Disorder and interference: localization phenomena

Cord A. Müller∗ and Dominique Delande†

Chapter 9 in: “Les Houches 2009 - Session XCI: Ultracold Gases and Quantum
Information”, C. Miniatura, L.-C. Kwek, M. Ducloy, B. Grémaud, B.-G. Englert,
L.F. Cugliandolo, A. Ekert, eds. (Oxford University Press, Oxford 2011)

This is v3 of arXiv:1005.0915 with minor corrections marked in color.

Communications from B. Nowak and C. Texier are gratefully acknowledged.

Contents
1 Introduction 3
1.1 Anderson localization with atomic matter waves . . . . . . . . . . . . . . . . . . 4

2 Transfer-matrix description of transport and Anderson localization in 1d sys-

tems 6
2.1 Scattering matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Transfer matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Chaining transfer matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 Incoherent transmission: Ohm’s law . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 Phase-coherent transmission: strong localization . . . . . . . . . . . . . . 10
2.3 Scaling equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Fokker-Planck equation and log-normal distribution . . . . . . . . . . . . . . . . 11
2.5 Full distribution function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Scaling theory of localization 15

3.1 What is a scaling theory? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Dimensionless conductance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Scaling in 1D systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 Quasi-1D systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Scaling in any dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.6 d = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.7 d = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.8 d > 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
∗ Centre for Quantum Technologies, National University of Singapore, Singapore 117543 and Department of

Physics, University of Konstanz, D-78457 Konstanz, Germany

† Laboratoire Kastler-Brossel, Université Pierre et Marie Curie, Ecole Normale Supérieure, CNRS; 4 Place

Jussieu, F-75005 Paris, France

1
4 Key numerical and experimental results 23
4.1 d = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.1 Localization of cold atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.2 Localization of light: a ten-Euro experiment . . . . . . . . . . . . . . . . . 24
4.1.3 Fluctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.4 The Anderson model: a free (numerical) experiment . . . . . . . . . . . . 26
4.2 d = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 d = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5 Microscopic description of quantum transport 35

5.1 Diagrammatic perturbation theory . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.1.1 Quantum propagator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.1.2 Ensemble average . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.1.3 Gaussian disorder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.1.4 Speckle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.5 Average propagator: self-energy . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 Intensity transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2.1 Density response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2.2 Quantum intensity transport . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.3 Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.4 Localization length in 1d systems . . . . . . . . . . . . . . . . . . . . . . . 45
5.2.5 Weak-localization correction . . . . . . . . . . . . . . . . . . . . . . . . . . 47

6 Coherent backscattering (CBS) 48

6.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.2 Live experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.3 Dephasing/decoherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.3.1 Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.3.2 Residual velocity of the scatterers . . . . . . . . . . . . . . . . . . . . . . 57
6.3.3 Non-linear atom-light interaction . . . . . . . . . . . . . . . . . . . . . . . 57
6.3.4 Internal atomic structure/spin-flip . . . . . . . . . . . . . . . . . . . . . . 57

7 Weak localization (WL) 59

7.1 d = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2 d = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.3 d = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.4 Self-consistent theory of localization . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.4.1 d = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.4.2 d = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.4.3 d = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

8 Kicked rotor 66
8.1 The classical kicked rotor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
8.2 The Quantum Kicked Rotor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.3 Dynamical Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.4 Link between dynamical and Anderson localizations . . . . . . . . . . . . . . . . 69
8.5 The quasi-periodically kicked rotor . . . . . . . . . . . . . . . . . . . . . . . . . . 70

References 74

2
1 Introduction
Although complex systems are ubiquitous in nature, physicists tend to prefer “simple” systems.
The reason if of course that simple systems obey simple laws, which can be represented by simple
mathematical equations, as expressed by Goldenfeld and Kadanoff [1]:
One of the most striking aspects of physics is the simplicity of its laws. Maxwell’s
equations, Schrödinger’s equation, and Hamiltonian mechanics can each be expressed
in a few lines. The ideas that form the foundation of our worldview are also very
simple indeed: The world is lawful, and the same basic laws hold everywhere. Every-
thing is simple, neat, and expressible in terms of everyday mathematics, either partial
differential or ordinary differential equations.
Everything is simple and neat—except, of course, the world.
Even though the world obviously is not simple, many systems can be split, be it only in Gedanken,
into simple components, each obeying simple laws. This is the viewpoint of standard reduction-
ism, upon which modern science has been built [2]:
The reductionist hypothesis may still be a topic for controversy among philosophers,
but among the great majority of active scientists I think it is accepted without question.
The workings of our minds and bodies, and of all the animate or inanimate matter of
which we have any detailed knowledge, are assumed to be controlled by the same set
of fundamental laws, which except under certain extreme conditions we feel we know
pretty well.
Reductionism was a key ingredient for the development of physics in the 19th and the first
half of the 20th century, when (classical and quantum) mechanics, electromagnetism, relativity,
thermodynamics, etc. came to huge success. But is there anything fundamental beyond the
simple laws of physics, or can one always reconstruct the properties of composed systems from the
workings of their parts? This credo of “constructivism” has been challenged by P. W. Anderson,
the author of the preceding quotation and one of the physicists who played a major role in the
analysis of complex physical systems:
The ability to reduce everything to simple fundamental laws does not imply the ability
to start from those laws and reconstruct the universe. In fact, the more the elementary
particle physicists tell us about the nature of the fundamental laws, the less relevance
they seem to have to the very real problems of the rest of science, much less to those
of society.
The constructionist hypothesis breaks down when confronted with the twin difficulties
of scale and complexity. The behavior of large and complex aggregates of elementary
particles, it turns out, is not to be understood in terms of a simple extrapolation
of the properties of a few particles. Instead, at each level of complexity, entirely new
properties appear, and the understanding of the new behaviors requires research which
I think is as fundamental in its nature as any other.
which he summarized shortly with:
More Is Different.
It is one aim of these lectures to show—in a restricted context—how intricate the interplay be-
tween the small and the large can be in complex systems. We did not yet give a precise definition
of the word “complex”. It turns out that the very same word may be used in different contexts

3
with different meanings. In these lectures, we will address a specific example of complexity, viz.,
disorder. In physics as in everyday life, disorder is associated with some lack of regularity. In a
disordered material, atoms are not arranged in crystalline periodic patterns, but appear in more
or less random positions. Randomness occurs because some agents, called “degrees of freedom”
by physicist, are not under control, either because we cannot or chose not to control them. It
means that we have to learn to deal not with a single specific complex system—that is called
a single realization of the disorder—but with a whole family of systems whose properties are
described in terms of distribution laws, correlation functions, etc. The goal of the game is not to
describe as accurately as possible a single system, but rather to predict global properties shared
by (almost) all systems, i.e. to acquire knowledge of universal features independent of the precise
realization of the disorder. These are instances of the “new behaviors” mentioned by Anderson.
And this is also the viewpoint taken long ago by classical thermodynamics where one forfeits the
microscopic description of a gas in terms of all positions and momenta, concentrating instead
on new concepts like entropy and temperature, which prove to be the relevant, and therefore
fundamental, concepts at this level of complexity.
The specific problem we address in these lectures is the problem of transport and localization
in disordered systems, when interference is present, as characteristic for waves. A wave propa-
gates in some medium (be it vacuum), and interference occurs when different waves overlap, for
example scattered from different positions with various wavevectors. The “simple laws” are the
wave equation in the homogeneous medium together with a microscopic description of scattering
by the impurities. The complex behavior we want to describe is, for example, the propagation
of the wave over long distances and for long times. Physical situations of this type cover the
propagation of sound in a concert hall with complicated shape, seismic waves multiply scattered
inside the earth, electronic matter waves in dirty semiconductor crystals, atomic matter waves in
the presence of a disordered potential, etc. In this context, it is good to keep in mind a warning
issued by W. Thirring [3]:
It is notoriously difficult to obtain reliable results for quantum mechanical scatter-
ing problems. Since they involve complicated interference phenomena of waves, any
simple uncontrolled approximation is not worth more than the weather forecast.

1.1 Anderson localization with atomic matter waves

To start with a specific, state-of-the-experimental-art example, imagine a one-dimensional non
relativistic particle evolving in a potential V (z) as depicted in Fig. 1. The evolution of the
wavefunction ψ(z, t) is given by Schrödinger’s equation:

i~∂t ψ(z, t) = Hψ(z, t) (1)

with the single-particle Hamiltonian

p2
H= + V (z). (2)
2m
Let us assume that the particle is initially prepared in a Gaussian wave-packet. In the absence
of any potential, the Gaussian wave-packet will show ballistic motion, where the center of mass
moves at constant velocity while the width increases linearly with time at long times. In the
presence of a certain realization V (z) of the disorder, the wave function will take a certain
form ψ(z, t). For different realizations, different wave functions will be obtained. But we are
not interested in the fine details of each wave function. Rather, we wish to understand the
generic, if not universal, properties of the final stationary density distribution |ψ(z)|2 obtained

4
at long times. We will see that not only averages, but also their fluctuations contain important
information.
Let us forget for a moment interference effects and try to guess what happens to a classical
particle. If its kinetic energy is much larger than the typical strength of the disorder V0 , the
particle will fly above the potential landscape, and the motion is likely to be ballistic on the
average. If on the other hand V0 is larger than the kinetic energy, the particle will be trapped
inside a potential well and transport over long distance is suppressed, i.e. localization takes place.
Quantum mechanics modifies this simple picture fundamentally: waves can both tunnel
through potential hills higher than the kinetic energy and be reflected even by small potential
fluctuations. So the initial wavepacket will split on each potential fluctuation into a transmitted
part and a reflected part, no matter how large the kinetic energy with respect to the potential
strength may in detail be. After many scattering instances, this looks like a random walk and
one naively expects that, on average, the motion at long times will be diffusive, with a diffusion
constant depending on some microscopic properties of particle and potential.
This simple model system has been recently realized experimentally [4] using a quasi-one-
dimensional atomic matter wave, interacting with an effective optical potential created by a
speckle pattern, see Fig. 1. The experimental result is the following: at short times, the
wavepacket spreads as expected, but at long times, its average dynamics freeze, and the wavepacket
takes a characteristic exponential shape:
|z|
 
2
|ψ(z)| ∝ exp − (3)
ξloc

where ξloc is called the localization length.1 Moreover, if a different realization of the disorder
is used (i.e. a microscopically different, but statistically equivalent speckle pattern), an almost
identical shape is obtained, meaning that the phenomenon is robust versus a change of the
microscopic details.
This surprising phenomenon is known as Anderson localization, sometimes also called strong
localization. Although it was predicted on theoretical grounds in the late 50’s—most famously
by P.W. Anderson himself [5]—it has only been observed directly rather recently. Cold atoms,
where an in situ direct observation of the wavefunction is possible, are from that point of view
highly valuable.
In these lecture notes, we present an introduction to transport properties in disordered sys-
tems, with a strong emphasis on Anderson localization. As an appetizer, we show in section 2
how the one-dimensional case can be exactly solved, providing us with useful physical pictures.
We then introduce in section 3 the scaling theory of localization, a typical illustration of the
appearance of new concepts and parameters relevant at the long distance and long time scale.
After reviewing some of the most important experimental and numerical results in section 4,
we develop a microscopic description of quantum transport in section 5, several applications of
which are discussed in sections 6, 7 and 8.
Many other, interesting questions will not be touched upon, foremost the impact of inter-
action between several identical particles. However, technically speaking averages over disorder
introduce an effective interaction. The relevant diagrammatic approach, originally introduced in
the context of quantum electrodynamics, is quite versatile and used equally well to describe, e.g.,
interacting electrons in solid state samples or interacting atoms in Bose-Einstein condensates. In
these lecture notes, we restrict the discusion to non-interacting particles in a disordered medium,
but the general framework and the technical tools introduced should provide our readers with
1 In the literature, the localization length is defined as the characteristic length for the decay of either |ψ|2

or of |ψ|. The two quantities of course differ by a factor 2. Usage of one or the other definition depends on the
community, but may also fluctuate from paper to paper. One has to live with this source of disorder.

5
Figure 1: Direct experimental observation of one-dimensional Anderson localization of an atomic matter
wave in a disorder potential. The disorder potential (represented in blue in the lower part of the figure)
is created by a speckle pattern. (a) An initially localized wave packet (prepared in a harmonic trap at
the center) evolves freely, diffuses and eventually freezes at long times in a characteristic exponential
shape (b). The pink tube represents the transverse-confinement laser beam that ensures an effectively
one-dimensional dynamics. Reprinted from [4] (courtesy of Ph. Bouyer).

solid foundations to follow also more advanced developments. Understanding the combined ef-
fects of interaction and disorder has been, still is, and doubtlessly will remain the subject of
fascinating research for a long time to come.

2 Transfer-matrix description of transport and Anderson

localization in 1d systems
In order to develop some intuition on transport in disordered systems, it is useful to study
solvable models, where one can identify relevant phenomena and mechanisms. It turns out that
a one-dimensional system, with the specific choice of δ point scatterers put at random positions,
provides such a solvable model, that is moreover sufficiently rich to teach us useful lessons for
more realistic disorder and higher dimensions.
Consider therefore a spinless particle confined to a 1d wave-guide geometry by tight transverse
trapping to its ground state. Free propagation along the z-direction with wave vector k is
described by amplitudes ψ±k (z) = exp{±ikz}. In the following, we discuss the transmission of a
single, fixed k-component through a series of obstacles j = 1, 2, . . . , N , well separated and placed
at randomly chosen distances ∆zj = zj − zj−1 as pictured in Fig. 2.

2.1 Scattering matrix

Each obstacle shall be described by a potential Vj (z). To fix ideas, we may assume that it is
sufficiently short-ranged to be well approximated by a δ-type impurity, Vj (z) = V (z − zj ) =
σ0 V0 δ(z − zj ), with an internal length scale σ0 that is not resolved by the propagating wave,
kσ0  1. Furthermore, we suppose that the obstacles are well separated, i.e., their density
n = N/L is small compared to the wavelength, n  k.

6
 e±ikz 1 t1 ?
...
r1
z1 z2 zN z

Figure 2: One-dimensional waveguide with randomly placed scatterers. We know the reflection and
transmission coefficients rj and tj of each scatterer. What is the total transmission TN across the whole
ensemble?

Consider first scattering by a single impurity at z = 0. We can decompose the wave functions
to the left (L, z < 0) and right (R, z > 0) into left- and right-moving components:

ψL (z) = ψLin e+ikz + ψLout e−ikz (4)

out +ikz in −ikz
ψR (z) = ψR e + ψR e . (5)

The outgoing amplitudes are linked to the incident amplitudes by the reflection and transmission
coefficients r and t from the left, and r0 , t0 from the right:
t
in
ψL out
ψR in
= t ψL + r 0 ψR
in

r r0 (6)
in 0 in out
t0 in
r ψL +t ψR = ψL ψR

Writing these relations in matrix form introduces the scattering or S-Matrix:

r t0
 out   in   
ψL ψL
=S with S= . (7)
out
ψR ψRin
t r0
For the present setting of a single-mode wave guide, the reflection and transmission coefficients are
complex numbers, and the probabilities for reflection and transmission from the left are R = |r|2
and T = |t|2 , respectively, and similarly from the right. In a more general setting of multi-mode
scattering with m modes or “channels” on the left and m0 modes on the right, r and t are matrices
with m × m and m0 × m entries, respectively. And for P example, the totalPtransmission probability
“all channels in to all channels out” then reads T = m,m0 tm0 m t∗m0 m = m0 (tt† )m0 m0 =: tr0 {tt† }.
Within this section, we have only use for the single-channel notation and refer to the literature
for the general case [6, 7, 8, 9]
Probability flux conservation requires that S be unitary, S† = S−1 . From S† S = 1, it follows
directly that reflection and transmission probabilities add up to unity: R+T = 1 and R0 +T 0 = 1.
One also finds r∗ t0 + t∗ r0 = 0 and its complex conjugate rt0∗ + tr0∗ = 0. From this, it follows for
the single-channel case that R = R0 , T = T 0 : the reflection and transmission probabilities are
the same from both sides.
Time reversal exchanges the roles of “in” and “out” states. For a time-reversal invariant
potential V (z), this implies S∗ = S−1 . With unitarity, this is equivalent to St = S or t = t0 , a
symmetry called reciprocity. This setting defines the so-called “orthogonal” symmetry class of
random matrix theory. Reciprocity is typically violated in presence of an external magnetic field
or magnetic impurities (see Secs. 6.3 and 7 below).

Exercise 1 – Consider an elementary impurity with V (z) = σ0 V0 δ(z). Solve the Schrödinger eigenvalue
equation −ψ 00 + (2m/~2 )V ψ = k2 ψ at fixed k (use the continuity of the free wave function and compute

7
its derivative discontinuity at z = 0) and show that the S-Matrix in terms of f = mσ0 V0 /~2 k is given
by [10]  
1 −if 1
S= . (8)
1 + if 1 −if

2.2 Transfer matrix

If we now have several impurities in series, in principle the total transmission can be calculated
from the S-matrix of the whole system. But the total S-matrix is not simply linked to the
individual S-matrices. Since the transmission depends on the incident amplitudes from both
left and right, adding a scatterer requires to recompute the entire sequence. So instead of
distinguishing in/out amplitudes, one prefers to decompose the wave function into right-/left
moving amplitudes, respectively: ψ(z) = ψ + e+ikz + ψ − e−ikz , and this on both sides R/L of the
obstacles. The transfer matrix M then maps the amplitudes from the left side of the obstacle to
the right:
+ +
ψL ψR
+
ψL+
   
ψR
or − =M − . (9)
M ψR ψL
− −
ψL ψR

One can easily determine its matrix elements in terms of t, t0 , r, r0 . For instance, ψLout = rψLin +
t 0 ψR
in
rewrites as ψL− = rψL+ + t0 ψR
−
, which we can immediately solve for ψR −
= t10 ψL− − tr0 ψL+ ,
+
and similarly for ψR . Eliminating r0 , t0 with unitarity relations in favor of r, t and their complex
conjugates yields a simple form:

1/t∗ −r∗ /t∗

 
M= . (10)
−r/t 1/t

Exercise 2 – Check the following interesting properties of the transfer matrix:

(i) det M = 1.
(ii) Current conservation (unitarity of S) implies now that Mσz M† = σz , with σz = 1 0


0 −1 the third
Pauli matrix.
(iii) Equivalently, M−1 = σz M† σz .
(iv) (M† M)−1 = σz M† Mσz . Thus, the hermitian matrices (M† M)−1 and M† M have the same (real)
eigenvalues. Verify this property by computing the eigenvalues directly using (10). Since these
eigenvalues must also be each other’s inverses, they can only be of the form λ+ = 1/λ− = e2x .
(v) As a matrix, 2 + (M† M)−1 + M† M = 4/T . Thus, the total transmission probability is T =
1/(cosh x)2 .

2.2.1 Chaining transfer matrices

By construction, the transfer matrix maps the amplitudes from left to right across each scatterer.
Therefore, the total transfer matrix across N scatterers is obtained by multiplying them:

M12...N = MN . . . M2 M1 . (11)

8
Consider the simplest case of two obstacles j = 1, 2 in series, for which M12 = M2 M1 . After
matrix multiplication, one finds the transmission coefficient
t1 t2
t12 = (12)
1 − r10 r2

This transmission amplitude contains the entire series of repeated internal reflection between the
two scatterers: t12 = t2 t1 + t2 r10 r2 t1 + t2 (r10 r2 )2 t1 + . . . . The transmission probability reads

T T
T12 = √ 1 2 iθ 2 (13)
|1 − R1 R2 e |

where θ is the total phase accumulated during one complete internal reflection. Since the scatter-
ers are placed with a random distance k∆z  2π, the phase θ will also be randomly distributed
in [0, 2π], independently of the details of the random distribution of distance between consecutive
scatterers or their reflection phases. One can calculate expectation values of any function of θ
by Z 2π
dθ
hf (θ)i = f (θ). (14)
0 2π
But as we will see in the following, a very important question is: “what quantity f (θ) should be
averaged?”

2.2.2 Incoherent transmission: Ohm’s law

The most natural thing seems to average directly the transmission probability (13). One finds

T1 T2
hT12 i = . (15)
1 − R1 R2
The same transmission probability is obtained for a purely classical model where only reflection
and transmission probabilities are combined:

Exercise 3 – Show that therule (15)

 is also obtained if one uses an S-matrix propagating probabilities
R T
instead of amplitudes, S̊ = , by determining the corresponding transfer matrix M̊ and chaining
T R
it.

In this case, the distance ∆z of ballistic propagation between the two scatterers is completely
irrelevant, and T12 is given by (15). This classical description applies to systems that are subject
to strong decoherence, where the phase of the particle is completely scrambled by coupling to an
external degree of freedom, while traveling between scatterers 1 and 2.
The so-called element resistance of the obstacles, (1 − T )/T , calculated with the classical
transmission, is additive:
1 − T12 1 − T1 1 − T2
= + . (16)
T12 T1 T2
Therefore, the classical resistance across N identical impurities distributed with linear density
n = N/L along a wire of length L grows like

R R1 L
(L) = N =: (17)
T T1 l1

where l1 = T1 /(nR1 ) is a length characterizing the backscattering strength of a single impurity.

9
 Within the context of electronic conduction, the result (17) is known as Ohm’s law, stating
that the total classical resistance of a wire grows linearly with its length L. Obviously, averaging
the transmission itself at each step has wiped out completely the phase-coherence and left us with
a purely classical transport process. This process can be formulated equivalently as a persistent
random walk on a lattice where the particle has uniform probability T1 to continue in the same
direction at each time step and probability R1 to make a U-turn. For long times, this random
walk leads to diffusive motion with diffusion constant [11]
T1
D = vl0 . (18)
2R1
Here v is the velocity of the particle and l0 = 1/n the distance between consecutive scatterers.
Since the diffusion constant is related to the transport mean free path l by the general relation
D = vl/d, with d = 1 the dimension of the system, one can identify l1 as twice the transport
mean free path, l1 = 2l.

2.2.3 Phase-coherent transmission: strong localization

The relation (15) cannot be easily generalized to more than two scatterers. Indeed, already
for three scatterers, the average of the complicated product of transmission matrices even over
independent, random phases θ12 and θ23 becomes very complicated. In order to predict the
behavior of transmission across long samples, it is advantageous to find a quantity that is additive
as new scatterers are added to the wire. There is such a quantity that becomes additive under
ensemble-averaging, namely the so-called extinction coefficient κ = − ln T = | ln T |. When
averaging the logarithm of (13), the denominator drops out since
Z 2π
dθ  p 
ln 1 − R1 R2 eiθ  = 0 (19)
 
0 2π
due to the analyticity of the complex logarithm for all 0 ≤ R1 R2 < 1. Thus, one immediately
finds that the average extinction across two consecutive scatterers is strictly additive:

hln T12 i = ln(T1 ) + ln(T2 ). (20)

The generalization to many scatterers is now easy because hln T i is additive: the total extinction
of a channel of length L grows on average like |hln T i| = nL| ln T1 |. With this scaling behavior,
one obtains that the log-averaged transmission

exp{hln T i} = e−L/ξloc (21)

drops exponentially fast with increasing sample length L. In the absence of absorption, this is
a hallmark of strong localization by disorder, and we have found the localization length ξloc =
1/(n| ln T1 |). In a weak scattering situation where nl1 = T1 /R1  1, we approximate | ln T1 | ≈
1/nl1 and thus find the localization length as ξloc = l1 = 2l.
What is the meaning of the log-averaged transmission (21)? It is important to realize that
the transmission T as a function of the microscopic realization of disorder is a random variable.
We will show in the following sections that T itself is not a self-averaging quantity, meaning
that its average hT i has no resemblance to the most likely found value of T , called the typical
transmission Ttyp . We will shortly see that for long samples, the probability distribution of ln T
is very close to a normal distribution (see (32) below), which is centered at hln T i = ln Ttyp .
And since the logarithm is monotonic, the most probable value of the transmission is indeed
Ttyp = exp{hln T i} = e−L/ξloc .

10
 How does one go about to identify a properly self-averaging quantity in the first place? —
By observing that the transmission matrices of several obstacles multiply, see eq. (11). The
logarithm of this random product therefore realizes a noncommuting variant of the central limit
theorem, known as Furstenberg’s theorem [12], which implies that the extinction or log-averaged
transmission is indeed a good candidate for a self-averaging quantity.

2.3 Scaling equations

In order to substantiate the previous arguments, we should find the full distribution function
P (T, L) that permits to derive expectation values for arbitrary functions of T at length L. This
distribution function can be found exactly by solving recursion relations that describe how the
transmission is changed when a small bit ∆L is added to a sample of length L:
TL T∆L

0 L L + ∆L

We know already that the transfer matrices multiply, ML+∆L = M∆L ML . The composition law
(13) then gives
TL T∆L
TL+∆L = iθ
√ . (22)
|1 − e RL R∆L |2
The idea is now to study the change in the expectation values of T as the “time” t = L/l1 grows.
For this, we assume that the added part ∆L is long enough such that an independent disorder
average h. . .i∆L in this section is meaningful. We also assume that the scatterers are weak such
that the backscattering probability remains small:

R∆L = ∆L/l1 =: ∆t  1. (23)

These assumptions are verified if ∆L is of the order of n−1 = l1 R1 with a single weak scatterer
on average. We may now expand (22) using T∆L = 1 − ∆t to leading order in ∆t, finding for the
change h p i
∆T = TL+∆L − TL = TL 2 RL ∆t cos θ + (4RL cos2 θ − 1 − RL )∆t . (24)

Thus, to order ∆t, we find by averaging h. . .i∆L = h. . .iθ that

(∆T )2 θ


h∆T iθ 2 h(∆T )n iθ
= −TL , = 2TL2 (1 − TL ), =0 (n ≥ 3) (25)
∆t ∆t ∆t
So only the first two moments of fluctuations contribute. From the first relation we can read
off that the transmission decreases with increasing length, an expected result. But the second
relation shows that the relative fluctuations, while small in the beginning where 1 − T = R  1,
will grow for long samples where T  1.

2.4 Fokker-Planck equation and log-normal distribution

The above equations (25) describe
a random quantity T (t) whose first two moments obey the
equations ∂hT i/∂t = hA(T )i and ∂ T 2 /∂t = h2T A(T ) + B(T )i with A(T ) = −T 2 and B(T ) =


2T 2 (1 − T ). Here, we use the continuous notations ∆t → dt and ∆T → dT while bearing in

mind the coarse-grained character of the averaged quantities. The theory of Brownian motion [13]

11
then teaches us that the probability distribution P (T, t) of T at time t obeys the Fokker-Planck
equation
1
∂t P = −∂T [AP ] + ∂T2 [BP ]. (26)
2
This equation of motion can be seen as a continuity equation, ∂t P + ∂T J = 0, for the locally
conserved probability density with current J = AP − 21 ∂T (BP ). Here, A describes the drift,
whereas B/2 plays the role of a diffusion constant.
By standard terminology, the Fokker-Planck equation (26) is called “non-linear”, because
B(T ) and A(T ) depend non-linearly on T . By changing the variable, one can try to simplify
these coefficients. A first option consists in using T = 1/(cosh x)2 (remember exercise 2(v )). In
the remainder of this section, let us explore the consequences of this choice. Knowing that the
change of variables for a probability density requires
 
−2
 dT 
P (T, t)dT = P ((cosh x) , t) 
  dx = P̃ (x, t)dx, (27)
dx 

we find the corresponding Fokker-Planck equation:

1 h i 1
∂t P̃ (x, t) = − ∂x coth(2x)P̃ + ∂x2 P̃ (28)
2 4
with initial condition P̃ (x, 0) = δ(x). With this choice of variable, the second derivative de-
scribing the fluctuations has become most simple. Although this equation is still (too) difficult
to solve exactly, we can extract the limiting distribution for long samples in the limit t  1 as
follows. First we rewrite the equation as
 
1 1 1
∂t + coth(2x)∂x P̃ (x, t) = P̃ + ∂x2 P̃ (29)
2 (sinh 2x)2 4

We may interpret the derivatives on the left hand side as a Lagrangian derivative ∂t + ẋ0 ∂x = Dt
in a co-moving frame defined by ẋ0 (t) = 21 coth(2x0 ), which is solved by x0 (t) = 21 arcosh(et ) ≈ 12 t,
for large t. Developing all terms for small deviations ∆ = x − x0 (t) from this point of reference,
we find a very simple equation for F̃ (∆, t) = P̃ (x0 (t) + ∆, t):
1 2
∂t F̃ (∆, t) = ∂ F̃ (∆, t). (30)
4 ∆
This is the elementary diffusion equation, and the solution, perhaps best known as the heat
kernel, is readily obtained by Fourier transformation:

∆2
 
1
F̃ (∆, t) = √ exp − (31)
2 πt 4t

Going back to the transmission using ln T = −2x valid for large x, we thus find the limiting
distribution
(ln T + t)2
 
1
Flog-norm (ln T, t) = √ exp − (32)
2 πt 4t
for small deviations around the most probable value ln T0 (t) = −2x0 (t) = −t. We have suc-
cessfully demonstrated that indeed the logarithm of the transmission is a normally distributed
random quantity. It is characteristic for disordered channels that the two defining moments
2
var(ln T ) = (ln T )2 − hln T i = 2t


|hln T i| = t, (33)

12
are determined by a single parameter, namely the length t = L/2l = L/ξloc of the one-dimensional
2
wire in units of the localization length. Clearly, the relative fluctuations var(ln T )/hln T i = 2/t
decay with system size. Thus we are assured that the transmission logarithm is a self-averaging
quantity, with moreover a normal probability distribution whose most probable value is equal to
the mean.
But careful! Even in the limit t  1, this does unfortunately not imply that one may use
the log-normal distribution (32) indiscriminately to calculate moments of the transmission. A
striking counterexample is
Z

ln T 
hT ilog-norm = e log-norm
= dyFlog-norm (y, t)ey = 1 (wrong), (34)

and quite obviously so. What goes wrong here? We will have a second look at the end of the
next section once we know the exact solution.

2.5 Full distribution function

Another choice of variable is ρ = T −1 , the dimensionless total resistance of the channel. The
Fokker-Planck equation (26) for its probability distribution W (ρ, t) = P (ρ−1 , t)ρ−2 reads

∂t W = ∂ρ [ρ(ρ − 1)∂ρ W ] (35)

with initial condition W (ρ, 0) = δ(ρ − 1), i.e., a wire of zero length has perfect transmission. The
solution can be calculated in closed form [14]:

exp{−t/4} ∞ exp{−y 2 /t}d(y 2 )

Z
W (ρ, t) = √ 3/2 √
p . (36)
πt arcosh ρ (cosh y)2 − ρ

Figure 3 shows how the distribution function F (κ, t) = W (eκ , t)eκ for the extinction κ = ln ρ =
− ln T moves from a δ-distribution with growing system size t to the log-normal distribution
(32), drawn as a dashed line at t = 10.
The full distribution function permits to calculate all moments hρ−n i = hT n i of the trans-
mission [14]. For t  1, one finds the asymptotic expression

π 3/2 Γ(n − 21 )2 −3/2 −t/4

hT n i = t e , (37)
2Γ(n)2

showing that all moments decay with the same dependence t−3/2 e−t/4 . The first two moments
are then
π 5/2 −3/2 −t/4
2 1
hT i = t e , T = hT i  1. (38)
2 4
Although the average transmission and its fluctuations decay exponentially, as expected for a
strongly localizing system, the relative fluctuations of the transmission itself grow very quickly:
2
var(T )/hT i ∝ t3/2 et/4 .
And why does the blind application (34) of the limiting log-normal distribution (32) predict
hT ilog-norm = 1 instead of the correct decrease (38)? Well, in (34), we have used the normal
distribution on the entire real axis for y = ln T , without paying attention to the constraint that
the physically admissible transmission is T ≤ 1. To take this into account, a popular recipe
consists in using a truncated log-normal distribution on the half-line κ ≥ 0 [9, 15]

(κ − t)2
 
+ Θ(κ)
Flog-norm (κ, t) = √ exp − (39)
C(t) πt 4t

13
 0.8

t = 0.5

0.6

F 0.4 1

0.2 5

10

0.0
0 5 10 15 20
κ = − ln T

Figure 3: Probability distribution function F (κ, t) = W (eκ , t)eκ for the extinction κ = ln ρ = − ln T
across a one-dimensional channel of length t = L/ξloc = 0.5, 1, 2, 5, 10. Dashed line at t = 10: log-normal
distribution (32).

√
with a normalization C(t) = 1+erf( t/2) ≈ 2. The moments √ of this distribution depend entirely
+
on the value at truncation, limκ→0+ Flog-norm (κ, t) = (2 π)−1 t−1/2 e−t/4 , that is the probability
for perfect transmission T = 1. The transmission moments for t  1 are predicted to be

+ 1
hT n ilog-norm ≈ √ t−1/2 e−t/4 . (40)
(2n − 1) π

Comparing them with the exact moments (37), we see that the truncated log-normal distribu-
tion can describe the leading exponential decay, but fails to capture the algebraic dependence
correctly. Mathematically, this is due to the fact that the log-normal distribution overestimates
the probability of perfect transmission T = 1, which is really only W (1, t) = (π 3/2 /2)t−3/2 e−t/4
for t  1.
On physical grounds, this limited applicability of the log-normal distribution emphasizes the
difficulties one faces when dealing with broad distributions. Figure 4 shows the transmission
probability distributions at different lengths t = L/l. It is instructive to look at the last curve
for the altogether moderate system size t = 10. The most probable or “typical” value for the
transmission is quite small, Ttyp = exp{hln T i} = e−10 ≈ 4.5 · 10−5 . The exact average value is
much bigger, hT i ≈ 1.06·10−2 , indicating that this average is to a large extent determined by very
rare events with anomalously large transmission. The truncated log-normal distribution, shown
in dashed, overestimates the frequency of large transmissions and predicts hT i ≈ 1.28 · 10−2 .
Let us close this section by emphasizing once more that the typical transmission — averaged
over all possible relative phases accumulated between consecutive scatterers — displays expo-
nential decay at large distance as stated in eq. (21). This is in sharp contrast to the classical
transmission (17), where classical probabilities, not amplitudes, are combined to an algebraic de-
cay. Clearly, averaging over disorder implies averaging over quantum mechanical phases globally,
but is not equivalent to removing phase-coherence and interference effects locally from the very
start. This is a striking example of a mesoscopic effect, a rather counter-intuitive phenomenon,
where microscopic phase coherence has macroscopic physical consequences that survive averaging
over quenched disorder.

14
3 Scaling theory of localization
3.1 What is a scaling theory?
A scaling theory describes the relevant properties of physical systems by considering their be-
havior under changes of size L 7→ bL. Quantitative scaling arguments were invented in quantum
field theory in the context of renormalization. Scaling arguments became widely popular in sta-
tistical physics by the mid-60’s for describing phase transitions and critical phenomena [16]. The
immense success of renormalization-group techniques developed in the 70’s [17] rapidly radiated
to the field of disorder-induced phase transitions that Anderson’s celebrated paper had founded
[5]. After pioneering work by Wegner [18], a scaling theory of localization was formulated by
Abrahams, Anderson, Licciardello, and Ramakrishnan [19], a quartet that became known as the
“gang of four”.
A scaling theory can hope to capture those features that are important on macroscopic
scales, but will be insensitive to microscopic details. This means that its predictions are only
semi-quantitative, in the sense that it cannot furnish the precise location of a critical point
in parameter space nor provide any system-specific data. In return, if one feeds it with the
microscopic data (such as the transport mean-free path), it can give general, and surprisingly
accurate, predictions of universal character.
Let us have a look at the different lengths characterizing quantum transport in a disordered
material:

λ ζ l ξloc
| | | | size

L L L

As short-scale lengths (in the left dotted box) one has the wavelength λ = 2π/k of the propagating
object and the correlation length ζ of the disorder. If ζ  λ, the details of the disorder are
unimportant, and models of δ-correlated scatterers are appropriate. If ζ  λ, the disorder
correlation can be resolved by the wave; this is typically the case for our example of optical
speckle potentials probed with ultracold atoms [4].

10.00
5.00
t = 0.5

1.00 1
0.50
2
P

0.10
0.05 5

0.01 10

0.0 0.2 0.4 0.6 0.8 1.0

T

Figure 4: Probability distribution function P (T, t) for the transmission T across a one-dimensional
channel of length t = L/ξloc . Dashed line at t = 10: log-normal distribution (32).

15
 10 10
8 8
6 6
g g −1
4 4
2 2
0 0
0 1 2 3 4 0 1 2 3 4
L/l L/l

Figure 5: Typical conductance (left panel) and resistance (right panel) of a 1d channel as function of
sample length L/l. Short channels have a resistance that grows linearly as expected by Ohm’s law (green
dashed, eqn (41)), whereas long channels show exponentially large resistance (full line, eqn (42)).

As larger scales (in the right dotted box) one has the transport mean free path l and the
localization length ξloc . We have already seen in section 2.2.3 that in d = 1 these two lengths are
practically identical, ξloc = 2l. In d = 2 the localization length is much larger than the transport
length, as will be discussed in sections 3.6 and 7.4 below, and in d = 3 it may well be infinite.
Depending on the system size L, one can distinguish three basic transport regimes: ballistic
transport through small samples with L < l, diffusive transport for l < L < ξloc with, possibly,
weak-localization corrections, and finally strong localization for large samples with ξloc < L. In
d = 1, there is no room for diffusion between l and ξloc , and strong localization is basically a
single-scattering effect.2 The scaling theory of localization has the purpose of describing the
transition between these regimes as function of system size L [22].

3.2 Dimensionless conductance

Traditionally, the scaling theory of localization is formulated in terms of a channel’s proper
conductance, a dimensionless parameter defined as g = T /R by transmission and reflection
probabilities. Equivalently, one may consider the channel’s proper resistance g −1 . A perfectly
transmitting channel T = 1 has a proper conductance of g = ∞, and a perfectly resisting
channel with T = 0 has g = 0, which seems a rather sensible definition. Moreover, we have seen
in section 2.2.2 that this resistance is additive when classical subsystems are chained in series.
Alternatively, one could define the total resistance as ρ = 1/T = 1 + g −1 , where the additional 1
represents the “contact resistance” due to the leads connecting the sample to the external world.
In the previous section, we also learned that the transmission of a disordered channel is
a random variable with a broad distribution around a most probable, typical value Ttyp =
exp hln T i. Therefore also g = T /(1 − T ) is a broadly distributed random variable, fluctuating
around the typical conductance gtyp = Ttyp /(1 − Ttyp ). In all of the following, we discuss the
behavior of gtyp , but in order not to overburden the notation, we will simply write gtyp = g.
In order to get used to this vocabulary, let us reformulate the results of section 2 for the
typical conductance. The exact exponential behavior (21) of the typical transmission translates
2 See [20, 21] for a situation where backscattering of an atom by a smooth speckle potential is zero at lowest

order, but localization still prevails due to higher orders in perturbation theory.

16
into
2l/L, L  l, (41)

1
g(L) = =
exp{L/2l} − 1 exp{−L/2l} L  l. (42)

This conductance together with the resistance g −1 is plotted in Fig. 5. Only the conductance of
short, ballistic channels is given by the classical expression (41), that we have already encountered
as Ohm’s law in section 2.2.2.
The results of scaling for the conductance can be reformulated for other quantities if those
seem more convenient. One of the most popular, and useful, quantities is the diffusion constant
D = vl/d, the product of velocity and transport mean free path divided by the number of
dimensions d, a convention whose rationale will become clearer below. For matter waves with
wave vector k, the velocity is v = ~k/m, and the diffusion constant can also be written D =
dm kl, i.e., the product of an elementary diffusion constant (~/dm) by the dimensionless quantity
~

kl = 2πl/λ. This ratio describes the effective disorderedness of the medium: kl  1 means that
the wave can travel over many periods before suffering scattering. We will see in Sec. 7 below
that kl is a crucial parameter for transport and localization properties.
In a metallic sample with usual electrical conductivity, the Drude formula σ̄ = ne2 τ /m
establishes the direct proportionality between the conductivity σ̄ and the classical diffusion con-
stant D = v 2 τ /d of charge carriers with mean free path l = vτ . We are thus led to define a
dimensionless classical conductivity3
2mD
σ̊ = . (43)
~
The conductance g̊ of a sample of linear size L in d dimensions is the ratio of σ̊ to L2−d . To
see this, picture a metallic block where the voltage U is applied along one dimension to give
E = U/L, whereas the current density j = σ̄E over the transverse area Ld−1 yields the total
current I = Ld−1 j, which results in the dimension-full conductance G = I/U = Ld−2 σ̄. In order
to define a dimensionless conductance g̊, one has to compensate the factors of L with another
length scale. The simplest choice is the inverse of the wave number k. One can thus define a
classical dimensionless conductance as

g̊(L) = (kL)d−2 σ̊ (44)

In 1d, this definition gives g̊(L) = 2l/L, i.e., is fully compatible with the known exact result
at short distance, eq. (41) (this is the reason for the factor 2 introduced in eq. (43)). It is of
course no accident that the two definitions of dimensionless conductance — through the diffusion
constant or through the transmission across a sample — coincide. The Landauer formula [23]
makes the connection explicit.
In any dimension, the classical dimensionless conductance can be rewritten as:
2kl
g̊(L) = (kL)d−2 (45)
d
In particular, in dimension 2 we have g̊ = kl itself, independently of the system size.
3 This definition, as well as (44), uses ~/m available for quantum matter waves. It should be adapted to any

other specific transport problem under study, along the same lines. The somewhat arbitrary factor of 2 is included
for future convenience.

17
 β
1

0
ln g
-4 -3 -2 -1 1 2 3 4

-1

-2

-3

-4

Figure 6: Conductance scaling β-function in d = 1, eq. (48). Arrows show the flow from the Ohmian
behavior (49) for g  1 in short samples to the exponential localization (50) for g  1 in long samples.

3.3 Scaling in 1D systems

Since we wish to follow how the dimensionless conductance g evolves with system size, we make
use of the β-function,
d ln g d ln g
β=L = . (46)
dL d ln(L/L0 )
β = 0 means that g does no change with L. Actually, β = cst implies purely algebraic dependence
g(L) ∝ Lβ . The celebrated function β(g) has been introduced originally by Callan and Symanzik
to describe the change of a coupling constant under a change of scale within quantum field theory
[24]. Let us familiarize ourselves with the β-function, arguably the most important single object
of scaling theory, in the case d = 1, for which we know already everything exactly. It is a matter
of elementary calculus to find
L 1
β(L) = − . (47)
2l 1 − exp{−L/2l}

Since g(L) is a monotonous function of L/2l, one can easily invert this dependence and express
β as function of the conductance alone:

β(g) = −(1 + g) ln 1 + g −1 .
 
(48)

This result can also be derived directly as follows: the linear scaling of the typical-transmission
−1
logarithm implies Ttyp (bL) = [Ttyp (L)]b = exp{b ln Ttyp (L)}. Writing Ttyp = 1 + g −1 , differenti-
ating with respect to b, and setting b = 1 at the end leads to (48). The fact that β(g) can be
expressed as function of g instead of the original length scale L does not seem very profound
in d = 1 [14]. However, in field theory this property is vital for renormalizability [24], and
in statistical physics it guarantees that β(g) can describe universal behavior close to a phase
transition.

18
 Figure 6 shows the β-function for d = 1, plotted as function of ln(g) together with its
asymptotics. For short samples L  l, the conductance g ∝ L−1 is large, and limg→∞ β(g) = −1.
More precisely, one has the following asymptotic behavior:
1
β(g) = −1 − + O(g −2 ) (49)
2g

which shows a weak-localization correction (see section 7 below). In the opposite limit of a large
sample, g  1 is exponentially small, and

β(g) = ln g − g(| ln g| + 1) + O(g 2 ). (50)

The transition between the two asymptotic regimes occurs around g = 1. The function β(g)
entirely describes how the dimensionless conductance evolves with system size L. Indeed, if g
is known for some small size L, it can be deduced for any other size by solving the differential
equation (46), moving along the arrows shown on the curve of figure 6. This is the so-called
“renormalization flow” followed by the system when its size is increased towards macroscopic
scales. For d = 1, β is always negative, implying that g always decreases with increasing system
size, and thus the renormalization flow is unidirectional from right to left. Characteristically,
when g decreases, β becomes more negative, which makes g decrease even faster, until it finishes
by dropping exponentially fast.
A crucial asset of scaling theory is that its predictions are valid for an arbitrary 1d system,
although the specific form of β(g) was deduced using a specific model. Suppose you have a
large-size complex disordered system and that you want to study its conductance. You may
start with a small system for which you can calculate the conductance microscopically using a
method of your choice. By following the renormalization flow, you are then provided, almost
magically, with the conductance at any scale. Moreover, you find the localization length ξloc
as the system size for which g(ξloc ) = O(1). Of course, there is no real magic here: your initial
calculation yields the mean free path l and, as this is the only macroscopic length scale relevant
for transport in a 1d system, you finally have everything.

3.4 Quasi-1D systems

The previous results may also be applied to quasi-one-dimensional systems that consist of several
parallel channels i = 1, . . . , N⊥ . One may either have in mind channels that are literally built
parallel to each other [25] or a multi-mode waveguide with spatially overlapping transverse modes.
If there is no coupling between channels, then the purely 1d description of section 2 applies. For
weakly coupled channels, which arises naturally by the disorder present, an equation of motion
for the full distribution function of transmission eigenvalues very similar to (35) has been derived
by Dorokhov and independently by Mello, Pereyra, and Kumar, known as the DMPK equation
[9]. Also the scaling picture remains essentially the same. Keeping the number of transverse
modes N⊥ fixed, the short-scale conductance is g = N⊥ 2l/L, as expected for parallel resistors.
Thus, the initial condition for the scaling flow on the curve β(g) is changed, but the transition to
the localized regime is the same. Since the crossover again occurs at L = ξloc with g(ξloc ) = O(1),
we simply find that the localization length is increased toward ξloc = 2N⊥ l.

3.5 Scaling in any dimension

In arbitrary dimension d, one changes the system size L 7→ bL in all directions, but still looks at
the transmission along one chosen direction. In the ballistic regime L  l, we start again from

19
the classical behavior, eq. (45), where g(L) ∝ Ld−2 . One therefore expects to find limg→∞ β(g) =
d − 2, and
cd
β(g) = d − 2 − + O(g −2 ) (51)
g
where a microscopic calculation is required to find the coefficient cd that describes weak local-
ization corrections.
In the strongly localized regime L  ξloc , exponential localization prevails. And since adding
parts to the system beyond the localization length in the perpendicular direction cannot change
its longitudinal transport, we still expect the power law Ttyp (bL) = [Ttyp (L)]b to hold in each
channel. Thence follows the asymptotic behavior

β(g) = ln(g/gd ) (52)

in the strongly localized regime in any dimension, with a constant gd of order unity.
Taking into account that the number of transverse channels scales as bd−1 , we would obtain
the simple scaling relation Ttyp (bL) ≈ bd−1 [Ttyp (L)]b , if there were strictly no coupling between
the channels. Then, the same calculation than for 1d would give

β(g) = (d − 1) − (1 + g) ln 1 + g −1 ,
 
(53)

that is a simple vertical shift of (48) by d − 1. In particular, this would imply that the weak
localization correction −cd /g is the same in all dimensions, a result known to be wrong, see Sec. 7.
It nevertheless remains true that the shape of the true β(g) curves, interpolating smoothly
between the known asymptotics, is qualitatively given by (53), see also Fig. 7. Although the
scaling description encompasses arbitrary dimensions, its consequences are radically different in
d = 2 and d = 3, meriting a separate discussion.

3.6 d=2
In the ballistic limit of short samples with typical conductance g  1, β(g) ≈ 0 describes
scale-independent conductance of N⊥ ∝ L transverse channels, each with element conductance
g ∝ L−1 . But then, β(g) is not exactly zero. Starting the flow at the finite conductance g0 of
a sample of length L0 , a slightly negative β(g) = −c2 /g makes g decrease with size (it will be
shown in section 7 below that c2 = 2/π). We can integrate the flow equation
c2 1 dg
β(g) = − = (54)
g g d ln(L/L0 )
by elementary means to find
g(L) = g0 − c2 ln(L/L0 ). (55)
To fix ideas, we can chose L0 = l, a scale on which transport is classical, such that, from eq. (45),
g(L0 ) = g̊ = kl  1. The transition to the strong localization regime occurs at g(ξloc ) = O(1).
Together with (55), this predicts an exponentially large localization length

ξloc ∼ L0 exp{g0 /c2 )} = l exp{kl/c2 }. (56)

The prediction of scaling theory for noninteracting particles in d = 2 is therefore that all states
are localized. This transcends also from the scaling flow depicted in figure 7. However, the
localization length can be extremely large if the system is only weakly disordered with kl  1.
Let us take some figures from the Orsay experiment [4]. With l = 100 µm and k = 2.5 µm−1 ,
one finds the rather large localization length ξloc ≈ le400 , which would surely overstretch the

20
 β
1

ln g
ln gc
0

-1

-2

-3

Figure 7: Schematic plot of the β-function in d = 1, 2, 3, showing a smooth interpolation from the
metallic regime (51) for g  1 in short samples to the localized regime (52) for g  1 in long samples.
Note the existence of an unstable fix-point at critical gc in d = 3.

possibilities of even the most capable experimentalist. The take-home message here is: In order
to observe 2d localization, kl should be chosen as close to unity as possible. In turn, this implies
that the classical diffusion constant D = ~kl/2m must be of the order of ~/m, which for a
typical cold atomic gas is a rather small quantity of the order of 10−9 m2 /s. In order to observe
2d Anderson localization with atomic matter waves, the experimentalist must be capable to
observe the dynamics for a long time while keeping phase coherence, a challenging task indeed.

3.7 d=3
In d = 3, we encounter a qualitatively new situation: the β-function is positive for large g.
So if we start with some g  1, the conductance flow will take us to even larger values of
g. In renormalization-group terms, the behavior of the system in the thermodynamic limit
L → ∞ is described by the “infra-red stable fix-point” g = ∞ (as in statistical physics, but
in contrast to quantum field theory, we are interested in the large-distance behavior, i.e., the
infra-red asymptotics with respect to momentum). Since large conductance characterizes a good
conductor, this is also known as the “metallic” fix-point.
By contrast, if we start with some g  1, the negative β-function will drive the system
towards the stable insulating fix-point g = 0 with exponentially small conductance at finite
length. Between these two extrema, the β-function, assumed to be continuous, must have a
zero at some gc . A zero of β(g) ∝ dg/dL is also a fix-point, but in this case an unstable one.
This unstable fix-point β(gc ) = 0 marks the critical point and shows the possibility of a metal-
insulator phase transition at some critical strength of disorder. Although a scaling theory, with
its roughly interpolating β-function, cannot predict the precise position of the critical point, it
can give a semi-quantitative estimate. Indeed, a microscopic calculation of the transport mean
free path l provides us with the dimensionless conductance g(l). At such a scale, interference
effect are unimportant, and thus, eq. (45) can be used, giving g(l) ≈ 2(kl)2 /3. As the critical
point is such that gc ∼ 1, we obtain that the threshold for Anderson localization is given by:

kl ∼ 1 (57)

21
an equation known as the Ioffe-Regel criterion for localization. The precise value of the critical
kl depends on microscopic details and is thus not universal.
Let us assume that the microscopic physics involves disorder whose strength is measured
by some parameter W , typically the width of the disorder probability distribution (cf. Sec.
4.1.4). Even though scaling theory does not predict the precise position of the critical point, the
behavior of the β-function around the critical point yields precious information about the large-
scale physics: it permits to calculate critical exponents that are the hallmark of universality. In
their 1979 paper [19], Abrahams et al. showed that the localization length diverges close to the
transition for W > Wc as
ξloc ∼ (W − Wc )−ν , (58)
where the critical exponent ν = 1/s is determined by the slope of the β-function at the transition,
s = [dβ/d ln g]gc .
The calculation leading to this prediction is elementary, but quite instructive in order to
appreciate the power of a scaling description. Let us start at some length L0 with some value
g0 < gc on the localized side of the fix-point. The β-function always allows us to calculate any
other g(L) implicitly by integration:
ln g
d ln g 0
 
L
Z
ln = . (59)
L0 ln g0 β(g 0 )

Using the linearized form β(g) = s ln(g/gc ) around the fix-point leads to
 s
L ln(gc /g)
= . (60)
L0 ln(gc /g0 )
−1/s
Now we are free to choose L0 = ξloc for which g0 = O(1) such that ξloc ∼ L [ln(gc /g)] .
Because the microscopic physics on small scales ignores the critical behavior on large scales and
can involve only smooth dependencies, one can always write ln(g/gc ) ≈ (g − gc )/gc ∝ (Wc − W )
close enough to the critical conductance gc , and we finally end up with (58).
For the simplest possible interpolation (53), one finds ν ≈ 1.68. This value is not disas-
trously far from the true value ν = 1.58 ± 0.01 that is known today from extensive numerical
simulations [26, 27], cf. Sec. 4.3.
The “metallic” side of the transition can also be studied using a similar approach, but fol-
lowing the metallic branch β > 0 of the renormalization flow. For smaller-than-critical disorder
strength W < Wc , the microscopically computed g at some size L0 will be slightly larger than
gc . It is left as an exercise for the reader to show that this results at large scale in a diffusive
(i.e., metallic) behavior with a diffusion constant

D ∝ (Wc − W )ν . (61)

The continuous (algebraic) vanishing of diffusion constant and conductance on the metallic side
of the Anderson transition is characteristic of a continuous second order phase transition.

3.8 d>3
The Anderson transition is expected to take place in any dimension d ≥ 3. According to the
simple scaling theory sketched above, the transition point will shift to lower and lower gc , requir-
ing a more strongly scattering medium to observe localization, and thus a Ioffe-Regel criterion,
eq. (57), with a smaller constant.

22
 Contrary to conventional phase transitions, the Anderson metal-insulator transition does not
have a finite upper critical dimension above which fluctuations would be unimportant and critical
exponents simply given by their mean-field values [15]. This is compatible with the observation
that as the dimension d increases, the zero of the β-function must shift more and more to the
asymptotic ln(g)-wing where the slope tends towards s = 1. Thus, from the scaling description
it is tempting to surmise that the critical exponent tends towards ν = 1 only continuously as
d → ∞. We will see in section 8 below that this observation is not only a theoretician’s spleen
but may be put to experimental testing.

4 Key numerical and experimental results

Over the past 50 years, a wealth of numerical and experimental results has been accumulated on
localization phenomena, especially on Anderson localization in dimension 1, 2, 3 and beyond. In
the following, we present a selection, necessarily subjective and limited, of the most remarkable
results.

4.1 d=1
Anderson localization is a generic feature in phase-coherent 1d and quasi-1d systems, as explained
in Section 2.2.3 above. Any amount of disorder, even very small, will eventually localize a
wavepacket, independently of how large its energy is. Of course, the localization length can be
huge if the energy is large compared to the disorder; see Section 5.2.4 for a quantitative estimate.

4.1.1 Localization of cold atoms

Concerning the experiment described in the Introduction, there is thus no surprise that a quasi-
1d atomic wavepacket displays localization in an optical speckle potential. Figure 8a) shows
the experimentally measured spatial shape of the wavepacket at various times. One clearly
distinguishes an exponential decrease in the wings, from which a localization length is extracted
by a fit to exp{−2|z|/Lloc }. As shown in b), this localization length first increases with time,
then settles for a stationary value after about 500 ms. In the stationary regime, the wavepacket
displays spatial fluctuations which are different for each single realization of the disorder. In
addition, there remains a large fraction of the atoms still trapped near the original location of
the wavepacket.
How can we understand these experimental results? The initial wavepacket is not monochro-
matic at all: it contains plane waves with a large dispersion in the wave-vector k and consequently
in the kinetic energy ~2 k 2 /2m (the added optical potential also contributes to the total energy,
but is a small correction here). The initial, free expansion of an interacting Bose-Einstein con-
densate released from a harmonic trap leads to a population of the various k classes that is given
by an inverted parabola [28]:
2
3(kmax − k2 )
Π0 (k) = 3
Θ(kmax − |k|), (62)
4kmax

where kmax is the maximum k value, related to the initial chemical potential by µ = ~2 kmax
2
/2m.
Because the disordered is “quenched” or stationary, energy is conserved, and each k-component
of the wavepacket evolves independently. When averaged over time, the interference terms be-
tween different energy components will be smoothed out, leaving the averaged wavepacket as
the incoherent superposition of all energy components. From Section 2.2.3, we know that each

23
 Atomic density (at
10

10

Atomic density (atoms µm–1)

NATURE | Vol 453 | 12 June 2008
–0.5 0.0 0.5
z (mm)
100 0.8 a
a b b
0.8 s
Atomic density (atoms µm–1)
1.0 s
2.0 s 0.6
10

Lloc (mm)
0.4
10 –1.010

0.2 10

Atomic density (atoms µm–1)

0.0
–0.5 0.0 0.5 0 1 2

~
~
z (mm) t (s)
–1.0 –0.5 0.5
0.8
Figure 2 | Stationarity of the localized profile. a, Three successive density
b b
Figure 8: a) The atomic density of a BEC profiles, from which
expanding in a the localization
quasi-1d opticallength Llocpotential,
speckle is extracted by fitting
shown in an
exponential, exp(22 | z | /L loc) (dotted black lines ), to the atomic density in Figure 4 |
logarithmic
0.6 scale at various times, displays clear exponential localization in the wings, from which the 10
the wings. b, Localization length L loc versus expansion time t. Error bars, potential.
localization length is extracted by a fit (dashed line). b) The localization length first increases linearly
95% regime.
confidence intervalsfrom for the fitted values (62 Bouyer).
s.e.m.). profiles, s
with time, then saturates in the stationary Reprinted [4] (courtesy of Ph.
Lloc (mm)

exponenti
0.4 kmaxsR 5
exponent of 1.95 6 0.10 (62 s.e.m.),10in agreement with the theor- –1
k-component localizes with a localization etical prediction thatequal
length ξloc (k) density decreases
to twice like 1/z2 in
the transport meanthe free
wings. The released B
fit to the w
path. For0.2the fastest atoms, this localization length is much larger than the initial spatial exten- work as
semi-log plot (inset) confirms that an exponential would not
(62 s.e.m
sion of the wave packet. Thus, we predict well. Forthe
comparison,
stationarywe present
spatial in Fig. 4b aonce
distribution log–log plot and a semi-
localization
right-hand
sets in to be roughly given by log plot (inset) for the case with kmaxsR 5 0.65 and VR/min 5 0.15, confirms t
0.0 Zwhere we2 concludein favourof exponential rather than algebraic
0 1 kmax (log–log a
tails. Π0 (k)
These dataexp
support |z|
2
− the existencedk. of a crossover from (63)an expo-

~
~
h|ψ(z)|
t (s) i = and kmaxs
nential 2ξ (k) ξ
−kmax to an algebraic regime
loc loc (k) in our speckle
–1.0 potential.
–0.1 0.1
Figure 2 | Stationarity of the localized profile. a, Three successive density z (mm)
Sincewhich
profiles, from ξloc (k)
theislocalization
an increasing function
length of |k|,2.5
Lloc is extracted see
bysection 5.2.4, the asymptotic decrease at large
fitting an Direct
2
distance is dominated by the largest |k| values, such
exponential, exp(22 | z | /Lloc) (dotted black lines), to the atomic density in that h|ψ(z)| i
Figure∝ |
exp[−|z|/ξ
4 Algebraic (k
and
loc max )]. The
exponential regimes disordere
in a one-d
low-k components have short localization lengths
the wings. b, Localization length Lloc versus expansion time t. Error bars, and thus potential.
produce the Log–log
large bumpand semi-log
near the plots of the stationary
of open q
origin in the final density. Using only
95% confidence intervals for the fitted values (62 s.e.m.).2.0 the wings of the profiles,
experimentally showing
measured the difference
density, between
it the algebraic
problems (k
is possible to estimate the localization length, for which we derive exponential
in Sec.(k5.2.4 maxsRa,theoretical
1) regimes. a, Density profile direct for
im
of 1.95 6 0.10 (62 s.e.m.), in agreement with the theor-
exponentprediction. kmaxsR 5 1.16 6 0.14 (62 s.e.m.). The momentum distr
ities to m
etical prediction that density decreases like 1/z2 in the wings. The released BEC has components beyond the effective
b Second,
mob
semi-log4.1.2
plot (inset) confirms that an exponential
Localization of light: a ten-Euro experiment
1.5 not work as
would fit to the wings with a power-law decay 1/ | z | yields
controlleb5
Lloc (mm)

(62 s.e.m.) for the left-hand wing and b 5 2.01 6 0.03 (

well. For comparison, we present in Fig. 4b a log–log plot and a semi- Bose glas
The reasoning in section 2.2.3 is entirely based on the construction right-hand
of a 2wing. The insetmatrix
× 2 transfer shows the same data in a sem
log plot (inset) for the case with kmaxsR 5 0.65 and VR/min 5 0.15, decay. b, For comparison, gases and
which can be chained; any randomness in the1.0 transfer matrix thenconfirms
leads tothe non-exponential
localization. The fact
where wethat conclude in favour of exponential rather than algebraic
our starting point was a quantum matter wave and the underlying (log–logwave-particle
and semi-log) in the exponential
duality of been
regime for the pred
tails. These data support
quantum mechanics the are
existence of a crossover
not central from an expo-
to this argument. kmaxsR 5 0.65 6 0.09
and transfer-matrix
Indeed, the (62 s.e.m.).
description measurem
nential toapplies
an algebraic regime in our speckle potential.
to all physical situations governed by a 1d (or quasi-1d) linear wave equation. Conse- ization in
0.5 in our se
quently, Anderson localization has been observed for many different types
2.5 Directofimagingnon-quantum waves:quantum gases
of atomic in co
microwaves, elastic waves in solids, optical waves, to cite a few. disordered potentials is a promising technique simulato
to31,32
in
Even a poor man’s experiment using viewgraph transparencies, of openi.e.,questions
plastic films on made of quantumsions
disordered systems
,
0.0 ition, an
polyester carbonate, allows to observe Anderson 0 localization.20 A problems
stack of 40 several transparencies
of condensed 60 matter80simulated using
2.0
parallel to each other, separated by air layers of randomly varying thickness, realizes the simple transition
directVimaging
R (Hz) of atomic matter waves offers unprec
exponent
model shown in fig. 2. The transmission and reflection coefficients for an individual film can be
ities to measure important properties, such as loc
controlle
computed from its index of refraction Figure 3 | Localization
and its thickness. If the randomness
length versus amplitude in the filmdisordered
of the spacing ispotential.
Second, our experiment can be extended to qua
larger than an optical wavelength, weLloc
1.5 is obtained
have by an exponential
a truly disordered system. fit to the wingsand
Transport of the stationary localized
localization
controlled interactions where localization of q
Lloc (mm)

density profiles, as shown in Fig. 2. Error bars, 95% confidence intervals for
14,15,28 29
METHOD
the fitted values (62 s.e.m.); 1.7Bose3 104glassatoms ; min and 5 219 Lifshits
Hz. Theglass
dash–dot are expected, Momentu as
line is plotted 24using equation (1),gases
where and kmax toisBose–Fermi
determined from mixtures
the where ization rich pha len
1.0 observed free expansion of the condensate
been predicted (see Methods).
30 The shaded
. The reasonable area
quantitative the agreem
maxim
represents uncertainty associated with the evaluations
measurements and the of kmax and of
theory sR. one-dimensional
We ning of the
note that the limited extension of the disordered potential (4 mm) allows
ization in a speckle potential demonstrates the high us BEC with
to measure values of Lloc up to about 2 mm. observing
0.5 in our set-up. We thus anticipate that it can be us
simulator for investigating Anderson localization
sions31,32, first©2008 Macmillan
to look for thePublishers
mobilityLimited.
edge ofAllthe
rig
0.0 ition, and then to measure important features
0 20 40 60 80
transition that are still under theoretical investigatio
VR (Hz)
exponents. It will also become possible to investi
 factors from all the matrices, and, defining the complex
variables zs = exp(2iφs ), write E as a 2N-fold contour
integral around unit circles:
! " # #
1 1 dz1 dz2N
E = 2 log + 2 lim 2N
· · ·
τ N→∞ N (2πi) z1 z2N
$% √ &
z1√ ± ρ
× log Tr ··· 227
±z1 ρ 1
% √ &'
z2N√ ± ρ
··· . (25)
±z2N ρ 1
right of the face.
The traces never vanish when |zs | < 1, because then
e can pull out phase the matrices in the product correspond to absorbing
efining the complex optical elements, whereas a zero-trace the matrix would
s a 2N-fold contour be elliptic and so represent a transparent medium. A
transfer matrix represents absorption if the eigenvalues
# # of S† S, where S is the scattering matrix relating
dz1 dz2N
··· incoming and outgoing waves, are less than unity,
z1 z2N because then the energy flowing out of the element is
always less than that flowing in. It follows that the
··· (a)
absorption condition is (b)
|T |2 + |R|2 < 1, |T |2 + |R− |2 < 1,
(25) Figure 3. Dots: measurements of logarithmic
(1 − |T |2 − |R|2 )(1 − |T |2 − |R− |2 )
Figure 9:∗ Logarithmic ∗ 2 transmitted intensity across transmitted
stacks of N intensity for stacks
plastic films N plastic
withofmean films (five
thickness (a)
> |T R− + T R | . (26) runs): full curve, best fit to the data; dashed curve,
< 1, because then 0.25 mm; (b) 0.1 mm. Dots: experimental
A short calculation now shows that the logarithm in (25) data. Full curve: best fit to the data. Dashed curve: prediction
predictions of naive ray theory. (a) mean thickness
spond to absorbing of
is incoherent
never singular, transmission (Ohm’sare
and the integrals law).givenReprinted
by their from d̄ =[29]
0.25(courtesy of M.V.
mm (best-fit slope,Berry).
−0.059); (b) d̄ = 0.1 mm
ce the matrix would residues at zs = 0, which vanish, giving the claimed (best-fit slope, −0.046).
sparent medium. A result (17).
n if the eigenvalues canThe benaive
observed ray theoryby illuminating the stackin of
can also be formulated transparencies with a plane light wave (or rather a
terms
of transfer matrices, andofwe can askwave, why the generalthe light7. Experiment
ng matrix relating good approximation a plane namely, from a simple commercial He-Ne laser)
re less than unity, exponential
and recording argument fails, giving instead
the transmission. Fig.the9 peculiar
shows theWe transmission
linear decay (10). The obvious adaptation of (18) gives measured thevs. the number
transmittance N of laser
in He–Ne filmslight
or of
ut of the element is thickness of the sample. It displays a clear exponential stacks decay,
of two one of
different thePPCsignatures
(polyester of Anderson
carbonate) plastic
the ray matrix for a single element as
It follows that the localization, 1and % 2markedly &differs from the linear decay films,(Ohm’s
from a thick
law)folder
predicted d̄ =the
cover (for 0.25 mm) and thin
incoherent
τ − ρ2 ρ viewgraph transparencies (d̄ = 0.1 mm). In each case,
transportmray = of intensities. . (27)
τ −ρ 1
2
< 1, It
ForFigureis important
transparent to
τ +realize
ρ = 1, that
films,measurements so that absorption
mray becomes in thewe cut the films into rectangles and formed these into
transparencies would also result in an expo-
3. Dots: of logarithmic a staircase whose treads were 2.5 mm wide, which was
) nential decay of the transmitted intensity. In
transmitted %
1 τ −intensityρ
&
ρ for stacksρof N
%
−1plastic
1
& films (fiveilluminated normally byexperiments,
all Anderson localization the laser beamit (1 is mmcrucial
wide).
(26) to
mray = full
ensure
runs): thatcurve,absorption
best fit=tois +
1thenegligible,
data; dashedespecially
. (28)
curve, when
The working
transmitted with electromagnetic
intensity was measured waves.
by In
allowing
τ −ρ τ +ρ τ −1 1
he logarithm in (25) this specific
predictions case,
of the
naive bulk
ray absorption
theory.
This a unimodular matrix with the peculiar feature that (a) meancoefficient
thickness of polyester
the beam to carbonate
enter the is known
aperture to be
(diameter negligible
10 mm) of
are given by their d̄ =degenerate
here.
it has 0.25
In mm (best-fit
principle, one slope,
eigenvalues, could −0.059);
also
so that (b) d̄ to
double-check
when raised = 0.1 mmanophotodetector
thethat photon is lost afterbypassing
measuringthrough thethe stack. By
reflection
giving the claimed (best-fit
coefficient
2N th power of slope,
(forthe −0.046).
N sample
films) it and growsverifying
linearly ratherthat thanR + T = 1. In condensed-matter experimentsmeasured
moving the staircase through the beam, we with
exponentially: since the transmittance of N = 0, 1, 2 . . . films.
electrons or cold atoms, number conservation of massive particles makes absorption less
Figure 3 shows the results. It is clear that the relevant
formulated in terms % &
and −1 1to2monitor.
7. simpler
Experiment transmittance decays exponentially as predicted by the
sk why the general =0 (29)
−1 1crucial requirement in the experiment is
Another thattheory,
wave it remains a 1dthe
and that system,
naive rayi.e.,theory
that there
is false.
instead the peculiar the
is We 2N th
a single power isthe transmittance
transverse electromagnetic mode This
involved. isLight
an observation
polarization of localization
is not anofissue light (for
caused
ptation of (18) gives measured % & in He–Ne laser light of
ρ by macroscopic coherence (that is, coherence between
as
perpendicular
stacksm2N ray = 1 + 2N
different−1
of twoincidence, scattering
PPC 1 (polyester
, is independent
carbonate) plastic of polarization), but surface roughness or lack
−1scattering
1 waves whose path differences are large in comparison
of films,
parallelism
from acan thickcause
τ folder cover (d̄ into
= 0.25 othermm) transverse
andwiththinthemodes. This coupling into loss channels
i.e. wavelength). Similar results were obtained with
(27) eventually
viewgraphdestroys localization.
transparencies ρ ( d̄ = Some
0.1 mm). indication
In each of this
case, loss is visible in Fig. 9(b): the decay is
white light.
2N
(m
notwereally
cutraythe)22 = 1 + into
exponential,
films 2N but bent upwards, towards (30) theThe prediction
measured for incoherent
exponents transmission.
τ rectangles and formed these into decay were
o that mray becomes a The
whose staircase whose
experimental
reciprocal treads
reproduces data were
the show 2.5fluctuations
transmission mm (10).wide,This which
of the was transmission for various realizations of the
− log(transmitted intensity)/N
ρ % −1 1 & illuminated
experiment.
result is an example normally
This isofnot bysurprising,
physics the (inlaser thisbeam
on
case the(1 mm wide).
contrary,
wrong according
= 0.059 (thick to Section
films) 2.5, the fluctuations
. (28) areThe
physics)
even transmitted
associated
expectedwith intensity
to degeneracies
be large. wasHowever,
measured
of non-Hermitian by allowing
it turns out that the observed fluctuations are smaller(31)
τ −1 1 matrices; otherto examples areaperture
given in(diameter
Berry (1994).
= 0.046 (thin films).
the predicted:
than beam enter
in thethe plot, the transmission 10 mm) ofshould appear as a cloud of points whose
logarithm
peculiar feature that (When there is absorption, the ray transfer matrix WeByalso measured the average transmitted intensity τ
a photodetector
variance, eqn. (33), after passing
increases likethrough
N , which theisis no
stack.
not of clearly the case. Most probably, this is due
at when raised to the longer degenerate, and the transmitted intensity decays the individual films over 20 runs; the result was
to moving the staircase
the experimental through the mentioned
imperfections beam, we measured
linearly rather than exponentially, but ray theory gives the wrong exponent.)above τ =that0.94couple
± 0.01 for several
the thicktransverse
films, and modes
τ = 0.94and ± 0.05
the transmittance
consequently attenuate = 0,
of Nthe 1, 2 . . . films.
fluctuations.
Figure 3 shows the results. It is clear that the
transmittance decays exponentially as predicted by the
(29) 4.1.3 Fluctuations
wave theory, and that the naive ray theory is false.
AsThis is anemphasized
already observation several
of localization
times, the of light caused
existence of large fluctuations of the transmission
by macroscopic coherence (that is, coherence between
in a characteristic feature of Anderson localization. A key advantage of measuring relative
, waves whose path they
differences
fluctuations is that are notaremuch largeaffected
in comparison
by absorption, which merely induces a global
with the wavelength). Similar results
decay of the whole transmission distribution. were obtained with
Consequently, in the last few years, much progress
white light.
(30) has been made in calculating and measuring fluctuations in diffusive and localized systems.
The measured decay exponents were
nsmission (10). This − log(transmitted intensity)/N
in this case wrong = 0.059 (thick films) 25
es of non-Hermitian = 0.046 (thin films). (31)
n in Berry (1994).
ransfer matrix is no We also measured the average transmitted intensity τ
ted intensity decays of the individual films over 20 runs; the result was
he wrong exponent.) τ = 0.94 ± 0.01 for the thick films, and τ = 0.94 ± 0.05
Figure 10: Microwave intensity transmitted across a copper tube filled with scattering aluminium
spheres vs. the microwave frequency. In the diffusive regime (a) (frequency around 17 GHz), fluctuations
are comparable to the average value. In the localized regime (b), huge fluctuations are visible, a hallmark
of Anderson localization. Reprinted from [30] (courtesy of A. Z. Genack).

Fluctuations provide us with an unambiguous way of characterizing Anderson localization, even

in the presence of absorption.
In order to illustrate this claim, we show in Fig. 10 the transmission of microwaves across
a quasi-1d sample composed of aluminium spheres randomly disposed in a long copper tube
(cooled with liquid nitrogen so that absorption is negligible), as a function of the microwave
frequency [30]. The transport mean free path depends on the resonant scattering cross section
of the aluminium spheres, and thus varies strongly with frequency, implying large changes in the
relative sample length t = L/l. Plot (a) is obtained in the diffusive regime, where the localization
length is longer than the sample size: there, the transmitted intensity fluctuates in an apparently
random way, but the fluctuations are relatively small, the rms deviation being comparable to the
mean. This is expected in the diffusive regime for relatively short samples, where the transmission
amplitude itself is expected to behave like a complex random number, whose real and imaginary
parts are independent, normally distributed variables. In contrast, in the localized regime shown
in plot (b), the fluctuations are much larger, the transmission being most of the time very small
with some rare events of exceptionally high transmission, as predicted in Section 2.5.
Visual inspection reveals immediately that plots (a) and (b) are obtained in different regimes.
While plot (a) has relatively small fluctuations, characteristic of a diffusive regime, where the
fluctuations are comparable to the mean, plot (b) suggests some kind of huge (log-normal) fluc-
tuations, typically associated with the localized or critical regime. The take-home message here
is: don’t rely solely on exponential decay to prove the existence of localization, look also at
the fluctuations, they are better indicators. Since the relative fluctuations are insensitive to
moderate absorption, they may even provide a quantitative criterion whether the strong local-
ization threshold has been reached or not, and this under circumstances when the exponentially
decreasing transmission alone could not be a reliable signature [31, 32].

4.1.4 The Anderson model: a free (numerical) experiment

Although the Anderson model was originally introduced as a tight binding model for electrons in
a disordered crystal, it is of broader interest and has become a paradigm for one-body localization
effects.
Up to now, we have considered continuous models where a wave propagates along a continuous
1d axis, encountering a set of discrete objects that scatter the wave backward and forward. The
transfer-matrix game is just to combine the discrete scatterers with the proper phases. One can

26
even go one step further, disregard the ballistic propagation, and build a completely discrete
model, where the wave lives on a 1d lattice. One of the simplest discrete models is certainly the
Anderson model whose Hamiltonian is
+∞
X
H= (wn |nihn| + t|nihn + 1| + t|n + 1ihn|) (64)
n=−∞

where the state |ni is the occupation amplitude of site n with wn its on-site energy. t is the so-
called “tunneling” matrix element coupling neighboring sites, traditionally taken with numerical
value t = 1. In a concrete physical realization, typically also the t’s are random variables and
thus define “off-diagonal disorder”. However, it is enough to take diagonal disorder to observe
localization. The precise value of t is irrelevant, as long as it is not zero, and may be used
to define the energy scale of the problem. If wn = 0, it is easy to check that the eigenstates
are discrete plane Bloch waves ψn = exp ikn, for k ∈ [−π, π[, and the energy is given by the
dispersion relation E(k) = 2 cos k of this single-band model.
Disorder is introduced by allowing the on-site energies wn to be random variables. The
standard choice is to take wn to be uncorrelated random variables uniformly distributed in
the interval [−W/2, W/2], with W 2 = 12 wn2 measuring the disorder strength. A noteworthy



property—simplifying analytic calculations—is that the spatial correlation length of the disorder
is zero (see Sec. 5.1.4 for a discussion of spatial correlations arising in optical speckle potentials).
Actually, the 1d tight-binding Anderson model can be solved for almost any distribution [33], with
simple closed expressions for the Cauchy-Lorentz on-site distribution. The techniques developed
in Sec. 5 permit to show that the localization length is at lowest order in W given by
4 sin2 k 12(4 − E 2 )
ξloc = = . (65)
hwn2 i W2

Exercise 4 – Few properties of the Anderson model.

(i) Show that the equations obeyed by an eigenstate of the Anderson model at energy E can be put
in the following matrix form:    
ψn+1 ψn
= Tn (66)
ψn ψn−1
with a transfer matrix:  
E − wn −1
Tn = . (67)
1 0
Show that the amplitudes of the left and right propagating plane waves in a disorder-free region
can be expressed as simple linear combinations of ψn and ψn+1 . Show that, in consequence, it
is possible to construct a transfer matrix M as in eq. (9). This shows that the general results of
section 2.2 can be used and that exponential localization is expected.
(ii) Consider a continuous model of a 1d particle in a disordered potential V (z). By discretizing the
Schrödinger equation on a lattice with sufficiently small spacing (much shorter than the de Broglie
wavelength and the correlation length of the potential), show than one recovers the Anderson
model, but with spatially correlated wn .

Numerical simulations of the Anderson model are extremely easy, at least in dimension 1.
Indeed, the previousP
exercise shows that the time-independent Schrödinger equation reduces, for
an eigenstate |ψi = n ψn |ni with energy E, to the three-term recurrence relation
ψn+1 + (wn − E)ψn + ψn−1 = 0 (68)
which can be solved recursively. In Fig. 11, we give an example of a simple script, written in the
Python language, that solves this equation at some arbitrary energy across a random sample of
arbitrary length.

27
# !/ usr / bin / python
from __future__ import print _functi on
import math
import random
import sys
# compute_AL1D . py
# Authors : Dominique Delande and Cord A . Mueller
# Release date : May 25 , 2016
# License : GPL3
# -----------------------------------------------------------------------------------------
# This script models localization in the 1 D Anderson model with box disorder ,
# i . e . uncorrelated on - site energies w_n uniformly distributed in [ - W /2 , W /2].
# The script computes the wave function and resulting transmission as function of energy ,
# system size L , and for nr number of realizations .
# Without disorder , the dispersion relation is energy =2* cos ( k ) , with support on [ -2 ,2].
# The equations to be solved are : psi_ { n +1} + psi_ {n -1} + ( w_n - E ) psi_n = 0
# They are solved in the backward direction , starting from an outgoing wave of unit modulus ,
# with normalization to unit incident flux at the end of the computation .
# Optionally , either the values of - log ( T ) are printed in the file logT . dat
# or the intensities | psi_n |**2 are printed in the file logpsi2 . dat .
# The localization length is 12*(4 - energy **2)/ W **2 to lowest order in W .
# -----------------------------------------------------------------------------------------
if len ( sys . argv ) != 6:
print ( ’ Usage (5 parameters ):\ n compute_AL1D . py L W energy nr savepsi ’)
sys . exit ()
# switch savepsi : ’1 ’ for saving log | psi_n |^2 ( with a small number of realizations )
# ’0 ’ for saving - logT ( with a large number of realizations )
L = int ( sys . argv [1])
W = float ( sys . argv [2])
energy = float ( sys . argv [3])
nr = int ( sys . argv [4])
savepsi = int ( sys . argv [5])
if savepsi == 1:
filename = ’ logpsi2 . dat ’
elif savepsi == 0:
filename = ’ logT . dat ’
else :
sys . exit ( " Please choose between ’1 ’ for saving log | psi |^2 or ’0 ’ for saving - logT " )

k = math . acos (0.5* energy )

exp_i_k = complex ( math . cos ( k ) , math . sin ( k ))
psi =[0.0]*( L +1)

def computepsi ():

psi [ L ] = exp_i_k
psi [L -1] = 1.0
for n in range (L -1 ,0 , -1):
w_n = W *( random . random () -0.5)
psi [n -1] = ( energy - w_n )* psi [ n ] - psi [ n +1]
return psi

def minuslogt ( psi ):

psi_minus_1 = energy * psi [0] - psi [1]
incident = (0.5* abs ( psi_minus_1 - exp_i_k * psi [0])/ math . sin ( k ))**2
return math . log ( incident )

f = open ( filename , ’w ’)
for j in range ( nr ):
psi = computepsi ()
if savepsi == 1:
logpsi2 = [2* math . log ( abs ( x )) for x in psi ]
for n in range ( L ):
f . write ( ’ %6 d %12.8 g \ n ’ % (n , logpsi2 [ n ] - logpsi2 [0]))
f . write ( ’\ n \ n ’)
else :
f . write ( ’ %6 d %12.8 g \ n ’ % (j , minuslogt ( psi )))
f . close ()
print ( " Done , result saved to " , filename , " ! " )

Figure 11: Python script for solving the 1D Anderson model, Eq. (68), also available at http://
www.lkb.upmc.fr/Anderson-localization-in-a-one. The script is run by in a shell by typing python
compute AL1D.py L W E NR svpsi where the parameters L,W,E,NR,svpsi are respectively the system size,
the disorder strength, the energy, the number of disorder realizations, and the switch between an output of
log |ψn |2 or − log T , respectively. Typical results are exemplified in Fig. 12.
 Note that the boundary condition used, ψN = 1 and ψN +1 = eik , describes a purely outgoing
wave with wavevector k and amplitude 1 on the right end of the sample. The boundary condition
on the left is actually more complicated, because there the incident wave interferes with the
reflected wave, whose amplitude depends on the microscopic realization of disorder of the entire
sample. The Schrödinger recursion equation is thus better solved backwards from the far end of
the sample, yielding on average an exponentially increasing solution toward the left. This is in
agreement with the fact, shown in exercise 2(iv) above, that the two eigenvalues of the transfer
matrix are of the form λ± = e±2x . So starting with this boundary condition and an arbitrary
value of k (and thus E), one has, with probability one, a finite overlap with the eigenvectors of
the larger eigenvalue and therefore numerically picks up an exponentially growing solution. This
solution is then at the same time physically acceptable for the transmission experiment, viz.,
decreasing on average exponentially from left to right. Since eq. (68) is linear, one can always
normalize the solution to unit incoming flux and thus finally find the transmission probability as
the outgoing flux on the right side, calculated as a linear combination of ψN and ψN +1 (cf. the
Python script in Fig. 11 and exercise 4(i)).
Figure 12 (left plot) shows the intensity ln |ψn |2 for three different realizations of the disorder
at energy E = 0.5 and disorder strength W = 0.6 for a moderately large sample of 1250 sites,
corresponding to a length L = 10ξloc . Although the decay is on the average exponential, huge
fluctuations from one realization to another are visible with the naked eye. Moreover, even for this
pure transmission experiment from left to right, the intensity is not monotonously decreasing
at all, sometimes increasing by factors larger than 10. The histogram of the extinction (or
transmission logarithm) over 10000 realizations is shown in the right plot. Its width clearly
visualizes the huge fluctuations characteristic for the localized regime. The agreement with the
predicted distribution function, eq. (36), shown with a red full line, is excellent. One also observes
the convergence toward the truncated normal distribution (dashed), implying the truncated log-
normal distribution (39) for T itself.
The reader is strongly encouraged to play with this script to experiment personally with
Anderson localization. We recommend the following numerical experiments:

Exercise 5 – Numerical study of 1d Anderson localization

(i) Use a single realization of the disorder and look at the wavefunction (or rather |ψ|2 ) inside the
medium. Try different sample lengths and different energies, but avoid the band center E = 0.
Indeed, the Anderson model is singular at this value. There is still Anderson localization, but the
localization length slightly differs from eq. (65), a so-called Kappus-Wegner singularity [33].
(ii) Using a few hundred or thousand realizations, compute the statistical distribution of the transmis-
sion. Compare with the exact prediction eq. (36) as well as with a truncated normal distribution.
(iii) Modify the script, for example for a Gaussian or Cauchy distribution of disorder and run additional
numerical experiments. You may also introduce correlated disorder to simulate e.g. cold atoms in
a speckle potential (see [20] for generation of a realization of the disorder with proper correlation
functions).

A slightly different approach consist in diagonalizing the Hamiltonian for a large system
numerically. Choosing strict boundary conditions on both ends of the sample yields normalizable
eigenstates, centered at random positions within the sample and decreasing from there in both
directions (similar to Fig. 8(a)), together with the discrete set of corresponding eigenvalues. In
the thermodynamic limit, these eigenenergies form a dense, but still pure-point spectrum.
An important message to keep in mind is that there is very little difference between a con-
tinuous and a discrete system, as far as localization on large spatial scales is concerned. For
example, the exercise shows how a particle in a continuous 1d random potential (e.g., an optical

29
 0.1
0

F (− ln T )
0.08
-5
ln |ψ|2
-10 0.06

-15
0.04
-20
0.02
-25

-30 0
0 2 4 6 8 10 0 5 10 15 20
z/ξloc − ln T

Figure 12: Results of numerical experiments on the 1d Anderson model, with uncorrelated uniform
distribution of disorder. The left plot shows the intensity |ψ(z)|2 inside the disordered medium—on a
logarithmic scale—for three different realizations of the disorder. Note the overall exponential decrease,
decorated by huge fluctuations: the transmission across a sample of size L = 10ξloc fluctuates by more
than 3 orders of magnitude. This illustrates why the typical transmission differs from the average one.
The right plot shows the full probability distribution F (− ln(T ), t) (histogram over 10000 realizations)
for t = z/ξloc = 10, together with the prediction eq. (36) (full red line), which is close to a truncated
Gaussian (dashed).

speckle) can be mapped to a variant of the Anderson model. This universality of the Anderson
model cannot really surprise because localization is an asymptotic property taking place at large
distance; whether the underlying configuration space is discrete or continuous plays only a minor
role.

4.2 d=2
Scaling theory predicts d = 2 to be the lower critical dimension for Anderson localization. In
dimension d = 2 +  (which can be numerically studied by constructing an Anderson model on
a fractal set), a critical point should exist where β(gc ) = 0, separating a diffusive phase from an
insulating one. Strictly at d = 2, scaling theory predicts localization provided there is a weak
localization correction with c2 > 0, see section 3.6. We will see in section 7 that this is indeed
what a microscopic approach predicts in spinless time-reversal invariant systems. Scaling theory
does not pretend to be an exact theory, there is thus a real interest in knowing whether there is
localization in 2 dimensions for specific systems.
Experiments with cold atoms are expected to be much more difficult than in 1d. Indeed, the
localization length, eq. (56), is predicted to increase exponentially with the parameter kl, instead
of linearly in 1d. Detailed theoretical studies [34] have shown that experimental observation
requires at the same time a speckle potential with a very short correlation length (comparable
to what has been done in 1d, but in 2 directions) and a long atomic de Broglie wavelength, that
is very cold atoms. Altogether, satisfying all conditions is far from easy, making 2d Anderson
localization of ultra cold atoms an interesting challenge.
A metal-insulator transition has been observed for electrons in clean semiconductor sam-
ples [35]. It is generally acknowledged that the Coulomb electron-electron interaction—much
stronger than the atom-atom interaction in a dilute cold atomic gas—plays a major role in this
transition, which is thus qualitatively different from the pure Anderson transition and sometimes
referred to as the Mott-Anderson transition [36].
Other types of waves have been successfully used in 2d systems. For example, using con-

30
Figure 13: Experimental results on the propagation of light across a transversally disordered 2d lattice of
photonic wave guides, mimicking the 2d Anderson model. As the strength of the disorder is increased, the
dynamics evolves from ballistic (a and c) to diffusive (b and d), characterized by a Gaussian shape of the
wavepacket, and eventually to Anderson localization, with a wavepacket of characteristic exponential
shape (e), when the disorder is sufficiently strong to make the localization length comparable to the
extension of the wavepacket. Reproduced from [37] (courtesy of S. Fishman).

veniently engineered optical fibers, one can create a 2d “photonic lattice” composed of parallel
optical guides along which the light can freely propagate. Thanks to the photorefractive mate-
rial used, its index of refraction can be adjusted by an external light source. Also the transverse
coupling between the optical guides can be adjusted at will, as well as the disorder due to small
variations of the index of refraction in each guide. As the light propagates at roughly constant
velocity along the guides, the spatial propagation mimics the temporal evolution of the Anderson
model, each guiding mode playing the role of a site. Using such a device, the evolution from
ballistic motion (on a scale shorter than the mean free path) to diffusive motion and eventually
to strong localization has been experimentally observed [37], see Fig. 13.
The Anderson model itself, described in section 4.1.4, can be trivially extended to any dimen-
sion by adding hopping terms to nearest neighbors in a (hyper)cubic lattice. The numerical study
is slightly more difficult than in 1d. The basic idea is to study first the quasi-1d propagation
on a strip with a fixed number M of transverse sites, imposing for example periodic boundary
conditions along this direction. One can write a 2M × 2M transfer matrix for this quasi-1d
system and calculate its asymptotic properties as the length N goes to infinity, extracting the
quasi-1d localization length ξloc (M ). Next, one studies the behavior of ξloc (M ) as M is sent to
infinity. If ξloc (M ) diverges without bounds, one concludes that the system is not localized. If on
the other hand ξloc (M ) tends to a finite limiting value, one concludes that the system localizes
with ξloc = limM →∞ ξloc (M ).
Powerful numerical techniques, such as finite-size scaling [38], make it possible to extrapolate
properties of the infinite system from numerical experiments on limited systems. Especially, the

31
Figure 14: Scaling function β(g) reconstructed from numerical simulations of the Anderson model. The
solid line is for dimension 1, the triangles for dimension 2 and the crosses for dimension 3. Different sets of
microscopic parameters produce data lying on the same curve, which can be considered an “experimental
proof” that the scaling approach is valid. In 2d, β(g) is always negative, proving that the system is in
the localized regime. In 3d, depending on the disorder strength, the system may be localized (β(g) < 0,
strong disorder) or diffusive (β(g) > 0, weak disorder). Reprinted from [39] (courtesy of A. McKinnon
and B. Kramer).

scaling function β(g) can be reconstructed, see Fig. 14. The fact that various data, computed
for various values of the system parameters (energy, disorder strength, system size), lead to the
very same β(g) strongly indicates that the scaling approach is valid, and thus corroborates the
existence of universal properties independent of the microscopic details. In 2d, the numerically
computed β(g) is always negative, as expected and its shape is in good agreement with the naive
prediction, eq. (53).

4.3 d=3
Dimension 3 is arguably the most interesting, because scaling theory there predicts a transition
between diffusive behavior for small disorder and Anderson localized behavior at large disorder.
Consequently, much experimental and numerical effort has been spent to observe this Anderson
transition. Numerical simulations of the 3d Anderson model are a very valuable tool, especially
to locate the critical point where β(gc ) = 0 and to characterize its vicinity. The results in Fig. 14
very clearly show the existence of the two regimes and the fact that β(g) behaves smoothly across
the transition. This constitutes a clear-cut proof that the Anderson transition is a continuous
phase transition of second order. Note the absence of data for g just below gc ; this corresponds
to localized systems with a localization length too large to be reliably measured in the numerical
simulations. As shown in Sec. 3.7, the slope dβ/d ln g|gc at the critical point is the inverse of the
critical exponent ν of the Anderson transition. Although the slope at the critical point cannot
be accurately measured on these data, it is without any doubt smaller than unity—the value
of the asymptotic slope in the deep localized regime ln g → −∞. This implies that the critical
exponent ν is larger than unity. Recent numerical studies on much larger systems fully confirm

32
 n ( ´ 10 16 /cm 3 )
2.4 2.6 2.8 3.0

0.4
3
H=0 H=4T
0.2
s (S/cm)
2
0
0 0.02
n/nc-1
H=0
1
T®0

Al 0.3 Ga 0.7 As

0
520 560 600 640
Total Exposure Time (units)

Figure 15: Experimentally measured conductivity of a Si-doped AlGaAs 3d crystal vs. electron con-
centration (upper horizontal scale), showing a clear metal-insulator transition. The critical exponent is
very close to unity, a value not compatible with the universal value ν = 1.58 of the pure 3d Anderson
transition. Electron-electron interaction is probably responsible for the difference. Adapted from [40]
(courtesy of S. Katsumoto).

this point, the current best estimate being ν = 1.58 ± 0.01 [26, 27].
Direct experimental observation of Anderson localization in 3d is even more difficult than
in 2d, because it requires an even more strongly scattering system (kl smaller than 1 from the
Ioffe-Regel criterion, eq. (57), instead of kl of the order of few units). Moreover, creation of
a sufficiently disordered potential can be technically much more difficult in 3d: for a speckle
potential, this would require to send plane waves with random phases from a large solid angle.
Thus, Anderson localization of atomic matter waves in a disordered potential has not yet been
observed. However, using the equivalence of a quasi-periodically kicked rotor with a 3d Anderson
model, the Anderson transition with atomic matter waves has been observed, and its critical
exponent experimentally measured, as discussed in section 8.
Electronic transport in solids, the field where localization theory was originally developed,
provides also interesting experimental results. Metal-insulator transitions can be observed in
solid state samples, but it is never easy to identify the microscopic mechanism. This is because
electron-electron interactions play an essential role. Whether the observed transition is a one-
body effect like the Anderson transition or a many-body one like the Mott transition [41] is
not easily proved. We are not aware of any unambiguous observation of the pure Anderson
transition. Figure 15 shows the experimentally measured conductivity of a Si-doped AlGaAs
crystal vs. a parameter essentially representing the electronic Fermi energy. A clear insulator-
to-metal transition is observed. It seems that the curve is almost linear in the metallic regime,
which—because conductivity is essentially a measure of the diffusion constant —means that

33
Figure 16: Experimentally measured transmission of microwaves through a 3d strongly scattering
medium vs. the size of the medium (doubly logarithmic scale). In the usual diffusive regime (open
symbols), a 1/L decrease is observed, in agreement with classical transport theory (Ohm’s law). At the
critical point of the Anderson transition (filled symbols) a characteristic 1/L2 behavior is observed, in
agreement with the scaling theory of localization. Note that, because of residual absorption, the signal
drops at large size. Reprinted from [46] (courtesy of A.Z. Genack).

the measured critical exponent is ν ≈ 1. This markedly differs from the exponent of the pure
Anderson transition, indicating that interaction effects are probably important.
One may also turn to other type of waves, for example ultrasonic waves [42] or electromag-
netic waves. Direct measurement of the electromagnetic field inside the disordered medium is
not straightforward, and transmission experiments are easier. As mentioned earlier, absorption
induces an exponential decay of the intensity, which must be carefully discriminated from the
same effect being produced by Anderson localization. Thus, experimentalists have turned to
measuring tell-tale properties right at the critical point. There, according to scaling theory, the
dimensionless conductance has the constant value gc , independently of the system size, whereas
the classical dimensionless conductance, eq. (45), increases linearly with the system size L. This
additional power of L makes the total transmission across the sample evolve from a 1/L behavior
(Ohm’s law) in the diffusive regime to a 1/L2 scaling law at the critical point, and eventually
to the exponential decay in the localized regime. Any spurious absorption is likely to transform
the critical 1/L2 behavior into an exponential decrease. Thus, the existence of an 1/L2 may be
considered a sensitive test of observing the Anderson transition. Fig. 16 shows the experimental
result obtained on the propagation of microwaves in a disordered medium, in the diffusive and
critical regimes. The existence of a range with 1/L2 power law—before absorption wins at even
larger size—is a convincing proof.
Similar results have been obtained in the optical regime [43], where strong scattering is
provided by oxide powders, but the role of absorption has been discussed controversially [44].
Recently, time-resolved transmission experiments, where absorption has less impact, have shown
a slowing down of classical transport [45], which gives strong evidence for Anderson localization.
In the last few years, several numerical and laboratory experiments have characterized the
fluctuations appearing in the vicinity of the Anderson transition. In particular, numerical exper-
iments on the 3d Anderson model have shown that the critical eigenstates have a multi-fractal
structure, implying the coexistence of regions where the wavefunction is exceptionally large
together with regions where it is exceptionally small. This is presently a very active field of
research [47], whose description is beyond the scope of these lectures. The reader may refer to
the recent review paper of Evers and Mirlin [15]. In the near future, it is very likely that exper-

34
iments on localization of atomic matter waves will concentrate on the existence and properties
of fluctuations.

5 Microscopic description of quantum transport

5.1 Diagrammatic perturbation theory
In section 2, we have seen that localization in (quasi-)one-dimensional systems can be very
efficiently described by a transfer-matrix approach. In higher dimensions, we have resorted
to the scaling arguments presented in section 3. We now wish to give an introduction to a
microscopic description of quantum transport in disordered systems. The main advantage of
diagrammatic perturbation theory lies in its versatility. It applies in arbitrary dimensions d and
to any model with a Hamiltonian of the form

H = H0 + V, (69)

in which H0 describes regular propagation in an ordered substrate, and V is the disorder po-
tential that breaks translational invariance. On a microscopic level, “disorder” refers to degrees
of freedom whose detailed dynamics are not of interest and whose properties are only known
statistically. In the following, we will consider static or quenched disorder that remains frozen
on the timescale of wave propagation under study (as sole exception of this rule, we mention in
section 6.3 the dephasing effect of moving impurities).
A first model of type (69) describes a single quantum particle in an external potential,

p2
H= + V (r), (70)
2m
with direct bearing on experiments with non-interacting matter waves like [4], but equally ap-
plicable to other massive particles like electrons, neutrons, etc. Note that the plain Hamiltonian
(70) operating in Hilbert space describes the same single-particle physics as the more fanciful
many-body version
~2 2
Z  
H = dd rΨ† (r) − ∇ + V (r) Ψ(r), (71)
2m
defined in terms of particle creation and annihilation operators Ψ(†) in Fock space.
Since we assume that H0 is translation-invariant (if only by discrete translation on a lattice),
the following equivalent formulation in Fourier space is also useful:

ε0k a†k ak + Vq a†k+q ak .

X X
H= (72)
k k,q

Here, wave vectors k are used as good quantum numbers labeling the eigenstates of H0 . If there
is an underlying lattice, one has to include also a Bloch band index. ε0k is Rthe free dispersion
relation; for matter waves, ε0k = ~2 k 2 /2m. Particle annihilators ak = L−d/2 dd reik·r Ψ(r) and
creators a†k fulfill the canonical commutation relations [ak , a†k0 ]± = δkk0 .
The disorder potential breaks translation invariance by scattering R d particles k → k 0 with an
−d −iq·r
amplitude given by its Fourier component Vq = hk + q|V |ki = L d re V (r), conveniently
represented by
q
Vq = k k+q . (73)

35
 This model Hamiltonian (72) is not limited to matter waves. By appropriate changes in ε0k ,
different physical systems can be described. For instance, photons and other mass-less excitations
have a linear dispersion ε = ~ck with characteristic speed c. Yet another realization is provided
by thepelementary excitations of Bose-Einstein condensates, featuring the Bogoliubov dispersion
εk = ε0k (ε0k + 2µ), that interpolates between a linear sound-wave dispersion at low energy and
a quadratic particle-like dispersion ε0k = ~2 k 2 /2m at high energy. The general formalism to be
introduced below applies to all these cases, provided the scattering potential Vq is known.
The basic model can be made richer, depending on the circumstances and effects one wishes
to describe. For example, spin often plays an important role, for instance via spin-orbit effects,
due to coupling of spin and direction of propagation. Also spin-flip processes can be of interest,
as in electronic spin-flips induced by magnetic impurities or photon polarization flips induced
by Zeeman-degenerate atomic dipole transitions. A typical spin-flip process (m, σ) 7→ (m0 , σ 0 )
changes the spin of the propagating object from σ to σ 0 , while the impurity spin undergoes
m 7→ m0 . Processes of this type can store information about the path traveled, and generally act
as a source of strong decoherence (as discussed in Sec. 6.3 below).
The main drawback of the diagrammatic Green function approach is its perturbative char-
acter. Most results are obtained from an expansion in powers of V and are valid only for
small enough potential strength. If one is interested in truly strong-disorder effects, it may be
worthwhile to start from the opposite situation where the propagation described by H0 is small;
Anderson’s original method of a “locator expansion” [5] is an example for such a weak-coupling
perturbative approach. In any case, it takes considerable effort to derive non-perturbative re-
sults using controlled approximations. Yet, the basic diagrammatic technique is a prerequisite for
more powerful, field-theoretic methods involving, e.g., replica methods, renormalization-group
analysis and supersymmetry [48].

5.1.1 Quantum propagator

Let us then start by calculating the Green function for the single-particle Hamiltonian (70) that
determines the time evolution of a state |ψi in Hilbert space according to the Schrödinger equation
i~∂t |ψi = H|ψi. For t > 0, the forward-time evolution operator GR (t) = − ~i θ(t) exp{−iHt/~}
solves the differential equation
[i~∂t − H] GR (t) = δ(t). (74)
Obviously, GR (t) is the retarded Green operator for the Schrödinger
Dh equation.
iE It encodes the
R i †
same information than its many-body version Gkk0 (t) = − ~ θ(t) ak (t), ak0 that one would
use starting from (72); in the following, we stick to the simpler form, referring the reader to the
literature for the more advanced presentation [49, 50, 51, 52]. Going from time to energy by
Fourier transformation, one defines the resolvent
Z
−1 −1
G (E) = lim+ dtei(E+iη)t/~ GR (T ) = lim+ [E − H + iη] =: [E − H + i0] .
R
(75)
η→0 η→0

The limiting procedure η → 0+ guarantees that indeed the retarded Green operator is obtained,
different from zero only for t > 0. The advanced Green operator GA (t) is obtained by taking
η → 0− . In the basis where H is diagonal, H|ni
P = εn |ni, the resolvent is also diagonal and thus
admits the spectral decomposition G(z) = n |ni[z − εn ]−1 hn| for any argument z ∈ C outside
the spectrum of H. The resolvent’s matrix elements are called “propagators”. For example, in
the position representation hr|ni = ψn (r),
X ψn (r0 )ψ ∗ (r)
GR (r, r0 ; E) = hr0 |GR (E)|ri = n
= r r0 . (76)
n
E − εn + i0

36
This propagator contains precious information: As function of E, it has singularities on the real
axis that are precisely the spectrum of H and thus encode all possible evolution frequencies.
Furthermore, the residues at these poles provide information about the eigenfunctions.
The total Hamiltonian (70) contains the disorder potential so that we cannot write down
its eigenfunctions and eigenvalues analytically (numerically, one may of course calculate eigen-
functions and eigenvalues for each realization of disorder). We start therefore with the free
Hamiltonian H0 . Its resolvent G0 (z) = [z − H0 ]−1 is diagonal in momentum representation,
hk 0 |G0 (z)|ki = δkk0 G0 (k, z) with
1
GR
0 (k; E) = = . (77)
E − ε0k + i0 k

Now we are ready to describe the perturbation due to the potential V : Using G(E) = [E − H0 −
V ]−1 = [(E − H0 ){1 − (E − H0 )−1 V }]−1 , we express

G(E) = [1 − G0 V ]−1 G0 (78)

= G0 + G0 V G0 + G0 V G0 V G0 + . . . (79)

as the Born series in powers of V . For notational brevity, we have already suppressed the energy
argument on the right-hand side. Still, if one tries to write out a matrix element hk 0 |G(E)|ki, the
operator products convert into cumbersome expressions that tend to obscure the series’ simple
structure:
X
hk 0 |G(E)|ki = δkk0 G0 (k) + G0 (k 0 )Vk0 −k G0 (k) + G0 (k 0 )Vk0 −k00 G0 (k 00 )Vk00 −k G0 (k) + . . . (80)
k00

At this point, we are well advised to use the graphical representation known as “Feynman
diagrams”, for which we have already all ingredients at hand:

hk 0 |G(E)|ki = δkk0 + + + ... (81)

k k k0 k k00 k0

Already, we achieve a much more compact notation, aided by the fact that we do not need to label
the dangling impurity lines, defined in (73), since their momentum is automatically determined
by the incident and scattered momenta. Also, we henceforth use the prescription that all internal
momenta have to be summed over, here k 00 in the last contribution.

5.1.2 Ensemble average

In principle, the Born series (81) permits to calculate the full propagator perturbatively. However,
the result will be different for each realization of disorder. We are really only interested in suitable
expectation values and thus have to understand how to perform the ensemble average over the
disorder distribution.
The potential V (r) as a function fluctuating in space is a random process. As such, it can
be completely characterized by its moments or correlation functions hV1 i, hV1 V2 i, hV1 V2 V3 i,
etc., with the short-hand notation Vi = V (ri ). We will assume that the process is stationary
or, preferring the spatial dictionary, statistically homogeneous, which means that correlation
functions can only depend on coordinate differences rij = ri − rj . We can therefore define the

37
following correlation functions and corresponding diagrams:

hV1 i = hV i (82)
hV1 V2 i = P (r12 ) = (83)
1 2

hV1 V2 V3 i = T (r12 , r23 ) = (84)

1 2 3

and so on for arbitrary n-point correlation functions. Without loss of generality, one may always
take hV i = 0 by defining a centered potential V 7→ V − hV i while redefining the zero of energy
E − hV i 7→ E. In Fourier representation, these correlation functions are

q q q0

P (q) = , T (q, q 0 ) = , etc. (85)

Depending on the specific type of disorder, these general correlation functions can take different
forms, and it may be instructive to discuss two of them in detail.

5.1.3 Gaussian disorder

As a first example, let us consider Gaussian-distributed disorder that is completely defined by
its first two moments
hV i= 0 and P (r) = V02 C(r). Here, one conveniently factorizes the one-
point variance V02 = V12 from the spatial correlation function C(r) that obeys C(0) = 1 by
construction. The characteristic property of a Gaussian process is that all higher-order correlation
functions completely factorize into pair correlations. Indeed, a simple property of the Gaussian
integral

implies
thatthe moments of a normally distributed, centered scalar random variable X
n
are X 2n = Cn X 2 where Cn = (2n)!/(2n n!) is the number of pairs that can be formed out


of 2n individuals. Similarly, the Gaussian moment theorem applies to a Gaussian-distributed

random potential:
1 X


hV1 · · · V2n i = Vπ(1) Vπ(2) · · · Vπ(2n−1) Vπ(2n) (86)
2n n! π

where π denotes the (2n)! permutations. Pictorially, this implies also a complete factorization of
2n-point potential correlation into products of pair correlations. The first interesting example is
n = 2 with

= + + (87)

and so on for higher orders.

Such a Gaussian potential can be constructed with arbitrary spatial correlation C(r). A
popular choice here is often to model it as a Gaussian as well, C(r) = exp{−r2 /2σ 2 }, such as
in [53], because this is easy to implement numerically (it suffices to draw uncorrelated random
variables Vi on a discrete grid and convolute by a Gaussian correlation function afterwards).
Moreover, this choice leads to simple analytical calculations because also the k-space pair cor-
relator is Gaussian, P (q) = V02 σ d (2π)d/2 exp{−q 2 σ 2 /2}. In the limit of low momenta qσ  1,
the potential details cannot be resolved and it appears δ-correlated. Then, everything can be
expressed in terms of P (0) = (2π)d/2 σ d V02 .

38
5.1.4 Speckle
A slightly more interesting example is provided by the optical speckle potential used recently for
matter-wave Anderson localization [4, 54]. The atoms are subject to an optical dipole potential
V (r) = K|E(r)|2 created by the local field intensity of far-detuned laser light. K contains the
frequency-dependent atomic polarizability besides some constants [55]. With a laser beam that
is blue-detuned from the optical resonance, one has K > 0 and thus expells atoms from high-
intensity regions. This potential landscape features repulsive peaks with hV i > 0. Conversely, a
red-detuned laser leads to K < 0, and one finds a potential landscape with attractive wells and
hV i < 0. To create a disorder potential, the laser beam is focused through a diffuse glass plate,
whose randomly positioned individual grains act as elementary sources for the emitted field. The
electric field E(r) at some far point then is the sum of a large number of complex amplitudes. By
virtue of the central limit theorem, it is a complex Gaussian random variable with normalized
pair correlator
hE ∗ (ri )E(rj )i
γij = γ(ri − rj ) = = ∗ (88)
h|E|2 i i j

∗
with the obvious properties γij = γji and γii = 1. In a 1d-geometry, the pair correlator takes its
simplest form in Fourier components:

γ(q) = πζΘ(1 − |q|/kζ ) (89)

where kζ = kα is the maximum wave-vector that can be built from a monochromatic laser source
with wave vector k seen under an optical aperture α. This eqn (89) simply says that the random
field contains all wave vectors inside the allowed interval with equal weight. In real space, this
Fourier transforms to γ(r) = sin(r/ζ)/(r/ζ) with the correlation length ζ = 1/kζ = 1/(αk).
Other pair correlations such as Ei Ej and Ei∗ Ej∗ have uncompensated random phases and
average to zero. The Gaussian moment decomposition now applies to arbitrary moments of
the speckle disorder potential Vi = KEi∗ Ei . An n-point potential correlation is really a (2n)-
field correlation, which decomposes into all possible pair correlations (88). As in a conventional
ballroom dancing situation involving n couples, all possible heterosexual pairings between the
Ei∗ s and Ej s are allowed. This gives for the 2-point potential correlator
2
hV1 V2 i = K 2 hE1∗ E1 E2∗ E2 i = hV i [γ11 γ22 + γ12 γ21 ] . (90)

2 2
Setting r1 = r2 shows that V = 2hV i , which means that the potential variance is equal to

2 2 2
its mean square, var(V ) = V − hV i = hV i .
The shift V 7→ V − hV i to the centered potential removes the first term in the bracket in
(90). The same applies to all diagrams with field self-contractions:

~ = 0. (91)

So henceforth, we can neglect those diagrams by considering a centered potential hV i = 0.

Altogether, we have as a first building block the speckle potential pair correlator

hV1 V2 i = P (r12 ) = V02 C(r12 ) = V02 ~ ~ (92)

1 2

Here, the potential strength V02 = var(V ) is factorized from the dimensionless correlation function
C(r) = |γ(r)|2 that is normalized to C(0)=1. In d = 1, from (89), we have the real-space intensity
correlator C(r) = [sin(r/ζ)/(r/ζ)]2 . In higher dimensions and in an isotropic setting, the Fourier

39
transformation of the simple k-space field correlator yields C(r) = [2J1 (r/ζ)/(r/ζ)]2 in d = 2
and C(r) = [sin(r/ζ)/(r/ζ)]2 again in d = 3 [34].
An interesting effect occurs for potential correlations of odd order (2n + 1). They are really
field correlation of twice the order, which is even and thus different from zero. The first example
of this kind is (since the fields ∗ and ◦ will always appear together, we note ~ = • from now on)

hV1 V2 V3 i = V03 2Re{γ12 γ23 γ31 } = V03 (93)

1 2 3

Diagrams of this type can only contain closed loops of field correlations (because field self-
contractions no longer appear). Since the loops can be closed both clockwise and counterclock-
wise, there are two contributions complex conjugate of each other.

5.1.5 Average propagator: self-energy

Now we are in position to take the ensemble average of the single-particle propagator (79):
hGi = G0 + G0 hV G0 V iG0 + G0 hV G0 V G0 V iG0 + . . . (94)
or
hGi = + + + ... (95)
The precise form of potential correlations depends on the model of disorder. As shown by the
example of the Gaussian model (87), starting from the fourth-order term there appear completely
factorized contributions. Before writing all possible combinations down, we had better introduce
one of the cornerstones of diagrammatic expansions: the self-energy Σ(E) defined by the Dyson
equation
hGi = G0 + G0 ΣhGi. (96)
Introducing the self-energy invariably prompts the following frequently asked questions:
1. Why is the self-energy convenient for perturbation theory?
2. How do I calculate Σ?
3. What is the physical meaning of Σ?
4. Is there a simple example?
Let us answer them in turn.
1. By iterating the Dyson equation (96), one finds that the average propagator expands as
hGi = + Σ + Σ Σ + Σ Σ Σ + ... (97)
By construction, there are no disorder correlations between the different self-energies appearing
here. In return, this implies that the self-energy contains exactly all correlations that cannot be
completely factorized by removing a free propagator G0 in between. These non-factorizable terms
are called “one-particle irreducible” (1PI). Moreover, the self-energy contains only the correla-
tions and internal propagators, but is stripped off the external propagator lines (“amputated”).
This makes the self-energy the simplest object describing all relevant disorder correlations.
2. Due to statistical homogeneity, the self-energy is diagonal in momentum and thus only
depends on k and E. The self-energy matrix element Σ(k, E) is calculated by applying so-
called Feynman rules to evaluate the diagrams. As a specific example,
let us give
 the Feynman
rules for the self-energy of the retarded single-particle propagator GR (k, E) in momentum
representation for the case of the speckle potential:

40
 (i) Draw all amputated 1PI diagrams with incident momentum k:

Σ(k, E) = + + ... (98)

(ii) Convert straight black lines to free propagators

−1
= GR 0
0 (k, E) = [E − εk + i0] .
k

q
(iii) Convert disorder correlation lines to • • = γ(q). The precise functional dependence
γ(q) depends on dimension and geometry.
(iv) For each scattering vertex, multiply by one power of the potential strength and the con-
servation of momentum:
q q0
= V0 δk+q,k0 +q0 (99)
k k0

(v) Sum over all free momenta after respecting momentum conservation.
For other types of disorder, these rules have to be adapted slightly to the precise shape of
diagrams, correlation functions and vertex factors. But in all cases, the general idea of writing all
possible combinations, respecting momentum conservation and integrating out the free momenta
is the same.
3. One can rewrite the Dyson equation (96) as [1 − G0 Σ]hGi = G0 and solve formally for the
average propagator: hGi = [1 − G0 Σ]−1 G0 = [G−1 0 − Σ]
−1
. Thus, its matrix elements are
1
GR (k, E) =


. (100)
E − ε0k − Σ(k, E)

We recognize that the self-energy modifies the free dispersion relation. Generally, the self-energy
is a complex quantity with a real as well as an imaginary part. The modified dispersion relation

Ek = ε0k + ReΣ(k, Ek ) (101)

is an implicit equation for the new eigen-energy Ek of the mode k. So one effect of the disorder
is to shift the energy levels. 4 But plane waves with fixed k are no longer proper eigenstates of
the disordered system. This is encoded in the imaginary part. Writing Γk = −2ImΣ(k, Ek ) and
using the fact that the self-energy varies smoothly with k and E, one finds a spectral density
Γk
A(k, E) = −2Im GR (k, E) =


. (102)
(E − Ek )2 + Γ2k /4

This spectral function is the probability density that an excitation k has energy E. Its wave-
number integral is the average density of states per unit volume,

1 dd k
Z
N (E) = A(k, E). (103)
2π (2π)d

For the free Hamiltonian, A0 (k, E) = 2πδ(E − ε0k ). The disorder introduces a finite spectral
width Γk , which translates into a finite lifetime τk = ~Γ−1
k . Equivalently, this finite lifetime
4 Alternatively, one can also solve for kE as the modified k-vector of an excitation with given energy E.

41
translates into a finite scattering mean-free path ls for the spatial matrix elements of the average
propagator,
dd k ik·(r0 −r)
Z
0
0
hG(r − r , E)i = d
e hG(k, E)i = G0 (r − r0 , E)e−|r −r|/2ls , (104)
(2π)

showing an exponential decay over ls = vk τk where group velocity vk = ~−1 ∂k ε0k and lifetime τk
are evaluated at the wave vector k such that Ek = E.
4. The simplest possible example is the calculation of the lifetime from the lowest-order,
so-called Born approximation
k − k0

Σ(k, E) = (105)
k k0 k

for some potential with correlation function P (q). To lowest order in V0 , we can use Ek = ε0k
and thus find the spectral width
dd k 0
Z
Γk = P (k − k 0 )2πδ(ε0k − ε0k0 ) = 2πP (0)N0 (ε0k ) (106)
(2π)d
in terms of the free density of states N0 (ε) and the low-k limit P (0) of potential correlation,
eq. (85), that is appropriate for the δ-correlated limit. This is precisely the result that one
gets from a straightforward application of Fermi’s Golden Rule for the average probability of
scattering out of the mode k by the external potential Vq . The interest of the full-fledged
diagrammatic expansion is of course that one is in principle able to calculate corrections to the
lowest-order estimate, and to tackle more complicated potentials. There exist literally hundreds
of other applications in the most diverse physical systems. Let us mention two examples from
our own experience.
For matter waves with quadratic dispersion ε0k = ~2 k 2 /2m in two-dimensional Gaussian
correlated potentials such as the one introduced in section 5.1.3, the scattering rate evaluates to

1 Γk 2πV 2 2 2
= 0 = 2 2 0 2 e−k σ I0 (k 2 σ 2 ) (Gauss, d = 2) (107)
kls 2εk k σ Eσ
where Eσ = ~2 /mσ 2 is a characteristic correlation energy and I0 a modified Bessel function.
For matter waves i In a one-dimensional speckle potential, we can use (105) and (106) with
the speckle potential correlation function (92). In d = 1, the only contributions can come from
forward scattering k 0 = k and backward scattering k 0 = −k, such that
1 Γk V 2k
= 0 = 0 2 [P (0) + P (2k)] (speckle, d = 1) (108)
kls 2εk ε0k
in terms of the k-space pair correlator P (2k) = πζ(1 − |kζ|) Θ(1 − |kζ|).
The estimates (107) and (108) can only be trusted if Γk /ε0k  1 or equivalently kls  1,
otherwise the assumption of a small correction to the free dispersion is no longer valid. Since
the scattering rates diverge at low k, we find that the perturbative approach breaks down at low
energy. A closer analysis shows that a sufficient criterion for weak disorder is Ek  V02 /Eσ [34].
Sometimes, also the real part of the self-energy is of importance. For example, one can
calculate the speed of sound in interacting Bose-Einstein condensates, and especially the shift
due to correlated disorder by the same Green function formalism [56]. Incidentally, for sound
waves the scattering mean-free path grows as k → 0, and the perturbative approach stays valid
even at very low energy.

42
5.2 Intensity transport
We would

like to calculate the ensemble-averaged density n(r, t) = hhr|ρ(t)|rii (or, its many-body
form Ψ† (r)Ψ(r) ) in the limit of long time. In the Schrödinger picture, the state evolves as


ρ(t) = U † (t)ρ0 U (t). After transforming the time evolution

operators to Green functions as in
(75), we need a theory for the ensemble-averaged product GA (E)GR (E 0 ) . This is known as


the average intensity propagator. In most experimental situations—be it with electromagnetic

or matter waves—one measures intensities (see for example the average transmission through a
1d disordered system of length L studied in section 2); the average intensity propagator is thus
the fundamental quantity of interest. Before going into details, we propose to have a look at
what we should expect to be the result.

5.2.1 Density response

The generic behavior that one may expect for transport in a disordered environment is diffusion.
Indeed, diffusion follows from two very basic and rather innocuous hypotheses. Firstly, one
generally has a local conservation law, for instance for particle number, taking the form of a
continuity equation:
∂t n + ∇ · j = s (109)
where j(r, t) is the current density associated with n(r, t), and s(r, t) is some source function.
Secondly, one assumes a linear response in the form of Fourier’s law
j = −D∇n, (110)
saying that a density gradient induces a current that tries to reestablish global equilibrium. The
diffusion constant D appears here as a linear response coefficient. Inserting (110) into (109), we
immediately find as a consequence the diffusion equation
[∂t − D∇2 ]n(r, t) = s(r, t). (111)
This equation can be solved by Fourier transformation5 . The solution for a unit source s(r, t) =
δ(r)δ(t) is the Green function for this problem, viz., the density relaxation kernel
1
Φ0 (q, ω) = . (112)
−iω + Dq 2
Its temporal version
dω −iωt
Z
Φ0 (q, t) = e Φ0 (q, ω) = θ(t) exp{−Dq 2 t} (113)
2π
shows that the relaxation exp{−t/τq } with characteristic time τq = 1/Dq 2 becomes very slow in
the large-distance limit q → 0 because of the local conservation law. In real space and time, the
relaxation kernel reads
dd q iq·r
Z
Φ0 (r, t) = e Φ0 (q, ω) = θ(t)[4πDt]−d/2 exp{−r2 /4Dt}. (114)
(2π)d
This relaxation kernel describes diffusive spreading with r2 = 2dDt.



This is the “hydrodynamic” description of dynamics on large distances and for long times,
accessed by small momentum q and frequency ω. A microscopic theory is then only required to
calculate the linear response coefficient D.
5 which was invented right for this purpose by Joseph Fourier, namesake of the French university in Grenoble

hosting the Les Houches school.

43
5.2.2 Quantum intensity transport
In complete analogy to the Dyson equation (96) for the average single-particle

propagator, one
may write a structurally similar equation for the intensity propagator Φ = GR GA , known as


the Bethe-Salpeter equation:

Φ = GR GA + GR GA U Φ.





(115)

Here, one splits off the known evolution with uncorrelated, average amplitudes

G (k, E) GA (k 0 , E 0 ) =

R 

. (116)

k0

The upper part of intensity diagrams describes the retarded propagator, called “particle channel”
in condensed-matter jargon, whereas the lower part contains the advanced propagator or “hole
channel”. All scattering events that couple these amplitudes are contained in the intensity
scattering operator U . By construction, this “particle-hole irreducible” vertex contains exactly
all diagrams that cannot be factorized by removing a propagator pair (116). Its detailed form
again depends on the model of disorder. In all cases, Ukk0 (E) is essentially the differential
cross-section for scattering from k to k 0 and generally has the following structure:

k k0
0
U (k, k ; E) = = + + + + ... (117)
0
k k

Linear-response theory shows that this scattering vertex permits to calculate the transport mean-
free path l, in close analogy to the calculation of the scattering mean-free path ls from the
self-energy. Their ratio is expressed as
ls
= 1 − hcos θiU (118)
l
where θ is the scattering angle between k and k 0 , and the brackets h.iU indicate an average over
the scattering cross-section U . The physical interpretation of the transport mean free path l
is the following: while the scattering mean free path ls measures the distance after which the
memory of the initial phase of the wave is lost, l is the distance over which the direction of
propagation is randomized.

5.2.3 Diffusion
The scattering processes encoded in U are perhaps more easily visualized in real space. We will
draw a full line for every amplitude ψ propagated by GR (upper lines in (117)) and a dashed line
for every ψ ∗ propagated by GA (lower lines in (117)). Impurities are represented by black dots
as before. Then, the first contribution to U describes the single-scattering process
r

UB : r1 (119)
r0

44
in which both ψ and ψ ∗ are being scattered by the same impurity at position r1 . This process
is insensitive to phase variations and could just as well take place for classical particles. So this
Boltzmann contribution UB describes classical diffusion with diffusion constant
vlB
DB = (120)
d
The Boltzmann transport mean-free path is calculated by inserting UBkk0 = V02 P (k − k 0 ) into
(118): R
ls (k) dΩd cos θP (2k| sin(θ/2)|)
=1− R (121)
lB (k) dΩd P (2k| sin(θ/2)|)
Depending on the microscopic scattering process, lB can be longer than ls , if forward scattering
is dominant, hcos θiUB > 0. This is the case for matter waves in spatially correlated potentials.
For isotropic scattering with hcos θiUB = 0, these two length scales coincide, ls = lB .
By combining eq. (121) with eq. (106) giving the scattering mean free path, one can easily
compute, in the Born approximation, the transport mean free path and consequently the classical
Boltzmann diffusion constant, using only microscopic ingredients: the dispersion relation of the
free wave and the correlation function of the scattering potential.

5.2.4 Localization length in 1d systems

As we have already seen repeatedly in previous sections, in 1d the transport mean free path
is (up a factor 2) equal to the localization length, ` = ξloc /2, an identity that can also be
verified microscopically [57], at least to lowest order V02 in perturbation theory.6 So now we
are in a position to give a microscopic prediction for the 1d localization length for arbitrarily
correlated potentials, namely taking twice the backscattering contribution from (108), selected
by the (1 − cos θ)-factor in (121):

1 V 2k
= 0 2 P (2k). (122)
kξloc 4ε0k

Figure 17, taken from [4], shows this prediction for Lloc = 2ξloc as a dashed line together
with the results of a fit to the intensity measured in the real experiment (see Fig. 8). Here
P (2k) = πζ(1 − kmax ζ)Θ(1 − kmax ζ) with kmax the largest k-value present in the expanding wave
packet, resulting in:
~4 kmax
2
ξloc = . (123)
πm2 V02 ζ(1 − kmax ζ)
There is no adjustable parameter, and the agreement is rather satisfactory. Significant devi-
ations are visible both for small disorder (there the localization length becomes too large and
experimental limitations start to show) and for large disorder, where the lowest-order theoreti-
cal estimate, or Born approximation (123) becomes insufficient. Moreover, for strong disorder,
atom-atom interaction for the strongly localized cloud may no longer be negligible and induce
some delocalization.
An interesting scenario occurs in speckle potentials when the fastest atoms have a wave vector
kmax larger than the most rapid spatial fluctuations, with wave vector 1/ζ. Then, P (2kmax ) = 0,
and (123) predicts prima facie ξloc = ∞ or absence of localization, which would signal the
existence of a mobility edge in this 1d random potential, contradicting rigorous mathematical
6 Whether this holds to all orders in perturbation theory is to our knowledge an open, and also interesting

question, especially in optical speckle potentials [20].

45
 profiles, from which the localization length Lloc is extracted by fitting an
exponential, exp(22 | z | /Lloc) (dotted black lines), to the atomic density in Figure 4 | Algebraic and exponential regimes in a one-dimensional speck
the wings. b, Localization length Lloc versus expansion time t. Error bars, potential. Log–log and semi-log plots of the stationary atomic density
95% confidence intervals for the fitted values (62 s.e.m.). profiles, showing the difference between the algebraic (kmaxsR . 1) and
exponential (kmaxsR , 1) regimes. a, Density profile for VR/min 5 0.15 and
exponent of 1.95 6 0.10 (62 s.e.m.), in agreement with the theor- kmaxsR 5 1.16 6 0.14 (62 s.e.m.). The momentum distribution of the
etical prediction that density decreases like 1/z2 in the wings. The released BEC has components beyond the effective mobility edge 1/sR. Th
semi-log plot (inset) confirms that an exponential would not work as fit to the wings with a power-law decay 1/ | z | b yields b 5 1.92 6 0.06
(62 s.e.m.) for the left-hand wing and b 5 2.01 6 0.03 (62 s.e.m.) for the
well. For comparison, we present in Fig. 4b a log–log plot and a semi-
right-hand wing. The inset shows the same data in a semi-log plot, and
log plot (inset) for the case with kmaxsR 5 0.65 and VR/min 5 0.15, confirms the non-exponential decay. b, For comparison, a similar set of plo
where we conclude in favour of exponential rather than algebraic (log–log and semi-log) in the exponential regime for the same VR/min 5 0.1
tails. These data support the existence of a crossover from an expo- and kmaxsR 5 0.65 6 0.09 (62 s.e.m.).
nential to an algebraic regime in our speckle potential.
2.5 Direct imaging of atomic quantum gases in controlled, optica
disordered potentials is a promising technique to investigate a variet
of open questions on disordered quantum systems. First, as in othe
2.0
problems of condensed matter simulated using ultracold atom
direct imaging of atomic matter waves offers unprecedented possibi
ities to measure important properties, such as localization length
Second, our experiment can be extended to quantum gases wit
1.5
controlled interactions where localization of quasi-particles26,2

Lloc (mm)
Bose glass14,15,28 and Lifshits glass29 are expected, as well as to Ferm
gases and to Bose–Fermi mixtures where rich phase diagrams hav
1.0 been predicted30. The reasonable quantitative agreement between ou
measurements and the theory of one-dimensional Anderson loca
ization in a speckle potential demonstrates the high degree of contro
0.5 in our set-up. We thus anticipate that it can be used as a quantum
simulator for investigating Anderson localization in higher dimen
sions31,32, first to look for the mobility edge of the Anderson trans
0.0 ition, and then to measure important features at the Anderso
0 20 40 60 80
transition that are still under theoretical investigation, such as critica
VR (Hz)
exponents. It will also become possible to investigate the effect o
Figure 3 | Localization length versus amplitude of the disordered potential. controlled interactions on Anderson localization.
Lloc is obtained by an exponential fit to the wings of the stationary localized
density profiles, as shown in Fig. 2. Error bars, 95% confidence intervals for METHODS SUMMARY
Figure 17: Comparison between
the fittedthe
valuesexperimentally measured
(62 s.e.m.); 1.7 3 104 atoms ; min 5 219localization
Hz. The dash–dotlength for adistribution
Momentum quasi-1d atomic
of the expanding BEC. To compare measured loca
wave-packet launched in theline is plotted using
disordered equationpotential
optical is determined
(1), where kmaxcreated by from the
a speckle pattern, andwith
ization lengths the theoretical
those calculated from equation (1), we need to know kma
observed free expansion of the condensate (see Methods). The shaded area the maximum amplitude of the k-vector distribution of the atoms, at the begin
prediction, eq. (123), when represents
the strength
uncertaintyof the disordered
associated potential
with the evaluations of kmax andissRvaried.
. We ningThere is noin adjustable
of the expansion the disordered potential. We measure kmax by releasing
note that the limited extension
parameter. Reprinted from [4] (courtesy of Ph. Bouyer). of the disordered potential (4 mm) allows us BEC with the same number of atoms in the waveguide without disorder, an
to measure values of Lloc up to about 2 mm. observing the density profiles at various times t. Density profiles are readi
89
©2008 Macmillan Publishers Limited. All rights reserved
theorems stating that all states are exponentially localized [58]. In fact, exponential localization
still prevails, but requires more than a single scattering event by the smooth random potential.
Going to higher orders in perturbation theory, beyond the Born approximation, one can show
in excellent quantitative agreement with numerics, that the localization length is always finite
[21, 20]. However, for weak disorder, the localization length can become much larger than the
system size, such that numerical or experimental results show an apparent mobility edge. For
other types of long-range correlations, one does find mobility edges [59]. There seems to be no
obvious way of deciding, for a certain class of potentials, whether true exponential localization
exists or not, and correlated potentials are still actively investigated in different contexts, see
[60] and references therein.
Despite the nice agreement shown in Fig. 17, some caution is indicated. The experimental
observation involves averaging over k, and also over several different realizations of the disorder
(single-shot results look similar, just more noisy). This means that the experimental data resem-
bles the average transmission hT (z)i as a function of sample thickness z.7 In section 2.2.3, we
showed that it is the typical transmission Ttyp (z) = exp(hln T (z)i) which decays exponentially,
not the average transmission. At very large z, this can make a huge difference, see section 2.5.
Fortunately, for z of the order of the localization length (t or order unity in the language of 2.5),
the fluctuations have not yet built up, and the difference between the typical and the average
value is still small, ln hT (z)i ≈ −z/ξloc , making the pure exponential decay an acceptable ap-
proximation. Further in the wings, one expects deviations of the average density from a pure
exponential decay, see eq. (38). This takes place however in the region where fluctuations are
huge, so that a typical experiment may not measure the average value of the density, but rather
its typical value.
7 The situation is actually more complicated, because this expansion experiment strictly speaking does not

measure the transmission across a sample. Instead, one starts with an atomic density inside the medium and see
how it propagates. The boundary conditions are thus different from those of a transmission experiment with its
connection to outside leads. Still, huge fluctuations must exist in the localized regime, implying that the average
transmission deviates from a pure exponential, as discussed in Section 2.5.

46
5.2.5 Weak-localization correction
The first corrections to the classical, incoherent scattering process (119) shown in (117) involve
one more scatterer and several possibilities of intermediate propagation. The most well-known
type of correction stems from the diagram with two crossed lines. In real space, the scattering
process is
r

r1

(124)
r2

r0

This is an interference correction with a phase shift ∆ϕ between ψ and ψ ∗ that depends on
the impurity positions r1 and r2 . Contributions of this type are ensemble-averaged to zero—or
rather, almost averaged to zero. Indeed, if the starting and final point of propagation come close,
r ≈ r0 , the phase shift picked up by the two counter-propagating amplitudes becomes smaller:
r1
r (125)
r0
r2

At exact backscattering r = r0 and in the absence of any dephasing mechanisms, the phase dif-
ference is exactly zero. Vanishing phase difference means constructive interference and therefore
enhanced backscattering probability to stay at the original position. This holds true no mat-
ter how many scatterers are visited on the path. One is led to consider all maximally crossed
diagrams:

UC = + + ... (126)

These diagrams were first considered in the electronic context [61] and became known as the
Cooperon contribution. This contribution is peaked around backscattering k = −k 0 . Therefore,
one may resort to a diffusion approximation and sum up all contributions with the help of the
diffusion kernel (112):

dd q
 
1 1 1 1
Z
= 1+ (127)
l lB πN0 (2π)d −iω + DB q 2 ω→0

Writing this in terms of the diffusion constant, one arrives at the weak-localization correction

dd q
 
1 1 1 1
Z
= 1+ (128)
D DB πN0 DB (2π)d q 2 − i0

The quantum correction of the Cooperon makes D < DB , and we have thus found the microscopic
reason for the weak-localization correction that was first mentioned in the scaling section 3.5.
Before looking in more details at this correction in section 7, we should like to understand it
better by selectively probing the Cooperon contribution. In Optics, this is indeed possible and
is developed in the next section.

47
 R

20 4

Figure 18: Half-space geometry of a backscattering experiment with cylindrical coordinates r = (R, z).
20 is the image point of the exit scatterer 2 used to construct the half-space propagator for the incoherent
intensity. As an example, the contribution of scattering from four scatterers is depicted.

6 Coherent backscattering (CBS)

One can probe the specific geometry of scattering paths like (124) by using a source of plane
waves together with a collection of randomly positioned scatterers in a half-space geometry (fig.
18). Hereafter, we suppose normal incidence and detection close to the backscattering direction;
generalizing to arbitrary incident and detection angles changes nothing to the central argument.

6.1 Theory
The picture in Fig. 18 shows scattering by four impurities, contributing to the incoherently
transported intensity. The corresponding intensity diagram is

. (129)

The sum of all such diagrams with a distinct ladder topology yields the intensity propagator or
diffuson, whose long-distance and long-time form is precisely the diffusion kernel (112), evaluated
with the Boltzmann diffusion constant.
1
ΦB = . (130)
−iω + DB q 2
The total back-scattered diffuse intensity per unit surface is given by summing contributions
from all possible starting and end points:
Z Z Z
−z1 /ls −z1 /ls
IL ∝ dz1 e dz2 e d2 R ΦB (r1 , r2 ). (131)

The exponential attenuation factors describe the propagation of

intensities
 from the surface to
the first scatterer and back out again with average propagator GR (zi ) , (104), featuring the
scattering mean free path ls . Moreover, translation invariance along the surface direction has
been used, leaving only the surface integral over the lateral distance R = R1 − R2 .
The propagation inside an infinite disordered medium would occur with the bulk kernel (130)
and thus have a time-integrated diffusion probability of ΦB (r) = dtΦB (r, t) = [4πDB r]−1 . This
R

expression leads to a diverging integral over R in (131). But the starting and end points r1 and

48
 CBS profile
1.0
R

0.8
1
θ
3
0.6

z
0.4

4
0.2
θ 2
q
(a) (b) -2 -1 0 1 2

Figure 19: Coherent backscattering (CBS) (a) Schematic picture of a four-scatterer path; the construc-
tive interference of path-reversed amplitudes in the backscattering direction θ = 0 leads to an observable
intensity enhancement. Away from backscattering, the phase differences average out and leave only the
background intensity. (b) CBS profile as function of reduced scattering angle q = klθ, normalized to the
value at θ = 0. Solid black: exact solution (136). Dashed blue: Diffusive solution (134). Dotted orange:
Linear solution (135) with characteristic slope discontinuity at backscattering.

r2 lie rather close to the surface, namely typically one scattering mean-free path ls away from
it. So for calculating the back-scattered intensity (131), we have to worry about appropriate
boundary conditions. The complete integral equation for intensity propagation in a half-space
geometry of a scalar wave and isotropic scatterers, known by the name Milne equation, can
be solved exactly [62, 63], albeit with considerable mathematical effort. For a simple solution
involving the diffusive bulk propagator valid far from the boundary, one can employ the method
of images that is often used in electrostatics. Since photons reaching the surface would escape
prematurely from the medium, one can exclude these events by subtracting the contribution of
propagation to an image point r20 = (R2 , −z2 ) mirrored to the outside of the sample:
 
1 1 1
ΦB (r1 , r2 ) = − . (132)
4πDB r12 r120

This half-space propagator behaves like R−3 at large R and thus permits to carry out the
integration. The final result is some number IL and gives the incoherent background on top
of which we now study the interference contribution.
Each multiple-scattering diagram like (129) has an interference-correction counterpart such
as
(133)

in which the conjugate amplitude travels along the same scatterers, but in opposite direction.
The contribution of such maximally crossed diagrams to the back-scattered intensity can be
accounted for along the same lines. First of all, incident and scattered amplitudes now pick up
a phase between the surface and the scattering end points. Namely, the amplitude ψ picks up
exp{ik · r1 } at the entrance and exp{ik 0 · r2 } at the exit of the medium. The path-reversed
complex conjugate amplitude picks up exp{−i[k · r2 + k 0 · r1 ]}. So after all, there is a total
phase difference of ∆ϕ = (k + k 0 ) · (r1 − r2 ). Exactly toward the backscattering direction,
k 0 = −k, this phase difference vanishes. Close to backscattering, for a small angle θ  1, one
has |k + k 0 | ≈ k⊥ = k sin θ ≈ kθ, and the phases differ by ∆ϕ = kRθ.

49
 Thus, each path acts like Young double-slit interferometer with the two end-point scatterers
playing the role of the two slits. The larger the transverse distance R between the scatterers, the
finer the interference fringes. The only point where all fringes are bright is the symmetry point
θ = 0 toward backscattering. Sufficiently far away from this direction, the sum of random fringe
patterns averages out to zero. The sum of all interference term is again the integral over all end
points with the appropriate weight furnished by the intensity propagator (132) (which must be
modified if some additional dephasing processes are at work, see section 6.3 below). This simple
calculation predicts a relative interference enhancement over the background
IC (θ) 1
≈ (134)
IL (1 + kl|θ|)2
The interference-induced enhancement, shown in figure 19 as a dashed blue line, survives in an
angular range ∆θ = 1/kl = λ/(2πl) around backscattering. Very characteristically, this peak
features a triangular cusp at backscattering (plotted in dotted orange),
IC (θ)
= 1 − 2|q| + O(q 2 ) (135)
IL
where q = klθ is the reduced momentum transfer.
The exact solution for scalar waves and isotropic point scatterers can be calculated solving
the Milne equation of intensity transport. The CBS profile can then be expressed as the integral
[63] ( " #)
2 π/2
p
IC (θ) 1 arctan q 2 + tan2 β
Z
= exp − dβ ln 1 − p (136)
IL C π 0 q 2 + tan2 β
n R π/2 o
where q = klθ and a constant C = exp − π2 0 dβ ln [1 − β cot β] ≈ 8.455 such that at the
origin IC (0) = IL . This profile is plotted as a black curve in Fig. 19. It becomes apparent that
the diffusive solution (134) gives a very good description for small scattering angles. Notably, its
slope at θ = 0 is precisely equal to the exact value that can be extracted from (136). This was to
be expected since diffusion should be valid for long-distance bulk propagation, and long scattering
paths have widely separated end points that contribute to the small transverse momenta making
up the top of the CBS peak. Indeed, the diffusion prediction (134) is precisely recovered by
replacing the exact propagation kernel under the logarithm by its diffusion approximation:
1
A(Q) = arctan(Q)/Q ≈ 1 − Q2 (137)
3
p
valid at small Q = q 2 + tan2 β.
The agreement between the diffusive profile (134) and the exact result (136) deteriorates at
larger angle q = klθ. This discrepancy is due to low scattering orders that are not accurately cap-
tured by the diffusion approximation and the imaging method used to mimic the exact boundary
conditions. Indeed, only the very tip of the CBS peak stems from long paths reaching far into
the bulk. The larger part of the total signal is due to contributions from rather short paths, for
which scattering inside the surface skin layer is crucial.
One can calculate the contribution of scattering orders n = 1, 2, . . . to the total incoherently
back-scattered intensity, measured in units of the incident flux by the so-called bistatic coeffi-
cient γ(cos θ0 , cos θ) [64] that depends on the angles θ0 and θ of incidencePand observation. For
exact backscattering and normal incidence (θ = θ0 = 0), one has γ = n≥1 γn . The largest
contribution comes from single scattering with γ1 = 1/2 followed by double-scattering with
γ2 = ln(2)/2 ≈ 0.35 and so on, with an asymptotic decrease as γn ∼ n−3/2 [63].

50
 When the CBS peak was first observed in the beginning of the 1980’s [65, 66, 67], the diffusive
theory used an image point placed at z20 = −(2z0 + z2 ) such that the diffuse propagator vanishes
at a distance z0 = 2/3 outside the sample. This translates to the boundary condition that the
total incident diffusive flux on the surface vanishes [62, 68] and leads to a diffusive CBS peak
shape of
1 − exp{−2z0 |q|}
 
IC (q) 1 1
= 1+ . (138)
IL (1 + |q|)2 1 + 2z0 |q|
This diffusive solution predicts a different slope at the origin, viz., −2[1 + z02 /(1 + 2z0 )], which
is off by more than 20% from the exact value, although one would expect the diffusion solution
to get this value right [63]. This is all the more disturbing as fits to the diffuse CBS peak shape
are generally used to measure the transport mean-free path. Also at larger angles this solution
cannot convince because the diffusion profile decays as q −2 , whereas the exact solution decreases
like |q|−1 . This asymptotic behavior is known to come from the double-scattering contribution.
Bart van Tiggelen has noticed [69] that the diffusion approximation becomes virtually exact
if single- and double scattering are included separately since

1 A(q)2 3α
= 1 + A(q) + ≈ 1 + A(q) + 2 (139)
1 − A(q) 1 − A(q) q

both for small and large q, with a numerical coefficient α = 1 for q → 0 and α = π 2 /12 ≈ 0.822
for q → ∞. Therefore, the best approximation to the exact solution is obtained by first taking
the exact double-scattering profile [63, 70]

1 π/2 2 cosh−1 (1/|q|) − cosh−1 1/q 2

Z q 

2 2
γ2 (q) = dβA q + tan β = p (140)
π 0 2 1 − q2

where cosh−1 (x) is the inverse hyperbolic cosine function, then adding the diffusive solution

1 − exp{−2z0 |q|}
 
3α
γdiff (q) = 1+ (141)
2(1 + |q|)2 |q|
and finally fitting the extrapolation length z0 and diffusion-constant multiplicator α such that
height and slope are equal to the exact values at the origin. Doing this, we find α∗ ≈ 0.86 within
the expected interval [0.822, 1] and z0∗ ≈ 0.81. Figure 20 shows the exact CBS profile (with
the single-scattering contribution subtracted as required) together with the double scattering
contribution plus the full approximated diffusive CBS profile that turns out to be in excellent
agreement, both for small and large angles.
The full width at half height of the CBS profile is ∆q ≈ 0.73kl ≈ 4.59l/λ, and observing the
CBS peak can be used to measure the transport-mean free path quite accurately. The explicit
occurrence of the wavelength λ emphasizes that CBS is a genuine interference effect. In many
circumstances, the mean-free path is much longer than the wavelength, such that kl ∼ 102 ...103 ,
and ∆θ is at most a couple of mrad. This makes CBS difficult to observe with the naked eye,
together with the constraint that one has to look exactly toward the backscattering direction,
but it can be easily imaged using standard optics, as schematically shown in Fig. 21.

6.2 Live experiment

Because the CBS cone is typically very narrow and its maximum height at best equal to the
average background, a source with large angular dispersion will broaden the signal too much
and reduce enhanced backscattering. Thus, it is highly desirable to use a quasi-parallel beam

51
 CBS enhancement

1.8

1.6

1.4

1.2

q
-3 -2 -1 0 1 2 3

Figure 20: CBS intensity enhancement in units of the background intensity as function of reduced
scattering angle q = klθ. Solid black: exact solution (136) minus the single-scattering value γ1 = 1/2.
Dashed blue: double scattering contribution (140). Dashed red: sum of double scattering and best
diffusive solution, (141) with α∗ ≈ 0.86 and z0∗ ≈ 0.81. Dotted orange: Traditional diffusive CBS peak,
eq. (141), with α = 1 and z0 = 2/3 shown for comparison. Even for this time-reversal invariant case,
the backscattering enhancement is slightly smaller than 2 because of the single-scattering contribution
to the background that is absent in the CBS signal.

BS
1
0
0
1
0
1
0
1 Sample
0
1
0
1
0
1
Laser p1 0
1
p2

CCD
Figure 21: Schematic view of a table-top CBS experiment. Light from a wide laser beam with small
angular dispersion is directed onto a disordered sample. The diffuse retro-reflected intensity is sent by a
beam-splitter (BS) in the focal plane of a lens and recorded by a CCD camera, thus imaging the angular
distribution. Polarization elements (p1 and p2) select a suitable polarization channel; generally, the
helicity-preserving channel of opposite circular polarization is recommended.

52
Figure 22: Intensity around the back-scattered direction for a piece of paper exposed to a parallel
laser beam. For a fixed paper (left figure), one observes a characteristic speckle pattern, due to random
interference of phase-coherent light scattered by a single configuration of disorder. By averaging over
various parts of the paper, small-scale random variations are averaged out, but an enhanced intensity
is clearly visible around exact backscattering. Original data from an experiment performed during the
Les Houches Summer School in Singapore on July, 15th, 2009. Special thanks to David Wilkowski, Kyle
Arnold, and Lu Yin from the Center for Quantum Technologies, National University of Singapore, for
generous support and invaluable help in setting up the experiment.

obtained from a laser source, with angular divergence smaller than a fraction of mrad. This
is turn requires a large spot, with a diameter larger than the mean free path, which is easily
obtained by expanding the output beam of a commercial diode laser with a telescope. The
scattering medium should scatter efficiently and must not absorb the light: a bright white object
is thus chosen. A piece of ordinary paper turns out to give the best results. A sheet of paper is
about 100 µm thick and obviously scatters most of the incoming beam, meaning that the mean
free path does not exceed a few tens of µm. A piece of teflon could also be used, but the mean free
path is significantly larger, meaning a narrower CBS cone, much harder to detect. White paint or
milk make also good samples, with the advantage that the concentration and thus the mean free
path can be varied and that the thermal motion of scatterers inside the solvent provides us with
configuration averaging for free; however, these samples must be put inside some transparent
container whose surface can produce specular reflection that is easily confounded with the CBS
signal.
A semitransparent plate (beamsplitter) can be used to send the back-reflected light into
a 1280×1024 pixel CCD camera with pixel size around 5 µm, located about 20 cm from the
scattering medium, thus ensuring a 0.025 mrad angular resolution. Fig. 22 (left) shows the
image recorded from a fixed piece of paper. This situation corresponds to a single realization of
the disorder. The electric field on each pixel is the coherent sum of the field amplitudes radiated
by each point of the sheet of paper. Because of its disordered nature, each contribution picks a
random amplitude and phase, resulting in a characteristic speckle pattern on the CCD camera,
the “optical fingerprint” of the paper. The angular size of the speckle grain is of the order of
1/kL where L is the size of the illuminated spot on the sheet of paper. For our case, it is about
0.1 mrad, i.e. slightly larger than the pixel size, in agreement with the experimental observation.
The attentive reader may notice that the bright spots look slightly brighter in a roughly circular
area on the right side of the figure. In order to see the CBS cone, one should perform configuration
averaging. This is easily done by mounting the piece of paper on a rotating device (in our case
a battery-powered computer fan). On the time-averaged intensity, shown in Fig. 22 (right), the

53
 Experimental signal (July, 15, 2009, 3.56 pm)
1/(1+|x|/180)^2
1 with decoherence

CBS signal/background
0.8

0.6

0.4

0.2

0
-800 -600 -400 -200 0 200 400
Position (pixel #)

Figure 23: Experimentally measured coherent backscattering signal, together with a fit to the simplest
theoretical formula, (134) (full red line), and to (146) with phenomenological decoherence included
(dashed blue). From the angular width of the CBS signal, one can extract the transport mean free path
inside the sheet of paper, here 25 µm. Experimental imperfections limit the coherence of the phenomenon
and are responsible for the deviation from the peaked shape at the center of the cone. When properly
taken into account (dashed curve), the agreement is very good.

fluctuating speckle pattern has been washed out, leaving a uniform background, on top of which
appears a smooth bright spot of approximate width 10 mrad due to coherent backscattering. The
effect is perhaps not dramatic, as the enhancement factor cannot be larger than 2 (it is 1.6 in
this live experiment), but clearly present and visible with the naked eye.
A cut across the spot center is presented in Fig. 23 together with a fit to the simplest
theoretical formula, eq. (134). The fit is quite good in the wings and allows us to extract the
mean free path inside the piece of paper, in our case 25 µm. The fact that the top of the CBS
peak is rounded can be attributed to various experimental imperfections such as the finite angular
resolution, geometrical aberrations, finite thickness and residual absorption of the piece of paper,
but could in principle also highlight the presence of a decoherence mechanism.

6.3 Dephasing/decoherence
The CBS phenomenon presented so far relies on perfect phase coherence of the multiply scattered
wave. What happens if some external agent—such as some degree of freedom inside the paper
coupled to the wave—affects the scattered amplitude in an uncontrolled way? Qualitatively, it
is clear that amplitudes of long scattering paths are more fragile than those of shorter paths. As
very long paths are responsible for the characteristic triangular shape of the CBS cone around
the exact backscattering direction, it is important to understand the effect of decoherence on the
CBS signal. In return, the CBS enhancement factor can serve as a sensitive measure for phase
coherence.
There is no universal way of breaking phase coherence, and the effect on CBS can be dif-
ferent depending on the specific mechanism at work. Nevertheless, a simple phenomenological
approximation may often be used, and we will see below that several physical processes are well
described by this approximation. This assumption is that phase coherence is lost at a constant

54
rate, characterized by a phase-coherence time τφ , also called dephasing time. Then, interfer-
ence terms associated with paths who are visited in a time t have to be multiplied by a factor
exp(−t/τφ ). An example of such a situation is provided by a Michelson interferometer oper-
ated with a classical light source, where the interference disappears once the optical path length
difference exceeds cτφ , with τφ the longitudinal coherence time of the source. Note that these
phenomena are typically called “dephasing” in the context of classical waves, and “decoherence”
for quantum mechanical matter waves. The bottom line is simply that interference is lost by
coupling to some external degree of freedom.
The exponential attenuation of interference as exp(−t/τφ ) applies especially often to the
Cooperon contribution. In Fourier space, the effect is simply tantamount to the replacement
i
ω 7→ ω + (142)
τφ

or, in the diffusive propagator (112):

1 1 1
2
7→ 1 =   (143)
−iω + DB q −iω + τφ + DB q 2
−iω + DB q 2 + 1
D B τφ

which can be also be obtained via the replacement

1
q 2 7→ q 2 + . (144)
L2φ

The phase coherence length, p

Lφ = DB τφ , (145)
is the average distance over which the wave propagates diffusively before losing its phase coher-
ence.
This simple replacement can be used to calculate the shape of the CBS cone in the presence
of decoherence effects. Indeed, Sec 6.1 discusses several approximate expressions for the shape,
all expressed as a function of the transverse momentum k⊥ ≈ k|θ|, which is nothing but the sum
of the incoming and outgoing momenta (in the limit of small angles θ  1). The substitution
2 2
k⊥ 7→ k⊥ + 1/DB τφ in eq. (134) yields

IC (θ) 1
≈h i2 . (146)
IL
q
1 + (klθ)2 + l2 /L2φ

This expression is now a smooth function of θ (no cusp at θ = 0 anymore). In the limiting
case l  Lφ , one recovers the previous expression, only slightly perturbed near the tip. In the
opposite limit Lφ  l, the CBS cone disappears completely, which is quite natural as interference
effects are washed out before the wave travels a single mean free path. The relative height of the
CBS peak, compared to the background at kl|θ|  1, is

IC (0) 1 l τl
r

IL
= 2 ≈ 1 − 2L = 1 − 2 τ
(147)
[1 + l/Lφ ] φ φ

where the last two expressions are valid in the limit of weak decoherence l  Lφ . Here, τl = DB /l2
is of the order of the mean free time separating two consecutive scattering events. This expression
emphasizes the sensitivity of the CBS cone to dephasing effects. Indeed, if the dephasing time
is say 10 times larger than the mean free time, its effect on the CBS cone is still very noticeable,

55
Figure 24: CBS of light by cold Strontium atoms [71]. The backscattering enhancement is close to 2,
which indicates full phase coherence. Signal observed in the helicity preserving channel (incoming light
circularly polarized, detection in the opposite circular polarization). Solid line: the result of a calculation
taking into account the finite geometry and inhomogeneous density of the atomic cloud.

reducing its height by almost 50%. For example, the experimentally observed CBS cone in the
live experiments is well fitted by eq. (146), with a decoherence time τφ = 12.5τl .
Several other types of decoherence have been studied in great detail in connection with light
and cold atoms. In the following, we present a few of them qualitatively, referring to the literature
for more details.

6.3.1 Polarization
For simplicity, we have up to now considered a scalar complex wave, describing e.g. a spinless
atomic matter wave. Many real atoms, all electrons and also electromagnetic waves are more
complicated because of their spin/polarization. Especially light scattering does not preserve po-
larization, as is obvious from the requirement of transversality. Thus the light back-scattered
along the direct and reverse paths generically emerges from the medium with different polar-
ization. But orthogonal polarizations do not interfere, and thus one may expect a reduced
enhancement factor.
Technically, one has to dress the multiple-scattering Cooperon contribution with the polar-
ization structure. Two independent polarizations of the propagating intensity must be taken
into account (one in the retarded, one in the advanced Green function), leading to a tensorial
structure for diffuson and Cooperon alike. The complete calculation of this effect is possible by
decomposing the intensity kernel into irreducible tensor moments [72]. To make a long story
short, it is enough to say that each contribution has a kernel of the type (143), with its own τφ
of the order of τl . The physical interpretation is clear: because scattering will on average lead to
depolarization, all channels associated with specific polarization correlations must decay during
propagation. Only the one channel measuring the total intensity is protected by conservation of
energy, with 1/τφ = 0, and propagates diffusively.
If the system is additionally time-reversal invariant, the same conserved intensity channel
also exists for the Cooperon. The population of the various contributions depends on the specific
choices for the incoming and outgoing polarizations used for recording the CBS signal. Using the
same linear polarization for excitation and analysis populates the conserved mode and ensures
an optimal interference contrast for long scattering paths, at least for classical point-like objects
(such as dye molecules) acting as Rayleigh scatterers (we will discuss in Sec. 6.3.4 the more general

56
case). The same is true if the incident field has circular polarization and the opposite circular
polarization is used for detection (helicity-preserving channel), with the additional advantage
that the single scattering background of the diffuson is filtered out, allowing in principle the
observation of a a perfect CBS enhancement by a factor of 2 [73, 71].

6.3.2 Residual velocity of the scatterers

The previous derivation assumed quenched disorder, i.e. scatterers at fixed positions. Moving
scatterers are a cause of decoherence: as light travels along two reciprocal paths, it visits the same
scatterers, but in opposite order, i.e. at different times. If, during the time delay separating the
scattering events on the direct and reversed path, the atom has moved by at last one wavelength,
the phase coherence between the two paths will be lost. This phenomenon can alternatively
be interpreted in the frequency domain, where moving scatterers induce a Doppler shift of the
scattered photon which is different along the direct and reversed path. Although this phenomenon
does not lead to a strict exponential decay of the phase coherence [74], it reduces the enhancement
factor, which has been notably observed with cold atoms [75].
In general, interference of waves is suppressed once the environment has acquired knowledge
of the path taken by the scattered object. This can be most simply seen in experiments of the
Young’s double-slit type [76], but applies equally to the CBS by light from moving atoms, where
moreover the storage of which-path information in the atomic recoil has been studied [77].

6.3.3 Non-linear atom-light interaction

Because atoms have extremely narrow resonance lines, they have large polarisabilities and al-
ready quite low laser intensities can saturate an atomic transition, in which case the atom scatters
photons inelastically. It is easy to understand that such a non-linear inelastic process will re-
duce the phase coherence of the scattered light and the enhancement factor. This indeed has
been observed [78]. A full quantitative understanding of multiple inelastic scattering is still not
available. For a model system of two atoms driven by a powerful laser field, a rather complete
understanding of the CBS signal has been achieved (see [79, 80] and references therein).
In the context of matter waves, one may study coherent backscattering of interacting matter
waves, obeying a non-linear equation such as the Gross-Pitaevskii equation, evolving in a random
optical potential. It has been shown that already a moderate non-linearity induces a phase-shift
between the direct and reversed paths and thus a decrease of the height of the CBS peak, and
may in some cases even create a negative contribution in the backward direction [53]. These
theoretical predictions still await experimental realization.

6.3.4 Internal atomic structure/spin-flip

The preceding description treats atoms as Rayleigh point scatterers that radiate a purely dipolar
electromagnetic field, with an induced dipole directly proportional to the incoming electric field.
This is an excellent approximation for atoms with a non-degenerate electronic ground state, such
as Strontium. The situation is radically different if the atomic ground state is degenerate: indeed,
when scattering a photon, the atom may stay in the same atomic state (Rayleigh transition) or
change to another state with the same energy (degenerate Raman transition), see Fig. 25(a).
The basic rules of quantum mechanics imply that orthogonal final state cannot interfere. In
other words, two multiple scattered paths will interfere only if they are associated with the same
initial and final states of all atoms.8 Note that there is no need for the initial and final states
8 Not taking this into account may lead to incorrect results, see for example [81], corrected in [82].

57
 |Fe me i
...
ε′

ε ε′′

...

(a) |Fg mg i (b)

Figure 25: (a) Light scattering by a degenerate atomic dipole transition Fg ↔ Fe can either preserve
spin (Rayleigh transition, full arrows) or change the spin (degenerate Raman transition, dotted arrows).
(b) Multiple light scattering by randomly placed atoms with internal spin states involves depolariza-
tion/decoherence from both spin-orbit (transversality) and spin-flip effects.

to be identical, so that Raman transitions can very well contribute to interference terms, if and
only if the same Raman transitions occur along the interfering paths.9 A commonly encountered
situation is that the degeneracy of the atomic ground state is due to its non-zero total angular
momentum, see Fig. 25(a): Raman transitions then involve different Zeeman sub-states. The
detailed calculation of the scattering vertex requires to incorporate also the angular momentum,
i.e. the polarization, of the light. Then, the whole structure of the diffuson and the Cooperon
boils down to various kernels of type (143), where the various depolarization/decoherence rates
are rotational invariants that depend only on the angular momenta Fg , Fe [70, 83].
For a typical alkali atom like 85 Rb with a Fg = 3 → Fe = 4 resonance line, the longest
19
decoherence time for the Cooperon is τφ = 21 τl [72], meaning that the CBS interference is
quite efficiently killed already by very few scattering events. An immediate consequence is that
the CBS cone observed on a cold Rb gas has a much reduced enhancement factor [85]. A less
trivial feature is that the best Rb CBS signal is not observed in the same channels than with Sr.
Detailed calculations can be performed and an excellent agreement between the measured and
the calculated CBS signals is observed [84], see Fig. 26.
The internal atomic degrees of freedom are here responsible for the loss of coherence. Infor-
mation flows from the light to the atoms; as long as we do not precisely measure the internal state
of each atom, this information is lost and the interference contrast is reduced. This information-
theoretic argument can be made quantitative by investigating how much which-path information
is stored within the atomic internal degrees of freedom. A quantitative measure of this wave-
particle-duality, developed originally in the context of Mach-Zehnder-type interferometers [86],
can be investigated analytically in the simplest cases and highlights the rôle of which-path infor-
mation in the loss of CBS interference visibility [87].
A simple way to restore phase coherence is to lift the atomic degeneracy by applying an
external magnetic field. Fields as small as a few Gauss are enough to detune some of the
atomic transitions far from resonance, thus reducing the effect of Raman transitions. It has
been experimentally observed and theoretically explained how this can increase the enhancement
factor [88]. We here face a seemingly paradoxical situation (in view of the negative magneto-
resistance discussed in Sec. 7), where adding an external magnetic field, which should break the
time-reversal symmetry, has the effect of increasing the interference between time-reversed paths!
9 Thus, one must take statements like “Raman scattered light is incoherent”, often made by quantum opticians,

with great care. It is true that Raman scattered light does not interfere with the incoming reference beam—because
the final states of the atom are different—but a single Raman-scattered photon along two different paths does
very well interfere with itself.

58
 1.20 1.20

1.15 1.15

CBS signal

CBS signal
1.10 1.10

1.05 1.05

1.00 1.00
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
θ (mrad) θ (mrad)
1.20 1.20

1.15 1.15

CBS signal
CBS signal

1.10 1.10

1.05 1.05

1.00 1.00
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
θ (mrad) θ (mrad)

Figure 26: CBS by a cloud of cold Rubidium atoms, in four polarization channels (upper/lower
left: parallel/perpendicular linear polarization; upper/lower right: circular polarization with non-
preserving/preserving helicity). Solid red line: calculation taking into account the geometry of the
atomic cloud. [84]. The enhancement factor is strongly reduced compared to ideal Rayleigh scatter-
ers such as Strontium, Fig. 24. This is due to the internal Zeeman structure of the Rubidium atom,
Fig. 25a). Note in particular that the helicity preserving channel—where the largest enhancement is
found for point scatterers such as Strontium—gives here the smallest enhancement.

Similarly, strong magnetic fields in electronic samples have been used to align free magnetic
impurities, reduced thus spin-flip effects and restore Aharonov-Bohm interference [89, 90].

7 Weak localization (WL)

As discussed in the preceding section, the Cooperon is responsible for enhanced backscatter-
ing, which implies an increased probability to return to the starting point. In the bulk of a
disordered system, diffusive transport is thus hindered. This phenomenon, known as weak lo-
calization, is quantitatively expressed by a reduction of the diffusion constant (or dimensionless
conductance/conductivity) with respect to the classical diffusion constant expected for phase-
incoherent transport.
The weak-localization effect of the Cooperon is expressed by eq. (128). For the sake of
concreteness, we will take in the following quantitative estimates the example of atomic matter
waves with a quadratic dispersion relation ε = ~2 k 2 /2m. The free density of states (103) is

Sd mk d−2
N0 (ε) = (148)
(2π)d ~2

where Sd = 2π d/2 /Γ(d/2) is the area of the unit sphere in dimension d: S1 = 2, S2 = 2π, S3 = 4π.
The Boltzmann diffusion constant DB = ~kl/dm is directly proportional to the transport mean
free path l. Because the Cooperon is isotropic, the d-dimensional integral in (128) can be reduced

59
to a trivial (d − 1)-dimensional angular integral and a radial integral over momentum q, such
that  Z ∞ d−1 
1 1 ~ q dq
= 1+ . (149)
D DB πmk d−2 DB 0 q 2 − i0
The result of the q-integral depends crucially on the dimensionality of the system. This is a
consequence of the fact, well known from classical random walks, that the return probability to
the origin is the higher, the lower the spatial dimension d. Therefore, weak—and consequently
also strong—localization are immediately seen to have the largest impact in low dimensional
systems.

7.1 d=1
In dimension d = 1, the integral (149) diverges for small q. However, for a system of size L, the
momentum q cannot take arbitrary small values, and a lower cutoff of the order of 1/L must be
used. A simple way of implementing it—following the recipe of Sec. 6.3 for including decoherence
effects—consists in replacing q 2 by q 2 + 1/L2 . One then gets:
 
1 1 L
= 1+ . (150)
D DB 2l

This expression is valid only if the weak localization contribution is a small correction, i.e. for
L  l. To lowest order, we recover D ≈ DB (1 − L/2l), which is the exact result (49) already
derived for 1d systems in Sec. 3.3. The interest of the present approach is that a full microscopic
theory provides us with the weak localization correction and thus puts the scaling theory of
localization on firm grounds.

7.2 d=2
In dimension 2, the integral diverges both for small and large q. A suitable cutoff at small q
is again 1/L, the inverse of the system size. Diffusive transport is a long time, large distance
behavior. It is not expected to give an accurate description on a scale shorter than the mean
free path. Performing the integral with a natural cutoff 1/l at large q thus leads to
  
2 L
D ≈ DB 1 − ln . (151)
πkl l

In terms of the dimensionless conductance g = 2mD/~, this implies the following scaling relation:
d ln g 2
β(g) = =− . (152)
d ln L πg
Our microscopic calculation thus gives an explicit prediction that can be readily incorporated
into scaling theory, as anticipated in Sec. 3.6.

How can weak localization be observed experimentally? A priori, any measured diffusion
constant incorporates already all interference corrections to the classically expected value. For-
tunately, the Cooperon contribution to weak localization is due to the constructive interference
between a multiply scattered path and its time-reversal. If one breaks time-reversal symmetry
on purpose, then the delicate interference is likely to disappear, and an enhancement of diffusive
transport should be observed.

60
Figure 27: Experimentally measured resistance of a thin Mg film exposed to a perpendicular magnetic
field. The magnetic field breaks the constructive interference of waves counter-propagation along closed
loops, reduces the weak localization effect, and thus results in negative magneto-resistance. Similarly,
larger temperatures reduce the phase coherence of the electronic wavefunction, and also reduce weak
localization corrections. Adapted from [91] (courtesy of G. Bergmann).

For charged particles—such as electrons in solid state samples—the simplest way is to add a
magnetic field perpendicular to the Rsample. In the presence of a vector potential A,~ a charged
particle picks an additional phase eA ~ · d~l/~ along a closed loop. This is nothing but e/~
times the enclosed magnetic flux. Along each closed loop, this additional phase appears in the
Cooperon contribution. If this phase fluctuates largely from one loop to the other, the resulting
interferential contribution will vanish. As the smallest area enclosed by a diffusive loop is l2 ,
the weak localization correction is expected to vanish above B ≈ ~/el2 . For a typical mean free
path of a fraction of µm, this is in the Tesla range. Figure 27 shows the measured resistance of
a 2D Mg film vs. magnetic field at various temperatures [91]. At the lowest temperature, prop-
agation is almost fully phase coherent and one observes a decreasing resistance, i.e. a increasing
conductance, when a magnetic field is applied. This negative magneto-resistance was a mystery
when first observed and only later explained as a manifestation of weak localization. When tem-
perature increases, the phase coherence of the electrons diminishes, and the weak localization
correction gets smaller. This is, in a different context, analogous to decoherence phenomena
discussed for coherent backscattering in Sec. 6.3. Note that, from such experimental data, it is
possible to measure the transport mean free path (via the width of the weak localization peak) as
well as the temperature-dependence of the decoherence time. In recent times, weak localization
measurements have been used as very sensitive detectors for minute concentrations of magnetic
impurities, which induce spin-flip decoherence and are responsible for finite decoherence times
even at zero temperature [92, 90, 93].

7.3 d=3
In dimension 3, the integral in eqn (149) requires only a cutoff at large q, which we take again
as 1/l and obtain  
3
D ≈ DB 1 − . (153)
π(kl)2
This diffusive Cooperon contribution to weak localization is found to scale as 1/(kl)2 . However,
this is not the whole story, because other diagrams, not included in the simple diffusive Cooperon,
give contributions that are actually more important for small disorder kl  1. Just as for the

61
CBS cone discussed in Sec. 6.1, also here the double-scattering diagrams appearing in eqn (117)
contribute to leading order 1/kl. In d = 3, the static electronic conductivity was found to be
given by [36]
σ(ω = 0) 2π π2 − 4
=1− − ln(kl) + O((kl)−2 ). (154)
σ̊ 3kl (kl)2
As long as kl  1, the weak localization is only a small correction, again providing us with a
macroscopic ground for the scaling theory of localization. It also gives an approximate criterion
for the onset of Anderson localization, which should set in approximately when the right hand
sides of eqns (153) or (154) vanish, i.e. (kl)c = O(1). This is Ioffe-Regel criterion, eq. (57).
However, the precise calculation of the critical point is a delicate endeavor. What precisely
happens at the 1/l scale is not universal. The same is true for the Ioffe-Regel criterion, but the
latter nonetheless yields a first estimate on where to expect the Anderson transition.

7.4 Self-consistent theory of localization

Weak localization describes how diffusive transport is affected by interference. In essence, how-
ever, weak localization is a perturbative result: first, because the Cooperon contribution is
evaluated using a diffusive kernel valid in the absence of interference; second, because this simple
approach takes into account only a specific type of diagrams. The first assumption is especially
questionable in 1d, where diffusive transport actually never occurs, because localization appears
at the very same scale (the localization length) than diffusion (the mean free path). Concerning
the second point, the dominant rôle of the Cooperon in large systems was recognized already by
Gorkov et al. [94] and Abrahams et al. [19]. However, a weak-disorder perturbation theory in
powers of 1/kl alone would never be able to describe the Anderson transition (in 3d) for strong
disorder, nor the crossover from weak to strong localization in 1d and 2d systems.
The self-consistent theory of localization, developed by Vollhardt and Wölfle in the 1980s
[95, 96, 97], is an attempt to escape this seemingly hopeless situation by applying a suitable self-
consistency scheme, as often employed with success to describe phase transitions in statistical
physics. Rather than a theory with rigorously controlled approximations, it must thought of
as a guess, albeit highly educated, about the most important contributions of diagrams to all
orders. The basic observation is that the diffusive contribution of large closed loops in eq. (149)
must itself be modified by weak localization: inside a large loop, the wave explores smaller loops,
leading to a decreased diffusion constant for propagation along the large loop. This argument
can of course be repeated: one should take into account loops within loops within loops..., all
the way down to the smallest loops, stopping at the scale of the transport mean free path.
The whole description must now be self-consistent, describing what happens at every scale
from the mean free path up to the size of the system—or toward infinity in the bulk. The simplest
idea would be to replace the static Boltzmann diffusion constant DB in the integral of (149) by
the renormalized diffusion constant D itself, thus providing us with an implicit equation for D.
It turns out that this is not enough: indeed, a single number—the static diffusion constant D—
cannot describe the full dynamics both for short times, where it is diffusive, and for long times
where localization may eventually set in. So we require a scale-dependent diffusion constant, and
it turns out that it is simpler to consider various time scales rather than various spatial scales.
We thus consider a diffusion constant D(ω) which depends on frequency ω. The self-consistent
expression for D(ω) just derives from (149) by re-introducing the ω dependence and replacing
DB by D(ω) in the integral:

q d−1 dq
Z
~
D(ω) + = DB . (155)
πmk d−2 q 2 − (iω/D(ω))

62
 1 1
D/DB
d=1 d=2
0.5 0.5

0 0
0 1 2 0 1 2
−iωτl −iωτl

Figure 28: Diffusion constant vs. (imaginary) frequency, in units of Boltzmann diffusion constant and
mean free time, respectively, as predicted by the self-consistent theory of localization, in 1d (left) and
2d (right, for kl=1.5). At large ω (short time), one recovers the classical Boltzmann diffusive behavior.
At small ω, the dependence is linear, which implies exponential localization in configuration space, in
agreement with scaling theory, numerical and experimental observations (in 1d). In 2d, the localized
regime is reached at much smaller frequency because the localization length and time are exponentially
large.

In the short-time limit ω → ∞, the contribution of the integral vanishes and one gets back to
classical Boltzmann diffusive propagation, as expected. The most interesting part takes place at
long times, i.e. in the limit ω → 0, whose consequences again depend crucially on the dimension.

7.4.1 d=1
π
p
At finite ω, the integral in (155) does not need any regularization, it is simply 2 D(ω)/(−iω),
and the implicit equation for D(ω) is easily solved [98]:
√
D(ω) 1 − 16iωτl − 1
=√ (156)
DB 1 − 16iωτl + 1
where τl = l2 /DB is the mean free time between two scattering events. This function is plotted
in the left panel of Fig. 28 as function of −iω.10 In the limit of small ω, it behaves linearly
D/DB ≈ −4iωτl . This in turns implies that the propagation kernel 1/(−iω + D(ω)q 2 ) is just
1/(−iω) × 1/(1 + 4l2 q 2 ). When going back from momentum to configuration space by inverse
Fourier transform, it implies that the intensity kernel is proportional to exp(−|z|/2l) at long
times. It successfully describes exponential localization with the localization length ξloc = 2l, i.e.
the exact result for the localization length! The elementary ingredients used for obtaining this
important result are: quantum kinetic theory, microscopic calculation of the weak localization
correction in the perturbative regime and its self-consistent extension. That the exact result is
eventually obtained is a strong hint that the self-consistent approach catches an important part
of the physics of localization.
But beware! This triumph is somewhat tarnished by the fact that it is the typical intensity
that decays with ξloc , whereas the average intensity calculated here should asymptotically decay
with 4ξloc , as shown in Sec. 2.5. The precise source for this discrepancy escapes our
present
understanding. Clearly, the self-consistent theory is built for the average intensity kernel GR GA
10 Imaginary frequency is only used for convenience, as it makes the diffusion constant purely real and thus

easier to plot. The most important transport property is the small-|ω| behavior, which is linear in the localized
regime, both for real or imaginary ω.

63
and thus cannot describe the huge fluctuations in the localized regime. One lacks a diagrammatic
expansion for the typical transmission, which would require to calculate contributions of advanced
and retarded Green functions to all orders.
Decoherence effects can be easily included in the self-consistent approach by the replacement
−iω 7→ −iω + 1/τφ explained in Sec. 6.3. In Figure 28, this replacement simply translates
the curve horizontally to the left. One immediately finds that the diffusion constant no longer
vanishes at ω = 0, but takes a finite value, implying diffusive motion at long times. In the limit of
2
weak decoherence τl  τφ , the residual diffusion constant is D ≈ 4τl DB /τφ = ξloc /τφ . It is much
smaller than the Boltzmann diffusion constant and allows for a simple physical interpretation: a
phase-breaking event, occurring on average every τφ , destroys the delicate interference responsible
for localization. This implies a restart of diffusion during time τl after which localization sets in
again, until the next phase breaking event, etc.

7.4.2 d=2
In 2d, the integral in (155) diverges in the large-q limit, requiring a regularization. The natural
short-distance cut-off is the mean free path l. Elementary manipulations shows that D(ω) is
implicitly determined by
 
D(ω) 1 D(ω) 1
=1− ln 1 − . (157)
DB πkl DB 2iωτl

In contrast with the 1d case, D/DB is not a universal function, it depends on the parameter
kl. The right panel of Fig. 28 plots it for kl = 1.5. It displays the classical diffusive behavior
D ≈ DB at large ω (short times), and localization at long times. Indeed, for ω → 0, one finds
2
D(ω) ≈ −iωξloc , i.e. exponential localization with the localization length
 
p πkl
ξloc = l exp (πkl) − 1 ≈ l exp . (158)
2

This provides us with a microscopic derivation of the result of scaling theory, eq. (56). The
self-consistent approach describes correctly the exponentially large localization length in 2d.11
Note that, even for strong disorder with a rather small value kl = 1.5, the linear regime in Fig. 28
is observed only at very small ω, i.e. for very long times.
Decoherence can be taken into account exactly like in 1d. Instead of a true metal-insulator
transition, one observes a cross-over from classical diffusion at large kl towards a residual diffusion
(triggered by decoherence) at small kl. Explicit calculations have been carried out in [99] for the
case of atomic matter waves in a speckle potential, where residual spontaneous emission is one
source of decoherence that can be experimentally tuned. Figure 29 shows typical results. Because
of the exponential dependence in 2d, the cross-over from quasi-localized behavior at small k to
diffusive behavior at large k is rather rapid. In any case, a crucial requirement is to have very
cold atoms, with de Broglie wavelength shorter than the speckle correlation length.
Considering an expanding BEC wave packet released from a harmonic trap, one can calculate
the expected stationary (for negligible decoherence) density distribution along the lines of (63).
Just as in 1d, the asymptotic decay
is governed by the wave vector kmax of the fastest atoms,
and the density is predicted to be |ψ(r)|2 ≈ Cr−5/2 exp{−r/ξloc (kmax )} [99].


64
 1

0.8
V0 = 0.2Eζ
0.6 V0 = 0.1Eζ
D
V0 = 0.05Eζ
DB
0.4

0.2 kc

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
kζ

Figure 29: Diffusion constant (normalized to the Boltzmann diffusion constant) computed from the self-
consistent theory of localization, for atomic matter waves with wave vector k exposed to a 2d speckle
potential with correlation length ζ and different amplitudes V0 . A stronger potential means a smaller
value of kl. Dashed lines are the prediction of the simple perturbative weak localization correction,
eqn (151), solid lines the result of the self consistent approach, eqn (157), including residual decoher-
ence due to spontaneous emission implemented via eqn (142). A rather sharp cross-over between the
Boltzmann diffusive behavior at high energy and the quasi-localized behavior at low energy is observed
around a critical value kc [99].

D/DB
0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.01 0.02
−iωτl
Figure 30: Diffusion constant vs. (imaginary) frequency, in units of Boltzmann diffusion constant and
mean free time, respectively, computed from the self-consistent theory of localization in 3d, eqn (159).
Dotted red line: metallic regime kl = 1.2 with finite diffusion constant at long times. Dashed blue line:
insulating regime kl = 0.8 with a finite localization length. Solid black line: critical point kl = (kl)c =
3/π of the metal-insulator Anderson transition, where the diffusion constant scales like (−iω)1/3 ,
p

implying an anomalous diffusion hr2 (t)i ∝ t2/3 at long times.

65
7.4.3 d=3
In 3d, the same short-distance regularization than in 2d is necessary, leading to the following
implicit equation, valid in the limit ωτl  1:12
s !
D(ω) 3 π −3iωτl
+ 1− = 1. (159)
DB π(kl)2 2 D(ω)/DB

The behavior of the solution, shown in Fig. 30, depends on the Ioffe-Regel parameter kl and
defines three distinct regimes:
p
Diffusive regime: For kl > (kl)c = 3/π, D/DB tends to a constant value in the limit ω → 0,
which means that the system behaves always diffusively, albeit with a diffusion constant smaller
than the Boltzmann diffusion constant. This is the regime of weak localization.

Localized regime: For kl < (kl)c , it is easy to see that D/DB → 0 in the limit ω → 0. More
precisely, one has exponential localization:

2 1
D(ω) ≈ −iωξloc with ξloc ∼ (160)
(kl)c − kl
immediately below the Anderson transition, on the insulating side. This means, see eq. (58),
that the critical exponent deduced from the self-consistent approach is ν = 1, quite far from the
true value ν = 1.58 known from numerical simulations [26]. The reason lies in the approximate
character of the self-consistent approach, which disregards the huge fluctuations in the vicinity
of the critical point. Field-theoretic approaches can in principle capture the effect of fluctuation
and have been quantitatively tested in d = 2 +  dimensions [15]. For d = 3, however, to our
knowledge no analytical theory is available that makes a better quantitative prediction than the
self-consistent theory.

Critical regime: At the critical point kl = (kl)c , it is easy to see that the solution of eq. (159)
scales like D(ω) ∼ (−iω)1/3 . Consequently, the critical behavior is anomalous diffusion where the
squared extension increases subdiffusively at long times: hr2 (t)i ∝ t2/3 . This anomalous diffusion
has been experimentally observed with the quasi-periodically kicked rotor, see Sec. 8.

8 Kicked rotor
The physics of Anderson localization is, as amply discussed in the preceding sections, highly
dependent on the dimension of the system. While the 1d situation is fairly well understood—
localization is the generic behavior, the localization length is comparable to the mean free path,
and the fluctuation properties in the localized regime are essentially well understood—the physics
of higher dimensions is much richer still. Dimension 3 is especially interesting, as one expects a
so-called mobility edge, separating, in the continuum case, localized states at low energy/strong
disorder from extended states at high energy/weak disorder.
As explained in Sec. 4.3, it is very difficult to find a clean experimental system to observe
this metal-insulator Anderson transition unambiguously. Cold atomic matter waves are very
11 Thesame caveats than in 1d exist, concerning average versus typical quantities.
12 Thisformula features only the static Cooperon contribution, which suffices for qualitative predictions, but
should be extended to include the full interference terms appearing in (154) when quantitative precision is neces-
sary.

66
attractive because they can be directly observed, and because most experimental imperfections
as well as atom-atom interactions can be precisely controlled, if not reduced to a minimum. The
main difficulty consists in reaching the Ioffe-Regel threshold kl = O(1), eq. (57), i.e. preparing a
sufficiently small k (low energy, large de Broglie wavelength) and short mean free path l. Indeed,
the latter cannot be shorter than the correlation length of the disordered potential, i.e. of the
order of 1 µm for optical speckle.
This limitation on the mean free path can be overcome using a different approach, where
disorder is not provided by an external potential in configuration space, but by classically chaotic
dynamics in momentum space. This idea has been realized experimentally with the atomic kicked
rotor, and Anderson localization in 1d has been observed as early as 1994 [100], 14 years prior
to the widely noticed Anderson localization in configuration space [4]! A key advantage of the
kicked rotor is that is does not require ultra-cold atoms from a Bose-Einstein condensate: a
standard magneto-optical trap suffices to prepare the initial state. The kicked rotor also has
permitted the clean observation of the metal-insulator Anderson transition in 3D, and the first
experimental measurement of the critical exponent, with non-interacting matter waves [101].

8.1 The classical kicked rotor

We consider a one-dimensional rotor whose position can be described by the angle x (defined
modulo 2π) and the associated momentum p, and kick it periodically with a position-dependent
amplitude. In properly scaled units, the Hamiltonian function can be written as
+∞
p2 X
H= + k cos x δ(t − nT ) (161)
2 n=−∞

where T and k are period and strength of the kicks, respectively.

Because of the time-dependence, energy is not conserved, but thanks to the time-periodicity,
we can analyze the motion stroboscopically and build a Poincaré map picturing the evolution
once every period. This map relates the phase space coordinates just before kick n + 1 to the
coordinates just before kick n:

In+1 = In + K sin xn
(162)
xn+1 = xn + In+1

where K = kT and In = T pn . This is nothing but the celebrated standard map (also known as
the Chirikov map) that has been widely studied [102, 103]: it is almost fully chaotic and ergodic
around K = 10 and above.
When the stochasticity parameter K is very large, each kick is so strong that the positions
of the consecutive kicks can be taken statistically uncorrelated. By averaging, one thus gets:
k2
hp2n+1 i ' hp2n i + k 2 hsin2 xn i ' hp2n i + (163)
2
It follows that the motion in momentum space is diffusive (hp2 i increases linearly with time) with
diffusion constant
k2
D= . (164)
2T
Numerical experiments [102] show that this expression works well for K ≥ 10. Note that the
kicked rotor is a perfectly deterministic system, without any randomness. It is the chaotic
nature of the classical motion, and thus its extreme sensitivity to perturbations, which renders
the deterministic classical motion diffusive on average.

67
8.2 The Quantum Kicked Rotor
The quantum Hamiltonian is obtained from the classical one, eq. (161), through the canonical
replacement of p by −i~∂x . The evolution operator over one period is the product of the free
evolution operator and the instantaneous kick operator:

i p2 T
   
i
U = U (T, 0) = exp − exp − k cos x (165)
~ 2 ~

The long-time dynamics is generated by successive iterations of U. Thus, one can use the eigen-
states of U as a basis set. U being unitary, its eigenvalues are complex numbers with unit
modulus:  
iEi T
U |φi i = exp − |φi i (166)
~
with 0 < Ei ≤ 2π~/T are defined modulo 2π~/T . They are not exactly the energy levels of the
system—the |φi i are not stationary states of the time evolution, but are only periodic—and are
called quasi-energy levels, the |φi i being the Floquet eigenstates. This Floquet description is the
time-analog of the Bloch theorem that applies to spatially periodic potentials.

8.3 Dynamical Localization

The quantum dynamics of the kicked rotor can be quite simply studied numerically by repeated
application of the one-period evolution operator U to the initial state, alternating free propaga-
tion phases with instantaneous scattering events
in momentum space induced by the kicks. The
free evolution between kicks, exp −ip2 T /2~ , is diagonal in momentum representation, such
that each momentum eigenstate, characterized by its momentum m~ with integer m, picks up
a different phase shift. The kick operator exp (ik cos θ/~), in contrast, is diagonal in position
representation and couples different momenta. Being unitary, it plays the role of a scattering
matrix in momentum space and contains the quantum amplitude for changing an incoming initial
momentum in an outgoing one, under the influence of one kick; k is the parameter controlling
the scattering strength. The dynamics of the kicked rotor can be seen as a sequence of scattering
events interleaved with free propagation phases.
For sufficiently large K = kT , the classical dynamics is diffusive in momentum space, but it
should come as no surprise to the reader now familiar with 1d Anderson localization, that the
quantum dynamics may be localized at long times. This localization was baptized “dynamical
localization” when it was observed in numerical simulations [104]. Only later, people realized
that it is nothing but the Anderson scenario of 1d localization, as explained below.
Dynamical localization has been experimentally observed in the dynamics of a Rydberg elec-
tron exposed to an external microwave field [105]. Arguably the simplest observation uses a cold
atomic gas, prepared in a standard magneto-optical trap with a typical velocity spread of few
recoil velocities [100, 106, 101]. After the trap is switched off, a periodic train of laser pulses
is applied to the atoms. Each pulse is composed of two far-detuned counter-propagating laser
beams producing a spatially modulated optical potential. Each laser pulse thus produces a kick
on the atom velocity, whose amplitude is proportional to the gradient of the optical potential.
If the kicks are infinitely short, we recover exactly the kicked rotor, eq. (161), where the
position of the atom in the standing wave plays the role of the x variable and its velocity is the
p variable. The kick strength k is proportional to the laser intensity divided by the detuning.
The spatial dimensions perpendicular to the laser beams do not play any role in the problem,
so that we have an effectively one-dimensional time-dependent problem. The mapping of the
dimensionfull Hamiltonian for cold atoms to the kicked rotor Hamiltonian, eq. (161), shows that

68
Figure 31: Experimental time evolution of the momentum distribution of the atomic kicked rotor [100],
from the initial Gaussian distribution until the exponentially localized distribution at long time; N is
the number of kicks (courtesy of M. Raizen).

2
the effective Planck’s constant of the problem [107] is ~eff = 4~kL T /M = 8ωr T where kL is the
laser wavenumber and M the atomic mass. Up to a numerical factor, it is the ratio of the atomic
recoil frequency ωr to the pulse frequency, and can be easily varied in the experiment, from the
semiclassical regime ~eff  1 to the quantum regime ~eff ∼ 1.
After the series of pulses is applied, the momentum distribution is measured either by a
time of flight technique [100] or velocity selective Raman transitions [101]. Figure 31 shows
the momentum distribution as a function of time. While, at short time, the distribution is
Gaussian—as expected for a classical diffusion—, its shape changes around the localization time
and evolves toward an exponential shape exp(−|p|/ξloc ) at long time, a clear-cut manifestation
of Anderson/dynamical localization.
Adding decoherence on the system—either by adding spontaneous emission [106] or by weakly
breaking the temporal periodicity [108]—induces some residual diffusion at long time, in accor-
dance with the discussion in Sec. 6.3 and 7.4. This is another proof that dynamical localization
is based on delicate destructive interference.

8.4 Link between dynamical and Anderson localizations

So far, we have only made plausible that dynamical localization with the quantum kicked rotor
is similar to Anderson localization in a spatially disordered medium. We now demonstrate
the connection between the two phenomena, following [109]. Consider the evolution operator,
eq. (165), and the associated eigenstate |φi with quasi-energy E. The part of the evolution
operator associated with the kick can be written as:
 
i 1 + iW (x)
exp − k cos x = (167)
~ 1 − iW (x)

where W (x) is a periodic Hermitean operator which can be Fourier-expanded:

∞
X
W (x) = Wr exp (irx). (168)
r=−∞

69
Similarly, the kinetic part can be written as:
i p2
   
1 + iV
exp − −E T = (169)
~ 2 1 − iV
The operator V is diagonal is the eigenbasis of p, labeled by the integer m (see above). If one
performs the following expansion in this basis set,
1 X
|φi = χm |mi, (170)
1 − iW (x) m

it is straightforward to show that the eigenvalue-equation (166) can be rewritten as

X
m χm + Wr χm−r = −W0 χm (171)
r6=0

where
E − 21 m2 ~2 T /2~ .


m = tan (172)
Equation (171) is the time-independent Schrödinger equation for a one-dimensional Anderson
model, cf. (68), with site index m, on-site energy m , coupling Wr to the nearest sites and total
energy W0 . Compared to (68), there are two new ingredients: firstly, there are additional hopping
amplitudes to other neighbors. But since they decrease sufficiently fast at large distance, they
do not play a major role. Secondly, the m values, determined deterministically by (172), are
not really random variables, but only pseudo-random13 with a Lorentzian distribution.14 Still,
localization is expected and indeed observed. The computation of the localization length follows
the general lines explained in section 5, and is in good agreement with experimental observations.
It should be emphasized that space and time play different roles in the Anderson model
and in dynamical localization. What plays the role of the sites of the Anderson model are the
momentum states. This is why dynamical localization is not observed in configuration space,
but in momentum space.

8.5 The quasi-periodically kicked rotor

How can the kicked rotor be used to study Anderson localization in more than one dimension?
The first idea is to use a higher-dimensional rotor with a classically chaotic dynamics and to
kick it periodically. It turns out that this is not easily realized experimentally, as it requires
to build a specially crafted spatial dependence[110]. Yet, remember that time and space have
switched roles, and so a simpler idea is to use additional temporal dimensions rather than spatial
dimensions. Instead of kicking the system periodically with kicks of constant strength, one may
use a temporally quasi-periodic excitation. Various schemes have been used [111], but the one
allowing to map on a multi-dimensional Anderson model uses a quasi-periodic modulation of the
kick strength, the kicks being applied at fixed time interval [112].
We will be interested in a 3d Anderson model, obtained by adding two quasi-periods to the
system:15
p2 X
Hqp = + K(t) cos x δ(t − n) , (173)
2 n
13 As is well known, “random-number generators” implemented in computers also generate deterministic, merely

pseudo-random sequences; most of them use formulae analogous to (172).

14 The non-random character appears for example, when the product ~T /2 is chosen as an integer multiple of

2π. Then all m are equal, the motion is ballistic and localization is absent. Similarly, when ~T is commensurate
with π (the so-called quantum resonances), the sequence m becomes periodic, and Anderson localization does
not take place, giving way to Bloch-band transport.
15 In this section, we take the kicking period T as unit of time

70
with
K(t) = K [1 + ε cos (ω2 t + ϕ2 ) cos (ω3 t + ϕ3 )] . (174)
Now where is the three dimensional aspect in the latter Hamiltonian? The answer lies in a formal
analogy between this quasi-periodic kicked rotor and a 3d kicked rotor with the special initial
condition of a “plane source”, as follows.
Take the Hamiltonian of a 3d, periodically kicked rotor:

p21 X
H= + ω2 p2 + ω3 p3 + K cos x1 [1 + ε cos x2 cos x3 ] δ(t − n), (175)
2 n

and consider the evolution of a wavefunction Ψ with the initial condition

Ψ(x1 , x2 , x3 , t = 0) ≡ ψ(x1 , t = 0)δ(x2 − ϕ2 )δ(x3 − ϕ3 ). (176)

This initial state, perfectly localized in x2 and x3 and therefore entirely delocalized in the con-
jugate momenta p2 and p3 , is a “plane source” in momentum space [98]. A simple calculation
shows that the stroboscopic evolution of Ψ under (175) coincides exactly with the evolution of
the initial state ψ(x = x1 , t = 0) under the Hamiltonian (173) of the quasi-periodically kicked
rotor (for details, see [107]). An experiment with the quasi-periodic kicked rotor can thus be seen
as a localization experiment in a 3d disordered system, where localization is actually observed
in the direction perpendicular to the plane source. In other words, the situation is comparable
to a transmission experiment where the sample is illuminated by a plane wave and the expo-
nential localization is only measured along the wave vector direction. Therefore, the behavior of
the quasi-periodic kicked rotor (173) matches all dynamic properties of the quantum 3d kicked
rotor.
The classical dynamics has been shown to be a chaotic diffusion, provided the parameter ε is
sufficiently large to ensure efficient coupling between the 3 degrees of freedom [113]. As for the
standard 3d kicked rotor (175), its quantum dynamics can be studied using the Floquet states
via mapping to a 3d Anderson-like model:
X
 m Φm + Wr Φm−r = −W0 Φm , (177)
r6=0

where m ≡ (m1 , m2 , m3 ) labels sites in a 3d cubic lattice, the on-site energy m is

m1 2
   
1
m = tan ω− ~ + ω2 m2 + ω3 m3 , (178)
2 2
and the hopping amplitudes Wr are the Fourier expansion coefficients of

W (x1 , x2 , x3 ) = − tan [K cos x1 (1 + ε cos x2 cos x3 )/ 2~] . (179)

A necessary condition for localization is obviously that m not be periodic. This is achieved
if (~, ω2 , ω3 , π) are incommensurate. When these conditions are verified, localization effects
as predicted for the 3d Anderson model are expected, namely either a diffusive or a localized
regime. Localized states would be observed if the disorder strength is large compared to the
hopping. In the case of the model (177), the amplitude of the disorder is fixed, but the hopping
amplitudes can be controlled by changing the stochasticity parameter K (and/or the modulation
amplitude ε): Wr is easily seen to increase with K. In other words, the larger K, the smaller
the disorder. One thus expects to observe diffusion for large stochasticity K and/or modulation
amplitude ε (small disorder) and localization for small K and/or ε (large disorder). It should

71
 1000
hp2 (t)i
800

600

400

200

0
0 50 100
time (number of kicks)

Figure 32: Experimentally measured temporal dynamics of the quasi-periodically kicked rotor, for
increasing values of the kick strength. The average kinetic energy hp2 (t)i tends to a constant in the
localized regime (lower curve, K = 4, ε = 0.1), increases linearly with time in the diffusive regime
(upper curve, K = 9, ε = 0.8). At the critical point K = Kc ≈ 6.4 (middle curve), anomalous diffusion
hp2 (t)i ∼ t2/3 (dashed curve) is clearly observed.

be emphasized that stricto sensu there is no mobility edge in our system that would separate
localized from delocalized eigenstates. Depending on the parameters K, ~, ε, ω2 , ω3 , either all
Floquet states are localized or all are delocalized. The boundary of the metal-insulator transition
is in the (K, ~, ε, ω2 , ω3 )-parameter space. As seen below, K and ε are the primarily important
parameters.
In the experiment performed at the University of Lille [101], kicks are applied to atoms with
an initially narrow momentum distribution, and the final momentum distribution is measured
using velocity-selective Raman transitions.16 Figure 32 shows the experimental data. For large
disorder, one clearly sees the initial diffusive phase and the freezing of the quantum dynamics in
the localized regime (lower curve). In the diffusive regime (upper curve), hp2 (t)i is seen to increase
linearly with time. The intermediate curve displays an anomalous diffusion hp2 (t)i ∼ t2/3 . The
anomalous exponent 2/3 is exactly the prediction of the self-consistent theory of localization,
section 7.4.3, which also fully agrees with the scaling theory of localization. Time here plays
the role of the system size L in the scaling theory: going to longer times means following the
renormalization flow in fig. 7. Only exactly at the unstable critical point will the anomalous
diffusion subsist for arbitrarily long times. At slightly larger (resp. smaller) K, the motion will
eventually turn diffusive (resp. localized) at long time. Experimental constraints prevent the
observation beyond 150-200 kicks. Numerical simulations may extend much beyond: it has been
checked that the anomalous diffusion with exponent 2/3 is followed for at least 108 kicks [107].
Since in numerical or experimental practice one always works in finite-size systems, we should
emphasize that there is an important difference between a true metal-insulator transition and
a cross-over between two limiting behaviors. For example, consider the simplest 1d situation
where the dynamics eventually localizes for sure, with a localization time depending on the kick
strength K. Over a finite experimental time, one may observe an apparently diffusive behavior if
16 Measuring the average hp2 (t)i is tedious and very sensitive to noise in the wings of the momentum distribu-
tion. It is much easier to measure the atomic population Π0 (t) at zero momentum. Because of atom number
conservation, hp2 (t)i is roughly proportional to 1/Π20 (t). The proportionality factor depends on the shape of the
distribution, but does not show large changes. As we are interested in scaling properties, 1/Π20 (t) or hp2 (t)i are
essentially equivalent.

72
 40
4000

3000 30

20
ξ
2000 ξ
1000 10

0 0
4 5 6 7 8 9 4 5 6 7 8 9
K K
Figure 33: Characteristic length (for localization in momentum space) extracted from numerical (left)
and real (right) experiments on the quasi-periodically kicked rotor, in the vicinity of the metal-insulator
Anderson transition [101]. Finite-size scaling is used. The characteristic length is proportional to the
localization length on the insulating side, and to the inverse of the diffusion constant on the metallic side.
It has an algebraic divergence 1/|K − Kc |ν at the transition, smoothed by finite size and decoherence,
which are more important in the real experiment (limited to 110 kicks) than in the numerical calculations
(up to one million kicks). In both cases, it is possible to extract a rather precise estimate of the critical
exponent ν ≈ 1.5.

the localization time is longer than the duration of the experiment.17 An intermediate situation
with the localization time comparable to the duration of the experiment could produce data
looking like anomalous diffusion. However, this could be only a transient behavior and a longer
measurement will eventually show localization. In contrast, the t2/3 behavior at the critical
point of the Anderson transition is not a transient behavior, it extends to infinity, highlighting
the scale-free behavior with fluctuations of all sizes present right at the critical point.
The unavoidable experimental limitation by finite size can also be turned into a powerful
tool of analysis. It is known as finite-size scaling [38] and has its roots in the scaling properties
observed in the vicinity of the transition. The idea is that all results, obtained for various
values of parameters and time, are described by a universal scaling law depending on a single
parameter, namely, the distance to the critical manifold. Close to the transition, there is only
one characteristic length (which diverges at the critical point) and all details below this scale are
irrelevant. Such an approach has been extremely successful to extract critical parameters from
numerical simulations of the Anderson model for various system sizes. The approach has been
transposed to the kicked rotor—see [107, 114] for details—and makes it possible to extract the
localization length (in momentum space) from numerical or experimental data acquired over a
restricted time interval.
The results are shown in Fig. 33 for both numerical simulations and experimental observa-
tions. One clearly sees the divergence of the characteristic length (the localization length on
the insulator side) in the vicinity of the transition. The divergence is smoothed by experimen-
tal imperfections and the finite duration of numerical and real experiments. The smoothing is
much more important in the latter case than in the former one, because the duration of the
real experiment (110 kicks maximum) is about 4 orders of magnitude shorter than in numerical
experiments. It is nevertheless possible to extract the critical exponent of the transition. For the
17 This also occurs for 1d Anderson localization in a speckle potential [4], as already pointed out in Sec. 5.2.4:

Because the localization length and time vary rapidly with energy, one observes localization at low energy and
apparently diffusive behavior at high energy. In between, an apparent mobility edge appears [20], which should
not be confounded with the true Anderson metal-insulator transition taking place in the thermodynamic limit,
although the experimental signatures may be similar.

73
numerical experiments, one finds ν = 1.58 ± 0.01 in perfect agreement with the best determina-
tion on the Anderson model. Moreover, it has been checked that this exponent is universal, i.e.
independent of the microscopic details such as the choice of the parameters ~, ω2 , ω3 [27]. This is
an additional confirmation that the transition observed is actually the metal-insulator Anderson
transition.
The critical exponent can also be determined—albeit with reduced accuracy—from the ex-
perimental data [101]. For the data of Fig. 33, one obtains:

νexp = 1.4 ± 0.3 (180)

These values are in excellent agreement with the numerical results. The key point is that the ex-
ponent significantly differs from unity, which is the value deduced from solid state measurements,
see Fig. 15, and the prediction of the self-consistent approach.
Since the atom-atom contact interaction in a cold dilute gas is much smaller than the electron-
electron Coulomb interaction in a solid sample, and since atoms are less easily lost than photons,
cold atoms appear particularly suitable for precise measurements of the Anderson transition.
Moreover, the possibility to picture wave functions directly opens the way to studies of fluctu-
ations in the vicinity of the critical point [113], and may even permit to observe multifractal
behavior [47] with matter waves. The flexibility of the kicked rotor could also be used to study
the Anderson transition in lower dimensions (by reducing the number of quasi-periods) or, why
not, even higher dimensions (by increasing it beyond 3). In any case, it is an attractive alternative
to experiments on spatially disordered systems.

References
[1] N. Goldenfeld and L. P. Kadanoff, “Simple Lessons from Complexity”, Science, 284, 87
(1999).
[2] P. W. Anderson, “More Is Different”, Science, 177, 393 (1972).

[3] W. Thirring, “Exact results for the scattering of three charged particles”, in Few Body
Systems and Nuclear Forces II, Lect. Notes Phys. 78, 353-361 (Springer, 1978).
[4] J. Billy, V. Josse, Z. Zuo, A. Bernard, B. Hambrecht, P. Lugan, D. Clément, L. Sanchez-
Palencia, Ph. Bouyer and A. Aspect, “Direct observation of Anderson localization of mat-
ter waves in a controlled disorder”, Nature 453, 891-894 (12 June 2008); Ph. Bouyer
et al., “Anderson localization of matter waves”, Proceedings of the XXI ICAP Conference,
R. Coté, Ph. L. Gould, M. Rozman and W. W. Smith eds., World Scientific (2008).
[5] P. W. Anderson, “Absence of Diffusion in Certain Random Lattices”, Phys. Rev. 109, 1492
(1958).

[6] P. A. Mello and N. Kumar, Quantum Transport in Mesoscopic Systems: Complexity and
Statistical Fluctuations (Oxford, 2004).
[7] Y. Imry, Introduction to mesoscopic physics (Oxford University Press, 2002).
[8] S. Datta, Electronic transport in mesoscopic systems (Cambridge University Press, 2002).

[9] C. W. J. Beenakker, “Random matrix theory of quantum transport”, Rev. Mod. Phys. 69,
731 (1997).

74
[10] P. A. Mello, “Theory of Random Matrices: Spectral Statistics and Scattering Problems”,
in: E. Akkermans et al. (eds), Mesoscopic Quantum Physics, Les Houches 1994 Session
LXI (North-Holland, Elsevier, 1995).
[11] S. Godoy, “Landauer diffusion coefficient: A classical result”, Phys. Rev. E 56, 4884 (1997).

[12] H. Furstenberg and H. Kesten, “Products of Random Matrices”, Ann. Math. Stat. 31,
457-469 (1960)
[13] N. G. van Kampen, Stochastic Proceses in Physics and Chemistry, (Elsevier, Amsterdam,
2007).
[14] A. A. Abrikosov, “The paradox with the static conductivity of a one-dimensional metal”,
Solid State Comm. 37, 997 (1981).
[15] F. Evers and A. D. Mirlin, “Anderson Transitions”, Rev. Mod. Phys. 80, 1355 (2008).
[16] L. P. Kadanoff et al., “Static Phenomena Near Critical Points: Theory and Experiment”,
Rev. Mod. Phys. 39, 395 (1967).

[17] K. G. Wilson, “The renormalization group and critical phenomena”, Rev. Mod. Phys. 55,
583 (1983).
[18] F. Wegner, “ Electrons in disordered systems. Scaling near the mobility edge”, Z. Phys. B
25, 327 (1976).

[19] E. Abrahams, P. W. Anderson, D. C. Licciardello and T. V. Ramakrishnan, “Scaling

Theory of Localization: Absence of Quantum Diffusion in Two Dimensions”, Phys. Rev.
Lett. 42, 673 (1979).
[20] P. Lugan, A. Aspect, L. Sanchez-Palencia, D. Delande, B. Grémaud, C.A. Müller and
C. Miniatura, “One-dimensional Anderson localization in certain correlated random po-
tentials”, Phys. Rev. A 80, 023605 (2009).
[21] E. Gurevich and O. Kenneth, “Lyapunov exponent for the laser speckle potential: A weak
disorder expansion”, Phys. Rev. A 79, 063617 (2009).
[22] P. W. Anderson, D. J. Thouless, E. Abrahams, and D. S. Fisher, “New method for a scaling
theory of localization”, Phys. Rev. B 22, 3519 (1980)

[23] R. Landauer, “Electrical resistance of disordered one-dimensional lattices”, Phil. Mag. 21,
863 (1970).
[24] M. E. Peskin and D. S. Schroeder, An Introduction to Quantum Field Theory (Addison-
Wesley, 1995).

[25] Y. Lahini, R. Pugatch, F. Pozzi, M. Sorel, R. Morandotti, N. Davidson and Y. Silberberg,

“Direct observation of a localization transition in quasi-periodic photonic lattices”, Phys.
Rev. Lett. 103, 013901 (2009).
[26] K. Slevin and T. Ohtsuki, “Corrections to Scaling at the Anderson Transition”, Phys. Rev.
Lett. 82, 382 (1999).

[27] G. Lemarié, B. Grémaud and D. Delande, “Universality of the Anderson transition with
the quasiperiodic kicked rotor”, Europhys. Lett. 87, 37007 (2009).

75
[28] L. Sanchez-Palencia, D. Clément, P. Lugan, P. Bouyer, G. V. Shlyapnikov, A. Aspect,
“Anderson Localization of Expanding Bose-Einstein Condensates in Random Potentials”,
Phys. Rev. Lett. 98, 210401 (2007).
[29] M. V. Berry and S. Klein, “Transparent mirrors: rays, waves and localization”, Eur. J.
Phys. 18, 222 (1997).

[30] A.Z. Genack and A.A. Chabanov,“Signatures of photon localization”, J. Phys. A: Math.
Gen. 38, 10465 (2005).
[31] A. A. Chabanov, M. Stoytchev and A. Z. Genack, “Statistical signatures of photon local-
ization”, Nature 404, 850 (2000).

[32] A. A. Chabanov and A. Z. Genack, “Photon localization in resonant media”, Phys. Rev.
Lett. 87, 153901 (2001).
[33] J.M. Luck, “Systèmes désordonnés unidimensionnels”, Commissariat à l’énergie atomique
(1992), in french.

[34] R.C. Kuhn, O. Sigwarth, C. Miniatura, D. Delande, C.A. Müller, “Coherent Matter Wave
Transport in Speckle Potentials” New J. Phys. 9, 161 (2007) .
[35] S.V. Kravchenko, G.V. Kravchenko, J.E. Furneaux, V.M. Pudalov and M. D’Iorio, “Possi-
ble metal-insulator transition at B = 0 in two dimensions”, Phys. Rev. B 50, 8039 (1994).
[36] D. Belitz and T. R. Kirkpatrick, “The Anderson-Mott-transition”, Rev. Mod. Phys. 66,
261 (1994).
[37] T. Schwartz, G. Bartal, S. Fishman and M. Segev, “Transport and Anderson localization
in disordered two-dimensional photonic lattices”, Nature 446, 52-55 (2007).
[38] M.E. Fisher and M.N. Barber, “Scaling Theory for Finite-Size Effects in the Critical Re-
gion”, Phys. Rev. Lett. 28, 1516 (1972).
[39] A. MacKinnon and B. Kramer, “One-Parameter Scaling of Localization Length and Con-
ductance in Disordered Systems”, Phys. Rev. Lett. 47, 1546 (1981).
[40] S. Katsumoto, F. Komori, N. Sano and S. Kobayashi, “Fine tuning of metal-insulator
transition in Al0.3 Ga0.7 As using persistent photoconductivity”, J. Phys. Soc. Jap. 56, 2259
(1987).
[41] M. Greiner et al., “Quantum phase transition from a superfluid to a Mott insulator in a
gas of ultracold atoms”, Nature 415, 39 (2002).
[42] H. Hu, A. Strybulevych, J.H. Page, S.E. Skipetrov and B.A. van Tiggelen, “Localization
of ultrasound in a three-dimensional elastic network”, Nature Physics 4, 945 (2008).
[43] D.S. Wiersma, P. Bartolini, A. Lagendijk and R. Righini, “Localization of light in a disor-
dered medium”, Nature 390, 671, (1997).
[44] F. Scheffold, R. Lenke, R. Tweer, G. Maret, “Localization or classical diffusion of light?”
Nature 398, 206, (1999).

[45] M. Störzer, P. Gross, G.M. Aegerter and G. Maret, “Observation of the Critical Regime
Near Anderson Localization of Light”, Phys. Rev. Lett. 96, 063904 (2006).

76
[46] A.Z. Genack, “Statistical approach to photon localization” in Waves and Imaging through
complex media, P. Sebbah (ed.) (Kluwer, 2001).
[47] S. Faez, A. Strybulevych, J.H. Page, A. Lagendijk and B.A. van Tiggelen, “Observa-
tion of multifractality at the Anderson localization transition of ultrasound in open three-
dimensional media”, Phys. Rev. Lett. 103, 155703 (2009).

[48] K.B. Efetov, “Supersymmetry and theory of disordered metals”, Advances in Physics 32,
53 (1983).
[49] H. Bruus and K. Flensberg, Many-Body Quantum Theory in Condensed Matter Physics
(Oxford University Press, 2004).

[50] J. W. Negele and H. Orland, Quantum Many-Particle Systems (Westview Press, 1998).
[51] G. D. Mahan, Many-Particle Physics (Kluwer, 2000).
[52] A. Altland and B. Simons, Condensed Matter Field Theory (Cambridge University Press,
2006).

[53] M. Hartung, T. Wellens, C. A. Müller, K. Richter, and P. Schlagheck, “Coherent backscat-

tering of Bose-Einstein condensates in two-dimensional disorder potentials”, Phys. Rev.
Lett. 101, 020603 (2008).
[54] D. Clément, A.F. Varón, J.A. Retter, L. Sanchez-Palencia, A. Aspect, P. Bouyer, “Ex-
perimental study of the transport of coherent interacting matter-waves in a 1D random
potential induced by laser speckle”, New J. Phys. 8, 165 (2006) .
[55] L. Allen and J. H. Eberly, Optical resonance and two-level atoms (Dover Publications, New
York, 1987).
[56] C. Gaul, N. Renner, and C. A. Müller, “Speed of sound in disordered Bose-Einstein conden-
sates”, Phys. Rev. A 80, 053620 (2009). C. Gaul, and C. A. Müller, “Bogoliubov excitations
of disordered Bose-Einstein condensates”, Phys. Rev. A 83, 063629 (2011).
[57] D. J. Thouless, “Localization distance and mean free path in one-dimensional disordered
systems”, J. Phys C.:Solid State Phys. 6, L49 (1973).
[58] S. Kotani and B. Simon, “Localization in general one-dimensional random systems”, Com-
mun. Math. Phys. 112, 103 (1987).
[59] F. A. B. F. de Moura and M. L. Lyra, “Delocalization in the 1D Anderson model with
long-range correlated disorder”, Phys. Rev. Lett. 81, 3735 (1998).
[60] F. M. Izrailev and N. M. Makarov, “Anomalous transport in low-dimensional systems with
correlated disorder”, J. Phys. A: Math. Gen. 38, 10613 (2005).
[61] J. S. Langer and T. Neal, “Breakdown of the concentration expansion for the impurity
resistivity of metals”, Phys. Rev. Lett. 16, 984 (1966).
[62] P. M. Morse and H. Feshbach, Methods of Theoretical Physics (McGraw-Hill, 1953).

[63] T. M. Nieuwenhuizen and J. M. Luck, “Skin layer of diffusive media”, Phys. Rev. E 48,
569-588 (1993).

77
[64] A. Ishimaru, Wave Propagation and Scattering in Random Media (Academic, New York,
1978), Vols. I and II.
[65] Y. Kuga and A. Ishimaru, “Retroreflectance from a dense distribution of spherical parti-
cles”, J. Opt. Soc. Am. A 1, 831-835 (1984).

[66] M. P. van Albada and A. Lagendijk , “Observation of Weak Localization of Light in a

Random Medium”, Phys. Rev. Lett. 55, 2692 (1985).
[67] P. E. Wolf and G. Maret, “Weak Localization and Coherent Backscattering of Photons in
Disordered Media”, Phys. Rev. Lett. 55, 2696 (1985).
[68] E. Akkermans and G. Montambaux, Mesoscopic Physics of Electrons and Photons, (Cam-
bridge University Press, 2007).
[69] B. A. van Tiggelen, Multiple Scattering and Localization of Light, PhD thesis, University
of Amsterdam (1992).
[70] C. A. Müller, T. Jonckheere, C. Miniatura and D. Delande, “Weak localization of light by
cold atoms: the impact of quantum internal structure”, Phys. Rev. A 64, 053804 (2001).
[71] Y. Bidel, B. Klappauf, J.-C. Bernard, D. Delande, G. Labeyrie, C. Miniatura, D. Wilkowski
and R. Kaiser, “Coherent light transport in a cold strontium cloud”, Phys. Rev. Lett. 88,
203902 (2002).
[72] C. A. Müller and C. Miniatura “Multiple scattering of light by atoms with internal degen-
eracy”, J. Phys. A: Math. Gen. 35, 10163 (2002).
[73] D. S. Wiersma, M. P. van Albada, B. A. van Tiggelen and A. Lagendijk, “Experimental
Evidence for Recurrent Multiple Scattering Events of Light in Disordered Media”, Phys.
Rev. Lett. 74, 4193 (1995).

[74] A. A. Golubentsev, “Suppression of interference effects in multiple scattering of light”,

JETP 59, 26 (1984).
[75] G. Labeyrie, D. Delande, R. Kaiser, and C. Miniatura, “Light Transport in Cold Atoms
and Thermal Decoherence”, Phys. Rev. Lett. 97, 013004 (2006).
[76] W. M. Itano, J. C. Bergquist, J. J. Bollinger, D. J. Wineland, U. Eichmann, M. G. Raizen,
“Complementarity and Youngs interference fringes from two atoms”, Phys. Rev. A 57,
4176 (1998).
[77] C. Wickles and C.A. Müller, “Thermal breakdown of coherent backscattering: a case study
of quantum duality”, Europhys. Lett. 74, 240 (2006).

[78] T. Chanelière, D. Wilkowski, Y. Bidel, R. Kaiser, and C. Miniatura, “Saturation-induced

coherence loss in coherent backscattering of light”, Phys. Rev. E 70, 036602 (2004).
[79] V. Shatokhin, T. Wellens, C. A. Müller, A. Buchleitner, “Coherent backscattering of light
from saturated atoms”, Eur. Phys. J. Special Topics 151, 51 (2007).
[80] V. Shatokhin, T. Wellens, B. Grémaud, A. Buchleitner, “Spectrum of coherently backscat-
tered light from two atoms”, Phys. Rev. A 76, 043832 (2007).

78
[81] O. Assaf and E. Akkermans, “Intensity Correlations and Mesoscopic Fluctuations of Dif-
fusing Photons in Cold Atoms”, Phys. Rev. Lett. 98, 083601 (2007); Phys. Rev. Lett. 100,
199302 (2008).
[82] B. Grémaud, D. Delande, C. A. Müller, C. Miniatura, “Comment on ‘Intensity correlations
and mesoscopic fluctuations of diffusing photons in cold atoms’”, Phys. Rev. Lett. 100,
199301 (2008).
[83] C. A. Müller, C. Miniatura, E. Akkermans, G. Montambaux, “Mesoscopic scattering of
spin s particles”, J. Phys. A: Math. Gen. 35, 10163 (2002).
[84] G. Labeyrie, D. Delande, C.A. Müller, C. Miniatura, and R. Kaiser, “Coherent backscat-
tering of light by cold atoms: Theory meets experiment”, Europhys. Lett. 61, 327 (2003).
[85] T. Jonckheere, C. A. Müller, R. Kaiser, C. Miniatura, D. Delande, “Multiple scattering of
light by atoms in the weak localization regime”, Phys. Rev. Lett. 85, 4269 (2000).
[86] B.-G. Englert, “Fringe visibility and which-way information: an inequality”, Phys. Rev.
Lett. 77, 2154 (1996).
[87] C. Miniatura, C. A. Müller, Y. Lu, G. Wang, B.-G. Englert, “Path Distinguishability in
Double Scattering of Light by Atoms”, Phys. Rev. A 76, 022101 (2007).
[88] O. Sigwarth, G. Labeyrie, T. Jonckheere, D. Delande, R. Kaiser, and C. Miniatura, “Mag-
netic field enhanced coherence length in cold atomic gases”, Phys. Rev. Lett. 93, 143906
(2004).
[89] S. Washburn and R. Webb, “Aharonov-Bohm effect in normal metal: Quantum coherence
and transport”, Adv. Phys. 35, 375 (1986).
[90] F. Pierre and N.O. Birge, “Dephasing by extremely dilute magnetic impurities revealed by
Aharonov-Bohm oscillations”, Phys. Rev. Lett. 89, 206804 (2002).
[91] G. Bergmann, “Weak localization in thin films”, Phys. Rep. 107, 1 (1984).
[92] C. A. Müller, “Diffusive spin transport”, in: A. Buchleitner, C. Viviescas, and M. Tiersch
(Eds.), Entanglement and Decoherence. Foundations and Modern Trends, Lect. Notes Phys.
768, 277-314 (Springer, 2009).
[93] F. Pierre, A.B. Gougam, A. Anthore, H. Pothier, D. Estève, and N. Birge, “Dephasing of
electrons in mesoscopic metal wires”, Phys. Rev. B 68, 085413 (2003).
[94] L. P. Gor’kov, A. I. Larkin, and D. E. Khmel’nitskiı̆, Pis’ma Zh. Eksp. Teor. Fiz. 30, 248
(1979) [JETP Lett. 30, 228 (1979)].
[95] D. Vollhardt and P. Wölfle, “Diagrammatic, self-consistent treatment of the Anderson
localization problem in d ≤ 2 dimensions”, Phys. Rev. B 22, 4666 (1980).
[96] D. Vollhardt and P. Wölfle, “Scaling equations from a self-consistent theory of Anderson
localization”, Phys. Rev. Lett. 48, 699 (1982).
[97] D. Vollhardt and P. Wölfle, “Self-consistent theory of Anderson localization”, in: W. Hanke
and Y. V. Kopaev, editors, Electronic phase transitions (Elsevier, Amsterdam, 1992).
[98] O.I. Lobkis and R.L. Weaver, “Self-consistent transport dynamics for localized waves”,
Phys. Rev. E 71, 011112 (2005).

79
 [99] C. Miniatura, R.C. Kuhn, D. Delande, C. A. Müller, “Quantum diffusion of matter waves
in 2D speckle potentials”, Eur. Phys. J. B 68, 353 (2009).
[100] F.L. Moore, J.C. Robinson, C. Bharucha, P.E. Williams and M.G. Raizen, “Observa-
tion of Dynamical Localization in Atomic Momentum Transfer: A New Testing Ground
for Quantum Chaos”, Phys. Rev. Lett. 73, 2974 (1994); F. L. Moore, J. C. Robinson,
C. F. Bharucha, B. Sundaram, and M. G. Raizen, “Atom Optics Realization of the Quan-
tum δ-Kicked Rotor”, Phys. Rev. Lett. 75, 4598 (1995).
[101] J. Chabé, G. Lemarié, B. Grémaud, D. Delande, P. Szriftgiser and J.C. Garreau, “Experi-
mental observation of the Anderson metal-insulator transition with atomic matter waves”,
Phys. Rev. Lett. 101, 255702 (2008).
[102] A.J. Lichtenberg and M.A. Lieberman, Regular and stochastic motion, Springer-Verlag,
New-York (1983).
[103] G. Casati, I. Guarneri and D. Shepelyansky, “Classical chaos, quantum localization and
fluctuations: A unified view”, Physica A 163, 205 (1990) and references therein.
[104] G. Casati, B.V. Chirikov, J. Ford and F.M. Izrailev, “Stochastic Behavior of Classical and
Quantum Hamiltonian Systems”, Lecture Notes in Physics, 334, G. Casati and J. Ford
eds., Springer, New York (1979).
[105] A. Buchleitner, D. Delande and J.C. Gay, “Microwave ionization of 3-d hydrogen atoms in
a realistic numerical experiment”, J. Opt. Soc. Am. B 12, 505 (1995).
[106] H. Ammann, R. Gray, I. Shvarchuck and N. Christensen, “Quantum Delta-Kicked Rotor:
Experimental Observation of Decoherence”, Phys. Rev. Lett. 80, 4111 (1998).
[107] G. Lemarié, J. Chabé, P. Szriftgiser, J.C. Garreau, B. Grémaud and D. Delande, “Ob-
servation of the Anderson Metal-Insulator Transition with Atomic Matter Waves: Theory
and Experiment”, Phys. Rev. A 80, 043626 (2009).
[108] B.G. Klappauf, W.H. Oskay, D.A. Steck and M.G. Raizen, “Observation of Noise and
Dissipation Effects on Dynamical Localization”, Phys. Rev. Lett. 81, 1203 (1998).
[109] D.R. Grempel, R.E. Prange and S. Fishman, “Quantum dynamics of a nonintegrable sys-
tem”, Phys. Rev. A 29, 1639 (1984).
[110] J. Wang and A.M. Garcia-Garcia, “The Anderson transition in a 3d kicked rotor”, Phys.
Rev. E 79, 036206 (2009).
[111] H. Lignier, J. Chabé, D. Delande, J.C. Garreau and P. Szriftgiser, “Reversible destruction
of dynamical localization”, Phys. Rev. Lett. 95, 234101 (2005).
[112] G. Casati, I. Guarneri and D.L. Shepelyansky, “Anderson transition in a one-dimensional
system with three incommensurable frequencies”, Phys. Rev. Lett. 62, 345 (1989).
[113] G. Lemarié, H. Lignier, D. Delande, P. Szriftgiser and J.C. Garreau, “Critical State of the
Anderson Transition: Between a Metal and an Insulator”, Phys. Rev. Lett. 105, 090601
(2010).
[114] G. Lemarié, “Transition d’Anderson avec des ondes de matière atomiques”, PhD the-
sis, Université Pierre et Marie Curie, Paris (2009), http://tel.archives-ouvertes.fr/
tel-00424399/fr/

80