Evolution of wave packets in quasi- ID and ID random media

diffusion versus localization

F.M. Izrailev 1,2 *, Tsampikos Kottos 2 , A. Politi 3 , and G.P. Tsironis 2

1 Budker Institute of Nuclear Physics,
630090 Novosibirsk, Russia
2 Department of Physics, University of Crete and Research Center
of Crete, P.O. Box 2208, 71003 Heraklion- Crete, Greece
3 Istituto Nazionale di Ottica 50125 Firenze and
INFN-Firenze, Raly
(February 7, 2008)

Abstract

We study numerically the evolution of wavepackets in quasi one-dimensional
random systems described by a tight-binding Hamiltonian with long-range
random interactions. Results are presented for the scaling properties of the
width of packets in three time regimes: ballistic, diffusive and localized. Par-
ticular attention is given to the fluctuations of packet widths in both the
diffusive and localized regime. Scaling properties of the steady-state distribu-
tion are also analyzed and compared with a theoretical expression borrowed
from the one-dimensional Anderson theory. Analogies and differences with
the kicked rotator model and the one-dimensional localization are discussed.

*email addresses: izrailev® inp.nsk.su; izrailev@physics.spa.umn.edu

1

I. INTRODUCTION

The main approach to a statistical description of the spectra in complex quantum systems
originates from the pioneering work of Wigner ffl], who conjectured that random matrices
could represent the simplest meaningful model for studying heavy nuclei. Currently, ran-
dom matrix theory (RMT) has become a very effective tool in a large variety of physical
applications. Until recently the matrices in this theory were assumed to be homogeneous,
i.e. all matrix elements were taken to have identical statistics. This simplification is mainly
dictated by mathematical reasons since the corresponding ensembles of random matrices are
rotational invariant, a property that simplifies the theoretical analysis.

For a long time, the RMT had no concrete physical basis, in the sense that conditions
for its applicability were not specified. It was believed that the systems under consideration
had to be extremely complex in order to have a good agreement with the predictions of the
RMT [0-0]. The situation has changed with the progress of the so-called quantum chaos
theory which deals with dynamical Hamiltonian systems exhibiting chaotic motion in the
classical limit. One of the main results of this theory is that in the extreme case, when
classical motion is strongly chaotic and no influence of quantum localization is taken into
account, statistical properties of both spectra and eigenfunctions are well described by the
RMT. This statement has been thoroughly studied and confirmed both for autonomous
systems like chaotic billiards [|J and for time- dependent models like the Kicked Rotator
model (KR) |||6j and the Kicked Tops 0. Moreover, there are many physical examples
where "complexity" of a quantum system is not maximal, but nevertheless a statistical
description applies pretty well.

To describe the consequences of quantum localization in the presence of strong classical
chaos, a new type of random matrices has been introduced when studying the KR [§]. The
distinctive feature of these matrices is their band-like structure, which is related to the finite
range of interactions in a given basis. In a sense, the ensembles of Band Random Matrices
(BRM) can be regarded as an extension of the conventional random ensembles, since the

2

latter are recovered by just setting the band size b equal to the matrix size N. Currently, the
interest for BRM raised significantly due to their close relationship with quasi-lD models
with random potentials ||, or ID systems with long-range hoppings between neighboring
sites. In the ID interpretation, the band size b corresponds to the hopping range while in
the quasi-lD case, it is associated with the number of transverse channels for electron wave
propagation along a thin wire. Recent numerical and analytical studies of BRM (see PJlO

and references therein), led to numerous results regarding the structure of eigenstates. In
particular, the localization length has been shown to depend only on the scaling parameter
b 2 /N, so as the statistical properties of the eigenstates that are directly related to the
fluctuation properties of the conductance.

However, much less is known about the evolution of wave packets in models described
by BRM. One should note that even in the "simple" case of Anderson-type models in ID
only the short and the long time scales, corresponding to ballistic spread and saturation of
the packet width, respectively, have been successfully studied [PTf| . In quasi-lD (or ID with
long-range hoppings) models, the picture is both more complicated and more interesting
with respect to the Anderson case. Indeed, while the diffusion time scale is absent in ID
models of Anderson type, since the mean free path is of the order of the localization length,
classical diffusion alters the ballistic spread in quasi ID systems, before being eventually
suppressed by localization effects.

In dynamical systems, the influence of strong localization on classical diffusion in mo-
mentum (or energy) space has been studied in detail in the framework of the KR. It was
found that the time scale of the wave-packet spreading that is analogous to classical diffu-
sion is much longer than the logarithmic time scale over which the complete correspondence
between classical and quantum description holds 0,^,|1^,|T^] . The entire diffusive process,

including the final saturation proved to exhibit remarkable scaling properties ||[T3| . In par-
ticular, the diffusive time scale is proportional to the localization length of those eigenstates
which are involved in the dynamics. In view of the analogies existing between dynamical
and Anderson-type localization, the results obtained in the study of wave packet evolution

in the KR will represent the touchstone for the present investigation of packet dynamics
in quasi-lD disordered models. However, because of the existence of basic differences as

ID geometry, with particular attention to their width and fluctuations on different time
scales. We hope that our detailed numerical investigations, providing direct information
on the dynamical properties of strong localization, will furnish also some insight for future
analytical studies.

The paper is structured in the following way. In Sec. 2, the model is introduced and
discussed together with the main properties of spectra and eigenstates of BRM. In Sec. 3
both ballistic and diffusive time scales are analyzed and the scaling properties of packets in
terms of the band-width b are established. The effect of noise in the destruction of coherence
is also briefly discussed. Section 3 terminates with the results for the suppression of classical
diffusion due to the localization of eigenstates. In section 4, we focus on the problem of
fluctuations of the shape of packets both for the diffusive and relaxation time scales. In
section 5, a phenomenological description of the asymptotic shape of the packets is given,
based on results for the ID Anderson model. Moreover, fluctuations along the asymptotic
profile of packets are studied. The conclusions are summarized in Sec. 6.

Since its birth, random matrix theory has been mainly dealing with statistical properties
of "full" random matrices, for which all off-diagonal matrix elements are independent and
distributed according to the same law. In physical applications this implies that interactions
are assumed to be so strong and complex that no other parameter, apart from the symmetry

presently clear to what extent the similarity between

dynamical and disordered systems can be pushed forward.

In this paper we extend a previous study jljj of initially <5-like wave packets in a quasi-

II. BAND RANDOM MATRICES: MAIN PROPERTIES

A. Definition and applications

4

of matrices, is to be taken into account. As a result, such matrices are associated with the
extreme case of maximal chaos which is known to appear in various physical systems such
as heavy nuclei, atoms, metallic clusters, etc. Furthermore, full random matrices represent
a good model for the description of local statistical properties of spectra and eigenstates in
some range of the energy spectrum, typically, in the semiclassical region.

On the other hand, the conventional RMT is both unable to describe important phenom-
ena such as localization of eigenstates, and to characterize the spectra of physical systems
influenced by strong localization effects. For this reason, much attention has been recently
paid to the so-called Band Random Matrices which are characterized by the free param-
eter b defining the effective band width of a Hamiltonian. Such random matrices with
elements decaying away from the main diagonal, appear to provide more realistic models
for the Hamiltonian of "complex" quantum systems (see, e.g., ||,|10| and references therein).
The simplest type of BRM is given by matrices H nm with zero elements outside the band
(|n — m\ > b), while inside the band (\n — m\ < b), matrix elements are assumed to be
independent and distributed according to a Gaussian law,

V(H nm ) = X -= ex V (-HlJ2a 2 nm ), a 2 nm =< H 2 nrn >= ^(1 + S nm ) (1)

a

inn

2^]- — nmi — nmn ~ nm — nm ^

where 5 nm is the Kronecker symbol, and Oq = 2 implying that the variance of the off-diagonal
matrix elements is equal to 1. The BRM ensemble can be regarded as a generalization of
the standard Gaussian orthogonal ensemble (GOE) as the former reduces to the latter
when b = N. Analogous generalizations can be introduced for the unitary and symplectic
ensembles of band random matrices .

The limit case b = corresponds to diagonal matrices, while 6=1 corresponds to
tridiagonal Hamiltonians with both diagonal and off-diagonal disorder. The latter case is
well known in the physics of disordered media; the main properties of such matrices are
relatively well understood. The general case of BRM where the variance of the matrix

elements decreases with the distance from the main diagonal, introduced in NM, is also

amenable to an analytic treatment. They are not, strictly speaking band matrices, but an

effective band size b can be denned from the shape of the envelope |19 .

In what follows, we consider large values of b 3> 1 and assume that iV > b 2 . The first
condition implies a large number of nearby states coupled by the interaction. The second
condition allows to neglect finite size corrections arising from the finiteness of the samples.

Considerable interest towards the ensemble of BRM was stimulated by the investigation
of the quantum behavior of periodically driven Hamiltonian systems. A paradigmatic system
in this class is the so-called Kicked Rotator model. Indeed, the unitary matrix U yielding the
time evolution between two consecutive kicks has, in the angular momentum representation,
a band-like structure with an effective band-size approximately equal to the strength k of the
kicks. Outside the band, the matrix elements of U decrease extremely fast, while inside the
band they can be treated as pseudo-random entries if the corresponding classical evolution
is chaotic 0. As a consequence, both spectra and eigenstates of BRM of the type ([[])
are expected to have statistical properties similar to those of the KR. A number of data
substantiate this belief .

BRM play also an important role in the understanding of quasi-lD disordered media.
The connection has been established through the nonlinear sigma model which, on the one
hand, is known to provide an excellent description of quasi-lD systems and, on the other,
has been rigorously shown to be reducible to the BRM model Such relationships gave

a boost to the investigations of statistical properties of eigenstates and eigenvalues of BRM
as they also allow a better understanding of the properties of the KR, as well as of quasi- ID
and ID models with long-range random interactions.

B. Density of States and Structure of Eigenstates

As was first numerically shown in [z(J and later analytically proved in ||18| , the density
of states p(E) for infinite BRM (N — > oo) obeys the semicircle law

1

p(E) = — -VSbv 2 - E 2 ; \E\ <R = vV8b (2)
Aixbv 1

6

with p(E) = for \E\ > Rq . The parameter v is just the standard deviation of the
distribution of off-diagonal elements, v 2 = (H 2 m ) = (Tq/2; it does not influence the statistical
properties of the spectra as well as the structure of the eigenstates, since it can be scaled
out. For b = N, expression (Q) reduces to the well known Wigner semicircle law derived in
conventional RMT.

In infinite BRM, all eigenstates <fE(n) are known to be exponentially localized [|9jl0,20

\<p E (n)\ ~ exp I- j j ; n^±oc (3)

where no is the "center" of an eigenfunction in the basis in which the random matrices
have been defined. The quantity loa(E) is the localization length defined as the inverse
of the asymptotic spatial decay-rate of the amplitude of the corresponding eigenfunction.
Numerically, loo{E) can be determined by implementing the transfer matrix method. It is
important to recall that the localization length describes the decay of the eigenfunction
only in the tail and not in the central region of size ~ b 2 . This region is characterized by
an effective number I ~ of "principal components" which are usually defined in terms of
the inverse participation ratio and of the so-called entropic localization length Ih (see
for details).

In finite samples, one more parameter comes into play, the rank N of the matrices.
In such a case, all relevant properties of spectra and eigenstates are parametrized by the
ratio A = b 2 /N PJlQl. Upon changing A, one can accurately follow the transition from

the completely localized (A <C 1) to the delocalized (A 3> 1) regime. Numerous studies of
BRM allowed to unravel the dependence of the statistical properties of eigenfunctions on
this scaling parameter. In particular, in finite bases of size N, all eigenstates are extended if
A>1, and all their properties are very similar to those for the standard RMT. Our interest
here, is devoted to the opposite limit A 1 of very localized eigenstates where finite-size
effects can be neglected. Based on the results obtained for the KR PJ2"0[|, it was predicted
that the localization length is proportional to b 2 . A rigorous analysis [TE| has confirmed

this prediction and established the dependence ~ p b of the localization length on the

energy. In the case of BRM with a general envelope function a(k) for matrix elements H nm ,

oo

a(k) k 2

UE) = 2[1 - {^) 2 ]B ■ B= k -^ ; k = n-m (4)

Ro [ £ a(k)} 2

k=—oo

where B is the second moment of the function a(k). In case ([TJ), i.e. for a sharp band of size
b, one obtains B = b 2 /3 and thus

U£) = jj[l-(^) 2 ]& 2 • (5)

At variance with the KR, the localization length of the eigenstates of BRM depends on the
energy E.

C. Numerical procedure

Although considerable progress has been made in the description of the eigenstate struc-
ture (see |5,ni] and references therein), the evolution of wave packets is still poorly under-

stood even in the limit of infinite samples. The mathematical model we consider below is
the time-dependent Schrodinger equation on a ID lattice,

dr (t) n+b

^jr= E H nmCm (t), (6)

' ' m=n—b

where c n (t) is the probability amplitude for an electron to be at site n and H nm is a symmetric
BRM of the type (|]). Equation (^) has been integrated numerically using a finite-time-step
(dt ~ 10~ 3 — 10~ 4 ) fourth order Runge-Kutta algorithm on a self-expanding lattice in order to

eliminate finite-size effects ||21|| . Whenever the probability of finding the particle at the edges
of the chain exceeded 10 -15 , 106 new sites were added to each edge. The initial condition
was taken to be a 5-like state located in the middle of the chain, i.e. c n (t = 0) = 5 n fi. At
each time step, the normalization condition for the total probability, J2 n l c n(^)| 2 = 1 ; was
checked observing fluctuations smaller than 10~ 4 .

A further check of the accuracy of our calculations has been performed by reversing the
time-axis direction after 2000 time units (for b = 10). The difference between the initial

8

probability distribution and that obtained after integrating for 2000+2000 units was found
to be less than 10~ 13 . In all the cases discussed below, a large number of disorder realizations
has been considered (more than 150) in order to get rid of sample-to-sample fluctuations.

III. DIFFUSION OF WAVE PACKETS

The time-evolution of a quantum wave packet in the lattice is naturally described by the
mean square displacement,

M(M) = («(*)> = ^> 2 M*)I 2 ) , (7)

where (...) stands for the average over different realizations of the frozen disorder H nm . The
time dependence of M(b, t) provides a qualitative description of the dynamical regime: a
power-law evolution, M(t) ~ t u , with v < 1 corresponds to a sub-diffusive behaviour (hinting
at a possible, eventual localization), v — 1 corresponds to ordinary diffusion, while v > 1 to
super- diffusion (y — 2 characterizes ballistic motion).

As mentioned above, BRM can be regarded as a good model for dynamical quantum
systems such as the KR in the region of strong classical chaos. In the classical limit, this
model exhibits an unbounded diffusion in angular momentum space if the kick strength
exceeds some critical value. It was discovered that even in the deep semi-classical domain,
quantum effects can suppress classical diffusion || giving rise to a phenomenon that is
closely related to Anderson localization of a quantum particle in random potentials |22]-|24

This effect of "dynamical localization" was claimed to be experimentally observed in the
ionization of hydrogen subject to a monochromatic field |25j. A formal connection with a

ID tight-binding model has been found in [pfj , thus reviving a general interest for localization
in one-dimensional systems.

We have investigated the behaviour of M(b, t) by numerically integrating Eq. (||) for
different values of b. The results of our analysis have been compared with both theoretical
predictions of the ID Anderson model and numerical findings for the KR.

9

A. Ballistic time scale

The essential difference between periodic and disordered quantum lattice structures
mostly lies in the localization properties of their electronic states. In the periodic case,
all the states are perfectly extended Bloch waves, while in strongly disordered samples, the
states are asymptotically localized in time because of quantum interference effects. How-
ever, even in the latter case, there exists a ballistic regime, occurring on time scales of the
order of the elastic scattering time i.e the time for an electron to move by an amount
corresponding to the mean free path l m . In quasi-lD systems, l m is known to be equal to
the number of transverse channels.

In order to investigate the scaling behavior of the packet size with b, we have numerically
integrated Eq. @ for very short times ~ tf, and different values of b in the range b = 20-=-45.
The behaviour of the mean-square displacement M(b, t) is reported in Fig. 1 with the scaling
assumption

M(b,t) =b 2 M(tVb) . (8)

The very good data collapse confirms the scaling Anzatz. Eq. (§) can be understood by
estimating the ballistic time scale t&. Let us start by noticing that the leading contribution
to the wavepacket spreading over short time scales comes from the 2b sites which are directly
coupled with the site n = where the wave packet is concentrated at time t = 0. By
evaluating the Schrodinger equation and thereby determining M(b,t), one obtains

M(M)= £ n 2 \H nfi \H 2 « bH 2 . (9)

n=— b

The above type of evolution terminates when the average width a/M of the packet becomes
of the same order as the band-size b, so that farther sites come into play. By substituting
back in Eq. (||), one finds that the ballistic spread occurs on the time scale

t<t b n—. (10)

10

Accordingly, the ballistic time scale shrinks to zero for increasing the interaction range b.
Notice that the ballistic regime is entirely new with respect to the analogous problem in the
KR where, at small times, an exponentially fast spread of the packet takes place.

B. Diffusive time scale

In ID Anderson-type models, wave packet saturation starts immediately beyond the
ballistic time scale, since the mean free path l m and the localization length are of the
same order (l^ « 2l m ). Accordingly, no intermediate diffusive regime can be observed. In
the case of large band-size, b ^> 1, the localization length ~ b 2 is much larger than the
mean free path l m ~ b, so that a diffusive time scale appears. In order to estimate tu,
we shall follow the scaling arguments developed from the theory of quantum chaos where
they have been successfully introduced to explain the well known "quantum suppression of

classical diffusion" in the KR [|T^].

The first crucial observation is that the eigenfunctions fE{n) are exponentially localized
in the standard basis for all energy values E in the spectrum (consequently, there is a pure-
point spectrum p{E)). Let us proceed by noticing that an initial state c n (t = 0) excites an
effective, finite number N e ff of eigenstates with corresponding energies Ei. Accordingly, the
spectrum of those eigenstates participating to the evolution of the packet is characterized
by a mean level spacing A e // ~ Ro/N e ff where = 8bv 2 is the radius of the semicircle,
see Eq.(|).

Therefore, for times t < 1/A e ff, the packet evolution does not "feel" the discreteness of
the spectrum, which is resolved over longer time scales and eventually leads to localization. A
typical evolution is reported in Fig. 2, where the dependence of the mean square displacement
M(b, t) for b = 12 is shown on a very large time scale. However, as long as tf, < t < tp, with

t D ^N eff /R (11)

the motion of the particle is analogous to the standard (classical) diffusion. This means that
M fa Dt, where D is the diffusion constant and M can be interpreted as the square of the

11

number of effectively excited, unperturbed states. This regime is clearly seen in the inset of
Fig. 2, where the evolution of M is reported in a doubly logarithmic plot.

The quantity M reaches its maximal value M max at t ~ to- The value M^ 2 X is °f t ne
same order as the total number N e jf of eigenstates that participate to the packet evolution.
Let us finally notice that the packet-width is asymptotically of the order of the localization
length of the eigenf unctions, i.e. N e ff ~ ~ b 2 (the energy dependence is here irrelevant
and can be dropped). Accordingly, the following scaling relations hold

t D ~loo/Ro D-l^Ro , (12)

where Rq is the width of the spectrum. The second estimate in ( |T2"D corresponds to the well
known relation between the localization length and the diffusion coefficient in the theory of
disordered solids, ~ npD where p is the density of states (see, for example, 0).

Eq. (|L2"|) seems to suggest that the proper scaling of the time axis is 6 3//2 . However, as
we shall see in the next section, there is convincing evidence that best scaling Ansatz, in
comparison with the direct integration of the Schrodinger equation (^), is t/b 2 . Whether the
discrepancy is to be attributed to some weakness of the above arguments or to an improper
choice of the time units in the numerical procedure it is not clear (in fact, everything works
perfectly as if we had to refer to dimensionless time units in Eq. @ normalized to <jq - see
Eq. ([[]) - rather than to the width of the spectrum R ).

C. Diffusion suppression and scaling properties

One of the most important peculiarities of the time evolution of wave packets in ID
and quasi-lD random potentials is the saturation of the width M for t — > oo. In analogy
with the evolution of wave packets in the KR 0] and from the localization properties of the
stationary problem (the localization length grows as b 2 ), one expects that for b ^> 1 the
limiting value M^b) grows as 6 4 . In order to confirm this prediction, we have performed
detailed numerical experiments in the range b = 4 4- 12. The asymptotic value M 0O (fo) has

12

been accurately determined by averaging M(b,t) over a long time after an initial transient.
From our data we have found that the dependence of on b is slightly slower than
expected, ~ b a with a ~ 3.87 ± 0.02. This anomalous behaviour is presumably to
be attributed to the presence of finite band-size corrections which are not negligible in the
range of b values that has been numerically investigated (b < 12). Our results are reported
in Fig. 3, where M(b,t) and t are divided by the asymptotic value and its square root,
respectively. While the scaling Ansatz for M follows straightforwardly from the detailed
knowledge of the localization properties, the rescaling of time axis is mainly justified a
posteriori from the resulting good data collapse that can be observed in Fig. 3.

Anyway, the nontrivial evolution during the late stages of the diffusive regime confirmed
by a direct investigation. Since generic properties of eigenstates in quasi-lD models have
been found to be similar to those of strictly ID disordered models 0, it is natural to expect
that the similarity extends to the dynamics of wave packets as well. However, even in 1D-
geometry, the analytical treatment is very difficult. Analytical results are available only in
the two opposite regimes, t <C 1 and t ^> t D . For example, the asymptotic dependence of
the mean squared displacement (u(t)) = (x 2 (t)) of packets in the long-time limit is given by

IT

(u(t)) ~ a l 2 m ( 1 - ln |/2tl m) ) ; t>>tm (13)

where l m and t m are the mean free path and the corresponding time between consecutive
back scattering processes. This estimate is based on the expression for the quantum diffu-
sion coefficient obtained in p4 |. The logarithmic singularity in (pT3|) follows from resonant

transitions occurring between pairs of the so-called Mott states |27|]28|| . Such states have a
peculiar structure characterized by two humps lying at a distance much larger than their
effective width. Since Mott states appear in pairs, the corresponding energies are very close
to each other and this results in a resonant tunneling over large distances. The influence of
Mott states on electronic properties of disordered ID models has been studied in , where
the clustering of energy levels was discovered and attributed to these states.

13

The effect of these states has been included in the study of the long-time behavior of

wave packets in the KR , where an expression similar to Eq. flT3| ) has been introduced to
describe the evolution of the mean square displacement M(t) in the momentum representa-
tion,

& m^Mt/m). t>>2D (14)

dt (t/(2D)f V ;

where D is the classical diffusion coefficient. Numerical data seems to confirm the expec-
tations (Eq. flliD ), although the presence of very large fluctuations prevent to draw a con-
vincing conclusion. Despite the better statistics of our data, the presence of the logarithmic
correction cannot be definitely assessed in BRM too.

Another approach to the problem of quantum diffusion in the presence of strong localiza-
tion has been recently suggested in (see further developments in fl4]|): it is essentially
based on a phenomenological diffusion equation for the Green function, which takes into
account backward scattering. At large times, the relaxation rate is given by |T3|

(t/2D) 2 K J

which differs from Eq. ( |T4"D by a further logarithmic factor. At the moment, all available
numerical data for the KR do not allow to draw a final conclusion in favor of either expression.
In any case, let us again remark that quantum localization in the KR is of dynamical nature
(there is no randomness in the model), so that it is not clear to what extent it is similar to
the localization of Anderson type.

Instead of focussing on the question of the exact asymptotic dependence {t — > oo) of the
mean square displacement M(b,t), it is, for the time being, more useful to limit ourselves
to provide an effective description of the wide time region that includes also the crossover
from classical diffusion to complete saturation. In the absence of any theory, we make use
of the phenomenological expression suggested in |32| (see also B)

MtM^wfi-^d— ), (16)

14

where M m , to and (3 are the three independent parameters to be determined. The first one
is obviously obtained from the asymptotic evolution, while the short-time classical diffusion
(here, we neglect the ballistic time scale which is indeed negligible for large b) provides a
further constraint to be fulfilled. In fact, for t -C to, Eq. (|IIJ) reduces to

M (b,t) = = Dt , (17)

tD

which allows expressing to in terms of the last unknown j3 which can be determined by
fitting the global behaviour of M(b,t).

The main idea behind the phenomenological expression (|i6|) is the repulsion of the energy
levels participating to the evolution of the packet. As it was previously discussed, the
diffusion rate is proportional to the mean spectral density and it remains unchanged for
t < to, according to the uncertainty principle. However, for t > tp it decreases, since
the only eigenstates that continue to contribute ("operative eigenstates" ) are those whose
energy level-spacing s satisfies the relation s < t^/t. The relative number of such spacings
(hence, the relative diffusion rate) is given by the spacing distribution p(s) for the operative
eigenstates,

where t ^> tjj is assumed. The above time dependence for large times is the core of the
phenomenological expression (|T6|) (see details in ||).

According to the above relation, the parameter (3 characterizes an effective repulsion
between those eigenstates which are excited by the initial wave packet. It is clear that
these eigenstates strongly overlap. As a result, the value of (3 can be expected to be quite
close to 1. Although these arguments are no longer valid for very long times, when the
level clustering fll6| , |33| due to the influence of Mott states becomes important, Eq. ([H]) can
still provide a sufficiently good description of the packet dynamics. Numerical experiments
done for the KR |6|,|32| yield quite a small value of (3 {(3 ~ 0.3). This result, which is
somehow contradictory with other studies (see the discussion in ) , is probably due to the
insufficiently long times considered in the simulations.

15

Our detailed numerical experiments with BRM, performed on a much longer time scale
and with high statistics, reveal quite a good correspondence with the scaling dependence
(|T6|) (see Fig. 3). The best fit of Eq. (|TJ) gives the following values: ft pa 0.8 and t D ~ 2b 2 .
Since the values of the band size b are not very large, the limiting value has been
purposely rescaled to the same level for the different 6-values. By neglecting the residual
weak deviations from a perfect scaling, we obtain pa 1.9b 2 . The most important point
of the above analysis is that the value of the repulsion parameter (5 is quite close to 1.
This means that even for very large times close to the relaxation, the approximate power
dependence l/£ ' 8 for the difference AM = M (X) (6) — M(b,t) mimics the correct dependence
AM ~ ln(t)/£ (see Eq. fll"3|)). As a result, one can treat the scaling dependence ( |16D as a
good description of both classical diffusion and its suppression due to strong localization of
eigenstates.

The asymptotic localization of the wave packet is entirely a consequence of the frozen
character of the disorder in the Hamiltonian. However, in reality, physical systems are also
subjected to time-dependent noise (this is, for instance, the case of applications to nuclear
physics [0). In the KR, the influence of a time-dependent noise was discussed for the first
time in |55|] (see also the detailed investigation in |3"6|), finding that if the strength of the
noise exceeds some critical value, then it destroys coherent effects of quantum localization
and pure classical diffusion is recovered.

The addition of noise to the BRM provides an alternative method to determine the
diffusion coefficient D from the direct computation of the linear growth of M(b,t). It is
first interesting to notice that the computation of D cannot be easily performed in practice
without the presence of noise. The reason is that in the absence of a time-dependent noise,
corrections to the linear behaviour of M(b, t) arise already at short times, thus preventing
an accurate determination of the coefficient of the linear growth. In other words, coherent
effects of quantum localization come into play even on the time scale of classical diffusion
and this results in a rather smooth transition from classical diffusion to complete relaxation
(see Fig. 3).

16

In order to estimate the critical value of the noise-strength g 2 r which destroys completely
quantum coherence and leads to pure classical diffusion, one needs to compare the shift
of levels induced by the additional noise, with the mean level spacing between operative
eigenstates. Since the latter turns out to be proportional to 1/b 2 , one can see that if the
shift AE « g 2 tn is larger than 1/b 2 , localization will be completely destroyed. Accordingly,
effects of quantum coherence should be observable only when the condition g 2 > 1/b is
satisfied. Our numerical simulations confirm this estimate: the data in Fig. 4, which refer
to b = 8, show that D/b 2 rs 1.8 in good agreement with the value found from the fit of the
scaling dependence given by (|16|). We would like to stress that diffusion due to the noise
occurs also for g < g cr however, the rate of such diffusion is different from that given by the
classical diffusion determined in the limit b — > oo (for details see, e.g., f36fl ).

In any statistical process, the analysis of average quantities provides a limited description
of the underlying properties. At least the variance should be considered in order to achieve a
more complete characterization of the phenomenon of interest. In the present case, we shall
consider the fluctuations of M(b,t) in both the diffusive (t& < £ < £d) and the relaxation
(t ^> t D ) regime. The relevant quantity to be determined is the size of sample-to-sample
fluctuations,

where the brackets (...) denote the average over different realizations of the Hamiltonian
H nm . A meaningful way to present the numerical data is through the relative amplitude
H = AM/M of the fluctuations. Its scaling properties with b are discussed in the two
following subsections.

IV. FLUCTUATIONS OF PACKET WIDTH

(19)

17

A. Diffusive time scale

As the diffusive time scale is relatively short, we have been able to determine \i for quite
large b- values (namely, b < 45), so that finite band-size corrections should be definitely
negligible. In our numerical experiments, we have integrated Eq. (|6|) up to a time t m = b 2 /10
for more than 1000 realizations of the disorder. The results are reported in Fig. 5 under the
scaling assumption

/i(M) -b^Jiit/b 2 ) . (20)

By performing a least square fit of log/i versus log b at fixed time t = b 2 /10, well inside the
diffusive regime, we have estimated 77 which turns out to be approximately 0.7 (see the inset
in Fig. 5). One should notice the substantial disagreement with the value of the exponent
recently obtained in the KR |37 ], rj « 1.0. This scaling parameter has been conjectured

to be related to the mesoscopic fluctuations of the diffusion coefficient p7fl - However, the
connection has not been entirely clarified.

B. Relaxation time scale

The next important issue concerns fluctuations around the so-called steady state distri-
bution of the wave packet in the asymptotic regime t^> tp. If one assumes that the steady
state distribution of c n (t — > 00) is characterized by an ergodic spread of the packet over
some finite size N s , and if the components c n are statistically independent, then

AM 1

a = — — « -= . 21

M \[W S

Since the components c n are directly related to the amplitudes of eigenstates and the latter
are expected to be random on the scale of their localization length, one can conclude that
N s ~ (^oo(fr)) ~ b 2 , i.e. fi w 1/6. Surprisingly, our numerical data in the range b = 4 4- 12
indicate that the scaling has a definitely different form, namely

^(t/b 2 ) « A (22)

with 5 ~ 0.7 (see Fig. 6). A detailed study shows that when the value of 5 is varied by
±0.1, the superposition of the various curves on the plateau gets appreciably worse. As
an additional check, we have performed a least square fitting of fi versus b after averaging
the curves over times t > 20b 2 . The fit presented in the inset of Fig. 6, confirms the value
5 ~ 0.7 ± 0.01 . One should note that this value is in perfect agreement with the result
found on the diffusive time scale for the factor rj, although there is no reason a priori to
expect such an equality over time scales where different physical mechanisms control the
packet evolution. We would like to stress that the anomalous scaling described by Eq. (p2|) ,
is in close agreement with the numerical data for the KR, where it has been observed that
ji ~ b~ s with an anomalous exponent 5 ~ 0.6 [14]. Note that in the KR, larger values of the
effective parameter b have been reached.

In order to better disentangle the question of fluctuations, we have investigated also the
temporal behavior of M(b, t). The main motivation for this study is the comparison between
sample-to-sample and temporal fluctuations for a typical realization of the disorder. In
practice, we have integrated the Schrodinger equation for a time up to t = 1256 2 , discarding
an initial transient time t < to = 20b 2 (which is sufficiently long for M(t) to saturate). The
Fourier power spectrum |[/(co>)| 2 of u(t) (let us recall that M(t) = (u(t))) signal has been
then averaged over more than 150 realizations. The results for b = 4, 6 and 8 are reported
in Fig. 7 in a doubly logarithmic plot. The best data collapse is obtained by assuming
that (|[/(c<j)| 2 |) ~ 6 ? with <; = 6.6. At "high" frequencies, (\U(lj)\ 2 ) exhibits a Lorentzian-
type behaviour, which turns, at low frequencies, into a weak divergence that reveals the
presence of non-trivial long-time correlations. In fact, by invoking the Wiener-Kintchin
theorem, the low-frequency tail in the spectrum of (|£7(u;)| 2 ) can be connected with the
relaxation properties of M(t) towards its asymptotic value M^. More precisely, the power-
law convergence of the type assumed in Eq. ( |16D implies a power-law divergence as
cd~ l+ P which is compatible with our data. However, the low-frequency cut-off due to the
finite time of our simulations prevents drawing a definite statement about the presence
of a truly power-law divergence. Nevertheless, we can, at least, determine the cross-over

19

frequency u c separating the two temporal regimes, which turns out to be u c ~ 0.02/6 2 .
Such a frequency is approximately 100 times smaller than the mean spacing between the
energy levels of the eigenstates which effectively participate to the evolution of the wave
packet. This observation can be taken as an indirect confirmation of the role played by
Mott states in the long-time evolution. As it has been already recalled, Mott states have
quite a specific structure: they appear in pairs and are characterized by two humps a distance
L apart. This leads to a quasi- degeneracy of the order of AE pa exp(— L/^), where L is
typically much larger than the localization length l^. Accordingly, over long time scales, a
few Mott states may dominate the packet dynamics.

For what concerns the scaling behavior of the spectrum (|£7(u;)| 2 ) with respect to b, the
estimated exponent q pa 6.6 is in perfect agreement with (p2[). Indeed, the relation between
the spectral density (or power spectrum) and the variance of the signal M(t), implies

(AM(t)) 2 = Y,(\U(lu)\ 2 ) = Y,b 6S (\U(iub 2 )\ 2 ) = b 6S (AM(t/b 2 )) 2 . (23)

By comparing Eq. (|23|) with Eq. fl22|), one obtains again 5 ~ 0.7.

In Ref. [14] it was conjectured that the above anomalous scaling can be considered as
an indication of the fractal structure of the quantum steady-state distribution c n (t — > oo).
More precisely, they argued that the asymptotic shape can be described by an ensemble
of only N s ~ statistically independent degrees of freedom (N s being the number of
"channels" where the amplitude of the wavepacket is essentially different from zero). The
same conjecture can be raised in the present case as well, although a direct check is an
extremely hard task.

V. STEADY-STATE DISTRIBUTION

A. General discussion

As mentioned above, the localization of all eigenf unctions (see Eq. (|J)) implies that for
t ^> t D the quantum steady-state f(n,t) = \c n (t)\ 2 fluctuates around an average profile

20

f s (n) = (f(n,t)). As the effective number of eigenstates composing a single wavepacket is
finite, the average profile does depend on the disorder realization. However, in the limit
b — ► oo, the number of statistically independent components diverges and sample-to-sample
fluctuations are expected to vanish. In that limit, temporal and ensemble averages should
coincide as long as the motion is ergodic.

On the basis of diagrammatic techniques, many results have been obtained for the steady
state distribution f s (x) in continuous ID models with white noise potential (here, x denotes
the position of the electron). In particular, an exponential decay f s (x) ~ exp(— \x\ /4l m )
has been predicted for the tails of f s (x) (l m , being the mean free path) |24|]. A subsequent
more accurate analysis |38| revealed the presence of the prefactor |xp 3 ^ 2 . Both findings are

included in the global expression derived in |3

j-w - i£r^ v+ *> • < 24 »

In fact, the above expression implies that, close to the origin, the spatial dependence is
purely exponential,

f s (x) ~ exp(-\x\/l m ) ; x < l m (25)

while the asymptotic decay is described by

f a (x) ~ |xr 3/2 exp(-|x|/4/ m ) ; x > 4Z m . (26)

Therefore, the above two equations reveal that the decay rate S(x) = (In f s (x))' (the prime
denotes derivative with respect to the argument), changes by a factor 4. One consequence
of the non purely exponential behavior is that the average size of the saturated packet is
two times smaller than the asymptotic localization length (\x\) = 2l m . It is interesting
to notice also that the asymptotic dependence (1261) is similar to that for the conductance

of ID samples of finite size [|0| in the strong localized regime; in this case x is the ratio
of the sample size with the localization length. As no analytical results are available for
quasi- ID systems, in the next section we shall compare our numerical results with the above
expressions, by fitting the only free parameter l m .

21

In any case, some information on the steady state distribution f s (n) can be obtained
from the structure of the eigenstates by exploiting the following equality,

fs(n) = J2\<p m (n )(p m (n)\ 2 , (27)

m

where tp m (n) is the n-th component of the eigenstate with energy E m and n is the position
of the initial 5— like packet. Therefore, determining the asymptotic shape of a wave packet is
tantamount to determining the average correlation properties of single eigenstates. Although
no rigorous results are known in this direction, a phenomenological approach allowed to shed
some light on the closely related KR problem P, p0| , p2| . Because of the analogies between
BRM and the KR, it is instructive to compare the results for the steady distribution as well.
A rough estimation of the tails of f s (n) can be obtained in the following way (see |T3|J41

for details). If one assumes the simple exponential form (f m (n) ~ exp(— |m — n|/ioo)> for the
m-th eigenstate, Eq. ( pT|) leads to the expression

/.(„)- .^-fc^l) , (28)

which, in turn, implies l s = l^, where l s is the localization length of the asymptotic packet
(defined from the probability amplitude, i.e. taking the square root of Eq. (|2"B|)). However,
this result is inconsistent with the numerical data for the KR which instead indicate that

l s ~ 2/oo |yj. To explain the latter result, it was suggested to take into account the large

fluctuations of the eigenstates around their shape,

1 / 1 771 — Tl I \

fm(n) = exp Vimn) , (29)

where £ mn is a Gaussian noise with the zero mean and a variance [II],

< (A£ m „) 2 > = D 3 \m-n\, D s ~ 1//^ . (30)

By inserting the Ansatz ( p9|) in Eq. (|27|) and averaging over the noise term, it was found
that

( \n — m\\

(\(Pm{n)\) ~expl I , (31)

22

implying that the linear average yields a different localization length compared to (|28|).

By repeating the same calculations for the expression (p7[), one obtains that f s ~ exp(\n—
no\/loc,), which is found to be in agreement with the numerical results for the KR model as
it implies l s = 21^.

However, relation (f27|) implies that the average < |0 m (n)| 2 > should be used rather than
{\4>m(n)\). In such a case, the result is

which implies that f s ~ exp(\n — nol/2/oo). The correctness of this expression is confirmed
by the relation l s = 41^ that it implies.

B. Numerical data

In order to determine the asymptotic shape of the wave packet, we have followed the
evolution of an initially 5-like packet for times t > 120ti> The distribution f s (n) has been
then obtained by averaging over more than 150 realizations for several 6-values in the range
b = 4 12. The results are reported in Fig. 8 with the by-now-standard scaling hypothesis,

f s (x) = b 2 f s (n), x = n/b 2 , (33)

that is once more confirmed by the good data collapse.

A peculiarity of all our simulations is that f s (no) is larger than the neighboring values
by approximately a factor 3. The reason of this apparent anomaly can be traced back to the
specific 5-like shape of the initial packet (that implies Eq. (|27|)) and to the spatial random
structure of the eigenvectors. Indeed, the latter assumption, together with the observation
that only a finite number L of channels effectively contribute to the sum in Eq. (p7|), leads
to

/.(no) « L ( V \n )) ; f s (n + n ) « L 2 (<p 2 (n)) (<p 2 (n)) . (34)

23

Since it is known that (ip A ) = 3/L, while (if 2 ) = 1/L for the eigenfunctions of matrices
belonging to the Gaussian Orthogonal Ensemble ||, we obtain that f s (n )/ 'f s (n) = 3 for n
close to but different from uq. This value of the ratio is in pretty good agreement with the
above mentioned numerical estimate.

Moreover, the numerical results reported in Fig. 8 strongly suggest that the decay of
f s {n) in the vicinity of no is definitely faster than in the tails. Therefore, it is very tempting
to compare this data with the theoretical dependence derived for ID disordered models
(Eq. (|24f) ). The best fit of the only free parameter gives / m ~ 0.29; the corresponding curve
is shown in Fig. 8 (see the solid line). The very good agreement between the numerical
results and the analytical curve over a broad range of x-values suggests that a properly
modified theory to include the determination of the mean free path from first principles
should be able to account for the asymptotic properties of packets in quasi-lD systems as
well.

One should also notice that the dependence of the slope on the distance from the center
n of the packet is an entirely new feature with respect to the analogous problem in the KR
4"T| , |T3"|] , where no evidence of two distinct regions of localization has been found. The reason

of this discrepancy is not clear: on the one hand it is possible that this reflects an actual
difference between the two models, on the other hand, it is possible that the accuracy of

the numerical data for the KR |4T| is not enough to reveal this peculiarity in the associated
profile f s (n).

A further difference is the variation of the localization length of the eigenstates with
the energy in BRM (see Eq. (|5D). Far from the center of the packet, we expect that the
decay is dominated by the longest localization length Zoo(0) = 2/3 (in b 2 units). On the
other hand, from Eq. (|26|), we find that l s = 8l m ~ 2.32, which is only slightly smaller
than 4^(0) ~ 2.66. Accordingly, the equality l s = 4/^ found in the phenomenological
theory for KR and explained by invoking the presence of strong fluctuations of the individual

eigenstates ET| appears to hold also in the present case. The small deviation is presumably to

be attributed to the not yet vanishing contributions of more localized eigenstates. Instead,

24

if we average over all energies, we obtain (l^E)) = 0.5 (in b 2 units) which results to a
localization length l s = 4(Z 0O (£ 1 )) = 2 for the total wavefunction. Moreover, it is interesting
to notice that a direct determination of l s by fitting the profiles reported in Fig. 8 with a
pure exponential law, yields l s ~ 2: this means that the multiplicative correction l/|x| 3 / 2
in expression (^Bj) is essential for a correct estimate of l s , if the range of n/b 2 is not large
enough.

A further more direct confirmation of the presence of strong fluctuations in the various
eigenstates is obtained by determining the localization length lj® of the asymptotic packet
from the logarithmic average ((ln(/ s ))) of the single packets. The resulting profile, reported
in Fig. 9, yields ~ 1.3 to be compared with the value l s ~ 2.32, obtained from the
arithmetic average. Interestingly enough, the ratio between the two lengths and l s for
the shape of saturated wavepacket, is approximately equal to 2 as known for the single
eigenstates (see the previous subsection).

C. Fluctuations of the steady-state

In this section we investigate directly the nature of wave-packet fluctuations in the asymp-
totic regime. This enables us to test the correctness of a conjecture relative to the single
eigenvectors raised in the context of the KR. Indeed, fluctuations are very important in that
they allow explaining the difference between the decay rate of the wave packet and that of
the eigenvectors.

Having in mind Eq. (p9j), we computed the logarithm of f s (n) and studied its variance
at a distance An = 6b 2 from the center of the packet (averaging also over a small window of
5 neighboring sites, under the assumption that the fluctuations are nearly constant in such
interval). As a result, we have found that the distribution function of y = ln(/ s ) is, with a
good accuracy, a Gaussian (see for instance the histogram reported in Fig. 10 which refers
to the case 6 = 5 and is the result of 5,000 simulations with independent realizations of the
disorder). This represents a first confirmation that the hypotheses made in the KR can be

25

profitably carried over to the present model.

A more complete information is obtained by studying the fluctuations of y for different
values of An. More precisely, we have computed the variance

a 2 (n) = ((ln/ s (n)) 2 )-(ln/,(n)) 2 (35)

in the steady-state regime. The results for the cases 6 = 5^-8 are reported in Fig. 11 with
the scaling assumption a 2 (n) = b 2 a 2 (x), where x = An/6 2 . The data shows that for large
x, the variance grows linearly with x, indicating that the logarithm of the profile diffuses
around the average value. This is a further confirmation of the validity of a relation of the
type (j29D for the wave packet and it is strengthened by the observation that the estimated
slope is approximately equal to 1 in scaled units.

This behaviour is also analogous to what happens in ID disordered models, where it is
assumed that the logarithm of the absolute value of the Green function In \G(m,n; E)\ has
a Gaussian probability distribution with the mean equal to — \m — nl/l^ and the variance
equal to \m — nl/l^ .

Another interesting observation concerns the central part of the packet (i.e., |x| < 1),
where the variance a 2 is almost constant, indicating that the amplitude values are essentially
independent of one another (see also the inset in Fig. 11). Accordingly, all numerical findings
do confirm the conjectures that have been so far utilized to present all the features of
wavepacket diffusion in a coherent manner. Unfortunately, so far there are no analytical
results concerning the structure of the eigenstates in the middle of their localization region,
even in the well-studied ID Anderson case.

VI. CONCLUSIONS

In the present paper we have studied the evolution properties of wavepackets in quasi
ID disordered media described by tight-binding Hamiltonians with long-range random in-
teractions. We have found that the wavepacket: (i) first spreads ballistically, over a time

26

scale of the order t ~ 1/5 ' 5 , which becomes negligible in the limit b — > oo; (ii) exhibits a
diffusive behaviour, for times of the order t ~ b 2 \ (iii) finally, for times larger than to > b 2 ,
stops spreading remaining asymptotically localized.

The scaling properties of the spread of the packet are different in the ballistic (see
Eq. ([TOD) and diffusive (see Eq. flT2"|)) regime. Beyond the ballistic regime, we propose
the heuristic formula (|16|) to effectively describe both the diffusive spreading and the even-
tual saturation. The most interesting feature of Eq. (|i~6|) is the prediction of a power-law
convergence of the wavepacket width M to its asymptotic value, the deviation going to zero
as 1/t 13 with (3 ~ 0.8. The parameter (3 accounts for the repulsion between those eigenstates
which are effectively excited by the initial wavepacket. More precisely, the slow convergence
of M is attributed to effect of Mott states, that are expected to give eventually rise to a
logarithmic time dependence, that cannot be numerically observed.

The "short" time diffusive regime has been investigated by computing dM/dt when a
small amount of time- dependent disorder was superimposed to the quenched disorder. As
a result we found that the diffusion coefficient in dimensionless units is approximately four
times larger than the localization length, after averaging over the energy dependence. The
presence of noise with strength g 2 of the order of a critical value g 2 r ~ 1/6, destroys quantum
coherence and recovers classical diffusion.

Another issue addressed in this paper concerns the fluctuations of the size of the packet
over different time scales. We found that for both diffusive and saturation time scales,
the relative amplitude of the fluctuations scales as fi = AM/M ~ l/b v where 77 ~ 0.7.
We confirmed this law by taking the Fourier power spectrum |[/(c<j)| 2 of sample to sample
fluctuations of M. This result, besides confirming the anomalous scaling behaviour provides
evidence for the ergodicity of the evolution.

The study of the Fourier power spectrum shows that the tails of |£7(u;)| 2 have a
Lorentzian-type form. At small frequencies, a weak singular behaviour for |£7(u;)| 2 has been
detected; this result is in agreement with the power-law convergence of M to its asymptotic
value. Furthermore, we have been able to extract the cross-over frequency u c , where devia-

27

tions from 1/uo 2 start in |£/(u;)| 2 . Such a frequency is two orders of magnitude smaller that
the mean average separation between the energy levels which participate to the evolution
of the wavepacket. This is a further confirmation of the participation of Mott states to the
long-time evolution.

For what concerns the asymptotic shape of the wavepackets, we found that the scaling
law (^) is well verified already for relatively large b values. Moreover, the analytic expression
derived in the context of the ID Anderson model (see Eq. (|24f) ) reproduces pretty well the
shape of the average profile, upon fitting a single parameter. A further interesting result
of our numerical analysis concerns the difference by a factor 4 between the decay rate of
the asymptotic profile around the origin and that along the tails. Moreover, the difference
between the localization length estimated from the average of the logarithm of the profile and
that obtained from the linear average of the profile confirms the relevance of fluctuations.
This observation is reinforced by the asymptotic linear growth of the variance a 2 of the
logarithm of the profile versus the distance from the center of the packet.

VII. ACKNOWLEDGMENTS

We acknowledge useful discussions with B. Chirikov, Y. Fyodorov, A. Mirlin, S. Ruffo and
D. Shepelyansky. (F.M.I) is grateful to A. Bulgac for illuminating discussions of the related
problems in nuclear physics application. (T. K.) and (G. P. T.) acknowledge discussions on
the Anderson localization with C. Soukoulis and E. N. Economou. They also acknowledge
partial support from Human Capital and Mobility grant CHRXCT930331. (T. K.) acknowl-
edge the support of Grant CHRX-CT93-0107 and also wishes to thank Instituto Nazionale di
Ottica for the kind hospitality during the 1995 fall. Partial support by grant INTAS-94-2058
is also acknowledge by (F.M.I).

28

REFERENCES

[1] Statistical Theories of Spectra: Fluctuations, ed. C.E.Porter, Academic Press, (1965).
[2] M.L.Metha, Random Matrices, Academic Press, New York, (1967).
[3] T.Brody et al, Rev.Mod.Phys., 53 385 (1981).

[4] O.Bohigas and M.Giannoni, Lecture Notes in Physics, 209 1 (1984).

[5] G.Casati, B.V.Chirikov, J.Ford and F.M.Izrailev, Lect. Notes Phys., 93 334 (1979).

[6] F.M.Izrailev, Phys. Rep., 196 299 (1990).

[7] F.Haake, Quantum Signatures of Chaos, (Springer, Berlin, 1991).
[8] G.Casati, I.Guarneri, F.M.Izrailev and R.Scharf, Phys. Rev. Lett., 64 5 (1990).
[9] Y.F.Fyodorov and A.D.Mirlin, Int. J.Mod.Phys., 8 3795 (1994).
[10] F.M.Izrailev, Chaos, Solitons and Fractals, 5 1219 (1995).

[11] E.P.Nachmedov, V.N.Prigodin and u.A.Firsov, Zh. Eksp. Teor. Fiz., 92 2133 (1987)
[Sov. Phys. JETP, 65 1202 (1987)].

[12] B.V.Chirikov, F.M.Izrailev and D.L.Shepelyansky, Sov. Sci. Rev., C 2 209 (1981).

[13] B.V.Chirikov, F.M.Izrailev, and D.L.Shepelyansky, Physica D, 33 77 (1988).

[14] G.Casati and B.Chirikov, in: Quantum Chaos: Between Order and Disorder, G.Casati
and B.Chirikov eds., Cambridge Univ. Press, Cambridge, 1994.

[15] G.Casati and B.Chirikov, Physica D, 86 220 (1995).

[16] T. Dittrich, Phys. Rep., 271 269 (1996).

[17] F.M.Izrailev, T.Kottos, A.Politi, S.Ruffo, and G.Tsironis, Europhys. Lett., 34 441
(1996).

[18] Y.V.Fyodorov and A.D.Mirlin, Phys. Rev. Lett., 67 2405 (1991).

29

[19] The only condition is that the amplitude of the matrix elements has to decrease suffi-
ciently fast in such a way that the second moment of the envelope is finite.

[20] G.Casati, L.Molinari and F.M.Izrailev, Phys. Rev. Lett., 64 1851 (1990).

[21] M.I. Molina and G.P. Tsironis Phys. Rev. Lett., 73 464 (1994).

[22] P.W.Anderson, Phys. Rev., 109 1492 (1958).

[23] N.F.Mott and W.D.Twose, Adv. Phys., 10 107 (1961).

[24] V.L.Berezinskii, Zh. Eksp. Teor. Fiz., 65 1251 (1973) [Sov. Phys. JETP, 38 620 (1974)].
[25] G.Casati, B.V.Chirikov, I.Guarneri and D.Shepelyansky, Phys. Reports, 154 77 (1987).
[26] S.Fishman, D.R.Grempel and R.E.Prange, Phys.Rev.Lett., 49 508 (1982).
[27] N.Mott, Phyl. Mag., 22 7 (1970).

[28] N.F.Mott and E.A.Davis, Electronic Processes in Non- Crystalline Materials, Oxford,
1971.

[29] L.P Gor'kov, O.N. Dorokhov and F.V. Prigara, Zh. Eksp. Teor. Fis., 84 1440 (1983);
85 1470 (1983) [Sov. Phys. JETP, 57 838 (1983); 58 852 (1983)].

[30] D.Cohen, Phys. Rev., A 44 2292 (1991).

[31] B.V.Chirikiv, Chaos, 1 95 (1991).

[32] G.P.Berman and F.M.Izrailev, Preprint 497-F, Institute of Physics, Krasnoyarsk, 1988.
[33] T.Dittrich and U.Smilansky, Nonlinearity, 4 59 (1991); ibid. 85.

[34] A. Bulgac, D.Kusnezov and G.Do Dang, Ann. Phys., 242 1 (1995); Phys. Rep., 264 67
(1966); in Recent Progress in Maby Body Theories, eds. E. Schachinger, H. Mitter and
H. Sormann, vol.4, p.293 (1995).

[35] E.Ott, T.M.Antonsen,Jr and J.D.Hanson, Phys.Rev.Lett., 53 2187 (1984).

30

[36] T.Dittrich and R.Graham, Ann. Phys. (N. Y.), 200 363 (1990).
[37] G.Casati, B.V.Chirikov and F.M.Izrailev, 1994, unpublished.

[38] A. A. Gogolin, V.I.Melnikov and E.L.Rashba, Zh. Eksp. Teor. Fis.,69 327 (1975) [Sov.
Phys. JETP, 42 168 (1976)].

[39] A. A. Gogolin, Zh. Eksp. Teor. Fis., 71 1912 (1976) [Sov. Phys. JETP, 44 1003 (1976)].

[40] M.R.Zirnbauer, Phy. Rev. Lett., 69 1584 (1992).

[41] B.V.Chirikov and D.L.Shepelyansky, Radiofizika, 29 1041 (1986).

[42] E. N. Economou, Green's Functions in Quantum Physics, Springer Series in Solid State
Physics, Vol. 7 (Springer- Verlag, Berlin, 1979).

31

Figure Captions

Fig. 1 Scaling of M(b, t) vs. time for the ballistic time scale. The reported values of b
are b = 20, 25, 30, 35, 40, 45. The inset shows the lost of scaling for t > i&.

Fig. 2 An example of wave-packet diffusion beyond the ballistic time scale for b = 12.
The same curve is reported (in a doubly logarithmic plot) in the inset, magnifying the early
stages of diffusion, to testify the linear behaviour.

Fig. 3 Mean square displacement M for b = 4, 5, 6, 7, 8, 10, 12. The smooth curve
corresponds to the phenomenological expression (16). In the inset the same quantities are
shown for shorter times.

Fig. 4 Diffusion of packets with additional noise for 6 = 8. The diffusive constant is
determined from simulations performed for different noise strengths g 2 .

Fig. 5 Relative fluctuations on the diffusive time scale. The data is scaled with respect
to the law (20). In the inset, a least square fit for the same quantity and different b- values
is shown; the time is fixed, t = 0.16 2 .

Fig. 6 Relative fluctuations in the saturation regime. The values of b are

4, 5, 6, 7, 8, 10, 12. In the inset, a least square fitting is shown for the same quantity which
now is averaged over time from t = 20b 2 up to t = 1206 2 .

Fig. 7 The Fourier spectrum (\U(u))\ 2 ) for the cases b = 4, 6, 8 and for times t > 20b 2 .

Fig. 8 Asymptotic average profile of the wavepacket for b = 4 4- 12 after rescaling. The
inset shows the behavior near to the maximum; smooth lines follow from the theoretical
expression (|24|) .

Fig. 9 Logarithm of the steady-state f n for b = 5.

Fig. 10 Distribution of log f n for 6 = 5.

Fig. 11 Variance of log/ n in the steady-state after rescaling the for b =

5, 6, 7, 8, 10. The straight line is the fit for large values of x. The inset shows the behavior
close to the maximum.

32

10'

Fig. 1

_Q

10 1 r

,0

10 r

10

-1

10

-2

10

-3

10

-2

j i i i

10

-1

10

o

j i

ob=20
□ b=25
ob=30
Ab=35
vb=40
*b=45

j i '

10 10'

tb 05

Fig.2

Fig.4

20 40 60 80 t 100

Fig.7

log(cob 2 )

Fig.8

_12 -8 -4 4 8 12

x=n/b 2

Fig.10