Introduction to Floquet theory

Konrad Viebahn∗
Institute for Quantum Electronics, ETH Zurich, 8093 Zurich, Switzerland
(Dated: November 13, 2020)

These are introductory lecture notes on Floquet theory periodic potential

for applications in quantum optics and ultracold quan-
tum gases, with a particular focus on driven optical lat- V (x) = V (x + a) , (3)
tices.
where a is the lattice spacing. Bloch’s theorem tells us
Floquet theory aims at finding solutions to the time-
that the solutions to Eq. 3 are of the form
dependent Schrödinger equation with a time-periodic
Hamiltonian: ψ n (x, q) = un (x, q) × eiqx (4)
d with un (x, q) = un (x + a, q) , (5)
i~ |ψ(τ )i = H(τ ) |ψ(τ )i (1)
dτ where ψ n (x, q) is sometimes called Bloch wave and
with H(τ ) = H(τ + T ) . (2) un (x, q) is called Bloch function. The argument q is called
quasimomentum or lattice momentum and n labels the
eigenstate (also called band index ). At first sight, we
Analogies between Bloch theory and Floquet theory have not gained anything but simply shifted the problem
of finding ψ n (x, q) to another unknown function un (x, q).
Before we tackle the time-dependent problem, let us However, the crucial advantage of the Bloch function is
review briefly the theory describing spatially periodic, its periodicity (Eq. 5) which allows us to Fourier-expand
but temporally static systems: Bloch theory. There are it into position-independent coefficients cnl (q):
some similarities between Floquet and Bloch theory that
+∞
can help us gain an intuition on the former. For ex- n
X
ample, the real momentum p loses its meaning in Bloch u (x, q) = cnl (q)ei2klat lx (6)
l=−∞
theory and must be replaced by ~q, the quasimomentum
+∞
q mod 2klat = 2π/a (a is the lattice spacing), while in X
Floquet theory the energy becomes undefined and must ⇒ ψ n (x, q) = cnl (q)ei(2klat l+q)x . (7)
be replaced by quasienergy mod ~ω, where ω = 2π/T is l=−∞

the driving frequency. Both theories reduce an originally We can now re-write the Hamiltonian resulting from
intractable problem to a relevant low-energy subspace. Eq. 3 in the basis of plane waves (with the coefficients
However, the analogies between Bloch theory and Flo- cnl (q)) which form the natural basis for spatially periodic
quet theory only hold up to a certain point and should problems.
not be overstretched. One should keep in mind that the
two descriptions are trying to solve two different prob-
lems. In Bloch theory, on the one hand, the underlying Optical lattices
challenge is an infinite-dimensional Hamiltonian which
can then be reduced to an approximate finite-dimensional Let us directly look at an optical lattice as an example.
one. In Floquet theory, on the other hand, the underly- The potential is given by
ing challenge is the evaluation of a time-ordered integral
(the time-evolution operator) which then turns out to be V (x) = V0 cos2 (klat x) . (8)
separable into slow-moving and fast-moving parts. This
is independent of the dimension of the Hamiltonian: as You showed in a previous exercise that, using Eq. 7, the
we will see, it applies both to a two-level system and an Hamiltonian can be written as
extended optical lattice.  2
~
 2m (q + 2lklat )
 2
for l = l0
Hl,l0 = V0 /4 for |l − l0 | = 1 (9)

0
I. BLOCH’S THEOREM otherwise,
neglecting any static energy offsets. While this Hamilto-
Here, we are trying to find solutions to the stationary nian is still infinite-dimensional, and hence intractable,
Schrödinger equation for a Hamiltonian with a spatially the natural basis of plane waves allows us to truncate
it at finite ±lmax , still capturing the proper low-energy
behaviour. The matrix in Eq. 9 is sparse and can eas-
ily be diagonalised numerically to yield the well-known
∗ [email protected] bandstructure (Fig. 1).
2

t.iq
%:

it i÷÷
:

€
%!
→

I
⇐
÷
÷
.

FIG. 1. Band structure of an optical lattice. By drawing free-particle dispersions separated by 2~klat and introducing a coupling
between them, the characteristic band structure emerges. This description can then be reduced to a single Brillouin-zone with
q ∈ [− πa , πa ].

The photon picture the time-evolution can be completely absorbed into the
rotating frame of reference. As we will see in an example,
An intuitive picture for the emergence of band gaps in these situations can be understood as ‘trivial’ cases of
the free-particle dispersion is the absorption and subse- Floquet theory, in which we do not rely on Floquet’s
quent re-emission of a photon from the optical lattice. In theorem. The power of Floquet theory comes from the
a process in which the absorption and emission processes fact that, even if there remains a time-dependence in the
happen in opposite directions, the atom gets a momen- rotating frame, we can still make the problem tractable
tum kick of 2~klat , thereby coupling these states in the using the periodicity of the Hamiltonian. In particular,
free-particle dispersion, leading to a gap opening. This we want to separate fast dynamics, within one period T
picture is illustrated in Fig. 1 using ‘extended zones’. of the Hamiltonian, from slow ones, which change from
one period to the next.
Our derivation starts with noting the following (‘semi-
II. FLOQUET’S THEOREM group’) identity of the time-evolution operator (Eq. 10)
that evolves a given state from an initial time τ0 to τ1 +τ2 :
Now, let us get to the time-periodic problem, i.e. U(τ1 + τ2 , τ0 ) = U(τ1 + τ2 , τ1 )U(τ1 , τ0 ) . (12)
i τ1 Now, for the periodically modulated Hamiltonian in
Z
U(τ1 , τ0 ) = T exp − H(τ ) dτ (10) Eq. 10, we have for the evolution from an initial time
~ τ0
τ0 to some final time τ + T
with H(τ ) = H(τ + T ) . (11)
U(τ + T, τ0 ) = U(τ + T, τ0 + T )U(τ0 + T, τ0 ) (13)
(T denotes time-ordering.) Here, a simple basis change à Ex
la Bloch or, equivalently, a change of reference frame will = U(τ, τ0 )U(τ0 + T, τ0 ) . (14)
only work in specific situations, namely, if the stationary Now, knowing that any U(τ2 , τ1 ) must be unitary
part of H(τ ) is rotationally invariant [1]. In these cases, (i.e. U † U = 1), we are able to write the second part
3

i →
e⇐ k→ .

i.
FIG. 2. Floquet’s theorem illustrated. The slow dynamics (from one period T to the next) is captured by Floquet Hamiltonian
HF , whereas the fast ‘micromotion’ within a period is governed by the fast motion operator e−iK̂F (τ )
.

of our time-evolution as This result is the mathematical way of separating the

−iHF ×T /~ slow dynamics (governed by HF ) from the fast ‘micro-
U(τ0 + T, τ0 ) ≡ e , (15)
motion’ during one period T (governed by e−iK̂F (τ ) ), as
which defines HF as a time-independent, hermitian op- illustrated in Fig. 2.
erator (i.e. (HF )† = HF , ensuring that U(τ0 + T, τ0 ) is
unitary). The hermitian operator HF [τ0 ] does, in gen-
eral, depend on the choice of initial time τ0 . With the
help of Eq. 15 we define the ‘fast-motion’ operator
e−iK̂F (τ ) ≡ U(τ, τ0 )e+iHF ×(τ −τ0 )/~ , (16)
where K̂F (τ ) is called the ‘stroboscopic kick operator’
(for reasons that will hopefully become clear later). The
fast-motion operator e−iK̂F (τ ) obeys the following iden-
tity
e−iK̂F (τ +T ) = U(τ + T, τ0 )e+iHF ×(τ +T −τ0 )/~
(14)

= U(τ, τ0 ) U(τ0 + T, τ0 )e+iHF ×T /~
×e+iHF ×(τ −τ0 )/~ (17)
Time
'
'

(15,16)
= e −iK̂F (τ )
(18) I. t.tt t c-

This means that the kick operator describes the motion

within one period T but does not change from one pe- FIG. 3. The action of a periodic drive (upper panel) on the
riod to the next (it just depends on the global choice of time-evolution of a quantum system. The slow dynamics of an
starting time τ0 ). This allows us to directly write down observable (red line) is captured by the Floquet Hamiltonian
Floquet’s theorem for periodically modulated Hamiltoni- HF whereas the exact dynamics (blue line) is composed of fast
ans: dynamics (micromotion) and slow dynamics. The micromo-
tion is described by the kick operators e−iK̂F (τ ) and vanishes
U(τ, τ0 ) = e−iK̂F (τ ) e−iHF ×(τ −τ0 )/~ (19) at stroboscopic times τ0 , τ0 + T , etc (blue points).
−iK̂F (τ +T ) −iK̂F (τ )
with e =e .
4

In particular, we can now calculate the slow dynamics (This is shown in ref. [1].) We can think of Eq. 21 as a
of a given quantum state over many periods T by simply gauge transformation and τ0 as a gauge choice, called the
evolving it with the static, hermitian operator HF , as if it ‘Floquet gauge’. Hence, knowing one particular member
was a time-independent problem. For this reason, HF is of the family of Floquet Hamiltonians {HF [τ0 ]} allows us
called a ‘Floquet Hamiltonian’. Keep in mind, however, to reconstruct all other members. Correspondingly, the
that for reaching arbitrary final times τ (and also for ar- spectrum of HF [τ0 ] will not depend on τ0 .
bitrary starting times τi 6= τ0 ) we need to additionally Therefore, we should write the definition of the Floquet
apply the stroboscopic kick operator K̂F (τ ). At ‘strobo- Hamiltonian and the stroboscopic kick operator with ex-
scopic’ times τ0 , τ0 + T , τ0 + 2T , etc. the kick operator plicit τ0 dependence as
is identically zero, which means U(τ0 + T, τ0 ) = e−iHF [τ0 ]×T /~ (22)
e−iK̂F (τ0 ) = e−iK̂F (τ0 +T ) (20) e−iK̂F [τ0 ](τ ) = U(τ, τ0 )e+iHF [τ0 ]×(τ −τ0 )/~ , (23)
−iK̂F (τ0 +2T ) leading to Floquet’s theorem in its gauge-dependent form
=e
= ... U(τ, τ0 ) = e−iK̂F [τ0 ](τ ) e−iHF [τ0 ]×(τ −τ0 )/~ (24)
=1. with K̂F [τ0 ](τ + T ) = K̂F [τ0 ](τ ) .
Consequently, the dynamics calculated by the Floquet The insight of Eq. 21 motivates us to look for a prudent
Hamiltonian is exact at stroboscopic times, as illustrated choice of τ0 that eliminates any explicit dependence of HF
in Fig. 3. on τ0 . This is possible, following ref. [2], by absorbing all
gauge dependence into new kick operators:

The Floquet gauge τ0 Heff = eiK̂(τ0 ) HF [τ0 ]e−iK̂(τ0 ) (25)

which are related to the old ones via
So far, we have tacitly assumed a certain starting time
τ0 in all derivations. However, it is clear that the result- e−iK̂F [τ0 ](τ ) = e−iK̂(τ ) eiK̂(τ0 ) . (26)
ing dynamics will not be independent of the value of τ0 , Now, we can write down the Floquet theorem in its
as illustrated in Fig. 4. Indeed, both HF [τ0 ] and K̂F [τ0 ] gauge-independent form
carry an explicit dependence on τ0 which will be denoted
by the square brackets [·] throughout the text. What is U(τf , τi ) = e−iK̂(τf ) e−iHeff ×(τf −τi )/~ eiK̂(τi ) . (27)
the relevance of τ0 ?
(The second kick operator comes about because we do
not want to fix the time-evolution to a particular starting
time τ0 .) Eq. 27 gives the full time-evolution of an arbi-
trary state and can be interpreted as follows. First, we
transform the starting state into a new frame of reference
in which the Hamiltonian becomes time-independent us-

i÷÷÷
ing eiK̂(τi ) . Then, the time-evolution under the static
Hamiltonian Heff takes place. Finally, the state is trans-
formed again using e−iK̂(τf ) which depends on the final
time τf within one period T (‘micromotion’). The Hamil-
tonian Heff is called ‘effective Hamiltonian’ and the op-
erators K̂(τ ) are the ‘non-stroboscopic kick operators’
which, unlike K̂F [τ0 ](τ ) do not vanish at stroboscopic
times τ0 , 2τ0 , etc.
Floquet’s theorem can also be written as

∂ −iK̂(τ )
FIG. 4. The Floquet gauge τ0 . The Floquet Hamiltonian Heff = eiK̂(τ ) Ĥ(τ )e−iK̂(τ ) − i~eiK̂(τ ) e ,
HF [τ0 ] depends on the choice of starting time. However, all ∂τ
Floquet Hamiltonians can be transformed into one another (28)
by a unitary transform (Eq. 21) and have the same spectrum. which is equivalent to Eq. 27.
The micromotion will depend on the choice of starting time.

Floquet states and quasienergies

We note that all choices of τ0 can be related via a
unitary transformation: We want to make use of Floquet’s theorem and calcu-
late the time-evolution of an arbitrary state in the pres-
HF [τ00 ] = U(τ00 , τ0 )HF [τ0 ]U(τ0 , τ00 ) . (21) ence of an external, periodic drive (H(τ ) = H(τ + T )).
5

Let us for the moment assume that we know the oper- the associated kick operator K̂F (τ ). Having diagonalised
ator U(τ0 + T, τ0 ) = e−iHF ×T /~ which evolves a state HF (or, more precisely, U(τ0 + T, τ0 )), we can expand the
from one driving period to the next. Since the opera- initial state |ψ(τ0 )i in eigenstates |ni with coefficients
tor U(τ0 + T, τ0 ) is unitary its eigenvalues are complex an and work out the Floquet modes |un (τ )i. Then, the
numbers that lie on the unit circle, allowing us to write time-evolution is directly given by Eq. 33. In particular,
its eigenvalues as {e−in T /~ }, which are called ‘Floquet we have found the basis in which all time-dependence has
multipliers’. The notation e−in T /~ reminds us that only been shifted into the basis states. The coefficients an do
the Floquet multipliers are uniquely defined, whereas the not depend on time, despite the periodic modulation in
values {n } are multi-valued, defined mod ~ω with the Hamiltonian H(τ ). Thus, knowledge of the Floquet
Hamiltonian HF allows us to treat the problem as if it
2π was time-independent. This description is exact at stro-
ω= . (29)
T boscopic times τ0 , τ0 +T , etc. since |un [τ0 ](τ0 )i = |n[τ0 ]i,
whereas at non-stroboscopic times we need to take spe-
Hence, the values {n mod ~ω} are called ‘quasienergies’
cial care of the Floquet modes (Eq. 34).
(in analogy to the quasimomentum). Taking the multi-
valuedness of the energies n into account, we can also
directly get the eigenstates and eigenvalues via Example I: two-level system with
circularly-polarised drive
HF |ni = n |ni . (30)

We should keep in mind that he eigenstates {|ni} of The archetypical spin in a circularly-driven field is nice
U(τ0 + T, τ0 ) in reality carry the gauge dependence τ0 example in which the Floquet Hamiltonian can be simply
and we should write |n[τ0 ]i. However, we will now often written down. In this easy case, the kick operator does
omit the square brackets for clarity. nothing else than going into the rotating frame.
Now, we expand an arbitrary initial state |ψ(τ0 )i in We start with the Hamiltonian of a circularly driven
the eigenbasis of U(τ0 + T, τ0 ) spin-1/2:
X
|ψ(τ0 )i = |ni hn|ψ(τ0 )i (31) ~ω0 µB0
Hc-1/2 (τ ) = σz + (σx cos ωτ + σy sin ωτ ) , (37)
n
| {z }
≡an
2 2
X
= an |ni . (32) where σi are the Pauli matrices, ω0 is the bare transition
n frequency, µ is the magnetic moment, B0 is the magnetic
field strength, and ω is the driving frequency. We trans-
Then, the state at later times will be form the spin-1/2 Hamiltonian into the rotating frame
|ψ(τ )i = U(τ, τ0 ) |ψ(τ0 )i using the kick operator
(24) X
= an e−iK̂F (τ ) e−iHF ×(τ −τ0 )/~ |ni ω(τ − τ0 )
K̂F [τ0 ](τ ) = − × (1 − σz ) . (38)
n 2
X
= an e−iK̂F (τ ) e−in (τ −τ0 )/~ |ni Then, we have
n
c
HF [τ0 ] eiK̂F [τ0 ](τ ) Hc-1/2 (τ )e−iK̂F [τ0 ](τ ) (39)
=
X
= an e−in (τ −τ0 )/~ |un (τ )i , (33)
n ∂ −iK̂F [τ0 ](τ )
− i~eiK̂F [τ0 ](τ ) e
where in Eq. 33 we have defined the ‘Floquet modes’ ∂τ
~ω ~
|un (τ )i ≡ e−iK̂F (τ ) |ni (34) = 1 + (ω0 − ω)σz
2 2
(18) µB0
with |un (τ )i = |un (τ + T )i . (35) + (σx cos ωτ0 + σy sin ωτ0 )
2
In the literature, the states c (τ0 ≡0) ~ω ~ µB0
⇒ HF [0] = 1 + (ω0 − ω)σz + σx
2 2 2
|ψn (τ )i = e−in (τ −τ0 )/~ |un (τ )i (36) (40)
are often called ‘Floquet states’ in analogy with the Bloch where we have chosen the Floquet gauge τ0 = 0 in the
theorem (Eq. 4). Again, the Floquet states |ψn [τ0 ](τ )i last line. The quasienergies are
and |un [τ0 ](τ )i carry gauge dependence.
Equation 33 is the reward of all our previous deriva-
q
~
tion, giving the time-evolution of an arbitrary state ± = ω ± ∆2 + Ω20 mod ~ω (41)
2
|ψ(τ0 )i as a function of time-independent coefficients an .
In order words, in order to solve the time-dependent with ∆ = ω0 − ω and ~Ω0 = µB0 . The Floquet states
problem we need to find the Floquet Hamiltonian HF and can be evaluated according to Eq. 36.
6

Example II: two-level system with linearly-polarised Then we find

drive
1 T
Z
(0)
Heff = H(τ )dτ ≡ H0
T 0
Only very few Floquet Hamiltonians can be directly ∞
computed as in the circularly driven spin-1/2 above. Our (1) 1 X1
Heff = [Hl , H−l ]
next example is the linearly-driven spin-1/2, ~ω l
l=1
"∞
(2) 1 X 1
~ω0 Heff = ([[Hl , H0 ], H−l ] + [[H−l , H0 ], Hl ])
Hlin-1/2 (τ ) = σz + µB0 σx cos ωτ , (42) (~ω)2 2l2
2 l=1

∞
which can be thought of as superposition of a left-
X 1
+ ([Hl , [Hk , H−l−k ]] − [Hl , [H−k , Hk−l ]] + h.c.)
and right-hand circularly polarised drive. It will retain 3lk
l,k=1
some time-dependence after applying the transformation
(Eq. 38) into the rotating frame since one of the two cir- and
cular drives is ‘rotated away’ but the ‘counter-rotating’ K̂ (0) = 0 (46)
one will oscillate twice the driving frequency. This time, ∞
1 X 1
we directly choose the Floquet gauge τ0 = 0: K̂ (1) = Hl eilωτ − H−l e−ilωτ

(47)
i~ω l
l=1
lin
Hrot [0](τ ) = eiK̂F [0](τ ) Hlin-1/2 (τ )e−iK̂F [0](τ )

∂ −iK̂F [0](τ ) Example III: Bloch-Siegert shift
− i~eiK̂F [0](τ ) e
∂τ
Ex c µB 0 Now, using the high-frequency expansion we can tackle
= HF [0] + (σx cos 2ωτ − σy sin 2ωτ ) .(43) the linearly driven two-level system where we left off in
2
example II. Our starting point is the Hamiltonian in the
It is evident from Eq. 43 that we require further as- rotating frame (Eq. 43),
sumptions to make progress in this situation. The ~∆ ~Ω0 ~Ω0
lin
answer will be to assume very high driving frequency Hrot (τ ) = σz + σx + [σx cos 2ωτ − σy sin 2ωτ ]
2 2 2
ω µB0 /2 in order to neglect the 2ω terms in Eq. 43, (48)
c
such that the Hamiltonian reduces to HF [0]. This is where we have chosen again τ0 = 0, neglecting global
the central idea behind the so-called ‘high-frequency energy offsets. Applying the high-frequency expansion,
expansion’[3] of which you already know a prominent ex- we find
ample: the famous rotating-wave approximation (RWA) ~∆ ~Ω0
(0)
in the linearly-driven two-level system. Heff = H0 = σz + σx (RWA) (49)
2 2
2
(1) Ex ~ω Ω0
Heff = σz (Bloch-Siegert) (50)
4 ω
III. HIGH-FREQUENCY EXPANSION (1)
where Heff is known as the Bloch-Siegert shift. We
see that the first-order correction to the time-averaged
In almost all practical cases, it is impossible to find Hamiltonian H0 becomes relevant for strong driving,
HF or Heff in a closed form. Instead, one can employ a i.e. Ω0 6 ω.
high-frequency expansion (HFE)[4] The high-frequency expansion is a good method to an-
alytically derive effective Hamiltonians. In some cases,
∞ ∞
X (n)
X the correct choice of reference frame will make higher-
Heff = Heff , K̂(τ ) = K̂ (n) (τ ) , (44) order terms in the series expansion vanish and the re-
n=0 n=0 sulting effective Hamiltonian becomes exact (exercise).
If, however, the high-frequency expansion cannot eas-
(n) (n)
in inverse powers of ω, i.e. Heff ∼ ω −n and K̂eff (τ ) ∼ ily be truncated, or the underlying Hamiltonian is too
ω −n . Thus, the HFE is particularly useful if the driving complicated, there are three other options to obtain the
frequency ω is much higher than all other energy scales, effective Hamiltonian numerically.
such that the series can be truncated.
We can expand the time-dependent H(τ ) in Fourier com-
ponents IV. NUMERICALLY SOLVING THE FLOQUET
PROBLEM
+∞
X
H(τ ) = Hl eilωτ . (45) In the following, we discuss three recipes to numeri-
l=−∞ cally calculate the effective Hamiltonian and/or the time-
7

evolution operator U(τ0 + T, τ0 ). T ] into N segments τi = τ0 + i∆τ with ∆τ = T /N :

" #
i τ0 +T
Z
U(τ0 + T, τ0 ) = T exp − H(τ ) dτ (53)
~ τ0
Exact time-evolution of U(τ0 + T, τ0 )
N −1
" #
i X
' exp − H(τi )∆τ (54)
If the Hilbert space of H(τ ) is not too big, one can ~ i=0
obtain U(τ0 + T, τ0 ) by integrating the time-dependent
N −1
Schrödinger equation over one period. First, expand the Y i
exp − H(τi )∆τ + O ∆τ 2

=
initial state in a certain basis ~
i=0
X
|ψ(τ0 )i = cn (τ0 ) |ni . (51) Here, T denotes time-ordering and the third line is the
n Trotter decomposition, i.e. going from the exponential
of a sum to the product of exponentials. Generally, this
Then, evolve ψ(τ0 ) over one period according to the time- can be a hard problem, because Hamiltonians at different
dependent Schrödinger equation times do not commute, [H(τi ), H(τj )] 6= 0. Trotter tells
us that the error we make by ignoring the commutators
d X is only O(∆τ 2 ), i.e. it goes down quadratically with the
i~ cn (τ ) = Hnm (τ )cm (τ ) (52) duration of one time-step. During each time-step the
dτ m Hamiltonian H(τi ) is taken to be constant, thus reducing
the time-ordered integral to a product of exponentiated
for each coefficient cn (τ0 ) from τ0 to τ0 + T . The time- matrices, as illustrated in Fig. 5. As before, diagonalising
evolved vectors {cn (τ0 + T )} directly give the columns U(τ0 + T, τ0 ) gives the Floquet multipliers e−in T /~ .
of U(τ0 + T, τ0 ) which, upon diagonalisation, give the
Floquet multipliers {e−in T /~ }.
The extended Hilbert space

HE ;) Finally, we will discuss a method of obtaining the

-

quasienergy spectrum via extending the Hilbert space

i÷
by multiples of ‘photon numbers’ with ~ω. More details
about this method can be found in refs. [5, 6] (among

*
others).
We start with the Floquet states from Eq. 36 which
we expand in its harmonics due to the periodicity of the
Floquet modes |un (τ )i = |un (τ + T )i (c.f. Eq. 7 for Bloch
waves).

|ψn (τ )i = e−in τ /~ |un (τ )i (55)

+∞
(35) −in τ /~ X
= e e−imωτ |n, mi , (56)
to ti
It T m=−∞

where |n, mi are the Fourier-coefficients of |un (τ )i and we

FIG. 5. Trotter decomposition of the time-evolution operator have again dropped the explicit dependence on the start-
U(τ0 + T, τ0 ). The time-axis is discretised in units of ∆τ , ing time τ0 . Plugging this ansatz into the time-dependent
taking the Hamiltonian H(τi ) to be constant at each instant. Schrödinger equation (Eq. 1) yields
+∞
Ex
X
(n + m~ω) |n, mi = Hm−m0 |n, m0 i . (57)
m0 =−∞

As before, the Hl are the Fourier-components of the time-

Trotter decomposition of U(τ0 + T, τ0 )
dependent Hamiltonian H(τ ) (Eq. 45). Here, we have
created an over-complete problem, as the n are only de-
If the Hilbert space is too large a numerical integration fined mod ~ω. The multi-valuedness of n is the reason
of Eq. 52 can become unfeasible. In this case, one can why we obtain an extended Hilbert space and the solu-
make use of the so-called Trotter decomposition of the tions to Eq. 57 will be ~ω-periodic. We can write Eq. 57
time-evolution operator by dividing the timespan [τ0 ,τ0 + explicitly as
8
    
.. .. ..
 . H −1 H −2  .   . 
H1 H0 − m~ω H−1 H−2    |n, mi  = n  |n, mi 
   

H2 (58)
H1 H0 − (m + 1)~ω H−1  |n, m + 1i
  
|n, m + 1i
 
.. ..

..
H2 H1 . . .

Note that the big matrix in Eq. 58 consists of many

blocks à d × d entries, where d is the dimension of the ii
Hilbert space of H(τ ). Likewise, the vectors |n, mi are "

IET
d-dimensional.
For a common time-dependent Hamiltonian of the kind
H(τ ) = H0 + V eiωτ + V † e−iωτ (59)
G-
the extended-space Hamiltonian takes the simple form has frame wooing frame
 
..
 † . V 0  FIG. 6. The moving lattice in the lab frame (left) can be
V H0 + ~ω V 0  transformed into an oscillating force (right) in the co-moving
 †

0 V H 0 V 0  , (60) frame.
†
 

 0 V H 0 − ~ω V 

† ..
0 V .
move the retro-reflecting mirror that defines the stand-
similar to a tight-binding Hamiltonian with nearest- ing wave. In this way, the position of the optical lattice
neighbour hopping. Since the periodic drive often has Vlat (x) can be modulated in time with a waveform xm (τ ).
the cosine-shape of Eq. 59, the numerical evaluation of Another way to generate the shaken lattice is to modu-
this block-diagonal matrix can be very efficient. But how late the frequency of one of the two beams that form the
can we truncate this matrix which is in general infinite- standing wave, resulting in time-periodic phase-shift on
dimensional? Here, we have to distinguish two regimes. the light.
In the weak driving regime, i.e. ~ω hV i, which is A generic hamiltonian describing these situations is
equivalent to the high-frequency regime, only one block
of this Hamiltonian is relevant, e.g. p̂2
Hlab (τ ) = + Vlat [x̂ − xm (τ )] . (62)
2m
H0 + ~ω V
, (61)
V† H0 We consider the unitary transformation R0 (τ ) =
e−ip̂xm (τ )/~ and apply R†0 (τ ) to Hlab according to
in which we recognise the rotating-wave approximation
(Eq. 49 for the spin-1/2). If instead the driving frequency

∂
is on the same order as the time-dependent Hamiltonian H̃(τ ) = R(τ )H(τ )R† (τ ) + i~ R(τ ) R† (τ ) . (63)
∂τ
~ω ∼ hV i, we reach the ‘strong-driving’ limit in which
many blocks have to be taken into account. with R(τ ) = R†0 (τ ), using
1
eA Be−A = B + [A, B] + [A, [A, B]] + . . . (64)
V. FLOQUET ENGINEERING WITH OPTICAL 2!
LATTICES
to get
Now we will bring together the two concepts, Bloch [p̂ − A(τ )]2
theory and Floquet theory, giving spatio-temporal Hrot (τ ) = + Vlat (x̂) (65)
2m
‘Floquet-Bloch waves’.
(neglecting a non-operator valued term involving A(τ )2 ,
which can be transformed away with another transfor-
Lattice shaking: reference frames and energy scales mation). This frame of reference is called the rotating
frame. We read off the ‘vector potential’ A(τ ) = mẋm (τ ).
Before we start tackling the Hamiltonian, we want to Now, we apply a second unitary transformation R1 (τ ) =
choose a convenient frame of reference in which to de- e−ix̂A(τ )/~ to reach the reference frame that is co-moving
scribe the problem. with the shaken lattice:
A common experimental method to implement lat- p̂2
tice shaking schemes employs a piezo-electric actuator to Hcm (τ ) = + Vlat (x̂) − F (τ )x̂ (66)
2m
9

i÷i÷÷?÷iii÷÷
Static Dion 4. coupling) Driven

FIG. 7. Resonant coupling of two Bloch bands. Left: static system. The dispersion has two bands which are not coupled in
the absence of any driving. Middle: driven system with infinitesimal driving amplitude. The dispersion relation is extended by
multiples of ~ω. Right: driven system with non-zero driving amplitude. The dispersion is folded back into a single Floquet-zone,
with interband couplings appearing where two bands used to cross. When the driving frequency ω is on the same order as the
band separation the two bands (say, ‘valence band’ and ‘conduction band’) become hybridised.

with the time-periodic force F (τ ) = −mẍm = −Ȧ(τ ), reads

as shown in Fig. 6. The co-moving frame (Eq. 66) is
V (x̂, τ ) = V0 cos2 [klat (x̂ − xm (τ ))] , (69)
particularly useful when describing the optical lattice as
tight-binding model (Exercise). leading to the time-dependent Hamiltonian (in the rotat-
ing frame)
 2
Driven two-band system  2m (q + 2lklat − A(τ )/~)
 ~ 2
for l = l0
Hl,l0 (τ ) = V0 /4 for |l − l0 | = 1

0
As a first example, we will only consider two Bloch otherwise,
bands (e.g. a two-site tight-binding model) in which we as in Eq. 9 in the beginning. It is convenient to introduce
introduce a resonant driving at the frequency of the band the dimensionless shaking amplitude
separation (Fig. 7). The resulting Floquet-Bloch disper-
sion can be calculated using any of the formalisms intro- mx0 ωd
Ex
K0 = . (70)
duce above (section IV). ~
The resulting Floquet-Bloch bandstructure is shown in
Fig. 8, taken from ref. [7] for three different values of
Shaken optical lattice K0 . They can be calculated numerically using the Trotter
decomposition (or the exact time-evolution method).
Using the shaking waveform

xm (τ ) = x0 cos(ωτ ) (67) Dynamical localisation

A(τ ) = mẋm (τ ) = −mx0 ω sin(ωτ ) (68)
Several experiments have realised the flattening of
in the lab frame (67), or in the rotating frame (68), we can the lowest Floquet-Bloch band, giving rise to the phe-
study the full optical lattice problem, including higher nomenon of ‘dynamical localisation’ [8]. As you will cal-
bands. The moving-lattice potential in the lab frame culate in the exercise, the flattening occurs around the
10

zero of the zeroth Bessel function J0 (K0 ), which is also

reflected in the full calculation in Fig. 8. The rescaling of
the tunnelling with the zeroth Bessel function is the fun-
damental manifestation of off-resonant lattice shaking.

i¥¥¥ H¥¥E ÷.

FIG. 9. The action of the real magnetic field on a tight-

binding lattice.

Artificial magnetic fields

Arguably the most important application of Floquet

theory in optical lattices is the generation of artificial
magnetic fields. The two ingredients required to achieve
an artificial magnetic field is the breaking of time-reversal
symmetry, on the one hand, and a non-trivial distribution
of tunnellings in the lattice, on the other. There is an
excellent review on this topic by Cooper et al. [9].

It -_ tw

at
<

FIG. 8. Floquet-Bloch bands for V0 = 4Erec , ~ω = 0.5Erec ,

Lattice stealing
and K0 = [0.20, 0.74, 1.21] × π (top to bottom). The driving
frequency is chosen higher than the bandwidth of the lowest FIG. 10. Induced Peierls phases ϕR in a resonantly-shaken
band, but lower than the first band gap. Consequently, the optical lattice.
lowest band stays mostly intact, but the higher bands become
hybridised for low driving strengths (K0 = 0.2π). The shape
of the lowest band is changed significantly by the off-resonant
driving, it becomes flat (middle) and inverted (bottom), albeit Let us first consider the action of a uniform magnetic
with many couplings to higher bands. These crossings can field on a lattice, as depicted in Fig. 9. In the tight-
be avoided by increasing the separation of the bands (deeper binding model it shows up as complex tunnelling ma-
lattices) [7]. In their definition, kL = π/a. trix elements that add up to a constant value ϕ, called
magnetic flux, when hopping around a unit cell. The
11

a-

# 5£
#H :
←

\
+
&
a

:÷ :÷ ÷ :÷ ÷
:

D
to
+

FIG. 11. Generating an artificial magnetic field in an optical lattice. Method a) employs a non-uniform distribution of Peierls
phase factors in order to achieve a net uniform flux. A secondary, moving optical lattice leads to a periodic modulation of
the on-site energies of the lattice. The frequency of this modulation is chosen to be resonant with a static tilt in both x-
and y-directions. Method b) uses circular lattice shaking in order to induce next-nearest-neighbour tunnellings. The shaking
frequency can be chosen resonant with a large AB-sublattice offset or it can be off-resonant. While hopping around the entire
unit cell does not lead to a magnetic flux, individual paths within the unit cell do. This pattern of staggered fluxes was first
proposed by Haldane [10] in 1988 and it was later realised in the laboratory using ultracold atoms.

argument ϕi of each complex tunnelling is called Peierls

phase,
ti = |ti |eiϕi (71)
4
X
ϕi = ϕ . (72)
i=1

As you will show in the exercise, non-zero Peierls

phases can be engineered via lattice shaking. For ex-
ample, resonant shaking with a static tilt ∆ in the lat-
tice gives rise to a complex tunnelling element, as shown
schematically in Fig. 10.
However, it is clear from Fig. 10 that simply adding a
tilt, even in a two-dimensional lattice, will not result in
a net flux ϕ, as hopping in the opposite direction cancels
the phase of hopping in the forward direction. Therefore,
a more complicated Peierls phase structure must be engi-
neered in order to achieve a non-zero flux. Two possible
options of doing this are sketched in Fig. 11.
12

[1] Marin Bukov, Luca D’Alessio, and Anatoli Polkovnikov, [7] Martin Holthaus, “Floquet engineering with quasienergy
“Universal high-frequency behavior of periodically driven bands of periodically driven optical lattices,” Journal of
systems: from dynamical stabilization to Floquet engi- Physics B: Atomic, Molecular and Optical Physics 49,
neering,” Advances in Physics 64, 139–226 (2015). 013001 (2016).
[2] N. Goldman and J. Dalibard, “Periodically Driven Quan- [8] H. Lignier, C. Sias, D. Ciampini, Y. Singh, A. Zen-
tum Systems: Effective Hamiltonians and Engineered esini, O. Morsch, and E. Arimondo, “Dynamical
Gauge Fields,” Physical Review X 4, 31027 (2014). Control of Matter-Wave Tunneling in Periodic Poten-
[3] Or the ‘Floquet-Magnus expansion’ in the stroboscopic tials,” Physical Review Letters 99 (2007), 10.1103/Phys-
case. RevLett.99.220403.
[4] The analogous expressions for HF , called ‘Floquet- [9] N. R. Cooper, J. Dalibard, and I. B. Spielman, “Topo-
Magnus’ expansion, can be found in ref. [1], together with logical bands for ultracold atoms,” Reviews of Modern
higher-order terms of the expansions. Physics 91 (2019), 10.1103/RevModPhys.91.015005.
[5] Jean Dalibard, “Réseaux dépendant du temps,” Collège [10] F. D. M. Haldane, “Model for a Quantum Hall Effect
de France, Cours 4 (2013). without Landau Levels: Condensed-Matter Realization
[6] Mark S. Rudner and Netanel H. Lindner, The Floquet of the ”Parity Anomaly”,” Physical Review Letters 61,
Engineer’s Handbook , Vol. 2003.08252 (arXiv, 2020). 2015–2018 (1988).