single point 𝑡 (𝑧) (cf. Proposition 25.A.2). We denote by 𝑽 the
𝜕2 𝐹
analytic subvariety of Ω defined by the property that 𝜕𝐹 𝜕𝑡 (𝑧, 𝑡) = 𝜕𝑡 2 (𝑧, 𝑡) = 0 for
every 𝑡 ∈ I; if I𝐹 (𝑧) = I then 𝑧 ∈ 𝑽. If 𝑽 = Ω then 𝐹 (𝑧, 𝑡) = 𝑓 (𝑧) and the
subharmonicity of 𝑓 is stated in (b). In the remainder of the proof we reason under
the hypothesis that dim 𝑽 < 𝑛 and therefore 𝑽 has measure zero.
Since (c) is the same as (ii2 ) in Proposition 25.A.3 we know that 𝑓 ∈ C 1,1 (Ω)
∞ (Ω). Therefore, it suffices to prove the subharmonicity of 𝑓 in the
hence Δ 𝑓 ∈ 𝐿 loc
open set Ω\𝑽. We point out that Ω\𝑽 may not be connected. We now take a closer
look at special unions of connected components of Ω\𝑽, and first

+ 𝜕𝐹
𝔄◦ = 𝑧 ∈ Ω; (𝑧, 𝑡◦ ) > 0 ,
𝜕𝑡

− 𝜕𝐹
𝔄1 = 𝑧 ∈ Ω; (𝑧, 𝑡1 ) < 0 .
𝜕𝑡

The quasiconvexity of I ∋ 𝑡 ↦→ 𝐹 (𝑧, 𝑡) has the following consequences. If 𝑧 ∈ 𝔄◦+

then necessarily 𝐹 (𝑧, 𝑡◦ ) < 𝐹 (𝑧, 𝑡) for all 𝑡 ∈ (𝑡 ◦ , 𝑡1 ] which is equivalent to 𝑓 (𝑧) =
𝐹 (𝑧, 𝑡◦ ), implying Δ 𝑓 (𝑧) = Δ𝐹 (𝑧, 𝑡◦ ) ≥ 0. Likewise, if 𝑧 ∈ 𝔄1− then necessarily
𝐹 (𝑧, 𝑡1 ) < 𝐹 (𝑧, 𝑡) for all 𝑡 ∈ [𝑡 ◦ , 𝑡1 ) and Δ 𝑓 (𝑧) = Δ𝐹 (𝑧, 𝑡1 ) ≥ 0. We define

𝜕𝐹 𝜕𝐹
𝔐 = 𝑧 ∈ Ω; (𝑧, 𝑡◦ ) < 0, (𝑧, 𝑡1 ) > 0 ;
𝜕𝑡 𝜕𝑡

if 𝑧 ∈ 𝔐 then I𝐹 (𝑧) ⊂ (𝑡 ◦ , 𝑡1 ). We have

Ω\𝑽 = 𝔄◦+ ∪ 𝔄1− ∪ 𝔐∪𝑾,

where
𝜕𝐹 𝜕𝐹
𝑾 = 𝑧 ∈ Ω\𝑽; (𝑧, 𝑡◦ ) = (𝑧, 𝑡1 ) = 0 .
𝜕𝑡 𝜕𝑡
2 2
If 𝑧 ∈ 𝑾 necessarily 𝜕𝜕𝑡𝐹2 (𝑧, 𝑡◦ ) 𝜕𝜕𝑡𝐹2 (𝑧, 𝑡1 ) ≠ 0. If 𝑾 contains an open subset of Ω\𝑽
then 𝜕𝐹 𝜕𝐹 + −
𝜕𝑡 (𝑧, 𝑡 ◦ ) = 𝜕𝑡 (𝑧, 𝑡 1 ) = 0 for every 𝑧 ∈ Ω and 𝔄◦ ∪ 𝔄1 ∪ 𝔐 = ∅, 𝑾 = Ω\𝑽;
otherwise dim 𝑾 < 𝑛.
If 𝑾 = Ω\𝑽 we have Ω\𝑽 = 𝔅◦ ∪ 𝔅1 , where

𝜕2 𝐹

𝔅◦ = 𝑧 ∈ 𝑾; (𝑧, 𝑡◦ ) < 0 ,
𝜕𝑡 2
𝜕2 𝐹

𝔅1 = 𝑧 ∈ 𝑾; (𝑧, 𝑡1 ) < 0
𝜕𝑡 2

are open and disjoint subsets of Ω\𝑽. If 𝑧 ∈ 𝔅◦ we have 𝑓 (𝑧) = 𝐹 (𝑧, 𝑡◦ ) and if
𝑧 ∈ 𝔅1 we have 𝑓 (𝑧) = 𝐹 (𝑧, 𝑡1 ). It follows that Δ 𝑓 ≥ 0 in 𝔅◦ ∪ 𝔅1 .
1128 25 Solvability of Constant Vector Fields of Type (1,0)

It remains to deal with the case dim 𝑾 ≤ 1 [i.e., (Ω\𝑽) ∩ 𝑾 has measure
zero] and prove that 𝑓 is subharmonic in 𝔐\ (𝔐 ∩ 𝑾). If 𝑧 ∈ 𝔐\ (𝔐 ∩ 𝑾) then
𝜕2 𝐹
I𝐹 (𝑧) = {𝑡 (𝑧)} ⊂ (𝑡◦ , 𝑡1 ) and we have 𝜕𝐹
𝜕𝑡 (𝑧, 𝑡 (𝑧)) = 0, 𝜕𝑡 2 (𝑧, 𝑡 (𝑧)) > 0; the
Implicit Function Theorem implies that 𝑡 (𝑧) ∈ C 𝜔 (𝔐\ (𝔐 ∩ 𝑾) ; (𝑡◦ , 𝑡1 )) and
𝑓 (𝑧) = 𝐹 (𝑧, 𝑡 (𝑧)) ∈ C 𝜔 (𝔐\ (𝔐 ∩ 𝑾)). We have
𝜕 𝜕𝐹 𝜕𝐹 𝜕𝑡 𝜕𝐹
𝐹 (𝑧, 𝑡 (𝑧)) = (𝑧, 𝑡 (𝑧)) + (𝑧, 𝑡 (𝑧)) (𝑧) = (𝑧, 𝑡 (𝑧))
𝜕𝑧 𝜕𝑧 𝜕𝑡 𝜕𝑧 𝜕𝑧
implying

𝜕2 𝜕2 𝐹 𝜕2 𝐹 𝜕𝑡
𝐹 (𝑧, 𝑡 (𝑧)) = (𝑧, 𝑡 (𝑧)) + (𝑧, 𝑡 (𝑧)) (𝑧) .
𝜕𝑧𝜕 𝑧¯ 𝜕𝑧𝜕 𝑧¯ 𝜕 𝑧¯ 𝜕𝑡 𝜕 𝑧¯
Since 𝑡 (𝑧) ∈ (𝑡 ◦ , 𝑡1 ) Hypothesis (c) implies

𝜕2 𝐹 𝜕2
2
(𝑧, 𝑡) 𝐹 (𝑧, 𝑡 (𝑧))
𝜕𝑡 𝜕𝑧𝜕 𝑧¯
2
𝜕2 𝐹 𝜕2 𝐹 𝜕2 𝐹 𝜕𝑡
≥ (𝑧, 𝑡) + (𝑧, 𝑡 (𝑧)) 2 (𝑧, 𝑡 (𝑧)) (𝑧)
𝜕𝑧𝜕𝑡 𝜕 𝑧¯ 𝜕𝑡 𝜕𝑡 𝜕 𝑧¯
2
𝜕2 𝐹 𝜕2 𝐹 𝜕2 𝐹

𝜕 𝜕𝐹
≥ (𝑧, 𝑡) + (𝑧, 𝑡 (𝑧)) (𝑧, 𝑡 (𝑧)) − (𝑧, 𝑡 (𝑧))
𝜕𝑧𝜕𝑡 𝜕 𝑧¯ 𝜕𝑡 𝜕 𝑧¯ 𝜕𝑡 𝜕 𝑧¯ 𝜕𝑡
2 2 2 2
𝜕 𝐹 𝜕 𝐹
≥ (𝑧, 𝑡) − (𝑧, 𝑡 (𝑧)) ≥ 0,
𝜕𝑧𝜕𝑡 𝜕 𝑧¯ 𝜕𝑡

whence Δ 𝑓 ≥ 0 in 𝔐\ (𝔐 ∩ 𝑾). □
It is now easy to prove the result we were aiming at.

Theorem 25.A.11 Let Ω be a domain in C𝑛 , I = [𝑡◦ , 𝑡1 ] ⊂⊂ R and 𝐹 ∈

C 𝜔 (Ω × I; R). If the function I ∋ 𝑡 ↦→ 𝐹 (𝑧, 𝑡) is quasiconvex whatever 𝑧 ∈ Ω
then, for 𝑓 (𝑧) = min𝐹 (𝑧, 𝑡) to be plurisubharmonic in Ω (Definition 11.2.6) it is
𝑡 ∈I
necessary and sufficient that the following three conditions be satisfied:
(a) The function 𝑡 ↦→ 𝜕𝑧 𝐹 (𝑧, 𝑡) is constant in I𝐹 (𝑧) = {𝑡 ∈ I; 𝐹 (𝑧, 𝑡) = 𝑓 (𝑧)} for
every 𝑧 ∈ Ω. 2
(b) If 𝑡 ∈ I𝐹 (𝑧), 𝑧 ∈ Ω, then the 𝑛 × 𝑛 self-adjoint matrix 𝜕𝑧𝜕𝑗 𝜕𝐹𝑧¯ 𝑘 (𝑧, 𝑡) is
1≤ 𝑗,𝑘 ≤𝑛
positive semidefinite.
(c) For all 𝑧 ∈ Ω, 𝑡 ∈ I𝐹 (𝑧) ∩ (𝑡 ◦ , 𝑡1 ), 𝑤 ∈ C𝑛 , we have
𝑛 𝑛
∑︁ 𝜕2 𝐹 ∑︁ 𝜕 2 𝐹 𝜕2 𝐹
(𝑧, 𝑡) 𝑤 𝑗 𝑤 𝑘 + 2 Re (𝑧, 𝑡) 𝑤 + 2 (𝑧, 𝑡) ≥ 0.
𝑗,𝑘=1
𝜕𝑧 𝑗 𝜕 𝑧¯ 𝑘 𝑗=1
𝜕𝑧 𝑗 𝜕𝑡 𝜕𝑡

If these conditions are satisfied then 𝑓 ∈ C 1,1 (Ω).

25.A Appendix: Minima of Families of Plurisubharmonic Functions 1129

Proof The function 𝑓 is plurisubharmonic if, whatever the affine complex plane
Π in C𝑛 , the restriction of 𝑓 to Ω ∩ Π is subharmonic. This shows that Theorem
25.A.11 is a direct consequence of Propositions 25.A.9, 25.A.10. □
Chapter 26
Pseudodifferential Solvability and Property (𝚿)

This chapter touches on the question of the microlocal solvability of a classical

analytic pseudodifferential operator of principal type, 𝑃. As shown in the first section
this calls into play Condition (𝚿) – which has already been encountered before in
this text, in the main result of Ch. 25 (Theorems 25.3.17, 25.3.18), as well as in Ch.
23 (Theorem 23.2.10), where (𝚿) is shown to be satisfied if the equation 𝑃𝑢 = 𝑓 is
(micro)locally solvable. In the C ∞ category, (𝚿) has been proved to be necessary,
first in 2 dimensions by R.D. Moyer (see [Moyer, 1978]) and, following Moyer’s
ideas, in higher dimensions by L. Hörmander (for details, see [Hörmander, 1983,
IV], pp. 96 et seq.); the proof of this result is much subtler than that of the same claim
in the C 𝜔 category, Theorem 23.2.10. Condition (𝚿) was proved to be sufficient,
still in the C ∞ category, first in 2 dimensions by N. Lerner (see [Lerner, 1988])
and in all dimensions by Nils Dencker (see [Dencker, 2006]); Dencker’s proof is
one of the most difficult feats, technically speaking, in this area of analysis. It is
commonly conjectured (see [Lerner, 2010], p. 284) that, in the C 𝜔 category, (𝚿)
entails solvability with loss of one derivative, which essentially means that if 𝑃 is
a first-order analytic pseudodifferential operator of principal type and if 𝑓 ∈ 𝐿 2
then there is a 𝑢 ∈ 𝐿 2 satisfying 𝑃𝑢 = 𝑓 locally. This is not true in the C ∞ case,
in dimension >2, as shown in [Lerner, 1994]. For a brief account of the historical
context of Condition (𝚿) and of so far unanswered questions about (𝚿), we refer the
reader to [Lerner, 2010], pp. 283 et seq.
Condition (𝚿) is necessary and sufficient for the microlocal solvability of an an-
alytic pseudodifferential equation of principal type in microfunctions. This is how
Trépreau’s theorem is stated in Subsection 26.2.5, where the reader may find an out-
line of the original proof. The significance of this result goes beyond the microlocal
solvability of 𝑃 in that it establishes an isomorphism between the solvability of 𝑃 and
the solvability of the vector field 𝜕/𝜕𝑧 1 at the boundary of a strictly pseudoconvex
domain in C𝑛 , characterized in full detail in Ch. 25. The isomorphism is achieved
through quantized contact transformations provided by the theory of Sato et al.; we
limit ourselves to referring the reader to the standard text on the matter, [Sato-Kawai-
Kashiwara, 1973]. For more on this topic see [Trépreau, 1984], also [Hörmander,
1994], Ch. VII. In the last section of this chapter we provide a completely different

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1131
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_26
1132 26 Pseudodifferential Solvability and Property (Ψ)

proof of Trépreau’s theorem but only for microdistributions. Our approach is purely
analytic: we construct a microlocal parametrix in the form of an FBI integral operator
(see Ch. 22) and use it under Condition (𝚿) reformulated in terms of the (complex)
phase-function.

26.1 Solvability: the Difference between Differential and

Pseudodifferential

26.1.1 Mizohata pseudodifferential operators

Simple examples reveal the fundamental difference between differential and pseu-
dodifferential equations of principal type from the viewpoint of solvability, such as
the following equations in R 𝑥 × R𝑡 (cf. Example 23.2.13):

D𝑡 𝑢 ± 𝑖𝑡 2𝑘−1 |D 𝑥 | 𝑢 = 𝑓 , (26.1.1)

where 1 ≤ 𝑘 ∈ Z, 𝑓 ∈ Cc∞ R2 . The heading is justified by the differential analogue

of (26.1.1), the Mizohata equation D𝑡 𝑢−𝑖𝑡 2𝑘−1 D 𝑥 𝑢 = 𝑓 (Examples 23.1.19, 23.1.24),
which does not satisfy Condition (P) (Definition 23.1.37) and is therefore not locally
solvable, at points of the 𝑥-axis. The natural method for solving (26.1.1) is to use the
Fourier transform in the 𝑥 variable:
𝜕b
𝑢
∓ 𝑡 2𝑘−1 |𝜉 | b 𝑓.
𝑢 =𝑖b (26.1.2)
𝜕𝑡
The ODE (26.1.2) has a solution
∫ 𝑡
1 2𝑘 −𝑠 2𝑘
𝑢 (𝜉, 𝑡) = 𝑖
b e± 2𝑘 | 𝜉 | ( 𝑡 )b
𝑓 (𝜉, 𝑠) d𝑠 (26.1.3)
𝑡◦

whose inverse Fourier transform provides a distribution solution of (26.1.1) only if

𝜉 ↦→ b 𝑢 (𝜉, 𝑡) is tempered.
If the sign is − in (26.1.1) and therefore also in (26.1.3) then b 𝑢 (𝜉, 𝑡) is tempered
if and only if 𝑠2 ≤ 𝑡 2 in the integral in (26.1.3). This shows that, in this case, taking
𝑡 ◦ = 0 yields a distribution solution to (26.1.1) in R2 .
If the sign is + in (26.1.1) and in (26.1.3) then b 𝑢 (𝜉, 𝑡) is tempered if and only if
𝑠2 ≥ 𝑡 2 in the integral in (26.1.3), requiring 0 < 𝑡 < 𝑡 ◦ or 𝑡 ◦ < 𝑡 < 0. For simplicity,
let us take 𝑡 ◦ = ±∞; there are solutions, separately, in each one of the halves of the
𝑡-axis:
∫ +∞
1 2𝑘 2𝑘
𝑢 + (𝜉, 𝑡) = 𝑖
b e 2𝑘 | 𝜉 | ( 𝑡 −𝑠 ) b 𝑓 (𝜉, 𝑠) d𝑠 if 𝑡 > 0,
𝑡
∫ 𝑡
1 2𝑘 2𝑘
𝑢 − (𝜉, 𝑡) = 𝑖
b e 2𝑘 | 𝜉 | ( 𝑡 −𝑠 ) b 𝑓 (𝜉, 𝑠) d𝑠 if 𝑡 < 0.
−∞
26.2 Property (𝚿) 1133

In order to get a distribution solution of (26.1.1) in a neighborhood 𝑈 of the origin

in R2 we must have
𝑢 + (𝜉, 𝑡) = lim b
lim b 𝑢 − (𝜉, 𝑡) . (26.1.4)
𝑡↘0 𝑡↗0

This can be rewritten as

∫ +∞
1 2𝑘
e− 2𝑘 | 𝜉 |𝑠 𝑓 (𝜉, 𝑠) + b
b 𝑓 (𝜉, −𝑠) d𝑠 = 0, (26.1.5)
0

which makes it easy to find right-hand sides 𝑓 ∈ Cc∞ (𝑈) such that (26.1.1) has no
distribution solution in any neighborhood of the origin in R2 .
The symbol of the operator in (26.1.1) is 𝜏 ± 𝑖𝑡 2𝑘−1 |𝜉 |. The null bicharacteristics
of its real part are the straight-lines R ∋ 𝑡 ↦→ ((𝑥, 𝑡) , (𝜉, 0)) in R2𝑥,𝑡 × R2𝜉 , 𝜏 , 𝜉 ≠ 0.
The imaginary part, ±𝑡 2𝑘−1 |𝜉 |, changes sign at 𝑡 = 0 along every one of these lines,
from + to − in the solvable case, from − to + in the unsolvable case.
This can be reformulated microlocally, in terms of the symbol 𝜏 + 𝑖𝑡 2𝑘−1 𝜉: for
𝜉 > 0 the function 𝑡 ↦→ 𝑡 2𝑘−1 𝜉 changes sign from − to + at 𝑡 = 0, whereas, if 𝜉 < 0,
the opposite change of sign occurs. As we have seen there is microlocal solvability
of D𝑡 + 𝑖𝑡 2𝑘−1 D 𝑥 in the latter case and not in the former. There is also a connection
with the considerations in Section 25.3: the function 𝑡 ↦→ −𝜉𝑡 2𝑘 is quasiconvex
(Definition 25.3.7) in an interval (−𝑇, 𝑇), 𝑇 > 0, if and only if 𝜉 < 0.

26.2 Property (𝚿)

26.2.1 Property (𝚿). Definition

The example of Section 26.1 suggests a conjecture for the local solvability of a
classical (C ∞ ) pseudodifferential equation 𝑃𝑢 = 𝑓 of principal type (Definition
23.1.2) at a point 𝑥 ◦ of a C ∞ manifold M.

Definition 26.2.1 Let 𝑃𝑚 be the principal symbol of 𝑃 and let U be a conic open set
in 𝑇 ∗ M\0. We shall say that the principal symbol 𝑃𝑚 of 𝑃, as well as the operator 𝑃
itself, has Property (𝚿) in U if Re 𝑃𝑚 is of principal type at every point of U and
if the restriction of Im 𝑃𝑚 to an arbitrary null bicharacteristic curve of Re 𝑃𝑚 in U
(Definition 23.1.15, loc. cit.) nowhere changes sign from − to +.

The bicharacteristic curve in Definition 26.2.1 is a one-dimensional immersed

submanifold of U, without self-intersections, oriented by the (tangent) Hamiltonian
vector field of Re 𝑃𝑚 , a vector field whose base projection is assumed not to have any
critical point. There is an obvious asymmetry between the roles of Re 𝑃𝑚 and Im 𝑃𝑚
in Definition 26.2.1; the proof that Property (𝚿) is invariant under multiplication of
𝑃𝑚 by a complex, nowhere vanishing, smooth homogeneous function in U (Theorem
26.2.8) is not trivial. In Subsections 26.2.2 and 26.2.3 we present the original proof
1134 26 Pseudodifferential Solvability and Property (Ψ)

in [Nirenberg-Treves, 1970], where it was conjectured that the validity of Property

(𝚿) at every point of 𝑇 ∗ M\0 over a neighborhood of a point ℘ in M is necessary
and sufficient for the local solvability at ℘ (in distributions) of the equation 𝑃𝑢 = 𝑓 .
It is obvious that Definition 26.2.1 can be restated in more general terms applicable
to functions on an arbitrary real, symplectic, C ∞ manifold M (where the concept of
homogeneity with respect to some variables might not make sense). This is has the
advantage of covering the role of (𝚿) in Subsection 25.3.4.

Definition 26.2.2 Let 𝑓 ∈ C ∞ (M) be such that d 𝑓 ≠ 0 at every point of M. We

shall say that 𝑓 satisfies Condition (𝚿) if there is a 𝑐 ∈ C such that d Re (𝑐 𝑓 ) ≠ 0
at every point of M and Im (𝑐 𝑓 ) does not change sign from − to + on any integral
curve of 𝐻Re(𝑐 𝑓 ) on which Re (𝑐 𝑝) vanishes identically. We shall say that 𝑓 satisfies
Condition (Ψ) at a point ℘◦ ∈ M if 𝑓 satisfies Condition (𝚿) in a neighborhood of
℘◦ in M.

The proof of Theorem 26.2.8 extends without any significant modification to the
general case in Definition 26.2.2: in it the claim is independent of the choice of 𝑐 ∈ C
such that d Re (𝑐 𝑓 ) ≠ 0.
From here on we shall limit our attention to analytic pseudodifferential operators,
and more generally to objects in the C 𝜔 class, mostly locally or microlocally, in
Ω × (R𝑛 \ {0}) 𝑇 ∗ Ω\ {0}, Ω a domain in R𝑛 , and toÍthe relevant extensions to
the complex domain. Throughout the sequel 𝑃 (𝑥, 𝜉) = ∞ 𝑗=0 𝑃 𝑚− 𝑗 (𝑥, 𝜉) shall be a
classical analytic symbol of order 𝑚 in Ω × (R𝑛 \ {0}) (Definition 17.2.28). We stress
what it means, in this chapter, for a homogeneous symbol 𝑝 to be of principal type
at a point (𝑥, 𝜉) ∈ Ω × (R𝑛 \ {0}): 𝑝 (𝑥, 𝜉) = 0 =⇒ d 𝜉 𝑝 (𝑥, 𝜉) ≠ 0.

26.2.2 Invariance of (𝚿). Preparatory results

We are going to need results in [Bony, 1969, 2] and [Brézis, 1970]; we follow the
proofs in the latter work. Let X be a domain in R 𝑁 and 𝐴 : X −→ R 𝑁 a Lipschitz
continuous map, meaning that its distribution first partial derivatives belong to
∞ (X). For each 𝑥 ◦ ∈ X let 𝑆 (𝑥 ◦ , 𝑡) (0 ≤ 𝑡 < 𝑇 ◦ ) denote the solution of the initial
𝐿 loc 𝑥
value problem for the system of ODEs defined by 𝐴,

d𝑆
= 𝐴 (𝑆) , 𝑆| 𝑡=0 = 𝑥 ◦ . (26.2.1)
d𝑡
We will use the semigroup properties of the map 𝑡 ↦→ 𝑆 (𝑥 ◦ , 𝑡): if 0 < 𝑡 + 𝜏 < 𝑇𝑥 ◦ ,
𝑡 > 0, 𝜏 > 0, then
𝑆 (𝑥 ◦ , 𝑡 + 𝜏) = 𝑆 (𝑆 (𝑥 ◦ , 𝜏) , 𝑡) . (26.2.2)
d𝑆
Indeed, both sides satisfy d𝑡 = 𝐴 (𝑆), 𝑆| 𝑡=0 = 𝑆 (𝑥 ◦ , 𝜏), and uniqueness applies.
26.2 Property (𝚿) 1135

Theorem 26.2.3 The following two properties of a closed subset 𝑭 of X are equiv-
alent:
∀𝑥 ∈ 𝑭, 𝑡 ∈ [0, 𝑇𝑥 ), 𝑆 (𝑥, 𝑡) ∈ 𝑭; (26.2.3)
1
∀𝑥 ∈ 𝑭, lim dist (𝑥 + 𝜀 𝐴 (𝑥) , 𝑭) = 0. (26.2.4)
𝜀↘0 𝜀

Proof Let 𝑥 ◦ be an arbitrary point of 𝑭. If (26.2.3) holds we have, for 𝜀 > 0

sufficiently small,

1 1
dist (𝑥 ◦ + 𝜀 𝐴 (𝑥 ◦ ) , 𝑭) ≤ (𝑆 (𝑥 ◦ , 𝜀) − 𝑥 ◦ ) − 𝐴 (𝑥 ◦ )
𝜀 𝜀

and the right-hand side tends to zero as 𝜀 does, proving (26.2.4).

Now suppose (26.2.4) holds. It suffices to prove (26.2.3) with 𝑇𝑥 ◦ replaced by 𝛿 >
0 taken to be as small as needed. Assume that 𝔅𝑟 (𝑥 ◦ ) = 𝑥 ∈ R 𝑁 ; |𝑥 − 𝑥 ◦ | < 𝑟 ⊂⊂
X. We avail ourselves of the Lipschitz continuity of the map 𝐴:

| 𝐴 (𝑥) − 𝐴 (𝑥 ′)|
𝑀= sup < +∞,
𝑥, 𝑥 ′ ∈𝔅𝑟 ( 𝑥 ◦ ) |𝑥 − 𝑥 ′ |

whence, if 0 ≤ 𝑡 < 𝛿, 𝑥, 𝑥 ′ ∈ 𝔅𝑟 (𝑥 ◦ ),

d𝑆 d𝑆 ′
(𝑥, 𝑡) − (𝑥 , 𝑡) ≤ 𝑀 |𝑥 − 𝑥 ′ |
d𝑡 d𝑡

and, as a consequence,

|𝑆 (𝑥, 𝑡) − 𝑆 (𝑥 ′, 𝑡)| ≤ e 𝑀𝑡 |𝑥 − 𝑥 ′ | . (26.2.5)

We use the notation 𝜙 (𝑡) = dist (𝑆 (𝑥 ◦ , 𝑡) , 𝑭). We require 𝛿 to be so small that

𝑡 < 𝛿 implies 𝑆 (𝑥 ◦ , 𝑡) ∈ 𝔅𝑟/4 (𝑥 ◦ ) and the existence of 𝑧 ∈ 𝑭 ∩ 𝔅𝑟/4 (𝑥 ◦ ) satisfying
|𝑆 (𝑥 ◦ , 𝑡) − 𝑧| = 𝜙 (𝑡); from here on we keep 𝑡 < 𝛿 fixed.
Let 𝑦 ∈ 𝑭 be arbitrary and 𝜀 > 0 be suitably small; we have

𝜙 (𝑡 + 𝜀) ≤ |𝑆 (𝑥 ◦ , 𝑡 + 𝜀) − 𝑦|
≤ |𝑆 (𝑥 ◦ , 𝑡 + 𝜀) − 𝑆 (𝑧, 𝜀)| + |𝑆 (𝑧, 𝜀) − 𝑦| .

We also have, by (26.2.2) and (26.2.5),

|𝑆 (𝑥 ◦ , 𝑡 + 𝜀) − 𝑆 (𝑧, 𝜀)| = |𝑆 (𝑆 (𝑥 ◦ , 𝑡) , 𝜀) − 𝑆 (𝑧, 𝜀)|

≤ e 𝑀 𝜀 |𝑆 (𝑥 ◦ , 𝑡) − 𝑧| = e 𝑀 𝜀 𝜙 (𝑡) ,

whence
𝜙 (𝑡 + 𝜀) ≤ e 𝑀 𝜀 𝜙 (𝑡) + |𝑆 (𝑧, 𝜀) − 𝑦| .
Let 𝑦 ∈ 𝑭 be such that |𝑆 (𝑧, 𝜀) − 𝑦| = dist (𝑆 (𝑧, 𝜀) , 𝑭). We use once again the fact
that 𝑆 (𝑧, 𝜀) = 𝑧 + 𝜀 𝐴 (𝑧) + 𝑜 (𝜀); we derive
1136 26 Pseudodifferential Solvability and Property (Ψ)

𝜙 (𝑡 + 𝜀) ≤ e 𝑀 𝜀 𝜙 (𝑡) + dist (𝑧 + 𝜀 𝐴 (𝑧) , 𝑭) + 𝑜 (𝜀)

and therefore
𝜙 (𝑡 + 𝜀) − 𝜙 (𝑡) e𝑀 𝜀 − 1 1 1
≤ 𝜙 (𝑡) + dist (𝑧 + 𝜀 𝐴 (𝑧) , 𝑭) + 𝑜 (𝜀) .
𝜀 𝜀 𝜀 𝜀
d𝜙
Letting 𝜀 go to zero yields d𝑡 ≤ 𝑀 𝜙; since 𝜙 (0) = 0 we reach the conclusion that
𝜙 (𝑡) = 0 for all 𝑡 < 𝛿. □

Remark 26.2.4 There is a more geometric formulation of Theorem 26.2.3 in terms

of the outer conormal bundle of 𝑭 (Definition 13.6.16), essentially saying that if the
flow defined by (26.2.1) is to leave 𝑭 at a point its tangent vector field must point
outward. On this approach we refer the reader to [Hörmander, 1983, I], Theorem
8.5.11.

We will avail ourselves of the following application of Theorem 26.2.3 (cf. [Brézis,
1970]).

Lemma 26.2.5 Consider a system of ODEs

d𝑦
= 𝐺 (𝑦, 𝑡) (26.2.6)
d𝑡
where 𝐺 : Ω × (𝑇1 , 𝑇2 ) −→ R𝑛 is Lipschitz continuous. Let the function 𝑏 ∈
C 2 (Ω × (−𝑇1 , 𝑇2 ) ; R) satisfy the following conditions at every point (𝑦, 𝑡) ∈ Ω ×
(−𝑇1 , 𝑇2 ):

i) if 𝑏 (𝑦, 𝑠) < 0 for some 𝑠 ∈ (𝑇1 , 𝑡) then 𝑏 (𝑦, 𝑡) ≤ 0;

ii) 𝑏 = 0 =⇒ 𝐺 · 𝜕𝑦 𝑏 ≤ 0;
iii) 𝑏 = 0, 𝑑𝑏 = 0 =⇒ 𝐺 = 0.
Under these hypotheses, if 𝑦 (𝑡) is a solution of (26.2.6) in an interval (𝑡1 , 𝑡2 ) ⊂
(𝑇1 , 𝑇2 ) and 𝑏 (𝑦 (𝑠) , 𝑠) < 0 for some 𝑠 ∈ (𝑡1 , 𝑡2 ) then 𝑏 (𝑦 (𝑡) , 𝑡) ≤ 0 for all
𝑡 ∈ (𝑠, 𝑡2 ).

Proof We apply Theorem 26.2.3 with X = Ω × (𝑇1 , 𝑇2 ), 𝑁 = 𝑛 + 1, 𝑥 = (𝑦, 𝑡),

𝐴 (𝑥) = (𝐺 (𝑦, 𝑡) , 1); with these choices (26.2.1) and (26.2.6) are equivalent. Let
𝑭 be the closure in Ω × (𝑇1 , 𝑇2 ) of the set of points (𝑦, 𝑡) with the property that
𝑏 (𝑦, 𝑠) < 0 for some 𝑠 ∈ (𝑇1 , 𝑡). Clearly, if (𝑦 ◦ , 𝑡 ◦ ) ∈ 𝑭 then (𝑦 ◦ , 𝑡) ∈ 𝑭 for
𝑡 ◦ ≤ 𝑡 ≤ 𝑡 ◦ + 𝜀, 𝜀 > 0 sufficiently small. It follows from i) that 𝑏 (𝑦, 𝑡) ≤ 0 for all
(𝑦, 𝑡) ∈ 𝑭. We are going to prove that (26.2.4) holds, in three cases:
◦
Case 1: 𝑏 (𝑦 ◦ , 𝑡 ◦ ) < 0. Thus (𝑦 ◦ , 𝑡 ◦ ) ∈ 𝑭, the interior of 𝑭; (26.2.4) holds trivially
in a neighborhood of (𝑦 ◦ , 𝑡 ◦ ).
Case 2: 𝑏 (𝑦 ◦ , 𝑡 ◦ ) = 0, d𝑏 (𝑦 ◦ , 𝑡 ◦ ) = 0. By iii) we have 𝐺 (𝑦 ◦ , 𝑡 ◦ ) = 0, hence
𝐴 (𝑦 ◦ , 𝑡 ◦ ) = (0, 1) and (𝑦 ◦ , 𝑡 ◦ ) + 𝜀 𝐴 (𝑦 ◦ , 𝑡 ◦ ) = (𝑦 ◦ , 𝑡 ◦ + 𝜀) for sufficiently small
𝜀 > 0; (26.2.4) holds at (𝑦 ◦ , 𝑡 ◦ ).
26.2 Property (𝚿) 1137

Case 3: 𝑏 (𝑦 ◦ , 𝑡 ◦ ) = 0, d𝑏 (𝑦 ◦ , 𝑡 ◦ ) ≠ 0. Taylor expansion to second order shows

that there exist a neighborhood 𝑈 of (𝑦 ◦ , 𝑡 ◦ ) in Ω × (−𝑇, 𝑇) and a constant 𝐶 > 0
such that, if (𝑦, 𝑡) ∈ 𝑈 then

(𝑦 − 𝑦 ◦ ) · 𝜕𝑦 𝑏 (𝑦 ◦ , 𝑡 ◦ ) + (𝑡 − 𝑡 ◦ ) 𝜕𝑡 𝑏 (𝑦 ◦ , 𝑡 ◦ ) + 𝐶 |𝑦 − 𝑦 ◦ | 2 + |𝑡 − 𝑡 ◦ | 2 < 0

implies 𝑏 (𝑦, 𝑡) < 0. In this case, dist (𝑦 ◦ + 𝜀𝐺 (𝑦 ◦ , 𝑡 ◦ ) , 𝑭) is equal to

|∇𝑏 (𝑦 ◦ , 𝑡 ◦ )| −1 𝜀∇ 𝑦 𝑏 (𝑦 ◦ , 𝑡 ◦ ) · 𝐺 (𝑦 ◦ , 𝑡 ◦ ) + 𝜀𝜕𝑡 𝑏 (𝑦 ◦ , 𝑡 ◦ ) + 𝐶𝜀 2 (1 + |𝐺 (𝑦 ◦ , 𝑡 ◦ )|)

if this number is positive, to zero otherwise. Property i) implies 𝜕𝑡 𝑏 (𝑦 ◦ , 𝑡 ◦ ) ≤ 0 and

Property ii) implies ∇ 𝑦 𝑏 (𝑦 ◦ , 𝑡 ◦ ) · 𝐺 (𝑦 ◦ , 𝑡 ◦ ) ≤ 0, whence

1 1 + |𝐺 (𝑦 ◦ , 𝑡 ◦ )|
dist ((𝑦 + 𝜀𝐺 (𝑦, 𝑡) , 𝑡) , 𝑭) ≤ 𝜀𝑐 ,
𝜀 |∇𝑏 (𝑦 ◦ , 𝑡 ◦ )|

proving that (26.2.4) holds at (𝑦 ◦ , 𝑡 ◦ ). □

It is helpful to have a more “balanced” and invariant formulation of Lemma
26.2.5. We note that i) is a property of the behavior of 𝑏 along the integral curves
of 𝜕/𝜕𝑡 in Ω × (𝑇1 , 𝑇2 ) whereas the conclusion is the analogue along the integral
curves of the vector field
𝑛
𝜕 ∑︁ 𝜕
+ 𝐺 𝑗 (𝑦, 𝑡) .
𝜕𝑡 𝑗=1 𝜕𝑦 𝑗

Note also that the hypotheses and conclusion in Lemma 26.2.5 are local properties.
Inspection of its proof shows that Lemma 26.2.5 can be restated as follows (a trivial
claim when vector fields and 𝑏 are C ∞ ).

Lemma 26.2.6 Let 𝑋, 𝑌 be two real vector fields with Lipschitz coefficients in a C ∞
manifold M. Let 𝑏 ∈ C 2 (M; R) be such that the following hold at an arbitrary point
℘ ∈ M:
(a) 𝑏 (℘) = 0 =⇒ 𝑋 𝑏 (℘) ≤ 0;
(b) 𝑏 (℘) = 0 & d𝑏 (℘) = 0 entails 𝑋 | ℘ = 𝑐 𝑌 | ℘ , 𝑐 > 0.
Under these hypotheses, if 𝑏 does not change sign from − to + along any integral
curve of 𝑋 the same is true along any integral curve of 𝑌 .

The integral curves of 𝑋 (resp., 𝑌 ) are oriented by 𝑋 (resp., 𝑌 ) pointing to the

positive direction.

Remark 26.2.7 To say that 𝑏 does not change sign from − to + along any integral
curve of 𝑋 is equivalent to the following property: if 𝐵 ∈ C 3 (M; R) is such that
𝑋 𝐵 = 𝑏 then the restriction of 𝐵 to an arbitrary integral curve of 𝑋 is quasiconvex
(cf. Definition 25.3.1, Proposition 25.3.10).
1138 26 Pseudodifferential Solvability and Property (Ψ)

26.2.3 Invariance of (𝚿). Statement and proof

Theorem 26.2.8 If 𝑃𝑚 has Property (𝚿) in the conic open subset U of 𝑇 ∗ M\0
the same is true of 𝑞𝑃𝑚 for every complex, nowhere vanishing, homogeneous C 𝜔
function 𝑞 in U.

Proof The claim is microlocal in nature: we assume that the base projection of U
is the domain of analytic coordinates 𝑥1 , ..., 𝑥 𝑛 with dual coordinates 𝜉1 ., ..., 𝜉 𝑛 . Our
“central point” (𝑥 ◦ , 𝜉 ◦ ) ∈ U shall satisfy 𝑃𝑚 (𝑥 ◦ , 𝜉 ◦ ) = 0; it is convenient to take
𝑥 ◦ = 0 and identify U with 𝑈 × Γ, 𝑈 an open ball centered at the origin in R𝑛 and
Γ a convex cone in R𝑛 \ {0}. In the proof we assume tacitly that 𝑈 and Γ ∩ S𝑛−1 are
contracted as much as needed about 0 and 𝜉 ◦ respectively.
We avail ourselves of Proposition 23.1.10 and assume that (23.1.7) holds: 𝑚 = 1
and 𝑃1 (𝑥, 𝜉) = 𝜉 𝑛 + 𝑖𝐵 (𝑥, 𝜉 ′) with 𝐵 ∈ C 𝜔 (U) real-valued, homogeneous of
degree 1, having the following property:
(𝜓) the function 𝑥 𝑛 ↦→ 𝐵 (𝑥, 𝜉 ′) does not change sign from − to + at any point of U.
As usual 𝑥 ′ = (𝑥1 , ..., 𝑥 𝑛−1 ), 𝜉 ′ = (𝜉1 , ..., 𝜉 𝑛−1 ). Let then 𝑞 = 𝑓 + 𝑖𝑔, 𝑓 , 𝑔 ∈
C𝜔 (U; R) homogeneous of degree 0, 𝑞 nowhere vanishing in U; thus

𝑞𝑃1 = 𝑓 𝜉 𝑛 − 𝑔𝐵 + 𝑖 ( 𝑓 𝐵 + 𝑔𝜉 𝑛 ) . (26.2.7)

We use the notation 𝐴♭ = Re (𝑞𝑃1 ), 𝐵♭ = Im (𝑞𝑃1 ). We reason differently, depending

on whether 𝑓 vanishes or not at (0, 𝜉 ◦ ).
Case 𝑓 (0, 𝜉 ◦ ) ≠ 0. Division by 𝑓 has no effect on whether 𝑞𝑃1 has Property (𝚿)
or not; we can take 𝑓 = 1:

𝐴♭ = 𝜉 𝑛 − 𝑔𝐵,

𝐵♭ = 𝐵 + 𝑔𝜉 𝑛 = 1 + 𝑔 2 𝐵 + 𝑔 𝐴♭ .

On every null bicharacteristic of 𝐴♭ we have 𝐵♭ = 1 + 𝑔 2 𝐵 and therefore it suffices

to prove that 𝐴♭ + 𝑖𝐵 has Property (𝚿) in U. Since 𝐵 is independent of 𝜉 𝑛 it suffices
to look at its values on the projection 𝔠 in (𝑥, 𝜉 ′)-space of the integral curve of the
Hamiltonian vector field of 𝐴♭ , 𝐻 𝐴♭ , through (0, 𝜉 ◦ ). We have
𝑛−1 𝑛
𝜕𝑔 𝜕 ∑︁ 𝜕 𝜕 ∑︁ 𝜕 𝜕
𝐻 𝐴♭ = 1+𝐵 − (𝑔𝐵) + (𝑔𝐵) .
𝜕𝜉 𝑛 𝜕𝑥 𝑛 𝑗=1 𝜕𝜉 𝑗 𝜕𝑥 𝑗 𝑗=1 𝜕𝑥 𝑗 𝜕𝜉 𝑗

We see that the curve 𝔠 is described (in R2𝑛−1 ) by(𝑥 ′ (𝑡) , 𝑡, 𝜉 ′ (𝑡)), with (𝑥 ′ (𝑡) , 𝜉 ′ (𝑡))
the solution of the system of ODEs
26.2 Property (𝚿) 1139
−1
d𝑥 𝑗 𝜕𝑔 𝜕𝑔 𝜕𝐵
=− 1+𝐵 𝐵+𝑔 , (26.2.8)
d𝑡 𝜕𝜉 𝑛 𝜕𝜉 𝑗 𝜕𝜉 𝑗
−1
d𝜉 𝑗 𝜕𝑔 𝜕𝑔 𝜕𝐵
= 1+𝐵 𝐵+𝑔 ,
d𝑡 𝜕𝜉 𝑛 𝜕𝑥 𝑗 𝜕𝑥 𝑗

(where 𝑗 = 1, ..., 𝑛 − 1 and 𝑥 𝑛 = 𝑡) satisfying 𝑥 ′ (0) = 0, 𝜉 ′ (0) = 𝜉 ◦′.

Since 𝐵 (0, 𝜉 ◦′) = 0 we can solve the equation 𝜏 = 𝑔 (𝑥, 𝜉 ′, 𝜏) 𝐵 (𝑥, 𝜉 ′) with
respect to 𝜏 under the condition 𝜏 (0, 𝜉 ◦′) = 0. We define

ℎ (𝑥, 𝜉 ′) = 𝑔 (𝑥, 𝜉 ′, 𝜏 (𝑥, 𝜉 ′)) 𝐵 (𝑥, 𝜉 ′) ,

whence 𝐴♭ (𝑥, 𝜉 ′, 𝜏 (𝑥, 𝜉 ′)) ≡ 0. We apply Lemma 26.2.5 with 𝑦 = (𝑥 ′, 𝜉 ′), 𝑡 = 𝑥 𝑛 ,

𝑏 (𝑦, 𝑡) = 𝐵 (𝑥 ′, 𝑡, 𝜉 ′) and 𝐺 (𝑦, 𝑡) the vector in R2𝑛−1 whose components are the
right-hand sides of (26.2.8). Hypothesis (𝜓) states that 𝑏 (𝑦, 𝑡) satisfies Condition
i) in Lemma 26.2.5. Condition ii) holds because 𝐵 (𝑥, 𝜉 ′) = 0 =⇒ ℎ (𝑥, 𝜉 ′) = 0;
Condition iii) holds because 𝐵 (𝑥, 𝜉 ′) = 0, d 𝑥′ , 𝜉 ′ 𝐵 (𝑥, 𝜉 ′) = 0 imply 𝐺 (𝑦, 𝑡) = 0
[by (26.2.8)]. The conclusion derived from Lemma 26.2.5 is the claim in Theorem
26.2.8 in the present case.
Case 𝑓 (0, 𝜉 ◦ ) = 0, which requires 𝑔 (0, 𝜉 ◦ ) ≠ 0 and therefore

d 𝜉 (𝑞𝑃1 ) (0, 𝜉 ◦ ) = 𝑖𝑔d 𝜉 𝑃1 (0, 𝜉 ◦ ) ≠ 0.

After division by −𝑔 and replacing 𝑓 /𝑔 by 𝑓 (26.2.7) reads

𝑞𝑃1 = 𝐵 − 𝑓 𝐴 − 𝑖 ( 𝐴 + 𝑓 𝐵) . (26.2.9)

Our aim is to prove that 𝐴 + 𝑓 𝐵 does not change sign from + to − along the null
bicharacteristic curve 𝔠◦ of 𝐵 − 𝑓 𝐴 passing through (0, 𝜉 ◦ ). By Definition 26.2.1 we
must assume that d 𝜉 Re (𝑞𝑃𝑚 ) = −𝑔d 𝜉 (Im 𝑃1 ) does not vanish in 𝔠◦ ; by (26.2.9)
this demands d 𝜉 𝐵 ≠ 0 in 𝔠◦ since 𝑓 𝐴 vanishes to second order at (0, 𝜉 ◦ ). Since

𝐴 + 𝑓 𝐵 = 1 + 𝑓 2 𝐴 + 𝑓 (𝐵 − 𝑓 𝐴)

the sign of 𝐴 + 𝑓 𝐵 along 𝔠◦ is the same as that of 𝐴. By the Darboux theorem 13.3.20
(Subsection 13.3.4) there is a symplectic change of variables preserving (0, 𝜉 ◦ ) and
transforming 𝐵 − 𝑓 𝐴 into one of the coordinates 𝜉 𝑗 , say 𝜉 𝑛 . To summarize: in the
new coordinates, on the one hand we have 𝑃1 = 𝐴 + 𝑖 (𝜉 𝑛 + 𝑓 𝐴) and Property (𝚿)
for 𝑃1 demands that 𝜉 𝑛 not change sign from − to + on any null bicharacteristic of
𝐴; on the other hand we must show that 𝐴 does not change sign from + to − on the
curve 𝔠◦ , now an integral curve of 𝜕𝑥𝜕𝑛 in which 𝐴 = 𝐴 (0, 𝑥 𝑛 , 𝜉 ◦ ). We have, for some
𝑎 ∈ R, 𝐴 (0, 𝑥 𝑛 , 𝜉 ◦ ) = 𝑎𝑥 𝑛𝑘 (1 + 𝑂 (𝑥 𝑛 )). If 𝑎 = 0 or if 𝑘 is even 𝐴 (0, 𝑥 𝑛 , 𝜉 ◦ ) does
not change sign along 𝔠◦ ; thus suppose 𝑎 ≠ 0 and 𝑘 = 2ℓ + 1. We have

𝐻 𝐴 𝜉 𝑛 = − (2ℓ + 1) 𝑎𝑥 𝑛2ℓ (1 + 𝑂 (𝑥 𝑛 ))
1140 26 Pseudodifferential Solvability and Property (Ψ)

and (𝚿) for 𝑃1 demands 𝑎 > 0, proving that 𝐴 (0, 𝑥 𝑛 , 𝜉 ◦ ) changes sign from − to +
at 𝑥 𝑛 = 0. □

Remark 26.2.9 In the last part of the proof of Theorem 26.2.8 we have used the
fact that Re 𝑃1 is analytic. This is not really needed, as shown in the slightly more
complicated proof in [Nirenberg-Treves, 1970] where Lemma 26.2.5 is used again.
Otherwise all results so far in this chapter remain valid in the C ∞ class. When needed
the Weierstrass Preparation Theorem must be replaced by the Weierstrass–Malgrange
Theorem (cf. Remark 23.1.11).

26.2.4 Microlocal solvability of 𝑷 in distributions. Necessity of

Condition (𝚿)

We now focus on the microlocal solvability of the analytic pseudodifferential equa-

tion 𝑃 (𝑥, D 𝑥 ) 𝑢 = 𝑓 .

Definition 26.2.10 We shall say that 𝑃 (𝑥, D 𝑥 ) is microlocally solvable at (𝑥 ◦ , 𝜉 ◦ )

in hyperfunctions (resp., distributions) if given any germ f 𝑥 ◦ of a hyperfunction
(resp., distribution) at 𝑥 ◦ there is the germ u 𝑥 ◦ of a hyperfunction (resp., distribution)
at 𝑥 ◦ such that P 𝑥 ◦ u 𝑥 ◦ − f 𝑥 ◦ is microanalytic at (𝑥 ◦ , 𝜉 ◦ ) (Definition 3.5.1).

We have denoted by P 𝑥 ◦ the germ operator defined by 𝑃 (𝑥, D 𝑥 ). Definition

26.2.10 could also be stated for microfunctions in the place of hyperfunctions.
In the remainder of this section we shall always reason under the hypothesis
that 𝑃 (𝑥, D 𝑥 ) is an analytic, classical pseudodifferential operator of principal type,
defined in a domain Ω of R𝑛 [or in a conic neighborhood of (𝑥 ◦ , 𝜉 ◦ ) in C𝑇 ∗ Ω]. The
necessity of Condition (𝚿) for local solvability of 𝑃 (𝑥, D 𝑥 ) has already been stated
in Ch. 23. Actually, inspection of the reasoning in Ch. 23 and, in fact, the statement
itself of Theorem 23.2.10 and its proof are wholly microlocal.

Theorem 26.2.11 If 𝑃 (𝑥, D 𝑥 ) is microlocally solvable in distributions at (𝑥 ◦ , 𝜉 ◦ ) ∈

Char 𝑃 then 𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) holds.

The necessity of (𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) ) for the microlocal solvability at (𝑥 ◦ , 𝜉 ◦ ) in hyperfunc-

tions is part of the result discussed in the following subsection.
26.2 Property (𝚿) 1141

26.2.5 Microlocal solvability of 𝑷 in microfunctions. Trépreau’s

Theorem

That (𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) ) is necessary and sufficient for the microlocal solvability of 𝑃 in

microfunctions was proved in [Trépreau, 1984].

Theorem 26.2.12 Let 𝑃 (𝑥, D 𝑥 ) be an analytic, classical pseudodifferential operator

of principal type. For the pseudodifferential operator 𝑃 (𝑥, D 𝑥 ) to be microlocally
solvable in microfunctions at (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 it is necessary and sufficient that
(𝚿 ( 𝑥, 𝜉 ) ) hold at every characteristic point (𝑥, 𝜉) of 𝑃 in a conic neighborhood of
(𝑥 ◦ , 𝜉 ◦ ) in 𝑇 ∗ R𝑛 \0.

The natural phase space in Trépreau’s proof is the vector bundle 𝑇 (1,0) C𝑛 \0
identified Í with C𝑛 × (C𝑛 \ {0}), a symplectic manifold for the fundamental two-form
d𝜁 ∧d𝑧 = 𝑗=1 d𝜁 𝑗 ∧d𝑧 𝑗 . It is convenient to identify the cotangent bundle 𝑇 ∗ R𝑛 with
𝑗=𝑛
√
the R-Lagrangian and I-symplectic subspace (Section 13.5) R𝑛𝑥 × −1R𝑛𝜂 of C2𝑛 𝑧,𝜁 . We
then have the choice of reasoning in the associated sphere bundles (thereby dealing
directly with microfunctions) or else reasoning in 𝑇 (1,0) C𝑛 \0 with the homogeneous
and conic terminology (thereby dealing with hyperfunctions). Here we shall follow
the latter path.
The first, relatively elementary, step is to determine a homogeneous complex-
analytic symplectomorphism of a conic neighborhood U◦ of (𝑥 ◦ , 𝑖𝜉 ◦ ) in C𝑛 ×
(C𝑛 \ {0}) onto another such neighborhood, mapping (𝑥 ◦ , 𝑖𝜉 ◦ ) onto (𝑥 ◦ , (0, ..., 0, −𝑖))
and transforming 𝜁1 +𝑖𝑞 (𝑥, 𝜁 ′′) into 𝜁1 . In the framework of [Sato-Kawai-Kashiwara,
1973], to this symplectomorphism there corresponds a quantized contact transforma-
tion that transforms 𝑃 microlocally into D 𝑥1 . This quantized contact transformation
is constructed in terms of relative cohomology classes (cf. Subsections 10.3.3 and
10.3.4). At first sight this seems to settle the solvability problem since D 𝑥1 is obvi-
ously solvable. Actually the problem remains, as it is not clear what the quantized
contact transformation does to microsupports (i.e., analytic wave-front sets). What
we know is that it cannot do what is needed to prove microlocal solvability, since we
know that the latter does not hold unless (𝚿 ( 𝑥, 𝜉 ) ) does.
This conundrum is resolved by exploiting Corollary 13.6.22: there is a homo-
geneous complex-analytic symplectomorphism 𝜒 of a conic neighborhood U of
(𝑥 ◦ , 𝑖𝜉 ◦ ) in C𝑛 × (C𝑛 \ {0}) onto a conic open subset 𝜒 (U) of the outer conormal
bundle 𝑁out ∗ 𝜕ΩC (Definition 13.6.16) of a strictly pseudoconvex domain ΩC in C𝑛

with C boundary 𝜕ΩC such that 𝜒 (𝑥 ◦ , 𝑖𝜉 ◦ ) = (𝑥 ◦ , (0, ..., 0, −𝑖)) ∈ 𝑁out

𝜔 ∗ 𝜕ΩC , pre-

serving 𝜁1 ; the quantized contact transformation defined by 𝜒 preserves 𝜕/𝜕𝑧1 (as a

microdifferential operator). To visualize things it helps to take 𝑧◦ = 0 and define ΩC
in a neighborhood of 0 in C𝑛 as in Lemma 25.2.7, by the inequality 𝜌 (𝑧) < 0 with
𝑛
∑︁
𝜌 (𝑧) = −2𝑦 𝑛 + |𝑧 ′ | 2 + Re 𝑐 𝑗,𝑘 𝑧 𝑗 𝑧 𝑘 + 𝑂 |𝑧| 3 ,
𝑗,𝑘=1
1142 26 Pseudodifferential Solvability and Property (Ψ)

where 𝑧 ′ = (𝑧1 , ..., 𝑧 𝑛−1 ) and 𝑐 𝑗,𝑘 = 𝑐 𝑘, 𝑗 ∈ C. We recall that the outer conormal
bundle of 𝜕ΩC over 𝑈 C ∩ 𝜕ΩC (𝑈 C : a neighborhood of the origin in C𝑛 ) consists of
the points (𝑧, 𝜆d𝜌 (𝑧)), 𝑧 ∈ 𝑈 C ∩ 𝜕ΩC , 𝜆 > 0. At 𝑧 = 0 we get the points (0, −2𝜆d𝑦 𝑛 )
which can be identified with the points (0, ..., 0, −2𝑖𝜆) ∈ ΩC . It follows that a conic
neighborhood of (0, (0, ..., 0, −𝑖)) in 𝑁out∗ 𝜕ΩC can be identified with the set of points

(𝑧, 𝜆𝜁), 𝑧 ∈ 𝑈 C ∩ 𝜕ΩC , 𝜆 > 0, |𝜁 − (0, ..., 0, −𝑖)| sufficiently small, 𝜁 = d𝜌 (𝑤) for
some 𝑤 ∈ 𝜕ΩC . Setting an upper bound on 𝜆 allows us to identify the resulting set
with an open subset of ΩC of the form 𝑈 C ∩ ΩC , with 𝑈 C a neighborhood of 0 in C𝑛 .
We may identify a wedge W𝛿 (𝑈, Γ) [cf. √ (3.3.1); Γ : as above, a cone in R𝑛 \ {0},
◦ ∗
Γ ∋ 𝜉 ] with the subset of 𝑇 R R 𝑥 × −1R𝑛𝜉 consisting of the points (𝑥, 𝑖𝜉)
𝑛 𝑛

such that 𝑥 ∈ 𝑈, 𝜉 ∈ Γ, |𝜉 | < 𝛿. We take the neighborhood 𝑈 of 0 in R𝑛 and the

cone Γ to be such that W𝛿 (𝑈, Γ) ⊂ U, the domain of definition of the complex
symplectomorphism
𝜒. It can be easily checked that 𝜒 (W𝛿 (𝑈, Γ)) contains a subset
ΩC𝑐 = 𝑧 ∈ ΩC ; 𝑦 𝑛 < 𝑐 , 𝑐 > 0 suitably small. The pullback of ℎ ∈ O ΩC𝑐 under
the map 𝜒 is a function 𝜒∗ ℎ defined and holomorphic in a wedge W𝛿1 (𝑈1 , Γ1 ) ⊂
W𝛿 (𝑈, Γ), 𝑈1 ∋ 0, Γ1 ∋ 𝜉 ◦ ; in turn, this determines a hyperfunction boundary value
𝑓 = 𝑏𝑈1 𝜒∗ ℎ (Definition 7.2.3). This correspondence is best described in terms of
germs, the germs at 0 of hyperfunction boundary values of holomorphic functions
such as ℎ, and the germs of holomorphic functions at the germ of set ΩC , 0 ; the
correspondence between these two kinds of germs is obviously one-to-one (in part,
because the boundary value map is injective as per Proposition 7.2.5). Since 𝑃
has been transformed into 𝜕/𝜕𝑥1 we have transformed the problem of solving the
equation 𝑃𝑢 = 𝑓 = 𝑏𝑈1 𝜒∗ ℎ, microlocally in a neighborhood of (𝑥 ◦ , 𝑖𝜉 ◦ ) in U, into
𝜕𝑢
that of solving the equation 𝜕𝑧 1
= ℎ in the ring O 0 Ω C of germs of holomorphic

functions in the germ of set ΩC , 0 . This is where we avail ourselves of the results
in Ch. 25, specifically Theorem 25.3.17 (cf. also Theorem 25.3.18) which states
that the appropriate version of Condition (𝚿) is necessary and sufficient for the
surjectivity of the derivation of O0 ΩC defined by 𝜕/𝜕𝑧 1 . Since (𝚿) is invariant
under (complex-analytic) symplectomorphisms this completes the proof of Theorem
26.2.12.
We must underline the fact that the preceding paragraphs give only an outline
of the proof of Trépreau’s Theorem. Two important elements in this proof are not
provided by the material presented in this book:
(1) The reader will not find a description of quantized contact transformations and
their links to pseudodifferential and Fourier integral operators.
(2) We have not introduced the hyperfunction boundary value of (arbitrary) holo-
morphic functions in a strongly strictly pseudoconvex domain in C𝑛 with a C 𝜔
boundary.

On point (1) we refer the reader to [Sato-Kawai-Kashiwara, 1973].

Point (2) should not be too difficult to work out, in analogy with what has been
done in wedges (see Lemma 7.2.2).
In the next section we use analytic tools to prove the microdistribution version of
Trépreau’s Theorem.
26.3 Microlocal Solvability in Distributions 1143

26.3 Microlocal Solvability in Distributions

26.3.1 The pseudodifferential operator under study

For the sake of clarity and connection with Section 26.1 we modify the notation of the
preceding section and revert to space variable 𝑥 = (𝑥 1 , ..., 𝑥 𝑛−1 ) [𝑧 = (𝑧1 , ..., 𝑧 𝑛−1 )
in complex space] and time 𝑡 (even when complex), dual variable 𝜉 = (𝜉1 , ..., 𝜉 𝑛−1 )
[𝜁 = (𝜁1 , ..., 𝜁 𝑛−1 ) in complex dual space], 𝜏. We are allowed to do this as long as we
limit ourselves to microlocal analysis. We can then take full advantage of (23.1.7)
and Proposition 23.1.12: elliptic factors are irrelevant in the discussion of local or
microlocal solvability and Egorov’s Theorem 18.5.31 provides us with the needed
conjugate T −1 𝑃T with T a Fourier integral operator with real-phase function. In
the space-time coordinates this means that, within a microanalytic framework, the
pseudodifferential operator under study can be taken to be

𝑃 (𝑥, 𝑡, D 𝑥 , D𝑡 ) = D𝑡 + 𝑖𝑞 (𝑥, 𝑡, D 𝑥 ) , (26.3.1)

√
(where D 𝑥𝑘 = − −1𝜕/𝜕𝑥 𝑘 and likewise for all variables, real or complex). We
assume that the real symbol 𝑞 (𝑥, 𝑡, 𝜉) extends as a holomorphic function 𝑞 (𝑧, 𝑡, 𝜁)
in ΩC × ℭ 𝜅 , 𝑞 (𝑧, 𝑡, 𝜆𝜁) = 𝜆𝑞 (𝑧, 𝑡, 𝜁) for 𝜆 > 0, with ΩC a domain in C𝑛 such that
Ω = ΩC ∩ R𝑛 and

ℭ 𝜅 = 𝜁 = 𝜉 + 𝑖𝜂 ∈ C𝑛−1 ; 𝜉 ∈ ℭ, |𝜂| < 𝜅 |𝜉 | , (26.3.2)

where ℭ is an open cone in R𝑛−1 \ {0} to be chosen later, 𝜅 > 0 is as small as needed.
Our “central point” will be ((𝑥 ◦ , 𝑡 ◦ ) , (𝜉 ◦ , 0)) ∈ Char 𝑷, meaning 𝑞 (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ ) = 0
[ |𝜏◦ | + |𝑞 (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ )| ≠ 0 entails ((𝑥 ◦ , 𝑡 ◦ ) , (𝜉 ◦ , 𝜏◦ )) ∉ Char 𝑷]. It will be convenient
to assume that |𝜉 ◦ | = 1. The analysis will take place, mostly, over a product 𝑈 ′C ×
Δ𝑟◦ (𝑡 ◦ ) ⊂⊂ ΩC , where 𝑈 ′C is a geometrically simple neighborhood of 𝑥 ◦ in C𝑛−1 ,
Δ𝑟◦ (𝑡 ◦ ) = {𝑡 ∈ C; |𝑡 − 𝑡 ◦ | < 𝑟 ◦ }; diam 𝑈 ′C and 𝑟 ◦ will be as small as needed. We
define 𝑈 ′ = 𝑈 ′C ∩ R𝑛−1 , I◦ = (𝑡 ◦ − 𝑟 ◦ , 𝑡 ◦ + 𝑟 ◦ ) ⊂ R, and

𝑈 C = 𝑈 ′C × Δ𝑟◦ (𝑡 ◦ ) , 𝑈 = 𝑈 ′ × I◦ . (26.3.3)

The action of the (analytic) pseudodifferential operator 𝑞 (𝑥, 𝑡, D 𝑥 ) on 𝑓 ∈ E ′ (𝑈) is

defined by the oscillatory integral (cf. Remark 3.5.14)
∫ ∫
1
𝑞 (𝑥, 𝑡, D 𝑥 ) 𝑓 (𝑥, 𝑡) = e𝑖 ( 𝑥−𝑦) · 𝜉 𝑞 (𝑥, 𝑡, 𝜉) 𝑓 (𝑦, 𝑡) d𝑦d𝜉, (26.3.4)
(2𝜋) 𝑛−1 ℭ
which defines a distribution in Ω (compactly supported with respect to 𝑡). Actually,
we shall mainly deal with compactly integrable functions in 𝑈, i.e., 𝑓 ∈ 𝐿 c1 (𝑈).
1144 26 Pseudodifferential Solvability and Property (Ψ)

26.3.2 Phase-function attached to the pseudodifferential operator 𝑷

We now associate to the pseudodifferential operator 𝑃 [assumed to have the expres-

sion (26.3.1)] a special phase-function 𝜑. To do this we apply the results of Section
13.4, with regular meaning complex-analytic, i.e., holomorphic. We refer the reader
to the Cauchy problem (13.4.9)–(13.4.10) with 𝑧 ∈ 𝑈 ′C in the place of 𝑥; 𝜑 is the
holomorphic solution in 𝑈 C × 𝑈 C × ℭ 𝜅 [𝑈 C and ℭ 𝜅 are defined in (26.3.3) and
(26.3.2) respectively]. We reason, throughout, under the proviso that diam 𝑈 ′C , 𝑟 ◦
and diam ℭ∩S𝑛−2 are as small as needed, even when we do not repeat it. In the
present context, the eikonal equation reads

𝜕𝜑
(𝑧, 𝑡, 𝑤, 𝑠, 𝜁) + 𝑖𝑞 (𝑧, 𝑡, 𝜕𝑧 𝜑 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁)) = 0 (26.3.5)
𝜕𝑡
while the Cauchy conditions are
1
𝜑 (𝑧, 𝑠, 𝑤, 𝑠, 𝜁) = (𝑧 − 𝑤) · 𝜁 + 𝑖 ⟨𝜁⟩ ⟨𝑧 − 𝑤⟩ 2 , (26.3.6)
2
where ⟨𝜁⟩ is the main branch square-root of 𝜁 · 𝜁 (𝜁 ∈ C𝑛−1 ). The existence and
uniqueness of the solution 𝜑 have been established in Theorems 13.4.22, 13.4.26.
At this point we introduce two functions that will play a crucial role in the sequel:

𝜋 𝜑 (𝑤, 𝑡, 𝑠, 𝜁) = 𝑖𝜑 (𝑤, 𝑡, 𝑤, 𝑠, 𝜁) , (26.3.7)

♮
𝜑 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) = 𝜑 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) + 𝑖𝜋 𝜑 (𝑤, 𝑡, 𝑠, 𝜁) . (26.3.8)

We shall sometimes refer to 𝜋 𝜑 as the exponent-weight. It follows from (26.3.5)

that
𝜕𝜋 𝜑
(𝑤, 𝑡, 𝑠, 𝜁) = 𝑞 (𝑤, 𝑡, 𝜕𝑧 𝜑 (𝑤, 𝑡, 𝑤, 𝑠, 𝜁)) ; (26.3.9)
𝜕𝑡
(26.3.6) implies

𝜕𝜋 𝜑
𝜋 𝜑 (𝑤, 𝑠, 𝑠, 𝜁) = 0, (𝑤, 𝑠, 𝑠, 𝜁) = 𝑞 (𝑤, 𝑠, 𝜁) , (26.3.10)
𝜕𝑡
and
1
𝜑 ♮ (𝑧, 𝑠, 𝑤, 𝑠, 𝜁) = (𝑧 − 𝑤) · 𝜁 + 𝑖 ⟨𝜁⟩ ⟨𝑧 − 𝑤⟩ 2 . (26.3.11)
2
Remark 26.3.1 The function (𝑧, 𝜃) ↦→ 𝜑 ♮ (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) is an FBI phase-function in
𝑈 ′C ×ℭ 𝜅 at (𝑤, −𝜉) ∈ 𝑈 ′C ×ℭ (Definition 22.1.1), for each (𝑡, 𝑠) ∈ Δ𝑟◦ (𝑡 ◦ ) ×Δ𝑟◦ (𝑡 ◦ )
provided diam 𝑈 ′C , diam ℭ ∩ S𝑛−2 , 𝑟 ◦ are sufficiently small.

Remark 26.3.2 Equation (26.3.9) and the Cauchy condition 𝜋 𝜑 (𝑤, 𝑠, 𝑠, 𝜁) = 0 de-
termine 𝜋 𝜑 unambiguously and are satisfied by 𝑖𝜑 (𝑤, 𝑡, 𝑤, 𝑠, 𝜁), whence (26.3.7).
26.3 Microlocal Solvability in Distributions 1145

We go to real space, replacing 𝑧, 𝑤, 𝜁 by 𝑥, 𝑦, 𝜉, respectively. Taylor expansion

with respect to 𝑥 about 𝑦 yields, in the notation of (26.3.8),

𝜑 ♮ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) = (𝑥 − 𝑦) · 𝜕𝑥 𝜑 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) + 𝑖𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) , (26.3.12)

with 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) ∈ C 𝜔 (𝑈 × 𝑈 × ℭ). Since 𝜕𝑥2 𝜑 (𝑦, 𝑠, 𝑦, 𝑠, 𝜉) = 𝑖 we have (as

usual, provided diam 𝑈 and diam ℭ ∩ S𝑛−2 are sufficiently small)

Re 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) ≥ 𝑐 ◦ |𝑥 − 𝑦| 2 |𝜉 | (26.3.13)

for some 𝑐 ◦ > 0 and all (𝑥, 𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈 × 𝑈 × ℭ. Of course, |𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)| ≲

|𝑥 − 𝑦| 2 |𝜉 |.

26.3.3 Properties of the exponent-weight 𝝅𝝋

We introduce the function

∫ 𝑡
𝑄 (𝑦, 𝑡, 𝑠, 𝜉) = 𝑞 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′. (26.3.14)
𝑠

The function 𝑄 is analytic in 𝑈 ′ × I◦2 × ℭ. We are going to use the notation

∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) = 𝜕𝑦 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) , |𝜉 | 𝜕 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) ,

(26.3.15)

the superscript h indicating that ∇h𝑦, 𝜉 preserves the homogeneity degree.

We are going to need the following result about the phase-function 𝜑; we use the
notation 𝜑 𝑥 = 𝜕𝑥 𝜑.

Lemma 26.3.3 After decreasing diam 𝑈 ′, diam ℭ ∩ S𝑛−2 , 𝑟 ◦ by an arbitrarily small

amount, there is a 𝐶 > 0 such that
∫ 𝑡
|𝜑 𝑥 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) − 𝜉 | ≤ 𝐶 ∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) + 𝐶 ∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) d𝑡 ′
𝑠

for all (𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈 ′ × I◦2 × ℭ.

Proof From (26.3.5) we derive

𝜕𝜑 𝑥
(𝑥, 𝑡, 𝑦, 𝑠, 𝜉) + 𝑖𝜕𝑥 𝑞 (𝑥, 𝑡, 𝜑 𝑥 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉))
𝜕𝑡
+ 𝑖𝜕𝑥2 𝜑 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝜕𝜃 𝑞 (𝑥, 𝑡, 𝜑 𝑥 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)) = 0

and from (26.3.6)

𝜑 𝑥 (𝑥, 𝑠, 𝑦, 𝑠, 𝜉) = 𝜉 + 𝑖 |𝜉 | (𝑥 − 𝑦) .
1146 26 Pseudodifferential Solvability and Property (Ψ)

If we define 𝑢 (𝑦, 𝑡, 𝑠, 𝜉) = 𝜑 𝑥 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) − 𝜉 we have

𝜕𝑢
(𝑦, 𝑡, 𝑠, 𝜉) = − 𝑖𝜕𝑥 𝑞 (𝑦, 𝑡, 𝜉 + 𝑢 (𝑦, 𝑡, 𝑠, 𝜉)) (26.3.16)
𝜕𝑡
− 𝑖𝜕𝑥2 𝜑 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) 𝜕𝜃 𝑞 (𝑥, 𝑡, 𝜉 + 𝑢 (𝑦, 𝑡, 𝑠, 𝜉))

and 𝑢 (𝑦, 𝑠, 𝑠, 𝜉) = 0. We derive

𝜕𝑢
(𝑦, 𝑡, 𝑠, 𝜉) = − 𝑖 𝜕𝑥 𝑞 (𝑦, 𝑡, 𝜉) + 𝜕𝑥2 𝜑 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) 𝜕𝜃 𝑞 (𝑥, 𝑡, 𝜉)
𝜕𝑡
+ 𝜇 (𝑦, 𝑡, 𝑠, 𝜉) 𝑢 (𝑦, 𝑡, 𝑠, 𝜉) ,

where 𝜇 is an (𝑛 − 1) × (𝑛 − 1) matrix with entries that are analytic functions in 𝑈 ′ ×

I◦2 × ℭ, homogeneous of degree zero. We obtain, using the notation 𝑀 (𝑦, 𝑡 ′, 𝑠, 𝜉) =
∫ 𝑡′
𝑠
𝜇 (𝑦, 𝑡 ′′, 𝑠, 𝜉) d𝑡 ′′,
∫ 𝑡
′
𝑖𝑢 (𝑦, 𝑡, 𝑠, 𝜉) = e 𝑀 ( 𝑦,𝑡 ,𝑠, 𝜉 ) 𝜕𝑦 𝑞 (𝑦, 𝑡 ′, 𝜉) + 𝜕𝑥2 𝜑 (𝑦, 𝑡 ′, 𝑦, 𝑠, 𝜉) 𝜕 𝜉 𝑞 (𝑥, 𝑡 ′, 𝜉) d𝑡 ′
𝑠
∫ 𝑡
′ 𝜕
= e 𝑀 ( 𝑦,𝑡 ,𝑠, 𝜉 )
𝜕 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) d𝑡 ′
′ 𝑦
𝑠 𝜕𝑡
∫ 𝑡
′ 𝜕
+ e 𝑀 ( 𝑦,𝑡 ,𝑠, 𝜉 ) 𝜕𝑥2 𝜑 (𝑦, 𝑡 ′, 𝑦, 𝑠, 𝜉) ′ 𝜕 𝜉 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) d𝑡 ′
𝑠 𝜕𝑡
= e 𝑀 ( 𝑦,𝑡 ,𝑠, 𝜉 ) 𝜕𝑦 𝑄 (𝑦, 𝑡, 𝑠, 𝜉)
+ e 𝑀 ( 𝑦,𝑡 ,𝑠, 𝜉 ) 𝜕𝑥2 𝜑 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) 𝜕 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉)
∫ 𝑡
𝜕 𝑀 ( 𝑦,𝑡 ′ ,𝑠, 𝜉 ) ′
− 𝜕𝑦 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) e d𝑡
𝑠 𝜕𝑡 ′
∫ 𝑡
𝜕 𝑀 ( 𝑦,𝑡 ′ ,𝑠, 𝜉 ) 2
− 𝜕 𝜉 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) e 𝜕 𝜑 (𝑦, 𝑡 ′
, 𝑦, 𝑠, 𝜉) d𝑡 ′,
𝑠 𝜕𝑡 ′ 𝑥

whence the claim. □

Lemma 26.3.4 After decreasing diam 𝑈 ′, diam ℭ ∩ S𝑛−2 , 𝑟 ◦ by an arbitrarily small

amount, we have
2
∫ 𝑡 2
Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) ≲ ∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) + ∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) d𝑡 ′.
𝑠

Proof We continue to use the notation 𝑢 (𝑦, 𝑡, 𝑠, 𝜉) = 𝜑 𝑥 (𝑦, 𝑡, 𝑦, 𝑠, 𝜉) − 𝜉. It follows

from (26.3.9) that
26.3 Microlocal Solvability in Distributions 1147

𝜕𝜋 𝜑 ∑︁ 1
(𝑦, 𝑡, 𝑠, 𝜉) = 𝑢 𝛼 (𝑦, 𝑡, 𝑠, 𝜉) 𝜕 𝜉𝛼 𝑞 (𝑦, 𝑡, 𝜉)
𝜕𝑡 𝑛−1
𝛼!
𝛼∈Z+

= 𝑞 (𝑦, 𝑡, 𝜉) + 𝑢 (𝑦, 𝑡, 𝑠, 𝜉) · 𝜕 𝜉 𝑞 (𝑦, 𝑡, 𝜉)

∑︁ 1
+ 𝑢 𝛼 (𝑦, 𝑡, 𝑠, 𝜉) 𝜕 𝜉𝛼 𝑞 (𝑦, 𝑡, 𝜉) .
𝛼!
| 𝛼 | ≥2

Integration yields
∫ 𝑡
Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) = 𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉) · 𝜕 𝜉 𝜕𝑡 𝑄 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′
𝑠
∑︁ 1 ∫ 𝑡
+ 𝑢 𝛼 (𝑦, 𝑡 ′, 𝑠, 𝜉) 𝜕 𝜉𝛼 𝜕𝑡 𝑄 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′.
𝛼! 𝑠
| 𝛼 | ≥2

Since 𝑢 (𝑦, 𝑠, 𝑠, 𝜉) ≡ 0 integration by parts yields

Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) = 𝜕 𝜉 𝑄 (𝑦, 𝑡, 𝜉) · Re 𝑢 (𝑦, 𝑡, 𝑠, 𝜉)

∫ 𝑡
− 𝜕 𝜉 𝑄 (𝑦, 𝑡 ′, 𝜉) · Re 𝜕𝑡 𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉) d𝑡 ′
𝑠
∑︁ 1 ∫ 𝑡
+ 𝑢 𝛼 (𝑦, 𝑡 ′, 𝑠, 𝜉) 𝜕 𝜉𝛼 𝜕𝑡 𝑄 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′.
𝛼! 𝑠
| 𝛼 | ≥2

From (26.3.16) we get

Re 𝜕𝑡 𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉) = Re −𝑖𝜕𝑥 𝜕 𝜉 𝑞 (𝑦, 𝑡 ′, 𝜉) 𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉) + 𝑂 |𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉)| 2

≲ |𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉)| ,

whence

Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) ≤ 𝜕 𝜉 𝑄 (𝑦, 𝑡, 𝜉) |𝑢 (𝑦, 𝑡, 𝑠, 𝜉)|

∫ 𝑡
+ 𝜕 𝜉 𝑄 (𝑦, 𝑡 ′, 𝜉) |𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉)| d𝑡 ′
𝑠
∫ 𝑡
+ 𝐶1 |𝑢 (𝑦, 𝑡 ′, 𝑠, 𝜉)| 2 d𝑡 ′,
𝑠

where 𝐶1 > 0 is independent of (𝑦, 𝑡, 𝑠, 𝜉). Combining the last inequality with
Lemma 26.3.3 proves the claim. □
We point out that the results so far have not required (𝚿) to hold.
1148 26 Pseudodifferential Solvability and Property (Ψ)

26.3.4 Partition of 𝑼 ′ × 𝕮 entailed by Condition (𝚿)

Condition (𝚿) in 𝑈 × ℭ can be stated as follows:

(𝚿) For every (𝑥, 𝜉) ∈ 𝑈 ′ × ℭ the function I◦ ∋ 𝑡 ↦→ 𝑞 (𝑥, 𝑡, 𝜉) does not change sign
from − to + at any point of I◦ .
We derive directly

Proposition 26.3.5 The validity of Condition (𝚿) in 𝑈 × ℭ is equivalent to the

property that 𝑈 ′ × ℭ is the union of the following four disjoint sets:
(1) The set N of points (𝑥, 𝜉) such that 𝑞 (𝑥, 𝑡, 𝜉) = 0 for all 𝑡;
(2) the set 𝔄 + (resp., 𝔄 − ) of points (𝑥, 𝜉) such that 𝑞 (𝑥, 𝑡, 𝜉) > 0 (resp., < 0) in the
whole interval I◦ except, possibly, on a discrete set;
(3) the set 𝔄 of points (𝑥, 𝜉) such that 𝑡 ↦→ 𝑞 (𝑥, 𝑡, 𝜉) changes sign from + to − at a
point 𝑡 (𝑥, 𝜉) ∈ I◦ .

In the sequel we preclude the trivial case 𝑞 ≡ 0. Then N is a proper analytic

subvariety of 𝑈 ′ × ℭ and therefore N has Lebesgue measure zero. The sets 𝔄 ± and
𝔄 ◦ are open; their boundaries in 𝑈 ′ × ℭ are the parts of N . If (𝑥, 𝜉) ∈ 𝔄 the point
𝑡 (𝑥, 𝜉) ∈ I◦ where 𝑡 ↦→ 𝑞 (𝑥, 𝑡, 𝜉) changes sign is unique. If (𝑥, 𝜉) ∈ 𝔄 + (resp., 𝔄 − )
the points of I◦ where we do not have 𝑞 (𝑥, 𝑡, 𝜉) > 0 (resp., < 0) are the zeros of
𝑞 (𝑥, 𝑡, 𝜉). After contracting 𝑈 × ℭ about (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ ) we may assume that the number
of these zeros is bounded independently of (𝑥, 𝜉) (on this last claim, see Ch. 14).

Lemma 26.3.6 If (𝚿) holds in 𝑈 × ℭ the function 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) [defined in (26.3.14)]

has the following properties:
(a) If (𝑦, 𝜉) ∈ N , 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) ≡ 𝑞 (𝑦, 𝑡, 𝜉) ≡ 0.
(b) If (𝑦, 𝜉) ∉ N , 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) < 0 in every one of the following subsets of 𝑈 ′ ×I◦2 ×ℭ:
(1) (𝑦, 𝜉) ∈ 𝔄 + , 𝑡 ◦ − 𝑟 ◦ < 𝑡 < 𝑠 < 𝑡 ◦ + 𝑟 ◦ , where 𝜕𝑄
𝜕𝑡 (𝑦, 𝑡, 𝑠, 𝜉) ≥ 0;
(2) (𝑦, 𝜉) ∈ 𝔄 − , 𝑡 ◦ − 𝑟 ◦ < 𝑠 < 𝑡 < 𝑡 ◦ + 𝑟 ◦ , where 𝜕𝑄
𝜕𝑠 (𝑦, 𝑡, 𝑠, 𝜉) ≥ 0;
(3) (𝑦, 𝜉) ◦ 𝜕𝑄
∈ 𝔄, 𝑡 (𝑦, 𝜉) ≤ 𝑠 < 𝑡 < 𝑡 + 𝑟 ◦ , where 𝜕𝑡 (𝑦, 𝑡, 𝑠, 𝜉) ≤ 0;
(4) (𝑦, 𝜉) ∈ 𝔄, 𝑡 ◦ − 𝑟 ◦ < 𝑡 < 𝑠 ≤ 𝑡 (𝑦, 𝜉), where 𝜕𝑄 𝜕𝑠 (𝑦, 𝑡, 𝑠, 𝜉) ≤ 0.

Proof If (𝚿) holds in 𝑈 × ℭ the following properties hold:

(1) 𝑞 (𝑦, 𝑡 ′, 𝜉) ≥ 0 if (𝑦, 𝜉) ∈ 𝔄 + for all 𝑡 ◦ − 𝑟 ◦ < 𝑡 ′ < 𝑡 ◦ + 𝑟 ◦
(2) 𝑞 (𝑦, 𝑡 ′, 𝜉) ≤ 0 if (𝑦, 𝜉) ∈ 𝔄 − and 𝑡 ◦ − 𝑟 ◦ ≤ 𝑡 ′ < 𝑡 ◦ + 𝑟 ◦ ;
(3) 𝑞 (𝑦, 𝑡 ′, 𝜉) ≤ 0 if (𝑦, 𝜉) ∈ 𝔄 and 𝑡 (𝑦, 𝜉) ≤ 𝑡 ′ < 𝑡 ◦ + 𝑟 ◦ ;
(4) 𝑞 (𝑦, 𝑡 ′, 𝜉) ≥ 0 if (𝑦, 𝜉) ∈ 𝔄 and 𝑡 ◦ − 𝑟 ◦ < 𝑡 ′ ≤ 𝑡 (𝑦, 𝜉).
These properties of 𝑞 have the following consequences:
26.3 Microlocal Solvability in Distributions 1149

If (𝑦, 𝜉) ∈ 𝔄 + and 𝑡 ◦ − 𝑟 ◦ < 𝑡 < 𝑠 < 𝑡 ◦ + 𝑟 ◦ , we have

∫ 𝑠
𝑄 (𝑦, 𝑡, 𝑠, 𝜉) = − 𝑞 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′ < 0
𝑡

and 𝜕𝑄
𝜕𝑡 (𝑦, 𝑡, 𝑠, 𝜉) ≥ 0. If (𝑦, 𝜉) ∈ 𝔄 − and 𝑡 ◦ − 𝑟 ◦ < 𝑠 < 𝑡 < 𝑡 ◦ + 𝑟 ◦ , we have
∫ 𝑡
𝑄 (𝑦, 𝑡, 𝑠, 𝜉) = 𝑞 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′ < 0
𝑠

and 𝜕𝑄 ◦
𝜕𝑠 (𝑦, 𝑡, 𝑠, 𝜉) ≥ 0. If (𝑦, 𝜉) ∈ 𝔄 and if 𝑡 (𝑦, 𝜉) ≤ 𝑠 < 𝑡 < 𝑡 + 𝑟 ◦ or if
◦
𝑡 − 𝑟 ◦ < 𝑡 < 𝑠 ≤ 𝑡 (𝑦, 𝜉),
∫ 𝑡
𝑄 (𝑦, 𝑡, 𝑠, 𝜉) = 𝑞 (𝑦, 𝑡 ′, 𝜉) d𝑡 ′ < 0.
𝑠

If 𝑡 (𝑦, 𝜉) ≤ 𝑠 < 𝑡 < 𝑡 ◦ + 𝑟 ◦ then 𝜕𝑄 ◦

𝜕𝑡 (𝑦, 𝑡, 𝑠, 𝜉) = 𝑞 (𝑦, 𝑡, 𝜉) ≤ 0; if 𝑡 − 𝑟 ◦ < 𝑡 < 𝑠 ≤
𝜕𝑄
𝑡 (𝑦, 𝜉) then 𝜕𝑠 (𝑦, 𝑡, 𝑠, 𝜉) = −𝑞 (𝑦, 𝑠, 𝜉) ≤ 0. □

Lemma 26.3.7 If (𝚿) holds in 𝑈 × ℭ there are a neighborhood 𝑈◦′ ⊂⊂ 𝑈 ′ of 𝑥 ◦ in

R𝑛−1 and an open cone ℭ◦ in R𝑛−1 , 𝜉 ◦ ∈ ℭ◦ ∩S𝑛−2 ⊂⊂ ℭ, such that

2 ∑︁
∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉)
𝛽
≤ −4 ­ sup 𝜕𝑦𝛼 𝜕 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) ® 𝑄 (𝑦, 𝑡, 𝑠, 𝜉)
© ª
′ 2
«𝑈 ×I◦ ×ℭ | 𝛼+𝛽 |=2 ¬
(26.3.17)
for every (𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈◦′ × I◦2 × ℭ◦ belonging to one of the subsets of 𝑈 ′ × I◦2 × ℭ,
(1),(2),(3),(4), in Lemma 26.3.6, or such that (𝑦, 𝜉) ∈ N .

Proof Follows directly from Lemma 26.3.6 and Lemma 1.1 in [Nirenberg-Treves,
1970], pp. 5 et seq., whose proof, based on the Implicit Function Theorem, need not
be reproduced here. In the sets (1), (2), in Lemma 26.3.6, the proof of the claim is
more routine. □

Corollary 26.3.8 If (𝚿) holds in 𝑈 × ℭ we have

2
∇h𝑦, 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) ≲ − |𝑡 − 𝑠| 𝑄 (𝑦, 𝑡, 𝑠, 𝜉)

for every (𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈◦′ × I◦2 × ℭ◦ belonging to one of the subsets of 𝑈 ′ × I◦2 × ℭ,
(1),(2),(3),(4), in Lemma 26.3.6, or such that (𝑦, 𝜉) ∈ N .

Proof We have 𝜕𝑦𝛼 𝜕 𝜉 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) ≲ |𝑡 − 𝑠| for all (𝛼, 𝛽) ∈ Z2(𝑛−1)

𝛽
+ . The claim then
follows from (26.3.17). □
1150 26 Pseudodifferential Solvability and Property (Ψ)

26.3.5 A new formulation of Condition (𝚿)

We continue to use the notation of Lemma 26.3.7.

Lemma 26.3.9 Assume that Condition (𝚿) holds in 𝑈 × ℭ. Possibly after decreasing
𝑟 ◦ we have
1
Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) ≤ − 𝑄 (𝑦, 𝑡, 𝜉) (26.3.18)
2
for every (𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈◦′ × I◦2 × ℭ◦ belonging to one of the subsets of 𝑈 ′ × I◦2 × ℭ,
(1),(2),(3),(4), in Lemma 26.3.6, or such that (𝑦, 𝜉) ∈ N .

Proof We combine Lemma 26.3.4 and Corollary 26.3.8

∫ 𝑡
Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) ≲ − |𝑡 − 𝑠| 𝑄 (𝑦, 𝑡, 𝑠, 𝜉)− |𝑡 ′ − 𝑠| 𝑄 (𝑦, 𝑡 ′, 𝑠, 𝜉) d𝑡 ′.
𝑠

In the subsets (1),(2),(3),(4) of 𝑈 ′ × I◦2 × ℭ in Lemma 26.3.6, 𝑄 is monotone

increasing and 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) < 0 if 𝑡 ≠ 𝑠, whence

Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) − 𝑄 (𝑦, 𝑡, 𝜉) ≲ −4𝑟 ◦ 𝑄 (𝑦, 𝑡, 𝑠, 𝜉) .

The claim ensues. □

Corollary 26.3.10 Assume that Condition (𝚿) holds in 𝑈 × ℭ. Possibly after de-
creasing diam 𝑈 ′, diam ℭ∩S𝑛−2 , 𝑟 ◦ , we have Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) ≤ 0 for every

(𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈◦′ ×I◦2 ×ℭ◦ belonging to one of the subsets of 𝑈 ′ ×I◦2 ×ℭ, (1),(2),(3),(4),
in Lemma 26.3.6, or such that (𝑦, 𝜉) ∈ N .

Proof Indeed (26.3.18) implies Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) ≤ 21 𝑄 (𝑦, 𝑡, 𝜉) and 𝑄 (𝑦, 𝑡, 𝜉) ≤ 0

in said subsets. □

We are now in a position to prove the result we were aiming for.

Proposition 26.3.11 Assuming that diam 𝑈◦′ , diam ℭ◦ ∩S𝑛−2 , 𝑟 ◦ , are sufficiently

small, Condition (𝚿) in 𝑈 × ℭ is equivalent to the following condition:

(R) Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) ≤ 0 for every (𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈◦′ × I◦2 × ℭ◦ belonging to one of the
subsets of 𝑈 ′ × I◦2 × ℭ, (1),(2),(3),(4), in Lemma 26.3.6, or such that (𝑦, 𝜉) ∈ N .

Proof Corollary 26.3.10 states that (𝚿)=⇒(R). We prove the converse. We assume
that (𝑦, 𝜉) ∉ N , implying 𝑞 (𝑦, 𝑠, 𝜉) ≠ 0 for all 𝑠 with the possible exception of the
zeros of 𝑡 ↦→ 𝑞 (𝑦, 𝑡, 𝜉) (which form a discrete set). We have, by (26.3.10),

Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) = (𝑡 − 𝑠) 𝑞 (𝑦, 𝑠, 𝜉) + 𝑂 (𝑡 − 𝑠) 2 . (26.3.19)
26.3 Microlocal Solvability in Distributions 1151

Suppose 𝑡 ↦→ 𝑞 (𝑦, 𝑡, 𝜉) changes sign from − to + at 𝑡 ∗ ∈ I◦ [this is the negation

of (𝚿)]; 𝑡 ∗ is an isolated zero of 𝑞 (𝑦, 𝑡, 𝜉). Take 𝑡 ∗ − 𝑟 < 𝑡 < 𝑠 < 𝑡 ∗ , 𝑟 > 0 arbitrarily
small, in which case 𝑞 (𝑦, 𝑠, 𝜉) (𝑡 − 𝑠) > 0 and Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) > 0 by (26.3.19);
and the same conclusion if 𝑡 ∗ < 𝑠 < 𝑡 < 𝑡 ∗ + 𝑟. Thus (R) does not hold in a
neighborhood of (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ ). This proves that (R) implies (𝚿) in 𝑈 × ℭ. □

Remark 26.3.12 Proposition 26.3.11 entails the invariance of (R) under coordinate
changes or multiplication of 𝑃 by an elliptic factor, since we know (Theorem 26.2.8)
this to be true of (𝚿 ( 𝑥 ◦ ,𝑡 ◦ , 𝜉 ◦ ) ).

26.3.6 Approximate microlocal parametrices

In the sequel 𝜑 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) will be the phase-function solution of (26.3.5)–(26.3.6)

and 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) will be a classical amplitude of degree zero, both defined in
𝑈 × 𝑈 × ℭ. We define, for 𝜀 > 0 and 𝑓 ∈ 𝐿 c1 (𝑈),

(2𝜋) 𝑛−1 𝑨 ( 𝜀) 𝑓 (𝑥, 𝑡) (26.3.20)

∫ ∫ 𝑡 2
= e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )−𝜀 | 𝜉 | 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
𝔄+ 𝑡 ◦ −𝑟 ◦
∫ ∫ 𝑡 ◦ +𝑟◦ 2
− e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )−𝜀 | 𝜉 | 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
𝔄− 𝑡
∫ ∫ 𝑡 2
+ e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )−𝜀 | 𝜉 | 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦,
𝔄 𝑡 ( 𝑦, 𝜉 )

where:
𝔄 and 𝔄 ± are the (conic) open ′
subsets of 𝑈 × ℭ in Proposition 26.3.5;
∞
𝜒 = 𝜙𝜓𝜌 , 𝜙 ∈ Cc ℭ∩S 𝑛−2 , 𝜙 ≡ 1 in a neighborhood of 𝜉 ◦ , extended to R𝑛 \ {0}
as a homogeneous function of degree zero, 𝜓𝜌 ∈ C ∞ (R𝑛 ) such that 𝜓𝜌 (𝜉) = 0 for
|𝜉 | < 𝜌, 𝜓𝜌 (𝜉) = 1 for |𝜉 | > 2𝜌, 𝜌 > 0 large, chosen in the next subsection.
Note that, in the third integral, we can have 𝑡 < 𝑡 (𝑦, 𝜉) as well as 𝑡 > 𝑡 (𝑦, 𝜉).
The amplitude 𝑎 is selected so that

(D𝑡 + 𝑖𝑞 (𝑧, 𝑡, D𝑤 )) e𝑖 𝜑 (𝑧,𝑡 ,𝑤,𝑠,𝜁 ) 𝑎 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) = 0, (26.3.21)

and
1
𝑎 (𝑧, 𝑠, 𝑤, 𝑠, 𝜁) = 1 + 𝑖 ⟨𝜁⟩ −1 (𝑧 − 𝑤) · 𝜁. (26.3.22)
2
Actually, we shall take 𝑎 to be a finite realization (Definition 19.1.7) of the formal
analytic series
∞
∑︁
𝑎 (𝑧, 𝑡, 𝑤, 𝑠, 𝜉) = 𝑎 ℓ (𝑧, 𝑡, 𝑤, 𝑠, 𝜉) (26.3.23)
ℓ=0
1152 26 Pseudodifferential Solvability and Property (Ψ)

with 𝑎 ℓ (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) ∈ O 𝑈C × 𝑈 C × ℭ 𝜅 homogeneous of degree −ℓ. The homo-
geneous parts 𝑎 ℓ ∈ O U ′C are recursively determined by the transport equations
[cf. (18.3.26)] under the initial conditions [cf. (26.3.22)]
1
𝑎 0 (𝑧, 𝑠, 𝑤, 𝑠, 𝜁) = 1 + 𝑖 ⟨𝜁⟩ −1 (𝑧 − 𝑤) · 𝜁, (26.3.24)
2
𝑎 ℓ (𝑧, 𝑠, 𝑤, 𝑠, 𝜁) = 0 if ℓ ≥ 1.

The estimates needed to establish that 𝑎 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) is a classical analytic amplitude

(Definition 17.2.28) are the analogues of those in the proof of Theorem 5.1.13, and are
derived recursively in the same manner. Note that 𝑎 0 (𝑧, 𝑠, 𝑤, 𝑠, 𝜁) = Δ1/2 (𝑧 − 𝑤, 𝜁)
in the notation (3.4.6).
For fixed 𝜀 > 0, 𝑨 ( 𝜀) 𝑓 (𝑥, 𝑡) ∈ C (𝑈), C 𝜔 with respect to 𝑥.

26.3.7 Convergence of 𝑨 (𝜺)

The convergence of 𝑨 ( 𝜀) will depend on the behavior of Re 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉), which

makes the validity of Property (R) (cf. Proposition 26.3.11) crucial.
We shall make use of the notation (26.3.12) and the following function:

𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) = ⟨𝜉 + 𝜕𝑥 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)⟩ 2 − 𝑖Δ 𝑥 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) (26.3.25)

with Δ 𝑥 the Laplacian in 𝑥-space R𝑛−1 . By (26.3.13) and requiring diam 𝑈 to be

sufficiently small we can ensure that
3 2
⟨𝜉 + 𝜕𝑥 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)⟩ 2 ≥ |𝜉 | .
4
Since Δ 𝑥 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) is homogeneous of degree 1 we can select the number 𝜌
sufficiently large that |𝜉 | > 𝜌 implies |Δ 𝑥 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)| < 41 |𝜉 | 2 , thereby getting

1 2
|𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)| ≥ |𝜉 | (26.3.26)
2
for all (𝑥, 𝑦, 𝑡, 𝑠, 𝜉) ∈ 𝑈 × 𝑈 × ℭ, |𝜉 | > 𝜌.
Let 𝑏 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) stand for a classical complex-analytic amplitude in 𝑈 C × 𝑈 C ×
ℭ 𝜅 of the same type as 𝑎 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) but of arbitrary degree 𝑚. A straightforward
calculation shows that
♮
− e−𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) Δ 𝑥 e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
♮

= 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) − Δ 𝑥 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)

+ 2D 𝑥 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) · (𝜉 + 𝜕𝑥 𝐹 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)) .
26.3 Microlocal Solvability in Distributions 1153

In turn, this implies

♮
− 𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) −1 Δ 𝑥 e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
♮ ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )
= e𝑖 𝜑 (𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) + 𝑏 1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)) ,

where 𝑏 1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) is a classical amplitude of degree 𝑚−1 extending as a complex-

analytic classical formal series 𝑏 1 (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) in 𝑈 C × 𝑈 C × ℭ 𝜅 , |Re 𝜁 | > 𝜌. We
derive easily
♮ ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )
e𝑖 𝜑 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)

𝑖 𝜑 ♮ ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
+ e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑏 1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) ,
♮
= −Δ 𝑥 e
𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)

with 𝑏♭1 similar to 𝑏 1 , deg 𝑏♭1 = 𝑚 − 1. Induction on 𝑁 = 1, 2, ..., leads to the formula
♮ ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )
e𝑖 𝜑 𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) (26.3.27)

𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
= (−1) 𝑁 Δ 𝑥𝑁 e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )
♮

𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁
𝑁
!
∑︁
𝑁 −ℓ 𝑁 −ℓ 𝑖 𝜑 ♮ ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )
𝑏♭ℓ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
+ (−1) Δ𝑥 e ,
ℓ=1 𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁 −ℓ

where 𝑏♭ℓ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) is a classical amplitude of degree −ℓ extending as a complex-

analytic classical formal series 𝑏 ℓ (𝑧, 𝑡, 𝑤, 𝑠, 𝜁) in 𝑈 C × 𝑈 C × ℭ 𝜅 , |Re 𝜁 | > 𝜌. We
point out that the degrees of the amplitudes

𝑏 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑏♭ℓ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)

,
𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁
𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁 −ℓ
are respectively 𝑚 − 2𝑁 and 𝑚 − 2𝑁 + ℓ ≤ 𝑚 − 𝑁.

Proposition 26.3.13 Suppose that Property (𝚿) holds in 𝑈 × ℭ ′. Let 𝑨 ( 𝜀) be defined

as in (26.3.20). If diam 𝑈 is sufficiently small and 𝜌 > 0 sufficiently large, then 𝑨 ( 𝜀)
converges, as 𝜀 ↘ 0, to the continuous linear operator 𝑨 : 𝐿 c1 (𝑈) −→ D ′ (𝑈)
defined by the oscillatory integrals

𝑨 𝑓 (𝑥, 𝑡) (26.3.28)
∫ ∫ 𝑡
1
= e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
(2𝜋) 𝑛−1 𝔄+ 𝑡 ◦ −𝑟◦
∫ ∫ 𝑡 ◦ +𝑟◦
1
− e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
(2𝜋) 𝑛−1 𝔄− 𝑡
∫ ∫ 𝑡
1
+ 𝑛−1
e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦.
(2𝜋) 𝔄 𝑡 ( 𝑦, 𝜉 )
1154 26 Pseudodifferential Solvability and Property (Ψ)

Proof Recall (26.3.7)–(26.3.8):

𝑖𝜑 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) = 𝑖𝜑 ♮ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) + 𝜋 𝜑 (𝑦, 𝑡, 𝑠, 𝜉) .

Let 𝑓 ∈ 𝐿 c1 (𝑈) be arbitrary. From (26.3.27) where we put 𝑏 = 𝑎 (and then 𝑏♭ℓ = 𝑏 ℓ ,
see above) we derive
∫ ∫ 𝑡
2
e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )−𝜀 | 𝜉 | 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
𝔄+ 𝑡 ◦ −𝑟◦
∫ ∫ 𝑡 2 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
= (−1) 𝑁 Δ 𝑥𝑁 e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )−𝜀 | 𝜉 | 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
𝔄+ 𝑡 ◦ −𝑟 ◦ 𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁
𝑁 ∫ ∫ 𝑡 2
∑︁
𝑁 −ℓ
+ (−1) Δ 𝑥𝑁 −ℓ e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 )−𝜀 | 𝜉 |
ℓ=1 𝔄+ 𝑡 ◦ −𝑟 ◦

𝑏 ℓ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
× 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦
𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁 −ℓ
If we take 𝑁 sufficiently large then (R) (cf. Proposition 26.3.11) entails that the
integrands in
∫ ∫ 𝑡
𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦,
𝔄 + 𝑡 ◦ −𝑟◦ 𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁
∫ ∫ 𝑡
𝑏 ℓ (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)
e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑓 (𝑦, 𝑠) 𝜒 (𝜉) d𝑠d𝜉d𝑦, ℓ = 1, ..., 𝑁,
𝔄 + ◦
𝑡 −𝑟◦ 𝐹1 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑁 −ℓ

are absolutely integrable functions of 𝜉 ∈ ℭ, |𝜉 | > 𝜌. Likewise for the second and
third integral in the right-hand side of (26.3.28), whence the claim. □

It is convenient to select arbitrarily a neighborhood of (𝑥 ◦ , 𝑡 ◦ ), 𝑈◦ = 𝑈◦′ ×

(𝑡 ◦− 𝑟 1 , 𝑡 ◦ + 𝑟 1 ) ⊂⊂ 𝑈 (thus 𝑟 1 < 𝑟 ◦ ), and assume that 𝑓 ∈ 𝐿 c1 (𝑈◦ ). We also select
a cutoff function 𝑔◦ ∈ Cc∞ (𝑈 ′), 𝑔◦ ≡ 1 in 𝑈◦′ , implying 𝑔◦ (𝑥) 𝑓 (𝑥, 𝑡) = 𝑓 (𝑥, 𝑡).
Taking the amplitude 𝑎 to be a finite realization of a formal classical amplitude has
the effect that the equation (26.3.21) will be only approximately satisfied, introducing
errors that decay exponentially as |𝜉 | ↗ +∞. We define an integral operator 𝑱 ◦ by
substituting

D𝑡 + 𝑖𝑞 𝑥, 𝑡, D 𝑦 𝑔◦ (𝑥) e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉)

for e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) in (26.3.28); 𝑱 ◦ defines a continuous linear operator

𝐿 c1 (𝑈◦ ) −→ D ′ (𝑈 ′), analytic regularizing with respect to 𝑥 in 𝑈◦′ and depending
in C 𝜔 fashion on 𝑡, meaning that its associated kernel distribution 𝐽◦ (𝑥, 𝑦, 𝑡) is of
class C 𝜔 in 𝑈◦′ × 𝑈◦ . Then these properties are also valid for

𝑲 ◦ = 𝑔◦ (𝑥) 𝑱 ◦ − [𝑞 (𝑥, 𝑡, D 𝑥 ) , 𝑔◦ (𝑥)] 𝑱 ◦ .

26.3 Microlocal Solvability in Distributions 1155

We reach the following conclusion:

𝜕
− 𝑞 (𝑥, 𝑡, D 𝑥 ) (𝑔◦ 𝑨 𝑓 ) (𝑥, 𝑡) = 𝑔◦𝔉𝜒 𝑓 + 𝑲 ◦ 𝑓 (𝑥, 𝑡) , (26.3.29)
𝜕𝑡

where
∫
1 1 2
𝔉𝜒 𝑓 (𝑥, 𝑡) = e𝑖 ( 𝑥−𝑦) · 𝜉 − 2 | 𝜉 | |𝑥−𝑦 | 𝑓 (𝑦, 𝑡) Δ1/2 (𝑥 − 𝑦, 𝜉) 𝜒 (𝜉) d𝜉d𝑦.
(2𝜋) 𝑛−1 R2(𝑛−1)
(26.3.30)
In (26.3.30) the phase-function and the determinant Δ1/2 are those of the standard
FBI transform in R𝑛−1 [cf. (3.4.8)]; of course, the cut-off function 𝜒 (𝜉) is a novelty.

26.3.8 Sufficiency of Condition (𝚿) for solvability of 𝑷 in

microdistributions

We avail ourselves of (26.3.29) to prove the distribution version of the Trépreau

Theorem. We apply the inversion formula (3.4.10) with 𝜅 = 1/2; we have, for
arbitrary 𝑢 ∈ E ′ R𝑛−1 ,
∫ ∫
1 𝑖 ( 𝑥−𝑦) · 𝜉 − 12 | 𝜉 | |𝑥−𝑦 | 2
𝑢 (𝑥) = e 𝑢 (𝑦) Δ1/2 (𝑥 − 𝑦, 𝜉) d𝑦 d𝜉.
(2𝜋) 𝑛−1 R𝑛−1 R𝑛−1

But now we let 𝑢 depend also on 𝑡 ∈ I◦ : 𝑢 shall be a compactly supported integrable

function of 𝑡 ∈ (𝑡 ◦ − 𝑟 1 , 𝑡 ◦ + 𝑟 1 ) valued in E ′ (𝑈 ′). We can select 𝑘 ∈ Z+ such that
(1 − Δ 𝑥 ) −𝑘 𝑢 ∈ 𝐿 1 (R𝑛 ). We define 𝑓 = (1 − Δ 𝑥 ) −𝑘 𝑢 in 𝑈1 = 𝑈1′ × (𝑡 ◦ − 𝑟 1 , 𝑡 ◦ + 𝑟 1 ),
𝑥 ◦ ∈ 𝑈1′ ⊂⊂ 𝑈◦′ , 𝑓 ≡ 0 in R𝑛 \𝑈1 ; thus 𝑓 ∈ 𝐿 c1 (𝑈◦ ) and therefore 𝑔◦ (𝑥) 𝑓 (𝑥, 𝑡) =
𝑓 (𝑥, 𝑡). We derive from (26.3.29)

𝜕
− 𝑞 (𝑥, 𝑡, D 𝑥 ) (𝑔◦ 𝑨 𝑓 ) − 𝑓 − 𝑲 ◦ 𝑓 = (26.3.31)
𝜕𝑡
∫
1 1 2

𝑛−1
𝑔 ◦ (𝑥) e𝑖 ( 𝑥−𝑦) · 𝜉 − 2 | 𝜉 | |𝑥−𝑦 | 𝑓 (𝑦, 𝑡) Δ1/2 (𝑥 − 𝑦, 𝜉) (1 − 𝜒 (𝜉)) d𝜉d𝑦.
(2𝜋) R2(𝑛−1)

Lemma 26.3.14 If |𝜏| ≤ 𝐶 |𝜉 |, 𝐶 > 0, the point ((𝑥 ◦ , 𝑡) , (𝜉, 𝜏)) ∉ 𝑊 𝐹a (𝑲 ◦ 𝑓 )

whatever 𝑡 ∈ I◦ .

Proof We look at the FBI transform

∫
˜ | 2 +|𝑡−𝑠 | 2 )
· 𝜉 +𝑖 𝜏 (𝑡−𝑠)− 21 | ( 𝜉 , 𝜏) | ( | 𝑥−𝑥
e𝑖 ( 𝑥−𝑥)
˜
𝑲 ◦ 𝑓 (𝑥, 𝑠) d𝑥d𝑠 (26.3.32)
𝑈
1156 26 Pseudodifferential Solvability and Property (Ψ)
√︃
(|(𝜉, 𝜏)| = |𝜉 | 2 + 𝜏 2 ; we suppose |𝜏| ≤ 𝐶 |𝜉 |). After contracting 𝑈 ′ about 𝑥 ◦ we can
assume that 𝑥 ↦→ 𝑲 ◦ 𝑓 (𝑥, 𝑠) extends holomorphically to a domain in C𝑛−1 containing
𝑈 ′; we take 𝜕𝑈 ′ to be smooth. We deform the contour of integration in 𝑥-space from
𝑈 ′ to the image of 𝑈 ′ under the map 𝑥 ↦→ 𝑥 − 𝑖𝜀𝜉/|𝜉 | with 𝜀 > 0 suitably small. The
Cauchy Integral Theorem equates (26.3.32) to the sum of two integrals,
∫ 𝑡 ◦ +𝑟◦ ∫ 𝜆=𝜀 ∫
2
1
· 𝜉 +𝑖 𝜏 (𝑡−𝑠)−𝜆| 𝜉 |− 2 | ( 𝜉 , 𝜏) | ( | 𝑥−𝑥+𝑖𝜆 +|𝑡−𝑠 | 2 )
e𝑖 ( 𝑥−𝑥)
˜ ˜ 𝜉/| 𝜉 | |
𝑡 ◦ −𝑟◦ 𝜆=0 𝜕𝑈 ′
× 𝑲 ◦ 𝑓 (𝑥, 𝑠) d𝜎 (𝑥) d𝜆d𝑠, (26.3.33)

in which d𝜎 (𝑥) is the appropriate volume element on 𝜕𝑈 ′, and

∫
· 𝜉 +𝑖 𝜏 (𝑡−𝑠)−𝜀 | 𝜉 |− 21 | ( 𝜉 , 𝜏) | ( | ⟨ 𝑥−𝑥+𝑖 𝜀 𝜉 / | 𝜉 | ⟩ | 2 +|𝑡−𝑠 | 2 )
e𝑖 ( 𝑥−𝑥)
˜ ˜
𝑲 ◦ 𝑓 (𝑥 − 𝑖𝜀𝜉/|𝜉 | , 𝑠) d𝑥d𝑠.
𝑈
(26.3.34)
The first integral, (26.3.33), decays exponentially as |𝜉 | ↗ ∞ provided 𝑥˜ stays in
a neighborhood of 𝑥 ◦ sufficiently distant from 𝜕𝑈 ′. The real part of the exponent in
(26.3.34) is equal to
1 1
−𝜀 |𝜉 | + 𝜀 2 |(𝜉, 𝜏)| − |(𝜉, 𝜏)| | 𝑥˜ − 𝑥| 2 + |𝑡 − 𝑠| 2 .
2 2
If 𝜀 > 0 sufficiently small then (26.3.34) decays exponentially as |𝜉 | ↗ +∞. Our
claims about these two integrals remain valid if we allow ( 𝑥,
˜ 𝑡) to vary a little in
C𝑛 . □
Keep in mind that (supp 𝜒) ∩ S𝑛−2 ⊂⊂ ℭ ∩ S𝑛−2 . Since 𝜒 (𝜉) ≡ 1 in a cone
in R𝑛−1 \ {0} containing 𝜉 ◦ it is readily proved (by the same reasoning as in the
proof of Lemma 26.3.14) that, if |𝜏| ≤ 𝐶 |𝜉 |, then ((𝑥 ◦ , 𝑡) , (𝜉 ◦ , 𝜏)) does not belong
to the analytic wave-front set of the right-hand side of (26.3.31), whatever 𝑡 ∈ I◦ .
Combining this with Lemma 26.3.14 shows that

◦ ◦ ◦ 𝜕
((𝑥 , 𝑡 ) , (𝜉 , 0)) ∉ 𝑊 𝐹a − 𝑞 (𝑥, 𝑡, D 𝑥 ) (𝑔◦ 𝑨 (𝑔◦ 𝑓 )) − 𝑓 . (26.3.35)
𝜕𝑡
−𝑘
Now let 𝑓 ∈ E ′ (𝑈) be arbitrary; there is a 𝑘 ∈ Z+ such that 1 − Δ 𝑥,𝑡 𝑓 ∈ 𝐿 1 (R𝑛 ).
Let 𝑔 ∈ Cc 𝑈1 , 𝑔 ≡ 1 in a neighborhood of 𝑥 . Substituting 𝑔 (1 − Δ 𝑥 ) −𝑘 𝑓 for 𝑓
∞ ′ ◦

in (26.3.35) yields

◦ ◦ ◦ 𝜕 −𝑘
((𝑥 , 𝑡 ) , (𝜉 , 0)) ∉ 𝑊 𝐹a − 𝑞 (𝑥, 𝑡, D 𝑥 ) 𝑮 ◦ 𝑓 − 𝑔◦ (1 − Δ 𝑥 ) 𝑓 ,
𝜕𝑡
(26.3.36)
where
𝑮 ◦ 𝑓 = 𝑔◦ 𝑨 𝑔 (1 − Δ 𝑥 ) −𝑘 𝑓 .
−𝑘
Since 𝑊 𝐹a ( 𝑓 ) = 𝑊 𝐹a 1 − Δ 𝑥,𝑡 𝑓 we reach the sought conclusion:
26.3 Microlocal Solvability in Distributions 1157

Theorem 26.3.15 Suppose Condition (𝚿) holds in 𝑈 × ℭ. Provided diam 𝑈 and

diam ℭ ∩ S𝑛−2 are sufficiently small we have

𝜕
((𝑥 ◦ , 𝑡 ◦ ) , (𝜉 ◦ , 0)) ∉ 𝑊 𝐹a − 𝑞 (𝑥, 𝑡, D 𝑥 ) 𝑮 ◦ 𝑓 − 𝑓
𝜕𝑡

for every 𝑓 ∈ E ′ (𝑈).

Remark 26.3.16 Recall that 𝑞 (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ ) = 0. At points ((𝑥 ◦ , 𝑡 ◦ ) , (𝜉, 𝜏)) such that
𝜏 2 + 𝑞 2 (𝑥 ◦ , 𝑡 ◦ , 𝜉) ≠ 0 the (microdifferential) operator 𝜕𝑡𝜕
− 𝑞 (𝑥, 𝑡, D 𝑥 ) is elliptic
and therefore microlocally invertible, meaning
that
there is a 𝑢 ∈ E ′ (𝑈) such that
◦ ◦ 𝜕
((𝑥 , 𝑡 ) , (𝜉, 𝜏)) ∉ 𝑊 𝐹a 𝜕𝑡 − 𝑞 (𝑥, 𝑡, D 𝑥 ) 𝑢 − 𝑓 .

If D𝑡 + 𝑖𝑞 (𝑥, 𝑡, D 𝑥 ) is the transform of a pseudodifferential operator 𝑃 of order

𝑚 ≥ 2 defined locally, we can retrace the transformations that brought us from 𝑃 to
D𝑡 +𝑖𝑞 (𝑥, 𝑡, D 𝑥 ): first a division by an elliptic factor of order 𝑚 − 1, then conjugation
with an elliptic Fourier integral operator with real phase. These operations are
strictly microlocal, in a conic neighborhood of the central point ((𝑥 ◦ , 𝑡 ◦ ) , (𝜉 ◦ , 0))
with 𝑞 (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ ) = 0 and so is the selection of the space-time coordinates. If we
revert to the generic coordinates 𝑥 𝑗 and cocoordinates 𝜉 𝑗 and denote the central point
by (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 we can rephrase Theorem 26.3.15 as follows:

Theorem 26.3.17 Let 𝑃 be an analytic pseudodifferential operator of principal type

in a domain Ω of R𝑛 satisfying Condition (𝚿) in a conic neighborhood U of
(𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑃 in Ω × (R𝑛 \ {0}). Provided diam U ∩ Ω × S𝑛−1 is sufficiently

small there is a continuous linear operator 𝑮 : E ′ (𝑈) −→ D ′ (𝑈) (𝑈: base
projection of U) such that

∀ 𝑓 ∈ E ′ (𝑈) , (𝑥 ◦ , 𝜉 ◦ ) ∉ 𝑊 𝐹a (𝑃𝑮 𝑓 − 𝑓 ) .

26.3.9 Open questions

We shall only mention the most evident of the open questions that come to mind in
connection with Theorems 26.3.15, 26.3.17.
(1) Can the parametrix 𝑮 ◦ (in Theorem 26.3.15) be adapted to analyze the exis-
tence (or nonexistence) and the propagation of the analytic singularities of the
distributions 𝑢 that satisfy (D𝑡 + 𝑖𝑞 (𝑥, 𝑡, D 𝑥 )) 𝑢 ∈ C 𝜔 (𝑈)?
(2) As a consequence of the result in [Dencker, 2006] we know that 𝑃𝑢 =
𝑓 ∈ 𝐿 c1 (𝑈◦ ) is solvable in 𝑈◦ (sufficiently small) under the hypotheses that
d 𝜉 𝑝 (𝑥 ◦ , 𝜉) ≠ 0 for all 𝜉 ∈ R𝑛 \ {0} (𝑝: principal symbol of 𝑃) and, needless to
say, that 𝑃 satisfies Condition (𝚿) at every point of Char 𝑃. In passing we point
out that 𝑃𝑢 = 𝑓 is not locally solvable, generally speaking, if 𝑃 is of principal
1158 26 Pseudodifferential Solvability and Property (Ψ)

type in the sense of Definition 23.1.2, even when 𝑃 is a real vector field. For
𝜕𝑢
instance, if 𝑓 (𝑥) = 𝑥 2 𝜕𝑥 − 𝑥 1 𝜕𝑢 in the unit disk, with 𝑢 a C 1 function in the
1
∫ 2 𝜋 𝜕𝑥2
disk, then 𝑃𝑢 = 𝑓 entails 0 𝑓 (𝑟 cos 𝜃, 𝑟 sin 𝜃) d𝜃 = 0.
Assuming d 𝜉 𝑝 ≠ 0 at every point of Char 𝑃, the question is how to construct
a local (as opposed to microlocal) parametrix for 𝑃 by adapting the techniques
introduced in this section. We can replace the cone ℭ of the previous subsections
by ℭ 𝑗 , one of a set of 𝜈(< +∞) open cones forming an open covering of
R𝑛−1 \ {0}, every ℭ 𝑗 satisfying the conditions
required from ℭ above. Í We then
select cutoff functions 𝜙 𝑗 ∈ Cc∞ ℭ 𝑗 ∩S𝑛−2 , 𝑗 = 1, ..., 𝜈, such that 𝜈𝑗=1 𝜙 𝑗 ≡ 1,
and replace 𝜒 = 𝜙𝜓𝜌 by 𝜒 𝑗 = 𝜙 𝑗 𝜓𝜌 . The construction carried out in the
preceding subsections, now under the hypothesis that (𝚿) holds in 𝑈 × ℭ 𝑗 for
every 𝑗, leads us to 𝜈 integral operators 𝑨 𝑗 : 𝐿 c1 (𝑈◦ ) −→ D ′ (𝑈) analogous to
(26.3.28):

𝑨 𝑗 𝑓 (𝑥, 𝑡) (26.3.37)
∫ ∫ 𝑡
1
= 𝑛−1
e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 𝑗 (𝜉) d𝑠d𝜉d𝑦
(2𝜋) 𝔄𝑗 𝑡 ◦ −𝑟 ◦
∫ ∫ 𝑡◦
1
− e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 𝑗 (𝜉) d𝑠d𝜉d𝑦
(2𝜋) 𝑛−1 𝔄 +𝑗 𝑡
∫ ∫ 𝑡
1
+ e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑠, 𝜉 ) 𝑎 (𝑥, 𝑡, 𝑦, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) 𝜒 𝑗 (𝜉) d𝑠d𝜉d𝑦,
(2𝜋) 𝑛−1 𝔄 −𝑗 𝑡◦

where 𝔄 𝑗 , 𝔄 +𝑗 , 𝔄 −𝑗 , are the analogues of 𝔄, 𝔄 + , 𝔄 − , for ℭ 𝑗 replacing ℭ (cf.

Proposition 26.3.5). The difficulty arises from the fact that the reduction of 𝑃
to the form (26.3.1) is strictly microlocal and, in general, so is the choice of
the space-time coordinates. In each conic set 𝑈 × ℭ 𝑗 we might be forced to use
different space-time coordinates. This is clearly visible if 𝑃 = D2𝑥1 − D2𝑥2 (in
R2 ). Then how do we patch together 𝑨 𝑗 𝑓 and 𝑨 𝑘 𝑓 , 𝑗 ≠ 𝑘?
(3) The difficulties indicated in the preceding paragraph would be resolved if we
knew that 𝑓 ∈ 𝐿 c2 (𝑈) =⇒ 𝑨 𝑗 𝑓 ∈ 𝐿 2 (𝑈) (assuming that 𝑃 is a first-order
operator) – as was done in Ch. 23. The rough estimates in Subsection 26.3.7
tell us that 𝑨 𝑗 𝑓 ∈ 𝐻 −𝑁 (𝑈), 𝑁 large. To reduce 𝑁 to zero, assuming that it is
possible, would require a finer harmonic analysis of operators (26.3.28). If this
were achieved it would prove that 𝐿 c2 (𝑈) ⊂ 𝑃𝐿 2 (𝑈), a property not valid if 𝑃
were a generic C ∞ pseudodifferential operator in a domain Ω in R𝑛 , 𝑛 ≥ 3.
(4) One may also want to explore extension in the opposite direction, replacing 𝑓 in
Theorem 26.3.17 by an analytic functional. This cannot be done directly in the
integrals (26.3.28), due to the fragmentation of the domain of integration 𝔄 ∪
𝔄 + ∪ 𝔄 − . A more reasonable approach could be to replace 𝑓 by a holomorphic
function in a wedge, restricted to a level Im 𝑧 = const., and investigate whether
𝑨 𝑗 𝑓 could be of the same type and have a boundary value.
Chapter 27
Pseudodifferential Complexes in Tube Structures

In this final chapter we introduce pseudodifferential complexes of principal type

and describe relatively simple normal forms of the type used in dealing with single
operators in Chapters 23 and 26. The remainder of the chapter is devoted to a special
case of pseudodifferential complexes of principal type, those defined by commuting,
classical, analytic, pseudodifferential operators 𝑃 𝑗 = D𝑡 𝑗 − 𝑖𝑄 𝑗 (𝑡, D 𝑥 ) of order 1
in R𝑛−𝜈
𝑥 × Ω ( 𝑗 = 1, ..., 𝜈 ≤ 𝑛 − 1, Ω a domain in R𝑡𝜈 ) invariant under translations
in R𝑛−𝜈
𝑥 . This type of system is related to the tube integrable structures familiar
in several complex variables theory, and in more general contexts. We extend the
tube appellative to the pseudodifferential case. Our main goal is to prove the local
Poincaré Lemma modulo functions and differential forms that are analytic in the
space variables 𝑥 𝑘 , in the pseudodifferential complex – in all degrees simultaneously
(the latter is a weakness, not a strength, of the theorem). All the proving is done
by hand, with essentially sole recourse to (many) results established earlier in the
book and with obvious links to our handling of the sufficiency of Condition (𝚿) for
microlocal solvability in distributions, in the last section of the preceding chapter.
We thought that a fairly “concrete” example, in a sense a model, of the general
set-up described in Section 27.1 could be appreciated by some readers and perhaps
provide an incentive to investigate further the vastly unknown territory of differential
complexes and over-(or under-)determined systems of PDEs. Differential operators
of the tube type have been used as models from the beginning of the general theory
of PDEs in the XXth century, as exemplified by the class of operators first studied
in [Grushin, 1970], later the Mizohata vector fields (Example 23.1.19), and also by
the pseudodifferential operators (26.1.1) in the first section of the preceding chapter,
introduced to justify the recourse to Condition (𝚿).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1159
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3_27
1160 27 Pseudodifferential Complexes in Tube Structures

27.1 Pseudodifferential Complexes of Principal Type

27.1.1 Involutive Systems of Pseudodifferential Operators of Principal

Type. Microlocal normal forms

In this section we take a look at an involutive system of classical analytic pseudodif-

ferential operators of principal type, 𝑷 = (𝑃1 , ..., 𝑃 𝜈 ) (𝜈 ≥ 2), in a domain Ω in R𝑛 .
The operators 𝑃 𝑗 shall all have the same order which it is convenient to take equal to
1 since first-order pseudodifferential operators form a Lie algebra with respect to the
2
commutation bracket. Here involutive means that there exist 𝜈 (𝜈−1) 2 classical analytic
pseudodifferential operators of order zero in Ω, 𝐶 𝑗,𝑘,ℓ (𝑥, D), 1 ≤ 𝑗 < 𝑘 ≤ 𝜈 ≤ 𝑛,
1 ≤ ℓ ≤ 𝜈, such that
𝜈
∑︁
𝑃 𝑗 , 𝑃𝑘 = 𝐶 𝑗,𝑘,ℓ 𝑃ℓ . (27.1.1)
ℓ=1

We denote by 𝑝 𝑗 (𝑥, 𝜉) the principal symbol of 𝑃 𝑗 , 𝑐 𝑗,𝑘,ℓ that of 𝐶 𝑗,𝑘,ℓ . Condition

(27.1.1) implies directly that {𝑝 1 , ..., 𝑝 𝜈 } form an involutive system of functions in
Ω× (R𝑛 \ {0}) 𝑇 ∗ Ω\0 in the sense of Definition 13.4.2: the Poisson brackets verify
𝜈
∑︁

𝑝 𝑗 , 𝑝𝑘 = 𝑐 𝑗,𝑘,ℓ 𝑝 ℓ (27.1.2)
ℓ=1

everywhere in Ω × (R𝑛 \ {0}). We will use the notation

Char 𝑷 = (𝑥, 𝜉) ∈ 𝑇 ∗ Ω\0; 𝑝 𝑗 (𝑥, 𝜉) = 0, 𝑗 = 1, ..., 𝜈 .

(27.1.3)

Definition 27.1.1 (cf. Definition 13.4.1) We say that the system 𝑷 is of principal
type at a point (𝑥 ◦ , 𝜉 ◦ ) ∈ Char 𝑷 if d 𝜉 𝑝 1 ∧ · · · ∧ d 𝜉 𝑝 𝜈 ≠ 0 at every point of a conic
neighborhood of (𝑥 ◦ , 𝜉 ◦ ) in Ω × (R𝑛 \ {0}).

We select a neighborhood 𝑈 C of 𝑥 ◦ in C𝑛 and 𝜀 > 0 such that every 𝑝 𝑗 extends

holomorphically to

𝜉◦

C C 𝑛 𝜁
U = 𝑈 × 𝜁 ∈ C ; 𝜁 ≠ 0, − <𝜀 , (27.1.4)
|𝜁 | |𝜉 ◦ |

and d 𝜁 𝑝 1 ∧ · · · ∧ d 𝜁 𝑝 𝜈 ≠ 0 at every point of U C . We set U = U C ∩ (Ω × R𝑛 );

U ∩ Char 𝑷 is an analytic subvariety of U. The complex characteristic set of 𝑷 in
UC,
Σ 𝑷, U C = (𝑧, 𝜁) ∈ U C ; 𝑝 𝑗 (𝑧, 𝜁) = 0, 𝑗 = 1, ..., 𝜈 , (27.1.5)

is a complex-analytic submanifold of U C of codimension 𝜈, i.e., dimC Σ 𝑷, U C =
2𝑛 − 𝜈.
27.1 Pseudodifferential Complexes of Principal Type 1161

We denote by 𝐻 𝑝 𝑗 the Hamiltonian vector field of 𝑝 𝑗 (𝑧, 𝜁) for the complex

symplectic structure induced by C𝑛𝑧 × C𝑛𝜁 (Definition 13.3.4); each 𝐻 𝑝 𝑗 is a vector
field of type (1, 0):
𝑛
∑︁ 𝜕𝑝𝑗 𝜕 𝜕𝑝𝑗 𝜕
𝐻𝑝𝑗 = (𝑧, 𝜁) − (𝑧, 𝜁) . (27.1.6)
𝑘=1
𝜕𝜁 𝑘 𝜕𝑧 𝑘 𝜕𝑧 𝑘 𝜕𝜁 𝑘

The 𝐻 𝑝 𝑗 are linearly independent at every point of U C and tangent to Σ 𝑷, U C
at every one of its points (𝑧, 𝜁) where they span the symplectic orthogonal
(1,0)
of 𝑇(𝑧,𝜁 )
Σ 𝑷, U C . By the complex Frobenius Theorem (Theorem 12.2.5) the

𝐻 𝑝 𝑗 define a complex-analytic
foliation of Σ 𝑷, U C . The integral manifolds of
𝐻 𝑷 = 𝐻 𝑝1 , ..., 𝐻 𝑝𝜈 have complex dimension 𝜈; they are the complex null bichar-
acteristic leaves (or simply complex null bicharacteristics)
of the system 𝑷 in U C .
Every complex-analytic submanifold of Σ 𝑷, U of (complex) dimension 2 (𝑛 − 𝜈)
C

transversal at each one of its points to theintegral manifold of 𝐻 𝑷 through that point
is a symplectic submanifold of Σ 𝑷, U C .
Possibly after contracting 𝑈 C about 𝑥 ◦ and decreasing 𝜀 we apply (13.4.6):
𝜈
∑︁
𝑝 𝑗 (𝑧, 𝜁) = 𝐸 𝑗,𝑘 (𝑧, 𝜁) (𝜁 𝑛−𝜈+𝑘 + 𝑞 𝑘 (𝑧, 𝜁 ′)) , 𝑗 = 1, ..., 𝜈,
𝑘=1

with 𝜁 ′ = (𝜁1 , ..., 𝜁 𝑛−𝜈

) if 𝜈 < 𝑛, 𝐸 𝑗,𝑘 (𝑧, 𝜁) ∈ O U C homogeneous of degree
C

zero, det 𝐸 𝑗,𝑘 (𝑧, 𝜁) 1≤ 𝑗,𝑘 ≤𝜈 ≠ 0 whatever (𝑧, 𝜁) ∈ U ; every 𝑞 𝑘 ∈ O U C is
homogeneous of degree 1.

Proposition 27.1.2 If diam 𝑈 C and 𝜀 are sufficiently small then we have, in U C ,

𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝜁 𝑛−𝜈+ 𝑗 ′ + 𝑞 𝑗 ′ (𝑧, 𝜁 ′) ≡ 0, 1 ≤ 𝑗 < 𝑗 ′ ≤ 𝜈.

(27.1.7)

Proof If 𝐹 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝜈 is the inverse of the matrix 𝐸 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝜈 we have, for
𝑗 = 1, ..., 𝜈,
∑︁𝜈
𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) = 𝐹 𝑗,𝑘 (𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁) ,
𝑘=1

whence
𝜈
∑︁
𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝑝 𝑗 ′ (𝑧, 𝜁) =

𝐹 𝑗,𝑘 (𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁) , 𝑝 𝑗 ′ (𝑧, 𝜁) .
𝑘=1

The right-hand side is a linear combination of 𝑝 1 (𝑧, 𝜁) , ..., 𝑝 𝜈 (𝑧, 𝜁), by (27.1.2). It
follows that the same is true of the right-hand side in the equality
1162 27 Pseudodifferential Complexes in Tube Structures

𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝜁 𝑛−𝜈+ 𝑗 ′ + 𝑞 𝑗 ′ (𝑧, 𝜁 ′)

𝜈
∑︁
𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝐹 𝑗 ′ ,𝑘 (𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁)

=
𝑘=1
𝜈
∑︁
𝐹 𝑗 ′ ,𝑘 (𝑧, 𝜁) 𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝑝 𝑘 (𝑧, 𝜁) ,

+
𝑘=1

i.e.,
𝜈
∑︁
𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝜁 𝑛−𝜈+ 𝑗 ′ + 𝑞 𝑗 ′ (𝑧, 𝜁 ′) = 𝐺 𝑗,𝑘 (𝑧, 𝜁) (𝜁 𝑛−𝜈+𝑘 + 𝑞 𝑘 (𝑧, 𝜁 ′)) .

𝑘=1
(27.1.8)
Since the left-hand side of (27.1.8) is independent of 𝜁 ′′ = (𝜁 𝑛−𝜈+1 , ..., 𝜁 𝑛 ), (27.1.8)
requires 𝐺 𝑗,𝑘 (𝑧, 𝜁) = 0 for all (𝑧, 𝜁) ∈ U C . □
We may as well, and shall, assume that

𝑝 𝑗 (𝑧, 𝜁) = 𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ′) , 𝑗 = 1, ..., 𝜈, (27.1.9)

for (𝑧, 𝜁) ∈ U C . By Proposition 27.1.2 we now have 𝑝 𝑗 , 𝑝 𝑘 ≡ 0 in U C for all
( 𝑗, 𝑘), 1 ≤ 𝑗 < 𝑘 ≤ 𝜈. An important point to keep in mind is that no change of
variables has been needed beyond a relabeling of the 𝜁 𝑛−𝜈+ 𝑗 (or, equivalently, of the
𝑧 𝑗 ), only invertible linear substitutions of the 𝑝 𝑗 with coefficients in O U C . Also,
◦ ◦ , 𝜉 ′◦ ), 𝑗 = 1, ..., 𝜈, with 𝜉 ′◦ = 𝜉 ◦ , ..., 𝜉 ◦

𝜉 𝑛−𝜈+ 𝑗 = 𝑞 𝑗 (𝑥 1 𝑛−𝜈 .
The involution hypothesis (27.1.1) has significant implications well beyond
(27.1.2), for the homogeneous terms of every degree in the total symbols
𝜈
∑︁ ∞
∑︁
𝑃 𝑗 (𝑧, 𝜁) = 𝐸 𝑗,𝑘 (𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁) + 𝑝 𝑗,ℓ (𝑧, 𝜁)
𝑘=1 ℓ=0

[𝑝 𝑗,ℓ (𝑧𝜁) = 𝜆−ℓ 𝑝 𝑗,ℓ (𝑧, 𝜁), 𝜆 > 0]. Let 𝐹 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝜈 be the inverse of the matrix


𝐸 = 𝐸 𝑗,𝑘 1≤ 𝑗,𝑘 ≤𝜈 ; the entries 𝐹 𝑗,𝑘 (𝑧, 𝜁) ∈ O U C are homogeneous of degree
zero; we can replace the total symbols 𝑃 𝑗 (𝑧, 𝜁) by
∞
∑︁
𝑃♭𝑗 (𝑧, 𝜁) = 𝑝 𝑗 (𝑧, 𝜁) + 𝑝 (0)
𝑗,ℓ
(𝑧, 𝜁) ,
ℓ=0

where
𝜈
∑︁
𝑝 (0)
𝑗,ℓ
(𝑧, 𝜁) = 𝐹 𝑗,𝑘 (𝑧, 𝜁) 𝑝 𝑘,ℓ (𝑧, 𝜁) .
𝑘=1

The system 𝑃1♭ (𝑧, 𝜁) , ..., 𝑃♭𝜈 (𝑧, 𝜁) is involutive since the Poisson brackets are
independent of 𝜁 ′′ = (𝜁 𝑛−𝜈+1 , ..., 𝜁 𝑛 ). We apply the Weierstrass Division Theorem
14.3.5, for each 𝑗 = 1, ..., 𝜈,:
27.1 Pseudodifferential Complexes of Principal Type 1163
𝜈
∑︁
𝑝 (0)
𝑗,0
(𝑧, 𝜁) = 𝑔 (0)
𝑗,𝑘
(𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁) + 𝑞 𝑗,0 (𝑧, 𝜁 ′)
𝑘=1

where 𝑔 (0) (0)

𝑗,𝑘 and 𝑞 𝑗,0 are holomorphic in U , homogeneous of degree −1 and 0
C

respectively. We derive
𝜈
∑︁ ∞
∑︁
𝑃♭𝑗 (𝑧, 𝜁) = 𝛿 𝑗,𝑘 + 𝑔 (0)
𝑗,𝑘
(𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁) + 𝑞 𝑗,0 (𝑧, 𝜁 ′
) + 𝑝 (0)
𝑗,ℓ
(𝑧, 𝜁 ′) .
𝑘=1 ℓ=1
(27.1.10)
The matrix with entries 𝛿 𝑗,𝑘 + 𝑔 (0)
𝑗,𝑘
(𝑧, 𝜁) is invertible in the algebra (with respect to
ordinary multiplication) of classical symbols of degree ≤ 0. Letting its inverse act
on both sides of (27.1.10) yields the total symbols
∞
∑︁
𝑃 (1) ′
𝑗 (𝑧, 𝜁) = 𝑝 𝑗 (𝑧, 𝜁) + 𝑞 𝑗,0 (𝑧, 𝜁 ) + 𝑝 (1)
𝑗,ℓ
(𝑧, 𝜁) , (27.1.11)
ℓ=1

where 𝑝 (1)

𝑗,ℓ ∈ O U
C is homogeneous of degree −ℓ. We have

n o
𝑃 (1)
𝑗 (𝑧, 𝜁) , 𝑃 (1)
𝑗 ′ (𝑧, 𝜁) ≡ 0, 1 ≤ 𝑗 < 𝑗 ′ ≤ 𝜈,

since the Poisson brackets are independent of 𝜁 ′′ = (𝜁 𝑛−𝜈+1 , ..., 𝜁 𝑛 ). We repeat the
procedure: we apply the Weierstrass Division Theorem to get
𝜈
∑︁
𝑝 (1) 𝑔 (1) (𝑧, 𝜁) 𝑝 𝑘 (𝑧, 𝜁) + 𝑞 𝑘,0 (𝑧, 𝜁 ′) + 𝑞 𝑗,1 (𝑧, 𝜁 ′) ,

𝑗,1
(𝑧, 𝜁) = 𝑗,𝑘
𝑘=1

where 𝑔 (1)
𝑗,𝑘 and 𝑞 𝑗,1 are holomorphic in U , homogeneous of degree −2 and −1
C

respectively. We derive
𝜈
∑︁
𝑃 (1) 𝛿 𝑗,𝑘 + 𝑔 (1) 𝑝 𝑘 (𝑧, 𝜁) + 𝑞 𝑘,0 (𝑧, 𝜁 ′)

𝑗 (𝑧, 𝜁) = 𝑗,𝑘
(𝑧, 𝜁) (27.1.12)
𝑘=1
∞
∑︁
+ 𝑞 𝑗,1 (𝑧, 𝜁 ′) + 𝑝 (1)
𝑗,ℓ
(𝑧, 𝜁) .
ℓ=2

The matrix with entries 𝛿 𝑗,𝑘 + 𝑔 (1)

𝑗,𝑘
(𝑧, 𝜁) is invertible in the algebra of classical
symbols of degree ≤ 0. Letting its inverse act on both sides of (27.1.12) yields the
total symbol
∞
∑︁
𝑃 (2) (0) ′ (1) ′
𝑗 (𝑧, 𝜁) = 𝑝 𝑗 (𝑧, 𝜁) + 𝑞 𝑗,0 (𝑧, 𝜁 ) + 𝑞 𝑗,1 (𝑧, 𝜁 ) + 𝑝 (2)
𝑗,ℓ
(𝑧, 𝜁)
ℓ=2

with 𝑝 (2)

𝑗,ℓ ∈ O U
C homogeneous of degree −ℓ ≤ −2. Thus
1164 27 Pseudodifferential Complexes in Tube Structures

𝑃 (2) ′ (0) ′ (1) ′

𝑗 (𝑧, 𝜁) = 𝜁 𝑛−𝜈+ 𝑗 + 𝑞 𝑗 (𝑧, 𝜁 ) + 𝑞 𝑗,0 (𝑧, 𝜁 ) + 𝑞 𝑗,1 (𝑧, 𝜁 )
+ terms of homogeneity degree ≤ −2,
n o
and 𝑃 (2) (2) ′
𝑗 (𝑧, 𝜁) , 𝑃 𝑗 ′ (𝑧, 𝜁) ≡ 0, 1 ≤ 𝑗 < 𝑗 ≤ 𝜈. By repeating the procedure ad
infinitum we end up with total formal symbols

𝑃 (∞)
𝑗 (𝑧, 𝜁) = 𝜁 𝑛−𝜈+ 𝑗 + 𝑄 𝑗 (𝑧, 𝜁 ′) , 𝑗 = 1, ..., 𝜈, (27.1.13)

with each
∞
∑︁
𝑄 𝑗 (𝑧, 𝜁 ′) = 𝑞 𝑗 (𝑧, 𝜁 ′) + 𝑞 𝑗,ℓ (𝑧, 𝜁 ′)
ℓ=0

a classical symbol of order 1. Of course, we must still prove that 𝑄 𝑗 is an analytic

classical symbol, i.e., that the homogeneous terms in 𝑄 𝑗 (𝑧, 𝜁 ′) satisfy the appropriate
version of Condition (FA), Definition 19.1.1. This is equivalent to saying that the
infinite product !
Ö∞
(ℓ)
𝛿 𝑗,𝑘 + 𝑔 𝑗,𝑘 (𝑧, 𝜁)
ℓ=0 1≤ 𝑗,𝑘 ≤𝜈

is invertible in the algebra of 𝜈 × 𝜈 matrices whose entries are analytic classical

symbols of order zero. This is routine but tedious work; we will not pursue the
matter here.

Remark 27.1.3 In the discussion of the preceding paragraph multiplication and

division were the ordinary ones but they could as well have been intended in the
sense of the composition law # [see (19.1.12)] since, in effect, each step involved
only the principal symbols of operators of degrees ↘ −∞.

27.1.2 The pseudodifferential complex defined by 𝑷

In this subsection it is convenient to change the notation of the preceding subsection

and, in a manner of speaking, distinguish the “space variables” 𝑥1 , ..., 𝑥 𝑛−𝜈 from
the “time variables” 𝑡1 , ..., 𝑡 𝜈 (the latter replacing 𝑥 𝑛−𝜈+1 , ..., 𝑥 𝑛 ) with corresponding
covariables 𝜉1 , ..., 𝜉 𝑛−𝜈 , 𝜏1 , ..., 𝜏𝜈 ; complex extensions of 𝑥 𝑘 , 𝜉 𝑘 , will be 𝑧 𝑘 , 𝜁 𝑘 (we
keep 𝑡 𝑗 and 𝜏 𝑗 even when complex-valued); 𝑥 = (𝑥1 , ..., 𝑥 𝑛−𝜈 ) will vary in a domain
Ω in R𝑛 , 𝑡 = (𝑡 1 , ..., 𝑡 𝜈 ) will vary in a multi-interval
n o
𝔔𝑟(𝜈)
◦
(𝑡 ◦ ) = 𝑡 ∈ R𝑛 ; 𝑡 𝑗 − 𝑡 ◦𝑗 < 𝑟 ◦ , 𝑗 = 1, ..., 𝜈 , 𝑟 ◦ > 0. (27.1.14)

The procedure of the preceding subsection transforms the system 𝑷 into the system
of classical, analytic pseudodifferential operators in Ω × 𝔔𝑟(𝜈)
◦
(𝑡 ◦ ) [cf. (27.1.11)]
27.1 Pseudodifferential Complexes of Principal Type 1165

𝑃 𝑗 (𝑥, 𝑡, D 𝑥 , D𝑡 ) = D𝑡 𝑗 + 𝑄 𝑗 (𝑥, 𝑡, D 𝑥 ) , (27.1.15)

∞
∑︁
𝑄 𝑗 (𝑥, 𝑡, D 𝑥 ) = 𝑞 𝑗 (𝑥, 𝑡, D 𝑥 ) + 𝑞 𝑗,ℓ (𝑥, 𝑡, D 𝑥 ) .
ℓ=0
√
As usual D𝑡 𝑗 = − −1𝜕/𝜕𝑡 𝑗 and likewise for D 𝑥𝑘 ( 𝑗 = 1, ..., 𝜈, 𝑘 = 1, ..., 𝑛 − 𝜈). We
shall also use the notation

𝑸 (𝑥, 𝑡, D 𝑥 ) = 𝑷 (𝑥, 𝑡, D 𝑥 , D𝑡 ) − 𝑫 𝑡 = (𝑄 1 (𝑥, 𝑡, D 𝑥 ) , ..., 𝑄 𝜈 (𝑥, 𝑡, D 𝑥 )) . (27.1.16)

The symbols 𝑞 𝑗,ℓ (𝑥, 𝑡, 𝜉) are as described in the preceding subsection; 𝑞 𝑗 (𝑥, 𝑡, 𝜉)
is homogeneous of degree 1; it is the principal symbol of 𝑄 𝑗 ; the principal symbol
of 𝑃 𝑗 is 𝑝 𝑗 (𝑥, 𝑡, 𝜏, 𝜉) = 𝜏 𝑗 + 𝑞 𝑗 (𝑥, 𝑡, 𝜉). As pointed out at the end of the preceding
subsection the involution property of the system

𝑷 (𝑥, 𝑡, D 𝑥 , D𝑡 ) = (𝑃1 (𝑥, 𝑡, D 𝑥 , D𝑡 ) , ..., 𝑃 𝜈 (𝑥, 𝑡, D 𝑥 , D𝑡 ))

translates here in the commutation 𝑃 𝑗 (𝑥, 𝑡, D 𝑥 , D𝑡 ) , 𝑃 𝑗 ′ (𝑥, 𝑡, D 𝑥 , D𝑡 ) = 0; and,
correspondingly, in the vanishing of the Poisson brackets of the symbols. This allow
us to define the pseudodifferential complex defined by the system 𝑷.
To do this we introduce the 𝑘-forms
∑︁
𝑓 (𝑥, 𝑡, d𝑡) = 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 (27.1.17)
|𝐼 |=𝑘

where 𝐼 = (𝑖 1 , ..., 𝑖 𝑘 ), 1 ≤ 𝑖1 < · · · < 𝑖 𝑘 ≤ 𝜈, |𝐼 | = 𝑖1 +· · ·+𝑖 𝑘 ; the coefficients 𝑓 𝐼 (𝑥, 𝑡)

may belong to any given space 𝑬 of functions or distributions in (or microfunctions
over) Ω × 𝔔𝑟(𝜈) ◦
(𝑡 ◦ ) on which pseudodifferential operators (27.1.15) do act. The
vector space of forms (27.1.17) shall be denoted by Λ 𝑘 𝑬; we define Λ0 𝑬 = 𝑬. We
define as follows the action of 𝑷 on Λ 𝑘−1 𝑬 (𝑘 ≥ 1):
𝜈
∑︁ ∑︁
d𝑷(𝑘) 𝑓 (𝑥, 𝑡, d𝑡) = 𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝑗 ∧ d𝑡 𝐼 . (27.1.18)
|𝐼 |=𝑘−1 𝑗=1

The commutation of the pseudodifferential operators 𝑃 𝑗 implies d𝑷(𝑘+1) ◦ d𝑷(𝑘) = 0.

As commonly done we omit the superscript 𝑘 and write d2𝑷 = 0. This defines the
“differential” complex
d𝑷 d𝑷 d𝑷 d𝑷 d𝑷
0 −→ 𝑬 −→ Λ1 𝑬 −→ · · · −→ Λ𝜈−1 𝑬 −→ Λ𝜈 𝑬 −→ 0. (27.1.19)

Take for instance 𝑬 = D′( 𝑥,𝑡 , 𝜉 , 𝜏) , the space of germs of microdistributions at a

point (𝑥, 𝑡, 𝜉, 𝜏) ∈ U, a conic neighborhood of ((𝑥 ◦ , 𝑡 ◦ ) , (𝜉 ◦ , 𝜏 ◦ )) (with 𝜏 ◦𝑗 = −
𝑞 𝑗 (𝑥 ◦ , 𝑡 ◦ , 𝜉 ◦ ), 𝑗 = 1, ..., 𝜈); thus the “pseudodifferential complex” is
1166 27 Pseudodifferential Complexes in Tube Structures

d𝑷 d𝑷
0 −→ D′( 𝑥,𝑡 , 𝜉 , 𝜏) −→ Λ1 D′( 𝑥,𝑡 , 𝜉 , 𝜏) −→ · · · (27.1.20)
d𝑷 d𝑷 d𝑷
−→ Λ𝜈−1 D′( 𝑥,𝑡 , 𝜉 , 𝜏) −→ Λ𝜈 D′( 𝑥,𝑡 , 𝜉 , 𝜏) −→ 0. (27.1.21)

27.2 Tube Pseudodifferential Complexes

27.2.1 The tube system 𝑷 and the associated (pseudo)differential

complexes

In the remainder of this chapter we limit our attention to a special case of systems
of pseudodifferential operators (27.1.15): we shall assume that the operators 𝑄 𝑗 do
not depend on the space variables 𝑥 𝑗 . Thus we deal with the system of pairwise
commuting pseudodifferential operators

𝑃 𝑗 (𝑡, D 𝑥 , D𝑡 ) = D𝑡 𝑗 + 𝑄 𝑗 (𝑡, D 𝑥 ) , (27.2.1)

∞
∑︁
𝑄 𝑗 (𝑡, D 𝑥 ) = 𝑞 𝑗 (𝑡, D 𝑥 ) + 𝑞 𝑗,ℓ (𝑡, D 𝑥 ) .
ℓ=0

To define a differential complex based on 𝑷 = (𝑃1 , ..., 𝑃 𝜈 ) we introduce the 𝑝-forms

∑︁
𝑓 (𝑥, 𝑡, d𝑡) = 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 (27.2.2)
|𝐼 |= 𝑝

where 𝐼 = 𝑖 1 , ..., 𝑖 𝑝 , 1 ≤ 𝑖 1 < · · · < 𝑖 𝑝 ≤ 𝜈, |𝐼 | = 𝑖 1 + · · · + 𝑖 𝑝 . Throughout the
remainder of this chapter 𝑈 shall be a domain in R𝑛−𝜈 , 𝑉 ⊂ Ω a neighborhood of 0
in R𝜈 , ℭ an open cone in R𝑛−𝜈 \{0}. We introduce ℭ because many of the definitions
and results will be valid under the assumption that the symbols 𝑄 𝑗 (𝑡, 𝜉) are merely
defined in Ω × ℭ, in which case the framework will be microlocal.
Let 𝑬 (𝑈) be a locally convex topological vector space of functions or distri-
butions in 𝑈. Let C ∞ (𝑉; 𝑬 (𝑈; Λ 𝑝 𝑇 ∗ Ω)) denote the space of forms (27.2.2) with
coefficients 𝑓 𝐼 ∈ C ∞(𝑉; 𝑬 (𝑈)), meaning smooth maps 𝑉 −→ 𝑬 (𝑈); we define
C ∞ 𝑉; 𝑬 𝑈; Λ0𝑇 ∗ Ω = C ∞ (𝑉; 𝑬 (𝑈)). Consider the choices 𝑬 = C ∞ or 𝑬 = D ′
(the space of distributions). We would like to define, for 𝑝 ≥ 0,
𝜈
∑︁ ∑︁
(
d𝑷𝑝) 𝑓 (𝑥, 𝑡, d𝑡) = 𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝑗 ∧ d𝑡 𝐼 . (27.2.3)
|𝐼 |= 𝑝−1 𝑗=1

Unless it is a vector field the pseudodifferential operator 𝑞 𝑗 (𝑡, D 𝑥 ) is not local (only
pseudolocal) and for (27.2.3) to be admissible the functions 𝑥 ↦→ 𝑓 𝐼 (𝑥, 𝑡) must be
compactly supported; even then the support of 𝑥 ↦→ 𝑃 𝑗 𝑓 𝐼 may not be compact,
generally speaking. It is therefore natural to introduce cut-off functions 𝜒 ∈ Cc∞ (𝑈),
𝜒 ≡ 1 in an open set 𝑈 ′ ⊂⊂ 𝑈, replace 𝑓 𝐼 (resp., 𝑃 𝑗 𝑓 𝐼 ) by 𝜒 𝑓 𝐼 (resp., 𝜒𝑃 𝑗 𝑓 𝐼 ) and
27.2 Tube Pseudodifferential Complexes 1167

(
then restrict consideration of d𝑷𝑝) 𝜒 (𝑥) 𝑓 (𝑥, 𝑡, d𝑡) to 𝑈 ′ × 𝑉. This procedure does
not produce a differential complex, since, in general,

( (
d𝑷𝑝+1) 𝜒 (𝑥) d𝑷𝑝) 𝜒 (𝑥) 𝑓 (𝑥, 𝑡, d𝑡) . 0 in 𝑈 ′ × 𝑉.

( (
As a matter of fact, d𝑷𝑝) 𝜒 (𝑥) 𝑓 (𝑥, 𝑡, d𝑡) . d𝑷𝑝) 𝑓 (𝑥, 𝑡, d𝑡) in 𝑈 ′ × 𝑉. However,

( (
Theorem 17.2.13 implies that the coefficients of d𝑷𝑝+1) 𝜒 (𝑥) d𝑷𝑝) 𝜒 (𝑥) 𝑓 (𝑥, 𝑡, d𝑡)
( (
as well as those of d𝑷𝑝) 𝜒 (𝑥) 𝑓 (𝑥, 𝑡, d𝑡) −d𝑷𝑝) 𝑓 (𝑥, 𝑡, d𝑡) are analytic with respect to 𝑥
′
in 𝑈 . It follows that the natural choices, in defining differential complexes (27.1.19)
for tube structures, are

𝑬 (𝑈) = C ∞ (𝑈) /C ∞ (𝑈) ∩ C 𝜔 (𝑈 ′) (27.2.4)

or the space of singularity distributions

𝑬 (𝑈) = D ′ (𝑈) /D ′ (𝑈) ∩ C 𝜔 (𝑈 ′) , (27.2.5)

the denominators meaning that restrictions to 𝑈 ′ are analytic functions. The sequence
of maps
d𝑷 𝜒 ( 𝑥)
d𝑷 𝜒 ( 𝑥)
0 −→ C ∞ (𝑉; 𝑬 (𝑈)) −→ C ∞ 𝑉; 𝑬 𝑈; Λ1𝑇 ∗ Ω −→ · · · (27.2.6)
d𝑷 𝜒 ( 𝑥) ∞
d𝑷 𝜒 ( 𝑥) d𝑷 𝜒 ( 𝑥)
−→ C 𝑉; 𝑬 𝑈; Λ𝜈−1𝑇 ∗ Ω −→ C ∞ (𝑉; 𝑬 (𝑈; Λ𝜈 𝑇 ∗ Ω)) −→ 0

defines, after restriction to 𝑈 ′ × 𝑉, a sort of differential complex; this complex is

independent of the choice of cut-off function 𝜒.

Remark 27.2.1 When the operators 𝑃 𝑗 are vector fields then (27.2.3) defines the
natural differential complex, namely (27.2.6) with 𝑬 (𝑈) = C ∞ (𝑈) or 𝑬 (𝑈) =
D ′ (𝑈) and without cutoffs 𝜒 nor restriction to 𝑈 ′.

Remark 27.2.2 Aside from the insertion of cutoff functions there is the question of
how to interpret the operators 𝑃 𝑗 . As defined in (27.2.1) they are formal classical
series. If we insist on having true operators we may use finite realizations (Definition
19.1.7) of the formal series. Different finite realizations will introduce errors that
are analytic regularizing, one more reason for introducing quotients like (27.2.4) or
(27.2.5). Until specified otherwise we deal with the 𝑃 𝑗 as formal series, many of the
statements applying to each term.

Remark 27.2.3 In (27.2.6) we have chosen the regularity with respect to 𝑡 to be C ∞ .

Nothing in the forthcoming reasoning prevents us from selecting subclasses of C ∞
such as C 𝜔 , Gevrey, etc. In most of the intermediary steps it will suffice to assume
the coefficients 𝑓 𝐼 in (27.2.2) to be of class C 1 with respect to 𝑡.
1168 27 Pseudodifferential Complexes in Tube Structures

A natural inclination is to let 𝑈 and 𝑈 ′ ⊂⊂ 𝑈 range over a basis of neighborhoods

of an arbitrary point 𝑥 ◦ in R𝑛−𝜈 and introduce the stalk E 𝑥 ◦ of germs at 𝑥 ◦ of
elements of 𝑬 (R𝑛−𝜈 ) properly topologized, and deal with 𝑝-forms (27.2.2) whose
coefficients belong to C ∞ (𝑉; E 𝑥 ◦ ) rending the cutoffs 𝜒 unnecessary. This produces
a true differential complex

d𝑷
d
𝑷
0 −→ C ∞ (𝑉; E 𝑥 ◦ ) −→ C ∞ 𝑉; E 𝑥 ◦ Λ1𝑇 ∗ Ω −→ · · · (27.2.7)
d𝑷 d𝑷 d𝑷

−→ C ∞ 𝑉; E 𝑥 ◦ Λ𝜈−1𝑇 ∗ Ω −→ C ∞ (𝑉; E 𝑥 ◦ (Λ𝜈 𝑇 ∗ Ω)) −→ 0.

We can go further and microlocalize the complex, ultimately taking E to be the

stalk of microdistributions at a point (𝑥 ◦ ,𝜉 ◦ ) ∈ 𝑆 ∗ R𝑛−𝜈−1 (the cosphere bundle of
R𝑛−𝜈−1 ); this can also be approached by taking E to be the space of continuous
sections, over a neighborhood U of (𝑥 ◦ ,𝜉 ◦ ) in 𝑆 ∗ R𝑛−𝜈−1 , of the sheaf of microdis-
tributions in R𝑛−𝜈 , in other words, dealing with the presheaf of microdistributions:
d𝑷
d
𝑷
0 −→ C ∞ (𝑉; D ′ (U)) −→ C ∞ 𝑉; D ′ U; Λ1𝑇 ∗ Ω −→ · · · (27.2.8)
d𝑷 d𝑷 d𝑷

−→ C ∞ 𝑉; D ′ U; Λ𝜈−1𝑇 ∗ Ω −→ C ∞ (𝑉; D ′ (U; Λ𝜈 𝑇 ∗ Ω)) −→ 0.

Contracting U about (𝑥 ◦ ,𝜉 ◦ ) leads to the microlocal differential complex

d d
𝑷 𝑷
0 −→ C ∞ 𝑉; D′micro ◦ ◦
(𝑥 , 𝜉 ) −→ C ∞
𝑉; D ′micro
◦ ◦
(𝑥 , 𝜉 ) Λ1 ∗
𝑇 Ω −→ · · · (27.2.9)
d𝑷 d𝑷 d𝑷

−→ C ∞ 𝑉; D′micro
( 𝑥◦ , 𝜉 ◦) Λ
𝜈−1 ∗
𝑇 Ω −→ C ∞ 𝑉; D′micro 𝜈 ∗
( 𝑥 ◦ , 𝜉 ◦ ) (Λ 𝑇 Ω) −→ 0.

Extension to microfunctions could be tried by replacing everywhere, in this

section, E ′ (R𝑛−𝜈 ) by O ′ (R𝑛−𝜈 ), i.e., the space of compactly supported distributions
by the space of analytic functionals in R𝑛−𝜈 . We do not pursue this matter further,
limiting our attention to distributions.

27.2.2 Preliminary reductions of the tube system 𝑷

We carry out a few simple reductions of the system of pseudodifferential

operators
(27.2.1). By equating to 0 the terms of the same order in 𝑃 𝑗 , 𝑃 𝑗 ′ = 0 we see that

𝑞 𝑗 (𝑡, D 𝑥 ) , 𝑞 𝑗 ′ (𝑡, D 𝑥 ) = 0 and that the operators
∞
∑︁
𝑅 𝑗 = D𝑡 𝑗 + 𝑞 𝑗,ℓ (𝑡, D 𝑥 )
ℓ=0
27.2 Tube Pseudodifferential Complexes 1169

also commute, implying Í∞that there is a classical symbol 𝐹 of order zero in Ω ×

(R𝑛−𝜈 \ {0}) such that ℓ=0 𝑞 𝑗,ℓ (𝑡, 𝜉) = −D𝑡 𝑗 𝐹 (𝑡, 𝜉) for every 𝑗. We can therefore
form the pseudodifferential operators of order zero exp (±𝐹 (𝑡, D 𝑥 )) such that

𝑅 𝑗 (𝑡, D 𝑥 ) exp 𝐹 (𝑡, D 𝑥 ) 𝑢 = (exp 𝐹 (𝑡, D 𝑥 )) D𝑡 𝑗 𝑢, 𝑗 = 1, ..., 𝜈,

for all 𝑢 ∈ E ′ (Ω×R𝑛−𝜈 ). Since 𝑞 𝑗 (𝑡, D 𝑥 ) and exp 𝐹 (𝑡, D 𝑥 ) commute we derive

exp (−𝐹 (𝑡, D 𝑥 )) 𝑃 𝑗 exp (𝐹 (𝑡, D 𝑥 )) = D𝑡 𝑗 + 𝑞 𝑗 (𝑡, D 𝑥 ) , 𝑗 = 1, ..., 𝜈.

Keep in mind that 𝑞 𝑗 (𝑡, 𝜉) is homogeneous of degree 1. The commutation of the

𝑃 𝑗 ’s implies

𝜕 𝜕
Re 𝑞 𝑗 ′ (𝑡, 𝜉) = Re 𝑞 𝑗 (𝑡, 𝜉) , 1 ≤ 𝑗 < 𝑗 ′ ≤ 𝜈.
𝜕𝑡 𝑗 𝜕𝑡 𝑗 ′

In turn this implies that there is a real-valued function 𝐹1 ∈ C 𝜔 (Ω × (R𝑛−𝜈 \ {0}))

homogeneous of degree 1 in Ω × (R𝑛−𝜈 \ {0}) such that Re 𝑞 𝑗 (𝑡, 𝜉) = 𝜕𝑡 𝑗 𝐹1 (𝑡, 𝜉)
for every 𝑗. This allows us to introduce the FIOs with real phase-function
exp (±𝑖𝐹1 (𝑡, D 𝑥 )); their action on E ′ (R𝑛−𝜈 ) is defined by the formula
∫
1
exp (±𝐹1 (𝑡, D 𝑥 )) 𝑢 (𝑥) = e𝑖 ( 𝑥· 𝜉 ±𝐹1 (𝑡 , 𝜉 )) b
𝑢 (𝜉) d𝜉.
(2𝜋) 𝑛−𝜈 R2𝑛−𝜈
It is immediate that

exp (𝑖𝐹1 (𝑡, D 𝑥 )) D𝑡 𝑗 + 𝑞 𝑗 (𝑡, D 𝑥 ) exp (−𝑖𝐹1 (𝑡, D 𝑥 )) = D𝑡 𝑗 + 𝑖 Im 𝑞 𝑗 (𝑡, D 𝑥 ) .

We may as well assume, from the start, that we are dealing with the reduced normal
forms
𝑃 𝑗 = D𝑡 𝑗 + 𝑖𝑞 𝑗 (𝑡, D 𝑥 ) , 𝑗 = 1, ..., 𝜈, (27.2.10)
with 𝑞 𝑗 (𝑡, 𝜉) ∈ C 𝜔 (Ω × (R𝑛−𝜈 \ {0})) real-valued, 𝑞 𝑗 (𝑡, 𝜆𝜉) = 𝜆𝑞 𝑗 (𝑡, 𝜉) if 𝜆 > 0
(cf. Remark 27.1.3).

27.2.3 First integrals of a system of partially Hamiltonian vector fields

Since 𝑃 𝑗 , 𝑃 𝑗 ′ = 0 we have

𝜕𝑞 𝑗 ′ 𝜕𝑞 𝑗
= , 1 ≤ 𝑗 < 𝑗 ′ ≤ 𝑛 − 𝜈. (27.2.11)
𝜕𝑡 𝑗 𝜕𝑡 𝑗 ′

We derive that there is a unique Φ (𝑡, 𝜉) ∈ C 𝜔 (Ω × (R𝑛−𝜈 \ {0})) such that

1170 27 Pseudodifferential Complexes in Tube Structures
𝜈
∑︁
𝑞 𝑗 (𝑡, 𝜉) d𝑡 𝑗 = −d𝑡 Φ, Φ (0, 𝜉) ≡ 0; (27.2.12)
𝑗=1

Φ is real-valued. We associate to the system 𝑷 the system of pairwise commuting

(partially Hamiltonian, since we disregard the 𝜕/𝜕𝜏 𝑗 ) vector fields
𝑛−𝜈
𝜕 ∑︁ 𝜕𝑞 𝑗 𝜕
𝐻𝑗 = +𝑖 (𝑡, 𝜉) , 𝑗 = 1, ..., 𝜈. (27.2.13)
𝜕𝑡 𝑗 𝑘=1
𝜕𝜉 𝑘 𝜕𝑥 𝑘

For each 𝑘 = 1, ..., 𝑛 − 𝜈, we define 𝑍 𝑘 (𝑥, 𝑡, 𝜉) = 𝑥 𝑘 + 𝑖 𝜕𝜕Φ 𝜉𝑘 (𝑡, 𝜉); thus 𝐻 𝑗 𝑍 𝑘 = 0

and 𝑍 𝑘 (𝑥, 0, 𝜉) = 𝑥 𝑘 for all 𝑗 = 1, ..., 𝜈, 𝑘 = 1, ..., 𝑛 − 𝜈. We shall use the notation
𝑍 = (𝑍1 , ..., 𝑍 𝑛−𝜈 ) and refer to the 𝑍 𝑘 as first integrals of the system of vector fields
𝑯 = (𝐻1 , ..., 𝐻 𝜈 ). Then the characteristic set of the system 𝑷 is

Char 𝑷 = {((𝑥, 𝑡) , (𝜉, 𝜏)) ∈ R𝑛−𝜈 × Ω × R𝑛 ; d𝑡 Φ (𝑡, 𝜉) = 0, 𝜉 ≠ 0, 𝜏 = 0} .

(27.2.14)
In the remainder of the section we shall assume that 𝜉 ↦→ 𝑞 𝑗 (𝑡, 𝜉) extends
holomorphically in a region |Im 𝜁 | ≤ 𝑐 |Re 𝜁 | (0 < 𝑐 ≪ 1); this allows us to assume

𝜕𝑞 𝑗 𝜕𝑞 𝑗
𝑞 𝑗 (𝑡, 𝜆𝜉) = 𝜆𝑞 𝑗 (𝑡, 𝜉) , (𝑡, 𝜆𝜉) = (𝑡, 𝜉) , (27.2.15)
𝜕𝜉 𝑘 𝜕𝜉 𝑘

for all 𝜆 ∈ C, |Im 𝜆| < 𝑐 Re 𝜆, (𝑡, 𝜉) ∈ Ω× (R𝑛−𝜈 \ {0}), 𝑘 = 1, ..., 𝑛−𝜈. It follows that
𝑍 (𝑥, 𝑡, 𝜉) and Φ (𝑡, 𝜉) also extend holomorphically in the region |Im 𝜁 | ≤ 𝑐 |Re 𝜁 |,
and are homogeneous [in the sense of (27.2.15)] of degree zero and 1 respectively.

27.3 Phase-function and Amplitude

27.3.1 Fourier-like phase function

We emphasize the fact that all functions of (𝑥, 𝑦, 𝑡, 𝑠, 𝜉) so far introduced in the
preceding subsections are well defined in the whole of R2(𝑛−𝜈) × Ω2 × (R𝑛−𝜈 \ {0}).
A constraint on the variables (𝑡, 𝑠) is introduced in the next statement, required by
the bounds on the extendibility of the holomorphic function 𝜁 ↦→ 𝑞 (𝑡, 𝜁). We define

𝜇◦ = sup max |Φ (𝑡, 𝜉) − Φ (𝑠, 𝜉)| . (27.3.1)
(𝑡 ,𝑠) ∈Ω2 𝜉 ∈S𝑛−𝜈−1
27.3 Phase-function and Amplitude 1171

Proposition 27.3.1 If 𝜇◦ is sufficiently small then the following function in

(R𝑛−𝜈 × Ω) 2 × (R𝑛−𝜈 \ {0}),

𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) = (𝑍 (𝑥, 𝑡, 𝜉) − 𝑍 (𝑦, 𝑠, 𝜉)) · 𝜉 (27.3.2)

= (𝑥 − 𝑦) · 𝜉 + 𝑖 (Φ (𝑡, 𝜉) − Φ (𝑠, 𝜉))

satisfies the eikonal equations

𝜕𝜑
(𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) + 𝑖𝑞 𝑗 (𝑡, 𝜕𝑥 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉)) = 0, (27.3.3)
𝜕𝑡 𝑗
𝜕𝜑
(𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) − 𝑖𝑞 𝑗 𝑠, −𝜕𝑦 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) = 0, (27.3.4)
𝜕𝑠 𝑗

and the Cauchy condition

𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉)| 𝑡=𝑠 = (𝑥 − 𝑦) · 𝜉.

Since 𝜕𝑥 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) ≡ 𝜉 the claim is a direct consequence of (27.2.12) and

the eikonal equations can be rewritten as

𝜕𝜑
(𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) + 𝑖𝑞 𝑗 (𝑡, 𝜉) = 0, (27.3.5)
𝜕𝑡 𝑗
𝜕𝜑
(𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) − 𝑖𝑞 𝑗 (𝑠, 𝜉) = 0, (27.3.6)
𝜕𝑠 𝑗

which, in turn, can be rewritten as

𝜕
(𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) − 𝑖Φ (𝑡, 𝜉)) = 0, (27.3.7)
𝜕𝑡 𝑗
𝜕
(𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) + 𝑖Φ (𝑠, 𝜉)) = 0, (27.3.8)
𝜕𝑠 𝑗

an evident consequence of (27.3.2).

The function 𝑖𝜑 (𝑥, 𝑡, 𝑥, 𝑠, 𝜉) will play an important role in the sequel [cf. (26.3.7)];
we give it a name.

Definition 27.3.2 By the exponent-weight of the phase-function 𝜑 we shall mean

the following (real-valued) C 𝜔 function in Ω2 × (R𝑛−𝜈 \ {0})

𝜋 𝜑 (𝑡, 𝑠, 𝜉) = − (Φ (𝑡, 𝜉) − Φ (𝑠, 𝜉)) .

We have 𝜋 𝜑 (𝑠, 𝑠, 𝜉) ≡ 0 and [cf. (27.2.12)]

𝜕𝜋 𝜑
(𝑡, 𝑠, 𝜉) = 𝑞 𝑗 (𝑡, 𝜉) , 𝑗 = 1.., 𝜈. (27.3.9)
𝜕𝑡 𝑗
1172 27 Pseudodifferential Complexes in Tube Structures

27.3.2 Amplitude

It is convenient to replace 𝜉 by 𝜆𝜉, 𝜉 ∈ S𝑛−𝜈−1 , 𝜆 > 0:

𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜆𝜉) = 𝜆𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) = 𝜆 (𝑍 (𝑥, 𝑡, 𝜉) − 𝑍 (𝑦, 𝑠, 𝜉)) · 𝜉.

We are going to determine a classical amplitude 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) of order zero,

satisfying the following equations:

D𝑡 𝑗 + 𝑖𝑞 𝑗 (𝑡, D 𝑥 ) e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) = 0, 𝑗 = 1, ..., 𝜈, (27.3.10)
𝑎 (𝑥 − 𝑦, 𝑠, 𝑠, 𝜉, 𝜆) = 1. (27.3.11)

Likewise, 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) will satisfy the following equations:

𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 )
D𝑠 𝑗 − 𝑖𝑞 𝑗 𝑠, −D 𝑦 e 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) = 0, 𝑗 = 1, ..., 𝜈.
(27.3.12)
We seek the solution of (27.3.10)–(27.3.11) as a formal series
∞
∑︁
𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) = 𝜆−𝑚 𝑎 𝑚 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) . (27.3.13)
𝑚=0

We have

𝑞 𝑗 (𝑡, D 𝑥 ) e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) = 𝑞 𝑗 (𝑡, D 𝑥 + 𝜆𝜉) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) .
(27.3.14)

Remark 27.3.3 Unless 𝑞 𝑗 (𝑡, D 𝑥 ) is a vector field for every 𝑗 = 1, ..., 𝜈, (27.3.13)
may not be a finite series in the powers of 𝜆−1 . When the 𝑞 𝑗 (𝑡, D 𝑥 ) are vector fields,
i.e., the 𝑞 𝑗 (𝑡, 𝜉) are linear functions of 𝜉, it is readily found that 𝑍 𝑘 = 𝑥 𝑘 + 𝑖Φ 𝑘 (𝑡),
𝑘 = 1, ..., 𝑛 − 𝜈, and 𝑎 ≡ 1.

The equations (27.3.10) become the system of transport equations

𝜕𝜑
𝑎 + D𝑡 𝑗 𝑎 = −𝑖𝑞 𝑗 (𝑡, D 𝑥 + 𝜆𝜉) 𝑎.
𝜕𝑡 𝑗

The eikonal equation (27.3.3) implies

𝜕𝑎
+ 𝑞 𝑗 (𝑡, 𝜉)𝑎 = 𝑞 𝑗 (𝑡, D 𝑥 + 𝜆𝜉) 𝑎.
𝜕𝑡 𝑗

We introduce the (formal) Taylor expansion of the right-hand side, whence

27.3 Phase-function and Amplitude 1173
+∞
∑︁ 𝜕𝑎 𝑚 ∑︁ 1 1−|𝛽 | 𝛽
𝜆−𝑚
𝛽
= 𝜆 𝜕 𝜉 𝑞 𝑗 (𝑡, 𝜉) D 𝑥 𝑎
𝑗=0
𝜕𝑡 𝑗 0≠𝛽 ∈Z 𝑛−𝜈 𝛽!
+
∞
∑︁ ∑︁ 1 1−|𝛽 |−𝑘 𝛽 𝛽
= 𝜆 𝜕 𝜉 𝑞 𝑗 (𝑡, 𝜉) D 𝑥 𝑎 𝑘
𝑘=0 0≠𝛽 ∈Z+𝑛−𝜈
𝛽!

and therefore
+∞ +∞
∑︁ 𝜕𝑎 𝑚 ∑︁ −𝑚 ∑︁ 1 𝛽
𝜆−𝑚
𝛽
= 𝜆 𝜕 𝜉 𝑞 𝑗 (𝑡, 𝜉) D 𝑥 𝑎 𝑚+1−|𝛽 | . (27.3.15)
𝑗=0
𝜕𝑡 𝑗 𝑚=0 0≠𝛽 ∈Z 𝑛−𝜈 𝛽!
+

We follow standard procedure: we equate the coefficients of 𝜆−𝑚 on both sides of

(27.3.15). Back to the formal series (27.3.13) we derive the transport equations, for
𝑗 = 1, ..., 𝜈,

𝐻 𝑗 𝑎 0 = 0, (27.3.16)
∑︁ 1 𝛽 𝛽
𝐻 𝑗 𝑎𝑚 = 𝜕 𝑞 𝑗 (𝑡, 𝜉) D 𝑥 𝑎 𝑚+1−|𝛽 | if 𝑚 = 1, 2, ...,
𝛽! 𝜉
2≤ |𝛽 | ≤𝑚+1

where 𝐻 𝑗 is the vector field (27.2.13).

The Cauchy conditions follow from (27.3.11):

𝑎 0 (𝑥 − 𝑦, 𝑠, 𝑠, 𝜉) = 1, (27.3.17)
𝑎 𝑚 (𝑥 − 𝑦, 𝑠, 𝑠, 𝜉) = 0 if 𝑚 = 1, 2, ....

We get right-away
𝑎 0 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) ≡ 1. (27.3.18)

Remark 27.3.4 There is no need to prove compatibility conditions for the right-hand
sides of the equations (27.3.16). Starting from (27.3.18) these equations determine
𝑎 𝑚 for 𝑚 = 1, 2, ..., as a consequence of (27.3.10)–(27.3.11).

The analogous conclusion can be reached by reasoning in 𝑠-space and starting

from the eikonal equation 27.3.4.
From (27.3.16)–(27.3.17) and the fact that the 𝑍 𝑘 (𝑥, 𝑡, 𝜉) are first integrals of the
vector fields 𝐻 𝑗 we derive the following result.

Proposition 27.3.5 For every 𝑚 ∈ Z+𝑛 , 𝑎 𝑚 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) is a polynomial in

𝑍1 (𝑥, 𝑡, 𝜉) − 𝑍1 (𝑦, 𝑠, 𝜉),...,𝑍 𝑛−𝜈 (𝑥, 𝑡, 𝜉) − 𝑍 𝑛−𝜈 (𝑦, 𝑠, 𝜉), in 𝜉1 , ..., 𝜉 𝑛−𝜈 , and in the
partial derivatives 𝜕 𝜉𝛼 𝑞 𝑗 (𝑡, 𝜉), 𝜕 𝜉 𝑞 𝑗 ′ (𝑡, 𝜉), 𝑗, 𝑗 ′ = 1, ..., 𝜈, 𝛼, 𝛽 ∈ Z+𝑛−𝜈 .
𝛽

A closer inspection of the right-hand side of (27.3.16) allows us to prove that, in

fact, (27.3.13) is a formal analytic series (Definition 19.1.18) in the ring of functions
defined by the properties in Proposition 27.3.5. It is a matter of checking that the
condition (GFA) (in Definition 19.1.18) is satisfied, using induction on 𝑚 and the
1174 27 Pseudodifferential Complexes in Tube Structures

standard procedure (cf. e.g. the proof of Proposition 23.2.7). Condition (27.3.1) and
(27.3.18) ensure that the classical amplitude 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) is elliptic of order
zero.

27.4 Approximate Homotopy Formulas

27.4.1 A sequence of Fourier integral operators with complex phase

We are going to reason under the assumption that 𝜇◦ [cf. (27.3.1)] is suitably small
and use the notation

𝑠 (𝑡, 𝜅) = (1 − 𝜅) 𝑡 ◦ + 𝜅𝑡, 0 ≤ 𝜅 ≤ 1, (27.4.1)

with 𝑡 ◦ ∈ Ω to be chosen below (independent of 𝑥, 𝑦, 𝑡); note that 𝑡 ∈ Ω =⇒ 𝑠 (𝑡, 𝜅) ∈

Ω.
We begin by defining, for each 𝑝 = 1, 2, ..., an integral operator acting on distri-
butions 𝑓 ∈ C 1 (Ω; E ′ (𝑈)):
(
𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) (27.4.2)
∫ ∫ 1
− e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 ,𝜅), 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜅) , 𝜉, 𝜆) 𝑓 (𝑦, 𝑠 (𝑡, 𝜅)) 𝜅 𝑝−1 d𝜅d𝑦
R𝑛−𝜈 0

with 𝜑 given by (27.3.2), (𝑥, 𝑡) ∈ 𝑈 × Ω, 𝜉 ∈ S𝑛−𝜈−1 , 𝜆 > 0. The “integral” with

respect to 𝑦 in (27.4.2) is a duality bracket. At this juncture we can choose between
two paths to follow:
(1) We replace the amplitude 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) by one of its finite realiza-
tions (Definition 19.1.7) and the signs = must be replaced by congruences
mod errors that decay exponentially as 𝜆 ↗ +∞ (when dealing with sym-
bols) or mod analytic regularizing operators (when dealing with operators). To
start with, the equations (27.3.10) must be replaced by inequalities

D𝑡 𝑗 + 𝑖𝑞 𝑗 (𝑡, D 𝑥 ) e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) ≤ 𝐶◦ e−𝑐◦ 𝜆 . (27.4.3)

(2) We continue to deal with formal series (in powers of 𝜆−1 , or of |𝜉 | −1 or |D 𝑥 | −1 )

and the signs = truly mean equality (between, say, formal classical symbols or
operators). In this case, the integral with respect to 𝑦 in (27.4.2) is a duality
bracket when 𝑎 is replaced by every one of its homogeneous terms.
Most often we will be following Path (2) since the use of the = sign is then
(
legitimate. However, in the needed estimates of the operators 𝑲 ◦ 𝑝) , it is much
simpler to deal with finite realizations of the amplitudes.
27.4 Approximate Homotopy Formulas 1175

Actually, on each one of the two approaches we encounter another problem at

(
the start: since R𝑛−𝜈 ∋ 𝑥 ↦→ 𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) is not, generally speaking, compactly
(
supported, it might seem that 𝑞 𝑗 (𝑡, D 𝑥 ) 𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) is not well-defined unless
𝑞 𝑗 (𝑡, D 𝑥 ) is a vector field. It is therefore convenient to define

( 𝜕 ( 𝑝)
𝑞 𝑗 (𝑡, D 𝑥 ) 𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) = 𝑲 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) (27.4.4)
𝜕𝑡 𝑗 ◦
∫ ∫ 1
𝜕 𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 )
− e 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑓 (𝑦, 𝑠) 𝜅 𝑝 d𝜅d𝑦.
R𝑛−𝜈 0 𝜕𝑠 𝑗 𝑠=𝑠 (𝑡 ,𝜅)

Then (27.3.10) implies

(
D𝑡 𝑗 + 𝑖𝑞 𝑗 (𝑡, D 𝑥 ) 𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) (27.4.5)
∫ ∫ 1
= D𝑠 𝑗 e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑓 (𝑦, 𝑠) 𝜅 𝑝 d𝜅d𝑦;
R𝑛−𝜈 0 𝑠=𝑠 (𝑡 ,𝜅)

(27.4.5) is literally valid, not just formally. If we were to follow Route (1) the
drawback would be that the definition (27.4.4) depends on the choice of the finite
realization of (27.3.13), as (27.4.2) already did. A change in the finite realization
results in an error exponentially decaying as 𝜆 ↗ +∞, uniformly with respect to the
other variables (possibly after contracting 𝑈 and 𝑉).
We resume dealing with amplitudes that are formal series (presently in the powers
of 𝜆−1 ).

Lemma 27.4.1 We have, for all 𝑓 ∈ C 1 (Ω; E ′ (𝑈)),

( (
𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 = 𝑲 ◦ 𝑝+1) 𝑃 𝑗 𝑓 , 𝑗 = 1, ..., 𝜈. (27.4.6)

Proof Since 𝑦 ↦→ 𝑓 (𝑦, 𝑡) is compactly supported, (27.3.3)–(27.3.4) and (27.4.4)

imply

(
D𝑡 𝑗 + 𝑖𝑞 𝑗 (𝑡, D 𝑥 ) 𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆)
∫ ∫ 1
= D𝑠 𝑗 e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑓 (𝑦, 𝑠 (𝑡, 𝜅)) 𝜅 𝑝 d𝜅d𝑦
R𝑛−𝜈 0 𝑠=𝑠 (𝑡 ,𝜅)
∫ ∫ 1
+ e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) D𝑠 𝑗 𝑓 (𝑦, 𝑠) 𝜅 𝑝 d𝜅d𝑦
R𝑛−𝜈 0 𝑠=𝑠 (𝑡 ,𝜅)
1176 27 Pseudodifferential Complexes in Tube Structures
∫ ∫ 1
=𝑖 𝑞 𝑗 (𝑠, D 𝑥 ) e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆)
R𝑛−𝜈 0 𝑠=𝑠 (𝑡 ,𝜅)
× 𝑓 (𝑦, 𝑠 (𝑡, 𝜅)) 𝜅 𝑝 d𝜅d𝑦
∫ ∫ 1
+ e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑃 𝑗 𝑓 (𝑦, 𝑠) 𝜅 𝑝 d𝜅d𝑦
R𝑛−𝜈 0 𝑠=𝑠 (𝑡 ,𝜅)
∫ ∫ 1
e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑞 𝑗 𝑠, D 𝑦 𝑓 (𝑦, 𝑠)

−𝑖 𝜅 𝑝 d𝜅d𝑦.
R𝑛−𝜈 0 𝑠=𝑠 (𝑡 ,𝜅)

Since 𝑞 𝑗 (𝑠,D 𝑥 ) ℎ (𝑥 − 𝑦) = 𝑞 𝑗 𝑠, −D 𝑦 ℎ (𝑥 − 𝑦) and 𝑞 𝑗 𝑠, −D 𝑦 is the transpose
of 𝑞 𝑗 𝑠, D 𝑦 , (27.4.6) ensues. □

27.4.2 The operators 𝑲 (𝒑) acting on differential forms

We now deal with a 𝑝-form (27.2.2), 𝑓 (𝑥, 𝑡, d𝑡), 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)), and the
(formal) pseudodifferential operator (27.2.3), d𝑷 . Let 𝑡 ◦ ∈ Ω be the same point as in
(27.4.1). We shall make use of the ( 𝑝 − 1)-form in R𝜈 ,
𝑝
∑︁
𝜛𝐼 (𝑡 ◦ ) = (−1) 𝛼−1 𝑡 𝑖 𝛼 − 𝑡 𝑖◦𝛼 d𝑡𝑖1 ∧ · · · ∧ d𝑡
d 𝑖 𝛼 ∧ · · · ∧ d𝑡 𝑖 𝑝 , (27.4.7)
𝛼=1

where 𝐼 = 𝑖 1 , ..., 𝑖 𝑝 is a multi-index such that 𝑖1 < · · · < 𝑖 𝑝 (𝑝 ≥ 1) and the hatted
factor must be omitted; we have d𝜛𝐼 = 𝑝d𝑡 𝐼 . We define [cf. (27.4.2)]
∑︁
(
𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) = 𝑲 ◦ 𝑝) 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) 𝜛𝐼 (𝑡 ◦ ) , (27.4.8)
|𝐼 |= 𝑝

where (𝑥, 𝑡) ∈ 𝑈 × Ω, 𝜉 ∈ S𝑛−𝜈−1 , 𝜆 > 0. We also introduce the following FIO

operator with a complex phase-function, acting on compactly supported distributions
in R𝑛−𝜈 : ∫
𭟋 𝑓 (𝑥, 𝜉, 𝜆) = e𝑖𝜆( 𝑥−𝑦) · 𝜉 𝑓 (𝑦) d𝑦; (27.4.9)
R𝑛−𝜈
recall that 𝜑 (𝑥 − 𝑦, 𝑡, 𝑡, 𝜉) = (𝑥 − 𝑦) · 𝜉.
Henceforth, the operators 𝑲 ( 𝑝) d𝑷 and d𝑷 𝑲 ( 𝑝) are defined by (27.4.8) and (27.4.5),
(27.4.6).

Remark 27.4.2 When dealing with the operators 𝑲 ( 𝑝) d𝑷 and d𝑷 𝑲 ( 𝑝) we can regard
𝜉 and 𝜆 as parameters until specified otherwise.

Lemma 27.4.3 If 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) we have, for all (𝑥, 𝑡, 𝜉) ∈ 𝑈 × Ω × S𝑛−𝜈−1 ,

𝜆 > 0,
27.4 Approximate Homotopy Formulas 1177

𭟋 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) − 𝑲 (1) d𝑷 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) (27.4.10)

∫
◦
= e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑡 , 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑡 ◦ , 𝜉, 𝜆) 𝑓 (𝑦, 𝑡 ◦ ) d𝑦.
R𝑛−𝜈

Proof We have, by (27.4.6) and (27.4.8),

𝜈 𝜈 𝜈
©∑︁ ª ∑︁ ∑︁
𝑲 (1) ­ 𝑃 𝑗 𝑓 d𝑡 𝑗 ® = 𝑡 𝑗 − 𝑡 ◦𝑗 𝑲 ◦(1) 𝑃 𝑗 𝑓 = 𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦(1) 𝑓 .
« 𝑗=1 ¬ 𝑗=1 𝑗=1

We take into account the fact that 𝜕

𝜕𝜅 𝑠 𝑗 (𝑡, 𝜅) = 𝑡 𝑗 − 𝑡 ◦𝑗 , 𝑠 𝑗 (𝑡, 1) = 𝑡 𝑗 , 𝑠 𝑗 (𝑡, 0) = 𝑡 ◦𝑗 ,
and apply (27.4.5):
𝜈
∑︁
𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦(1) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆)
𝑗=1
𝜈
∑︁
= 𝑡 𝑗 − 𝑡 ◦𝑗
𝑗=1
∫ ∫ 1
𝜕 𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 )
× e 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑓 (𝑦, 𝑠) d𝜅d𝑦
R𝑛−𝜈 0 𝜕𝑠 𝑗 𝑠=𝑠 (𝑡 ,𝜅)
𝜈 ∫ ∫ 1
∑︁ 𝜕 𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 ,𝜅), 𝜉 )
= e 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜅) , 𝜉, 𝜆)
𝑗=1 R
𝑛−𝜈 0 𝜕𝜅

× 𝑓 (𝑦, 𝑠 (𝑡, 𝜅))) d𝜅d𝑦

∫
= e𝑖𝜆( 𝑥−𝑦) · 𝜉 𝑓 (𝑦, 𝑡) d𝑦
R𝑛−𝜈
∫
◦, 𝜉)
− e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑡 𝑎 (𝑥 − 𝑦, 𝑡, 𝑡 ◦ , 𝜉, 𝜆) 𝑓 (𝑦, 𝑡 ◦ ) d𝑦. □
R𝑛−𝜈

Lemma 27.4.4 Let 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 with 𝑝 ≥ 1, 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)).

Í

We have, for all (𝑥, 𝑡, 𝜉) ∈ 𝑈 × 𝑉 × S𝑛−𝜈−1 , 𝜆 > 0,

∑︁
𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝐼 = d𝑷 𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) + 𝑲 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) .
|𝐼 |= 𝑝
(27.4.11)

Proof Given 𝐼 = 𝑖1 , ..., 𝑖 𝑝 we have
𝑝
!
∑︁ ∑︁
( (
𝑃𝑖 𝛼 𝑲 ◦ 𝑝) 𝑓 𝐼 (𝑦, 𝑡, 𝜉, 𝜆) d𝑡𝑖 𝛼 ∧ 𝜛𝐼 = 𝑡 𝑗 − 𝑡 ◦𝑗 𝑃𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 (𝑦, 𝑡, 𝜉, 𝜆) d𝑡 𝐼
𝛼=1 𝑗 ∈𝐼

and d𝑷 𝜛𝐼 (𝑡 ◦ ) = 𝑝d𝑡 𝐼 , whence

1178 27 Pseudodifferential Complexes in Tube Structures
∑︁ ∑︁
(
d𝑷 𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) = 𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 d𝑡 𝐼
|𝐼 |= 𝑝 𝑗 ∈𝐼
∑︁ ∑︁
(
+ 𝑃𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 d𝑡 𝑗 ∧ 𝜛𝐼 (𝑡 ◦ ) + 𝑝𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) ,
|𝐼 |= 𝑝 𝑗∉𝐼

i.e.,
𝜈
∑︁ ∑︁
( 𝑝) (
d𝑷 𝑲 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) = 𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 d𝑡 𝐼 (27.4.12)
|𝐼 |= 𝑝 𝑗=1
∑︁ ∑︁
(
+𝑝𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) − 𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 d𝑡 𝐼
|𝐼 |= 𝑝 𝑗∉𝐼
∑︁ ∑︁
(
+ 𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 d𝑡 𝑗 ∧ 𝜛𝐼 (𝑡 ◦ ) .
|𝐼 |= 𝑝 𝑗∉𝐼

We apply (27.4.5):
𝜈
∑︁
(
𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆)
𝑗=1
𝜈
∑︁
= 𝑡 𝑗 − 𝑡 ◦𝑗
𝑗=1
∫ ∫ 1
𝜕 𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 )
× e 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) 𝑓 𝐼 (𝑦, 𝑠) 𝜅 𝑝 d𝜅d𝑦
R𝑛−𝜈 0 𝜕𝑠 𝑗 𝑠=𝑠 (𝑡 ,𝜅)
∫ ∫ 1
𝜕 𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 ,𝜅), 𝜉 )
= 𝜅𝑝 e 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜅) , 𝜉, 𝜆) 𝑓 𝐼 (𝑦, 𝑠 (𝑡, 𝜅)) d𝜅d𝑦
R𝑛−𝜈 0 𝜕𝜅

∫
= e𝑖 𝜑 ( 𝑥,𝑡 ,𝑦,𝑡 , 𝜉 ,𝜆) 𝑎 (𝑥, 𝑡, 𝑦, 𝑡, 𝜉, 𝜆) 𝑓 𝐼 (𝑦, 𝑡) d𝑦
R𝑛−𝜈
∫
−𝑝 e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 ,𝜅), 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜅) , 𝜉, 𝜆) 𝑓 𝐼 (𝑦, 𝑠 (𝑡, 𝜅)) 𝜅 𝑝−1 d𝜅d𝑦,
R𝑛−𝜈

whence
𝜈
∑︁ ∑︁ ∑︁
(
𝑡 𝑗 − 𝑡 ◦𝑗 𝑃 𝑗 𝑲 ◦ 𝑝) 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝐼 = 𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝐼
|𝐼 |= 𝑝 𝑗=1 |𝐼 |= 𝑝
∑︁
(
−𝑝 𝑲 ◦ 𝑝) 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝐼 .
|𝐼 |= 𝑝

Putting this into (27.4.12) and applying (27.4.6) yields

27.4 Approximate Homotopy Formulas 1179
∑︁
d𝑷 𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) = 𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝐼 (27.4.13)
|𝐼 |= 𝑝
∑︁ ∑︁
(
− 𝑲 ◦ 𝑝+1) 𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) 𝑡 𝑗 − 𝑡 ◦𝑗 d𝑡 𝐼 − d𝑡 𝑗 ∧ 𝜛𝐼 .
|𝐼 |= 𝑝 𝑗∉𝐼

A straightforward computation yields

𝑲 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡)

∑︁ ∑︁
(
= 𝑲 ◦ 𝑝+1) 𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) 𝑡 𝑗 − 𝑡 ◦𝑗 d𝑡 𝐼 − d𝑡 𝑗 ∧ 𝜛𝐼
|𝐼 |= 𝑝 𝑗∉𝐼

and putting this into (27.4.13) yields (27.4.11). □

Corollary 27.4.5 Let

∑︁
𝑓 (𝑥, 𝑡, d𝑡) = 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 with 𝑝 ≥ 0, 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)) .
|𝐼 |= 𝑝

We have
𝜈
∑︁ ∑︁
d𝑷 𝑲 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) = 𝑃 𝑗 𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝑗 ∧ d𝑡 𝐼 .
|𝐼 |= 𝑝 𝑗=1

Proof Let d𝑷 act on both sides of (27.4.11) if 𝑝 ≥ 1 [(27.4.10) if 𝑝 = 0]. □

Corollary 27.4.6 Let 𝑓 (𝑥, 𝑡, d𝑡) be as in Corollary 27.4.5; then

∑︁
d𝑷 𝑲 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) = 𭟋𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝑗 ∧ d𝑡 𝐼 .
|𝐼 |= 𝑝

Proof Replace 𝑓 by d𝑷 𝑓 and 𝑝 by 𝑝 + 1 in (27.4.11). □

If we combine the two preceding corollaries, in which we put 𝑝 = 0, we obtain
∫ 𝑛−𝜈
1 𝜉
e𝑖𝜆( 𝑥−𝑦) · 𝜉 𝑃 𝑗 𝑓 (𝑦, 𝑡) 1 + 𝑖 (𝑥 − 𝑦) ·

𝑃 𝑗 𭟋 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) = d𝑦
R𝑛−𝜈 2 |𝜉 |
(27.4.14)
for every 𝑓 ∈ C 1 (Ω; E ′ (𝑈)); if 𝑝 ≥ 1,
∑︁ 𝑛
∑︁ ∑︁
d𝑷 𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝐼 = 𭟋𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡, 𝜉, 𝜆) d𝑡 𝑗 ∧ d𝑡 𝐼 (27.4.15)
|𝐼 |= 𝑝 |𝐼 |= 𝑝 𝑗=1

for every 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 with 𝑝 ≥ 1, 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)). Of

Í
course, Formulas (27.4.14) and (27.4.15) can be derived directly from (27.4.9).
1180 27 Pseudodifferential Complexes in Tube Structures

(𝒑)
27.4.3 Basic bounds on 𝑲 ◦ 𝒇

We prove rough bounds for the integrals (27.4.2), based on the simple identity, valid
for holomorphic functions ℎ in a domain in C𝑛−𝜈 :

𝑛−𝜈
!𝑁 𝑛−𝜈
!𝑁
∑︁ ∑︁ 2
−𝑖𝜆𝑧· 𝜉
e D2𝑧𝑘 e 𝑖𝜆𝑧· 𝜉
ℎ = D𝑧𝑘 − 𝜆𝜉 𝑘 ℎ (27.4.16)
𝑘=1 𝑘=1

(𝑁 ∈ Z+ ). We will take 𝑧 = 𝑍 (𝑥, 𝑡) − 𝑍 (𝑦, 𝑠 (𝑡, 𝜅)) and make use of the identities

Δ 𝑥 ℎ (𝑍 (𝑥, 𝑡) − 𝑍 (𝑦, 𝑠)) = Δ 𝑦 ℎ (𝑍 (𝑥, 𝑡) − 𝑍 (𝑦, 𝑠))

𝑛−𝜈 2
∑︁ 𝜕 ℎ
= (𝑍 (𝑥, 𝑡) − 𝑍 (𝑦, 𝑠)) .
𝑘=1
𝜕𝑧 2𝑘

In this context it simplifies the notation if we apply (27.4.16) to a preselected finite

realization of (27.3.13); this is made possible by Proposition 27.3.5. We denote by
ℎ 𝑁 (𝑧, 𝑡, 𝜉, 𝜆) the right-hand side in (27.4.16); since 𝑎 0 ≡ 1 we see that to each
𝑁 ∈ Z+ there is an 𝑅 𝑁 > 0 such that 𝜆 > 𝑅 𝑁 implies

|ℎ 𝑁 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆)| ≥ 2−𝑁 𝜆2𝑁 . (27.4.17)

We return to (27.4.2) with now 𝑎 a finite realization of (27.3.13). To shorten the

formulas it is convenient to use, just for the remainder of this subsection, the notation
𝑋 = (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜅) , 𝜉). By (27.3.2) we have
(
𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆)
∫ ∫ 1 𝑓 (𝑦, 𝑠 (𝑡, 𝜅))
= Δ 𝑥𝑁 e𝑖𝜆𝜑 (𝑋) 𝑎 (𝑋, 𝜆) 𝜅 𝑝−1 d𝜅d𝑦,
R𝑛−𝜈 0 ℎ 𝑁 (𝑋, 𝜆)

where 𝑓 ∈ C 1 (𝑈 × Ω). If we apply the transpose Leibniz rule (1.1.5) we obtain

∑︁
(
𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) = Δ 𝑥𝑁 ℑ 𝑁 (𝑥, 𝑡) + 𝜕𝑥𝛼 ℑ 𝑁 ,𝛽 (𝑥, 𝑡), (27.4.18)
| 𝛼+𝛽 |=2𝑁
𝛽≠0

where
∫ ∫ 1
𝑎 (𝑋, 𝜆)
ℑ 𝑁 (𝑥, 𝑡) = e𝑖𝜆𝜑 (𝑋) 𝑓 (𝑦, 𝑠 (𝑡, 𝜅)) 𝜅 𝑝−1 d𝜅d𝑦,
R𝑛−𝜈 0 ℎ 𝑁 (𝑋, 𝜆)

ℑ 𝑁 ,𝛽 (𝑥, 𝑡)
∫ ∫ 1
𝑎 (𝑋, 𝜆)
= e𝑖𝜆𝜑 (𝑋) 𝑔 𝑁 ,𝛽 (𝑋, 𝜆) 𝑓 (𝑦, 𝑠 (𝑡, 𝜅)) 𝜅 𝑝−1 d𝜅d𝑦,
R𝑛−𝜈 0 ℎ 𝑁 (𝑋, 𝜆)
27.4 Approximate Homotopy Formulas 1181

and 𝑔 𝑁 ,𝛽 (𝑋, 𝜆) is bounded in (𝑈 × Ω) 2 × S𝑛−𝜈−1 , 𝜆 > 1. In (27.4.18) the absolute

values of the integrands are ≲ |𝜆| −2𝑁 exp (− Im 𝜆𝜑 (𝑋)); more precisely, to every
𝑁 ∈ Z+ there is a 𝐶 𝑁 > 0 such that, for all (𝑥, 𝑦, 𝑡, 𝜉) ∈ 𝑈 2 × Ω × S𝑛−𝜈−1 , 0 ≤ 𝜅 ≤ 1,
𝜆 > 𝑅𝑁 ,

𝑎 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜅) , 𝜉, 𝜆)
e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 ,𝜅), 𝜉 ) ≤ 𝐶 𝑁 𝜆−2𝑁 exp 𝜆𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜅) , 𝜉)

ℎ 𝑁 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆)
(27.4.19)
(cf. Definition 27.3.2).
It is also obvious that we are allowed to take 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) since finitely
many integrations by parts with respect to 𝑦 in the right-hand side of (27.4.18) are
permitted and yield amplitudes

∑︁
−𝑖𝜆𝜑 (𝑋) 𝛼 𝑖𝜆𝜑 (𝑋) 𝑔 𝑁 ,𝛽 (𝑋, 𝜆)
e 𝜕𝑦 e 𝑎 (𝑋, 𝜆) ,
𝛼∈Z𝑛−𝜈
ℎ 𝑁 (𝑋, 𝜆)
+

that have good bounds – provided 𝑁 is suitably large.

In the remainder of this section, unless specified otherwise, we continue to deal
with functions (or distributions) in a convex domain Ω (or 𝑉 ⊂⊂ Ω) of 𝑡-space
R𝜈 valued in the space of compactly supported distributions in a domain 𝑈 of 𝑥-
space R𝑛−𝜈 . With this understanding all integrals with respect to 𝑦 stand for duality
brackets.

27.4.4 Approximate homotopy formulas

We introduce a convex neighborhood 𝑉 ⊂⊂ Ω of 0 in R𝜈 , whose size will be

adjusted as needed, an open cone ℭ in R𝑛−𝜈 \ {0}, possibly equal to R𝑛−𝜈 \ {0}, and a
Lebesgue-measurable map ℭ ∋ 𝜉 ↦→ 𝑡 (𝜉) ∈ Ω, homogeneous of degree zero, whose
range is contained in a compact subset of Ω. We apply all the results established so
far, with 𝑡 ◦ = 𝑡 (𝜉) and we shall frequently use the notation

𝑠 (𝑡, 𝜉, 𝜅) = (1 − 𝜅) 𝑡 (𝜉) + 𝑡. (27.4.20)

We denote by Λ (𝑡, 𝜉) the affine line R ∋ 𝜅 ↦→ 𝑠 (𝑡, 𝜉, 𝜅). Since 𝑉 is convex Λ (𝑡, 𝜉) ∩
𝑉 ≠ ∅ is a nonempty (open) interval if 𝑡 = 𝑠 (𝑡, 𝜉, 1) ∈ 𝑉; but 𝑠 (𝑡, 𝜉, 0) = 𝑡 (𝜉)
might not belong to 𝑉, not even to 𝑉. The integral with respect to 𝜅 over [0, 1] can
be regarded as an integral in 𝑠-space, over the segment connecting 𝑡 (𝜉) to 𝑡; (27.4.2)
becomes
1182 27 Pseudodifferential Complexes in Tube Structures
∫ ∫ 1
(
𝑲 ◦ 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) = e𝑖𝜆𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉, 𝜆)
R𝑛−𝜈 0
× 𝑓 (𝑦, 𝑠 (𝑡, 𝜉, 𝜅)) 𝜅 𝑝−1 d𝜅d𝑦,
(27.4.21)
(
where 𝑓 ∈ C 1 (Ω; E ′ (𝑈)). With this definition of 𝑲 ◦ 𝑝) Formula (27.4.8) still defines
𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆, d𝑡) if 𝑓 (𝑥, 𝑡, d𝑡) is a differential form
At this stage 𝜉 becomes a bona fide variable. We introduce the integrals
∫
1 2
f◦ ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉) d𝜉,
𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡) = 𝑛−𝜈 e−𝜀 | 𝜉 | 𝑲 (27.4.22)
(2𝜋) ℭ

f◦ ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉) = 𝑲 ◦( 𝑝) 𝑓 𝑥, 𝑡, 𝜉 , |𝜉 | – which

where 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) and 𝑲 |𝜉 |
makes us deal with amplitudes that are formal analytic series in the powers of |𝜉 | −1 .
We extend coefficientwise the notation (27.4.22) to an arbitrary 𝑝-form (27.2.2),
𝑓 (𝑥, 𝑡, d𝑡) (𝑝 ≥ 1):
∑︁
𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡) = 𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 𝐼 (𝑥, 𝑡) 𝜛𝐼 , (27.4.23)
|𝐼 |= 𝑝

where 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)) for all multi-indices 𝐼. We have

∫
1 2
f◦ ( 𝑝) d𝑷 𝑓 (𝑥, 𝑡, 𝜉) d𝜉,
𝑨 ( 𝑝, 𝜀) (ℭ) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) = 𝑛−𝜈 e−𝜀 | 𝜉 | 𝑲 (27.4.24)
(2𝜋)
∫ℭ
1 2
f◦ ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉) d𝜉,
d𝑷 𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡) = e−𝜀 | 𝜉 | d𝑷 𝑲
(2𝜋) 𝑛−𝜈 ℭ

with 𝑲 ( 𝑝) d𝑷 and d𝑷 𝑲 ( 𝑝) still defined by (27.4.8), (27.4.5), (27.4.6). We write 𝑨 ( 𝑝, 𝜀)

instead of 𝑨 ( 𝑝, 𝜀) (R𝑛−𝜈 \ {0}) .
We also introduce the following notation, for arbitrary 𝑓 ∈ E ′ (R𝑛−𝜈 ),
∫
𝜉
𭟋 𝑓 (𝑥, 𝜉) = 𭟋 𝑓 𝑥, , |𝜉 | =
e e𝑖 ( 𝑥−𝑦) · 𝜉 𝑓 (𝑦) d𝑦. (27.4.25)
|𝜉 | R𝑛−𝜈

We derive directly from Lemmas 27.4.3, Lemma 27.4.4:

Theorem 27.4.7 Let 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) be arbitrary; we have

∫
2
e−𝜀 | 𝜉 | e 𭟋 𝑓 (𝑥, 𝑡, 𝜉) d𝜉 = (2𝜋) 𝑛−𝜈 𝑨 (1, 𝜀) (ℭ) d𝑷 𝑓 (𝑥, 𝑡) (27.4.26)
ℭ
∫ ∫
2 𝜉
+ e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 )−𝜀 | 𝜉 | 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 | 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉.
ℭ R 𝑛−𝜈 |𝜉 |

𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 with 𝑝 ≥ 1, 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)). We have

Í
Let 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝
27.4 Approximate Homotopy Formulas 1183
∫
∑︁ 1 2
e−𝜀 | 𝜉 | e𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉) d𝜉 d𝑡 𝐼 (27.4.27)
(2𝜋) 𝑛−𝜈 ℭ
|𝐼 |= 𝑝

= d𝑷 𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡) + 𝑨 ( 𝑝+1, 𝜀) (ℭ) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) .

The case ℭ = R𝑛−𝜈 \ {0} is noteworthy, due to the Fourier inversion formula:
∫
1
𝑓 (𝑥) = 𭟋 𝑓 (𝑥, 𝜉) d𝜉.
e (27.4.28)
(2𝜋) 𝑛−𝜈 R𝑛−𝜈

Combining Theorem 27.4.7 where ℭ = R𝑛−𝜈 \ {0} with 27.4.28 yields

Theorem 27.4.8 Let 𝑓 ∈ C 1 (Ω; E ′ (𝑈)); then, as 𝜀 ↘ 0,

𝑨 (1, 𝜀) d𝑷 𝑓 (𝑥, 𝑡, d𝑡)

∫
1 𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 )−𝜀 | 𝜉 | 2 𝜉
+ e 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 |
(2𝜋) 𝑛−𝜈 R2(𝑛−𝜈) |𝜉 |
× 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉

converges uniformly Í on compact subsets of 𝑈 × Ω to 𝑓 (𝑥, 𝑡).

Let 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 with 𝑝 ≥ 1, 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)); then, as
𝜀 ↘ 0,
d𝑷 𝑨 ( 𝑝, 𝜀) 𝑓 (𝑥, 𝑡, d𝑡) + 𝑨 ( 𝑝+1, 𝜀) d𝑷 𝑓 (𝑥, 𝑡, d𝑡)
converges uniformly on compact subsets of 𝑈 × 𝑉 to 𝑓 (𝑥, 𝑡, d𝑡).

The conclusions in Theorem 27.4.8 do not mean that 𝑨 ( 𝑝+1, 𝜀) d𝑷 𝑓 and, if 𝑝 ≥ 1,

d𝑷 𝑨 ( 𝑝, 𝜀) 𝑓 separately converge in C (𝑈; Λ 𝑝 C𝑇 ∗𝑉). It has, however, the following

Corollary 27.4.9 Let 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) satisfy d𝑷 𝑓 = 0; then 𝑓 is the limit of

∫
1 𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 )−𝜀 | 𝜉 | 2 𝜉
e 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 |
(2𝜋) 𝑛−𝜈 R2(𝑛−𝜈) |𝜉 |
× 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉

in C (R𝑛−𝜈 × 𝑉) as 𝜀 ↘ 0.
Í
Corollary 27.4.10 Let 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑝 ≥ 1, be as in Theorem
27.4.8. If d𝑷 𝑓 = 0 then 𝑓 is the limit of d𝑷 𝑨 ( 𝑝, 𝜀) 𝑓 in C (R𝑛−𝜈 × 𝑉) as 𝜀 ↘ 0.

If we take Proposition 27.3.5 into account we can regard Corollary 27.4.9 as a

special case of the approximation formula in [Baouendi-Treves, 1981] and Corol-
lary 27.4.10 as a special case of the Approximate Poincaré Lemma (see [Treves,
1992], II.6). Corollary 27.4.10 does not say that if d𝑷 𝑓 = 0 then 𝑨 ( 𝑝, 𝜀) 𝑓 converges
uniformly on compact subsets of R𝑛−𝜈 × Ω.
1184 27 Pseudodifferential Complexes in Tube Structures

27.5 Homotopy Formulas

27.5.1 A sufficient condition for convergence

We assume that (27.4.19) holds with 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) and 𝑁 as large as needed
(depending on 𝑓 ). We start from the decomposition
∫
1 2
e ◦( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉) d𝜉 (27.5.1)
𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡) = e−𝜀 | 𝜉 | 𝑲
(2𝜋) 𝑛−𝜈 𝜉 ∈ℭ, | 𝜉 |<𝑅 𝑁
∫
1 2
e ◦( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉) d𝜉.
+ e−𝜀 | 𝜉 | 𝑲
(2𝜋) 𝑛−𝜈 𝜉 ∈ℭ, | 𝜉 |>𝑅 𝑁

The first term in the right-hand side converges, as 𝜀 ↘ 0, to a C 1 function in

𝑈 × 𝑉. By (27.4.18) the convergence of the second term (to a finite sum of distri-
bution derivatives of continuous functions in 𝑈 × 𝑉) will depend on the behavior
of the exponent-weight 𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) [cf. Definition 27.3.2; keep in mind that
𝑠 (𝑡, 𝜉, 𝜅) = (1 − 𝜅) 𝑡 (𝜉) + 𝜅𝑡 ]. We come now to the key condition on 𝜋 𝜑 that will
enable us to prove the homotopy formulas and the versions of Poincaré lemma that
they entail.
Definition 27.5.1 Let the convex open set 𝑉 in R𝜈 , 𝑉 ⊂⊂ Ω, and 𝜉 ∈ R𝑛−𝜈 \ {0} be
arbitrary. We shall say that Condition 𝚷 (𝑉, 𝜉) is satisfied (implicitly, by the system
of pseudodifferential operators 𝑷) if there is a 𝑡 (𝜉) ∈ Ω such that

∀ (𝑡, 𝜅) ∈ 𝑉 × [0, 1], 𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) ≤ 0.

Given an open cone ℭ in R𝑛−𝜈 \ {0} we shall say that Condition 𝚷 (𝑉, ℭ) is
satisfied if 𝚷 (𝑉, 𝜉) is satisfied for every 𝜉 ∈ ℭ, with ℭ ∋𝜉 ↦→ 𝑡 (𝜉) ∈ 𝐾 ⊂ Ω
measurable, homogeneous of degree zero, 𝐾 compact.
Given arbitrarily (𝑡 ◦ , 𝜉 ◦ ) ∈ Ω× (R𝑛−𝜈 \ {0}) we shall say that Condition 𝚷 (𝑡 ◦ , 𝜉 ◦ )
is satisfied if there is a basis of convex neighborhoods of 𝑡 ◦ , 𝑉 𝑗 ⊂⊂ Ω, 𝑗 = 1, 2, ...,
such that 𝚷 𝑉 𝑗 , 𝜉 ◦ is satisfied for each 𝑗.
The important point in Condition 𝚷 (𝑉, 𝜉) is that 𝑡 (𝜉) is independent of 𝑡 ∈ 𝑉. It is
also worth underlining the fact that 𝚷 (𝑉, ℭ) does not rely on a particular orientation
of the segment joining 𝑡 (𝜉) to 𝑡 ∈ 𝑉.
Remark 27.5.2 If Condition 𝚷 (𝑉, ℭ) is satisfied then 𝚷 (𝑉 ′, ℭ ′) is satisfied for
every convex open set 𝑉 ′ in R𝜈 , 𝑉 ′ ⊂ 𝑉, and every open cone ℭ ′ in R𝑛−𝜈 \ {0},
ℭ ′ ⊂ ℭ.
We shall write 𝚷 (𝑛−𝜈) (𝑉) for 𝚷 (𝑉, R𝑛−𝜈 \ {0}); 𝚷 (𝑛−𝜈) (𝑉) =⇒ 𝚷 (𝑛−𝜈−1) (𝑉) if
𝑛 − 𝜈 ≥ 2.
Theorem 27.5.3 Let 𝑈 be a domain in R𝑛−𝜈 , 𝑉 a convex open subset of R𝜈 , 𝑉 ⊂⊂ Ω,
and ℭ an open cone in R𝑛−𝜈 \ {0}. Suppose that Condition 𝚷 (𝑉, ℭ) is satisfied. Under
these hypotheses the following properties hold:
27.5 Homotopy Formulas 1185

If 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) then 𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡) converges in C 1 𝑉; D 𝑥′ (𝑈) , as

𝜀 ↘ 0, to 𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡).

1 ′
( 𝑝, 𝜀) 1 ′
to C (Ω; E (𝑈))
If 𝑝 ≥ 1 and if the coefficients 𝑓 𝐼 of the 𝑝-form (27.2.2) belong
then those of 𝑨 (ℭ) 𝑓 (𝑥, 𝑡, d𝑡) converge in C 𝑉; D 𝑥 (𝑈) to the coefficients of
a ( 𝑝 − 1)-current 𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡).
Proof By (27.4.18) it suffices to prove the claim for 𝑓 ∈ C 1 (𝑈 × 𝑉) such that
𝑥 ↦→ 𝑓 (𝑥, 𝑡) is compactly supported; (27.4.17)–(27.4.19) entail, for 𝜉 ∈ R𝑛−𝜈−1 ,
|𝜉 | > 𝑅 𝑁 ,

𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝜉
e 𝑎 𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , , |𝜉 |
|𝜉 |
−2𝑁
≤ 𝐶 𝑁 |𝜉 | exp 𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) ,

with the understanding that 𝑎 is a finite realization of the formal analytic series
(27.3.13). If 𝚷 (𝑉, ℭ) is satisfied, and if we take (27.4.17) and (27.4.19) into account
in (27.4.21) then, for (𝑥, 𝑡) ∈ 𝑈 × 𝑉, 𝜉 ∈ ℭ, we obtain
∫ ∫ 1
( 𝜉
𝑲 ◦ 𝑝) 𝑓 𝑥, 𝑡, , |𝜉 | ≤ 𝐶 𝑁 |𝜉 | −2𝑁 | 𝑓 (𝑦, 𝑠 (𝑡, 𝜉, 𝜅))| 𝜅 𝑝−1 d𝜅d𝑦.
|𝜉 | R𝑛−𝜈 0

Applying this last estimate, with 2𝑁 > 𝑛 − 𝜈, to (27.4.22) and (27.4.23) proves the
claim (by the Lebesgue Dominated Convergence Theorem). □
When Condition 𝚷 (𝑉, ℭ) is satisfied the formulas (27.4.18) and the estimates
(27.4.19) allow us to interpret

𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡) (27.5.2)

∫ ∫ ∫ 1
1 𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝜉
= e 𝑎 𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , , |𝜉 |
(2𝜋) 𝑛−𝜈 ℭ R𝑛−𝜈 0 |𝜉 |
× 𝑓 (𝑦, 𝑠 (𝑡, 𝜉, 𝜅)) 𝜅 𝑝−1 d𝜅d𝑦d𝜉

as an oscillatory integral.
By combining Theorems 27.4.7 and 27.5.3 we obtain the following “true” homo-
topy formulas:
Theorem 27.5.4 Under the same hypotheses as in Theorem 27.5.3 the following
properties hold:
If 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) then we have, in 𝑈 × 𝑉,
∫
1
𭟋 𝑓 (𝑥, 𝑡, 𝜉) d𝜉 = 𝑨 (1) (ℭ) d𝑷 𝑓 (𝑥, 𝑡) +
e (27.5.3)
(2𝜋) 𝑛−𝜈 ℭ
∫ ∫
1 𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝜉
e 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 | 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉.
(2𝜋) 𝑛−𝜈 ℭ R𝑛−𝜈 |𝜉 |

𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)), 𝑝 ≥ 1, then we have, in

Í
If 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝
𝑈 × 𝑉,
1186 27 Pseudodifferential Complexes in Tube Structures
∫
∑︁ 1
𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉) d𝜉 d𝑡 𝐼
e (27.5.4)
(2𝜋) 𝑛−𝜈 ℭ
|𝐼 |= 𝑝

= d𝑷 𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡) + 𝑨 ( 𝑝+1) (ℭ) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) .

Corollary 27.5.5 (cf. Corollary 27.4.5) Let 𝑓 (𝑥, 𝑡, d𝑡), 𝑈 and 𝑉 be as in Theorem
27.5.3. If Condition 𝚷 (𝑉, ℭ) is satisfied then we have, in 𝑈 × 𝑉,
𝜈 ∫
∑︁ ∑︁ 1
d𝑷 𝑨 ( 𝑝+1) (ℭ) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) = 𭟋𝑃 𝑗 𝑓 𝐼 (𝑥, 𝑡, 𝜉) d𝜉 d𝑡 𝑗 ∧ d𝑡 𝐼 .
e
(2𝜋) 𝑛−𝜈 ℭ
|𝐼 |= 𝑝 𝑗=1
(27.5.5)
Proof Replace 𝑓 by d𝑷 𝑓 and 𝑝 by 𝑝 + 1 in (27.5.4). □
We take a closer look at the regularity of 𝑨 ( 𝑝) (ℭ) 𝑓 with respect to 𝑡. Back to
the definition (27.4.2) it is clear that if 𝑓 ∈ C (Ω; E ′ (𝑈)) then 𝑲 ( 𝑝) 𝑓 𝑥, 𝑡, | 𝜉𝜉 | , |𝜉 |
is continuous with respect to 𝑡 (and C 𝜔 with respect to 𝑥, 𝜉). It then follows from
(27.4.2) and the estimates implied by (27.4.18) that if 𝑓 ∈ C 𝑘 (Ω; E ′ (𝑈)) with
𝑘 ∈ Z+ or 𝑘 = ∞, then 𝑲 ( 𝑝) 𝑓 (𝑥, 𝑡, 𝜉, 𝜆) is C 𝑘 with respect to 𝑡. We can state
Proposition 27.5.6 Assume the same hypotheses as in Theorem 27.5.3. If 𝑓 ∈
C 𝑘 (Ω; E ′ (𝑈)) (1 ≤ 𝑘, 1, 2, ..., +∞) then 𝑡 ↦→ 𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡) converges in
C 𝑘 Ω; D 𝑥′ (𝑈) , as 𝜀 ↘ 0, to a distribution 𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡). If 𝑝 ≥ 1 and if the
coefficients 𝑓 𝐼 of the 𝑝-form (27.2.2) belong to C 𝑘 (Ω; E ′ (𝑈)) then the coefficients
of 𝑨 ( 𝑝, 𝜀) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡) converge in C 𝑘−1 Ω; D 𝑥′ (𝑈) , as 𝜀 ↘ 0, to those of a
( 𝑝 − 1)-current 𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡).

27.5.2 Conditions 𝚷 (𝑽, 𝝃) and (𝚿)

We now reinterpret Condition Π (𝑉, 𝜉) (Definition 27.5.1); as before, 𝑉 ⊂⊂ Ω is a

convex open subset of R𝜈 , 𝜉 ∈ S𝑛−𝜈−1 , 𝑡 (𝜉) ∈ Ω. Recall that

𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) = − (Φ (𝑡, 𝜉) − Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉)) .

Since Ω is convex we have 𝑠 (𝑡, 𝜉, 𝜅) = (1 − 𝜅) 𝑡 (𝜉) +𝜅𝑡 ∈ Ω for all (𝑡, 𝜅) ∈ 𝑉 × [0, 1].
A reminder:
Π (𝑉, 𝜉) We have, for all (𝑡, 𝜅) ∈ 𝑉 × [0, 1],

Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) ≤ Φ (𝑡, 𝜉) . (27.5.6)

Keep in mind that Φ ∈ C 𝜔 (Ω × (R𝑛−𝜈 \ {0}); R).

Proposition 27.5.7 For Π (𝑉, 𝜉) to hold it is necessary and sufficient that
𝜕
∀ (𝑡, 𝜅) ∈ 𝑉 × [0, 1] , Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) ≥ 0. (27.5.7)
𝜕𝜅
27.5 Homotopy Formulas 1187

Proof Sufficiency of the condition. If (27.5.7) holds then, for 𝑡 ∈ 𝑉 arbitrary, the
function
(0, 1) ∋ 𝜅 ↦→ Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) − Φ (𝑡, 𝜉) (27.5.8)
is monotone increasing. Since 𝑠 (𝑡, 𝜉, 1) = 𝑡, (27.5.6) must hold.
Necessity of the condition. We assume that (27.5.6) holds. Let 𝑡 ∈ 𝑉 be arbitrary.
Suppose there exist 𝜅 ∗ ∈ (0, 1), 𝜀 ∈ (0, 𝜅 ∗ ) such that

𝜕
𝜅 ∗ − 𝜀 < 𝜅 < 𝜅 ∗ =⇒ 𝑠 (𝑡, 𝜉, 𝜅) ∈ 𝑉 & Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) < 0. (27.5.9)
𝜕𝜅
We define 𝑡 ∗ = 𝑠 (𝑡, 𝜉, 𝜅 ∗ ); (27.5.9) implies that the function

𝛿 ↦→ Φ (𝑡 (𝜉) + 𝛿𝜅 ∗ (𝑡 ∗ − 𝑡 (𝜉)) , 𝜉)

is strictly decreasing to Φ ((1 − 𝜅 ∗ ) 𝑡 (𝜉) + 𝜅 ∗ 𝑡) = Φ (𝑡 ∗ , 𝜉) as 𝛿 ↗ 1. This is the

same as saying that the function

Φ (𝑡 ∗ , 𝜉) − Φ (𝑡 (𝜉) + 𝛿𝜅 ∗ (𝑡 ∗ − 𝑡 (𝜉)) , 𝜉)

is strictly increasing to zero as 𝛿 ↗ 1. The latter implies, for 𝛿 < 1, 1 − 𝛿 sufficiently

small,
Φ (𝑡 ∗ , 𝜉) − Φ (𝑡 (𝜉) + 𝛿𝜅 ∗ (𝑡 ∗ − 𝑡 (𝜉)) , 𝜉) < 0,
thereby contradicting (27.5.6). We conclude that there are no positive numbers
𝜅 ∗ < 1, 𝜀 < 𝜅 ∗ , such that (27.5.9) holds. □
We may rephrase Proposition 27.5.7 as follows:
Proposition 27.5.8 Let 𝑉 b (𝜉) be the convex hull of 𝑡 (𝜉) ∪ 𝑉. For Π (𝑉, 𝜉) to hold
it is necessary and sufficient that, given any affine line Λ such that 𝑡 (𝜉) ∈ Λ ∩ Ω
and Λ ∩ 𝑉 ≠ ∅, the function Λ ∩ 𝑉 b (𝜉) ∋ 𝑡 ↦→ Φ (𝑡, 𝜉) be quasiconvex (Definition
25.3.7).
Proof Let Λ be as in the statement and let 𝑡 ◦ and 𝑡 ∗ be two points in Λ ∩ 𝜕𝑉 (recall
that 𝑉 is open and convex). We assume that 𝑡 (𝜉) ∈ Λ and we consider two cases:
Case 1: 𝑡 (𝜉) ∉ 𝑉. The condition in Proposition 27.5.7 implies that 𝑡 ↦→ −Φ (𝑡, 𝜉)
is monotone in Λ ∩ 𝑉 b (𝜉).
Case 2: 𝑡 (𝜉) ∈ 𝑉. The condition in Proposition 27.5.7 implies that Φ (𝑡, 𝜉) is
monotone decreasing to Φ (𝑡 (𝜉) , 𝜉) when 𝑡 moves along Λ to 𝑡 (𝜉) from either 𝑡 ◦ or
𝑡∗. □
The condition in Proposition 27.5.8 is that the tangential derivative of the function
𝑡 ↦→ −Φ (𝑡, 𝜉) along every interval Λ ∩ 𝑉 b (𝜉) has Property (𝚿) [cf. Proposition
25.3.10, also (27.2.12)]. The orientation of Λ is immaterial.
Example 27.5.9 We take 𝑷 to consist of a single operator, actually the Mizohata
𝜕 𝜕
vector field 𝜕𝑡 + 𝑖 (𝑘 + 1) 𝑡 𝑘 𝜕𝑥 in the plane (Example 23.1.19). We have Φ (𝑡, 𝜉) =
−𝑡 𝜉; we let 𝑡 vary in 𝑉 = [−1, 1], and let 𝜉 = +1 or −1. Here Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) is
𝑘+1

the function
1188 27 Pseudodifferential Complexes in Tube Structures

𝜅 ↦→ 𝜓 (𝑡, 𝜉, 𝜅) = − 𝑡 𝑘+1 − 𝑠 𝑘+1 (𝑡, 𝜉, 𝜅) 𝜉. (27.5.10)

Recalling that 𝑠 (𝑡, 𝜉, 𝜅) = (1 − 𝜅) 𝑡 (𝜉) + 𝜅𝑡 we have

𝜕𝜓
= (𝑘 + 1) 𝜉 (𝑡 − 𝑡 (𝜉)) 𝑠 𝑘 (𝑡, 𝜉, 𝜅) .
𝜕𝜅
Let 𝑘 be even. If 𝜉 = 1 we select 𝑡 (𝜉) = −1; if 𝜉 = −1 we select 𝑡 (𝜉) = 1; in both
cases (27.5.7) holds for all 𝜅 ∈ R, 𝑡 ∈ [−1, 1]: (27.5.10) is monotone increasing.
Now let 𝑘 be odd. We select 𝑡 (𝜉) = 0, whence
𝜕𝜓
= (𝑘 + 1) 𝜉𝜅 𝑘 𝑡 𝑘+1 .
𝜕𝜅
If 𝜉 = 1 the function (27.5.10) has a minimum at 𝜅 = 0; it is convex; if 𝜉 = −1
(27.5.10) has a maximum at 𝜅 = 0 and therefore it is concave, not quasiconvex. This
last property implies the nonsolvability of the equation 𝑃𝑢 = 𝑓 in any neighborhood
of points (𝑥 ◦ , 0) ∈ R2 when 𝑘 is odd.
Remark 27.5.10 In the entire construction of the homotopy operators in Theorem
27.5.4 we have taken advantage of the convexity of the domains Ω and 𝑉, crucially
when joining 𝑡 to 𝑡 (𝜉) by a straight-line segment. Actually we could have used a
fibration of Ω, say, by continuous curves [0, 1] ∋ 𝜅 ↦→ 𝛾 (𝑡, 𝜉, 𝜅) (with some kind
of regularity with respect to 𝜉). This slightly complicates the argument. It suggests,
however, what the hypothesis in a result such as Proposition 27.5.8 ought to be,
namely that the sublevels of the function 𝑉 ∋ 𝑡 ↦→ Φ (𝑡, 𝜉) be homological trivial
(cf. Proposition 25.3.8, also [Treves, 1976]).

27.5.3 Consequences of Property 𝚷 (𝑽, 𝕮) for the analytic wave-front

sets

We recall the definition of a wedge in C𝑛−𝜈 with edge a domain 𝑈 in R𝑛−𝜈 , height
𝛿 > 0:
W𝛿 (𝑈, Γ) = {𝑧 ∈ C𝑛−𝜈 ; Re 𝑧 ∈ 𝑈, Im 𝑧 ∈ Γ, |Im 𝑧| < 𝛿} ;
Γ is an open cone in R𝑛−𝜈 \ {0}. As before, 𝑉 ⊂ Ω is a convex domain in R𝜈 and ℭ
an open cone in R𝑛−𝜈 \ {0}. We recall an elementary property of Fourier integrals:
Proposition 27.5.11 Let Γ be an open cone in R𝑛−𝜈 \ {0} such that Im 𝑧 ∈ Γ =⇒
𝜉 ·∫Im 𝑧 > 0 for every 𝜉 ∈ ℭ, and let 𝑓 ∈ E ′ (𝑈) be arbitrary. The restriction
of ℭ e𭟋 𝑓 (𝑥, 𝜉) d𝜉 [see (27.4.25)] to 𝑈 × 𝑉 is the distribution boundary value of a
holomorphic function in W𝛿 (𝑈, Γ) depending continuously on 𝑡 ∈ 𝑉.
Proof Let v ∈ Γ ∩ S𝑛−𝜈−1 and 𝜏 ∈ (0, 𝛿); we have
∫ ∫ ∫ ∫
𭟋 𝑓 (𝑥 + 𝑖𝜏v, 𝜉) d𝜉 =
e e𝑖 ( 𝑥−𝑦) · 𝜉 −𝜏v· 𝜉 𝑓 (𝑦) d𝑦d𝜉 = e−𝜏v· 𝜉 b
𝑓 (𝜉) d𝜉.
ℭ ℭ R𝑛−𝜈 ℭ
27.5 Homotopy Formulas 1189

The result follows directly since v · 𝜉 > 0. □

Corollary 27.5.12 Assume the same hypotheses as in Proposition 27.5.11; then

∫
(𝑥 ◦ , −𝜉 ◦ ) ∉ 𝑊 𝐹a ℭ e
𭟋 𝑓 (𝑥, 𝑡, 𝜉) d𝜉 .

Remark 27.5.13 The hypothesis in Proposition 27.5.11 requires that ℭ be contained

in the interior of the polar Γ◦ of Γ, hence that ℭ be acute.

Proposition 27.5.14 Suppose Condition 𝚷 (𝑉, ℭ) is satisfied (Definition 27.5.1) and

let Γ be an open cone in R𝑛−𝜈 \ {0} such that Im 𝑧 ∈ Γ =⇒ 𝜉·Im 𝑧 > 0 for every 𝜉 ∈ ℭ.
Under these hypotheses the following property holds, for every 𝑓 ∈ C (Ω; E ′ (𝑈)):
The restriction of 𝑨 ( 𝑝) (ℭ) 𝑓 [1 ≤ 𝑝 ∈ Z+ , cf. Theorem 27.5.3] to 𝑈 × 𝑉 is
the distribution boundary value of a holomorphic function in W𝛿 (𝑈, Γ) depending
continuously on 𝑡 ∈ 𝑉.

Proof We must stress the fact that, in the notation of this proof, 𝑦 is not related
to 𝑧 (below 𝑦 = Re 𝑤). In accordance with (27.4.2) and (27.4.22) we must look at
the “integral” (actually a duality bracket in 𝑦-space and an oscillatory integral in
𝜉-space):
∫ 1∫ ∫
𝑖 𝜑 (𝑧−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 )−𝜀 | 𝜉 | 2 𝜉
e 𝑎 𝑧 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , , |𝜉 | (27.5.11)
0 ℭ R𝑛−𝜈 |𝜉 |
× 𝑓 (𝑦, 𝑠 (𝑡, 𝜉, 𝜅)) 𝜅 𝑝−1 d𝑦d𝜉d𝜅,

where 𝑠 (𝑡, 𝜉, 𝜅) = (1 − 𝜅) 𝑡 (𝜉) + 𝜅𝑡. We have

𝑖𝜑 (𝑧 − 𝑦, 𝑡, 𝑠, 𝜉) = 𝑖 (𝑧 − 𝑦) · 𝜉 + 𝜋 𝜑 (𝑡, 𝑠, 𝜉)

and 𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) ≤ 0 by Property Π (𝑉, ℭ). Then, if 𝜉 · Im 𝑧 > 0, 𝜉 ∈ ℭ,

0 < |Im 𝑧| < 𝛿, we derive

− Im 𝜑 (𝑧 − 𝑦, 𝑡, 𝑠, 𝜉) ≤ −𝜉 · Im 𝑧 + 𝜋 𝜑 (𝑡, 𝑠, 𝜉) < 0.

It follows that, as 𝜀 ↘ 0, (27.5.11) converges uniformly on compact subsets of

W𝛿 (𝑈, Γ) to a holomorphic function ℎ in W𝛿 (𝑈, Γ) depending continuously on
𝑡 ∈ 𝑉. The hyperfunction boundary value 𝑏𝑈 ℎ is a distribution in 𝑈 (by the estimates
in Subsection 27.4.3). It follows then from Theorem 7.4.15 that 𝑏𝑈 ℎ is a distribution
boundary value of ℎ, i.e, the growth of ℎ at the edge is tempered, meaning that
|ℎ (𝑧)| ≲ |Im 𝑧| −𝑚 for some 𝑚 < +∞, if 𝑧 ∈ W𝛿 (𝑈 ′, Γ ′), 𝑈 ′ ⊂⊂ 𝑈, Γ ′ ∩ S𝑛−𝜈−1 ⊂⊂
Γ. □

Corollary 27.5.15 Suppose that Condition 𝚷 (𝑉, ℭ) is satisfied and let Γ be as

Í
in Theorem 27.5.16. Given arbitrarily 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈
C 1 (Ω; E ′ (R𝑛−𝜈 )) for all 𝐼, |𝐼 | = 𝑝 ≥ 1, the restriction to 𝑈 × 𝑉 of every co-
(
efficient of the 𝑝-form d𝑷 𝑨ℭ𝑝) 𝑓 (𝑥, 𝑡, d𝑡) is the distribution boundary value of a
holomorphic function in W𝛿 (𝑈, Γ) continuous with respect to 𝑡 ∈ 𝑉.
1190 27 Pseudodifferential Complexes in Tube Structures

Proof Combine Propositions 27.5.11, 27.5.14, with Theorem 27.5.4 and the homo-
topy formula (27.5.4). □
In the next statement, when referring to the analytic wave-front set of a function
𝑓 ∈ C 1 Ω; E ′ R𝑛−𝜈 𝑥 we regard 𝑓 as a distribution in 𝑥-space and 𝑡 ∈ Ω as a
parameter, andÍwe denote it by 𝑊 𝐹a ( 𝑓 ). By the analytic wave-front set of a 𝑝-current
𝑓 (𝑥, 𝑡, d𝑡) = Ð 1 ′
|𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C (Ω; E (R
𝑛−𝜈 )) for all 𝐼, |𝐼 | = 𝑝 ≥ 1, we

mean the union |𝐼 |= 𝑝 𝑊 𝐹a ( 𝑓 𝐼 ). We recall Definition 7.4.7: if 𝑢 is the hyperfunction

boundary value of a holomorphic function in W𝛿 (𝑈, Γ) then (𝑥, −𝜉) ∉ 𝑊 𝐹a (𝑢)
whatever 𝑥 ∈ 𝑈, 𝜉 ∈ R𝑛−𝜈 such that 𝜉 · 𝑦 > 0 for every 𝑦 ∈ Γ. Also recall that
pseudodifferential operators decrease the analytic wave-front set of distributions,
and therefore so does d𝑷 . Theorem 27.5.16 has the following direct consequence:

Theorem 27.5.16 Let 𝑥 ◦ ∈ 𝑈, 𝜉 ◦ ∈ ℭ, be arbitrary. If diam 𝑈 and max |𝑡 − 𝑡 (𝜉 ◦ )|

𝑉
are sufficiently small and if Condition 𝚷 (𝑉, ℭ) is satisfied then the following prop-
erties hold, for every 𝑡 ∈ 𝑉:
If 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) then (𝑥 ◦ , −𝜉 ◦ ) does not belong to the analytic wave-front
set of any of the two terms on the right in
∫
1
𭟋 𝑓 (𝑥, 𝑡, 𝜉) d𝜉 = 𝑨 (1) (ℭ) d𝑷 𝑓 (𝑥, 𝑡)
e (27.5.12)
(2𝜋) 𝑛−𝜈 ℭ
∫ ∫
1 𝜉
+ e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 | 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉 .
(2𝜋) 𝑛−𝜈 ℭ R𝑛−𝜈 |𝜉 |

If 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)), 𝑝 ≥ 1, then (𝑥 ◦ , −𝜉 ◦ )

Í
does not belong to the analytic wave-front set of any of the two terms on the right in

1 ∑︁ ∫
𭟋 𝑓 𝐼 (𝑥, 𝑡, 𝜉) d𝜉 d𝑡 𝐼 = d𝑷 𝑨 ( 𝑝) (ℭ) 𝑓 (𝑥, 𝑡, d𝑡)
e (27.5.13)
(2𝜋) 𝑛−𝜈 ℭ
|𝐼 |= 𝑝
( 𝑝+1)
+𝑨 (ℭ) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) .

Notice that, in Theorem 27.5.16, there are no hypotheses about 𝑊 𝐹a ( 𝑓 ) or

𝑊 𝐹a ( 𝑓 𝐼 ) and the conclusions are unchanged if we interpret the operators 𝑃 𝑗 as finite
realizations of the classical series (27.2.1) and the amplitude 𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉, 𝜆) as a
finite realization of (27.3.13).

27.6 Poincaré Lemmas

27.6.1 The effect of local regularity

In this section we assume that 𝚷 (𝑛−𝜈) (𝑉) is satisfied. We will need results on the
effect on 𝑨 ( 𝑝) 𝑓 of the analyticity of 𝑓 with respect to 𝑥 in a neighborhood of 𝑥 ◦ in
R𝑛 . We shall take advantage of (27.4.18)–(27.4.19) and assume that, if |𝜉 | > 𝑅 𝑁 .
27.6 Poincaré Lemmas 1191

𝑎 𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , | 𝜉𝜉 | , 𝜉
𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 )
e
ℎ 𝑁 𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , | 𝜉𝜉 | , 𝜉

≤ 𝐶 𝑁 |𝜉 | −2𝑁 exp 𝜋 𝜑 (𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) .

(27.6.1)

We return to (27.5.2). The integration with respect to 𝜅 is irrelevant for our present
purposes and can be ignored. It is evident that
∫ ∫
𝜉
𝑈 ∋ 𝑥 ↦→ e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝑎 𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , , 𝜉
| 𝜉 | ≤𝑅 𝑁 R𝑛−𝜈 |𝜉 |
× 𝑓 (𝑦, 𝑠 (𝑡, 𝜉, 𝜅)) d𝑦d𝜉

can be extended as a holomorphic function of 𝑧 in a suitably small neighborhood of

𝑥 ◦ in C𝑛−𝜈 . We shall therefore focus on the integral

𝑱 ( 𝑝) (R𝑛−𝜈 ) 𝑓 (𝑥, 𝑡) (27.6.2)

∫ ∫
1 𝜉
= e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝑎 𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , , 𝜉
(2𝜋) 𝑛−𝜈 | 𝜉 |>𝑅 𝑁 R𝑛−𝜈 |𝜉 |
× 𝑓 (𝑦, 𝑠 (𝑡, 𝜉, 𝜅)) d𝑦d𝜉.

We reason under the assumption that 𝑓 ∈ C 1 Ω; 𝐿 c1 (𝑈) .

We take 𝑁 > 𝑛 − 𝜈 in (27.6.2); this and the holomorphic extendibility of 𝜑 and
𝑎 allows us to deform the domain of 𝜉-integration from R𝑛−𝜈 to the image of R𝑛−𝜈
under the map 𝜉 −→ 𝜁 = 𝜉 + 𝑖𝜀 (𝑥 − 𝑦) |𝜉 |, with 𝜀 > 0 suitably small. We need,
however, a certain degree of regularity of the map S𝑛−𝜈−1 ∋ 𝜉 ↦→ 𝑡 (𝜉) = 𝑡 (Re 𝜁).
Our assumption shall be:
(Lip) The map S𝑛−𝜈−1 ∋ 𝜉 ↦→ 𝑡 (𝜉) ∈ Ω is Lipschitz continuous.
If (Lip) holds we can apply Stokes’ Theorem and derive

(2𝜋) 𝑛−𝜈 𝑱 ( 𝑝) (R𝑛−𝜈 ) 𝑓 (𝑥, 𝑡) (27.6.3)

∫ ∫
= e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠,𝜁 ) 𝑎 ♯ (𝑥 − 𝑦, 𝑡, 𝑠, 𝜁) 𝑓 (𝑦, 𝑠) d𝑦d𝜁
| 𝜉 |>𝑅 𝑁 R 𝑛−𝜈
∫ ∫
e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠,𝜁 ) 𝑎 ♯ (𝑥 − 𝑦, 𝑡, 𝑠, 𝜁) 𝜕 𝜉 𝑡 (𝜉) 𝜕𝑡 𝑓 (𝑦, 𝑠) d𝑦d𝜁

+ (1 − 𝜅)
| 𝜉 |>𝑅 𝑁 R𝑛−𝜈
∫ ∫
e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠,𝜁 ) 𝑓 (𝑦, 𝑠) 𝜕 𝜉 𝑡 (𝜉) b (𝑥 − 𝑦, 𝑡, 𝑠, 𝜁) d𝑦d𝜁,

+ (1 − 𝜅)
| 𝜉 |>𝑅 𝑁 R𝑛−𝜈

where
1192 27 Pseudodifferential Complexes in Tube Structures

♯ 𝜁
𝑎 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜁) = 𝑎 𝑥 − 𝑦, 𝑡, 𝑠, ,𝜁 ,
⟨𝜁⟩

b (𝑥 − 𝑦, 𝑡, 𝑠, 𝜁) = e−𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠,𝜁 ) 𝜕𝑠 e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠,𝜁 ) 𝑎 ♯ (𝑥 − 𝑦, 𝑡, 𝑠, 𝜁, ) ;

𝑠 is the abbreviation of 𝑠 (𝑡, 𝜉, 𝜅)

and 𝜕 𝜉 𝑡 (𝜉) is a 𝜈 × (𝑛 − 𝜈) matrix with (real)
entries belonging to 𝐿 ∞ S𝑛−𝜈−1 .
The exponent in the integrand in (27.6.3) is

𝑖𝜑 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜁) = 𝑖 (𝑥 − 𝑦)·𝜉−𝜀 |𝑥 − 𝑦| 2 |𝜉 |−(Φ (𝑡, 𝜁) − Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜁)) .

Using the analyticity and homogeneity of 𝜁 ↦→ Φ (𝑡, 𝜁) [for |Im 𝜁 | ≪ |Re 𝜁 |] we get,
for some 𝐶 > 0 and all (𝑥, 𝑦, 𝑡, 𝜉) ∈ 𝑈 2 × 𝑉 × (R𝑛−𝜈 \ {0}),

|Re (Φ (𝑡, 𝜁) − Φ (𝑡, 𝜉))| ≤ 𝐶𝜀 2 |𝑥 − 𝑦| 2 |𝜉 | .

We derive, after suitably increasing 𝐶 and decreasing 𝜀,

1
− Im 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜁) ≤ − 𝜀 |𝑥 − 𝑦| 2 |𝜉 | + 𝜛𝜑 (𝑡, 𝑠, 𝜉) .
2
If 𝚷 (𝑛−𝜈) (𝑉) holds we reach the conclusion that we have, in (27.6.2),
1
− Im 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜁) ≤ − 𝜀 |𝑥 − 𝑦| 2 |𝜉 | . (27.6.4)
2
This implies directly the following result:
Proposition 27.6.1 Let 𝑓 ∈ C 1 Ω; 𝐿 c1 (𝑈) and suppose that 𝚷 (𝑛−𝜈) (𝑉) and (Lip)

hold. If 𝑓 (𝑥, 𝑡) ≡ 0 in a neighborhood 𝑈 ′ of 𝑥 ◦ in R𝑛 whatever 𝑡 ∈ 𝑉 then 𝑥 ↦→
𝑨 ( 𝑝) 𝑓 (𝑥, 𝑡) extends as a holomorphic function of 𝑧 in a neighborhood of 𝑥 ◦ in C𝑛−𝜈
whatever 𝑡 ∈ 𝑉.
Proof It follows from (27.6.1) and (27.6.4) that the integrand in (27.6.2) decays
faster than exp (−𝑐 |𝜉 |) for some 𝑐 > 0 and all 𝜉 ∈ R𝑛−𝜈 , |𝜉 | > 𝑅 𝑁 , provided
|𝑥 − 𝑦| > 𝑑, 𝑑 > 0 independent of (𝑦, 𝑡) ∈ supp 𝑓 , and |𝑥| is sufficiently small. This
remains true if we allow 𝑥 to vary in a small neighborhood of 𝑥 ◦ in C𝑛−𝜈 . □
Next we modify the hypothesis about 𝑓 . We assume that 𝚷 (𝑛−𝜈) (𝑉) holds and
that, for some 𝛿 > 0 and every 𝑡 ∈ Ω, 𝑥 ↦→ 𝑓 (𝑥, 𝑡) extends as a holomorphic function
of 𝑧 if Re 𝑧 ∈ 𝑈 ′, |Im 𝑧| < 𝛿, with 𝑈 ′ a neighborhood of 𝑥 ◦ in R𝑛 , 𝑈 ′ ⊂⊂ 𝑈. Let
𝑈 ′′ be a neighborhood of 𝑥 ◦ in R𝑛 , 𝑈 ′′ ⊂⊂ 𝑈, with a smooth boundary 𝜕𝑈 ′′ and
0 < 𝜀 < 𝛿. We start from the following decomposition

(2𝜋) 𝑛−𝜈 𝑱 ( 𝑝) (R𝑛−𝜈 ) 𝑓 (𝑥, 𝑡) (27.6.5)

∫ ∫
= e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 ♯ (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) d𝑦d𝜉
| 𝜉 |>𝑅 𝑁 R𝑛−𝜈 \𝑈 ′′
∫ ∫
+ e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 ♯ (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) d𝑦d𝜉.
| 𝜉 |>𝑅 𝑁 𝑈 ′′
27.6 Poincaré Lemmas 1193

The Cauchy Integral Theorem implies

∫
e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠, 𝜉 ) 𝑎 ♯ (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) 𝑓 (𝑦, 𝑠) d𝑦 (27.6.6)
𝑈 ′′
∫
𝑖 𝜑 𝑥−𝑦+𝑖 𝜀 | 𝜉𝜉 | ,𝑡 ,𝑠, 𝜉 ♯ 𝜉 𝜉
= e 𝑎 𝑥 − 𝑦 + 𝑖𝜀 , 𝑡, 𝑠, 𝜉 𝑓 𝑦 + 𝑖𝜀 , 𝑠 d𝑦
𝑈 ′′ |𝜉 | |𝜉 |
∫ ∫ 𝜀
𝜉

𝑖 𝜑 𝑥−𝑦+𝑖 𝜏 | 𝜉 | ,𝑡 ,𝑠, 𝜉 ♯ 𝜉 𝜉
+ e 𝑎 𝑥 − 𝑦 + 𝑖𝜏 , 𝑡, 𝑠, 𝜉 𝑓 𝑦 + 𝑖𝜏 , 𝑠 d𝜎 (𝑦) d𝜏
𝜕𝑈 ′′ 0 |𝜉 | |𝜉 |

with d𝜎 (𝑦) the appropriate measure on S𝑛−𝜈−1 . The first integral in the right-hand
side of (27.6.5) is taken care of by Proposition 27.6.1; moreover, the argument in the
proof of Proposition 27.6.1 enables us to prove, possibly after decreasing 𝜀, that the
integral
∫ ∫ ∫ 𝜀
𝑖 𝜑 𝑥−𝑦+𝑖 𝜏 | 𝜉𝜉 | ,𝑡 ,𝑠, 𝜉 ♯ 𝜉
e 𝑎 𝑥 − 𝑦 + 𝑖𝜏 , 𝑡, 𝑠, 𝜉 (27.6.7)
| 𝜉 |>𝑅 𝑁 𝜕𝑈 ′′ 0 |𝜉 |

𝜉
× 𝑓 𝑦 + 𝑖𝜏 , 𝑠 d𝜎 (𝑦) d𝜏d𝜉
|𝜉 |

extends as a holomorphic function of 𝑧 in a neighborhood of 𝑥 ◦ in R𝑛 whatever

𝑡 ∈ 𝑉.
It remains to deal with the integral
∫ ∫
𝑖 𝜑 𝑥−𝑦+𝑖 𝜀 | 𝜉𝜉 | ,𝑡 ,𝑠, 𝜉 ♯ 𝜉 𝜉
e 𝑎 𝑥 − 𝑦 + 𝑖𝜀 , 𝑡, 𝑠, 𝜉 𝑓 𝑦 + 𝑖𝜀 , 𝑠 d𝑦d𝜉.
| 𝜉 |>𝑅 𝑁 𝑈 ′′ |𝜉 | |𝜉 |

Here we have

𝜉
𝑖𝜑 𝑥 − 𝑦 + 𝑖𝜀 , 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉 = 𝑖 (𝑥 − 𝑦)·𝜉−𝜀 |𝜉 |−(Φ (𝑡, 𝜉) − Φ (𝑠 (𝑡, 𝜉, 𝜅) , 𝜉)) ,
|𝜉 |

whence
− Im 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜉) ≤ −𝜀 |𝜉 | + 𝜛𝜑 (𝑡, 𝑠, 𝜉) .
If 𝚷 (𝑛−𝜈) (𝑉) holds we reach the conclusion that we have, in (27.6.2),

− Im 𝜑 (𝑥 − 𝑦, 𝑡, 𝑠 (𝑡, 𝜉, 𝜅) , 𝜁) ≤ −𝜀 |𝜉 | .

This completes the proof of the following

Proposition 27.6.2 Let 𝑓 ∈ C 1 Ω; 𝐿 c1 (𝑈) and suppose that 𝚷 (𝑛−𝜈) (𝑉) and (Lip)

hold. If for some 𝛿 > 0 and every 𝑡 ∈ Ω, 𝑥 ↦→ 𝑓 (𝑥, 𝑡) extends as a holomorphic
function of 𝑧 if Re 𝑧 ∈ 𝑈 ′, |Im 𝑧| < 𝛿, with 𝑈 ′ a neighborhood of 𝑥 ◦ in R𝑛 , 𝑈 ′ ⊂⊂ 𝑈,
then 𝑥 ↦→ 𝑨 ( 𝑝) 𝑓 (𝑥, 𝑡) extends as a holomorphic function of 𝑧 in a neighborhood of
𝑥 ◦ in C𝑛−𝜈 whatever 𝑡 ∈ 𝑉.
1194 27 Pseudodifferential Complexes in Tube Structures

A direct consequence of Proposition 27.6.2 is that if 𝑎♭ and 𝑎 ♮ are different

finite realizations
and 𝑨
of (27.3.13) ♭( 𝑝)
and 𝑨♮ ( 𝑝) are the corresponding operators
then 𝑥 ↦→ 𝑨♭( 𝑝) − 𝑨♮ ( 𝑝) 𝑓 (𝑥, 𝑡) extends as a holomorphic function of 𝑧 in a
neighborhood 𝑈 ′′ of 𝑥 ◦ in R𝑛 whatever 𝑓 ∈ C 1 Ω; 𝐿 c1 (𝑈) and 𝑡 ∈ 𝑉.

27.6.2 Local Poincaré Lemma modulo analytic errors

We now apply Theorems 27.5.3, 27.5.4, taking ℭ =R𝑛−1 \ {0}. Of course, in this
case the expressions (27.2.1) cannot be regarded as microlocal; they define (formal)
pseudodifferential operators in R𝑛−𝜈 × Ω. Availing ourselves of 27.4.28 we can state
(in terms of formal analytic series)

Theorem 27.6.3 Let 𝑈 be a domain in R𝑛−𝜈 , 𝑉 a convex open subset of R𝜈 , 𝑉 ⊂⊂ Ω.

If 𝚷 (𝑛−𝜈) (𝑉) is satisfied then the following properties hold:
If 𝑓 ∈ C 1 (Ω; E ′ (𝑈)) then we have, in 𝑈 × 𝑉,

𝑓 (𝑥, 𝑡, 𝜉) = 𝑨 (1) d𝑷 𝑓 (𝑥, 𝑡) + (27.6.8)

∫ ∫
1 𝜉
e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 | 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉.
(2𝜋) 𝑛−𝜈 R𝑛−𝜈 R𝑛−𝜈 |𝜉 |

𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)), 𝑝 ≥ 1, then we have, in

Í
If 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝
𝑈 × 𝑉,

𝑓 (𝑥, 𝑡, d𝑡) = d𝑷 𝑨 ( 𝑝) 𝑓 (𝑥, 𝑡, d𝑡) + 𝑨 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) . (27.6.9)

Corollary 27.6.4 Assume the same hypotheses as in Theorem 27.6.3. If 𝑓 (𝑥, 𝑡, d𝑡) =
𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)), 𝑝 ≥ 0, then we have, in 𝑈 × 𝑉,
Í
𝑓
|𝐼 |= 𝑝 𝐼 (𝑥,

d𝑷 𝑨 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) = d𝑷 𝑓 (𝑥, 𝑡, d𝑡) . (27.6.10)

At first sight these statements have severe drawbacks: First of all, 𝑨 ( 𝑝+1) acts on
distributions that are compactly supported with respect to 𝑥. As we have indicated
in introducing the differential complex (27.2.6) we are willing to insert cut-off
functions 𝜒 ∈ Cc∞ (𝑈). But we seek results that ought to be (locally) independent
of the choice of the cut-off 𝜒, after we mod off C 𝜔 terms. That this is so is a
consequence of Proposition 27.6.1. For arbitrary pseudodifferential systems 𝑷 we
can substitute 𝜒 𝑓 for 𝑓 but this entails that we must then deal with composites such
as 𝑨 ( 𝑝+1) 𝜒 [𝑷, 𝜒]. Assuming that 𝜒 ≡ 1 in a domain 𝑈 ′ this makes 𝜒 [𝑷, 𝜒] 𝑓 ∈
C 𝜔 (𝑈 ′). Here Proposition 27.6.2 helps: since 𝜒 [𝑷, 𝜒] 𝑓 ∈ E ′ (𝑈) we can conclude
that 𝑨 ( 𝑝+1) 𝜒 [𝑷, 𝜒] 𝑓 ∈ C 𝜔 (𝑈 ′′) in a domain 𝑈 ′′ ⊂ 𝑈 ′. And of course, the same
reasoning applies if we modify the finite realization of 𝑷 or that of the amplitude 𝑎
(cf. the end of preceding subsection). To formalize these kinds of relations a little
we shall write
27.6 Poincaré Lemmas 1195

d𝑷 𝑨 ( 𝑝) 𝑓 𝑔, resp., 𝑨 ( 𝑝+1) d𝑷 𝑓 ♭ 𝑔♭ , (27.6.11)

𝑈 ′′ ×𝑉 𝑈 ′′ ×𝑉

to mean that the restriction to 𝑈 ′′ of d𝑷 𝜒 𝑨 ( 𝑝) ( 𝜒 𝑓 ) − 𝑔, resp., 𝑨 ( 𝑝+1) d𝑷 𝑓 ♭ − 𝑔♭ ,
belongs to C 𝜔 (𝑈 ′′) for all 𝑡 ∈ 𝑉; and obviously the projections into 𝑥-space of the
supports of 𝑓 and 𝑓 ♭ need not be compact. With this notation we can state:
Theorem 27.6.5 Let 𝑈, 𝑈 ′′ be domains in R𝑛−𝜈 such that 𝑈 ′′ ⊂⊂ 𝑈, 𝑉 a convex
open subset of R𝜈 , 𝑉 ⊂⊂ Ω. If 𝚷 (𝑛−𝜈) (𝑉) and (Lip) are satisfied then the following
properties hold:
If 𝑓 ∈ C 1 (Ω; D ′ (𝑈)) then we have

𝑓 (𝑥, 𝑡, 𝜉) 𝑨 (1) d𝑷 𝑓 (𝑥, 𝑡) + (27.6.12)

𝑈 ′′ ×𝑉
∫ ∫
1 𝜉
e𝑖 𝜑 ( 𝑥−𝑦,𝑡 ,𝑠 (𝑡 , 𝜉 ,𝜅), 𝜉 ) 𝑎 𝑥 − 𝑦, 𝑡, 𝑡 (𝜉) , , |𝜉 | 𝑓 (𝑦, 𝑡 (𝜉)) d𝑦d𝜉.
(2𝜋) 𝑛−𝜈 R 𝑛−𝜈 R 𝑛−𝜈 |𝜉 |

𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C 1 (Ω; D ′ (𝑈)), 𝑝 ≥ 1, then we have

Í
If 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝

𝑓 (𝑥, 𝑡, d𝑡) d𝑷 𝑨 ( 𝑝) 𝑓 (𝑥, 𝑡, d𝑡) + 𝑨 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) . (27.6.13)

𝑈 ′′ ×𝑉

Corollary 27.6.6 Assume the same hypotheses as in Theorem 27.6.3; let 𝑓 (𝑥, 𝑡, d𝑡) =
𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈ C 1 (Ω; E ′ (𝑈)), 𝑝 ≥ 0. If d𝑷 𝑓 (𝑥, 𝑡, d𝑡) ′′ 0 then we have
Í
𝑓
|𝐼 |= 𝑝 𝐼 (𝑥,
𝑈 ×𝑉

d𝑷 𝑨 ( 𝑝+1) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) d𝑷 𝑓 (𝑥, 𝑡, d𝑡) . (27.6.14)

𝑈 ′′ ×𝑉

In (27.6.11), the coefficients of 𝑓 , 𝑓 ♭ , 𝑔, 𝑔♭ may belong to C 𝑘 (Ω; D ′ (𝑈)) for

any 𝑘 = 1, 2, ..., 𝑘 = +∞ or 𝑘 = 𝜔 (among other possible choices; this also applies
to 𝑓 and the coefficients 𝑓 𝐼 in Theorem 27.6.3 and in most of what precedes in this
chapter).
We can also go the other way and weaken the regularity of the coefficients of
𝑓 (𝑥, 𝑡, d𝑡) with respect to 𝑡, to prove the version of the Poincaré Lemma for d𝑷 acting
on currents, i.e., differential forms whose coefficients are distributions in (𝑥, 𝑡)-space.
We must, however, modify the meaning of ′′ to congruent mod C 𝜔 (𝑈 ′′; D ′ (𝑉)).
𝑈 ×𝑉
Theorem 27.6.7 Let the domains 𝑈, 𝑈 ′′, 𝑉 be as in Theorem Í 27.6.5 and suppose
that 𝚷 (𝑛−𝜈) (𝑉) and (Lip) are satisfied. If 𝑓 (𝑥, 𝑡, d𝑡) = |𝐼 |= 𝑝 𝑓 𝐼 (𝑥, 𝑡) d𝑡 𝐼 , 𝑓 𝐼 ∈
D ′ (𝑈 × Ω), 𝑝 ≥ 0, is such that d𝑷 𝑓 (𝑥, 𝑡, d𝑡) = 0 in 𝑈×Ω then there is a 𝑢 (𝑥, 𝑡, d𝑡) ∈
Λ𝑞−1 D ′ (𝑈 × 𝑉) such that d𝑷 𝑢 (𝑥, 𝑡, d𝑡) ′′ 𝑓 (𝑥, 𝑡, d𝑡).
𝑈 ×𝑉
Proof Since every element ofE ′ (𝑈 × Ω) is a finite sum of derivatives of functions
belonging to Cc1 Ω; E ′ R𝑛−𝜈 𝑥 it suffices to prove the claim when each coefficient
𝑓 𝐼 of 𝑓 (𝑥, 𝑡, d𝑥) is equal in 𝑈 ×Ω, to a current 𝜕𝑡𝛼 𝑔 𝐼 , 𝑔 𝐼 ∈ Cc1 Ω; E ′ R𝑛−𝜈

𝑥 ; 𝛼 ∈ Z+𝜈
can be selected independently of 𝐼. Let 𝑘 = |𝛼|; we have, in U,
∑︁
𝛽
𝑓 𝐼 = 𝜕𝑡𝛼 𝑔 𝐼 = 𝑃 𝛼 𝑔 𝐼 + 𝜕𝑡 ℎ 𝐼,𝛽 ,
|𝛽 |<𝑘
1196 27 Pseudodifferential Complexes in Tube Structures

where 𝑃 𝛼 = 𝑃1𝛼1 · · · 𝑃 𝜈𝛼𝜈 , ℎ 𝐼,𝛽 ∈ Cc1 (𝑉; E ′ (𝑈)). Induction on 𝑘 shows that it suffices
to prove the Íclaim when 𝑓 𝐼 = 𝑃 𝛼 𝑔 𝐼 , i.e., 𝑓 (𝑥, 𝑡, d𝑡) = 𝑃 𝛼 𝑔 (𝑥, 𝑡, d𝑡) in U, where
𝑔 (𝑥, 𝑡, d𝑡) = |𝐼 |=𝑞 𝑔 𝐼 (𝑥, 𝑡) d𝑡 𝐼 with coefficients 𝑔 𝐼 in Cc1 (𝑉; E ′ (𝑈)); here 𝑃 𝛼 acts
coefficientwise and obviously commutes with d𝑷 . The system of operators 𝑷 acting
on E ′ (R𝑛−𝜈 × Ω) is injective: indeed, if 𝑷𝑢 = 0 Fourier transform in 𝑥-space yields

𝜕 𝑢ˆ 𝜕Φ
+ (𝑡, 𝜉) 𝑢ˆ = 0, 𝑗 = 1, ..., 𝜈,
𝜕𝑡 𝑗 𝜕𝑡 𝑗

whence 𝑢ˆ (𝜉, 𝑡) = 𝑢ˆ (𝜉, 0) e−Φ(𝑡 , 𝜉 ) · 𝜉 , which is not compactly supported with respect
to 𝑡 ∈ Ω. We deduce that d𝑷 𝑓 = 0 implies d𝑷 𝑔 = 0 in U. At this point we apply
Corollary 27.6.6 to 𝑔. □

27.6.3 Concluding remarks

Remark 27.6.8 When 𝑷 is a system of commuting vector fields in R𝑛 ,

𝑚
𝜕 ∑︁ 𝜕
+𝑖 𝑏 𝑗,𝑘 (𝑡) , 𝑗 = 1, ..., 𝜈, (27.6.15)
𝜕𝑡 𝑗 𝑘=1
𝜕𝑥 𝑘

all the congruences can be replaced by =. In part, this is due to the fact that, when
the 𝑞 𝑗 (𝑡, 𝜉) are linear functions of 𝜉, the amplitudes can be taken to be homogeneous
(with respect to 𝜉) of degree zero (cf. Remark 27.3.3). It is also due the fact that,
when dealing with systems (27.6.15) there is no need to mod off terms analytic in a
full neighborhood of 0 in R𝑛−𝜈 since such error terms can be absorbed by applying
the Cauchy–Kovalevskaya Theorem – even when 𝑏 𝑗,𝑘 (𝑡) ∉ C 𝜔 (Ω) by applying the
result in [Nirenberg, 1972].

Remark 27.6.9 The last sentence in the preceding remark points to an interesting
open question: Is the Poincaré Lemma valid in the differential complex (27.2.6) when
𝑬 = C 𝜔 ? This is related to another interesting question: Is there a scale of Banach
spaces {𝑬 𝑠 }0≤𝑠 ≤1 of holomorphic functions in domains of 𝑧-space C𝑛−𝜈 in which
pseudodifferential operators such as 𝑞 𝑗 (𝑡, D𝑧 ) act analytically in the Ovsyannikov
sense (Ch. 5, Sect. 1)? If that were the case we could prove rather easily, assuming
analyticity with respect to 𝑡, local exactness at every degree, say in the differential
complex (27.2.7) when 𝑬 = C 𝜔 .
Note that, in formulating these last two questions, we tacitly take it for granted that,
under the analyticity hypothesis, no condition of the kind of 𝚷 (𝑛−𝜈) (𝑉) (Definition
27.5.1) or (𝚿) is needed. And indeed, in the cases (such as that of a system of vector
fields) where the amplitudes are not formal series but “true” analytic functions (of
the relevant variables), any misbehavior of the (imaginary part of) the phase-function
in operators such as 𝑨 ( 𝑝) (cf. Theorem 27.5.3) is made irrelevant by the exponential
decay of the integrands ensuing from analyticity, as frequencies −→ ∞.
27.6 Poincaré Lemmas 1197

Remark 27.6.10 It must be strongly underlined that the Poincaré Lemma represented
by Theorem 27.6.3 is a very rudimentary result. Local exactness in all degrees
is a very strong demand. Of course, it is true for the De Rham complex in 𝑡-
space, meaning 𝑞 𝑗 ≡ 0 for all 𝑗, and for the 𝜕 complex; in the latter 𝑛 = 2𝜈 and
𝑞 𝑗 (𝑡, 𝜉) = −𝜉 𝑗 , 𝑗 = 1, ..., 𝜈. But in general, a differential complex such as, say,
(27.2.7), is exact in some degrees and not in others. Here is the simplest example:
Example 27.6.11 Take 𝑛 = 3, 𝜈 = 2, and
𝜕 𝜕
𝑃𝑗 = + (−1) 𝑗 𝑖𝑡 𝑗 , 𝑗 = 1, 2. (27.6.16)
𝜕𝑡 𝑗 𝜕𝑥

There is a single first integral 𝑍 (𝑥, 𝑡) = 𝑥 + 𝑖 𝑡 12 − 𝑡 22 (the system 𝑷 is then said to

have corank one); in this case the phase-function (27.3.2) is

𝜑 (𝑥 − 𝑦, 𝑡, 𝑠, 𝜉) = 𝑥 + 𝑖 𝑡12 − 𝑡22 − 𝑦 − 𝑖 𝑠12 − 𝑠22 𝜉

where 𝜉 ∈ R, 𝜉 ≠ 0, while the function (27.5.8) is

((1 − 𝜅) 𝑡 1 (𝜉) + 𝜅𝑡1 ) 2 𝜉 − ((1 − 𝜅) 𝑡2 (𝜉) + 𝜅𝑡2 ) 2 𝜉 − 𝑡 12 − 𝑡 22 𝜉. (27.6.17)

It is readily checked that, for a generic choice of (𝑡, 𝑡 (𝜉)), (27.6.17) is not quasi-
convex. Yet, if 𝑓 ∈ C ∞ R3 there are functions 𝑢 𝑗 ∈ C ∞ R3 , 𝑗 = 1, 2, satisfying

𝑃1 𝑢 1 + 𝑃2 𝑢 2 = 𝑓 in any geometrically simple domain in R3 . But if 𝑓 𝑗 ∈ C ∞ R3 ,

𝑗 = 1, 2, verify 𝑃1 𝑓2 = 𝑃2 𝑓1 , in general there is no function 𝑢 ∈ C ∞ R3 verifying
𝑃 𝑗 𝑢 = 𝑓 𝑗 , 𝑗 = 1, 2, say in an open ball centered at the origin in R3 . The complex
(27.2.7) is exact in degree 2 but not in degree 1. For the complete description in the
corank 1 case we refer the reader to [Cordaro-Hounie, 2001], and for a survey of the
question, to [Berhanu, Cordaro, Hounie, 2008], VIII.
Remark 27.6.12 Concerns over the length of this chapter have prevented us from
establishing the necessity of Condition 𝚷 (𝑛−𝜈) (𝑉) for the conclusion in Theorems
27.6.3, 27.6.5, to be valid. The author believes that this should not be too difficult
to obtain, considering that it is known to be necessary when 𝜈 = 1, in which case
(Π)=(𝚿), and that local exactness in all degrees puts such a strong demand on the
system 𝑷. That necessity is an open question might perhaps be an incentive to prove
it and then tackle the much finer question of (necessary and sufficient) conditions for
the validity of the Poincaré Lemma for one of the complexes introduced in Subsection
27.1.2, in a given degree regardless of what happens in the other degrees. For an
approach to the latter question we refer the reader to [Treves, 1976].
Remark 27.6.13 The approach followed in Trépreau’s proof1 of Theorem 26.2.12,
suggests an extension to involutive systems 𝑷 = (𝑃1 , ..., 𝑃 𝜈 ) of analytic pseudodif-
ferential operators of principal type since 𝑷 should be transformable into the system
of vector fields (𝜕/𝜕𝑧1 , ..., 𝜕/𝜕𝑧 𝜈 ) at the boundary of a strictly pseudoconvex domain
1 This approach seems to have been suggested by M. Kashiwara c. 1978.
1198 27 Pseudodifferential Complexes in Tube Structures

in C𝑛 . This is feasible at the level of symplectic biholomorphic transformations of the

symbols (cf. Corollary 13.6.22). The differential complex associated to the system
(𝜕/𝜕𝑧1 , ..., 𝜕/𝜕𝑧 𝜈 ) is described in Ch. 25. Necessary and sufficient conditions for
the local exactness at a given degree 𝑑 are not known if 𝜈 ≥ 1 but perhaps could be
found. Once found they would have to be reformulated as conditions on 𝑷, thereby
providing the version of (𝚿) for microlocal exactness in degree 𝑑, in the differential
complexes associated to 𝑷.

It should be mentioned that, when 𝜈 = 𝑛−1, necessary and sufficient conditions for
local exactness have been proved for special classes of systems: in [Cordaro-Hounie,
1990] in each degree, for locally integrable C ∞ systems of vector fields, and in [Han,
1997], for top degree only, for involutive systems of analytic pseudodifferential
operators of principal type.
Finally, and needless to say, the bigger question is whether the methods of this sec-
tion can be somehow extended to involutive systems of pseudodifferential operators
𝑃 𝑗 = D𝑡 𝑗 − 𝑞 𝑗 (𝑥, 𝑡, D 𝑥 ), or even to a single such pseudodifferential operator.
For a different, more algebraic, viewpoint on some of the concepts touched upon
in this chapter we refer the reader to [Schapira, 1985].
References

[Alinhac-Gérard, 1991] Alinhac, S. and Gérard, P., Opérateurs pseudo-différentiels

et théorème de Nash–Moser, InterEditions/Editions du CNRS Paris, 1991.
[Ancona, 1994] Ancona, A., Ombres, Convexité, régularité et sous-harmonicité,
Ark. Mat. 33 (1995), 1–44.
[Arnold, 1989] Arnol’d, V.I., Mathematical Methods of Classical Mechanics, Grad-
uate Texts in Math. 80, 2nd edition, Springer 1989.
[Aroca-Hironaka-Vicente, 2018] Aroka, J., Hironaka, H. and Vincente, J.L., Com-
plex Analytic Desingularization, Springer 2018.
[Atiyah-Singer, 1963] Atiyah, M.F. and Singer, I.M., The index of differential oper-
ators on compact manifolds, Bull. Amer. Math. Soc. 69 (1963), 422–433.
[Atiyah, 1970] Atiyah, M.F., Resolution of Singularities and Division of Distribu-
tions, Comm. Pure Appl. Math. 23 (1970), 145–150.
[Baouendi-Goulaouic, 1972] Baouendi, M.S. and Goulaouic, Ch., Nonanalytic-
hypoellipticity for some degenerate elliptic operators, Bull. Amer. Math. Soc.
78 (1972), 483–486.
[Baouendi-Treves, 1981] Baouendi, M.S. and Treves, F., A property of functions
and distributions annihilated by a locally integrable system of complex vector
fields, Ann. of Math. 113 (1981), 387–421.
[Baouendi-Métivier, 1982] Baouendi, M.S. and Métivier, G., Analytic Vectors of
Hypoelliptic Operators of Principal Type, Amer. J. of Math., Vol. 104, No. 2
(Apr., 1982), 287–319.
[Baouendi, Ebenfeld, Rothschild, 1999] Real Submanifolds in Complex Space and
Their Mappings, Princeton University Press, Princeton N.J. 1999.
[Barone, Boschi-Filho, Farina, 2002] Barone, F.A., Boschi-Filho, H. and Farina,
C., Three methods for calculating the Feynman propagator, Amer. J. Phys. 71
(5/2003), 483–491.
[Beals-Fefferman, 1973] Beals, R. and Fefferman, C., On local solvability of linear
partial differential equations, Ann. of Math. 97 (1973), 482–498.
[Beals-Fefferman, 1974] Beals, R. and Fefferman, C., Spatial inhomogeneous
pseudo-differential operators 1, Comm. Pure Appl. Math. 17 (1974), 1–24.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1199
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3
1200 References

[Berhanu, Cordaro, Hounie, 2008] Berhanu, S., Cordaro, P.D. and Hounie, J., An
introduction to Involutive Structures, Cambridge University Press, Cambridge
U.K., 2008.
[Bernstein-Gelfand, 1969] Bernstein, I.N. and Gelfand, S.I., Meromorphy of the
function P𝜆 (in Russian), Funktsional’nyi Analiz 3 (1969), 84–85.
[Bierstone-Milman, 1988] Bierstone, E. and Milman, P.D., Semianalitic and suban-
alytic sets, Publications mathématiques de l’I.H.E.S. 67 (1988), 5–42.
[Bloom-Graham, 1977] Bloom, T. and Graham, I., On ‘Type’ Conditions for
Generic Real Submanifolds of C𝑛 , Invent. Math. 40 (1977), 217–243.
[Boas, 1954] Boas Jr., R.P., Entire functions, Academic Press, New York, 1954.
[Bogoljubov-Medvedev-Polivanov, 1958] Bogoljubov, N., Medvedev, B.V. and Po-
livanov, M.K., Voprosy teorii dispersionnykh sootnosheniy, Firzmatgiz 1958.
[Bonfiglioli-Fulci, 2012] Bonfiglioli, A., Fulci, R., Topics in Noncommutative Alge-
bra: The Theorem of Campbell, Baker, Hausdorff and Dynkin, Springer (2012).
[Bony, 1969] Bony, J.-M., Une extension du théorème de Holmgren sur l’unicité du
problème de Cauchy, C. R. Acd. des Sci., Paris 268 (1969), 1103–1106.
[Bony, 1969, 2] Bony, J.-M., Principe du maximum, inégalité de Harnack et unicité
du problème de Cauchy pour des opérateurs elliptiques dégénérés, Ann. Inst.
Fourier Grenoble 19 (1969), 277–304.
[Bony, 1976] Bony, J.-M., Equivalence des diverses notions de spectre singulier
analytique, Sémin. Goulaouic-Schwartz 1976–1977, Exposé no. III.
[Bony-Schapira, 1976] Bony, J.-M. and Schapira, P., Propagation des singularités
analytiques pour les solutions des équations aux dérivées partielles, Ann. Inst.
Fourier, Grenoble 26 (1976), 81–140.
[Boutet, 1972] Boutet De Monvel, L., Opérateurs pseudo-différentiels analytiques
et opérateurs d’ordre infini, Ann. Inst. Fourier, Grenoble 22 (1972), 229–268.
[Boutet-Grigis-Helffer, 1976] Boutet De Monvel, L., Grigis, A., and Helffer, B.,
Paramétrixes d’opérateurs pseudo-différentiels à caractéristiques multiples,
Astérisque (1976), 34–36, 93–121.
[Boutet-Guillemin, 1981] Boutet De Monvel, L. and Guillemin, V., The Spectral
Theory of Toeplitz Operators, Annals of Math. Studies Vol. 99, Princeton Uni-
versity Press, Princeton N.J. 1981.
[Boutet-Krée, 1967] Boutet De Monvel, L. and Krée, P., Pseudodifferential opera-
tors and Gevrey classes, Ann. Inst. Fourier, Grenoble 17 (1967), 295–323.
[Boutet-Treves, 1974] Boutet De Monvel, L. and Treves, F., On a Class of Pseu-
dodifferential Operators with Double Characteristics, Invent. Math. 25 (1974),
1–34.
[Bove-Mughetti, 2020] Bove, A., and Mughetti, M., Analytic regularity for solutions
to sums of squares: an assessment. Complex Anal. Synerg. 6 (2020), no. 2, Paper
No. 18, 18 pp.
[Bredon, 1997] Bredon, E.G., Sheaf Theory, Graduate Texts in Math, Springer 1997,
3rd edition.
[Brézis, 1970] Brézis, H., On a characterization of flow-invariant sets, Comm. Pure
Appl. Math. XXIII (1970), 261–263.
References 1201

[Brós-Iagolnitzer, 1973] Brós, J. and Iagolnitzer, D., Causality and local analyticity;
mathematical study, Proceed. Conf. sur la théorie de la renormalisation, Ann.
Inst. Poincaré 18 (1973), 147–184.
[Brós-Iagolnitzer, 1975] Brós, J. and Iagolnitzer, D., Support essentiel et structure
analytique des distributions, Sémin. Goulaouic-Schwartz 1975–1976, Exposé
no. XVIII.
[Carathéodory, 1909] Carathéodory, C., Untersuchungen über die Grundlagen der
Thermodynamik, Math. Ann. 67 (1909), 355–386.
[Chevalley, 1946] Chevalley, Cl., Theory of Lie Groups, Princeton University Press,
Princeton N.J. 1946.
[Coifman-Meyer, 1978] Coifman, R.R. and Meyer, Y., Au-delà des opérateurs
pseudo-différentiels, Astérisque 57, Paris 1978.
[Cordaro-Hounie, 1990] Cordaro, P.D. and Hounie, J.G., On local solvability of
undetermined systems of vector fields, Amer. J. Math. 112 (1990), 243–270.
[Cordaro-Hounie, 2001] Cordaro, P.D. and Hounie, J.G., Local solvability for a
class of differential complexes, Acta Math. 187 (2001), 191–212.
[Cordaro-Trépreau, 1998] Cordaro, P.D. and Trépreau, J.-M., On the solvability of
linear partial differential equations in spaces of hyperfunctions, Ark. Mat. 36
(1998), 41–71.
[Cordaro-Treves, 1994] Cordaro, P.D. and Treves, F., Hyperfunctions on Hypo-
Analytic Manifolds, Annals of Math. Studies 136, Princeton University Press,
Princeton N.J. 1994.
[Cordaro-Treves, 1995] Cordaro, P.D. and Treves, F., Necessary and sufficient con-
ditions for the local solvability in hyperfunctions of a class of systems of complex
vector fields, Invent. Math. 120 (1995), 339–360.
[Conway, 1973] Conway, J.B., Functions of One Complex Variable, Springer 1973.
[D’Angelo, 1979] D’Angelo, J.P., Finite type conditions for real hypersurfaces, J.
Diff. Geom. 14 (1979), 59–66.
[Dencker, 2006] Dencker, Nils, The resolution of the Nirenberg–Treves conjecture,
Ann. of Math. 163 (2006), 405–444.
[De Rham, 1955] de Rham, G., Variétés Différentiables, Hermann Paris 1955.
[Derridj, 1971] Derridj, M., Un problème aux limites pour une classe d’opérateurs
du second ordre elliptiques, Ann. Inst. Fourier Grenoble 21.4 (1971), 99–148.
[De Wilde, 1978] De Wilde, M., Closed graph theorems and webbed spaces, Re-
search Notes in Math. 19, Pitman, London 1978.
[Dickey, 2003] Dickey, L.A., Solitons Equations and Hamiltonian Systems, Ad-
vanced Series in Math. Phys. 26, World Sci. Publishing, Singapore 2003.
[Dieudonné, 1960] Dieudonné, J., Foundations of Modern Analysis, Academic
Press New-York 1960.
[Dolbeault, 1953] Dolbeault, P., Sur la cohomologie des variétés analytiques com-
plexes, C. R. Acad. Sci. Paris 236 (1953), 175–177.
[Duistermaat, 1973] Duistermaat, J.J., Fourier Integral Operators, Lecture Notes,
Courant Institute of Math. Sci, New York 1973.
[Duistermaat-Hörmander, 1972] Duistermaat, J.J. and Hörmander, L., Fourier Inte-
gral Operators, Acta Math. 128 (1972), 183–269.
1202 References

[Egorov, 1969] Egorov, Yu.V., On canonical transformations of pseudodifferential

operators, Uspehi Mat. Nauk. 5 (1969), 235–236.
[Ehrenpreis, 1954] Ehrenpreis, L., Solution of some problems of division I, Amer.
J. Math. 76 (1954), 883–903.
[Ehrenpreis, 1955] Ehrenpreis, L., Mean periodic functions I, Amer. J. Math. 77
(1955), 293–328.
[Eilenberg-Steenrod, 1952] Eilenberg, S. and Steenrod, N., Foundations of Alge-
braic Topology, Princeton University Press, Princeton N.J. 1952.
[Einstein-Podolsky-Rosen, 1935] Einstein, A., Podolsky, B., and Rosen, N., Can
Quantum-Mechanical Description of Physical Reality Be Considered Com-
plete?, Phys. Rev. 47, 777 (15 May 1935).
[Faà di Bruno, 1855] Faà di Bruno, F., “Sullo sviluppo delle funzioni” [On the
development of the functions], Annali di Scienze Matematiche e Fisiche (in
Italian), 6 (1855), 479–480.
[Fantappiè, 1943] Fantappiè, L., Teoría de los functionales analíticos y sus appli-
caciones, Consejo Supr. de Investig. Cientif., Barcelona 1943.
[Feynman, Hibbs, 1965] Feynman, R.P. and Hibbs, A.R., Quantum Mechanics and
Path Integrals, McGraw-Hill, New York 1965.
[Folland-Kohn, 1972] Folland, G.B. and Kohn, J.J., The Neumann Problem for the
Cauchy–Riemann Complex, Ann. of Math. Studies 75, Princeton University
Press, Princeton N.J. 1972.
[Friedman, 1969] Friedman, A., Partial Differential Equations, Holt, Rinehart and
Winston, Inc., New York 1969.
[Gårding, 1953] Gårding, L., Dirichlet problem for linear elliptic partial differential
equations, Math. Scand. 1 (1953), 55–72.
[Gelfand-Dickey, 1978] Gel’fand, I.M. and Dickey, L.A., Jet calculus and nonlinear
Hamiltonian Systems, Funktsional’nyi Analiz i Ego Prilozheniya 12 (1978),
8–23 (also in English translation).
[Geller, 1990] Geller, D., Analytic Pseudodifferential Operators for the Heisen-
berg Group and Local Solvability, Math. Notes 37, Princeton University Press,
Princeton N.J. 1990.
[Godement, 1964] Godement, R., Théorie des faisceaux, Hermann, Paris 1964.
[Goldschmidt, 1986] Goldschmidt, H., An Introduction to overdetermined systems
of partial dirreferential equations, M.S.R.I., Berkeley California 1986.
[Goldschmidt-Spencer, 1976] Goldschmidt, H. and Spencer, D. C., On the non-
linear cohomology of Lie equations. I, II, Acta Math. 136 (1976), 103–239.
[Goldschmidt-Spencer, 1978] Goldschmidt, H. and Spencer, D. C., On the non-
linear cohomology of Lie equations. III, IV, J. Diff. Geom. 13 (1978), 409–526.
[Grauert, 1958] Grauert, H., On Levi’s problem and the imbedding of real-analytic
manifolds, Ann. Math. 68 (1958), 460–478.
[Grothendieck, 1952] Grothendieck, A., Produits tensoriels topologiques et espaces
nucléaires, Memoirs of the Amer. Math. Soc., No. 16, Providence, R.I. 1955.
[Grushin, 1970] Grushin, V.V., On a class of hypoelliptic operators, Mat Sbornik
83 (1970), 456–473; in English, Math. USSR Sbornik 12 (1970), 458–476.
References 1203

[Gunning and Rossi, 1965] Gunning, R.C. and Rossi, H., Analytic Functions of Sev-
eral Complex Variables, Princeton University Press, Princeton N.J. 1965.
[Hadamard, 1923] Hadamard, J.S., Lectures on Cauchy’s problem in linear partial
differential equations, Yale University Press, Oxford University Press 1923,
Reprint Dover Publications, New York 2003.
[Hall, 2015] Hall, Brian C., Lie Groups, Lie Algebras, and Representations. An El-
ementary Introduction, Graduate Texts in Mathematics, 222 (2nd ed.), Springer
2015.
[Han, 1997] Han, Z., Local solvability of analytic pseudodifferential complexes in
top degree, Duke Math. J. 87 (1997), 1–28.
[Hanges, 1981] Hanges, N., Propagation of analyticity along real bicharacteristics,
Duke Math. J. 48 (1981), 269–277.
[Hanges-Sjöstrand, 1982] Hanges, N. and Sjöstrand, J., Propagation of analyticity
for a class of non-microcharacteristic operators, Ann. of Math. 116 (1982), 559–
577.
[Hilbert, 1910] Hilbert D., Grundzüge einer allgemeinen Theorie der linearen Inte-
gralgleichungen (sechste Mitteilung), Nachr. Kgl. Ges. Wiss. Göttingen Math.-
Phys. Kl. (1910) 355–417. (The six communications collected in book form,
Leipzig: Teubner 1912. Also as Chelsea Reprint, New York 1953.)
[Himonas-Misiolek, 2003] Himonas, A.A. and Misiolek, G., Analyticity of the
Cauchy problem for an integrable evolution equation, Math. Ann. 327 (2003),
575–584.
[Hironaka, 1965] Hironaka, H., Resolution of singularities of an algebraic variety
over a field of characteristic zero, I, II, Ann. of Math. 79 (1964), 109–203; ibid.,
205–326.
[Hironaka, 1973] Hironaka, H., Subanalytic sets, In: Number Theory, Algebraic
Geometry and Commutative Algebra, in honor of Y. Akizuki, Kinokuniya, Tokyo
1973, p. 453–493.
[Hochschild, 1965] Hochschild, G., The structure of Lie Groups, Holden-Day San
Fracisco 1965.
[Holmgren, 1901] Holmgren, E.A., Über Systeme von linearen partielle Differen-
tialgleichungen, Öfversigt af Kongl. Vetenskaps-Akad. Förh. 58 (1901), 91–105.
[Hörmander, 1955] Hörmander, L., On the theory of general partial differential
operators, Acta Math. 94 (1955), 161–248.
[Hörmander, 1958] Hörmander, L., On the division of distributions by polynomials,
Ark. Mat. 3 (1958), 555–568.
[Hörmander, 1963] Hörmander, L., Linear Partial Differential Operators, Springer,
Berlin 1963.
[Hõrmander, 1966, 2] Hörmander, L., Pseudo-differential operators and non-elliptic
boundary problems, Ann. of Math. 83 (1966), 129-209.
[Hörmander, 1966] Hörmander, L., An introduction to complex analysis in several
variables, Van Nostrand Princeton, N.J., 1966. Third revised edition, North
Holland Amsterdam 1990.
[Hörmander, 1967] Hörmander, L., Hypoelliptic second-order differential equa-
tions, Acta Math. 119 (1967), 147–171.
1204 References

[Hörmander, 1969] Hörmander, L., Linear Partial Differential Operators, Grund.

Math. Wiss., Springer 1969 (3rd printing).
[Hörmander, 1979] Hörmander, L., The Weyl Calculus of Pseudo-Differential Op-
erators, Comm. Pure and Applied Math. XXXII (1979), 359–443.
[Hörmander, 1983, I] Hörmander, L., The Analysis of Linear Partial Differential
Operators I, Grund. Math. Wiss., Springer 1983.
[Hörmander, 1983, III] Hörmander, L., The Analysis of Linear Partial Differential
Operators III, Grund. Math. Wiss., Springer 1983.
[Hörmander, 1983, IV] Hörmander, L., The Analysis of Linear Partial Differential
Operators IV, Grund. Math. Wiss., Springer 1983.
[Hörmander, 1994] Hörmander, L., Notions of convexity, Birkhäuser 1994.
[Iagolnitzer-Stapp, 1972] Iagolnitzer, D. and Stapp, H.P., Macroscopic Causality
and physical region analyticity in S-matrix theory, Comm. Math. Phys. 14 (1969),
15–55.
[Ishimura, 1978] Ishimura, R., Homorphismes du faisceau des germes de fonctions
holomorphes dans lui-même et opérateurs différentiels, Memoirs of the Faculty
of Sci., Kyushu University 32 (1978), 301–312.
[Ishimura, 1980] Ishimura, R., Théorèmes d’existence et d’approximation pour les
équations aux dérivées partielles linéaires d’ordre infini, Publ. Res. Inst. Math.
Sci. 16 (1980), no. 2, 393–415.
[Jacobowitz-Treves, 1983] Jacobowitz, H. and Treves, F., Nowhere solvable partial
differential equations, Bulletin (New Series) Amer. Math. Soc. 8 (1983), 467–469.
[Johnson, 2002] Johnson, W.P., The Curious History of Faà di Bruno’s Formula,
Amer. Math. Monthly, 109 (March 2002), 217–234.
[Kaneko, 1972] Kaneko, A., Representation of hyperfunctions by measures and
some applications, J. Fac. Sci. Univ. Tokyo Sect. IA, 19 (1972), 321–352.
[Kantor-Schapira, 1978] Kantor, J-M. and Schapira, P., Hyperfonctions associées
aux faisceaux analytiques réels cohérents, An. Acad. Brasil. Ci. 43 (1971), 299–
306.
[Kappeler-Pöschel, 2003] Kappeler, T., Pöschel, J., KdV & KAM, Results in Math.
and Related Areas (3rd Series). A Series of Modern Surveys in Mathematics,
45, Springer, Berlin 2003.
[Kashiwara-Kawai, 1970] Kashiwara, M. and Kawai, K., Pseudo-
differentialoperators in the theory of hyperfunctions, Proc. Japan Acad.
46 (1970), 1130–1134.
[Kashiwara, 1971] Kashiwara, M., On the flabbiness of the sheaf C and the Radon
transformation, Sûrikaiseki-kenkyûsho Kôkyûroku 114 (1971), 1–4, in Japanese.
[Kelley, 1955] Kelley, L., General Topology, D. Van Nostrand, Princeton, N.J.,
1955.
[Kohn, 1972] Kohn, J.J., Boundary behavior of 𝜕 on weakly pseudoconvex mani-
folds of dimension two, J. Diff. Geom. 6 (1972), 523–542.
[Kohn-Nirenberg, 1965] Kohn, J.J. and Nirenberg, L., Noncoercive boundary values
problems, Comm. Pure Appl. Math. 18 (1965), 443–492.
[Kollar, 2007] Kollar, J., Lectures on Resolution of Singularities (AM-166), Annals
of Math. Studies, Princeton University Press, Princeton N.J. 2007.
References 1205

[Komatsu, 1973] Komatsu, H., Ultradistributions, I: Structure theorems and a char-

acterization, J. Fac. Sci. Univ. Tokyo, Sect. IA, 20 (1973), 25–105.
[Komatsu, 1992] Komatsu, H., An elementary theory of hyperfunctions and micro-
functions, Banach Center Publ. 27, Institute Math. Polish Acad. Sci., Warsaw
1992.
[Korobeinik, 1969] Korobeinik, J.F., The existence of an analytic solution of an
infinite order differential equation and the nature of its domain of analyticity,
Mat. Sbornik 80 (1969), 53–71.
[Köthe, 1979] Köthe, G., Topological vector spaces II, Grund. Math. Wiss.,
Springer, Berlin 1979.
[Kupershmidt 2000] Kupershmidt, B.A., Noncommutative Mathematics of La-
grangian, Hamiltonian and Integrable Systems, Math. Surveys and Monographs,
Amer. Math. Soc. 78, Providence R.I. 2000.
[Lax, 1968] Lax, P.D., Integrals of nonlinear equations of evolution and solitary
waves, Comm. Pure Appl. Math. 21 (1968), 467–490.
[Leray, 1981] Leray, J., The meaning of Maslov’s asymptotic method: the need of
Planck’s constant in mathematics, Bull. Amer. Math. Soc. 5 (1981), 15–27.
[Lerner, 1988] Lerner, N., Sufficiency of condition (𝜓) for local solvability in two
dimensions, Ann. of Math. 128 (1988), 243–258.
[Lerner, 1994] Lerner, N., Nonsolvability in 𝐿 2 for a first-order operator satisfying
condition (𝜓), Ann. of Math. 139 (1994), 363–393.
[Lerner, 2010] Lerner, N., Metrics on the Phase-Space and Non-Selfadjoint Pseu-
doDifferential Operators, Birkhäuser Verlag, Basel-Boston-Berlin 2010.
[Levi, 1907] Levi E.E., Sulle equazioni lineari totalmente ellittiche alle derivate
parziali, Rend. Circ. Mat. Palermo 24, (1907), 275–317 (Also with corrections
in: Eugenio Elia Levi, Opere, Vol. II, 28–84. Roma: Edizioni Cremonese 1960.)
[Levin, 1964] Levin, B., Distributions of zeros of entire functions, Amer. Math. Soc.
Translations. Providence R.I. 1964.
[Lewy, 1956] Lewy, H., On the local character of the solution of an atypical linear
differential equation in three variables and a related theorem for regular functions
of two complex variables, Ann. of Math. 64 (1956), 514–522.
[Lions-Magenes, 1968] Lions, J.-L. and Magenes, E., Problèmes aux Limites Non
Homogènes et Applications, 3 vol. Dunod Paris 1968. English translation:
Grund. Math. Wiss., Band 181, 182, Springer Verlag Berlin-Heidelberg-New
York 1972.
[Lojasiewicz, 1959] Lojasiewicz, S., Sur le problème de la division, Studia Math.
XVIII (1959), 87–136.
[Lojasiewicz, 1964] Lojasiewicz, S., Triangulation of semianalytic sets, Ann. Scuola
Norm. Sup. Pisa 18 (1964), 449–474.
[Malgrange, 1955] Malgrange, B., Existence et approximation des solutions des
équations aux dérivées partielles et des équations de convolution, Ann. Inst.
Fourier Grenoble 6 (1955–1956), 271–355.
[Martineau, 1960] Martineau, A., Fonctions analytiques et distributions; support
des fonctionelles analytiques, Séminaire Schwartz, 4e année (1959/60) n𝑜 19.
1206 References

[Martineau, 1961] Martineau, A., Les hyperfonctions de M. Sato, Séminaire Bour-

baki, 13e année (1960/61), n𝑜 214.
[Martineau, 1964] Martineau, A., Distributions et valeurs au bord des fonctions
holomorphes, In: Proceed. Intern. Summer Institute Lisbon 1964, 195–326.
[Martineau, 1967] Martineau, A., Équations différentielles d’ordre infini, Bull. Soc.
Math. France 95 (1967), 109–154.
[Martineau, 1967/68] Martineau, A., Théorèmes sur le prolongement analytique du
type “Edge of the Wedge Theorem”, Séminaire Bourbaki, 20e année, 1967/68.
n◦ 340.
[Martineau, 1968] Martineau, A., Sur la notion d’ensemble fortement linéellement
convexe, Ann. Acad. Brasil. Ciênc. 40 (1968) 427–435.
[Martineau, 1970] Martineau, A., Le “edge of the wedge theorem” en théorie des
hyperfonctions de Sato, In: Proceed. Intern. Confer. Anal. and Related Topics,
Univ. of Tokyo Press, Tokyo 1970, 95–106.
[Maslov, 1965] Maslov, V.P., Theory of perturbations and asymptotics methods,
Moskov. Gos. Univ., Moskw 1965 (Russian). French translation in Études Math-
ématiques, Dunod, Paris 1972.
[Métivier, 1980] Métivier, G., Analytic hypoellipticity for operators with multiple
characteristics, Comm. in PDE 6 (1980), 1–90.
2
2 2
2 Métivier, G., Non-hypoellipticité Analytique pour 𝐷 𝑥 +
[Métivier, 1981]
𝑥 + 𝑦 𝐷 𝑦 , C. R. Acad. Sci. Paris Sér. I Math. 292 (1981), no. 7, 401–404.
[Métivier, 1985] Métivier, G., Uniqueness and approximation of solutions of first-
order non-linear equations, Invent. Math. 82 (1985), 263–282.
[Mizohata, 1962] Mizohata, S., Solutions nulles et solutions analytiques, J. Math.
Kyoto Univ. 1 (1962), 271–302.
[Morimoto, 1970] Morimoto, M., Sur la décomposition du faisceau des germs de
singularités d’hyperfonctions, J. Fac. Sci. University Tokyo 15 (1970), 215-239.
[Morimoto, 1973] Morimoto, M., Edge of the wedge theorem and hyperfunction,
In: [Sato-Kawai-Kashiwara, 1973], 41–86.
[Moyer, 1978] Moyer, R., Local solvability in two dimensions: Necessary conditions
for the principal-type case, mimeograph ms, University of Kansas 1978.
[Nagano, 1966] Nagano, T., Linear differential systems with singularities and ap-
plications to transitive Lie algebras, J. Math. Soc. Japan 18 (1966), 398–404.
[Nirenberg, 1971] Nirenberg L., A proof of the Malgrange preparation theorem, In:
Proceed. Liverpool Singularities, Symposium I, Lecture Notes in Mathematics,
192 (1971), Springer, Berlin, Heidelberg, 97–105.
[Nirenberg, 1972] Nirenberg, L., An abstract form of the nonlinear Cauchy–
Kowalewski theorem, J. Differential Geometry 6 (1972), 561–576.
[Nirenberg, 1974] Nirenberg, L., On a question of Hans Lewy, Russian Math. Sur-
veys 29 (1974), 251–262.
[Nirenberg-Treves, 1963] Nirenberg, L. and Treves, F., Solvability of a first order
linear partial differential equation, Comm. Pure Appl. Math. XVI (1963), 331–
351.
References 1207

[Nirenberg-Treves, 1970] Nirenberg, L. and Treves, F., On local solvability of linear

partial differential equations. I. Necessary conditions. II. Sufficient conditions.
Correction, Comm. Pure Appl. Math. XXIII (1970), 1–38 and 459–509, XXIV
(1971), 279–288.
[Okaji, 1988] Okaji, T., Gevrey-hypoelliptic operators which are not 𝐶 ∞ hypoellip-
tic, J. Math. Kyoto Uni. 28 (1988), 311–322.
[Oleinik, 1973] Oleinik, O., On the analyticity of solutions of partial differential
equations and systems, Astérisque 2, 3 (1973), 272–285.
[Osgood, 1929] Osgood, W.P., Lehrbuch der Funktionentheorie, II, 1, Leipsig 1929.
[Ovsyannikov, 1965] Ovsyannikov, L.V., Singular operators in Banach spaces scales
(in Russian), Doklady Acad. Nauk. 163 (1965), 819–822.
[Peetre, 1960] Peetre, J., Rectification à l’article “Une caractérisation abstraite des
opérateurs différentiels”, Math. Scand. 8 (1960), 116–120.
[Petrowsky, 1939] Petrowsky, I.G., Sur l’analyticité des solutions des systèmes
d’équations différentielles, Mat. Sb. 5 (1939), 3–70.
[Sato, 1960] Sato, M., Theory of hyperfunctions, J. Fac. Sci., Univ. Tokyo, Part I in
I (1959), 139–193; Part II in I (1960), 387–437.
[Sato, 1969] Sato, M., Hyperfunctions and partial differential equations, Proceed.
Intern. Confer. Anal. and Related Topics, Univ. of Tokyo Press, Tokyo 1970,
91-94.
[Sato, 1970] Sato, M., Regularity of hyperfunction solutions of partial differential
equations, Actes International Math. Congress 1970, Vol. 2, 785–794.
[Sato-Kawai-Kashiwara, 1973] Sato, M., Kawai, T. and Kashiwara, M., Micro-
functions and pseudo-differential equations, In: Hyperfunctions and pseudo-
differential equations, Springer Lecture Notes in Math. 287 (1973), 265–529.
[Schapira, 1970] Schapira, P., Théorie des hyperfonctions, Lecture Notes in Math.
126, Springer, Berlin 1970.
[Schapira, 1981] Schapira, P., Conditions de positivité dans une variété symplec-
tique complex. Applications à l’étude des microfunctions, Ann. Sci. École Norm.
Sup. (4), 14 (1981), 121–139.
[Schapira, 1985] Schapira, P., Microdifferential Systems in the Complex Domain,
Grund. Math. Wiss., Springer 1985.
[Schwartz, 1955] Schwartz, L., Division par une fonction holomorphe sur une var-
iété analytique complexe, Summa Brasiliensis Mathematicae, 3 (1955), reprinted
in Oeuvres Scientifiques (II), Documents Mathématiques, Soc. Math. France,
Paris 2011, 13–41.
[Schwartz, 1966] Schwartz, L., Théorie des distributions, Hermann Paris 1966.
[Shubin, 1978] Shubin, M.A., Pseudodifferential Operators and Spectral Theory,
Springer Series in Soviet Mathematics, Springer, Berlin 1987.
[Sigal-Soffer, 1987] Sigal, I.M. and Soffer, A., The N-particle scattering problem:
asymptotic completeness for short range systems, Ann. of Math. 126 (1987),
35–108.
[Sjöstrand, 1982] Sjöstrand, J., Singularités analytiques microlocales, Astérisque 95
(1982), Soc. Math. France.
1208 References

[Spivak, 1979] Spivak, M. A comprehensive introduction to Differential Geometry

Vol. I, Publish or Perish, Inc. Wilmington De. 1979.
[Stein, 1993] Stein, E.M. , Harmonic Analysis, Princeton Math. Series 43, Princeton
University Press, Princeton N.J. 1993.
[Stein-Weiss, 1971] Stein, E.M. and Weiss, G., Introduction to Fourier Analysis on
Euclidean Spaces, Princeton University Press, Princeton N.J. 1971.
[Stein-Shakarchi, 2011] Stein, E.M. and Shakarchi, R., Functional Analysis, Prince-
ton Lectures in Analysis, Princeton University Press, Princeton N.J. 2011.
[Sussmann, 1973] Sussmann, H.J., Orbits of families of vector fields and integra-
bility of distributions, Trans. Amer. Mat. Soc. 180 (1973), 171–188.
[Tartakoff, 1980] Tartakoff, D.S., On the local real-analyticity of solutions to □𝑏
and the 𝜕-Neumann problem, Acta Math. 145 (1980), 117–204.
[Taylor, 1981] Taylor, M.E., Pseudodifferential Operators, Princeton University
Press, Princeton N.J. 1981.
[Trépreau, 1984] Trépreau, J.-M., Sur la résolubilité analytique microlocale des
opérateurs pseudodifférentiels de type principal, Thèse, Université de Rennes
1984.
[Trépreau, 1992] Trépreau, J.-M., On the minimum of subharmonic functions with
respect to a real parameter, Preprint 1992.
[Treves, 1966] Treves, F., Linear Partial Differential Equations with Constant Co-
efficients, Gordon & Breach, New York 1966.
[Treves, 1967] Treves, F., Topological Vector Spaces, Distributions and Kernels,
Academic Press New York 1967, also Dover Publications, Mineola N.Y. 2006.
[Treves, 1970] Treves, F., An abstract nonlinear Cauchy–Kowalevska theorem,
Trans. A. M. S. 150 (1970), 77–92.
[Treves, 1971] Treves, F., Hypoelliptic Partial Differential Equations of Principal
Type with Analytic Coefficients, Comm. Pure Appl. Math. XXIII (1970), 637–
651.
[Treves, 1970, 2] Treves, F., A New Method of Proof of the Subelliptic Estimates,
Comm. Pure Appl. Math. XXIV (1971), 71–115.
[Treves, 1971, 2] Treves, F., Analytic-Hypoelliptic Partial Differential Equations of
Principal Type, Comm. Pure Appl. Math. XXIV (1971), 537–570.
[Treves, 1973] Treves, F., Concatenations of Second-Order Evolution Equations
Applied to Local Solvability and Hypoellpticity, Comm. Pure Appl. Math. XXVI
(1973), 201–250.
[Treves, 1975] Treves, F., Basic Linear Partial Differential Equations, Academic
Press New York 1975, also Dover Publications, Mineola N.Y. 2006.
[Treves, 1976] Treves, F., Study of a model in the theory of complexes of pseudod-
ifferential operators, Ann. of Math. 104 (1976), 269–324.
[Treves, 1978] Treves, F., Analytic hypo-ellipticity of a class of pseudodifferen-
tial operators with double characteristics and applications to the 𝜕-Neumann
problem, Comm. in PDE 3 (1978), 476–642.
[Treves, 1980] Treves, F., Introduction to Pseudodifferential and Fourier Integral
Operators, Volumes I and II, Plenum New York 1980.
References 1209

[Treves, 1983] Treves, F., On the local solvability and the local integrability of
systems of vector fields, Acta Math (1983), 1–38.
[Treves, 1992] Treves, F., Hypo-Analytic Structures. Local Theory, Princeton Uni-
versity Press, Princeton N.J. 1992.
[Treves, 1995] Treves, F., Parametrices for a Class of Schrödinger Equations, Comm.
Pure Appl. Math. XLVIII (1995), 13–78.
[Valiron, 1949] Valiron, G., Lectures on the general theory of integral functions,
Chelsea Publishing Company, New York, 1949.
[Van der Waerden, 1953] Van der Waerden, B.L., Modern Algebra, Frederick Ungar
Publishing Co. New York 1953.
[Wallach, 1977] Wallach, N.R., Symplectic Geometry and Fourier Analysis, Math
Sci Press Brookline Mass. 1977.
[Weyl, 1940] Hermann Weyl, The method of orthogonal projections in potential
theory, Duke Math. J. 7 (1940), 411–444.
Notation Index

(A1), 995 A (𝑈 × Θ), 217

⊤ , 29 ♮
𝐴1,2 𝑨𝑈,𝑦 ◦ ℎ, 905
[𝑎 1 , 𝑎 2 ] # , 589 Ae(𝑚) (𝔓), A eform (𝔓), 809
form
𝑎 1 #𝑎 2 , 771 𝑎 𝑅 (𝑧, 𝑤, 𝜁), 620
e
𝑎 1 #𝑎 2 , 588 e−1◦ ◦ ◦
e ( 𝑥 ◦ , 𝜉 ◦ ), ( 𝑥 ◦ , 𝜃 ◦ ) , A
A , 930
( 𝑥 , 𝜃 ), ( 𝑥 , 𝜉 ◦ )
(A2), 998 𝑎 𝑤 (𝑥, D), 603
𝑨𝑏𝑈 ℎ (𝑥, 𝜆), 913 A 𝑥 ◦ , 𝜃 ◦ , 924, 925
𝑨𝔠 𝑓 (𝜃, 𝜆), 𝑨ℭ 𝑓 (𝑧, 𝜆), 825 𝑨 𝑥 ◦ , 𝜃 ◦ , 𝑨−1 𝜃 ◦ , 𝑥 ◦ , 920
𝑨ℭ,Ω′ ℎ, 672 micro
A ( 𝑥 ◦ , 𝜃 ◦ ) , Amicro , 217
𝑨⊤ ℭ,Ω′ 𝜇, 672 A 𝑥 ◦ (𝜁), 911
𝑨 ( 𝜀) , 1151
𝐴 𝜀 (𝑥, 𝑦), 692 [B, B] ⊂ B, 355
Aform 𝑬, 772 𝛽 ⪯ 𝛼, 𝛽 ≺ 𝛼, 4
Aform (Ω), Aform (Ω), A eform (Ω), 𝑆 ⊥𝛽 , 𝑾 ⊥𝛽 , 377
765 B| 𝐸 , 266
micro (𝑈 × Σ), Bmicro (R𝑛 ), 776
Bform
Aform (R𝑛 ), 774 form
sing sing
𝔞#𝔟, 810 Bform (𝑈), Bform (R𝑛 ), 776
AΓ [f] 𝑥 ◦ , A [f] ( 𝑥 ◦ , 𝜉 ◦ ) , 914 Bform (𝑈), Bform (R𝑛 ), 775
𝛼!, 3 B micro , 349
|𝛼|, 3 Bmicro (M), 350
Amicro (𝑆 ∗ M), Amicro Bmicro (R𝑛 ), 218
𝑥 ◦ , 𝜃 ◦ , 350
𝑨𝜇 (𝜃, 𝜆), 862 B micro (𝑈 × Θ), 218
𝑨𝜇 (𝑧, 𝜃, 𝜆), 907 B (Ω) , B (Ω), 165
𝑨 ( 𝑝) (ℭ), 1186 𝑏 Ω ℎ, 48, 183
𝑨 𝜑[𝑁 ] , 𝑹 𝜑[𝑁 ] , 1011 B (R𝑛 ), 170
𝑎 ⊤ , 𝑎 ∗ , 569 B sing (Ω), 179
Bsing (R𝑛 ), 179
A (U), 459
𝔅𝑠,𝑅 (𝑢 ◦ ), 𝔅𝑅 (𝑢 ◦ ), 112
𝑨𝑈 𝜇 (𝜃, 𝜆), 908
Cb∞ (Ω), 515
♮
𝑨𝑈 ℎ, 907

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1211
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3
1212 Notation Index

Cc∞ (𝐾),
Cc∞ (Ω),
6 𝔇𝔦𝔣𝔣∞ class (Ω), 776
Cc∞ Ω, |d𝑥| 𝜆 , 330 𭟋1 , 51
𭟋 𝑓 , 1176
C◦∞ (S𝑛 ), 547
𭟋 𝜅 , 54, 185
c𝑬 , 431
𭟋𝑛 , 380
Char (2) 𝑃, 600
𭟋, 1182
e
Char∞ 𝑃, Charfin 𝑃, 950
dimK 𝑽, 467
Char𝑃, 15, 599, 684 ′ (S𝑛 ), 547
D∞
Char𝑷, 1161, 1170 dist ( 𝐴, 𝐵), 39
C C𝑛 \𝐾; 𝑇 (𝑛,𝑛−1) , 155
∞
div, 358
C ∞ (Ω), 6 D 𝐿′ 1 (R𝑛 ), 51
C 𝑘 ,19 (D 𝜔 ), 776
C 𝑘 Ω, |d𝑥| 𝜆 , 329 d𝑷 , 1165, 1167
D ′ (M; CΛ 𝑝 𝑇 ∗ M), 331
C 𝑘 (U; 𝑇M), 273
′
D M, |d𝑥| 𝜆 , 330
𝑚
𝑆class,form (Ω), 𝑆class,form (Ω), 596
−∞
𝑆class (Ω), 595 D ′ (Ω), 17, 18
𝑞
ℭlift (U, G), Zlift 𝑞
(U, G), D ′ (Ω1 × Ω2 ), 28
𝑞
Blift (U, G), 𝐻lift𝑞
(U, G), 305 D′ (R𝑛 ), 171
𝐶 ,6
𝜔 D ′ (R𝑛 ; H), 𝐻 𝑠 (R𝑛 ; H),
C 𝜔 (R𝑛 ), 178 𝐻c𝑠 (R𝑛 ; H), 𝐻loc
𝑠 (R𝑛
; H), 1021
Constloc (M; C), 288 D𝑥 , D𝑥 , 4
𝛼

CP𝑛 , Can CP𝑛 , 1071 d𝜉 ∧ d𝑥, 399

ℭ 𝑞 (U; F), 300 |d𝑥|, |d𝑥| 𝜆 , 329
ℭ𝜌 (𝜉 ◦ ), 64 d𝜁 ∧ d𝑧, 441
Crit 𝑋, 361 𝑑𝜁 ∧ d𝑧, 443
Cslow (Ω, 𝑺), 515 𝐷 (𝑧 ′), 484, 485
∞ (Ω), 515
Ctemp
(𝑞)
C {𝑧1 , ..., 𝑧 𝑛 }, 475 𝔈clos (Ω, 𝜅, 𝑀), 556
E 𝑗 (𝑧, d𝑧, d𝑧¯), 333
𝛿∗ , 305 E ′, E ′ ( Ω), 18
Δ 𝜅 (𝑧, 𝜉), 53 𝑬 𝑄 , 435
𝛿Λ , 533 Exp1,0 (C𝑛 ), 222
𝛿𝑃 exp 𝑎 (𝑥, D), 653
𝛿𝑢 , 𝐼 𝑃 (𝑢), 805
𝛿𝑃 Exp (C𝑛 ), 146
𝛿𝑢 𝑗 , 806
exp D 𝑦 ·𝜕 𝜉 , 584
𝛿 (𝑞) , 300 ∗
exp 𝑡𝐻 𝑓 , 401
Δ𝑟𝑛 , 120 (exp 𝑡 𝑋) ∗ 𝑓 , (exp 𝑡 𝑋) ∗ 𝑌 , 363
Δ𝑅(𝑛) (𝑧◦ ), 8 exp 𝑡 𝑋, Φ𝑋 (𝑡), 362
Δ𝑟𝑛 (𝑧 ◦ ), 471
Δ𝑥 , 7 (FA), 765
′ (Ω), 528
Dext (FA𝔓), 809
𝔡𝑃, 802 𝐹∗ 𝑢, 329
(𝑞)
d 𝑓 𝑋0 , 𝑋1 , ..., 𝑋𝑞 , 354 𝔉clos Ω𝑞 , 𝜅, 𝑀 , 𝔉clos (Ω, 𝜅, 𝑀), 556
diag Ω × Ω, 31 { 𝑓 , 𝑔}, 401
Diff ∞a (Ω), 781 𝐹 # (𝑎), 652
Notation Index 1213
√
𝑓 (𝑥, d𝑥), 285 ℑ, 478
f𝑧 ◦ , 1116 I𝑧 ◦ , 842
F𝑧 ◦ ,𝜁 ◦ , 878
𝑓 (𝑧, d𝑧¯), 292 |𝐽 | 𝜀 , 251
𝐹 (𝑧, d𝑧, d𝑧¯), 297 |𝐽 | 𝜀 , 249
F∗𝜁 ◦ ,𝑧 ◦ , 879 𝐽 (𝜕𝑧 ), 223
F 𝑍 , 𝐻 𝑍𝑞 (𝑋; F), 313 𝐽 (𝑧, 𝜕𝑧 ), 𝐽 (𝑧, 𝜁), 221
(
Γ 𝜀 (𝜃 ◦ ), 213 𝑲 ◦ 𝑝) , 1175
𝑲 ( 𝑝) , 1176
𝚪𝛾(𝑛) , 185 (𝑞) (𝑞)
KΩ , K𝜕Ω , 334
Γ ( 𝑿, F), 13
𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ), 𝔑𝔞𝔤 (U, 𝐻 𝐴, 𝐻 𝐵 ), ΛC𝑇 ∗ M, Λ𝑞 C𝑇 ∗ M, 285
946 Λ 𝑘 D′( 𝑥,𝑡 , 𝜉 , 𝜏) , 1165
𝔤 (𝔉), 𝑽 (𝔉), 411 Λ 𝜑 , 699
𝐺 𝑅 (𝑥), 𝑔 𝑅 (D), 659 Λ 𝜓 ◦ Λ 𝜑 , 732
𝑚 (Ω; 𝔓 [ℎ]),
G𝑆class Λ𝑞 𝑨 𝛿 (𝑈, Γ1 , ..., Γ𝜈 ), 322
G𝑆class (Ω; 𝔓 [ℎ]), 814 𝑳 (CB1 ; CB2 ), 𝜋 ∗ 𝑳 (CB1 ; CB2 ),
𝐺 𝑠 (Ω), 9 281
𝔤 (𝑋1 , ..., 𝑋𝑟 ), 353 (Lip), 1192
𝐿 𝑢𝑗 , 420
𝐻 (0,𝑞) (M), 292
𝐻◦𝑚 ( Ω), 𝐻 −𝑚 ( Ω), 25 𝔏 𝑘 (𝑬 𝑠 ; 𝑬 𝑠′ ), 113
𝐻c𝑠 (M), 𝐻loc 𝑠 (M),
964 L𝜇, 146
𝐻c (Ω), 26
𝑠 loc ℑ, 477
Hess 𝑓 , HessC (𝑄), 342 𝐿 ( 𝜑) (𝑧, D, 𝜆), 783
𝐻 𝑓 , 400 𝑳 𝜌(𝑧) (𝑤), 340
L 𝑥 ◦ , 782
𝐻 R𝑓 , 𝐻 I𝑓 , { 𝑓 , 𝑔}R , { 𝑓 , 𝑔}I , 441
L 𝑥 ◦ (𝜃), 800, 802
𝐻 𝑗 , 1170
L 𝑥 ◦ , 𝜃 ◦ , 783
𝐻𝐾 (𝜁), 239
𝑠 (Ω), L 𝜁 ◦ (𝑧, D𝑧 − 𝜆𝜁 ◦ , 𝜆), 891
𝐻loc 26
𝐻 𝑚 (Ω), 17 microsupp 𝐴, 592
𝐻 𝑚 ( Ω), 25 microsupp 𝑓 , 217
H𝑚 (𝑥), 103 M𝑚 (K), GL (𝑚, K), 382
H𝑛 [M], 316, 350 𝔐𝑛𝑘 , 475
𝐻 (𝑞) (M), 288 (M, 𝜛), 398
𝐻 𝑞 (𝑋, 𝑌 ; F), 313 Mp (𝑛, R), 388, 395
H1rel R𝑛 × S𝑛−1 , 322

𝔪𝔭 (𝑛, R), 395
𝐻 𝑠 (𝐾), 25
ℎ,v , 180
𝜇𝑈
𝐻 𝑠 (R𝑛 ), 17, 24
Hyperdiff (𝐾), 228 𝑁 ∗ L, 274
Hyperdiff (Ω), 229 ∗ 𝑭, 451
𝑁out
Hyperdiff ΩC , 226
O + (U), 214
𝔦𝔡𝑽, 477 O0(𝑛) , 475
I𝐹 (𝑥), 1121 Ob (𝑈 𝜀 ), 176
1214 Notation Index

Ob (𝑈), Ob,𝑧 ◦ , Ob , 767 𝑷 (𝑥, 𝑡, D 𝑥 , D𝑡 ), 𝑸 (𝑥, 𝑡, D 𝑥 ), 1164

O◦ (C\𝐾), 156 P𝑧 ◦ ,𝜁 ◦ , P−1
𝑧 ◦ ,𝜁 ◦ , 892
O (C𝑛 ), O𝑧 ◦ , 144 𝑃 (D), 32
1
𝛀 2 (Λ), 755 𝑃 [ℎ], 803
O (𝑚, K), O+ (𝑚, K), 383 𝜑∗ , 286
O (𝑚) (W𝛿 (Ω, Γ)), 42 𝜑∗ , 272, 286
O (𝑛) , 312 𝚽℘ (Λ), 𝚽◦℘ (Λ), 749
O (−𝜔) (Ω), 768 Φ♭◦ (𝜁), 926
O (−𝜔) (Ω), O𝑧(−𝜔) ◦ , 767 Φ♭ (𝜃), 922
Φ♭𝐾 (𝜃), 923

O Ω ,8 C
Φ (𝑡, 𝜉), 1170

O ΩC ; Exp1,0 (C𝑛 ) ,
O ΩC b

⊗Exp1,0 (C𝑛 ), 225 𝜛𝑛C ((𝑧, 𝜁) , (𝑧 ′, 𝜁 ′)), 429
Op, 566, 598 𝚷 (𝑛−𝜈) (𝑉), 1184
Op𝑊 𝑎, 606 𝜛𝑛 (𝑧, 𝑧 ′), 377
𝔒𝔭, 586 𝜋 𝜑 , 1144, 1170
O (Φ) (Ω), 836 𝚷 (𝑉, 𝜉), 𝚷 (𝑉, ℭ), 1184
′
Oform (Ω), 774 e 400, 410
𝜛,
O ′ (Ω), O ′ (𝐾), 141 𝑃 𝑗 = D𝑡 𝑗 + 𝑖𝑞 𝑗 (𝑡, D 𝑥 ), 1170
Otemp (W𝛿 (Ω, Γ)), 42 𝑝 𝜑 , 𝑞 𝜑 , 690
O𝑇 ( 𝑝,0) (Ω, 𝒁), 1115 Psh (Ω), 337
( ( 𝑝,0) (𝚿), 1134, 1147
OT𝑧 ◦𝑝,0) (Ω, 𝒁), OT𝝏Ω (Ω, 𝒁), 𝑚
Ψa,class (Ω, ℭ), Ψa,class (Ω, ℭ), 656
1117
𝑆 𝜓a,form (Ω × ℭ), 𝑆 𝜓a,form (Ω × ℭ),
𝑚
O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) ∩ O 𝑥 ◦ , 318
634
O (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))) ∩ O 𝑥 ◦ , 213
eb (Ω), 768 𝑆 −𝜔
𝜓a,form (Ω × ℭ), 634
O
O 𝑥+◦ (W𝛿 (𝑈, Γ 𝜀 (𝜃 ◦ ))), 213 Ψa𝑚 (Ω1 × Ω2 , ℭ), Ψa (Ω1 × Ω2 , ℭ),
O+( 𝑥 ◦ , 𝜃 ◦ ) , 213 623
𝑆𝑚 (Ω 1 × Ω2 × ℭ), 619
O𝑧 ◦ (Ω), 1080 𝜓a
𝑆𝑚 (Ω × ℭ), 𝑆 𝜓a (Ω × ℭ),
O𝑧(−𝜔) , O (−𝜔) , 768 𝜓a
◦
𝑆 −∞ −𝜔
𝜓a (Ω × ℭ), 𝑆 𝜓a (Ω × ℭ),
O𝑧(Φ)
◦ , O
(Φ) , 836
633
Ψa𝑚 (U, F), Ψa (U, F), 631
(𝑃1 #𝑃2 ) (𝑥, 𝜉), (𝑃1 #𝑃2 ) (𝑥, D 𝑥 ), 15 Ψa−𝜔 (Ω1 × Ω2 ), 625
𝝏 Z f𝑧 ◦ , 1117 𝑆 𝜓a (Ω1 × Ω2 ) , 𝑆 −∞ 𝜓a (Ω1 × Ω2 ), 620
𝜕𝜕, 436 𝑆 𝜓a (Ω1 × Ω2 × ℭ),
𝜕, 𝜕𝑧 , 276, 291 𝑆 −∞
𝜓a (Ω1 × Ω2 × ℭ), 620
𝜕𝑥 , 𝜕𝑥𝛼 , 4 𝜓∗ 𝑋, 272
𝜕𝒁 , 1115 Ψa (U, F), 𝚿a,F , 631
𝜕𝑧¯ , 𝜕, 291 𝚿micro
a (R𝑛 ), Ψamicro U † , 684
𝜕𝑧¯ , 𝜕, 297 Ψclass (M), Ψclass 𝑚 (M), 599

𝜕𝑧(𝑚) , 1115 Ψclass (Ω), Ψclass (Ω), 598

𝑚

(P), 951 Ψ−∞ (Ω1 × Ω2 ), Ψ−∞ (Ω), 574

𝑷𝔠,𝔠∗ , 𝑰𝔠,𝔠∗ , 890 Ψ𝑚 (Ω1 × Ω2 ), Ψ (Ω1 × Ω2 ), 573
𝑷ℭ ℎ, P𝑧 ◦ ,𝜁 ◦ , 873 Ψ𝑚 (Ω), Ψ (Ω), 573
𝑷 = (𝑃1 , ..., 𝑃 𝜈 ), 1160 (Ψ𝑈), 1109
Notation Index 1215

Ψ (U), Ψ (M), Ψ𝑚 (U), 580 𝑆 −∞ (Ω1 × Ω2 , Γ), 689

Ψ𝑊 ,𝑚 (R𝑛 ), Ψ𝑊 (R𝑛 ), 606 𝑆 −∞ (Ω, U), 591
(𝚿 ( 𝑥 ◦ , 𝜉 ◦ ) ), 971 singsupp, 20
𝑃⊤ (𝑥, D 𝑥 ), 14 singsuppa , 20, 205
pv 1𝑥 , 19 𝔰𝔩 (2, K), SL (2, K), 383
∥𝑃∥ 𝔓 , 809 𝑆 𝑚 (Ω1 × Ω2 , Γ), 𝑆 (Ω1 × Ω2 , Γ),
P 𝑥 , 175 689
𝑃 (𝑥, D 𝑥 ), 14 𝑆 𝑚 (Ω1 × Ω2 ), 𝑆 −∞ (Ω1 × Ω2 ),
𝑃∗ (𝑥, D 𝑥 ), 14 𝑆 (Ω1 × Ω2 ), 566
𝑃 (𝑥, 𝜉), 14 𝑆 𝑚 (Ω), 𝑆 (Ω), 𝑆 −∞ (Ω), 581
𝔰𝔭 (𝑛, K), 385
(Q), 992 Sp (𝑛, R), Sp (𝑛, C), Sp (𝑛, K), 383
(QCX), 1103 S ′ (R𝑛 ), 20
(∼ QCX), 1104 S ′ (R𝑛; H), 𝐻 𝑠 (R𝑛 ; H), 1021
𝑄 𝑗 (𝑧, 𝜁 ′), 1164 S ′ Z+𝑛 , 105
𝑞 𝑗 (𝑧, 𝜁 ′), 1161 𝑆 −𝜔 −𝜔
𝜓a (Ω1 × Ω2 × ℭ), 𝑆 𝜓a (Ω1 × Ω2 ),
𝔔 (𝑞) (℘), 556 622
𝔔𝑟𝑛 , 𝔔𝑟𝑛−𝜈
(𝜈) , 495 □, 686
𝑚 (Ω), 𝑆 𝑚 (Ω; L (H)), 1021
𝑆 𝜌,
𝑅 𝑎 (𝜆; 𝑥, 𝜉), 650 𝛿 𝜌, 𝛿
rank℘𝔤 (𝑋1 , ..., 𝑋𝑟 ), 368 S (R𝑛 ), 20
Res 𝔞, 810 S R𝑛 ; C𝑑 , 81
𝑚 (Ω × Ω ), 𝑆
ℜ (𝑬), 𝔖 (𝑬), ℜ𝑘 (𝑬), 459 𝑆sub 1 2 sub (Ω1 × Ω2 ), 566
𝑚 (Ω), 𝑆
ℜ 𝜑 , 708 𝑆sub sub (Ω), 581
𝑚
𝑆temp (R𝑛 ), 603
𝜌𝑈𝑉 , 10
R (Ω), R (𝑈), 260 𝑠 (𝑡, 𝜅), 1174
R (U; B), C 𝑘 (U; B), 269 𝑠 (𝑡, 𝜅, 𝜉), 1181
𝑔 𝑅 (𝜁), 659 supp, 5, 18, 265
supp [ 𝑓 ], 349
𝑚
𝑆a,class (Ω × ℭ), 𝑆a,class (Ω × ℭ), 655 supp𝜇, 162
−𝜔
𝑆a (Ω × Θ) , 𝑆a𝑚 (Ω × Θ), 829 supp𝑠, 269
𝑆a (Ω × Θ), 828 symb (𝑃),
598
SB, PB, 270 S Z+𝑛 , 105
𝑚 (Ω; 𝔓 [ℎ]), 𝑆
𝑆class class (Ω; 𝔓 [ℎ]),
814 𝑇 (1,0) , 𝑇 (0,1) , 291
𝑚 (Ω), 𝑆
𝑆class class (Ω), 595 𝑇 ∗ M, 𝑇℘∗ M, 273
𝑚 (Ω × Θ), 𝑆
𝑆class class (Ω × Θ), 769 tauA𝜏𝐴, 438
(𝑞)
𝑆c𝑚 (Ω1 × Ω2 , Γ) , 𝑆c (Ω1 × Ω2 , Γ), 𝜏F, U, V , 302
∗ N , 403
689 𝑇M
𝑆form (Ω), 𝑆form
𝑚 (Ω), 584 𝑇℘ M, 𝑇M, 272
𝝈, 𝝕, 698 𝑇℘ N ⊤𝜛 , 401
Σ 𝑷, U C , 1161 Tr𝐴, 383
Σ 𝜑 , 690, 869 𝑇𝑥∗ Ω, 15
𝜎sub (𝑃), 600, 602 𝑇𝑥 Ω, 15
⟨𝜎, 𝑋⟩, 353
U 𝐿 2 (R𝑛 ) , 388

sign, 342
1216 Notation Index

U (𝑚), 383 𝜕𝑧 , 𝜕/𝜕𝑧 𝑗 , 8

𝑢 ∧ 𝜑, 332 ∥𝑥∥, 556
𝑢, 6
b 𝑥 = (𝑥 1 , ..., 𝑥 𝑛−𝜈 ), 𝑡 = (𝑡 1 , ..., 𝑡 𝜈 ),
1164
e [1] , 𝑽
𝑽 e [2] , 486 [𝑋, 𝑌 ], 353
𝑽 , 𝒁 [𝜈 ] , 495
[𝜈 ]

(𝑽, ℘◦ ), V℘◦ , 465 [1]

𝒁 , 485
e
W𝛿 (Ω, Γ), 42 𝜕𝑧¯ , 𝜕, 8
𝑊 𝐹a ,57,198 𝒁 = (𝑍1 , ..., 𝑍 𝑚 ), 1115
𝜁 ′′ = (𝜁 𝑛−𝜈+1 , ..., 𝜁 𝑛 ), 1162
𝑊 𝐹a 𝑓 ♮ , 347
𝜁 ′ = (𝜁1 , ..., 𝜁 𝑛−𝜈 ), 1161
𝑊 𝐹 (𝑢), 23, 198 𝑍 𝑘 (𝑥, 𝑡, 𝜉), 1170
wght 𝑃, 803 ⟨𝑧⟩, 5, 184
𝑾 K , 470 [𝜈 ]
e [𝜈 ] , 489
𝑾 R , 465 𝒁 ,𝑽
e
c𝛿,𝑟 (𝑈, Γ 𝜀 (𝜃 ◦ )), 318
W 𝑍 (U; F), 𝐵𝑞 (U; F), 𝐻 𝑞 (U; F),
𝑞

300
Í
𝑥max (𝜃), 920 𝑍 = 𝑛𝑗=1 𝑎 𝑗 𝜕𝑧𝜕 𝑗 , 1076
𝑋 ⌟𝜔, 365 Z, Z𝑧 ◦ , 1080
Index

acyclic (sequence of sheaves), 309 analytic singular support (of a

adapted change of variables [to a hyperfunction), 198, 347
(1,0)-vector field], 1083 analytic stratification (of a set), 459
Ad derivation, 353 analytic subvariety, 462
adjoint (of a differential operator), 14 analytic symbol, 827
Ad operation, 352 analytic variety, 464
affinization (of a vector bundle), 268 analytic wave front set (of a
amplitude, 566, 689 hyperfunction in an analytic
analytic amplitude, 827 manifold), 346
analytic wave-front set (of a
analytic differential operator of
distribution), 57
infinite order, 832
analytic wave-front set (of a
analytic foliation (of a manifold), 365
hyperfunction), 198
analytic functionals, 143
analytic wave-front set (of a
analytic hypoelliptic (differential singularity hyperfunction), 347
operator), 36 analytic-singularity distributions,
analytic hypoelliptic 903
hyperdifferential operator, 244 approximate homotopy formula,
analytic hypoelliptic, in a conic 1181
subset, 58 asymptotic extension, 616, 620
analytic parametrix, 36 atlas (of a manifold), 265
analytic polyhedron, 506
analytic pseudodifferential operator, Baouendi–Goulaouic operator, 70
622 base projection, 266
analytic pseudolocal, 641 basis of the topology (of a
analytic regularizing (operator), 625 topological space), 10
analytic singular spectrum of a bicharacteristic curves, 409
hyperfunction, 198 bicharacteristic foliation, 405
analytic singular support (of a bicharacteristic leaves, 405
distribution), 20 bicharacteristics, 409, 415

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1217
F. Treves, Analytic Partial Differential Equations, Grundlehren der mathematischen
Wissenschaften 359, https://doi.org/10.1007/978-3-030-94055-3
1218 Index

biholomorphism, 230, 260 clean phase-function, 691

Bochner–Martinelli formulas, 335 closed (differential forms), 287
boundary of a set, 458 closed convex hull of a compact set,
boundary value (distribution), 48 147
boundary value (hyperfunction), 183 coboundaries, 300
boundary value map, 197 coboundary operator, 300
bounded symbols, 829 cochain, 300
cocycles, 300
Camassa–Holm equation, 136 codimension (of a submanifold), 262
Campbell–Hausdorff formula, 362 cohomology group, 300, 303
canonical bundle of projective space, coisotropic (submanifold), 404
1071 coisotropic (vector subspace), 377
canonical coordinates (in a commutator subalgebra, 395
Euclidean space), 264 commutator subgroup, 387
canonical phase-function, 852 complete system of first integrals,
carrier (of an analytic functional), 417
144 complex characteristic set of a
Cauchy formula, differential operator, 16
multivariable Cauchy formula, 8 complex involutive structure (on a
Cauchy kernel, 332 manifold), 359
Cauchy transform, 156 complex Lagrangian subspace, 433
Cauchy–Kovalevskaya Theorem, 120 complex manifold, 477
Cauchy–Riemann operator, 8 complex null bicharacteristic leaves
caustics, 754 (of a system of
C-convex (domain in complex pseudodifferential operators),
space), 1068 1161
chain-rule derivation, 802 complex symplectomorphism, 447
characteristic hypersurface at a point, complex-analytic (function), 8
for a differential operator, 16 complex-analytic (symbol), 827
characteristic set (of a complexification (of a real manifold),
pseudodifferential operator), 264
599 complexification (of a vector
characteristic set of a differential bundle), 269
operator, 15 composite (of Fourier integral
classical analytic amplitude, 769 operators), 729
classical analytic differential composition (of symbols), 588
operator of infinite order, 776 composition law #, 588, 810, 829
classical analytic pseudodifferential concatenations, 95
operator, 656 Condition (P), 951
classical analytic symbol, 655, 769 Condition (Q), 992
classical pseudodifferential operator, conically compact support
597 (amplitudes having a), 689
classical pseudodifferential operator conic sets, 57
(on a manifold), 598 conic subset (of a vector bundle), 271
classical symbol, 595 connecting homomorphism, 306
Index 1219

conormal bundle of a submanifold, differential form of bidegree (p,q),

274, 410 296
conormal covector (to a differential (of a function), 276
hypersurface, at a point), 15 differential (of a map), 276, 396
conserved polynomial, 816 differential operator (acting on
constant of motion, 815, 816 sections of a vector bundle),
contour of integration, 834 269
convex function, 1101 differential operator (in a manifold),
convex hull, 59 265
convolution of analytic functionals, differential operator of infinite order
149 with constant coefficients, 223
coordinate chart (or patch), 261 differential operators (of a vector
coprime polynomials, 476 bundle into another), 280
cosphere bundle (of a manifold), 274 differential polynomial, 4
cotangent bundle, 15, 271, 273 differential q-form, 285
cotangent space (at a point), 15 dimension of a manifold, 261
cotangent vector (to a manifold) at a dimension (of an analytic variety),
point, 273 467
countable at infinity (topological dimension of an immersed
space), 260 submanifold, 277
Cousin problem, 297 Dirac measure, 19
Cousin property, 151 Dirac measure associated to a
covector, 273 submanifold, 527, 533
critical points (of a vector field), 361 discriminant (of a polynomial), 484
critical set (of a phase-function), 690 distribution kernel, 28, 569
current kernel, 334 divergence free (vector field), 357
current of bidegree (p,q), 333 divergence (of a vector field), 357
current (on a manifold), 330 division of distributions by analytic
functions, 526
Darboux coordinates, 405 Dolbeault complex, 293
Darboux Theorem, 407 domain of holomorphy, 317
d-bar complex, 291 doubly characteristic point, 600
defining function (of a hypersurface), dual (of a vector bundle), 268
340 dual Sjöstrand pair, 851
degenerate phase-functions, 708
densities, 329 E-conjugation, 431
density distributions, 329 edge of a wedge, 42
De Rham cohomology, 288 Edge of the Wedge Theorem, 61
De Rham complex, 288 effective phase-function, 904
De Rham complex (for currents), 332 Egorov’s theorem, 746
derivation [of the sheaf O induced by Ehrenpreis cutoff functions (or
a (1,0)-vector field], 1080 cutoffs), 38
descending chain condition, 478 Ehrenpreis sequence relative to a pair
diffeomorphism, 260 of subsets, 39, 634
differential complex, 282, 1115, 1167 eikonal equations, 416, 1171
1220 Index

elliptic amplitude, 589 exponential (of an analytic

elliptic classical analytic differential pseudodifferential operator),
operator of infinite order, 781 653
elliptic classical analytic symbol, exponential order 1 and type 0
655, 771 (entire functions of), 222
elliptic classical microdifferential exponential type (entire functions
operator, 683 of), 146
elliptic classical pseudodifferential extendible distributions, 528
operator, 599 extendible distributions (across
elliptic differential complex, 283 analytic varieties), 530
elliptic differential operator, 15 exterior algebra, 284
elliptic formal classical symbol, 597 exterior derivative (of a current), 332
exterior derivative (of a function),
elliptic Fourier integral operator, 742
276
elliptic hyperdifferential operators,
exterior derivative (of a q-form), 287
242
exterior multiplication (of a p-current
elliptic pseudoanalytic amplitude,
by a smooth q-form), 331
symbol, 648
elliptic pseudodifferential operator,
589 FBI phase-function, 915, 1050
elliptic symbol, 589 FBI transform of a distribution, 51
elliptic system of vector fields, 369 FBI transform of an analytic
elliptic systems of differential functional, 185
operators, equations, 79 FBI transform of distributions
embedding (of a manifold into (inverse), 53
another), 263 fibre bundles, 266
fine resolution of a sheaf, 309
envelope of holomorphy, 317
fine sheaf, 308
equivalent open coverings, 302
finite order (analytic symbols of),
equivalent phase-functions, 722
829
essential support of a hyperfunction,
finite order (distributions of), 19
198
finite realization, 766, 773
exact (differential complex), 282 finite type (point of), 949
exact (differential form), 287 finite type (set of functions of), 411
exact sequence (of maps), 304 finite type (system of vector fields
excess of a phase-function, 691 of), 368, 413
exhaustive sequence of compact sets, first integrals, 417, 815, 816
9, 143, 148 first integrals (of a vector subbundle
exponent-weight, 836, 1144 of the tangent bundle), 360
of a phase-function, 1171 flabby sheaf, 171, 349
exponential decay, 55, 192, 767 flux, 816
exponentially decaying (amplitude), formal analytic functional series, 774
620 formal analytic series, 764, 772, 774
exponentially decaying (symbol), formal classical amplitude, 596
828 formal classical symbol, 596
Index 1221

formal exponentially decaying germ FBI-like transforms, 924

symbol, 633 germ FBI transform, 936
formal hyperfunction series, 775 germ FIO, 867
formal microfunction series, 776 germ Fourier integral operator, 867
formal pseudoanalytic symbol, 633 germ Fourier-like transform, 858
formal singularity hyperfunction germ (of a classical analytic
series, 775 differential operator of infinite
formal symbol, 583 order), 783
four-clover curve, 278 germ (of a holomorphic function at a
Fourier distribution kernel, 695 compact set), 144
Fourier integral operator, 706 germ (of a hyperfunction
Fourier transform of tempered microanalytic at a point), 217
distributions, 21 germ (of a mathematical entity) at a
Fourier–Brós–Iagolnitzer transform, point, 11
51 germ (of a set at a point), 14
frame (in a vector bundle), 268 germ (of a subvariety at a point), 464
Fréchet derivative, 113 germ pseudodifferential operator,
Fréchet differentiable map, 112 874
Fréchet–Montel space, 173 germs of analytic pseudodifferential
Fréchet space, 8, 173 operators (sheaf of), 631
Friedrichs Lemma, 989 Gevrey class, 9
Frobenius’ Theorem, 355 global solvability, 1078
functions whose derivatives of all Grassman algebra (of a manifold),
orders are bounded, 514 285
fundamental matrix, 380 Grushin operator, 81, 398
fundamental solution of a differential
operator, 31 Hamilton–Jacobi equations, 417
fundamental symplectic matrix, 380, hamiltonian, 808
851 Hamiltonian vector field, 400
fundamental symplectic two-form, Hamilton–Jacobi equations, 409
398 harmonic oscillator, 398
fundamental two-form, 409 harmonic oscillator operator, 70, 80
Hartog Theorem, 156
Gårding inequality, 74 height (of a wedge), 42
Gelfand–Dickey hierarchy, 824 Heisenberg group, 373, 391
general linear group, 381 Heisenberg Lie algebra, 69, 372
generalized Leibniz rule, 4 Hermite functions, 102
generalized mean value theorem, 520 Hessian matrix, 342
general linear group, 264 Hessian of a function, 341
geodesics of the torus, 279 Hironaka Theorem, 548
germ analytic hypoelliptic Holmgren Theorem, 134
(differential operator), 58 Holmgren Theorem for
germ at a subset, 13 hyperfunctions, 216
germ (classical analytic differential holomorphic De Rham complex, 290
operators of infinite order), 782 holomorphic functions, 8
1222 Index

homogeneous (change of variables), involutive system (of

940 pseudodifferential operators),
homogeneous coordinates in 1160
projective space, 1071 involutive vector subbundle (of the
homogeneous equation, 72, 1048 tangent bundle), 354
homogeneous functions (in irreducible components (of a germ of
cotangent bundles), 595 variety), 479
homotopy formula (for d-bar), 335 irreducible polynomial, 476
homotopy operators, 1188 irreducible unitary representation (of
hyperdifferential operator, 223, 226, the Heisenberg group), 391
827 irreducible variety, 464
hyperdifferential operators at a isotropic (submanifold), 404
compact set, 228 isotropic (vector subspace), 377
hyperdifferential operators in real I-symplectic (submanifold), 442
space, 228, 234 I-symplectic (vector subspace), 430
hyperfunction, 167, 170
Jacobi identity, 353, 401
hyperfunctions (sheaf of), 170
hypoelliptic differential operator, 33 KdV equation, 820
KdV Hierarchy, 820
I-coisotropic (submanifold), 442 Keller–Maslov line bundle, 755
I-coisotropic (vector subspace), 430 kernel current, 333
I-isotropic (submanifold), 442
Lagrangian (submanifold), 404, 408
I-isotropic (vector subspace), 430
Lagrangian submanifold (defined by,
I-Lagrangian (submanifold), 442
or associated to a
I-Lagrangian (vector subspace), 430
phase-function), 699
immersed submanifold, 277
Lagrangian (vector subspace), 377
immersion (of a manifold), 277 Laplace–Borel transform, 146, 223
infinite type (point of), 949 Laplacian, 7
infinite type (set of functions of), 411 leading term (of an amplitude), 743
infra-exponential type (entire Leibniz rule, 4
functions of), 222, 237 Leibniz rule (for hyperdifferential
inhomogeneous equation, 97 operators), 224, 226
integral curve (of a vector field), 361 Leibniz rule (for the variational
integral domain, 481 derivative), 808
integral manifold (of a system of Lenard operator, 819
vector fields), 354, 365 Levi form (of a domain at a
integral manifold (of a vector boundary point), 340
subbundle of the tangent Levi form of a function, 338
bundle), 354 Lie algebra (of a linear group), 382
interaction representation of the Lie bracket, 352
potential, 611 linear group, 266, 382
inverse (of an elliptic pseudoanalytic linearly convex, 1070
symbol), 648 local analytic stratification (of a set),
involutive (system of functions), 414 459
Index 1223

local chart (in a manifold), 261 microanalytic at a point

local coordinates, 261 (hyperfunction), 198
local flow (of a vector field), 362 microanalytic hyperfunctions (germs
locally convex topology on a vector of), 349
space, 156 microdifferential operator, 683
locally exact (differential complex), microdistribution, 219
290, 293 microdistribution (over a manifold),
locally holomorphically solvable, 349
1079 microfunctions, 218
locally integrable (vector subbundle microfunctions (over a manifold),
of the tangent bundle), 359 349
locally solvable (differential microlocally solvable in distributions
operator), 963 (pseudodifferential operator),
locally solvable at a boundary point 1140
[(1,0)-vector field], 1080 microlocally solvable in
locally solvable at a point hyperfunctions
(differential operator), 963 (pseudodifferential operator),
locally well-shaped (contour of 1140
integration), 834 microsupport (of a hyperfunction in
local operator, 227 a manifold), 348
local regular isomorphism, 278 microsupport (of a hyperfunction),
local representation (of a 198, 217
distribution), 19 microsupport (of a pseudodifferential
local representation (of operator), 592
hyperfunctions), 239 Mizohata pseudodifferential
local solvability (in distributions), operators, 1132
974, 983 Mizohata vector fields, 948
local solvability (in hyperfunctions), monic polynomial, 250
1055 Montel Theorem, 174
local solvability (in smooth multibracket, 353, 362
functions), 984
local symplectic graph, 705 Nagano foliation (of an analytic
locus (of an ideal of analytic manifold), 367
functions), 478 Nagano leaf, 367
Lojasiewicz inequality, 516, 521, 550 natural ordering, 303
long exact sequence (of maps), 306 natural partition (of the regular part
of a set), 461
Noetherian ring, 481
Mehler’s formulas, 92 noncharacteristic complex
metaplectic group, 388, 394 hypersurface at a point, for an
microanalytic at a point analytic differential operator,
(distribution), 57 16
microanalytic at a point noncharacteristic hypersurface at a
(hyperfunction in an analytic point, for a differential
manifold), 346 operator, 16
1224 Index

nondegenerate at a point Parseval–Plancherel identity, 7

(hypersurface), 447 partition of the identity subordinate
nondegenerate (critical point), 758 to an open covering, 308
nondegenerate (phase-function), 691, partition of unity (subordinate to a
719, 749, 869 covering), 6
normal bundle (of a submanifold of a partition of unity subordinate to a
Riemannian manifold), 274 covering by cubes, 560
normal convergence, 8 p-current (on a manifold), 330
normal convergence (in complex Petrowski Theorem, 71, 246
space), 875 phase-function, 389, 690, 869
normal form, 844, 944, 1169 phase-function of the identity, 705
nuclear spaces, 83 pluriharmonic function, 341
null bicharacteristic curve, 947, 1046 plurisubharmonic convex (compact
null bicharacteristic leaf, 947 set), 339
null bicharacteristic leaves (of a plurisubharmonic function, 337
system of pseudodifferential plurisubharmonic quadratic forms,
operators), 1161 341
null bicharacteristic surface, 947 Poincaré Lemma, 290
Poisson bracket, 352, 401
order (of a differential operator), 15
polar of a cone, 58
order (of an amplitude), 566, 689
polydisk, 8
order (of an analytic symbol), 828
polynomial growth at infinity
order (of a polynomial in an infinite
(sequences of), 105
sequence of indeterminates),
polynomially convex hull of a
803
compact subset of complex
order (of a symbol), 581
space, 150
orientable (manifold), 265
polynomially convex subsets of
orientation (of a manifold), 264
complex space, 150
orthogonal group, 382
polynomials in an infinite sequence
oscillatory integral, 566, 695
of indeterminates, 802
outer conormal bundle, 451
Ovsyannikov analytic map at a point, presheaf, 11
114 primary ideal, 482
Ovsyannikov class of k-th prime ideal, 482
continuously differentiable prime, same as irreducible, 476
functions, 128 principal part (of a classical symbol),
Ovsyannikov Theorem, 115 596
principal symbol (of a differential
Paley–Wiener–Schwartz Theorem, operator), 15, 280, 598
21 principal symbol (of a Fourier
Paley–Wiener Theorem, 147 integral operator), 755
paracompact (topological space), 307 principal symbol (of a
parametrix of a differential operator, pseudodifferential operator),
32 598
parametrix of a pseudodifferential principal type (pseudodifferential
operator, 998 operator of), 941
Index 1225

principal type (system of functions radial vector field, 409, 941

of), 414 radical (of an ideal), 478, 482
principal type (systems of rank (of a bilinear form), 404
pseudodifferential operators), rank (of a map), 277
1160 rank (of a module over a ring), 481
principal value distribution, 19 rank (of a vector bundle), 268
product manifold, 262 rapidly decaying at infinity
projective bundle, 271 (functions), 20
propagation of analytic singularities, rapidly decaying sequences, 105
1048 R-coisotropic (submanifold), 442
propagation set, 613 R-coisotropic (vector subspace), 430
properly supported (amplitude), 570, real Weierstrass Preparation
695 Theorem, 474
properly supported (distribution real-analytic (function), 6
kernel, operator), 30 real-analytic functions (topology on
Property (P), 951, 957 the space of), 9, 37
Property (Psi), 1133, 1187 reciprocal of an analytic function,
pseudoanalytic amplitude, 619, 756 550
pseudoanalytic symbol, 619, 830 reducible variety, 464
pseudoconvex (open subset of refinement (of an open covering),
complex space), 339, 1074 300
pseudodifferential complex, 1165 regular (distribution kernel), 29
pseudodifferential operator between regular (fiber bundle), 267
vector bundles, 580 regular (function, manifold), 259,
pseudodifferential operator in a conic 260
open subset of the cotangent regularizing operator,
bundle, 594 smoothing operators, 30
pseudodifferential operator on a regularly separated sets, 538
manifold, 580 regular (map), 262
pseudolocal operator, 570 regular part (of a set), 458
pullback (of a function), 261 regular point (of a set), 458
pullback (of a q-form), 286 relative cohomology classes, 313
pullback (of a sheaf), 13 relative sheaf cohomology groups,
purely monic, 812 313
pushforward (of a distribution), 328 Rellich’s Lemma, 26
pushforward (of a sheaf), 13 restriction map, 10
pushforward (of a tangent vector), Riemann Extension Theorem, 473
277, 286 Riemann surface, 962
q-form, 285 ring of square matrices, 381
quantized contact transformation, R-isotropic (submanifold), 442
1141 R-isotropic (vector subspace), 430
quasiconvex function, 1102, 1103 R-Lagrangian (submanifold), 442
quotient vector bundle, 269 R-Lagrangian (vector subspace), 430
R-symplectic (submanifold), 442
radial dilations, 409 R-symplectic (vector subspace), 430
1226 Index

Runge open (or compact) set, 145 spaces of distributions, 19

Runge property, 145, 151 special linear group, 382
Runge Theorem, 132 sphere bundle, 270
square system of PDEs, 121
saddle point (of a function), 835 stalk (of a sheaf, at a point), 12
scale of Banach spaces, 111 standard amplitude, 566
Schrödinger equation, 611 standard pseudodifferential operator,
section (of a disjoint union over its 573
base), 11 standard symbol, 581
section (of a fiber bundle), 266 stationary phase formal expansion,
selfadjoint (pseudodifferential 760
operator), 711 stationary phase formula (complex),
semianalytic set, 507 790
semiglobal solvability (in Stein manifold, 264, 341
hyperfunctions), 1055 stratifiable set, 460
semiregular (distribution kernel), 29 strictly convex function, 1101
sharp Gårding Inequality, 577 strictly plurisubharmonic function,
sheaf defined by a presheaf, 12 338
sheaf isomorphism, 12 strictly pseudoconvex (domain at a
sheaf map, sheaf morphism, 12 boundary point), 340
sheaf of germs of microfunctions, strictly subharmonic function, 337
218 strictly Z-pseudoconvex at a
sheaf topology, 12 boundary point (domain in
short exact sequence (of maps), 306 complex space), 1095
singular part (of a set), 458 strongly nondegenerate phase
singular point (of a set), 458 function, 704
singular support (of a distribution), strongly pseudoconvex (domain at a
20 boundary point), 340
singularity hyperfunction boundary strongly well-shaped (contour of
value, 184 integration), 834
singularity hyperfunctions, 179, 204 structure group (of a fiber bundle),
singularity hyperfunctions (in a 266
manifold), 346 Sturm–Liouville symbol, 816
Sjöstrand pair, 844 subanalytic sets, 513
Sjöstrand triad, 864 subelliptic (differential operator),
slow decay (at a set), 514 986
smooth, 6 subharmonic function, 336
smoothing in a conic open set submanifold, 262
(operator), 592 submersion (of a manifold), 280
smoothing operator, 30, 573 subprincipal symbol, 602, 607
Sobolev inequalities, 28 substandard amplitude, 567
Sobolev spaces, 24, 575 substandard symbol, 581
solvability of hyperdifferential subvariety, 462, 464
operators with constant support (of a distribution), 18, 162
coefficients, 253, 255 support (of a function), 5
Index 1227

support (of a hyperfunction), 170, totally real vector subspace (of a

346 complex vector space), 343
support (of an analytic functional in total symbol (of a differential
real space), 142, 159, 162 operator), 15
symbol, 581 total symbol (of an analytic
symbol map, 597 pseudodifferential operator),
symbol of a differential operator, 15 656
symplectic basis (of a symplectic total symbol (of a pseudodifferential
vector space), 378 operator), 598
symplectic form, 376 transfer operators, 930
symplectic group, 383 transport equations, 714, 966, 1172
symplectic isomorphism, 378 transpose Leibniz rule, 4
symplectic manifold, 398 transpose (of a differential operator),
symplectic map, 400 14
symplectic structure (on a manifold), transpose (of a hyperdifferential
398 operator), 230
symplectic (submanifold), 404, 407 transverse intersection (of a real
symplectic vector space, 376 hypersurface in complex space
symplectic (vector subspace), 377 by a complex line), 1072
symplectomorphic (symplectic transverse to a hypersurface (vector
manifolds), 400 field), 16, 1081
symplectomorphism, 400 trapped (or confined) bicharacteristic
leaf, 1054
Trépreau’s Theorem, 1141
tangent bundle, 271 Tricomi operator, 942
tangent vector (to a manifold) at a trimming (of a Weierstrass
point, 272 polynomial), 484
tautological one-form, 274, 410, 940 trim Weierstrass polynomial, 484
tempered distributions, 20, 547 trivialization (of a fiber bundle), 266
tempered fundamental solutions of true symbol (associated to a formal
linear PDEs with constant symbol), 596
coefficients, 554 twisted symplectic one-form, 698
tempered growth (of a function) at twisted symplectic two-form, 698
the boundary, 514 type (of a set of functions), 411
tempered growth (of a function) at type (rho,delta) (pseudodifferential
the edge, 43 operator of), 1022
tempered sequences, 105 typical fiber (of a fiber bundle), 266
tempered symbol, 603
test-densities, 328 unique factorization, 476
test-functions, 6 unique factorization domain, 481
Theorem of the Edge of the Wedge, unit (in a ring), 475, 481
206 unitary Fourier integral operator, 742
torus, 358 unitary group, 382
totally real submanifold (of a unitary representation (of the
complex manifold), 264 Heisenberg group), 391
1228 Index

unitary transformations (of Weierstrass–Malgrange Preparation

square-integrable functions), Theorem, 944
388 Weierstrass polynomial, 472
upper semicontinuous function, 336 Weierstrass Preparation Theorem,
471, 475
variational derivative (of a weight (of a polynomial in an infinite
polynomial in an infinite sequence of indeterminates),
sequence of indeterminates), 803
804 well-shaped (contour of integration),
vector bundle, 268 834
vector subbundle, 269 Weyl pseudodifferential operator,
velocity vector (of a curve), 361 606
Whitney condition, 460
Whitney stratifiable, 460
wave operator, 686 Whitney umbrella, 469
wave-front set of a distribution, 23
weakly linearly convex (domain in Z-convex at a boundary point
complex space), 1070 (domain in complex space),
weakly linearly convex (open subset 1111
of complex space), 1072 Z-line, 1080
wedge, 42 Z-pseudoconvex at a boundary point
Weierstrass Division Theorem, 472, (domain in complex space),
475 1093, 1095