AN INTRODUCTION TO THE

MATHEMATICS OF ANDERSON LOCALIZATION

arXiv:1104.2317v1 [math-ph] 12 Apr 2011

GÜNTER STOLZ

Abstract. We give a widely self-contained introduction to the mathematical

theory of the Anderson model. After defining the Anderson model and deter-
mining its almost sure spectrum, we prove localization properties of the model.
Here we discuss spectral as well as dynamical localization and provide proofs
based on the fractional moments (or Aizenman-Molchanov) method.
We also discuss, in less self-contained form, the extension of the fractional
moment method to the continuum Anderson model. Finally, we mention major
open problems.
These notes are based on several lecture series which the author gave at
the Kochi School on Random Schrödinger Operators, November 26-28, 2009,
the Arizona School of Analysis and Applications, March 15-19, 2010 and the
Summer School on Mathematical Physics, Sogang University, July 20-23, 2010.

1. Introduction
In 1958 the physicist P. W. Anderson introduced the model which is now named
after him to explain the quantum mechanical effects of disorder, as present in mate-
rials such as alloys and amorphous media [9]. The most famous phenomena which
arise in the context of this model are Anderson localization, i.e. the suppression of
electron transport due to disorder, and the Anderson transition in three-dimensional
disordered media which predicts the existence of a mobility edge separating energy
regions of localized states from an extended states region. Anderson localization
has important consequences throughout physics, in theory and experiment. Ander-
son’s work, and that of N. F. Mott and J. H. van Vleck, won the 1977 physics Nobel
prize “for their fundamental theoretical investigations of the electronic structure of
magnetic and disordered systems”1.
Mathematically rigorous studies of the Anderson Model and other models of
random operators started in the 1970s, with the first proof of Anderson localization
for a related one-dimensional model provided by I. Goldsheid, S. Molchanov and
L. Pastur in 1977 [23], followed several years later by a proof of localization for the
actual Anderson model by H. Kunz and B. Souillard [29], also initially for dimension
one. Since then the study of random operators has become an important field of
mathematical physics, which has led to a tremendous amount of research activity
and many mathematical results.
While the Anderson transition and extended states are still an open mathemat-
ical challenge, by now a good rigorous understanding of Anderson localization has
been achieved. Several powerful methods have been found to prove Anderson local-
ization. Important differences exist between one-dimensional and multi-dimensional

This work was supported in part by NSF grant DMS-0653374.

1
[Link] prizes/physics/laureates/1977/
1
2 GÜNTER STOLZ

models, where different physical mechanisms are responsible for localization effects.
In these notes we will focus on methods which allow to prove Anderson localization
in arbitrary dimension. Two such methods are available: The method of multiscale
analaysis (MSA) developed in 1983 by Fröhlich and Spencer [20], and the fractional
moments method (FMM) introduced by Aizenman and Molchanov in 1993 [5].
MSA has produced results in situations which are out of reach for an approach
through the FMM, see Section 9.1 for some related discussion. However, the FMM
is mathematically more elementary, in particular for the case of the classical dis-
crete Anderson model which will be our main focus here. Also, under suitable
assumptions, the FMM allows to prove stronger results on dynamical localization
than can be obtained by MSA. Therefore, in these lectures, after an introduction to
the Anderson model and its basic spectral properties, we will discuss how to prove
Anderson localization based on the FMM.
After more than 50 years of physical research and more than 30 years of math-
ematical work a vast literature with results on Anderson localization and, more
generally, the physics of disordered quantum mechanical systems, is available. In
these introductory lectures we ignore most of the literature as it can not be our goal
to provide a comprehensive survey, not even of the mathematical research which
has been done. Some book length presentations, or parts of such, which provide
very good further reading and many more references are [18, 14, 34, 43, 27].
It is assumed below that the reader is familiar with measure and integration
theory as presented in [37], with basic probabilistic concepts such as independence,
and with the foundations of the theory of linear operators in Hilbert spaces, up to
the spectral theorem for self-adjoint operators and consequences such as spectral
types (absolutely continuous, singular continuous and pure point spectrum) and the
abstract solution of the time-dependent Schrödinger equation via Stone’s theorem,
e.g. [45] or [35]. Otherwise, we have tried to keep these notes mostly self-contained.
For much of the first seven sections we provide full proofs.
We do not aim at the most general known results, but rather want to demonstrate
that simple and natural mathematical ideas can be used to rigorously establish
Anderson localization. Many further developments of the ideas discussed here can
be found in the literature. The references provided below can serve as a starting
point for further reading. An ideal source for continued reading and learning the
state of the art of much what is discussed here will be the upcoming book [8] by
M. Aizenman and S. Warzel.
In Section 2 we introduce the Anderson model and, as a warm-up, prove its first
important property, namely that its spectrum is almost surely deterministic. The
rest of these notes exclusively deals with the phenomenon of Anderson localization.
Section 3 introduces the concepts of spectral localization and dynamical localization,
followed by a discussion of what is known on the physics level of rigor.
In Sections 4 and 5 we prove localization in the large disorder regime of the
Anderson model. This is done via the fractional moments method, by first proving
in Section 4 that fractional moments of Green’s function decay exponentially, and by
then showing in Section 5 that this implies dynamical as well as spectral localization.
In these sections we use methods which were developed in some of the first papers
on the fractional moments method, e.g. [5] and [24]. In particular, these methods
work directly for the Anderson model in infinite volume.
 ANDERSON LOCALIZATION 3

Subsequently, other methods were introduced in the literature, e.g. [1] or [6],
which use finite volume restrictions of the Anderson model. A central concept
here are so-called eigenfunction correlators. These methods have proven to be very
powerful in further-reaching work, for example in dealing with the continuum An-
derson model or multi-particle Anderson models. Thus we introduce this approach
in Section 6 and Appendix A and show how they yield an alternative proof of
localization.
Section 7 discusses the second main regime in which multi-dimensional localiza-
tion has been established rigorously, the band edge regime. Among the new ideas
needed here are the phenomenon of Lifshits tails of the integrated density of states
near spectral edges and a geometric decoupling method to control correlations in
Green’s function. Parts of this section have the character of an outline, referring
to the literature for some of the results used.
Entirely written in form of an outline is Section 8, in which we discuss the ex-
tension of the FMM to continuum Anderson models, as accomplished in [2] and
[12]. This requires considerable technical effort and we merely point out the diffi-
culties which had to be overcome and mention some of the tools which allowed to
accomplish this.
The Anderson model and, more generally, the quantum mechanics of disordered
media, provides many difficult future challenges for mathematicians. We discuss
some of them in our concluding Section 9.

Acknowledgement: The author’s knowledge of random operators and, in par-

ticular, of the Anderson model has benefitted from many other mathematicians,
through their works as well as through personal contact. We apologize for not be-
ing able to properly give credit for the origins of all of these benefits. But we need
and want to make an exception for the contributions of Michael Aizenman, who
was the driving force in the development of the fractional moments method and has
influenced the author’s way of thinking about random operators in multiple ways.
Much of what we have to say here is based on ideas of Michael and his collaborators.
In particular, special thanks are due to Michael and Simone Warzel for letting the
author use some preliminary material from [8] in Section 6 and Appendix A below.
Thanks are also due to a referee for useful suggestions which improved our pre-
sentation.
Finally, the author would like to thank the organizers of the Kochi School on
Random Schrödinger Operators, the Arizona School of Analysis and Applications
and the Summer School on Mathematical Physics at Sogang University for their
invitations. Without these opportunities to lecture on the material covered here
these notes would never have been written.

2. The Anderson Model

2.1. The Discrete Laplacian. Below we will introduce the Anderson model as a
discrete Schrödinger operator, acting, for dimension d ≥ 1, on the Hilbert space
X
ℓ2 (Zd ) = {u : Zd → C : |u(n)|2 < ∞},
n∈Zd
P
with inner product hu, vi = n u(n)v(n).
4 GÜNTER STOLZ

The usual negative Laplacian −∆ = − j ∂ 2 /∂ 2 xj is replaced by its discrete

P

analogue h0 , which acts on u ∈ ℓ2 (Zd ) by

X
(1) (h0 u)(n) = − u(n + k), n ∈ Zd ,
k∈Zd , |k|=1

where |k| = |k1 | + . . . + |kd | is the graph distance on Zd . More appropriately, a

finite difference approximation of −∆ would be given by h0 + 2d, but we neglect the
mathematically trivial shift by 2d (which should still be kept in mind for physical
interpretations). In physics, the Hamiltonian h0 (or its negative) most frequently
arises more directly, i.e. not as a discretization of a differential operator, in the
context of the so-called tight-binding approximation.
Just as the continuum Laplacian, the discrete Laplacian is unitarily equivalent
to a multiplication operator via Fourier transform. Here we consider the Fourier
transform
F : L2 ([0, 2π)d ) → ℓ2 (Zd ),
which is the unitary operator given by
Z
(F g)(n) = (2π)−d/2 g(x)e−ix·n dx,
[0,2π)d

with inverse X
(F −1 u)(x) = l.i.m. (2π)−d/2 u(n)eix·n .
n∈Zd ,|n|≤N

Here x · n = x1 n1 + . . . xd nd and l.i.m. denotes the limit N → ∞ in ℓ2 (Zd ).

A calculation shows that
Xd
F −1 h0 F = −2 cos(xj ),
j=1

where the right-hand side is understood as a multiplicationP operator on L2 ([0, 2π)d )

in the variable x = (x1 , . . . , xd ). The function g(x) = −2 j cos(xj ) is real-valued
and bounded. Thus h0 is bounded and self-adjoint (which also can be checked
directly from the definition (1) without use of the Fourier transform). The range
of g gives the spectrum of h0 ,
(2) σ(h0 ) = [−2d, 2d].
With a bit more effort one can show that inverse images of Lebesgue-nullsets in
R under the function g are Lebesgue-nullsets in Rd . Thus the spectrum of h0 is
purely absolutely continuous.
Another similarity of h0 with the continuum Laplacian is that is has plane waves
as generalized eigenfunctions. To see this, let x ∈ [0, 2π)d and set
(3) φx (n) := ein·x .
While φx 6∈ ℓ2 (Zd ), h0 acts on it via (1) as
X
(h0 φx )(n) = − ei(n+k)·x
|k|=1
 
d
X
= − 2 cos(xj ) φx (n).
j=1
 ANDERSON LOCALIZATION 5

Thus
P φx is a bounded generalized eigenfunction of h0 to the spectral value
−2 j cos(xj ).
2.2. The Anderson Model. Let ω = (ωn )n∈Zd be a set of independent, identically
distributed (i.i.d.) real-valued random variables indexed by n ∈ Zd . Recall that
this means the following, where we denote probabilities by P:
• The (ωn ) are identically distributed, i.e. there exists a Borel probability
measure µ on R such that, for all n ∈ Zd and Borel sets A ⊂ R,
P(ωn ∈ A) = µ(A).
• The (ωn ) are independent. Thus, for each finite subset {n1 , . . . , nℓ } of Zd
and arbitrary Borel sets A1 , . . . , Aℓ ⊂ R,
ℓ
Y
P(ωn1 ∈ A1 , . . . , ωnℓ ∈ Aℓ ) = P(ωnj ∈ Aj )
j=1
ℓ
Y
= µ(Aj ).
j=1

It is sometimes useful to think of a concrete way in which i.i.d. random variables

can be realized as measurable functions on a probability space (Ω, A, P). The
standard construction is the infinite product space
O
(Ω, A, P) = (R, BR , µ),
n∈Zd

with A and P denoting the σ-algebra and measure generated by the pre-measure
d
induced by µ on the Borel cylinder sets in Ω = RZ . This is consistent with the
notation ω = (ωn )n∈Zd as the components ωn of ω ∈ Ω are now i.i.d. random
variables on Ω with common distribution µ.
It is also convenient to work on a complete probability space (Ω, A, P), which in
the above realization is achieved by completing the product algebra A under P, for
which the same notation will be kept.
The Anderson Model is a random Hamiltonian hω on ℓ2 (Zd ), defined for ω ∈ Ω
by
(4) (hω u)(n) = (h0 u)(n) + ωn u(n), n ∈ Zd .
Introducing the random potential Vω : Zd → R by Vω (n) = ωn , we may also
write
hω = h0 + Vω .
Note here that hω is not a single operator, but rather an operator-valued function
on a probability space. It’s operator theoretic properties will generally depend on
ω. Our goal will typically be that a certain property of hω holds almost surely or
with probability one, meaning that it holds for ω ∈ Ω0 , a measurable subset of Ω
with P(Ω0 ) = 1. It lies within the nature of random operator theory that the most
interesting properties will only hold almost surely rather than for all ω ∈ Ω.
One may think of the Anderson model hω as the Hamiltonian governing the
quantum mechanical motion of a single electron in a discretized alloy-type random
medium. In this view the random potential Vω (n) = ωn , n ∈ Zd , represents a
solid formed by nuclei located at the sites n of the lattice Zd and carrying random
electrical charges ωn . Assuming that hω is self-adjoint, the dynamics of the electron
6 GÜNTER STOLZ

is given through the unitary group e−ithω , defined via the spectral theorem, which
provides the solution ψ(t) = e−ithω ψ0 of the time-dependent Schrödinger equation
hω ψ(t) = iψ ′ (t), ψ(0) = ψ0 . The possible energies of the electron are given by the
spectrum σ(hω ) of the Anderson Hamiltonian hω .
In the above discussion we have assumed self-adjointness of the Hamiltonian,
which is the first mathematical fact to be checked. This is particularly easy for
discrete Schrödinger operators such as hω because the discrete Laplacian h0 is
bounded and self-adjoint.
d
Theorem 1. For every ω ∈ RZ , the operator hω is self-adjoint on
X
D(Vω ) = {u ∈ ℓ2 (Zd ) : |ωn u(n)|2 < ∞},
n
the domain of the maximal multiplication operator by the potential Vω .
Proof. Self-adjointness of the maximal multiplication operator by a real-valued
function is a standard fact. Perturbation of the self-adjoint maximal multiplication
operator Vω by the bounded self-adjoint operator h0 preserves self-adjointness with
same domain, e.g. [45]. 
If we assume, as will be done later, that the distribution µ of the ωn has bounded
support, i.e. that
supp µ := {t ∈ R : µ((t − ε, t + ε)) > 0 for all ε > 0}
is bounded in R, then the potential Vω is bounded and therefore defines a bounded
multiplication operator. Thus hω is a bounded self-adjoint operator on ℓ2 (Zd ) as
well. On the other hand, if supp µ is unbounded, then it is not hard to see that Vω ,
and thus hω , is almost surely unbounded.
2.3. The spectrum of the Anderson model. Our next goal is to determine
the spectrum of hω . It follows as a consequence of the general theory of so-called
ergodic operators (e.g. [14]), of which the Anderson model is a special case, that
σ(hω ) is almost surely deterministic, i.e. there exists a closed subset Σ of R such
that
σ(hω ) = Σ almost surely.
Rather than proving this within the general theory of ergodic operators we will
give a direct proof of the following result, which explicitly describes the almost sure
spectrum of the Anderson model:
Theorem 2. The spectrum of the Anderson model is almost surely given by
(5) σ(hω ) = σ(h0 ) + supp µ.
Here the sum of two subsets A and B of R is defined by A + B := {a + b : a ∈
A, b ∈ B}. In particular, this means that the almost sure spectrum of hω is a union
of intervals, namely of translates of [−2d, 2d] by the points in supp µ. If supp µ
doesn’t have large gaps, then the almost sure spectrum of hω is a single interval.
Proof. We begin with the easy part of the proof, namely that
(6) σ(hω ) ⊂ [−2d, 2d] + supp µ
almost surely.
We first argue that σ(Vω ) = {ωn : n ∈ Zd } ⊂ supp µ almost surely. In fact, as
µ(supp µ) = 1, for fixed n ∈ Zd , ωn ∈ supp µ holds almost surely, i.e. on a set
 ANDERSON LOCALIZATION 7

Ωn ⊂ Ω with P(Ωn ) = 1. The countable intersection Ω′ of the Ωn also has measure

one and for ω ∈ Ω′ we have σ(Vω ) ⊂ supp µ as supp µ is closed.
By a general fact from spectral theory, easily proven using a Neumann series
argument, a bounded self-adjoint perturbation B does not shift the spectrum of a
self-adjoint operator A by more than kBk, i.e. σ(A + B) ⊂ σ(A) + [−kBk, kBk].
Thus (6) holds for ω ∈ Ω′ by (2).
The proof of
(7) [−2d, 2d] + supp µ ⊂ σ(hω )
with probability one is more involved and falls into a probabilistic part and a
spectral theoretic part.
For the probabilistic part, start with fixed t ∈ supp µ, ε > 0 and N ∈ N and let
ω ∈ Ω : There exists a cube ΛN ⊂ Zd of side length N such that

Ωt,N,ε :=
|ωn − t| < ε for all n ∈ ΛN } .
As t ∈ supp µ, we have p := µ((t − ε, t + ε)) > 0. Thus, for each fixed cube Λ of
d
side length N in Zd , the probability that |ωn − t| < ε for all n ∈ Λ is pN > 0. We
can cover Zd by infinitely many disjoint cubes of side length N , where these events
are independent. It follows that P(Ωt,N,ε ) = 1.
Next, let Ωt,ε := ∩N ∈N Ωt,N,ε . Thus P(Ωt,ε ) = 1 and we will prove the following
below: For each ω ∈ Ωt,ε and a ∈ [−2d, 2d] it holds that
(8) [a + t − ε, a + t + ε] ∩ σ(hω ) 6= ∅.
Assuming that (8) is true, we proceed as follows: Define Ωt := ∩ℓ∈N Ωt,1/ℓ , such
that P(Ωt ) = 1. For ω ∈ Ωt we have by (8) that, for all a ∈ [−2d, 2d],
1 1
(a + t − , a + t + ) ∩ σ(hω ) 6= ∅
ℓ ℓ
for all ℓ ∈ N. As σ(hω ) is closed, this implies that a+t ∈ σ(hω ) for all a ∈ [−2d, 2d],
and thus [−2d, 2d] + t ⊂ σ(hω ).
For one last argument involving countable intersections of full measure sets, let
B be a countable subset of supp µ which is dense in supp µ and let Ω0 := ∩t∈B Ωt .
Then P(Ω0 ) = 1 and for ω ∈ Ω0 we have [−2d, 2d] + B ⊂ σ(hω ). Using again that
σ(hω ) is closed completes the proof of (7).
We still need to show (8), which is the spectral-theoretic part of the proof. Let
ω ∈ Ωt,ε . Thus, by assumption, for each N ∈ N there exists a cube ΛN of side
length N such that |ωn − t| < ε for all n ∈ ΛN . P
To a ∈ [−2d, 2d] pick x = (x1 , . . . , xd ) such that a = −2 j cos(xj ) and consider
the corresponding generalized eigenfunction φx from (3). Then ψN := χΛN φx has
finite support and, in particular, lies in ℓ2 (Zd ). We claim that
k(hω − (a + t))ψN k
(9) lim sup ≤ ε.
N →∞ kψN k
To find a norm bound for (hω − (a + t))ψN = (h0 − a)χΛN φx + (Vω − t)χΛN φx , we
first note that by assumption k(Vω −t)χΛN φx k ≤ εkψN k. Moreover, as (h0 −a)φx =
0, it follows that ((h0 − a)χΛN φx )(n) is non-zero only for n close to the boundary
of ΛN , where its values are bounded by a constant independent of N . Thus
k(hω − (a + t))ψN k ≤ CN (d−1)/2 + εkψN k.
On the other hand we have kψN k = N d/2 . This proves (9).
8 GÜNTER STOLZ

We conclude by a standard argument: If hω − (a + t) is invertible, then, by (9),

1
k(hω − (a + t))−1 k ≥ .
ε
This implies (8) by using the fact that for general self-adjoint operators A it holds
that
1
k(A − z)−1 k = .
dist(z, σ(A))


3. Localization Properties
We will be interested in localization properties of the Anderson model, which
can be described either by spectral properties or by dynamical properties of the
Hamiltonian.
To be more precise, let I ⊂ R be an open interval. We say that hω exhibits
spectral localization in I if hω almost surely has pure point spectrum in I, i.e.
I does not contain any continuous spectrum of hω , and its eigenfunctions to all
eigenvalues in I decay exponentially.
If I is a non-trivial interval contained in the almost sure spectrum of hω , which
is a union of intervals, then spectral localization in I necessarily means that the
spectrum consists of a dense set of eigenvalues (whose closure fills all of I). This
phenomenon is very different and much more subtle than the appearance of dis-
crete isolated eigenvalues, which is the classical situation encountered in atomic
or molecular hamiltonians. In fact, the possibility of dense pure point spectrum
historically can be considered as the biggest mathematical surprise provided by the
investigation of the Anderson model.
On the other hand, we say that hω exhibits dynamical localization in I if there
exist constants C < ∞ and µ > 0 such that
 
(10) E sup |hej , e−ithω χI (hω )ek i| ≤ Ce−µ|j−k| ,
t∈R

for all j, k ∈ Zd . Here {ej }j∈Zd is the canonical orthonormal basis in Zd , ej (k) = δjk ,
and E(·) Rdenotes the expectation with respect to the probability measure P, meaning
E(X) = Ω X dP for random variables X on Ω. Both, e−ithω as well as χI (hω ), are
defined via the functional calculus for self-adjoint operators. By χI we denote the
characteristic function of I, so that χI (hω ) is the spectral projection for hω onto I.
Dynamical localization in the form (10) is a strong form of asserting that so-
lutions of the time-dependent Schrödinger equation hω ψ(t) = i∂t ψ(t) are staying
localized in space, uniformly for all times, and thus shows the absence of quantum
transport. Let us illustrate this by showing that dynamical localization implies that
all moments of the position operator are bounded in time, i.e. for all p > 0 and all
finitely supported ψ ∈ ℓ2 (Zd ),

(11) sup k|X|p e−ithω χI (hω )ψk < ∞ almost surely,

t∈R
 ANDERSON LOCALIZATION 9

where the position operator |X| is defined by (|X|φ)(n) = |n|φ(n). To see how (11)
follows from (10), assume that ψ(k) = 0 for |k| > R. Then
X 2
k|X|p e−ithω χI (hω )ψk2 = hej , |X|p e−ithω χI (hω )ψi
j
2
X X
2p −ithω
= |j| hej , e χI (hω )ek iψ(k)
j |k|≤R
X X 2
≤ |j|2p hej , e−ithω χI (hω )ek i kψk2 ,
j |k|≤R

where the last step used the Cauchy-Schwarz inequality. We can drop the square
from |hej , e−ithω χI (hω )ek i|2 (as this number is bounded by 1) and then take expec-
tations to get
 
E sup k|X|p e−ithω χI (hω )ψk2
t
X X  
≤ |j|2p E sup |hej , e−ithω χI (hω )ek i| kψk2
t
j |k|≤R
X X
≤ C |j|2p e−µ|j−k| kψk2
j |k|≤R
< ∞.
This implies the almost sure statement in (11) (with square at the norm and there-
fore also without).
Dynamical localization is not only the physically more interesting statement
than spectral localization (as physicists usually have little patience and limited
appreciation for spectral theory), it is also the mathematically stronger property:
We will show later that dynamical localization in I implies spectral localization in
I.
Let us discuss situations in which localization, spectral or dynamical, is expected
physically. For this it will help to introduce an additional disorder parameter λ > 0
in the Anderson model and define
(12) hω,λ = h0 + λVω ,
with Vω (n) = ωn as above. Formally, this fits into the same framework as (4), using
the re-scaled distribution
(13) P(λωn ∈ B) = µλ (B) := µ(B/λ)
of the i.i.d. random variables λωn . The distribution µλ is spread out over larger
supports for larger λ, corresponding to a wider range of possible random charges
in an alloy-type medium. Thus λ >> 1 is the case of large disorder and λ << 1
represents small disorder.
Physicists know all of the following:
In dimension d = 1 the entire spectrum of hω,λ is localized for any value of the
disorder λ > 0.
In dimension d ≥ 2 the entire spectrum is localized at large disorder, i.e. for
λ >> 1.
10 GÜNTER STOLZ

For small disorder λ different behavior arises in dimensions d = 2 and d = 3. For

d = 2 one still has localization of the entire spectrum, but possibly in a weaker form
than for d = 1, e.g. a small amount (or weak type) of quantum transport might
be possible. On the other hand, in dimension d = 3 one observes the Anderson
transition. There are localized regions near the band edges of the almost sure
spectrum, separated by mobility edges from a region of extended states in the center
of the spectrum. Extended states are interpreted as the existence of quantum
transport in the sense that the moments (11) should be infinite for sufficiently
large p. The physical expectation for d = 3 is that this starts at p = 1/2, which
corresponds to the presence of diffusive motion.
Mathematically, localization has been proven for three different regimes: (i)
for all energies and arbitrary disorder in d = 1, (ii) in any dimension and for all
energies at sufficiently large disorder, and (iii) near band edges of the spectrum in
any dimension and for arbitrary disorder.
The mechanisms which cause localization in the Anderson model are fundamen-
tally different for the one-dimensional and multi-dimensional case, which is also
reflected in the mathematical methods which have been used to prove this. In
d = 1 strong tools from the theory of one-dimensional dynamical systems are avail-
able, in particular results on the asymptotics of products of independent random
variables which allow to prove positivity of Lyapunov exponents. Large parts of the
books [14] and [34] are devoted to the presentation of the one-dimensional theory.
A complete presentation of the Kunz-Souillard proof of localization for the one-
dimensional Anderson model can be found in [18]. For a somewhat later survey of
results on one-dimensional localization see [44].
As discussed in the introduction, we will focus here on methods which allow to
prove multi-dimensional localization and, among the two methods which have been
shown to accomplish this, focus on the fractional moments method. Using this
method we will give a detailed proof of large disorder localization and also explain
how it works to show band edge localization, in each case in arbitrary dimension.
We will not discuss localization proofs via multiscale analysis. Excellent intro-
ductions to this method can be found in [27] and [43], while the state of the art of
what can be obtained from Fröhlich-Spencer-type multiscale analysis is presented in
[21] and the review [28]. We also mention the recent powerful extension of the ideas
behind multiscale analysis in [13], which allow to prove localization for continuum
Anderson models (see Section 8) with discretely distributed random couplings, a
result which is beyond what can be obtained by the fractional moments method.

4. Localization at large disorder

Consider the Anderson model (12) at disorder λ > 0 and in any dimension d ≥ 1.
Throughout the rest of these notes we will work with a stronger assumption
on the distribution µ of the random parameters ωn , namely that µ is absolutely
continuous with density ρ,
Z
(14) µ(B) = ρ(v) dv for B ⊂ R Borel, ρ ∈ L∞ 0 (R),
B

i.e. ρ is bounded and has compact support. In particular, this means that the
Anderson hamiltonian hω,λ is a bounded self-adjoint operator in ℓ2 (Zd ).
 ANDERSON LOCALIZATION 11

Introduce the Green function as the matrix-elements of the resolvent of hω,λ ,

(15) Gω,λ (x, y; z) := hex , (hω,λ − z)−1 ey i.
Our first goal is to prove
Theorem 3 ([5]). Let 0 < s < 1. Then there exists λ0 > 0 such that for λ ≥ λ0
there are C < ∞ and µ > 0 with
(16) E (|Gω,λ (x, y; z)|s ) ≤ Ce−µ|x−y|
uniformly in x, y ∈ Zd and z ∈ C \ R.
It is the appearance of fractional moments of the form E(| · |s ), 0 < s < 1, in
the above theorem which prompted the name “fractional moments method” for
the circle of ideas which we want to present here. The method is also frequently
called the “Aizenman-Molchanov method”, as Aizenman and Molchanov did not
only realize that results such as Theorem 3 hold, but that they imply spectral and
dynamical localization. These implications will be discussed in the next section.
The following proof of Theorem 3 closely follows the original ideas from [5]. We
start with two lemmas, an a-priori bound on the fractional moments of Green’s
function and a decoupling lemma, which contain central ideas behind the method
and, in increasing degree of sophistication, have been used in all subsequent devel-
opments of the method.
Lemma 4.1 (A priori bound). There exists a constant C1 = C1 (s, ρ) < ∞ such
that
(17) Ex,y (|Gω,λ (x, y; z)|s ) ≤ C1 λ−s
for all x, y ∈ Zd , z ∈ C \ R, and λ > 0.
Here Z Z
Ex,y (. . .) = . . . ρ(ωx ) dωx ρ(ωy ) dωy
is the conditional expectation with (ωu )u∈Zd \{x,y} fixed. After averaging over ωx
and ωy the bound in (17) does not depend on the remaining random parameters.
Thus we get also that
E(|Gω,λ (x, y; z)|s ) ≤ C1 λ−s .
Proof. We first prove (17) for the case x = y, which demonstrates the simplicity
of the fundamental idea underlying the FMM. For fixed x ∈ Zd , write ω = (ω̂, ωx )
where ω̂ is short for (ωu )u∈Zd \{x} . With Pex := hex , ·iex , the orthogonal projection
onto the span of ex , we can separate the ωx and ω̂ dependence of hω,λ as
hω,λ = hω̂,λ + λωx Pex .
The resolvent identity yields
(18) (hω,λ − z)−1 = (hω̂,λ − z)−1 − λωx (hω̂,λ − z)−1 Pex (hω,λ − z)−1 .
Taking matrix-elements we conclude for the corresponding diagonal Green functions
that
(19) Gω,λ (x, x; z) = Gω̂,λ (x, x; z) − λωx Gω̂,λ (x, x; z)Gω,λ (x, x; z)
or
1 1
(20) Gω,λ (x, x; z) = with a = Gω̂,λ (x,x;z) .
a + λωx
12 GÜNTER STOLZ

Note that the latter is well-defined since one can easily check the Herglotz property
Im Gω̂,λ (x, x; z)/Im z > 0 of the Green function.
The important fact is that
R a is a complex number which does not depend on ωx .
Thus, writing Ex (. . .) := . . . ρ(ωx ) dωx , we find that
kρk∞ dωx C(ρ, s)
Z
(21) Ex (|Gω,λ (x, x; z)|s ) ≤ a ≤ ,
λs |
supp ρ λ + ω x | s λs
with C(ρ, s) independent of λ and a, and thus independent of ω̂, z and x.
The proof of (17) for x 6= y is based on the same idea, replacing the rank-
one-perturbation arguments above with rank-two-perturbation arguments. Write
ω = (ω̂, ωx , ωy ), P = Pex + Pey and
hω,λ = hω̂,λ + λωx Pex + λωy Pey .
Using the resolvent identity similar to above one arrives at
  −1
ωx 0
(22) P (hω,λ − z)−1 P = A + λ ,
0 ωy
where
A = (P (hω̂,λ − z)−1 P )−1 ,
both to be read as identities for 2×2-matrices in the range of P . This is a special case
of the Krein formula which characterizes the resolvents of finite-rank perturbations
of general self adjoint operators. For the matrix A one can check that Im A =
1 ∗
2i (A − A ) < 0 if Im z > 0 and Im A > 0 if Im z < 0. It is also independent of ωx
and ωy .
Using that Gω,λ (x, y; z) is one of the matrix-elements of P (hω,λ − z)−1 P , we find
  −1 s !
ω x 0
Ex,y (|Gω,λ (x, y; z)|s ) ≤ Ex,y A+λ
0 ωy
  −1 s !
−s 1 ωx 0
= λ Ex,y − A−
λ 0 ωy
−1 s
kρk2∞ r r
 
1
Z Z
ωx 0
≤ − A− dωx dωy ,
λs −r −r λ 0 ωy

where [−r, r] is an interval containing supp ρ. In the double integral we change

variables to
1 1
u = (ωx + ωy ), v = (ωx − ωy ),
2 2
which gives a Jacobian factor of 2. As (ωx , ωy ) ∈ [−r, r]2 implies (u, v) ∈ [−r, r]2
we arrive at the bound
Ex,y (|Gω,λ (x, y; z)|s )
−1 s
2kρk2∞ r r
  
1
Z Z
−v 0
≤ − A + − uI du dv
λs −r −r λ 0 v
4rkρk2∞ C(s, ρ)
≤ C(r, s) = .
λs λs
 ANDERSON LOCALIZATION 13

That the latter bound is uniform in x, y and z as well as in the random parameters
(ωu )u∈Zd \{x,y} follows from the fact that the matrix
 
1 −v 0
− A+
λ 0 v
has either positive or negative imaginary part and the following general result:
For every s ∈ (0, 1) and r > 0 there exists C(r, s) < ∞ such that
Z r
(23) k(B − uI)−1 ks du ≤ C(r, s)
−r
for all 2 × 2-matrices B such that either Im B ≥ 0 or Im B ≤ 0.
Let us reproduce an elementary proof of this fact, e.g. Lemma 4.1 in [25], starting
with the observation that, by Schur’s Theorem, B may be assumed upper triangular.
We also may assume without loss that Im B ≥ 0.
Thus
 
b11 b12
(24) B=
0 b22
and
!
1 b12
−1 b11 −u − (b11 −u)(b 22 −u)
(25) (B − uI) = 1 .
0 b22 −u

The bound (23) follows if we can establish a corresponding fractional integral bound
for the absolute value of each entry of (25) separately. For the diagonal entries this
is obvious.
We bound the upper right entry of (25) by
b12 |b12 |
≤
(b11 − u)(b22 − u) |Im ((b11 − u)(b22 − u))|
1
(26) = .
Im(b11 b22 )
u Im b11|b+ Im b22
12 |
− |b12 |

The positive matrix

1
 
Im b11 2i b12
Im B = 1
− 2i b̄12 Im b22
has positive determinant, i.e. det Im B = Im b11 Im b22 − |b12 |2 /4. We thus get
2
Im b11 + Im b22 2Im b11 Im b22 1
≥ 2
≥ .
b12 |b12 | 2
The latter allows to conclude the required integral bound for (26).

The other result needed for the proof of Theorem 3 is
Lemma 4.2 (Decoupling Lemma). For a compactly supported and bounded density
function ρ as above there exists a constant C2 < ∞ such that
R 1
|v−β|s ρ(v) dv
(27) R |v−η|s ≤ C2
|v−β|s ρ(v) dv

uniformly in η, β ∈ C.
14 GÜNTER STOLZ

This can be understood as a consequence of the following two facts: (i) The two
integrals on the left hand side of (27) are continuous functions of η and β. As both
of them neither vanish nor diverge, the same is true for the ratio of the integrals.
(ii) As |β| and |η| become large, the left hand side of (27) has finite limits. This
combines to give a uniform bound in β and η. The details are left as an exercise,
or can be found in [24].
We are now prepared to complete the proof of Theorem 3:

Proof. Given the a-priori bound from Lemma 4.1 we may assume y 6= x. Then
(28) 0 = hex , ey i
= hex , (hω,λ − z)−1 (hω,λ − z)ey i
D  X E
= ex , (hω,λ − z)−1 − eu + (λωy − z)ey
u:|u−y|=1
X
= − Gω,λ (x, u; z) + (λωy − z)Gω,λ (x, y; z).
u:|u−y|=1

Note that Gω,λ (x, y; z) is the upper left entry of the matrix on the left hand side
of the Krein formula (22). Explicitly inverting the right hand side of (22) we find
that
α
Gω,λ (x, y; z) = ,
λωy − β
where α and β do not depend on ω y (and it will
P not matter that they depend on
λ). Using Lemma 4.2, the bound ( n |an |)s ≤ n |an |s and (28) we find
P

1  α s
(29) E(|Gω,λ (x, y; z)|s ) = E β
λs ωy − λ
C2  s |ωy − λz |s 
≤ E |α|
λs |ωy − βλ |s
C2
= E(|λωy − z|s |Gω,λ (x, y; z)|s )
λs
C2 X
≤ E(|Gω,λ (x, u; z)|s ).
λs
u:|u−y|=1

If none of the lattice sites u are equal to x, then the argument can be iterated. For
given x and y one can iterate |x − y| times, in each step picking up a factor 2dC2 /λs
after a maximum is taken over the 2d terms in the sums over next neighbors. This
results in a bound
 |x−y|
2dC2
E(|Gω,λ (x, y; z)|s ) ≤ sup E(|Gω,λ (x, u; z)|s ).
λs u∈Zd

For the last term we use the a-priori bound C1 /λs provided by Lemma 4.1. We get
the exponential decay in (16) for λ ≥ λ0 if we choose λ0 such that 2dC2 /λs0 < 1.


We conclude this section by remarking that the exponential decay bound found
in Theorem 3 also holds for finite volume restrictions of the Anderson Hamiltonian.
More precisely, let L ∈ N and ΛL := [−L, L]d ∩ Zd . By hΛ ΛL
ω,λ and Gω,λ we denote
L
 ANDERSON LOCALIZATION 15

the restriction of hω,λ to ℓ2 (ΛL ) as well as its Green function. By the same proof
as above one finds that, for λ ≥ λ0 ,
(30) E(|GΛ s
ω,λ (x, y; z)| ) ≤ Ce
L −µ|x−y|
,
where the constants C < ∞ and µ > 0 are now also uniform in L.
Moreover, in the finite volume case the bound (30) is uniform in z ∈ C, allowing
for real energy. The reason for this is that the operators hΛ
ω,λ are finite-dimensional
L

and that any given real number E is almost surely not one of their eigenvalues, which
implicitly follows from the above proof. In the finite volume case this also holds
for the a-priori bound in Lemma 4.1. This explains why such bounds play a role in
the FMM similar to the role played by Wegner estimates in localization proofs via
MSA. They demonstrate that eigenvalues are sensitive to the disorder parameters.
A good way to think of the main idea behind the FMM is that eigenvalues are
singularities of the resolvent which move linearly under the random parameters.
Thus the Green function can be made integrable by taking fractional moments.

5. From Fractional Moment Bounds to Localization

We will now discuss methods which show that exponential decay of fractional
moments of Green’s function as shown in Theorem 3 implies spectral as well as
dynamical localization. For the sake of stating a general result of this form we
will absorb the disorder parameter into the random parameters ωx (re-scaling their
distribution as in (13)). Thus we consider the Anderson Hamiltonian in its original
form (4) with single-site distribution µ satisfying (14).
From now on we will generally leave the dependence of various quantities on the
random variable ω implicit and write h = hω , G = Gω , etc.
Our goal is to prove
Theorem 4. Let I ⊂ R be an open bounded interval. If there exist s ∈ (0, 1),
C < ∞ and µ > 0 such that
(31) E(|G(x, y; E + iε)|s ) ≤ Ce−µ|x−y|
uniformly in E ∈ I and ε > 0, then dynamical localization in the form (10) holds
on the interval I.
As a first consequence, by Theorem 3 this implies that at sufficiently large dis-
order λ the Anderson model is dynamically localized in the entire spectrum. In
Section 7 below, we will also use the criterion provided by Theorem 4 to prove
band edge localization.
The most direct way to conclude spectral localization, i.e. pure point spectrum
with exponentially decaying eigenfunctions, from bounds such as (31) is by the
Simon-Wolff method. It was developed in [42] to serve a similar purpose in the
context of multiscale analysis, where it showed that the Green function bounds
established in [20] indeed imply spectral localization. A short argument, showing
that the Simon-Wolff criterion also can be combined with (31) to show spectral
localization, is provided in [5].
Here we will instead discuss the proof of Theorem 4, i.e. focus on how (31)
implies dynamical localization. We have two reasons for doing so: First, dynamical
localization is the physically more relevant property. Second, as we will show at
the end of this section, dynamical localization implies spectral localization with a
straightforward argument using the RAGE theorem.
16 GÜNTER STOLZ

There are two substantially different arguments which prove Theorem 4. In this
section we will present a modification of an argument provided by Graf in [24].
This version of the argument has recently also been used in [25] to prove dynamical
localization for the so-called unitary Anderson model.
The second method, via the use of eigenfunction correlators, will be discussed in
the next section.
Graf’s argument starts with the realization that fractional moments of Green’s
functions of the Anderson model can be used to bound the second moment of
Green’s function as long as a small factor (the imaginary part of the energy) is
introduced to control the singularities of Green’s function at real energy.

Proposition 5.1. For every s ∈ (0, 1) there exists a constant C1 < ∞ only de-
pending on s and ρ such that

(32) |Im z| Ex (|G(x, y; z)|2 ) ≤ C1 Ex (|G(x, y; z)|s )

for all z ∈ C \ R and x, y ∈ Zd .

Here Ex denotes averaging over ωx as in the proof of Lemma 4.1. Integrating

over the remaining variables, we see that (32) also holds with Ex replaced by E.
Our proof follows the proof of Lemma 3 in [24] almost line by line.

Proof. As in the proof of Lemma 4.1 write ω = (ω̂, ωx ). Keep ω̂ fixed and consider
the Hamiltonian

h(α) = h(ω̂,ωx +α) = hω + αPex

obtained by “wiggling the potential at x”. Its Green function will be denoted by
G(α) . Similar to (18) to (20) we find

(hω − z)−1 = (h(α) − z)−1 + α(hω − z)−1 Pex (h(α) − z)−1 ,

and

Gω (x, y; z)
(33) G(α) (x, y; z) =
1 + αGω (x, x; z)
1 Gω (x, y; z)
= · .
α + Gω (x, x; z)−1 Gω (x, x; z)

For the special case x = y and α̃ = −Re Gω (x, x; z)−1 we get from (33) that

1 1
−1
= G(α̃) (x, x; z) ≤ ,
Im G(x, x; z) |Im z|

i.e. |Im G(x, x; z)−1 | ≥ |Im z|. Inserting this into (33) gives

|Im Gω (x, x; z)−1 | |Gω (x, y; z)|2

(34) |Im z||G(α) (x, y; z)|2 ≤ · .
|α + Gω (x, x; z)| |Gω (x, x; z)|2
 ANDERSON LOCALIZATION 17

On the other hand, we can bound the same expression by

X
(35) |Im z||G(α) (x, y; z)|2 ≤ |Im z| |G(α) (x, y ′ ; z)|2
y ′ ∈Zd

= |Im z|hex , (h(α) − z)−1 (h(α) − z)−1 ex i

1
= |Im z|hex , [(h(α) − z)−1 − (h(α) − z)−1 ]ex i
z−z
= Im G(α) (x, x; z)
|Im Gω (x, x; z)−1 |
= ,
|α + Gω (x, x; z)−1 |2
where the last step used (33) with x = y.
For t ≥ 0 one has min(1, t2 ) ≤ ts . Using this to interpolate between (34) and
(35) we get
|Im Gω (x, x; z)−1 | |Gω (x, y; z)|s
(36) |Im z||G(α) (x, y; z)|2 ≤ · .
|α + Gω (x, x; z)−1 |2 |Gω (x, x; z)|s
We will now use the following “re-sampling trick”, which has the effect of creating
an additional random variable (here α) to average over. For a non-negative Borel
function f on R,
Z Z
(37) f (ωx + α)ρ(ωx + α) dα ρ(ωx ) dωx
Z Z
= f (ωx + α)ρ(ωx + α)ρ(ωx ) dωx dα
Z Z
= f (ωx )ρ(ωx )ρ(ωx − α) dωx dα
Z Z 
= f (ωx )ρ(ωx ) ρ(ωx − α) dα dωx
Z
= f (ωx )ρ(ωx ) dωx ,

where the integration order was interchanged in the first and third steps and trans-
lation invariance of Lebesgue measure was used in the second.
Choose f (ωx ) = |G(ω̂,ωx ) (x, y; z)|2 , then (37) and (36) yield

(38) |Im z|Ex (|Gω (x, y; z)|2 )

Z 
= |Im z|Ex |G(α) (x, y; z)|2 ρ(ωx + α) dα

|Gω (x, y; z)|s

 
ρ(ωx + α)
Z
≤ Ex |Im Gω (x, x; z)−1 | dα .
Gω (x, x; z)|s |α + Gω (x, x; z)−1 |2
We now use Lemma 5.2 below with w = Gω (x, x; z)−1 to conclude
|Im z|Ex (|Gω (x, x; z)|2 ) ≤ CEx (|Gω (x, y; z)|s )
with a constant C < ∞ which only depends on supp ρ, but not on x, y and z. 

In the above proof we have used

18 GÜNTER STOLZ

Lemma 5.2. There exists a constant C = C(ρ) < ∞ such that

ρ(ωx + α)
Z
s
|Im w| · |w| dα ≤ C
|α + w|2
uniformly in w ∈ C and ωx ∈ supp ρ.
Proof. Using |w|s ≤ |α|s + |α + w|s , we need two estimates:
(i)
|α|s ρ(ωx + α)
Z
|Im w| dα ≤ πk|α|s ρ(ωx + α)k∞
|α + w|2
≤ π(|ωx |s kρk∞ + k|λ|s ρ(λ)k∞ ).
(ii)
ρ(ωx + α) 1
Z
|Im w| dα ≤ min( , Ckρk∞ |Im w|s )
|α + w|2−s |Im w|1−s
≤ Ckρk1−s
∞ .


We now complete the proof of Theorem 4:
Proof. Consider the mixed spectral measures µx,y of h, the complex Borel measures
defined by
(39) µx,y (B) = hex , χB (h)ey i
for Borel sets B ⊂ R. The total variation |µx,y | of µx,y is a regular bounded Borel
measure which can be characterized by
Z
(40) |µx,y |(B) = sup g(λ) dµx,y (λ)
g : R → C Borel
|g| ≤ 1
= sup |hex , g(h)χB (h)ey i|,
|g|≤1

e.g. [37]. The particular choice gt (x) = e−itx in (40) shows that
(41) |µx,y |(I) ≥ sup |hex , e−ith χI (h)ey i|.
t∈R

Therefore Theorem 4 will follow from a corresponding exponential decay bound for
E(|µx,y |(I)).
As I is an open bounded interval, it follows from Lusin’s Theorem ([37]) that one
can replace Borel functions in (40) by continuous functions with compact support
in I,
(42) |µx,y |(I) = sup |hex , g(h)ey i|.
g ∈ Cc (I)
|g| ≤ 1

For g ∈ Cc (I) it follows by elementary analysis (using that g is bounded and

uniformly continuous) that, uniformly in λ ∈ R,
ε g(E)
Z
g(λ) = lim dE.
ε→0+ π (λ − E)2 + ε2
 ANDERSON LOCALIZATION 19

By the spectral theorem this implies

ε
Z
hex , g(h)ey i = lim g(E)hex , (h − E − iε)−1 (h − E + iε)−1 ey i dE.
ε→0+ π I

This allows to estimate the expected value of (42) by

E(|µx,y |(I))
 
ε
Z X
≤ E lim inf |hex , (h − E − iε)−1 ez i||hez , (h − E + iε)−1 ey i| dE 
ε→0+ π I
z∈Z d

1
Z X

1/2
≤ lim inf E(ε|hex , (h − E − iε)−1 ez i|2 )
ε→0+ π I
z

1/2
· E(ε|hez , (h − E + iε)−1 ey i|2 ) dE,
where, in this order, Fatou, Fubini and Cauchy-Schwarz (on E) have been used.
Now Proposition 5.1 can be applied, allowing to bound further by
1
Z X
≤ lim inf (E(|G(x, z; E + iε)|s ))1/2 (E(|G(z, y; E − iε)|s )1/2 dE
ε→0+ π I
z
C1 C|I| X −µ|x−z|/2 −µ|z−y|/2
≤ e e .
π z

In the last step the assumption of Theorem 4 was used (which also applies to
|G(z, y; E − iε)| = |G(y, z; E + iε)|). The elementary bound, based on the triangle
inequality,
e−µ|x−z|/2 e−µ|z−y|/2 ≤ e−µ|x−z|/4 e−µ|x−y|/4 e−µ|z−y|/4
and another use of Cauchy-Schwarz (on the z-summation) complete the proof of
Theorem 4.


It deserves mentioning here that we have actually proven a stronger result than
dynamical localization in the form (10). The above proof shows that for an open
interval I on which (31) holds there are constants C < ∞ and µ > 0 such that
|hex , g(h)χI (h)ey i| ≤ Ce−µ|x−y|


(43) E(|µx,y |(I)) = E sup
g : R → C Borel
|g| ≤ 1

for all x, y ∈ Zd . An interesting special case is g = 1, where (43) establishes

exponential decay of correlations in the spectral projection χI (h). Consequences
of this for the conductivity of an electron gas in response to an electric field have
been discussed in [4]. Another consequence is mentioned at the end of this section.
Next we show that dynamical localization implies pure point spectrum via the
RAGE-Theorem. The underlying idea is very simple: The RAGE-Theorem char-
acterizes states in the continuous spectral subspace as scattering states (in time-
mean). Dynamical localization excludes scattering states and thus continuous spec-
trum.
Proposition 5.3. Suppose that dynamical localization in the form (10) holds in
an open interval I. Then hω almost surely has pure point spectrum in I.
20 GÜNTER STOLZ

Proof. For a discrete Schrödinger operator h = h0 + V in ℓ2 (Zd ) let Pcont (h) be the
projection onto its continuous spectral subspace. Then the RAGE-Theorem, e.g.
[18], says that for every ψ ∈ ℓ2 (Zd ),
Z T
dt
(44) kPcont (h)χI (h)ψk2 = lim lim kχ{|x|≥R} e−ith χI (h)ψk2 .
R→∞ T →∞ 0 T

If ψ has finite support, say supp ψ ⊂ {|x| ≤ r}, then

kχ{|x|≥R} e−ith χI (h)ψk2 ≤ kχ{|x|≥R} e−ith χI (h)χ{|x|≤r} kkψk2
X
≤ |hex , e−ith χI (h)ey i|kψk2 ,
|x|≥R,|y|≤r

where dropping a square is allowed as kχ{|x|≥R}e−ith χI (h)χ{|x|≤r} k ≤ 1.

Taking expectations in (44) implies, after using Fatou and Fubini,
(45) E(kPcont (hω )χI (hω )ψk2 )
Z T
dt X
E |hex , e−ithω χI (hω )ey i| kψk2 .


≤ lim
R→∞,T →∞ 0 T
|x|≥R,|y|≤r

−ithω
By (10) we have E(|hex , e χI (hω )ey i|) ≤ Ce−µ|x−y| uniformly in t, which
bounds the right hand side of (45) by
X
≤ lim C̃ e−µ|x−y| = 0.
R→∞
|x|≥R,|y|≤r

We conclude that Pcont (hω )χI (hω )ψ = 0 for almost every ω and every ψ of finite
support. The latter are dense in ℓ2 (Zd ) and thus Pcont (hω )χI (hω ) = 0 almost
surely, meaning that the spectrum in I is pure point.


We note that the above proof of pure point spectrum does not imply exponential
decay of corresponding eigenfunctions. It is shown in [5] how this follows directly
from exponential decay of fractional moments (31), using the Simon-Wolff-method
[42]. It can also be deduced from (43) by considering g(h) = δE (h), E ∈ I, using
the result from [39] that almost surely all eigenvalues of hω in I are non-degenerate.
For details on this see Section 2.5 of [2], where a corresponding argument for the
continuum Anderson model is provided which also applies to the discrete Anderson
model considered here.

6. Finite Volume Methods

The localization proof provided in Section 5 proceeds directly in infinite volume,
i.e. does not require to consider restrictions of the hamiltonian h to finite subsets of
Zd . However, it is also possible to work in finite volume, prove the relevant bounds
with volume-independent constants, and then deduce localization by taking the
infinite volume limit (sometimes referred to as “thermodynamical limit”). This
has conceptual advantages such as having to only deal with discrete spectra, thus
allowing to express functions of the hamiltonian by eigenfunction expansions and, as
described at the end of Section 4, to directly study Green’s function at real energy.
Moreover, finite volume methods have proven very robust under generalizations,
for example in the extension to continuum Anderson models which we will discuss
 ANDERSON LOCALIZATION 21

in Section 8. For these reasons we will use this section to provide a different proof
of dynamical localization, i.e. Theorem 4 above, using finite volume methods.
Many of the ideas involved here can already be found in the Kunz-Souillard ap-
proach to localization [29] for the one-dimensional Anderson model. They were first
combined with the fractional moment method in [1] to prove dynamical localiza-
tion for the multi-dimensional Anderson model. A central object are so-called finite
volume eigenfunction correlators, arising from eigenfunction expansions. Eigenfunc-
tion correlators are also used in similar form in proofs of dynamical localization via
multiscale analysis, see [43] or [28] and references therein.
The methods to be described here are not completely disjoint from the methods of
Section 5. As before, we consider the mixed spectral measures µx,y of h introduced
in (39) as well as their total variation |µx,y | given by (40). As will become clear in
(49) below, |µx,y | can be considered as an infinite volume eigenfunction correlator
for h. We will find bounds for it by finding bounds for finite volume eigenfunction
correlators which hold uniformly in the volume.
Let hΛ d d
ω be the restriction of hω to ΛL = [−L, L] ∩ Z and denote its Green
L

ΛL
function by Gω .
Proposition 6.1. Let 0 < s < 1 and I an open bounded interval. Then there exists
C = C(s, ρ, d) < ∞ such that
Z 1
 2−s
ΛL s
(46) E(|µx,y |(I)) ≤ C lim inf E(|Gω (x, y; E)| ) dE .
L→∞ I

Results of this form were first used in implicit form in [1] and later stated more
explicitly in [6]. The exact statement given here as well as its proof below and in
Appendix A follow notes provided to us in private communication by M. Aizenman
and S. Warzel. They used similar results also in [7].
Based on (41), we see that Proposition 6.1 may be applied to provide a proof of
dynamical localization in I in situations where it can be shown that
(47) E(|GΛ s
ω (x, y; E)| ) ≤ Ce
L −µ|x−y|

with constants which are uniform in L and E ∈ I. In fact, dynamical localization

follows under the somewhat weaker assumption that the energy average over I of
the fractional moments of Green’s function is exponentially decaying. However, in
all our applications we have uniform pointwise bounds available. For example, as
discussed at the end of Section 4, a bound of the form (47) holds on the entire
spectrum for sufficiently large disorder, thus providing a second proof of dynamical
localization in this regime.

Proof of Proposition 6.1. We start by reducing the claim (46) to properties of finite-
volume spectral measures. We again use the characterization (42) of |µx,y (I)| for
open bounded intervals I. Strong resolvent convergence of hL to h implies for
continuous g of compact support that hex , g(hΛL )ey i → hex , g(h)ey i and thus, by
(42),
(48) |µx,y |(I) ≤ lim inf sup |hex , g(hΛL )ey i|.
L→∞ |g|≤1

Here the regularity assumption on g can be dropped since hΛL has discrete spec-
trum.
22 GÜNTER STOLZ

Let hL
x be the restriction of h
ΛL
to the reducing subspace Hx for hΛL generated
by ex and let Px be the orthogonal projection onto Hx . Then ex is a cyclic vector
for hL L
x and all eigenvalues E of hx are simple. Thus we may label the corresponding
L L
normalized eigenvectors by ψE . We use the notation ψE also for ψxL ⊕ 0 in ℓ2 (ΛL ) =
⊥
Hx ⊕ Hx .
By expanding into eigenvectors we get
|hex , g(hΛL )ey i| = |hex , g(hL
x )Px ey i|
X
L L
= g(E)hex , ψE ihψE , ey i
E∈I∩σ(hL
x)
X
L L
≤ |ψE (x)||ψE (y)|
E∈I∩σ(hL
x)

=: QL (x, y; I),
and, in particular,
sup |hex , g(hΛL )ey i| ≤ QL (x, y; I).
|g|≤1
The latter will be referred to as eigenfunction correlators. Using Fatou’s lemma we
conclude from (48) that
(49) E(|µx,y |(I)) ≤ lim inf E(QL (x, y; I)).
L→∞

In order to establish a relation to the fractional moments of Green’s function we

will also introduce fractional eigenfunction correlators through
X
L
(50) QL (x, y; I, r) := |ψE (x)|2−r |ψE
L
(y)|r
E∈I∩σ(hL
x)

for 0 < r ≤ 2, noting that QL (x, y; I) = QL (x, y; I, 1). We claim that for 0 < s < 1,
1
(51) EQL (x, y; I) ≤ (EQL (x, y; I, s)) 2−s .
To see this, interpolate s < 1 < 2 via 1 = ps + q2 with the conjugate exponents
p = 2 − s and q = 2−s
1−s . Applying Hölder to expectation as well as to summation
yields
1 1−s
EQL (x, y; I, 1) ≤ (EQL (x, y; I, s)) 2−s (EQL (x, y; I, 2)) 2−s .
L
(y)|2 ≤ 1.
P
This implies (51) after observing that QL (x, y; I, 2) = E∈I∩σ(hLx ) |ψE
We will now be able to relate the fractional eigenfunction correlators to fractional
moments of Green’s function by showing that there exists a constant C = C(s, ρ, d)
such that
Z
(52) EQL (x, y; I, s) ≤ C E(|GΛL (x, y; E)|s ) dE.
I
This, combined with (49) and (51), implies (46).
In the proof of (52) we will use the fractional eigenfunction correlators
QL,v (x, y; I, s) which are defined as in (50), but with the summation being over
the eigenvalues and eigenfunctions of hL x + vPex . Note that, as ex is a cyclic vector
for hL
x , h L
x + vP ex is the same as the restriction of hΛL + vPex to Hx and that ex is
a cyclic vector for this operator for all values of v ∈ R. This makes Proposition A.2
in Appendix A applicable to our situation, which we will now use to finish the proof
of Proposition 6.1 by invoking a resampling argument.
 ANDERSON LOCALIZATION 23

R ρ(u)
For this note that |u−α| s du is continuous and non-vanishing as a function of

α ∈ R. Thus there exists a constant C = C(s, ρ) < ∞ such that

ρ(α)
(53) R ρ(u)
≤C for all α ∈ R.
|u−α|s du

Writing ω = (ω̂, ωx ) and denoting the expectation over ω̂ by Ê, we get

Z
ω (ω̂,ω )
(54) E(QL (x, y; I, s)) = Ê QL x (x, y; I, s)ρ(ωx ) dωx
R
Z Z 
(ω̂,ωx ) dωx
≤ C Ê QL (x, y; I, s) ρ(u) du.
R R |u − ωx |s
After the change of variable ωx 7→ v := ωx − u we see that the inner inte-
R (ω̂,u) dv
gral is equal to QL,v (x, y; I, s) |v| s. By Proposition A.2 this coincides with
ΛL s
R
I |G(ω̂,u) (x, y; E)| dE. Inserting into (54) we arrive at
Z Z
E(Qω L (x, y; I, s)) ≤ C Ê |GΛ s
(ω̂,u) (x, y; E)| dE
L

ZR I
= CE |GΛ s
ω (x, y; E)| dE.
L

I


7. Lifshits Tails and Band Edge Localization

7.1. Band edge localization. So far the only regime in which we have proven
localization is the case of large disorder λ >> 1 in (12). In this section we consider
the Anderson model in its original form (4), i.e. at fixed disorder. Our goal is a
proof of localization at energies near the bottom of the spectrum. The arguments
involved can be easily modified to show the same near the upper spectral edge.
For notational convenience we will assume that the density ρ of the distribution
of the ωx satisfies supp ρ = [0, ωmax ]. We also write E0 = −2d, which according to
Theorem 2 becomes the bottom of the almost sure spectrum
Σ = [E0 , 2d + ωmax ]
of hω .
Our localization proof will again proceed via showing exponential decay of the
fractional moments of Green’s function:
Theorem 5. For every s ∈ (0, 1) there exist δ > 0, µ > 0 and C < ∞ such that
(55) E(|Gω (x, y; E + iǫ)|s ) ≤ Ce−µ|x−y|
for all x, y ∈ Zd , E ∈ [E0 , E0 + δ) and ǫ > 0.
As discussed at the end of Section 4, our methods again yield a bound on the
finite volume Green function as in (30), uniform in the volume and in E ∈ [E0 , E0 +
δ], allowing for ǫ = 0. Thus we can conclude spectral and dynamical localization
at the bottom of the spectrum from either of the methods discussed in Sections 5
of 6, working in infinite volume or in finite volume.
Compared to the case of large disorder, we face essentially two new difficulties,
which are illustrated by the shortcomings of Lemmas 4.1 and 4.2. The a-priori
bound from Lemma 4.1 is still valid and will be used. But, as the disorder λ is fixed,
24 GÜNTER STOLZ

we can not hope that the a-priori bound also provides a “smallness mechanism”,
which can be used to iteratively prove exponential decay. We will again proceed by
iteration, but a different mechanism will be needed to get it started. Also, it will
turn out that we need a different decoupling argument. Lemma 4.2, which was used
in this context in the proof of Theorem 3, is too case-specific and will not work for
the geometric situations which we will encounter here.
7.2. Lifshits tails. Physically, the new smallness mechanism is provided by the
fact that the bottom of the spectrum E0 = inf Σ is a fluctuation boundary. This
means that finite volume restrictions of hω rarely have eigenvalues close to 0. To
make this precise, as before let hΛ ω
L
be the restriction of hω to ℓ2 (ΛL ), ΛL =
d d
[−L, L] ∩ Z .
Lemma 7.1. For every β ∈ (0, 1) there are η > 0 and C < ∞ such that
βd/2
(56) P(inf σ(hΛ
ω ) ≤ E0 + L
L −β
) ≤ CLd e−ηL
for all L ∈ N.
To illustrate why this means that small eigenvalues are rare, let us assume that
the |ΛL | = (2L + 1)d eigenvalues of hLω are uniformly distributed in Σ. Then the
smallest eigenvalue should be no larger than C/Ld . But by (56) this is extremely
rare for large L. In fact, the methods used to prove (56) can also be used to prove
that the integrated density of states N (E) of hω satisfies Lifshits-tail asymptotics
near the bottom of the spectrum:
−d/2
(57) N (E) ≤ Ce−η|E−E0 | ,
which is much “thinner” near E0 than the corresponding IDS N0 (E) = C|E−E0 |d/2
of the Laplacian h0 .
For detailed proofs of Lemma 7.1 as well as (57) we refer to [27] or [43], with
the latter working in the setting of the continuum Anderson model (but applicable
to the discrete model as well). Here we only briefly outline the reasons behind
Lemma 7.1. By the variational principle
(58) inf σ(hΛ
ω − E0 ) =
L
inf h(hΛ
ω − E0 )ϕ, ϕi
L
kϕk=1
 X 
= inf h(hΛ 0 − E0 )ϕ, ϕi +
L
ωi |ϕ(i)|2 .
kϕk=1
i∈ΛL

Note that both terms on the right hand side of (58), the kinetic and potential
energy, are non-negative. In order to find a low lying eigenvalue, they both need
to be small. By reasons of the uncertainty principle, small kinetic energy requires
that ϕ is approximately constant, ϕ(i) ∼P C = |ΛL |−1/2 , to be normalized. For such
ϕ the potential energy is approximately i∈ΛL ωi /|ΛL |, which by the central limit
theorem with large probability is close to the expected value E(ω0 ) > 0. The event
−β
P
i ω i /|Λ L| < L < E(ω0 ) is a large deviation and has probability exponentially
small in |ΛL |.
The weakest part of the above heuristics is the reference to the uncertainty
principle. Slightly different ways to make this rigorous are provided in [27] and
[43], both requiring arguments which control the separation of the second lowest
eigenvalue from the lowest eigenvalue. The proof provided in [27] (going back to
work from the 1980s) uses Temple’s inequality in this context, while [43] uses an
argument based on analytic perturbation theory.
 ANDERSON LOCALIZATION 25

In the context of proving Theorem 5, Lemma 7.1 provides a first step, a so-called
initial length estimate:
Lemma 7.2 (Initial Length Estimate). For every s ∈ (0, 1) there exist C < ∞ and
η > 0 such that
d/(d+2)
(59) E(|GΛL (x, y; E + iǫ)|s ) ≤ CLd e−ηL
for all L ∈ N, x, y ∈ ΛL with |x − y| ≥ L/2, E ∈ [E0 , E0 + 12 L−2/(d+2) ] and ǫ > 0.
Proof. Let β ∈ (0, 1) and, motivated by Lemma 7.1, define the “good” and “bad”
sets as ΩG := {ω : inf σ(hΛ ω − E0 ) ≤ L
L −β
} and ΩB := ΩcG . Then
(60)
E(|GΛL (x, y; E + iǫ)|s ) = E(|GΛL (x, y; E + iǫ)|s χΩG ) + E(|GΛL (x, y; E + iǫ)|s χΩB ).
Pick p > 1 sufficiently small such that sp < 1 and let q be conjugate to p,
1
p + q1 = 1. Hölder applied to the second term on the right hand side of (60) gives

1/p
(61) E(|GΛL (x, y; E + iǫ)|s χΩB ) ≤ E(|GΛL (x, y; E + iǫ)|sp ) P(ΩB )1/q
η βd/2
≤ CLd/q e− q L ,
where we have used the a-priori bound from Lemma 4.1 as well as the probability
bound from Lemma 7.1. The first term on the right hand side of (60) concerns
the event where E has distance at least 12 L−β from the bottom of the spectrum,
which allows to conclude exponential decay of |GΛL (x, y; E + iǫ)| in |x − y| from a
Combes-Thomas estimate (see e.g. [27]), giving
β
(62) E(|GΛL (x, y; E + iǫ)|s χΩG ) ≤ CLβs e−sη|x−y|/L
1−β
≤ CLβs e−η1 L
for constants η > 0, η1 > 0 and C < ∞.
The choice β = 2/(2 + d) leads to equal exponents in (61) and (62), which
combine to give (59). 
7.3. Geometric decoupling. We will eventually fix L = L0 , choosing L0 such
that the right hand side of (59) is sufficiently small (how small still to be deter-
−2/(d+2)
mined). After making this choice we will pick δ = 21 L0 , thus determining
the interval [E0 , E0 + δ] in which Theorem 5 establishes localization. In order to
derive the exponential decay bound (55) from this we have to develop a decoupling
method which will allow to proceed iteratively, splitting the path from x to y into
segments of length L0 .
The description of this so-called geometric decoupling method will fill the remain-
der of this section. Our argument will closely follow a construction introduced in
[6].
ΛcL
In addition to hΛ 2 c
ω , consider hω , the restriction of hω to ℓ (ΛL ), where ΛL =
L c
d
Z \ ΛL . Let
ΛcL
h(L) ΛL
ω = hω ⊕ hω .
This means that
(63) hω = h(L)
ω +T
(L)
,
where T (L) is the operator containing the “hopping terms” introduced by the dis-
(L) (L)
crete Laplacian between sites of ΛL and ΛcL . We write Gω (z) = (hω − z)−1 .
26 GÜNTER STOLZ

More precisely, the matrix-elements of T (L) are

−1, if (u, u′ ) ∈ ΓL ,

(64) T (L) (u, u′ ) =
0, else,
where ΓL is the boundary of ΛL defined as the set of pairs (u, u′ ) with |u − u′ | = 1
and u ∈ ΛL , u′ ∈ ΛcL , or vice versa.
We now perform a double decoupling of the resolvent Gω = Gω (z) by using the
resolvent equation for (63) twice, first at L and then at L + 1:

(L)
(65) Gω = Gω − G(L)
ω T
(L)
Gω
= G(L) (L) (L) (L+1)
ω − Gω T Gω + G(L)
ω T
(L)
Gω T (L+1)G(L+1)
ω .
Equations of this form are often referred to as geometric resolvent identities.
By translation invariance it suffices to prove (55) for x = 0. If |y| ≥ L + 2,
then the first two terms on the right hand side of (65) do not contribute to the
matrix-element Gω (0, y; z) and thus
Gω (0, y; z) = he0 , G(L)
ω T
(L)
Gω T (L+1)G(L+1)
ω ey i
X X
= G(L) ′ (L+1) ′
ω (0, u; z)Gω (u , v; z)Gω (v , y; z).
(u,u′ )∈ΓL (v,v ′ )∈ΓL+1

For s ∈ (0, 1) we get

(66) E(|Gω (0, y; z)|s )
ΛcL+1 ′
X  
≤ E |GΛ
ω
L
(0, u; z)Gω (u ′
, v; z)Gω (v , y; z)| s
.
(u, u′ ) ∈ ΓL
(v, v ′ ) ∈ ΓL+1
(L) (L+1)
Here we have replaced Gω by GΛ ω as 0 and u are both in ΛL . Similarly, Gω
L

ΛcL+1
was replaced by Gω . For fixed (u, u′ ) and (v, v ′ ) consider the corresponding term
on the right hand side of (66) and note that the first and last of the three factors
are independent of ωu′ and ωv . Thus, in taking the expectation we may integrate
over ωu′ and ωv first and use Lemma 4.1 to conclude
ΛcL+1 ′
 
(67) E |GΛ ′
ω (0, u; z)Gω (u , v; z)Gω
L
(v , y; z)|s
ΛcL+1 ′
 
(68) ≤ CE |GΛ ω
L
(0, u; z)| s
|Gω (v , y; z)| s

Λc
(69) = CE(|GΛ s
ω (0, u; z)| )E(|Gω
L L+1
(v ′ , y; z)|s ).
In the last step we have used that the remaining two factors in (68) are stochastically
independent. Now let z = E + iǫ with E ∈ [E0 , E0 + 21 L−2/(2+d) ]. Then we may
estimate the first factor in (69) by the bound obtained in Lemma 7.2 and, after
inserting into (66), find
d/(d+2) X Λc
(70) E(|Gω (0, y; z)|s ) ≤ CL2d−1 e−ηL E(|Gω L+1 (v ′ , y; z)|s ).
kv ′ k∞ =L+2

We want to use (70) as the first step in an iteration. The second step
Λc
would consist in finding a bound for E(|Gω L+1 (v ′ , y; z)|s similar to the bound for
E(|Gω (0, y; z)|s ) given by (70), with v ′ serving as the new origin. A problem arises
from the fact that the underlying domain is not any longer Zd , but ΛcL+1 . Iterating
 ANDERSON LOCALIZATION 27

this would result in more and more complex geometries and we would be faced with
the problem to check if all the constants involved in the estimates leading to (70)
can be chosen uniform in those geometries.
An elegant way around this is the following result of [6], see Lemma 2.3 there,
(L+1)
which allows to bound the depleted Green function Gω in terms of the full
Green function Gω :
Lemma 7.3. There exists a constant C = C(s, ρ) < ∞ such that
X
E(|G(L+1)
ω (v ′ , y; z)|s ) ≤ E(|Gω (v ′ , y; z)|s ) + C E(|Gω (u′ , y; z)|s ).
ku′ k∞ =L+2

(L+1)
The proof of this starts from the geometric resolvent identity Gω = Gω +
(L+1) (L+1)
Gω T Gω . In the resulting Green function expansion over (u, u′ ) ∈ ΓL+1
crucial use is made of the bound
E(|G(L+1)
ω (v ′ , u; z)|s |Gω (u′ , y; z)|s ) ≤ CE(|Gω (u′ , y; z)|s ).
The proof of this uses another special case of the Krein formula similar to (22) (but
this time tracking the dependence on all four variables ωv′ , ωu , ωu′ and ωy ) and
a decoupling argument similar to Lemma 4.2 (but for functions of two variables
which are linear separately in each variable). For the details we refer to [6].
Lemma 7.3 combines with (70) to yield
d/(d+2)
(71) E(|Gω (0, y; z)|s ) ≤ CL4d e−ηL sup E(|Gω (u1 , y; z)|s )
ku1 k∞ ≤L+2

for some constant C < ∞. With that constant we fix L = L0 such that ρ :=
d/(d+2) −2/(2+d)
CL4d0 e
−ηL0
< 1. We also choose δ := 21 L0 now. For E ∈ [E0 , E0 + δ]
we can use (71) to start an iteration,
E(|Gω (u1 , y; z)|s ) ≤ ρ sup E(|Gω (u2 , y; z)|s ),
ku2 k∞ ≤2(L0 +2)

and so forth. This iteration can be carried out approximately |y|/L0 times before
the chains u1 , u2 , . . . may reach y. After this number of steps we use the a-priori
bound from Lemma 4.1 to bound the last fractional moment in the chain. We have
proven Theorem 5 with exponential decay rate µ = | log ρ|/L0 .

8. The Continuum Anderson Model

It took somewhat more than a decade to find a generalization of the fractional
moment method to continuum Anderson models. Our goal in this section is to
explain why this took so long and how it was eventually done. Here our presentation
will be less self-contained than in previous sections. We will outline the new ideas
which were needed and refer to the literature for details.
The main difficulty is that the rather elementary arguments from rank-one and
rank-two perturbation theory, which worked so well for the discrete Anderson
model, fall far short of applying in the continuum. In the latter, each single site
potential is a perturbation of infinite rank, which at best has certain compactness
properties relative to the Laplacian. To make the central ideas behind the frac-
tional moment method work in this setting required a much deeper understanding
of some of the operator-theoretic aspects involved. Here we will follow the works [2]
and [12], where these questions were settled. Earlier work in [26] extended certain
28 GÜNTER STOLZ

aspects of the fractional moment method to continuum models, but still relied on
finite-rank perturbation arguments by, for example, considering continuum models
with random point interactions.
For our presentation here we choose to work with the deterministic background
operator
(72) H0 = −∆ + V0
in L (R ), where V0 is a real-valued, Zd -periodic potential in L∞ (Rd ). Let E0 :=
2 d

inf σ(H0 ) denote its spectral minimum.

A continuum Anderson-type model is then given by
X
(73) H ω = H0 − ωn Un ,
n∈Zd

where ω = (ωn )n∈Zd is an array of i.i.d. random variables with bounded density ρ
such that supp ρ = [0, ωmax ].
The single-site potentials Un (x) = U (x−n) are translates of a non-negative bump
function U characterized by the existence of 0 < r1 ≤ r2 < ∞ and 0 < c1 ≤ c2 < ∞
such that
(74) c1 χ{|x|≤r1 } ≤ U ≤ c2 χ{|x|≤r2 } .
The spectrum of Hω is almost surely deterministic,
σ(Hω ) = Σ a.s.,
and X
E1 := inf Σ = inf σ(H0 − ωmax Un )
n
is characterized by choosing all couplings maximal and thus, due to our sign-
convention, the potential minimal. It can be shown under the assumption (74)
that the spectral minimum is strictly decreased by the random potential: E1 < E0 .
We will use the notation χn = χΛ1 (n) , where Λ1 (n) refers to the unit cube in Rd
centered at n ∈ Zd .
The following theorem is a special case of a result in [12]. Similar results were
first obtained in [2], where a “covering condition” of the form
(75) U ≥ cχ0 , c > 0,
was required for the single-site potential.
Theorem 6. Let d ≤ 3 and 0 < s < 13 . Then there exist δ > 0, µ > 0 and C < ∞
such that
(76) E(kχk (Hω − E − iǫ)−1 χℓ ks ) ≤ Ce−µ|k−ℓ|
for all E ∈ [E1 , E1 + δ], ǫ > 0 and k, ℓ ∈ Zd .
In Theorem 6 we use the norm of the localized resolvent χk (Hω − E − iǫ)−1 χℓ
(sometimes called a “smeared Green function”) as a continuum analogue of the
discrete Green function G(x, y; E + iǫ). This has also been found to be the correct
object to consider in continuum extensions of multiscale analysis.
Without going into the details here (which for the continuum case can be done
similar to what was described in Section 6, see [2]), we state that exponential decay
of fractional moments of the smeared Green function, as established in (76), implies
spectral and dynamical localization:
 ANDERSON LOCALIZATION 29

Corollary 8.1. Under the assumptions of Theorem 6 the following holds:

(a) For almost every ω, Hω has pure point spectrum in [E1 , E1 + δ] with expo-
nentially decaying eigenfunctions.
(b) There are constants µ > 0 and C < ∞ such that
!
(77) E sup kχk g(Hω )χ[E1 ,E1 +δ] (Hω )χℓ k ≤ Ce−µ|k−ℓ|
|g|≤1

for all k, ℓ ∈ Zd , with the supremum taken over Borel functions g : R → C.

The overall approach to proving Theorem 6 is similar to the proof of Theorem 5
in the previous section. The main steps are:

(i) A priori-bound: It can be shown that to every E2 ∈ (E1 , E0 ) and 0 < s < 1
there exists C < ∞ such that
(78) E(kχk (Hω − E − iǫ)−1 χℓ ks ) ≤ C
uniformly in E ∈ [E1 , E2 ], ǫ > 0 and k, ℓ ∈ Zd .
Note here that, as opposed to the discrete case Lemma 4.1, the a-priori bound
is only shown for energies below the spectrum of the unperturbed operator H0 .
This is a consequence of not requiring the covering condition (75) for the single-site
potential. If a covering condition holds, then it was shown in [2] that the bound
(78) holds at all energies, with a constant C on the right which grows polynomially
in E.

(ii) Lifshits tails: The bottom E1 of the almost sure spectrum is again a
fluctuation boundary and close analogues to Lemmas 7.1 and 7.2 as well as Lifshits
tail asymptotics (57) of the IDS hold in the continuum, see e.g. [43] and [2]. As in
the discrete case, this provides the start of an iterative procedure for the proof of
exponential decay in (76).

(iii) Geometric decoupling: The geometric decoupling procedure described

at the end of Section 7 can be carried out similarly in the continuum. Additional
technical difficulties arise mostly due to the fact that the required geometric re-
solvent identities (compare (65)) are less straightforward in the continuum. One
consequence of this is the restriction of Theorem 6 to s < 1/3, which is due to the
need of an additional three-factor Hölder bound used in the decoupling procedure.
Also, elementary decoupling bounds such as Lemma 4.2 have to be replaced by a
more systematic construction involving resampling of the random variables ωn near
the surfaces at which the decoupling is carried out. For details in the setting of
Theorem 6 see [12].

The only one of the above three points which we want to address in some more
detail is the a-priori bound (78), as the existence of such a bound can be seen as
the crucial test for the possibility of using the fractional moment method in the
continuum.
For simplicity, we only consider the “diagonal” case k = ℓ = 0 here and will
assume the covering condition (75). We will discuss reasons why we could hope
that
(79) sup E(kU (Hω − E − iǫ)−1 U ks ) < ∞
ε>0
30 GÜNTER STOLZ

for energies near inf Σ. Under the covering condition, this implies the same result
with U replaced by χ0 .
When trying to implement ideas similar to the ones used in the proof of
Lemma 4.1, we are faced with having to find an analogue to the Krein formula.
It turns out that this is done by the identities known from Birman-Schwinger the-
ory. Write
ω = (ω̂, ω0 ), Hω = Hω̂ − ω0 U.
Then, at least formally, it is easy to derive by the resolvent identity that
(80) U 1/2 (Hω − z)−1 U 1/2 = (ABS − ω0 I)−1
in L2 (supp U ), with the Birman-Schwinger operator
 −1
(81) ABS = U 1/2 (Hω̂ − z)−1 U 1/2 .

It can be justified that the inverses in (81) and (80) exist and that ABS is
maximally dissipative. Here an operator A is called maximally dissipative if it
is dissipative, i.e. Imhφ, Aφi ≥ 0 for all φ in its domain, and it has no proper
dissipative extension. This can also be characterized by the fact that {eitA }t≥0 is
a contraction semigroup.
The identity (80) looks promising since the right hand side separates the de-
pendence on ω0 from the dependence on ω̂. Indeed, if the bound (23) could be
generalized from dissipative 2 × 2-matrices to general maximally dissipative opera-
tors B, then it would immediately give us (79). While (23) extends to dissipative
N × N -matrices, the bound C(r, s) on the right will become N -dependent and di-
verge for N → ∞, as is seen by choosing B to be a diagonal matrix with entries
1, . . . , N . Thus it is not possible to directly extend (23) to the Hilbert space setting.
However, the extension to the Hilbert space setting becomes possible if additional
Hilbert-Schmidt multipliers are introduced. This is most naturally stated in terms
of a closely related weak-L1-bound:
Theorem 7. Let H0 and H1 be separable Hilbert spaces, let A be maximally dissi-
pative in H0 , and let M : H0 → H1 be a Hilbert-Schmidt operator. Then
(a) the boundary value
M (A − v + i0)−1 M ∗ := lim M (A − v + iǫ)−1 M ∗
ǫ→0

exists in Hilbert-Schmidt norm for almost every v ∈ R,

(b) there exists a constant C < ∞ (independent of A and M ) such that
CkM k2HS
(82) |{v ∈ R : kM (A − v + i0)−1 M ∗ kHS > t}| ≤
t
for all t > 0.
In (82) | · | denotes Lebesgue measure and k · kHS the Hilbert-Schmidt norm.
Part (a) is well known in mathematical physics and has been frequently used in
scattering theory. Less well known is part (b), which describes the value-distribution
of the boundary values guaranteed to exist by (a). This was proven in [33] (see also
an appendix in [2] for a reproduction of the proof), based on the weak-L1-property
of the Hilbert transform of Hilbert space-valued functions, the latter being a quite
classical result in harmonic analysis.
 ANDERSON LOCALIZATION 31

The weak-L1-bound (82) can be turned into the fractional moment bound
Z
(83) kM (A − v + i0)−1 M ∗ ksHS ρ(v) dv ≤ C(s, ρ)kM k2s
HS ,

where the constant C(s, ρ) can be chosen uniform for all Hilbert-Schmidt operators
M and maximally dissipative A. This is done by the standard layer-cake integration
argument: If F (v) := kM (A − v + i0)−1 M ∗ kHS , then
Z Z
|F (v)|s ρ(v) dv ≤ kρk∞ |F (v)|s dv
supp ρ
Z ∞
= kρk∞ |{v ∈ supp ρ : |F (v)|s > t}| dt.
0

By (82) the integrand is bounded by min{C̃, CkM k2HS /t1/s }, where C̃ = |supp ρ|.
Splitting the integral at the t-value where C̃ = CkM k2HS /t1/s leads to (83).
When trying to use (83) for a proof of (79) we see from (80) that
(84) U (Hω − z)−1 U = U 1/2 (ABS − ω0 I)−1 U 1/2 .
This leaves us with one more problem to deal with: The multiplication operator
U 1/2 in L2 (Rd ) is not Hilbert-Schmidt. In fact, multiplication operators with non-
vanishing functions in the continuum are never compact.
The key to solving this last problem is that U 1/2 is relatively Hilbert-Schmidt
with respect to −∆ (meaning that U 1/2 (−∆ + 1)−1 is Hilbert-Schmidt), at least for
d ≤ 3, see e.g. [38]. Arguments as typical in relative perturbation theory allow to
split the left hand side of (84) into a sum of terms, some of which trivially satisfy
a fractional moment bound, while others include additional multipliers which lead
to the Hilbert-Schmidt property required in (83). These arguments only work at
energies below the spectrum of the unperturbed operator H0 , which is the reason
for the corresponding assumption which we made when stating (78).
For further details on these relative perturbation arguments as well as on the
“off-diagonal” case k 6= ℓ in (78) we refer to [12] and conclude our sketch of the
proof of Theorem 6 here.

9. Open Problems and directions for future work

To conclude this introduction into the theory of Anderson localization, we men-
tion some open problems and discuss some wide open issues which mathematicians
need to understand better in the future. Here we will not restrict ourselves to
further developments of the fractional moment method, but will address broader
aspects of the quantum mechanical description of disordered media. We will be
relatively brief here and note that a more complete and more detailed recent dis-
cussion of open problems in this field can be found in [10]. In particular, we do not
attempt here to give complete references to related works.
9.1. Singular distributions. Consider the discrete and continuous Anderson
models hω and Hω , but allow for singular distributions of the random coupling
parameters ωi , i ∈ Zd . The most extreme case would be the case of independent
Bernoulli variables, i.e. P(ωi = a) = p, P(ωi = b) = 1−p. This models the physically
interesting case of a two-component alloy. Both, the fractional moments method
and the Fröhlich-Spencer multiscale analysis, fail to provide localization proofs in
this situation. The reason for this is that both methods to a large extend use local
32 GÜNTER STOLZ

averaging arguments in the random parameters, as demonstrated very clearly by

the proof of Lemma 4.1 above. While it is possible to deal with Hölder-continuous
distributions, the Bernoulli case it out of reach for the traditional approaches.
However, Bourgain and Kenig [13] have shown
Theorem 8. Consider the continuum Anderson model Hω defined by (72) and
(73) with V0 = 0 and independent Bernoulli random variables (ωi ). Then Hω is
spectrally localized near E1 = inf Σ.
Their proof is based on a substantial enhancement of the multiscale analysis
approach and, in particular, a deeper understanding of the underlying averaging
mechanisms (such as the role of the so-called Wegner estimates). It has also been
shown in [3] how the argument provided in [13] can be used to handle (ωi ) with
arbitrary non-trivial distribution.
However, somewhat surprisingly, the same question remains open for the discrete
Anderson model (4) with Bernoulli distributed random couplings. The technical
reason for this is that [13] uses subtle unique continuation properties of the eigen-
functions of Schrödinger operators which are not available for lattice models.
More generally, one can easily imagine various other models of random operators
where the random parameters naturally have discrete distribution and where the
available mathematical methods fail to prove localization. One such model would
be discrete Laplacians on random subgraphs of the edges of Zd . An open question
is to decide if in the supercritical percolation regime, where the graph has a unique
infinite component, the Laplacian has localized spectrum. For a recent survey on
these models see [32]

9.2. Extended states. Every list of open problems in random operator theory
needs to mention the somewhat embarrassing fact that mathematicians are still
far from understanding the physically conjectured extended states regime in the
three-dimensional Anderson model.
A proof of the existence of continuous (or absolutely continuous) spectrum or of
diffusive solutions to the time-dependent Schrödinger equation for this model would
be an important break-through. Here we would like to mention another way to
characterize the Anderson transition from localized to extended states, namely the
level statistics conjecture. In fact, this is how physicists can numerically distinguish
the two regimes, which provides the most important evidence for the correctness of
the physical heuristics explaining the transition.
According to the level statistics conjecture it is possible to distinguish the lo-
calized and delocalized regimes by considering the statistical distribution of the
eigenvalues (viewed as point processes) of finite volume restrictions of the Ander-
son model. Localized states should be characterized by Poisson statistics of the
eigenvalues, while in spectral regions with extended states the finite volume eigen-
values should show GOE statistics. The latter it a special kind of level repulsion
observed for Gaussian orthogonal ensembles in random matrix theory.
In the spectral regimes where mathematicians can establish localization, it has
also been verified that the finite volume eigenvalues are Poisson distributed, see
[31], [30], [17] and [22]. However, regarding GOE statistics in the Anderson model,
as little is known as for other possible characterizations of extended states.
As discussed in the lectures by L. Erdos at this School [19], GOE statistics is
a rather universal phenomenon observed in large classes of random matrices, e.g.
 ANDERSON LOCALIZATION 33

so-called Wigner random matrices. The most apparent difference between Wigner
matrices and the Anderson model is that for the latter randomness is restricted
to the diagonal matrix-elements while in Wigner matrices all entries are random.
Understanding the transition between Anderson models and random matrices, for
example by considering random band matrices with an increasing amount of off-
diagonal random entries, could provide important insights into the localization-
delocalization transition in the Anderson model.

9.3. Electron-electron interactions and many-body systems. The Anderson

models discussed above are one-electron models, which ignore interactions between
electrons (as well as interactions between nuclei, which are considered as affixed to
the lattice sites). Quite recently, Anderson-type models for a fixed number N of
interacting electrons in a random background have been shown to have localization
properties. Chulaevsky and Suhov [15, 16] have done this by an extension of mul-
tiscale analysis, while [7] povides similar results based on the fractional moments
approach.
Let us give one example of a result which can be obtained by both approaches,
where we do not try to state the most general result. An N -particle discrete
Anderson-type model can be defined as
X N
X
(h(N )
ω φ)(x) = φ(y) + (U (x) + λ ωxj )φ(x),
y:|y−x|=1 j=1

where φ ∈ ℓ2 (ZN d ), x = (x1 , . . . , xN ) ∈ ZN d and y ∈ ZN d . As above, the (ωx )x∈Zd

are i.i.d. random variables with bounded, compactly supported density. Assume,
for simplicity, that U (x) is a two-particle interaction term of finite range,
X
U (x) = Φ(xj − xk ), supp Φ finite.
1≤j<k≤N

Then localization holds for large disorder:

(N )
Theorem 9. If λ is sufficiently large, then hω is spectrally and dynamically
localized at all energies.
Note that a version of dynamical localization suitable for N -particle systems has
to be used here, see [7].
While such results provide important mathematical progress, condensed matter
physicists will object to the above model because it keeps the number of electrons
fixed. The physically correct system to look at would be a model with positive
(N )
electron density. For example, one could consider the restriction of hω to ℓ2 (ΛN L ),
N
where ΛL = ΛL × . . . × ΛL , and study its properties if N goes to infinity together
with the volume, i.e. N ∼ Ld → ∞. This is mathematically wide open. It is not
even clear how the concepts of localization and extended states should be defined
in this setting, with spectral theoretic terms most likely not being the correct
language any more. Instead dynamical descriptions will have to be used. One
particularly interesting question, which is still challenging even to physicists, is if
electron interaction enhances or reduces localization effects. It has been argued in
the physics literature that such effects could be crucial to understand localization
for two-dimensional disordered systems, the critical case.
34 GÜNTER STOLZ

Appendix A. Basic Rank-One Perturbation Theory

Here we first provide some facts from rank-one perturbation theory. They can
be considered as finite-dimensional, elementary versions of some of the much more
profound insights behind Simon-Wolff theory, e.g. [42] or Chapter 11 to 13 of [40].
Then we establish a general relation between fractional moments of Green’s function
and eigenfunction correlators, which was used in Section 6 above.
Let H be a finite-dimensional Hilbert space, dim H = N , h0 a self-adjoint oper-
ator in H, and ϕ a normalized cyclic vector for h0 . By P = hϕ, ·iϕ we denote the
orthogonal projection onto span{ϕ}. Consider the family of rank-one perturbations
hv := h0 + vP, v ∈ R,
of h0 . Then ϕ is a cyclic vector for all hv . All hv have simple eigenvalues, which
we label as
E1 (v) < E2 (v) < . . . < EN (v).
Also consider the self-adjoint operator
h∞ := P h0 P on D(h∞ ) = {ϕ}⊥ .
Application of the Gram-Schmidt procedure to ϕ, Aϕ, A2 ϕ, . . . shows that Aϕ −
hϕ, Aϕiϕ is a cyclic vector for h∞ . In particular, h∞ has simple eigenvalues which
we label as
E1 (∞) < E2 (∞) < . . . < EN −1 (∞).
In the following we will also use the notation E0 (∞) := −∞, EN (∞) := ∞.
Lemma A.1. (i) For every k ∈ {1, . . . , N }, Ek (v) is analytic and strictly increas-
ing as a function of v ∈ R.
(ii) For every v ∈ R, the eigenvalues of hv and h∞ are intertwined as
(85) E1 (v) < E1 (∞) < E2 (v) < E2 (∞) < . . . < EN −1 (v) < EN −1 (∞) < EN (v).
(iii) For every k ∈ {1, . . . , N },
lim Ek (v) = Ek−1 (∞), lim Ek (v) = Ek (∞).
v→−∞ v→∞

Proof. (i) Let ψk (v) denote normalized eigenvectors of hv to Ek (v), k = 1, . . . , N .

That ϕ is a cyclic vector for hv means that hψk (v), ϕi =
6 0 for all k ∈ {1, . . . , N }.
By analytic perturbation theory, the functions Ek (v) are analytic with
(86) Ek′ (v) = hψk (v), P ψk (v)i = |hψk (v), ϕi|2 6= 0.
(ii) We will prove this by the variational characterization of eigenvalues of hv
and h∞ , e.g. Theorem XIII.2 in [36], which says that for k ∈ {1, . . . , N },
(87) Ek (v) = sup inf hf, hv f i
V ⊂H f ∈ V⊥
dim V = k − 1 kf k = 1

and for k ∈ {1, . . . , N − 1},

(88) Ek (∞) = sup inf hg, h∞ gi.
Ṽ ∈ {ϕ}⊥ g ∈ Ṽ ⊥
dim Ṽ = k − 1 kgk = 1

Note that in the infimum in (87) the orthogonal complement V ⊥ is taken with
respect to H, while Ṽ ⊥ in (88) is taken with respect to {ϕ}⊥ . By definition of h∞
and hv we also have that
(89) hf, h∞ f i = hf, hv f i for all f ∈ {ϕ}⊥ .
 ANDERSON LOCALIZATION 35

We first show that

Ek (v) ≤ Ek (∞) for all k ∈ {1, . . . , N }.
This is trivial for k = N . For k ≤ N −1, let V be a subspace of H with dim V = k−1.
Then V ⊥ ∩ {ϕ}⊥ is a subspace of {ϕ}⊥ of dimension at least N − k. Thus it is
the orthogonal complement of a subspace W of {ϕ}⊥ of dimension at most k = 1.
Therefore, by (88) and (89),
Ek (∞) ≥ inf hf, h∞ f i = inf hf, hv f i
f ∈ W⊥ f ∈ V ⊥ ∩ {ϕ}⊥
kf k = 1 kf k = 1

≥ inf hf, hv f i.
f ∈ V⊥
kf k = 1

As this holds for every subspace V of H with dim V = k − 1, (87) implies Ek (∞) ≥
Ek (v).
Next we will show that
Ek (∞) ≤ Ek+1 (v) for all k ∈ {0, . . . , N − 1},
which is trivial for k = 0. Let k ≥ 1 and Ṽ ⊂ {ϕ}⊥ with dim Ṽ = k − 1. Then
V := span{ϕ} ⊕ Ṽ ⊂ H with dim V = k and V ⊥ = {0} ⊕ Ṽ ⊥ . Thus, by (87) and
(89),
Ek+1 (v) ≥ inf hf, hv f i = inf hg, h∞ gi.
f ∈ V⊥ g ∈ Ṽ ⊥
kf k = 1 kgk = 1

As Ṽ ⊂ {ϕ}⊥ with dim Ṽ = k − 1 was arbitrary, (88) implies Ek+1 (v) ≥ Ek (∞).
Strictness of all inequalities in (85) now is a consequence of (i).
(iii) Here we use the following general fact, which can be proven using Schur
complementation (see e.g. [11] for a description of this method): For the self-adjoint
2 × 2-block matrix  
A B
,
B∗ D
let E 6∈ σ(D), then
 −1  
A + vI − EI B 0 0
lim = .
|v|→∞ B∗ D − EI 0 (D − EI)−1
Applying this to the 2 × 2-block representation of h0 in span{ϕ} ⊕ {ϕ}⊥ shows that
(hv − EI)−1 → 0 ⊕ (h∞ − EI)−1 as |v| → ∞ for every E 6∈ σ(h∞ ).
Using that for self-adjoint operators A,
1
k(A − EI)−1 k = ,
dist(E, σ(A))
we conclude that for every E ∈ σ(h∞ ) there exists a function E(v) such that
E(v) ∈ σ(hv ) for all v and limv→∞ E(v) = E. If E = Ek (∞) for k = 1, . . . , N −
1, it follows from the results of (i) and (ii) that E(v) = Ek (v) for v sufficiently
large, i.e. limv→∞ Ek (v) = Ek (∞). Similarly, it follows that limv→−∞ Ek+1 (v) =
Ek (∞). E1 (v) → −∞ as v → −∞ and EN (v) → ∞ as v → ∞ follows easily by
minimizing/maximizing the quadratic form of hv .

36 GÜNTER STOLZ

To given χ ∈ H, s ∈ (0, 1) and open interval I ⊂ R we consider the fractional

eigenfunction correlators
X
(90) Qv (ϕ, χ; I, s) := |hψk (v), ϕi|2−s |hψk (v), χi|s .
k:Ek (v)∈I

Proposition A.2. The fractional eigenfunction correlators satisfy the identity

Qv (ϕ, χ; I, s)
Z Z
(91) dv = |hϕ, (h0 − E)−1 χi|s dE.
R |v|s I

Proof. As ϕ is cyclic for hv and thus hψk (v), ϕi 6= 0 for all k, we can rewrite (90)
as
s
X hψk (v), χi
(92) Qv (ϕ, χ; I, s) = Ek′ (v) ,
hψk (v), ϕi
k:Ek (v)∈I

where we have also used (86). Observe that

(h0 − Ek (v))ψk (v) = (hv − Ek (v) − vP )ψk (v) = −vhϕ, ψk (v)iϕ.
Thus, using that Ek (v) 6∈ σ(h0 ) for all v 6= 0,
hψk (v), χi = h(h0 − Ek (v))ψk (v), (h0 − Ek (v))−1 χi
= −vhψk (v), ϕihϕ, (h0 − Ek (v))−1 χi.
This allows to further rewrite (92) as
Qv (ϕ, χ; I, s) X
(93) = Ek′ (v)|hϕ, (h0 − Ek (v))−1 χi|s .
|v|s
k:Ek (v)∈I

This will allow to prove (91) by integration. Here we may assume that I ⊂
(Ek−1 (∞), Ek (∞)) for a fixed k ∈ {1, . . . , N } (from which the general case fol-
lows easily). In this case (93) says that
 ′
Qv (ϕ, χ; I, s) Ek (v)|hϕ, (h0 − Ek (v))−1 χi|s , if Ek (v) ∈ I,
=
|v| s 0, else.
Integration yields
Qv (ϕ, χ; I, s)
Z Z
dv = Ek′ (v)|hϕ, (h0 − Ek (v))−1 χi|s dv
R |v|s v:Ek (v)∈I
Z
= |hϕ, (h0 − E)−1 χi|s dE,
I

which used the substitution v 7→ E = Ek (v). 

We conclude by noting that it is possible to prove a result corresponding to

Proposition A.2 without assuming that the Hilbert space is finite-dimensional, as
long as the spectral measure for h0 corresponding to χ is purely singular and bound-
ary values hϕ, (h0 − E − i0)−1 χi of Green’s function are used. A corresponding
streamlining of Aizenman’s original arguments in [1] has been provided by Simon
in [41] for the unitary models considered there and will be presented for the self-
adjoint setting in [8].
 ANDERSON LOCALIZATION 37

References
[1] M. Aizenman, Localization at weak disorder: some elementary bounds, Rev. Math. Phys., 6
(1994), 1163–1182
[2] M. Aizenman, A. Elgart, S. Naboko, J. Schenker and G. Stolz, Moment analysis for localiza-
tion in random Schrödinger operators, Invent. math., 163 (2006), 343–413
[3] M. Aizenman, F. Germinet, A. Klein and S. Warzel, On Bernoulli Decompositions for Random
Variables, Concentration Bounds, and Spectral Localization, Probab. Theory Related Fields
143 (2009), 219–238
[4] M. Aizenman and G. M. Graf, Localization bounds for an electron gas, J. Phys. A: Math.
Gen. 31 (1998), 6783–6806
[5] M. Aizenman and S. Molchanov, Localization at large disorder and at extreme energies: an
elementary derivation, Comm. Math. Phys. 157 (1993), 245–278
[6] M. Aizenman, J. Schenker, R. Friedrich and D. Hundertmark, Finite volume fractional mo-
ment criteria for Anderson localization, Comm. Math. Phys. 224 (2001), 219–253
[7] M. Aizenman and S. Warzel, Localization bounds for multiparticle systems, Comm. Math.
Phys. 290 (2009), 903–934
[8] M. Aizenman and S. Warzel, Quantum Dynamics with Disorder; a Mathematical Introduc-
tion, Lecture Notes, in preparation
[9] P. W. Anderson, Absence of dissusion in certain random lattices, Phys. Rev. 109 (1958),
1492–1505
[10] J. Bellissard, P. D. Hislop, A. Klein and G. Stolz, Random Schrödinger operators: universal
localization, correlations and interactions, Report on a Workshop held at Banff International
Research Station from April 19 to 24, 2009, available at

[Link]/workshops/2009/09w5116/[Link]

[11] J. Bellissard, P. D. Hislop and G. Stolz, Correlation estimates in the Anderson model. J.
Stat. Phys. 129 (2007), 649–662
[12] A. Boutet de Monvel, S. Naboko, P. Stollmann and G. Stolz, Localization near fluctuation
boundaries via fractional moments and applications, J. d’Analyse Math. 100 (2006), 83–116
[13] J. Bourgain and C. Kenig, On localization in the continuous Anderson-Bernoulli model in
higher dimension, Invent. Math. 161 (2005), 389–426
[14] R. Carmona and J. Lacroix, Spectral theory of random Schrödinger operators, Probability
Theory and its Applications, Birkhäuser, Boston, 1990
[15] V. Chulaevsky and Y. Suhov, Wegner bounds for a two-particle tight binding model, Comm.
Math. Phys. 283 (2008), 479–489
[16] V. Chulaevsky and Y. Suhov, Multi-particle Anderson localisation: induction on the number
of particles, Math. Phys. Anal. Geom. 12 (2009), 117–139
[17] J.-M. Combes, F. Germinet and A. Klein, Poisson Statistics for Eigenvalues of Continuum
Random Schrödinger Operators, Anal. PDE 3 (2010), 49–80
[18] H. L. Cycon, R. G. Froese, W. Kirsch and B. Simon, Schrödinger Operators with Application
to Quantum Mechanics and Global Geometry, Texts and Monographs in Physics Springer,
Berlin, 1987
[19] L. Erdos, Universality of Wigner Random Matrices: a Survey or Recent Results, Lectures
held at the Arizona School of Analysis with Applications, March 15 to 19, 2010, available at
[Link]
[20] J. Fröhlich and T. Spencer, Absence of diffusion in the Anderson tight binding model for
large disorder or low energy, Comm. Math. Phys. 88 (1983), 151–184
[21] F. Germinet and A. Klein, Bootstrap multiscale analysis and localization in random media,
Comm. Math. Phys. 222 (2001), 415–448
[22] F. Germinet and F. Klopp, Spectral statistics for random Schrdinger operators in the localized
regime, Preprint 2010, arXiv:1011.1832
[23] I. Ya. Goldsheid, S. A. Molchanov and L. Pastur, A random homogeneous Schrödinger oper-
ator has a pure point spectrum, (Russian) Funkcional. Anal. i Prilozen. 11 (1977), 96.
[24] G. M. Graf, Anderson localization and the space-time characteristic of continuum states, J.
Stat. Phys. 75 (1994), 337–346
[25] E. Hamza, A. Joye, G. Stolz, Dynamical localization for unitary Anderson models, Math.
Phys. Anal. Geom. 12 (2009), 381–444
38 GÜNTER STOLZ

[26] P. D. Hislop, W. Kirsch and M. Krishna, Spectral and dynamical properties of random models
with nonlocal and singular interactions. Math. Nachr. 278 (2005), 627–664
[27] W. Kirsch, An invitation to random Schrdinger operators. With an appendix by Frédéric
Klopp, Panor. Synthèses 25, Random Schrödinger operators, 1–119, Soc. Math. France,
Paris, 2008
[28] A. Klein, Multiscale analysis and localization of random operators, Panor. Synthèses 25,
Random Schrdinger operators, 121–159, Soc. Math. France, Paris, 2008
[29] H. Kunz and B. Souillard, Sur le spectre des opérateurs aux différences finies aléatoires,
Comm. Math. Phys. 78 (1980/81), 201–246
[30] N. Minami, Local fluctuation of the spectrum of a multidimensional Anderson tight binding
model, Comm. Math. Phys. 177 (1996), 709–725
[31] S. A. Molchanov, The local structure of the spectrum of the one-dimensional Schrödinger
operator, Comm. Math. Phys. 78 (1980/81), 429–446
[32] P. Müller and P. Stollmann, Percolation Hamiltonians, Preprint 2010, arXiv:1002.5000
[33] S. N. Naboko, The structure of singularities of operator functions with a positive imaginary
part, Funkts. Anal. Prilozh. 25 (1991), 1–13, 96
[34] L. Pastur and A. Figotin, Spectra of Random and Almost-Periodic Operators, Springer,
Berlin-Heidelberg-New York, 1992
[35] M. Reed and B. Simon, Methods of modern mathematical physics. I. Functional analysis,
2nd Edition, Academic Press, New York, 1980
[36] M. Reed and B. Simon, Methods of modern mathematical physics. IV. Analysis of operators,
Academic Press, New York-London, 1978
[37] W. Rudin, Real and Complex Analysis, 3rd Edition, McGraw-Hill, Boston, 1987
[38] B. Simon, Schrödinger semigroups, Bull. Am. Math. Soc., 7 (1982), 447–526
[39] B. Simon, Cyclic vectors in the Anderson model, Rev. Math. Phys. 6 (1994), 1183–1185
[40] B. Simon, Trace ideals and their applications, 2nd Edition, Mathematical Surveys and Mono-
graphs, 120, American Mathematical Society, Providence, RI, 2005
[41] B. Simon, Aizenman’s theorem for orthogonal polynomials on the unit circle, Const. Approx.
23 (2006), 229–240
[42] B. Simon and T. Wolff, Singular continuous spectra under rank one perturbations and local-
ization for random Hamiltonians, Commun. Pure Appl. Math. 39 (1986), 75–90
[43] P. Stollmann, Caught by disorder: bound states in random media, Progress in Mathematical
Physics 20, Birkhäuser, Boston, 2001
[44] G. Stolz, Strategies in localization proofs for one-dimensional random Schrödinger operators,
Proc. Indian Acad. Sci. (Math. Sci.) 112 (2002), 229–243
[45] J. Weidmann, Linear operators in Hilbert spaces, Graduate Texts in Mathematics, 68,
Springer-Verlag, New York-Berlin, 1980

Department of Mathematics, University of Alabama at Birmingham, Birmingham,

AL 35294, USA
E-mail address: stolz@[Link]