Notes on

Functional Analysis and Partial Differential

Equations
Herbert Koch
Universität Bonn
Winter term 2016

These are short incomplete notes. They do not substitute textbooks. The
following textbooks are recommended.

• H. W. Alt, Linear functional analysis, Springer, 2016.

• D. Werner, Funktionalanalysis, Springer, 2011.

Correction are welcome and should be sent to [email protected]

or told me during office hours. The notes are only for participants of
the course V3B1/F4B1 PDG und Funktionalanalysis /PDE and Functional
Analysis at the University of Bonn, Winter term 2016/2017. A current
version can be found at
http://www.math.uni-bonn.de/ag/ana/WiSe1617/V3B1 WS 16.html.

Examination: Saturday 18.02.2017 and Wednesday, 29.03.2017

Assistant: Xian Liao
Tutors:
Dimitrije Cicmilovic (from 05.12.2016)
Tobias Friesel
QiCheng Hua (until 02.12.2016)
Leona Schlöder
Florian Schweiger

1 [February 10, 2017]

Contents
1 Introduction 3

2 Hilbert spaces 9
2.1 Definition and first properties . . . . . . . . . . . . . . . . . . 9
2.2 The Riesz representation theorem . . . . . . . . . . . . . . . . 13
2.3 Orthonormal systems . . . . . . . . . . . . . . . . . . . . . . . 15

3 Lebesgue spaces 21
3.1 Review of measure spaces . . . . . . . . . . . . . . . . . . . . 21
3.2 Construction of measures . . . . . . . . . . . . . . . . . . . . 23
3.3 Jensen’s and Hölder’s inequalities . . . . . . . . . . . . . . . . 24
3.4 Minkowski’s inequality . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Hanner’s inquality . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6 The Lebesgue spaces Lp (µ) . . . . . . . . . . . . . . . . . . . 29
3.7 Projections and the dual of Lp (µ) . . . . . . . . . . . . . . . . 30
3.8 Young’s inequality and Schur’s lemma . . . . . . . . . . . . . 32
3.9 Borel and Radon measures . . . . . . . . . . . . . . . . . . . . 37
3.10 Compact sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.11 The Riesz representation theorem for Cb (X) . . . . . . . . . . 43
3.12 Covering lemmas and Radon measures on Rd . . . . . . . . . 47

4 Distributions and Sobolev spaces 56

4.1 Baire category theorem and consequences . . . . . . . . . . . 56
4.2 Distributions: Definition . . . . . . . . . . . . . . . . . . . . . 57
4.3 Schwartz functions and tempered distributions . . . . . . . . 64
4.4 Sobolev spaces: Definition . . . . . . . . . . . . . . . . . . . . 67
4.5 (Whitney) extension and traces . . . . . . . . . . . . . . . . . 70
4.6 Finite differences . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.7 Sobolev inequalities and Morrey’s inequality . . . . . . . . . . 75
4.8 Applications to PDE . . . . . . . . . . . . . . . . . . . . . . . 79

5 Linear Functionals 82
5.1 The Theorem of Hahn-Banach . . . . . . . . . . . . . . . . . 83
5.2 Consequences of the theorems of Hahn-Banach . . . . . . . . 86
5.3 Separation theorems . . . . . . . . . . . . . . . . . . . . . . . 88
5.4 Weak* topology and the theorem of Banach-Alaoglu . . . . . 89
5.5 The direct method in the calculus of variations . . . . . . . . 97

6 Linear Operators 102

6.1 Dual and adjoint operators . . . . . . . . . . . . . . . . . . . 102
6.2 Compact operators . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2.1 Examples of compact operators . . . . . . . . . . . . . 110
6.2.2 Eigenvalues and spectrum . . . . . . . . . . . . . . . . 112

2 [February 10, 2017]

[19.10.2016]

1 Introduction
Functional analysis is the study of normed complete vector spaces (called
Banach spaces) and linear operators between them. It is built on the struc-
ture of linear algebra and analysis. Functional analysis provides the natural
frame work for vast areas of mathematics including probability, partial dif-
ferential equations and numerical analysis. It expresses an important shift
of viewpoint: Functions are now points in a function space.
Let Ω ⊂ Rn be an open bounded set with smooth boundary. One of
the deepest results in Einführung in die PDG was that the Green’s function
g(x, y) provides a map
f →u
given by ˆ
u(x) = g(x, y)f (y)dy =: T f

so that
−∆u = f in Ω
u=0 on ∂Ω
whenever f is sufficiently regular. It is not hard to see that

T : L2 (Ω) → L2 (Ω)

and T is one of the most relevant operators in functional analysis.

The main abstract objects are topological vector spaces over K, K = R or
C. We will focus on normed spaces, the most important class of topological
vector spaces.

Definition 1.1. Let X be a K vector space. A map k.k : X → [0, ∞) is

called norm if
kxk = 0 ⇐⇒ x = 0, (1.1)
if for all x, y ∈ X

kx + yk ≤ kxk + kyk (triangle inequality) , (1.2)

and if for all λ ∈ K and x ∈ X

kλxk = |λ|kxk. (1.3)

It is called normed space and Banach space if it is complete as metric space.

3 [February 10, 2017]

Remark 1.2. A norm defines a metric by d(x, y) = kx − yk.
First examples.

1. Rn and Cn equipped with the Euclidean norm are real resp. complex
Banach spaces.

2. Let X be a set. The space of bounded functions B(X) equipped with

the supremum norm is a Banach space.

3. Let (X, d) be a metric spaces. The space of bounded continuous func-

tions Cb (X) equipped with the supremum norm is a Banach space, or
more precisely a closed sub vector space of B(X).

4. Let U ⊂ Rd be open. Cbk (U ) is the vector space of k times differentiable

functions on U which are together with there derivatives bounded. The
norm
kukC k (U ) = max k∂ α ukB(U )
|α|≤k

turns Cbk into a Banach space. Exercise

5. Let U ⊂ C be open. The space of bounded holomorphic functions

H ∞ (U ) is a Banach space when equipped with the supremum norm.

Lemma 1.3. Suppose that X is a Banach space and U ⊂ X is a vector

space which is a closed subset of X. Then U is a Banach space.
Definition 1.4. Let X and Y be normed spaces. We define L(X, Y ) as the
set of all continuous linear maps from X to Y .
Theorem 1.5. Let T : X → Y be in L(X, Y ). Then

kT kX→Y := sup kT (x)kY < ∞

kxkX ≤1

and k.kX→Y defines a norm on L(X, Y ). A linear operator T : X → Y is

continuous if and only if its norm kT kX→Y is finite. L(X, Y ) is a Banach
space if Y is a Banach space.
Proof. Continuous linear maps from X to Y are a vector space with the
obvious sum and multiplication. Let T : X → Y be a continous linear map.
We choose ε = 1 and x0 = 0. Then there exists δ > 0 so that

kT xkY ≤ 1 if kxkX ≤ δ,
δx
and hence, if x ∈ X, x 6= 0, then kxkX ≤ δ and
X

kxkX δx
kT xkY = T ≤ δ −1 kxkX
δ kxkX Y

4 [February 10, 2017]

and thus
kT kX→Y ≤ δ −1 .
Vice versa: Let T : X → Y be linear so that kT kX→Y < ∞. For ε > 0 we
choose δ = ε/kT kX→Y . Then

kT x − T ykY = kT (x − y)kY ≤ kT kX→Y kx − ykX ≤ ε

provided kx − ykX ≤ δ. In particular T is uniformly continuous.

Now assume that Y is a Banach space and let Tn ∈ L(X, Y ) be a Cauchy
sequence. For all x, Tn x is a Cauchy sequence in Y since

kTn x − Tm xkY ≤ kTn − Tm kX→Y kxkX .

Let
T x := lim Tn x.
n→∞
The convergence is uniform on bounded sets, and hence the limit T is con-
tinuous and in L(X, Y ). Moreover

kT − Tn kX→Y = sup k(T − Tn )xkY = sup lim sup k(Tm − Tn )xkY

kxkX ≤1 kxkX ≤1 m→∞

≤ lim sup kTm − Tn kX→Y → 0

m→∞

as n → ∞. Here we used continuity of addition and the map to the norm.

Definition 1.6 (Dual space). Let X be a normed space. We define the dual
(Banach) space as X ∗ = L(X, K).

Example: Let X = Rn with the Euclidean norm. The map

n
X
Rn 3 y → (x → xj yj ) ∈ (Rn )∗
j=1

is isometric and surjective. It allows to identify Rn and (Rn )∗ .

[19.10.2016]
[21.10.2016]

Lemma 1.7. The space C0 (Rn ) ⊂ Cb (Rn ) of functions converging to 0 at

∞ is a Banach space. Similarly the space c0 of sequences converging to 0
equipped with the sup norm is a Banach space.

Lemma 1.8. Let X be a normed space. Addition, scalar multiplication and

the map to the norm are continuous.

Lemma 1.9. The closure of a subvector space of a normed spaces is a

subvector space.

5 [February 10, 2017]

Further examples: Let 1 ≤ p ≤ q ≤ ∞. We define the sequence spaces
Definition 1.10. A K sequence (xj )j∈N is called p summable if
∞
X ∞
|xj |p < ∞ if p < ∞, sup |xj | < ∞ if p = ∞.
j=1 j=1

The set of all p summable sequences is denoted by lp (N) = lp .

Theorem 1.11. The set of p summable sequences is a vector space. The
expressions
 1/p
∞
X
k(xj )klp =  |xj |p  , p < ∞
j=1
resp.
k(xj )kl∞ = sup |xj |
j∈N

are norms on lp (N), which turn lp (N) into Banach spaces. If p1 + 1q = 1,

1 ≤ p, q ≤ ∞ and (xj ) ∈ lp , (yj ) ∈ lq then (xj yj ) is summable and Hölder’s
inequality holds:
∞
X ∞
X
xj yj ≤ |xj yj | ≤ k(xj )klp k(yj )klq .
j=1 j=1

Remark 1.12. We may replace N by Z, by a finite set, or even an arbitrary

set. Then l∞ (X) = B(X). The triangle inequality is called Minkowski
inequality.
We recall Young’s inequality
1 1
|xy| ≤ |x|p + |y|q
p q

for p1 + 1q = 1, 1 < p, q < ∞ and x, y ∈ R. Without loss of generality we

assume x, y > 0 and this can be proven by searching the maximum of
1
x → xy − xp
p

for y > 0 which is attained at x0 = y 1/(p−1) :

1 p − 1 p−1
p 1
x0 y − xp0 = y = yq .
p p q
As a consequence
X1 1 1 1
|xj |p + |yj |q = k(xj )kplp + k(yj )kqlq
X
|xj yj | ≤
p q p q
j j

6 [February 10, 2017]

and we obtain Hölder’s inequality
X X
|xj yj | =k(xk )klp k(yk )klq |xj |k(xk )k−1 −1
lp |yj |k(yk )klq
j j
1 1
=k(xk )klp k(yk )klq ( + )
p q
=k(xj )klp k(yj )klq .

Proof. Since l∞ (N) = B(N) there is nothing to show if p = ∞. Moreover

the triangle inequality is obvious if p = 1. Then, if 1 < p, q < ∞
∞
X ∞
X
p
|xj + yj | ≤ |xj + yj |p−1 |xj + yj |
j=1 j=1
X∞ ∞
X
≤ |xj + yj |p−1 |xj | + |xj + yj |p−1 |yj |
j=1 j=1

≤k(|xj + yj |p−1 )klq k(xj )klp + k(yj )klp

=k(xj + yj )kp−1
l p k(x) k
j l p + k(y )k
j l p

and
k(xj + yj )klp ≤ k(xj )klp + k(yj )klp .
provided we can devide by k(xj + yj )klp . There is nothing to show if this
quantity is 0 and it is finite whenever we sum over a finite number of indices.
Then a limit argument gives the full statement.
In particular we obtain the triangle inequality. One easily sees that
k(xj )klp = 0 implies (xj ) = 0 and

k(λxj )klp = |λ|k(xj )klp .

Thus the spaces lp are normed vector spaces. Now suppose that xn = (xn,j )
is a Cauchy sequence in lp . Then for every j, n → xn,j is a Cauchy sequence
in K. Let yj = limn→∞ xn,j and y = (yj ). Then, for every m > 1 (assuming
p < ∞, since l∞ = B(N))
N
xm kplp
X
ky − = lim |yj − xm,j |p
N →∞
j=1
N
X
= lim lim |xn,j − xm,j |p
N →∞ n→∞
j=1

≤ lim kxn − xm kplp

n→∞
→ 0 as m → ∞.

7 [February 10, 2017]

Definition 1.13. Two norms k.k and |.| on a normed spaces X are called
equivalent, if there exits C ≥ 1 so that for all x ∈ X
C −1 kxk ≤ |x| ≤ Ckxk.
Theorem 1.14. All norms on finite dimensional spaces are equivalent. Fi-
nite dimensional normed spaces are Banach spaces.
Proof. Let |.| be the Euclidean norm on Kd and k.k a second norm. Let
{ej }j=1,··· ,d be the standard basis. Then
d
X d
X √ d
X
aj ej ≤ |aj | max kek k ≤ ( d max kek k) aj e j .
k k
j=1 j=1 j=1

Thus v → k.k is continuous with respect to |.|. It attains the infimum on the
Euclidean unit sphere (which is compact). This minimum has to be positive
and we call it λ−1 . Then
|v| = |v||v/|v|| ≤ λ|v|kv/|v|k = λkvk.
The two inequalities imply the equivalence of the norms k.k and |.| by choos-
ing √
C = max d max kek k, λ, 1 .
Thus every norm on Kd is equivalent to the Euclidean norm, and any
two norms are equivalent.
A Cauchy sequence vm = (vm,j ) with respect to k.k is also a Cauchy
sequence with respect to the Euclidean norm, hence it converges to a vector
v with respect to the Euclidean norm, and hence also kvm − vk → 0.
This proves the claim for Kd . Now let X be a vector spaces of dimension
d. Then there is a basis of d vectors, and a bijective linear map φ from Kd
to X. If k.kX is a norm on X then x → kφ(x)kX is a norm on Kd . Thus the
first part follows. Since φ(xn ) is a Cauchy sequence with respect to k.kX iff
(xn ) is a Cauchy sequence in Kd with respect to the second metric, and one
converges iff the second converges.This completes the proof.

[21.10.2016]
[26.10.2016]

Lemma 1.15. Let X be a Banach space and U be a closed subvector space.

Then X/U is a vector space,
kx̃kX/U = inf ky − xk
y∈U

defines a norm (here x̃ is the equivalence class of x) and X/U is a Banach

space.

8 [February 10, 2017]

Proof. Exercise

Lemma 1.16. Let X and Y be normed spaces. Their direct sum X ⊕ Y (=

X × Y ) is a vector space. If 1 ≤ p ≤ ∞ then

k(x, y)kp = k(|x|X , |y|Y )klp

defines a norm with which X ⊕ Y becomes a Banach space.

Proof. Exercise

2 Hilbert spaces
2.1 Definition and first properties
Definition 2.1. Let X be a K vector space. A map h., .i : X × X → K is
called inner product if

hx1 + x2 , yi = hx1 , yi + hx2 , yi for all x1 , x2 , y ∈ X (2.1)

hλx, yi = λhx, yi for all x, y ∈ X, λ ∈ K (2.2)

hx, yi = hy, xi for all x, y ∈ X (2.3)
In particular hx, xi ∈ R for all x ∈ X and

hx, xi ≥ 0 for all x ∈ X (2.4)

hx, xi = 0 ⇐⇒ x=0 (2.5)

Examples:

1. Euclidean vector spaces over K, Euclidean inner product.

2
P and complex square summable sequences space l (N) with h(xj ), (yj )i =
2. Real
x j yj .
´
3. Let U ⊂ Rn be measurable, X = Cb (U ), hf, gi = U f gdmn where mn
denotes the Lebesgue measure.

Lemma 2.2 (Cauchy-Schwarz). Let X be a vector space with inner product.

Then 1
|hx, yi| ≤ (hx, xihy, yi) 2
for all x, y ∈ X.

9 [February 10, 2017]

Proof. Let x, y ∈ X and λ ∈ K. Then

0 ≤hx − λy, x − λyi

=hx, xi − λhy, xi − λ̄hx, yi + |λ|2 hy, yi.

hx,yi
If y = 0 there is nothing to show, so we assume y 6= 0 and define λ = hy,yi .
Then
|hx, yi|2
0 ≤ hx, xi −
hy, yi
which implies the Cauchy-Schwarz inequality.

Lemma 2.3. The map

p
x → kxk := hx, xi

defines a norm.

With this notation the Cauchy-Schwarz inequality becomes

|hx, yi| ≤ kxkkyk.

Proof. Clearly kxk ≥ 0, kxk = 0 iff x = 0. Moreover

kλxk2 = |λ|2 kxk2

and by the Cauchy-Schwarz inequality

kx + yk2 =kxk2 + hx, yi + hy, xi + kyk2

≤kxk2 + 2kxkkyk + kyk2
=(kxk + kyk)2 .

Definition 2.4. A vector space X with an inner product is called pre-Hilbert

space. It is a Hilbert space if it is a Banach space.

Lemma 2.5. The inner product defines a continuous map from X × X to

K.

Proof. Exercise

It is not hard to verify the parallelogram identity

kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 (2.6)

for x, y ∈ H some pre-Hilbert space.

10 [February 10, 2017]

Theorem 2.6 (Jordan von Neumann). Let X be a normed K vector space
with norm k.k. We assume that the norm satisfies the parallelogram identity

kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 (2.7)

Then
1
hx, yi = kx + yk2 − kx − yk2 (2.8)
4
if K = R and
1
hx, yi = kx + yk2 − kx − yk2 + ikx + iyk2 − ikx − iyk2 (2.9)
4
otherwise defines an inner product such that the norm is the norm of the
preHilbert space. Vice versa: The norm of a prehilbert space defines the
parallelogram identity.

As a consequence we could define a Hilbert spaces as a Banach space

whose norm satisfies the paralellogram identity. By an abuse of notation
we call a normed space pre-Hilbert space if it satisfies the parallelogram
identity.

Proof. We begin with a real normed spaces whose norm satisfies the paral-
lelogram identity. We define
1
hx, yi = (kx + yk2 − kx − yk2 ).
4
Then
hx, yi = hy, xi.
Since by the parallelogram identity

2kx + zk2 + 2kyk2 = kx + y + zk2 + kx − y + zk2

hence
kx + y + zk2 =2kx + zk2 + 2kyk2 − kx − y + zk2
=2ky + zk2 + 2kxk2 − ky − x + zk2

and
1 1
kx+y +zk2 = kxk2 +kyk2 +kx+zk2 +ky +zk2 − kx−y +zk2 − ky −x+zk2
2 2
1 1
kx+y −zk2 = kxk2 +kyk2 +kx−zk2 +ky −zk2 − kx−y −zk2 − ky −x−zk2
2 2

11 [February 10, 2017]

and we arrive at
1
hx + y, zi = (kx + y + zk2 − kx + y − zk2 )
4
1 1
= (kx + zk2 − kx − zk2 ) + (ky + zk2 − ky − zk2 )
4 4
=hx, zi + hy, zi.

We claim
hλx, yi = λhx, yi
for all x, y ∈ X and λ ∈ R. It obviously holds for λ = 1 by checking the
definition, and for all λ ∈ N be the previous step, hence for all λ ∈ Z. But
then it holds for all rational λ and by continuity for λ ∈ R.
We complete the proof for complex Hilbert spaces: We define
3
1X k
hx, yi = i kx + ik yk2
4
k=0

and observe that hix, yi = ihx, yi, hx, yi = hy, xi by definition, Rehx, yi is
the previous real inner product and Imhx, yi = Rehx, iyi.

Corollary 2.7. A normed space is a pre-Hilbert space if and only if all two
dimensional subspaces are pre-Hilbert spaces.
Proof. It is a pre-Hilbert space if and only if its norm satisfies the parallelo-
gram identity which holds if and only if the parallelogram identity holds for
all two dimensional subspaces.

This has geometric consequences.

Lemma 2.8. Let H be a Hilbert space, K ⊂ H compact, and C ⊂ H closed
and convex, C and K disjoint. Then there exist x ∈ K and y ∈ C so that

kx − yk = d(C, K)

Proof. Let xj ∈ K and yj ∈ C be a minimizing sequence. Since K is compact

there is a subsequence which we denote again by (xj , yj ) and x ∈ K so that
xj → x. By the triangle inequality

kx − yj k → d(C, K).

Then
kyn − ym k2 =k(x − yn ) − (x − ym )k2
=2kx − yn k2 + 2kx − ym k2 − k2x − (yn + ym )k2
≤2(kx − yn k2 + kx − ym k2 ) − 4d2 (C, K)
→0 as n, m → ∞

12 [February 10, 2017]

since by convexity 12 (yn + ym ) ∈ K. Thus (yn ) is a Cauchy sequence with
limit y ∈ C. Moreover

d(C, K) = lim kx − yn k = kx − yk.

n→∞

[26.10.2016]
[28.10.2016]

Definition 2.9. We call two elements x, y ∈ H orthogonal if hx, yi = 0 .

Lemma 2.10. Suppose that C is a closed and convex subset of a Hilbert
space H and x ∈ H. Then the closest point in C to x is unique and we
denote it by p(x). Moreover, if y = p(x) then

Rehx − y, z − yi ≤ 0 (2.10)

for all z ∈ C. If y ∈ C satisfies this inequality for all z ∈ C then y = p(x).

If C is a closed subspace then for all z ∈ C

hx − p(x), zi = 0. (2.11)

The point y = p(x) ∈ C is uniquely determined by this orthogonality condi-

tion. Moreover
kxk2 = kx − p(x)k2 + kp(x)k2 .
Proof. Uniqueness is a consequence of the proof of Lemma 2.8. Let y ∈ C.
Then by the triangle inequality, if z ∈ C then for 0 ≤ t ≤ 1

y(t) = y + t(z − y) ∈ C

and hence if y = p(x),

kx − yk2 ≤ kx − y − t(z − y)k2 = kx − yk2 − 2t Rehx − y, z − yi + t2 kyk2 .

This implies (2.10) and also the converse. In the case that C is a closed
subspace (2.10) is equivalent to the orthogonality relation.

2.2 The Riesz representation theorem

Theorem 2.11 (Riesz representation theorem). Let H be a Hilbert space.
Then
J : H 3 x → (y → hy, xi) ∈ H ∗
is an R linear isometric isomorphism. It is conjugate linear:

J(λx) = λJ(x).

13 [February 10, 2017]

Proof. By the Cauchy-Schwarz inequality
kJ(x)kH ∗ = sup |hy, xi| ≤ kxkH
kyk≤1

and the map is well defined and antilinear. Since

kxkH kJ(x)kH ∗ ≥ hx, J(x)i = hx, xi = kxk2H
we see that
kJ(x)kH ∗ ≥ kxkH .
Thus J is an isometry: kJ(x)kH ∗ = kxkH . In particular J is injective. To
show that J is surjective we assume that x∗ ∈ H ∗ and try to find x so that
x∗ = J(x). Let
N = {y ∈ H : x∗ (y) = 0}.
Then N is a subvector space and it is closed. Let p be the orthogonal
projection to N as above. We choose y0 ∈ H with x∗ (y0 ) = 1 and define
x0 = y0 − p(y0 ).
Then x∗ (x0 ) = 1 and for all y ∈ N by (2.11) hy, x0 i = 0. Moreover x∗ (x −
x∗ (x)x0 ) = 0 hence x − x∗ (x)x0 ∈ N and by (2.11)
hx − x∗ (x)x0 , x0 i = 0.
Since x = [x − x∗ (x)x0 ] + x∗ (x)x0 , then
hx, x0 i = hx∗ (x)x0 , x0 i = x∗ (x)kx0 k2H
and D x0 E x0
x∗ (x) = x, 2
= J( )(x).
kx0 k kx0 k2

Theorem 2.12 (Lax-Milgram). Let H be a Hilbert space and

Q : H × H 3 (x, y) → Q(x, y) ∈ K
be linear in x, antilinear in y, bounded in the sense that
|Q(x, y)| ≤ Ckxk|yk
and coercive in the sense that there exists δ > 0 so that
Re Q(x, x) ≥ δkxk2 .
Then there exists a unique continuous linear map A : H → H with contin-
uous inverse A−1 so that
Q(x, y) = hAx, yi.
Moreover
kAkH→H ≤ C, kA−1 kH→H ≤ δ −1 .

14 [February 10, 2017]

Proof. Let x ∈ H. Then

y → Q(x, y) ∈ H ∗ .

By the Riesz representation theorem there exists a unique z(x) ∈ H so that

hz(x), yi = Q(x, y)

for all y ∈ H. Then

kz(x)k = sup |hz(x), yi| = sup Q(x, y) .

kyk≤1 kyk≤1

Clearly z(x1 + x2 ) = z(x1 ) + z(x2 ) and z(λx) = λz(x) and we define the
continuous linear operator Ax = z. Since

RehAx, xi ≥ δkxk2

we obtain
kAxk ≥ δkxk.
It particular A is injective and the range is closed. If it is not surjective
there exists z with kzk = 1 and z is orthogonal to the range, i.e.

hAx, zi = 0

for all x ∈ X. In particular we reach the contradiction

0 = hAz, zi ≥ δkzk2 .

2.3 Orthonormal systems

We recall that N elements xj of a vector space X are called linearly inde-
pendent if
XN
λ j xj = 0
j=1

implies λj = 0. Let xj be N linearly independent vectors of a Hilbert space.

By the Gram-Schmidt procedure we obtain an orthonormal system with
1
y1 = x1
kx1 k

ỹ2 = x2 − hx1 , y1 iy1

1
y2 = ỹ2
kỹ2 k

15 [February 10, 2017]

recursively. We can do this with N = ∞.
Examples: Orthonormal polynomials. Let µ : R → (0, ∞) be measurable
so that all moments exists, i.e.
ˆ
|x|N µdx < ∞
R

for all N ≥ 0. Let

X = {f ∈ C(R) : (1 + |x|)−N |f | is bounded for all N }.
Then ˆ
X × X 3 (f, g) → hf, gi := f gµdx

defines an inner product. The monomials

fn = xn
are linearly independent.
We consider the case

1/2 if |x| ≤ 1
µ(x) =
0 otherwise
with X = C([−1, 1]). It leads to (multiples) of the Legendre polynomials.
Definition 2.13 (Legendre-polynomial). The Legendre polynomial Pn is
the unique polynomial of degree n with Pn (1) = 1 and
ˆ 1
xm Pn (x)dx = 0
−1

for all 0 ≤ m < n.

There is a very compact formula for them.
Lemma 2.14 (Rodrigues formula).
1 dn
Pn (x) = [(x2 − 1)n ]
2n n! dxn
Proof. The degree of Pn defined by the right hand side is obviously n and
the leading term of Pn (x) reads 2n(2n)!
(n!)2
xn . If 0 ≤ m < n then after m
integrations by parts
ˆ 1 n ˆ 1 n−m
m d 2 n m d
x n
(x − 1) dx = (−1) m! n−m
(x2 − 1)n dx = 0.
−1 dx −1 dx

Finally
dn 2 dn
(x − 1)n = 2n (x − 1)n = 2n n!.
dxn x=1 dxn x=1

16 [February 10, 2017]

A more difficult calculation gives
ˆ 1 ˆ 1 n
2 1 (2n)! n d
(Pn (x)) dx = n n (n!)2
x n
(x2 − 1)n dx
−1 2 n! 2 −1 dx
ˆ 1
(2n)! n
= 2n (−1) (x2 − 1)n dx
2 (n!)2 −1
ˆ 1
(2n)! 2
= 2n 2 · 22n sn (1 − s)n ds = .
2 (n!)2 0 2n +1

and r r ˆ 1
2n + 1 2m + 1
Pn (x)Pm (x) dx = δn,m
2 2 −1
and the functions r
2n + 1
Pn (x)
2
are an orthonormal system.
[28.10.2016]
[02.11.2016]

We consider the Banach space C([0, 2π]; C) with inner product

ˆ
1
hf, gi = f gdx.
2π
It is not hard to see that this is an inner product.

Lemma 2.15. The functions einx are an orthonormal system.

Proof. We compute
ˆ 2π ˆ 2π
1 inx −imx 1
e e dx = ei(n−m)x dx = 0
2π 0 2π 0
if n 6= m and = 1 if n = m.

Sturm-Liouville problems lead to something very similar. Let q ∈ C([0, 1], R)

and let λ ∈ C. We consider the boundary value problem

− u00 + qu = λu in (0, 1) u(0) = u(1) = 0 (2.12)

We have proven in Einführung in die PDG:

Theorem 2.16. Given λ ∈ C the space of solutions to (2.12) is vector

space of dimension 0 or 1. There exists a monotone sequence λn → ∞ and
a sequence of real valued functions un ∈ C 2 ([0, 1]) which satisfy

−u00n + qun = λn un

17 [February 10, 2017]

ˆ
un um dx = δnm .

The functions un have n − 1 zeroes in (0, 1).If λ 6= λn for some j then (2.12)
has only trivial solutions.

We repeat the proof of orthogonality:

ˆ ˆ ˆ ˆ
λn un um dx = (−u00n +qun )um dx = un (−u00m + qum )dx = λ̄m un ūm dx
´
which implies either λn = λm (since λm is real) or un um dx = 0.

Lemma 2.17 (Bessel inequality). Let xj be an orthonormal system. Then

N
X N
X
0 ≤ kxk2 − |hx, xn i|2 = kx − hx, xn ixn k2
n=1 n=1

Proof. Let M ⊂ H be the N dimensional subspace spanned by the elements

xj and let p be the projection to the closest point. Then by Lemma 2.10

kxk2 = kx − p(x)k2 + kp(x)k2 .

Moreover
N
X
p(x) = λj xj
j=1

and
N
X
hx, xn i = hp(x), xn i = λm hxm , xn i = λn
m=1

and
∞
X
2
kp(x)k = |λn |2 .
n=1

Definition 2.18. A subset A of a metrix space X is called dense if its

closure is X. A metric space X is called separable, if there is a countable
dense subset.

Examples:

1. N is countable.

2. X and Y countable implies X × Y is countable. In particular Qd is

countable.

3. If Xj are countable sets then there union is countable.

18 [February 10, 2017]

4. Subsets of countable sets are countable.

5. Q is countable.

6. QN is countable.

7. RN is separable since QN is countable and dense.

8. l2 (N) is separable.

Definition 2.19. A orthonormal set xj of a Hilbert space is called orthonor-

mal basis if
hx, xj i = 0 for all j ∈ N
implies x = 0.

Theorem 2.20. The following properties are equivalent for a Hilbert space
H which is not finite dimensional.

• The space H is separable.

• There exists an orthonormal basis and kxk2 = |hx, xj i|2 .

P

• The exists a surjective isometry l2 → H

Proof. Suppose that H is separable. Let (yn ) be a dense sequence and XN

the span of the first N yn . Its dimension is at most N . We use the Gram-
Schmidt procedure to find an orthonormal basis (xn ) of XN . We do this
recursively by increasing n. This leads to a countable orthonormal sequence
(xn ) so that its span is dense. Now let x ∈ X. By the Bessel inequality
N
X N
X
2
hx, xj i + kx − hx, xj ixj k2 = kxk2 .
j=1 j=1

Thus
N
X
N → kx − hx, xn ixn k
n=1

is monotonically decreasing. Since (yn ) is dense and N

P
n=1 hx, xn ixn = pN x
is the closed point in the span of (xn )n≤N it converges to 0, which is equiv-
alent to
XN
hx, xn ixn → x
n=1

in L2 , which in turn implies

∞
X
kxk2 = |hx, xn i|2 .
n=1

19 [February 10, 2017]

Now suppose that (xn ) is an orthonormal basis. We want to define
∞
X
2
l 3 (an ) → an xn ∈ H.
n=1

We claim that for M ≥ N

N
X M
X M
X
an xn − an xn = an xn
n=1 n=1 n=N +1

the norm of which is given by

v
u M
u X
t |aj |2 .
n=N +1

This implies that the partial sums are a Cauchy sequence and we define
N
X
x := lim an xn
N →∞
j=n

Then
N
X
2
kxk = lim |an |2 = k(an )k2 .
N →∞
n=1
The map is clearly linear, surjective (by the previous part) and an isometry.
To complete the proof it suffices to prove that l2 (N) is separable. Al-
most by definition the (countable) union of the subspaces of dimension N of
sequence being 0 for indices > N are dense: The truncated series converge
in L2 . It suffices now to find a dense countable subset of RN . QN is an
obvious choice.

In particular a Hilbert space is either isomorphic (there exists an isomet-

ric surjective linear map) to Rd resp. Cd , to l2 , or it is not separable.
Example: The space l2 (R) with inner product
X
hf, gi = f (x)g(x)
x∈R

is not separable since the vectors

x 0 if y 6= x
ey =
1 if y = x
are√an uncountable orthonormal system. In particular the pairwise distance
is 2 and there cannot be a countable dense set.
There are natural questions:

20 [February 10, 2017]

• Which of the concrete orthonormal systems constructed above are a
basis? We will see that the answer is all of them, but we need more
tools to prove this.

• Is there a good theory of not necessarely orthonormal basis? This is

more tricky.

[02.11.2016]
[04.11.2016]

3 Lebesgue spaces
3.1 Review of measure spaces
Reference:

1. Alt: Linear functional analysis, Springer.

2. Lieb and Loss: Analysis, AMS 2001.

3. Sharkarchi and Stein: Real Analysis: Measure theory, Integration and

Hilbert spaces. Princeton University Press. 2009.

Theorem 3.1 (Banach-Tarski). There exists finitely many pairwise dis-

joints sets An , Bm of R3 and isometric maps φj , ψj : R3 → R3 so that
N
[ M
[ N
[ M
[
B1 (0) = φn (Bn ) = ψm (Am ) = Bn ∪ Am
n=1 m=1 n=1 m=1

Remark: Makes use of the axiom of choice.

Definition 3.2. Let X be a set. A family of subset A is called a σ algebra

if

1. {} ∈ A

2. A ∈ A implies X\A ∈ A
∞
S
3. An ∈ A implies An ∈ A.
n=1

A map µ : A → [0, ∞] is called a measure if whenever An ∈ A are pairwise

disjoint then
[∞ X ∞
µ An = µ(An ).
n=1 n=1

The triple (X, A, µ) is called a measure space.

21 [February 10, 2017]

Examples:
1. X a set, A = 2X the set of all subsets, and µ(A) the number of
elements.

2. If (X, d) is a metric space then there is a smallest σ algebra containing

all open sets. It is called the Borel σ algebra of X.

3. In probability theory the σ algebra encodes the available information

on a system.

4. X = Rn , A the Borel sets, µ the Lebesgue measure restricted to the

Borel sets.

5. X = Rn , A the Lebesgue sets, µ the Lebesgue measure.

6. X = Rn , 0 ≤ s ≤ n, Hs , A the Borel sets, Hs the Hausdorf measure.

Definition 3.3. Let X be a set and A a σ algebra. A map f : X →
R ∪ {−∞, ∞} is called measurable if

f −1 ((t, ∞]) ∈ A

for all t ∈ R. If (X, A, µ) is a measure space and f : X → [0, ∞] then we

define the Lebesgue integral by the Riemann integral
ˆ ˆ ∞
f dµ = µ(f −1 ((t, ∞]))dt ∈ [0, ∞]
0

We call a measurable function f integrable if |f | is integrable. Let 1 ≤ p <

∞. We call a measurable function f p integrable if |f |p is integrable and
denote ˆ 1/p
kf kLp = |f |p dµ .

We call a measurable function ∞ integrable or essentially bounded if there

is a constant C so that

µ({x : |f (x)| > C}) = 0.

The best constant is denoted by kf kL∞ .

There are convergence theorems about the relation between the limit of
integrals, and the integral over limits: The theorem of Lebesgue on domi-
nated convergence, the Lemma of Fatou and the theorem of Beppo Levi on
monoton convergence.
Definition 3.4. A measure space (X, A, µ) is called sigma finite if there
∞
S
exists a sequence of measurable sets Aj of finite measure so that X = An .
n=1

22 [February 10, 2017]

3.2 Construction of measures
Measures are often constructed by first constructing outer measures.
Definition 3.5. Let X be a set. An outer measure µ maps subsets of X to
[0, ∞] so that
1. µ({ }) = 0.

2. A ⊂ B implies µ(A) ≤ µ(B).

∞
S P
3. µ Aj ≤ ∞ j=1 µ(Aj ).
j=1

Examples:

• Let X = Rd . We define the measure of a coordinate rectangle as the

product of the sidelengths and the measure of a countable disjoint
union of coordinate rectangles as the sum over the measures of the
rectangles. Finally we define the outer measure of a general set as the
infimum of all measures of coverings by unions of coordinate rectancles.

• If (X, A, µ) and (Y, B, ν) are σ finite measure spaces we define rectan-

gles in the cartesian product as the cartesian product of measurable
sets and their measure as the product of the measures. Then we pro-
ceed in the same way as above to obtain an outer measure on X × Y .

• The Hausdorff measure: Let X be a metric space and s ≥ 0. We define

the premeasure of a set A of diameter r

π s/2
φ(A) = 2−s rs
Γ( 2s + 1)

and the Hausdorff measure

∞
nX ∞
[ o
A = inf φ(An ) : A ⊂ An .
n=1 n=1

Definition 3.6. Let X be a metric space. We call µ an outer metric measure

if it is an outer measure which satisfies

µ(A ∪ B) = µ(A) + µ(B)

for all A, B ⊂ X with dist(A, B) > 0.

Definition 3.7. Let µ be an outer measure on X. We call a subset A ⊂ X
Caratheodory measurable if for all B ⊂ X

µ(B) = µ(B ∩ A) + µ(B ∩ (X\A)).

23 [February 10, 2017]

Theorem 3.8 (Caratheodory). Let µ be an outer measure on the set X.
Then the Caratheodory measurable sets C are a σ algebra and (X, C, µ|C ) is
a measure space. Moreover C contains all sets of exterior measure 0. If X
is a metric space and µ is a metric outer measure than C contains all open
sets. In the case of the Cartesian product C contains all Cartesian products
of measurable sets.

Theorem 3.9 (Fubini-Tonelli). Let (X, A, µ) and (Y, B, ν) be σ-finite mea-

sure spaces, A × B the product σ algebra and µ × ν the product measure. Let
f be ´µ × ν integrable. Then for almost of x ∈ X y → f (x, y) is ν integrable,
x → Y f (x, y)dν(y) is µ integrable and
ˆ ˆ ˆ
f (x, y)dµ × ν = f (x, y)dν(y)dµ(x).
X×Y X Y

[04.11.2016]
[09.11.2016]

3.3 Jensen’s and Hölder’s inequalities

Lemma 3.10. Suppose that f : R → R is convex. Then both one sided
derivatives exist and if x < y then

df + df − df +
(x) ≤ (y) ≤ (y)
dx dy dy
and for all z

df + df −
f (z) ≥ max{f (x) + (x)(z − x), f (x) + (x)(z − x)}.
dx dx
Proof. If x0 < x1 < x2 then

f (x1 ) − f (x0 ) f (x2 ) − f (x0 ) f (x2 ) − f (x1 )

≤ ≤
x1 − x0 x2 − x0 x2 − x1
and
f (x + h) − f (x)
h→
h
is monotonically increasing. This implies the differentiability from the right,
and similarly from the left and the relation between the derivatives.

Lemma 3.11. Let (X, A, µ) be a measure space, µ(X) = 1, F : R → R

convex and f real valued and integrable. Then F ◦ f is measurable, (F ◦ f )−
is integrable and ˆ ˆ
F◦ f dµ ≤ F ◦ f dµ
X X

24 [February 10, 2017]

´
Proof. Let t0 = X f dµ. Since F : R → R is convex, we have for any t

dF +
F (t) ≥ F (t0 ) + (t0 )(t − t0 ).
dt
Thus
dF +
µ({F ◦ f ≤ s}) ≤ µ({F (t0 ) + (t0 )(f − t0 ) ≤ s})
dt
+
and min{F ◦ f, 0} is integrable since x → F (t0 ) + dF dt (t0 )(f − t0 ) (which is
affine in f ) is integrable. Then
ˆ ˆ
dF +
F ◦ f dµ ≥ F (t0 ) + (t0 )(f − t0 )dµ
X X dt
ˆ
dF +

=F (t0 ) + (t0 ) f dµ − t0 = F (t0 ).
dt X

We call a function f : X → C integrable if the real and imaginary parts

are both integrable. We say a property holds almost everywhere, if it holds
outside a set of measure 0.
1 1
Lemma 3.12. Let 1 ≤ p, q ≤ ∞ and p + q = 1. If f ∈ Lp and g ∈ Lq then
f g is integrable and ˆ
f gdµ ≤ kf kLp kgkLq .

If 1 < p < ∞, then the equality implies that g = λ|f |p−2 f¯ almost everywhere
for some λ ∈ K with |λ| = 1.
Proof. We copy the proof basically from the one for the sequence space. As
there it suffices to consider f and g with kf kLp = 1 and kgkLq = 1 and prove
ˆ ˆ
1 p 1 q
|f ||g|dµ ≤ |f | + |g| dµ = 1.
p q
The inequality is strict unless
1 1
|f g(x)| = |f (x)|p + |g(x)|q
p q
almost everywhere, which implies |g| = |f |p−1 . Now
ˆ ˆ
f gdµ ≤ |f g|dµ ≤ kf kLp kgkLq

and in the case of equality all inequalities must be equalities. Hence |g| =
|f |p−1 . Now suppose for some integrable function h
ˆ ˆ
hdµ = |h|dµ.

25 [February 10, 2017]

´
Then there exits λ ∈ C with |λ| = 1 so that hdµ ∈ [0, ∞) and
ˆ ˆ ˆ
λ−1 hdµ = Re λ−1 hdµ = |h|dµ

and hence
h = λ|h|
almost everywhere. Back to our situation above this implies

g = λ|f |p−2 f¯

almost everywhere for some complex number of modulus 1.

3.4 Minkowski’s inequality

Theorem 3.13. Let 1 ≤ p < ∞ and let X and Y be spaces with σ finite
measures µ and ν respectively. Let f be µ × ν measurable. Then
ˆ ˆ p 1/p ˆ ˆ 1/p
|f (x, y)|dν(y) dµ(x) ≤ |f (x, y)|p dµ(x) dν(y).
X Y Y X

If 1 < p, if the integrals above are finite, and if

ˆ ˆ p 1/p ˆ ˆ 1/p
p
f (x, y)dν(y) dµ(x) = |f (x, y)| dµ(x) dν(y).
X Y Y X

then there exist a µ-measurable function α and a ν-measurable function β

so that
f (x, y) = α(x)β(y)
almost everywhere. A special case is the triangle inequality (which holds
without assuming σ finiteness)

kf + gkLp (µ) ≤ kf kLp (µ) + kgkLp (µ) ,

whenever f and g are p-integrable, with equality for p > 1 iff f and g are
linearly dependent.

Proof. We assume first that f is nonnegative and omit the absolute value.
We claim that
ˆ ˆ
p
y→ f (x, y)dµ(x) and H(x) := f (x, y)dν(y)
X Y

are measurable functions. This follows from the Theorem of Fubini if f resp
f p are µ × ν integrable, and by an approximation argument in the general

26 [February 10, 2017]

case. Then
ˆ ˆ ˆ
p
H (x)dµ(x) = f (x, y)dν(y)H(x)p−1 dµ(x)
X ˆ ˆ
X Y

= f (x, y)H p−1 (x)dµ(x)dν(y)

Y X
ˆ ˆ 1/p ˆ p−1
p
p p
≤ f (x, y)dµ(x) H dµ(x) dν(y)
Y X
ˆ ˆ 1/p ˆ p−1
p
p p
= f (x, y)dµ(x) dν(y) H dµ(x)
Y X
p
where we used Hölder’s inequality with q = p−1 .
We want to divide by the right hand. We can do that whenever the left
hand side is neither 0 nor ∞, and we can achieve that in the same fashion
as for sequences.
Now assume that p > 1, f is complex valued and integrable. Then, with
ˆ ˆ p ˆ ˆ p
f (x, y)dν(y) dµ(x) ≤ |f (x, y)|dν(y) dµ(x)
X Y X Y

and we continue as in the previous step, assuming equality. Then we have

equality in the application of Hölder’s inequality
ˆ
|f (x, y)| = α0 (x) |f (x, y 0 )|dν(y 0 )
Y

for almost all x and y. Since we must have also equality in the equality
above we must have
f (x, y) = α(x)β(y)
for some measurable function α and β.
For the last part we apply the first part with the counting measure on
Y = {0, 1}. The product measure is defined in the obvious fashion, even
without assuming that µ is σ finite. If f is p integrable then by the definition
of the integral
µ({x : |f (x)| > t}) ≤ t−p kf kpLp
Let
∞
[ 1
A= {x : |f (x)| + |g(x)| > }
j
j=1

which is a countable union of sets of finite measure. We replace X by A,

take as σ algebra the sets in A which are contained in A, and µ restricted
to this σ algebra as measure. This is σ additive.

27 [February 10, 2017]

3.5 Hanner’s inquality
There is an improvement of the triangle inequality.

Theorem 3.14 (Hanner’s inequality). Let (X, A, µ) be a measure space and

f , g be p-integrable functions, 1 < p < ∞. If 1 ≤ p ≤ 2 then
p
kf + gkpLp + kf − gkpLp ≥ (kf kLp + kgkLp )p + kf kLp − kgkLp , (3.1)
p
(kf +gkLp +kf −gkLp )p + kf +gkLp −kf −gkLp ≤ 2p (kf kpLp +kgkpLp ). (3.2)
If 2 ≤ p < ∞ all inequalities are reversed.

The inequalities reduce to the parallelogram identity if p = 2. Both are

equivalent: The second is obtained from the first by replacing f by f + g
and g by f − g. It suffices to prove the first inequality.

Proof. We may assume that kgkLp ≤ kf kLp (otherwise we exchange the two)
and kf kLp = 1 (otherwise we multiply f and g by the inverse of the norm).
The first inequality follows from the following pointwise inequality: Let

α(r) = (1 + r)p−1 + (1 − r)p−1 , β(r) = [(1 + r)p−1 − (1 − r)p−1 ]r1−p .

We claim that

α(r)|f |p + β(r)|g|p ≤ |f + g|p + |f − g|p (3.3)

for 1 ≤ p ≤ 2, 0 ≤ r ≤ 1 and complex numbers f and g (and the reverse

inequality for 2 ≤ p < ∞). Indeed, (3.3) implies

α(r)|f (x)|p + β(r)|g(x)|p ≤ |f (x) + g(x)|p + |f (x) − g(x)|p

and by integration

α(r)kf kpLp + β(r)kgkpLp ≤ kf + gkpLp + kf − gkpLp .

We apply the inequality with r = kg|Lp and recall that kf kLp = 1. The left
hand side becomes
h i
(kf kLp + kgkLp )p−1 + (kf kLp − kgkLp )p−1 kf kLp
h i
+ (kf kLp + kgkLp )p−1 − (kf kLp − kgkLp )p−1 ) kgkLp
=(kf kLp + kgkLp )p + (kf kLp − kgkLp )p .

It remains to prove (3.3).

Let for 0 ≤ R ≤ 1,

FR (r) = α(r) + β(r)Rp .

28 [February 10, 2017]

We claim that it attains its maximum at r = R if 1 ≤ p < 2 and resp. its
minimum if p > 2. We compute

FR0 = α0 + β 0 R = (p − 1)[(1 + r)p−2 − (1 − r)p−2 ](1 − (R/r)p )

and the derivative vanishes only at r = R and changes sign there. Thus

α(r) + β(r)Rp ≤ (1 + R)p + (1 − R)p

if 0 ≤ R ≤ 1 and p ≤ 2 with the opposite inequality if p ≥ 2. Now let R ≥ 1.

Since β ≤ α if p ≤ 2 we obtain

α(r)+β(r)Rp ≤ Rp α(r)+β(r) ≤ Rp [(1+R−p )+(1−R−p )] = (1+Rp )+(Rp −1)

and the reverse inequality if p > 2. This implies (3.3) for real f and g. We
claim that (3.3) holds for complex f and g. It suffices to consider f = a > 0
and g = beiθ . Since

(a2 + b2 + 2ab cos θ)p/2 + (a2 + b2 − 2ab cos θ)p/2

has its minimum at θ = 0 (resp. its maximum if p ≥ 2) since x → xp/2 is

concave if p ≤ 2 (convex if p ≥ 2).

[09.11.2016]
[11.11.2016]

3.6 The Lebesgue spaces Lp (µ)

Lemma 3.15. The set of p-integrable functions is a vector space. The
Minkowski inequality

kf + gkLp ≤ kf kLp + kgkLp

holds. Moreover
kλf kLp = |λ|kf kLp
and
kf kLp = 0
if and only if f vanishes outside a set of zero

µ({f 6= 0}) = 0.

These functions are p integrable for all p. They are a subvector space.

Proof. The vector space property follows from the Minkowski inequality.
The other statements are obvious.

29 [February 10, 2017]

Definition 3.16. We call two measurable functions equivalent f ∼ g if
µ({f 6= g}) = 0. We define Lp (µ) as the space of equivalence classes of p
integrable functions.
If f ∼ g then kf − gkLp = 0. The equivalence relation is compatible with
the vector space structure.
Theorem 3.17. [Fischer-Riesz] The space Lp (µ) is Banach space.
Proof. It is straight forward to verify that Lp (µ) is a vector space (using
Minkowksi’s inequality), and that k.kLp is a norm. Completeness is more
involved.
Let fn be representatives of a Cauchy sequence. By taking subsequences
if necessary we may assume
kfn − fm kLp ≤ 2− min{m,n} .
We define the monotone sequence of functions
n−1
X
Fn (x) = |f1 (x)| + |fm+1 (x) − fm (x)|
m=1

and F = limn→∞ Fn (x). F is measurable and by monotone convergence

ˆ ˆ
p
|F | dµ = lim |Fn |p dµ ≤ kf1 kpLp + 1
n→∞

and in particular it is finite almost everywhere. Thus

n−1
X
fn = f1 + (fm+1 − fm )
m=1

converges if F (x) < ∞. Let f be the limit if F (x) < ∞, and 0 otherwise. It
is measurable. Since max{f, fn } ≤ F we obtain by dominated convergence
ˆ
p
kf − fn kLp = |f − fn |p dµ → 0.

3.7 Projections and the dual of Lp (µ)

Lemma 3.18. Let 1 < p < ∞ and let K be a closed convex set in Lp (µ).
Let f ∈ Lp (µ). Then there exists a unqiue g ∈ K with
kf − gkLp (µ) = dist(f, K).
Moreover
ˆ
Re (h − g)(f¯ − ḡ)|f − g|p−2 dµ ≤ 0, ∀h ∈ K.
X

30 [February 10, 2017]

Proof. Let hn be a minimzing sequence. Since 12 (hn + hm ) ∈ K and
khn − f + hm − f kLp ≤ khn − f kLp + khm − f kLp , we see that

khn − f + hm − f kLp → 2 dist(f, K).

Now let p ≤ 2, from the second Hanner’s inequality we obtain

p
khn − f + hm − f kLp + khn − hm kLp
p
+ khm − f + hm − f kLp − khn − hm kLp
≤2p khn − f kpLp + khm − f kpLp .

Let A = lim supn,m→∞ khn − hm kLp . This limsup is obtained along two
subsequences n, m → ∞. Let D = dist(f, K). Then

(2D + A)p + |2D − A|p ≤ 2p+1 Dp

which implies A = 0 by the strict convexity of A → |2D + A|p .

If p > 2 we argue similarly with the first inequality.
Now let g ∈ K be the point of minimal distance and let h ∈ K. Let
ˆ
N (t) = |f − (g + t(h − g))|p dµ.

Then N (t) attains its minimum at t = 0 on the interval [0, 1]. We claim
that its derivative at t = 0 is
ˆ
d
N |t=0 = p Re |f (x) − g(x)|p−2 (f (x) − g(x))(ḡ(x) − h̄(x))dµ.
dt
This implies the assertion.
To calculate the derivative we assume that f, g ∈ Lp (µ) and define

N (t) = kf + tgkpLp .

Since almost everywhere

d
|t=0 |f + tg|p = p|f |p−2 Re f ḡ
dt
and the pth power is convex
1
|f |p − |f − g|p ≤ (|f + tg|p − |f |p ) ≤ |f + g|p − |f |p ,
t
the formula follows by dominated convergence.
1
Theorem 3.19. Let (X, A, µ) be a measure space, 1 < p, q < ∞, p + 1q = 1.
Then ˆ
j : L 3 g → (f → f gdµ) ∈ (Lp (µ))∗
q

is a conjugate linear isometric isomorphism.

31 [February 10, 2017]

The proof is the same as for Hilbert spaces: By Hölder’s inequality the
map is well defined and

kj(f )k(Lp )∗ ≤ kf kLq .

Since ˆ
j(f )(|f |q−2 f¯) = |f |q dµ

we conclude as for Hilbert spaces that

kj(f )k(Lp )∗ ≥ kf kLq .

Surjectivity is proven exactly as for Hilbert spaces.

[11.11.2016]
[16.11.2016]

Corollary 3.20. Suppose that µ is σ finite. Then

ˆ
L 3 g → (f → f gdµ) ∈ (L1 (µ))∗
∞

is an isometric isomorphism.

The proof is an exercise on sheet 5.

3.8 Young’s inequality and Schur’s lemma

Let (X, A, µ) be a measure space and suppose that 1 ≤ p, q, r ≤ ∞ and
1 1 1 p q r
p + q + r = 1. If f ∈ L (µ), g ∈ L (µ) and h ∈ L (µ), then f gh is integrable
and ˆ
f ghdµ ≤ kf kLp (µ) kgkLq (µ) khkLr (µ) .

This is a consequence of a multiple application of Hölder’s inequality:

ˆ
f ghdµ ≤ kf kLp kghk p−1p
L

and
ˆ p p
p p p
|g| p−1 |h| p−1 dµ ≤ k|g| p−1 k q(p−1) khk r(p−1) = kgkLp−1 p−1
q khkLr
L p L p

since
p 1 1 p 1
+ = (1 − ) = 1.
p−1 q r p−1 p
We denote Lp (Rd ) (or even Lp ) for Lp (md ) where md is the Lebesgue
measure.

32 [February 10, 2017]

Lemma 3.21. Suppose that 1 ≤ p, q, r ≤ ∞ satisfy
1 1 1
+ + =2
p q r

and that f ∈ Lp (Rd ), g ∈ Lq (Rd ) and h ∈ Lr (Rd ). Then

Rd × Rd 3 (x, y) → f (−x)g(x − y)h(y)

is integrable and
ˆ
I(f, g, h) := f (−x)g(x − y)h(y)dm2d (x, y)
Rd ×Rd

satisfies
|I(f, g, h)| ≤ kf kLp kgkLq khkLr
and
I(f, g, h) = I(g, f, h) = I(f, h, g) = I(h, g, f ).

Proof. We assume 1 < p, q, r < ∞ since the limit cases are simpler, and
follow by obvious modifications. Measurability is a consequence of the the-
orem of Fubini. It suffices to prove the statement for nonnegative functions
since ˆ ˆ
f ghdm2d ≤ |f gh|dm2d

and the integrability of f gh follows from the integrability of |f gh|. We define

p
p0 , q 0 and r0 by p1 + p10 = 1, i.e. p0 = p−1 etc. Let
0 0
α(x, y) =|f (−x)|p/r |g(x − y)|q/r ,
0 0
β(x, y) =|f (−x)|p/q |h(y)|r/q ,
0 0
γ(x, y) =|g(x − y)|q/p |h(y)|r/p .
1 1 1
Then p0 + q0 + r0 = 1 and
ˆ
I= α(x, y)β(x, y)γ(x, y)dm2d

≤kαkLr0 kβkLq0 kγkLp0

p q p r q r
0 0 0 0 0 0
=kf kLr p kgkLr q kf kLq p khkLq r kgkLp q khkLp r
=kf kLp kgkLq khkLr .

The second last equality is a consequence of the theorem of Fubini.

33 [February 10, 2017]

Theorem 3.22 (Young’s inequality). Suppose that 1 ≤ p, q, r0 ≤ ∞ and
1 1 1
+ = 1 + 0.
p q r
If f ∈ Lp and g ∈ Lq , then for almost all x

f (x − y)g(y)

is integrable and
´
f (x − y)g(y)dmd (y) if integrable
f ∗ g(x) :=
0 otherwise
0
defines a unique element in Lr (Rd ) and

kf ∗ gkLr0 (Rd ) ≤ kf kLp kgkLq .

2
Proof. We have e−|x| ∈ Lr for all 1 ≤ r ≤ ∞. Then
2
e−|x| f (x − y)g(y)
´
is m2d integrable by Lemma 3.21. We apply Fubini to see that f (x −
y)g(y)dmd (y) exists for almost all x. By Theorem 3.19 the estimate follows
once we prove ˆ
f ∗ ghdmd ≤ kf kLp kgkLq khkLr

for
1 1
+ 0 =1
r r
and all h ∈ Lr . Since then
1 1 1
+ + =2
p q r
and, by Fubini and Lemma 3.21
ˆ ˆ
f ∗ gh(x)dx ≤ |f | ∗ |g||h|dmd
ˆ
= |f (x − y)||g(y)||h(x)|dmd (x, y)

≤kf kLp kgkLq khkLr .

There is a particular case: if q = 1 and p = r0 :

kf ∗ gkLp ≤ kf kLp kgkL1 .

Schur’s lemma gives a criterium for an integral kernel to define a linear

map from Lp (ν) to Lp (µ) for 1 ≤ p ≤ ∞.

34 [February 10, 2017]

Theorem 3.23 (Schur’s lemma). Let (X, A, µ) and (Y, B, ν) be σ finite
measure spaces and k : X × Y → R be µ × ν measurable. Suppose that
C1 , C2 ∈ [0, ∞) and
ˆ ˆ
sup |k(x, y)|dν(y) ≤ C1 , sup |k(x, y)|dµ(x) ≤ C2 .
x y

If 1 ≤ p ≤ ∞ and f ∈ Lp (ν), then

ˆ
k(x, y)f (y)dν(y)

exists for almost all x and

ˆ
1− p1 1
k(x, y)f (y)dν(y) ≤ C1 C2p kf kLp (ν) .
Lp (µ)

The map ˆ
p
L (ν) 3 f → T f := k(x, y)f (y)dν(y) ∈ Lp (µ)

is a continuous linear map which satisfies

1− p1 1
kT kLp (ν)→Lp (ν) ≤ C1 C2p .

Proof. Repeating the argument of Young’s inequality we have to prove that

ˆ
1− 1 1
|g(x)||k(x, y)||f (y)|dµ × ν ≤ C1 p C2p (3.4)

where
1 1
+ =1
p q
if kf kLp = kgkLq = 1. If p = 1 or q = ∞ this is an immediate consequence
of the theorem of Fubini. So we assume 1 < p, q < ∞. It suffices to prove
(3.4) for nonnegative functions f , g and k where f and g are bounded and 0
outside a set of finite measure, since then (3.4) follows by an approximation
and monotone convergence.
For z ∈ C with 0 ≤ Re z ≤ 1, we define

|f |pz−1 f if f 6= 0

fz =
0 otherwise,

and ( p
(1−z)−1
|g| p−1 g if g 6= 0
gz =
0 otherwise .
Then for σ ∈ R,
kfiσ kL∞ = kg1+iσ kL∞ = 1,

35 [February 10, 2017]

and
kf1+iσ kL1 = kf kpLp = kgiσ kL1 = kgkqLq = 1,
and hence ˆ
|giσ (x)||k(x, y)||fiσ (y)|dµ × ν ≤ C1
ˆ
|g1+iσ (x)||k(x, y)||f1+iσ (y)|dµ × ν ≤ C2 .

Moreover
f 1 = f, g 1 = g.
p p

Notice that fz and gz are bounded and zero outside a set of finite mea-
sure. By dominated convergence
ˆ
z → H(z) = gz (x)k(x, y)fz (y)dµ × ν

is continuous in the strip C = {z : 0 ≤ Re z ≤ 1}, differentiable and satisfies

the Cauchy-Riemann differential equations. The claim follows from the three
lines inequality:.
Lemma 3.24 (Three lines inequality). Suppose that u ∈ C(C) is bounded
and holomorphic in the interior. Then

sup |u| = sup |u|

C ∂C

We apply the lemma to

u(z) = C1z−1 C2−z H(z).

[16.11.2016]
[18.11.2016]

Proof. 1) Let U ⊂ C be a bounded open connected set and u ∈ C(U ; C) be

a holomorphic function in the interior. We claim that then

sup |u(x)| = sup |u(x)|.

x∈U x∈∂U

We prove this by contradiction. Suppose that |u| attains its maximum M

at some interior point z0 and suppose that this is larger than sup∂U |u(z)|.
Then
f (z) = Re u(z)/u(z0 )

36 [February 10, 2017]

satisfies 0 ≤ f ≤ M and f (z0 ) = M . Moreover f is harmonic. Let

fε (x + iy) = f (x + iy) + ε|x − Re z0 |

where ε is so small that fε (z) < M for z ∈ ∂U . Then fε has a maximum

in an interior point z1 . At this point the Hessian is negative semidefinite by
its trace ∆fε (z1 ) = 4ε. This is a contradiction.
2) Let u be as in the lemma amd let
2
uε (z) = eεz u(z).

Since uε (z) → 0 as | Im z| → ∞

sup |uε (z)| = sup |uε | ≤ eε sup |u(z)|.

z∈C z∈∂C z∈∂C

Now we let ε tend to zero.

3.9 Borel and Radon measures

Let (X, d) be a metric space. We recall that the Borel sets B(X) are the
smallest σ algebra containing all open sets.

Definition 3.25. Let (X, d) be a metric space. A Borel measure is a mea-

sure on the Borel sets. A Radon measure is a Borel measure, such that for
every x ∈ X there exists an open environment U 3 x so that µ(U ) < ∞ and
such that for every Borel set A

µ(A) = sup{µ(K) : K ⊂ A, K compact }

Definition 3.26. We call a measure complete, if the σ algebra contains

every subset of a set of measure zero.

The theorem of Fubini in the form stated holds for µ×ν with the smallest
σ algebra containing all cartesian product of measurable sets. The Lebesgue
measure restricted to the Borel sets is not complete. We can easily complete
σ algebras.

Lemma 3.27. Let µ be a Radon measure. Then the measure of compact

sets is finite and for ε > 0 and K compact there exists an open set U ⊃ K
of finite measure with µ(U ) ≤ µ(K) + ε.

Proof. Let K be compact. For every x ∈ K exists an open set Ux containing

x with µ(Ux ) < ∞. Since K is compact and
[
K⊂ Ux

37 [February 10, 2017]

there exists a finite subcovering
N
[
K⊂ Uxj =: U
j=1

and
N
X
µ(K) ≤ µ(U ) ≤ µ(Uxj ).
j=1

We define Uj = U ∩ {x : d(x, K) < 1j }. By the theorem of Lebesgue

µ(Uj ) → µ(K).

Definition 3.28. We call X locally compact if for every point x there is

a neighborhood whose closure is compact. We call (X, d) σ compact if it is
locally compact and if it is a countable union of compact sets.
Let (X, d) be σ compact. Then there exists a sequence of compact
S∞ set
˚
Kj so that Kj is contained in the interior Kj+1 of Kj+1 and X = j=1 Kj .
This is proven on exercise sheet 6.
Lemma 3.29. Let µ be a Borel measure on a σ compact space (X, d) and
let B be a Borel set with µ(B) < ∞ and ε > 0. Then there exists a closed
set C ⊂ B with µ(B\C) < ε. If µ is in addition Radon then there exists an
open set U containing B with µ(U \B) < ε.
Proof. For the first part we may assume µ(X\B) = 0 - otherwise we define
ν(A) = µ(A ∩ B). we define

A ⊂ Rd : A is Borel and for every ε > 0 there exists a closed set C

F=
with µ(A\C) < ε.

It contains all closed sets. We claim:

T
1. If Aj ∈ F then Aj ∈ F.
S
2. If Aj ∈ F then Aj ∈ F.

3. Since open sets are countable unions of closed sets every open set is in
F.
We define
G = {A : X\A, A ∈ F}
Then G contains complements of elements and countable unions of elements
of G. Hence it is a σ algebra containing all open sets, and thus it is the
Borel σ algebra. This implies the first claim.

38 [February 10, 2017]

Let now µ be Radon and Kj as above. Then Kj \B is Borel with
µ(K̊j \B) < ∞. Then there exists a closed set Cj ⊂ K̊j \B with µ((K̊j \Cj )\B) <
ε2−j . Let
∞
[
U= (K̊j \Cj ).
j=1

It is open and
∞
[ [
B= (K̊j ∩ B) ⊂ K̊j \Cj = U
j=1

Moreover [
µ(U \B) = µ( (K̊j \Cj )\B) < ε.

Lemma 3.30. Let (X, d) be σ compact, and µ a Borel measure such that
any compact set is of finite measure. Then µ is Radon and it is outer regular.

Proof. Only inner regularity has to be proven since outer regularity fol-
lows then by Lemma 3.29. Let A be Borel with finite measure (why does
this suffice?). By Lemma 3.29 there exists a closedS set C ⊂ A such that
µ(A\C) < ε. Let Kj be compact subsets with X = Kj and Kj contained
in the interior of Kj+1 . Then

µ(C ∩ Kj ) → µ(C)

and C ∩ Kj is compact.

The most important example is the Lebesgue measure. A Radon measure

on a compact metric space is finite. If (X, d) is a countable union of compact
sets and µ is a Radon measure then µ is σ finite.
The counting measure on R is not a Radon measure.

Remark 3.31. Continuous functions on compact metric spaces are inte-

grable with respect to Radon measures.

Lemma 3.32. Let (X, d) be a σ compact metric space, µ a Radon measure

on X and 1 ≤ p < ∞. Then continuous functions with compact support are
dense in Lp (µ).

Proof. Let f be integrable. We decompose it into real and the imaginary

part and it suffices to prove the assertion for real functions. Similar we
decompose a real valued function into positive and negative part, and it
suffices to approximate a nonnegative integrable function f .
Since ˆ ˆ ∞
f dµ = µ({f > t})dt
0

39 [February 10, 2017]

given ε > 0 there exists 0 = t0 < t1 < . . . tj < tj+1 < tN < ∞ so that (with
t0 = 0)
ˆ ∞ N
X
0< µ({f > t})dt − (tj − tj−1 )µ({f > tj }) < ε
0 j=1

Let
Aj = {x : f (t) > tj }.
Then X
kf − (tj − tj−1 )χAj kL1 < ε
and it suffices to approximate a characteristic function of a measurable set
A of finite measure by a continuous function. Let ε > 0. By inner and outer
regularity there exists a compact set K and an open set U so that
K⊂A⊂U µ(U ) ≤ µ(K) + ε.
Then d(K, X\U ) := d0 > 0 and we define
fL (x) = max{1 − Ld(x, K), 0} ∈ C(X)
Then if d0 L ≥ 1
1
kfL − χA kLp < ε p .
If L is sufficiently large then supp f is compact. Thus continuous functions
with compact support are dense.

[18.11.2016]
[23.11.2016]

3.10 Compact sets

Lemma 3.33. If (X, d) is σ compact and µ is Radon measure, then Lipschitz
continuous functions with compact support are dense in Lp (µ) for 1 ≤ p <
∞.
Proof. We prove that for every ε > 0 and f ∈ C(X) with compact support,
there exists fε Lipschitz continuous with
supp fε ⊂ supp f
and
sup |fε − f | < ε.
It suffices to do this for f ≥ 0. Since supp f is compact it is uniformly
continuous: There exists δ > 0 so that |f (x) − f (y)| < ε if d(x, y) < δ. With

|f (x) − f (y)|
L = sup : d(x, y) ≥ δ
d(x, y)

40 [February 10, 2017]

which is finite since it is the supremum of a continuous function on a compact
set, we obtain the inequality

|f (x) − f (y)| ≤ ε + Ld(x, y), ∀x, y ∈ X.

We define
g(x) = min{f (y) + 2Ld(x, y)}.
y

One easily checks that g has Lipschitz constant 2L, and the mimimum is
attained in Bδ (x) and

max{0, f (x) − ε} ≤ g(x) ≤ f (x).

Theorem 3.34 (Arzela-Ascoli). Let (X, d) be a compact metric space. Then

a closed set A ⊂ Cb (X) is compact if and only if

1. A is bounded.

2. A is equicontinuous, i.e. for ε > 0 there exists δ > 0 so that

|f (x) − f (y)| < ε if f ∈ A and d(x, y) < δ.

Proof. Let A be compact. Since

Cb (X) 3 f → kf kCb (X)

is continuous and hence attains its maximum in A, we deduce that A is

bounded. Let ε > 0. For every f there exists δf > 0 and an open neighbor-
hood Uf ⊂ Cb (X) so that

|g(x) − g(y)| < ε if g ∈ Uf and d(x, y) < δf .

Then A ⊂f ∈A Uf , and since A is compact there is a finite subcovering,

A ⊂N j=1 Ufj . We define δ = min δfj .
Now assume that A is closed, bounded and equicontinuous. Let fj ∈ A
be a sequence. Given ε > 0 we claim that there exists g such that B3ε (g) ⊂
Cb (X) contains infinitely many fj . Let ε > 0 and δ > 0 as in the second
condition. Then there exist a finite number N of points xk so that Bδ/2 (xk )
cover X since X is compact. There exists a subsequence so that fjl (xk )
converges for all xk . In particular, after relabeling, there are infinitely many
{fjl }l∈N so that |fjl (xk ) − fjm (xk )| < ε. Then

fjl ⊂ B3ε (fj1 ).

41 [February 10, 2017]

Lemma 3.35. Let (X, d) be a compact set. Then it is separable.

Proof. Given ε there exists a finite number of points xεn , 1 ≤ n ≤ N (ε) so

that the union of the balls Bε (xεn ) cover X. Take a sequence ε = 2−j . This
leads to a dense sequence.

Corollary 3.36. Let (X, d) be compact. Then Cb (X) is separable.

Proof. By the proof of Lemma 3.33 the Lipschitz continuous functions are
dense. The countable union of separable sets is separable and its closure is
separable. Hence it suffices to prove that

K = {f ∈ Cb (X) : kf (x)kCb (X) ≤ n, |f (x) − f (y)| ≤ nd(x, y)}

is separable. This set is compact by Theorem 3.34 and hence separable.

Corollary 3.37. Let (X, d) be σ compact and µ a σ finite Borel measure.

If 1 ≤ p < ∞ then Lp (µ) is separable.

Proof. Since Lipschitz continuous functions with compact support are dense
we argue as for Cb (X).

Corollary 3.38. Suppose that 1 ≤ p < ∞, f ∈ Lp (Rd ), ε > 0. Then there

exist δ > 0 and R > 0 so that for all |h| < δ

kf (. + h) − f (.)kLp < ε, kχRd \BR (0) f kLp < ε.

Proof. The second claim is a consequence of monotone convergence. For the

first we approximate f by a Lipschitz continuous function g with compact
support, kg − f kLp < ε/4 and estimate

kf (. + h) − f kLp ≤kf (. + h) − g(. + h)kLp + kf − gkLp + kg(· + h) − gkLp

1/p
≤ε/4 + ε/4 + |h|kgkLip md (supp g)
≤ε

by choosing |h| ≤ r for some small r.

We want to characterize compact subsets of Lp spaces.

Theorem 3.39. Let 1 ≤ p < ∞. A closed subset C ⊂ Lp (Rn ) is compact

iff

1. C is bounded.

2. For every ε > 0 there exists δ so that for all |h| < δ and all f ∈ C

kf (. + h) − f kLp (Rd ) < ε

42 [February 10, 2017]

3. For every ε > 0 there exists R so that for all f ∈ C

kχRd \BR (0) f kLp < ε.

Proof. Let C be compact. Since f → kf kLp is continuous it attains its

maximum and hence C is bounded. Suppose there exists ε > 0 and hj → 0
and fj ∈ C so that
kfj (. + hj ) − fj kLp ≥ ε.
Since C is compact we may assume that fj is a Cauchy sequence with limit
f . Then there exists δ > 0 so that

kf (. + h) − f kLp < ε/2

for |h| < δ. This contradicts the previous inequality. Similarly we deduce
the third part.
Vice versa: Suppose that C ⊂ Lp (Rd ) is closed, bounded, and satisfies
the three claims. We choose a smooth
´ function η supported in the unit ball
with values between 0 and 1 and η = 1, define ηr (x) = r−d η(x/r) and we
fix ε > 0. Then there exists δ so that by Minkowski’s inequality and the
second assumption

kfr − f kLp ≤ sup kf (. + h) − f kLp < ε, fr = ηr ∗ f,

|h|≤r

for all f ∈ C and r ≤ δ. Moreover fr is Lipschitz continuous with Lipschitz

constant depending on δ. By Theorem 3.34 the set Cδ is compact, and we
can cover it by a finite number of balls of radius ε/2. But then the balls with
radius ε cover C. Thus C is precompact, and compact since it is closed.

[23.11.2016]
[25.11.2016]

3.11 The Riesz representation theorem for Cb (X)

Definition 3.40. Let (X, d) be a σ compact metricSspace and Kj ⊂ X
compact with Kj in the interior of Kj+1 and X = Kj . We denote by
C0 (X) ⊂ Cb (X) the continuous functions f with limit 0 at ∞, i.e. for all
ε > 0 there exists j so that f is at most of size ε outside Kj . We define
Cc (X) as the subspace of continuous functions with compact support.

Definition 3.41. Let (X, d) be a metric space. We call L ∈ (Cb (X))∗

nonnegative if
L(f ) ≥ 0 whenever f ≥ 0.

43 [February 10, 2017]

Theorem 3.42. Let (X, d) be a sigma compact metric space and let L :
Cc (X) → K satisfy
|L(f )| ≤ CK kf kCb (X)
for supp f ⊂ K compact. Then there exists a Radon measure µ and a
measureable function σ : X → {±1} so that
ˆ
L(f ) = f σdµ.
X

Definition 3.43. Let L be as above. We define the variation measure of L

by
µ∗ (U ) = sup{L(f ) : f ∈ Cc (X), supp f ⊂ U, |f | ≤ 1}
for open sets U and for general sets

µ∗ (A) = inf{µ(U ) : A ⊂ U, U open }.

Proof. We prove the theorem by several steps.

1. µ∗ is a outer metric measure, which defines a Radon measure on the

Borel sets.

2. For f ∈ Cc (X) nonnegative we define

λ(f ) = sup{|L(g)| : |g| ≤ f }

and prove for f nonnegative

ˆ
λ(f ) = f dµ

3. As a consequence ˆ
|L(f )| ≤ |f |dµ

and we can extend L to L1 (µ), hence L ∈ (L1 (µ))∗ and there exists
σ ∈ L∞ (µ) so that ˆ
L(f ) = f σdµ

for all f ∈ Cc (X) with kσkL∞ (µ) ≤ 1.

4. We complete the proof by |σ(x)| = 1 for almost all x. Since we may

change on a set of µ measure 0, we obtain |σ| = 1.

We observe that by multiplying g by a constant of size 1 we may always

assume that L(g) ∈ [0, ∞) in the definition of µ∗ and λ.

44 [February 10, 2017]

Step 1: We claim that µ∗ is an outer metric measure. To show that it
is an outer measure, let Uj be open sets, U = ∪Uj and we have to show that
X
µ∗ (U ) ≤ µ∗ (Uj ).

Let 0 ≤ f ≤ 1 with supp f ⊂ U . We have to show that

X
L(f ) ≤ µ∗ (Uj ).

Let K = supp f which is compact. Thus K is covered by finitely many

∪N
j=1 Uj for some N < ∞. Moreover we may assume that the Uj ’s are
contained in a fixed compact set, or even replacing X by this compact set,
X is compact. We claim that there exist gj , 0 ≤ gj ≤ 1, supp gj ⊂ Uj
that P
and gj = 1 on K. We define fj = gj f . Then

X N
X
L(f ) = L(fj ) ≤ µ∗ (Uj ).
j j=1

To see the existence of the gj , take U0 = X\K. Then X = ∪N j=0 Uj and we

take a subordinate partition of unity, i.e. functions ηj ∈ Cb (X) with 0 < ηj
and supp ηj ∈ Uj so that
XN
1= ηj .
j=0

The functions gj = ηj for 1 ≤ j ≤ N have this property. More precisely,

let A0 = X\ N
S
j=1 Uj . It is compact and satisfies A0 ⊂ U0 . There is η̃0 ∈
Cb (X) supported in U0 , identically 1 on A0 . Let A1 = X\({x : η̃0 (x) >
1 SN
2 } j=2 Uj ) ⊂ U1 and we repeat the contruction. Recursively we obtain η̃j
with
N
X 1
ρ= η̃j ≥
2
j=0

in X. We define
η̃j
ηj = .
ρ
Finally, if A, B are Borel sets with positive distance there exist disjoint
open sets V and W containing A resp. B. Then

µ∗ (A ∪ B) = inf µ∗ (U ) = inf µ∗ (U ∩ V ) + µ∗ (U ∩ W ) = µ∗ (A) + µ∗ (B).

Let µ be the measure defined by the Caratheodory construction. Then

µ(U ) = µ∗ (U )

for open sets. By construction µ is bounded on compact sets and thus its
restriction to Borel sets is a Radon measure.

45 [February 10, 2017]

1

Figure 1: Formula (3.5)

Step 2: Let f ∈ Cc (X) be non negative. We define

λ(f ) = sup{|L(g)| : g ∈ Cc (X), |g| ≤ f }.

Clearly 0 ≤ f1 ≤ f2 implies λ(f1 ) ≤ λ(f2 ) and for c > 0, λ(cf ) = cλ(f ). We

claim that
λ(f1 + f2 ) = λ(f1 ) + λ(f2 )
for f1 , f2 ∈ Cc (X) nonnegative. Indeed, if |g1 | ≤ f1 and |g2 | ≤ f2 then
|g1 + g2 | ≤ f1 + f2 , and, if in addition L(g1 ), L(g2 ) ∈ [0, ∞),

|L(g1 ) + L(g2 )| ≤ λ(f1 + f2 ).

This gives
λ(f1 ) + λ(f2 ) ≤ λ(f1 + f2 ).
Now let |g| ≤ f1 + f2 . We define
(
f1 g
f1 +f2 if f1 + f2 > 0
g1 =
0 otherwise

and similarly g2 . Then |gi | ≤ fi and hence

|L(g)| ≤ λ(f1 ) + λ(f2 )

which gives
λ(f ) = λ(f1 ) + λ(f2 ).

We claim that ˆ
λ(f ) = f dµ.

It suffices to consider 0 ≤ f ≤ 1. We approximate f by step function so that

N −1
1 X 1
kf − χUj ksup < (3.5)
N N
j=1

46 [February 10, 2017]

with
j
Uj = {x : f (x) >
}.
N
By continuity Uj+1 ⊂ Uj . We approximate the characteristic function by
continuous functions so that supp ηj ⊂ Uj−1 , ηj = 1 on Uj and µ(supp ηj \Uj ) <
1/j. It suffice to verify
ˆ
|λ(ηj )| − ηj dµ ≤ 1/j

which follows from

µ(Uj ) ≤ |λ(ηj )| ≤ µ(supp ηj ).
Step 3: Now ˆ
|L(f )| ≤ λ(|f |) = |f |dµ.

We extend L to an element in (L1 (µ))∗ , which is represented by an infinite

integrable function σ by Corollary 3.20 . Moreover
kσkL∞ (µ) ≤ kLk(L1 (µ))∗ = 1.
Step 4: We claim that |σ| = 1 almost everywhere. By definition
ˆ
µ(U ) = sup{ f σdµ = L(f ) : f ∈ Cc (X), |f | ≤ 1, supp f ⊂ U }

We choose a sequence of functions with

ˆ
fj σdµ = L(fj ) → µ(U ).
´ ´
Since fj σdµ ≤ |σ|dµ and |σ| ≤ 1 we deduce |σ| = 1 almost everywhere.

[25.11.2016]
[30.11.2016]

3.12 Covering lemmas and Radon measures on Rd

The space Rd is σ compact and the Lebesgue measure is σ finite Radon
measure.
Theorem 3.44 (Covering theorem of Besicovitch). There exists Md de-
pending only on d so that every family F of closed balls with bounded radii
contains Md subfamilies Gm , 1 ≤ m ≤ Md so that each Gm consists disjoint
balls and if A is the set of the centers then
[ [
A⊂ B.
m B∈Gm

47 [February 10, 2017]

The same statement with the same proof holds for open balls.

Proof. We assume first that A is bounded and define D as the supremum

of the radii. There exists a ball B1 = Br1 (x1 ) with r1 ≥ 3D
4 . We choose
recursively Bn = Brn (xn ) with xn in
n−1
[
An = A\ Brj (xj )
j=1

so that
3
rn ≥ sup{r : Br (x) ∈ F, x ∈ An }.
4
We stop if An = { }. For simplicity we consider the case when the procedure
does not stop. Then whenever j ≥ n, we have rj ≤ 34 rn (otherwise we would
not have chosen Brn (xn )) and
rn + rj
|xj − xn | ≥ rn ≥
3
and the balls Brj /3 (xj ) are all disjoint and
[
A⊂ Bj .

We fix k > 1 and define

I = {j : 1 ≤ j < k, Bj ∩ Bk 6= { }}.

We claim that there is a bound for the number of balls in I: #I ≤ Md with

Md depending only on d and D.
We first bound the number of small balls. Let K := I ∩ {j : rj ≤ 3rk }.
Then #K ≤ 20d .

48 [February 10, 2017]

xk

To see that we consider j ∈ K and choose x ∈ Brj /3 (xj ) ⊂ B5rk (xk ).

The #K balls Brj /3 (xj ) are all disjoint and hence
X
(5rk )d ≥ (rj /3)d ≥ (rk /4)d #K.
j∈K

Next we bound the number of large balls, i.e #(I\K). Let now i, j ∈
I\K, i 6= j. We will give a upper bound on
hxi − xk , xj − xk i
cos(∠(xk xi , xk xj )) = .
|xi − xk ||xj − xk |

This gives a lower bound on the distance of the points |xxnn −x

−xk | for n < k,
k

n ∈ I\K, and hence a upper bound on their numbers Ld depending only

on the dimension since the unit sphere is compact. Therefore we can take
Md = 20d + Ld + 1.
To simplify the notation we assume that xk = 0. Let θ be the angle
between the centers ∠xi , xj . Since Bi ∩ Bk 6= { } and Bj ∩ Bk 6= { }, we
have without loss of generality

|xi | ≤ |xj |, |xi | ≤ ri + rk , |xj | ≤ rj + rk .

We claim that xi ∈ Bj if cos θ > 56 . Firstly we notice that if |xi − xj | ≥

|xj |, then
|xi |2 + |xj |2 − |xi − xj |2 |xi |2 |xi | 1 5
cos θ = ≤ = ≤ ≤ .
2|xi ||xj | 2|xi ||xj | 2|xj | 2 6

49 [February 10, 2017]

Hence if we assume cos θ ≥ 56 then |xi − xj | ≤ |xj |. We suppose by contra-
diction that xi ∈
/ Bj . Then rj ≤ |xi − xj | and

|xi |2 + |xj |2 − |xi − xj |2

cos θ =
2|xi ||xj |
|xi | (|xj | − |xi − xj |)(|xj | + |xi − xj |)
= +
2|xj | 2|xi ||xj |
1 |xj | − |xi − xj |
≤ +
2 |xi |
1 rj + rk − rj
≤ +
2 ri
1 rk 5
= + ≤ .
2 ri 6
Now it suffices to derive the upper bound for cos θ when xi ∈ Bj , since
otherwise cos θ ≤ 56 has already a upper bound. So we assume xi ∈ Bj
from now on and we may restrict to rk = 1 by scaling. Then i < j, since
otherwise Bi would not have been chosen, and thus xj ∈ / Bi , and

4
3 ≤ ri < |xi − xj | < rj ≤ ri , ri < |xi | ≤ 1 + ri , rj < |xj | < 1 + rj .
3

xi

0 θ xj

50 [February 10, 2017]

The proof becomes now an exercise in planar geometry. We have
2
3 1 rj 1 ri ri − 1 ri + ri − rj − 1
≤ ≤ ≤ 3 ≤
16 4 |xj | 3 |xj | |xj | |xj |
|xi − xj | + |xi | − |xj |
≤
|xj |
|xi − xj | + |xi | − |xj | |xi − xj | − |xi | + |xj |
≤
|xj | |xi − xj |
2
|xi − xj | − ||xi | − |xj || 2
=
|xj ||xi − xj |
|xi ||xj |
= 2(1 − cos θ)
|xj ||xi − xj |
|xi |
= 2(1 − cos θ)
|xi − xj |
ri + 1
≤ 2(1 − cos θ)
ri
8
≤ (1 − cos θ)
3
119
and hence cos θ ≤ 128 .
It remain to define the sets Gm . We do this by defining a map
σ : N → {1, . . . Md }.
We choose it to be the identity for j ≤ Md . After that we proceed recursively,
which we can do since
n o
# j ≤ k : Bj ∩ Bk+1 6= { } < Md .

It remains to extend the result to unbounded sets. We do this by applying

the first part in the annuli 6(m − 1)D ≤ |x| < 6mD.

Theorem 3.45. Let µ be a Radon measure on Rd and let F be a family of

closed balls and let A be a Borel set which is the union of the centers. We
assume µ(A) < ∞ and inf{r : Br (x) ∈ F} = 0 for x ∈ A. Let U ⊂ Rd be
open. Then there exists a countable collection of disjoint closed balls G ⊂ F
so that [
µ (A ∩ U )\ B = 0.
B∈G

Proof. We fix θ so that 1 − M1d < θ < 1 and claim that there is a finite
collection of disjoint balls Bj , 1 ≤ j ≤ M in F so that
 
M
[
µ (A ∩ U )\ Bj  ≤ θµ(A ∩ U ).
j=1

51 [February 10, 2017]

Suppose this is true. Then we define
M
[
A1 = A\ Bj
j=1

and repeat the argument with F1 the subset of balls with center in A1 .
After the kth step the complement has a measure at most θj µ(U ∩ A). So it
remains to prove the claim. Let F0 be the subset of balls with radii at most
1. Then we apply the Besicovitch covering theorem and obtain Gm . Then
M
[d [
A∩U ⊂ B
j=1 B⊂Gj

and
Md
X [
µ(A ∩ U ) ≤ µ A∩U ∩ B .
j=1 B∈Gj

There exists J so that

1 [
µ(A ∩ U ) ≤ µ A ∩ U ∩ B .
Md
B∈GJ

By monotone convergence there exist finitely many balls in GJ so that the

claim holds.

We turn to derivatives of Radon measures.

Definition 3.46. Let µ and ν be Radon measures on Rd . For x ∈ Rd we

define
(
ν(Br (x))
lim supr→0 µ(Br (x))
if µ(Br (x)) > 0 for all r > 0
Dµ ν(x) =
∞ if for some r > 0, µ(Br (x)) = 0
(
ν(Br (x))
lim inf r→0 µ(Br (x)) if µ(Br (x)) > 0 for all r > 0
Dµ ν(x) =
∞ if for some r > 0, µ(Br (x)) = 0.

We say that ν is differentiable with respect to µ and x if Dµ ν(x) = Dµ ν(x).

Then we write Dµ ν(x) and call this quantity the density of ν with respect to
µ.

Remark 3.47. Let f ∈ Cc (Rd ) and µ Radon measure. Then

ˆ
x → f (y − x)dµ(y)

is continuous and hence Borel measurable.

52 [February 10, 2017]

Since the characteristic function of open and closed balls can be obtained
as pointwise limit of continuous functions with compact support, the map

x → ν(Br (x)), x → µ(Br (x))

are measurable. Thus

(
ν(Br (x))
µ(Br (x)) if µ(Br (x)) > 0
x→
∞ if µ(Br (x)) = 0

is Borel measurable. The map

(
ν(Br (x))
µ(Br (x)) if µ(Br (x)) > 0
r→
∞ if µ(Br (x)) = 0

is continuous from the left and right if µ(Br (x)) > 0 by inner and outer
regularity. Thus also Dµ ν(x) and Dµ ν(x) are Borel measurable since we
can write them as inf ’s and sup’s over rational radii. Moreover by inner
and outer regularity we obtain the same Dµ ν(x) and Dµ ν(x) if we use closed
balls.

Theorem 3.48. Let µ and ν be Radon measures on Rd . Then

1. Dµ ν(x) exists and is finite µ almost everywhere.

2. Dµ ν is Borel measurable.

[30.11.2016]
[02.12.2016]

Proof. We may assume that µ(Rd ) < ∞ and ν(Rd ) < ∞.

Step 1: We claim that for all Borel sets B and all t > 0

ν(B ∩ {x : Dµ ν(x) < t}) ≤ tµ(B ∩ {x : Dµ ν(x) < t})

and
ν(B ∩ {x : Dµ ν(x) > t}) ≥ tµ(B ∩ {x : Dµ ν(x) > t}).
By outer regularity (µ(B) = inf{µ(U ) : B ⊂ U }) and it suffices to prove the
assertion for B = U open. Let A = {x ∈ U : Dµ ν(x) < t}. Let

F = {Br (a) : a ∈ A, Br (a) ⊂ U, ν(Br (a)) < tµ(Br (a))}.

For every x ∈ A, F contains arbitrarily small balls and we apply Theorem

3.45 to obtain a sequence of disjoint closed balls Bj in F so that
[
ν(A\ Bj ) = 0

53 [February 10, 2017]

Then X X
ν(A) = ν(Bj ) ≤ t µ(Bj ) ≤ tµ(U ).
Since
µ(A) = inf{µ(U ) : A ⊂ U }
we obtain the first inequality. The second one is proven similarly.
Step 2: We claim that Dµ ν(x) < ∞ outside a set of µ measure 0. Let
A = {x : Dµ ν(x) = ∞}. Then

ν(A) ≥ tµ(A)

for all t, hence µ(A) = 0.

Step 3: For s < t we define

R(s, t) = {x : Dµ ν(x) < s < t < Dµ ν(x)}

Then
tµ(R(s, t)) ≤ ν(R(s, t)) ≤ sµ(R(s, t))
which implies µ(R(s, t)) = 0. Since
[
{x : Dµ ν(x) < Dµ ν(x)} = R(s, t)
s<t,s,t∈Q

we see that Dµ ν(x) = Dµ ν(x) for µ almost all x.

Definition 3.49. Let µ and ν be Borel measures on Rd . We say the measure

ν is absolutely continuous with respect to µ, ν << µ, if µ(A) = 0 implies
ν(A) = 0. The mearures ν and µ are mutally singular with respect to µ if
there exists a Borel set B such that µ(X\B) = ν(B) = 0. We write then
ν ⊥ µ.

Theorem 3.50 (Radon-Nikodym). Let ν and µ be Radon measures on Rd

with ν << µ. Then ˆ
ν(A) = Dµ νdµ
A
for all Borel sets.

Proof. It suffices to consider the case µ(Rd ) < ∞ and ν(Rd ) < ∞. We have
seen that
µ({Dµ ν(x) = ∞}) = 0
and hence, since ν << µ, ν(({Dµ ν(x) = ∞}) = 0. In the same fashion

ν({Dµ ν(x) = 0}) = 0.

54 [February 10, 2017]

Let A be a Borel set. For t > 1 we define

An = A ∩ {tn ≤ Dµ ν < tn+1 }.

Then
∞
X ∞
X ˆ ∞ ˆ
m+1
ν(A) = ν(Am ) ≤ t µ(Am ) ≤ t µ({Dµ ν > s})ds = t Dµ νdµ
m=−∞ m=−∞ 0 A

and
∞
X ∞
X ˆ ∞ ˆ
m −1 −1
ν(A) = ν(Am ) ≥ t µ(Am ) ≥ t µ({Dµ ν > s})ds = t Dµ νdµ.
m=−∞ m=−∞ 0 A

We let now t → 1.

Theorem 3.51 (Lebesgue points). Let µ be a Radon measure on Rd and

f˜ ∈ L1loc (µ). Then
ˆ
−1
f (x) := lim µ(Br (x)) f˜dµ
r→0 Br (x)

exists almost everywhere and we define f (x) = 0 if it does not exist. Then
f is in the equivalence class of f˜. If f˜ ∈ Lploc (µ) then f ∈ Lploc and
ˆ
−1
lim µ(Br (x)) |f (y) − f (x)|p dµ(y) = 0
r→0 Br (x)

almost everywhere.

Proof. It suffices to consider nonnegative f˜ and µ(Rd ) < ∞. We define

ˆ
ν(A) = f˜ dµ.
A

This is a Radon measure by Lemma 3.30 which is absolutely continuous

with respect to µ. Thus
ˆ ˆ
ν(A) = Dµ νdµ = f˜dµ

and Dµ ν lies in the equivalence class. Now the first claim follows from
Theorem 3.48.
For every t, |f (x) − t|p is integrable. From the first part
ˆ
−1
lim µ(Br (x)) |f (y) − t|p dµ(y) = |f (x) − t|p
r→0 Br (x)

55 [February 10, 2017]

almost every where. There is even a set N of µ measure zero so that this is
true for all t ∈ Q outside the same set of measure zero. Let ε > 0. Thus the
set of all x such that
ˆ
lim sup µ(Br (x))−1 |f (y) − f (x)|p dµ(y) > ε
r→0 Br (x)

is contained in N . To see this, chose t ∈ Q so that |f (x) − t|p < ε. This

completes the proof.

Corollary 3.52. Let µ be a Radon measure and f ∈ Lp (µ). Then there

exists a cannonical representative of the equivalence class.

4 Distributions and Sobolev spaces

4.1 Baire category theorem and consequences
Lemma 4.1 (Baire category theorem). A countable intersection of dense
open subsets of a complete metric space is dense.

Proof. Let (X, d) be a complete metric space and Aj open dense sets. Let
x ∈ X and ε > 0. Let x1 ∈ A1 so that d(x, x1 ) < ε/3 and 0 < δ1 < ε/3 so
that B2δ1 (x1 ) ⊂ A1 . We pick recursively xn , δn so that d(xn−1 , xn ) < δn ,
2δn < ε/3n and B2δn (xn ) ∈ An ∩ Bδn−1 (xn−1 ).
By construction, d(xn−1 , xn ) < ε/(2 · 3n−1 ) and, if n < m, d(xn , xm ) <
ε P m−n −j ≤ 3ε and (x ) is a Cauchy sequence with limit y. Since
2·3n j=0 3 4·3n n
xm ∈ Bδn (xn ) for m ≥ n the same is true for y, and y ∈ An for all n.

Theorem 4.2 (Banach-Steinhaus). Let X and Y be Banach spaces, F ⊂

L(X, Y ). Suppose for each x ∈ X

sup{kT xkY : T ∈ F} < ∞.

Then
sup{kT kX→Y : T ∈ F} < ∞.

Proof. Let
Cn = {x ∈ X : sup kT xkY ≤ n}.
T ∈F

This set is closed since both map and normS are continuous, and Cn is an
intersection of closed sets. By assumption Cn = X. We claim that some
Cn has nonempty open interior. If not then the sets Un = S X\Cn are open
and dense, with nonempty intersection, a contradiction to Cn = X. Let
U ⊂ Cn0 be nonempty and open. It contains a ball Br (x0 ). If kxk < r then

kT xkY ≤ kT (x − x0 )kY + kT (x0 )kY ≤ n0 + sup kT (x0 )kY =: R.

T ∈F

56 [February 10, 2017]

Then
kT kX→Y ≤ R/r
for all T ∈ F.

[02.12.2016]
[09.12.2016]

The Baire category theorem has interesting further consequences.

Theorem 4.3. Let X and Y be Banach spaces and T ∈ L(X, Y ). T is
surjective if and only if it is open, i.e. if the image of open sets is open.
Proof. Let T be open. S Then T (B1 (0)) is open. In particular it contains a
ball Br (0). Then Y = T (Bn (0)) and T is surjective. Now suppose that
T is surjective. It suffices to show that T (B1 (0)) contains a ball around 0.
(Why?). Let
Yn = T (Bn (0)) = {T x|kxkX ≤ n}.
S
It is closed by continuity, and Y = Yn . As above we conclude that one (and
hence all) of the Yn contains an open ball. Hence there exists Br (0) ⊂ Y1 .

Corollary 4.4. Suppose that T ∈ L(X, Y ) is injective and surjective. Then

T −1 ∈ L(Y, X).
Thus continuous linear maps which are invertible as maps between sets
are invertible as continuous linear maps.

Proof. Linearity of the inverse map is immediate. By Theorem 4.3 T is

open. So T (B1X (0)) contains a ball BrY (0) and hence

kT xkY ≥ r−1 kxkX .

Theorem 4.5. The set of nowhere differentiable functions in Cb (0, 1) is

dense.
Proof. Exercise.

4.2 Distributions: Definition

We need a preliminary result.
Lemma 4.6. Let U ⊂ Rd be open and k ∈ N. Then for every f ∈ Cck (U )
there exists a compact set K ⊂ U and a sequence fn ∈ C ∞ (U ) supported in
K so that ∂ α fn → ∂ α f in Cb (U ) for n → ∞ and |α| ≤ k.
It suffices to consider U = Rd . We will later prove a more general result
and make sure that the reasoning is not circular.

57 [February 10, 2017]

Definition 4.7. Let U ⊂ Rd be open, and D(U ) = Cc∞ (U ) be the vector
space of infinitely differentiable functions with compact support called test
functions. We say fj → f in Cc∞ (U ) = D(U ) if there is a compact set
K ⊂ U and supp fj ⊂ K for all j and for all multiindices α
∂ α fj → ∂ α f in Cb (U ).
A distribution T on U is a continuous linear map from C0∞ (U ) → K. We
denote the space of distributions by D0 (U ).
By continuous we mean that
T fj → T f
if fj → f in the sense of test functions. It is immediate that the distributions
define a K vector space.
Lemma 4.8. Let T ∈ D0 (U ). For every compact set K ⊂ U there exists k
and C > 0 so that, if f ∈ D(U ) with supp f ⊂ K then
|T (f )| ≤ Ckf kC k (U ) .
b

Proof. We define for K ⊂ U compact

XK = {f ∈ D(U ) : supp f ⊂ K}.
We define a metric on XK
d(f, g) = sup 2−k min{1, kf − gkC k (U ) }.
b
k≥0

With this metric XK is a complete metric space:

d(fj , f ) → 0 iff fj → f in C k (U ) for all k ≥ 0.
Then T ∈ D0 (U ) define a continuous linear map from XK → K. More-
over
|T f | < ∞
for all f ∈ XK . Now we argue as for the uniform boundedness principle of
Banach-Steinhaus: There exists m so that the set
{f ∈ XK : |T f | < m}
contains an open ball. Then there exists r > 0 so that
|T f | < m for all f ∈ XK with d(f, 0) < r.
Let k be so that 21−k < r. Then
d(f, 0) < r if kf kC k (U ) < 2−k and supp f ⊂ K.
we continue as for the theorem of Banach-Steinhaus.

58 [February 10, 2017]

Definition 4.9. We say that
Tn → T in D0 (U )
if
Tn (f ) → T (f ) for all f ∈ D(U ).
If f ∈ D(u) and g ∈ C ∞ (U ) then f g ∈ D(U ).
Definition 4.10. Let φ ∈ C ∞ (U ) and T ∈ D0 (U ). We define their product
by
(φT )(f ) = T (φf )
and the derivative
(∂xj T )(f ) = T (−∂xj f ).
It is easy to see that the right hand side of the formulas defines a distri-
bution. We can easily calculate Leibniz’ formula in the form
∂xj (φT )(f ) = −T (φ∂xj f ) = T ((∂xj φ)f ) − T (∂xj (φf ))
= [(∂xj φ)T ](f ) + (φ∂xj T )(f )
and the associative and distributive law:
φ(ψT ) = ψ(φT ).
Similarly the theorem of Schwarz holds
∂xj ∂xk T = ∂xk ∂xj T.
Let L1loc (U ) be the set of measurable functions on U which are integrable
on compact subsets. We say fj converges to f in L1loc if fj |K → f |K in L1 (K)
for all compact subsets K.
Definition 4.11. We define L1loc (U ) 3 f → Tf ∈ D0 (U ) by
ˆ
Tf (φ) = f φdmd
U
for φ ∈ D(U ).
Lemma 4.12. The map L1loc → D0 is linear, continuous and injective.
Proof. Only injectivity has to be proven. After multiplying by a character-
istic function of a ball we consider f ∈ L1 (B). Suppose that
ˆ
f φdx = 0

for all φ ∈ Cc∞ supported in B1 (0). Then

f ∗ φ(x) = 0
for all φ supported in Br (0) and |x| < 1 − r. But we have seen that there
is such a sequence φj so that f ∗ φj → f in L1 . Then f |B1−r (0) = 0. This
implies the full statement.

59 [February 10, 2017]

Similarly, any Radon measure µ on U ⊂ Rd defines a linear map from
the continuous functions with compact support to K, and we identify it with
the restriction to D.

Examples:

1. The Dirac measure δ0

0 if x < 0
2. The Heaviside function H(x) = satisfies
1 if x ≥ 0

∂x H = δ0 .

Firstly H ∈ L1loc . If φ ∈ D(U )

ˆ ∞
0
(∂x H)(φ) = (∂x TH )(φ) = −TH (φ ) = − H(x)φ0 (x)dx = φ(0) = δ0 (φ).
0

3. Let
1/2 if |x| < −t
f=
0 if |x| ≥ −t
Then
ˆ 0 ˆ −t
2 2 1
(∂tt − ∂xx )f (φ) = (∂t2 − ∂x2 )φ(t, x)dxdt
2 −∞ t
=φ(0, 0) = δ0 (t, x),

which is the contents of exercise 3 on sheet 8.

4. Let d > 2 and

2 Γ( d2 ) 2−d
g(x) = |x|
d − 2 π d/2
In Einführung in die PDG we have seen that
ˆ
−∆g ∗ φ = g(x − y)(−∆φ(y))dy = φ(x).

Thus
−∆g = δ0 .

[09.12.2016]
[14.12.2016]

Lemma 4.13. The following identities hold

Tφψ = φTψ for φ, ψ ∈ C(U ),

T∂xj φ = ∂xj Tφ for φ ∈ C 1 (U ).

60 [February 10, 2017]

Proof. We use Fubini, integration by parts and the fact that f ∈ D(U ) has
compact support, to get
ˆ
Tφψ (f ) = φψf dmd = (φTψ )(f ),
ˆU ˆ
T∂xj φ f = f ∂xj φ dmd = (−∂xj f )φ dmd = (∂xj Tφ )(f ).

Definition 4.14. Let T ∈ D0 (U ). We say that T vanishes near x ∈ U if

there exists r > 0 so that T (f ) = 0 for all f ∈ D(U ) with support in Br (x).
We define the support of T as the complement of the points near which T
vanishes.
Lemma 4.15. Let φ ∈ D(U ) and T ∈ D0 (U ) with disjoint supports. Then

T (φ) = 0.

Proof. Let K be the support of φ. Given x ∈ K there exists r so that

T ψ = 0 for every ψ ∈ D(U ) supported in Br (x). Since K is compact there
is a finite covering of such balls Brj (xj ) with 1 ≤ j ≤ N . We choose a
partition of unity ηj ∈ C ∞ (U ) supported in Brj (xj ) so that
N
X
ηj (x) = 1
j=1

for x ∈ K. Then
N
X
Tφ = T (ηj φ) = 0
j=1

Definition 4.16. Let T ∈ D0 (Rd ) and φ ∈ D(Rd ). We define their convo-

lution by
(φ ∗ T )(x) = T (φ(x − ·)), ∀x ∈ Rd .
The righthand side denotes T acting on φ(x − y) as a function of y.
Lemma 4.17. With the notation above, we know φ ∗ T ∈ C ∞ (Rd ) and

∂xj (φ ∗ T ) = (∂xj φ) ∗ T = φ ∗ (∂xj T ).

If ψ ∈ L1 (Rd ) then
φ ∗ Tψ (x) = φ ∗ ψ(x).
Moreover, if supp φ = K1 and supp T = K2 then

supp φ ∗ T ⊂ K1 + K2 = {x + y : x ∈ K1 and y ∈ K2 }.

61 [February 10, 2017]

Proof. For v ∈ Rd , we have to prove that

(φ ∗ T )(x + tv) − (φ ∗ T )(x) 1

= T ( (φ(x + tv − ·) − φ(x − ·)))
 t  t
Xd Xd
→T  vj (∂j φ)(x − ·) =
 vj ∂j φ ∗ T (x)
j=1 j=1
 
d
X d
X
≡ −T  vj ∂j (φ(x − ·)) = φ ∗ vj ∂j T (x).
j=1 j=1

This is a consequence of Lemma 4.8 and that the difference quotient

d
(∂ α φ)(x + tv) − (∂ α φ)(x) X
→ vj ∂j (∂ α φ)(x) as t → 0
t
j=1

uniformly in x since the support is compact.

The remaining properties are not hard to verify. (Exercise)

Remark 4.18. We can equivalently define the convolution of φ ∈ D(Rd )

and T ∈ D0 (Rd ) as the distribution:

(φ ∗ T )(f ) = T (φ̃ ∗ f ), ∀f ∈ D(Rd ),

where φ̃(x) = φ(−x).

In the same way we can easily define the convolution φ ∗ T ∈ D0 (Rd )
when φ ∈ C k (Rd ) has compact support. By an abuse of notation we write
the convolution evaluated at x whenever it is defined, even if it is not defined
on all of Rd .

Definition 4.19. Let T ∈ D0 (Rd ) and let S ∈ D0 (Rd ) with compact support.
We define their convolution by

(S ∗ T )(φ) = T (S̃ ∗ φ)

for φ ∈ D(Rd ). Here S̃(ψ) = S(ψ̃), ψ ∈ D(Rd ).

It is an exercise to formulate and prove reasonable properties of the

convolution of distributions.

Example 4.20.
Let d > 2 and
2 Γ( d2 ) 2−d
g(x) = |x| ∈ L1loc (Rd ).
d − 2 π d/2
Then
−∆g = δ0

62 [February 10, 2017]

and
−∆(g ∗ φ) = δ0 ∗ φ = φ.
We identify ∆δ0 with the distribution

∆δ0 (f ) = ∆f (0), ∀f ∈ D(Rd ).

Its support is {0}. Then

φ ∗ (∆δ0 )(x) = ∆δ0 (φ(x − ·)) = ∆φ(x).

The same construction works for all differential operators with constant co-
efficients.
Lemma 4.21. Suppose that Tn → T in D0 (U ) and that K ⊂ U is compact.
Then there exists k and C so that

sup |Tn (f )| ≤ Ckf kC k (U )

n b

for all f ∈ XK = {f ∈ D(U ) : suppf ⊂ K} and

sup{|Tn (f ) − Tm (f )| : f ∈ XK , kf kC k ≤ 1} → 0
b

as n, m → ∞.
Proof. The proof of the first part is the same is for Lemma 4.8. So, given
K, there exist k ≥ 1 and C so that for any n

|Tn (f )| ≤ Ckf kC k−1 .

b

Since K is compact and

Cck (K) 3 f → (∂ α f )|α|≤k ∈ Cb (Σk × K) = Cb (K; K#Σk )

is an isometry where Σk is the set of all multiindices of length at most k.

By a multiple application of Theorem 3.34

B̄1 (0) = {f ∈ XK |kf kC 1 (K;K#Σk−1 ) ≤ 1} ⊂ Cb1 (K; K#Σk−1 )

b

is compact in Cck−1 (K). Let ε > 0. Then there exist finitely many functions
fm ∈ Cbk (K) so that the ε balls in Cbk−1 centered at fm cover B̄1 (0). We
may assume that they are in Cc∞ by Lemma 4.6. There exists n0 so that

|(Tn − T )(fm )| < ε

if n ≥ n0 . Then, for any f ∈ B̄1 (0) there exists fm such that kf −fm kC k−1 <
b
ε and hence for n ≥ n0 there hold

|(Tn − T )f | ≤ |(Tn − T )(fm )| + |Tn (f − fm )| + |T (f − fm )| ≤ ε + 2Cε.

63 [February 10, 2017]

Theorem 4.22. Let U ⊂ Rd be open. Then D(U ) ⊂ D0 (U ) is dense.
Proof. There are several steps.
Step 1: Distributions with compact support are dense. We choose a se-
quence φj ∈ D(U ) so that
∂ α φj → ∂ α 1
on compact subsets. Then φj T has compact support and
(φj T )(g) = T (φj g) → T g
for all g ∈ D(U ).
Step 2: Construction of the φj . Let Kj be a monotone sequence S of
compact sets so that Kj is contained in the interior of Kj+1 and U = Kj .
Then for any j there exists rj > 0 so that
min{dist(Kj−1 , Rd \Kj ), dist(Kj , Rd \Kj+1 )} ≥ rj > 0.
´
Let ϕ ∈ Cc∞ (B1 (0)) be radial with ϕdx = 1 and let ϕr (x) = r−d ϕ(x/r).
Then
φj = ϕrj ∗ χKj ∈ C ∞ ,
supp φj ⊂ Kj+1 , φj = 1 on Kj−1 .
Step 3 Let T ∈ D0 (U )
have compact support. Then, for r small and
g ∈ D(U )
ϕr ∗ T (g) = T (ϕ˜r ∗ g) → T (g), r → 0
by an abuse of notation.

The same argument gives Lemma 4.6.

Lemma 4.23. Suppose that U is connected and ∂j T = 0, j = 1, · · · , d.
Then there eixsts a constant c so that T = Tc .
Proof. Exercise.

[14.12.2016]
[16.12.2016]

4.3 Schwartz functions and tempered distributions

We briefly cover the definition of Schwartz functions and tempered distri-
butions, which are the proper frame work for the Fourier transform.
Definition 4.24. The Schwartz space S(Rd ) consists of all Schwartz func-
tions, which are functions f so that for all multiindices α, β
kxα ∂ β f ksup < ∞.
We say fj → f as Schwartz functions if for all multiindices α and β
xα ∂ β fj → xα ∂ β f uniformly.

64 [February 10, 2017]

Remark 4.25. It is easy to see that if f ∈ S(Rd ), then for any N there
exists CN such that |f (x)| ≤ CN (1 + |x|)−N , and hence f ∈ Lp (Rd ), for any
p ∈ [1, ∞]. So is ∂ α f for any multiindex α.
Roughly speaking, the Schwartz functions have two properties: they have
infinite bounded derivatives and they decay fast at infinity. Recalling ∂d
xj f (ξ) =
2πiξj fˆ(ξ) and x
d i ˆ
j f (ξ) = 2π ∂ξj f (ξ), the Fourier transform should work well
in the framework of Schwartz space and tempered distributions (see below).
Since ηn f → f in S(Rd ), ηn (x) = η(n−1 x) where η ∈ Cc∞ (B2 (0)) takes
value 1 on B1 (0), the inclusion D(Rd ) ⊂ S(Rd ) is dense.

Lemma 4.26. Let f be a Schwartz function.

1. If α is a multiindex then ∂ α f ∈ S(Rd ).

2. If g ∈ C ∞ and for any multiindex α there exist c|α| and κ|α| so that

|∂ α g| ≤ c|α| (1 + |x|)κ|α|

then gf ∈ S(Rd ).

3. If g ∈ C(Rd ) satisfies for any multiindex α

kxα g(x)ksup < ∞,

then g ∗ f ∈ S(Rd ).

Proof. The first property follows from the definition. By the first property,
in order to prove the second property it suffices to show

sup |xα (∂ β g)f | < ∞,

which follows from the definition.

Since
∂ α (g ∗ f ) = g ∗ ∂ α f
and since ∂ α f is Schwartz, by the first property the proof of the third prop-
erty is reduced to bounding

kxα (g ∗ f )ksup .

We observe that
xj (g ∗ f ) = (xj g) ∗ f + g ∗ (xj f )
and the claim follows by induction on the length of α.

Definition 4.27. We define d : S × S → [0, ∞) by

d(f, g) = sup 2−k min{1, sup kxα ∂ β (f − g)ksup }.

k |α|+|β|=k

65 [February 10, 2017]

Lemma 4.28. The expression d(f, g) defines a metric on S which turns it
into a complete metric space.

Proof. Easy exercise.

Definition 4.29. A tempered distribution is a continuous linear map from

S(Rd ) to K. We denote the space of tempered distributions by S 0 (Rd ). We
say that Tj → T in S 0 (Rd ) if Tj (φ) → T (φ) for all φ ∈ S(Rd ).

Lemma 4.30. Let T ∈ S 0 (Rd ). Then there exist k and c so that

|T φ| ≤ c sup kxα ∂ β φksup .

|α|+|β|≤k

If Tn → T in S 0 (Rd ) then there exist C and k so that

sup |Tn φ| ≤ C sup kxα ∂ β φksup

n |α|+|β|≤k

and
|Tn (φ) − T φ|
→ 0.
sup|α|+|β|≤k kxα ∂ β φksup

Proof. The proof is similar as that of Lemma 4.21. The existence of C and k
follows from the idea of the proof of Banach-Steinhaus theorem. The conver-
gence result follows from the compactness of the ball {φ ∈ S(Rd ) | sup|α|+|β|≤k kxα ∂ β φksup ≤
1} in the space {φ ∈ S(Rd ) | sup|α|+|β|≤k−1 kxα ∂ β φksup < +∞}, which is
easy to see if we notice that

sup kxα ∂ β φksup ≤ R−1 sup kxα ∂ β φksup((BR (0))C ) +Rk−1 kφkC k−1 (BR (0))
b
|α|+|β|≤k−1 |α|+|β|≤k

and we can choose R big enough.

Remark 4.31. We define the derivative and the multiplication by a smooth

function with controlled derivatives for a tempered distribution as we did it
for distributions. Similarly, since compactly supported distributions S can
act on Schwartz functions f , we can define the convolution (S ∗ f )(x) ∈
S(Rd ). We then can define the convolution of a tempered distribution with
Schwartz functions and with compactly supported distributions.
Let 1 ≤ p ≤ ∞. There are the embeddings

D(Rd ) ⊂ S(Rd ) ⊂ Lp (Rd ) ⊂ S 0 (Rd ) ⊂ D0 (Rd ).

The embeddings are dense if p < ∞.

66 [February 10, 2017]

4.4 Sobolev spaces: Definition
Definition 4.32. Let U ⊂ Rd be open, k ∈ N and 1 ≤ p ≤ ∞. The Sobolev
space W k,p (U ) ⊂ Lp (U ) is the set of all Lp (U ) functions, so that for all
multiindices α of length at most k there exists fα ∈ Lp (U ) so that

∂ α Tf = Tfα .

We define (identifying Tf and f and ∂ α Tf with fα by an abuse of notation)

 1
p

f kpLp (U ) 
X
α
kf kW k,p = k∂
|α|≤k

with the usual modification if p = ∞.

We have
g = ∂xj f, f, g ∈ Lp (U )
if and only if ˆ ˆ
d
gφdm = − f ∂xj φdmd

for all φ ∈ D(U ).

Lemma 4.33. Let g ∈ Cbk (U ) and f ∈ W k,p (U ). Then gf ∈ W k,p (U ).
Proof: Easy exercise.
Definition 4.34. Let k ∈ N, 1 ≤ p < ∞. We define W0k,p (U ) as the closure
of Cc∞ (U ) with respect to the norm k · kW k,p (U ) . If V ⊂ U the extension
defines a canonical (nonsurjective) isometry from W0k,p (V ) to W0k,p (U ).
Theorem 4.35. Let k ∈ N and 1 ≤ p ≤ ∞. The Sobolev space W k,p (U )
is a Banach space. If V ⊂ U then the restriction defines a map of norm 1
from W k,p (U ) to W k,p (V ). Moreover W0k,p (Rd ) = W k,p (Rd ) if 1 ≤ p < ∞.
Proof. Let Σk = {α : |α| ≤ k}. There is an obvious isometry

W k,p (U ) 3 f → (fα )|α|≤k ∈ Lp (U × Σk ).

Let fj be a Cauchy sequence in W k,p (U ) with limit f ∈ Lp (U ). Then

∂ α fj → fα in Lp (U ) and in D0 (U ). It is easy to check that fα = ∂ α f in
D0 (U ): for any φ ∈ D(U ),
ˆ ˆ
|α| α d |α|
Tfα (φ) = lim (−1) fj ∂ φ dm = (−1) f ∂ α φ dmd = T∂ α f (φ).
j→∞ U U

Thus f ∈ W k,p (U ) and W k,p (U ) is complete and we can identify W k,p (U )

with a closed subspace of Lp (U × Σk ).

67 [February 10, 2017]

The restriction map with norm ≤ 1 follows from the definition, and the
norm is indeed 1 since

kf kW k,p (V ) = kfekW k,p (U ) , f ∈ W0k,p (V ),

and fe ∈ W0k,p (U ) is the trivial extension of f by 0.

Now let 1 ≤ p < ∞. Density of D(Rd ) ⊂ W k,p (Rd ) follows by the same
argument as in the proof of Theorem 4.22.

[16.12.2016]
[21.12.2016]

Definition 4.36. Let k ∈ N and 1 < p < ∞. We define W −k,p (U ) =

0
(W0k,p (U ))∗ .
Lemma 4.37. The map J: Lp (U × Σk ) → W −k,p (U ) defined by
X ˆ 0
J((fα ))(u) = fα ∂ α u dmd , ∀u ∈ W0k,p (U )
|α|≤k U

has norm 1.
Proof. Exercise.

Lemma 4.38. Let U ⊂ Rd be bounded and open and k ∈ N. We assume

that f ∈ Cbk (U ) and its derivatives of order up to k − 1 extend to continous
functions in Ū which vanish at ∂U . Then f ∈ W0k,p (U ) for 1 ≤ p < ∞.
Proof. We proceed as in Theorem 4.22. Let Kj = {x ∈ U : dist(x, ∂U ) ≥
2−j }. We extend χKj by 0 to Rd and convolve it with a smooth function (say
ϕ2−j−1 ) of integral 1 supported in B2−j−1 (0), to obtain ηj ∈ Cc∞ (U ). Then
supp ηj ⊂ {x ∈ U : dist(x, ∂U )} ≤ 2−j−1 and ηj (x) = 1 for dist(x, ∂U ) ≥
21−j and
|∂ α ηj | ≤ c(|α|)2|α|j .
Since for |α| ≤ k, by the Taylor formula,

|∂ α f (x)| ≤ c dist(x, ∂U )k−|α| .

Thus the sequence ηj f is uniformly bounded in Cbk (U ), and hence in W k,p (U ).

Moreover
∂ α (ηj f ) → ∂ α f
for every x ∈ U and by dominated convergence

ηj f → f in W k,p (U ).

We complete the proof by regularizing ηj f as in Lemma 4.6.

68 [February 10, 2017]

The cofactor matrix cof A of an n × n matrix has as (i, j) entry (−1)i−j
times the determinant of the (n − 1) × (n − 1) matrix obtained from A by
removing the ith row and the jth column. It is the same as the partial
derivative of det(A) with respect to the (i, j)th entry. Then (linear algebra)

AT cof A = det A1n×n (4.1)

Lemma 4.39. Let U ⊂ Rd be open and φ ∈ C 2 (U ; Rd ). Then

d
X
∂xj cof(Dφ)ij = 0.
j=1

Proof. ¿From (4.1) with A = Dφ,

d
X
∂xi det(Dφ) = δij ∂xj det(Dφ)
j=1
d
X d
X
=δij (cof Dφ)km ∂x2m xj φk
j=1 k,m=1
d
X d
X
k
= (∂xi ∂xj 0 φ )(cof Dφ)kj 0 + ∂xi φk ∂j cof(Dφ)kj
j 0 ,k=1 j,k=1
d
X
∂x2i xj φk (cof Dφ)kj + ∂xi φk ∂xj (cof Dφ)kj .

=
j,k=1

which implies  
d
X d
X
∂xi φk  ∂xj (cof Dφ)kj  = 0.
k=1 j=1

This implies the claim if det Dφ 6= 0. If det Dφ(x) = 0 we apply the reason-
ing to φ + εx and send ε to 0.

Lemma 4.40. If φ : V → U is a Cbk diffeomorphism (bounded derivatives

of φ and φ−1 ) then there exists C > 1 so that

kf ◦ φkW k,p (V ) ≤ Ckf kW k,p (U ) .

Moreover the chain rule holds

d
X
∂yj (f ◦ φ) = (∂xk f ◦ φ)∂yj φk .
k=1

69 [February 10, 2017]

Proof. The first claim follows from the chain rule and the transformation
formula. We prove the chain rule for a smooth diffeomorphism. The general
case follows by approximating the diffeomorphism and taking limits. We
write ψ ◦ φ(y) = ψ̃(y). Then
ˆ ˆ d
X ∂xk
− d
(f ◦ φ(y))∂yj ψ̃(y)dm (y) = − f (x) ∂xk ψ det(Dφ−1 )dmd (x)
V U ∂yj
k=1
d ˆ
X ∂xk
=− f (x)∂k ( ψ det(Dφ−1 ))dmd (x)
∂yj
k=1 U
d ˆ
X ∂xk
+ f (x)ψ∂xk ( det(Dφ−1 ))dmd (x)
U ∂yj
k=1

The second term vanishes by Lemma 4.39 and we continue, assuming that
φ is smooth (which requires an approximation argument)
d ˆ
X ∂xk
= (∂xk f ) det(Dφ−1 )dmd (x).
U ∂yj
k=1

The chain rule follows by another application of the transformation formula.

4.5 (Whitney) extension and traces

Definition 4.41. Let U ⊂ Rd be open, bounded and connected. We say that
U is a Lipschitz domain if there exist a continuous vector field ν on ∂U and
a Lipschitz continuous function ρ and c > 0 so that ∂U = ρ−1 ({0}) and

ρ(x + tν) − ρ(x + sν) ≥ t − s

for x ∈ ∂U and −c < s < t < c.

Examples:

1. Bounded connected open sets with C 1 boundary.

2. Let h : Rd−1 → R be Lipschitz continuous with Lipschitz constant L.

The set below the graph is not compact, but the other conditions are
satisfied with ρ(x) = xd − h(x1 , · · · , xd−1 ) and ν = ed .

Theorem 4.42 (Whitney). Let 1 ≤ p < ∞ and U ⊂ Rd be a Lipschitz

domain. Then there exists a linear extension map

W k,p (U ) → W k,p (Rd ).

70 [February 10, 2017]

Proof. We prove the theorem under the stronger assumption that ∂U is a
C k manifold. By use of rotation, compactness and partition of unity, it suf-
fices to consider the extension problem for U = {x : xd < ψ(x1 , . . . , xd−1 )}
where ψ is a function in Cbk . By Lemma 4.40 we can choose φ(x) =
(x1 , . . . , xd−1 , xd − ψ(x1 , x2 , . . . xd−1 )) to reduce the problem to extending
Sobolev functions on the lower half space V = {x|xd ≤ 0}. Let f be defined
on V . We make the Ansatz
if xd ≤ 0

 f (x)
k+1
F (x) = P
 aj f (x1 , . . . xd−1 , −jxd ) if xd > 0.
j=1

If f ∈ C k (V ) we want to choose the aj so that ∂ α F is continuous for

|α| ≤ k. It clearly suffices to do this for ∂xj d for 0 ≤ j ≤ k when {xd = 0},
which leads to the Vandermonde matrix
  
1 1 1 ... 1 a1 1

 1 2 3 ... k+1  a2   −1 
  
 
12 22 32 . . . (k + 1)2   a3

= 1  .
 
 
 .. .. .. . . .. .. ... 
  
. . . . .  . 
1k 2k 3k . . . (k + 1)k ak+1 (−1)k

The Vandermonde matrix is invertible and we can solve this system. The
coefficients aj hence exist and depend only on k. Then

kF kW k,p (Rd ) ≤ Ckf kW k,p (V ) .

Now we would like to pass from the assumption f ∈ C k (U ) to f ∈

W k,p (U ).We define the extension F in the same way. Then we have to prove
that for |α| ≤ k the distributional derivative is given by the distribution
defined by the distributional derivatives on both sides. By an application of
the theorem of Fubini and using distributional derivatives in d − 1 variables
this is true for all α with αd = 0. Again by Fubini it suffices to consider
d = 1. Let 1 ≤ κ ≤ k and φ ∈ D(R). Then, with
 κ
 ∂ f if x < 0
κ k+1
∂ F = P dκ
 aj dx κ (f (−jx)) if x > 0
j=1

71 [February 10, 2017]

we obtain
ˆ ˆ 0 k+1 ˆ ∞
κ dκ dκ X dκ
(−1) F κ φdx =(−1)κ f κ φdx + (−1)κ aj f (−jx) φ(x)dx
R dx −∞ dx 0 dxκ
j=1
ˆ 0 k+1
dκ h κ
X
κ−1
i
= f (−1) φ + j aj φ(−x/j) dx
−∞ dxκ
j=1
ˆ 0 k+1
dκ X
= f (x)[φ(x) − aj (−j)κ−1 φ(−x/j)]dx
−∞ dxκ
j=1
ˆ
dκ
= F φ(x)dx.
R dxκ
In last equality we observe that
0 k+1
dκ X
0 (φ(x) − aj (−j)κ−1 φ(−x/j)) = 0
dxκ
j=1

for x = 0 and κ0 < κ by the definition of the aj : Indeed, we observe that

k+1 0
0
X
1− aj (−j)κ−κ −1 φ(κ ) (0) = 0.
j=1

Hence it is in W0κ,p ((−∞, 0)) by Lemma 4.38 and we can approximate it by

functions in D((−∞, 0)). For those we can move the derivatives to f .

Corollary 4.43. Suppose that U is a open, bounded with Lipschitz bound-

ary, k ∈ N and 1 ≤ p < ∞. Then the restrictions of C0∞ (Rd ) functions is
dense in W k,p (U ).

[21.12.2016]
[23.12.2016]

Theorem 4.44 (Traces). Let U be a bounded domain with C 1 boundary and

let f ∈ W 1,p (U ), 1 ≤ p ≤ ∞. Then there is a unique trace g ∈ Lp (∂U ) so
that
ˆ d
X ˆ d
X ˆ X
d
F j ν j gdHd−1 = f ∂xj F j dmd + ∂xj f F j dmd , (4.2)
∂U j=1 U j=1 U j=1

where F j ∈ C 1 (U ) and ν denotes the outer normal vector of ∂U . It satisfies

p−1 1
kgkLp (∂U ) ≤ ckf kLpp (U ) kDf kLp p , kDf kLp := k |(∂xj f )| kLp .

72 [February 10, 2017]

Proof. If f ∈ C 1 (U ) then g = f |∂U has the desired properties. By a partition
of unity we may assume that U is below the graph of a Cb1 function φ, and
that f has compact support. Then, if p < ∞,
ˆ ˆ
1
kgkpLp (∂U ) = |g|p dmd−1 = |g|p (1 + |∇φ|2 ) 2 dmd−1 (x0 )
ˆ Rd−1
1
=p Re (1 + |∇0 φ(x0 )|2 ) 2 |f |p−1 f¯∂xd f dmd
U
1
≤p sup(1 + |∇0 φ|2 ) 2 kf kp−1
Lp k∂xd f kL .
p

We approximate f by smooth functions, and go to the limit. Then we obtain

a function g at the boundary so that (4.2) holds. We fix a vector field F so
that F · ν ≥ 0 at the boundary. Let h ∈ C 1 (∂U ) and it has an extension to
C 1 (U ), which we denote again by h. We apply (4.2) with F̃ = hF . Then
the right hand side of (4.2) (with F replaced by F̃ ) determines
ˆ
(F · ν)hgdHd−1
∂U

for h ∈ C 1 (∂U ). This determines g uniquely. The case p = ∞ is simpler.

4.6 Finite differences

Here we want to relate the analogue of finite differences to Sobolev functions.
Theorem 4.45. a) Let 1 < p ≤ ∞. If f ∈ Lp and
kf (. + h) − f kLp (Rd )
sup ≤C (4.3)
h |h|

then f ∈ W 1,p and

f (. + tej ) − f
k∂xj f kLp ≤ sup ≤ C.
t6=0 t Lp (Rd )

b) Now let 1 ≤ p ≤ ∞ and f ∈ W 1,p . Then

Lp

f (. + tej ) − f (.) if p < ∞
→ ∂j f in
t D0 if p = ∞

Proof. Suppose that (4.3) holds. Then

f (x + te1 ) − f (x)
lim → ∂x1 f
t→0 t
as distribution:
ˆ
1 1
Tft (φ) = (f (x + te1 ) − f (x))φ(x)dmd = Tf ( (φ(x − te1 ) − φ(x)))
t t

73 [February 10, 2017]

and in D(Rd )
1
(φ(x − te1 ) − φ(x)) → −∂x1 φ(x).
t
p
The difference quotient defines an element in L p−1 (Rd )∗ . It is bounded by
p
C uniformly with respect to t. Then also ∂x1 f defines an element in (L p−1 )∗
with norm at most C. By the representation theorem there exists ∂x1 f ∈ Lp .
This proves the first direction and

f (tej + .) − f
k∂xj f kLp (Rd ) ≤ lim inf .
t→0 t Lp

Consider now f ∈ W 1,p (U ), 1 ≤ p ≤ ∞ and we would like to show (4.3).

By the fundamental theorem of calculus and Minkowski’s inequality
ˆ d
1X
kf (. + h) − f (.)kLp (Rd ) = k hj ∂j f (. + sh)dskLp (Rd ) ≤ |h|k|Df |kLp
0 j=1

for C 1 ∩ W 1,p functions. Density completes the argument for p < ∞. For
p = ∞ we use that by the previous argument
ˆ ˆ 1ˆ d
X
d
(f (x + h) − f (x))φ(x)dm (x) = ∂xj f (x + th)hj φ(x)dmd (x)dt
0 Rd j=1

≤|h|k|∇f |kL∞ kφkL1

and hence
f (x + h) − f (x)
k kL∞ ≤ k|∇f |kL∞ .
|h|

Corollary 4.46. [Poincaré inequality] Let U ⊂ {x ∈ Rd : a < x1 < b} and

f ∈ W01,p (U ). Then

kf kLp (U ) ≤ |b − a|k|∇f |kLp (U ) .

Proof. Extend f by 0 to Rd and apply the previous theorem.

We define ˆ
fB = f dmd = (md (B))−1 f dmd .
B B

Lemma 4.47 (Poincaré inequality on ball). If 1 ≤ p ≤ ∞ and f ∈

W 1,p (BR (0)) then
d
kf − fBR (0) kLp (BR (0)) ≤ 2 p Rk|Df |kLp (BR (0)) .

74 [February 10, 2017]

Proof. It suffices to consider R = 1 by replacing f by f (x/R). We calculate
again by Minkowski’s and Jensen’s inequality
ˆ
d p
(m (B1 (0)) |f − fB1 (0) |p dmd (x)
B1 (0)
ˆ ˆ p
= (f (x) − f (y))dm (y) dmd (x)
d
B1 (0) B1 (0)
ˆ ˆ
≤ |f (x) − f (y)|p dmd (x)dmd (y)
B1 (0)×B1 (0)
ˆ ˆ 1 p
= ∇f (x + t(y − x))dt |x − y|p dmd (x)dmd (y)
B1 (0)×B1 (0) 0
ˆ 1ˆ
p
≤2 |∇f (x + t(y − x))|p dmd (x)dmd (y)dt
0 B1 (0)×B1 (0)
ˆ 1 ˆ
2
p
=2 · 2 |∇f (x + t(y − x))|p dmd (x)dmd (y)dt
0 B1 (0)×B1 (0)

where we used symmetry in x and y in the last equality. However, if y ∈

B1 (0) and 0 ≤ t ≤ 21 then
ˆ ˆ
|∇f (x + t(y − x))|p dmd (x) ≤ 2d |∇f (x)|p dmd (x)
B1 (0) B1 (0)

by the transformation formula.

[23.12.2016]
[11.01.2017]

There has been an omission in the proof of Theorem 4.45: We have not
shown that the difference quotient converges in Lp to the derivative - only
as a distribution.

4.7 Sobolev inequalities and Morrey’s inequality

Lemma 4.48. Let f ∈ C0 (R), f 0 ∈ L1 . Then the Sobolev inequality
1
kf ksup ≤ kf 0 kL1 (4.4)
2
holds. If f 0 ∈ Lp , 1 ≤ p ≤ ∞ then every point is a Lebesgue point and

|f (x) − f (y)|
sup ≤ kf 0 kLp . (4.5)
1− p1
x6=y |x − y|

75 [February 10, 2017]

Proof. It suffices to prove the estimate for smooth functions with compact
support. The inequality (4.4) is a consequence of the fundamental theorem
of calculus: ˆ x ˆ ∞
0
f (x) = f (y)dy = − f 0 (y)dy.
−∞ x
and we derive (4.5) by Hölder’s inequality
ˆ y
1− p1
|f (x) − f (y)| ≤ |f 0 (z)|dz ≤ kf 0 kLp kχ[x,y] k p = |x − y| kf 0 kLp .
x L p−1

This proof applies to f ∈ C 1 . The general statement follows as in Morrey’s

inequality below.

The Sobolev inequality and Morrey’s inequality are the versions of these
inequalities (4.4), (4.5) in higher space dimension.
Theorem 4.49 (Morrey). Let U be open. Suppose that p > d and f˜ ∈
W 1,p (U ). Every point is a Lebesgue point and the canonnical representative
f is continuous. There exists c depending on p and d so that the following
is true: Let x, y ∈ U with

|x − y| < dist(x, Rd \U ).

Then
1− dp
|f (x) − f (y)| ≤ c|x − y| k|∇f |kLp (B|x−y| (x)) .
Proof. The inequality follows from
1− pd
f (y) − fBR (x0 ) ≤ CR k|∇f |kLp (BR (x0 )) (4.6)

for |y − x0 | < R (and Lebesgue points y) with a constant C(p, d) which is

bounded as p → ∞ by

|f (y) − f (x0 )| ≤ f (y) − fBR (x0 ) + f (x0 ) − fBR (x0 )

and two applications of (4.6) if y and x0 are Lebesgue points.

It remains to prove (4.6) and to prove that every point is a Lebesgue
point. We have for BR/2 (y) ⊂ BR (x0 )
ˆ
d −1
(m (BR )) f (x0 + z) − f (y + z/2)dmd (z)
BR (0)
ˆ ˆ d
1X
d −1
= (m (BR )) (x0 + z/2 − y)j ∂xj f (x0 + (1 − t/2)z + t(y − x0 ))dtdmd (z)
BR (0) 0 j=1
ˆ
d d −1
≤ R2 (m (BR )) |∇f |dmd (x)
BR (x0 )
− p1 1− dp
≤ 2d (md (B1 (0))) R k∇f kLp (BR (x0 ))

76 [February 10, 2017]

first for smooth functions, and then by approximation for Sobolev functions.
By a geometric series and an iterative application of the above inequality
with (x0 , y) = (xj−1 , xj ):

∞
X
|f (y) − fBR (x0 ) | = fB2−j R (xj ) − fB21−j R (xj−1 )
j=1
∞
X
≤ |fB2−j R (xj ) − fB21−j R (xj−1 ) |
j=1
∞
1− pd −j(1− pd )
X
≤cR 2 k|∇f |kLp (BR (x0 ))
j=1

provided y is a Lebesgue point and for a suitable converging sequence of xj .

This however follows from the convergence for every point.

Theorem 4.50 (Rademacher). I) Lipschitz continuous functions are almost

everywhere differentiable.
II) Functions in W 1,p (U ) with p > d are almost everywhere differentiable.
The derivative almost everywhere is the same as the weak derivative.

Proof. Part I is an exercise and we prove Part II. Let f ∈ W 1,p . By Morrey’s
theorem 4.49 we know that every point is a Lebesgue point and there is a
uniformly continuous representative. By Theorem 3.51 there exists a set A
whose complement has zero measure so that every x ∈ A is a Lebesgue point
and ˆ
lim md (Br (x))−1 |∇f (y) − ∇f (x)|p dmd (y) = 0.
r→0 Br (x)

We apply the Morrey’s inequality Theorem 4.49 to

d
X
v(y) = f (x + y) − f (x) − ∂j f (x)yj
j=1

on Br (0) where again x ∈ A . Then

ˆ !1
p
1− dp
|v(y)| ≤cr |∇f (x + z) − ∇f (x)|p dmd (z)
Br (0)
ˆ !1
p
d −1 p d
=cr m (Br (x)) |∇f (z) − ∇f (x)| dm (z)
Br (x)

=o(r).

This implies that f is differentiable at x ∈ A.

77 [February 10, 2017]

Theorem 4.51 (Sobolev). Suppose that 1 ≤ p < d and
1 1 1
+ = .
q d p
Then
kf kLq (Rd ) ≤ ck|Df |kLp (Rd )
whenever f ∈ Lq (Rd ) and |Df | ∈ Lp (Rd ).
d
Proof. We prove the estimate first for p = 1 and q = d−1 . More precisely
we prove the estimate
d
Y
kf kd d ≤ 2−d k∂j f kL1 (Rd ) (4.7)
L d−1 (Rd )
j=1

by induction on the dimension. The case d = 1 has been contained in Lemma

4.48. Suppose we have proven the estimate for d ≤ k − 1. Then by Fubini
1
and Hölder’s inequality (since k−1 + k−2k−1 = 1)
k
ˆ ˆ
1
kf k k−1k = |f | k−1 |f |dmk−1 dm1 (x1 )
L k−1 (Rk ) k−1
ˆR R 1
≤ kf (x1 , . . . )kLk−11 (Rk−1 ) kf (x1 , . . . )k k−1 dm1 (x1 )
L k−2 (Rk−1 )
R
1
ˆ
≤ sup kf (x1 , .)kLk−1
1 (Rk−1 ) kf (x1 , .)k k−1 dm1 (x1 )
x1 R L k−2 (Rk−1 )
1
ˆ
 
k k−1
1
− k−1
Y
≤2 k∂x1 f kL1 (Rk ) 2−(k−1) k∂xj f (x1 , .)kL1 (Rk−1 )  dm1 (x1 )
R j=2

We take the inequality to the power k − 1, and apply Hölder’s inequality

in the form
ˆ Y k ˆ
 k−1
k
1 Y
1
 |gj | k−1 dm  ≤ |gj |dm1
j=2 j=2

to arrive at
k
Y
k −k
kf k k ≤2 k∂j f kL1 (Rk ) .
L k−1 (Rk )
j=1
(d−1)p
Now let 1 < p < d. We apply the above inequality (4.7) to |f | d−p .
Then
(d−1)p (d−1)p
kf kLqd−p
(Rd )
=k|f | d−p k d
L d−1 (Rd )
ˆ
(d−1)p (d − 1)p d(p−1)
≤kD|f | d−p kL1 (Rd ) ≤ |f | d−p |Df |dmd
d−p
d(p−1)
(d − 1)p
≤ kf kLqd−p
(Rd )
kDf kLp (Rd )
d−p

78 [February 10, 2017]

where we first argue for smooth functions, and where we used Hölder’s in-
equality in the last step.

[11.01.2017]
[13.01.2017]

4.8 Applications to PDE

Let U ⊂ Rd be open and bounded, f : U × R × Rd → R be continuous and
suppose that there exists 1 ≤ p ≤ r < ∞ such that
1 1 1
+ ≥
r d p
and
|f (x, u, P )| ≤ c(1 + |u|r + |P |p ). (4.8)
If u ∈ W 1,p (U ) then
x → f (x, u(x), ∇u(x))
is measurable and integrable since
ˆ ˆ ˆ
d d r d
|f (x, u(x), ∇u(x))|dm ≤ c(m (U ) + |u| dm + |∇u|p dmd )
U U U
r p
≤ c̃(1 + kukLr (U ) + kukW 1,p (U ) ),

where we put all the constants into c̃. If 1 ≤ p < d by Hölder’s, Poincaré’s
and Sobolev’s inequality, with 1q + d1 = p1 and
1 1
p − r
α= 1 1,
p − q

we have
1−α α
kukLr (U ) ≤ kukLp (U ) kukLq (U ) ≤ CkukW 1,p (U ) ,

where q is the exponent of the Sobolev inequality. Here we assume that

either U ⊂ Rd such that Whitney extension Theorem 4.42 holds true, or
u ∈ W01,p (U ).
Definition 4.52. Let U ⊂ Rd be open and bounded,

F j , f ∈ C(U × R × Rd )

for 1 ≤ j ≤ d and suppose these functions to satisfy (4.8). Then u ∈

W 1,p (U ) is called weak solution to
d
X
∂j F j (x, u, ∇u) = f (x, u, ∇u)
j=1

79 [February 10, 2017]

if this identity holds in the sense of distributions. This holds iff
ˆ X
d
F j (x, u, ∇u)∂j φ + f (x, u, ∇u)φ dmd = 0
U j=1

for all φ ∈ D(U ).

Remark 4.53. We may replace the condition φ ∈ D(U ) by φ ∈ C 1 (U ) with

φ|∂U = 0. If 1 < p we may replace it by φ ∈ W01,q (U ) where 1q + p1 = 1.

We turn to a particular case. Let

aij ∈ L∞ (U )

F j , f ∈ Lp (U )
for 1 ≤ i, j ≤ d. Then u ∈ W 1,p (U ) is a weak solution to
d
X d
X
∂i (aij (x)∂j u) = ∂j F j + f
i,j=1 j=1

if and only if
ˆ d
X d
X
ij
a (x)∂j u∂i φ − F j (x)∂j φ + f φ̄ dmd = 0
U i,j=1 j=1

for all φ ∈ C 1 (U ) with φ|∂U = 0.

If g ∈ W 1,p (U ) we say u = g on ∂U in the W 1,p sense if

u − g ∈ W01,p (U ).

Definition 4.54. We call the (aij )1≤i,j≤d elliptic if there exists κ > 0 so
that
Xd
Re aij (x)ξi ξj ≥ κ|ξ|2
i,j=1

for almost all x ∈ U and all ξ ∈ Cd .

Let F j , f ∈ L2 (U ), g ∈ W 1,2 (U ). We consider the boundary value

problem
Xd d
X
∂i (aij ∂j u) = ∂j F j + f in U
i,j=1 j=1

u=g on ∂U

80 [February 10, 2017]

Theorem 4.55. There exists exactly one weak solution u ∈ W 1,2 (U ) which
satisfies the boundary condition in the W 1,2 sense. The map

(L2 (U ))d+1 × W 1,2 (U ) 3 (F, f, g) → u ∈ W 1,2 (U )

is a continuous linear map.

Proof. Step 1: Reduction Formally setting w = u − g it suffices to find
w ∈ W01,2 (U ) such that
d
X d
X d
X
∂i (aij ∂j w) = ∂j (F j − aji ∂i g) + f in U
i,j=1 j=1 i=1

which reduces the problem to the case g = 0.

Step 2: The Hilbert space We recall that W01,2 (U ) is a Hilbert space
- we can consider it as a closed subspace of (L2 (U ))d+1 , and closed subspaces
of Hilbert spaces are Hilbert spaces. The inner product is
ˆ X
d
hu, vi = ∂j u∂j vdmd
U j=1

Step 3: The quadratic form We define the quadratic form

ˆ d
X
A(u, v) = aij ∂j u∂i vdmd .
U i,j=1

The quadratic form is continuous - there exists C > 0 so that

|A(u, v)| ≤ CkukW 1,2 (U ) kvkW 1,2 (U )

and it satisfies ˆ
Re A(u, u) ≥ κ |∇u|2 dmd
U

for all u, v ∈ By the Poincaré inequality in W01,2 (U ) (since U is

W 1,2 (U ).
bounded) there exists c > 0 so that

kukL2 (U ) ≤ ck∇ukL2

and thus there exists κ̃ > 0 so that

Re A(u, u) ≥ κ̃kuk2W 1,2 (U ) .

0

Step 4: The linear form Let F and f be as in the theorem. Then

ˆ d
X
v → Lv = − F j ∂j v + f v
U j=1

81 [February 10, 2017]

satisfies
kLk(W 1,2 (U ))∗ ≤ kF k(L2 (U ))d + kf kL2 (U )
0

Step 5: The lemma of Lax-Milgram. By the Lemma of Lax-Milgram

there is a unique u ∈ W01,2 (U ) such that

A(u, v) = L(v), ∀v ∈ W01,2 (U ),

which is then a weak solution. The map L → u is linear and continuous.

Step 6: The bound Let v = u and take the real part. Then

κk∇uk2L2 ≤kF kL2 k∇ukL2 + kf kL2 kukL2

≤kF kL2 k∇ukL2 + Ckf kL2 k∇ukL2

and hence
κk∇ukL2 ≤ kF kL2 + Ckf kL2 .
we complete the bound by a second application of the Poincaré inequality.

There is a particular case aij = δ ij and F = 0. It becomes the Poisson

problem
−∆u = f in U
u=g on ∂U
We obtain a unique weak solution for U open and bounded, f ∈ L2 (U ) and
g ∈ W 1,2 (U ).
If there is a barrior at every boundary points then the problem

−∆u = 0 in U

u=g on ∂U
has a unique solution u ∈ C 1 (U ) ∩ C(U ) for every continuous g. We call
this solution classical.
A classical solution is not necessarily in W 1,2 , and a weak solution is not
necessarily in C(U ). In most important cases both solutions are identical.
[13.01.2017]
[18.01.2017]

5 Linear Functionals
In this section we will study the dual space X ∗ of Banach spaces X.

82 [February 10, 2017]

5.1 The Theorem of Hahn-Banach
Definition 5.1. Let X be a K vector space. A map p : X → R is called
sublinear if

1. p(λx) = λp(x) for x ∈ X and λ ≥ 0,

2. p(x + y) ≤ p(x) + p(y) for x, y ∈ X.

Examples:

1. The norm of a normed space is sublinear.

2. If K = R, any element of X ∗ is sublinear.

3. The Minkowski functional of a convex set. Let K ⊂ X be convex such

that for every x ∈ X there exists λ > 0 so that λx ∈ X. we define
1
pK (x) = inf{λ > 0 : x ∈ K} ∈ [0, ∞).
λ
It is not difficult to verify that pK is sublinear. A norm is the Minkowski
functional of the unit ball.

Theorem 5.2 (Hahn-Banach, real case). Let X be a real vector space, Y ⊂

X a subvector space, p : X → R sublinear and l : Y → R linear such that

l(y) ≤ p(y) for all y ∈ Y.

Then there exists L : X → R linear so that

1. l(y) = L(y) for all y ∈ Y

2. l(x) ≤ p(x) for all x ∈ X.

Proof. There are two very different steps.

Suppose that Y 6= X. Then there exists x0 ∈ X\Y . Let Y1 be the space
spanned by Y and x0 . Every element of Y1 can uniquely be written as

y + rx0 , y ∈ Y, r ∈ R.

We want to find a linear map l1 : Y1 → R such that

1. l1 (y) = l(y) for y ∈ Y

2. l1 (y + sx0 ) ≤ p(y + sx0 ) for s ∈ R and y ∈ Y .

83 [February 10, 2017]

By the first condition and linearity we have to find t = l1 (x0 ) so that

l(y) + st ≤ p(y + sx0 )

for all y ∈ Y and s ∈ R. We consider s > 0. Then this inequality is

equivalent to
st ≤ s(p(y/s + x0 ) − l(y/s))
for all s > 0 and y ∈ Y

⇐⇒ t ≤ inf p(y + x0 ) − l(y).

y

Similarly the inequality holds for s < 0 if and only if

t ≥ sup l(y) − p(y − x0 ).

y

We can find t if and only if

l(y) − p(y − x0 ) ≤ p(ỹ + x0 ) − l(ỹ) for all y, ỹ ∈ Y

which follows from the inequality on Y and sublinearity of p:

l(y) + l(ỹ) = l(y + ỹ) ≤ p(y + ỹ) ≤ p(y + x0 ) + p(ỹ − x0 ).

This completes the first step.

For the second step we need the axiom of choice in the form of Zorn’s
lemma.
Let Z be a partially ordered set which contains an upper bound for every
chain. Then there is a maximal element.
A chain is a totally ordered subset, i.e. a subset A so that always either
a ≤ b or b ≤ a. An element b is an upper bound for the chain A, if a ≤ b for
all a ∈ A. An element a ∈ Z is maximal if b ∈ Z, b ≥ a implies b = a.
We define

Z = {(W, lW ) : Y ⊂ W, lW |Y = l, lW (w) ≤ p(w) for w ∈ W },

with the ordering

(W, lW ) ≤ (V, lV ) if W ⊂ V and lV |W = lW .

This is a partial order. If Z̃ is a chain then

[
V = W
(W,lW )∈Z̃

with the obvious lV being an upper bound for the chain. Now let (V, lV )
be a maximal element. If V = X we are done. Otherwise we obtain a
contradiction by the first step.

84 [February 10, 2017]

There is a complex version. It relies on the observation
Lemma 5.3. Let X be a complex vector space and l : X → R be R linear.
Then it is the real part of the linear map

lC (x) = l(x) − il(ix).

The real part determines lC .

Proof. We have to show complex linearity. Real linearity is obvious. We
compute

lC (ix) = l(ix) − il(iix) = i(l(x) − il(ix)) = ilC (x).

Theorem 5.4 (Complex Hahn Banach). Let X be a complex vector space,

Y a subvector space, p sublinear and l : Y → C linear so that

Re l(y) ≤ p(y) for y ∈ Y.

Then there exists L : X → C linear so that

1. L|Y = l

2. Re L(x) ≤ p(x).
Proof. We apply the real theorem of Hahn Banach to the real part, and
extend it to a complex linear map by Lemma 5.3. To complete the proof we
observe that L and L̃ are the same iff the real parts are the same.

We formulate the consequences for normed vector spaces, making use of

the fact that norms are sublinear.
Theorem 5.5. Let X be a normed K vector space, Y a subspace and l :
Y → K a continuous linear map. Then there exists L : X → K linear and
continuous so that
1. L|Y = l

2. kLkX ∗ = klkY ∗ .
Proof. We define
p(x) = klkY ∗ kxkX .
Then
Re l(y) ≤ p(y)
for all y ∈ Y . We apply the theorems of Hahn-Banach to obtain L ∈ X ∗ so
that L|Y = l and
Re L(x) ≤ p(x).

85 [February 10, 2017]

This implies

|L(x)| = Re αL(x) = Re L(αx) ≤ p(αx) = p(x) = klkY ∗ kxkX

for some α ∈ C with |α| = 1. Thus

kLkX ∗ = klkY ∗ .

5.2 Consequences of the theorems of Hahn-Banach

Lemma 5.6. Let X be a normed space and x ∈ X. There exists x∗ ∈ X ∗
with kx∗ kX ∗ = 1 and x∗ (x) = kxkX .

Proof. Let x ∈ X\{0} and let Y be the span of x. Y is one dimensional

and we define y ∗ ∈ Y ∗ by y ∗ (rx) = rkxkX . Then ky ∗ kY ∗ = 1. We apply
Theorem 5.5 to obtain x∗ . If x = 0 we choose x0 6= 0 and find x∗ for x0 .

Corollary 5.7. Let X be a normed space. If x ∈ X then

kxkX = sup{Re x∗ (x) : x∗ ∈ X ∗ , kx∗ kX ∗ = 1}.

If x∗ ∈ X ∗ then

kx∗ kX ∗ = sup{Re x∗ (x) : x ∈ X, kxkX = 1}.

Proof. The first claim is a consequence of Lemma 5.6. The second statement
is an immediate consequence of the definition.

Lemma 5.8. If X is a normed space and Y ⊂ X is a closed subspace,

Y 6= X, then there exists x∗ ∈ X ∗ , x∗ 6= 0 with x∗ |Y = 0.

Proof. Let x0 6∈ Y . We define

l(y + tx0 ) = t

and extend it to X by Theorem 5.5.

Corollary 5.9. Let X ∗∗ be the dual space of X ∗ . There is a cannonical

isometry
J : X → X ∗∗ , J(x)(x∗ ) = x∗ (x).

Proof. Only the isometry has to be shown

kJ(x)kX ∗∗ = sup Re J(x)(x∗ ) = sup x∗ (x) = kxkX .

kx∗ kX ∗ =1 kx∗ kX ∗ =1

86 [February 10, 2017]

p
Remark 5.10. If X = Lp (µ), 1 < p < ∞ then X ∗ is isomorphic to L p−1 (µ)
and J is surjective.

Lemma 5.11. (l∞ )∗ 6= l1 .

Proof. The space of converging sequences c is a closed subspace of l∞ . Let

l : c → K be defined by
l((xj )) = lim xj .
j→∞

Then
|l((xj ))| ≤ k(xj )kl∞
for every converging sequence. By Theorem 5.5 it has an extension L to l∞ .
Clearly L(ej ) = l(ej ) = 0. We claim that it cannot be represented in the
form
∞
X
L((xj )) = yj xj
j=1

for (yj ) ∈ l1 - if it were represented in this fashion then all yj would have to
vanish.

In particular J : l1 → (l1 )∗∗ is not surjective.

[18.01.2017]
[20.01.2017]

Lemma 5.12. If X is a normed space and X ∗ is separable then X is sepa-

rable.

Proof. Since X ∗ is separable, and a subset of a separable set is separable

also the unit sphere is separable. Let x∗j be a dense sequence of unit vectors.
We choose a sequence of unit vectors xj with x∗j (xj ) ≥ 21 . We claim that the
span U of the xj is dense. Otherwise, by Lemma 5.8 there exists x∗ ∈ X ∗
of norm 1 which vanishes on the closure of U , and in particular x∗ (xj ) = 0
for all j. By density there exists j such that kx∗ − x∗j kX ∗ < 21 and hence

1 1
≤ Re x∗j (xj ) = Re(x∗ (xj ) + (x∗j − x∗ )(xj )) <
2 2
This is a contradiction.

Lemma 5.13. Let U ⊂ Rd be open, 1 ≤ p < ∞ and k ∈ N. The map

p X
J : L p−1 (U × Σk ) 3 (gα ) → (f → gα ∂ α f ) ∈ (W k,p (U ))∗
|α|≤k

is bounded and surjective.

87 [February 10, 2017]

Proof. The map

W k,p (U ) 3 f → (∂ α f )|α|≤k ∈ Lp (U × Σk )

map W k,p isometrically to a closed subspace. Any y ∗ ∈ (W k,p )∗ defines a

linear functional on this closed subspace. By Theorem 5.5 we can extend
it to the whole of Lp (U × Σk ). This can be represented by a function in
p
L p−1 (U × Σk ). We have seen that J has norm 1 in an exercise.

5.3 Separation theorems

Lemma 5.14. Let X be a normed space and K ⊂ X convex. If 0 is in
the interior of K then for every x ∈ X there exists λ > 0 so that λx ∈ K.
Moreover there exists C > 0 so that

pK (x) ≤ CkxkX .

If X is a Banach space, and for every x there exists λ > 0 so that λx ∈ K

then 0 is in the interior of K.

Proof. If 0 is in the interior of K there exists ε > 0 so that Bε (0) ⊂ K. An

easy calculation shows that then pK (x) ≤ ε−1 kxkX .
Suppose that X is Banach and that for every x there exists λ > 0 so
that λx ∈ K. In particular 0 ∈ K. Let
1
An = {x ∈ X : x ∈ K}.
n
S
The set An are closed, convex, 0 ∈ An and X = An . By the Baire
category theorem one and hence all of the An have nonempty interior. In
particular there exist x and ε > 0 so that Bε (x) ⊂ A1 . There exists An
so that −x ∈ An hence −x/n ∈ A1 . The convex hull of Bε (x) and −x/n
contains a ball around 0. Hence 0 is in the interior of A 1 ⊂ K.
2

Lemma 5.15. Let X be a normed vector space and V convex, open with
/ V . Then there exists x∗ ∈ X ∗ with
0∈

Re x∗ (x) < 0 if x ∈ V

Proof. It suffices to consider K = R. The complex case is a consequence of

Lemma 5.3. Let x0 ∈ V and define the translate U = V − x0 . Let pU be the
Minkowski functional of U . It is sublinear. Let y0 = −x0 ∈/ U and Y the
span of y0 . We define

l(ty0 ) = tpU (y0 ) t ∈ R.

88 [February 10, 2017]

Then l(y) ≤ p(y) for all y ∈ Y . By Theorem 5.2 there exists L ∈ X ∗ with
L|Y = l, L(x) ≤ p(x) for x ∈ X (here we use Lemma 5.14). In particular
L(y0 ) ≥ 1 and for x ∈ V and u = x + y0

L(x) = L(u) − L(y0 ) ≤ pU (u) − 1 < 0.

Theorem 5.16 (Separation theorem 1). Let X be a normed space, V and

W disjoint convex sets with V open. Then there exists x∗ ∈ X ∗ such that

Re x∗ (v) < Re x∗ (w) for every v ∈ V, w ∈ W

Proof. Let U = V − W . It is convex and open. Since V and W are disjoint

0∈/ U . By Lemma 5.15 there exist x∗ ∈ X ∗ so that Re x∗ (x) < 0 for x ∈ U .
This implies the desired inequality.

Theorem 5.17 (Separation theorem 2). Let X be a normed space, V convex

/ V . Then there exists x∗ ∈ X ∗ such that
and closed, x ∈

x∗ (x) < inf x∗ (v). (5.1)

v∈V

Proof. We may assume x = 0. Since V is closed there exists ε > 0 so that

Bε (0) ∩ V = {}. We apply Theorem 5.16 to see that there is x∗ ∈ X ∗ with

Re x∗ (u) < Re x∗ (v) for u ∈ Bε (0), v ∈ V

There exists x ∈ Bε (0) so that

Re x∗ (x) ≥ εkx∗ kX ∗ /2 > 0 = x∗ (0)

which implies (5.1) .

Corollary 5.18. Let X be a normed vector space, K ⊂ X comvex and

x ∈ ∂K. Then there exists a half space containing K with x a boundary
point.

Proof. Exercise

5.4 Weak* topology and the theorem of Banach-Alaoglu

Let X be a normed space. The dual space X ∗ is a Banach space, and hence
a metric space. The metric defines open sets, and hence a topology which
we call norm topology.

Definition 5.19. Let A be a set. A family τ of sets is called topology if

1. { }, A ∈ τ .

89 [February 10, 2017]

2. B, C ∈ τ implies B ∩ C ∈ τ .
S
3. If Λ is a set, and for every λ ∈ Λ there is a set Bλ ∈ τ then λ∈Λ Bλ ∈
τ.

We call the elements of τ open. A map is called continuous if the preimage

of open sets is open. A set is called compact, if it is Borel and every open
covering contains a finite subcover.

We want to define a topology on X ∗ . Desired properties are

1. For every x the map X ∗ 3 x∗ → x∗ (x) ∈ K is continuous. Equivalently,

for every open set U ∈ K and x ∈ X

Ux∗ = {x∗ : x∗ (x) ∈ U }

is open.

2. The weak∗ topology is the weakest topology with this property. This
means that the open sets are the smallest subset of the power set, such
that all sets above are contained in it, and arbitrary unions and finite
intersections are contained in it.

Finite intersections of sets of the first type are sets

N
\
(Uj )∗xj (5.2)
j=1

and open sets are arbitrary unions of such sets. This follows from a multiple
application of the distributive law of union and intersection.

Definition 5.20. A local base S of a topology is a family of open sets so

that for every x and every open set U there exists V ∈ S so that x ∈ V ⊂ U .
A subbase is a collection of open sets so that finite intersections form a local
base.

Examples.

1. In metric space the balls {B1/n (x) : x ∈ X, n ∈ N} are a base.

2. The sets (5.2) form a base.

3. The sets {Ux∗ : x ∈ X, U ⊂ K open } are a subbase.

Lemma 5.21 (Alexander). If S is a subbase of a topology, then X is com-

pact if every S cover has a finite subcover.

90 [February 10, 2017]

Proof. We argue by contrapositive and assume that X is not compact. Let
P be the collection of all covers without finite subcover. By assumption P
is not empty. We take the partial order by inclusion. The union of every
element of a chain is an upper bound. By Zorn’s lemma there is a maximal
cover Γ without finite subcover.
Now let Γ̃ = Γ ∩ S. It has no finite subcover since it is a subset of Γ. We
show that Γ̃ covers X, which gives the conclusion.
Arguing by contradiction assume that x ∈ X is in none of the elements
of Γ̃. Γ covers X hence there is W ∈ Γ so that x ∈ W . Since S is a subbase
there are Vj ∈ S so that ∩Nj=1 Vj ⊂ W . Since x is not covered by Γ̃, Vj ∈
/ Γ.By
maximality for each j Γ ∪ {Vj } has a finite subcover
Mj
N [
[
X= Yjk ∪ Vj
j=1 k=1

Hence
Mj
N [
[
X=W ∪ Yjk
j=1 k=1

is a finite subcover of Γ which is a contradiction.

[20.01.2017]
[25.01.2017]

Let D be a set, suppose that for every α ∈ D there is a set Xα . The

cartesian product Y
X= Xα
α∈D

is the set of all ’maps’ which assign to each α the element of Xα . There are
the obvious projections
πα : X → Xα
Suppose that all spaces Xα are topological spaces. Let τ be the smallest
topology (subset of the power set) containing all preimages of open sets in
Xα under πα .
Lemma 5.22. The preimages of open sets in Xα under πα define a subbase
S of τ .
Proof. We define the collection of arbitrary unions of finite intersections of
such sets. Then arbitrary unions and finite intersections have this form.
Thus every open set in τ is a union of finite intersections of such sets. Thus
these sets are a subbase.

Theorem 5.23 (Tychonoff). Any cartesian product X of compact sets Xα

with topology τ is compact.

91 [February 10, 2017]

Proof. Let Γ be a S cover. Let Γα consists of the sets whose preimages
under πα are in Γ. Assume that no Γα covers Xα . Then for all α there exists
xα ∈ Xα so that xα is not covered by Γα . Then x ∈ X with πα (x) = xα
is not covered by Γ . This is a contradiction. Hence at least one Γα covers
Xα . Since Xα is compact, a finite subset of Γα covers Xα and hence the
preimages of this finite subset under πα covers X. By Alexander’s theorem
X is compact.
Theorem 5.24 (Banach-Alaoglu). Let X be a normed space. The closed
unit ball Λ ⊂ X ∗ is compact in the weak∗ topology.
Proof. Denote B1X (0) = {x ∈ X | kxkX ≤ 1} and B1 (0) ⊂ K. Then
Z = {f : B1X (0) → B1 (0)}
is the Cartesian product of compact sets. We consider Λ as subset of Z.
Then Λ carries two topologies: The weak topology, and the topology as
subset of Z. We claim that the two are the same. But this is an immediate
consequence of the definitions of the topologies.
Now let f0 be in the closure of Λ (closed sets are complements of open
sets, and the closure is the smallest closed set containing Λ). We claim that
f0 is linear. Let kxkX , kykX ≤ 1 such that kαx + βykX ≤ 1 and ε > 0. The
set of all f ∈ Λ with
max{|f (x) − f0 (x)|, |f (y) − f0 (y)|, |f (αx + βy) − f0 (αx + βy)|} < ε
is open and contains a linear f of norm 1. Hence
|f0 (αx + βy) − αf0 (x) − βf0 (y)| < ε + |α|ε + |β|ε
for all ε > 0. Hence
f0 (αx + βy) = αf0 (x) + βf0 (y)
for all such x, y, α, β and thus f0 is linear. Similarly the norm of f0 ≤ 1.
Thus Λ is closed in the compact set Z. Any closed subset of a compact set
is compact.
Definition 5.25. Let X be a Banach space and X ∗ its dual. We say a
sequence x∗n ∈ X ∗ is weak* convergent to x∗ if
x∗n (x) → x∗ (x)
for all x ∈ X. We write
x∗n *∗ x∗ .
We say xn ∈ X converges weakly to x ∈ X if
x∗ (xn ) → x∗ (x)
for all x∗ ∈ X ∗ . We then write
xn * x.

92 [February 10, 2017]

Lemma 5.26. Weak and weak* convergent sequences are bounded.

Proof. The statement follows for the weak* converging subsequences by the
uniform boundedness principle. We may consider weakly converging se-
quences as weak∗ converging sequences in X ∗∗ , hence they are bounded.

Theorem 5.27. Let X be a Banach space and K ∈ X a closed bounded

convex set. Let (xj ) be a weakly convergent sequence in K with weak limit
x. Then x ∈ K.

/ K. By Theorem 5.17 there exists x∗ ∈ X ∗ such that

Proof. Let x0 ∈

x∗ (x0 ) < inf x∗ (x).

x∈K

Thus x0 cannot be a weak limit.

Thus
x∗n *∗ x∗ =⇒ kx∗ kX ∗ ≤ lim inf kx∗n kX ∗ (5.3)
n→∞

and
xn * x =⇒ kxkX ≤ lim inf kxn kX . (5.4)
n→∞

Again (5.4) follows form (5.3) by the cannonical embedding J : X → X ∗∗ .

Inequality (5.3) follows from

kx∗ kX ∗ ≤ sup kx∗n kX ∗ ,

which is a consequence of Theorem 5.27 applied with K a closed ball.

Lemma 5.28. Suppose that X is separable. Then there exists a metric on

the closed unit ball of X ∗ so that the topology as metric spaces is the weak∗
topology.

Proof. Let xj be a dense sequence in B1X (0). We define

d(x∗ , y ∗ ) = max 2−j min{1, |x∗ (xj ) − y ∗ (xj )|}.

j

This is a metric. Convergence of bounded sequences in this metric is the

same as weak convergence. Moreover open balls are open in the weak topol-
ogy, and they form a base of the weak topology.

Theorem 5.29. [Weak* compactness of bounded sequences] Every bounded

sequence x∗j ∈ X ∗ where X is separable contains a weakly convergent subse-
quence.

We will provide two proofs of this important fact.

93 [February 10, 2017]

Proof 1. Banach Alaoglu implies that the closed balls are weak* compact.
By Lemma 5.28 the weak* topology comes from a metric which makes X ∗
with the weak* topology a metric space. For metric spaces the different
notions of compactness are equivalent.

Proof 2. Let (xl ) be a dense sequence on the unit ball of X. Let (x∗j ) be a
bounded sequence of element of X ∗ . Then there is a subsequence such that

(x∗jk (x1 ))

converges to y1 such that

|x∗jk (x1 ) − y1 | < 2−k .

Taking repeated subsequences we find a subsequence and a sequence yl such

that
|x∗j̃ (xl ) − yl | < 2−k
k

for k ≥ l. Since x∗j is bounded the yl define a unique continuous linear map
x∗ : X → K. Moreover x∗j *∗ x∗ .

[25.01.2017]
[27.01.2017]

Definition 5.30. We call a Banach space X reflexive, if J : X → X ∗∗ is

surjective.

Hilbert spaces are reflexive. If 1 < p < ∞, Lp (µ) is reflexive as a

consequence of Theorem 3.19.
We define the weak topology in the same fashion as the weak* topology.

Lemma 5.31. Suppose that X is reflexive. Then the weak toplogy of X is

the same as the weak* topology of X ∗∗ .

Proof. Both are the smallest topology such that x → x∗ (x) = J(x)(x∗ ) are
continuous for all x∗ ∈ X ∗ .

Corollary 5.32. Let (X, d) be a σ compact metric space, µ a Radon measure

on X and 1 < p < ∞. Then every bounded sequence in Lp (µ) has a weakly
convergent subsequence. If µ is σ-finite any bounded sequence in L∞ (µ) has
a weak* convergent subsequence.

Proof. By Corollary 3.37 Lp (µ) is separable if 1 ≤ p < ∞. By Theorem

p
3.19 and Corollary 3.20 L p−1 (µ) is isomorphic to the dual space of Lp (µ)
and vice versa when 1 < p < ∞. By Lemma 5.31 the weak and the weak*
topology are the same when 1 < p < ∞. By Theorem 5.29 every bounded
sequence has a weak* convergent subsquence.

94 [February 10, 2017]

Lemma 5.33. Suppose the closed unit ball of a Banach space X is compact.
Then its dimension is finite.
Proof. Let X be not of finite dimension. We claim that there is a sequence of
unit vectors of distance > 21 . We construct them recursively. Let (xn )n≤N
be chosen and let XN denote their span. It is a closed subspace and by
Theorem 5.5 there is x∗ ∈ X ∗ of norm 1 so that x∗ |XN = 0. Thus there
exists xN +1 of norm 1 so that Re x∗ (xN +1 ) > 43 . Then
3
≤ Re x∗ (xN +1 − xn ) ≤ kxN +1 − xn k
4
for n ≤ N . This yields the result.

We collect a number of examples and remarks.

1. A closed subspace Y of a reflexive space X is reflexive. This is seen

as follows: Let 0 6= y ∗∗ ∈ Y ∗∗ . It defines an element x∗∗ ∈ X ∗∗ by
x∗∗ (x∗ ) = y ∗∗ (x∗ |Y ). Since X is reflexive there exists x ∈ X with
x∗∗ (x∗ ) = x∗ (x). We claim that x ∈ Y . If not there exists x∗ ∈ X ∗
such that x∗ (x) = 1 and x∗ |Y = 0. This would imply

0 = y ∗∗ (x∗ |Y ) = x∗∗ (x∗ ) = x∗ (x) = 1

which is a contradiction, and hence y := x ∈ Y . If y ∗ ∈ Y ∗ and

x∗ ∈ X ∗ satisfies x∗ |Y = y ∗ then

y ∗ (y) = x∗ (x) = x∗∗ (x∗ ) = y ∗∗ (y ∗ ).

Thus J : Y → Y ∗∗ is surjective.
2. Let 1 < p < ∞, U ⊂ Rd , un , u ∈ Lp (U ). Then

un * u ⇐⇒ sup kun kLp < ∞ and um → u as distributions.

n

To see this, assume that un * u. Then supn kun kLp < ∞ and for all
p
φ ∈ D(U ) ⊂ L p−1 (U )
ˆ ˆ
un φdx → uφdx.

Vice versa: Suppose that supn kun kLp < ∞. The same is true for every
subsequence. Then every subsequence (unj ) has a weakly converging
subsequence, unjk * ũ. Then for all φ ∈ D(U )
ˆ
(ũ − u)φdx = 0

and thus ũ = u. This is independent of the subsequence, and hence

un * u. A similar statement holds for L∞ (U ) and weak∗ convergence.

95 [February 10, 2017]

´
3. Let φ ∈ C0∞ (R) with φdx = 1. We define

un (x) = φ(x − n).

Then un (x) → 0 for all x and

´ kun kLp = ku0 kLp (R) 6= 0 for 1 ≤ p ≤ ∞.
Clearly, for all ψ ∈ D(R), un ψdx → 0 as n → ∞,

um → 0 as distributions,

hence un * 0 in Lp for 1 < p < ∞. However un does not converge

in Lp (R) since the norm limit is also the weak limit if the first exists,
but 0 cannot be the norm limit since the norm of un does not depend
on n.

4. Let un (x) = einx χ[0,1] . Then kun kLp = 1. If φ ∈ D(R) then

ˆ ˆ 1
un φ dx = einx φdx
0
ˆ
1 1 inx 0 1
=− e φ dx + (ein φ(1) − φ(0))
in 0 in
→0 as n → ∞.

Thus un → 0 as distribution and as above un * 0 as n → ∞.

5. Let 1 < p < ∞ and

1
un (x) = n p e−(nx) χ(0,∞) .

Then
ˆ ˆ ∞ ˆ ∞
|un |p dx = n e−p(nx) dx = e−px dx = ku1 kpLp .
0 0

It is not hard to see that for all φ ∈ D(R)

ˆ
− p−1
un φdx = n p (φ(0) + O(n−1 )) → 0 as n → ∞.

Thus again un * 0 as n → ∞.

6. Let
− p1 −(x/n)
un = n e χ(0,∞) .
Again kun kLp = ku1 kLp ,
− p1
kun ksup = n →0

and as above un * 0 as n → ∞.

96 [February 10, 2017]

5.5 The direct method in the calculus of variations
To motivate the study below we consider a specific problem. Let U ⊂ Rd
be open and bounded, f ∈ L2 (U ), g ∈ W 1,2 (U ). We define
ˆ
1
E(u) = |Du|2 + f udx
U 2

for u ∈ W 1,2 (U ). Let

A = {u ∈ W 1,2 (U ) : u − g ∈ W01,2 (U )}.

Suppose that v ∈ A minimizes E(u) in A. If w ∈ W01,2 (U ) and t ∈ R then

E(v) ≤E(v + tw)

ˆ
1
= |D(v + tw)|2 + f (v + tw)dx
U 2
ˆ X d ˆ
1 2
=E(v) + t ∂xj v∂xj w + f wdx + t |Dw|2 dx.
U 2 U
j=1

This is true for all t hence

ˆ X
d
∂xj v∂xj w + f wdx = 0
U j=1

for w ∈ D(U ) ⊂ W01,2 (U ). Thus

∆v = f

in the sense of distributions, and v = g at ∂U in the Sobolev sense.

We attempt to find solutions to such equations by minimzing an energy
like E.

Lemma 5.34. Let U ⊂ Rd be open and bounded, 1 < p < ∞ and either
X = W01,p (U ) or X = W 1,p (U ) under the assumption that the Whitney
extension property holds. Let (un ) ∈ X be a bounded sequence. Then there
is a subsequence (un ) and u ∈ X satisfying

1. un converges weakly to u in W 1,p (U ).

2. ∂xk un converges weakly to ∂xk u.

3. un converges to u in Lp (U ).

Proof. We consider W 1,p (U ) resp. W01,p (U ) as closed subspace of Lp (U ×

Σ). Let (un ) be a bounded sequence. Then there is a weakly convergent

97 [February 10, 2017]

subsequence by Corollary 5.32 with weak limit u. A linear space is convex
and by Theorem 5.27 we can consider u as element of X.
In both cases there is an extension to W 1,p (Rd ) functions with fixed sup-
port, see Theorem 4.42 and Corollary 4.43. This yields a bounded sequence
support in a fixed ball in W 1,p (Rd ). By Theorem 4.45 there exists C such
that
kuj (. + h) − uj kLp (Rd ) ≤ C|h|.
Now we apply the Theorem 3.39 of Kolmogorov and obtain convergence of
a subsequence in Lp (U ). A norm limit is also a weak limit, and hence the
limit is u.

[27.01.2017]
[01.02.2017]

Lemma 5.35. Let 1 ≤ p < ∞, F : Rd → R be nonnegative and suppose

that c ≥ 1 and

F (x) ≤ c(1 + |x|)p for all x ∈ Rd .

Suppose that x ∈ Rd , v ∈ Rd and

F (x + h) − F (x) ≥ hv, hi (5.5)

for all h ∈ Rd . Then

|v| ≤ 2p c(1 + |x|)p−1
Proof. There is nothing to show if |v| ≤ 1. Suppose that |v| ≥ 1 and let
h = |x|+1
|v| v. Then

|x| + 1 |x| + 1 p
|v|(|x| + 1) ≤ F (x + v) ≤ c 1 + x + v ≤ 2p c(1 + |x|)p
|v| |v|
and we obtain the desired conclusion.

Let 1 < p < ∞, U ⊂ Rd open and bounded and let

F : U × R × Rd → R ∪ {∞}

be measurable. We assume that for all x ∈ U

(u, P ) → F (x, u, P )

is continuous and for every x ∈ U and u ∈ R the map

P → F (x, u, P )

is convex. We assume that there exists C > 0 so that

0 ≤ F (x, u, P ) ≤ C(1 + |P |p ) (5.6)

98 [February 10, 2017]

Theorem 5.36 (Weak lower semicontinuity). If u ∈ W 1,p (U ) then F (x, u(x), ∇u(x))
is measurable and integrable. Moreover, if un * u in W 1,p (U ) then
ˆ
E(u) ≤ lim inf E(un ) with E(u) = F (x, u(x), ∇u(x)) dmd (x).
n U

Proof. We prove the result under the additional assumption that

P → F (x, u, P )

is differentiable for all x and that

∂F
(u, P ) → (x, u, P )
∂Pj

is continuous for all x and j.

Measurability and integrability follows as in Subsection 4.8. Let un be
as in the assumptions. By Lemma 5.34

un → u in Lp (U )

and by the proof of Theorem 3.17 a subsequence converges almost every-

where. Without relabelling we assume that un converges almost everywhere.
We write
ˆ
E(un )−E(u) = F (x, un , ∇un )−F (x, un , ∇u)+F (x, un , ∇u)−F (x, u, ∇u)dmd
U

Then
F (x, un (x), ∇u(x)) → F (x, u(x), ∇u(x))
almost everywhere. Moreover

C(1 + |∇u(x)|p )

is majorant and by the convergence theorem of Lebesgue

ˆ
F (x, un , ∇u) − F (x, u, ∇u)dmd → 0.

For every x the map

P → F (x, un (x), P )

is convex. Hence

F (x, un (x), ∇un (x)) − F (x, un (x), ∇u(x))

∂F
≥ (x, un (x), ∇u(x)) · (∇un (x) − ∇u(x)).
∂P

99 [February 10, 2017]

By Lemma
∂F
≤ c(1 + |∇u(x)|p−1 ).
∂Pj
We observe that
∂F ∂F
(x, un (x), ∇u(x)) → (x, u(x), ∇u(x))
∂Pj ∂Pj

for almost all x. Moreover

∂F ∂F
(x, un , ∇u(x)) − (x, u, ∇u(x)) ≤ c(1 + |∇u(x)|p−1 )
∂Pj ∂Pj

and hence by Lebesgue

∂F ∂F
(x, un , ∇u(x)) − (x, u, ∇u(x)) p
→ 0.
∂Pj ∂Pj L p−1

Since
∇un − ∇u * 0 in Lp
we obtain
ˆ
∂F
lim (x, un , ∇u(x))∂j (un − u)dmd (x)
n→∞ ∂Pj
ˆ
∂F
= lim (x, u, ∇u(x))∂j (un − u)dmd (x) = 0.
n→∞ ∂Pj

Essentially without changing the proof we may relax (5.6) by

− C|u| ≤ F (x, u, P ) ≤ C(1 + |P |p + |u|p ) + f (x) (5.7)

for some integrable function f .

Theorem 5.37. Suppose that in addition to the assumptions of Theorem

5.36 the function F is coercive: There exists C > 0 sich that

F (x, u, P ) ≥ C|P |p .

Let g ∈ W 1,p (U ). Then there exists u which satisfies

u − g ∈ W01,p (U )

and
E(u) = inf{E(v) : v − g ∈ W01,p (U )}.

100 [February 10, 2017]

Proof. Let
A = {u ∈ W 1,p (U ) : u − g ∈ W01,p (U )}.
Let uj be a minimizing sequence. Then
ˆ
E(uj ) ≥ C |∇uj |p dmd ,

and hence by Poincaré’s inequality

kuj kW 1,p ≤kuj kLp + k∇uj kLp

≤kgkLp + kuj − gkLp + k∇uj kLp
≤kgkLp + dk∇(uj − g)kLp + k∇uj kLp
1

≤c kgkW 1,p + E(uj ) p

and hence
sup kuj kW 1,p (U ) < ∞.
j

By Theorem 5.36 uj − g has a weakly convergent subsequence, which con-

verges in Lp (U ), and thus the same is true for uj . Thus

E(u) ≤ lim inf E(uj ) = inf{E(v) : v ∈ A}

and u ∈ A since subspaces are convex.

Theorem 5.38 (Euler-Lagrange equations). Suppose in addition that for

every x, F is continuously differentiable with respect to u and P and that
∂F
(x, u, P ) ≤ c(1 + |P |)p . (5.8)
∂u
Let u be a minimizer in A. Then u is a weak solution to
d
X ∂F ∂F
− ∂ xj (x, u, ∇u) + (x, u, ∇u) = 0.
∂Pj ∂u
j=1

Proof. We argue as for the Dirichlet integral, choose v ∈ D(U ) and define
η(t) = E(u+tv). Then η(t) ≥ η(0) = E(u). We claim that it is differentiable
with respect to t:
d
1 X ∂F ∂F
(F (x, u+tv, ∇u+t∇v)−F (x, u, ∇u)) → (x, u, ∇u)∂j v+ (x, u, ∇u)v
t ∂Pj ∂u
j=1

almost everywhere. Moreover

1
(F (x, u + tv, ∇u + t∇v) − F (x, u, ∇u + t∇v)) ≤ c(1 + |∇u| + |∇v|)p
t

101 [February 10, 2017]

by the fundamental theorem of calculus and assumption (5.8). Moreover
1
(F (x, u, ∇u + t∇v) − F (x, u, ∇u)) ≤ c(1 + |∇u| + |∇v|)p−1
t
by the fundamental theorem of calculus and Lemma 5.35. Then by the
convergence theorem of Lebesgue
ˆ
0 ∂F ∂F
0 = η (0) = (x, u, ∇u)∂j v + (x, u, ∇u)vdmd
∂Pj ∂u

Lemma 5.39. The minimizer is unique if in addition

(u, P ) → F (x, u, P )

is striclty convex for every x.

We call F strictly convex if for all 0 < λ < 1 and x, y

λF (x) + (1 − λ)F (y) > F (λx + (1 − λ)y).

Proof. We claim that

A 3 u → E(u)
is strictly convex. It suffices to verify that
u+v 1
E( ) < (E(u) + E(v))
2 2
unless u = v which is equivalent to
ˆ
u + v ∇u + ∇v
F (x, u, ∇u) + F (x, v, ∇v) − 2F (x, , )dmd > 0.
Rd 2 2
The integrant is nonnegative, and the integral is positive unless it vanishes
almost everywhere. But this is equivalent to u = v.

[01.02.2017]
[03.02.2016]

6 Linear Operators
6.1 Dual and adjoint operators
Definition 6.1. Let X and Y be Banach spaces, T ∈ L(X, Y ). We define
0
the dual operator T ∈ L(Y ∗ , X ∗ ) by
0
(T y ∗ )(x) = y ∗ (T x) for y ∗ ∈ Y ∗ , x ∈ X.

102 [February 10, 2017]

This formula defines a unique map X → K. It is clearly linear and
bounded, hence it is in X ∗ .
Examples and properties
p p
1. 1 ≤ p < ∞, X = Y = Lp (µ), g ∈ L∞ , T f = gf , T 0 : L p−1 → L p−1 ,
T 0 h = gh.

2. X = Y = lp (Z), T : X → X the shift operator, T ej = ej+1 . Then

T 0 ej = ej−1 .

3. X = Y = lp (N), T : X → X the shift operator, T ej = ej+1 . Then

T 0 ej = ej−1 if j ≥ 2 and T 0 e1 = 0.

4. J : X → X ∗∗ , J 0 : X ∗∗∗ → X ∗ is the restriction, if we identify X with

a closed subspace of X ∗∗ .

5. Consider K = R,
−∆u = f u|∂U = 0
T : L2 3 f → u ∈ L2
Then T 0 = T .

6. (λT + µS)0 = λT 0 + µS 0 , i.e. the map T → T 0 in complex linear.

7. kT 0 kY ∗ →X ∗ = kT kX→Y

Definition 6.2. Let H1 and H2 be Hilbert spaces, T ∈ L(H1 , H2 ). The

adjoint operator is defined by

hT x, yi = hx, T ∗ yi

Definition 6.3. Let T ∈ L(X, Y ). We define its range ran T as

ran T = {y ∈ Y : there exists x with T (x) = y} ⊂ Y

and
ker T = {x ∈ X : T x = 0}.
If X0 ⊂ X is a subspace then

X0⊥ := {x∗ ∈ X ∗ : x∗ (x) = 0 for x ∈ X0 }

and if X0∗ ⊂ X ∗ is a subspace then

(X0∗ )⊥ := {x ∈ X : x∗ (x) = 0 for all x∗ ∈ X0∗ }.

Lemma 6.4. Let X and Y be Banach spaces and T ∈ L(X, Y ). Then

ran T = (ker T 0 )⊥

103 [February 10, 2017]

Proof. Let T xj → y ∈ ran T and y ∗ ∈ Y ∗ with T 0 y ∗ = 0. Then

y ∗ (y) = lim y ∗ (T xj ) = lim T 0 y ∗ (xj ) = 0.

j→∞ j→∞

Thus
ran T ⊂ (ker T 0 )⊥ .
Now suppose that
y ∈ (ker T 0 )⊥
By Hahn Banach there exists y ∗ such that y ∗ |ran T = 0 and

y ∗ (y) = dist(y, ran T ).

But y ∗ ∈ ker T 0 since y ∗ |ran T = 0, hence y ∗ (y) = 0 = dist(y, ran T ).

Corollary 6.5. Suppose that T ∈ L(X, Y ) has closed range. Then

Tx = f

is sovable if and only if

T 0 y ∗ = 0 =⇒ y ∗ (f ) = 0.

Theorem 6.6. Let X and Y be Banach spaces and T ∈ L(X, Y ). The

following statements are equivalent:

1. ran T is closed,

2. ran T 0 is closed,

3. ran T = (ker T 0 )⊥ ,

4. ran T 0 = (ker T )⊥ .

Proof. Step 1: ’1 ⇐⇒ 2’
Let ỹ ∗ ∈ (ran T )∗ . By Hahn-Banach we may extend it to y ∗ ∈ Y ∗ . Since

T 0 y ∗ (x) = y ∗ (T x) = ỹ ∗ (T x) := T10 ỹ ∗ (x),

∗
the range of T 0 ∈ L(Y ∗ , X ∗ ) and T10 ∈ L(ran T , X ∗ ) are the same. Without
loss of generality we may assume that ran T = Y .
Step 1a: Now suppose that ran T is closed. We want to prove that ran T 0
is closed. By the consideration above it suffices to consider the case that T
is surjective. By the open mapping theorem Theorem 4.3 there exists c so
that for every y ∈ Y there exists x ∈ X with

kxkX ≤ ckykY and T x = y.

104 [February 10, 2017]

Thus

|y ∗ (y)| = |y ∗ (T x)| = |T 0 y ∗ (x)| ≤ kxkX kT 0 y ∗ kX ∗ ≤ ckykY kT 0 y ∗ kX ∗

and
ky ∗ kY ∗ ≤ ckT 0 y ∗ kX ∗
and hence ran T 0 is closed.
Step 1b: Now suppose that ran T 0 is closed. We want to prove that ran T
is closed. Again it suffices to consider the case that ran T is dense in Y .
Suppose there is sequence yn → 0 such that yn ∈ / T (B1 (0)). By the
∗ ∗ ∗
separation theorems there exists yn ∈ Y with kyn kY ∗ = 1 such that

Re yn∗ (yn ) > sup Re yn∗ (T x) = sup Re T 0 yn∗ (x).

kxkX ≤1 kxkX ≤1

Thus
kT 0 yn∗ kX ∗ ≤ kyn kY kyn∗ kY ∗ = kyn kY → 0.
However, by the Open Mapping Theorem 4.3, there exists c > 0 such that
if x∗ ∈ ran T 0 there exists y ∗ ∈ Y ∗ so that

ky ∗ kY ∗ ≤ ckx∗ kX ∗ , T 0 y ∗ = x∗ .

However T 0 is injective since ran T = Y by Lemma 6.4. Hence

1 = kyn∗ kY ∗ ≤ ckT 0 yn∗ kX ∗ → 0.

This is a contradiction. Thus there exists r > 0 so that

BrY (0) ⊂ T (B1 (0)).

Let y ∈ BrY (0). There exists x0 with kx0 kX < 1 and ky−T x0 kY ≤ r/2. Then
there exists x1 with kx1 kX < 21 and ky − T (x0 + x1 )kY < 2−2 r. Recursively
we obtain a sequence xj with kxj kX < 2−j and
N
X −1
ky − T xn kY < 2−N r.
n=0

Then
∞
X
y=T xn
n=0

and the range intersected with BrY (0) is closed. Hence the range is closed.
Step 2: ’1 ⇐⇒ 3’:
This is a consequence of Lemma 6.4.
Step 3: ’2 ⇐⇒ 4:
’2 ⇐= 4’ is trivial since (ker T )⊥ is closed.

105 [February 10, 2017]

So assume that ran T 0 is closed and hence ran T is closed. It is obvious
that ran T 0 ⊂ (ker T )⊥ . Let x∗ |ker T = 0. If y = T x we define

y ∗ (y) = x∗ (x)

This is well defined since x∗ vanishes on the null space of T . Since we may
assume that T is surjective we obtain from the open mapping theorem that
for y ∈ Y there exists x ∈ X with T x = y and kxkX ≤ CkykY . Thus

|y ∗ (y)| ≤ Ckx∗ kX ∗ kykY

and there is a unique continuous extension and we may consider y ∗ ∈ Y ∗ .

But then x∗ = T 0 y ∗ ∈ ran T 0 .

[03.02.2016]
[08.02.2017]

Theorem 6.7. If T ∈ L(X, Y ) is invertible iff T 0 ∈ L(Y ∗ , X ∗ ) is invertible.

Proof. Suppose that T is invertible. Thus the range is closed. By Theorem
6.6 also ran T 0 is closed. Again by Lemma 6.6 T 0 is injective since Y =
ran T = (ker T 0 )⊥ . Moreover ran T 0 = (ker T )⊥ = X ∗ . The same argument
gives the reverse implication.

6.2 Compact operators

Definition 6.8. Let X and Y be Banach spaces, T ∈ L(X, Y ). We call
T compact if for every bounded sequence (xj ), (T xj ) contains a convergent
subsequence.
Lemma 6.9. The following are equivalent
1. T is compact.

2. The image of the closed unit ball is relatively compact (the closure is
compact).

3. The image of a bounded set is relatively compact.

4. The image of a bounded set is precompact.

Proof. Suppose that T is compact. Let K = T (B1 (0)). Let xj ∈ B1X (0) and
yj = T xj . Then (xj ) is a bounded sequence and since T is compact (T xj )
contains a convergent subsequence. Thus the image of the ball of radius 1
is relatively compact.
The second statement obviously implies the third. The closure of a
relatively compact set is compact, and hence precompact. Now assume

106 [February 10, 2017]

that T B1X (0) is precompact. Then the closure is compact, and if (xj ) is a
sequence with kxj kX < 1 then (T xj ) is a sequence in a precompact set, and
hence there is a convergent subsequence and T is compact.

Lemma 6.10. If T is compact, S and U are continuous, then ST U is

compact. If (Tj ) ∈ L(X, Y ) are compact and

Tj → T in L(X, Y )

then T is compact. If T has finite rank it is compact. If an invertible map

is compact then X and Y are finite dimensional.

Proof. Let S ∈ L(Y, Z), T ∈ L(X, Y ) and U ∈ (V, X) as above and (vj ) a
bounded sequence in V . Then (U vj ) is a bounded sequence in X since U is
bounded, (T U vjl ) is a converging subsequence in Y since T is compact and
(ST U xjl ) is a convergent sequence in Z since S is continuous.
Let Tj → T a convergent sequence of compact operators and let (xn ) be
a bounded sequence. For simplicity we assume that it is bounded by 1.
We claim that for every ε > 0 there exists a subsequence xnj so that

kT xnj − T xnl kY < ε.

Suppose the claim is true. We apply it iteratively with ε = 2−j . We choose

x̃n then from the nth iteration. Then T x̃n is a Cauchy sequence.
Let ε > 0. There exists n0 so that kT − Tn kX→Y < ε/3 for n ≥ n0 .
Since Tn0 is compact there exists a subsequence so that Tn0 xnj is a Cauchy
sequence and l0 so that

kTn0 (xjl − xjm )kY < ε/3

for l, m ≥ l0 . Thus
kT (xjl − xjm )kY < ε
for l, m ≥ l0 .
Now let T be an operator of finite rank. Let (xj ) be a bounded sequence.
Then (T xj ) is a bounded sequence in a finite dimensional space and there
exists a convergent subsequence.
An invertible linear operator is compact if and only if every bounded
sequence has a convergent subsequence, which by Heine-Borel holds iff the
space has finite dimension.

Lemma 6.11. Let T ∈ L(X, X) be compact. Then ran(1X − T ) is closed.

Proof. Suppose that (xj − T xj ) is a convergent sequence where kxj kX = 1.

Then
xj = T xj + (xj − T (xj ))

107 [February 10, 2017]

The second term converges by assumption and the first has a convergent
subsequence. Without changing the notation we assume that xj → x. Then

xj − T xj → x − T x ∈ ran(1 − T ).

Adding an element of the null space ker(1 − T ) we may assume that

kxn k ≤ 2 dist(xn , ker(1 − T )).

If kxn k is unbounded, there is a subsequence so that

dist(xnj , ker(1 − T )) ≥ nj .
1
Let x̃j = kxnj k xnj . Then

xnj − T xnj
x̃j − T x̃j = →0
kxnj k

and, since kx̃j k = 1 there is another subsequence T x̃jl converges to x̃∞ ∈ X

x̃jl = x̃jl − T x̃jl + T x̃jl → x̃∞ .

Then kx̃∞ kX = 1, dist(x̃∞ , ker(1 − T )) ≥ 1

2 which is a contradiction to
x̃∞ = T x̃∞ .

Theorem 6.12 (Schauder). Let X, Y be Banach spaces, T ∈ L(X, Y ).

T compact ⇐⇒ T0 compact

Proof. Let T be compact, yj∗ ∈ Y ∗ a bounded sequence and

K = T B1 (0) ⊂ Y.

K is compact since T is compact and the functionals yj∗ define uniformly

continuous bounded functions on K. By the theorem of Arzela Ascoli 3.34
there exists a convergent subsequence yj∗l . But then

T 0 yjl ∈ X ∗

is a Cauchy sequence. If T 0 is compact then also T 00 and also T since

JY T = T 00 JX .

Lemma 6.13. Suppose that T ∈ L(X) is compact and ker(1 − T ) = {0}.

Then there exists ε > 0 so that

εkxk ≤ kx − T xk.

108 [February 10, 2017]

1
Proof. If not there exists (xn ) with kxn k = 1 and kxn − T xn k < n kxn k.
There is a subsequence and x∞ ∈ X with T xnj → x∞ . Then

xnj = (xnj − T xnj ) + T xnj → x∞

and kx∞ kX = 1 and x∞ − T x∞ = 0. This is a contradiction with the

assumption of the lemma.

Lemma 6.14. Let X be a Banach space, T ∈ L(X) such that ker T = {0}
and T is not surjective. Then ran T n+1 ⊂ ran T n but ran T n+1 6= ran T n .

Proof. We have always

ran T n+1 = T n (T X) ⊂ ran T n X.

Since ran T 6= X there exists x ∈/ ran T . If T n x = T n+1 y for some y ∈ X

then
T n (x − T y) = 0
and since T is injective x = T y, a contradiction.

Lemma 6.15. Let T ∈ L(X) be compact. If ker(1 − T ) = {0} then ran(1 −

T ) = X and 1 − T is invertible.

Proof. Let S = 1 − T . Then

n n
n n
X n j
X n
S = (1 − T ) = (−T ) = 1 + (−T )j = 1 − T̃
j j
j=0 j=1

with T̃ compact. Then ker S n = {0} and the range of S n is closed by Lemma
6.11. If S is not surjective then by Lemma 6.14 there exists a sequence yn
so that kyn k = 1, yn ∈ ran S n and
1
dist(yn , ran S n+1 ) ≥ .
2
We claim that
1
kT yn − T ym k ≥
2
if n < m. This is a contradiction to compactness and it remains to prove
the claim. It follows from

T yn − T ym = yn − ym − S(yn − ym )

and
ym + S(yn − ym ) ∈ ran S n+1 .
Thus S is injective and surjective, and invertibility follows now from the
theorem of the inverse, Corollary 4.4.

109 [February 10, 2017]

Theorem 6.16 (Riesz-Schauder). Let T ∈ L(X, X) be compact. λ 6= 0 is
an eigenvalue of T iff it is an eigenvalue of T 0 .

Proof. By Lemma 6.15 T − 1 is invertible iff 1 is not an eigenvalue. By

Theorem 6.12 T 0 is compact iff T is compact. By Theorem 6.7 T − 1 is
invertible iff T 0 − 1X ∗ is invertible. Again by Lemma 6.15 T 0 − 1 is invertible
iff 1 is not an eigenvalue of T 0 . This completes the proof since T − λ1, λ 6= 0
is invertible iff λ−1 T − 1 is invertible.

Theorem 6.17 (Fredholm alternative). Let T : X → X be compact. Then

either for f ∈ X the problem

x − Tx = f

has a unique solution or

x − Tx = 0
has a nontrivial solution.
In general
x − Tx = f
is solvable iff x∗ (f ) = 0 for all x∗ ∈ ker T 0 .

Proof. If x−T x = 0 has a nontrivial solution and if x−T x = f has a solution

then it has infinitely many solutions. If x−T x = 0 has no nontrivial solution
then 1 − T is invertible by Lemma 6.15. The last statement follows from
Lemma 6.6.

[08.02.2017]
[10.02.2017]

6.2.1 Examples of compact operators

Let 1 ≤ p ≤ ∞ and X = lp (N; C), (an ) ∈ c0 . We define T ∈ L(X) by

T ((xn )n∈N ) = (an xn )n∈N .

We claim that T is compact. Let aN N

n = an if n ≤ N and an = 0 otherwise
and TN (xn ) = (aN
n xn ). Then rk TN ≤ N , hence TN is compact and TN → T
in L(X). Thus T is compact by Lemma 6.10.

Lemma 6.18 (Compact embeddings). Let U be bounded and 1 ≤ p < ∞.

Then the embedding W01,p (U ) → Lp (U ) is compact. The same is true for the
embedding W 1,p (U ) → Lp (U ) under the Whitney extension assumption.

110 [February 10, 2017]

Proof. In either case there is an extension operator. We focus on W01,p (U ):

E : W01,p (U ) → W 1,p (Rd )

such that every function in the range is supported in a fixed ball. By the
difference characterisation of Sobolev functions
1
sup sup kEf (. + h) − Ef kLp (Rd ) ≤ ckf kW 1,p (U )
kf kW 1,p (U ) ≤1 h6=0 |h| 0

and
sup kEf kLp (Rd ) ≤ ckf kW 1,p (U ) .
0
kf kW 1,p (U ) ≤1

By Theorem 3.39 and Corollary 4.46 the unit ball in W 1,p (U ) is compact in
Lp (U ).

Lemma 6.19. Let U be open and bounded. If f ∈ L2 (U ) there exists a

unique u ∈ W01,2 (U ) which satisfies

−∆u = f

as distribution. The map f → u defines a compact operator

T : L2 (U ) → L2 (U ).

If z ∈
/ (0, ∞) the equation

− ∆u = zu + f (6.1)

has a unique solution in W01,2 (U ) for all f ∈ L2 (U ).

Proof. By the Lemma of Lax-Milgram, the operator T̃ : L2 (U ) 3 f →
u ∈ W01,2 (U ) is continuous. Hence the composition of T̃ with the compact
embedding W01,2 (U ) → L2 (U ) is compact. That is, T : L2 (U ) 3 f → u ∈
L2 (U ) is compact.
We rewrite (6.1) as
u − zT u = f˜ := T f. (6.2)
Since T is compact, (6.2) is uniquely solvable for every f ∈ L2 (U ) iff u −
zT u = 0 has only the trivial solution. This equation u = T (zu) can be
reformulated as weak solution to

−zu − ∆u = 0

which is equivalent to
ˆ X
d
0= ∂xj u∂xj v − zuv̄ dmd
j=1

111 [February 10, 2017]

which with v = u gives
ˆ ˆ
2 d
0= |Du| dm − z |u|2 dmd .

This equation has only the trivial solution u = 0 unless z ∈ (0, ∞). Thus
(6.2) is uniquely solvable and the solution is the unique weak solution to
(6.1).

Finally let X = L2 ([0, 1]) and

ˆ x
(T f )(x) = f (y)dy, x ∈ [0, 1].
0

Then
T : L2 ((0, 1)) → W 1,2 ((0, 1))
and it is compact as an operator to L2 ((0, 1)) since the embedding is com-
pact. We claim that there is no eigenvalue. If
ˆ x
zf (x) = f (y)dy
0

then, if z 6= 0
f 0 = z −1 f, f (0) = 0.
By Gronwall’s lemma
−1 x
|f (x)| ≤ e|z| |f (0)| = 0.

Hence f = 0 and z 6= 0 is not an eigenvalue. On the other hand if z = 0

then ˆ x
f (y)dy = 0.
0
This implies that any antiderivative of f is constant, and hence f = 0. We
obtain

Lemma 6.20. The operator T is compact. It has no eigenvalues.

6.2.2 Eigenvalues and spectrum

Lemma 6.21. The set of invertible operators in L(X, Y ) is open.

Proof. Suppose that T ∈ L(X, Y ) is invertible. Take S ∈ L(X, Y ) with

kSkL(X,Y ) kT −1 kL(Y,X) < 1. We define
∞
X
A= (T −1 S)n T −1
n=0

112 [February 10, 2017]

Since

k(T −1 S)n T −1 kL(Y,X) ≤ (kT −1 kL(Y,X) kSkL(X,Y ) )n kT −1 kL(Y,X)

the partial sums converge. A straightforward calculation shows that

(T − S)A = 1Y , A(T − S) = 1X .

Thus the operator T −S is invertible and hence the set of invertible operators
is open.

Theorem 6.22 (Spectrum of compact operators, Riesz-Schauder). Let T ∈

L(X) be compact and 0 6= λ. If λ is not an eigenvalue then T −λ is invertible.
Suppose that λ is an eigenvalue, then there exists k0 ∈ N such that

1. ker(T − λ)k = ker(T − λ)k0 =: N if k ≥ k0 .

2. ran(T − λ)k = ran(T − λ)k0 =: R if k ≥ k0 .

3. T : N → N and T : R → R. The second map is invertible.

4. dim N = dim ker(T 0 − λ)k0 < ∞ .

5. Every x ∈ X can uniquely be written as x = y + z with y ∈ N and

z ∈ R.

6. 0 is the only possible accumulation point of the eigenvalues. T − λ is

invertible if λ is neither an eigenvalue nor zero.

Proof. Let λ be an eigenvalue. We claim that dim ker(T − λ) < ∞. If the

eigenspace of the eigenvalue λ is infinite dimensional, then the closed ball
in the eigenspace with radius |λ| - the image of the closed ball with radius
1 under T - is not compact.
Let Rk = ran(T − λ)k . Then there exists k0 so that Rk = Rk0 if k > k0 .
We argue as in Lemma 6.15.
Thus ker(T 0 − λ)k = ker(T 0 − λ)k0 . Reversing the role of T and T 0 we
see that there exists k0 so that also ker(T − λ)k = ker(T − λ)k0 for k ≥ k0 .
Let x ∈ N , then (T − λ)k0 T x = T (T − λ)k0 x = 0 and hence T x ∈ N . If
x ∈ R then (T − λ)x ∈ (T − λ)R = R and hence T x ∈ R.
If x ∈ R = ran(T − λ)k0 then there exists y such that (T − λ)k0 y = x.
Suppose that also x ∈ N = ker(T − λ)k0 . Then

(T − λ)2k0 y = 0 = (T − λ)k0 y = x.

Thus R ∩ N = {0} and T : R → R is injective and surjective, and hence

invertible.
We apply the same reasoning to (T 0 − λ)k0 and obtain N 0 and R0 in X ∗ .
Then N 0 → N 0 |N ∈ N ∗ is injective (otherwise there would be a nontrivial

113 [February 10, 2017]

element in N 0 ∩R0 ). Similarly N → (N 0 )∗ is injective (otherwise N ∩R would
be nontrivial). Since dim N and dim N 0 < ∞ we must have dim N = dim N 0
and N 0 → N 0 |N is bijective.
Now let x ∈ X. Then there exists n ∈ N so that x∗ (x − n) = 0 for all
x ∈ N 0 . But then x − n ∈ R.
∗

We denote by TR : R → R and TN : N → N . Then TR −µ1R is invertible

for |µ − λ| small. Moreover TN − µ1N is invertible for µ 6= λ since TN − λ
is nilpotent. Thus T − µ is invertible for |µ − λ| small µ 6= λ. Also T − µ is
invertible unless µ is an eigenvalue.

This we can apply to problem (6.1) and conclude that there is a monotone
sequence of eigenvalues, with ∞ as only possible limit, and (6.1) is uniquely
solvable unless z is an eigenvalue.

114 [February 10, 2017]