a

r
X
i
v
:
c
o
n
d
-
m
a
t
/
9
7
0
7
3
0
1
v
1


2
9

J
u
l

1
9
9
7
Random Matrix Theories in Quantum Physics: Common Concepts
Thomas Guhr, Axel M ullerGroeling, and Hans A. Weidenm uller
MaxPlanckInstitut f ur Kernphysik, Postfach 103980, 69029 Heidelberg, Germany
(July 10, 1997)
We review the development of randommatrix theory (RMT) during the last decade. We em-
phasize both the theoretical aspects, and the application of the theory to a number of elds. These
comprise chaotic and disordered systems, the localization problem, manybody quantum systems,
the CalogeroSutherland model, chiral symmetry breaking in QCD, and quantum gravity in two
dimensions. The review is preceded by a brief historical survey of the developments of RMT and of lo-
calization theory since their inception. We emphasize the concepts common to the abovementioned
elds as well as the great diversity of RMT. In view of the universality of RMT, we suggest that
the current development signals the emergence of a new statistical mechanics: Stochasticity and
general symmetry requirements lead to universal laws not based on dynamical principles.
PACS numbers: 02.50.Ey, 05.45.+b, 21.10.-k, 24.60.Lz, 72.80.Ng
Keywords: Random matrix theory, Chaos, Statistical manybody theory, Disordered solids
MPI preprint H V27 1997, submitted to Physics Reports
Contents
I Introduction 4
II Historical survey: the period between 1951 and 1983 9
A RMT: the early period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
B From 1963 to 1982: consolidation and application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
C Localization theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
D The supersymmetry method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
E RMT and classical chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
III Theoretical aspects 22
A Results of classical Random Matrix Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1 Correlation functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 Form of the probability density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Analytical results for the Gaussian ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4 Analytical results for the rotation noninvariant case . . . . . . . . . . . . . . . . . . . . . . . . 27
5 Twolevel correlation functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
B Analysis of data and spectral observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1 Unfolding procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2 Nearest neighbor spacing distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Longrange spectral observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4 Fourier transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5 Superposition of independent spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6 GSE test of GOE data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7 Ergodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
C Introduction to supersymmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1 Supersymmetric representation of the generating function . . . . . . . . . . . . . . . . . . . . . . 40
2 Saddlepoint approximation and nonlinear model . . . . . . . . . . . . . . . . . . . . . . . . 42
D Scattering systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1
1 General aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2 Scattering and Random Matrix Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
E Wave functions and widths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1 Results derived in the framework of classical RMT . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2 Results derived in the framework of the supersymmetry method . . . . . . . . . . . . . . . . . . 48
F Crossover transitions in spectral correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1 Diusion in the space of ordinary matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2 Diusion in the space of supermatrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3 Numerical and analytical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
G Crossover transitions in other observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
H Parametric level motion and parametric level correlations . . . . . . . . . . . . . . . . . . . . . . . . 56
1 Level motion and curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2 Parametric correlation functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3 Gaussian random processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
IV Manybody systems 62
A Atomic nuclei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
1 Basic features of nuclei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2 Spectral uctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3 Resonances and scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 Model systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5 Invariances and symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
B Atoms and molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
1 Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2 Molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
C General aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
1 To what extent do the Gaussian ensembles model manybody systems properly? . . . . . . . . . 75
2 Are manybody systems chaotic? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
V Quantum chaos 76
A Historical development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
B Dynamical localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
C Billiards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
D Demonstrative experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
1 Microwave cavities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2 Acoustics and elastomechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
E Quantum dots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
1 Magnetoconductance and level correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2 Coulomb blockade regime and wave function statistics . . . . . . . . . . . . . . . . . . . . . . . . 90
F Hydrogen atom in a strong magnetic eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
G Model systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
1 Coupled oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2 Anisotropic Kepler problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
H Classical phase space and quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
I Towards a proof of the Bohigas conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
1 Periodic orbit theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
2 Supersymmetric eld theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3 Structural invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
J Summary: quantum chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
VI Disordered mesoscopic systems 100
A Energy eigenvalue statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
1 Diusive regime and WignerDyson statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
2 Ballistic and localized regimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3 Critical distribution at the metal insulator transition . . . . . . . . . . . . . . . . . . . . . . . . 109
B Persistent currents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
1 Early theory, and the three experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
2 Analytical theory for noninteracting electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
2
3 Theories with electronelectron interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
C Transport in quasi onedimensional wires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
1 Universal conductance uctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
2 DorokhovMelloPereyraKumar equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3 Nonlinear model for quasi 1d wires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4 Equivalence of nonlinear model and DMPK equation . . . . . . . . . . . . . . . . . . . . . . 128
D Random band matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
E Transport in higher dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
1 Scaling theory and distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
2 Renormalization group analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3 Supersymmetry approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
F Interactionassisted coherent transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
1 Idea and scaling picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
2 Microscopic approaches and renements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
3 Finite particle density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
G RMT in condensed matter physics a summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
VII Field theory and statistical mechanics 144
A RMT and interacting Fermions in one dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
1 Continuous matrix model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
2 Use of Dysons Brownian motion model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
3 Further results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
B QCD and chiral Random Matrix Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
1 Evidence for generic features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
2 Chiral random matrix ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
3 The chiral phase transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4 Current lines of research and open problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
C Random matrices in eld theory and quantum gravity . . . . . . . . . . . . . . . . . . . . . . . . . . 152
1 Preliminary remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
2 Planar approximation to eld theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3 RMT and two-dimensional quantum gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
VIII Universality 156
IX Common concepts 160
A Synopsis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
B Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3
FIG. 1. Nearest neighbor spacing distribution for the Nuclear Data Ensemble comprising 1726 spacings (histogram)
versus s = S/D with D the mean level spacing and S the actual spacing. For comparison, the RMT prediction labelled GOE
and the result for a Poisson distribution are also shown as solid lines. Taken from Ref. 1.
I. INTRODUCTION
During the last ten years, Random Matrix Theory (RMT) underwent an unexpected and rapid development: RMT
has been successfully applied to an ever increasing variety of physical problems.
Originally, RMT was designed by Wigner to deal with the statistics of eigenvalues and eigenfunctions of complex
manybody quantum systems. In this domain, RMT has been successfully applied to the description of spectral
uctuation properties of atomic nuclei, of complex atoms, and of complex molecules. The statistical uctuations of
scattering processes on such systems were also investigated. We demonstrate these statements in Figs. 1, 2 and 3,
using examples taken from nuclear physics. The histogram in Fig. 1
1
shows the distribution of spacings of nuclear
levels versus the variable s, the actual spacing in units of the mean level spacing D. The data set comprises 1726
spacings of levels of the same spin and parity from a number of dierent nuclei. These data were obtained from
neutron timeofight spectroscopy and from highresolution proton scattering. Thus, they refer to spacings far from
the groundstate region. The solid curve shows the randommatrix prediction for this nearest neighbor spacing
(NNS) distribution. This prediction is parameterfree and the agreement is, therefore, impressive. Typical data
used in this analysis are shown in Fig. 2
2
. The data shown are only part of the total data set measured for the
target nucleus
238
U. In the energy range between neutron threshold and about 2000 eV, the total neutron scattering
cross section on
238
U displays a number of wellseparated (isolated) resonances. Each resonance is interpreted as
a quasibound state of the nucleus
239
U. The energies of these quasibound states provide the input for the statistical
analysis leading to Fig. 1. We note the scale: At neutron threshold, i.e. about 8 MeV above the ground state,
the average spacing of the swave resonances shown in Fig. 2 is typically 10 eV! What happens as the energy E
increases? As is the case for any manybody system, the average compound nuclear level spacing D decreases nearly
exponentially with energy. For the same reason, the number of states in the residual nuclei (which are available for
decay of the compound nucleus) grows strongly with E. The net result is that the average width of the compound
nucleus resonances (which is very small compared to D at neutron threshold) grows nearly exponentially with E.
In heavy nuclei, D already a few MeV above neutron threshold, and the compoundnucleus resonances begin
to overlap. A few MeV above this domain, we have D, and the resonances overlap very strongly. At each
bombarding energy, the scattering amplitude is a linear superposition of contributions from many (roughly /D)
resonances. But the lowenergy scattering data show that these resonances behave stochastically. This must also
apply at higher energies. Figure 3
3
conrms this expectation. It shows an example for the statistical uctuations
(Ericson uctuations
4
) seen in nuclear cross sections a few MeV above neutron threshold. These uctuations are
stochastic but reproducible. The width of the uctuations grows with energy, since ever more decay channels of the
compound nucleus open up. Deriving the characteristic features of these uctuations as measured in terms of their
variances and correlation functions from RMT posed a challenge for the nuclear physics community.
These applications of RMT were all in the spirit of Wigners original proposal. More recently, RMT has found
a somewhat unexpected extension of its domain of application. RMT has become an important tool in the study
of systems which are seemingly quite dierent from complex manybody systems. Examples are: Equilibrium and
transport properties of disordered quantum systems and of classically chaotic quantum systems with few degrees of
freedom, twodimensional gravity, conformal eld theory, and the chiral phase transition in quantum chromodynamics.
4
FIG. 2. Total cross section versus neutron energy for scattering of neutrons on
238
U. The resonances all have the same spin
1/2 and positive parity. Taken from Ref. 2.
FIG. 3. Dierential cross section at several lab angles versus proton c.m. energy (in MeV) for the reaction
26
Mg(p, p)
26
Mg
leaving
26
Mg in its second excited state. Taken from Ref. 3.
5
FIG. 4. The nearest neighbor spacing distribution versus s (dened as in Fig. 1) for the Sinai billiard. The histogram
comprises about 1000 consecutive eigenvalues. Taken from Ref. 5.
FIG. 5. Nearest neighbor spacing distribution versus s (as in Fig. 1) for the hydrogen atom in a strong magnetic eld. The
levels are taken from a vicinity of the scaled binding energy

E. Solid and dashed lines are ts, except for the bottom gure
which represents the GOE. The transition Poisson GOE is clearly visible. Taken from Ref. 6.
6
0 1 2 3 4
s
0.0
0.3
0.6
0.9
p
(
s
)
FIG. 6. Nearest neighbour spacing distribution for elastomechanical modes in an irregularly shaped quartz crystal. Taken
from Ref. 7.
FIG. 7. The dierence G of the conductance and its mean value, in units of e
2
/h, of a 310 nm wire at a temperature of
0.01 K versus magnetic eld strength H in Tesla. Taken from Ref. 8.
7
Figures 4 to 7 show several cases of interest. The Sinai billiard is the prime example of a fully chaotic classical system.
Within statistics, the NNS distribution for the quantum version shown
5
as the histogram in Fig. 4 agrees perfectly
with the RMT prediction (solid line). The successful application of RMT is not conned to toy models like the Sinai
billiard. Rydberg levels of the hydrogen atom in a strong magnetic eld have a spacing distribution
6
which once
again agrees with RMT (Fig. 5). The same is true
7
of the elastomechanical eigenfrequencies of irregularly shaped
quartz blocks (Fig. 6). And the transmission of (classical or quantum) waves through a disordered medium shows
patterns similar to Ericson uctuations. Figure 7
8
shows the uctuations of the conductance of a wire of micrometer
size caused by applying an external magnetic eld. Such observations and independent theoretical progress paved the
way for the rapid development of RMT.
Some of these developments were triggered or at least signicantly advanced by the introduction of a new tool into
RMT: Efetovs supersymmetry technique and the ensuing mapping of the randommatrix problem onto a nonlinear
supersymmetric model. This advance has, in turn, spurred developments in mathematical physics relating, among
other items, to interacting Fermion systems in one dimension (the CalogeroSutherland model), to supersymmetric
Lie algebras, to Fourier transformation on graded manifolds, and to extensions of the ItzyksonZuber integral.
As the title suggests, we use the term Random Matrix Theory in its broadest sense. It stands not only for the
classical ensembles (the three Gaussian ensembles introduced by Wigner and the three circular ensembles introduced
by Dyson) but for any stochastic modelling of a Hamiltonian using a matrix representation. This comprises the
embedded ensembles of French et al. as well as the random band matrices used for extended systems and several
other cases. We attempt to exhibit the common concepts underlying these Random Matrix Theories and their
application to physical problems.
To set the tone, and to introduce basic concepts, we begin with a historical survey (Sec. II). Until the year 1983,
RMT and localization theory (i.e., the theory of disordered solids) developed quite independently. For this reason,
we devote separate subsections to the two elds. In the years 1983 and 1984, two developments took place which
widened the scope of RMT enormously. First, Efetovs supersymmetric functional integrals, originally developed for
disordered solids, proved also applicable to and useful for problems in RMT. This was a technical breakthrough. At
the same time, it led to a coalescence of RMT and of localization theory. Second, the Bohigas conjecture established
a generic link between RMT and the spectral uctuation properties of classically chaotic quantum systems with few
degrees of freedom. It was these two developments which in our view largely triggered the explosive growth of RMT
during the last decade.
Subsequent sections of the review are devoted to a description of this growth. In Sec. III we focus attention on formal
developments in RMT. We recall classic results, dene essential spectral observables, give a short introduction into
scattering theory and supersymmetry, and discuss wavefunction statistics, transitions and parametric correlations. It
is our aim to emphasize that RMT, in spite of all its applications, has always been an independent eld of research
in its own right. In Sec. IV we summarize the role played by RMT in describing statistical aspects of systems
with many degrees of freedom. Examples of such systems were mentioned above (atoms, molecules, and nuclei) but
also include interacting electrons in ballistic mesoscopic devices. Section V deals with the application of RMT to
quantum chaology, i.e. to the quantum manifestations of classical chaos. Here, we restrict ourselves to systems
with few degrees of freedom. Important examples are mesoscopic systems wherein the electrons move independently
and ballistically (i.e., without impurity scattering). Section VI deals with the application of RMT to disordered
systems, and to localization theory. This application has had major repercussions in the theory of mesoscopic systems
with diusive electron transport. Moreover, it has focussed theoretical attention on novel issues like the spectral
uctuation properties at the mobility edge. Treating the Coulomb interaction between electrons in such systems
forms a challenge which has been addressed only recently. The need to handle the crossover from ballistic to diusive
electron transport in mesoscopic systems has led to a common view of both regimes, and to the discovery of new and
universal laws for parametric correlation functions in such systems. However, we neither cover the integer nor the
fractional quantum Hall eect. Section VII deals with the numerous applications of RMT in onedimensional systems
of interacting Fermions, in QCD, and in eld theory and quantum gravity. Section VIII is devoted to a discussion of
the universality of RMT. This concerns the question whether or not certain statistical properties are independent of
the specied probability distribution function in matrix space.
In our view, the enormous development of RMT during the last decade signals the birth of a new kind of statistical
mechanics (Dyson). The present review can be seen as a survey of this emergent eld. The evidence is growing
that not only disordered but also strongly interacting quantum systems behave stochastically. The combination of
stochasticity and general symmetry principles leads to the emergence of general laws. Although not derived from rst
dynamical principles, these laws lay claim to universal validity for (almost) all quantum systems. RMT is the main
tool to discover these universal laws. In Sec. IX, we end with general considerations and speculations on the origins
and possible implications of this new statistical mechanics.
The breadth of the eld is such that a detailed account would be completely beyond the scope of a review paper.
As indicated in the title, we focus attention on the concepts, both physical and mathematical, which are either basic
8
to the eld, or are common to many (if not all) of its branches. While we cannot aim at completeness, we attempt
to provide for the last decade a bibliography which comprises at least the most important contributions. For earlier
works, and for more comprehensive surveys of individual subelds, we refer the reader to review articles or reprint
collections at appropriate places in the text.
We are painfully aware of the diculty to give a balanced view of the eld. Although we tried hard, we probably
could not avoid misinterpretations, imbalances, and outright oversights and mistakes. We do apologize to all those who
feel that their work did not receive enough attention, was misinterpreted, or was unjustifyably omitted alltogether.
In each of the nine sections, we have adopted a notation which has maximum overlap with the usage of the literature
covered in that section, often at the expense of consistency between dierent sections. We felt that this approach
would best serve our readers. We could not always avoid using the same symbol for dierent quantities. This is
the case, for instance, for the symbol P. Most often, P denotes a probability density. Context and argument of the
symbol should identify the quantity unambiguously.
We could not have written this review without the continuous advice and help of many people. Special thanks are
due to David Campbell, editor of Physics Letters, for having triggered the writing of this paper. We are particularly
grateful to J. Ambjrn, C. Beenakker, L. Benet, R. Berkovits, G. Casati, Y. Gefen, R. Hoerbert, H. Koppel, I.
Lerner, K. Lindemann, E. Louis, C. Marcus, A. D. Mirlin, R. Nazmitdinov, J. Nygard, A. Richter, T. Seligman,
M. Simbel, H. J. Stockmann, G. Tanner, J.J.M. Verbaarschot, and T. Wettig for reading parts or all of this review,
and/or for many helpful comments and suggestions. Part of this work was done while the authors were visiting CIC,
UNAM, Cuernavaca, Mexico.
II. HISTORICAL SURVEY: THE PERIOD BETWEEN 1951 AND 1983
As mentioned above, in this period RMT and localization theory developed virtually independently, and we therefore
treat their histories separately.
After a period of rapid growth during the 1950s, RMT was almost dormant until it virtually exploded about ten
years ago. Our history of RMT therefore consists of three parts. These parts describe (i) the early period (1951 till
1963) in which the basic ideas and concepts were formulated, and the classical results were obtained (Sec. II A); (ii) the
period 1963 till 1983 in which the theory was consolidated, relevant data were gathered, and some fundamental open
problems came to the surface (Sec. II B); (iii) the almost simultaneous introduction of the supersymmetry method
(Sec. II D) and of the Bohigas conjecture (Sec. II E) around 1983.
For a long time, applications of RMT have essentially been conned to nuclear physics. The reason is historical:
This was the rst area in physics where the available energy resolution was ne enough, i.e. of the order of D, and
the data set large enough, to display spectral uctuation properties relevant for tests of RMT. This is why aside
from a description of the technical development of RMT Secs. II A and II B deal almost entirely with problems in
nuclear theory. We will not mention again that here and in other systems, spectroscopic tests of the theory always
involve levels of the same spin and parity or, more generally, of the same symmetry class.
After its inception by Anderson in 1958, localization theory received a major boost by the work of Mott and, later,
by the application of scaling concepts. This is the history described in Sec. II C. For pedagogical reasons, this section
precedes the ones on supersymmetry and chaos.
In this historical review, we refer to review papers wherever possible. We do so at the expense of giving explicit
references to individual papers.
A. RMT: the early period
Most references can be found in Porters book
9
and are not given explicitly. Our account is not entirely historical:
History serves as an introduction to the relevant concepts.
In 1951, Wigner proposed the use of RMT to describe certain properties of excited states of atomic nuclei. This
was the rst time RMT was used to model physical reality. To understand Wigners motivation for taking such a
daring step, it is well to recall the conceptual development of nuclear theory preceding it.
In the scattering of slow neutrons by mediumweight and heavy nuclei, narrow resonances had been observed.
Each of these resonances corresponds to a longlived compound state of the system formed by target nucleus and
neutron. In his famous 1936 paper, N. Bohr
10
had described the compound nucleus as a system of strongly interacting
neutrons and protons. In a neutroninduced nuclear reaction, the strong interaction was thought to lead to an almost
equal sharing of the available energy between all constituents, i.e. to the attainment of quasiequilibrium. As a
consequence of equilibration, formation and decay of the compound nucleus should be almost independent processes.
9
FIG. 8. Photograph of Niels Bohrs wooden toy model for compoundnucleus scattering. Taken from Ref. 10.
In his appeal to such statistical concepts, Bohr prepared the ground for Wigners work. In fact, RMT may be seen
as a formal implementation of Bohrs compound nucleus hypothesis. At the same time, it is remarkable that the
concepts and ideas formulated by Bohr have a strong kinship to ideas of classical chaotic motion which in turn are
now known to be strongly linked to RMT. This is most clearly seen in Fig. 8
11
which is a photograph of a wooden
model used by Bohr to illustrate his idea. The trough stands for the nuclear potential of the target nucleus. This
potential binds the individual nucleons, the constituents of the target, represented as small spheres. An incoming
nucleon with some kinetic energy (symbolized by the billard queue) hits the target. The collision is viewed as a
sequence of nucleonnucleon collisions which have nearly the character of hardsphere scattering.
In the absence of a dynamical nuclear theory (the nuclear shell model had only just been discovered, and had not
yet found universal acceptance), Wigner focussed emphasis on the statistical aspects of nuclear spectra as revealed
in neutron scattering data. At rst sight, such a statistical approach to nuclear spectroscopy may seem bewildering.
Indeed, the spectrum of any nucleus (and, for that matter, of any conservative dynamical system) is determined
unambiguously by the underlying Hamiltonian, leaving seemingly no room for statistical concepts. Nonetheless, such
concepts may be a useful and perhaps even the only tool available to deal with spectral properties of systems for which
the spectrum is suciently complex. An analogous situation occurs in number theory. The sequence of prime numbers
is perfectly well dened in terms of a deterministic set of rules. Nevertheless, the pattern of occurrence of primes
among the integers is so complex that statistical concepts provide a very successful means of gaining information
on the distribution of primes. This applies, for instance, to the average density of primes (the average number of
primes per unit interval), to the rootmeansquare deviation from this average, to the distribution of spacings between
consecutive primes, and to other relevant information which can be couched in statistical terms.
The approach introduced by Wigner diers in a fundamental way from the standard application of statistical
concepts in physics, and from the example from number theory just described. In standard statistical mechanics,
one considers an ensemble of identical physical systems, all governed by the same Hamiltonian but diering in initial
conditions, and calculates thermodynamic functions by averaging over this ensemble. In number theory, one considers
a single specimen the sequence of primes and introduces statistical concepts by performing a running average
over this sequence. Wigner proceeded dierently: He considered ensembles of dynamical systems governed by dierent
Hamiltonians with some common symmetry property. This novel statistical approach focusses attention on the generic
properties which are common to (almost) all members of the ensemble and which are determined by the underlying
fundamental symmetries. The application of the results obtained within this approach to individual physical systems
is justied provided there exists a suitable ergodic theorem. We return to this point later.
Actually, the approach taken by Wigner was not quite as general as suggested in the previous paragraph. The
ensembles of Hamiltonian matrices considered by Wigner are dened in terms of invariance requirements: With every
Hamiltonian matrix belonging to the ensemble, all matrices generated by suitable unitary transformations of Hilbert
space are likewise members of the ensemble. This postulate guarantees that there is no preferred basis in Hilbert
space. Many recent applications of RMT use extensions of Wigners original approach and violate this invariance
principle. Such extensions will be discussed later in this paper.
It is always assumed in the sequel that all conserved quantum numbers like spin or parity are utilized in such
a way that the Hamiltonian matrix becomes blockdiagonal, each block being characterized by a xed set of such
quantum numbers. We deal with only one such block in many cases. This block has dimension N. The basis states
in Hilbert space relating to this block are labelled by greek indices like and which run from 1 to N. Since Hilbert
space is innitedimensional, the limit N is taken at some later stage. Taking this limit signals that we do
not address quantum systems having a complete set of commuting observables. Taking this limit also emphasises the
generic aspects of the randommatrix approach. Inasmuch as RMT as a new kind of statistical mechanics bears
some analogy to standard statistical mechanics, the limit N is kin to the thermodynamic limit.
10
Using early grouptheoretical results by Wigner
12
, Dyson showed that in the framework of standard Schrodinger
theory, there are three generic ensembles of random matrices, dened in terms of the symmetry properties of the
Hamiltonian.
(i) Timereversal invariant systems with rotational symmetry. For such systems, the Hamiltonian matrix can be
chosen real and symmetric,
H
mn
= H
nm
= H

mn
. (2.1)
Timereversal invariant systems with integer spin and broken rotational symmetry also belong to this ensemble.
(ii) Systems in which timereversal invariance is violated. This is not an esoteric case but occurs frequently in
applications. An example is the Hamiltonian of an electron in a xed external magnetic eld. For such systems, the
Hamiltonian matrices are Hermitean,
H
mn
= [H

]
nm
. (2.2)
(iii) Timereversal invariant systems with halfinteger spin and broken rotational symmetry. The Hamiltonian
matrix can be written in terms of quaternions, or of the Pauli spin matrices

with = 1, 2, 3. The Hamiltonian has

the form
H
(0)
nm
1
2
i
3

=1
H
()
nm

, (2.3)
where all four matrices H
()
with = 0, . . . , 3 are real and where H
(0)
is symmetric while H
()
with = 1, 2, 3 are
antisymmetric.
In all three cases the probability of nding a particular matrix is given by a weight function P
N
(H) times the
product of the dierentials of all independent matrix elements. Both the symmetry properties (2.1), (2.2), and
(2.3), and the weight functions P
N
(H) for = 1, 2, 4 are invariant under orthogonal, unitary, and symplectic
transformations of the Hamiltonian, respectively. The index is often used to specify the ensemble altogether. In
a sense specied by Dyson, the three ensembles are fundamentally irreducible and form the basis of all that follows.
Novel ensembles not contained in this list may arise
13,14
when additional symmetries or constraints are imposed. Such
ensembles have recently been discussed in the context of Andreev scattering
15
, and of chiral symmetry
16
.
The choice P
N
(H) = 1 would be consistent with the symmetry requirements but would lead to divergent integrals.
For the Gaussian ensembles considered by Wigner, the weight functions P
N
are chosen to have Gaussian form,
P
N
(H) exp
_

2
trH
2
_
. (2.4)
We have suppressed a normalization factor, see Sec. III A2. The constant is independent of N. The factor N
ascertains that the spectrum of the ensemble remains bounded in the limit N . The independent elements of
the Hamiltonian are independent random variables; the distributions factorize. For a more formal discussion of the
three Gaussion ensembles we refer to Sec. III A.
The choice (2.4) denes the three canonical ensembles: The Gaussian orthogonal ensemble (GOE) with = 1. the
Gaussian unitary ensemble (GUE) with = 2, and the Gaussian symplectic ensemble (GSE) with = 4. These
ensembles have similar weight functions P
N
but dierent symmetries, and consequently dierent volume elements in
matrix space. As mentioned before, this review covers a variety of Random Matrix Theories. For clarity, we refer to
the ensembles introduced above jointly as to Gaussian Random Matrix Theory (GRMT).
The introduction of RMT as a theory of physical systems poses two questions. (a) How are predictions on observables
obtained from RMT? (b) How are such predictions compared with physical reality? A third question arises when we
note that the requirements (i) to (iii) dening the three classical ensembles are based on general symmetry principles
and their quantummechanical implementation, while the choice of the Gaussian functions (2.4) was dictated by
convenience. We must therefore ask: (c) Are the conclusions derived from the Gaussian ansatz generally valid, i.e.
independent of the Gaussian form (which would then indeed serve only as a convenient vehicle)?
Basic for most applications of GRMT is the distinction between average quantities and their uctuations. Because
of the cuto due to the Gaussian weight factors, all three Gaussian ensembles dened above have a spectrum which in
the limit N is bounded and has length 4. In the interval 2 E 2, the average level density has the shape
of a semicircle (cf. Eq. (2.7) below). For most physical systems, a bounded spectrum with semicircle shape is totally
unrealistic. Therefore, GRMT is generically useless for modelling average properties like the mean level density. The
situation is dierent for the statistical uctuations around mean values of observables. Since there are N levels in the
11
spectrum, D N
1
tends to zero as N . In this limit, uctuation properties may become independent of the
form of the global spectrum and of the choice of the Gaussian weight factors, and may attain universal validity. This
is the expectation held upon employing GRMT, see in particular Sec. VIII. The wide range of successful applications
of RMT in its Gaussian form has for a long time suggested that this expectation is justied. Balian
17
derived the
three Gaussian ensembles from a maximum entropy principle, postulating the existence of a second moment of the
Hamiltonian. His derivation shows that in the absence of any further constraints, the Gaussian weight factors are
the must natural ones to use. This result and numerical studies gave further credence to the belief that uctuation
properties should not depend on the Gaussian form of the weight factors. There is now also direct evidence for this
assertion. It is presented in Sec. VIII.
GRMT can thus be used to predict uctuations of certain observables. The general scheme is this: The average of
the observable over the ensemble serves as input, and the uctuation properties are derived from GRMT. Examples
are local level correlation functions (with the average level density as input to dene the physical value of N/),
distribution laws and local correlation functions for eigenvectors (which are usually measured by coupling to some
external eld, the average of the square of the coupling matrix element serving as input), etc. We will encounter
numerous examples of this type as we proceed. In some select cases, it is possible to work out the uctuation
properties of the observable completely, and to calculate the entire distribution function over the ensemble. In most
cases, however, this goal is too ambitious, and one has to settle for the rst few moments of the distribution.
In comparing such predictions with experiment, we must relate the average over the (ctitious) ensemble given by
GRMT with information on a given system. Here, an ergodic hypothesis is used. It says that the ensemble average
is equal to the running average, taken over a suciently large section of the spectrum, of almost any member of the
ensemble. In specic cases, this ergodic hypothesis has been proved
18
.
It is clear that RMT cannot ever reproduce a given data set in its full detail. It can only yield the distribution
function of, and the correlations between, the data points. An example is the nuclear level spectrum at neutron
threshold, the rst case ever studied in a statistically signicant fashion. The actual sequence of levels observed
experimentally remains inaccessible to RMT predictions, while the distribution of spacings and the correlations
between levels can be predicted. The same holds for the stochastic uctuations of nuclear cross sections, for universal
conductance uctuations of mesoscopic probes, and for all other applications of RMT.
GRMT is essentially a parameterfree theory. Indeed, the only parameter () is xed in terms of the local mean
level spacing of the system under study. The successful application of GRMT to data shows that these data uctuate
stochastically in a manner consistent with GRMT. Since GRMT can be derived from a maximum entropy principle
17
,
this statement is tantamount to saying that the data under study carry no information content. It is an amazing fact
that out of this sheer stochasticity, which is only conned by the general symmetry principles embodied in GRMT,
the universal laws alluded to above do emerge.
The distribution of the eigenvalues and eigenvectors is obviously a central issue in applications of GRMT. To
derive this distribution, it is useful to introduce the eigenvalues E

and the eigenvectors of the matrices as new

independent variables. The grouptheoretical aspects of this procedure were probably rst investigated by Hua
19
, cf.
also Ref. 20. After this transformation, the probability measures for the three Gaussian ensembles factorize. One
term in the product depends only on the eigenvalues, the other, only on the eigenvectors (angles). Moreover, because
of the assumed invariance properties, the functions P
N
depend on the eigenvalues only. Therefore, eigenvectors and
eigenvalues are uncorrelated random variables. The form of the invariant measure for the eigenvectors implies that in
the limit N , these quantities are Gaussian distributed random variables.
Following earlier work by Scott, Porter and Thomas suggested this Gaussian distribution for the eigenvectors
in 1956. Their paper became important because it triggered many comparisons between GRMT predictions and
empirical data. In nuclear reaction theory, the quantities which determine the strength of the coupling of a resonance
to a particular channel are the reduced partial width amplitudes, essentially given by the projection of the resonance
eigenfunction onto the channel surface. Identication of the resonance eigenfunctions with the eigenfunctions of the
GOE implies that the reduced partial width amplitudes have a Gaussian probability distribution centered at zero,
and that the reduced partial widths (the squares of the partial width amplitudes) have a
2
distribution with one
degree of freedom (the PorterThomas distribution). This distribution obtains directly from the volume element in
matrix space and holds irrespective of the choice of the weight function P
N1
(H). It is given in terms of the average
partial width which serves as input parameter.
For the eigenvalues, the invariant measure takes the form ( = 1, 2, 4)
P
N
(E
1
, . . . , E
N
)

m>n
[E
m
E
n
[

l=1
dE
l
. (2.5)
This expression displays the famous repulsion of eigenvalues which becomes stronger with increasing . (Eigenvalue
repulsion as a generic feature of quantum systems was rst discussed by von Neumann and Wigner
21
. The arguments
12
used in this paper explain the dependence of the form (2.5)). We note that this level repulsion is due entirely to
the volume element in matrix space, and independent of the weight function: It inuences local rather than global
properties. It turned out to be very dicult to deduce information relevant for a comparison with experimental
data from Eq. (2.5). Such a comparison typically involves the correlation function between pairs of levels, or the
distribution of spacings between nearest neighbors. The integrations over all but a few eigenvalues needed to derive
such information were found to be very hard.
Lacking exact results, and guided by the case N = 2, Wigner proposed in 1957 a form for the distribution p(s) of
spacings of neighboring eigenvalues, i.e. for the nearest neighbor spacing distribution mentioned in the Introduction.
Here, s is the level spacing in units of the local mean level spacing D. This Wigner surmise, originally stated for
= 1, has the form
p

(s) = a

exp(b

s
2
). (2.6)
The constants a

and b

are given in Eq. (3.51) of Sec. III B2. The Wigner surmise shows a strong, dependent level
repulsion at small spacings (this is a reection of the form (2.5)), and a Gaussian fallo at large spacings. This fallo
has nothing to do with the assumed Gausian distribution of the ensembles and stems directly from the assumed form
of the volume element in matrix space. The solid line in Fig. 1 shows the sdependence of p(s) for the orthogonal
case.
In the same paper, Wigner also derived the semicircle law mentioned above. As a function of energy E, the average
level density = D
1
is given by
(E) =
N

_
1 (E/(2))
2
. (2.7)
Later, Pastur
22
derived a quadratic equation for the averaged retarded or advanced Green function. This equation
coincides in form with the scalar version of the saddle point equation (3.89) of the nonlinear model and yields for
(E) the semicircle law (2.7).
The diculties encountered in calculating spectral properties of GRMT were overcome in 1960 when M. L. Mehta
introduced the method of orthogonal polynomials, summarized in his book
23
. Mathematically, his methods are
related to Selbergs integral. Mehtas work provided the longmissing tool needed to calculate the spectral uctuation
properties of the three canonical ensembles, and therefore had an enormous impact on the eld. Mehta could not
only prove earlier assertions but also obtained many new results. For instance, Gaudin and Mehta derived the exact
form of the nearest neighbour spacing distribution. The Wigner surmise turns out to be an excellent approximation
to this distribution. The method of orthogonal polynomials has continued to nd applications beyond the canonical
ensembles.
GRMT is conceived as a generic theory and should apply beyond the domain of nuclear physics. Early evidence
in favor of this assertion, and perhaps the earliest evidence ever in favor of the nearest neighbor spacing distribution,
was provided in 1960 by Porter and Rosenzweig. These authors analyzed spectra in complex atoms.
In a series of papers written between 1960 and 1962, Dyson developed RMT much beyond the original ideas of
Wigner. Aside from showing that there are three fundamental ensembles (the unitary, orthogonal and symplectic one),
he contributed several novel ideas. As an alternative to the Gaussian ensembles dened above, he introduced the three
circular ensembles, the circular unitary ensemble (CUE), the circular orthogonal ensemble (COE), and the circular
symplectic ensemble (CSE). In each of these ensembles, the elements are unitary matrices of dimension N rather
than the Hamiltonian matrices used in Wigners ensembles. Since the spectra of all three ensembles are automatically
conned to a compact manifold (the eigenvalues exp[i

] with = 1, . . . , N of unitary matrices are located on the

unit circle in the complex plane), there is no need to introduce the (arbitrary) Gaussian weight factors of Eq. (2.4).
This removes an ambiguity of the Gaussian ensembles and simplies the mathematical problems. Unfortunately,
these advantages are partly compensated by the fact that the physical interpretation of the eigenvalues of the circular
ensembles is not obvious.
We briey give the mathematical denitions of the three circular ensembles here because they will not be discussed
in detail in Sec. III. Each circular ensemble is dened by the same invariance postulate as the corresponding Gaussian
ensemble.
(i) The COE consists of symmetric unitary matrices S of dimension N. Each such matrix S can be written as
S = U
T
U where U is unitary and T denotes the transpose. To dene the measure, we consider an innitesimal
neighborhood of S given by S + dS = U
T
(1
N
+ idM)U where dM is innitesimal, real and symmetric. Let d
ij
with 1 i j N denote the dierentials of the elements of the matrix M. Then the measure (dS) is dened by
(dS) =

ij
d
ij
, and the COE has the probability measure
P(dS) =
(dS)
_
(dS)
. (2.8)
13
The COE is invariant under every automorphism S U
T
SU

where U

is unitary.
(ii) The CUE consists of unitary matrices S of dimension N. Each such matrix can be written as the product
S = UV of two unitary matrices U and V so that S +dS = U(1
N
+idH)V where dH is innitesimal Hermitean. The
measure (dS) is dened by (dS) =

ij
dRe
ij

i<j
dIm
ij
where d
ij
denotes the dierentials of the elements of
the matrix H. With this changed denition of (dS), the probability measure again has the form of Eq. (2.8). The
CUE is invariant under every automorphism S U

SV

where U

, V

are unitary.
(iii) The CSE consists of unitary matrices S of dimension N which have the form S = U
R
U where U is unitary, and
where R denotes the dual, dened by the combined operation of timereversal (K) and transposition, U
R
= KU
T
K
1
.
We have S + dS = U
R
(1
N
+ idM)U where dM is innitesimal, quaternionreal and selfdual. The measure (dS)
is dened in terms of the dierentials d

ij
, = 0, . . . , 3 of the four matrices M

of the quaternion decomposition of

M as (dS) =

i
d
0
ii

3
=0

ij
d

ij
. With this changed denition of (dS), the probability measure again has the
form of Eq. (2.8). The CSE is invariant under every automorphism S W
R
SW where W is unitary.
The phase angles

of Dysons ensembles are expected to have the same local uctuation properties as the eigenval-
ues E

of the corresponding Gaussian ensembles. This expectation is borne out by calculating the invariant measure.
It carries the factor

>
[ exp(i

) exp(i

)[

, in full analogy to Eq. (2.5). Using the GaudinMehta method,

Dyson was able to derive the klevel correlation function (i.e. the probability density for k phase angles

with
= 1, . . . , k N, irrespective of the position of all others). These functions coincide with the corresponding ex-
pressions for the Gaussian ensembles found by Mehta, strengthening the belief that the Gaussian weight factor is
immaterial for local uctuation properties. The explicit formulas for the twolevel correlation functions can be found
in Sec. III A5.
Aside from furnishing useful alternatives to the Gaussian ensembles, the circular ensembles have found direct
application in stochastic scattering theory since the elements of the ensembles the unitary matrices S can be
viewed as scattering matrices.
Dyson also noticed an interesting connection between GRMT and a classical Coulomb gas. For the Gaussian
ensembles, the factors appearing in Eq. (2.5) can be written in the form
exp
_
N
N

n=1
_
E
n

_
2
+

2

m>n
ln
_
E
m
E
n

_
2
_
. (2.9)
From a thermodynamic point of view, this expression is the free energy of a static Coulomb gas of N particles in one
dimension with positions E
1
, . . . , E
N
and temperature
1
. The gas is conned by a harmonic oscillator potential,
which is absent for the circular ensembles originally considered by Dyson. The parameter plays the role of an
inverse temperature. This analogy helps to understand the uctuation properties of the spectrum. Both the level
repulsion at short distance and the longrange stiness of the spectrum (see Eq. (2.10) and the remarks following it)
characteristic of GRMT become intuitively obvious.
Another important idea in Dysons work relates to cases of slightly broken symmetries or invariance properties.
The ensembles considered so far all correspond to exactly obeyed symmetries and invariances. In many applications
of RMT, it is necessary to consider cases of slight symmetry breaking. This is true, for instance, for isospinmixing
in nuclei, for parity violation in nuclei, or for timereversal invariance breaking in solids. These eects are caused by
the Coulomb interaction, the weak interaction, and an external magnetic eld acting on the electrons, respectively.
Such cases can be modeled by adding to the ensemble describing the conserved symmetry (or invariance) another one,
which violates the symmetry (or invariance) and has relative strength t with respect to the rst. More precisely, the
variances of the matrix elements in the second ensemble dier by a factor t
2
from those in the rst. By generalizing the
static Coulomb gas model of the last paragraph to a dynamical one, where the N particles, in addition to their mutual
repulsion, are also subject to dissipative forces, Dyson arrived at a Brownian motion model for the eigenvalues of a
random matrix. Starting at time t = 0 with the case of pure symmetry, and letting the Brownian motion proceed, the
eigenvalues move under the inuence of the symmetrybreaking ensemble. The Brownianmotion model has found
important applications in recent years.
We refer to the Gaussian and the circular ensembles jointly as to classical Random Matrix Theory (cRMT).
In comparing cRMT results with experiment, it is useful to have available statistical measures tailored to the fact
that the data set always comprises a nite sequence only. In 1963, Dyson and Mehta introduced several statistical
measures (the DysonMehta statistics) to test a given sequence of levels for agreement with cRMT. These measures
have found wide application. One important example is the
3
statistic, which will be discussed in detail in Sec. III B3.
Let (E) be the number of energy levels in the interval [, E], and let us assume that the spectrum has been
unfolded so that the local average level density is constant (independent of E). On this unfolded, dimensionless
energy scale , the number of levels in the interval [, ] is given by (). The
3
statistic
14

3
(L) =
_
min
A,B
1
L
_
s+L
s
( () A B)
2
d
_
s
. (2.10)
measures how well the staircase function () can be locally approximated by a straight line. The angular brackets in
Eq. (2.10) indicate an average over
s
. If subsequent spacings of nearest neighbors were uncorrelated,
3
(L) would
be linear in L,
3
(L) = L/15. An essential property of RMT is the logarithmic dependence of
3
(L) on L. It is often
referred to as the stiness of the spectrum (cf. the remarks below Eq. (2.5)). Explicit expressions for
3
(L) in all
three symmetry classes are given in Sec. III B3.
As remarked above, the compoundnucleus resonances observed in slow neutron scattering are prime candidates
for applications of GRMT. By implication, the neutron scattering cross section in this energy range is expected to
be a random process. In the Introduction, it was pointed out that already a few MeV above neutron threshold, the
resonances overlap strongly. Nonetheless, the cross section should still be stochastic. Using a simple statistical model
for the nuclear scattering matrix, Ericson proposed in 1963 that in this domain, crosssection uctuations with well
dened properties (Ericson uctuations) should exist. This proposal was made before the necessary experimental
resolution and the data shown in Fig. 3 became available. The proposal gave a big boost to nuclear reaction studies,
see the review in Ref. 24. For quite some time, the connection between Ericsons model and cRMT remained an open
question, however.
B. From 1963 to 1982: consolidation and application
In this period, nuclear reaction data became available which permitted the rst statistically signicant test of GRMT
predictions on level uctuations. Ericsons prediction of random crosssection uctuations was conrmed, and many
reaction data in the domain of strongly overlapping resonances were analyzed using his model. Theorists tackled the
problem of connecting Ericson uctuations with GRMT. Twobody random ensembles and embedded ensembles were
dened as a meaningful extension of GRMT. Symmetry breaking became a theoretical issue in GRMT. The results
were used to establish upper bounds on the breaking of timereversal symmetry in nuclei.
Much of this work is reviewed in Refs. 18,25.
A signicant test of GRMT predictions requires a suciently large data set. Collecting such a set for the sole
purpose of testing stochasticity may seem a thankless task. The opposite is the case. It is important to establish
whether stochasticity as formulated in GRMT does apply in a given system within some domain of excitation energy,
and to what extent this is the case. Moreover, once the applicability of GRMT is established, there is no room left
for spectroscopic studies involving a levelbylevel analysis. Indeed, Balians derivation of GRMT from a maximum
entropy principle
17
shows that spectroscopic data taken in the domain of validity of GRMT carry no information
content.
In the 1970s, a determined experimental eort produced data on lowenergy neutron scattering by a number of
heavy nuclei, and on proton scattering near the Coulomb barrier by several nuclei with mass numbers around 50.
The number of levels with the same spin and parity seen for any one of these target nuclei ranged typically from
100 to 200. In 1982, the totality of these data was combined into the Nuclear Data Ensemble comprising 1726
level spacings. This ensemble was tested by Haq, Pandey and Bohigas
26,1
for agreement with the GOE (the relevant
ensemble for these cases). They used the nearest neighbor spacing distribution and the
3
statistic. This analysis
produced the rst statistically signicant evidence for the agreement of GOE predictions and spectral properties of
nuclei, see Fig. 1.
The advent of electrostatic accelerators of suciently high energy and of suciently good energy resolution (which
must be better than the average width of the compoundnucleus resonances) made it possible at that time to
detect Ericson uctuations
24
in nuclear cross sections, see Fig. 3. As a function of incident energy, the intensity
of particles produced in a nuclear reaction at xed scattering angle shows random uctuations. An evaluation of
the intensity autocorrelation versus incident energy yields the average lifetime h/ of the compound nucleus, and the
autocorrelation versus scattering angle yields information on the angular momenta relevant for the reaction. From the
variance one nds the ratio of direct reactions, i.e. of particles emitted without delay, to the typically longdelayed
compoundnucleus reaction contribution. The elastic enhancement factor favors elastic over inelastic compound
nucleus reactions and is the forerunner of the weak localization correction in mesoscopic physics. The authors of
reference
24
were well aware of the fact that the phenomena discovered in nuclear reactions are generic. And indeed,
most of the concepts developed within the theory of Ericson uctuations have later resurfaced in dierent guise when
the theory of wave propagation through disordered media was developed, and was applied to chaotic and disordered
mesoscopic conductors, and to light propagation through media with a randomly varying index of refraction.
15
At the same time, intense theoretical eorts were undertaken to give a solid foundation to Ericsons model, and to
connect it with cRMT
25
. This was necessary in order to obtain a unied model for uctuation properties of nuclear
cross sections in the entire range of excitation energies extending from D (i.e. neutron threshold where GRMT
was known to work) till the Ericson domain D. The approaches started from two dierent hypotheses.
(i) The random Hamiltonian approach.
Formal theories of nuclear resonance reactions express the scattering matrix S in terms of the nuclear Hamiltonian
H, so that S = S(H). Since RMT is supposed to model the stochastic properties of H, an ensemble of scattering
matrices was obtained by replacing the nulear Hamiltonian by the GOE, i.e. by writing S = S(GOE). For example,
the shellmodel approach to nuclear reactions
27
yields for the elements S
ab
(E) of the scattering matrix
S
ab
(E) =
ab
2i

W
a
(D
1
(E))

W
b
. (2.11)
Here, E is the energy, and a, b refer to the physical channels. The indices and of the inverse propagator D

refer
to a complete set of orthonormal compound nucleus states, and D (not to be confused with the mean level spacing)
has the form
D

(E) = E

+i

c
W
c
W
c
. (2.12)
We have omitted an irrelevant shift function. The symbol H

stands for the projection of the nuclear Hamiltonian

onto the set of compound nucleus states, and the matrix elements W
a
(E) describe the coupling between these states
and the channels a. Equations (2.11,2.12) are generically valid. By assuming that H is a member of a randommatrix
ensemble, S becomes an ensemble of scattering matrices, and the challenge consists in calculating ensembleaveraged
cross sections and correlation functions. Work along these lines was carried out by Moldauer, by Agassi et al., by
Feshbach et al., and by many others. Inter alia, this approach led to an asymptotic expansion kin to impurity
perturbation theory in condensed matter physics. It is valid for D and uses the assumption that the GRMT
eigenvalue distribution can be replaced by a model with xed nearest neighbor spacings (picket fence model). In
this framework, Ericsons results could be derived. However, all attempts failed to obtain a unied theoretical GRMT
treatment valid in the entire regime from D till D. The methods developed by Mehta which had proven so
successful for level correlations did not seem to work for the more complex problem of crosssection uctuations.
Much theoretical attention was also paid to reactions at energies beyond the Ericson regime where an extension
of RMT is needed to describe physical reality. Indeed, the application of RMT to transport processes (like cross
sections) implies some sort of equilibrium assumption: By virtue of the orthogonal invariance of the GOE, all states
of the compound system are equally accessible to an incident particle, and a preferred basis in Hilbert space does not
exist. This assumption is justied whenever the internal equilibration time
eq
needed to mix the states actually
populated in the rst encounter of projectile and target with the rest of the system is small compared with the time
h/ for particle emission. But
eq
changes little with energy while h/ decreases nearly exponentially over some
energy interval. At energies above the Ericson regime there exists a domain where decay happens while the system
equilibrates. This is the domain of precompound or preequilibrium reactions. It requires an extension of RMT
using a preferred basis in Hilbert space, reecting the ever increasing complexity of nuclear congurations reached
from the incident channel in a series of twobody collisions. Formally, this extension of RMT is a forerunner of and
bears a close relationship to the modelling of quasi onedimensional conductors in terms of random band matrices.
In both cases, we deal with nonergodic systems where a novel energy (or time) scale the equilibration time or, in
disordered solids, the diusion time comes into play.
(ii) The random S matrix approach.
A maximumentropy approach for the scattering matrix kin to Balians derivation of GRMT was developed by
Mello, Seligman and others. A closed expression for the probability distribution of the elements of the scattering
matrix was derived for any set of input parameters (the average elements of S). Unfortunately, the result was too
unwieldy to be evaluated except in the limiting cases of few or very many open channels. Moreover, parametric
correlation functions, e.g. the correlation between two elements of the scattering matrix at dierent energies, seem
not accessible to this approach. Under the name random transfer matrix approach, a similar approach has later
found wide application in the theory of disordered quasi onedimensional conductors, cf. Sec. VI C2. Here again, the
calculation of parametric correlation functions has been fraught with diculties.
The random S matrix approach has the virtue of dealing directly with the elements of the S matrix as stochastic
variables. Thereby it avoids introducing a random Hamiltonian in the propagator as done in Eqs. (2.11,2.12) and the
ensuing diculties in calculating ensemble averages. This very appealing feature is partly oset by the diculties in
calculating parametric correlations.
16
A basic criticism leveled against GRMT relates to the fact that in most physical systems, the fundamental interaction
is a twobody interaction. In a shellmodel basis, this interaction has vanishing matrix elements between all states
diering in the occupation numbers of more than two singleparticle states. Therefore, the interaction matrix is
sparse. In an arbitrary (nonshell model) basis, this implies that the number of independent matrix elements is
much smaller than in GRMT, where the coupling matrix elements of any pair of states are uncorrelated random
variables. This poses the question whether GRMT predictions apply to systems with twobody forces. It is also
necessary to determine whether the twobody matrix elements are actually random. Another problem consists in
understanding how a random twobody force acts in a manyFermion Hilbert space. This problem leads to the
question of how information propagates in spaces of increasing complexity. These problems were tackled mainly by
French and collaborators
18
and led to statistical nuclear spectroscopy as an approach to understand the workings
of the twobody residual interaction of the nuclear shell model in large, but nite, shellmodel spaces. Specically,
it was shown that in the shell model, the matrix elements of the twobody force are random and have a Gaussian
distribution. The mathematical problem of handling such random twobody forces was formulated by introducing
the embedded ensembles: A random mbody interaction operates in a shellmodel space with K Fermions where
m < K. The propagation of the twobody force in complex spaces was investigated with the help of moments methods.
It was shown numerically that for m = 2, the spectral uctuation properties of such an ensemble with orthogonal
symmetry are the same as for the GOE. The last result shows that GOE can meaningfully be used in predicting
spectral uctation properties of nuclei and other systems governed by twobody interactions (atoms and molecules).
Nonetheless, embedded ensembles rather than GRTM would oer the proper way of formulating statistical nuclear
spectroscopy. Unfortunately, an analytical treatment of the embedded ensembles is still missing.
In the early 80s, Mehta and Pandey
28,29
made a signicant advance in the theoretical treatment of GRMT. Their
work established for the rst time the connection between GRMT, eld theory, and the ItzyksonZuber integral. They
succeeded in extending the orthogonal polynomial method to the problem of symmetry breaking. In this way, they
showed that the nearest neighbor spacing distribution (NNS) could be used as a test of timereversal invariance in
nuclei. The basic idea is that for small spacings, the presence of an interaction breaking timereversal symmetry would
change the linear slope of NNS characteristic of the GOE into a quadratic one typical for the GUE, cf. Eq. (2.6). To
work out this idea, it is necessary to dene a random ensemble which allows for the GOE GUE crossover transition.
The ensemble dened by Mehta and Pandey for this purpose has the form
H
nm
= H
GOE
nm
+i
_
t/NH
A
nm
. (2.13)
Here, H
GOE
is the GOE, and H
A
is the ensemble of real antisymmetric Gaussian distributed matrices having the
same variance as the GOE. For t = 0, the ensemble (2.13) coincides with the GOE, while for t = N, it coincides
with the GUE. Writing the strength parameter in the form (t/N)
1/2
is motivated by the following consideration. The
perturbation (t/N)
1/2
H
A
is expected to inuence the local uctuation properties of the spectrum (those on the scale
D) for values of t such that the meansquare matrix element of the perturbation is of order D
2
. From Eq. (2.7),
this implies (t/N)(
2
/N) D
2
(/N)
2
, or t 1. This argument shows that the local uctuation properties are
extremely sensitive to a symmetrybreaking perturbation. The results obtained by Mehta and Pandey were used by
French et al. to obtain an upper bound on timereversal symmetry breaking in nuclei.
A similar problem arises from the Coulomb interaction. This interaction is weak compared to the nuclear force and
leads to a small breaking of the isospin quantum number. The randommatrix model needed for this case is dierent
from Eq. (2.13), however. Indeed, without Coulomb interaction the Hamiltonian is block diagonal with respect to the
isospin quantum number. In the simplest but realistic case of the mixing of states with two dierent isospins, it consists
of two independent GOE blocks. The isospinbreaking interaction couples the two blocks. The strength parameter of
this coupling again scales with N
1/2
. For spectral uctuations, this crossover transition problem has not yet been
solved analytically (except for the GUE case) whereas an analytical treatment has been possible for nuclear reactions
involving isospin symmetry breaking. This case is, in fact, the simplest example for the precompound reaction referred
to above. The results provided an ideal tool for the study of symmetry breaking in statistical nuclear reactions. The
data gave strong support to the underlying statistical model
30
.
We conclude this historical review with a synopsis. RMT may be viewed as a new kind of statistical mechanics, and
there are several formal aspects which support such a view. First, RMT is based on universal symmetry arguments.
In this respect it diers from ordinary statistical mechanics which is based on dynamical principles; this is the
fundamental novel feature of RMT. It is linked to the fact that in RMT, the ensembles consist of physical systems
with dierent Hamiltonians. Second, the Gaussian ensembles of RMT can be derived from a maximum entropy
approach. The derivation is very similar to the way in which the three standard ensembles (microcanonical, canonical
and grand canonical ensemble) can be obtained in statistical mechanics. Third, the limit N is kin to the
thermodynamic limit. Fourth, in relating results obtained in the framework of RMT to data, an ergodic theorem is
used. It says that the ensemble average of an observable is for almost all members of the ensemble equal to the
17
running average of the same observable taken over the spectrum of a single member of the ensemble. In statistical
mechanics, the ergodic theorem states the equality of the phasespace average of an observable and the average of
the same observable taken along a single trajectory over suciently long time. Finally, the elements of the GRMT
matrices generically connect all states with each other. This is a consequence of the underlying (orthogonal, unitary
or symplectic) symmetry of GRMT and is reminiscent of the equal a priori occupation probability of accessible states
in statistical mechanics. The assumption is valid if the time scale for internal equilibration is smaller than all other
time scales.
In the period from 1963 till 1982, the evidence grew strongly that both in nuclear spectrocopy and in nuclear
reaction theory, concepts related to RMT are very successful. Two elements were missing, however: (i) A compelling
physical argument was lacking why GRMT was such a successful model. Of course, GRMT is a generic theory. But
why does it apply to nuclei? More precisely, and more generally: What are the dynamical properties needed to ensure
the applicability of GRMT to a given physical system? (ii) Both in nuclear spectroscopy and in nuclear reaction
theory, the mathematical tools available were insucient to answer all the relevant questions analytically, and novel
techniques were needed.
C. Localization theory
It may be well to recall our motivation for including localization theory in this historical survey. Localization
theory deals with the properties of electrons in disordered materials. Disorder is simulated by a random potential.
The resulting random singleparticle Hamiltonian shares many features with an ensemble of random matrices having
the same symmetry properties. In fact, after 1983 the two elds RMT and localization theory began to coalesce.
A review of localization theory was given by Lee and Ramakrishnan
31
in 1985. Here, we focus on those elements of
the development which are pertinent to our context, without any claim of completeness.
The traditional quantummechanical description of the resistance of an electron in a metal starts from Bloch
waves, the eigenfunctions of a particle moving freely through an ideal crystal. The resistance is due to corrections
to this idealized picture. One such correction is caused by impurity scattering. The actual distribution of impurity
scatterers in any given sample is never known. Theoretical models for impurity scattering therefore use statistical
concepts which are quite similar to the ones used in RMT. The actual impurity potential is replaced by an ensemble
of impurity potentials with some assumed probability distribution, and observables are calculated as averages over
this ensemble. It is frequently assumed, for instance, that the impurity potential V (r) is a Gaussian random process
with mean value zero and second moment
V (r
1
)V (r
2
) =
1
2
(r
1
r
2
). (2.14)
Here, the ensemble average is indicated by a bar, = (V )
1
is the density of singleparticle states
1
per
volume V , and is the elastic mean free time, i.e.
1
measures the strength of the impurity potential. In 1958,
Anderson
32
realized that under certain conditions, the standard perturbative approach to impurity scattering fails.
He introduced the concept of localization. Consider the eigenfunctions of the singleparticle Hamiltonian H = T +V
with T the kinetic energy operator and V the impurity potential dened above. For suciently strong disorder,
these eigenfunctions, although oscillatory, may be conned to a nite domain of space. More precisely, their envelope
exp(r/) falls o exponentially at the border of the domain. The scale is given by the localization length ,
and r is the distance from the center of the domain. For probes larger than the localization length, the contribution
to the conductance from such localized eigenfunctions is exponentially small. This important discovery was followed
by the proof
33
that in one dimension all states are localized no matter how small the disorder. In higher dimensions,
localization occurs preferentially in the tails of the bands where the electrons are bound in deep pockets of the impurity
potential. If there are extended states in the band center, they are energetically separated from the localized states
in the tails by the mobility edges. These characteristic energies mark the location of the metalinsulator transition.
In the 1970s, Thouless and collaborators applied scaling ideas to the localization problem
34
. Information about
localization properties can be obtained by the following Gedankenexperiment. Consider two cubic blocks of length
L in d dimensions of a conductor with impurity scattering. Connect the two blocks to each other to form a single
bigger conductor, and ask how the eigenfunctions change. The answer will depend on the ratio /L. For L , the
localized eigenfunctions in either of the smaller blocks are aected very little, and the opposite is true for L .
Thouless realized that there is a single parameter which controls the behavior of the wave functions. It is given by
the dimensionless conductance g, dened in terms of the actual conductance G as
G =
e
2
h
g. (2.15)
18
Thouless expressed g in terms of the ratio of two energies, g = g(E
c
/). Here, is the mean singleparticle level
spacing in each of the smaller blocks, not to be confused with the mean level spacing of the manybody problem
used in previous sections. The Thouless energy E
c
is dened as the energy interval covered by the singleparticle
energies when the boundary conditions on opposite surfaces of the block are changed from periodic to antiperiodic.
The Thouless energy is given by
E
c
=
hT
L
2
. (2.16)
where T is the diusion constant. We note that L
2
/T is the classical diusion time through the block. It is remarkable
that the Thouless energy, dened in terms of a classical time scale, plays a central role in localization, manifestly a
pure quantum phenomenon. For length scales L such that E
c
, an electron is multiply scattered by the impurity
potential and moves diusively through the probe. (This is true provided L is larger than the elastic mean free path
l. Otherwise, the motion of the electron is ballistic). In this diusive regime, we have
g = E
C
/ = /L. (2.17)
The last equality applies in quasi onedimensional systems only. For E
c
, L is of the order of the localization
length , and the conductance is of order unity. For even larger values of L, L, the multiple scattering of the
electron by the impurity potential leads to annihilation of the wave function by interference (localization), and the
conductance falls o exponentially with L. This picture was developed further by Wegner who used the analogy to
the scaling theory of critical phenomena.
Quantitative scaling theory asks for the change of g with the length L of the probe in d dimensions and expresses
the answer in terms of the function (L) = d ln g(L)/d ln(L). The calculation of (L) makes use of diagrammatic
perturbation theory, or of eldtheoretical methods (renormalization theory, loop expansion). The answer depends
strongly on d. In one dimension, all states are localized, there is no diusive regime, and g falls o exponentially
over all length scales L l. Abrahams et al. argued that (L) is a function of g(L) only (oneparameter scaling).
This implied that not only in one but also in two dimensions all states should be localized. The view of these authors
was largely supported later by perturbative and renormalization calculations, cf., however, Sec. VI E. For d 2,
the perturbative approach to localization made use of an expansion of the observable in powers of the impurity
potential. The resulting series is simplied with the help of the small parameter (k
F
l)
1
1 where k
F
is the Fermi
wave length and l is the elastic mean free path due to impurity scattering. This is the essence of diagrammatic
impurity perturbation theory. In the presence of a magnetic eld, some diagrams are not aected (those containing
diusons) while others (containing cooperons) become suppressed ever more strongly as the strength of the
magnetic eld increases. This discussion carries over, of course, to the eldtheoretical approach. We do not discuss
here the results of these approaches.
Prior to the development of the supersymmetry method, the eldtheoretical approach to the localization problem
made use of the replica trick, introduced by Edwards and Anderson in 1975
35
. The replica trick was originally
introduced to calculate the average of lnZ where Z is the partition function of a disordered system, for instance,
a spin glass. It was applied later also to the calculation of the ensemble average of a product of resolvents (or
propagators).
Observables like the density of states or the scattering matrix depend on the impurity Hamiltonian H typically
through advanced or retarded propagators (E

H)
1
where the energy E

carries an innitesimalIy small positive or

negative imaginary part. It is notoriously dicult to calculate the ensemble average of a product of such propagators
directly. Instead, one writes the (trace or matrix element of) the propagator as the logarithmic derivative of a
suitable generating function F. For instance, in case H is a simple Hermitean matrix of dimension N rather than a
spacedependent operator,
tr
1
E

H
=
ln F
C
(J
C
)
J
C

JC=0
(2.18)
where the generating function F
C
has the form
F
C
(J
C
) =
_
d[S] exp
_
iS

(E

1
N
H +J
C
1
N
)S
_
. (2.19)
We have introduced an N component vector S with complex entries. The integration is performed over real and
imaginary part of each of the N complex variables S
n
n = 1, . . . , N. This procedure, analogous to standard methods
in statistical mechanics, has the advantage of bringing H into the exponent of F
C
. Therefore, it is easy to average
19
F
C
if H has a Gaussian distribution. Unfortunately, calculating the observable requires averaging the logarithm of
F
C
(or of a product of such logarithms). This cannot be done without some further trick. We write
ln F
C
= lim
n0
1
n
(F
n
C
1). (2.20)
The replica trick consists in calculating the average of F
n
C
for integer n only and in using the result to perform the
limit in Eq. (2.20). Many of the subsequent steps used in the replica trick have later been applied analogously in the
supersymmetry method, see Sec. II D. We refrain from giving further details here. Suce it to say that the trick
of conning the variable n in Eq. (2.20) to integer values often limits the scope of the method: It yields asymptotic
expansions rather than exact results. A comparison between the replica trick and the supersymmetry method
36
has
shown why this is so.
An important step in the development of the eldtheoretical method was the introduction of the Norbital model
by Wegner and its subsequent analysis
3739
. In this discrete lattice model, each site carries N orbitals, and an
asymptotic expansion in inverse powers of N is possible for N . In the replica formulation, this limit denes
a saddle point of the functional integral. It was found that for observables involving a product of retarded and
advanced propagators, the saddle point takes the form of a saddlepoint manifold with hyperbolic symmetry. This
fact reects the occurrence of the mean level density which carries opposite signs in the retarded and the advanced
sector, respectively, and breaks the symmetry between retarded and advanced elds. This implies the existence of a
massless or Goldstone mode which causes singularities in the perturbation series. The resulting eld theory is closely
related to a nonlinear model. All of these features reappear in the supersymmetric formulation, except that the
topology of the saddlepoint manifold is modied because of the simultaneous presence of bosonic and fermionic
degrees of freedom.
Prior to many of the developments described above, Gorkov and Eliashberg
40
realized that the WignerDyson
approach (RMT) could be applied to the description of level spectra of small metallic particles, and to the response
of such particles to external electromagnetic elds. The success of this description implied the question whether a
derivation of RMT from rst principles was possible. This question played an important role in Efetovs work on
supersymmetry to which we turn next.
D. The supersymmetry method
At the end of Sec. II B, it was mentioned that in order to apply RMT to a number of problems of practical interest,
novel methods were needed. What was the diculty? An observable relating to transport properties typically consists
of the square of a quantummechanical amplitude, or of a product of such squares. For example, the cross section
is a sum of squares of S matrix elements, and the correlation function of the cross section is a product of sums of
squares. Each quantummechanical amplitude occurring in a transport observable contains a propagator, i.e. a factor
(EHi)
1
. Here H stands for the random matrix, and denotes a width matrix which describes the coupling to
external channels. An example is given by Eqs. (2.11, 2.12). The occurrence of is typical for transport problems.
No such matrix arises in spectral uctuation problems (i.e., in the calculation of the nlevel correlation functions).
While Mehtas orthogonal polynomial method makes it possible to calculate the latter, his method apparently fails
for the former problem. (At least we are not aware of a successful application of the method in the former case). The
reason probably is that the matrix massively breaks the symmetry which would otherwise exist between retarded
and advanced propagators. Hence, a novel method is needed to calculate ensemble averages over observables relating
to transport.
The replica trick mentioned in Sec. II C seemed to oer such a novel method. It had been applied to technically
very similar problems (scattering of electrons in a random impurity potential). It had led to an understanding of the
singularities encountered in diagrammatic perturbation theory, and to the hope of overcoming these problems in terms
of an exact integration over the Goldstone mode (the saddlepoint manifold). The rst applications of this method
to RMT scattering problems were disappointing, however. A calculation of the Smatrix autocorrelation function
41
gave results which were in agreement with but did not go beyond those obtained from an asymptotic expansion in
powers of D/ mentioned in Sec. II B under (i). This was later understood as a generic shortcoming of the replica
trick
36
. Remedy came from the supersymmetry method.
The fundamental problem in using a generating functional of the form of Eq. (2.19) or its analogue in extended
solids is a problem of normalization. Indeed, for J
C
= 0, we have F
C
(0) = det[2i/(E

H)]. The occurrence of the

logarithm in Eq. (2.18) is due to the need to remove this normalization factor F
C
(0). The replica trick accomplishes
the same aim by taking the limit n 0 of det
n
[2i/(E

H)]. Several authors

42,43
noted that this factor can
also be removed, and F
C
(0) thereby normalized to unity, by integrating in Eq. (2.18) over both normal (commuting
20
or bosonic) and anticommuting (Grassmann or fermionic) variables. This approach was developed in a series of
papers by Efetov and coworkers and led to the supersymmetry method, summarized in Refs. 44,45. A short technical
introduction to supersymmetry is given in Sec. III C.
In developing the supersymmetry method, Efetov was mainly interested in the theory of disordered metals. It was
soon realized, however, that the method is eminently suitable also for all scattering and transport problems of RMT
type
46
. In the supersymmetry formulation, the width matrix which generically appears in these problems does not
lead to diculties. The supersymmetry approach to scattering and transport problems applies likewise to problems
of disorder in d > 1 dimensions and has played an important role in making the supersymmetry method such an
ubiquitous tool for stochastic quantum problems. With the help of this method it was possible, for instance, to
calculate average nuclear cross sections and the Smatrix autocorrelation function versus energy in the entire domain
D till D
46
. This provided the solution of a longstanding problem in statistical nuclear reaction theory
described in Sec. II B and, for D, conrmed Bohrs picture of the compound nucleus.
It was mentioned above that the supersymmetry method is not applicable to Hamiltonians containing explicitly
interacting Bosons or Fermions in the language of second quantization. This is because the symmetry properties of
the fermionic or bosonic creation and annihilation operators clash with the introduction of supersymmetry. There is a
natural way to avoid this problem: Introduce a complete set of (properly symmetrized) manybody states, and write
the Hamiltonian as a matrix in the associated Hilbert space. Nothing prevents us from applying the supersymmetry
method to the resulting matrix problem. However, stochastic properties of onebody or twobody operators lead
to stochastic properties of the resulting Hamiltonian matrix which are very dicult to handle. These diculties
are similar to the ones encountered in the treatment of the embedded ensembles, see Sec. II B. Thus it is not the
supersymmetry method as such which fails in these cases but rather the complexity of the statistical properties of the
Hamiltonian which renders the problem unmanageable.
Aside from this shortcoming, the supersymmetry method suers from two additional technical restrictions. (i) The
dimension of the supermatrices in the nonlinear model grows linearly with the number k of propagators in the
kpoint function. This makes the exact (rather than asymptotic) calculation for k 4 so cumbersome that until
now very few calculations for k = 4 and none for larger k, have been published. (ii) For a long time, calculations for
disordered solids based on the nonlinear model were restricted to quasi onedimensional probes. Only recently has
it been possible to overcome this restriction.
E. RMT and classical chaos
As mentioned several times in previous sections, RMT received a signicant boost by the discovery of its connection
with classical chaos. A review may be found in Ref. 47.
In the late 1970s and early 1980s, several authors investigated the quantum spectra of conservative systems
which behave chaotically in the classical limit (i.e., are K systems in this limit). The interest in this question arose
naturally from the great attention paid at that time to classical chaotic motion. Sequences of levels pertaining to
the same symmetry class were generated numerically for the Bunimovich stadium by McDonald and Kaufman
48
and
by Casati et al.
49
, and for the Sinai billiard by Berry
50
. These data suggested agreement of the NNS distribution
with the Wigner surmise. Interest in this question was further kindled by a paper by Zaslavsky
51
which also contains
references to earlier work. In 1984, Bohigas, Giannoni and Schmit
5
produced a statistically signicant set of data,
consisting of more than 700 eigenenergies of the Sinai billiard, larger than what had been available before. Perhaps
more importantly, these authors employed the methods of statistical analysis developed for nuclear and atomic spectra
to this data set. The good agreement between both the NNS distribution and the
3
statistic calculated from the data
and the corresponding GOE results (see Fig. 8) caused these authors to formulate the following conjecture: Spectra
of timereversal invariant systems whose classical analogues are K systems show the same uctuation properties as
predicted by the GOE. The K systems mentioned in the conjecture are the most strongly mixing classical systems.
The most unpredictable K systems form a subclass, they are referred to as Bernouilli systems. An alternative,
stronger version of the conjecture, also formulated in Ref. 5, replaces K systems by less chaotic systems provided
they are ergodic. In both its forms, this conjecture is commonly referred to as the Bohigas conjecture. For systems
without timereversal invariance, the GOE is replaced by the GUE. replaced by the GUE. Originally, Bohigas et
al. formulated the conjecture without referring to the semiclassical regime, i.e. to the limit h 0. However, all
attempts to proof the conjecture are based on some kind of semiclassical approximation, see the discussion in Sec. V.
Further evidence for this conjecture was soon forthcoming from the numerical study of other chaotic twodegrees of
freedom systems, and from experimental and theoretical studies of the hydrogen atom in a strong magnetic eld.
The conjecture also suggested a (perhaps naive) physical explanation why RMT works in nuclei and other complex
quantum systems. Such an explanation, which refers to chaotic classical motion, goes beyond the mere statement of
21
generic properties which were used to introduce RMT by Wigner. In fact, in the light of the Bohigas conjecture and
later ndings, systems with classically mixed phase space, where layers of chaotic and regular motion are intimately
interwoven, would not be expected to display RMT properties in a clean fashion.
There are two problems connected with the Bohigas conjecture. The analytical understanding of the connection
between GOE and fully developed classical chaos is still incomplete, and the conjecture must be complemented by
a number of caveats. As for the rst point: A proof of the Bohigas conjecture is still missing. Apart from very
recent attempts to use a model approach
52,53
or structural invariance
54,55
to substantiate this conjecture (see
Secs. VI 2 and VI 3), the only extant analytical link between classical chaos and RMT is based on the semiclassical
approximation. Using this approximation, Berry
56
showed that for classically chaotic systems, the
3
statistic dened
in Eq. (2.10) above has the same logarithmic dependence on the interval length L as for RMT. His argument also
showed a limitation of RMT: The correspondence between chaotic systems and RMT applies only up to a maximum L
value dened by the shortest classical periodic orbit. Beyond this value,
3
becomes constant for dynamical systems.
Unfortunately, Berrys arguments could so far not be extended to address other RMT predictions such as the NNS
distribution. This is because short energy intervals relate to long periodic orbits. In chaotic systems these orbits
increase exponentially in density and thus are dicult to handle. Progress on this problem has recently been made,
however, see Sec. VI 1. The caveat that GOE properties apply only up to a maximum L value is practically relevant
only for systems with few degrees of freedom. Indeed, the semiclassical approximation is of order h while the mean
level spacing is of order h
d
with d the number of degrees of freedom. For large values of d, the length of the level
sequence needed to observe deviations from GOE is beyond experimental or numerical possibilities. Another caveat
concerns the structure of classical phase space. Fully developed chaos implies that in the course of time, a single
chaotic trajectory is ergodic (i.e., approaches every point in phase space arbitrarily closely). But this requirement is
not sucient to guarantee agreement with GOE uctuations. It is necessary that, in addition, all parts of phase space
are equally accessible. A counterexample is provided by two chaotic billiards connected by a thin channel. Here, the
conjecture would apply only in the semiclassical limit h 0. More generally, when geometry and/or Cantori cause
the system to have more than one intrinsic time scale, the connection with RMT becomes more complicated, see
Sec. VH. Some of these issues are reviewed in Ref. 57.
III. THEORETICAL ASPECTS
Although developed in the framework of Nuclear Physics and with the aim of providing a statistical approach
to the analysis and description of spectra of complex manybody systems, RMT has become a branch sui generis
of Theoretical Physics, with its own concepts and mathematical methods. RMT studies analytically the uctuation
properties and correlation functions of levels and wave functions of ensembles of random matrices, quite independently
of possible applications to physical systems. The present section is written with this fact in mind. Without going into
details of the derivations, we give a selfcontained presentation of the central concepts of RMT, and we summarize the
most important formal results. We restrict ourselves to the Gaussian ensembles, because on the physically relevant
unfolded scale, the circular ensembles give identical results. The enormous wealth of the published material forced
us to restrict ourselves to those topics which are of particular importance either for the analysis of data or for the
understanding of the relationship between RMT and elds such as chaos theory and condensed matter physics. Thus,
some important contributions of mainly mathematical interest had to be left out. The most prominent example is
the Pechukas gas, whose eldtheoretical aspects are only briey discussed in Sec. VII.
The section is organized as follows. In Sec. III A, we begin with the classical analytical study of level uctuation
properties. In Sec. III B, we discuss the analysis of data and the spectral observables. The supersymmetry method
which is used in numerous modern applications of RMT is introduced in Sec. III C. Scattering systems and the
statistics of wave functions and widths are discussed in Secs. III D and III E, respectively. The study of crossover
transitions in spectral correlations and other observables is presented in Secs. III F and III G, respectively. Parametric
level motion and the ensuing correlations are studied in Sec. III H.
A. Results of classical Random Matrix Theory
Some of the results collected below were mentioned already in the Historical Survey. We do not reiterate the intuitive
arguments given there and focus attention on the formal aspects. In our notation we follow as far as possible Mehtas
book on Random Matrices
23
. In Sec. III A1, we give the formal denition of correlation functions and discuss some
general aspects. The choices for the probability density distributions and the statistical implications and consequences
thereof are discussed in Sec. III A2. Analytical results for the rotation invariant Gaussian ensembles and the rotation
22
noninvariant ensembles are presented in Secs. III A3 and III A4. Results for the particularly important twolevel
correlations are summarized in Sec. III A5.
1. Correlation functions
We consider a spectrum of N energy levels x
n
, n = 1, . . . , N with values on the entire real axis. An ensemble
of innitely many such spectra is dened in terms of a normalized probability density P
(E)
N
(x
1
, . . . , x
N
). For the
volume element, we use the at measure

N
n=1
dx
n
. It is assumed that the levels are equivalent. This implies that
P
(E)
N
(x
1
, . . . , x
N
) is invariant under permutations of the arguments. The correlation functions describe the statistical
properties of this ensemble. According to Dyson
5860
, the kpoint correlation function is obtained by integrating the
probability density over N k arguments,
R
k
(x
1
, . . . , x
k
) =
N!
(N k)!
_
+

dx
k+1

_
+

dx
N
P
(E)
N
(x
1
, . . . , x
N
) . (3.1)
The functions R
k
in Eq. (3.1) measure the probability density of nding a level around each of the positions x
1
, . . . , x
k
,
the remaining levels not being observed. These quantities are independent of the labeling of the levels. Since
the function P
(E)
N
(x
1
, . . . , x
N
) is normalized to unity, the klevel correlation functions R
k
are normalized to the
combinatorial factors N!/(N k)!.
Frequently, correlation functions are introduced in a dierent way. Let H denote a Hamiltonian dened in a Hilbert
space with a cuto so that H has nite dimension N with N 1. The spectrum of H contains N levels and is given
by the trace of the imaginary part of the Green function, i.e. of the matrix (x

p
H)
1
where the energy x

p
= x
p
i
has a small imaginary increment. We consider an innite ensemble of such Hamiltonians. The ensemble is dened by
a probability density P
N
(H) and a at measure in matrix space. The kpoint correlation functions can be dened as
R
k
(x
1
, . . . , x
k
) =
1

k
_
P
N
(H)
k

p=1
Imtr
1
x

p
H
d[H] (3.2)
where the at measure d[H] is the Cartesian volume element of H, i.e. the product of the dierentials of all independent
elements of H. With a proper identication of P
N
(H) and of P
(E)
N
, the two denitions (3.1) and (3.2) are basically
equivalent. There is a minor dierence, however. It is easily seen that the denition (3.2) yields terms containing
factors (x
p
x
q
), independent of the form of the probability density. This is due to identities of the type
1

2
Imtr
1
x

p
H
Imtr
1
x

q
H
=
N

n=1
(x
p
x
n
)
N

m=1
(x
q
x
m
)
= (x
p
x
q
)
N

n=1
(x
p
x
n
) +

n=m
(x
p
x
n
)(x
q
x
m
) (3.3)
where the x
n
are the eigenvalues of H. (We dier in this one instance from Mehtas notation. Usually, the eigenvalues
are also denoted by x
n
.) Inserting Eq. (3.3) into the denition (3.2) for k = 2, one obtains two terms. After
integration over the eigenvalues x
1
and x
2
, the second term has the structure of the denition (3.1). In the general
case for arbitrary k, identities of the form (3.3) yield for R
k
a sum of terms. All terms but one contain at least one
function of the form (x
p
x
q
). The argument of this function does not depend on the eigenvalues x
n
. Equations (3.2)
and (3.1) are fully equivalent only when all these terms are omitted in Eq. (3.2). Unfortunately, the two denitions
(3.2) and (3.1) are sometimes not clearly distinguished in the literature.
Some general results apply without any specic assumptions on the form of the probability density. For systems
described by the Schrodinger equation, it was shown by Wigner
61
that there are three symmetry classes. For systems
which are invariant under time reversal and under space rotation, the Hamiltonian matrix H can be chosen real
symmetric. For systems with broken timereversal invariance, H is complex Hermitean, irrespective of whether
rotation invariance holds or not. For timereversal invariant systems with broken rotation invariance and halfodd
integer spin, H is quaternion real, i.e. selfdual Hermitean. Thus, H can be viewed as an N N matrix with 2 2
23
entries or as an 2N 2N matrix. According to Kramers
62
, all eigenvalues of this last type of system are doubly
degenerate. The label = 1, 2, 4 indicates the number of real parameters H
()
nm
, = 0, . . . , 1 needed to specify a
matrix element H
nm
in each of the three classes. For = 1, 2, 4, the volume element has the form
d[H] =

nm
dH
(0)
nm
1

=1

n>m
dH
()
nm
. (3.4)
We emphasize two points. First, the limit N of innite dimension has to be taken. This compensates the
eect of the technically motivated cuto in Hilbert space. Second, in order to eliminate the dependence on the mean
level density R
1
(x
1
), the dimensionless scaled variables

p
=
p
(x
p
) =
_
xp

R
1
(x

p
)dx

p
, p = 1, . . . , k (3.5)
are introduced. The correlation functions R
k
with k > 1 have to be rescaled accordingly. In analytical calculations, this
unfolding procedure is usually combined with the limit N . The unfolded correlation functions X
k
(
1
, . . . ,
k
)
are then obtained by equating the dierential probabilities on both energy scales
X
k
(
1
, . . . ,
k
)d
1
d
k
= R
k
(x
1
, . . . , x
k
)dx
1
dx
k
(N ). (3.6)
In particular we have X
1
(
1
) = 1 by construction. In a generic situation, it is sucient to perform the unfolding in a
small region of the spectrum, provided that in the limit N this region contains innitely many levels. Let the
region be centered at zero. We then put
p
= x
p
/D. The mean level spacing D is dened by D = 1/R
1
(0). We note
that the mean level spacing D depends on N. The unfolded correlation functions can be written as
X
k
(
1
, . . . ,
k
) = lim
N
D
k
R
k
(D
1
, . . . , D
k
) . (3.7)
The new energy variables
p
are held xed while the limit is taken.
2. Form of the probability density
To specify the probability density, we conne ourselves to the two classical cases which were of paramount
importance for the development of Random Matrix Theory. This will allow us to establish the relation between the
functions P
(E)
N
(x
1
, . . . , x
N
) and P
N
(H). We use the denition (3.2) which is based on the Hamiltonian.
(i) Rotation invariant density.
The probability density P
N
(H) is taken to be invariant under arbitrary rotations of matrix space, P
N
(H) =
P
N
(U
1
HU) where U is in the global unitary group U(N; ). Physically, this means that all states in matrix space
interact with each other and that no preferred direction exists. With H = U
1
XU, we introduce the elements of the
diagonal matrix X = diag(x
1
, . . . , x
N
) of the eigenvalues and the independent elements of U as new variables. Then
the volume element (3.4) has the form
23
d[H] = [
N
(X)[

d[X]d(U)

N
(X) =

n>m
(x
n
x
m
) (3.8)
where
N
(X) is the Vandermonde determinant. (To keep with the usual notation, the eigenvalues are denoted by
x
n
, in contrast to Eq. (3.3).) The integration over the group U with the invariant Haar measure d(U) is trivial and
we can identify
P
(E)
N
(x
1
, . . . , x
N
) =

C
N
P
N
(X)[
N
(X)[

. (3.9)
The constant

C
N
ensures the normalization.
(ii) Rotation noninvariant density.
All states in matrix space are uncoupled, and the rotation invariance of the probability density is maximally broken.
This situation is modeled by assigning the value zero to all odiagonal matrix elements of H, and we can write
24
P
N
(H) =
N

n=1
p
(0)
(H
(0)
nn
)
1

=0

n>m
(H
()
nm
) (3.10)
with a normalized function p
(0)
(z). The matrices of this ensemble are diagonal by construction. The level positions
x
n
in denition (3.1) are given by the diagonal elements, x
n
= H
(0)
nn
. The probability density is
P
(E)
N
(x
1
, . . . , x
N
) =
N

n=1
p
(0)
(x
n
) . (3.11)
We refer to the two cases (i,ii) as to the pure cases. Various crossover transitions are discussed later. The rotation
invariant case always contains the Vandermonde determinant and thus the famous Wigner repulsion [x
n
x
m
[

between each pair of levels. The absence of the levelrepulsion factor is typical for the rotation noninvariant case.
The occurrence or absence of the Vandermonde determinant is independent of the functional form of P
N
(X) or of
p
(0)
(z) but strongly inuences the character of the correlation functions. They are WignerDysonlike (Poissonlike)
in the presence (absence) of this determinant. This connects to the issue of universality taken up in Sec. VIII.
For the Gaussian ensembles, the functional form of the rotation invariant density P
N
(H) = P
N
(X) is determined
by the following condition. The matrix elements of H not connected by symmetry are assumed to be statistically
independent. Together with rotation invariance, this implies that P
N
(H) is a normalized Gaussian
23
,
P
N
(H) =
_

2
N+N(N1)/2
exp
_

2
tr H
2
_
. (3.12)
This yields the Gaussian Orthogonal (GOE), Unitary (GUE) and Symplectic (GSE) Ensembles. The joint probability
density of the eigenvalues has the form
P
(E)
N
(x
1
, . . . , x
N
) = C
N
exp
_

2
N

n=1
x
2
n
_
[
N
(X)[

. (3.13)
The normalization constant is given by
23
C
N
=

N/2+N(N1)/4
(2)
N/2

N
(1 +/2)

N
n=1
(1 +n/2)
. (3.14)
The variances of the Gaussians in Eqs. (3.12) and (3.13) are simply given by 1/, and the energies x
n
are measured on
a dimensionless scale. Sometimes, another scaling is employed in the literature where energies E
n
and variances 2v
2
/
are used. Here v has the dimension energy. Very often, the two cases can be mapped onto each other with the simple
substitution x
n
= E
n
/

2v
2
. In dealing with crossover transitions in Sec. III F, we use 2v
2
/ as the variance of the
Gaussian distribution. In Eq. (2.4), we have used another form for P
N
(H). This form implies the value 2
2
/N for
the variance of the Gaussian distributions. This Ndependent choice is frequently made within the supersymmetry
method to be discussed in Sec. III C. It renders the radius of the Wigner semicircle (see the Historical Survey)
independent of N. This has some technical advantages in the supersymmetry method if the limit of innitely many
levels is taken through a saddlepoint approximation, see Sec. III C. This choice yields for the mean level spacing at
the origin D 1/N. The Nindependent choice of the variance used throughout the present section makes the radius
of the Wigner semicircle grow like

N. This implies for the mean level spacing at the origin D 1/

N.
3. Analytical results for the Gaussian ensembles
The considerable mathematical diculties involved in the complete analytical calculation of the klevel correlation
function for the three Gaussian ensembles were overcome by Mehta, Gaudin and Dyson
23
. The GUE is technically the
simplest of the three Gaussian ensembles. The GOE and the GSE are much more dicult and, somewhat surprisingly
at rst sight, closely related. We use the denition (3.1) of the klevel correlation functions. These functions have a
determinant structure. For all three ensembles, they can be written as a quaternion determinant of k k quaternion
matrices built on 2 2 matrix entries
N
(x
p
, x
q
),
R
k
(x
1
, . . . , x
k
) = qdet [
N
(x
p
, x
q
)]
p,q=1,...,k
. (3.15)
25
The quaternion determinant qdet of a k k selfdual quaternion matrix Q is dened as
qdet Q =
_
det C(Q) . (3.16)
Here, C is the 2k 2k matrix constructed from Q by using the Pauli matrices as basis for the quaternion entries of
Q. If Q is selfdual, det C(Q) is a perfect square and the square root can be taken. The determinant on the right
hand side of Eq. (3.15) has precisely this property. To further illustrate the remarkably compact result (3.15), we give
explicit results for the 2 2 matrices
N
(x
p
, x
q
) which depend on only two energies x
p
and x
q
. All entries involve
the oscillator wave functions

n
(z) =
1
_
2
n
n!

exp
_

z
2
2
_
H
n
(z) (3.17)
where H
n
(z) are the standard Hermite polynomials. We dene the kernel
K
N
(x
p
, x
q
) =
N1

n=0

n
(x
p
)
n
(x
q
) (3.18)
and the functions
S
N
(x
p
, x
q
) = K
N
(x
p
, x
q
) +
1
2
_
N
2

N1
(x
p
)
_
+

sgn(x
q
z)
N
(z)dz
DS
N
(x
p
, x
q
) =

x
q
S
N
(x
p
, x
q
)
IS
N
(x
p
, x
q
) =
1
2
_
+

sgn (x
p
z)S
N
(z, x
q
)dz
JS
N
(x
p
, x
q
) = IS
N
(x
p
, x
q
)
1
2
sgn(x
p
x
q
) (3.19)
where sgn(z) is the signum function. In the case of even N, the result for the GOE is

N1
(x
p
, x
q
) =
_
S
N
(x
p
, x
q
) DS
N
(x
p
, x
q
)
JS
N
(x
p
, x
q
) S
N
(x
q
, x
p
)
_
. (3.20)
The (slight) modication for odd N can be found in Mehtas book
23
. For the GUE one has for all N

N2
(x
p
, x
q
) =
_
K
N
(x
p
, x
q
) 0
0 K
N
(x
q
, x
p
)
_
. (3.21)
The result for the GSE reads

N4
(x
p
, x
q
) =
1

2
_
S
2N+1
(

2x
p
,

2x
q
) DS
2N+1
(

2x
p
,

2x
q
)
IS
2N+1
(

2x
p
,

2x
q
) S
2N+1
(

2x
q
,

2x
p
)
_
. (3.22)
The similarity of the structures for the GOE and the GSE is apparent. In the GUE case, the correlation functions
can be written as ordinary k k determinants
R
2,k
(x
1
, . . . , x
k
) = det [K
N
(x
p
, x
q
)]
p,q=1,...,k
. (3.23)
This is due to the simple structure of Eq. (3.21) and implies that the quaternion structure of Eq. (3.15) is not essential
for = 2.
For the unfolded correlation functions X
k
(
1
, . . . ,
k
) dened in Eqs. (3.6) and (3.7), the beautiful determinant
structure (3.15) survives and the problem is reduced to unfolding the 22 matrices
N
(x
p
, x
q
) and their entries (3.18)
and (3.19). One nds on the unfolded scale in the limit of innitely many levels
23
X
k
(
1
, . . . ,
k
) = qdet [

(
p

q
)]
p,q=1,...,k
. (3.24)
The unfolded correlation functions depend only on the dierences
p

q
and are, thus, translation invariant. The
entries of the unfolded matrices

(r) are
26
s(r) =
sin r
r
Ds(r) =
d
dr
s(r) and Is(r) =
_
r
0
s(r

)dr

. (3.25)
The function s(
p

q
) = lim
N
DK
N
(x
p
, x
q
) is the unfolded kernel (3.18). Explicitly, one has

1
(r) =
_
s(r) Ds(r)
Is(r) 1/2 s(r)
_

2
(r) =
_
s(r) 0
0 s(r)
_

4
(r) =
_
s(2r) Ds(2r)
Is(2r) s(2r)
_
. (3.26)
The result for the GUE can be written in the simple form
X
2,k
(
1
, . . . ,
k
) = det [s(
p

q
)]
p,q=1,...,k
. (3.27)
The determinant structure of Eqs. (3.15) and (3.23) and, therefore, also of Eqs. (3.24) and (3.27) is entirely due to
the Vandermonde determinant [
N
(X)[

in Eq. (3.9). Therefore, this structure prevails for all probability densities
P
N
(X) which factorize in the arguments x
n
, n = 1, . . . , N. Moreover, in many cases, the unfolding will yield exactly
the same functional forms (3.25). Actually, it is shown in Sec. VIII on universality that the results given above apply
to a wide class of probability densities P
N
(X).
It is apparent from the determinant structure that the correlation functions R
k
with k > 1 contain terms which
reect the clustering of the k levels into subgroups of k

= 1, . . . , k 1 levels. In analogy to the Green functions or

propagators in eld theory, it is possible to remove such terms by introducing a cumulant or cluster expansion. The
klevel cluster function T
k
(x
1
, . . . , x
k
) is accordingly dened as the true or connected part of R
k
(x
1
, . . . , x
k
). This
means that it cannot be written in terms of lower order correlation functions. The completely disconnected part of
R
k
(x
1
, . . . , x
k
) is always the product of the k mean level densities. For k = 2, we have, in particular,
R
2
(x
1
, x
2
) = R
1
(x
1
)R
1
(x
2
) T
2
(x
1
, x
2
) . (3.28)
In general, the klevel cluster function can be written as
T
k
(x
1
, . . . , x
k
) =

1
2
tr
k

p=1

N
(x
(p)
, x
(p+1)
) (3.29)
where the sum is over all (k 1)! distinct cyclic permutations of the indices 1, 2, . . . , k. Every cycle is closed by the
denition (k + 1) = (1). On the unfolded scale, the klevel cluster functions are dened in analogy to Eqs. (3.6)
and (3.7),
Y
k
(
1
, . . . ,
k
) = lim
N
D
k
T
k
(D
1
, . . . , D
k
) . (3.30)
In the case of the Gaussian ensembles, these functions can be written as
Y
k
(
1
, . . . ,
k
) =

1
2
tr
k

p=1

(
(p)

(p+1)
) . (3.31)
Again, there are simplications for the GUE. The completely disconnected part of X
k
(
1
, . . . ,
k
) is simply unity,
i.e. the product of all unfolded mean level densities.
4. Analytical results for the rotation noninvariant case
In this case, the probability densities have the forms (3.10) and (3.11). The Vandermonde determinant is absent,
and the calculation of the correlation functions is almost trivial. As in the case of the Gaussian ensembles, we use the
denition (3.1). For the level density one nds
27
R
1
(x) = Np
(0)
(x) . (3.32)
The klevel correlation functions are simply kfold products of mean level densities,
R
k
(x
1
, . . . , x
k
) =
N!
(N k)!
k

p=1
p
(0)
(x
p
) =
N!
(N k)!N
k
k

p=1
R
1
(x
p
)
=
k

p=1
_
1
p 1
N
_
R
1
(x
p
) (3.33)
The choice of the factor on the right hand side of Eq. (3.33) guarantees the normalization of the klevel correlation
function to N!/(Nk)!. By denition, this normalization applies to arbitrary ensembles. For the Gaussian ensembles,
the completely disconnected part in the cumulant or cluster expansion of the klevel correlation function is simply the
product of all k level densities. The prefactor in front of this term is unity. The functional form of the expression (3.33)
coincides with this completely disconnected part. However, to ensure the overall normalization, its prefactor is dierent
from unity. In the limit N , all factors (1 (p 1)/N) in the last of Eqs. (3.33) yield unity. This leads to a
normalization problem on the unfolded scale since the limit N and the normalization do not commute. We will
discuss this for the twolevel cluster function in Sec. III A5. Thus, it is not very convenient to dene klevel cluster
functions.
If on the scale of the mean level spacing the function p
(0)
(z) has no structure, one has
X
k
(
1
, . . . ,
k
) = 1 for all k (3.34)
which coincides with the fully disconnected part in the case of the Gaussian ensembles. The absence of all correlations
in the spectrum is referred to as Poisson regularity and is formally expressed in terms of the klevel cluster functions
as
Y
k
(
1
, . . . ,
k
) = 0 for k > 1 . (3.35)
The opposite case of a regular but maximally correlated spectrum is represented by the harmonic oscillator. There,
p
(0)
(z) =
1
N
+N/2

n=N/2
(z nD) . (3.36)
The condition of smoothness of p
(0)
(z) is strongly violated. Hence Eqs. (3.34) and (3.35) do not apply. Instead one
nds
X
1
(
1
) =
+

n=
(
1
n) (3.37)
for the level density of an innite spectrum and
X
k
(
1
, . . . ,
k
) =
k

p=1
X
1
(
p
) (3.38)
for the correlation functions. For k > 1, these correlation functions contain terms proportional to (
p

q
). This is
due to the form (3.36) of the density p
(0)
(z).
5. Twolevel correlation functions
Twolevel correlation functions are especially important for the analysis of data. We summarize results for these
functions on the unfolded scale and for the pure ensembles. For translation invariant spectra, these functions depend
only on the energy dierence r =
1

2
and one has
X
2
(r) = 1 Y
2
(r) . (3.39)
28
0 1 2 3
r
0.0
0.5
1.0
1.5
1
-
Y
2
(
r
)
FIG. 9. The twolevel correlation functions X
2
(r) = 1 Y
2
(r) on the unfolded energy scale. The solid line is the GOE
result ( = 1), the dashed line is the GUE result ( = 2) and the dotted line is the GSE result ( = 4). We notice that
1 Y4,2(r) overshoots the value one due to strong oscillations.
In the Poisson case the connected part vanishes, Y
2
(r) = 0. For the Gaussian ensembles, one has with s(r) = sinr/r
Y
1,2
(r) = s
2
(r) +
ds(r)
dr
_

r
s(r

)dr

Y
2,2
(r) = s
2
(r)
Y
4,2
(r) = s
2
(2r)
ds(2r)
dr
_
r
0
s(2r

)dr

(3.40)
where the rst index labels the GOE, GUE and GSE, respectively. In all three cases, the normalization integral of
Y
2
(r) over the real axis yields unity. This sum rule reects the normalization mentioned in Sec. III A4. In the
Poisson case, on the other hand, this sum rule yields zero since Y
2
(r) = 0. This normalization problem arises
because the normalization for nite N and the limit N do not commute. In the Poisson case, the sum rule
is only satised for nite N. This issue will be taken up in Sec. VI A3. In Fig. 9, the twolevel cluster function is
shown for the three Gaussian ensembles. In various applications, one needs the Fourier transforms or twolevel form
factors
b
2
(t) =
_
+

Y
2
(r) exp(i2rt)dr . (3.41)
In the Poisson case, one has b
2
(t) = 0. For the Gaussian ensembles, one nds
b
1,2
(t) =
_
1 2[t[ +[t[ ln(2[t[ + 1) if [t[ 1
1 +[t[ ln
2|t|+1
2|t|1
if [t[ > 1
b
2,2
(t) =
_
1 [t[ if [t[ 1
0 if [t[ > 1
b
4,2
(t) =
_
1
1
2
[t[ +
1
4
[t[ ln [[t[ 1[ if [t[ 2
0 if [t[ > 2
(3.42)
The functions 1 b
2
(t) are shown in Fig. 10. The discontinuities of these functions or there derivatives reect the
dierent oscillatory structures of the twolevel correlation functions. We notice the singularity at [t[ = 1 in b
4,2
(t).
29
FIG. 10. Relevant parts 1 b
2
(t) of the Fourier transforms of the unfolded twolevel correlation functions
X
2
(r) = 1 Y
2
(r). The functions b
2
(t) are referred to as the twolevel form factors. The solid line is the GOE result
( = 1), the dashed line is the GUE result ( = 2) and the dotted line is the GSE result ( = 4). The dierent oscillatory
structure of the twolevel correlation functions is most strikingly reected by the behavior of the Fourier transforms at |t| = 1.
B. Analysis of data and spectral observables
In this section, we review some methods used in the statistical analysis of experimental data, and we summarize
the relevant predictions of Random Matrix Theory. The unfolding procedure is described in Sec. III B1. Then, we
discuss several statistical observables: The nearest neighbor spacing distribution (Sec. III B2), longrange spectral
observables (Sec. III B3), and Fourier transform methods (Sec. III B4). The superposition of several independent
spectra is treated in Sec. III B5. Specic properties of Random Matrix Theory may yield additional tests of data.
In Sec. III B6, we present the GSE test of GOE data as an example. The predictions of Random Matrix Theory are
based on ensemble averages while experimental results are obtained from a running average over the spectrum of a
single sample. The justication for this procedure is nontrivial and raises the question of ergodicity. This question
is briey discussed in Sec. III B7.
1. Unfolding procedure
A measurement or a calculation yields an ordered sequence of energies E
1
, E
2
, . . . , E
N
which form the stick
spectrum or spectral function,
S(E) =
N

n=1
(E E
n
) . (3.43)
To analyze the uctuation properties, this spectrum has to be unfolded, i.e. the systemspecic mean level density
R
1
(E) must be removed from the data. We dene the cumulative spectral function
(E) =
_
E

S(E

)dE

=
N

n=1
(E E
n
) . (3.44)
This function counts the number of levels with energy less than or equal to E and is also referred to as the staircase
function. It is decomposed into a smooth part (E) and a uctuating part

(E),
30
600 700 800 900
E [kHz]
0
500
1000
1500
(
E
)
820 825
E [kHz]
10
-5
10
-3
700 702 704
E [kHz]
370
380
390
(
E
)
FIG. 11. Example of an experimentally obtained staircase function. The top gure shows the cumulative spectral function
(E) for a spectrum of 1428 elastomechanical eigenfrequencies of a resonating quartz block in the frequency range between
600 kHz and 900 kHz. To keep with the notation in the text, the frequency is denoted by E. The inset shows a small section
of the measured spectrum between 820 kHz and 825 kHz. Due to the high number of levels, the staircase function appears as
a smooth line. The smooth part (E) is a polynomial whose coecients were found by a t. The bottom part shows a small
section of the staircase function between 699 kHz and 705 kHz, containing about 25 levels. The corresponding section of (E),
i.e. the polynomial t, is drawn as a thin line. We notice that (E) is obtained by tting to the entire cumulative spectral
function, not only to this section. Adapted from Ref. 7.
31
(E) = (E) +

(E) . (3.45)
The smooth part is given by the cumulative mean level density,
(E) =
_
E

R
1
(E

)dE

. (3.46)
An example for an experimentally obtained staircase function and its smooth part is shown in Fig. 11. To unfold the
spectrum, the sequence E
1
, E
2
, . . . , E
N
is mapped onto the numbers
1
,
2
, . . . ,
N
with

n
= (E
n
) , n = 1, . . . , N . (3.47)
In these new variables the cumulative spectral function simply reads
() = +

() . (3.48)
The mean level density of the unfolded spectrum, i.e. the derivative of the smooth part with respect to , is unity, as
required. Comparing Eqs. (3.5) and (3.46) we see that the present unfolding procedure coincides with the analytical
unfolding in Random Matrix Theory. Therefore, measurement and theoretical prediction can be directly compared
with each other. In practice, the separation in Eq. (3.45) of a spectrum into a smooth and a uctuating part can be
nontrivial task.
2. Nearest neighbor spacing distribution
The nearest neighbor spacing distribution p(s) is the observable most commonly used to study the shortrange
uctuations in the spectrum. This function is equal to the probability density for two neighboring levels
n
and
n+1
having the spacing s. The function p(s) involves all correlation functions R
k
with k 2 and diers from the twolevel
correlation function X
2
(r). The latter gives the probability of nding any two levels at a distance r from each other.
Only for small arguments, p(s) is approximated by X
2
(s). The function p(s) and its rst moment are normalized to
unity,
_

0
p(s)ds = 1 and
_

0
sp(s)ds = 1 . (3.49)
For the uncorrelated or Poisson case, p(s) = exp(s) while for the harmonic oscillator, p(s) = (s 1). For the
Gaussian ensembles, the analytical calculation of p(s) from the correlation functions given in Sec. III A is possible
but highly nontrivial
23
. It leads to expressions involving innite products. An excellent approximation is given by
the Wigner surmises which are the exact spacing distribution for a 2 2 matrix model
63
with Gaussian probability
density function. In the most general form the Wigner surmises reads
p

(s) = a

exp
_
b

s
2
_
(3.50)
for all three symmetry classes with = 1, 2, 4. The level repulsion factor s

reects the Vandermonde determinant

in the Nlevel probability density (3.13). The constants
a

= 2

+1
(( + 2)/2)

+2
(( + 1)/2)
and b

=

2
(( + 2)/2)

2
(( + 1)/2)
(3.51)
are explicitly given by a
1
= /2, b
1
= /4 (GOE), a
2
= 32/
2
, b
2
= 4/ (GUE), and a
4
= 262144/729
3
, b
4
= 64/9
(GSE), respectively. These functions p

(s) are shown in Fig. 12. The Gaussian fallo of p(s) with large s is
unrelated to the assumed Gaussian probability density of the three ensembles. Because of universality, the spacing
distribution (3.50) is valid for a wide class of probability densities, see Sec. VIII.
A statistical argument due to Wigner
64
leads to an interesting heuristic formula for the spacing distribution,
p(s) = (s) exp
_

_
s
0
(s

)ds

_
. (3.52)
The repulsion function (s) models the presence or absence of the Vandermonde determinant in the eigenvalue
probability density. For a linear repulsion (s) = s/2 one obtains the Wigner surmise, for a constant value (s) = 1,
the Poisson distribution. Choosing (s) = c
q
s
q
with 0 q 1 leads to the Brody distribution
65
,
32
0 1 2 3
s
0.0
0.4
0.8
1.2
p
(
s
)
FIG. 12. The Wigner surmises p

(s) for the nearest neighbor spacing distribution. The solid line is the result for orthogonal
symmetry ( = 1), the dashed line is the result for unitary symmetry ( = 2) and the dotted line is the result for symplectic
symmetry ( = 4). We notice the importance of the repulsion law s

for small spacings.

p
q
(s) = c
q
s
q
exp
_

c
q
q + 1
s
q+1
_
with c
q
=

q+1
(1/(q + 1))
q + 1
. (3.53)
In the case of orthogonal symmetry ( = 1), this distribution interpolates between the Poisson case and the Wigner
surmise (3.50). The free parameter q serves as a purely phenomenological measure for the degree of mixing between
Poisson and GOE statistics. Izrailev
66,67
proposed a dierent phenomenological interpolation formula,
p

eff
(s) = A

eff
s

eff
exp
_

e
16
s
2

_
B

eff


e
4
_
s
_
. (3.54)
The parameter
e
varies smoothly between zero, the Poisson repulsion factor, and four, which applies to symplectic
symmetry. The normalization constants A

eff
and B

eff
have to be determined from Eqs. (3.49). Caurier et al.
68
and Lenz and Haake
69
calculated yet another interpolation formula for the spacing distribution starting from two
dimensional matrix models.
Whenever the agreement of one of the distributions p(s) with data or numerical calculations is tested, it is helpful
to use the integral
F(s) =
_
s
0
p(s

)ds

(3.55)
which is referred to as the cumulative spacing distribution.
3. Longrange spectral observables
The nearest neighbor spacing distribution contains information about the spectrum which involves short scales (a
few mean level spacings). Longrange correlations are measured by quantities such as the level number variance
2
(L)
and the spectral rigidity
3
(L). The level number variance is given by

2
(L) =
2
(L,
s
)) (L,
s
))
2
. (3.56)
33
Here, (L,
s
) counts the number of levels in the interval [
s
,
s
+ L] on the unfolded scale. The angular bracket in
Eq. (3.56) denotes the average over starting points
s
. By construction, i.e. unfolding, one has (L,
s
)) = L. Thus,
in an interval of length L one expects on average L
_

2
(L) levels. The spectral rigidity, introduced by Dyson and
Mehta
70
, is closely related to
2
(L). In an interval of length L. it is dened as the least square deviation of the
staircase function () from the best t to a straight line,

3
(L) =
1
L
_
min
A,B
_
s+L
s
( () A B)
2
d
_
. (3.57)
The angular bracket is dened as in Eq. (3.56). The denition (3.57) is very natural since by construction, the smooth
part of the staircase function () is itself. The bottom part of Fig. 11 illustrates the meaning of the spectral rigidity

3
(L). Since the few stairs below 702 kHz come quite regularly, there is only a small rootmeansquare deviation
from the thin line describing the smooth behavior. This implies a small contribution to
3
(L). Some of the levels
above 702 kHz are nearly degenerate which makes the succession of the stairs rather irregular. Thus, the deviation
from the smooth part is larger, yielding a larger contribution to
3
(L). We notice that the staircase function in
Fig. 11 is shown on the original scale prior to unfolding. However, since the section in the bottom part is small, the
smooth part is almost a straight line and the statements just made apply qualitatively also to the unfolded case. For
xed
s
, Bohigas and Giannoni
71,47
have given a very useful procedure to obtain
3
(L) from a measured or calculated
spectrum. A proper error analysis for
2
(L) and
3
(L) is nontrivial because the results for dierent values of L are
strongly correlated. A consistent method of estimating errors was given by Shriner and Mitchell
72
.
In contrast to the spacing distribution, both the number variance and the spectral rigidity probe only twolevel
correlations. Indeed,
2
(L) can be reduced to the form
47

2
(L) = L 2
_
L
0
(L r)Y
2
(r)dr , (3.58)
provided the spectrum is translation invariant. As a consequence the number variance is related to the twolevel form
factor b
2
(t) dened in Eq. (3.41),
lim
L
1
L

2
(L) = 1 b
2
(0) . (3.59)
For the Poisson spectrum without correlations and for the harmonic oscillator one obtains
2
(L) = L and
2
(L) = 0,
respectively. The results for the Gaussian ensembles lie between these limiting cases. For large L and = 1, 2, 4 one
nds the following approximations
70,23
, valid to order 1/L,

2
1
(L) =
2

2
_
ln(2L) + + 1

2
8
_

2
2
(L) =
1

2
(ln(2L) + + 1)

2
4
(L) =
1
2
2
_
ln(4L) + + 1 +

2
8
_
. (3.60)
Here = 0.5772 . . . is Eulers constant. These approximations display the famous logarithmic behavior and the
increasing stiness of the spectrum with increasing . Exact expressions can be found, for example in Sec. V C of
Ref. 18, they are shown in Fig 13. There is a minor misprint in Eq. (5.13) of Ref. 18 for
2
4
(L), the argument of
2
2
on
the right hand side of this equation should read 2r instead of r. Moreover, we notice that there are visible oscillatory
modulations in
2
4
(L) which are not borne out in the approximation (3.60).
The spectral rigidity can similarly be expressed as an integral over the twopoint function, and the minimization in
Eq. (3.57) can be done analytically
70,73,23
. For a spectrum which is invariant under a continuous set of transformations,
one arrives at

3
(L) =
L
15

1
15L
4
_
L
0
(L r)
3
(2L
2
9Lr 3r
2
)Y
2
(r)dr (3.61)
which shows a close formal similarity to Eq. (3.58). Alternatively, one can write
73

3
(L) =
2
L
4
_
L
0
(L
3
2L
2
r +r
3
)
2
(r)dr . (3.62)
34
FIG. 13. The level number variances
2

(L) for the Gaussian ensembles. These are the exact results obtained from the
integral (3.58). The solid line is the GOE result ( = 1), the dashed line is the GUE result ( = 2) and the dotted line is the
GSE result ( = 4). We notice the oscillations in
2
4
(L) which are not contained in the logarithmic approximation of Eq. (3.60).
This equation can be viewed as an integral transform of the level number variance. Remarkably, the parabola is in
the kernel of this transform, i.e. inserting
2
(r) = r
2
yields
3
(L) = 0. For the limiting cases of the Poisson spectrum
and the harmonic oscillator, one has
3
(L) = L/15 and
3
(L) = 1/12, respectively. The results for the Gaussian
ensembles can be found by numerical integration. For large L and = 1, 2, 4, one has to order 1/L,

3,1
(L) =
1

2
_
ln(2L) +
5
4


2
8
_

3,2
(L) =
1
2
2
_
ln(2L) +
5
4
_

3,4
(L) =
1
4
2
_
ln(4L) +
5
4
+

2
8
_
. (3.63)
These approximations show once more the close relation between
3,
(L) and the level number variance
2

(L).
We do not address here observables which probe threelevel, fourlevel or even higherorder correlations. We refer
the reader to Refs. 70,47. In real systems, nongeneric and systemspecic properties exist which are not described
by Random Matrix Theory. These show up on large scales and are discussed in later sections.
4. Fourier transforms
Extracting the positions of the levels from a measured spectrum can be dicult for at least two reasons. Either the
resolution is so poor that the extracted sequence is incomplete, or the data set is so large that the task of nding all
levels is simply too tedious. In these cases it is desirable to introduce a statistical observable which is directly related
to the measured spectrum and in some way separates the statistics of the level positions from the uctuations due to
the line shapes of the levels. The correlationhole method introduced by Leviandier et al.
74
is a rst successful step
in this direction.
We consider a measured spectrum I() on the unfolded scale . The observable of interest is the decay function,
i.e. the square modulus of the Fourier transform of the spectrum
35
[c(t)[
2
=

+
_

d I() exp(2it)

2
. (3.64)
The Fourier coordinate t denes an unfolded time. Expression (3.64) can be rewritten as the Fourier transform
[c(t)[
2
=
+
_

A(r) exp(2irt)dr (3.65)

of the autocorrelation function
A(r) =
+
_

I(R r/2)I(R +r/2)dR . (3.66)

This autocorrelation function still contains nongeneric information about the system. To obtain the generic statistical
properties, A(r) or, equivalently, [c(t)[
2
have to be smoothed properly
75
. The resulting function A(r)) is generically
translation invariant and can be related to the twolevel cluster function Y
2
(r) of Random Matrix Theory. The decay
function [c(t)[
2
is correspondingly related to the twolevel form factor b
2
(t) and at short times reects longrange
spectral correlations, see Eq. (3.59).
To be denite we discuss the model spectrum
I() =
N

n=1
y
n
L(
n
) (3.67)
studied in Ref. 76 and investigated in a similar form in Refs. 74,7779. Here, the levels
n
, n = 1, . . . , N on the
unfolded scale have a common line shape L() and random intensities y
n
, N is assumed to be a very large number.
For nonnegative times the decay function is given by
74,7779,76
[c(t)[
2
= Ny
2
(t) +Ny
2
(1 b
2
(t)) [

L(t)[
2
, (3.68)
with = y
2
/y
2
the ratio of the rst two moments of the intensity distribution. The function

L(t) is the Fourier
transform of the line shape L(). For a Lorentzian line shape, [

L(t)[
2
reects the exponential decay of the resonances.
If L() = () and if all intensities y
n
are unity, the decay function [c(t)[
2
equals 1 b
2
(t) for nonzero t. For an
uncorrelated Poisson spectrum we have b
2
(t) = 0 and the decay function is constant. In a correlated spectrum derived
from one of the Gaussian ensembles the decay function approaches zero with vanishing t. This behavior is referred
to as the correlation hole. If the intensities are random numbers, the correlation hole is described by the function
1 b
2
(t) and does not approach zero. In particular, if the intensity distribution is the PorterThomas distribution
(see Sec. III E1), one has = 1/3. The function in Eq. (3.68) occurs since we assumed N to be very large. For a
nite number of levels, this contribution will acquire a width, see Refs. 77,79.
The correlationhole method separates the statistics of the level positions from that of the intensities and line
shapes. This is its main advantage. For realistic spectra, i.e. with inclusion of the line widths, the theory of the
correlation hole was worked out in Ref. 77,80. There, a scattering model was used. The application of the Fourier
transform to statistical spectra is summarized and a qualitative discussion given in Ref. 78. A short version of the
theory of the correlation hole is presented in Appendix A of Ref. 79. In Ref. 81, the correlation hole is related to its
classical analogue, the survival probability.
5. Superposition of independent spectra
We consider a system with a set of good quantum numbers such as spin and parity. In such a system, the Hamiltonian
H assumes blockdiagonal form, H = diag(H
1
, H
2
, . . . , H
M
). So far we have always dealt with the spectrum generated
by only one of the subblocks H
m
, m = 1, . . . , M. Now we consider the superposition of M such spectra. The total
level density R
1
(E) is simply the sum of the M subblock densities, R
1
(E) =

M
m=1
R
1m
(E). We assume that
these M densities have roughly the same energy dependence, and that the statistical measures for the M individual
spectra are known. How can we calculate the corresponding statistical measures for the total spectrum? For the
spacing distribution, the answer is already contained in the seminal article by Rosenzweig and Porter
82
of 1960, see
36
also Mehtas book
23
. The unfolded observables are expressed in units of the total mean level spacing D. Assuming
that energy interval of interest is centered around the origin E = 0, we have D = 1/R
1
(0). We also introduce the
fractional densities g
m
= R
1m
(0)/R
1
(0) with

M
m=1
g
m
= 1. Let p
m
(g
m
s), m = 1, . . . , M be the nearest neighbor
spacing distribution of the mth subspectrum. No assumption is made on the form of the distributions p
m
(g
m
s).
The spacing distribution p(s) of the superposition of the M spectra is found to be
p(s) =
d
2
ds
2
E(s) = E(s)
_
M

m=1
g
2
m
p
m
(g
m
s)
E
m
(g
m
s)
+
_
M

m=1
g
m
1 F
m
(g
m
s)
E
m
(g
m
s)
_
2

m=1
_
g
m
1 F
m
(g
m
s)
E
m
(g
m
s)
_
2
_
, (3.69)
where E(s) =

M
m=1
E
m
(g
m
s) and
F
m
(g
m
s) =
_
gms
0
p
m
(s

)ds

, E
m
(g
m
s) =
_

gms
(1 F
m
(s

))ds

. (3.70)
Whenever the longrange spectral observables of interest are pure twopoint measures, it is easy to relate those of the
full spectrum and those of the M individual spectra. Let Y
2m
(g
m
r) be the twolevel cluster function,
2
m
(g
m
L) the
level number variance, and
3m
(g
m
L) the spectral rigidity of the m-th subspectrum, respectively. These observables
are completely arbitrary. Just like the p
m
(g
m
s), they may coincide with but are not restricted to any of the specic
forms presented in previous sections. From the denition of the twopoint correlation function, one nds immediately
for the total twolevel cluster function
Y
2
(r) =
M

m=1
g
2
m
Y
2m
(g
m
r) . (3.71)
It then follows directly from Eqs. (3.58), (3.61) and (3.62) that

2
(L) =
M

m=1

2
m
(g
m
L) and
3
(L) =
M

m=1

3m
(g
m
L) . (3.72)
The rst property (3.72) is not surprising because
2
(L) is a variance. All results given here apply only to M strictly
noninteracting spectra. Crossover transitions due to the breaking of a quantum number are discussed later.
Very important for the analysis of data is the superposition of a large number M of spectra. For all individual
distributions p
m
(g
m
s) which play a role in realistic spectra the superposition p(s) tends rather quickly towards the
Poisson limit. The superposition of only six or so spectra which individually follow the Wigner surmises yields a
spacing distribution which is experimentally hard to distinguish from a Poisson distribution. This is illustrated in
Fig. 14 for spectra with equal fractional densities g
m
= 1/M, m = 1, . . . , M. In this case, formula (3.69) implies
p(0) = 11/M, reecting the increasing number of degeneracies. For the interval lengths L which are usually available
for the analysis of data, a similar statement applies to
2
(L) and
3
(L) as superpositions of the relevant individual
distributions
2
m
(g
m
L) and
3m
(g
m
L). There are of course exceptions such as, for example, the superposition of pure
harmonic oscillator spectra.
The property (3.71) of the twolevel cluster function can be used to estimate the number M of superimposed
independent spectra, i.e. of symmetries in a measured spectrum. This is important in situations where an assignment
of the set of quantum numbers to every individual level in the spectrum is impossible or too tedious. In a molecular
physics context, Leviandier et al.
74
argued that the decay function [c(t)[
2
introduced in Sec. III B4 is rescaled by
a factor 1/M so that the correlation hole becomes narrower by a factor M. In a more general discussion
77
, a
superposition of M independent spectra was considered. Each spectrum has a line shape L
m
() and a set of intensities
with rst and second moments y
m
and y
2
m
. One nds
[c(t)[
2
= f
0
(t) +f
1
M

m=1
g
m
y
2
m
(1
m
b
2
(t/g
m
)) [

L
m
(t)[
2
(3.73)
where f
0
and f
1
are systemspecic constants, and where
m
= y
m
2
/y
2
m
. If all M mean level densities R
1m
(E) are
roughly equal, one has g
m
1/M, and the correlation hole is indeed narrowed by a factor M.
37
0 1 2 3
s
0.0
0.5
1.0
p
(
s
)
FIG. 14. Nearest neighbor spacing distribution p(s) as given in Eq. (3.69) for the superposition of M spectra with equal
fractional densities gm = 1/M, m = 1, . . . , M. Each individual spectrum follows the Wigner surmise. One has p(0) = 1 1/M
at s = 0. The cases M = 2, 4, 8 are shown as solid lines. The Wigner surmise and the Poisson distribution are drawn as dashed
lines.
Finally, we mention that Berry and Robnik proposed a spacing distribution supposed to describe the transition
from Poisson behavior to the Wigner surmise. This distribution is obtained from formula (3.69) by taking M = 2
spectra, one satisfying the Poisson spacing distribution, the other one, the Wigner surmise. Strictly speaking this
formula is only applicable to two noninteracting spectra. It is nonetheless often used as a phenomenological guess
for mixed systems.
6. GSE test of GOE data
We consider a GOE spectrum E
1
, E
2
, E
3
, ... with E
n
E
n+1
. We divide the sequence E
1
, E
2
, E
3
, ... into
two sequences with odd and even indices, i.e. into E
1
, E
3
, E
5
, ... and E
2
, E
4
, E
6
, ... and consider each as a new
spectrum. Then, a theorem due to Mehta and Dyson
83
states that after proper unfolding, each of the two spectra
obeys GSE statistics. This theorem was recently used by Lombardi, Bohigas and Seligman
84
for a GSE test of GOE
data. In this test, the twolevel form factor b
2
(t) and the correlation hole [c(t)[
2
are particularly useful observables.
In contrast to the GOE case, the GSE form factor has a singularity at t = 1 caused by the oscillatory structure
of the twolevel cluster function Y
2
(r). Thus, one expects a characteristic peak at [t[ = 1 which gives information
on correlations on the scale of about two mean level spacings. As pointed out in Ref. 84, the twolevel observables
calculated for the two subsequences contain information on higher level correlations (k > 2) of the original spectrum.
7. Ergodicity
In Random Matrix Theory, observables are calculated by averaging over an ensemble of Hamiltonians. This average
is indicated by a bar. Experimentally, the observables are calculated as running averages over part of the spectrum of
a given system, containing N
I
levels in the interval [E, I + E]. This average is indicated by an angular bracket. For
any function F(E) of energy E, it is given by
F(E))
I
=
1
I
_
E+I
E
F(E

)dE

=
1
N
I
_
NI
0
F(E +Dr)dr = F)
NI
(3.74)
38
where D denotes the mean level spacing. A comparison of the results of both procedures is meaningful only if they
are equivalent, i.e. if
F(E) = F(E)) (3.75)
In this case, one speaks of ergodicity. Mathematical discussions of ergodicity can be rather involved. Here, we only
summarize a very instructive discussion by French, Mello and Pandey
85
. It is assumed that the spectrum is generic
so that after proper unfolding averaging will give translation invariant results. As a test of ergodicity, French et al.
consider the variance
varF)
NI
= [F F)[
2
. (3.76)
This quantity is the ensemble average of the squared dierence between running and ensemble average. It can be cast
into a form similar to Eq. (3.58) for the level number variance,
varF)
NI
=
1
I
2
_
+I
I
(I [[)C
F
()d =
1
N
2
I
_
+NI
NI
(N
I
[r[)C
F
(Dr)dr . (3.77)
The auto-covariance function C
F
() = (F

(E) F

)(F(E +) F) is, after proper unfolding, a translation invariant

twopoint function independent of E. With the help of a result of the theory of random functions
86
, Eq. (3.77) leads
to
lim
r
C
F
(Dr) = 0 = lim
NI
varF)
NI
= 0 . (3.78)
or equivalently
lim
r
C
F
(Dr) = 0 = lim
NI
F)
NI
= F . (3.79)
This is the desired ergodic behavior for almost all members of the ensemble, except for a set of measure zero. We
note that ergodicity is attained only for a large, preferably innite, number N
I
of levels. The result (3.78) can be
generalized fairly easily: The kpoint function is ergodic if in the limit of innite energy dierences. the ensemble
averaged 2kpoint function vanishes. We have made no specic assumptions concerning F(E). Thus, the argument
just presented also applies to correlations of scattering matrix elements and all other level correlators of Random
Matrix Theory.
C. Introduction to supersymmetry
The description of all technical aspects of supersymmetry would be far beyond the scope of this review. We
attempt to give the interested reader an idea of the main elements of the method. For further reading, we mention
Berezins book
87
where mathematical aspects of superanalysis are described, see also Ref. 88. Several papers and
review articles explain the use of the standard supersymmetry method in problems of chaos and disorder
44,46,89,90
,
and Efetov
45
has recently devoted a book to the subject. The term supersymmetry was rst introduced by Wess
and Zumino in relativistic eld theory
91
. In the present context, supersymmetry does involve a symmetry between
bosonic and fermionic integration variables but lacks the invariance properties of a relativistic eld theory. Perhaps
more importantly, the fermionic variables lack any physical meaning and are introduced merely as bookkeeping devices.
Wherever applicable, the supersymmetry method has been found to be more powerful than the replica trick: It yields
exact results where the replica trick only yields asymptotic expansions. The price one has to pay consists in using
analysis on supermanifolds. The replica trick, on the other hand, nds a much wider range of applications.
In Sec. III C1 we derive the supersymmetric representation of the generating function for level correlators. In
Sec. III C2, the saddlepoint approximation is used to derive the standard supersymmetric nonlinear model. An
alternative technique which avoids the saddlepoint approximation and is useful for purely spectral observables is
discussed in Sec. III F2.
39
1. Supersymmetric representation of the generating function
We develop the method for an observable given by the product of the retarded and advanced propagators tr(E
+
1

H)
1
and tr(E

2
H)
1
, with H a member of the GUE, and E

p
= E
p
i with positive innitesimal and p = 1, 2.
It is highly nontrivial to average a product of propagators over an ensemble of matrices H since H appears in the
denominators. In Sec. III A, we have shown that the MehtaDyson method accomplishes this goal very elegantly in
the case of spectral correlations in the pure cases. If, however, an additional matrix appears in the propagator,
rotation invariance in Hilbert space is broken, and almost all of the convenient features of the MehtaDyson method
disappear. Scattering systems require such an addition to the Hamiltonian to describe the coupling to the channels,
see Sec. III D. Prior to the advent of supersymmetry, one was restricted to perturbative or asymptotic expansions in
these situations, see Refs. 18,25.
To keep the presentation transparent, we do not discuss scattering systems and restrict ourselves to the spectral
twopoint correlation function, even though in this particular case, the supersymmetry method only reproduces the
results obtained earlier by Dyson and Mehta. We emphasize again that the supersymmetry method develops its
full power in those cases where the MehtaDyson method does not apply. Later, we indicate the modications for
other cases (more than two propagators, dierent symmetry of H, propagators in higher dimension). Our choice of
observable is dictated by the fact that the occurrence of a pair of complex conjugate propagators is typical of many
problems. Our notation is that of Ref. 46. We use the form
2
/N for the variance of the Gaussian distribution of H,
see Sec. III A2.
To perform the ensemble average exactly, i.e. nonperturbatively, one expresses the propagator as the derivative of
a generating function F
p
(J
p
) with respect to a source variable J
p
such that
tr
1
E

p
H
=

J
p
F
p
(J
p
)

Jp=0
. (3.80)
It is easily seen that the choice
F
p
(J
p
) =
det(E

p
H +J
p
)
det(E

p
H J
p
)
(3.81)
represents such a generating function. We observe that the logarithms of the numerator and the denominator separately
generate the propagator, too. In contrast to the latter, however, the function F
p
(J
p
) has the very useful property
of being normalized to unity, i.e. to a constant independent of H: We have F
p
(0) = 1. Still, an exact average of
F
p
(J
p
) over the ensemble seems out of question. This is the point where supersymmetry enters. The inverse of an
N N determinant can be expressed as the Gaussian integral over an N component vector S
p
of complex commuting
variables, while the determinant itself is given as the Gaussian integral over an N component vector
p
of complex
anticommuting variables. Thus we have
F
p
(J
p
) =
_
d[S
p
] exp
_
iS

p
(E

p
H J
p
)S
p
_
_
d[
p
] exp
_
i

p
(E

p
H +J
p
)
p
_
. (3.82)
Calculating the average over the Gaussian ensemble is now straightforward since H appears in the exponent.
With these tools, we can derive the generating function of the twopoint function. It is convenient to write
E = (E
1
+E
2
)/2 and = E
1
E
2
. To be consistent with most of the literature, we use the symbol throughout this
section and the symbol in its stead in Sec. VI. We already know that the unfolded twopoint function will only
depend on /D where D is the mean level spacing. We have to take into account that the imaginary increment i lies
on dierent sides of the real axis for p = 1 and p = 2. To ensure convergence of the integrals one proceeds as follows.
We dene a metric L = diag(+1, +1, 1, 1), reecting the location of i, and the direct product L = L1
N
. Here,
1
N
denotes the N N unit matrix. Moreover we set J = diag(J
1
, +J
1
, J
2
, +J
2
) and introduce the supervector
= (S
1
,
1
, S
2
,
2
)
T
. The proper generating function is then given by
F
1
(J
1
)F
2
(J
2
) =
_
d[] exp
_
i

L
1/2
_
1
4
(E H)
+
_

2
+i
_
L 1
N
+J 1
N
_
L
1/2

_
. (3.83)
40
Some authors prefer a dierent form for expression (3.83): The symmetric arrangement of the two factors L
1/2
is
dropped in favor of a single factor L appearing right behind

.
For an observable containing k > 2 propagators (usually an equal number of advanced and of retarded propagators),
the form of Eq. (3.83) would be the same, but the dimension of the vectors and of the supermatrices would be 2kN
rather than 4N. The dimensions of the supermatrices and supervectors and their internal symmetries would change
for real symmetric or quaternion (rather than Hermitean) matrices H. The source term J 1
N
must be altered,
if matrix elements (rather than traces) of the propagators are considered. For Hamiltonians in d > 0 dimensions,
the fourcomponent super vectors

n
and
n
with n = 1, . . . , N are replaced by spacedependent vectors (r)

and
(r), respectively, and the bilinear forms in the matrix space of H are replaced by integrations over ddimensional
space. The generating function becomes a generating functional, and the theory turns into a eld theory.
Only the Hdependent term in expression (3.83) is aected by averaging over the GUE. It is easy to calculate
_
d[H]P
N2
(H) exp
_
i

(L H)
_
= exp
_
1
2N
trgA
2
_
(3.84)
where
A = iL
1/2
N

n=1

n
L
1/2
(3.85)
is a four by four supermatrix. The symbol trg denotes the supertrace. The term trgA
2
in the exponent is of fourth
order in the integration variables contained in
n
, and a direct integration over these variables is, therefore, impossible.
This diculty is overcome with the help of the HubbardStratonovich transformation. The fourthorder term in the
exponent is reduced to second order at the expense of introducing a set of auxiliary integration variables,
exp
_
1
2N
trgA
2
_
=
_
d[] exp
_

N
2
trg(
2
) trg(A)
_
. (3.86)
The matrix has essentially the same symmetries as the matrix iA. The dierential d[] denotes the product of
the dierentials of the independent matrix elements. The integration over the
n
can now be done. The procedure
sketched in Eqs. (3.84) and (3.86) can also be viewed as a double Fourier transform, the rst in ordinary, the second
in superspace
92
.
Collecting everything, we nd for the ensemble average F
1
(J
1
)F
2
(J
2
) of the generating function (3.83)
F
1
(J
1
)F
2
(J
2
) =
_
d[] exp L() (3.87)
where the Lagrangean L has the form
L() =
N
2
trg(
2
) N trg ln
_
E1
4
+J +

2
L
_
. (3.88)
In the case of a disordered system which might, for instance, be described by a singleparticle Hamiltonian H = T +V
where T is the kinetic energy operator, and V is an impurity potential of the form (2.14), one obtains similarly a
generating functional where is a function of r, and where the operator T appears as argument under the logarithm.
Equations (3.87,3.88) embody the rst important result of the supersymmetry technique: The ensemble average
of F
1
(J
1
)F
2
(J
2
) has been calculated, and the resulting integral expressed in terms of a small number of integration
variables, determined by the number of independent elements in the matrix . The 4N original complex integration
variables have disappeared.
The matrix possesses the same dimension and the same symmetries as the matrix iA. The matrix iA, in turn,
comprises all the unitary invariants which can be constructed from the elements of each, equivalent, vector
n
. Only
such invariants survive the ensemble average. Hence, both iA and directly mirror the underlying unitary symmetry
in superspace. Moreover, the dimension of is as small as is consistent with both the form of the observable, and the
symmetry of the problem.
These statements apply equally to observables involving kpoint functions, see for example Ref. 92, and to problems
with orthogonal
46
or symplectic symmetry. For Hamiltonians in d 1 dimensions, the matrix attains a dependence
on a spatial variable r. With proper modications, the statements just made carry over to this case, too, and the
resulting expression for the generating functional constitutes a genuine eld theory
44
. In a conceptually important
work, Zirnbauer
93
recently extended the supersymmetry method to to the circular ensembles, i.e. to systems with
unitary disorder.
41
2. Saddlepoint approximation and nonlinear model
The second important step of the supersymmetry technique consists in an approximate evaluation of the remaining
integrations in the ensembleaveraged supersymmetric generating function in Eqs. (3.87,3.88), or at least in the
reduction of the problem to a nonlinear model. This is accomplished by using the saddlepoint approximation,
suggested by the occurrence of the factor N multiplying all terms in the exponent. We recall that in RMT, we always
take the limit N . In some cases, it is advantageous to avoid the saddlepoint approximation, and to use an
alternative technique, see Sec. III F.
The generating function (3.87,3.88) possesses an important symmetry: Except for its dependence on J and , it
remains invariant under the group G of those global transformations T with T
1
T which do not change the
symmetry of . The form of the supermatrices T which fulll this requirement depends, of course, on the dimension
of the matrix and on the symmetry (orthogonal, unitary or symplectic) of the underlying Hamiltonian ensemble.
The saddle point is found after neglecting in the expression (3.88) for L both the term and the source term J.
This is justied since scales with 1/N. Indeed, the uctuations depend on the unfolded energy dierence /D. Since
D vanishes like 1/N, the same must be true for . The source term J may be viewed as being innitesimally small
since the derivatives with respect to J are taken at J = 0. The saddlepoint equation is
=
1
4
E
. (3.89)
This matrix equation is solved in two steps. First, a diagonal matrix Q
0
is found which obeys Eq (3.89). For [E[ 2,
i.e. if E lies within Wigners semicircle, the elements Q

of Q
0
have the values Q

= E/(2) i
_
1 (E/(2))
2
.
For a twopoint function, i.e. a product of the trace of an advanced and of a retarded propagator, it is necessary to
choose Q
0
= diag(Q
+
, Q
+
, Q

, Q

).
This choice breaks the symmetry under the group G dened above. This is most easily seen by writing Q
0
in the
form
Q
0
=
E
2
1
4
+
(E)
N
L (3.90)
with dened in Eq. (2.7). The matrix L does not commute with all elements T of G: The symmetry is broken by
the mean level density . We will see shortly that as a consequence of this broken symmetry, there occurs a Goldstone
or zero mode.
All matrices Q = T
1
Q
0
T obtained from Q
0
by application of a transformation T G are also solutions of the
saddlepoint equation, and the totality of all these matrices denes the saddlepoint manifold. Let G
R
with elements
R be the maximum subgroup of G which obeys [R, L] = 0 for all R G
R
. This subgroup leaves Q
0
invariant. Let
G/G
R
be the associated coset space with elements T
0
. Every T G can be written in the form T = RT
0
with R G
R
and T
0
G/G
R
. The saddlepoint manifold is then given by Q = T
1
0
Q
0
T
0
. This manifold has dimension d > 0
precisely because the symmetry under G is broken by Q
0
, and it therefore constitutes the Goldstone mode caused by
the broken symmetry.
In a seminal paper, Sch afer and Wegner
39
investigated the structure of the saddlepoint manifold for bosonic
variables. In Refs. 44,46, their work was extended to the supersymmetry formalism. It turns out that the saddle
point manifold possesses mixed (hyperbolic and compact) symmetry. This property of the saddlepoint manifold is
not reproduced in the replica trick, and is the cause of its failure to produce exact results
36
.
The integration over the massive modes (i.e. in directions orthogonal to the saddlepoint manifold) can be carried
out. For Gaussian RMT, these modes have a mass proportional to N, and yield corrections to the saddlepoint integral
which vanish for N . The remaining integrations extend over the coset space G/G
R
and yield a zerodimensional
nonlinear model (because the saddlepoint solutions Q obey the nonlinear Eq. (3.89)). The generating function
attains the form
F
1
(J
1
)F
2
(J
2
) =
_
d(t)(...) exp
_
i

4D
trg(QL)
_
. (3.91)
The dots indicate the Jdependent terms. Notice that the mean level spacing is given by D = /N at E = 0. The
integration measure d(t) is dened in terms of a suitable parameterization of the elements T
0
of the coset space
G/G
R
.
Mutatis mutandis, these arguments carry over to the case of disordered solids in d 1 dimensions. For the potential
model of Eq. (2.14), the large parameter which formally corresponds to N is k
F
l where l = v
F
is the elastic mean
free path, is the elastic mean free time, and k
F
(v
F
) is the Fermi wave number (the Fermi velocity, respectively).
The need to use this parameter shows that the supersymmetry technique is conned to problems with weak disorder.
42
Minimization of the Lagrangean in d 1 leads to a diagonal saddlepoint solution Q
0
which has the same form as
in RMT. The saddlepoint manifold now comprises matrices Q(r) = T
1
0
(r)Q
0
T
0
(r) with a very slow rdependence
(long wavelength limit). The resulting generating functional F(Q) has the form
F(Q) =
_
d(t(r))(...) exp
_

8
_
d
d
r
_
h T trg (Q(r))
2
+2i trg(QL)
__
. (3.92)
Here, T = v
2
F
/d is the diusion constant. Except for the additional space dependence, all other symbols have the
same meaning as in Eq. (3.91).
It is instructive to compare the terms in the exponents of Eqs. (3.91,3.92). If we disregard the additional space
dependence in Eq. (3.92), the symmetrybreaking term proportional to is identical in both cases. Indeed, the
space integration yields a factor V , so that V =
1
, and we surely must identify with the Gaussian RMT
mean level spacing D = /N. The gradient term in Eq. (3.92) (obviously missing in Eq. (3.91)) has roughly the
magnitude hT/L
2
= E
C
. For E
C
< , this term is more important than the symmetrybreaking term, and conversely,
see Sec. VI.
We distinguish a diusive regime, dened by the condition g 1, or L with the localization length, and
a localized regime, where L . Since g = E
C
/ = /L in the diusive regime, see Eq. (2.17), the conditions
L and E
C
are equivalent. In the diusive regime, the symmetrybreaking term is the dominant zero mode.
This implies that all correlation functions of the disordered solid are approximately of Gaussian RMT type in this
regime! This statement holds for energy dierences [[ < E
C
and establishes the close link between Gaussian RMT
and localization theory, see Sec. VI. In the localized regime, on the other hand, the gradient term is the dominant zero
mode. It may be surprising that in this regime, Eq. (3.92) retains its validity. The reason is that the characteristic
length scale L
0
over which Q(r) varies must obey the inequality l L
0
. This is why a nonlinear model does
not apply for d = 1 where l = .
In concluding this introduction to supersymmetry, we notice that the supersymmetric representation of the gener-
ating function for the RMT correlators provides an exact reformulation of the RMT problem. Therefore, all results
derived from this function are fully equivalent to the results of classical RMT. Starting from a whitenoise potential,
on the other hand, one obtains a supersymmetric nonlinear model. In the zerodimensional limit, this model
is fully equivalent to th limit N of the supersymmetric representation of RMT. However, the supersymmetry
method continues yielding new results which, in many cases, could not be obtained with other techniques.
D. Scattering systems
Scattering systems play a very important role in the application of Random Matrix Theory, particularly in the
context of transport phenomena in mesoscopic systems. The theory of these systems requires a generalization of the
concepts introduced so far for bounded systems and the associated spectral uctuations. We keep our presentation
brief since various well written papers and reviews are available. We mention the early reviews by Eckhardt
94,95
and Smilansky
96,97
which contain discussions of chaotic scattering, especially from a semiclassical point of view.
Likewise, Doron, Smilansky and Frenkel
98
mainly focus on semiclassical aspects, while Lewenkopf and Weidenm uller
99
concentrate on aspects of Random Matrix Theory and on the comparison with semiclassical results. Very recently, a
review of chaotic scattering was presented by Jung and Seligman
100
, and one on scattering and transport in mesoscopic
and disordered systems by Beenakker
101
. In Sec. III D1, we summarize general aspects of scattering theory, before
we establish the link to Random Matrix Theories in Sec. III D2.
1. General aspects
In the context of RMT, scattering processes are important in both, cavities and/or mesoscopic wires connected to
the outside world by leads or antennae, and nonrelativistic manybody systems as encountered in nuclear, atomic
and molecular physics. In all these cases, the participants in the scattering process move freely at asymptotically large
distances but encounter strong interactions in the interaction region. We idealize a scattering system as consisting of a
compact interaction region, i.e. a region of nite volume, and of channels through which this region is accessible and in
which the propagation is free. A general scattering theory based on this idealization was constructed by Wigner and
Eisenbud
102
. These authors made the physically well justied further assumption that only twobody fragmentation
43
is allowed at low energies. They also postulated shortrange interactions. However, the formalism can be extended
to the Coulomb interaction, too.
To illuminate the physical content of this approach, we rst consider a boundstate problem dened by restricting
the motion of all particles to the compact interaction region. Then, the spectrum is discrete. As we allow for the
coupling to the channels, many of the bound eigenstates turn into resonances. These resonances may dominate the
scattering process. Following Ref. 27, we present here a variant of the general theory which focuses attention from
the outset onto such resonances. We consider N orthonormal functions

, = 1, . . . , N localized on the interaction

region, where eventually the limit N has to be taken. These functions represent a basis for the quasibound
eigenstates. The channels are represented by the channel wave functions
c
(E), c = 1, . . . , where E is the total
energy. These states obey the orthonormality condition
a
(E)[
b
(E

)) =
ab
(E E

). They do not vanish in the

interaction region but are orthogonal to the boundstate wave functions,
a
(E)[

) = 0 for all E, a, . In this basis,

the total Hamiltonian has the form
H =
N

,=1
[

)H

[
+
N

=0

c=1
_

c
dE ([
c
(E))W
c

[ +[

)W
c

c
(E)[)
+

c=1
_

c
dE[
c
(E))E
c
(E)[ . (3.93)
Here
c
is the threshold energy in channel c (for E <
c
only evanescent waves exist, and for E >
c
the channel is
said to be open). The rst term in Eq. (3.93) involves the N N matrix H which describes the mutual coupling
of the bound states. The second term contains the N coupling matrix W between channels and bound states.
This term causes the bound states to acquire widths, and to become resonances. The third term is assumed to be
diagonal in the channels. This assumption can be lifted but provides a useful simplication and is realistic in later
applications. In the framework of Eq. (3.93), transitions between dierent channels are possible only via intermediate
population of the bound states. We assume that elastic background phase shifts in the channels can be neglected so
that the functions
a
(E) are essentially plane waves.
The physical assumptions made in this model become even more transparent as one works out
27
the scattering
matrix S
ab
(E),
S
ab
(E) =
ab
i2W

a
D
1
(E)W
b
(3.94)
where the N component vector W
a
is the ath column of the matrix W. The propagator D
1
(E) is the inverse of
D(E) = E1
N
H +i

c open
W
c
W

c
. (3.95)
This notation should not be confused with that for the mean level spacing D between the eigenvalues of H. All
threshold eects, including the dependence of W on E have been neglected. The scattering matrix element S
ab
(E)
has N poles corresponding to N resonances. For N = 1, the scattering matrix has BreitWigner form, with W
c1
denoting the partial width amplitude. For N > 1, Eqs. (3.94) and (3.95) constitute Nlevel unitary generalizations
of the BreitWigner formula. In the limit

c=1
[W
c
[
2
D for all , the model describes N isolated resonances
= 1, . . . , N, with widths very small compared to the spacing between resonances. Here, D denotes the mean level
spacing. As the ratios

c=1
[W
c
[
2
/D increase, the resonances begin to overlap. Eventually, one arrives at the Ericson
regime of strongly overlapping resonances where the uctuations of the cross section are no longer related to individual
resonances. Increasing the ratios

c=1
[W
c
[
2
/D even further, one returns to the regime of isolated resonances.
2. Scattering and Random Matrix Theory
Apart from the general constraints mentioned above, no specic assumptions on the boundstate Hamiltonian H
have been made in the derivation of the formal result (3.94). We now connect this model to RMT by assuming that
the Hamiltonian H in the interaction region can be represented by a random matrix drawn from one of the Gaussian
ensembles. In particular, for a timereversal invariant system, H is a member of the GOE. With this assumption, the
uctuation properties of S can be predicted, given the ensemble average of S which serves as an input parameter.
44
Anticipating the discussion in later sections, we argue that this choice for H is justied in three typical physical
situations: Scattering at complex manybody systems, chaotic scattering in fewdegreesoffreedom systems, and
scattering by a disordered quantum dot. Indeed, the spectral uctuation properties of a wide class of complex many
body systems such as nuclei, atoms and molecules are known to be well described by the spectral observables discussed
in Sec. III B. Therefore, it is very natural to choose H from the GOE. For chaotic systems with few degrees of freedom,
such as billiards, the justication is based on the Bohigas conjecture. For a disordered quantum dot, the random
Hamiltonian directly simulates the eect of the random potential.
The autocorrelation function of the scattering matrix S is dened as an ensemble average,
C
abcd
() = S
ab
(E)S

cd
(E + ) S
ab
(E) S

cd
(E) . (3.96)
In a generic situation, this correlator is, after proper unfolding, independent of the energy E. The mathematical
techniques developed by Mehta and Dyson cannot be used to work out the correlator (3.96). Extending the super-
symmetric method which Efetov
44
had developed in the framework of disordered systems, Verbaarschot, Weidenm uller
and Zirnbauer
46
derived the exact result
C
abcd
() =
1
8
_

0
d
1
_

0
d
2
_
1
0
d
(1 )[
1

2
[
_
(1 +
1
)
1
(1 +
2
)
2
( +
1
)
2
( +
2
)
2
exp
_
i

D
(
1
+
2
+ 2)
_

e=1
(1 T
e
)
_
(1 +T
e

1
)(1 +T
e

2
)
_

ab

cd
S
aa
S

cc
T
a
T
c
_

1
1 +T
a

1
+

2
1 +T
a

2
+
2
1 T
a

_
_

1
1 +T
c

1
+

2
1 +T
c

2
+
2
1 T
c

_
+(
ac

bd
+
ad

bc
)T
a
T
b
_

1
(1 +
1
)
(1 +T
a

1
)(1 +T
b

1
)
+

2
(1 +
2
)
(1 +T
a

2
)(1 +T
b

2
)
+
(1 )
(1 +T
a
)(1 +T
b
)
__
. (3.97)
The coecients T
c
= 1 [S
cc
(E)[
2
are called transmission coecients and dene the strength of the coupling between
resonances and channels. Each T
c
depends nonlinearly on the quantity

W
2
c
. If all coecients T
c
are unity,
the uctuations attain maximal size. An expansion of C
abcd
() to rst order in the inverse of

c=1
T
c
yields the
HauserFeshbach formula
103
, see Eq. (4.1) of Sec. IVA1.
In the stochastic approach to scattering just described, the boundstate Hamiltonian H occurring in expres-
sions (3.94) and (3.95) is replaced by an ensemble of random matrices. Alternatively, it is possible to consider
the scattering matrix S itself as a stochastic object, without the intermediate step of choosing a random Hamil-
tonian. In this case, one directly constructs an ensemble of scattering matrices S. This is the approach by Dyson
59
who dened the three circular ensembles: the Circular Orthogonal Ensemble (COE), the Circular Unitary Ensemble
(CUE), and the Circular Symplectic Ensemble (CSE), see Sec. II A. These three ensembles correspond, respectively,
to the GOE, the GUE, and the GSE. The invariance properties used by Dyson imply that the ensemble averages of
the S matrices vanish for all three ensembles, S = 0. This is equivalent to saying that the transmission coecients
are unity, or that the coupling to the channels is maximal. In that sense, the three ensembles of Dyson are subsets
of the more general ensembles introduced above via the random Hamiltonian approach. We do not go into further
details here of the random scattering matrix approach. We return to this point in Sec. VI.
Using the random Hamiltonian approach and the supersymmetry method, Fyodorov and Sommers
104106
recently
worked out several other statistical measures. They studied the case of broken timereversal invariance (GUE Hamil-
tonian). In particular, they established a link to the random scattering matrix approach by showing that in both
approaches the paircorrelation functions of phase shifts at xed energy coincide. Moreover, they also calculated
analytically the distribution of the poles of the scattering matrix. Because of the last term in Eq. (3.95) which repre-
sents the coupling to the channels, these poles occur in the lower half of the energy plane. Their results apply to any
breaking of Hermitecity while previous studies are only valid the case of strongly broken Hermitecity. We mention in
passing that there is a relation to Ginibres
107
ensemble of matrices with arbitrary complex eigenvalues. A discussion
can be found in Haakes book
57
.
45
E. Wave functions and widths
For the understanding of the connection between classical and quantum chaos, the study of the stochastic properties
of wave functions has become very important. For an investigation of wave packet dynamics, of scars and of the
relationship with periodic orbits, systems with few degrees of freedom are particularly well suited. A review can
be found in chapter 15 of Gutzwillers book
108
. Here, we focus on the statistical properties of wave functions in
chaotic systems and on the way these properties are modeled in Random Matrix Theory (Sec. III E1), and on recent
developments in mesoscopic physics, see Sec. III E2. The limits of the RMT description have been explored by
various authors. As one example we mention a conceptually important contribution due to Tomsovic
109
. He showed
the existence of parametric correlations (see Sec. III H) between states and level velocities in a particular system.
Random Matrix Theory cannot describe these correlations.
1. Results derived in the framework of classical RMT
To characterize the distribution of the eigenfunctions of an ensemble of random matrices, one considers a xed basis
and introduces the probability density P
N
(a) of nding for the components a value between a and a + da, with N
the dimension of the matrices. Rotation invariance of the ensemble implies that P
N
(a) is determined by the Haar
measure d(U) of the group of diagonalizing matrices, H = U
1
XU. For the orthogonal case, a quick and elementary
derivation
9
of P
N
(a) proceeds as follows. We consider one component of one of the N eigenvectors of H. In polar
coordinates, it can be represented by cos . The associated part of the measure is sin
N2
. For 1 < a < +1, we
have
P
N
(a) = C
N
_

0
(a cos ) sin
N2
d . (3.98)
The constant C
N
is determined by the normalization of P
N
(a) to unity. A straightforward calculation yields
P
N
(a) =
(N/2)

((N 1)/2)
_
1 a
2
_
(N3)/2
(3.99)
which is symmetric in a. This expression is exact for every integer N. For the second moment of this distribution,
one nds a
2
= 1/N. In the limit of large N, the distribution approaches the Gaussian
P
N
(a) =
_
N
2
exp
_

N
2
a
2
_
(3.100)
with variance 1/N.
In many cases, the distribution of the wave function components is not accessible to a direct measurement. However,
for isolated resonances the reduced partial width amplitudes
nc
for level n and scattering channel c can be determined
from a scattering experiment. Let J
mc
be the overlap integral of the channel wave function in channel c and the mth
wave function of the basis set. Then,

nc
=
N

m=1
U
nm
J
mc
(3.101)
with U as dened above. Consider L levels n = 1, . . . , L where L < N. As observed by Ullah
110
, the joint probability
distribution P(
1c
, . . . ,
Lc
) for the L partial width amplitudes can be written as
P(
1c
, . . . ,
Lc
) =
_
L

n=1

nc

m=1
U
nm
J
mc
_
d(U) . (3.102)
With the help of a proper orthogonal transformation
110
of the eigenvectors, this expression can be reduced to a form
which generalizes Eq. (3.98). One nds
P(
1c
, . . . ,
Lc
) =
(N/2)
_
N
2
c
_
L/2
((N L)/2)
_
1
1
N
2
c
L

n=1

2
nc
_
(NL2)/2
(3.103)
46
0 1 2 3
y
c
0.0
0.5
1.0
1.5
p
(
y
c
)
FIG. 15. The distributions p

(yc) of the normalized partial widths yc = c/c for the Gaussian ensembles. The solid line is
the GOE result ( = 1), the dashed line is the GUE result ( = 2) and the dotted line is the GSE result ( = 4). A singularity
at yc = 0 arises in the orthogonal case ( = 1) only.
where N
2
c
= N
2
nc
=

N
n=1
J
2
nc
. Again, this is exact for any N. The functional forms of Eqs. (3.99) and (3.103)
agree for L = 1 as expected. In the sequel, we consider only the case L = 1 and drop the level index.
Typically, the quantity measured in an experiment is the partial width
c
for the decay of a level into a channel
c, rather than the partial width amplitude
c
. Apart from proportionality constants such as penetration factors etc.,
we have
c
=
2
c
. For large N, Eq. (3.103) then yields the famous PorterThomas law
P(
c
)d
c
=
1

exp(
c
/2
c
)
_

c
/2
c
d
c
2
c
(3.104)
where
c
=
2
c
. These considerations can easily be generalized to unitary and symplectic symmetry. The value of
the symmetry parameter with = 1, 2, 4 is equal to the dimension of the number space on which the random
matrices are constructed. The reduced width amplitudes
c
accordingly possess components
cj
, j = 1, . . . , . If all
amplitudes
cj
are distributed with the same Gaussian probability distribution, the absolute square
c
=

j=1

2
cj
of
c
is distributed according to
P

(
c
)d
c
=
1
(/2)
_

c
2
c
_
/21
exp
_

c
2
c
_
d
c
2
c
. (3.105)
This implies that for the three symmetry classes, the distributions P

(
c
) are special cases of a
2

distribution for
= degrees of freedom. We note that only the orthogonal case has a singularity for vanishing width
c
. One often
introduces a normalized partial width y
c
=
c
/
c
and its distribution p

(y
c
) dened by p

(y
c
)dy
c
= P

(
c
)d
c
. This
function is shown in Fig. 15 for the three Gaussian ensembles.
In a system with open channels, there may exist correlations between the partial width amplitudes
c
, c = 1, . . . , .
Typically, such correlations have a dynamical origin. An example is provided by direct reactions in nuclear physics.
The correlation matrix of the partial width amplitudes is dened as
M
cc
=

c
. (3.106)
In the limit of large level number N and under the condition that N, the distribution of the partial width
amplitudes was found to be
111,112
47
P() =
1
det
/2
M
exp
_

M
1

_
. (3.107)
Here, = (
1
, . . . ,

) is a component vector containing the partial width amplitudes. In Ref. 110 it is also shown
how to derive multilevel and multichannel distributions for nite level number N.
Having investigated the distribution of the partial widths
c
, we now turn to the distribution function P() of the
total width . For isolated resonances, we have =

c=1

c
where is the number of open channels. Using and
extending the methods described in Porters review
9
and Ullahs book
113
, Alhassid and Lewenkopf
114
derived in a
direct calculation a very general result for the width distribution. Let M be the correlation matrix (3.106) of the
partial width amplitudes. For = 1, 2, 4, the distribution of the total width is given by
P() =
1
2
_
+

exp(it)
det
/2
(1

i2tM/)
dt (3.108)
where 1

is the unit matrix. We note that P() depends only on the eigenvalues w
2
c
of M. Since M is Hermitean
and positive denite, the w
2
c
s are all positive. In the unitary case ( = 2), the integral (3.108) can be done and yields
P() =
1

c=1
w
c

c=1
exp(/w
2
c
)

c =c

_
w
2
c
w
2
c

_ . (3.109)
For two equivalent channels, i.e. = 2 and M
11
= M
22
= /2 one nds
P() =
2
[f[
exp
_

2
(1 [f[
2
)
_
sinh
_

2[f[
(1 [f[
2
)
_
. (3.110)
Here f = M
12
/

M
11
M
22
measures the correlation between the two channels. This is in agreement with Ref. 115
where this result was derived using the supersymmetry method, see Sec. III E2. For = 1 one nds
114
P() =
1
_
1 f
2
exp
_

(1 f
2
)
_
I
0
_

f
(1 f
2
)
_
. (3.111)
where I
0
is the modied Bessel function of zeroth order. The methods described in Refs. 9,110 and used above allow
for a very ecient computation of the distribution of wave functions and widths.
2. Results derived in the framework of the supersymmetry method
The results summarized in the previous section were mainly prompted by research in scattering theory, and were
obtained in the framework of random matrix theory. Recently, condensed matter physicists became also interested
in distributions of wave functions and widths, particularly in the context of transport properties of quantum dots.
Prigodin et al
116
studied the statistics of conductance uctuations in a quantum dot. Inter alias, this led to a re
derivation of the PorterThomas distribution within the supersymmetry formalism, cf. also the discussion by Efetov
and Prigodin
117
.
In many situations there is interest in probability measures which go beyond the distribution of wavefunction
components. This applies, in particular, to the spatial correlations of wave functions for chaotic systems
118120
.
Berry
118,108
assumed that the eigenfunction
E
(r) at energy E in a timereversal invariant, ergodic system is an
innite superposition of random plane waves. For the spaceaveraged correlation function of this eigenfunction in a
ddimensional system, taken at two dierent points in space, he found

E
(r
1
)

E
(r
2
))
space
[
E
((r
1
+r
2
)/2)[
2
)
space
= 2
d/21

_
d
2
_
J
d/21
(kr)
(kr)
d/21
. (3.112)
Here

k is the wave vector, and k the inverse de Broglie wave length, r = [r
1
r
2
[ is the distance between the two
points.
Closely related questions arise in disordered mesoscopic systems. For a singleparticle model with random impurity
potential, cf. Sec. VI,
H =
p
2
2m
+V (r), (3.113)
48
Prigodin et al.
121
and Prigodin
120
recently calculated wave function correlation functions with the help of the super-
symmetry method. In a system of volume V , they obtained
[
E
(r
1
)
E
(r
2
)[
2
V
2
= 1 +c

[f(r)[
2
, (3.114)
with = 1, 2, 4 and c
1
= c
4
= 1, and c
2
= 2. Under the assumption of ergodicity, the function f(r) is closely related
to the averaged spatial correlator (3.112),
f(r) =
E
(r
1
)

E
(r
2
)V . (3.115)
The existence of spatial correlations of chaotic wave functions demonstrated in this work is essentially caused by the
nite (nonzero) wave length.
Equations (3.114) and (3.115) connect the fourth moment of the wave function with the square of the second
moment, suggesting a Gaussianlike probability density. To work out this probability density in its most general
form, let us consider M local densities of a given eigenfunction, a
j
= V [
E
( r
j
)[
2
, j = 1, . . . , M, and the spatial
correlator
120
,
P(a
1
, . . . , a
M
, r
1
, . . . , r
M
) =
M

j=1
(a
j
V [
E
(r
j
)[
2
) . (3.116)
The partial widths discussed in the previous section can be viewed as special cases of the quantities a
j
. Therefore,
the correlation function (3.116) has to reproduce the results of that section for M = 1.
The distribution function (3.116) is constructed from its moments. These moments can be expressed in terms of
retarded (and advanced) Green functions
122,123
dened as
G
R
(E, r
1
, r
2
) =
N

n=1

En
(r
1
)

En
(r
2
)
E E
n
+i
(3.117)
where is innitesimal. We consider the case M = 2. Let R
1
(E) denote the mean level density. The moments can
be written as
[
E
(r
1
)[
2n
[
E
(r
2
)[
2m
=
i
nm
(n 1)!(m1)!
2R
1
(E)
n+m
(n +m2)!
lim
0

n+m1
(G
R
(E +i/2, r
1
, r
1
))
n
(G
A
(E +i/2, r
2
, r
2
))
m
. (3.118)
For vanishing , the only contributions come from terms in which the energy E coincides with one of the eigenenergies
E
n
.
In the unitary case = 2 one obtains
120
the distribution function
P(a
1
, a
2
, r) =
1
1 f
2
(r)
exp
_

a
1
+a
2
1 f
2
(r)
_
I
0
_
2f(r)

a
1
a
2
1 f
2
(r)
_
(3.119)
which contains the averaged spatial correlator f(r), corresponding to the quantity f in Eq. (3.110). The esti-
mate (3.112) shows that f(r) takes values between 0 and 1. The function f(r) can be calculated in special situations.
In a quantum dot
120
, it is given by the Friedel function. In any realistic physical system, there will be a critical
distance r
c
beyond which the spatial correlator vanishes, f(r) = 0, r > r
c
. Then, the twopoint function becomes a
product of two onepoint functions, P(a
1
, a
2
, r) P(a
1
)P(a
2
), see also Ref. 116. For very short distances r 0, on
the other hand, we have f(r) 1 so that P(a
1
, a
2
, r) P(a
1
)(a
1
a
2
).
The mathematically much harder but physically very important case of orthogonal symmetry ( = 1) was also
worked out by Prigodin et al.
124
. Srednicki
125
showed that these results can also be obtained in a direct calculation
without supersymmetry. Using Berrys conjecture
118
he found a distribution for the wave functions which formally
coincides with the distribution (3.107) for the partial width amplitudes. From that, he immediatley obtains Eq. (3.119)
for = 2 and, in the notation of Ref. 124,
P(a
1
, a
2
, r) =
1
2
_
1 f
2
(r)

a
1
a
2
exp
_

a
1
+a
2
2(1 f
2
(r))
_
cosh
_
f(r)

a
1
a
2
1 f
2
(r)
_
(3.120)
49
for = 1. In the case of = 2 equivalent channels with
1
=
2
, the distribution of the total width is given by the
integral
115
P() =
_

0
da
1
_

0
da
2
P(a
1
, a
2
, r)
_

a
1
+a
2
2

_
(3.121)
where is the average value. Inserting Eqs. (3.119) and (3.120), one can obtain the results (3.110) and (3.111).
This does not imply, however, that the supersymmetry method has become obsolete in this context. It is not clear
how to use the methods of Refs. 9,110 in a crossover region, such as, for example, the gradual breaking of timereversal
invariance. In the supersymmetry method, such calculations are possible, see Sec. III G.
F. Crossover transitions in spectral correlations
The concept of fully developed chaos and, likewise, the strict applicability of the orthogonal, unitary or symplectic
symmetry class are often an idealization. Frequently, we deal with systems which undergo a crossover transition
between dierent statistics, or dierent universality classes. The physically most interesting situations are (i) tran-
sitions from regular to chaotic uctuation properties, (ii) breaking of timereversal invariance, and (iii) breaking of
symmetries.
We stress the dierence between the breaking of timereversal invariance and symmetry breaking. In quantum
mechanics, the timereversal operator is antiunitary while an operator corresponding to a symmetry of the Hamiltonian
is unitary. If a symmetry such as angular momentum, parity etc. holds, the Hamiltonian can be reduced and can
be written in blockdiagonal form. Each block belongs to a given quantum number, i.e. a representation of the
symmetry such as positive or negative parity. On the other hand, the timereversal operator cannot be used to reduce
the Hamiltonian.
In this section we study crossover properties of energycorrelation functions. These functions depend on one or
several external parameters which represent the physical situation. The functions must be distinguished from the
parametric correlations which involve two or more values of the same external parameter. The latter are discussed
in Sec. III H.
In Sec. III F1 we discuss transitions in terms of a diusion process for the joint probability density in matrix space.
Thereby, we summarize some of the results of Dysons Brownian motion model. In Sec. III F2 we show that a similar
diusion equation exists for the generating functions of the correlators in the much smaller space of supermatrices.
In Sec. III F3 we summarize the numerical and analytical results.
1. Diusion in the space of ordinary matrices
We recall the denition (3.12) of the probability distribution for the Gaussian ensembles. The volume element (3.8)
factorizes in an eigenvalue part and an angular part. Because of rotation invariance, the same statement holds for
the joint probability density of the eigenvalues which is given by Eq. (3.9) and the eigenvectors which is the Haar
measure of the diagonalizing group. The local uctuation properties are governed by the Vandermonde determinant
in Eqs. (3.8) and (3.9). This quantity turns out to be also crucial for the investigation of crossover transitions. From
a mathematical viewpoint, this is not surprising: The Jacobian of a transformation reects the topology of the spaces
in question. However, the various breakings of rotation invariance employed in studying crossover transitions destroy
the convenient factorization in an eigenvalue part and an angular part.
Dyson
126,127
was the rst to realize that crossover transitions can be formulated as Brownian motion of the eigen-
values. He constructed the Smoluchowski equation for the joint probability density. Recent reviews can be found in
the books by Haake
57
and Mehta
23
, see also the discussion by Lenz and Haake
128
. The connection to the Pechukas gas
and the CalogeroSutherlandMoser model
57
will be discussed in Sec. VII. Here, we want to take a slightly dierent
route and use explicitly the universality of the unfolded correlations, see Sec. VIII. It allows us to restrict ourselves
to the Gaussian probability density (3.12).
To formulate the problem, we consider a random Hamiltonian
H = H
(0)
+H
(1)
(3.122)
as a function of the transition parameter . Each of the three matrices belongs to the same symmetry class. The
elements of H
(1)
are Gaussian distributed according to Eq. (3.12). The distribution P
(0)
N
(H
(0)
) of the elements of
H
(0)
, however, is completely arbitrary. In particular, P
(0)
N
(H
(0)
) can restrict the symmetry of H
(0)
to a subsymmetry
50
of H
(1)
. In the case of a Hermitean H
(1)
, for example, P
(0)
N
(H
(0)
) can put all imaginary parts of the elements of
H
(0)
to zero, thereby eectively restricting H
(0)
to a real symmetric matrix. The general problem is: How does the
distribution P
(0)
N
develop into the full Gaussian (3.12) as increases?
We absorb the parameter into the probability density by writing /4
2
instead of /2 in the exponent of
Eq. (3.12). The resulting function is denoted by P
(1)
N
. The squared transition parameter may be viewed as a ctitious
time t =
2
/2 which governs the evolution from the arbitrary into the Gaussian form. Introducing the matrix gradient
/H and the Laplacean
= tr

2
H
2
(3.123)
in the Cartesian spaces of the three symmetry classes, we can identify the Gaussian P
(1)
N
(H, t) as a diusion kernel
dened through

4
P
(1)
N
(H, t) =

t
P
(1)
N
(H, t) and lim
t0
P
(1)
N
(H, t) = (H) . (3.124)
The evolution of the total probability density is given as the solution of the diusion equation

4
P
N
(H, t) =

t
P
N
(H, t) with lim
t0
P
N
(H, t) = P
(0)
N
(H) . (3.125)
According to Eq. (3.124), the solution can be written as the convolution
P
N
(H, t) =
_
P
(1)
N
(H H
(0)
, t)P
(0)
N
(H
(0)
)d[H
(0)
] . (3.126)
For initial conditions which depend only on the eigenvalues X
(0)
of H
(0)
, the diusion process is restricted to the
curved space of the eigenvalues. For the joint probability density, one obtains

X
P
N
(X, t) =

t
P
N
(X, t) and lim
t0
P
N
(X, t) = P
(0)
N
(X) . (3.127)
The radial part of the Laplacean is given by

X
=
1
[
N
(X)[

n=1

x
n
[
N
(X)[

x
n
. (3.128)
The notation
N
(X) for the Vandermonde determinant should not be confused with the symbol
X
standing for the
radial part of the Laplacian. This diusion equation (3.127) is solved by the convolution
P
N
(X, t) =
_

N
(X, X
(0)
, t)P
(0)
N
(X
(0)
)[
N
(X
(0)
)[

d[X
(0)
] (3.129)
and the kernel
N
(X, X
(0)
, t) is given by the group integral

N
(X, X
(0)
, t) =
_
P
(1)
N
(U
1
XU X
(0)
, t)d(U) . (3.130)
In the unitary case = 2, this is the famous Harish-Chandra-Itzykson-Zuber integral
129,130
. For the joint probability
density of the eigenvalues, it yields
P
(E)
N2
(X, t) = P
N2
(X, t)
2
N
(X) =
1

2t
N
_
d[X
(0)
]P
(0)
N
(X
(0)
)
exp
_

1
2t
tr(X X
(0)
)
2
_

N
(X)

N
(X
(0)
)
. (3.131)
Since Eq. (3.1) applies also in the case of transitions, we can use Eq. (3.131) to compute the klevel correlation
functions in the unitary case. An alternative procedure for the Circular Unitary Ensemble was recently given by
51
Pandey
131
. He starts from an exact formal solution of the diusion equation in terms of representation functions of
the unitary group U(N). His method is equivalent to calculating the Harish-Chandra-Itzykson-Zuber integral. As
discussed in Sec. III A1, the correlation functions must be properly unfolded by introducing
132
an unfolded transition
parameter = /D and an unfolded ctitious time =
2
/2.
An important property of the correlation functions was obtained by French et al.
133
. Starting from the Fokker
Planck or Smoluchowski equation closely related to the diusion equation discussed above, these authors derived
hierarchic relations among the correlation functions. The unfolded correlation functions are dened by extending
Eqs. (3.6) and (3.7). On this unfolded scale, the hierarchic equations involve the ctitious time . They acquire the
form

X
k
(
1
, . . . ,
k
, ) =
k

p=1
1
[
k
()[

p
[
k
()[

p
X
k
(
1
, . . . ,
k
, )

p=1

p
_
+

X
(k+1)
(
1
, . . . ,
k
,
k+1
, )

k+1
d
k+1
(3.132)
where
k
() =

p<q
(
p

q
). For k = 1 one obtains a closed equation which coincides with the famous Pastur
equation
22
.
Als already mentioned, Brownian motion and diusion also exist in the circular ensembles
126,131
. We mention this
here only briey since on the unfolded scale, the uctuation properties agree with those of the Gaussian ensembles.
Pandey and Shukla
134
have shown that the ensuing hierarchic relations coincide with Eq. (3.132).
2. Diusion in the space of supermatrices
Remarkably, the diusion equation (3.127) for the joint probability density in ordinary space has an analogue in
superspace
135
. We consider an ensemble of random Hamiltonians of the form (3.122). We are interested in the
correlation functions

R
k
(x
1
, . . . , x
k
, t). In contrast to Eq. (3.2), these functions are dened in terms of the full Green
functions and do contain the principal value parts. Following Refs. 44,46, we express these correlation functions as
derivatives

R
k
(x
1
, . . . , x
k
, t) =
1
(2)
k

k
p=1
J
p
Z
k
(x +J, t)

J=0
(3.133)
of a normalized generating function Z
k
(x + J, t). Here, x and J contain the k energies and the k source variables,
x = diag(x
1
, x
1
, x
1
, x
1
, . . . , x
k
, x
k
, x
k
, x
k
), J = diag(J
1
, J
1
, +J
1
, +J
1
, . . . , J
k
, J
k
, +J
k
, +J
k
) for = 1 and =
4, and x = diag(x
1
, x
1
, . . . , x
k
, x
k
), J = diag(J
1
, +J
1
, . . . , J
k
, +J
k
) for = 2. The physically interesting functions
R
k
(x
1
, . . . , x
k
, t) of Eq. (3.2) are constructed by the operation Z
k
(x + J, t) which produces the proper linear
combination
135,92
without the principal value parts. With the help of standard techniques of the supersymmetry
method
44,46
, the average over the Gaussian distributed random part H
(1)
can be performed directly and the generating
function acquires the form
Z
k
(x +J, ) =
_
d[H
(0)
]P
(0)
N
(H
(0)
)
_
d[]Q
k
(, t)
detg
1
_
(x

+J ) 1
N
1

H
(0)
_
. (3.134)
Here, 1
N
and 1

k
are unit matrices of dimension N and

k, respectively, and

= 4 for the GOE and the GSE

while

= 2 for the GUE. The supermatrix reects the symmetries of H

(1)
and has dimension

k. With c
k
a
normalization constant, the graded probability density is given by
Q
k
(, t) = c
k
exp
_

4t
trg
2
_
(3.135)
In the limit t 0 it reduces to the superspace function (). For vanishing transition parameter t, Z
k
therefore
becomes the generating function
52
Z
(0)
k
(x +J) =
_
d[H
(0)
]P
(0)
N
(H
(0)
)
detg
1
_
(x

+J) 1
N
1

k
H
(0)
_
(3.136)
for the correlations of the ensemble of matrices H
(0)
.
Because of the Gaussian form of the graded probability density (3.135) we can formulate a diusion process in
superspace. We shift x J and diagonalize the supermatrix = u
1
su where s is the diagonal matrix of
the eigenvalues of . There are 2k eigenvalues for the GUE and 3k eigenvalues for the GOE and the GSE. We now
replace the matrix x +J by an diagonal matrix r having the same form as s. These steps yield
Z
k
(r, t) =
_
Q
k
( r, t) Z
(0)
k
(s) d[] . (3.137)
There are only 2k energies and source variables in x + J, whereas the matrix r contains 3k variables for = 1 and
= 4. The convolution integral (3.137) is related to a diusion equation. Since Z
(0)
k
(s) depends only on s, the
diusion takes place in the curved space of the eigenvalues of the supermatrix. In analogy to the previous section,
the diusion equation has the form

r
Z
k
(r, t) =

t
Z
k
(r, t) with lim
t0
Z
k
(r, t) = Z
(0)
k
(r) (3.138)
and holds for all three symmetry classes and for arbitrary initial generating functions Z
(0)
k
(r). The Laplacean
r
is the radial part of the full Laplacean = trg
2
/
2
, dened in terms of the matrix gradient /. As usual,
the radial part involves the Jacobians or Berezinians B
k
(r) of the transformation to radial coordinates. Explicit
expressions for these quantities are not given here and may be found in Ref. 135. We notice that Eq. (3.127) describes
a diusion of the probability density of the eigenvalues in the space of ordinary matrices whereas Eq. (3.138) describes
a diusion of the generating function for the correlations in the space of supermatrices.
The corresponding diusion kernel is given by the angular average

k
(s, r, t) =
_
Q
k
(u
1
su r, t) d(u) (3.139)
where d(u) is the Haar measure. It also satises the diusion Eq. (3.138), and an initial condition at t = 0 not given
here. This implies that we can write the solution of Eq. (3.138) in the form
Z
k
(r, t) =
_

k
(s, r, t)Z
(0)
k
(s)B
k
(s)d[s] . (3.140)
The diusion process cannot be formulated for the k variables x but only for the 2k or 3k variables r, respectively.
Therefore, the role of the joint probability density in ordinary space is here played by the generating function rather
than the correlation functions.
In the unitary case = 2, the kernel
k
is explicitly known: It is the supersymmetric Harish-Chandra-Itzykson-
Zuber integral
92
. The solution of the diusion equation reads
Z
2,k
(r, t) =
1
B
2,k
(r)
_
G
k
(r s, t)Z
(0)
k
(s)B
2,k
(r)d[s] (3.141)
where the relevant part of the kernel is the Gaussian
G
k
(s r, t) =
1

2t
2k
exp
_

1
2t
trg(s r)
2
_
. (3.142)
The correlation functions can now be calculated straightforwardly
135,136
. They are given by the 2kdimensional
integral representation
R
2,k
(x
1
, . . . , x
k
, t) =
(1)
k

k
_
G
k
(s x, t)Z
(0)
k
(s)B
2,k
(r)d[s] (3.143)
which is exact for all initial conditions and also exact for any nite level number N.
53
It can easily be seen
135
that the transition on the unfolded scale is a diusive process as well. With = r/D,
z
(0)
k
() = lim
N
Z
(0)
k
(r), and z
k
(, ) = lim
N
Z
k
(r, t) one nds

z
k
(, ) =

z
k
(, ) with lim
0
z
k
(, ) = z
(0)
k
() . (3.144)
The diusion processes on the original and the unfolded scale are the same, only the initial conditions are dierent,
see Ref. 135. Therefore, the integral representation (3.140) must hold on the unfolded scale as well,
z
k
(, ) =
_

k
(s, , ) z
(0)
k
(s) B
k
(s)d[s] . (3.145)
In particular, in the unitary case = 2 we nd
z
2,k
(, ) =
1
B
2,k
()
_
G
k
( s, t)z
(0)
k
(s)B
2,k
(r)d[s] (3.146)
for the generating function. From that, the 2kdimensional integral representation
X
2,k
(
1
, . . . ,
k
, ) =
(1)
k

k
_
G
k
(s , )z
(0)
k
(s)B
2,k
(r)d[s] (3.147)
for the correlation functions on the unfolded scale is obtained. In the case k = 2, a further simplication arises because
the twolevel correlation function depends only on r =
1

2
. The remaining four integrals can be trivially reduced
to two integrals and one has
135,136
X
2,2
(r, ) =
4

+
_

+
_

exp
_

1
4
(t
2
1
+t
2
2
)
_
t
1
t
2
(t
2
1
+t
2
2
)
2
sinh
rt
1
2
sin
rt
2
2
z
(0)
2
(t
1
, t
2
)dt
1
dt
2
. (3.148)
The trigonometric and the hyperbolic functions can be directly traced back to the eigenvalues in the BosonBoson
and the FermionFermion block.
We compare the hierarchic equations (3.132) and the diusion equation (3.144). The former couple the correlation
functions with index k to all those with index up to k +1 and thus cannot be viewed as describing a diusion process.
The diusion equation (3.144) in superspace, on the other hand, does not couple dierent values of k. In this sense,
the diusion equation (3.144) diagonalizes the hierarchic equations (3.132).
Starting from Eqs. (3.138) and (3.144), one can also derive stationary equations for the pure ensembles on the
original and the unfolded scale
135
.
In the case of symmetry breaking, the standard model is that of a blockdiagonal Hamiltonian H
(0)
to which the
perturbation H
(1)
is added,
H
(0)
=
_
H
(0,1)
0
0 H
(0,2)
_
and H
(1)
=
_
0 H
(1,P)
H
(1,P)
0
_
. (3.149)
The matrices H
(0,1)
and H
(0,2)
represent the system for two xed values 1 and 2 of some quantum number such as
parity or isospin. The perturbation H
(1)
breaks the blockdiagonal structure. Although symmetry breaking is not
fully compatible with diusion in superspace, formula (3.148) was found to apply to this case
137
, too. Thus, it is valid
beyond the diusion model.
3. Numerical and analytical results
We follow the scheme mentioned in the introduction to Sec. III F and begin with the transition from regularity
to chaos. Classical integrability may be destroyed when some system parameter such as shape, external eld, etc.,
varies. Then, the structure of classical phase space changes and chaotic regions appear. Such change aects the
uctuation properties of the analogous quantum system. Modeling this process is not trivial, however, since uctuation
properties of regular systems are not generic. Therefore, the crossover transition may not have generic properties either.
54
Nonetheless, much work has been devoted to this problem. Poisson regularity and harmonic oscillator regularity have
served as main examples of classically integrable systems, see Secs. III B2 and III B3.
Several scenarios have been used to study the transition from Poisson statistics to WignerDyson (WD) statistics:
(i) An ensemble of blockdiagonal Hamiltonian matrices consisting of two blocks, one with Poisson statistics and the
other taken from one of the three classical ensembles, will show a transition from Poisson to WD statistics as the
ratio of the dimensions of the regular and the random matrix block changes from innity to zero. Following work
by Mehta
23
, Berry and Robnik
138
calculated a spacing distribution which describes this transition. (ii) An ensemble
of Hamiltonians of the form H
(0)
+ H
(1)
discussed in the previous section where H
(0)
is diagonal and the diagonal
elements obey Poisson statistics, while H
(1)
is a member of an ensemble of random matrices of proper symmetry. The
transition parameter denes the relative strength of H
(1)
. On the unfolded scale, the parameter = /D where
D is the mean level spacing has to be used. The phenomenological Brody formula
65
interpolates between the two
limiting cases = 0 and without being explicitly derived from the model, see Sec. III B2. (iii) Random band
matrices where all matrix elements located beyond a certain distance from the main diagonal, vanish. For xed matrix
dimension and decreasing (increasing) band width, the uctuation properties approach Poisson (WD) statistics
67,139
,
respectively. These, the related chains of Gaussian ensembles
140,141
and other models involving crossover transitions
are studied in Secs. VI D, VI C and VI A, respectively. (iv) Moshe et al.
142
studied a rotation invariant ensemble
which interpolates between Poisson and WD statistics.
A large body of work exists on models of type (ii). Caurier et al.
68
and Lenz and Haake
69
calculated the spacing
distribution of twodimensional matrix models. Numerical studies for timereversal symmetric systems were presented
in Refs. 143146. For a small chaotic admixture, closed formulae were obtained analytically in dierent, but equivalent,
perturbation schemes by French et al.
133
, by Leyvraz and Seligman
147
, and by Leyvraz
148
for the transition from
Poisson regularity to all three Gaussian ensembles. The exact computation of the twopoint function for the entire
transition has up to now been possible only in the unitary case. Unlike the GOE and the GSE, the GUE can be
understood as an interactionfree statistical model, cf. Eq. (6.79) of Sec. VI C2. This is reected in the fact that the
diusion kernel (3.130) is known explicitly as the HarishChandraItzyksonZuber integral. Lenz
149
calculated the
correlation functions with the MehtaDyson method
23
. He used Eq. (3.131) in Eq. (3.1) for k = 2 and found a four
dimensional integral representation. His result contains a ratio of determinants which depend on the level number.
Unfortunately, this representation does not seem to be amenable to further analytical treatment. Pandey
131
worked
out an exact expression for the twopoint function on the unfolded scale in terms of a double integral. Avoiding
formula (3.131) he used an exact formal solution of the Fokker-Planck equation in terms of representation functions
of the unitary group U(N). Another exact derivation of the twopoint function on both scales, the original and the
unfolded one, was given in Refs. 135,136 with the help of the supersymmetry method. Equation (3.148) gives the
twopoint function as a double integral for arbitrary initial conditions, and one simply needs to compute the proper
initial condition of the Poisson ensemble. This can be done straightforwardly. Rather compact exact expressions
for the kpoint functions on the original and the unfolded scale were given in terms of a 2kdimensional integral in
Refs. 135,136.
Most of the concepts just discussed can be adopted to the transition from harmonic oscillator regularity to chaos.
In the spirit of the BerryRobnik procedure, one can write down a formula for the spacing distribution in a block
diagonal model. The transition in a model of the type H
(0)
+ H
(1)
was simulated in Ref. 143. As in the case of
the transition from Poisson regularity, analytical results for the twopoint function for the entire transition are only
available in the unitary case
131,150
.
For the breaking of timereversal invariance, the random matrix ensemble has to interpolate between the GOE
and the GUE limits, see Eq. (2.13). This particular choice for the total Hamiltonian H is often referred to as the
MehtaPandey Hamiltonian. In a perturbative scheme, analytical results for the twopoint correlations and the level
number variance were given by French et al.
133,151
. Using the Harish-Chandra-Itzykson-Zuber integral, Mehta and
Pandey
28,29
derived exact analytical results on the original and on the unfolded scale in a pioneering study as early as
1983. They employed Eq. (3.131) in Eq. (3.1) and found that the klevel correlation function can be written as a kk
quaternion determinant. This expression reproduces the quaternion determinant in the GOE limit and reduces to the
ordinary determinant in the GUE limit. The calculation benets from this determinant structure which is absent in
the Poisson to GUE transition. In Ref. 152 the result for the twopoint function on the unfolded scale was reproduced
by using the supersymmetry method and the saddlepoint approximation which leads to Efetovs nonlinear model.
Pandey and Shukla
134
calculated all correlation functions of the corresponding transition in the circular ensembles.
Symmetry breaking is stuidied in terms of the model dened in Eq. (3.149). Both numerical simulations and
analytical results were reported in Ref. 137. The latter were derived for the unitary case and for equal dimensions
of the symmetryconserving blocks. The supersymmetry technique was used which led to an integral representation
of the type (3.148). A generalization to the case where the blocks have dierent dimensions is due to Pandey
131
. A
detailed discussion in the framework of perturbation theory was given by Leitner
153
. His work includes a surprisingly
ecient formula for the spacing distribution.
55
A remarkable speed with which the WignerDyson uctuations are reached on short scales is common to all
transitions discussed here. More precisely, as the transition parameter = /D increases, observables such as the
spacing distribution p(s) become in almost all cases indistinguishable from the WignerDyson expectation when the
transition parameter is in the regime 0.7 1 or so. In the spectral long range observables, such as the level
number variance
2
(L) and the spectral rigidity
3
(L), one can identify, as a function of , the maximal interval
length L
max
within which the uctuations are of WignerDyson type. This issue is very important in manybody
systems and in disordered systems, it relates to the spreading width and to the Thouless energy. This discussion will
be taken up in Secs. IV and VI.
Finally, we mention a model which employs the concept of transitions in random matrix theory for a deeper
understanding of classical and semiclassical systems. Aiming at an understanding of how the the classical phase
space in a mixed system manifests itself in quantum mechanics, Berry and Robnik
138
and later, in much more detail,
Bohigas et al.
154
constructed random matrix models whose block structure can be viewed as generalizations of certain
transition models. This is discussed in Sec. VH. Formally, these models are in the spirit of the RosenzweigPorter
model
82
which was introduced as early as 1960.
G. Crossover transitions in other observables
Crossover transitions also aect the distributions of wave functions and widths. Most authors considered the
transition from orthogonal to unitary symmetry, i.e. the breaking of timereversal invariance. To obtain a family of
distributions which interpolates between the
2

distributions of the widths in Eq. (3.105) for the two limiting cases
= 1 and = 2, one could simply view as a continuous variable. This is mathematically possible. It is not
clear, however, how to relate the continuous variable to the strength of an interaction which breaks timereversal
invariance. Another interpolating formula due to Zyczkowski and Lenz
155
has similar shortcomings.
A meaningful interpolating formula has to be derived from a Hamiltonian of the form H = H
(0)
+ H
(1)
where
must be a physically welldened parameter. Such a derivation is nontrivial since the joint probability densities
of eigenvalues and eigenvectors of the random matrix H factorize only in the two limiting cases = 0 and = .
Sommers and Iida
156
and Falko and Efetov
157
gave exact integral representations for the distributions of wave
functions and widths which are valid for the entire transition regime. The supersymmetry method is used in both
works. Although equivalent, the starting points are slightly dierent. In Ref. 156, the full distribution of eigenvalues
and eigenvectors in the transition regime was used, whereas in Ref. 157, the moments of the eigenvector distribution
were calculated.
Falko and Efetov
158
calculated correlations of wave functions of the type (3.116) in the crossover transition from
orthogonal to unitary symmetry and found a surprising result. In the two limiting cases, the correlations at distant
points in space vanish as expected. For = 1 and = 2, the correlation functions become products of two
2

distributions. However, the correlations do not vanish in the transition regime. Recently, van Langen et al.
159
studied
these unexpected longrange correlations by relating them to phaserigidity uctuations of chaotic wave functions.
H. Parametric level motion and parametric level correlations
Do chaotic systems possess observables other than energy correlators or wave function and width distributions
which also show a high degree of universality? This question, answered in the armative in recent years, is treated
in the present section. In Sec. III H1, we summarize results on the motion of energy levels when a parameter of the
system varies. The correlations associated with this motion are presented in Sec. III H2. In Sec. III H3, we discuss a
generalization of the Gaussian ensembles, the Gaussian processes, which models these parametric correlations.
1. Level motion and curvature
Wilkinson
160,161
studied a classically chaotic quantum system with Hamiltonian H(X) which is in the same symme-
try class for all values of X. On the scale of the mean level spacing, the spectral uctuations of H(X) have the same
statistics for all X and are given by one of the three Gaussian ensembles. However, both eigenvalues E
n
= E
n
(X)
and eigenfunctions depend on X. For a chaotic system, there is always level repulsion, in contrast to the regular case.
Hence, a plot of the spectrum of H versus X shows many avoided crossings, see Fig. 16. We consider the parameter
X as a function of a ctitious time t, X = X(t). Suppose that at time t = 0, the system is prepared in a highlying
eigenstate of H(X(0)). Wilkinson showed that with increasing t, the occupation probability of eigenstates of H(X(t))
56
FIG. 16. The motion of eigenenergies as a function of an external parameter X. Many avoided crossings due to level
repulsion can be seen, indicating that the level statistics at any xed value of X is predominantly of WignerDyson type. This
picture illustrates results of a calculation for the hydrogen atom in a strong magnetic eld by Goldberg et al.
162
. Taken from
Ref. 163.
spreads diusively from the initially prepared state. The parameter dX/dt measures the nonadiabaticity of the
change of H. For small , the diusion process is governed by LandauZener transitions at the avoided crossings.
The diusion constant is proportional to
(+2)/2
where labels the three symmetry classes. This reects the power
law s

of the nearest neighbor spacing distributions for small spacings. It is argued that, for large , the diusion is
governed by Ohmic dissipation and the diusion constant is proportional to
2
, independent of the symmetry class.
The statistics of the avoided crossings was worked out in Ref. 164. In recent, related work, Wilkinson and Austin
165
found a diusion constant proportional to
2
for = 1. This is in agreement with the Kubo formula (see Sec. VI C1).
However, there is a limiting value of
2
beyond which the rate is lower. To test these results, these authors also
performed numerical studies involving random band matrices. This study suggests that the transition mechanism is
dierent from the above mentioned discussion. Simultaneously, Bulgac, Do Dang and Kusnezov
166
performed closely
related studies.
A useful local measure of the parametric dependence of eigenvalues is the curvature
K
n
=
d
2
E
n
(X)
dX
2

X=0
(3.150)
of the level motion. Indeed, [K
n
[ takes large values at the avoided crossings. Several authors
167169
used K
n
or related
quantities to investigate the transition from regular to chaotic motion. Another measure is the distribution P(K) of
curvatures. We drop the index n because P(K) must be the same for all eigenvalues. Gaspard et al.
170,171
investigated
P(K) in the fully chaotic regime. These authors used arguments involving the Pechukas gas
172
and the generalized
CalogeroMoser system
173,174
to derive the asymptotic behavior of P(K) for large K. They found that this form is
universal and given by 1/[K[
+2
. As in the case of the diusion of occupation probabilities discussed above, this form
is due to level repulsion since roughly [K[ 1/s. The universal character of this result was numerically conrmed in
Refs. 175178 for several dierent systems. In studying the kicked top, Zakrzewski and Delande
177
conjectured that
the full distribution is given by
P(k) =
c

(1 +k
2
)
(+2)/2
(3.151)
where the normalization constants read c
1
= 1/2, c
2
= 1/2 and c
4
= 8/3. The dimensionless curvature k =
DK/(dE
n
/dX)
2
) is given in terms of the mean level spacing D and the mean square level velocity (dE
n
/dX)
2
).
57
This quantity is independent of the index n. Using the supersymmetry method, von Oppen
179,180
could prove that
the conjecture (3.151) is indeed true. He used a random matrix model dened through the Hamiltonian
H(X) = H
1
cos X +H
2
sin X (3.152)
in which both H
1
and H
2
belong to the same symmetry class. Several authors have stressed that at this point a
caveat is in order: Second order perturbation theory yields an exact expression
179
for the curvature (3.150) in the
model (3.152). Consequently, formula (3.151) is universal only for those models which yield an equivalent expression
for the curvature. The class of, in this sense, equivalent models, however, is obviously very large. Furthermore, the
tail of the distribution is expected to be rather insensitive to the particular model because it is directly related to
the spacing distribution. In numerical simulations of billiard systems, Li and Robnik
181
found deviations from the
formula (3.151). Recently, Yurkevich and Kravtsov
182
calculated analytically corrections to this formula.
2. Parametric correlation functions
In contrast to the previous section we consider now two dierent points X and X

in the space of the parameter

which governs the level motion. We study the correlation of E
n
(X) and E
m
(X

) as function of the separation in

energy and in parameter value. For X = X

we have to recover the universal spectral correlations of RMT. New

universal features can only be expected on scales on which the gross features of the system play no role, i.e. on the
scale of the mean level spacing D. We therefore dene the unfolded energies
n
(X) = E
n
(X)/D. We must also unfold
the level motion. The appropriate scale is given by the velocity d
n
(X)/dX. We dene the new parameter
x =

_
_
_
d
n
(X)
dX
_
2
_
X . (3.153)
The average is performed either over the spectrum or, in an analytical calculation, over the ensemble. We recall that
the same average is used for the normalization of the curvature parameter k below Eq. (3.151).
Goldberg et al.
162
introduced the parametric number variance as a probe of correlations between dierent points
in parameter space. The cumulative spectral function or staircase function,
(, x) =
N

n=1
(
n
(x)) (3.154)
serves to dene the parametric level number variance
v(x) =
_
_
(, x x/2) (, x +x/2)
_
2
_
(3.155)
which is obtained by averaging over both the spectrum (or the ensemble) and over the position of the midpoint x
between two points in parameter space, keeping the distance x xed. Note that v(x) is not directly related to the
spectral level number variance
2
(L): At x = 0, the function v(x) vanishes by denition. With increasing x, the two
staircase functions in Eq. (3.155) get more and more out of phase. Hence, we expect that v(x) is a monotonically
increasing function of x.
Another natural measure of parametric level correlations is the correlation function of the velocity of a xed
eigenvalue
n
taken at two dierent points x x/2 x +x/2
c(x) =
_

n
(x x/2)
x

n
(x +x/2)
x
_
. (3.156)
This function was introduced by Szafer and Altshuler
183
who studied a disordered electronic system in the form of
a ring threaded by a magnetic ux. Using diagrammatic perturbation theory, these authors
184
found that beyond a
critical regime the correlator (3.156) becomes independent of any system properties and inversely proportional to the
squared ux (x
2
in the present notation). Together with the investigations by Goldberg et al., this gave the rst clear
indication that parametric correlations of level motion possess universal features.
Beenakker
185
argued that it ought to be possible to extract such universal features from the standard concepts of
Random Matrix Theory. Indeed, Dysons Brownian motion
126,127
and the Pechukas gas
172
both model the motion of
58
levels as function of a ctitious time t. Beenakker studied Dysons Brownian motion under the crucial assumption
that t can be identied with the square of the parameter which governs the level motion. He found that the smooth
behavior in x of the correlator (3.156) is universally given by
c

(x)
2

2
x
2
for x 1 . (3.157)
This conrmed the result of Szafer and Altshuler in the unitary case = 2. The restriction x 1 is not imposed
by Dysons Brownian motion model itself, but is only due to the particular method chosen in Ref. 185 to derive
Eq. (3.157). Beenakker pointed out that the universality of his result conrms the fact that parametric correlations
are dominated by level repulsion and thus uniquely determined by the symmetry class. The specic physical realization
of the parameter X is immaterial. In the unitary case, the result (3.157) was also obtained by Brezin and Zee
186
who
used a diagrammatic expansion.
Within the supersymmetric model, Simons and Altshuler
187,188
simultaneously derived exact expressions for the
parametric densitydensity correlator at two dierent energies /2 and +/2,
k

(, x) =
N

n,m=1
_
( /2
n
(x x/2))
( +/2
m
(x + x/2))
_
1 (3.158)
and the parametric currentcurrent correlator at xed energy separation
c

(, x) =
1

N
n,m=1
(
n
(x x/2)
m
(x +x/2)) ))
N

n,m=1
_
(
n
(x x/2)
m
(x +x/2)) )

n
(x x/2)
x

m
(x +x/2)
x
_
.
(3.159)
The two correlators (3.158) and (3.159) are connected through a Ward identity. Hence, we restrict ourselves to the
densitydensity correlator (3.158). Simons and Altshuler studied a disordered system. They used Efetovs supersym-
metric nonlinear model
44
. They point out that the same results can be derived starting from a random matrix
model in the spirit of Ref. 46. The integral representations in the three universality classes read
k
1
(, x) = Re
_
+1
1
d
_

1
d
1
_

1
d
2
(1
2
)(
1

2
)
2
(2
1

2
1

2
2
+ 1)
2
exp
_

2
x
2
4
(2
2
1

2
2

2
1

2
2
+ 1) i
+
(
1

2
)
_
k
2
(, x) =
1
2
Re
_
+1
1
d
_

1
d
1
exp
_

2
x
2
2
(
2
1

2
) i
+
(
1
)
_
k
4
(, x) =
1
2
Re
_

1
d
_
+1
1
d
1
_
+1
1
d
2
(
2
1)(
1

2
)
2
(2
1

2
1

2
2
+ 1)
2
exp
_
2
2
x
2
(
2
1
+
2
2
+
2
2
2
1

2
2
1) +i2
+
(
1

2
)
_
(3.160)
where the energy dierence is given an imaginary increment. The symplectic case was worked out in Ref. 189. For
x = 0, Efetovs integral representations of the spectral twopoint functions
44
are recovered. For = 2, Beenakker and
Rejai
163
performed a calculation in the framework of Dysons Brownian motion model which yields full agreement with
the results of Simons and Altshuler. From the denition (3.159) it is obvious that for large x, the currentcurrent
correlator c(0, x) at zero energy separation has to approximate the velocity correlator c(x) dened in Eq. (3.156).
Using the Ward identity mentioned above, the result (3.157) can indeed be rederived. Remarkably, no integral
representation of the form (3.160) valid for arbitrary x could be derived for c(x) yet. Mucciolo
190,191
obtained c

(x)
by extensive numerical simulations for orthogonal ( = 1) and unitary ( = 2) symmetry. He used random matrices
of the form (3.152). The results are shown in Fig. 17.
Just as its spectral analogue, the parametric level number variance can be expressed in terms of the twopoint
function,
59
0.0 0.5 1.0 1.5 2.0
x
-0.5
0.0
0.5
1.0
c
(
x
)
FIG. 17. The parametric velocity correlators c

(x), obtained from numerical simulations. The solid line is the GOE result
( = 1), the dashed line is the GUE result ( = 2). The curves have a very small numerical error. Taken from Ref. 190.
v

(x) = 2 lim
x

0
_
0

d
_

(k

, x

) k

, x)) . (3.161)
For small values of x, these integrals can be worked out. For the unitary and the orthogonal case, one nds
v

(x)
_
2

[x[ for x 0 . (3.162)

For larger values of x, the parametric level number variance can be obtained by numerical integration of Eq. (3.161). It
turns out that v
1
(x) and v
2
(x) begin to dier. A discrepancy with the semiclassical formula in Ref. 162 was attributed
by Simons and Altshuler to the failure of the semiclassical method for small x.
The results derived within the supersymmetric nonlinear model give the most general proof of the universality of
parametric correlations. Although experiments and numerical simulations have veried this universality, see Ref. 192,
deviations have also been found.
Naturally, the question arises whether observables involving the wave functions also possess universal features.
Using the Gaussian processes discussed in the following section, Attias and Alhassid
193
studied the normalized energy
dependent parametric overlap of wave functions
n
(r, x) at dierent values of the dimensionless parameter x,
o

(, x) =
1

N
n,m=1
(
n
(x x/2)
m
(x +x/2)) ))
N

n,m=1
_
[
n
(r, x x/2)[
m
(r, x +x/2))[
2
(
n
(x x/2)
m
(x +x/2)) )
_
. (3.163)
With the supersymmetry method, exact expressions were given for = 1 and = 2 and compared to numerical
simulations of the Anderson model. In the orthogonal case and for zero energy separation, the overlap is Lorentzian.
As increases, the shape changes, and the peak moves from x = 0 to larger values of x.
Mucciolo et al.
194
also studied the dependence of wave functions in the chaotic regime on a parameter, using the
full joint distribution function
60
w

(a, a, X) =
_
(a a/2 V [
n
(r, x x/2)[
2
)
(a +a/2 V [
n
(r, x +x/2)[
2
)
_
. (3.164)
They showed that in the limit of small x, this distribution has asymptotically the universal form
w

(a, a, X) P

(a)
1
2x

a
B

_
a
2x

a
_
for x 0 (3.165)
where P

is the
2
distribution dened in Eq. (3.105). The function B

is also universal and depends only on the

symmetries of the system. In the unitary case, the function B
2
was given explicitly.
3. Gaussian random processes
The universality of the parametric correlation functions suggests the existence of a proper generalization of the
Gaussian ensembles. This generalization should lead naturally to the parametric correlations. In the context of
parametric level motion, and for numerical simulations of the diusion constants discussed in Sec. III H1, Wilkinson
161
introduced a parameter dependence in the GOE. Let the Gaussian distributed whitenoise matrices W
nm
(X) have
zero mean values and a second moment given by
W
nm
(X)W
n

m
(X

) = (
nn

mm
+
mn

nm
)(X X

) . (3.166)
A parametric dependence of the random Hamiltonian H can then be dened by convoluting these white noise functions
with a smoothly varying function h(X),
H
nm
(X) =
_
+

W
nm
(X

)h(X X

)dX

. (3.167)
For the second moments of the matrix elements this yields
H
nm
(X)H
n

m
(X

) = (
nn

mm
+
mn

nm
)

f(X X

f(X) =
_
+

h(X X

)h(X

)dX

. (3.168)
The function

f(X) is translation invariant by construction but not necessarily symmetric in X. In studying related
questions in a dierent context, Ko et al.
195
and Brink et al.
196
had earlier introduced distributions dened by an
equation similar to Eq. (3.168). In these cases, however, the function f(X) had an additional dependence on the matrix
indices, leading to a band structure of the matrices. More recently, related studies have been performed by Bulgac et
al.
166
. From the point of view of statistical mechanics, the introduction of the dependence on such a parameter X
in Eqs. (3.166) and (3.168) is tantamount to the introduction of another dimension: The zerodimensional Random
Matrix Theory becomes onedimensional. In Sec. VII A1, we discuss a related important process, the free propagation
of a timedependent random matrix.
A suitable unifying generalization of Wilkinsons ideas was introduced by Alhassid and Attias
193,197
in the form
of the Gaussian random process. Let H(X) denote matrices of dimension N, with volume element d[H(X)], which
depend on the parameter X. For all X, the matrices belong to the same symmetry class labeled by . The Gaussian
Orthogonal, Unitary and Symplectic Processes, GOP, GUP and GSP, respectively, are dened through the probability
density
P(H(X)) exp
_

2v
2
_
dX
_
dX

trH(X)K(X, X

)H(X

)
_
(3.169)
where K(X, X

) is a scalar function. Since the distribution is Gaussian, it is completely determined by the rst two
moments
H
nm
(X) = 0
H
nm
(X)H
n

m
(X

) =
v
2
2
f(X, X

)g
,nmn

m
(3.170)
61
where g
1,nmn

m
= (
nn

mm
+
mn

nm
) in the orthogonal and g
2,nmn

m
= 2
nm

n

m
in the unitary case. Equa-
tion (3.170) bears a close formal relationship to Eq. (3.168). It is assumed, however, that f be both translation
invariant and symmetric, f(X, X

) = f([X X

[), and that f(0) = 1. This implies that K(X, X

) also obeys
K(X, X

) = K([X X

[). The kernel K and the function f are related via their Fourier transforms

K(k) and

f(k)
by

K(k) = 1/

f(k) . (3.171)
All Gaussian processes can be constructed from the elementary Gaussian process dened by the white noise weight
functions
f(X X

) = (X X

) and K(X X

) = (X X

) . (3.172)
The elementary Gaussian process formally includes also the case of a Gaussian white noise potential. To see this, we
interpret the parameter X as the ddimensional position vector and insert the Fourier transform of Eq. (3.171) into
Eq. (3.169) for matrix dimension N = 1.
As an illustration, we present one of the examples given in Refs. 197,193. We consider the OrnsteinUhlenbeck
Gaussian process
198
f(X) = exp([X[) and
K(X X

) =
1
2
_

2
X
2
_
(X X

) (3.173)
where is a positive constant. If the parameter X is identied with time t, the weight functions (3.173) lead to
the action of a freely propagating matrix with quadratic potential, see Eq. (7.6). Naturally, both the physical and
mathematical signicance of X can be very dierent for dierent choices of weight functions. In Refs. 193,197, the
relationship between the OrnsteinUhlenbeck process and Dysons Brownian motion model is also discussed from the
viewpoint of Gaussian processes.
IV. MANYBODY SYSTEMS
We deal with selfbound systems consisting of several or many particles which interact through twobody forces.
In the case of atomic nuclei, the particles are nucleons and in the case of atoms and molecules, they are ions and
electrons. We can cover only a small fraction of the enormous wealth of published material. Atomic nuclei display
considerable complexity and have been studied extensively. This is why we put emphasis on this eld which is
discussed in Sec. IVA, while atoms and molecules are the subject of Sec. IVB. Some general questions arising in the
context of manybody systems are addressed in Sec. IVC.
A. Atomic nuclei
Mediumweight or heavy nuclei are complex manybody systems par excellence
63,199
. A nucleus with mass number
A = N +Z contains N neutrons and Z protons. The size of a nucleon and the average distance between the nucleons
are both about 1 fm. The range of the strong interaction is also of the order of 1 fm. The energy scale is MeV (Mega
electron volts).
As outlined in the Historical Survey, the high complexity of nuclei prompted the introduction of statistical concepts
already in the early days of nuclear physics. This eventually led to Random Matrix Theory. This development
has had enormous impact on statistical manybody physics at large. To prepare the ground for the introduction
of statistical concepts, we present in Sec. IVA1 a very brief introduction into the most essential aspects of nuclear
physics. In Secs. IVA2 and IVA3, we collect, respectively, the evidence for random matrix properties of spectral
uctuations, and data on the distribution of scattering parameters and widths. All theoretical models for the nucleus
are eective, i.e., do not start from the basic nucleonnucleon interaction. Thus, it is important to see whether the
uctuation properties displayed by the data are also borne out in model calculations. This is discussed in Sec. IVA4.
In Sec. IVA5 we summarize some of the work which employs Random Matrix Theory as a tool to obtain upper
bounds or values for the strengths of invariance and/or symmetrybreaking forces.
The are numerous reviews on Random Matrix Theory in nuclear physics. We mention Refs. 9,200,25,201,18,202205.
62
1. Basic features of nuclei
The cross section for the elastic scattering of slow neutrons on heavy nuclei displays sharp resonances with widths
ranging from about 3 meV to 1 eV, and with spacings of about 10 eV, see Fig. 2. These resonances cannot be due to
a single interaction between the incident neutron and a target nucleon. Indeed, a straightforward estimate shows that
the inverse reaction time for such a twobody collision yields a value for the resonance width which is in the MeV
range. The small observed widths correspond to much longer interaction times. This requires that many nucleons in
the target participate in the reaction, and that each resonance has a complex wave function. While it may thus be
dicult to account for the properties of individual resonances, the average of the cross section over many resonances
can be obtained by the following statistical argument. The available energy is assumed to be spread uniformly over
all participating nucleons in such a way that the target nucleus and the incident neutron together form a new system,
the compound nucleus. This system equilibrates and loses memory of its mode of formation before it decays. As a
consequence, formation and decay of the compound nucleus are independent processes. This is the essence of Bohrs
picture
10
of the compound nucleus.
The picture yields a formula for the average cross section
ab
). The incident channel (in our example, the slow
neutron) is denoted by a. The average probability for formation of the compound nucleus from the incident channel
is given by the transmission coecient T
a
. The decay of the compound nucleus into other channels c is likewise
governed by associated transmission coecients T
c
. Depending on the available energy, c may stand for an inelastically
scattered neutron, a proton, an particle, etc. In Bohrs statistical picture,
ab
) is proportional to T
a
G
b
where G
b
is the probability for the decay into channel b. Since the nucleus has to decay into one of the open channels, we
have

b=1
G
b
= 1. On the other hand, G
b
must be proportional to T
b
. Thus, G
b
= T
b
/

c=1
T
c
. This yields the
HauserFeshbach formula
103
,

ab
) = C
kin
T
a
T
b

c=1
T
c
. (4.1)
Kinematic factors have been absorbed in the constant C
kin
.
These stochastic concepts received strong support when in 1966 Ericson
24
showed that the nuclear cross section can
be viewed as a random variable. His ideas apply in the regime where the average resonance width is much larger than
the mean level spacing, i.e. in the regime of strongly overlapping resonances. At any one energy, the cross section
is due to the superposition of many resonances. If each resonance contributes a random amplitude, the cross section
uctuates randomly versus energy. Minima and maxima are not related to individual resonances. Nevertheless, the
cross section is fully reproducible. Therefore, the uctuations do not represent white noise, but internal stochastic
behavior of the compound system.
Prior to Ericsons observation, the discovery of shell structure in nuclear binding energies had led to quite dierent
a picture of the nucleus. At certain values of the neutron and proton numbers N and Z, the nuclear binding energy
displays maxima. As in atomic physics, these maxima can be explained in the framework of a shell model
206
. In
contrast to the atomic case, a strong spinorbit force has to be invoked to describe the data. In the nuclear shell
model, each nucleon moves independently in a mean eld U generated by all other nucleons. The Hamiltonian takes
the form
H =
A

k=1
(T(k) +U(k)) +H
res
H
res
=

k<l
W(k, l)
A

k=1
U(k) . (4.2)
Here, T(k) is the kinetic energy and W(k, l) is the twobody interaction between nucleons k and l. The dierence
H
res
between the true twobody interaction and the mean eld is referred to as the residual interaction. (This
picture is oversimplied and does not account for the strong repulsion between pairs of nucleons.) The model is very
successful in accounting for the lowenergy spectra of many nuclei, particularly when the residual interaction is chosen
phenomenologically. In the space of Slater determinants of shellmodel states, the Hamilton matrix acquires a highly
complex structure, making it a wellsuited object for a statistical approach. In this basis, the matrix elements of
H
res
have a nearly Gaussian distribution. This fact links the shell model to the random matrix approach to nuclear
reactions.
Nuclei are capable of yet another type of motion. This was discovered shortly after the emergence of the shell
model. A nucleus can be viewed as a piece of dense but elastic matter with spherical or ellipsoidal shape. This object
63
can vibrate and, in the case of deformation, rotate. Collective modes, the quantal manifestations of such motion,
occur in many nuclei. The minimum rotational energy E(I) a nucleus must have for a given total spin I denes a
lower bound for its total energy E. In the (E, I) plane, the curve of E(I) versus I has parabolic shape. It is called
the yrast line. This curve denes rotation without inner excitation (cold rotation). Excitations heat up the system.
On the yrast line, the collective motion decays through the emission of electric quadrupole (E2) radiation. A state of
spin I on the yrast line decays into a state of spin I 2, and so on. States connected by a sequence of such strong E2
transitions form a rotational band. Near the yrast line, the members of a rotational band are characterized by the K
quantum number. This is the projection of the intrinsic angular momentum (caused by singleparticle excitations)
onto the symmetry axis of the deformed nucleus. Collective excitations often follow an easily predictable pattern,
leaving little room for the kind of stochastic behavior modeled by random matrices.
Collective excitations at low energy and low angular momentum values are often modeled in terms of the Interacting
Boson Model (IBM)
207
. Due to the strong pairing force in the nuclear Hamiltonian, pairs of protons or of neutrons
couple to bosonic pairs which form the basic ingredients of the model. In its basic version, the model contains Bosons
with spin zero (s Bosons) and Bosons with spin two (d Bosons). For nuclei with even A, the IBM Hamiltonian is
solely expressed in terms of these Bosons. Nucleons do not appear. The Hamiltonian has the form
H = c
0
n
d
+c
2
Q

+c
1

L
2
. (4.3)
Here

L is the total angular momentum operator and c
0
, c
1
and c
2
are parameters. The operator n
d
counts the
number of d Bosons. The quadrupole operator Q

models the deformation of the nucleus. It depends on a parameter

. The Hamiltonian (4.3) possesses a U(6) and an O(3) symmetry, corresponding to the conservation of Boson number
and angular momentum. Three dierent group chains connect U(6) with O(3). Each chain represents a dynamical
symmetry which is realized in the Hamiltonian (4.3) for a certain choice of parameters. Physically, these dynamical
symmetries correspond to vibrational, rotational and deformationinstable nuclei. Every IBM Hamiltonian describing
a real nucleus corresponds to a point in the space of parameters of the IBM. The IBM can be viewed as a mapping
of the space of shellmodel Slater determinants onto a drastically smaller subspace spanned by bosonic wave
functions.
In summary, independentparticle motion with coupling through the residual interaction, collective modes due to
deformation, and a stochastic approach form the main ingredients of nuclear theory. These elements have dierent
relative weights at dierent excitation energies and mass numbers A.
For the statistical analysis, the concept of the spreading width has proven very useful. We consider a nucleus
modeled by a Hamiltonian H = H
0
+ V . The rst part, H
0
, describes the subsystem which is pure with respect
to symmetries, invariances or other properties. The second part, V , is a perturbation which destroys this property.
For example, H
0
conserves timereversal invariance and V breaks it, or, H
0
models purely collective motion while
V contains singleparticle excitations. The inuence of the perturbation can most conveniently be measured if the
eigenbasis of H
0
is taken as the reference frame. As V increases, an eigenstate of H spreads out, starting from an
initial eigenstate of H
0
. The averaged energy scale of this spreading in the spectrum can be shown to be given by the
spreading width
= 2
V
2
)
D
0
(4.4)
where D
0
is the mean level spacing due to H
0
and V
2
) is the mean square perturbation matrix element. The spreading
width should not be confused with the decay width of an open quantum system. It is conceptually closely related to
the Thouless energy which is used in condensed matter physics, see Sec. VI. The spreading width is frequently used
to characterize the strength of a perturbation V . This is done because , in contrast to both the matrix elements of V
and H
0
, depends only very little on excitation energy or mass number. Indeed, D
0
decreases essentially exponentially
with excitation energy E, and the same is true for V
2
) because the complexity of eigenstates of H
0
grows strongly
with E. It can be shown
18
that these exponential dependences cancel in the ratio of Eq. (4.4).
2. Spectral uctuations
Data of suciently high quality to test spectral uctuations in nuclei are available only in restricted energy windows.
It is crucial that sequences of levels of the same spin and parity be complete (no missing levels) and pure (no levels with
wrong quantum numbers). Spectra which contain levels of all spins and parities yield a Poisson distribution
208,209
, see
Sec. III B5. One such window is dened by the scattering of slow neutrons
2,210212
, see Fig. 2, and of protons at the
Coulomb barrier
213
. Earlier than in other areas of physics, the energy resolution and the size of the data set attained
64
FIG. 18. Nearest neighbor spacing distributions and cumulative spacing distributions for all levels in six dierent regions
of the nuclear mass parameter A. The dashed lines show the behavior for the Poisson distribution, and the smooth solid lines
show it for the Wigner surmise. Taken from Ref. 222.
here were sucient for a statistical analysis. In 1982, Haq, Pandey and Bohigas
26
analyzed the set of 1726 nuclear
spacings then available from such data. This nuclear data ensemble is a combination of data on neutron
214,215
and
proton
216
resonances. For neutrons, about 150 to 170 s wave resonances per nucleus located typically about 9 MeV
above the ground state, contribute to the ensemble. For protons, the sequences consist of 60 to 80 levels per nucleus.
After separate unfolding, the sequences were lumped together and analyzed. Remarkable agreement of both spacing
distribution (see Fig. 1) and spectral rigidity with GOE predictions was found. Three and fourlevel correlations
also agree well
217
. Recently, Lombardi, Bohigas and Seligman
84
analyzed the spectral twolevel form factor of the
nuclear data ensemble and applied the GSE test discussed in Sec. III B6. Again, consistence with random matrix
predictions was found.
Another window for a test of spectral uctuations exists in the groundstate domain. This region was rst in-
vestigated by Brody et al.
218
who found some evidence for GOElike behavior. A much more detailed study is due
to von Egidy, Behkami and Schmidt
219,220
. The data set of this analysis consists of 1761 levels from 75 nuclides
between
20
F and
250
Cf. The set is obtained from the analysis of (n, ) reactions. The data are separated into pure
sequences. The combined nearest neighbor spacing distribution lies between a Poisson distribution and the Wigner
surmise. AbulMagd and Weidenm uller
221
divided the data set of Ref. 219 into a number of groups and analyzed
the spacing distribution and the spectral rigidity within each group. It was found that the collective states with spin
J and parity values of J

= 2
+
and 4
+
in rotational nuclei show Poisson characteristics whereas the uctuations
of the remaining levels are much closer to GOE predictions. A more complete study along similar lines is due to
Shriner, Mitchell and von Egidy
222
. These authors analyzed a data set of 988 spacings from 60 nuclides for which
the level schemes are believed to be complete in an energy interval starting with the ground state, and over some
range of angular momentum values. The results show a strong mass dependence. The spacing distributions of the
lighter nuclides are close to GOE while those of the heavier ones are closer to a Poisson distribution, see Fig. 18. This
suggests that the coupling of independent particles by the residual interaction generically leads to GOE behavior, and
that collectivity is the origin of Poisson statistics. Further support for this hypothesis comes from the investigation of
65
spin and shape eects in Ref. 222, although the number of states available was rather small. The spacing distributions
of the 2
+
and 4
+
states in heavier and deformed nuclides (where collectivity is strongest) are close to Poisson while
those of the 0
+
and 3
+
states in the same nuclides and the 2
+
and 4
+
states in spherical nuclides are close to GOE. A
Poisson distribution would be expected for a superposition of states with dierent quantum numbers
23
, see Sec. III B5.
Candidates for such unidentied conserved quantum numbers are isospin T and the K quantum number introduced
in the previous section. However, even a weak breaking of the symmetry associated with T or K quickly leads to
GOE statistics, see Sec. III F3. In the data of Shriner et al. there is a hint towards an interuption of the gradual
transition from WignerDyson to Poisson statistics in the mass region 180 < A 210. AbulMagd and Simbel
223
argued that this could be related to the occurence of oblate deformation in this mass region.
Another window exists in deformed rareearth nuclei with mass numbers A between 155 and 185. In contrast to
the cases discussed above, the data are taken near the yrast line and contain much higher spin values. Garrett and
coworkers
224,225
analyzed 2522 level spacings. Very recently, the analysis was extended
226
and includes now 3130
spacings. The average spacing was calculated for the total ensemble, before dividing it into pure sequences of xed
spin and parity, because the properties of the deformed nuclei in this mass region are known to be similar. The
authors state that, for this and other reasons, the usual method to determine the average spacing for individual pure
sequences can be replaced by this procedure of combining the sequences in the yrast region. The resulting spacing
distribution is the closest to a Poisson distribution yet obtained except for a statistically signicant suppression of
very small spacings. It is argued that this suppression indicates that K is (nearly) a good quantum number for the
lowlying states in deformed rareearth nuclei.
Little is known about the spectral uctuations of individual nuclei. The most complete data sets available are
probably the ones for
26
Al, analyzed by Mitchell et al.
227,228
, and for
116
Sn, analyzed by Raman et al.
229
. The results
for
26
Al are important for the study of symmetry breaking and are discussed in Sec. IVA5. The authors of Ref. 229
present an almost complete level scheme of
116
Sn from the ground state to a maximum excitation energy of 4.3 MeV.
The resulting spacing distribution lies between a Poisson and a Wigner distribution. Bybee et al.
230
applied the
correlationhole method (see Sec. III B4) to spectra of
57
Co and
167
Er. However, these authors conclude, that the
limited sample size of their data yields ambiguous results whereas the spacing distribution and the spectral rigidity
are statistically more signicant.
In Refs. 225,226 it is attempted to draw a coherent picture of the results summarized above. In the work of
Refs. 224,225, the value of the mean level spacing, D = 297 keV, considerably exceeds the typical size of the interaction
matrix elements, V
int
< 35 keV, between the states. Thus, the uctuations are expected to be close to Poisson. For
small spacings s < 70 keV/D, a deviation from the Poisson distribution is seen because here s is comparable with
V
int
/D. The other extreme is provided by the nuclear data ensemble analyzed in Refs. 26,217,84. Here, the mean level
spacing D is much smaller than a typical interaction matrix element, D V
int
. The states are strongly mixed and the
uctuations are GOE. In the case of
116
Sn, the mean level spacing D = 110 keV is comparable with the interaction
matrix element V
int
, causing the spacing distribution to lie between the Poisson and the Wigner distribution. Figure 19
shows the three windows in which the aforementioned data were taken. This interpretation suggests a statistical model
of the type H
(0)
+H
(1)
where H
(0)
and H
(1)
model Poisson and GOE uctuations, respectively, see Sec. III F. The
parameter = /D is of the order of V
int
/D.
3. Resonances and scattering
Aside from the resonance positions, data for resonance reactions also yield partial and total widths. RMT predicts
the distribution of partial widths to be of PorterThomas form, see Sec. III E1. Prior to a comparison of data with this
prediction, it is necessary to remove the penetration factors due to the Coulomb and/or angular momentum barriers
from the measured partial widths. Early experimental evidence for the validity of the PorterThomas distribution
for the resulting reduced widths is summarized in Refs. 9 and 200. Later, the distributions of reduced widths
for neutron
215
and proton
231
resonance data were shown to be consistent with the PorterThomas law. However,
Harney
232
pointed out that a statistically signicant test requires many more data than were available at the time.
Therefore, Shriner, Mitchell and Bilpuch
233,234
used another statistical observable. The PorterThomas distribution
for the reduced widths results from the Gaussian distribution for the reduced width amplitudes, see Sec. III E1, and
that latter distribution is tested. With n the running index of a set of resonances, the reduced width amplitudes
na
and
nb
for two dierent channels a and b are taken as two data sets. If two sets x
1
, . . . , x
N
and y
1
, . . . , y
N
have
a Gaussian distribution, the linear correlation coecient
(x, y) =

N
n=1
(x
n
x)) (y
n
y))
_

N
n=1
(x
n
x))
2

N
n=1
(y
n
y))
2
(4.5)
66
FIG. 19. The range of excitation energies Ex and angular momenta I associated with the three data sets mentioned in
the text. The point in this diagram from which the levels for the nuclear data ensemble are taken is labeled with Haq.
The shaded area labeled
116
Sn indicates the area in the plane from which the data for the analysis of Raman et al.
229
were
gathered. The shaded area designated present work refers to the analysis of Garrett et al.
224226
. The yrast line is shown as
a solid line. Taken from Ref. 226.
has the property (x
2
, y
2
) =
2
(x, y). This test is far more sensitive than a comparison with the PorterThomas
law because it involves the joint moments of the partial width amplitudes
na
and
nb
and, thus, the relative signs
of the amplitudes. This information was provided by the analysis of interference phenomena in the exit channels in
protoninduced reactions. The analysis yields a statistically signicant conrmation of GOE predictions.
The successful analysis of the distribution of the positions and widths of wellresolved isolated resonances in terms of
the GOE implies that compound nucleus scattering itself is a stochastic process and can be modeled by the stochastic
scattering problem introduced in Sec. III D. It is assumed that this modeling applies not only in the domain of isolated
resonances but also at higher energies where resonances overlap. Apart from kinematic factors, the compound nucleus
cross section is then given by the square modulus of the scattering matrix in Eq. (3.94) of Sec. III D, and the uctuation
properties of compound nucleus cross sections can be predicted. The history of this problem is mirrored in the reviews
in Refs. 200,25,18. For a long time, only perturbative solutions were available, valid either for small or for large values
of the ratio /D. Here, is the average total width of and D the mean spacing between resonances. Use of the
supersymmetry method
46
made it possible to work out the uctuations of the scattering matrix in all regimes, from
isolated resonances to the Ericson regime of strongly overlapping resonances, see Sec. III D2. The result is also of
practical use. An example is the application to reactor physics at low energies, see Ref. 235. The leading term of
an expansion of the result in inverse powers of

c=1
T
c
yields the HauserFeshbach formula (4.1). Nishioka and
Weidenm uller
236
showed that direct reactions can easily be included in the formalism.
The use of the GOE in this model for compound nucleus scattering implies that the internal equilibration time
of the nucleus is small in comparison with the decay time h/. This is true at low incident energies. However, in
reactions induced by light ions, h/ decreases strongly with energy and becomes comparable to the equilibration
time for incident energies higher than 10 MeV or so. Then, the compound nucleus can decay before it is completely
equilibrated, and the model of Sec. III D2 must be extended to account for this fact. This is done by using the nuclear
shell model and dividing the compound system into classes of states. Each class comprises states with xed particle
hole number and is represented by a random matrix. Owing to the twobody character of the residual interaction, only
neighboring classes are coupled. This coupling is also modeled by random matrices. The total Hamiltonian is then a
band matrix whose entries are random matrices. Following earlier work by Agassi, Weidenm uller and Mantzouranis
237
and Feshbach, Kerman and Koonin
238
, Nishioka et al.
239
succeeded in calculating the average cross section for this
model. They used the supersymmetry method and a loop expansion
240
. The strong breaking of rotation invariance in
Hilbert space of this random matrix model amounts to the introduction of a dimension. This is why this model later
became important for mesoscopic systems, see Sec. VI C3. The ideas used in the model have been tested perhaps
67
most stringently in nuclear reactions with isospin mixing, see Sec. IVA5.
Another approach to compound nucleus scattering considers the scattering matrix itself as the basic random entity
and uses the maximum entropy approach to determine the uctuation properties, see Sec. III D2. The random transfer
matrix technique used in mesoscopic physics (see Sec. VI C2) can be viewed as an extension of this approach.
4. Model systems
As mentioned above, it is important to test whether spectral uctuation properties predicted by various nuclear
models correctly reproduce the properties of nuclei reviewed in the last two sections. This is true, in particular,
because all models of the nuclear manybody problem are eective models based on meaneld approaches.
There are numerous studies of spectral uctuation properties within the nuclear shell model. Early work is reviewed
in Ref. 18. We only mention the studies of Soyeur and Zuker
241
and of Wong
242
which date back to the early seventies.
The former authors found good agreement of the spin J = 2 states in
24
Mg with the Wigner surmise. Wong calculated
spectra including several J values whose spacing distributions follow the Poisson law. We can also mention only a few
of the recent results. Ormand and Broglia
243
performed realistic shellmodel calculations for light nuclei in the mass
region 20 < A < 35 and for
46
V. Very good agreement of the spacing distribution and the spectral rigidity with the
GOE prediction was found in the region near the ground state. However, in these calculations isospin conservation
was assumed, see Sec. IVA5.
The inuence of the mean eld was investigated by Dro zd z et al.
244
in the spectrum of twoparticle twohole
states. As expected on general grounds, there is no level repulsion within a pure meaneld approach. GOE type
uctuations occur when the residual interaction is switched on. Level repulsion arises from particlehole rescattering
eects. Zelevinsky et al.
205,245
calculated spectra of light nuclei as a function of the strength of the residual interaction.
A relatively weak interaction suces to produce a Wigner type spacing distribution. This is agreement with the RMT
expectations, see Sec. III F.
The distribution of the amplitudes of shellmodel wave functions in a xed basis has also been compared to the GOE
prediction
18
. In the groundstate domain, Whitehead et al.
246
and Verbaarschot and Brussaard
247
found deviations
from the Gaussian behavior predicted by the GOE. The singleparticle states within a major shell are not completely
degenerate. The splitting of the singleparticle energies counteracts the mixing of states by the residual interaction.
Interestingly, Whitehead et al.
246
set up a random matrix with a block structure in the spirit of the Rosenzweig
Porter model to be discussed in Sec. IVB1. Brown and Bertsch
248
showed that the distribution of the amplitudes
becomes Gaussian at higher excitations where the level density is larger. It was also argued by Zelevinsky et al.
249
that calculations with degenerate singleparticle energies yield GOE features at lower energies than models which are
believed to describe the nucleus in a more realistic fashion. A discussion can be found in Zelevinskys review
205
. The
decay in time of the compound nucleus has also been the object of several studies. For many (few) open channels,
RMT predicts exponential (algebraic) decay
99,250,251
, respectively. This was compared with model calculations in
various regimes
252,253
.
The interplay between collective motion and singleparticle eects can ideally be studied in hot rotating nuclei, cf.
Sec. IVA1. The results reviewed in Sec. IVA2 suggest that collective motion is likely to yield Poisson or harmonic
oscillator statistics whereas singleparticle eects should lead to GOE behavior. Reviews were given by

Aberg
254
and
Dssing et al.
255
. Rotational bands exist not only near the yrast line (see Sec. IVA1), but also at excitation energies
of a few MeV above it. At these energies, the rotational bands are mixed by the residual interaction. Thus, a state
with spin I does not decay into a single state of spin I 2 within the same band, but into several or even many states
with spin I 2. This phenomenon is referred to as rotational damping and is characterized by a spreading width, see
Sec. IVA1. Experimental and theoretical evidence for rotational damping was given by Leander
256
as early as 1982.
Numerous investigations have since been added. We only mention the work of Lauritzen, Dssing and Broglia
257
.
The picture was conrmed experimentally
258,259
. Guhr and Weidenm uller
143
interpreted rotational damping in terms
of the random matrix model H
(0)
+H
(1)
discussed in Sec. III F. The collective part, modeled by H
(0)
, is chosen to
have a Poisson or harmonic oscillator spectrum. The mixing is modeled by H
(1)
which is drawn from the GOE. The
inuence of the perturbation rises with excitation energy.

Aberg
260
addressed the same issue in a still schematic but
more realistic model and obtained similar results. Person and

Aberg
144
and Mizutori and

Aberg
145
studied a random
matrix model based on

Abergs approach. They identied a critical scale L
max
in the energy spectrum. Below this
scale, the longrange correlations of levels are of GOE type and above it, they become Poissonlike. This length
scale is related to the spreading width or energy localization length of the wave functions. The problem is similar
to problems in disordered systems, see Sec. VI F. In the random matrix model H
(0)
+ H
(1)
with H
(1)
drawn from
the GUE, Guhr and M ullerGroeling
261
studied such a critical length analytically. Dssing et al.
255,262
discussed
microscopic models of rotational damping and compared with GOE predictions.
68
Various studies address singleparticle motion in deformed potentials. This motion may itself be chaotic. Of the
many works on this subject, we mention, inter alia, the early papers by Arvieu et al.
263,264
which focus on classical
and semiclassical aspects. Milek, N orenberg and Rozmej
265
studied a twocenter shell model. Such a model is used in
the theory of nuclear ssion. It describes the separation of a spherical nucleus via deformation and formation of a neck
into two separated spherical nuclei. Poissonlike spacing distributions were found in the two limiting cases. When a
neck is formed and rotational symmetry is broken, the spacing distribution agrees with the Wigner surmise. Heiss,
Nazmitdinov and Radu
266,267
investigated an axially symmetric potential of quadrupole shape. Such a potential is
used to model deformed nuclei. They found that the addition of an octupole deformation can render the motion
chaotic. In the case of a very strong deformation, the spacing distribution is quite close to the Wigner surmise. In
similar work, Heiss and Nazmitdinov
268
studied other deformed potentials which are often used to describe deformed
nuclei and metallic clusters. Such and other studies of singleparticle motion cannot answer the question whether the
nuclear manybody problem is chaotic. In fact, the inuence of the singleparticle motion on chaotic features of nuclei
is not yet understood, particularly since the residual interaction is known to mix the singleparticle states already at
rather low excitation energies. We must also remember that shell structure, one of the central elements of theoretical
nuclear physics, is due to regular motion. We mention in passing that the study of shell eects
269,270
led Strutinsky
and Magner
271
as early as 1976 to an interpretation of shell structure in terms of periodic orbits, a fact little known
outside the nuclear physics community.
Certain features of the Interacting Boson Model (IBM), see Sec. IVA1, which can be related to chaotic motion,
have been studied intensely by Alhassid, Novoselsky and Whelan
272
and Alhassid and Whelan
273,274
. A semiclassical
approximation to the Hamiltonian H dened in Eq. (4.3) is used to dene the classical limit, and to check whether
the classical dynamics is chaotic. The inverse number of Bosons plays the role of h. The construction uses coherent
states. We recall that H depends on a number of parameters. For those values where one of the three group chains
applies, the classical motion is regular, and the spectral uctuations are close to Poisson. For values of the parameters
where the dynamical symmetries are broken, there is a tendency toward chaotic classical motion and GOE statistics.
Remarkably, a regular strip was found within this domain which probably corresponds to a previously unknown
approximate symmetry. However, the coordinates and momenta of the semiclassical approximation to H do not
represent in any direct way the spatial coordinates and momenta of the shellmodel Hamiltonian which the IBM
Hamiltonian is supposed to approximate. Thus, no conclusion can be drawn as to whether the nuclear manybody
problem is regular or chaotic. Lopac, Brant and Paar
275
investigated numerically the Interacting Boson Fermion
Model. Here, the odd nucleon is coupled to the Bosons of the IBM. Analyzing the spectral rigidity, they found
uctuation properties between Poisson and GOE.
5. Invariances and symmetries
In previous sections, we summarized some of the evidence showing that in certain domains of excitation energy and
mass number, spectral uctuation properties of nuclei agree with RMT. We also showed that these random matrix
features are consistent with certain nuclear models. We now use these results as tools to address the breaking of
symmetry and/or invariance. We focus on a domain of excitation energies and mass number where RMT concepts do
apply, and we address the breaking of timereversal invariance, of the isospin quantum number, and of parity. The
important facts are: (i) In this domain, there is a strong enhancement of symmetry breaking. Indeed, a noticeable
eect appears whenever the relevant matrix element is of the order of the mean level spacing D, in agreement with
the general consideration of Sec. III F3. We recall that at neutron threshold, D 10 eV in heavy nuclei, whereas
D 100 keV near the ground state. (ii) In this domain, the nuclear wave functions are stochastic. Meaningful
information on the strength of the symmetrybreaking interaction can only be gained by determining the rootmean
square (rms) matrix element from a large number of measurements. This is done under the assumption that the
uctuations are described by RMT. There is a bonus to this approach: By denition, the rms matrix element does
not depend on details of the wave functions of the individual states. In the analysis, analytical and numerical results
on crossover transitions mentioned in Sec. III F are used.
The breaking of timereversal invariance is modeled as a crossover transition from GOE to GUE. Two tests have been
used to obtain an upper bound on the violation of this invariance. (i) In the nearest neighbor spacing distribution,
the slope for small s changes from linear to quadratic, and the level number variance changes accordingly, as the
GOE is replaced by the GUE. Absence of this eect provides a test. French et al.
276,133
employed a perturbative
calculation of the level number variance
2
(L) valid for small admixtures of the GUE. This is physically justied
because the breaking of timereversal invariance is expected to be weak. They used the nuclear data ensemble
discussed in Sec. IVA1, see Ref. 26. The value of the level number variance at L = 1 is plotted as a function of the
parameter measuring the invariancebreaking GUE admixture and compared with the data, and an upper bound for
69
0.4
0.3
0.2
0.1
0.0
0 5 10 15
L

3

(
L
)
FIG. 20. Experimental spectral rigidity 3(L) in
26
Al. Solid dots represent the experimental results of the 100 levels
between 0 and 8 MeV excitation energy. The solid line is the result of a numerical simulation of the random matrix model
incorporating a symmetry breaking. This value of the symmetry breaking gave the best t to the experimental data. For
comparison the theoretical results for no symmetry and one fully conserved symmetry are shown as lower and upper dashed
lines, respectively. Adapted from Ref. 137.
timereversal invariance breaking is found. (ii) The principle of detailed balance is tested in the regime of Ericson
uctuations. The cross section for a nuclear reaction and its inverse are compared. The average cross section is
timereversal symmetric, and the test applies to the uctuations of the cross section. Blanke et al.
277
measured the
cross section of the reaction
24
Mg +
27
Al +p and its inverse. Boose, Harney and Weidenm uller
278,279
used the
random matrix model for the GOE to GUE transition within the random Smatrix approach described in Sec. VI C3.
The observable is the correlation function
ab

ba
)/
ab
)
ba
) relating the cross sections of the two reactions.
The analysis yielded an upper bound for the spreading width of 10
2
eV. This is consistent with the value found
in the analysis of French et al.
276,133
. From this value an upper bound for the relative strength of the timereversal
invariance breaking interaction in the shellmodel Hamiltonian of 10
3
can be derived.
As shown in Sec. III F, the breaking of symmetries (which are always associated with quantum numbers) involves
a structure of the Hamiltonian matrices which diers from the one for timereversal violation. Nevertheless, certain
features, particularly the remarkable sensitivity of the uctuations to the crossover transition, are the same. In nuclei,
isospin is only an approximate symmetry because the Coulomb interaction, although much weaker than the nuclear
force, is not completely negligible. Harney, Richter and Weidenm uller
30
analyzed isospin breaking in compound
nucleus reactions by comparing the data with analytical calculations based on a random matrix model mentioned in
Sec. III F. The strength of symmetry breaking is measured in terms of the spreading width introduced in Sec. IVA1.
Over the entire mass range 25 < A < 240 and for all energies investigated, the spreading widths for isospin mixing
resulting from this statistical analysis lie in a narrow band between 1 keV and 100 keV. Mitchell et al.
227,228
tested
isospin violation in the spectrum of a single nucleus,
26
Al. In this nucleus a nearly complete scheme of about
100 levels is available ranging from the ground state up to 8 MeV. The levels are classied by spin, parity, and
isospin
280282
. Of these 100 levels, 75 have isospin T = 0 and 25 have isospin T = 1. Mitchell et al. obtained a spacing
distribution and a spectral rigidity which lie between the two limiting RMT cases of perfectly conserved and fully
broken isospin. Guhr and Weidenm uller
137
compared the spectral rigidity found experimentally in Ref. 227 with a
numerical simulation of a fulledged random matrix model, see Fig. 20. Within relatively large errors, the resulting
rms Coulomb matrix element of about 20 keV agrees with the data mentioned above. In Ref. 228 a slightly dierent
experimental analysis was presented. An interpretation in the framework of the RMT model for symmetry breaking
is still possible. Due to the relatively large errors, the resulting rms Coulomb matrix did not change signicantly.
However, improved experimental data are highly desirable. Extensive shellmodel calculations by Endt et al.
282
and
Ormand and Broglia
243
which included the Coulomb interaction, yielded good agreement with both experiment and
70
the statistical analysis. Very recently, Adams et al.
283
calculated the distribution of reduced transition probabilities
in
22
Na using the shell model. Magnetic dipole (M1) and electric quadrupole (E2) transitions where studied among
the rst 25 states with positive parity, spin values ranging from J = 0 to J = 8, and isospin values T = 0 and
T = 1. The results are consistent with a PorterThomas distribution. Remarkably, no dependence on spin, isospin,
electromagnetic character or excitation energy was found.
The violation of parity conservation is modeled in the same way as isospin symmetry breaking. Naively, one might
expect a very weak eect: The weak interaction is about 10
7
times weaker than the strong interaction. Whereas
the spreading width of nuclear states due to the strong interaction is of the order of MeV, the spreading width for
the weak interaction is, therefore, expected to be of the order of 10
7
eV. There are, however, two mechanisms which
strongly enhance parity violation in the regime of isolated compound nucleus resonances
284,285
. First, in this regime
resonances of opposite parity have typical spacings D of a few eV. A spreading width of 10
7
eV thus implies a rms
weak interaction matrix element V
weak
of about 10
4
eV and a ratio V
weak
/D of about 10
4
. This value is much larger
than the naive estimate 10
7
and yields an enhancement factor of about 10
3
. Second, if an incident neutron populates
a p wave resonance which by parity violation is mixed with an s wave resonance, decay of the latter by s wave neutron
emission is not hindered by the angular momentum barrier. Comparing this decay with the decay of the p wave
resonance back into the entrance channel, one obtains a second enhancement factor which is also about 10
3
. It is
therefore not surprising that signals for parity violation amounting to several per cent were subsequently observed
286
in the helicitydependence of neutron transmission through several p wave resonances. A statistical analysis requires
many such data, however. The TRIPLE collaboration has furnished and analyzed such a data base, see Refs. 287,288.
The spreading width deduced from the data is of the order of 10
7
eV, in agreement with theoretical expectations.
We refer to a recent theoretical review on parity and timereversal violation by Flambaum and Gribakin
289
.
B. Atoms and molecules
Complex atoms and polyatomic molecules are governed by the Coulomb interaction. The relevant energy scale is
very dierent from the nuclear case: The ionization energies for atoms and the dissociation energies for molecules are
of the order of several electron volts. In Secs. IVB1 and IVB2, we discuss atoms and molecules, respectively.
1. Atoms
It is not clear a priori whether atomic spectra are expected to show random matrix uctuations. Complex atoms
dier from nuclei in various ways. The nuclear charge Ze is concentrated in the center. The meaneld approximation,
the basis of the atomic shell model, is better justied and a better approximation than in the nuclear case. Nevertheless,
the screened interaction between the electrons is not negligible. But is it strong enough to induce RMT behavior?
Rosenzweig and Porter
82
seemingly were the rst to investigate the uctuation properties of spectra of complex atoms.
This seminal work was done as early as 1960, only a few years after the rst evidence for random matrix uctuations
in nuclear spectra had been found, see Porters review
9
.
Rosenzweig and Porter analyzed spectra of neutral atoms, members of three chains. The rst chain consists of
the atoms Sc, Ti, V, Cr, Mn, Fe, Co, Ni with Z = 21, 22, . . . , 28, the second one of Y, Zr, Nb, Mo, Tc, Ru, Rh, Pd
with Z = 39, 40, . . . , 46 and the third one of Lu, Hf, Ta, W, Re, Os, Ir, Pt with Z = 71, 72, . . . , 78. The levels were
grouped according to parity and total angular momentum J. The angular momentum values ranged from J = 0 to
J = 17/2. No data were available for Tc, Lu and Pt. In the three chains, there was a total of between 1000 and 2000
spacings for each parity. For each pair of (J, ) values, the spacing distributions were analyzed in each of the three
chains. For the odd parity levels, the rst chain gave a spacing distribution close to Poisson, the third one a spacing
distribution close to the Wigner surmise, and the second one an intermediate result, see Fig. 21. To explain their
results, Rosenzweig and Porter used arguments related to the superposition of spectra, see Sec. III B5. In the three
chains, the subshells 3d, 4d and 5d, respectively, are lled. The chains dier in the strength of the spinorbit force.
In the rst long chain, this force is very weak, and the RussellSaunders or LS coupling scheme applies. For larger Z
values, the spinorbit force becomes stronger and the coupling scheme changes. In the second long chain, LS coupling
gives a relatively poor but still acceptable description. In the third chain, the coupling scheme is very dierent. In
the rst chain, L and S can be viewed as almost good quantum numbers. The spacing distribution for a given J is a
superposition of as many nearly independent spectra as there are (L, S) pairs for that J, and the spacing distribution
is very close to Poisson. This interpretation is strongly supported by an analysis which separates levels also by L and
S quantum numbers. The ensuing spacing distribution is very close to the Wigner surmise. In the case of the third
chain, there are no nearly conserved symmetries, and the spacing distribution is Wigner. In the intermediate case of
71
FIG. 21. Nearest neighbor spacing distribution for the odd parity levels in the rst, second and third chain (histograms (a),
(b) and (c), respectively). Separate distributions were constructed for the J sequences of each element and then the results
were combined. For comparison, the Poisson distribution and the Wigner surmise are also drawn. Taken from Ref. 82.
72
the second chain, the symmetries are not completely destroyed yet. The picture diers for levels with even parity.
For various reasons, the even parity spectra are simpler and closer to the ground state. To support their argument,
Rosenzweig and Porter set up a random matrix model. For xed J, they used a blockdiagonal Hamiltonian. Each
block represents a possible (L, S) pair and is modeled by a GOE. The remaining matrix elements are independent
Gaussian distributed random variables with rms value . The spacing distribution was calculated numerically. As
increases, the spacing distribution changes from Poisson to Wigner. The relevant parameter is N
2
, where N is the
matrix dimension. It was interpreted as the rms symmetrybreaking strength normalized to the mean level spacing,
see Sec. III F.
We have discussed this work in some detail because in many ways it was a truly pioneering contribution. Aside
from its technical aspects which later found application in many other papers, this work and a preliminary study
by the same authors constitutes the rst application of RMT outside the eld of nuclear physics. It provided
the rst strong evidence for the universal applicability of RMT, independently of the specic interaction governing
the physical system. In 1961, only a year later, Trees
290
extended this work by analyzing spacing distributions of
calculated spectra.
In 1983, Camarda and Georgopulos
291
analyzed new level schemes of atoms in the beginning of the third chain, from
La with Z = 57 to Lu with Z = 71, using the spectral rigidity and the covariance of adjacent spacings as observables.
(For each pair of consecutive spacings, the latter is given by the expression (4.5)). This test of longrange correlations
is also consistent with GOE predictions.
Of the more recent work, we only mention the extensive theoretical study of the rareearth atom Ce by Flambaum
et al.
292
. A realistic model is used to calculate the structure of complex atomic states. Above 1 eV excitation energy,
these states acquire a structure similar to that of compound states in heavy nuclei (a linear superposition of very
many basis states). Various spectral observables were worked out in this model and compared with GOE predictions.
2. Molecules
In a polyatomic molecule, there are three types of excitations: Electronic excitations, vibrations, and rotations.
The energy scale of the electronic excitations is a few electron volt. Vibrational states built on top of every electronic
state have an energy scale of 0.1 eV. On top of the vibrational states, there are rotational states with an energy scale
of 0.1 meV. There is no spherical symmetry. The spectra are characterized by quantum numbers which dier from
the ones used in atomic physics. Modern laser spectroscopy is capable of resolving the enormously rich and complex
spectra of such molecules even at very high level density. Traditional theoretical techniques often fail in this regime,
a description of individual levels is out of the question, and an analysis of the data with the aid of statistical concepts
is called for.
The molecule NO
2
provides a particularly interesting example and has played a prominent role in the application
of RMT concepts. Between the ground state and the dissociation threshold at about 3 eV, there are four electronic
states. The BornOppenheimer approximation (based on an adiabatic separation of nuclear and electronic motion)
breaks down in certain parts of the spectrum. It is this feature which makes NO
2
such an interesting object of
experimental and theoretical studies of spectral uctuations, although the molecule consists of only three atoms.
In 1983, Haller, Koppel and Cederbaum
293
studied experimental and theoretical vibronic spectra of the molecule
NO
2
and of the acetylene kation C
2
H
+
4
in a regime where the BornOppenheimer approximation fails. Here, the
term vibronic means that the excitation mechanism is only vibrational and electronic. Rotations are absent. Good
evidence for GOE uctuations was found in both the short and longrange spectral observables. Zimmermann
et al.
294
extended the analysis of NO
2
data, including a comparison of the measured intensity distribution with
the PorterThomas law. Theoretical studies were presented by Zimmermann, Koppel and Cederbaum
295
. Leitner,
K oppel and Cederbaum
296
compared the spectral statistics and the classical dynamics of both NO
2
and C
2
H
+
4
. The
dynamics of those systems with coupled potential surfaces can only approximately be described in classical terms
since the electronic degrees of freedoms remain quantized. However, it turns out that the phase space is dominated by
one of the potential surfaces. This allowed the authors of Ref. 296 to empirically establish a correspondence between
this classical dynamics and the spectral uctuation properties. In 1985, Mukamel, Sue and Pandey
297
investigated
the spectral uctuations and the distribution of resonance intensities in data on vibrationally highly excited acetylene.
Deviations from the GOE predictions are attributed to missing levels. Sundberg et al.
298
extended this analysis of
acetylene spectra.
It is apparent from this discussion that NO
2
has played a very important role in the investigation of spectral
uctuations in molecular physics. This was possible because the Grenoble group succeeded in improving the quality
of experimental data substantially. Georges, Delon and Jost
299
have given a detailed account of the measurements on
NO
2
. They summarize the evidence for GOE uctuations obtained from short and longrange spectral observables
73
FIG. 22. Correlationhole analysis for a stimulated emission pumping spectrum of acetylene (left) and for an anticrossing
spectrum of methylglyoxal (right). The properly smoothed decay function dened in Sec. III B4 is shown. The variable t/
should be identied with the variable t of Sec. III B4. The hole is due to the correlation of the level positions while the
exponential decay is due to the resonance shapes. Notice the drastically dierent widths of the two holes, indicating a very
dierent number of symmetries. Taken from Ref. 74.
as follows. The vibronic states built on the rst and the second electronic state show GOE uctuations. This can
be attributed to the mixture of these states due the breakdown of the BornOppenheimer approximation. Higher
up in the spectrum, GOE uctuations occur for another reason in rovibronic states, i.e. in states where rotations
on top of vibronic states become important. The mixing of these states is due to a symmetry breaking which has
specic molecular physics causes. Very recently, parametric correlations as discussed in Sec. III H were addressed
experimentally and theoretically in NO
2
by Nygard, Delon and Jost
300
.
The correlationhole method (Sec. III B4) introduced by Leviandier et al.
74
brought a conceptually new aspect into
the analysis of spectral uctuations, see also the review
301
. This technique is designed for the analysis of raw spectra.
It is less sensitive to the eect of missing levels than other observables. It does not require a linebyline identication
of the resonances. Hence, it can deal with the enormous amount of data supplied by modern laser spectroscopy. The
correlationhole method eectively separates the statistical uctuations of the level positions and of the resonance
shapes, see Fig. 22. Leviandier et al. applied this technique to anticrossing spectra of methylglyoxal. Consider a singlet
state [s) populated from the singlet ground state [s
0
). A triplet state [t) couples to [s) by some interaction V
st
. If the
energy dierence between the two states obeys [E
t
E
s
[ V
st
, the interaction can be neglected and the resonance
uorescence light back into the ground state [s
0
) attains a maximum. If, however, [E
t
E
s
[ and V
st
are comparable,
the admixture of the triplet state leads to a suppression of the uorescence yield, since transitions from [t) to [s
0
) are
forbidden. The energy of the triplet states is shifted by a magnetic eld B, and the resonance uorescence yield I(B)
is measured as a function of B. For each isolated triplet state, I(B) has Lorentzian shape. In methylglyoxal, there are
many such triplet states with overlapping Lorentzians, and the measured resonance uorescence intensity I(B) has a
complicated dependence on B. Leviandier et al. found a deep correlation hole in the decay function constructed from
I(B). The uctuations of the anticrossing spectrum must therefore at least to some extent reect GOE behavior.
The exponential fallo of the decay function is due to the Lorentzian shapes of the individual resonances. The
arguments presented in Sec. III B4 were used to conclude that the measured I(B) must be a superposition of several
noninteracting spectra. Further experimental work is due to Delon, Jost and Lombardi
75
. Pique et al.
302
applied
the same type of analysis to spectra on acetylene. In the framework of scattering theory, a fulledged random
matrix model simulating these experiments was studied analytically
77
and numerically
80
. Detailed accounts of the
correlationhole method which also include other aspects of the statistical analysis were given by Lombardi et al.
78
and by Lombardi and Seligman
79
.
The distribution of decay rates and related questions were investigated theoretically by Zimmermann, Koppel and
Cederbaum
295
, by Miller et al.
303
and by Peskin, Reisler and Miller
304
, cf. also the work by Persson, Gorin and
Rotter
253
, see Sec. IVA4.
Davis
305
introduced the interesting concept of hierarchical analysis: The structure of peaks in a spectrum is studied
as a function of the resolution. In this way, hierarchical trees are generated which correspond to states of ever higher
74
complexity. This analysis is intimately related to the constraints imposed by the experimental setup and can yield
signicant statistical information. It seems worthwhile to see whether links can be established to RMT.
C. General aspects
This short presentation of the evidence for random matrix uctuations in manybody systems would be incomplete
without a brief discussion of two questions posed by the results: (i) To what extent do the Gaussian ensembles properly
model the properties of manybody systems? This question, asked in Sec. IVC1, addresses the very foundations of
the application of Random Matrix Theory to real systems. (ii) Are manybody systems chaotic? This question
connects to the discussion of chaos in Sec. V and is treated in Sec. IVC2.
1. To what extent do the Gaussian ensembles model manybody systems properly?
We recall that the distribution P(H) for the Gaussian ensembles is uniquely derived from two assumptions: (i) P(H)
is invariant under orthogonal, unitary or symplectic transformations, respectively, and (ii) the independent matrix
elements are uncorrelated random variables. There is evidence that both in atoms and in nuclei, the distribution
of matrix elements diers from what is assumed in the Gaussian ensembles. The key argument is based on the
observation that in both systems, the residual interaction is of twobody type. Thus, in a shellmodel basis, the
Hamiltonian matrix is sparse. Put dierently, in an arbitrary basis, the elements of the Hamiltonian matrix are
strongly correlated. In contradistinction, the Gaussian ensembles may be said to contain matrix elements of kbody
type, with k ranging all the way up to N. This is because in these ensembles, all matrix elements not connected by
symmetry are uncorrelated. To what extend does this dierence aect spectral uctuations? The question can be
formulated more precisely as follows. We consider a system of Fermions with a twobody
306,307
and, more generally,
a kbody random interaction. The matrix elements of the kbody operator between kparticle states are assumed to
be independent Gaussian distributed random variables. There is evidence from nuclear physics
18
that for k = 2, this
is a realistic assumption. There are m Fermions which occupy N degenerate singleparticle levels. Here, k m N.
This denes the embedded Gaussian kbody ensembles
308
. Which are the spectral uctuation properties of these
ensembles? Very little is known analytically. This is why the embedded ensembles receive scarce attention in the
present review. Numerical simulations
18
have shown, however, that the local uctuation properties of the embedded
twobody random ensembles do agree with RMT predictions.
The Gaussian ensembles do not only fail to take account of the twobody character of the residual interaction.
They also misrepresent the actual distribution of the matrix elements in a nonshellmodel basis. This is shown by
work of Flambaum et al.
292
who investigated the distribution of the nondiagonal matrix elements of the Hamiltonian
for the J = 4 states of odd and even parity in Ce. The basis is given by all possible singledeterminant states which
are rst constructed for a given value of the total azimuthal quantum number M of the atom and then rotated such
that the Hamiltonian is block diagonal in the total angular momentum J. These states are labeled by an index n.
For m ,= n, Flambaum et al. found the empirical law
P(H
nm
) = C
exp ([H
nm
[/V )
_
[H
nm
[
(4.6)
where C and V are constants. For m n, a peak occurs on top of this distribution. Nonetheless, the spectral
uctuations calculated/measured in this atom agree with the GOE.
In spite of these shortcomings, the Gaussian ensembles generically model the uctuation properties correctly. The
reason is that uctuations are studied on the unfolded, i.e. local, scale dened by the mean level spacing. This is in
contrast to average or global properties which cannot be modeled by these ensembles. The local uctuation properties
are very insensitive to details of the distribution P(H). For a class of ensembles, this is shown in Sec. VIII. More can
be found in the review by Brody et al.
18
. This issues have received renewed interest in recent years, cf. the review by
Flambaum and Gribakin
289
and the work of M uller and Harney
309,310
and Granzow, Harney and Kalka
311
.
2. Are manybody systems chaotic?
We have presented evidence to the eect that the complex structure of the manybody system often yields uctuation
properties which are consistent with RMT. There is also evidence for Poissontype uctuations, especially in systems
75
capable of collective motion. In the description of these phenomena, we have tried to avoid the terms chaotic and
regular, respectively, for these two scenarios. It is shown in Sec. V that for systems with few degrees of freedom,
there is strong and growing evidence for the correctness of the Bohigas conjecture. As stated in the Introduction, this
conjecture links classical chaos to the spectral uctuations described by the Gaussian ensembles. Does this conjecture
also apply to manybody systems? We are not aware of any arguments which would establish a link between many
body systems and classical chaos. Various studies have shown chaotic features of singleparticle motion in realistic
potentials, see Sec. IVA4. Therefore, it is intriguing and may even be justied to think of complex manybody
systems such as nuclei, atoms or molecules as chaotic. Such thinking must come to terms, however, with the existence
of regular collective and singleparticle properties in the same systems which may show up as gross structures at the
very energies where the underlying uctuations are of GOE type. Thus, a theoretical understanding of why RMT is
so successful in manybody systems has yet to be found. The evidence we have at hand is either experimental or
phenomenological. RMT was developed and used in the context of manybody physics long before the connection
to classical chaos in fewdegreesoffreedom systems had been suggested. Thus, in retrospective, it is no surprise
that phrases such as unreasonable eectiveness were used to characterize the predictive power of Random Matrix
Theory in manybody physics.
V. QUANTUM CHAOS
RMT originated in nuclear physics and was conceived as a statistical approach to systems with many degrees
of freedom. However, RMT applies also to fewdegreesoffreedom systems with chaotic classical dynamics. This
observation is of recent origin, and it forms the content of the present section. In 1984, Bohigas, Giannoni and Schmit
5
stated the famous conjecture: Spectra of timereversal invariant systems whose classical analogues are K systems
show the same uctuation properties as predicted by the GOE. We recall that K systems are the most strongly mixing
classical systems. The most unpredictable K systems are referred to as Bernouilli systems. An alternative, stronger
version of the conjecture, also formulated in Ref. 5, replaces K systems by less chaotic systems provided they are
ergodic. In both its forms, this conjecture is commonly referred to as the Bohigas conjecture. For systems without
timereversal invariance, the GOE is replaced by the GUE. In its original version, the Bohigas conjecture was stated
without a reference to the semiclassical regime, i.e. to the limit h 0. However, all attempts to proof the conjecture
are based on some kind of semiclassical approximation, see Sec. VI.
We begin with a brief review of the historical development prior and up to the formulation of the Bohigas conjecture
(Sec. VA). In Sec. VB, we briey outline the main concepts of dynamical localization. We then turn to systems
which were crucial to the development (billiards, the hydrogen atom in a strong magnetic eld, and others). These
systems form the rst major part of the present section (Secs. VC to VG). Although representatives of mesoscopic
systems, quantum dots are included here (Sec. VE) because they share many properties with billiards. The examples
given and many others strongly support the Bohigas conjecture. This accumulation of convincing evidence has led
to several attempts at establishing a theoretical link between classical chaos and RMT. These are reviewed in the
second part of this section, i.e. in Secs. VH to VI. Lack of space forces us to present only the key elements of
this discussion. In Sec. VH, we outline a hypothesis which has particular bearing on RMT. It concerns the way
in which the structure of classical phase space inuences quantum properties. Three approaches towards a formal
proof of the Bohigas conjecture are presented in Sec. VI: Periodic orbit theory, a eldtheoretic approach using the
supersymmetry method, and a probabilistic argument based on structural invariance. In Sec. VJ, a brief summary
of the present section is given. There are many excellent reviews on classical and quantum chaos. With respect to
classical and semiclassical aspects, we mention only, in the order of appearance, the books by Moser
312
, Lichtenberg
and Liebermann
313
, Schuster
314
and Gutzwiller
108
.
A. Historical development
In its beginnings, the development of quantum chaos was strongly inuenced by chemists who tried to understand
simple molecules. The phenomenon of avoided level crossings was of particular interest. We follow partly the survey
given by Robnik
315
.
In 1929, von Neumann and Wigner
21
had shown that level crossings cannot occur in generic systems and that
avoided crossings appear instead. Teller
316
stated the same theorem again in 1937. Percival
317
argued in 1973 that, in
the semiclassical limit, the energy spectrum ought to consist of a regular and an irregular part reecting the regular
and irregular parts of classical phase space. In the early eighties Ramaswamy and Marcus
318
, Berry
319
. Zaslavsky
51
and Marcus
320,321
showed that level repulsion and avoided crossings are typical for nonintegrable systems. Already in
76
the late seventies and early eighties, Zaslavsky
322,323,51
, Berry
324,325,50,319
and Berry and Tabor
326
expected a drastic
change in the spacing distribution to accompany the transition from integrability to nonintegrability. In 1979,
McDonald and Kaufman
48
published their pioneering numerical study of the quantum analogues of the classically
integrable circular billiard and of the classically chaotic Bunimovich stadium. This work was closely followed by
numerical work of Casati et al.
49
on the Bunimovich stadium. In these papers, the abovementioned expectations
came true. Further and improved evidence, also in billiard systems, was presented by Berry
50
and by Bohigas et
al.
47,5,327
. All this work is reviewed in more detail in Sec. VC. The behavior of eigenfunctions near avoided crossings
was simultaneously studied by Noid et al.
328,168
, Marcus
320,321
and Ramaswamy and Marcus
318
who showed that
the eigenfunctions are strongly mixed. According to McDonald and Kaufman
48
and Noid
328
this causes a random
pattern of the wave functions whose nodal lines do not cross. In these papers, evidence was found that the lowlying
eigenvalues and the corresponding eigenfunctions do not follow this pattern and have mainly regular features.
For small spacings, level repulsion was expected to lead to a spacing distribution proportional to s

. Using ge-
ometrical arguments, Berry
50
suggested the value unity for the critical exponent , in keeping with the Wigner
surmise for the spacing distribution in timereversal invariant manybody systems. Zaslavsky
51
related to the
Kolmogorov entropy for the classical system. His work was a pioneering attempt to nd a quantitative relationship
between classical chaos and spectral uctuations, although no convincing evidence for his assertion was found.
More formal arguments were used in the framework of statistical mechanics. Pechukas
172
, Yukawa
329
and Nakamura
and Lakshmanan
174
showed that the probability distribution of the energy eigenvalues of the Pechukas gas coincides
with the GOE probability density. We do not follow this important line here because there is no easy link to chaos
in fewdegreesoffreedom systems.
B. Dynamical localization
Dynamical localization is a completely unexpected and novel feature of quantum chaos. In a classical system, driven
by some external timedependent force, the occupation probability in phase space spreads forever diusively. Not
so in the quantum analogue: After some characteristic time, the wave packet stops spreading in momentum space.
This phenomenon is referred to as dynamical localization. It is dual to the localization in conguration space
typical of disordered systems, see Sec. VI. The close relationship between both types of localization phenomena was
discovered by Fishman et al.
330332
. In both cases, there is quantum suppression of classical diusion. The wave
packet relaxes into a nonergodic localized state. For the case of the kicked rotor, a step towards formally establishing
the correspondence between both types of localization was recently made by using the supersymmetric nonlinear
model
333
, see Sec. VI D. A detailed account of dynamical localization was given by Casati and Chirikov
334,335
, see
also Ref. 336. We partly follow Ref. 334.
From a conceptual point of view, the study of wave functions is quite dierent from that of level statistics. Levels
are eigenvalues of the Hamiltonian and can be studied experimentally. The spectral uctuation properties of the
quantum analogues of classically chaotic systems have accordingly received much interest, and they form the content
of the bulk of the present section. The wave function , on the other hand, is not an observable. Investigating the
evolution of (t) in time t, one disregards all aspects of quantum mechanics related to the measurement process, and
the collapse of the wave function. However, the wave function (t) obeys the Schrodinger equation, a deterministic
equation. Thus, direct comparison of the time evolution of (t) with the dynamics of the corresponding classical
system is useful. Some authors prefer to reserve the term quantum chaos for investigations conducted in this spirit.
It should be emphasized that quantum dynamical localization in classically conservative systems is due to quantum
suppression of classically chaotic diusion. Many semiclassical studies address questions such as scars around periodic
orbits or similar phenomena which are often also referred to as localization. These two eects, however, should clearly
be distinguished.
Two time scales govern dynamical localization. The relaxation time scale t
R
is proportional to the density
0
of
those eigenstates (quasi energy levels) which are admixed to the original wave packet at t = 0. This is the time scale
on which periodic orbit theory can be applied. The time scale t
e
is characteristic of the classical relaxation to the
ergodic steady state. The quantity
=
_
t
R
t
e
(5.1)
is a measure for the ergodicity of the state. Shnirelman
337
argued that for 1 the nal steady state and all
eigenfunctions are ergodic, and the Wigner functions are close to the classical microcanonical distribution in phase
space. If, however, 1, all states are nonergodic, and their structure is governed by quantum mechanics.
77
On a very short time scale (which, according to the uncertainty principle, involves very high energies) the wave
packet spreads quite like a classical cloud of particles in phase space would. Then quantum eects set in, the wave
packet spreads in a manner not allowed by the constraints of Liouvilles theorem in classical mechanics, and eventually
may stop spreading and remain in a nonergodic localized state. To measure the size of localized states, Izrailev
67
introduced the entropy localization length
l
h
= exp(h) (5.2)
where the entropy is given by
h =

n
[(n)[
2
ln [(n)[
2
. (5.3)
Here, (n) are the wave functions in momentum space. In various cases, the wave functions are exponentially localized,
i.e. one has (n) exp([n[/l), [n[ . In those cases, l
h
turns out to be close to l. The spacing distribution
of localized states obeys, in the extreme case, the Poisson law whereas the Wigner surmise describes that of fully
ergodic systems. Thus, the spacing distribution can give information about the degree of ergodicity and localization.
Remarkably, it has been found in numerical studies that the repulsion parameter
e
in Izrailevs phenomenological
crossover formula (3.54) can be approximated by

e
=
l
h
l
e
(5.4)
where l
h
is the average of l
h
over all states and l
e
l
h
is the maximal entropy localization length corresponding
to ergodic states. As usual, labels the three symmetry classes. Another measure is the inverse participation ratio

1
m
=

n
[
m
(n)[
4
(5.5)
which is related but not identical to l
h
.
One of the rst examples to show dynamical localization was the kicked rotor. Another case, which is experimentally
accessible, is that of the hydrogen atom in a microwave eld. Experimental results by Bayeld and Koch
338
led to a
series of activities. Casati et al.
339
predicted localization eects in this system which were experimentally conrmed
by Bayeld et al.
340
. We refer the reader to the review in Ref. 336. Recently, these studies have been extended to the
hydrogen atom in both magnetic and microwave elds by Benvenuto et al.
341
.
C. Billiards
A classical billiard consists of a point particle which moves freely in a compact domain of ddimensional space
and which is reected elastically by the boundary of this domain. The energy and the modulus of the momentum of
the particle are constants of the motion which, therefore, is determined by geometrical optics. Hence, the dynamics
(regular, mixed, or chaotic) is the same for all energies. We restrict ourselves to d = 2. This case was and is
conceptually of great importance. Examples of integrable (regular) billiards are furnished by the circle and the
ellipse. Among the many examples of chaotic billiards, two have received particular attention. The Sinai billiard
consists of a square and an inserted concentric circle. The particle moves in the domain between square and circle.
The Bunimovich stadium is of oval shape and consists of two equally long sections of parallel straight lines which at
both ends are joined by two semicircles. In the Sinai billiard, chaos is due to the defocusing eect of reection at
the inner circle. In the Bunimovich stadium, chaos arises because the straight lines destroy the rotation invariance of
the two semicircles. Alternatively, chaotic dynamics can be attributed to an overfocussing eect in this geometry.
Four important billiards are shown in Fig. 23. Sinai and Bunimovich
342344
have shown that both billiards are fully
chaotic and belong to the Bernoulli class, i.e. to the strongest form of chaotic dynamics. Mathematical aspects were
discussed by Katok and Strelcyn
345
. Physicsoriented reviews on classical chaos in billiards can be found in the article
by Bohigas et al.
47
and in Gutzwillers book
108
.
The quantum analogue of a classical billiard is dened by the stationary Schrodinger equation with Dirichlet
boundary conditions (the wave function vanishes at the boundary). Apart from factors, the Hamilton operator for
a free particle is simply the Laplacean. Therefore, the problem is mathematically equivalent to determining the
vibrations of a membrane. Independently of the connection to quantum mechanics, such and more general systems
are of considerable interest and have been studied by mathematicians for a long time. For example, the Dirichlet
78
FIG. 23. Four important billiards: the Bunimovich stadium (top left), the Sinai billiard (top right), the Pascalian snail
(bottom left) and the annular billiard (bottom right). To reduce the symmetries, these billiard have to be cut along the
symmetry axes which are drawn as dashed lines.
79
condition may be replaced by the Neumann boundary condition (the normal derivative of the wave function vanishes
at the boundary). In these systems, the smoothed part of the level density and the smoothed part of the cumulative
density (k), see Sec. III B, show some very general features. As a function of the wavenumber k, (k) is given by
(k) =
A
4
k
2

L
4
k +C . (5.6)
Here A is the area and L the length of the perimeter of the billiard. The constant C describes curvature corrections,
cusps and other topological properties. The minus and the plus sign in front of the perimeter term apply in the case of
Dirichlet and Neumann boundary conditions, respectively. In the case of a particle of mass m in a quantum billiard,
the energy is E = ( hk)
2
/2m. Formula (5.6) is valid for arbitrary geometries. The area term was found by Weyl
346
in
1911 and 1912, the perimeter term by Kac
347
in 1966. More general results for arbitrary linear boundary conditions
were derived by Balian and Bloch
348
, cf. also Ref. 349. A review of the interesting history of these results can be
found in Refs. 47,108,350.
We focus attention on the spectral uctuation properties of quantum billiards. In sharp contrast to the level density,
these properties depend sensitively on the shape of the boundary. The spectrum of a billiard can be unfolded very
eciently with the help of formula (5.6). To analyze the spectral uctuations, it is important to use only levels which
pertain to the same set of quantum numbers. This is done by removing all symmetries from the problem. The Sinai
billiard, for instance, has four symmetry axes. One considers instead a billiard comprising 1/8th of the area. In the
case of the Bunimovich billiard, a billiard of 1/4th of the area is considered, see Fig. 23.
McDonald and Kaufman
48
investigated numerically the spectrum of the circle, a classically integrable system. The
results for the spacing distribution are shown in Fig. 24. Level clustering and degeneracies due to the absence of level
repulsion could clearly be seen. The rectangle, also regular, was studied with comparably high statistics (more than
10 000 eigenvalues) by Casati et al.
351
. To avoid degeneracies, the squared ratio of the side lengths was chosen as an
irrational number. Simultaneously, Feingold
352
performed similar numerical work. The spacing distribution probes
correlations on the scale of up to three or four mean level spacings and agrees very well with the Poisson distribution.
The spectral rigidity
3
(L) for the system considered agrees with the Poisson behavior L/15 only up to a critical
value of L 100 and becomes constant for larger interval lengths L. The spectrum of the rectangle is fully random
only on scales below this critical value. On larger scales nontrivial correlations appear which cannot be seen in
the spacing distribution. Casati et al.
353
argue that such correlations occur generically in integrable systems. The
very large number of levels needed to see this saturation of the spectral rigidity is often not accessible in numerical
simulations, let alone experiments.
Chaotic billiards behave very dierently. In their pioneering work of 1979, McDonald and Kaufman
48
calculated
eigenenergies and eigenfunctions of the Bunimovich billiard numerically. They showed that the spacing distribution is
of Wigner type and stressed the qualitative dierence to the spacing distribution of the circle calculated in the same
paper, see Fig. 24. McDonald and Kaufman related this dierence to the classical integrability or nonintegrability
of the two cases. They also observed that the nodal lines of eigenfunctions in the Bunimovich stadium do not cross.
The stadium was also investigated by Casati et al.
49
. Berry
50
studied the Sinai billiard in 1981, calculated about
500 eigenvalues and found linear level repulsion for small spacings. He gave a simple analytical argument for this
behavior and mentioned the connection to RMT. Berrys work demonstrates that accurate numerical methods for
the calculation of a large number of levels are not only technically important, they often prompt new conceptual
developments. In 1984, Bohigas et al.
5,47
computed more than 700 eigenvalues of the Sinai billiard and found very
good agreement of the spacing distribution with the Wigner surmise, see Fig. 4. This very accurate calculation
provided the rst statistically signicant evidence of the validity of the Wigner surmise not only for small, but for all
spacings. Bohigas et al. also showed that the spectral rigidity
3
(L) agrees very well with the GOE prediction up
to L 15. Bohigas et al.
327
also calculated about 800 eigenvalues of the Bunimovich stadium and found the same
excellent agreement of the data with GOE predictions. Mainly on the basis of these results, they formulated the
famous conjecture quoted in the introduction to the present section. Beginning at that time, many more examples
were studied numerically, comprising billiards and other systems with few degrees of freedom. Some of these cases are
treated in other parts of this review. Most examples essentially support the conjecture, some mark its limitations.
The rst study of the transition from regular to chaotic spectral uctuations in a billiard is probably due to
Robnik
315
in 1984. His billiard is dened by the conformal quadratic map z

= az + bz
2
of the unit circle, described
by the complex coordinate z. This is also referred to as the Pascalian snail, see Fig. 23. The parameter of interest
is = b/a with 0 1/2. For = 0 the billiard is integrable. For = 1/2, a cusp occurs and the map is no
longer conformal. For 1/4 1/2, the classical billiard shows strong chaos
354
. The quantum analogue displays a
transition
315
from Poisson to WignerDyson statistics as increases from 0 to 1/2. Robnik also introduced another
billiard dened by the cubic map
80
FIG. 24. Nearest neighbor spacing distributions of odd parity levels in the circle (left) and the Bunimovich stadium (right).
We notice that the spacing E is not normalized to the mean level spacing and that the distributions N(E) are normalized
to the total number of spacings. The energy ranges are 2500 < E < 10000 for the circle and 2500 < E < 4900 for the stadium
where the units are xed by the choice h
2
/2m = 1 in the Schrodinger equation. Taken from Ref. 48.
z

=
z +bz
2
+cz
3
exp(i)

1 + 2b
2
+ 3c
2
. (5.7)
It is called the Africa billiard since for some values of the parameters b, c and , the shape resembles that of the
continent Africa. Ishikawa and Yukawa
355
studied the transition from Poisson to WignerDyson statistics in yet
another billiard.
Of the numerous studies of billiards since 1984, we can mention but a few. A strong line of research uses semiclassical
quantization and the theory of periodic orbits to investigate quantum properties of nonintegrable billiards. Reviews
are found in Refs. 94,108,350. Sieber et al.
356
investigated the eect of the bouncing ball orbits on the spectral
uctuations of the Bunimovich stadium. The family of orbits which are perpendicular to the parallel straight sections
has a strong eect, such that the longrange spectral uctuations appear more regular. Sieber et al. constructed a
semiclassical method to remove the contributions of these orbits from the spectral uctuations. The general aspects
of these eects had earlier been studied by Keating and Berry
357
. Bohigas et al.
358
introduced the annular billiard. It
consists of the area between two circles with dierent radii, see Fig. 23. The billiard is integrable only when the circles
are concentric. The properties of the system depend on two parameters, the ratio of the two radii and the distance
between the two centers. The phase space is mixed. In the quantum system, tunneling from one regular region in
phase space to another may be enhanced by the intermediate chaotic domain (chaosassisted tunneling). This
prediction of the BohigasTomsovicUllmo model for mixed systems
154
has been conrmed analytically by Doron and
Frischat
359
. Dittrich and Smilansky
360,361
and Dittrich et al.
362
studied a chain of pairwise coupled billiards, with the
aim to study localization, see also the review in Ref. 363. The universal form of the curvature distribution (Sec. III H1)
was numerically conrmed for the Bunimovich stadium by Takami and Hasegawa
176
, and for Robniks billiard dened
above by Li and Robnik
181
. In the latter case, deviations for small curvatures from the analytical form (3.151) were
attributed to peculiarities of the billiard. Tomsovic
109
studied parametric correlations in the Bunimovich stadium to
investigate phase space localization properties of chaotic eigenstates. He introduced a correlator between state overlap
intensities and level velocities. Tomsovic found sizable correlations due to scarring of specic states Those correlations
are not described by RMT. Sieber et al.
364
discussed parametric statistics in billiards with mixed boundary conditions.
Schanz and Smilansky
365
studied the Sinai billiard from a scattering point of view. They found good agreement of the
spacing distribution of the eigenphases of the scattering matrix with Random Matrix predictions, see also the review
in Ref. 366. Primack and Smilansky
367
performed numerical simulations for the threedimensional Sinai billiard.
They found good agreement of the level statistics with RMT.
Pseudointegrable billiards form a separate class, with specic uctuation properties. A pseudointegrable system
81
is dened as a dynamical system whose phasespace trajectories live on surfaces with genus higher than one. The
classical dynamics is not chaotic
368
. Billiards in the form of certain polygons fall into this class. Detailed studies of
the spectral uctuations in pseudointegrable models can be found, for example, in the works of Richens and Berry
369
,
Cheon and Cohen
370
and Shudo et al.
371
. The spectral uctuation properties of those systems have confusing features:
on shorter scales, uctuations properties such as the spacing distribution are often indistinguishable from the Wigner
surmise. Only on longer scales, in the spectral rigidity for example, sizeable deviations from the properties of a chaotic
system become visible. Balazs et al.
372,373
introduced another class of billiards: These authors demonstrated that
the free motion of a twodimensional rigid body bouncing between two walls is equivalent to a billiard system. Thus,
coin tossing can be viewed as a billiard problem. Billiards involving a nonEuclidean metric are of great interest
for the problem of quantization. Schmit
374
has given a review on billiards on the hyperbolic plane. In chapter 19
of his book
108
, Gutzwiller discusses the motion on a surface of constant negative curvature. The structure of wave
functions can give rich information on the relationship between properties of the classical and the quantum billiard, see
Hellers review
375
and Gutzwillers book
108
. In some billiard systems, geometrical optics does not suce to understand
the motion. Examples are ray splitting eects
376
, see Sec. VD2, and Andreev billiards
333,377379
which prompted
the introduction of new random matrix ensembles, see Sec. VI A2. Recently, Borgonovi et al.
380
and Frahm and
Shepelyansky
381,382
studied localization eects of wave functions in billiards which partly destroy level repulsion. In
such regimes, the uctuations lie between Poisson and GOE. These authors studied diusive eects in the Bunimovich
stadium and rough billiards, respectively. Again, in the semiclassical limit h 0, pure GOE statistics is found. This
relates to the general discussion of dynamical localization in Sec. VB, i.e. to the quantum suppression of classically
chaotic diussion.
D. Demonstrative experiments
The experiments discussed below show that billiards are not purely theoretical constructs but can be realized
experimentally. For instance, quantum billiards can be simulated by microwave cavities. Such simulations are valuable
because parameters such as the size and the shape of the cavity, the range of frequencies etc. can be altered, and
they have been used to study specic aspects of quantum chaos. Simulations are not as fundamental, of course, as
experiments on systems such as atoms, molecules, or nuclei where entirely new and unexpected phenomena may occur.
Still, in some cases the simulation experiments exceed their original purpose and turn into interesting research objects
of their own right. Moreover, some experiments (such as the ones on elastomechanical resonances in threedimensional
solids) open new frontiers: The wave equation and the boundary conditions in these bodies diers qualitatively from
the Schrodinger equation.
The eld was pioneered by Schroder
383
who, about 40 years ago and long before the connection to quantum chaos
was established, realized the connection between acoustic waves and microwaves. He experimented with microwaves
in order to study the acoustic properties of rooms. He used rectangular cavities and, interestingly, analyzed the
nearest neighbor spacing distribution. For spacings larger than a mean level spacing or so, he found agreement with
the exponential law now referred to as the Poisson distribution. He also discussed connections of his experiments with
number theory. This and much more can be found in a fascinating book
384
he wrote on number theory in science and
communication.
Demonstrative experiments with microwaves and elastomechanical waves are discussed in Secs. VD1 and VD2,
respectively.
1. Microwave cavities
The stationary Helmholtz wave equations for the electric and the magnetic eld

E(r) and

B(r), respectively, in a
threedimensional metal cavity read
( +

k
2
)

E(r) = 0 and ( +

k
2
)

B(r) = 0 (5.8)
where

k is the wave vector. If the boundary of the cavity is perfectly conducting, the tangential component of the
electric and the normal component of the magnetic eld vector vanish at the boundary. We consider a at cavity
formed by two plane, congruent and parallel metal plates. The cavity is uniform in the direction perpendicular to
the planes, the z axis. With h the distance between the plates, modes with wave number k < /h cannot have a
node in the z direction and are therefore eectively twodimensional. These modes are transverse magnetic, i.e. the
electric eld vector is parallel to the z direction. For such modes, the relevant part of the wave equation reduces to a
twodimensional equation for the z component of the electric eld,
82
_

2
x
2
+

2
y
2
+k
2
x
+k
2
y
_
E
z
(x, y) = 0 . (5.9)
The tangential component of

B can be calculated from E
z
. If we identify E
z
(x, y) with the wave function (x, y)
and k
2
with 2mE/h
2
, Eq. (5.9) becomes identical to the Schrodinger equation for a twodimensional billiard. In the
electromagnetic case, the ray limit is attained if the wave length is small compared to the dimension of the cavity.
This corresponds to the classical limit in quantum mechanics.
Almost simultaneously but independently, Stockmann and Stein
385
and Doron et al.
386
in 1990 published the rst
experimental studies of microwaves in such at metal cavities and interpreted their results from the viewpoint of
classical chaos. Stockmann and Stein measured about 1000 eigenmodes of a Bunimovich stadium and of a Sinai
billiard. Both cavities were about half a meter in size and 8 mm thick. The experimental spacing distribution
agrees very well with the Wigner surmise. In a rectangle of similar size, Stockmann and Stein obtained a Poisson
distribution. Doron et al. investigated scattering in an elbow shaped cavity and compared the autocorrelation
function of the scattering matrix versus frequency to a semiclassical formula
387
which predicts a Lorentzian shape.
This shape applies when many channels are open. The autocorrelation function is nonLorentzian in the case of few
open channels
99,388
. Lewenkopf et al.
388
improved the analysis of the data in the light of this fact. Doron et al. also
discussed the enhancement of the Wigner time delay due to timereversal symmetry. Another scattering experiment
in microwave cavities was performed by Schultz et al.
389
.
Sridhar
390
, Sridhar and Heller
391
and Sridhar et al.
392
used a cavity to study wave functions of the Sinai billiard.
To measure the distribution of the electrical eld strength within the cavity, Sridhar et al.
392
used a perturbative
technique. These authors demonstrated the connection between the wave functions and classical periodic orbits and
identied scarred states predicted by Heller
375
. They also showed that a symmetry breaking in the Sinai billiard
leads to localized eigenfunctions. Related experiments were simultaneously done by Stein and Stockmann
393
in the
Bunimovich stadium. The measured microwave power is proportional to the imaginary part of the Green function.
The latter can be expressed semiclassically in terms of periodic orbits by Gutzwillers trace formula
108
. Thus, via
Fourier transform, the measured power gives information about the classical orbits. Stein et al.
394
extended these
investigations and thereby experimentally veried an insight due to Tomsovic and Heller
395,396
. These authors had
found that the semiclassical dynamics has a strong impact on the quantum mechanical behavior in billiards over very
long time scales.
The statistics of wave functions was measured directly by Kudrolli et al.
397
. For the Bunimovich stadium, slight
deviations from the PorterThomas law are caused by the bouncing ball orbits. The correlator of wave functions at
dierent space points (see Sec. III E2) agrees well with theoretical predictions for chaotic systems. Tiles in the cavity
which act as hard scatterers change the cavity into a disordered system. Deviations from the predictions for chaotic
systems due to localization eects are found both for wave function statistics and correlators. The results agree well
with a formula by Fyodorov and Mirlin
398
which accounts for localization eects. Prigodin et al.
124
also compared
experimental and theoretical results for wave functions, see Sec. III E2.
All these experiments suer from a broadening of the resonances due to absorption by the walls of the cavity. This
makes it impossible to separate closelying resonances. Missing levels pose a serious problem for the analysis of spectral
uctuations. To cure this problem, Graf et al.
399
used superconducting niobium cavities at a temperature of 2 K.
This results in a dramatic improvement of the resolution. About 1000 well resolved resonances were measured in the
Bunimovich stadium. The eect of the bouncingball orbits (Sec. VC) is clearly visible in the spectral rigidity, and can
be removed with the help of the semiclassical analysis of Sieber et al.
356
. With the same superconducting technique,
Alt et al.
400
measured eigenmodes in the hyperbola billiard and compared the results with various transition formulae
for the spectral uctuations. Because of the high resolution, Alt et al.
401
could compare the shape of individual
resonances with the BreitWigner formula and found excellent agreement over the full dynamical range. Likewise,
Alt et. al
402
found good agreement of the distribution of partial widths with the PorterThomas law, see Sec. III E1.
These authors also conrmed the abovementioned theoretical predictions of Lewenkopf and Weidenm uller
99
and
Harney et al.
251
for the autocorrelation function of the scattering matrix in the limit of isolated resonances. The
nonLorentzian shape of this function implies that the Fourier transform decays algebraically and not exponentially.
A detailed investigation of the decay properties of the Bunimovich stadium billiard was performed by Alt et al.
403
.
Measurements of spectral observables at a crossover transition are particularly interesting. For the case of gradual
breaking of timereversal invariance, such experiments were conducted by So et al.
404
and Storegen et al.
405
. So
et al. used a cavity with a thin ferrite strip adjacent to one of the walls. Timereversal invariance is broken when a
magnetic eld is applied to the ferrite. In the
3
statistic, the GOE GUE transition was observed with surprisingly
good resolution. Storegen et al. placed a wave guide inside a microwave cavity. The wave guide contained a
microwave isolator controlled by a magnetic eld. One side of the cavity could be moved. This made it possible to
study the motion of levels versus the length of the cavity. Storegen et al. determined the nearest neighbor spacing
distribution, the tail of the curvature distribution of the eigenvalues, and the distance of closest approach at avoided
83
crossings and found agreement of all three observables with GUE predictions. In an earlier study, Stockmann et al.
406
had already shown experimentally that in the case of timereversal invariance, the tail of the curvature distribution
agrees well with theoretical predictions, see Sec. III H1. A detailed theoretical study of microwave billiards with
broken timereversal invariance was given by Haake al.
407
. Recently, Kollmann et al.
408
and Stockmann et al.
409
compared experimental results on the level motion with the predictions of the Pechukas gas
172
. They also study
velocity and curvature distributions, see Sec. III H. Moreover, a link to periodic orbit theory is discussed.
Richter
410
discusses numerous further projects with microwaves: The Pascalian snail for the study of the regularity
chaos transition, the annular billiard for the investigation of chaos assisted tunneling and experiments addressing
localization.
In addition to the spectral rigidity, there exists another statistical observable, the correlation hole, to measure
longrange correlations, see Sec. III B4. The Fourier transform of the spectral twopoint correlation function maps
longrange properties onto shortrange ones in Fourier space. For RMT correlations, this function has a hole at small
values of the variable. Kudrolli et al.
411
found good agreement with the GOE prediction. Alt et al.
76
investigated the
correlation hole in more detail, including an analysis of scattering data.
As early as 1991, Haake et al.
412
used microwave experiments to study a certain pseudointegrable billiard, see
Sec. VC. Shudo et al.
371
investigated pseudointegrable billiards in the shape of polygons.
Cavities have been used also for another purpose. Cavities with simple boundaries show good agreement with the
Weyl formula for the mean density of eigenvalues. This changes when tiles are placed within the cavity to roughen
it. Deviations from the Weyl formula can be viewed as indicating a change of dimensionality
413
. In another study,
Sridhar and Kudrolli
414
conrm experimentally the theorem of isospectrality. It states that pairs of billiards with
dierent, but related, shapes exist which have identical spectra.
RMT was designed to describe uctuation properties of quantum systems. Does it also work for the spectral
uctuation properties of other wave equations? Deus et al.
415
, Alt et al.
416,417
and Dorr et al.
418
studied the statistics
of eigenmodes in threedimensional microwave cavities. In a truly threedimensional system, the vector Helmholtz
equations (5.8) are structurally dierent from the scalar Schrodinger equation. Deus et al. found very good agreement
of the spacing distribution with the Wigner surmise but deviations of the longrange correlations from the GOE
prediction. They attributed the dierence to a remnant of regularity in the irregularly shaped cavity. Alt et al.
416
used
a more regular cavity and found statistics intermediate between Poisson and GOE even for the spacing distribution.
Alt et al.
417
study a threedimensional Sinai billiard. In both investigations, periodic orbit theory is applied and
bouncing ball modes are extracted. The same geometry was investigated by Dorr et al.
418
who studied scarring and
various statistical observables of the electromagnetic eld distributions.
2. Acoustics and elastomechanics
Weaver
419
was the rst author to study the question raised in the previous paragraph: Does RMT apply to the
spectral uctuations of wave equations dierent from the Schrodinger equation? In 1989, a year before the rst
microwave experiments were reported, he studied the elastomechanical eigenmodes of aluminum blocks. The motion
of the displacement vector u(r, t) of a mass element at position r is governed by Naviers equation
420
. For an isotropic
material, the stationary wave equation
420
for the displacement vector u(r) can be decomposed into two Helmholtz
equations for the longitudinal (L) and transverse (T) parts. These equations read
_
+

2
c
2
X
_
u
X
(r) = 0 for X = L, T . (5.10)
Here c
L
and c
T
are the longitudinal and the transverse velocities, respectively, and is the frequency. Free boundary
conditions (vanishing stress at the surface) apply. The longitudinal mode represents a pressure wave, the two transverse
modes, shear waves. These modes are coupled through reection at the boundary. Since always c
L
> c
T
, Snells
reection law leads to ray splitting: The reection of an incoming longitudinal or transverse wave at the boundary
yields both a longitudinal and a transverse wave traveling in dierent directions. This phenomenon is also referred
to as mode conversion. Thus, elastomechanical wave equations are structurally very dierent from the Schrodinger
equation and even more complicated than the wave equation for the electromagnetic eld in three dimensions.
Weaver used rectangular aluminum blocks a few centimeters in size with angled slits. The slits break the symmetry
of the blocks. Dropping steel balls onto the blocks and measuring the response, he found about 150 eigenmodes.
The spacing distribution agrees well with the Wigner surmise, but the longrange correlations deviate from GOE
statistics. Why are the results for these nearly regular blocks so close to GOE? Bohigas et al.
421
attributed this fact
to the angled slits which act as defocusing structures. Delande et al.
422
used periodic orbit theory to analyze the
problem and reached a similar conclusion. Weaver and Sornette
423
showed that these systems can be viewed as being
84
pseudointegrable. On short frequency scales, the correlations are indistinguishable from GOE predictions. These
authors supported their conclusion by numerical simulations of a rectangular membrane with a point scatterer. Cheon
and Cohen
370
reached a similar conclusion for polygons and pseudointegrable billiards.
Using Weavers technique on bricksized aluminum blocks, Ellegaard et al.
424
measured about 400 eigenmodes in
the interval from 0 to 100 kHz. Rectangular blocks display Poisson statistics. The spectrum is a superposition of eight
spectra pertaining to dierent symmetries. In such a case, Poisson statistics is known
23
to apply almost irrespective
of the statistics of the individual spectra, see Sec. III B5. Using blocks in the shape of a threedimensional Sinai
billiard, Ellegaard et al. observed the Poisson GOE transition for the rst time in a threedimensional system.
Monocrystalline quartz blocks a few centimeters in size are even better suited objects because about 1500 eigenvalues
can be measured
7
in the range from 600 to 900 kHz. Unlike aluminum, quartz is an anisotropic substance. Both
the longitudinal and transverse velocities dier in the direction of dierent axes. Thus, the wave equation is even
more complicated than Eqs. (5.10) which apply to isotropic materials. Quartz belongs to the discrete crystal group
D
3
. The rectangular blocks used by Ellegaard et al.
7
were cut in such a way that all symmetries are fully broken
except a ip symmetry which has two representations, just like parity. The spectra were measured in transmission.
The resolution was close to that of superconducting microwave cavities. The existence of the ip symmetry makes it
possible to measure a gradual symmetry breaking which is statistically fully equivalent to the breaking of a quantum
number such as parity or isospin
227,137
, see Secs. III F and IVA5. The symmetry violation is achieved by removing
successively bigger octants of a sphere from one corner, thereby again creating a threedimensional Sinai billiard. The
results for the spacing distribution are shown in Fig. 25. They are in agreement with the random matrix simulations
of Ref. 137. The tiny symmetry violation induced by small octants immediately lifts all degeneracies. The spectral
rigidity deviates from the GOE prediction. This is probably because the blocks used have features similar to those of a
scalar pseudointegrable system, see Sec. VC. On short scales, however, the spectral uctuations are indistinguishable
from RMT.
The experiments show that RMT applies far beyond quantum mechanics to classical wave equations which are very
dierent from the Schrodinger equation. This is true even for anisotropic materials and for nontrivial eects such
as crossover transitions. As Weaver points out in his original paper
419
, statistical analysis has had a long history
in acoustics, especially room acoustics and instrument calibration, see Ref. 425. Vibrations of complex structures
have likewise been analyzed for a long time with statistical methods
426
. We are under the impression, however, that
Weaver was the rst author to recognize the importance of RMT for the understanding of uctuation properties in
acoustics and elastomechanics. (Schroders work
383
addressed a regular system).
We turn to twodimensional systems. Berry
50
cites unpublished experimental work by Ede on the vibrations of
metal plates in the shape of a Sinai billiard. Sapoval et al.
427
investigated numerically and experimentally vibrations
of membranes and found that fractal boundaries lead to localization eects. This is related to the work of Sridhar
et al.
413
mentioned above. In 1992, Legrande et al.
428
studied numerically the exural vibrations (bending modes)
of thin metal plates in the shape of a Bunimovich stadium and extended in the (x, y) plane. The stationary wave
equation for the displacement u
z
(x, y) perpendicular to the plate is of fourth order and can be written as
( +k
2
)(k
2
)u
z
(x, y) = 0 (5.11)
where is the Laplacean in two dimensions. The wave number is given by k
4
= 3
2
(1
2
)/Eh
2
where is the
frequency, the density, h the thickness, the Poisson number and E the elasticity module. Longitudinal modes
obeying the second order Helmholtz equation exist as well in elastic plates, but are not studied in Ref. 428. Legrande
et al. used the boundary condition of a clamped plate. Because of the form (5.11) the waves are a superposition of
Helmholtz waves, and of exponentially decreasing and increasing solutions due to the operator k
2
. Nevertheless, the
spectral uctuations agree well with GOE predictions. Recently, Bogomolny and Hugues
429
calculated the smooth
part of the cumulative level density, i.e. the Weyl formula, for exural vibrations in elastic plates. They used a
semiclassical approximation. Moreover, they worked out trace formulae for the integrable and the chaotic case.
Ray splitting occurs also in twodimensional systems and causes problems similar to those of elastomechanics.
Prange et al.
376
calculated the cumulative eigenfrequency density in a twodimensional system with ray splitting.
In three dimensions, the technical diculties are such that the cumulative eigenvalue density of a solid with free
boundaries has not yet been calculated. Dupois et al.
430
solved a simplied problem by introducing periodic boundary
conditions.
E. Quantum dots
During the last decade, it has become possible to manufacture microstructures of m size. In the case of semicon-
ductors, the electrons form a twodimensional gas located between the two layers of a heterostructure. The potential
85
0 1 2 3 4
0.0
0.3
P(s)
x
0.0
0.3
f
0.0
0.3
e
0.0
0.3
d
0.0
0.3
c
0.0
0.3
b
0.0
0.3
0.6
0.9
s
a
FIG. 25. Nearest neighbor spacing distribution P(s) for the dierent radii r of the removed octant: (a) r = 0 mm, (b)
r = 0.5 mm, (c) r = 0.8 mm, (d) r = 1.1 mm, (e) r = 1.4 mm, (f) r = 1.7 mm, (x) a huge radius of r 10 mm. The dotted
and the dashed curves are the theoretical predictions for a chaotic system containing no or one symmetry, respectively. Taken
from Ref. 7.
86
conning this twodimensional gas can be controlled extremely well, and it is possible to fabricate devices with widely
varying shapes. At temperatures below 1 K or so, the phase coherence length is larger than the system size, and
quantum coherence plays an important role. The elastic mean free path is typically 10 m or larger. The electrons
move ballistically, i.e. they scatter only at the boundaries dened by the conning potential. The device can be
viewed as a billiard. When coupled to external leads, it is called a quantum dot. The number of electrons on the dot
varies typically between a few and several hundred. The coupling between dot and leads may be controlled by gates
which create a potential barrier. A review of electron transport in quantum dots was recently given by Kouwenhoven
et al.
431
. Here, we are concerned with level statistics and wave function correlations in ballistic quantum dots. In our
discussion, we disregard the Coulomb interaction between electrons, save for the charging energy which is a mean eld
eect. Interaction eects in quantum dots have drawn interest only recently and are not discussed here, a detailed
discussion was presented by Marcus et al.
432
.
It is very useful to discuss quantum dots in the framework of scattering theory. Let G denote the conductance and
the total number of channels in each lead, dened by the number of transverse modes below the Fermi energy. The
dimensionless conductance g = (h/e
2
)G is the central object of theoretical interest. The Landauer formula
g =

a=1

b=1
_
[S
LR
ab
[
2
+[S
RL
ba
[
2
_
(5.12)
expresses g as the total transmission. The latter can be written in terms of the scattering matrix S describing the
electron transport through the device. Here, S
LR
and S
RL
are those blocks of the scattering matrix S which describe
scattering from the left to the right lead and from the right to the left lead, respectively. A more detailed discussion
of this formula is given in Sec. VI C1. The dimensionless conductance g is bounded by twice the number of channels
in one lead, g 2. The properties of the dot depend on the strength of its coupling to the leads. Without barriers,
we have g 2. The electrons move freely between dot and leads, the quasibound states in the dot become strongly
overlapping resonances, and g shows conductance uctuations. This regime is addressed in Sec. VE1. For g < 1 or
so, the electrons must tunnel through the barriers between dot and leads, and g displays isolated resonances. This
Coulomb blockade regime forms the topic of Sec. VE2.
1. Magnetoconductance and level correlations
In 1990, Jalabert et al.
433
suggested that uctuations of observables such as the conductance might contain in-
formation about the shape of quantum dots. The spectral uctuations of a regular system such as the circle dier
qualitatively from those of a chaotic system such as the Bunimovich stadium. This dierence might also aect the
conductance uctuations of quantum dots with either shape. It was known that magnetotransport properties in small
twodimensional conductors do depend on the geometry
434436
, and Doron et al.
386
investigated almost simultane-
ously the open elbow billiard with a similar aim, cf. Sec. VD1. But Jalabert et al. established the connection
between classical chaos and ballistic quantum dots. These authors investigated both semiclassically and numerically
the uctuations of g versus an external magnetic eld (magnetoconductance uctuations). Jalabert et al. used the
Landauer formula (5.12) and calculated semiclassically the autocorrelation function of g versus magnetic eld strength
B,
c(B) = g(B B/2)g(B +B/2)) (5.13)
where g = g g). The average is taken over an appropriate interval of values of B. For a chaotic dot, they found
c(B) =
c(0)
(1 + (eB/hc
cl
)
2
)
2
(5.14)
where the constant 1/
cl
denotes the rootmeansquare area enclosed by trajectories traversing the dot. The power
spectrum (Fourier transform) of c(B) decays exponentially,
s() = s(0)(1 +hc
cl
) exp(hc
cl
) . (5.15)
As mentioned in Sec. VD1, Bl umel and Smilansky
387
had computed a related quantity, the energy correlator of
the scattering matrix, also semiclassically. The result (5.15) holds only for weak elds where the cyclotron radius is
larger than the size of the dot. Moreover, the semiclassical approximation used in the derivation is justied only for
large channel number, 1. Jalabert et al. compare formula (5.14) with numerical simulations in two classically
87
FIG. 26. Electron micrograph of a quantum dot in the shape of a Bunimovich stadium, with 1 m bar for scale. This device
was used in the experiments of Ref. 442. The electrons can move in the black area. Two leads are coupled to the stadium.
Adapted from Ref. 442.
chaotic scattering systems: a Bunimovich stadium with two external leads and a four disk junction. The latter
consists of four identical disks whose centers form the corners of a square. The distance between these disks is a
quarter of the disk radius. Thus the four disks form an open scattering system
437,438
, extending the famous three disk
scattering system
439
. Good agreement is found. Deviations in the tail of the Lorentzian are ascribed to nonuniversal
behavior due to short trajectories. Moreover, to check the uniqueness of their ndings for chaotic systems, Jalabert et
al. perform numerical simulations on a rectangle which is integrable. There are uctuations, but the Fourier transform
indicates that they are qualitatively dierent and have nonuniversal shape. Related theoretical work was presented
by Doron et al.
98
, Jensen
440
and Oakeshott and MacKinnon
441
.
In 1992, Marcus et al.
442
performed an experimental test of the dierence between regular and chaotic quantum
dots. These authors fabricated two quantum dots in the form of a circle of radius 0.44 m, and two quantum dots in
the shape of a Bunimovich stadium of length 1.2 m and width 0.6 m, see Fig. 26. All dots had the same area of
0.41 m
2
. Both the electron density of 3.8 10
11
cm
2
and the mobility were measured and yielded an elastic mean
free path of 2.6 m, several times the sizes of the dots, so that the electrons move ballistically. The power spectra
show a striking dierence for the two geometries. The semiclassical prediction (5.15) for chaotic motion agrees with
the data for the Bunimovich stadium and not with those for the circle. The results are shown in Fig. 27. The number
of channels in this experiment was rather small, 3. Hence the semiclassical analysis is not very well justied.
However, Marcus et al. provided convincing experimental evidence that, as Jalabert et al. had predicted, the dierence
between regular and chaotic motion reveals itself in the conductance uctuations of quantum dots.
At zero magnetic eld, the resistance displays sizeable peaks. The widths dier for the two geometries. At 20 mK,
Marcus et al. found a few milli Tesla for the width of the peak of the Stadium and a much smaller value for that of
the circle. A connection between these peaks and the weak localization eect
31
in disordered systems was suggested.
Baranger et al.
443
investigated this phenomenon semiclassically. At zero magnetic eld, the system is timereversal
invariant. The amplitudes which contribute to the reection coecient come in pairs. Semiclassically, the members of
a pair correspond to timereversed classical trajectories and interfere constructively. As the magnetic eld is turned
on and timereversal symmetry is broken, the two amplitudes get out of phase. Hence at zero eld, the reection
coecient has a maximum and the transmission coecient has a minimum. This leads to a maximum of the resistance.
Baranger et al. showed that this weak localization eect depends on whether the dynamics in the dot is regular or
chaotic. The calculation accounts for the existence of the peaks and for the dierence in widths. The authors pointed
out that the diagonal approximation
108
typically used in the semiclassical approximation (Sec. VI 1) does not yield
all relevant contributions, in contrast to a fulledged random matrix model such as that of Iida et al.
140
.
In further experimental work, the weak localization dip of the conductance in the Bunimovich stadium was mea-
sured
444
for temperatures between 1.5 K and 4.25 K. In keeping with semiclassical arguments, the dip becomes less
pronounced as the temperature increases, see also Ref. 445.
The experiments on chaotic quantum dots had shown that the shape of the weak localization peak is Lorentzian.
Pluhar et al.
446,89
reproduced this behavior in a random matrix model, using the random Hamiltonian approach to
88
FIG. 27. Averaged power spectra (here denoted by Sg(f)) of conductance uctuations g(B) for stadium (solid diamonds)
and circle (open circles) with approximately three transverse modes. The left and the right gures display the results for two
dierent samples. Solid curves are ts of the semiclassical theory. The autocorrelation functions (here denoted by C(B)) of
stadium (solid) and circle (dashed) are shown as insets. Taken from Ref. 442.
the scattering matrix as discussed in Sec. III D. The dependence on magnetic eld was modeled as a GOE GUE
crossover transition. The ctitious time t in Eq. (2.13) of Sec. II B is proportional to the square of the magnetic ux
through the dot. In the notation of Sec. III F one has t
2
. Assuming ideal coupling to the leads, these authors
obtain an exact expression in terms of a threedimensional integral for g(t, )). Numerical integration shows that
the expression
g(t, ))

2 + 1
1
1 + 2(2 1)t/
2
(5.16)
is an excellent approximation to g(t, )) for all 2. This expression has Lorentzian shape. Gerland and Wei-
denm uller
447
extended this calculation to the case of nonideal coupling to the leads (barriers). They show that the
peak shape is always Lorentzian but that the parameters depend sensitively on the transmission coecients. Since
electronelectron interaction is not included in this model, these results are not valid in the Coulomb blockade regime.
Using a Brownian motion model in the space of scattering matrices, Frahm and Pichard
448
and Rau
449
investigated
the same problem, now formulated as a crossover transition from COE to CUE in the random Smatrix approach (cf.
Sec. VI C). Frahm and Pichard calculated all correlators of of g as functions of the eigenvalues of the transmission
matrix and of a ctitious time t. Unfortunately, it has not been possible to relate t with the ux through the dot.
Moreover, the analytical dependence of g(t, )) on t diers from Eq. (5.16). This suggests that the Brownian motion
model in the space of S matrices may not be a suitable means to model the crossover transition. This is in line with a
diculty often encountered in attempts to model parametric correlations in the random S matrix (or transfer matrix)
approach, cf. Sec. VI C.
The calculation of the correlator c(B) of the magnetoconductance dened in Eq. (5.13) within a random matrix
approach poses a much harder problem. Frahm and Pichard
448
point out that it is highly doubtful whether c(B)
can be worked out at all in the random S matrix approach. Within an appropriate extension of the model of Pluhar
et al., Frahm
450
calculated c(B) using an asymptotic expansion in the inverse channel number and obtained the
result (5.13). Recently, Gossiaux et al.
451
tackled the full problem for arbitrary channel number.
Prigodin et al.
116
investigated the conductance distribution P(g) for a chaotic quantum dot weakly coupled to
leads. The Hamiltonian has the form
H = H
c
+
i
2
+
iV
2
(
1
(r r
1
) +
2
(r r
2
)) . (5.17)
89
Here, H
c
is the disorder Hamiltonian of the chaotic dot. The second, nonHermitean term on the right hand side of
Eq. (5.17) accounts for incoherent processes such as electronphonon scattering which produce a width = /2
measured in units of the mean singleparticle level spacing . Prigodin et al. focus on the regime E
C
where
E
C
is the Thouless energy, see Sec. VI. In this regime, an interactionfree theory is appropriate. The last term on
the right hand side of Eq. (5.17) describes
141
the coupling to the leads via point contacts at r
1
, r
2
with V the volume
and
i
, i = 1, 2 dimensionless parameters. Prigodin et al. study this model with Efetovs supersymmetry formalism.
In the zero mode approximation and in the weak tunneling limit
i
, i = 1, 2, they compute the moments
g
n
) =
_
P(g)g
n
dg , n = 0, 1, 2, . . . . (5.18)
of the conductance. If the magnetic eld is suciently large, weak localization eects can be neglected and the unitary
symmetry class is appropriate. The distribution P(g) decreases monotonically with g and g = 0 is the most probable
value. This is in contrast to the case of systems strongly connected with leads. Prigodin et al. also rederive the
PorterThomas distribution for the unitary case, see Sec. III E2. Marcus et al.
452
measured the magnetoconductance
in the tunneling regime where the dot is nearly isolated. The power spectrum is enhanced at high magnetic frequencies.
This is consistent with the results of Prigodin et al.
116
. Moreover, the magnetoconductance displays a crossover from
aperiodic uctuations at lower elds to periodic, AharonovBohmlike uctuations at higher elds. At the crossover
point, the cyclotron radius of the electron motion has roughly the size of the Bunimovich stadium. For elds larger
than this critical value, the electron bounces along the boundary of the stadium, and the motion becomes regular.
Baranger and Mello
453
and Jalabert et al.
454
used the randomS matrix approach to calculate statistical observables,
particularly the distribution of the dimensionless conductance g. They discuss their results as functions of the channel
number and the universality class . In the case of only one channel in each of the two leads, they nd
P(g) =

2
g
/21
(5.19)
which shows a qualitative dierence for the three cases = 1, 2, 4. With increasing channel number, P(g) quickly
approaches a Gaussian with nonzero mean. These results were obtained for ideal leads and for maximal coupling
between dot and leads, and do not contradict those of Prigodin et al. Brouwer and Beenakker
455
bridged the gap
between these two approaches by calculating, in the random S matrix model, the distribution of g for arbitrary and
for all . For = 2 and small , the result of Prigodin et al. is reproduced.
2. Coulomb blockade regime and wave function statistics
In this regime, the barriers between dot and leads are so high that the intrinsic mean width of the quasibound
states in the dot is much smaller than the resonance spacing. The conductance is measured as function of the gate
voltage applied to the dot. At low temperature, a series of welldened isolated resonances is seen. Such measurements
were made by Meirav et al.
456
, Kouwenhoven et al.
457
, McEuen et al.
458
, Weis et al.
459
and Foxman et al.
460
, see
also the review by Kastner
461
. At each resonance, another quasibound level of the quantum dot passes through
the Fermi surface from above as the gate voltage is increased, and the number of electrons on the dot increases by
one. Beenakker
462
has shown that the resonance spacing E
n
E
n1
+ e
2
/C is the sum of the spacing E
n
E
n1
between adjacent quasibound singleparticle levels in the dot, and the charging energy e
2
/C. This charging energy
suppresses tunneling processes between resonance peaks and causes the Coulomb blockade. For suciently small
values of the capacity C, the charging energy dominates the expression E
n
E
n1
+e
2
/C. In a plot of conductance
versus gate voltage, equally spaced peaks appear. Several theoretical investigations
458,462,463
address the Coulomb
blockade regime. Here we focus on a feature which is of particular interest for RMT. In many of the experiments
456,458
,
the peak widths are dominated by temperature broadening, so that k
B
T e
2
/C, and all peaks have the same
width. Nevertheless, the peak heights uctuate by orders of magnitude. This behavior can be modeled within a
random matrix approach. Several aspects of this problem are of general interest and are presented in Sec. III E. Here
we discuss the applications specic to quantum dots.
We are interested in the regime k
B
T < (E
n
E
n1
). The height of a resonance peak is given by
g
max
=

4k
B
T
with =
1

(L)

(R)

(L)
+
(R)
(5.20)
where
(L)
and
(R)
are the partial decay widths of the resonance into the channels of the left and the right lead,
respectively. RMT had been shown to give the correct description for the spectral uctuations in irregularly shaped
90
quantum dots. In 1992, Jalabert et al.
464
used RMT to determine the statistical properties of the quantity in
Eq. (5.20). For an asymmetric dot with a single decay channel per lead, the distribution function P

() is obtained
as
P
1
() =
_
2

exp(2)
P
2
() = 4
_
K
0
(2) + K
1
(2)
_
exp(2) (5.21)
for GOE and GUE, respectively. Here, asymmetric means that possible reection symmetries have been removed by
proper deformations of the system geometries. If such symmetries are present, one has
(L)
=
(R)
and = /2.
Then the distribution function coincides with the PorterThomas law. In Eq. (5.21), K
n
is the nth modied Bessel
function of the second kind. Stone and Bruus
465,466
performed extensive numerical tests of these results. They used
the Africa billiard (5.7) as a model for the quantum dot and found good agreement with both of the predictions (5.21).
Experimental verications were presented by Chang et al.
467
and Folk et al.
468
. Stone and Bruus
465,466
also computed
the parametric velocity correlator (3.156) and showed its universality. This work was extended both analytically
and numerically by several other authors
116,115,114
. Prigodin et al.
116
discuss to which extend their results can be
applied to the Coulomb blockade regime. Bruus et al.
191
and Alhassid and Attias
469
studied parametric correlations
(see Sec. III H2) of the conductance peaks as a function of the magnetic eld separation. They predict a squared
Lorentzian in the case of fully broken timereversal invariance. This was also experimentally veried in Ref. 468.
F. Hydrogen atom in a strong magnetic eld
The hydrogen atom in a strong magnetic eld has played a central role in the conceptual development of quantum
chaos. This is mainly due to three properties of this system. (i) It is a real system for which experimental data
were available when theorists became interested in quantum chaos. (ii) Eectively, it has two degrees of freedom
and is, therefore, calculable. (iii) It has an extremely useful scaling property which simplied the calculations and
helped in the understanding of experimental results. Our brief discussion centers on the random matrix aspects
of this system. Detailed presentations, particularly on the periodic orbit aspects, can be found in the reviews by
Friedrich and Wintgen
470
, Hasegawa et al.
471
, and in Gutzwillers book
108
. A detailed discussion of the present status
of experimental work on the hydrogen atom in magnetic and electric elds was recently given by Neumann
472
.
The Hamiltonian of the hydrogen atom in a homogeneous constant magnetic eld B in z direction reads
H =
p
2
2m

e
2
r

L
L
z
+
1
2

2
L
(x
2
+y
2
) (5.22)
with r =
_
x
2
+y
2
+z
2
and where
L
= eB/2mc is the Larmor frequency. The rst two terms on the right hand
side correspond to the Hamiltonian of the hydrogen atom without magnetic eld. The third term represents the
paramagnetic Zeeman contribution, the fourth one is the diamagnetic term. The last two terms break rotation
invariance and destroy full integrability. With increasing magnetic eld strength, ever larger parts of classical phase
space become chaotic. We express the magnetic eld in natural units dened by dividing the Rydberg energy me
4
/2 h
2
by the Larmor energy h
L
. Then, B = 2.35 10
5
T where the magnetic eld strength parameter is dimensionless.
We use atomic units. Because of axial symmetry with respect to the z axis, the azimuthal quantum number M is
a good quantum number, and so is parity . It is convenient to rotate the frame with the Larmor frequency
L
.
This removes the paramagnetic term. In cylindrical coordinates, the momenta in =
_
x
2
+y
2
and z direction are
denoted by p

and p
z
, respectively. For xed M

values, the Hamiltonian reads

H =
1
2
_
p
2

+p
2
z
_
+
M
2
2
2

1
_

2
+z
2
+
1
8

2
. (5.23)
Precise calculations which yield an excellent description of the experimental data were performed by various groups.
As an example, we show results obtained by Holle et al.
473
in Fig. 28. The Hamiltonian (5.23) has the remarkable
scaling property
474
H( p, r, ) =
2/3
H(
1/3
p,
2/3
r, 1) . (5.24)
Wintgen
475
and Wintgen and Friedrich
6,476
realized that the scaling property (5.24) can be used for an ecient
computation of the eigenvalues. This step was crucial. It was used later also by Holle et al.
477
for a much improved
91
FIG. 28. Comparison of theoretical and experimental results for the deuterium Rydberg atom. The magnetic eld strength
is 5.96 T. In spectroscopic units, the energies of the Rydberg states are between -80 cm
1
and -20 cm
1
, where 1 cm
1
corresponds to 0.12 meV. At the end of the energy range, i.e. above -25 cm
1
, the corresponding classical system becomes
completely chaotic. Taken from Ref. 473.
92
understanding of experimental data, see the review
470
. Wintgen and Friedrich
476
calculated the spectrum some meV
below the ionization threshold at magnetic eld strengths of a few Tesla. All properties depend only on the scaled
energy =
2/3
E. This variable increases with both excitation energy and magnetic eld strength. One has < 0
for bound states and > 0 for resonances. Between = 0.8 and = 0.2, the spacing distribution for bound
states changes from Poisson to Wigner, see Fig. 5. In the same range, the fraction of classical phase space lled
by regular trajectories decreases drastically. The coincidence of these facts strongly promoted the understanding of
quantum chaos. As a function of and for L values below a critical value L
max
, the spectral rigidity follows the
same pattern but saturates for L > L
max
. This is in agreement with Berrys
56
general argument based on periodic
orbit theory, see Sec. VI 1. The hydrogen atom in a strong eld has been of great importance for the development
of periodic orbit theory. Although the relevance of classical periodic orbits for the quantum spectrum was already
known
478
, Wintgen
479
was the rst author to realize the full applicability of Gutzwillers theory. For reviews see
Refs. 470,471,108.
Although a real system, the hydrogen atom in a strong magnetic eld is intimately connected to one of the toy
models studied theoretically. Upon the introduction of parabolic coordinates

with
2

= r z, the Hamilto-
nian (5.22) can be written as the sum of three terms. Each of the rst two terms represents a harmonic oscillator
with angular momentum barrier, while the third term provides a sixth order coupling in

between the two. This is

actually the form used by Delande and Gay
480
and Wintgen and Friedrich
470
to calculate eigenstates.
Simons et al.
481
veried the existence of universal parametric correlations, see Sec. III H2, in the spectrum of the
hydrogen atom versus magnetic eld strength, see also Ref. 162. These authors demonstrated the universality of the
parametric level number variance and of the velocity correlator.
Zakrzewski et al.
482
made the following interesting observation. Very close to ionization threshold, the nearest
neighbor spacing distribution deviates from the Wigner surmise. This is caused by an almost complete absence of
spacings larger than about 1.5 mean level spacings. The classical motion is chaotic. However, the diamagnetic term
in Eq. (5.23) connes the motion only in the (x, y) plane. In the z direction, the electron can move very far away from
the proton. At such large distances, the Hamiltonian is the sum of two integrable parts. Zakrzewski et al. showed
that this eect can be described by a regular Hamiltonian coupled to a chaotic one modeled by a random matrix.
G. Model systems
Among a large variety of theoretical models, we subjectively select a few which, in our judgement, had impact
on RMT. Regarding other systems, we refer the reader to the reviews by Bohigas
47,350
, Eckhardt
94
, Gutzwiller
108
,
and Haake
57
. Molecular chemists introduced many such models systems which later became interesting for quantum
chaos. In Secs. VG1 and VG2, we discuss coupled oscillators and the anisotropic Kepler problem, respectively.
1. Coupled oscillators
In 1973, Percival
317
predicted that, in the semiclassical limit, the energy spectrum of coupled oscillators consists of
a regular and an irregular part which reect the classically regular and irregular motion, respectively, see Sec. VA. To
test this prediction, Pullen and Edmonds
483
in 1981 investigated numerically the classical and the quantum properties
of two onedimensional harmonic oscillators coupled by a fourthorder interaction,
H = H
1
+H
2
+ 4kx
2
1
x
2
2
H
i
=
1
2
(p
2
i
+x
2
i
), i = 1, 2 (5.25)
where k is the coupling parameter. This system has several advantages over other model systems. For example, it is
bound at all energies. Pullen and Edmonds studied the levels E
n
(k) as functions of the transition parameter k, and
worked out the second dierences

2
n
= [(E
n
(k + k) E
n
(k)) (E
n
(k) E
n
(k k))[ . (5.26)
Large values of the
2
n
s are due to avoided crossings and occur in the regime where the classical dynamics is chaotic.
Today we would say that
2
n
measures the curvature of the parametric level motion and indicates chaotic dynamics,
see Sec. III H1. Aside from their interest for chaotic systems, pairs of coupled harmonic oscillators became interesting
also as models for a new class of integrable systems with a second hidden integral of motion
484,485
.
A year later, Haller et al.
486
studied the system (5.25) numerically. This system possesses a C
4v
pointgroup
symmetry. It also has a scaling property, and the levels can be computed eciently. Haller et al. worked out
93
FIG. 29. Numerically calculated spectral rigidities and nearest neighbor spacing distributions for the two coupled oscillators
dened in Eq. (5.27). The parameters are 1 = 0, 1 = 122.1, 1 = 0, 2 = 0, 2 = 24.1, 2 = 0, 12 = 0, 12 = 50 and
12 = 0. The transition is studied as a function of the interaction strength . Figures (a) to (e) correspond to interaction
strengths = 0.10, 0.04, 0.02, 0.01 and 0. We notice the saturation of the spectral rigidities. The lines were obtained from a
model of random band matrices. Taken from Ref. 488.
the spacing distribution for a set of k values and found a transition from Poisson to Wigner behavior which was
tted with the BerryRobnik formula. The classical analogue of this system was known to undergo a transition
from predominantly regular to chaotic motion. The work of Haller et al. was one of the rst studies to show the
close relationship between the transitions from regular to chaotic motion in classical mechanics, and from Poisson
to WignerDyson statistics in the quantum case. Meyer et al.
487
added further evidence in the frame of classical
mechanics.
In a conceptually important contribution, Seligman et al.
488
investigated a related system which has a more complex
structure and shows richer features. They chose two onedimensional oscillators with fourthorder potentials coupled
by an interaction of the same form,
H =
1
2
(p
2
1
+p
2
2
) +V
1
(x
1
) +V
2
(x
2
) +V
12
([x
1
x
2
[)
V
i
(x) =
i
x
2
+
i
x
4
+
i
x
6
, i = 1, 2, 12 . (5.27)
The extensive numerical simulations show that, controlled by the parameters
i
,
i
and
i
, the spacing distribution
exhibits a transition from Poisson to GOE statistics when the classical system makes the related crossover from
predominantly regular to chaotic motion. Seligman et al. also worked out the spectral rigidity and found a saturation
beyond some interval length L
max
. Their results are shown in Fig. 29 for a certain choice of parameters. They also
investigated the behavior of the corresponding classical system. Shortly thereafter, Berry
56
expressed the spectral
94
rigidity in terms of periodic orbits and explained the saturation, see Sec. VI 1. To explain the transition, Seligman et
al. also constructed a model of random band matrices. Very similar models are nowadays used in localization theory,
see Sec. VI D.
Zimmermann et al.
489
calculated the spectral rigidity for the system (5.25) and compared the saturation eect
quantitatively to Berrys prediction. Berry argues that the maximal interval length L
max
for universal GOE behavior
is given by L
max
= 2h/DT
min
where D is the mean level spacing and T
min
is the period of the shortest periodic
orbit, see Sec. VI 1. Zimmermann et al. found that this estimate is consistent with the tendency of the numerically
obtained L
max
as function of the coupling parameter.
Bohigas et al.
490,491,154
used two coupled quartic oscillators,
H =
1
2
(p
2
1
+p
2
2
) +a()
_
1
b
x
4
1
+bx
4
2
+ 2x
2
1
x
2
2
_
(5.28)
with parameters , b and a() as a model system to study the eect of classical transport on quantum properties. In
a parameter range where the system is mixed, there exists a classical ux in phase space through imperfect barriers
which are due to Cantori. Cantori are remnants of tori in the classical phase space of a mixed system. A semiclassical
theory for the level number variance was derived and applied to this system by Smilansky et al.
492
. Because of its
conceptual importance, we discuss this system and the concepts emerging from this study in Sec. VH.
2. Anisotropic Kepler problem
Another model system introduced by condensedmatter physicists is the anisotropic Kepler problem. The Hamil-
tonian is
H =
p
2

2m

+
p
2
z
2m
z

e
2
_

2
+z
2
(5.29)
with =
_
x
2
+y
2
. The anisotropy is caused by the dierence between the masses m

and m
z
. This model describes
donorimpurity levels in semiconductors. In 1971, Gutzwiller
493,494
had shown that the classical problem exhibits
hard chaos. Later this problem became very important in periodic orbit theory. For most classical systems, the
KolmogorovArnoldMoser theorem prevents the abrupt transition from integrable to ergodic behavior: The structure
of most invariant surfaces in phase space changes smoothly under small perturbations. However, this theorem does not
apply to the pure Coulomb system. Therefore, it was not clear how the system (5.29) behaves if the mass anisotropy
1 m

/m
z
diers slightly from zero. Gutzwiller
495
had found strong evidence for an abrupt transition from regular
to ergodic motion. His results and the question whether the spectral uctuations change abruptly, too, led Wintgen
and Marxer
496
to a detailed numerical study of the classical anisotropic Kepler problem and its quantum analogue.
At a mass anisotropy of 0.2, classical phase space is densely lled with remnants of tori, i.e. with Cantori. Thus, the
system is only weakly chaotic. At suciently high excitation energies, the nearest neighbor spacing distribution agrees
perfectly with the Wigner surmise. But the spectral rigidity follows the GOE prediction only up to an interval length
of L 7 and then grows linearly with L in a Poissonlike fashion. The speed of convergence to the GOE prediction
is dierent in dierent regions of the spectrum. It is conjectured that the intermediate scale deviations and the slow
convergence to the GOE predictions are connected to the pronounced Cantori structure of classical phase space. This
example shows that, in its gross features, even a somewhat special system such as the anisotropic Kepler problem
is consistent with the overall picture developed throughout this review: RMT type correlations become visible very
quickly on short scales in the spectrum. On larger scales, however, there is stronger resistance to RMT correlations.
H. Classical phase space and quantum mechanics
We now leave the discussion of special systems and return to the general aspects: The connection between quantum
chaos and RMT. How does the structure of classical phase space manifest itself in quantum mechanics? About
ten years ago, Berry and Robnik
138
conjectured that the spectrum of a mixed system should be a superposition of
independent spectra, each associated with a chaotic (or regular) region in phase space. This is related to Percivals
317
picture, see Sec. VA. Under the crucial assumption of independence of all these spectra, the nearest neighbor spacing
distribution and other uctuation measures can be predicted. In the simplest case, one obtains the BerryRobnik
formula for the spacing distribution (Sec. III B). In a more realistic picture, boundaries partitioning the chaotic part
of classical phase space must be taken into account. Such boundaries may, for instance, be due to Cantori. Classical
95
trajectories do pass such boundaries. Nevertheless, the existence of the boundaries denes a new time scale for the
system. Bohigas, Tomsovic and Ullmo
154
extended the BerryRobnik surmise by accounting for this time scale in
terms of the classical ux through the boundary. The ensuing random matrix model acquires block structure. Each
region of phase space is represented by a Gaussian or Poissonian ensemble, as the case may be. The corresponding
matrix blocks are located on the diagonal of the total Hamiltonian matrix. The diagonal blocks are coupled through
blocks of random matrices. The strength (second moment) of the latter is determined by the classical ux between
the regions. The random matrix model for symmetry breaking (Sec. III F) is a special case of this model. The
mathematical concept is very similar to early work by Rosenzweig and Porter on the spacing distribution in complex
atoms, see Sec. IVB1. In the model of Bohigas, Tomsovic and Ullmo, there exists, in addition to the coupling between
dierent regions of the chaotic part of phase space due to classical ux, also a coupling between the regular and the
chaotic blocks and among the latter which is due to quantummechanical tunneling processes. This relates to the
idea of chaosassisted tunneling (see Sec. VC) where separated regular regions in phase space are coupled via a
chaotic region.
One can view the BohigasTomsovicUllmo model as an extension of the original Bohigas conjecture: Since Cantori
lead to more than one intrinsic time scale the ensuing random matrix model acquires a more complicated structure.
I. Towards a proof of the Bohigas conjecture
The most remarkable feature of RMT is the prediction of universal statistical behavior of quantum systems. After the
formulation of the Bohigas conjecture, it became a challenge to nd analytical arguments linking classical chaos with
RMT. Three approaches towards a formal proof of the Bohigas conjecture are presented here: Periodic orbit theory,
Berrys argument, and subsequent theoretical work (Sec. VI 1), a eldtheoretic approach using the supersymmetry
method (Sec. VI 2), and a grouptheoretical and probabilistic argument based on structural invariance (Sec. VI 3).
All three approaches involve, in one way or the other, a semiclassical approximation to quantum mechanics which
links quantum and classical behavior, i.e. formally, the limit h 0 is taken.
1. Periodic orbit theory
A natural framework for limking RMT and classical chaos is provided by periodic orbit theory. Developed by
Gutzwiller
497499,493,500
and by Balian and Bloch
501,502
, periodic orbit theory is based on the saddlepoint approx-
imation to Feynmans path integral for h 0 and is one of the means to implement the semiclassical approxima-
tion. In an important paper, Berry
56
used periodic orbit theory to investigate universal aspects of the
3
statistics
(Sec. III B3). We recall that
3
(L) is dened in terms of the staircase function (). This function is the unfolded
integral over the level density (E), and
3
(L) is obtained by minimizing the mean square deviation of () from
a straight line over an energy interval of length L, see Eq. (3.57). Berry writes (E) as the sum of a smooth and a
uctuating part,
(E) = (E)) +

(E) . (5.30)
The smooth part (E)) is identical to the mean level density R
1
(E), see Sec. III B1. Periodic orbit theory can be
used to write the uctuating part as the sum over classical periodic orbits (trace formula),

(E) =
1
h
+1

j
A
j
(E) exp
_
i
h
S
j
(E)
_
(5.31)
where j labels all distinct orbits, including multiple traversals. The phase factor in Eq. (5.31) is determined by the
classical action S
j
(E) of the orbits. The amplitude A
j
(E) is related to the monodromy matrix of the orbit. In an
integrable system, periodic orbits form d 1 parameter families lling ddimensional phasespace tori. A chaotic
system is dened as an ergodic system in which all periodic orbits are isolated and unstable. It can be shown that
in Eq. (5.31), this implies = (d 1)/2 for integrable and = 0 for chaotic systems. The length of the averaging
interval needed for a separation of (E)) and of

(E) must obey two conditions. (i) It must be small on the classical
scale, i.e. small compared to the energy E of the system. (ii) It must be large compared to two intrinsic energy scales
of the system, the mean level spacing D = 1/R
1
(E) and the energy h/T
min
given by the period T
min
of the shortest
periodic orbit. The associated energy DL
max
= h/T
min
sets an important scale. Since D is of order h
d
and T
min
is of
order h, we have D h/T
min
in the semiclassical regime. The largest uctuations in Eq. (5.31) stem from h/T
min
.
The energy range DL is classically small, and we can write S
j
(E + ) S
j
(E) +T
j
(E) where T
j
(E) = dS
j
(E)/dE
96
is the period of the orbit labeled j. Ignoring the dependence of the amplitudes A
j
and of the mean level spacing D,
the energy average in the denition (3.57) of
3
(L) can be performed trivially, resulting in a double sum over periodic
orbits. Inspection of the energy scales allowed Berry to simplify the result further and to write it as an integral over
all periods,

3
(L) =
2
h
2
_

0
dT
T
2
(T)G
_
DT
2 h
L
_
(5.32)
where the function
(T) =
_

j,+
A
i
A
j
cos
_
S
i
S
j
h
_

_
T
1
2
(T
i
+T
j
)
_
_
(5.33)
is the averaged double sum over the periodic orbit contributions. The plus sign in the sum over j denotes the restriction
to positive traversals, T
j
> 0. The function
G(y) = 1
_
sin y
y
_
2
3
_
d
dy
siny
y
_
2
(5.34)
does not depend on individual orbits j. The shape of G(y) is very similar to (but G(y) is dierent from) the twolevel
correlation function X
2
(r) of the GUE. Because of its shape, G(y) selects from (T) only those pairs of orbits whose
average period exceeds 2h/DL. According to Berry, this feature reects the fact that
3
(L) measures the deviations
from the linear behavior of the staircase. Closer inspection of the steps leading to Eq. (5.33) shows that the linear
behavior is determined by orbits with T
j
< h/DL while the deviations are due to orbits with T
j
> h/DL. Berry then
argues that the twopoint function (T) can be written in the form
(T) =
h
2+1
2D
_
1 b
sc
2
_
DT
h
__
. (5.35)
Here b
sc
2
(t) is the semiclassical approximation to the twolevel form factor dened in Eq. (3.41) of Sec. III A5,
i.e. the Fourier transform of the unfolded twolevel cluster function Y
2
(r). Collecting everything and introducing the
dimensionless time t = DT/h as new integration variable, one arrives at the semiclassical result

3
(L) =
1
2
2
_

0
dt
t
2
(1 b
sc
2
(t))G(Lt) . (5.36)
According to Eq. (5.36), the spectral rigidity can be obtained from a semiclassical approximation to the twolevel
form factor b
sc
2
(t).
It remains to show that the semiclassical approximation to b
sc
2
(t) leads to an expression for
3
(L) which is consistent
with RMT. To this end, Berry uses sum rules. The rst sum rule for the diagonal part of the sum in Eq. (5.33),

dia
(T) =
_

j,+
A
2
j
(T T
j
)
_
(5.37)
is due to Hannay and Ozorio de Almeida
503
. The sum rule applies for large values of T. Periodic orbit theory shows
that, in this limit, the density of periodic orbits increases, while the intensities A
2
j
decrease algebraically for integrable
and exponentially for chaotic systems. Hence the limiting relations

dia
(T)
1
2D
_
h
d
integrable
ht chaotic
(5.38)
follow for large T where t = DT/h is the dimensionless time of the periods. The second sum rule due to Berry
56
says
(T)
h
2+1
2D
(5.39)
in the regime DT h or t 1. This sum rule guarantees that the amplitudes and phases of very long orbits in the
trace formula correctly generate the mean level density. Comparison with Eq. (5.35) shows that the second sum rule
97
implies b
sc
2
(t) 0 for t 1. This is equivalent to the vanishing of the twolevel correlations on scales much larger
than the mean level spacing.
One further ingredient from periodic orbit theory is still needed. It can be shown that for T T
min
and because of
destructive interference, the diagonal sum (5.37) is a good approximation to the averaged double sum (5.33). We recall
that T
min
is the period of the shortest classical periodic orbit. In the regime t
min
t 1 where t
min
= DT
min
/h,
the limit relations (5.38) of the diagonal sum can be used. In the case of integrable systems, the limit relation (5.38)
coincides with the sum rule (5.39) implying that the diagonal approximation is also valid in the regime t 1. Thus,
b
sc
2
(t) = 0 is a good approximation for the twolevel form factor down to values of t
min
and Eq. (5.36) yields

3
(L) =
1
2
2
_

0
dt
t
2
G(Lt) =
L
15
, (5.40)
the random matrix result for Poisson regularity. For fully chaotic systems, the diagonal approximation must break
down in the regime t 1, because the sum rule (5.39) is in conict with the limit relation (5.38). The o-diagonal
terms enforce the physically required cuto of the limit relation (5.38), expressed by b
sc
2
(t) = 0 for t 1. In order to
work out the integral (5.36), Berry chooses a linear interpolation 1 b
sc
2
(t) t for t
min
< t < 1 as suggested by the
limit relation (5.38), implying

3
(L) =
1

2
ln L +c

(5.41)
in the regime 1 L L
max
= 1/t
min
. As usual we have = 1 and = 2 for timereversal invariant and non
invariant systems, respectively. This logarithmic behavior is in agreement with the random matrix prediction (3.63)
of Sec. III B3. Berry also calculates the numerical values of the constants c

. For = 2, he nds exact agreement

with the GUE prediction. This is so because the linear interpolation coincides with the GUE result (3.42) for the
twolevel form factor. For = 1, there is no numerical agreement with the GOE prediction. This is because the GOE
result (3.42) diers form the linear interpolation. Periodic orbit theory predicts universal behavior of the spectral
rigidity only in the regime L L
max
. Beyond this scale, the spectral rigidity saturates at a value
3
(). Berry
138
works out
3
() and discusses examples.
It is very dicult to go beyond Berrys semiclassical approximation to the twolevel form factor, i.e. beyond the
diagonal approximation to the periodic orbit double sum (T). In some special, nongeneric cases
504506
, and for
the Riemann function
507,508
, such a step has been possible. Recently, progress has been made by Bogomolny and
Keating
509
in the general case. These authors use a relationship discovered by Andreev and Altshuler
510
to connect
the nondiagonal contribution in the double sum (5.33) to the diagonal one. The key assumption is that in generic
systems and modulo exact degeneracies, the orbits up to a cuto period of the order of h/D can be treated as
statistically independent. Bogomolny and Keating point out that their results are closely related to the ones found
by using the supersymmetric nonlinear model treated in the following section.
2. Supersymmetric eld theory
We discuss two approaches
52,53
based on semiclassical eld theory. The central aim in both approaches is the same:
To derive, within a semiclassical framework, a supersymmetric generating functional for conservative systems. For
systems with fully chaotic classical phase space and in the long wavelength limit, this functional should yield RMT
statistics for the eigenvalues. In addition, the derivation should display the limits of validity of RMT level statistics.
Although based on a semiclassical approximation, both approaches avoid the use of periodic orbits or similar entities.
The two approaches are technically demanding. Moreover, they touch upon dicult issues in ergodic theory and are
still under discussion. Therefore, we restrict ourselves here to a discussion of the central ideas.
The approach by Muzykantskii and Khmelnitzkii
52
, related to work by the same authors
511
discussed in Sec. VI E3,
starts from a disordered system and considers the limit of vanishing disorder. While it retains the usual semiclassical
condition k
F
l 1 of Efetovs supersymmetry approach (where k
F
is the Fermi wave number and l the elastic mean
free path), it does not use the long wavelength limit in its usual form, ql 1. The wave number q is typically given
by the inverse of the linear dimension L
s
of the sample, and the paper focuses on the limit l L
s
. (In Sec. VI, the
symbol L is used for the sample size. Here we use L
s
to distinguish it from the symbol L for the metric used in the
supersymmetry method. Apart from this, we use the same notation as in Sec. VI.)
The authors consider a supersymmetric generating functional (averaged over disorder) of the form of Eq. (3.88).
For a disordered system, the Lagrangean takes the form
L() =

8
_
d
d
r trg(
2
)
1
2
_
d
d
r trg ln(iK) (5.42)
98
with K = ET +L/2, where T is the operator of the kinetic energy and L is the metric dened in Sec. III C1.
The source term is omitted. The authors circumvent the usual saddlepoint approximation. Instead, they construct
the Green function G(r, r

[) of the operator K, a matrixvalued function in superspace. Let

G(r, p) denote the
Wigner transform of G. Integrating

G(r, p) over the modulus [ p[ of the momentum p = n[ p[ denes a function g
n
(r)
for which a dynamical equation is derived. With v
F
the Fermi velocity, this equation reads
2v
F
n
g
n
(r)
r
= i
_

2
L , g
n
_
. (5.43)
A functional of g
n
(r) is introduced with the property that the stationary points of (with respect to a variation of
g
n
(r)) are the solutions of Eq. (5.43). It is shown that in the limit k
F
l 1, the action of Eq. (5.42) can be rewritten
in terms of the functional . This technically somewhat complex procedure allows taking the limit k
F
l 1 without
implying L
s
l. Further condence in the derivation stems from the fact that in the long wavelength limit, the
resulting action L() reduces to Efetovs form, Eq. (3.92). Moreover, it is found that L() remains welldened in the
limit l . This suggests that in this limit, L() yields the correct level statistics for systems without disorder. It
is shown that for fully chaotic systems, limitations of the range of validity of WignerDyson statistics are determined
by properties of the classical Liouville operator.
The approach by Andreev et al.
53,512
constructs a eld theory where the eective action is associated with ow in
classical phase space. Starting point is a supersymmetric generating functional containing the Hamiltonian operator
H for a system without disorder. A Gaussian average over energy centered at E for the single system described
by H replaces the ensemble average typical for disordered systems. It leads naturally to the HubbardStratonovich
transformation, and to the introduction of the matrix, see Sec. III C. The resulting Lagrangean is given by (we
again omit the source terms)
L() =
1
2
trg
q

2
+ trg
q
ln K (5.44)
where K = E H N + L/2. Here, depends on two sets of coordinates, trg
q
indicates a graded trace as well
as an integration over both sets, while N denotes the number of levels (eigenvalues of H) in the averaging interval.
We note the formal similarity of Eq. (5.42) and of Eq. (5.44). A saddlepoint approximation based on N 1 leads
to a saddlepoint manifold and the occurrence of Goldstone modes. The resulting nonlinear model contains the
classical Liouville operator. It is equivalent
53
to the model obtained in the zero disorder limit by Muzykantskii and
Khmelnitzkii, although the two theories dier in form.
If H is the Hamiltonian of a classically chaotic system, the nonlinear model requires regularization. This is
achieved by adding a small noise to the classical Liouville operator. This procedure renders the classical motion irre-
versible. In the limit of vanishing strength of the noise, the spectrum of the resulting timeevolution operator (known
as the PerronFrobenius operator) reects intrinsic irreversiblity properties of the underlying classical dynamics. The
spectral properties of this operator determine the statistics of the eigenvalues of H. If the PerronFrobenius operator
describes an exponential relaxation towards equilibrium, then its spectrum has a gap between the lowest and the next
eigenvalue, and the spectrum of H displays WignerDyson statistics in an energy interval determined by the size of
the gap. Corrections to RMT can be obtained using the spectral properties of the PerronFrobenius operator.
3. Structural invariance
In order to furnish the Bohigas conjecture with a formal justication, Leyvraz and Seligman
54,55
take an alternative
approach reminiscent of the use of probability theory in statistical mechanics. The three key steps of the argument
are: (i) A given object is identied as a typical representative of some ensemble. (ii) This ensemble is known to have
a particular property p with probability one. (iii) As a member of the ensemble, the given object possesses property p
with probability one. Leyvraz and Seligman use this general probabilistic reasoning to establish a formal link between
chaos and RMT. In this context, the three steps read: (i) A given classical system is identied as a typical member of
a set of classical systems which can be embedded in an ensemble E. (ii) Upon quantization, the members of E are
known to possess random matrix uctuations with probability one. (iii) Therefore, the given system also possesses
random matrix uctuations with probability one. We note that the term ensemble here has a meaning which diers
from its use in RMT.
The set of classical systems is dened in terms of a number of characteristic properties common to all members
of . Such properties may be dynamical symmetries, behavior under time reversal, etc. The key concept in the
construction of the ensemble E is structural invariance. It is based on a grouptheoretical notion. Let G be the group
of transformations which map a system in onto another such system. If G possesses an invariant measure, then
99
G induces an invariant measure on , and this denes the ensemble E. The properties dening must be ergodic,
i.e. hold with probability one for any element of E.
Examples show that, in general, G has innite dimension and, therefore, does not possess an invariant measure. To
overcome this diculty, Leyvraz and Seligman consider systems with compact phase space and with phasespace
volume [[. The associated Hilbert space has nite dimension N = [[/h
d
where d is the number of degrees of freedom
of the system. For systems without any discrete symmetries, invariant tori etc., i.e. for the fully chaotic systems, the
elements of G are given by the set of all canonical maps of onto itself. According to an observation by Dirac
513
,
a unitary transformation can be assigned to every element in . Thus, is mapped onto the unitary group U(N) in
N dimensions. This group U(N) possesses an invariant measure d(U(N)) which is precisely the probability density
of the Circular Unitary Ensemble (CUE). The third step is trivial and amounts to saying that quantization of any
map C in yields the random matrix uctuations of the CUE for the eigenphases of the quantum system. Leyvraz
and Seligman also show that the inclusion of timereversal invariance as a relevant property leads to the Circular
Orthogonal Ensemble (COE). Moreover, they apply the same line of reasoning to timeindependent Hamiltonians and
obtain random matrix uctuations for the eigenvalues of the quantum system. A critical discussion of the diculties
related to the construction of the ensemble E, the quantization procedure, and the necessity to consider a nite
dimensional Hilbert space, is given in Ref. 55.
By constructing a surprising example, Leyvraz, Schmit and Seligman
514
show that in this approach, symmetries are
easily dealt with. They consider a billiard in the shape of an equilateral triangle. It has threefold rotation symmetry
and three mirror symmetries with respect to the three symmetry axes. The mirror symmetries can be removed by a
proper nonsymmetric rounding of the edges. The resulting billiard still possesses the threefold rotation symmetry.
The system is timereversal invariant. This suggests that the spectral uctuations are described by the GOE. However,
surprisingly, the GUE applies. Leyvraz, Schmit and Seligman
514
show this with the help of structural invariance. A
semiclassical explanation was given by Keating and Robbins
515
: The characters of the corresponding symmetry group
are complex which leads to phase contributions in the traces of the Green function in the irreducible representation.
Formally similar to a breaking of timereversal invariance due to an AharanovBohm ux, this yields GUE statistics.
J. Summary: quantum chaos
There exists overwhelming evidence that for fully chaotic systems governed by a single time scale, the Bohigas
conjecture applies. A body of analytical arguments lends theoretical plausibility to this statement. However, a fully
satisfactory proof of the conjecture is still lacking. For strongly chaotic systems with more than one intrinsic time scale
(example: a chain of pairwise weakly coupled billiards, each fully chaotic), RMT in its pure form cannot apply. Such
systems are kin to quasi onedimensional disordered mesoscopic systems and possess the same statistical properties.
A detailed discussion of such systems and their spectral statistics is given in Sec. VI. The situation is more dicult
and less clear for the generic classical systems with mixed phase space. Here, the hypothesis of Sec. VH is likely to
apply. However, a derivation of this hypothesis is lacking, and so is a nonphenomenological determination of the
coupling between blocks.
VI. DISORDERED MESOSCOPIC SYSTEMS
The investigation of disordered systems in one, two or three dimensions is almost exclusively a domain of condensed
matter physics. The motion of electrons in a crystal with random impurities provides a key example. Particularly
important for our purposes is the mesoscopic regime to which we will restrict ourselves in the following. In this
regime, the phase coherence length L

is the largest length scale of the problem. Phase coherence is lost by inelastic
scattering (in the case of electrons, through scattering by phonons or by other electrons), while elastic scattering does
not destroy phase coherence. Thus, L

is essentially given by the inelastic mean free path. For electrons, L

grows
strongly with decreasing temperature. For typical sample sizes L in the m regime (see Fig. 30), the condition L

> L
can be met at temperatures below 100 mK or so. Here, quantum coherence is a dominant feature of experiment and
theory. Closely related coherence phenomena occur when classical waves (light or sound) pass through a disordered
medium. The theoretical treatments of phenomena in these dierent elds of physics bear a close analogy. Unless
otherwise stated, we conne ourselves to the case of electrons.
Stochastic features enter into mesoscopic physics in two ways. The rst source of randomness is provided by real
spatial disorder. Schrodinger waves propagate through disordered media diusively and/or show localization. Disorder
is theoretically described in terms of a random disorder potential (see Eq. (2.14)) with a suitably chosen correlation
length and probability distribution. Such stochastic modeling leads directly to a theoretical description in terms of
100
FIG. 30. Picture of a mesoscopic gold loop taken with a scanning transmission electron microscope (STEM). Taken from
Ref. 516.
an ensemble of Hamiltonians and constitutes a natural extension of classical RMT. The essential dierence to the
latter is due to the central roles played by dimension d, and by the spatial extension of the sample. The second, more
implicit source of randomness arises from elastic wave reection at suitably shaped boundaries which in the classical
limit would lead to chaotic motion. For electrons, this situation is realized in ballistic systems where the elastic
mean free path is larger than the sample size L. This source of randomness is the same as described in the preceding
section. Again, dimension d and spatial extension of the sample play a key role.
In Sec. III B7 we have stressed the distinction between two averaging symbols, the angular brackets . . .) for spectral
averages and the bar (. . .) for ensemble averages. The equality of these two averages can be proven in certain cases,
see Sec. III B7. In many other cases, however, the equality of an experimental average (obtained by averaging over the
spectrum, an external magnetic eld, or the Fermi energy) and a theoretical ensemble average remains a hypothesis,
cf. Sec. VI C1. In this section on disordered systems we mainly use the symbol . . .) for average quantities, in spite of
the fact that these quantities have typically been calculated by averaging over the random potential. In the few cases
where we deal with a purely theoretical quantity which cannot be observed in an experiment (like, e.g., an average
generating functional) we employ the bar.
For a systematic approach to the properties of spatially extended mesoscopic systems we have to consider various
relevant length and energy (or time) scales. We assume that the Fermi energy E
F
is always the largest energy
scale. The length scales and L were dened above. A third important length scale is provided by the localization
length . Except for strictly onedimensional samples, we always have < , and we assume in the following that
. Related to the two length scales and L are two time (and corresponding energy) scales, the elastic scattering
time = /v
F
and the timeofight t
f
= L/v
F
through the sample, with v
F
the Fermi velocity. Further relevant
energy (and related time) scales are the singleparticle level spacing at the Fermi surface and the Thouless energy
E
C
= hT/L
2
= ht
1
d
, where T is the diusion constant and t
d
is the classical diusion time through the sample. It
was already remarked in Sec. II C that E
C
plays a central role in mesoscopic physics.
Four dierent regimes are distinguished.
(i) The localized regime with L . The system size exceeds the localization length, and the amplitude for
propagation of an electron through the probe is exponentially small.
(ii) The diusive regime with L . The system size is intermediate between the elastic mean free path and
the localization length. Electrons are multiply scattered by random impurities and propagate diusively. (Almost)
all states of the system are extended and Ohms law applies. In this regime, we have a further inequality. The
dimensionless conductance g is given by g = E
C
/ 1, see Eq. (2.17). Hence, we also have E
C
and, moreover,
E
C
h
1
(this denes the diusive regime). Thus, the diusion time through the sample is too short to resolve
individual levels but (trivially) long enough for the electron to be multiply scattered. The particular signicance of
the inequality E
C
is discussed further below.
(iii) The condition L or, equivalently
1
t
1
f
characterizes both the ballistic and the nearly clean regimes.
The system size is smaller than the elastic mean free path. The Thouless energy is replaced by the inverse time of
ight t
1
f
through the sample. The two regimes are distinguished by the degree of disorder.
101
(iii a) The ballistic regime where h
1
ht
1
f
. The disorder, characterized by
1
, is strong enough to
thoroughly mix many energy levels of the system.
(iii b) The nearly clean regime where h
1
, ht
1
f
. The disorder is so weak that it can be taken into account
by loworder perturbation theory.
All subsequent discussions of mesoscopic systems make use of this classication scheme. It should be evident already
that the existence of four mesoscopic regimes vastly extends the range of problems dealt with in classical RMT. For
instance, spectral uctuation properties may dier in the four regimes. And the existence in d > 2 of a mobility edge
separating localized and extended states requires special attention.
In the following we discuss equilibrium (Secs. VI A and VI B) and transport (Secs. VI C, VI D, VI E, and VI F)
properties of mesoscopic systems, respectively. Sec. VI A deals with the statistics of energy eigenvalues. Sec. VI B
is devoted to persistent currents. Sec. VI C discusses the important case of quasi onedimensional wires. This
leads naturally to the investigation of random band matrices in Sec. VI D. In Sec. VI E we turn to systems in two
and higher dimensions with special emphasis on the distribution function of conductance uctuations in mesoscopic
systems. Sec. VI F deals with the rather recent topic of two interacting electrons in a random impurity potential.
Some concluding remarks are contained in Sec. VI G. We emphasize those facts, phenomena, concepts, and theoretical
developments which are closely related to Random Matrix Theory in its widest sense. We pay special attention to the
role of dimension, and of the spatial extension of the systems under study. We do so at the expense of a systematic
introduction into and account of mesoscopic physics.
A. Energy eigenvalue statistics
This and the following section are devoted to the spectral properties of isolated mesoscopic systems. Why has
this topic found such wide interest? After all, transport properties have so far been the main source of experimental
information on mesoscopic systems. And the study of the electrical polarizability of an ensemble of small metallic
particles by Gorkov and Eliashberg
40
in terms of spectral correlations was a rather isolated occurrence.
The very strong interest in spectral uctuation properties results from the intimate connection between transport
properties and spectral properties of mesoscopic systems. In a seminal paper, this connection was pointed out already
in 1977 by Thouless
517
, see Sec. II C. His insight lies at the heart of the modern scaling theory of localization and
helped to understand and interpret numerous mesoscopic uctuation phenomena. Altshuler and Shklovskii
184
used
the Thouless energy and a postulated connection to RMT to obtain a semiquantitative understanding of universal
conductance uctuations, one of the main manifestations of quantum coherence in mesoscopic systems. Finally,
persistent currents in mesoscopic rings, which are the topic of Sec. VI B, measure the sensitivity of the spectrum to
an external magnetic ux and provide a direct link between experiment and spectral uctuation properties.
In Ref. 40, it was suggested that GRMT and WignerDyson (WD) statistics in the form described in Sec. III A
might be used to also address the spectral statistics in small disordered metallic particles at low temperatures. This
suggestion could, however, not have been correct in general since WD statistics contains no parameter related to the
dimensionality (or even the spatial extension) of the system. Research on spectral uctuations in disordered metals
has therefore to a large extent been driven by the following questions: Under which circumstances does WD statistics
apply, where should one expect deviations, and how can these be calculated?
We approach the subject as follows. First, we recall in Sec. VI A1 how the validity of WD statistics was established
in the diusive regime in some limiting case, and how corrections to this limit were calculated. Second, in Sec. VI A2,
we summarize attempts to go beyond diusive systems, in the direction of both the ballistic and the localized regime.
Third, we deal with spectral statistics in the proximity of the metalinsulator transition in Sec. VI A3.
A recent review by Dittrich
363
nicely complements our discussion in the present section in that it focuses on the
semiclassical approach to spectral statistics in (quasi) onedimensional disordered systems.
1. Diusive regime and WignerDyson statistics
The abovementioned conjecture by Gorkov and Eliashberg
40
was proved by Efetov
44
. Efetov showed that many
statistical uctuation measures of small disordered metallic particles can be calculated in terms of a functional integral
over a eld of supermatrices Q(r). This nonlinear model was described in Sec. III C. For instance, the twopoint
correlation function R(s) with s = / is apart from an additive constant given by
R(s) Re
_
d[Q](. . .) exp (S[Q]) (6.1)
102
where (with h = 1)
S[Q] =

8
_
trg
_
T(Q)
2
+ 2iLQ

d
d
r . (6.2)
Here, = 1/(V ) is the density of states (with V the volume and the mean singleparticle level spacing of the
system), T the diusion constant, L a diagonal supermatrix, and Q a supermatrix whose detailed structure depends
on the symmetry class. The precise denitions can be found in Sec. III C, cf. also Ref. 44. The quantity is the
dierence between two eigenvalues. The preexponential terms in Eq. (6.1) are not explicitly specied. They typically
involve certain components of Q. In the context of Random Matrix Theory the correlation function R(s) is typically
denoted by X
2
(r s), see Eq. (3.39).
The eective Lagrangean S[Q] comprises two terms, one with and one without spatial derivative. The gradient
term owes its existence to spatial uctuations in the Q eld and would disappear if the system under consideration
were pointlike. The second or symmetrybreaking term trg[LQ] survives even in the zerodimensional limit
Q(r) = Q
0
= const.
Efetov showed that in the zerodimensional limit, his nonlinear model yields WD statistics for all three symmetry
classes. He did so by reproducing the twolevel correlation functions R(s) of GRMT from the symmetrybreaking
term of the nonlinear model. This result established the equivalence of the zerodimensional model and RMT
(here in the narrow sense of the three Gaussian ensembles). Deviations from this limit are due to the spatial extension
of the system.
To determine the range of validity of GRMT, one has to compare the energy scales associated with the two terms
in S[Q]. For the gradient term this scale is given by the Thouless energy E
C
= T/L
2
, while the symmetrybreaking
term is proportional to the dierence of two eigenvalues. For E
C
the symmetrybreaking term gives the
dominant zero mode, and one expects WD statistics, while for E
C
corrections to WD statistics should arise.
First corrections of this type were calculated by Altshuler and Shklovskii
184
within the perturbative approach of the
impurity diagram technique. (We briey turn to this method in our discussion of transport in quasi onedimensional
systems, see Sec. VI C). Their study was partly motivated by the phenomenon of universal conductance uctuations,
i.e. by the fact that the variance g
2
) of the dimensionless conductance g is of order unity in mesoscopic systems, see
Sec. VI C1. The argument by Thouless
517
(to which we have already alluded in the beginning of this section) linked
spectral statistics and the statistics of the conductance through the relation g = E
C
/. This relation is rewritten in
the form
g = N(E
C
)) , (6.3)
where N()) is the mean number of levels within an energy interval of size . In the diusive regime, where E
C
,
we have g 1 (good conductor). Through Eq. (6.3), the variance g
2
) became linked with the number variance
2
(see Sec. III B3). Following common usage in much of the condensed matter literature we denote the number variance
by N
2
) in this section. Since Eq. (6.3) limits the spectral range of interest to the Thouless energy E
C
, WD statistics
seemed appropriate. However, WD statistics implies N
2
) lnN), and the nite width of the energy levels due to
the coupling to the leads had to be invoked
184
to reduce this value to N
2
) 1 as required. For E
C
the result
of the perturbative analysis in Ref. 184 was
N()
2
) =
c
d
k s
2

_

E
C
_
d/2
, (6.4)
where c
d
is a constant which depends on dimension d, where = 1, 2, and 4 for orthogonal, unitary and symplectic
symmetry, respectively, where k is the number of independent spectra considered, and where s characterizes the
degeneracy of the levels. This result shows that with increasing energy , the stiness in the spectra of disordered
metals is reduced and the uctuations in the number of levels increases. The physical picture behind these results is
the following. For energies smaller than E
C
, i.e. for time scales larger than the diusion time through the sample,
the wave packet of a diusing electron has reached the boundary of the sample and is, in a sense, equilibrated. The
spatial extension of the sample is then irrelevant and, therefore, WD statistics applies. This is precisely the same
situation as in compound nucleus scattering, see Sec. IVA. In the opposite limit of energies larger than E
C
we deal
with a wave packet that has not yet reached the boundaries of the sample. The corresponding level statistics has
therefore not yet attained the universal limit and depends on the dimensionality of the system.
Numerically, the crossover to the regime described by Eq. (6.4) was observed by Braun and Montambaux
518
, see
Fig. 31.
The corrections found by Altshuler and Shklovskii
184
in the diusive regime could be rened in later investigations,
in particular by avoiding their perturbative analysis. Their calculation had focussed on the twolevel correlation
103
FIG. 31. The number variance N
2
for a system with 20
3
sites versus the energy E in units of the mean level spacing
(open circles). The Thouless energy is approximately EC 2.5. For comparison, the RMT behavior (dotted line), the
asymptotic E
3/2
behavior of Eq. (6.4) (full circles), and the full perturbative result derived in Ref. 184 are also shown. Taken
from Ref. 518.
function R(s) as a necessary prerequisite to determine the number variance. In the regime 1 s g, the perturbative
approach
184
was only capable of yielding the smooth, nonoscillating behavior of R(s). In two papers
519,510
, corrections
to R(s) were calculated in the nonperturbative framework of the nonlinear model. In the method used by Kravtsov
and Mirlin, the parametrization Q = T
1
0
LT
0
for a constant Q matrix was modied by the substitution L

Q(r)
with an ensuing nontrivial change in the integration measure of the supersymmetric functional
520
. Fluctuations of

Q(r) around the origin L of the coset space (leading to contributions from the gradient term in Eq. (6.2)) were taken
into account perturbatively and were integrated out in the spirit of a renormalization group treatment, see Sec. VI E2.
For g s this led to the result
R(s) = 1
sin
2
(s)
(s)
2
+
a
d

2
g
2
sin
2
(s) (6.5)
with a
d
a constant depending on dimension. Equation (6.5) holds for the unitary case. Expressions for orthogonal
and symplectic symmetry are also given in Ref. 519. The rst two terms in Eq. (6.5) represent the result for the GUE,
i.e. the universal limit. Since g E
C
/, see Eq. (6.3), the condition g s means E
C
. Hence, the third term
in Eq. (6.5) is a very small correction to the zerodimensional limit of Eq. (6.2) originating from the comparatively
massive modes of the gradient term. Therefore, corrections to the result of cRMT are quite small in the entire diusive
regime. The smooth part of Eq. (6.5),
R(s) = 1
1
2
2
s
2
+
a
d
2
2
g
2
, (6.6)
had already been found in the perturbative calculation
184
.
Andreev and Altshuler
510
rederived Eq. (6.5) for 1 s g. They also showed that for s g the oscillations in
R(s) are exponentially damped and only the perturbative corrections of Ref. 184 survive. Their calculation was based
on a new and interesting technical point. Following Ref. 510 we consider the generalization of Eq. (6.1) and Eq. (6.2)
to parametric perturbations of the impurity potential characterized by the strength parameter x,
R(, x)

2
J
2
_
d[Q] exp(S
J
[Q])

J=0
,
S
J
[Q] =

8
_
trg
_
D(Q)
2
+ 2iLQ+iJLkQx
2
(LQ)
2
/2

. (6.7)
The derivative with respect to J produces the preexponential terms omitted in Eq. (6.1), and k is a diagonal matrix
with elements equal to 1 and 1 in the BosonBoson block and FermionFermion block, respectively.
Let us focus attention on the case of unitary symmetry. The conventional treatment takes into account the usual
saddle point L of the eective Lagrangean. There exists, however, one additional saddle point
510
. It has the form
kL and is well dened for any nite nonzero x. This saddle point breaks supersymmetry and had previously been
104
overlooked. Taking into account the contributions of both saddle points leads to Eq. (6.5). The orthogonal and
symplectic cases could be treated similarly
510
.
An interesting observation concerning the corrections to WD statistics for s g was made by Kravtsov and
Lerner
521
. We recall that s g corresponds to E
C
, i.e. to the limit in which Eq. (6.4) is valid. In this regime,
the general (smooth) result for the twopoint correlation function derived in Ref. 184 reads
R(s) =
b
d
g
d/2
[s[
2d/2
. (6.8)
But in two dimensions we have b
d
= 0, and the next order in g
1
has to be invoked (weak localization corrections)!
As a result, R(s) for d = 2 is given by
R(s) =
(g, )

1
g
2
1
[s[
, (6.9)
where (g, ) = ( 2)
1
for = 1, 4 and (g, ) = 1/2g for = 2. Hence, in contrast to the number variance (6.4),
the level correlation function was found
521
to be totally governed by weak localization corrections in this regime.
Motivated by the validity of WD statistics in the metallic regime, Akkermans and Montambaux
522
reexamined
Thouless arguments
523,517
connecting dissipation with level statistics, and the Thouless relation (6.3). (In the original
work, the inuence of WD statistics had not been taken into account, and this had meanwhile been found to be
manifestly incorrect). Let the closed system with eigenvalues E
n
be subject to an external perturbation , and dene
the dimensionless quantity g
C
by
g
C
=
1

_
_

2
E
n

2
_
2
=0
_
1/2
. (6.10)
Obviously, g
C
measures the sensitivity of the levels E
n
to an external perturbation. (The perturbation can, for
example, be realized by a magnetic ux with = 2/
0
). Therefore, we may view Eq. (6.10) as a denition of the
Thouless energy E
C
such that g
C
= E
C
/. The statement equivalent to Eq. (6.3) is then simply g = g
C
523
where g
is the dimensionless conductance as given by the Kubo formula. Akkermans and Montambaux succeeded in deriving
this Thouless formula. Using scattering theory, the Friedel sum rule, and very general assumptions concerning the
number correlator N(E
F
, )N(E
F
,

)), they could write g in the form

g =
1
4

2
N
2
(E
F
, ))

=0
. (6.11)
Assuming WD statistics for the energy levels gave the central result
g g
C
. (6.12)
Thus, the Thouless formula was derived in a much more rigorous fashion. In addition, it was found that
522
_
_
E

_
2
_

_
_

2
E

2
_
2
=0
_
1/2
(6.13)
(with E a typical energy level). The bar on the l.h.s. of Eq. (6.13) denotes an average over all ux values. We note
that Eq. (6.13) connects a ux average with a local quantity at = 0.
As the last item, we address the eect of timereversal symmetry breaking in the diusive regime. This prob-
lem arises in isolated mesoscopic rings threaded by a magnetic ux . In Refs. 524 and 152, this problem was
investigated and compared to the GOE GUE crossover transition rst considered by Pandey and Mehta
28
, who an-
alytically calculated all spectral correlation functions for a certain random matrix model, see Sec. III F3. Dupuis and
Montambaux
524
established numerically that the crossover transition is governed by the parameter (E
C
/)(/
0
)
2
.
In the ergodic regime < E
C
, the transition is well described by the PandeyMehta random matrix model
28
, see
Eq. (2.13), provided the parameter t used there is identied with the parameter 4(E
C
/)(/
0
)
2
. Altland, Iida
and Efetov
152
calculated the twopoint correlation function R(s) in the crossover regime analytically. The calculation
used the supersymmetry method and a novel parametrization of the Q matrix, Q = T
1
LT, tailored to the breaking
of timereversal symmetry. The matrices T were decomposed into two (Cooperon and Diuson) parts,
105
T = T
C
T
D
, (6.14)
with
[T
D
,
3
] = 0. (6.15)
where
3
= diag(1, 1) is the matrix which describes timereversal symmetry breaking. With this construction,
the symmetrybreaking term in the eective Lagrangean depends only on T
C
. This fact simplies the calculation
considerably. As a result, the eective Lagrangean in the supersymmetric functional (6.1) for the correlation function
R(s) has the form
S[Q] =
A
8
_
trg
_
T(
x
i(2/L)(/
0
)[
3
, Q])
2
+ 2iLQ

dx . (6.16)
Here, x is the longitudinal coordinate in the ring and A the cross section of the ring. In the regime E
C
, i.e in
zeromode approximation, an explicit result for R(s) was derived, which coincides with the ndings in Ref. 28.
This concludes our discussion of the diusive regime. We have seen that in an energy range dened by the Thouless
energy E
C
, the spectral uctuation properties (including the crossover induced by the breaking of timereversal
symmetry) are well described by the Gaussian ensembles. Deviations from this universal limit can be systematically
calculated and are found to be generically small.
2. Ballistic and localized regimes
We summarize some important work on the crossover transition from the diusive to either the ballistic or the
localized regime, cf. the classication given at the beginning of this section. This crossover can be realized by
changing either the system size or the strength of disorder. The extreme examples of a practically clean and a
strongly localized system show that the regime chosen has a drastic inuence on the level statistics. In the former
case the levels are determined by the quantization conditions dictated by the boundary of the system while in the
latter case the system is composed of essentially independent localization volumes, and level repulsion is suppressed.
Altland and Gefen
525
addressed the crossover from the diusive to the ballistic regime. They studied noninteracting
electrons in a twodimensional square geometry with a random white noise potential. Within a suitable perturbative
framework an equation for the twopoint level correlation function R(s; B
1
, B
2
) depending on two magnetic elds B
1
and B
2
was derived,
R(s; B
1
, B
2
) =

2
2
2

_
tr
_

ln[1
(D)
(, B

)]
+

ln[1
(C)
(, B
+
)] S
1
()
__
. (6.17)
Here, B

= B
1
B
2
,

ln(1 x) is dened by the power series of ln(1 x) + x, S
1
() is an essentially unimportant
contribution and the coordinate representation of the (Cooperon and Diuson) operators
(C,D)
reads

(D)
(, B

; r, r

) =

2
G
+
(E
1
, B
1
; r, r

)G

(E
2
, B
2
; r

, r) ,

(C)
(, B

; r, r

) =

2
G
+
(E
1
, B
1
; r, r

)G

(E
2
, B
2
; r, r

) , (6.18)
where G

are averaged singleparticle Green functions. The spectral properties are entirely governed by the eigen-
values of the operators
(C,D)
. Both the diusive and the ballistic regime can now be investigated. In the diusive
regime, three energy ranges were considered, D1 (0 < E
C
), D2 (E
C
< < 1/), and D3 (1/ < ). It turns
out that the range D1 coincides with the domain of WD statistics established in Ref. 44. In the range D2, the
results of Altshuler and Shklovskii
184
apply. These two cases have been treated in detail in the preceding section.
In the range D3, we probe the dynamics of the system for times smaller than the elastic scattering time . In this
range the direction of the particle momentum is not yet completely randomized and the microscopic features of the
random potential matter. For a standard Gaussian white noise potential the number variance was found to increase
logarithmically,
106
N
2
) =
1

_
L
2
_
2
ln(E) . (6.19)
This result is nonuniversal since it depends on the details of the random potential. In the ballistic regime, the
Thouless energy E
C
loses its physical signicance because the electron typically traverses the sample nondiusively.
The extension of the sample enters the description now via the time of ight t
f
through the system. Again, three
energy ranges were considered: B1 (0 < 1/), B2 (1/ < < 1/t
f
), and B3 (1/t
f
< ). The range B1
corresponds to very large times. The electron traverses the system many times and is multiply scattered in the
process. This range is therefore similar to the diusive range D1. The range B3, on the other hand, corresponds to
the case where the electron does not have the time to randomize its momentum or even to reach the boundary of
the system. Therefore, B3 is analogous to D3. Finally, the range B2 is typical for the ballistic regime and has no
analogue in the diusive regime. Here, the electron has sucient time to explore the whole system without undergoing
complete momentum relaxation via scattering by the random potential. Therefore, the boundary of the system plays
an important role. In the case of a twodimensional square (or rectangle)
525
the number variance for the range B2 is
given by
N
2
) =
2

2
ln()
2

2
(E)
2
1 + (E)
2
, (6.20)
where is a phenomenological level broadening of the order of . Interestingly, in this regime the uctuations in the
level number decrease upon increasing the energy window E. A discussion of both, the dierence between energy
and disorderaveraging in ballistic systems, and the eect of nite magnetic elds is also contained in Ref. 525.
Another interesting point was made in Ref. 526. There, it was shown that several observables are still characterized
by the Thouless energy E
C
in the ballistic regime, although this energy scale can no longer be interpreted as the
inverse transport time through the sample.
Building on ideas by Oppermann
527
, Altland and Zirnbauer showed
333,379
that new universality classes arise in
mesoscopic structures with normally conductingsuperconducting (NS) interfaces. They found that the ergodic limit
of a quantum dot in contact with superconducting regions (cf. regimes D1 and B1 above) is not characterized by
the three Gaussian ensembles GOE, GUE, and GSE, which dene WD spectral statistics. Instead, the NS system is
described by four new ensembles dubbed C, D, CI, and DIII after the associated four symmetric spaces in Cartans
classication scheme, see Ref. 528. These four ensembles correspond to the four possibilities of combining good or
broken time reversal invariance with good or broken spin rotation invariance. Shortly after Ref. 333 became available
a microscopic model for class C was derived by Frahm et al.
377
. To justify their claims the authors of Refs. 333,379
started from the Bogoliubovde Gennes Hamiltonian
H =
_
h

h
T
_
(6.21)
with the diagonal block h

= h

and the pairing eld

. Depending on the symmetry class considered,

further constraints on h and are necessary. With the two additional assumptions that the phase shift for Andreev
scattering vanishes on average over the NS interface, and that the classical dynamics of the N region is chaotic, the
Hamiltonian (6.21) could be replaced by a random matrix of appropriate symmetry, whose independent entries are
Gaussian distributed. Several properties of the ensuing four random matrix ensembles were derived in Ref. 379, in
particular the spectral npoint correlation functions for classes C and D, and for an open NS dot the weak
localization corrections to the conductance which had earlier been considered by Brouwer and Beenakker
529
. Surpris-
ingly, some of these corrections persist in an external magnetic eld
529,333
. Whether or not universal conductance
uctuations in the presence of Andreev scattering depend on timereversal symmetry breaking seems to depend on
the precise physical situation
530,379
.
The authors of Ref. 379 argue that the four new ensembles together with the three Gaussian and the three chiral
ensembles (see Sec. VII B2) exhaust Cartans classication scheme for symmetric spaces. Therefore they do not expect
that additional universality classes will be found.
The four ensembles just discussed apply to the case where the proximity eect, i.e. an excitation gap in the spectrum
of the normal metal induced by the superconductor, is suppressed. A RMT treatment of the proximity eect has
been given by Melsen et al.
378
.
Having explored the properties of various ballistic or nearly ballistic systems, we now turn our attention to the
transition to localization. In the localized regime, wave functions of nearly degenerate states may have an exponentially
small overlap. Therefore, level repulsion is suppressed and in the strongly localized regime, one expects Poisson
statistics for the eigenvalues. The study of the crossover to the localized regime requires nonperturbative methods.
In a numerical study, Sivan and Imry
531
employed a tightbinding model with diagonal disorder,
107
FIG. 32. Schematic drawing of the local density of states correlation function (normalized by the product n
2
0
of the density
of states) as a function of r = |r| (in units of ) for . Taken from Ref. 531.
H =

i
h
i
[i)i[ +

ij
V
ij
[i)j[ , (6.22)
where h
i
is Gaussian distributed and V
ij
connects only nearest neighbors. According to the analytical results of
Ref. 44, it is expected that in the diusive regime, the nearest neighbor spacing distribution P(s) should agree (in a
certain energy range) with WD statistics. This was explicitly conrmed in one and three dimensions. In the localized
regime, the authors investigated the local density of states correlation function
P
E
(, r) =
_

ij
(E E
i
)(E + E
j
)
_
V
d
3
x
2
i
(x)
2
j
(x +r)
_
, (6.23)
which had earlier been calculated by Gorkov, Dorokhov, and Prigara
532
for onedimensional chains. This function
contains information about the correlations of electronic states which are energetically separated by and spatially
separated by r = [r[. The general behavior of P
E
(, r) for d = 1, 2, 3 can be characterized as follows
532,531
, see
Fig. 32. For r 0 P
E
assumes a nite, independent value, from which it decays to exponentially small values for
< r <

(level repulsion). At the scale

the function crosses over to its disconnected part (no correlations). It

was pointed out in Ref. 531 that this behavior follows from earlier ideas by Mott
533
, who had already introduced the
frequencydependent length scale

= ln(c/) as the typical tunneling range between almost degenerate states.

Here, is the localization length and c a constant.
Analytical progress in understanding level statistics in the localized regime is possible with the help of suitable
model systems. Here, we consider the quasi onedimensional mesoscopic wire with unitary symmetry for which
transfer matrix techniques should be useful. (This system is further discussed in the context of transport properties,
cf. Sec. VI C). Altland and Fuchs
534
used the results of Ref. 44 and, in the spirit of the transfer matrix approach,
reduced the calculation of R(, L) (see Eq. (6.1)) to the solution of the dierential equations
[
t
+O] Y
0
(, t) = 0, Y
0
(, 0) = 1 ,
[
t
+O] Y

(, t) = Y
0
(, t), Y

(, 0) = 0 . (6.24)
Here, = (
1
,
2
) are the radial coordinates for the Q matrix in the unitary case, t = r/ with r the coordinate
along the wire, and O =
r
/16 + V () with
r
the radial part of the Laplacian on the manifold of Q matrices and
V () = i(/

)(
1

2
) (

is the level spacing in one localization volume). The ratio /

governs the relative

importance of the two terms in O. The correlation function R(, L) can be expressed in terms of the eigenvalues and
eigenfunctions of O
534
. In general, Eqs. (6.24) cannot be solved analytically, and one has to calculate these eigenvalues
and eigenfunctions on a computer. The main results were as follows. For t < 1 the WD and the AltshulerShklovskii
regimes were recovered, albeit with superimposed oscillations on the scale in the latter. This nonperturbative
eect had not been calculated earlier. For t > 1 a gradual crossover to the Poisson limit (R(, L) = 0 as L ) was
observed, obeying the scaling law R(, L) (/L)
d
f(/

). The function f(x) is proportional to lnx for very small

108
frequencies (x 1) and proportional to x
3/2
for larger frequencies (x > 1). The latter behavior is reminiscent of
the AltshulerShklovskii regime, see Eq. (6.8).
We have seen that in the various regimes relevant for disordered mesoscopic systems, a fairly thorough understanding
of spectral uctuations has been attained. The use of WD statistics frequently served as a reference standard which
helped to identify and interpret relevant length and energy scales. The greatest challenge is probably still posed by
the phenomenon of localization, in particular by the existence of a metal insulator transition to which we turn next.
3. Critical distribution at the metal insulator transition
For d > 2 (or, in the symplectic symmetry class, for d 2) there occurs generically a metal insulator transition
(MIT). It is characterized by the existence of a mobility edge E
M
which separates extended and localized states. With
the help of the Anderson model
32
the mobility edge can be investigated as follows. For disorder values W exceeding a
certain critical disorder W
M
all states in the spectrum are localized. For W < W
M
, however, extended states appear
in the center of the band. This means that states in the band center are critical for W = W
M
. We address the
question: What are the properties of the energy eigenvalue statistics in the vicinity of or, in the thermodynamic limit,
at the MIT? We use the index M to label quantities which refer to the mobility edge.
An early argument due to Altshuler et al.
535
suggesting that the MIT might signicantly inuence the energy
eigenvalue statistics used the following interpretation of the result (6.4) obtained in Ref. 184 for > E
C
. In the
time h/ an electron propagates over a distance L

= (T/)
1/2
. This distance is smaller than the system size
L = (T/E
C
)
1/2
. Therefore, cubes of size L

possess independently uctuating spectra, and N

2
) is proportional to
the number of such cubes. For d = 3, this yields
N
2
)
_
L
L

_
3
=
_

E
C
_
3/2
, (6.25)
in keeping with Eq. (6.4). At the MIT, two important eects have to be taken into account. First, we have E
C
.
Therefore, the range of validity of WD statistics is conned to a few levels at most, and the argument leading to
Eq. (6.25) applies practically to all energies. Second, at the MIT the dimensionless conductance g is expected to
approach a critical value g
M
wich is independent of the system size L. Imposing this condition, one nds that the
diusion constant T becomes scaledependent, T = T(L) with T(L) 1/L for d = 3. Taking these points into
account, Altshuler et al.
535
could estimate the behavior of N
2
),
N
2
)
N())
= 0.25 . (6.26)
This result was closer to the Poisson distribution (which gives N
2
) = N)) than to Random Matrix Theory (which
gives N
2
) lnN)). The corresponding estimate for the nearest neighbor spacing distribution P(s) with s = /
gave
P(s) exp(s/) , (6.27)
which again is reminiscent of the Poisson distribution. It turned out, however, that these arguments were oversimplied
and that the estimates (6.26) and (6.27) had to be revised.
Shklovskii et al.
536
were the rst to formulate the idea that the nearest neighbor spacing distribution P
M
(s) at
the MIT might be universal, thus representing a third possibility besides the WignerDyson statistics P
WD
(s) and
the Poisson law P
P
(s). In the band center of a threedimensional Anderson model, and in the thermodynamic limit
L , one expects that P(s) = P
WD
(s) below the critical disorder (W < W
M
) and P(s) = P
P
(s) above it
(W > W
M
). For W > W
M
, this is because P(s) is the superposition of the contributions from innitely many and
statistically independent localization volumes. For W < W
M
, on the other hand, the range of validity of Wigner
Dyson statistics is extended to innity since E
C
/ diverges in the thermodynamic limit. Precisely at the MIT these
arguments do not apply since the localization length (W) diverges and E
C
. This fact led to the postulate
536
of a third critical distribution, which was believed to be a hybrid of P
WD
(s) and P
P
(s). Linear shortrange level
repulsion was expected to be followed by Poissonlike behavior, in agreement with Ref. 535.
To check the hypothesis of a third universal distribution, an interesting test was developed and performed in
Ref. 536. We consider the integral A =
_

2
P(s)ds over the tail of P(s). We recall that s is the energy dierenc in
units of the level spacing . Moreover, P
WD
(s) and P
P
(s) coincide near s = 2. The quantity
109
FIG. 33. The quantity (W, L) as a function of W for dierent system sizes L from numerical diagonalization. Inset:
schematic gure illustrating the smooth crossover behavior for nite system sizes. Taken from Ref. 536.
(W, L) =
A A
WD
A
P
A
WD
(6.28)
vanishes for WD statistics, and equals unity for the Poisson distribution. In the thermodynamic limit, (W, L) exhibits
a sharp crossover from = 0 to = 1 as W crosses the critical value W
M
. (For nite L the crossover is smooth). If a
critical, sizeindependent distribution at the MIT were to exist, the curves of (W, L) as functions of W and for xed
but dierent L should all intersect at one critical point W
M
(in Fig. 33 this point is denoted by W
c
). Furthermore
these curves should obey the scaling law
(W, L) = f(L/(W)) (6.29)
with (W) the correlation length of the metal insulator transition. These expectations were numerically veried in
Ref. 536, and estimates for both W
M
and the critical exponent of the localization length were given.
Further numerical conrmation for the scenario developed in Ref. 536 was provided by Hofstetter and Schreiber
537
.
In particular it was veried that while for W < W
M
and W > W
M
the function P(s) tends to the WignerDyson and
the Poisson distribution, respectively, as the system size increases, a sizeindependent third distribution emerges for
W = W
M
. This fact and the critical behavior of the quantity dened in Eq. (6.28) were exploited in Refs. 538,539
to determine, as done in Ref. 536, W
M
and the critical exponent from the spectral statistics alone. The results
(W
M
= 16.5, = 1.34 .10
538
and W
M
= 16.35, = 1.45 .08
539
) were consistent with each other and with
earlier independent ndings
540
. A similar exponent, = 1.35 0.10, was found by Berkovits and Avishai
541
in a
threedimensional quantum bond percolation system. This suggests that the Anderson and the quantum percolation
model are in the same universality class.
Analytical work by Kravtsov et al.
542
led to a rst indication that the result (6.26) was not completely correct.
From the analytical derivation of the asymptotic behavior of the twolevel correlation function
R(s) s
2+
(s 1) (6.30)
where = 1 (d)
1
it was concluded
542
that
N
2
) N)

(6.31)
instead of Eq. (6.26). The constants of proportionality in Eqs. (6.30) and (6.31) depend only on d and the symmetry
class (unitary, orthogonal, or symplectic), thus conrming the universality of the statistics at the MIT. The absence
of a term linear in N) was attributed to the sum rule
_

R(s)ds = 0
542,543
(see also Refs. 133,147), which together
with
dN
2
)
dN)
=
_
N
N
R(s)ds (6.32)
110
implies that the coecient of the linear term vanishes for suciently large N). With the help of a certain plasma
model
544
, which reinterprets the distribution of levels as the distribution of interacting classical particles in one
dimension, cf. Eq. (2.9), it was possible
543
to deduce the asymptotic form of P
M
(s),
P
M
(s) exp( cs
2
) (s 1) , (6.33)
with c some constant. This was again at variance with previous claims
535,536
that P
M
(s) should be Poissonlike.
Moreover, numerical work
545
indicated the presence of both a linear term and a term proportional to N)

in the
number variance. The situation obviously called for clarication.
Progress was made when the abovementioned sum rule was critically examined
546,521
. Although valid (in the limit
N) ) for any nite system size, the sum rule may be violated
521
in the thermodynamic limit. In other words,
the limits N) and L do not commute. This insight invalidated the claim that there could be no linear
term in N
2
). The number variance at the critical point was therefore now expected to have the form
N
2
) = AN) +BN)

. (6.34)
But why had earlier investigations
535
missed the nontrivial exponent ? Some light was shed on this question by a
transparent derivation of the relation (6.30) by Aronov, Kravtsov and Lerner who combined a semiclassical approach
with scaling ideas
547
. We consider the spectral form factor (cf. the closely related quantity b
2
(t) in Eq. (3.41))
K(t) =
1
2
_
R(s) exp(ist)ds . (6.35)
This form factor can be related to the return probability P(t) for a classically diusing particle
548
,
K(t) =
2tP(t)
4
2

. (6.36)
This relation is quite powerful. We recall that t
d
= h/E
C
is the diusion time through the system. In the ergodic
regime t t
d
we have P(t) = const., K(t) t and hence R(s) 1/s
2
, which corresponds to WignerDyson statistics.
In the diusive regime t < t
d
, on the other hand, P(t) (Dt)
d/2
, K(t) t
1d/2
/T
d/2
and Eq. (6.8), i.e. the result
derived by Altshuler and Shklovskii
184
, is recovered. As in Ref. 535, the transition from the diusive to the critical
regime was made with the help of the idea that the diusion constant T becomes scaledependent. Close to the MIT,
the dimensionless conductance scales as g(L) = g
M
(1 + (L/)
1/
) with some critical value g
M
31
. On the other hand,
g TL
d2
and, therefore,
T(L) g
M
L
2d
_
1 + (L/)
1/
_
. (6.37)
In Ref. 535 the approximation g(L) g
M
was used, and the square bracket in the relation (6.37) was absent. This
leads to a timeindependent spectral form factor and, asymptotically, to R(s) = 0. Therefore the predictions in
Ref. 535 were close to the Poisson case. With the more accurate expression (6.37) for T(L) both time dependence and
the nontrivial exponent enter the form factor K(t), and Eq. (6.30) is recovered
547
. In this way, the controversial
issues associated with the two terms in the number variance (6.34) were resolved.
Several numerical investigations added new aspects to the discussion of the critical level distribution. Evangelou
549
and Varga et al.
550
found that P
M
(s) can be parametrized accurately by the function
P
M
(s) = c
1
s exp
_
c
2
s
2
_
, (6.38)
with = 1 (d)
1
as above and c
1
and c
2
two dependent normalization constants. This was in good agreement
with both the expected shortrange level repulsion
536
and the asymptotic form (6.33). (We note, however, that
Zharekeshev and Kramer
551
tted numerical data with P
M
(s) exp( cs), with c some constant, and identied only
the linear term in N
2
)). Detailed direct numerical calculations of the spectral twopoint function R(s) by Braun
and Montambaux
518
provided the probably most complete picture of the behavior of this quantity. In the metallic
phase both the random matrix regime and (for the rst time) the AltshulerShklovskii regime could be identied (cf.
Fig. 31). At the mobility edge, the shortrange correlations were found to be weakened and the ensuing powerlaw
behavior for larger s was consistent with the picture of anomalous (scaledependent) diusion discussed above. Some
evidence was found, however, that the linear term in the number variance might have a rather small, or even vanishing,
coecient. In Ref. 552 the asymptotic form (6.33) and the relation = 1 (d)
1
were conrmed for the center of
the lowest Landau band in a quantum Hall system.
111
Does the breaking of timereversal symmetry aect the critical distribution? This is a controversial issue. While
no eect of a magnetic eld has been seen by Hofstetter and Schreiber
553
, a recent investigation by Batsch et al.
554
reported a universal (i.e. sizeindependent) crossover behavior from the critical orthogonal distribution discussed
so far to a new and distinct critical unitary distribution. A third type of critical distribution was found in two
dimensional systems with spinorbit scattering
555,556
. These systems have symplectic symmetry and do exhibit a
MIT, in contrast to twodimensional unitary or orthogonal systems.
Chalker, Lerner, and Smith
557
used a Brownian motion model to describe the energy level statistics at the mobility
edge. For the model
H() = H
0
+
_

0
d

V (

, r) ,
H
0
=
h
2
2m

2
+ U(r) , (6.39)
where both U(r) and V (, r) are random white noise potentials, one can derive a Langevin equation governing the
parametric dependence of the eigenvalues E
n
on
557
,
dE
n
()
d
= v
2

l=0
c
n,n+l
()
E
n
() E
n+l
()
+
n
() . (6.40)
Here,
n
() is a random force and the coecients c
nm
(), which are given by c
nm
() = L
d
_
d
d
r[
n
(, r)[
2
[
m
(, r)[
2
,
contain information on eigenfunction correlations. By solving Eq. (6.40) approximately, a generalization of the ex-
pression (6.36) for the spectral form factor could be derived, linking spectral statistics with correlations in the wave
functions (through the coecients c
nm
). The two main results derived within this approach were an asymptotic power
law for the parametric twopoint function R(s,
2
) at the MIT
557
,
R(0, )
2/(1+/d)
(6.41)
and an exact result for the spectral compressibility
558
,
dN
2
)
dN)
=

2d
(N) 1) , (6.42)
i.e. the coecient of the linear term in Eq. (6.34). Here, = d d
2
, where d
2
is the multifractal exponent governing
the behavior of the inverse participation ratio. These results demonstrate that spectral statistics at the MIT are
inuenced by the multifractal properties of critical eigenstates.
In spite of considerable progress, the eigenvalue statistics at the mobility edge must, at least in its details, still be
considered an open question.
B. Persistent currents
An isolated mesoscopic ring threaded by a magnetic ux carries a persistent current I as a thermodynamic
equilibrium property. This amazing fact comes about as follows. An electron moving once around the ring picks up
a phase factor exp(2i/
0
) where
0
= hc/e is the elementary ux quantum. Therefore, all observables depend
periodically on . This applies, in particular, to the free energy F =
1
ln tr exp(H), so that F/ diers from
zero. The equilibrium current I is then given by the thermodynamic relation
I = c
F

. (6.43)
The current does not decay in time, hence the name persistent current.
Persistent currents have played a central role in mesoscopic physics, for several reasons. First, their existence
obviously hinges on the condition that the wave function of the system be coherent over a length scale given by
the circumference L of the ring. This must be true even in the case of multiple scattering by impurities. At the
beginning of the mesoscopic era, the existence of phase coherence in the diusive regime was a controversial issue.
The positive answer to this question paved the way to experimental and theoretical investigations of a number of
interference phenomena. Second, Eq. (6.43) shows the close connection between the persistent current, the sensitivity
112
of eigenvalues to boundary conditions, and energy level statistics, putting the study of the persistent current at the
heart of mesoscopic physics. Third, the problem of persistent currents brought to the fore the important dierences
between open and isolated (closed) systems. It turned out, in particular, that the grand canonical ensemble commonly
used in the calculation of transport properties is inappropriate to describe persistent currents. The grand canonical
ensemble has to be replaced by the canonical ensemble, and this caused technical problems. Fourth, experimental
values of the persistent current could not be explained in the framework of theories for independent electrons. This was
unexpected since theories neglecting the electronelectron interaction had been very successful in modeling transport
properties of mesoscopic systems. Seemingly, electronelectron interactions are more important in closed than in
open systems. A satisfactory theory with interactions that both reproduces the measured properties of the persistent
current and explains why interactions are less critical for open systems is still lacking.
After explaining the early insights and ideas leading to the theoretical prediction of persistent currents, we describe
the three experiments performed up to now. In a third part we review analytical calculations neglecting interactions.
In this section, aspects of Random Matrix Theory come into play. Therefore, we devote considerable space to this
discussion. Finally, we briey summarize the present status of theories which include the interaction between electrons.
A more comprehensive review of the topic can be found in Ref. 559.
1. Early theory, and the three experiments
It has been realized a long time ago
560,561
that persistent (dissipationless) currents occur in ideal rings free of impu-
rities and interactions. This fact was rediscovered and reformulated more than twenty years later
562,563
. Expanding
on our introductory remarks above, we consider an ideal onedimensional ring threaded by a magnetic ux . The
magnetic eld can be gauged away from the Hamiltonian
562
at the expense of introducing a twisted boundary condi-
tion for the wavefunction, (x+L) = exp(2i/
0
)(x), where x is the coordinate along the ring. The eigenfunctions
at zero eld are given by exp(2imx/L), and the corresponding eigenvalues are ( h
2
/2m
e
)(2/L)
2
m
2
, where m is the
quantum number of the z component of angular momentum (perpendicular to the plane of the ring). At nite eld
we have to replace m by m/
0
. This leads to a quadratic dependence of the singleparticle energies on magnetic
ux. Equation (6.43) shows that at zero temperature, the persistent current is given by the sum over all occupied
singleparticle levels of the ux derivatives of the singleparticle energies E
m
,
I =

m
E
m

. (6.44)
This relation explicitly shows the close connection between persistent currents, energy level statistics, and the sensi-
tivity of the spectrum to an external perturbation alluded to above.
Interest in this topic was greatly enhanced by a seminal paper of B uttiker, Imry, and Landauer
564
. These authors
claimed that persistent currents exist in normalmetal rings even at nite temperature and in spite of multiple
elastic scattering. Their work was followed by a series of papers on 1d
565,566
and multichannel
567569
rings with
noninteracting electrons. These investigations tried to assess the chances to actually observe persistent currents in
experiments by considering the inuence of channel number, temperature and disorder. The main results of these
papers can be summarized as follows. In ideal 1d rings with a xed number N of particles, the persistent current is a
sawtoothshaped function of jumping from I
0
= ev
F
/L to +I
0
(v
F
being the Fermi velocity) at even (N even) or
odd (N odd) multiples of
0
/2. The typical current I
typ
= I
2
)
1/2
is given by I
0

M, i.e. it grows with the number

M of channels. The relevant energy scale for temperature eects is the singleparticle (rather than the manyparticle)
level spacing . In the presence of disorder, the Fermi velocity v
F
in the expression for I
0
is replaced by the diusion
velocity L/t
d
, and the dependence on the channel number vanishes, I
typ
= I
0
/L. The average current, on the other
hand, was found to decay exponentially with L on the scale of the elastic mean free path .
Bouchiat, Montambaux et al.
570
showed that this last assertion is incorrect, for a very fundamental reason. For
closed systems in the mesoscopic regime, it is of vital importance to keep the particle number in the system exactly
constant as one averages over the disorder congurations. Thus it is necessary to employ the canonical ensemble.
The exponentially small currents found in Ref. 568 were shown to be an artifact produced by the use of the grand
canonical ensemble, i.e. by keeping the chemical potential (rather than the particle number) xed in the averaging
process. The thermodynamic limit does not apply to mesoscopic rings, and the two ensembles are not equivalent. By
a combination of analytical arguments and extensive numerical simulations, the following results for an ensemble of
isolated rings were obtained
570
. The typical current I
typ
= (I))
1/2
is independent of channel number M while the
average current decreases with increasing disorder and channel number like I) = I
0
(/ML)
1/2
. Furthermore, the
persistent current was found to be periodic in the ux with period
0
/2. This period halving comes about because
the ensemble considered contained rings with even and odd particle numbers. The most important result of Ref. 570
113
FIG. 34. The second (top) and third (bottom) zerotemperature harmonic of the magnetic moment (induced by persistent
currents in an ensemble of 10
7
copper rings) versus the dc magnetic eld. The Fourier decomposition was made with respect
to the frequency of a small ac component in the eld. One ux quantum through a single ring corresponds to 130 G. The
magnitude of the oscillatory moment is estimated to be roughly 10
15
Am
2
. Taken from Ref. 571.
was the comparatively large value for the average current. Experimental conrmation of the theoretically predicted
phenomenon of persistent currents suddenly seemed to be within reach.
The rst experiment on persistent currents was performed by Levy et al.
571
who used an ensemble of 10
7
copper
rings mounted on a single wafer. For technical reasons, these rings were actually squares, with a circumference of
L = 2.2m. With L , the experiment probed the diusive regime. A magnetic ux with a small, slowly varying
ac component was applied to the sample and the resulting magnetization measured. The ac component allowed for
an eective background rejection. The measured current was indeed
0
/2periodic, in agreement with the numerical
simulations
570
, see Fig. 34. The average current per ring was 0.4 nA, corresponding to 3 10
3
ev
F
/L. This rst
observation of persistent currents was a striking conrmation of the theoretical ideas developed earlier. However, the
current was roughly one order of magnitude larger than the values estimated in Ref. 570.
Two more experiments followed
572,573
. Both provided interesting additional information that went beyond Ref. 571.
Chandrasekhar et al.
572
investigated three single gold loops. Two of these actually were rings with diameters of 2.4 m
and 4.0 m, respectively. The third loop was a rectangle of dimensions 1.4 m 2.6 m. Transport measurements for
gold lms fabricated like the loops gave an elastic mean free path = 70 nm. Therefore, all three probes were in the
diusive regime. The observed current was uxperiodic with period
0
. The measured amplitudes at a temperature
of 4.5 mK were in the range 0.23.0 ev
F
/L. The absence of period halving is an immediate consequence of measuring
the response of single rings. No averaging over even and odd particle numbers as in the previous experiment was
involved. The large amplitudes, however, pose a serious problem. Even when compared with the theoretical estimates
for the typical current cited above, these values are larger by one to two orders of magnitude. In comparison with the
theoretical average current the discrepancy becomes even worse.
The third persistent current experiment by Mailly, Chapelier, and Benoit
573
poses less of a problem for theory. In
this experiment, the magnetic response of a single mesoscopic GaAlAs/GaAs ring was measured. This ring had an
internal diameter of 2 m and an elastic mean free path = 11m. It was therefore in the ballistic regime. In contrast
to the previous two experiments, where the channel number M was of the order of 10
4
, M was quite small, M = 4.
A very interesting feature of the semiconductor ring was the possibility to connect it via an electrostatic gate to two
external leads. A second gate was placed on one of the arms of the ring, making it possible to suppress all interference
eects. With this unprecedented control over the ring it was possible to measure transport and thermodynamic
properties in the same system. The current observed in the ring had the value I
0
= ev
F
/L, exactly as expected for
an almost clean sample with very few channels.
In summary, the persistent currents predicted theoretically were observed in three independent experiments. In the
114
diusive regime, the measured values exceeded theoretical expectations by at least one to two orders of magnitude.
This poses a serious challenge for theory.
2. Analytical theory for noninteracting electrons
The need to calculate the disorder average for the persistent current within the canonical (rather than the grand
canonical) ensemble
570
led to three almost simultaneous publications by Schmid
574
, v. Oppen and Riedel
575
, and
Altshuler, Gefen, and Imry
576
. The calculations for the average persistent current used the impurity perturbation
technique. In Refs. 574,576, the constraint of constant particle number N was met by letting the chemical potential
uctuate, = + , and by choosing in such a way that to rst order in N remained xed. In
Ref. 575, this approximation was avoided by relating the average current to the typical uctuation of a single level,
and by calculating the latter quantity using standard Green function techniques. Strictly speaking, the perturbative
results derived in
574576
are divergent at zero temperature T and for vanishing ux . However, this divergence is
regularized either by nite values of T, or by the energy scale 1/

set by inelastic processes. Both of these scales

can realistically be assumed to be at least of the order of the singleparticle level spacing . With this assumption,
all three investigations yielded for the rst even harmonics I
m
of the persistent current the values
I
m

ev
F
LM
=
I
0
M
, (6.45)
where hv
F
/LM. At rst sight this seemed to contradict earlier numerical results
570
according to which the
average current behaves as I) I
0
(/LM)
1/2
. However, it was pointed out in Ref. 576 that the rst
_
E
C
/
harmonics contribute to the current provided the regularizing cutos discussed above are of order . (Here, as
always, E
C
= hT/L
2
is the Thouless energy and T the diusion constant). With E
C
/ = g = M/L and g the
dimensionless conductance, these considerations led to a maximal amplitude of
I
max

E
C

ev
F
L
_

ML
, (6.46)
in good agreement with Ref. 570. Unfortunately, these analytical results did not remove the discrepancy between
theory and experiment: As stated above, the result (6.46) is too small by one to two orders of magnitude.
Altland et al.
577
used Random Matrix Theory and the supersymmetry method to calculate the persistent current
nonperturbatively at zero temperature in the model of independent electrons. In this work, a novel approach was
used to calculate the impurityensemble average within the canonical ensemble. The condition that all rings in the
ensemble carried an integer (but not the same) number of electrons was imposed. The resulting expressions turned
out to be equivalent to those obtained from the rstorder expansion in referred to above. The nonperturbative
treatment removed the divergence typical for perturbative results and yielded the current at T = 0, where all cutos
lose their physical signicance. This work
577
also contained the rst calculation of the crossover from orthogonal to
unitary symmetry within the supersymmetry formalism. The crossover occurs as the ux increases from = 0 (where
orthogonal symmetry applies) to nonzero values. In keeping with the scope of our review, we discuss this application
of Random Matrix Theory in more detail than previous work.
The persistent current I() can be expressed as follows,
I() = 2
N

n=1
E
n

= 2
_
E
0
dE

n=1
E
n

(E

E
n
)
_
. (6.47)
The factor 2 accounts for spin degeneracy. To account for the constraint of xed particle number N, the upper
integration limit E has to be adjusted as a function both of ux and of disorder realization. This is impossible in
practice. The diculty is avoided as follows. We recall that the rst experiment
571
used a wafer with 10
7
rings. It is
natural to expect that the particle numbers on these rings, although integer, were dierent. This suggests averaging
I over a range K of particle numbers centered at some value N
0
K. The associated eigenvalues E
N
dening the
upper limit of integration lie in an energy interval S,
I() =
2
K
_
S
dE
_
E
0
dE

n=1
E
n

(E

E
n
)
_ _

N=1
(E E
N
)
_
. (6.48)
Here, E
N
is the energy for which there are exactly N electrons on the ring. The second function with innitesimal
positive increment in Eq. (6.48) guarantees the discreteness of the particle number in the ensemble of rings. At rst
115
sight nothing has been gained, the problem of adjusting E has just been replaced by the problem of adjusting the
full interval S. However, as is explained in detail in
577
, the dependence of S on ux and disorder can be neglected
for K 1. Expressing the functions in Eq. (6.48) in terms of Green functions we get for the ensembleaveraged
persistent current
I()) =
2
K
_
S
dE
_
E
0
dE

g(E, E

)
g(E, E

) =
1
4
2
tr
_
H

1
E
+
H
_
tr
_
1
E

H
_
+ c.c. . (6.49)
This is equivalent
577
to the expression I) =

__

0
dE tr[(E H)]
_
2
employed in Ref. 576 (apart from a sign error
in Ref. 576).
In Ref. 577, the model Hamiltonian (cf. Sec. II C),
H() =
1
2m
_
p e

A
_
2
+V (x)
V (x) = 0; V (x)V (y) =
1
2
(x y) (6.50)
was used. Here, = 1/V is the density of states (with V the volume of the ring) and is the elastic scattering
time. (Instead of a Gaussian random white noise potential as in Eq. (6.50), the Hamiltonian could have been chosen
as a chain of coupled GOE matrices, see the IWZ model in Sec. VI C3. Both models lead to the same nonlinear
model). Introducing a supersymmetric generating functional for the Green functions as well as for the ux derivatives
of H, one can write the average current as
I()) =
1
8
2
K
_
S
dE
_
E
0
dE

_
_

s
_

j
Z[
s
,

, j]

j=0;s==
+ c.c.
_
,
Z[
s
,

, j] =
_
d[] exp
_
i
2
_
d
2
x

(x)L(D +jI
2
)(x)
_
. (6.51)
The denitions of the supersymmetric quantities appearing in Eq. (6.51) can be found in Ref. 577. For the present
purpose the following explanations hopefully suce, cf. Sec. III C. The inverse propagator D is basically given by
E H with suitable additional matrix structure to account for (i) the dierent energy arguments E, E

and (ii) the

necessary extension to the 8dimensional superspace. The quantities
s
,

are the ux arguments of those copies

of H which are associated with the BosonBoson block and the FermionFermion block, respectively. The matrix I
2
is diagonal with elements 1 and 1. After the usual steps (averaging, HubbardStratonovich transformation, gradient
expansion) we obtain the following nonlinear model,
I()) =
1
8
2
K

2
T e
8L
_
S
dE
_
E
0
dE

_
d[Q] trg[QI
2
] trg[Q [

K, Q]]
exp
_

8
_
trg[T (Q)
2
+ 2E LQ]d
3
r
_
, (6.52)
where

K is a suitable source matrix, where E = EE

, and where = i

+(e/L)[
3
, ]e

is a covariant derivative
with the gradient

and the tangential unit vector e

. We note that Eq. (6.52) is of the form of Eq. (6.1) except for
the redenition Q iQ. The matrix
3
= diag (1, 1) introduced at the end of Sec. VI A1 breaks the symmetry
between those parts of the 88 supermatrix Q that are connected by the operation of time reversal. As in the case of
the spectral uctuations discussed there, this matrix reects the necessity to calculate the crossover from orthogonal
to unitary symmetry for ux values near = 0. The divergence in the perturbative approaches discussed above
originates from this crossover, more precisely from the zero mode in Eq. (6.52) (i.e. from the contribution with
Q Q
0
= const.) which cannot be treated adequately in a perturbative framework. Actually, symmetry breaking in
this zero mode also poses a technical problem in the context of the supersymmetric formulation. The problem arises
from the term trg([
3
Q
0
]
2
) in the covariant derivative, cf. the exponent in Eq. (6.52). This problem was overcome
116
by a trick which is a precursor of the parametrization of the T matrix in Eq. (6.14). Focussing attention on the zero
mode, one gets
I
0
()) =
1
8
3/2

E
C

0
K(y) (6.53)
with
K(y) =

y
_

0
dx
_
d[Q
0
] exp
_
y trg [Q
0

3
Q
0

3
] +
x
4
trg [Q
0
L]
_
trg [Q
0
I
2
]
_
trg [Q
0

3
Q
0

K] 4
_
+ c.c. , (6.54)
where y = (E
C
/) (/
0
)
2
and x = E/. After the nal integration over Q
0
the result can be well parametrized
by the expression I) = (

E
C
/
0
) (
1

y/(1+
2
y)) with
1
2.25 and
2
12.40. This gives a maximal amplitude
of I
max
0.3

E
C
/
0
, in agreement with the estimates given in Ref. 576. The modes with spatially varying Q elds
can be treated perturbatively and added to the zero mode result to obtain a full picture of the current
577
.
In a closely related study, Efetov and Iida
578
calculated the persistent current from a dynamic response. This point
of view has the advantage that the dynamic response is not sensitive to the use of the canonical or the grand canonical
ensemble. Within the model (6.50) of noninteracting electrons in a ring threaded by a magnetic ux and with a
slowly varying eld
B(t) = B
0
+B

cos(t), (6.55)
the density of the oscillating current can be written in linear response as
j

= K()A

. (6.56)
Here, K() is the currentcurrent correlation function and A

is the vector potential corresponding to the magnetic

eld B

. At rst sight it is plausible that the current calculated from Eq. (6.56) in the limit of vanishing frequency
0 should be basically equivalent to the thermodynamic current (see, however, the discussion below). The quantity
K() is essentially given by the correlator

R() = p
r
G
+
E
(r, r

)p
r
G

E
(r

, r) , (6.57)
where p
r
= (1/m)[i+eA] is the velocity operator and G

E
the retarded (advanced) Green function. This correlator
can be expressed in the framework of the supersymmetry method as a functional average . . .)
Q
over a eld of Q
matrices
578
,
. . .)
Q
=
_
(. . .) exp (S[Q]) DQ (6.58)
with
S[Q] =

8
_
trg
_
D[QieA[Q,
3
]]
2
+ 2iLQ
_
dr . (6.59)
This is, of course, absolutely analogous to Eq. (6.52). Without specifying the precise form of the terms in the angular
brackets in Eq. (6.58) we conne ourselves to saying that the procedure is analogous to the one explained above in
the context of the thermodynamic approach. Again, the calculation may be restricted to the zero mode, i.e. to a
spatially constant Q eld. The nal result for the current has a maximal amplitude of the order of

E
C
/
0
, in
agreement with Ref. 577. However, the shape of the current as a function of the magnetic ux diers considerably
from the one found in the thermodynamic calculation.
The relation between the thermodynamic and the dynamic approaches was addressed by Kamenev and Gefen
579
.
These authors investigated a whole variety of averaging procedures, diering (i) in the initial preparation of the
system under study and (ii) in the constraints imposed while the external eld is varied adiabatically. According to
Ref. 579, the thermodynamic calculation corresponds to the situation where both the initial occupation of levels and
the evolution of the system with varying ux are determined within the canonical ensemble. The dynamic response,
on the other hand, corresponds to a particular procedure: The initial occupation of levels is dened grandcanonically
but is then held xed as determined by the canonical ensemble. Hence there is no reason why the persistent currents
calculated in both cases should coincide exactly. On physical grounds, however, one does not expect drastic (order
ofmagnitude) discrepancies.
The same authors also investigated the magnetoconductance
580
and the conductance distribution
581
of isolated rings
in the quantum regime, where the level widths are smaller than the level spacing. Signicant dierences between the
canonical and the grand canonical ensemble were found.
117
3. Theories with electronelectron interaction
After it had generally been recognized that the experimental values for the persistent current could not be explained
in the framework of theories which neglect the electronelectron interaction, numerous investigations appeared which
addressed the role of these interactions. A detailed discussion of this vast body of work is completely beyond the
scope of the present review which focusses on Random Matrix Theory. We restrict ourselves to a brief enumeration
of relevant work. The diculty in attempts to treat both disorder and interaction lies in the fact that disorder
alone poses severe theoretical problems (witness the present section) while interactions add the full complexity of the
nonrelativistic manybody problem.
In an early paper by Ambegaokar and Eckern
582
(actually preceding the full solution of the noninteracting
problem
574577
), the grand potential of a mesoscopic normal metal ring including electronelectron interactions was
calculated to rst order in the interaction. Evaluating both the Hartree and Fock diagrams, the authors found that
the average persistent current is parametrically given by E
C
/
0
, in apparent agreement with Ref. 571. The same
result was later derived in Ref. 574 from the condition of local charge neutrality. Also, the temperature dependence
of the result, see Ref. 583, was closer to experiment than obtained in theories without interactions. However, dia-
grams of higher order in the interaction renormalize the Cooper channel and were estimated to reduce the result to
roughly one order of magnitude below measured values
584
. A calculation of the typical current along similar lines
585
resulted in a very large value, I
typ
ev
F
/L, but was later corrected
586
to I
typ
ev
F
/L
2
= E
C
/
0
. Work using
density functional theory
587
also led to relatively large currents of order E
C
/
0
, but the absence of renormalizing,
higherorder corrections could again not be established. Recently, the results of Ref. 582 could be reproduced within
a simple (nondiagrammatic) HartreeFock picture
588
.
In a series of papers, the persistent current problem was approached from the point of view of Luttinger liquid theory.
From very early work
589
it was known that in a Luttinger liquid the eect of a single impurity is exponentially enhanced
(suppressed) for repulsive (attractive) interactions. It was possible to generalize the formalism to include the case of
twisted boundary conditions necessary to describe a clean mesoscopic ring in a magnetic eld
590
(see also Ref. 591).
Nevertheless the Luttinger liquid, predicting enhanced impurity scattering for interacting electrons, did not seem to be
a promising candidate for explaining the discrepancy between theory and experiment. Interacting Fermions in clean,
onedimensional rings were also investigated in Ref. 592 by means of the Bethe ansatz. In a rened treatment
593
of
the eect of impurities which improved on the work of Ref. 589 by going beyond simple perturbation theory for the
disorder, spinless Fermions in one dimension and with repulsive interaction were investigated. An algebraic (rather
than exponential) decay of the current with the ring circumference was found. On the other hand, MonteCarlo
simulations
594
indicated that for a strong impurity potential, the current may be enhanced. A signicant new aspect
was contributed to the discussion by Giamarchi and Shastry
595
, who took into account the spin degree of freedom.
In earlier work
596
on the localization properties of onedimensional disordered electron systems a combination of
the replica trick (to average over disorder) and renormalization group methods was used. The renormalization group
equations derived in Ref. 596 were employed in Ref. 595 to prove that repulsive interactions can enhance the persistent
current, provided the carriers have spin (see Ref. 597, and also Refs. 598,599). This work emphasized the important
role of the additional spin degree of freedom and (indirectly) gave rise to the interesting question whether a second
channel (i.e., a model beyond the strictly onedimensional Luttinger liquid) would play a similar role. It was pointed
out in Ref. 600 that the enhancement just mentioned does not lead back to the value for the clean ring.
Another approach
601
invoked the classical electromagnetic energy associated with charge uctuations and long
range Coulomb forces to derive a very large average current of the order of ev
F
/L. The critique
602
leveled against
this work basically points out that the assumed coexistence of large uctuations and longrange forces is highly
implausible.
A new mechanism explaining why the Coulomb interaction might indeed counteract the inuence of the disorder
potential was suggested and investigated in Ref. 603. The key point was a phasespace argument: By mixing many
particle states in a rather large energy interval, the Coulomb interaction redistributes the strength of the impurity
potential to highlying congurations, thus reducing the inuence of impurity scattering on the ground state relevant
for the persistent current. Numerical simulations supported this claim, but the relation of the model used to real
physical systems remained controversial.
Much important work used computer simulations of lattice models of interacting particles. Small onedimensional
systems of spinless Fermions could be treated by numerical diagonalization of the Hamiltonian, but no enhancement
of the persistent current was found
604607
. In Ref. 604 this fact was (at least partly) attributed to the onset of the
MottHubbard transition which sets in as the interaction strength increases. These results were conrmed in Ref. 608
in the framework of a HartreeFock approach. Similar conclusions were also reached in a recent study using the
densitymatrix renormalizationgroup algorithm
609
. Numerical investigations of one and threedimensional rings
610
which included rstorder corrections due to interactions, found (among other results) agreement with Ref. 582 in
118
the diusive regime.
Exact diagonalization is possible only when the system sizes are severely restricted. It is an important task to
devise reliable numerical approximation schemes capable of dealing simultaneously with interactions and disorder.
One step in this direction was the development of the selfconsistent HartreeFock method (SHF)
611,612
. Through
a selfconsistency condition, higherorder terms in the interaction are eectively resummed. Therefore, the method
goes beyond the analytical calculations in Ref. 582. With this method, an enhancement of the persistent current has
been found in threedimensional rings and for longrange interactions
612
. A similar enhancement was observed after
exact diagonalization of small twodimensional cylinders with longrange interactions in Ref. 613. For shortrange
(onsite and nearestneighbor) interactions in two dimensions, on the other hand, spin again seems to be an important
factor
614
: the current was found to be strongly suppressed in the spinless case but to be signicantly enhanced once
spin was included.
Why do interactions aect the persistent current but not the conductance? This very important question was
addressed by Berkovits and Avishai
615
. The authors expressed the conductance as the compressibility of the system
times the ux derivative of the ensembleaveraged persistent current. This formula leads to the plausible scenario
that an enhancement of the persistent current is compensated by a corresponding reduction of the compressibilty,
with zero net eect on the conductance.
C. Transport in quasi onedimensional wires
The main experimental information on mesoscopic systems is derived from transport studies: Each such system is
connected to voltmeters and/or external voltage dierences by a number of leads, and both the resistance tensor and,
by implication, the conductance tensor are measured. Historically, these experiments raised questions of a conceptual
nature. What is the meaning of a transport coecient like the conductance in a system where inelastic eects and,
hence, energy dissipation are absent by denition? This is not the place to review the heated discussion of nearly
a decade ago. An overview over this subject is given in an article by Stone and Szafer
616
. Suce it to say that
there are two approaches to quantum transport in mesoscopic systems, the Kubo formula
617
based on linear response
theory, and the scattering approach pioneered by Landauer
618
, see Eqs. (6.60) and (6.61) below. The problem of
dissipation in mesoscopic transport was mainly discussed in the framework of Landauers ideas. The key issue here
were the socalled reservoirs, ctitious sources and sinks for electrons. These reservoirs were introduced to represent
the macroscopic contacts attached to the mesoscopic probe. Energy dissipation occurs in, and irreversibility is caused
by, these reservoirs. In this way, the conductance of a mesoscopic system could be meaningfully dened even though
neither dissipation nor any irreversible processes occur in that system itself. This result can be seen as another
manifestation of quantum coherence in mesoscopic systems: The wave function is coherent in the entire domain
connecting the reservoirs.
The study of electron transport in quasi onedimensional (1d) wires is particularly interesting, for several reasons.
In contrast to strictly 1d systems, these wires possess a genuine diusive regime. At the same time, they share many
properties with strictly 1d systems. This is true, in particular, for the applicability of nonperturbative analytical
methods. In the last decade, both the DMPK equation
619,620
and the nonlinear model
44,140
have been successfully
used to describe such wires. This has led to an (almost) complete understanding of the physical properties of these
wires for all length scales L.
By denition, the transverse dimensions of quasi 1d wires are large enough to accomodate a number () of transverse
quantized modes (often referred to as channels) below the Fermi energy. Typically, is of the order 10
2
or 10
3
, hence
1. At the same time, the transverse dimensions are small enough (i.e., of the order of the elastic mean free path
) so that no diusion takes place in the transverse direction. The ballistic regime L < and the localized regime
L > (with = ) are separated by the diusive regime < L < where Ohms law applies.
In a rst part we briey discuss the phenomenon of universal conductance uctuations. This phenomenon triggered
most of the interest in mesoscopic physics and quasi 1d systems. Subsequently, we describe the two nonperturbative
approaches mentioned above. Both originate from or are at least intimately related to the concept of Random
Matrix Theory. Lastly, we review the development leading to the insight that these two approaches, in spite of their
considerable technical dissimilarity, are in fact equivalent.
1. Universal conductance uctuations
In 1984, peculiar uctuations in the conductance of small metallic wires
621,622
and of small metallic rings
623626
were observed experimentally, see Fig. 7. These uctuations occurred as a function of an external magnetic eld. The
119
FIG. 35. Top: The dimensionless conductance g (i.e. the conductance in units of e
2
/h) of a mesoscopic gold loop versus
magnetic eld. Bottom: The Fourier transform of the data in the top gure. Peaks in the spectral weight are abserved near
1/H = 0 (UCF), 130/T (principal AharonovBohm periodicity), and 260/T (higher harmonic). Taken from Ref. 516.
same phenomenon was found in quasi 1d MOSFETs
627630
as a function of gate voltage. The phenomena are reviewed
in Refs. 8,631. The uctuations occurred at Kelvin temperatures in systems that were much larger than the elastic
mean free path , but smaller than the phase coherence length L

, i.e. in what became known as the mesoscopic

regime. Most strikingly, the amplitude of the uctuations at low T was of the order of e
2
/h, independent of the
mean conductance of the sample and of samplespecic details like length, material or geometric shape. Therefore,
these uctuations were dubbed universal conductance uctuations (UCF). The uctuation pattern turned out to
be typical for a given sample and reproducible in consecutive measurements at xed temperature, hence the name
magnetongerprints. This observation ruled out any explanation in terms of timedependent uctuation processes.
In rings like the one shown in Fig. 30, a superposition of UCF and periodic oscillations of principal frequency h/e was
observed, see Fig. 35. This was a demonstration of the mesoscopic AharonovBohm eect
8
.
UCF can be qualitatively explained by estimating the level number uctuations, see Sec. VI A1, but compared
with expectations based on classical physics, the UCF are abnormally large: From Ohms law and in d dimensions,
we have g) L
d2
for the dimensionless mean conductance g). UCF implies var(g) constant. Therefore, the
relative uctuations var(g)/g)
2
of the conductance behave as L
2(2d)
. For a network of classical resistors of size L in
d dimensions, on the other hand, one can show
632
that the relative uctuations var(g)/g)
2
of the conductance behave
as L
d
: The relative uctuations vanish as the system size goes to innity. This is referred to as selfaveraging.
The existence of UCF implies that there is no selfaveraging for d = 1, 2.
Theoretical investigations by Altshuler, Lee, and Stone revealed that these uctuations were due to complex quan-
tum interference eects
633
typical for the diusive regime. The amplitude was indeed found to be universal and of
order e
2
/h
634636
. Shortly thereafter, Imry
637
further elucidated the origin of UCF by assuming that the longrange
stiness in the spectra of Gaussian Random Matrix Theory carries over to certain parameters x
j
characterizing the
transfer matrix of a disordered segment. With this assumption the universality of the conductance uctuations fol-
lowed immediately. While the assumption was not quantitatively correct, it helped to develop a simple and intuitive
120
understanding of UCF (cf. the discussion of Ref. 184 at the beginning of Sec. VI A1). We mention that UCF have
much in common with but are dierent from Ericson uctuations, cf. Sec. IVA.
The analytical modeling used an ensemble of impurity potentials of the kind described in Sec. II C. A general
problem of stochastic modeling (rst encountered in RMT in the context of nuclear physics, see Sec. II A) resurfaced
in the present context: To compare ensemble averages with data, an ergodic hypothesis had to be invoked. It states
that an ensemble average over the impurity potential is equal to the running average (taken over a range of values
of the magnetic eld strength or of the gate voltage) on a xed member of the ensemble (i.e., the physical sample).
This assumption is discussed in some detail in the comprehensive paper by Lee, Stone, and Fukuyama
632
.
For a quantitative analysis, the multichannel Landauer formula has been the most useful starting point. This
formula has acquired considerable importance in the interpretation of transport properties of mesoscopic systems. It
expresses the dimensionless conductance g in terms of the quantummechanical transmission amplitudes through the
mesoscopic device.
To derive this formula from linear response theory, i.e., from the Kubo formula
617
, we consider a nite quasi 1d
disordered region in d dimensions with volume V = L
d1

L. Innite ideal leads with identical cross sections L

d1

are
attached to both ends of this rod of length L. In this model geometry, the zero temperature Kubo formula for the
conductance reads
g =
_
2
L
_
2

k,l

_
V
d
3
r
1
2im
(

l
)

2
(E
F
E
k
)(E
F
E
l
) .
(6.60)
Here, E
F
is the Fermi energy,
k
are the eigenstates of the system (including the leads) and E
k
the corresponding
eigenvalues. The integral extends over the disordered region, and the derivatives are to be taken in the longitudinal
direction of the quasi 1d rod. Lee and Fisher
638
have shown that Eq. (6.60) is equivalent to
g =

a,b
([S
RL
ab
[
2
+[S
LR
ab
[
2
)
= 2

a,b
[S
RL
ab
[
2
, (6.61)
where S
RL
ab
and S
LR
ab
are those elements of the scattering matrix that connect the right lead (R) with the left one (L)
and vice versa, respectively. Eq. (6.61) is known as the multichannel Landauer formula
639,638
. It establishes a direct
link between the conductance and scattering theory, and has been extended to the case of many external leads by
B uttiker
640
. The proof of equivalence of the Kubo and the Landauer approach was further generalized by Stone and
Szafer
616
and by Baranger and Stone
641
. (In the rst of these papers, there are some misleading statements about
the lack of unitarity of the scattering matrix which are due to an incorrect normalization of the wave functions).
At rst sight, Eq. (6.61) seems absolutely obvious: Aside from a factor 2 for spin, g is given by the total quantum
mechanical transmission probability through the mesoscopic device. At second sight, one begins to wonder, however: g
is expressed in terms of the scattering matrix. This quantity describes the dissipationless passage of electrons through
the sample. How is it possible, then, that g is a transport coecient which accounts for dissipation? This is the issue
discussed at the beginning of this subsection. Here, we only emphasize that Eq. (6.61) is another manifestation of
quantum coherence in mesoscopic devices. Because of this coherence, it is problematic to dene a conductivity, which
would require local dissipation in the probe. Rather, properties of the entire device enter into the denition of the
conductance.
The actual calculations employed impurity perturbation theory. An early overview is given in Ref. 632. This
technique is tailored towards the calculation of (products of) Green functions, and the conductance must rst be
expressed in terms of such functions. This can be done either by expressing the S matrices in Eq. (6.61) in terms of
such functions
140
, or by using in Eq. (6.60) the identity
r[G
+
(E) G

(E)[r

) = 2i

k
(r)
k
(r

)(E E
k
) , (6.62)
where G

(E) = (E

H)
1
. The resulting expression for g is bilinear in the Green functions. After averaging over
the ensemble, the only nonvanishing contribution has the form G
+
(r, r

; E
F
)G

(r, r

; E
F
). In essence, the calculation
of such a term uses an expansion of the Green functions in powers of the impurity potential V (r). (For the denition
of V (r), cf. Eq. 2.14). The ensuing perturbation series is well dened in the diusive regime and can be reordered to
yield a systematic expansion in powers of the disorder parameter (k
F
)
1
. To leading order in (k
F
)
1
, the Fourier
representation of a single averaged Green function is given by
642
121
G
+
(p, p

) =
(p p

)
E
F
p
2
/2m+i/2
. (6.63)
The average Green function acquires a nite lifetime = /v
F
(the elastic mean free time). In coordinate space, the
Green function decays on the scale of the elastic mean free path. Again to leading order in (k
F
)
1
one can show
642
that the expression G
+
(r, r

; E
F
)G

(r, r

; E
F
) is given by the diusion propagator (r, r

), a solution of the equation

T(r, r

) = (r r

). (6.64)
Here, T = v
2
F
/3 is the diusion constant. Calculating higher moments of the conductance amounts to calculat-
ing averages of higher products of Green functions and becomes increasingly complicated. For the variance of the
conductance it turns out
634,635
that
var(g)
_
d
3
r
_
d
3
r

(r, r

)(r

, r) = tr
2
. (6.65)
For a quasi 1d wire without spinorbit scattering and obeying timereversal symmetry, Eq. (6.65) yields the universal
result 8/15
635
. For all three symmetry classes with = 1, 2, 4, the result is var(g) = 8/(15).
The impurity perturbation technique sketched above cannot give a complete picture of the transport properties of
quasi 1d systems because it fails (as does any perturbative approach) in the localized regime. It is therefore necessary
to consider alternative, nonperturbative approaches. This is the program for the remainder of this section on quasi
1d transport.
2. DorokhovMelloPereyraKumar equation
A rst nonperturbative approach deals directly with the transfer matrix M of a disordered conductor. This matrix
is dened as follows. We consider a disordered sample connected to two ideal leads. The number of channels in each
lead is denoted by . The matrix M expresses the dimensional vectors

in
and

out
in one lead (which describe
the incoming and outgoing wave amplitudes in that lead) in terms of the corresponding dimensional vectors (
in
and
out
) in the other lead,
_

out

in
_
= M
_

in

out
_
. (6.66)
The matrix M diers from the scattering matrix S. The latter connects the outgoing wave amplituds in either lead
with the incoming ones,
_

out

out
_
= S
_

in

in
_
. (6.67)
The two matrices M and S are, of course, not independent but are connected by a nonlinear equation not given
here. Both matrices yield equivalent descriptions of the scattering process. The virtue and usefulness of the transfer
matrix lies in a multiplicative composition rule: The transfer matrix of a number of disordered regions arranged in
series and pairwise connected by ideal leads, is the product of the transfer matrices of the individual regions. This
property is used in the construction of the transfer matrix in the local approach described below. It relates directly
to the scaling theory of localization.
Combining Eqs. (6.66) and (6.67) with the Landauer formula (6.61) one can show that the conductance is given in
terms of the radial eigenparameters
i
(to be dened below) of M as
g = 2

i=1
1
1 +
i
2

i=1
T
i
, (6.68)
where we have dened the transmission eigenvalues T
i
= 1/(1 +
i
). The task is then to determine M and, thus, g)
or higher moments of g.
Every microscopic realization of a disordered quasi 1d wire is associated with a particular transfer matrix M. This
fact suggests the basic idea of the transfer matrix approach: To dene an appropriate ensemble of transfer matrices
that correctly reects the transport properties of an ensemble of disordered wires. There are two dierent ways
to do so, the global and the local approach. In the global approach developed by Pichard, Stone et al.
643645
, the
122
statistical properties of the full transfer matrix are determined in a single step: The maximum entropy approach
is used in conjunction with physically motivated constraints to determine an ensemble of transfer matrices. In the
local approach by Dorokhov
619
, Mello, Pereyra, and Kumar
620
, the maximum entropy approach is employed to dene
an ensemble of transfer matrices for a small (but still macroscopic) piece of wire. The decomposition of the wire
associated with this model is similar to the one employed in Sec. VI C3 for the IWZ model, see Fig. 36. The full
ensemble of transfer matrices for the entire system is then obtained by putting many of these small pieces in series,
i.e. by multiplying the corresponding transfer matrices. This procedure denes a Brownian motion model for the
transfer matrix and leads to a FokkerPlanck equation for its eigenparameters
i
. This is the DorokhovMello
PereyraKumar (DMPK) equation. In the following we focus attention on this latter theory. The global approach
has been reviewed in detail in Ref. 646.
We briey sketch the derivation of the DMPK equation. We follow a formulation in Ref. 647 which was applied
by Frahm and Pichard
448,648
in a slightly dierent context, namely for a Brownian motion model for the S matrix.
We restrict ourselves to the case of unitary symmetry ( = 2). The relevant symmetry of M is dictated by current
conservation and reads
646
M

z
M =
z
, (6.69)
where
z
=diag(1, 1) in the block representation dened by Eq. (6.66). Putting M = exp(X) this implies
X =
_
a b
b

d
_

(6.70)
with a

= a and d

= d.
Using this symmetry, we now consider M(x) as a function of a variable x which describes the length of the wire.
We wish to nd a dierential equation for M(x). We dene a random process by
M(x +x) = e
M
M(x) (6.71)
where M is a random matrix which must have the same symmetry as X. For simplicity, we identify M with X and
introduce the following statistical assumptions (which can be justied in terms of a maximum entropy approach
620
):
M has a Gaussian probability distribution with vanishing rst moment and uncorrelated matrix elements a, b, d. The
second moments are given by
tr(a

a) = tr(d

d) = W
f

2
x ,
tr(b

b) = W
b

2
x. (6.72)
We recall that is the number of channels. The quantities W
f
and W
b
dene the relative strength of the diagonal
and nondiagonal elements, respectively.
It is useful to decompose the transfer matrix into radial and angular degrees of freedom,
M(x) = Ue
H
V , M(x +x) =

Ue
H+H

V , (6.73)
where U = diag(u
1
, u
2
), V = diag(v
1
, v
2
), and H =
x
(and similarly for

U,

V , and H). Here u
i
and v
i
are unitary
matrices, is a diagonal matrix of radial parameters, and
x
is the rst Pauli matrix. The quantities
i
are related
to the eigenparameters
i
in Eq. (6.68) through
i
= sinh
2
(
i
). It is the aim to express the variation H in the
radial paramters as a function of the statistical input (6.72) for the random process (6.71). This is achieved by an
expansion to second order in the small quantities b and b

. The quantities of central interest are the increments

i
of the radial parameters. Their detailed properties follow from Eq. (6.72). For = 2 the result reads

j
=
1
2
W
b
x
_
_

k(=j)
(coth(
j

k
) + coth(
j
+
k
)) + 2 coth(2
j
)
_
_
,

k
=
1
2
W
b
x
jk
. (6.74)
This denes a standard Brownian motion process leading immediately
126
to a FokkerPlanck equation for the distri-
bution function p(, x) of the radial parameters (DMPK equation)
123

x
p(, x) =
W
b
4

j
_

j
2

k(=j)
(coth(
j

k
) + coth(
j
+
k
))
2 coth(2
j
)
_
p(, x) . (6.75)
The dierential operator on the r.h.s. of Eq. (6.75) is the radial part of the Laplacian on the space of transfer
matrices
649
. Introducing as new variables the eigenparameters
i
= sinh
2
(
i
), we get an alternative form of the
DMPK equation,

x
p(, x) =
2

j
J()
j
(1 +
j
)
j
J
1
() p(, x). (6.76)
With J() =

j<k
[
j

k
[

, the denition for the elastic mean free path = 1/(W

b
( + (2 )/)) and the
localization length = ( + 2 ), Eq. (6.76) is valid for all three symmetry classes.
The DMPK equation has been solved by Beenakker and Rejaei
650,651
in the unitary case ( = 2). It is obviously
of considerable interest to have an exact solution for the distribution function p(, x). The actual motivation for this
work had another source, however: A puzzling discrepancy between the global approach
643645
and diagrammatic
perturbation theory. In the global approach it is assumed that all correlations among the eigenvalues
i
stem from
the Jacobian J() introduced above. (Formally, this is just the square of the Vandermonde determinant discussed in
Sec. III A2). Therefore the distribution function must have the form
p
glob
= e
(

i<j
u(i,j)+

i
V (i))
(6.77)
with the logarithmic twobody repulsion u(
i
,
j
) = ln[
i

j
[ coming from the Jacobian. It was shown by
Beenakker
652
, however, that Eq. (6.77) leads to var(g) = 1/2 (instead of the correct value 8/15), irrespective of
the form of the conning potential V (
i
). Hence the logarithmic repulsion of eigenparameters in Eq. (6.77) had
to be incorrect. The correct interaction between the
i
was obtained with the help of a variant of the Sutherland
transformation
650,651
. This transformation maps the DMPK equation (6.75) onto a Schrodinger equation in imaginary
time. For = 2, it has the form

x
= (HU) (6.78)
with
H =
1
2

i
_

2
i
+
1
sinh
2
2
i
_
, (6.79)
= + 2 , and U = /2 ( 1)/ ( 1)( 2)
2
/6. The Hamiltonian in Eq. (6.79) is the
sum of singleparticle Hamiltonians. The corresponding singleparticle Green functions can be explicitly calculated
and combined to give G(, x[), the manyparticle Green function of Eq. (6.78). From this function, one obtains
p(, x)
650,651
. Simplifying this expression for the metallic regime < x < , and transforming back to the
i
, one
nds
650,651
,
p(, x) e
(

i<j
u(i,j)+

i
V (i,x))
u(
i
,
j
) = ln [
j

i
[/2 ln

arsinh
2
_

j
arsinh
2
_

/2 ,
V (
i
, x) =

2x
arsinh
2
_

i
(1 +O(
1
)) . (6.80)
As the transmission eigenvalues T
i
= 1/(1 +
i
) go from 1 to 0, the
i
go from 0 to 1, and the twobody interaction
term crosses over from ln[
j

i
[ to half that value. This shows that it is the failure of the global approach to
correctly describe the behavior of weakly transmitting channels which is responsible for the slight discrepancy in the
UCF.
The exact solution p(, x) was used by Frahm
653
to calculate all mpoint correlation functions
R
m
(
1
, . . . ,
m
, x) =
!
( m)!
_

0
d
m+1
. . . d

p(, x) . (6.81)
124
This was possible with the help of a generalization of the method of orthogonal polynomials
23
. With m, n, j = 1, . . . , ,
p(, x) could be written as the product of two determinants of matrices of dimension ,
p(, x) det[W
n1
(
j
, x)]det[h
m1
(
j
, x)], (6.82)
The functions W
n
and h
m
are dened in Ref. 653 and fulll the biorthogonality relation
_

0
dW
n1
(, x)h
m1
(, x) =

nm
. Dening the functions
K

(,

; x) =
1

m=0
W
m
(, x)h
m
(

, x) (6.83)
one nds that the mpoint functions have the form of a determinant,
R
m
(
1
, . . . ,
m
, x) = det[K

(
i
,
j
; x)
1i,jm
] . (6.84)
From R
1
and R
2
, Frahm
653
calculated explicitly the rst two moments of the conductance g =

i
1/(1 +
i
). The
results turned out to be equivalent to those obtained in the framework of the nonlinear model described in the
next section.
Extending this formalism, Frahm and M ullerGroeling
654
calculated in the unitary symmetry class the autocorre-
lation function g(L)g(L + L)) for all length scales. Within the framework of the DMPK approach, this was the
rst calculation of a fourpoint correlation function. Explicit analytical results were obtained in the diusive regime
and in the strongly localized regime. In the diusive regime where t = L/2 1 and for large channel numbers
the correlation function is given by a squared Lorentzian,
g(t)g(t + t)) =
4
15
1
(1 +
t
t
)
2
+O(t) . (6.85)
This result can be viewed as a generalization of the universality of var(g). Equation (6.85) is independent of any
absolute length scale, and the scale on which the correlation function decays with t is set by the system size t itself.
In the strongly localized regime t, t 1, on the other hand, one gets
g(t)g(t + t)) (t)
3/2
t
3/2
e
t
e
t
. (6.86)
In this limit the localization length sets an explicit scale for the decay of the correlations. In the crossover regime
between metallic and localized behavior the analytical expressions derived in Ref. 654 had to be evaluated numerically.
Examination of these results explicitly conrmed the view that a wire of length L can be thought of as being
composed of independently uctuating segments of length .
3. Nonlinear model for quasi 1d wires
The second approach to transport in quasi onedimensional systems uses statistical properties of the Hamiltonian
of the system under study rather than of the transfer matrix. For this reason, it is sometimes called microscopic, in
contrast to the macroscopic DMPK approach. We prefer another terminology. We will refer to the DMPK approach
and the model approach also as to the random transfer matrix approach and the random Hamiltonian approach,
respectively. In the latter approach, ensemble averages are calculated with the help of Efetovs supersymmetry
method
44,46
, see Sec. III C. This is close to Efetovs original application of his method. In fact, starting from a
singleparticle Hamiltonian with a random white noise potential, Efetov
44
proved among other things that all
wave functions in innite quasi onedimensional systems are localized for all three symmetry classes. The rst model,
however, to properly take into account the geometry discussed above including the coupling to the two leads, was
introduced by Iida, Weidenm uller, and Zuk (IWZ)
140
. We briey describe its main features.
The disordered region is modeled as a chain of K segments, each roughly of the size of the mean free path , see
Fig. 36. The leads attached to the right and the left of the disordered region support
1
and
2
channels, respectively.
Without the leads, the system is described by the Hamiltonian H
ij

, where the indices i, j refer to the segments, while

the indices , denote the N states per segment. As usual, the limit N is taken. The elements of H are real
( = 1), complex ( = 2), or quaternion ( = 4) numbers. States within one segment labelled i are represented by
the matrix H
ii

with the probability distribution

P(H
ii
) exp
_

N
4v
2
1
tr(H
ii
)
2
_
. (6.87)
125
FIG. 36. Schematic illustration of the IWZ model. The disordered region (shaded) is divided into K segments. Ideal leads
are attached to the right and the left of the disordered region.
States in adjacent segments are coupled by the matrices H
ij

([i j[ = 1) with the probability distributions

P(H
ij
) exp
_

N
2
2v
2
2
tr(H
ij
H
ji
)
_
. (6.88)
Matrix elements H
ij

with [i j[ > 1 vanish identically. Matrix elements carrying dierent upper indices (i, j) ,= (k, l)
are uncorrelated. The Hamiltonian of the entire disordered region, a matrix of dimension KN, is denoted by H.
The coupling to the leads is eected by a xed (nonrandom) rectangular matrix W = W
1
+ W
2
of dimension
KN (
1
+
2
). It has elements W
i
n
, where i and identify a segment and a state, respectively, in the disordered
region and where n refers to the channels in the leads. It is assumed that only the rst (the last) segment couple to
the left (the right) lead, repectively. Thus the elements of W
1
are nonzero only for i = 1 and 1 n
1
, while the
elements of W
2
are nonzero only for i = K and
1
< n
1
+
2
. For this model, the scattering matrix S can be
written down explicitly,
S = 1 2i W

(E H+iWW

)
1
W . (6.89)
The ensemble of Hamiltonians denes an ensemble of scattering matrices. The latter determines the statistics of the
conductance via the Landauer formula (6.61).
The supersymmetric generating functional for a product of eective propagators (E H iWW

)
1
appearing
in Eq. (6.89) has the form
Z[J] =
_
d[H]P(H) (6.90)
_
D exp
_
1
2
i

L(E H+iW
1
W

1
L +iW
2
W

2
L +iL +J)
_
.
After averaging, HubbardStratonovich transformation, integration over the elds and saddlepoint
approximation
140
, the generating functional takes the form of a generalized nonlinear model,
Z[J] =
_
D[Q] exp
_
v
2
2
2v
4
1

i
trg[Q
i
Q
i+1
]

1
2
Trg ln(E Q+iW
1
W

1
L +iW
2
W

2
L +J)
_
. (6.91)
Setting E = 0 the matrices Q
i
obey the constraint Q
2
i
= v
2
1
. The matrix Q is blockdiagonal, with diagonal entries
Q
1
, Q
2
, . . . , Q
K
. The symbol Trg implies a trace over the segment label i and the state label .
The functional (6.91) has served as the starting point of several papers dealing with the diusive regime and/or
the localized regime of quasi onedimensional mesoscopic systems. In the diusive regime, where the dimensionless
conductance g is much larger than unity, the functional (6.91) can be evaluated perturbatively as follows
140
. The
saddlepoint manifolds of the matrices Q
i
, i = 1, . . . , K are parametrized in terms of a suitable set of independent
variables. Terms of second order in these variables are proportional to g
140
and are kept in the exponent. The
exponential containing all terms of higher order than the second is expanded in a Taylor series. This reduces all
integrations to an application of Wicks theorem. The expansion generated in this way is an asymptotic expansion
in inverse powers of g. In this way, conductance uctuations were investigated in the orthogonal
140
, unitary and
symplectic
655
symmetry classes. Phase breaking processes could be taken into account in a phenomenological way
656
.
The inuence of a magnetic eld on the conductance was studied in ring structures by calculating the weak localization
correction
657
as well as the autocorrelation function versus magnetic eld strength
658
. Resistance uctuations were
investigated in detail in three
659
, four
660
and multilead
661
devices. Some, but not all, of the results in these papers
have been obtained earlier by means of the impurity perturbation technique (see Ref. 632 and references therein). The
126
perturbative evaluation of the supersymmetric functional for the IWZ model is involved but conceptually relatively
simple. In its results, it is equivalent to the impurity diagram technique.
A nonperturbative treatment requires considerably more eort but leads to results which go beyond the domain
of applicability of the impurity diagram technique. The appropriate technique, Fourier analysis on supermanifolds,
was developed by Zirnbauer
662
. This technique was used by Zirnbauer to calculate the mean conductance
141
and by
Mirlin, M ullerGroeling, and Zirnbauer to calculate the variance of the conductance
663
in quasi onedimensional wires
for all length scales and for all three symmetry classes. We now sketch the calculation for the orthogonal symmetry
class.
For
1
=
2
= and in the limit of large channel number 1, the rst two moments of the conductance can
be expressed as integrals over the coset space of the Q matrices (with the redenition Q Q

= iQ/v
1
and omitting
the prime)
663
,

1
= g) =

2
2
_
D[Q] Q
51
1
Q
51
K
exp
_

8
trg[(Q
1
+Q
K
)L]
_
exp
_

32

i
trg(Q
i+1
Q
i
)
2
_
,

2
= g
2
) =

4
4
_
D[Q] Q
51
1
Q
62
1
Q
51
K
Q
62
K
exp
_

8
trg[(Q
1
+Q
K
)L]
_
exp
_

32

i
trg(Q
i+1
Q
i
)
2
_
. (6.92)
The superscripts on the preexponential terms indicate the appropriate element of the 8 8 supermatrix Q. We have
introduced the quantity = 8(v
2
1
/v
2
2
). By comparison with Ref. 44 it can be identied with the localization length
for the orthogonal case. Integrating over all Q matrices except for those referring to the rst and the last segment,
Q
1
and Q
K
, we obtain the heat kernel

W(Q
1
, Q
K
; L/2) =
_
dQ
2
. . . dQ
K1
exp
_

32

i
trg(Q
i+1
Q
i
)
2
_
,
(6.93)
where L = K is the system size (not to be confused with the supermatrix L occuring, e.g., in Eqs. (6.90) and (6.91)).
In the continuum limit and with the metric dened by the length element trg(dQdQ)/16 one can show
663
that

W
fullls the heat equation

W(Q, Q

; s) =
Q

W(Q, Q

; s); lim
s0

W(Q, Q

; s) = (Q, Q

). (6.94)
Here, s = L/2, (Q, Q

) is the function and

Q
the LaplaceBeltrami operator (Laplacian) on the coset space.
It turns out that all the information about the rst two conductance moments
1
and
2
is contained in the heat
kernel. Using for the Q matrices the parametrization Q = exp(X)Lexp(X), one obtains the following expansion of
the heat kernel in powers of X,

W(e
X
Le
X
, L; s) = 1
1
2

1
trgX
2
4
+
1
8

2
_
trgX
2
4
_
2
+. . . (6.95)
We notice that the coecients of the two leading terms in this expansion are the rst two moments of the conductance.
Fourier analysis was invented to solve the heat equation and is expected to work here as well. Equation (6.94)
presents a particularly complicated case, however, since it is dened on a curved supermanifold. Nonetheless, the
basic idea remains the same. Here, we only give the result. For GOE symmetry, one nds

n
(L) =

2
_

0
d tanh
2
(/2)(
2
+ 1)
1
p
n
(1, , )T(1, , )
+2
4

l2N+1
_

0
d
1
d
2
l(l
2
1)
1
tanh(
1
/2)
2
tanh(
2
/2)
p
n
(l,
1
,
2
)

,1,2=1
(1 +l + i
1

1
+i
2

2
)
1
T(l,
1
,
2
), (6.96)
127
FIG. 37. The variance of the conductance var(g) for the GOE (solid line), GUE (dashed line), and GSE (dotted line) versus
L/.
where T(l,
1
,
2
) = exp(2

(l,
1
,
2
)L/4),

(l,
1
,
2
) = l
2
+
2
1
+
2
2
and
p
1
(l,
1
,
2
) =

(l,
1
,
2
) ,
p
2
(l,
1
,
2
) = (
4
1
+
4
2
+ 2l
4
+ 3l
2
(
2
1
+
2
2
) + 2l
2

2
1

2
2
2)/2 . (6.97)
Corresponding results for the GUE and GSE were also calculated in Ref. 663. While var(g) =
2

2
1
showed
exponential localization as a function of system length for GOE and GUE symmetry, the curve for the GSE case
approached a nite limiting value as L/ . This suggested that for the GSE there is no localization. A similar
behavior had also been obtained for the mean conductance
141
. However, as discussed below, this unusual result
turned out to be due to a subtle technical mistake
664
. In Fig. 37 we show the correct behavior of var(g) for all three
symmetry classes
663,664
.
4. Equivalence of nonlinear model and DMPK equation
The coincidence in both approaches of the rst two moments of the conductance for the GUE
653
suggested at least
a very close connection between the two methods. Further evidence in this direction was provided in Ref. 665. Here
the average density of transmission eigenvalues (see Eq. (6.68)) in quasi 1d wires,
(T) =

i
(T T
i
)) , (6.98)
was calculated for = 2 within the model approach. The result was identical with the one obtained from the
DMPK equation
665
. This eectively extended the equivalence of the two approaches for the GUE to the average of
all linear statistics, i.e. of all quantities A that can be written as A =

i
A(T
i
). On the other hand, the surprising
results for g and var(g) obtained in the symplectic symmetry class within the nonlinear model
141,663
were in
striking disagreement with results obtained from the DMPK equation (which ruled out any delocalization
653
). Thus,
a clarication of the relation between the DMPK equation and the model for all was of considerable interest.
128
This problem was solved by Brouwer and Frahm
664
who proved the equivalence of the two approaches for all three
symmetry classes. We give a brief account of the main steps in the proof.
Before doing so we point out that the claimed equivalence holds only for the distribution function of the conductance
but not for all parametric correlations of g (like the autocorrelation function versus an external magnetic eld, or the
weak localization correction). This is because the DMPK approach is more limited than the model approach, for
the following reason. Within the random Hamiltonian approach, the calculation of parametric correlations is always
possible, at least in principle, via a suitable modication of the random Hamiltonian (the actual calculation may
be very cumbersome, of course). In contradistinction, the calculation of parametric correlations within the random
transfer matrix approach leads to principal diculties: In general it is not clear how to implement a parametric
dependence into the transfer matrix in a dynamically meaningful way
449,448
.
The main idea of the proof consists in dening a suitable generating functional Z(

, ) which combines the key

quantities of both methods, namely the eigenparameters
i
of the transfer matrix and the radial paramters
j
of the
supersymmetric Q matrix. The latter are dened by
Q = T
1
LT ,
T =
_
u
1
0
0 v
1
_
exp
_
0

/2

/2 0
_ _
u 0
0 v
_
, (6.99)
where

=diag(

a
,

b
) and

a
=
_

1

2

2

1
_
;

b
=
_
i
3
i
4
i
4
i
3
_
. (6.100)
Additional symmetry restrictions are imposed on the
j
for = 1, 2, 4
664
. The proof of equivalence proceeds by
showing that Z(

, ) obeys both, the DMPK equation (6.75) with respect to the variables , and the heat equation
(6.94) with respect to the variables

. If in addition Z(

, ) fullls the proper initial conditions, the proof is complete.

A variant of the supersymmetric generating functional for the Green functions of the IWZ model serves this purpose,
Z[Q] =
_
D exp
_
1
2
i

L(E H+iW
1
W

1
Q+iW
2
W

2
L +iL)
_
.
(6.101)
The precise denitions of the quantities involved here can be found in Ref. 664. It can be shown
664
that Z depends on
Q only through

. Moreover, the dependence of Z on the Hamiltonian H and on the coupling matrix elements W
1
, W
2
can be replaced by a dependence on the transmission eigenvalues T
i
= 1/(1 +
i
). Thus, Z has the required form
Z = Z(

, ). Furthermore, all transport quantities that can be obtained within the minimal model (i.e. without
increasing the dimension of Q) can be generated from Eq. (6.101) by taking proper derivatives with respect to the
j
.
We average Z over the ensemble of Hamiltonians H and denote the result by a bar (Z). We then use the heat
kernel

W(Q
1
, Q
K
; s) dened in Eq. (6.93), and we dene the functions
f
1(2)
(Q, Q
1(K)
) = exp
_

d
2

n
trg ln(1 +x
n
QQ
1(K)
)
_
, (6.102)
with d = 1(1, 2) for = 1(2, 4). The quantities x
n
are essentially the eigenvalues of W
1
W

1
(for f
1
) and W
2
W

2
(for
f
2
), respectively. As in Eq. (6.92), the functional Z can be expressed as an integral over the two Q matrices which
refer to the rst and the last segment, respectively,
Z[Q] =
_
dQ
1
_
dQ
K
f
1
(Q, Q
1
)f
2
(L, Q
K
)

W(Q
1
, Q
K
; s) . (6.103)
Because of the important symmetry property
Q
f
1
(Q, Q

) =
Q
f
1
(Q, Q

), the averaged supersymmetric functional

Z also fullls the heat equation with the radial part

of the full Laplacian

Q
,

s
Z(

, ) =
Q
Z(

, ) =

Z(

, ) . (6.104)
We now onsider another average of the generating functional Z(

, ), dened by the use of the probability distri-

bution p(, s), a solution of the DMPK equation (6.76), as weight factor. This average is indicated by the superscript
T,
129
Z(

, )
T
=
_
d
1
. . . d
N
Z(

, ) p(, s). (6.105)

It is obvious that Z
T
obeys the DMPK equation,

s
Z(

, )
T
= (2/)D

Z(

, )
T
, (6.106)
where D

is the dierential operator on the r.h.s. of Eq. (6.76) (without the factor 2/).
To prove the identity of both approaches, it is shown
664
that

Z(

, ) D

Z(

, ) . (6.107)
Substituting this result into Eq. (6.106) and interchanging the average over and the dierentiation with respect to

, one nds that Z and Z

T
obey the same dierential equation (6.104). Consideration of the initial conditions in both
cases
664
establishes the equivalence of both methods.
This insight implies, however, that the results for the GSE in Refs. 141,663 are incorrect. (We recall that these
results dier from those of the DMPK approach). Inspection shows that the nite (i.e., nonzero) limiting values
(L ) for the rst two moments of the conductance are technically due to a zero mode of the Laplacean (6.94).
In Ref. 664 it was shown that in the GSE case, this zero mode is not a singlevalued function on the space of Q
matrices, and that it does not exist when the Kramers degeneracy is correctly taken into account. Therefore, the
contribution due to this mode in Refs. 141,663 to the Fourier expansions and to the expressions for the moments
of the conductance, has to be omitted. In this way, corrected expressions for the symplectic symmetry class were
obtained in Ref. 664. These expressions show localization as expected from the DMPK equation.
D. Random band matrices
Originally, Random Matrix Theory was introduced by dening a number of matrix ensembles which describe
generic features of many microscopically dierent systems. An important ingredient was the postulated invariance
(orthogonal, unitary or symplectic) in Hilbert space. This concept was shown to have enormous power: Physical
systems are grouped into few universality classes characterized by very general symmetry properties. Many examples
were given in the previous section where the three Gaussian ensembles played a prominent role. It is natural to ask
whether generalized ensembles of random matrices exist from which the generic properties of extended systems can
likewise be generated. These ensembles must, of course, contain the linear dimensions of the system in parametric
form. The IWZ model described in Sec. VI C3 suggests that such ensembles do indeed exist, at least for quasi 1d
wires, and that they dier from the classical ensembles by relinquishing the requirement of (orthogonal, unitary or
symplectic) invariance. We now show that quasi 1d wires can be modeled by an ensemble of random band matrices
(RBMs). Early investigations of RBMs are summarized in chapter 20 of Mehtas book
23
.
A band matrix is characterized by the dimension N and by the bandwidth b. The latter denes the distance from the
main diagonal beyond which the matrix elements become either zero or are negligibly small. An ensemble of RBMs
is obtained by assuming that the matrix elements within the central band are uncorrelated Gaussian distributed
random varibles with zero mean and a variance which eectively denes the energy scale. As long as b N the RBM
ensembles are clearly equivalent to Wigners Gaussian ensembles, while for b = O(1) we deal with an almost diagonal
matrix with uncorrelated eigenvalues and strongly localized eigenfunctions. This shows that RBMs represent one way
to interpolate between chaotic systems with WignerDyson statistics and Poisson regularity (see Refs. 135,136,261
for another possibility). The wish to describe intermediate level statistics was indeed the original motivation which
led Seligman, Verbaarschot, and Zirnbauer
488
to the investigation of RBM ensembles. Many numerical investigations
of these ensembles by Casati et al. and Wilkinson et al. followed
666669
.
In Sec. VI C3, we have introduced the Hamiltonian of the IWZ model
140
. As mentioned above, this Hamiltonian,
although endowed with some ne structure, has the general form of a random band matrix. Another physical system
that has been mapped on a random band matrix is the kicked rotor
47,67,667
. It has been argued
47
that in a certain
basis the evolution operator S connecting values of the wavefunction before and after a single kick, exhibits the
structure of a RBM. For the kicked rotor, one classically expects unbounded diusion in angular momentum space
once the strength of the kick exceeds a critical value. Quantum mechanically, however, dynamical localization
sets in and the diusion is eventually suppressed, see Sec. VB. A rst connection between dynamical and Anderson
localization was established in Ref. 330. Recently, the correspondence between the kicked rotor and RBMs (and
hence the correspondence between dynamical and Anderson localization) has been made more exact. Altland and
Zirnbauer
670
have developed a eld theory for the timedependent Hamiltonian
130
H(t) =

l
2
2
+k cos(

+a)

n=
(n t) , (6.108)
which denes the quantum kicked rotor. This approach is still under discussion, see Ref. 671. The quantities

l and

are the operators corresponding to the angular momentum l and the angular coordinate , k determines the kick
strength and its period, and a is a symmetry breaking parameter similar to a magnetic eld. Using a novel variant
of the Hubbard Stratonovich transformation the authors of Ref. 670 claim that the long wavelength behavior of
Eq. (6.108) is described by the onedimensional nonlinear model. As we will see, the same is true for RBMs.
These two examples suce to show how much insight can be gained by analysing RBMs.
Numerical simulations
668
have shown that the uctuation properties of observables derived from RBMs are governed
by the scaling parameter x = b
2
/N. For completeness we mention that a second scaling parameter arises if the diagonal
elements of the RBM are allowed to increase linearly in size with increasing index
669
. Fyodorov and Mirlin
139
studied
RBMs analytically and derived a nonlinear model which turned out to be identical to the one found by Efetov
44
for noninteracting electrons in a quasi 1d wire. This proved the close correspondence between RBMs and random
quasi 1d systems. If one associates the matrix size N with the length L of the sample and uses the result (see below)
that the localization length scales like b
2
, it becomes clear that x is proportional to the conductance g = /L. We
now sketch the proof.
The ensemble of NN real and symmetric band matrices H has independent Gaussian distributed random elements
H
ij
= H
ji
with zero mean value and variance
H
2
ij
=
1
2
A
ij
[1 +
ij
] . (6.109)
Here, A
ij
decays (suciently fast) on the scale b. The asymptotic behavior for large [lm[ of the twopoint correlation
function
K
lm
=
_
(E +i H)
1
_
ll
_
(E i H)
1
_
mm
(6.110)
denes the localization length . With the help of the averaged supersymmetric functional
Z[J] =
_
d[] exp (S[, J]) ,
S[, J] =
1
2
trg

ij

i
(A
1
)
ij

j
+
1
2
trg

i
ln(E
i
+iL +J
i
) (6.111)
the average of the correlation function K
lm
can be expressed as
K
lm
=
1
4

2
Z[J]
J
1
l
J
2
m
. (6.112)
In Eq. (6.111), the
i
(i = 1, . . . , N) are a set of 8 8 supermatrices (not yet restricted to the saddlepoint manifold)
and J is a suitably dened
139
source eld with elements J
(1,2)
l
(l = 1, . . . , N). The saddlepoint approximation is
justied because N 1 and b 1. In the continuum limit one denes x/d = [l m[ where d is some suitable small
length scale, e.g. the lattice constant and arrives at
K
lm
=
_
d
4
_
2
_
d[Q] F(Q(0), Q(x)) exp (S[Q]) ,
S[Q] =

8
_
trg
_
T(Q)
2
+ 2iLQ

dx. (6.113)
Here is the density of states per unit length. To dene the diusion constant T, we write A
ij
= a([i j[) = a(x/d),
dene B
2
= (1/2)

a(r)r
2
b
2
and have T = dB
2
. The precise form of the preexponential term F(Q(0), Q(x))
in Eq. (6.113) can be found in Ref. 139. The 8 8 supermatrix Q belongs to the saddlepoint manifold discussed
in Refs. 44,46. Thus, the onedimensional model of Eq. (6.113) is identical to the one of Ref. 44. The localization
length for this model has been calculated in Ref. 44 and is given by
l
loc
/d = 4D/d b
2
. (6.114)
131
Invoking the oneparameter scaling hypothesis
672
we see that the scaling properties of a RBM can depend only on
the ratio of localization length and matrix size N. According to Ref. 139, this explains the scaling parameter b
2
/N
found numerically in
668
.
In Refs. 673675,398, the investigation of RBMs was extended to the statistical properties of the corresponding
eigenfunctions. The main results are summarized in Ref. 676. First, Fyodorov and Mirlin showed how to formulate
various statistical observables involving eigenfunctions in the framework of the nonlinear model, i.e. for arbitrary
dimension of the disordered sample. Second, they solved the model in one dimension and calculated the distri-
bution functions of the following three quantities for arbitrary system length L: the nth component
(k)
n
of the kth
eigenfunction, the inverse participation ratio
P
2
=

n
[
(k)
n
[
4
, (6.115)
and the product r(L) = [
(k)
0

(k)
L
[
2
of amplitudes taken from opposite ends of the wire. The distribution of r(L)
turned out to be lognormal in the insulating regime. This implies a Gaussian distribution for the localization lengths.
Technically, these results were derived by expressing the desired observable, in particular its moments, in terms of
arbitrarily high products of Green functions. Employing the supersymmetry formalism these moments could then be
expressed in terms of (the discrete variant of) the model in Eq. (6.113). The onedimensional chainlike structure of
the Q matrices suggested the use of a method
677
patterned like the transfer matrix approach. This led to a recurrence
relation and eventually to a comparatively simple dierential equation
673
. In subsequent papers
674,398,675
appropriate
generalizations of this method were used.
E. Transport in higher dimensions
Many of the rigorous results reviewed so far in this section on transport properties are restricted to quasi one
dimensional systems. This is mainly due to the fact that the model has been well under control in the one
dimensional case but not in higher dimensions. Very recently, however, important progress has been made, and
interesting information was obtained, from the model for d = 2 and d = 3. It is useful to briey recall the history
of the nonlinear model and of scaling theory prior to a discussion of this development. We thereby extend the
historical remarks of Sec. II C. For simplicity of notation, and deviating from former usage, we use in the present
subsection g for the average conductance and g for the dierence between the conductance and its average value.
1. Scaling theory and distributions
In Sec. II C it was mentioned that early scaling ideas by Thouless
34
were developed into a oneparameter scaling
theory by Abrahams et al.
672
. This theory received strong support from perturbative calculations by Anderson, Abra-
hams, and Ramakrishnan
678
, Abrahams and Ramakrishnan
679
, Gorkov, Larkin, and Khmelnitskii
680
, and from eld
theoretical considerations involving a nonlinear model by Wegner
681
. His eldtheoretical approach, subsequently
developed in a series of papers
682,39,683686
, marked the beginning of a development that led to the omnipresence of
the model (now in its supersymmetric form
44
) in modern discussions of disordered systems.
Scaling theory determines the dependence of the dimensionless conductance g(L) on the linear dimension L of the
system. The initial condition is set by the strength of the disorder as measured by the conductance g
0
= g() at the
scale of the elastic mean free path . The following general picture emerged. For large g the function has the form
(g) =
d ln g
d ln L
=
L
g
dg
dL
= (d 2)
a
g
(6.116)
where the constant a is of order unity. The function is seen to depend on g as the single parameter, hence the
term oneparameter scaling. For (g) > 0 the conductance increases with increasing system size and we deal with a
conductor supporting extended electron states. For (g) < 0, however, the conductance scales to zero with growing
L, and we are in an insulating regime with localized eigenfunctions. As long as d 2, (g) is always negative and one
expects all states to be localized. In three dimensions, however, the asymptotic value of g depends on the initial value
g
0
of the conductance. We introduce (for d = 3) the critical conductance g
c
a, which is dened by the condition
(g
c
) = 0. For g
0
> g
c
we have a conductor and for g
0
< g
c
, an insulator. The point g
0
= g
c
denes the mobility
edge or metalinsulator transition. This picture is qualitative but in a nutshell represents todays view of disordered
132
systems. However, reasons have emerged to call into question the basic tenets of oneparameter scaling. This has
come about as follows.
Altshuler, Kravtsov, and Lerner
687689
emphasized that the transport properties of a disordered sample cannot be
characterized by the mean conductance g alone for which the abovementioned scaling theory had been formulated.
Instead, anomalously large sampletosample uctuations have to be accounted for even in the weak localization
regime (see Sec. VI C1 on universal conductance uctuations, UCF). It was therefore proposed in Ref. 687 to study
the whole distribution of the conductance, and of the density of states, and to develop a scaling theory for all
moments of these quantities. Before we briey outline the method used to pursue this program we summarize the
main results
687,689
on conductance uctuations some of which contradict the oneparameter scaling hypothesis.
We focus attention on the nth cumulants (G)
n
c
) of the conductance G = (e
2
/h)g. For n < g
0
it was found that
(G)
n
c
) g
2n
_
e
2
h
_
n
. (6.117)
The rst cumulant is equal to g itself and changes with L according to Eq. (6.116). The second cumulant, i.e. the
variance of G, does not depend on g and therefore stays constant in the metallic regime (UCF). Higher moments are
small by powers of g
1
and can be neglected for g 1. Hence, for not too large uctuations g and for g 1, the
distribution function is almost Gaussian,
f(g) exp
_
(g)
2

, (6.118)
and oneparameter scaling applies: The distribution function is completely determined by the mean conductance.
Close to the localized regime with g 1, however, all higher moments are of comparable size and the distribution
function is strongly nonGaussian. It was emphasized in Refs. 687,689 that this picture is not entirely correct
because it neglects certain classes of perturbative corrections to the highorder cumulants of the conductance. Thus,
deviations from the Gausssian form (6.118) must occur in the distant tails of the distribution even in the regime of
weak localization where g 1. In these tails, the distribution turned out to be logarithmically normal,
f(g)
1
g
exp
_

1
4u
ln
2
_
g
g
1

__
(6.119)
where u = ln (
0
/), where is the conductivity of the system with linear dimension L, and where
0
is the classical
(Drude) conductivity. (We recall that is the elastic mean free time and is the average singleparticle level spacing).
In two dimensions, is identical to the conductance g(L) while
0
corresponds to g
0
.
The nonGaussian tails in Eq. (6.119) can be traced back to additional contributions to the cumulants of g arising
from the abovementioned perturbative corrections,
(G)
n
c
)
add
g
0
_

L
_
2nd
e
u(n
2
n)
_
e
2
h
_
n
. (6.120)
For n < g
0
these terms are insignicant but for n > g
0
they dominate those in expression (6.117) and lead to the
lognormal tails in the expression (6.119). The tails dominate the distribution function (i) in the metallic regime
g 1 for uctuations g

g
0
and (ii) in the region of strong localization for uctuations g

g
0
/ lng
0
. Exact
results for the distribution function in one dimension
690692
coincide with formula (6.119). An important point in this
expression is the dependence on u, i.e. on both and
0
. This introduces another microscopic parameter in addition
to the scaledependent conductance g = L
d2
. This is at variance with the hypothesis of oneparameter scaling in
its strict form. We briey return to this issue later. A formally similar violation of oneparameter scaling had earlier
been observed by Kravtsov and Lerner
693
in their investigation of the frequency dependence of the diusion constant
T. With the substitutions g and g ), the asymptotic form of the distribution of the density of states is
also given by Eq. (6.119).
2. Renormalization group analysis
First formal investigations employed the replica trick
35
described in Sec. II C. We recall that this method circum-
vents the problem of normalization of the generating functional for the Green functions by considering k replicas of
the system under study. At some point in the calculation the limit k 0 is taken. This procedure is fully justied
only within perturbative treatments
36
. This, however, suces for the present purpose. In Refs. 687689,694697,
133
moments of both the conductance and the density of states were considered. For the former, the replica functional
reads
G
n
)
n

i=1
tr
_

2
h
2
i
_
d[Q] exp(S[Q, h])
_
d[Q] exp(S[Q, h = 0])
_

h=0,k=0
,
S[Q, h] =
1
t
_
tr
_
Q
1
L
[h, Q]
_
2
d
d
r . (6.121)
Here, t
1
= T/8 g, Q is a matrix eld in replica space containing commuting entries. The matrix Q is analogous
to the supermatrix Q in previous sections, and h is a source matrix with components h
i
. The detailed structure of Q
and h is given in Ref. 687. The moments of the density of states are represented analogously,

n
())
n

i=1

i
_
d[Q] exp(S[Q, ])
_
d[Q] exp(S[Q, = 0])

=0,k=0
,
S[Q, ] =
1
2
_
tr(LQ)
d
d
r
L
d
. (6.122)
The components
i
of the source matrix are dened in Ref. 687. The replica matrix Q in Eqs. (6.121) and (6.122)
is subject to the same constraints as its supersymmetric counterpart, namely Q
2
= 1 and tr Q = 0. The combination
of the two actions in Eqs. (6.121) and (6.122) yields
S[Q, h, ] =
_
tr
_
1
t
_
Q
1
L
[h, Q]
_
2

1
2L
d
LQ
_
d
d
r (6.123)
Both formally and in physical content the action S in Eq. (6.123) corresponds to the supersymmetric action (6.16) in
the presence of a magnetic eld. The source matrix h plays the role of the symmetrybreaking term
3
, and is
analogous to the frequency in Eq. (6.16).
Equations (6.121) to (6.123) are based on the usual gradient expansion (long wavelength limit). The authors of
Refs. 687689,694697 made the important observation that additional higher order terms, in particular higherorder
gradient terms, have to be included in the action (6.123). To dene the basic structure of these terms we introduce
the covariant derivative
T

1
L
[h

, ] , (6.124)
where the index refers to the direction in coordinate space. We suppress details of the spatial dependence. Then
the additional contributions to the action (6.123) are given by
S
add
n
[h] = z
n
_
tr(TQ)
2n
d
d
r (6.125)
S
add
m
[] = Y
m
_
tr(LQ)
m
d
d
r (6.126)
and
S
n,m
= z
n,m
_
tr
_
(TQ)
2n
(LQ)
l
d
d
r

(6.127)
with certain (bare) coupling constants z
n
, Y
m
, and z
n,m
.
The perturbative analysis of the replica functional proceeds as indicated after Eq. (6.91) for the supersymmetric
functional. In the perturbation series generated from the action (6.123) without the additional terms in Eqs. (6.125)
to (6.127), a certain class of terms appears. In two dimensions, these terms are logarithmically divergent (in L/). In
principle, such terms can be summed eectively with the help of a renormalization group analysis of the generating
functional. This would lead to a change of the prefactors of the vertices in Eq. (6.123). However, considerations
involving particle number conservation, the Einstein relation, and the continuity equation prove that the action (6.123)
does not change its form under renormalization . It therefore depends on the single parameter t
1
g. This result
is consistent with the oneparameter scaling hypothesis
687
. However, higher moments of both the conductance and
134
FIG. 38. Schematic current response of a disordered conductor to an electric eld of the form E(t) = E(t), where (t)
is the step function. For times greater than the diusion time t
d
, higher quantum corrections lead to lognormal (rather than
exponential) decay towards the equilibrium value E. Taken from Ref. 702.
the density of states are determined
687,689
by the higherorder vertices (6.125), (6.126), and (6.127) which introduce
a large number of parameters, z
n
, Y
m
, and z
n,m
. The additional contributions (6.120) to the higher moments of the
conductance and, therefore, to the asymptotics (6.119) of the distribution function for g are due to such higherorder
terms. The number of diagrams originating from these terms is so high and their structure so complicated that a
direct summation appears to be hopeless. Therefore, one has to perform the renormalization group analysis for the
extended model including the terms (6.125), (6.126), and (6.127).
The principle of such an analysis is the following. In order to distinguish between slow and fast variables, the Q
matrix is parametrized in the form
Q =

U

Q
0

U =

U

0
LU
0

U . (6.128)
Here, the fast variables are given by Q
0
= U

0
LU
0
and the slow ones by

Q =

U

U. Integration over the fast variables

yields a functional of the slow modes

Q alone which serves as a basis for the next iteration of this procedure. Lowest
order perturbation theory for the fully renormalized functional results in an eective summation of the higherorder
terms in the expansion of the bare functional.
Using the results obtained in this way
687689
, several authors
687689,698700,643
have addressed the question whether
or not oneparameter scaling applies to mesoscopic uctuations. To some extent this is a question of denition
689
.
As we have seen, the asymptotic tails of the distribution functions for g and depend on the bare conductivity
0
as
an additional microscopic parameter. Therefore, these tails are not determined by (or g) alone and one might say
that oneparameter scaling is violated. It turns out, on the other hand, that renormalization of the parameters z
n
,
Y
m
, and z
n,m
in the higherorder terms does not inuence the coupling constant 1/t in Eq. (6.121). One could argue
that this is the essence of oneparameter scaling. A more detailed discussion can be found in Ref. 689.
The mesoscopic uctuations just discussed are intimately related to currentrelaxation processes
701,688
. This is
seen by considering the distribution function of the timedependent conductance G(t) which describes the response of
the current to a function pulse (in time) in the potential. This distribution function also exhibits lognormal tails
as in Eq. (6.119), cf. Fig. 38. We return to this point in the context of the supersymmetric model, see Sec. VI E3.
The local density of states (r, ) = ImG
+
(r, r; )/, of particular interest for the metalinsulator transition, was
investigated by Lerner, Altshuler, and Kravtsov
703,689
. In the metallic regime the distribution of (r, ) turned out to
be Gaussian with lognormal tails, as discussed above for the conductance and the density of states. In the vicinity
of the metalinsulator transition, however, the distribution was found to be completely lognormal
703,689
. Recently,
McCann and Lerner
704
showed that this lognormal distribution is linked to multifractal exponents for the cumulants
of the local density of states. Multifractal analysis had been pioneered by Wegner
122
who found a whole set of critical
exponents for the moments of the inverse participation ratio at the metalinsulator transition, and by Castellani and
Peliti
705
who suggested that the electron wave functions at the transition are multifractal. We now describe recent
work which has brought multifractal analysis and properties of asymptotic distributions within reach of the nonlinear
model.
135
3. Supersymmetry approaches
Muzykantskii and Khmelnitskii
511
introduced the novel idea that a nontrivial saddlepoint conguration of the
usual model in Eqs. (6.1) and (6.2) determines the asymptotic tails of the distribution of mesoscopic quantities in
one, two, and three dimensions. Starting from Ohms law in its timedependent form,
I(t) =
_
t

dt

G(t t

)V (t

) , (6.129)
these authors expressed the average timedependent conductance G(t) in terms of the usual supersymmetric functional,
G(t) =
_
d
2
exp(it)
_
d[Q]P(Q) exp(S[Q]) , (6.130)
where the action S[Q] is given by Eq. (6.2). The classical (Drude) term has been suppressed. The next steps can
probably be best understood when viewed as arising from a saddlepoint approximation with respect to both Q and
of the combined action
S
comb
[Q, ] =

8
_
trg[D(Q)
2
+ 2iLQ]d
d
r +it . (6.131)
Variation of Eq. (6.131) with respect to yields the selfconsistency equation
511
_
d
d
r
V
trg[LQ] =
4t

. (6.132)
Variation with respect to Qunder the constraint Q
2
= 1 leads to an equation reminiscent of the Eilenberger equation
706
for dirty superconductors,
2D(QQ) +i[L, Q] = 0 . (6.133)
This saddlepoint approximation should not be confused with the one necessary to derive the model of Eqs. (6.1)
and (6.2). While the latter restricts the Q matrices to the coset space with Q
2
= 1, the present one xes the spatial
variation of the Q matrices within this coset space. The matrix equation (6.133) as well as Eq. (6.132) can be reduced
to two equations involving only one scalar function (r)
511
. Boundary conditions on determine the solutions for
d = 1, 2 and 3.
Two dierent regimes are distinguished by values of the parameter t which denes the size of the r.h.s. of the
selfconsistency equation (6.132). Physically, t determines whether individual states in the mesoscopic probe can
be resolved. For t 1 and with t
D
the diusion time, G(t) is proportional to exp(t/t
D
), irrespective of the
dimensionality. In the opposite limit t 1 the results depend on d. For d = 1 one nds G(t) exp(g ln
2
(t))
and for d = 2, G(t) (t)
g
. Here g is the dimensionless conductance. These results do not apply for arbitrarily
long times because beyond a certain limit, t > t

, the diusion approximation was found to break down

511
. In this
extremely asymptotic regime the function uctuates on a scale exceeding the elastic mean free path: The spatial
derivative obeys

> 1/, and this contradicts the conditions under which is constructed. Therefore, the ultralong
time limit in d = 1, 2 and the regime t 1 in d = 3 seemed to be inaccessible and the comparison with the results
of Altshuler, Kravtsov, and Lerner
689
was inconclusive.
Mirlin
707
observed, however, that the regime t > t

can be reached if the additional constraint

< 1/ is taken
into account in solving Eqs. (6.132) and (6.133). This modies the boundary conditions employed in Ref. 511. For a
twodimensional disk of radius R Mirlin found the following asymptotic behavior in time,
G(t) exp
_

g
4
ln
2
(t/g)
ln(R/)
_
(t (R/)
2
) . (6.134)
This expression agrees exactly with the (slightly amended) result of Ref. 689.
The calculation of eigenfunction statistics, in particular the distribution of eigenfunction components, in the frame-
work of the supersymmetric nonlinear model has been pioneered by Fyodorov and Mirlin
398,676,708,520,709
. We
have already mentioned some of their work in Sec. VI D. The main results for the distribution function f(u) of the
eigenfunction amplitudes u = [(r
0
)[
2
are the following
709
. We consider the function
136
Y (Q
0
) =
_
Q(r0=Q0
d[Q] exp(S[Q]) , (6.135)
where Q(r
0
) = Q
0
is held xed at the observation point r
0
, and the action S[Q] is essentially given by Eq. (6.2) (see
709
for the precise notational conventions). In the limit of vanishing frequency , and for both orthogonal ( = 2) and
unitary ( = 1) symmetry, this function depends on Q
0
only via the radial parameter
1
, Y (Q
0
)

Y (2i
1
).
In terms of

Y (z) the desired distribution functions can be expressed as
398,708
f(u) =
1
V
d
2
du
2

Y (u) ( = 2) ,
f(u) =
2

2
V u
1/2
d
2
du
2
_

0
dz
z
1/2

Y (z +u/2) ( = 1) , (6.136)
where V is the volume of the system. These general formulas are valid for arbitrary dimension. In the limit of
a constant supermatrix Q (zero mode approximation) the functional Eq. (6.135) becomes equivalent to Gaussian
Random Matrix Theory, see Sec. VI A1. In this case we obtain the PorterThomas distribution discussed in detail in
Sec. III E1,
f(u) = V exp(uV ) ( = 2) ,
f(u) =
_
V
2u
exp(uV/2) ( = 1) . (6.137)
For quasi 1d systems, the integral in Eq. (6.135) can be solved exactly
398
. With x = L/ 1 the ratio of system
length and localization length, and the coecient = 2(1 3L

L
+
/L
2
), where L

and L
+
are the distances from r
0
to the sample edges, the distribution function for y = uV reads
398,676,520,709
f
(=2)
(y) = exp(y)
_
1 +
x
6
(2 4y +y
2
) +. . .
_
(y x
1/2
) ,
f
(=1)
(y) =
exp(y/2)

2y
_
1 +
x
6
(3/2 3y +y
2
/2) +. . .
_
(y x
1/2
) ,
f(y) exp
_

2
_
y +

6
y
2
x +. . .
_
_
(x
1/2
y x
1
) ,
f(y) exp
_
2
_
y/x
_
(y x
1
) . (6.138)
In d > 1 dimensions only approximate methods to evaluate Eq. (6.136) are available. Let us dene the parameter
=

q
(2Tq
2
V )
1
, where the summation extends over the eigenmodes of the Laplace operator in the sample.
This sum depends strongly on dimensionality. A perturbative treatment of the nonzero modes in Eq. (6.136) leads
to results
520
identical to the rst two lines in Eq. (6.138) with x/6 replaced by .
For large amplitudes y
1/2
perturbative methods are no longer admissible. Falko and Efetov
710
employed a
saddlepoint approximation similar to the one introduced in Ref. 511 to derive the following result for d = 2, 3,
f(y) exp
_

2
(y +y
2
+ . . .)
_
, (
1/2
y
1
)
f(y) exp
_

8
ln
d
(y)
_
, (y
1
) . (6.139)
Here, = ln(L/)/4
2
T for d = 2 and (k
F
)
2
for d = 3, where k
F
is the Fermi momentum and the elastic
mean free path. In particular it was possible to investigate the inverse participation numbers u
n
=
_

0
u
n
f(u)du.
By studying the dependence of these quantities on system size one can address the question of multifractality of the
wave functions. In two dimensions and for system sizes L smaller than or of the order of the localization length, the
multifractal nature of the wave functions had been established numerically
711,712
, with the result
t
n
L
(n)2
, (n) = (n 1)d

(n) . (6.140)
The fractal dimension d

(n) ,= 2 had been found to depend on the moment considered. In Ref. 710, this fractal
dimension was derived analytically,
137
d

(n) = 2
1

n
2
2
T
, (6.141)
together with the corresponding power law decay in the tails of the wave functions
[(r
0
+r)[
2
r
2
. (6.142)
In Eq. (6.141) the order n of the moment is restricted by the requirement that d

(n) > 0. The exponent in

Eq. (6.142) is smaller than unity and varies from wavefunction to wavefunction.
We see from Eq. (6.139) that the asymptotic behavior of f(y) is again logarithmically normal, in close analogy to
the distribution functions for the conductance
689,511,707
and for the density of states
687,689
discussed above. What
is the origin of all these lognormal tails? Already in the metallic regime, they appear as an early manifestation of
localization, and they become more pronounced as one approaches the localized regime. This suggests the following
interpretation. Even in the metallic regime there exist rare states which are localized. As we approach the localized
regime, there occur rare states with abnormally strong localization. The lognormal tails are due to the occurrence
of such atypical states at any value of g. This connection between atypical states and asymptotic distribution
functions is an active eld of research
709,713715
.
In conclusion it is satisfying to see how recent progress in the evaluation of the model helped to provide strong
support for earlier calculations
687689
of mesoscopic distribution functions based on a combination of perturbation
theory and renormalization group methods.
F. Interactionassisted coherent transport
In 1990, Dorokhov
716
considered particles moving in a strictly onedimensional random potential. He showed that
the localization properties of a pair of particles bound tightly by a harmonic potential dier from those of non
interacting particles. In particular, he predicted a sizeable enhancement of the pair localization length for the case
where the pair energy is sucient to excite highlying modes of the twoparticle system. Four years later, it was
suggested by Shepelyansky
717
that the wavefunction of two interacting particles in a onedimensional disordered
chain might extend over a length scale
2
much larger than the oneparticle localization length
1
, irrespective of
the sign of the twoparticle interaction. Shepelyanskys result indicated that interactions among electrons might
signicantly alter or modify our present understanding of transport and localization properties of mesoscopic systems.
This important development should be seen in connection with the persistent current problem in Sec. VI B. We recall
that this problem has led to the general belief that interactions are vital for the understanding of the equilibrium (i.e.
ground state) properties of mesoscopic rings.
Technically, the problem of two interacting particles in a random potential can be mapped onto an eective Hamil-
tonian which turns out to be a random band matrix with strongly uctuating diagonal entries. Matrices of this type
had not been studied previously, and the twoparticle problem is, therefore, also interesting from the point of view of
Random Matrix Theory.
1. Idea and scaling picture
Shepelyansky
717
considered the following Schrodinger equation for the twoparticle wavefunction
n1n2
in one
dimension,
[E
n1
+E
n2
+U
n1n2
]
n1n2
+ V [
n1+1,n2
+
n11,n2
+
n1,n2+1
+
n1,n21
] = E
n1n2
. (6.143)
Here, n
1
und n
2
are the (discrete) coordinates of the rst and second electron, respectively, E
n1
and E
n2
are onsite
energies drawn randomly from the interval [W, W], U characterizes the strength of the Hubbard interaction, and V
is the hopping integral. All lengths are measured in units of the lattice spacing and are, therefore, dimensionless. For
U = 0, and not too strong disorder, the oneparticle localization length
1
in the middle of the energy band is roughly
given by
1
25(V/W)
2718720
. We dene the twoparticle localization length
2
as the typical scale on which the
twoparticle wavefunction decays. It was predicted
717
that in the interacting case,

2

1
32
_
U
V
_
2

2
1
, (6.144)
138
FIG. 39. Time dependence of the second moments + and dened in the text for W = 0.7 and V = 1 and (from top to
bottom): U = 1 (+), U = 1 (), and U = 0 (+). Taken from Ref. 717.
provided
1
1 and U/V = O(1). This result is noteworthy in two respects: First, for suciently large
1
(i.e.
suciently weak disorder) one can have a signicant enhancement of
2
as compared to
1
. Second, since the
interaction strength U enters the expression (6.144) quadratically, the eect does not depend on the sign of the
interaction.
A rst qualitative numerical conrmation of this eect came from the study of the wave packet dynamics for
Eq. (6.143). The timedependent quantities
+
= ([n
1
[ + [n
2
[)
2
)/4 and

= ([n
1
[ [n
2
[)
2
) (characterizing center
of mass and relative motion, respectively) are shown in Fig. 39. (The angular brackets denote an average over the
wave packet probability distribution at a given time t). An enhancement eect is clearly visible.
The result (6.144) was obtained in two rather independent ways. In Ref. 717 the problem was mapped onto the
abovementioned random band matrix with strongly uctuating diagonal elements which was studied numerically.
The result (6.144) was obtained from a t to numerical data. Imry
721
, on the other hand, invoked a generalization of
Thouless famous scaling block picture to rederive Eq. (6.144). We treat both approaches in turn.
It is useful to rewrite Eq. (6.143) in a basis dened by the singleparticle eigenfunctions of the noninteracting part
of the system. This yields
H
ij,kl
= (
i
+
j
)
ki

jl
+

Q
ij,kl
,

Q
ij,kl
= U

i
(n)
j
(n)
k
(n)
l
(n) . (6.145)
In this representation the Hamiltonian is naturally decomposed into two pieces: a diagonal matrix carrying the
eigenvalues
i
[B, B] of the noninteracting system with bandwidth B 2(2V +W), and a nondiagonal matrix

Q
ij,kl
originating from the interaction operator and dened by the overlap of four eigenfunctions
i
. A typical matrix
element

Q can be estimated
717
as

Q U/
3/2
1
.
Since the wave functions
i
are localized the Hamiltonian (6.145) possesses band structure. Shepelyansky used the
crucial assumption that all elements of the matrix

Q
ij,kl
in Eq. (6.145) can be chosen as independent random variables
centered at zero. In this way, the Hamiltonian in Eq. (6.145) is replaced by an eective random Hamiltonian. The
latter is the sum of two random matrices: (i) A random band matrix with bandwidth
2
1
and Gaussian distributed
elements with variance U/
3/2
1
. (ii) A random diagonal matrix with Gaussian distributed elements having variance
B = 2(2V + W). The addition of the random diagonal matrix to the random band matrix denes a new class of
random band matrices. This new class diers signicantly from an ordinary random band matrix if the diagonal
elements uctuate much more strongly than the odiagonal elements. This condition is met in the present case since

1
1 and U is typically of the order of V . Studying the localization properties of the eigenfunctions of this eective
Hamiltonian by means of the transfer matrix method, Shepelyansky arrived at the result (6.144).
139
In an alternative approach, Imry
721
generalized the ideas of Thouless
517
, see Sec. II C, to the twobody problem.
He divided the system into blocks of size
2
1
. The interaction operator (6.145) couples the twoparticle levels of
one block with those of its neighbor. The generalization of the Thouless argument then says: If the twoparticle
Thouless energy E
(2)
C
is larger (smaller) than the twoparticle level spacing
2
in each of the blocks, the system
is conducting (localized), respectively. Using Fermis golden rule and the estimate

Q U/
3/2
1
given above, Imry
obtained E
(2)
C
=

Q
2
/
2
U
2
/B
1
. Here, B =
2

2
1
is the bandwidth which for
1
1 is proportional to V . In
analogy to the oneelectron case, Imry dened a twoparticle conductance g
2
by
g
2
(
1
) =
E
(2)
C

2
=
U
2
B
2

1
. (6.146)
For
1
1 one can clearly have g
2
1 and the system is far from being localized for electron pairs. Assuming
that g
2
as a function of the length L of the system behaves like a proper conductance, i.e. that g
2
(L) L
1
in one
dimension, and using g
2
(
2
) 1 as the denition of
2
, Imry obtained nally

1
= g
2
(
1
) =
U
2
B
2

1

2
=
_
U
B
_
2

2
1
. (6.147)
Apart from numerical factors which cannot be determined in this approach, the result (6.144) has been reproduced
in a few lines.
2. Microscopic approaches and renements
Both approaches described above rest on rather important assumptions and approximations. We mention, in
particular, the statistical assumptions of the eective random band matrix approach (which are also implicit in the
scaling picture, and which neglect all correlations among matrix elements), and the assertion that g
2
has the scaling
behavior of a proper conductance. Therefore, the work of Shepelyansky and Imry was followed by a number of
investigations
722724
all of which started from the microscopic problem dened by Eq. (6.143).
Frahm et al.
722
employed the transfer matrix method to solve the Schrodinger equation (6.143). These authors
found that in the parameter range considered, V = U = 1 and W = 0.7 . . . 2.0, the twobody localization length
behaved as
2

1
with = 1.65. This was the rst direct conrmation of the enhancement eect. The exponent
diered from the predicted value ( = 2), however. In Ref. 722 this discrepancy was attributed to the strongly
nonGaussian distribution of the interaction matrix elements

Q
ij,kl
in Eq. (6.145). We will see later, however, that
the discrepancy may also be an artifact of tting an exponent in the crossover regime from
2

1
to
2

2
1
.
A very direct demonstration of interactionassisted coherent pair propagation is due to Weinmann et al.
723
. These
authors considered a strongly localized, onedimensional, disordered ring threaded by a magnetic ux. The idea was
to identify coherently propagating electron pairs by the periodic ux dependence of their eigenvalues. The period
should be given by h/2e instead of h/e, the value for single electrons. Indeed, many states with h/2e admixtures
were found in Ref. 723, some of them with almost perfect h/2e periodicity. The associated wave functions showed the
characteristic cigar shape expected when a pair of typical size
1
travels coherently over a distance
2
, cf. Fig. 40.
This result provided compelling evidence for Shepelyanskys eect.
A Green function approach introduced by v. Oppen et al.
724
made it possible to investigate in a very ecient
way the dependence of the twoparticle localization length on the interaction strength U. For the problem dened
by Eq. (6.143), a new scaling parameter could be identied. Irrespective of the symmetry of the twobody wave
function (Bosons or Fermions), it was found numerically that the ratio
2
/
1
depends exclusively on the parameter
combination [ U [
1
/V ,

1
=
1
2
+c
[ U [
V

1
, (6.148)
where c is a number of the order 10
1
. It is remarkable that the dependence on U is linear instead of quadratic as
predicted in Refs. 717,721. Attempts to explain this result are discussed further below. Equation (6.148) implies that
asymptotically,
2

2
1
, in contrast to the exponent = 1.65 found in Ref. 722. The obvious explanation suggested
by Eq. (6.148) is that the results of Ref. 722 apply only in the crossover region where the dependence of
2
on
1
changes from linear to quadratic.
It would be very desirable to have information on the enhancement eect in d > 1 dimensions. Because of the
metalinsulator transition and the question how the mobility edge might be aected by coherent pair propagation,
140
0 40 80 0 40 80
0
40
80
n n
1 1
n
2
FIG. 40. A particular twoelectron wave function (left) of a onedimensional, disordered ring and the corresponding local
current distribution (right) demonstrating interactionassisted coherent transport. In the left gure, dark and bright regions
indicate high and low probability density, respectively. In the right gure, dark and bright regions indicate local currents in
the direction and against the direction of increasing electron coordinates n1, n2, respectively. Taken from Ref. 723.
data for d > 2 dimensions would be of particular interest. However, direct simulations for d = 3 run into severe
numerical diculties since the size of the basis grows with the linear dimension N like N
6
. The only numerical
information on the delocalization eect for d > 1 comes from simulations of the kicked rotor model by Borgonovi and
Shepelyansky
725,726
. If in a 1d model, the time dependence of the perturbing kick is modeled as the superposition
of periodic oscillations with incommensurate frequencies, the resulting problem corresponds to a hopping model in
d = +1 dimensions. A similar correspondence for the case of two interacting kicked rotors can be used to study two
interacting particles in higher dimensions. The gain in numerical eciency is impressive. It amounts to a factor N
4726
.
Two cases were investigated. They correspond to an interacting pair in eective dimensions d
e
= 2 and d
e
= 2.5,
respectively. The main results of Refs. 725,726 can be summarized in terms of the behavior of two quantities,
+
and

, dened as the distance travelled by the pair and as its typical size, respectively. For d
e
= 2, a saturation of
+
as function of time was observed. In the case d
e
= 2.5, however,
+
grew without saturation even in a regime where
single particles are localized. This was a rst numerical indication that pairs might be delocalized earlier than single
particles. The numerical data in Ref. 726 also conrmed qualitative arguments in Ref. 725. It had been predicted
that the size of the diusing electron pairs grows logarithmically with time, while their motion is, correspondingly,
slightly subdiusive: the diusion constant is inversely proportional to ln

t, with of the order of 1.

The linear dependence (6.148) of
2
on the interaction strength U was addressed again in a recent study by
Weinmann and Pichard
146
. Equations (6.146) and (6.147) show that it suces to investigate the Udependence of
the twoparticle Thouless energy E
(2)
C
. Analytical arguments and numerical simulations
146
show that for system
sizes of the order of
1
, there are two regimes for E
(2)
C
: a linear (E
(2)
C
U) and a quadratic one (E
(2)
C
U
2
). The
crossover occurs for E
(2)
C

2
. Unfortunately, this result does not directly lead to the appropriate scaling parameter
[U[
1
/V , and additional statistical assumptions are necessary to explain Eq. (6.148)
146
. Analytical investigations of
this problem in the framework of a suitable random matrix model were performed by Guhr and M ullerGroeling
261
,
who derived a novel analytical expression for the spectral twopoint correlation function. Jacquod, Shepelyansky,
and Sushkov
727
have pointed out that the linear Udependence should only apply close to the band center of the
twoparticle problem. In all other regions of the spectrum they predict an essentially quadratic dependence of
2
on
U.
We have seen that the problem of two interacting particles in a random potential naturally leads to a random band
matrix with strongly uctuating diagonal elements. This class of random matrices generalizes the ordinary random
band matrices studied in the context of chaotic quantum systems and transport in quasi 1d wires
139,673,674,398
. It has
led to partly
728
and fully
729,730
analytical work on the properties of the following ensemble of matrices,
H
nn
=
n

nn
+
nn
/

2b + 1 (n, n

= 1, ..., N) . (6.149)
Here, b is the bandwidth and
n
and
nn
are uncorrelated random variables. In Ref. 728 the
n
s were drawn from the
interval [W
b
, W
b
] and the
nn
s from [1, 1] provided that [n n

[ b. Otherwise,
nn
= 0. In Refs. 729,730 the
distribution of the
n
s was rather arbitrary but characterized by a strength W
b
while the
nn
s had a zerocentered
Gaussian distribution with variance [
nn
[
2
= (1 +
nn
( 2))A
nn
/2. Here A
nn
diers from zero essentially only in
141
the interval [n n

[ b where it has the typical value unity. In both cases, the typical size of the variables is given
by the parameter W
b
and by unity, respectively.
In Ref. 728, two main results for the regime 1 W
b
<

b were established by analytical arguments and numerical

simulations: (i) For a xed unperturbed state with energy
n
, the local density of states
n
(E) = (1/)G
+
nn
(E) is
always of BreitWigner form,

n
(E) =
/2
(E
n
)
2
+
2
/4
, (6.150)
where = /3W
b
. (ii) The scale dened by the localization length = b
2
/2W
2
b
diers from the one set by the
inverse participation ratio
IPR
= (

n
[

(n)[
4
)
1
)

= b
2
/4W
4
b
, where

(n) is an eigenfunction of the Hamiltonian

(6.149) and ...)

denotes a spectral average. This means that typical wave functions have very strong sitetosite
uctuations (We recall that the ensemble (6.149) is of interest only if W
b
1).
Simultaneously and independently, Fyodorov and Mirlin
729
and Frahm and M ullerGroeling
730
investigated the
ensemble (6.149) by means of the supersymmetric nonlinear model. In these papers, the ndings in Ref. 728 were
analytically conrmed and extended. In addition to the localization length and the local density of states
n
(E), the
full set of generalized inverse participation ratios P
q
=

n
[
(k)
n
[
2q
, and the distribution of the eigenvector components
were calculated. These latter quantities were shown to be connected with those for ordinary random band matrices
by an appropriate rescaling with a factor
n
(E)/
729,730
, where is the average density of states. Specialization of
these results to the cases considered in Ref. 728 conrmed all major results obtained there.
A variant of the Hamiltonian ensemble (6.149) was used by Frahm, M ullerGroeling, and Pichard
731,732
to investi-
gate the pair dynamics in arbitrary dimension. Generally speaking, the main result of these papers was the analytical
proof that the concepts familiar from transport theory of noninteracting electrons also apply mutatis mutandis to the
case of electron pairs. A supersymmetric nonlinear model for the twopoint function of the eective Hamiltonian
was derived. Under omission of source terms, the result is
Z
pair
=
_
d[Q] exp(S
pair
[Q]) , (6.151)
S
pair
[Q] =

2
B
2
8B
2
0
_
dRtrg(
R
Q)
2
i

4
h
_

0
(E)
_
dRtrg(QL) .
Here, B
0
and B
2
are numbers characterizing the band structure of the Hamiltonian,
0
(E) is the density of unperturbed
(noninteracting) states, and is the interactioninduced spreading width of these states at the scale
1
. The function
h(/) is dened in Ref. 731,732. Equation (6.151) must be compared with the corresponding expression for transport
of noninteracting electrons in metals
44
introduced in Sec. VI A1,
Z =
_
d[Q] exp(S[Q]) ,
S[Q] =

8
_
trg
_
T(Q)
2
+ 2iLQ

dr . (6.152)
Here, the scaling parameter of localization theory is the conductivity = T. It is clear that Eqs. (6.151) and (6.152)
are very closely related. This led to the following statements: (i) The eective conductivity
e
=
e
()D
e
()
U
2

2
1
/W
2
b
of the pairs plays exactly the same role as the oneparticle quantity . This rigorously justies Imrys scaling
approach
721
. For d = 1 one has
e

2
so that Eq. (6.144) is recovered. (ii) The mean quadratic displacement of
the pairs increases weaker than linearly, R
2
(t)) t/ ln(t)
d
for some time scale so that the pairs move subdiusively.
This conrms numerical ndings
725,726
discussed above. (iii) It is possible to dene a pair Thouless energy E
(2)
C
below
which the spectral statistics of the diusing pair states is of WignerDyson type. For energy scales larger than E
(2)
C
one recovers the AltshulerShklovskii regime
184
.
In summary, we have seen that in its essentials, Shepelyanskys discovery, the phenomenon of interactionassisted
coherent transport of pairs of particles, has been corroborated by a wealth of subsequent work.
3. Finite particle density
The problem of two interacting particles addressed in the previous paragraph is academic, at least for electrons.
Indeed, in the manybody case the exclusion principle may prevent virtual transitions into states which are essential
142
for coherent pair propagation. Hence it is necessary to study coherent pair propagation at nite particle density. Only
two papers have so far (partly) addressed this important issue
721,733
. In Ref. 733, von Oppen and Wettig considered
two interacting quasiparticles above the Fermi surface in one dimension. Formally, the only dierence to previous
work was the blocking of all singleparticle states up to the Fermi energy E
F
. The interaction of particles within the
Fermi sea with each other and with the two quasiparticles was not taken into consideration. It is argued
733
that this
procedure amounts to a lowdensity approximation. It was found that slightly above E
F
, the quasiparticle pair was
localized on the scale
1
. With increasing excitation energy the enhancement eect was eventually recovered, but the
energy scale for this to happen is given by the bandwidth. This result appears to rule out the possibility to observe
coherent pair propagation in an experiment at low temperature in one dimension. Imry
721
has emphasized that the
situation might be more favorable in higher dimensions. He argued that for d = 3 the mobility edge for pairs could be
located below the mobility edge for single particles. Let
m1
and
m2
be the excitation energies of the mobility edges
for single particles and pairs, respectively. Then, according to Ref. 721,

m2

d/2
m1
, (6.153)
where is the critical exponent dened by
1


m1
. Invoking the Harris criterion
734
, d/2 > 1, it is clear that

m2
0 faster than
m1
, as claimed.
Much more work is necessary to clarify the phenomenon of interactionassisted coherent transport in manybody
systems. The whole eld is rather new and progresses rapidly. The most recent developments include a detailed
study of the enhancement mechanism in 1d
735
, a numerical investigation
736
of two quasiparticles in 2d and 3d, and
an unfortunate controversy concerning the role of nite size eects
737,738
.
G. RMT in condensed matter physics a summary
To conclude our overview over developments in condensed matter physics where RMT has played a signicant role
we now try to summarize and judge the importance and prospects of RMT in this eld.
The most direct and therefore also historically rst application of RMT in solid state theory was to spectral
uctuations. Roughly, we can distinguish two periods: One in which research focussed on proving the validity of
WignerDyson (WD) statistics for some range of parameters, and a second one where eorts were concentrated
on identifying deviations from this universal limit. Thus, WD statistics was rst established as a valid means of
description and then used as a standard of reference. Identifying the limits of validity of WD statistics in the
description of spectral uctuations has turned out to be very fruitful. It has led to a better understanding of the
various regimes (clean, ballistic, ergodic, diusive, critical, and localized) in mesoscopic physics. In this context the
model has served at least two purposes. First, it provided the proof that an ergodic regime with WD type of spectral
correlations exists and, second, it turned out to be the proper tool to calculate a number of corrections to this result.
We have seen that such corrections were also discovered with the help of diagrammatic perturbation theory and/or
scaling arguments. The discovery of a third class of spectral statistics at the mobility edge (dierent from both RMT
and Poisson statistics) poses an interesting challenge: The construction of a stochastic Hamiltonian which reproduces
critical spectral uctuations. Although experimental verication of the theoretical predictions seems to be out of
reach for years to come, nding such an ensemble would undoubtedly improve our understanding of the mobility edge.
An interesting recent step in this direction is due to Kravtsov and Muttalib
739
.
Persistent currents in mesoscopic rings are closely connected with spectral uctuations. It is therefore natural to
expect RMT to play an important role in this problem, too. However, the contribution of RMT to the persistent
current problem has been comparatively moderate so far, for several reasons. Initially, one of the main problems was a
conceptual one: How to calculate averages in the canonical ensemble. When this question had been answered and rst
perturbative results had become available, nonperturbative calculations based on RMT did provide the complete and
singularityfree solution of the persistent current problem for noninteracting electrons at zero temperature. This was
an interesting technical achievement, but it did not alter the discrepancy between theory and experiment. Moreover it
is now widely held that the important eect to be understood is the interplay between disorder and electronelectron
interactions. Unfortunately this task is at present beyond the scope of RMT. One may speculate that the embedded
ensembles investigated some 20 years ago in nuclear physics could one day prove useful.
Applications of RMT to transport in quasi 1d systems have been particularly successful. The paradigm quasi 1d
wire has united three formerly separate research activities, the random transfer matrix approach, the random Hamil-
tonian approach of the IWZ model, and random band matrices. Within welldened limits these three approaches
are indeed equivalent. The random transfer matrix method illustrates that RMT is not restricted to modeling the
Hamiltonian of a system. This fact has been known at least since Dyson introduced the circular ensembles of S
matrices. For spectral problems the distribution of eigenvalues and, thus, a modeling of the Hamiltonian are of central
143
importance. In transport problems, on the other hand, it is the distribution of transmission coecients that matters.
These coecients are directly related to the eigenparameters of the transfer matrix. The ensemble of random transfer
matrices corresponding to an ensemble of disordered quasi 1d wires therefore appears to oer the most direct approach
to transport properties. However, an important and sometimes decisive advantage of the Hamiltonian formulation is
its great exibility. Any modication of the system under study is dened in terms of its Hamiltonian and there is
absolutely no question about how to incorporate this change into the Hamiltonian formalism. It is far less obvious
how the transfer matrix has to be adapted to such a modication. The autocorrelation function of the conductance
versus magnetic eld strength illustrates this point. This quantity can easily be formulated within the Hamiltonian
approach and has been calculated in the diusive regime. A similar treatment is, unfortunately, still missing in the
framework of the transfer matrix approach.
Techniques of the model have proven to be particularly valuable. Both the equivalence of the transfer matrix
approach and the IWZ model, and the relation between random band matrices and quasi 1d wires were established
with this technique. Random band matrices are found to play the same role for quasi 1d systems as do the Gaussian
ensembles for zerodimensional systems like atoms, molecules, nuclei, or quantum dots. It is of interest to ask whether
generic ensembles of random matrices exist which model properties of higherdimensional systems.
Problems in d > 1 dimensions have taken center stage in recent work on the nonlinear model. A systematic
improvement of techniques has led from the zerodimensional case (with perturbative corrections) to exact solutions
in quasi 1d systems and, via the recently discovered nontrivial saddlepoint solutions, to substantial progress in
the understanding of two and higherdimensional systems. It has been possible to establish contact with, and to
reproduce results of, earlier perturbative renormalization group calculations. This is the most recent success of a
program which goes ever further beyond the universal limit of the Gaussian ensembles.
Another step towards more general matrix ensembles has been taken in the context of interactionassisted coherent
pair propagation. The eective Hamiltonian describing two particles in a 1d random potential turned out to be a
random band matrix with strong uctuations on the diagonal. The necessity to introduce a new matrix ensemble
originated from the presence of both disorder and interactions. The disorder denes the diagonal entries (localized
states) while the interaction induces couplings between these basis states. Again, the model turned out to be the
appropriate tool to investigate the newly dened ensemble. Dierences and similarities between ordinary band matri-
ces and those with strongly uctuating diagonal elements could be identied, leading to an improved understanding
of coherent pair propagation. It would be desirable to extend this work and to construct a matrix ensemble which cor-
rectly accounts for generic features of Hamiltonians with one and twobody operators. The improved understanding
of the interplay between disorder and interactions might also be helpful in the persistent current problem.
We have seen that RMT has found wide applications in condensed matter physics. These applications go far beyond
the scope of the original Gaussian ensembles. In this development, the supersymmetry method has been an invaluable
tool. Mapping a random matrix model or a microscopic, eective Hamiltonian onto a nonlinear model reduces the
physical content to its very essence: Seemingly dissimilar models turn out to be equivalent, at least in certain regimes.
Several examples were mentioned above. The supersymmetry approach is also useful to establish the correspondence
between dynamical localization in the kicked rotor and Anderson localization in a thick metallic wire, or to show that
interactionassisted propagation of pairs can be described in terms similar to oneparticle diusion. The list can be
continued. Still, there are many open questions. We have mentioned phenomena for which no adequate matrix model
has been constructed up to now. Doing so remains a challenge for the future.
VII. FIELD THEORY AND STATISTICAL MECHANICS
In previous sections (Secs. I to III) of our review we have dealt with the history and with mathematical aspects of
RMT. With this insight into the foundations and the scope of RMT the subsequent applications to many body physics
(Sec. IV), quantum chaos (Sec. V), and disordered systems (Sec. VI) probably came as no surprise. In the present
section we discuss the importance of RMT for elds of physics which at rst sight seem to have little in common with
the more obvious elds of applications just mentioned. In particular we summarize the relation of RMT to interacting
Fermions in one dimension (Sec. VII A), to QCD (Sec. VII B), and to eld theory and quantum gravity (Sec. VII C).
The existence of these exotic relations once again demonstrates the universality and ubiquity of RMT.
A. RMT and interacting Fermions in one dimension
A deep-lying connection exists between classical Random Matrix Theory and special systems of interacting Fermions
in one dimension. This fact was rst demonstrated in work by Calogero
740
and by Sutherland
741
. Recently, Simons,
144
FIG. 41. Connections between various branches of physics. Taken from Ref. 743.
Lee, and Altshuler
189,742
have extended this result to the case of correlation functions. This important work has led to
a whole network of connections between dierent branches of physics. Moreover, it has made it possible to calculate
the (hitherto unknown) timedependent correlation functions of such Fermionic systems using RMT.
We consider the Nparticle Hamiltonian
H
C
=
h
2
2m
_
_
N

i=1

2
i

2
( 2)

i<j
1
(
i

j
)
2
_
_
+

i
V (
i
) (7.1)
where the position of particle i with i = 1, . . . , N in one dimension is given by
i
. It was shown by Calogero
740
that
H
C
is integrable provided the potential is Gaussian, V (
i
)
2
i
. Moreover, for this choice of V the ground state of H
C
denes a probability distribution P
N
(
i
) of the particle positions
i
which coincides
740
with the joint probability
density P
N
(
i
) for the rescaled eigenvalues
i
(see below) with i = 1, . . . , N of the three Gaussian ensembles with
= 1, 2, 4. The positions
i
have to be measured in units of the mean interparticle distance. There is a certain
freedom in dening the statistics of the particles described by Eq. (7.1). Without entering into a detailed discussion
we will think of these particles as spinless Fermions.
Putting the conning potential V (
i
) equal to zero and constraining the motion of the N particles to a circle leads
to the Sutherland Hamiltonian
741
H
S
=
h
2
2m
_
_
N

i=1

2
i

2
( 2)
_

N
_
2
i<j
1
sin
2
[(
i

j
)/N]
_
_
. (7.2)
The square of the groundstate wave function of H
S
is equal to the equilibrium distribution of Dysons circular
ensembles. In the limit N the correlation functions of the Hamiltonians (7.1) and (7.2) are expected to coincide.
We focus attention on the Sutherland Hamiltonian.
To formulate the conjecture made in Ref. 189,742 we recall some results concerning parametric correlation functions
of RMT, see Sec. III H2. We consider the random Hamiltonian
H(U) = H
1
+UH
2
, (7.3)
where H
1
belongs to one of the three Gaussian ensembles, and where H
2
is a xed traceless matrix drawn from the
same ensemble. The dimensionless parameter U measures the strength of the perturbation H
2
. The eigenvalues E
i
of
H are obviously functions of this strength, E
i
= E
i
(U). To display universal aspects of the correlation between the
145
eigenvalues, it is necessary to rescale both U and the E
i
. This is done by dening the rescaled eigenvalues
i
= E
i
/
where is the mean level spacing, and the rescaled strength
u
2
=
_
_

i
(U)
U
_
2
_
U
2
. (7.4)
The resulting universal expressions for the leveldensity autocorrelation functions have been given in Sec. III H2, see
Eq. (3.160). Simons, Lee, and Altshuler
189,742
conjectured that these universal functions coincide with the time
dependent particle density correlation functions of H
S
. To establish such a connection, one must identify dierences
in imaginary time and position with those in perturbation strength u and energy according to the rules
u
2
= 2, = . (7.5)
The rst evidence for the conjecture came from expansions of the timedependent particle density correlation functions
k

(, ). To leading order in and 1/, respectively, these functions were shown to agree with the corresponding limits
of the leveldensity autocorrelators (3.160). Subsequent investigations by the same authors
744,743
and by Narayan
and Shastry
745
established the equivalence of both types of correlation functions exactly. At the same time, Simons
and coworkers
744,743
set up and completed a network of interrelations between dierent branches of physics that is
schematically shown in Fig. 41. We explain and summarize the various connections in the sequel.
The link between the Sutherland Hamiltonian and the universal parametric correlations of Random Matrix Theory
has been established in three dierent ways. First, there are the asymptotic expansions of k

(, ) just mentioned
(connection IV in Fig. 41). These expansions provide a very direct link between the two elds but are not exact and,
in addition, restricted to those correlators for which explicit expressions are known. In practice, these are the two
point correlators. Second, it can be shown that both H
S
and the nonlinear model for the universal correlators are
related to a continuous matrix model introduced in Refs. 744,743 (connections V and VI). Third, Dysons Brownian
motion model can be invoked to relate H
S
with the random matrix Hamiltonian (7.3) and consequently also with
the universal correlations (connections III, IX, and X). We will treat the two exact approaches in turn. In a nal
paragraph we briey summarize important new results for 1d Fermions.
1. Continuous matrix model
Following Refs. 744,743 we consider the partition function
Z =
_
D() exp (S[()])
S[()] =
1
h
_

d tr
_
m
2
_
()

_
2
+V ()
_
(7.6)
describing the propagation of N N random matrices from one of the Gaussian ensembles in a potential V ().
The Hamiltonian corresponding to the path integral (7.6) reads
H = tr
_

h
2
2m
_

_
2
+V ()
_
. (7.7)
With the help of the diagonalization () = ()()()

the Laplacian / in Eq. (7.7) can be separated into

radial () and angular () degrees of freedom,
_

_
2
=

i
1
J(
i
)

i
J(
i
)

i
+Y (,
i
) . (7.8)
Here J(
i
) =

i<j
[
i

j
[

is the Jacobian known from Random Matrix Theory. In the limit N the angular
part Y (,
i
) can be neglected, and we arrive at the Hamiltonian H
C
of Eq. (7.1) provided (i) the potential V ()
does not depend on angular degrees of freedom and (ii) a factor [J(
i
)]
1/2
is absorbed into the wavefunctions. Such
a connection between an NFermion Hamiltonian and a continuous matrix model had originally been discussed by
Brezin, Itzykson, Parisi, and Zuber
746
. In particular, groundstate expectation values of H
C
can be expressed in
terms of the path integral (7.6),
146
0[ . . . [0) = Z
1
_
D() . . . exp(S[()]) . (7.9)
The proof of equivalence between the timedependent mpoint groundstate correlators of H
C
(or, equivalently, H
S
)
and the mpoint leveldensity autocorrelation functions of the random matrix model of Eq. (7.3) is carried out as
follows
744,743
. Supersymmetric nonlinear models are derived for both types of quantities in the largeN limit. The
two nonlinear models are shown to coincide.
The mpoint groundstate correlators of H
C
are given by
0[f
m
(

)[0) =
_
0

=1
tr
_
1

)
_

0
_
. (7.10)
The quantities f
m
can be expressed in a standard way in terms of a supersymmetric generating function,
Z
m
(

, J

) =
_
d[] exp
_
i

() +

JI)
_
. (7.11)
According to Eq. (7.9), calculating the groundstate expectation value of Z
m
amounts to averaging over () with the
probability density exp(S[()]). Since we are intersted in the universal largeN limit of the correlation functions we
can restrict ourselves to the Gaussian case and choose V () tr
2
. Then the average can be trivially performed. As
in other derivations of a nonlinear model, subsequent steps involve
744,743
a HubbardStratonovich transformation
and a saddlepoint approximation. The nal expression in terms of 4m4m supermatrix elds reads
0[Z
m
[0) =
_
d[Q] exp(F[Q])
F[Q] = i

trg[(

+J

I)Q

] +

2

8 h

[trg[Q

] , (7.12)
where the mean interparticle distance and h
2
/2m have been set equal to unity. In a similar way, the correlators of
the random matrix Hamiltonian (7.3)
f
m
(

, U

) =
m

=1
tr
_
1

H
1
U

H
2
_
(7.13)
can also be represented in terms of a nonlinear model involving the appropriately rescaled parameters

/
and u

(see above). For m = 2, this model can be shown to coincide with that of Eq. (7.12) after the formal
replacements

, (u

)
2
2[

[/h . (7.14)
For m = 2, this establishes the correspondence between the two types of correlation functions. For m > 2 the
replacements (7.14) cannot be realized by a single unique set of transformations u

= u

). Instead a slightly
more complicated argument has to be invoked, see Refs. 744,743. This completes the proof in the general case.
2. Use of Dysons Brownian motion model
To establish the link between onedimensional interacting Fermions and universal parametric correlations via
Dysons Brownian motion model, Narayan and Shastry
745
started from the Calogero Hamiltonian (7.1) with
V (
i
) =
2
i
/4a
2
. For imaginary time = it, the timedependent Schrodinger equation involving H
C
yields for
the quantity P(
i
; ) =
0
(
i
)(
i
; t) the FokkerPlanck equation
P(
i
; )

i
_
P

i
+

i
a
2
P
_

i=j

i
_

j
P
_
. (7.15)
Here,
0
is the groundstate wave function of H
C
. To interpret Eq. (7.15), we may consider the
i
as the position
coordinates of N classical particles. Then, Eq. (7.15) describes the dynamics of the WignerDyson Coulomb gas
58
, the
147
N particles being conned by a harmonic potential and interacting via a pairwise logarithmic repulsion (connection
X).
In a next step, Narayan and Shastry considered the evolution of the eigenvalues of a perturbed random matrix
model similar to Eq. (7.3),
H(x) = H
0
cos(x) +V sin(x)/ , (7.16)
where H
0
and V are drawn from one of the Gaussian ensembles. For any x, H(x) is perturbed by H

(x) = H/x
and the evolution of the eigenvalues
i
(x) can be expressed as a set of rstorder dierential equations,
d
i
dx
= H

ii
,
dH

ii
dx
=
2

i
+

i=j
2[H

ij
[
2

j
,
dH

ij
dx
=

k=i,j
H

ik
H

kj
_
1

k
+
1

j

k
_
+H

ij
(H

jj
H

ii
)
1

j
. (7.17)
Interpreted classically, these equations describe the dynamics of N point particles at the positions
i
(i = 1, . . . N)
and constitute the socalled Pechukas gas
172
. With certain assumptions about the properties of irregular systems,
in particular their eigenfunctions, Pechukas derived
172
linear level repulsion from the above equations (connection
VII). Moreover, the classical system corresponding to Eqs. (7.17) has been shown to be integrable
172,174
. In Ref. 745,
innitesimal increments
i
of the eigenvalues
i
were considered. With the identication x
2
= 2, the following
statistical properties of the
i
could be derived in the largeN limit,

i
) =

i
a
2
+

i=j

j
,

2
i
) = 2 . (7.18)
After setting
i
=
i
these two equations imply the FokkerPlanck equation (7.15) for the distribution function
P(
i
; ). This completes connection IX in Fig. 41. Taking into account the close relation between the Hamiltonians
(7.3), (7.16) and the nonlinear model
44,187,188,192
(connection III), we see that the second exact link between
universal correlations and the Sutherland Hamiltonian has been established. Prior to the publication of Ref. 745,
Beenakker
185
already noticed the relation between the universal parametric correlators and Dysons Brownian motion
model within the hydrodynamic limit.
Finally, it remains to comment on the connections I, II, and VIII in Fig. 41. The corresponding lines are dashed
because one has to assume ergodicity, i.e. the equality of the ensemble average in Random Matrix Theory and
the running average over energy in a real microscopic system. In many cases this hypothesis has been checked
numerically. Numerical conrmations for connection I include the case of quantum billiards
187,188,192
, hydrogen in
a magnetic eld
481
, and strongly correlated manyparticle systems
747
. Connection II is also based on numerical
evidence, and connection VIII is discussed in Ref. 57.
3. Further results
The calculation of the dynamical densitydensity correlators for = 1, 2 and 4 and the discovery of the various
relations discussed in the preceding Secs. VII A1 and VII A2 have stimulated important further progress in the
understanding of interacting Fermions in one dimension. Using the information provided by Simons et al., Haldane
748
conjectured the form of the densitydensity correlator for the CalogeroSutherland model for all rational values of
. In these cases there is, of course, no known relation to Random Matrix Theory. Independently, Minahan and
Polychronakos
749
used grouptheoretical arguments to guess the twopoint density correlator for integral values of
/2. Even earlier, Haldane and Zirnbauer
750
calculated the retarded part of the singleparticle Green function for
= 4 and later
751
extended their results to = 1 and to the advanced part. Haldanes conjecture was proven by
Ha
752
with the help of the theory of Jack polynomials
753
. At the same time Ha extended the results derived in
Refs. 750,751 to arbitrary rational /2, thus conrming another conjecture by Haldane
754
. Forrester
755
seems to have
been the rst author to introduce methods involving Jack polynomials in the present context. He, too, used such
methods
756
to prove Haldanes second conjecture.
This shows how new results in Random Matrix Theory have eventually led to novel dynamical correlation functions
for a whole class of interacting Fermion systems.
148
FIG. 42. Comparison of lattice data and RMT predictions for the number variance 2, for the 3 statistic, and for the
nearest neighbor spacing distribution p(s). Left: Wilson Fermions, right: Kogut-Suskind Fermions. Taken from Ref. 759.
B. QCD and chiral Random Matrix Theory
During the last few years, RMT has been applied successfully to certain aspects of quantum chromodynamics
(QCD). More precisely, some properties of data generated from lattice gauge calculations have been shown to agree
with RMT predictions. This is true, in particular, for level correlation functions and for lowenergy sum rules. Other
data like those on the chiral phase transition have been semiquantitatively reproduced by RMT. These discoveries
are important, and have generated much interest, because they show that QCD possesses generic features. RMT is
expected to be the ideal tool to identify and quantify such features. Thus, application of RMT to lattice QCD data
will hopefully help to separate the generic from the systemspecic properties, and to identify the latter. The work
was pioneered by Verbaarschot and his colleagues at Stony Brook. Our references are incomplete. Fairly complete
lists can be found in Refs. 757,758.
Lack of space forces us to treat this topic but briey, without paying attention to the historic development, and
without attempting to put these developments into the overall context of QCD. We conne ourselves to a description
of those aspects which are most relevant for RMT. We show evidence that lattice data do have generic features
(Sec. VII B1), we sketch the connection between QCD and RMT and introduce the chiral ensembles (Sec. VII B2),
we describe the application of chiral Random Matrix Theory (chRMT) to the chiral phase transition (Sec. VII B3),
and we mention some current lines of interest and open problems (Sec. VII B4).
1. Evidence for generic features
The rst piece of evidence comes from studies of the spectrum of the Euclidean (and Hermitean) Dirac operator
iD/ = i / +g

a
2
A/ (7.19)
for massless Fermions. Here, g is the coupling constant, the matrices
a
are the generators of the gauge group, and
A/ are the gauge elds.
Kalkreuter
760
has calculated the spectrum of iD/ for an SU(2) gauge theory on the lattice, both for Wilson and
for staggered Fermions. These data have been used to calculate the NNS distribution p(s) and the
3
statistic
149
FIG. 43. Dependence of the chiral condensate on the quark mass. Dots: Lattice data. Solid curve: RMT prediction. Taken
from Ref. 761.
(cf. Eqs. (3.50) and (3.57)). The results are shown and compared with chRMT predictions in Fig. 42, taken from
Ref. 759. (The symmetry of the Dirac operator on the lattice diers from the case of the continuum. Wilson
and staggered Fermions correspond to the chGOE and chGSE, respectively, both dened in Sec. VII B2). The
agreement is impressive. It shows that the spectral uctuation properties of the Dirac operator are generic. Somewhat
unfortunately, these properties have virtually no bearing on physical observables in QCD.
Another universal property (absent in classical RMT) occurs because the spectrum of iD/ is symmetric about zero.
The operator iD/ anticommutes with
5
and the eigenvalues
n
therefore occur in pairs
n
with opposite signs. Thus,
level repulsion should aect directly the average spectral density () near zero energy ( = 0) in a generic way. Here,
() is dened as
() =

n
(
n
)). (7.20)
The average (indicated by angular brackets) is taken over all gauge eld congurations. The weight factor is the
exponential of the Euclidean action. The expected generic dependence of () on near = 0 has direct consequences
for QCD because the value of (0) is related to the vacuum expectation value qq) of the chiral condensate by the
BanksCasher formula qq) = (0)/V where V is the spacetime volume. Chiral symmetry is broken if qq) is
nite (nonzero). In this case, (0) V , and the spacing of eigenvalues is 1/V rather than 1/V
1/4
as would be
the case for a noninteracting system. Chiral symmetry is restored for (0) = 0.
First evidence in favor of this assertion (universality of () near = 0) came from the LeutwylerSmilga sum
rules
762
for the eigenvalues
n
. For any integer k, the sum rules yield closed expressions for V
2k

2k
n
. This
sum can in turn be written as an integral over (). It was shown
16
that the sum rules can be derived from chRMT.
Moreover, it was conjectured that the microscopic limit
S
(z) = (z/(V [qq)[))/(V [qq)[) of the spectral density,
obtained by rescaling the argument by 1/(V [qq)[), should in the thermodynamic limit be a universal function. This
conjecture was validated when it was shown
761
that lattice results for the dependence of the chiral condensate on the
mass of the valence quarks can be reproduced using the chRMT result for
S
as the only input, see Fig. 43.
Perhaps the most compelling evidence so far for the universality of () near = 0 is obtained from a comparison of
lattice data for staggered Fermions in the quenched approximation with chRMT predictions. The lattice calculations
were performed for strong coupling ( = 4/g
2
= 2.0) and for four dierent lattice sizes V = L
4
with L = 4, 6, 8,
and 10, resulting in 9979, 9953, 3896, and 1416 congurations, respectively
763
. The comparison is shown in Fig. 44.
Higher correlation functions also display universal features.
2. Chiral random matrix ensembles
The chiral random matrix ensembles can best be introduced by writing the Euclidean Dirac operator of Eq. (7.19)
in matrix form. We choose a chiral basis (consisting of eigenstates of
5
). In this basis, iD/ only couples states of
opposite chirality and has the form
_
0 iD/
(iD/)

0
_
. (7.21)
150
FIG. 44. Microscopic spectral density (upper row) and distribution of the smallest eigenvalue of the Dirac operator for four
dierent lattice sizes. The histograms represent lattice data, the dashed curves correspond to chRMT predictions. Taken from
Ref. 763.
This suggests dening a random matrix model with matrices of the form
_
0 W
W

0
_
(7.22)
where W is a random matrix. In the chiral random matrix models, the average over gauge elds with the exponential
of the Euclidean action as weight factor is replaced by the average over the matrices W with a suitable weight factor
P(W) as in Sec. III A1. Normally, P(W) is chosen as a Gaussian. The limit N of the dimension N of W is
taken.
Depending on the symmetry of iD/ (which must be reected in the symmetry of W), there are
764
three chiral
ensembles, in complete analogy to Sec. III A1. They depend on the number N
c
of colors and on the representation
(fundamental [f] or adjoint [a]) of the Fermions. The chiral Gaussian unitary, orthogonal and symplectic ensembles
are dened, respectively, and apply as follows: (i) chGUE (W complex): for [f] and N
c
3, (ii) chGOE (W real):
for [f] and N
c
= 2, and (iii) chGSE (W real quaternion): for [a] and N
c
2. In the case of lattice calculations with
SU(2) gauge symmetry, the chGOE (the chGSE) applies for Wilson Fermions (for staggered Fermions, respectively).
The chiral random matrix ensembles dier from the classical ensembles introduced in Sec. III A2 mainly in their
behavior near = 0. As a consequence of the chiral symmetry of iD/, in these ensembles the point = 0 plays a
special role, and the translational symmetry of the classical ensembles is broken. Near = 0 there is repulsion of
pairs of eigenvalues of opposite sign. This causes
S
(x) to become generic. A very similar mechanism operates in the
unitary ensemble describing Andreev scattering, see Sec. VI A2. By way of example, we give here the form of
S
(x)
for the chGUE. It depends on the number N
f
of avors and on the topological charge ,

S
(x) =
x
2
_
J
2
N
f
+
(x) J
N
f
++1
(x)J
N
f
+1
(x)
_
. (7.23)
The special role of the point = 0 does not, however, aect the form of the NNS distribution p(s), or of the
3
statistic which in the bulk are the same as for the classical ensembles, see Fig. 42.
3. The chiral phase transition
As shown in Sec. VII B1, for zero quark masses chiral random matrix models yield a nonzero value of the chiral
condensate. This describes the spontaneous breaking of chiral symmetry at zero temperature. Two models have been
put forward which modify Eq. (7.22) in such a way that temperature is included in a schematic way
765768
. These
models aim at a description of the chiral phase transition. In both models, a temperaturedependent diagonal matrix
D is added to W in Eq. (7.22).
What happens qualitatively to () as D is increased, starting from D = 0? We focus attention on the global
properties of and disregard the local suppression of at = 0 displayed in Sec. VII B1. In the simplest case where
151
D is a multiple of the unit matrix, D = D
0
1
N
, two limiting cases are obvious. At D
0
= 0, () has the shape
of a semicircle. For W = 0 and D
0
> 0, on the other hand, all eigenvalues have values D
0
. Keeping D
0
xed
and increasing W from W = 0 will lift the degeneracy of eigenvalues of xed sign and generate two semicircles with
identical radii and with centers at D
0
. The two semicircles do (do not) overlap if their radii are smaller (bigger) than
D
0
. We now apply these insights to the chiral phase transition. We use the fact that according to the BanksCasher
formula, the chiral condensate is proportional to (0). We keep W xed and increase D
0
, starting from D
0
= 0. The
semicircle becomes deformed. At the value of D
0
where it splits into two semicircles, (0) vanishes, and the chiral
phase transition occurs. This is basically the mechanism employed in both Refs. 765,768.
We have discussed this point in some detail because it shows that in describing the chiral phase transition in terms
of chRMT, use is made of the semicircle law, not one of the generic properties of chRMT. And the critical exponents
for the phase transition are determined by the shape of the semicircle at the end of the spectrum. Is this procedure
trustworthy? Several authors have investigated nonGaussian probability distributions P for classical RMT, where
P(H) has the form exp(trV (H)) with V (H) some polynomial, see Sec. VIII. It was found that in the limit of innite
dimension, the spectra are generically conned to a nite interval, and that the behavior of the mean level density
near the end points is always of semicircle type. More precisely: Near the end point a of the spectrum, the mean
level density (E) behaves like (E a)

with a rational number between zero and one. This lends support to the
models of Refs. 765,768. Still, it must be kept in mind that RMT does not necessarily give a correct prediction of the
details of the chiral phase transition.
The two models mentioned above dier in detail. In Ref. 765, the temperature is included schematically in the
form of the lowest Matsubara frequency, and D is accordingly chosen as T1. With T
c
the critical temperature,
the temperature dependence of the chiral condensate qq)(T) is given by qq)(T) = qq)(0)
_
1 (T/T
c
)
2
. The chiral
phase transition is of second order with mean eld critical exponents. The inclusion of all Matsubara frequencies
766,767
yields very similar results. In the model of Ref. 768 (which was inspired by the instanton liquid model), a fraction
of the diagonal elements of D vanishes (with 0 1 trivially), while the rest has identical values d. Both and d
are (unknown) monotonically increasing functions of T. For a phase transition to occur, must be equal to one, and
d must exceed a critical value d
c
. The order as well as the critical exponents of the chiral phase transition depend
on the way in which and d approach 1 and d
c
, respectively. A comparison between these results and lattice and/or
instanton model calculations may shed light on the temperaturedependence of and d.
4. Current lines of research and open problems
The eld is still quite young. Yet, a fair number of papers has appeared. For recent reviews see Refs. 757. Aside
from the topics mentioned above, these papers have addressed the following points. (a) The spectrum of the Dirac
operator at nite temperature
769
; (b) the calculation of kpoint functions at nite T (with k > 2)
770
(c) the nite
volume QCD partition function for nite quark masses
771
; (d) the spectrum of the lattice Dirac operator for Wilson
Fermions; (e) inclusion of the chemical potential in chRMT
772775
. The last point is probably the most dicult but
also the most challenging because on the lattice, unquenched calculations with a chemical potential seem beyond
reach. As mentioned above, this list is incomplete.
C. Random matrices in eld theory and quantum gravity
After some preliminary remarks (Sec. VII C1) we briey explain the use of topological concepts and RMT in general
eld theory (Sec. VII C2) before we discuss the special example of twodimensional quantum gravity in Sec. VII C3.
1. Preliminary remarks
In the stochastic models discussed through most of this review, random matrices were used successfully to simulate
generic properties of the Hamiltonian or of a scattering matrix in a quantum problem described by the Schrodinger
equation. In Sec. VII B we have seen that this substitution also works for certain aspects of relativistic quantum
mechanics where the Dirac equation applies. Demonstrative experiments present strong evidence for the usefulness
of this same procedure in the study of classical wave phenomena, see Sec. VD. This holds true, in particular,
for electromagnetic waves which, in three dimensions, are described by a vectorial Helmholtz equation, and for
elastomechanical waves which are described by Naviers equation.
152
We now turn to stochastic models where the use of random matrices has a dierent origin. The representative
examples given below stand for a whole class of problems where it is the elds rather than the Hamiltonian or
another linear operator which are replaced by an ensemble of random matrices. Often, the replacement has topo-
logical reasons. Cases in point are eld theory (Sec. VII C2), and quantum gravity in two dimensions (Sec. VII C3).
Although dierently motivated, these random matrix models are, at least on the formal and on the mathematical
level, often closely related to the ones discussed in other sections of this review. Because of this connection, a fruitful
transfer of insights and methods between both areas exists. We expect that in future years this transfer will become
even stronger in both directions.
2. Planar approximation to eld theory
Consider a nonAbelian gauge theory with internal symmetry group SU(N). We treat N as a parameter and study
the limit of large N. Such a study is motivated by quantum chromodynamics where the gauge group is SU(3) because
quarks come in three colors. In the absence of exact solutions for SU(3), it is hoped that the results for large N might
be of help.
A useful classication of the diagrams generated by the perturbation expansion of the eld theory in powers of
the coupling constant uses topological concepts. Surfaces can be classied topologically in terms of their genus. For
instance, a sphere has genus h = 0, a torus has genus h = 1, etc. The diagrams are classied by the minimum
genus of the surface needed to draw them. Diagrams with genus zero are called planar. It was shown by t Hooft
776
that the asymptotic expansion of the eld theory in inverse powers of N yields terms which correspomd exactly to
this topological classication. Only planar diagrams survive in the limit N , provided the coupling constant is
properly scaled with N. The higher order terms in 1/N correspond to nonplanar diagrams which have to be drawn
on surfaces of genus h higher than zero.
The generality of this result suggests that a model study of the expansion of eld theories might be very useful.
Brezin et al.
746
studied models for the planar approximations of
3
and
4
theories. We focus here on the
4
theory.
In a d-dimensional space x

, = 1, . . . , d such a theory is dened by a Lagrangean of the type

L =

+
1
2

2
+ g
4
(7.24)
where g is the coupling constant. A model for this theory is obtained by replacing the eld (x) with a Hermitean
N N matrix M(x). The model Lagrangean reads
L = tr (

M) +V (M, g)
V (M, g) =
1
2
trM
2
+
g
N
trM
4
. (7.25)
The global invariance group of this Lagrangean is the group SU(N). Thus, it is possible to study the planar limit
by taking the limit N while the parameter g is held xed. The Feynman rules for the diagramatic expansion
of this matrix model in powers of g involve the two matrix indices and are a straightforward extension of those for
scalar or vector elds. To make plausible the explicit occurrence of the factor 1/N in the coupling constant g/N, we
refer to the discussion in Sec. VIII where more general nonGaussian weight factors with a polynomial potential in
the exponent will be considered. Within this model one can generically solve the so-called counting problem, i.e. one
can nd the number E
(d)
(g) of connected planar vacuum diagrams of a given order in g in d dimensions. We discuss
this for d = 0 and d = 1.
In zero dimensions, each diagram is unity, apart from an overall weight. This implies
746
that, up to a g-independent
normalization constant, E
(0)
(g) can be obtained from the largeN limit of the expression
exp
_
N
2
E
(0)
(g)
_
=
_
d[M] exp (V (M, g)) (N ) (7.26)
with the usual at measure d[M]. To connect to standard random matrix theory, the integrand can be viewed as a
nonGaussian probability density. For large N, the normalization integral yields the counting function versus g. It
is shown in Ref. 746 that this counting function coincides with the one of the
4
theory. In the largeN limit, there
are no contributions due to the matrix structure of the elds M.
Using the U(N) invariance of the integrand, we reduce the problem to integrations over the eigenvalues
n
, n =
1, . . . , N where the squared Vandermonde determinant
2
N
() appears in the integrand. To compute the largeN
limit of Eq. (7.26), the method of steepest descent is used and yields saddlepoint equations for the eigenvalues
n
.
153
The counting function E
(0)
(g) is approximated by the logarithm of the integrand of the right hand side of Eq. (7.26)
taken at these saddlepoints. In a proper continuum limit, the resulting sums are replaced by integrals. This yields
E
(0)
(g) =
_
+2a
2a
du()
_
1
2

2
+g
4
_

_
+2a
2a
du()
_
+2a
2a
d

u(

) ln [

[
(7.27)
where u() is the mean level density, and where 2a and +2a with a = a(g) are the end points of the spectrum. The
rst term on the right hand side of Eq. (7.27) is due to the potential V (M, g), the second one, due to the Vandermonde
determinant. The factor 1/N in the coupling constant in Eq. (7.25) has disappeared. This is because the continuum
limit involves a rescaling of the eigenvalues by a factor of

N. Brezin et al.
746
nd an explicit result for u(). In the
limit g = 0, this result reduces to the Wigner semicircle law. Hence, a closed expression for the counting function
E
(0)
(g) is obtained. In the planar limit, the coecients of an expansion of E
(0)
(g) in powers of g give the desired
combinatorical factors.
For d = 1 the planar approximation provides an interesting connection to the CalogeroSutherland Model and
related theories, see Sec. VII A. We replace the Lagrangean of Eq. (7.25) by the Hamiltonian and quantize the latter.
This yields
H =
1
2
+V (M, g) (7.28)
where the kinetic term is the usual Laplacian in matrix space,
=
N

n=1

2
M
2
nn
+
1
2

n<m
_

2
Re
2
M
nm
+

2
Im
2
M
nm
_
. (7.29)
It can be shown
746
that for large N, the counting function E
(1)
(g) is given by the lowest eigenvalue of H,
H = N
2
E
(1)
(g) . (7.30)
The value of E
(1)
(g) is found by varying the expectation value of H with respect to the functions in the limit of
large N. The groundstate wave function can only depend on the radial degrees of freedom. This is so because
the U(N) invariance of the potential V (M, g) implies the condition (M) = (UMU

). Hence, is a symmetric
function of the eigenvalues
n
of M. To construct such symmetric wave functions , Brezin et al. make the ansatz
(
1
, . . . ,
N
) =
(
1
, . . . ,
N
)

N
()
. (7.31)
The Vandermonde determinant
N
() is antisymmetric in the eigenvalues
n
. Since is symmetric, has to be
antisymmetric as well. This ansatz reduces the kinetic part of Eq. (7.28) to the Euclidean Laplacian of the N matrix
eigenvalues
n
, n = 1, . . . , N. If, moreover, the interaction is of
4
or related type, the Hamiltonian acquires the
form of N noninteracting Fermions, described by the antisymmetric wavefunction (
1
, . . . ,
N
). This construction
yields explicit results for the counting function. It also shows the connection to the CalogeroSutherland model, see
Sec. VII A.
By adding to the potential the terms trM
k
M
k
, Itzykson and Zuber
130
extended these investigations to K coupled
matrix elds M
k
, k = 1, . . . , K. In the resulting model, the U(N) invariance is broken. In the case K = 2, it was
possible to integrate analytically over the diagonalizing groups. The result, the famous Itzykson-Zuber integral, was
later identied as a special case of a more general formula due to Harish-Chandra
129
.
3. RMT and two-dimensional quantum gravity
As in eld theory, the use of Random Matrix Theory in twodimensional quantum gravity is motivated topologically.
Again, the basic idea derives from the planar approximation. The starting point is string theory, providing a link
to the theory of random surfaces where further applications exist. A review was recently given by Abdalla et al.
777
.
Here, we only try to motivate the use of RMT in quantum gravity.
We consider a d-dimensional space in which functions X

(
0
,
1
), = 1, . . . , d parametrize the position of a string
on the two-dimensional world sheet (
0
,
1
). For the description of the string dynamics, it is convenient to use the
Polyakov string action
778
,
154
FIG. 45. Left hand side: A triangulated torus. Right hand side: A piece of a triangulated surface (dashed lines) superposed
with its dual (solid lines). Taken from Ref. 777.
S =
1
2
_
d
2

_
[g[g

. (7.32)
Here g

(
0
,
1
) with , = 0, 1 is the twodimensional gravitational eld. Minimization of the functional S with
respect to the X

and to the elds g

gives the correct equations of motion for the string with all required constraints.
However, further contributions like a cosmological term have to be added to the action S. The partition function of
a string action of this type is then given by the functional integral
Z =
_
D[X]
_
D[g] exp(S) . (7.33)
The integration over the gravitational eld amounts to a summation over all topologies of the two-dimensional surface
described by the string. Therefore, the partition function can be expanded in terms of the genus h and the area A of
the surface,
Z =

h=0

2h
0
_

0
dAexp (
B
A) Z
h
(A) . (7.34)
Here
B
is the bare cosmological constant and
B
A is the cosmological term of the action. The expansion proceeds
in powers of the bare string coupling constant
0
. The expansion coecients Z
h
(A) are the partition functions for
xed genus h and area A. At this point, the connection to the previous section becomes apparent where the genus
expansion was motivated by the topology of diagrams. Hence, the h = 0 term in Eq. (7.34) corresponds to a planar
approximation.
A way to proceed in actual calculations is to discretize the surfaces of given genus h, i.e. the sphere, the torus and
so on, using polygons with a xed number n of edges. Discretized models for random surfaces were introduced by
David
779
, Kazakov
780
and Ambjrn, Durhuus and Frohlich
781
. Apparently, these authors were inspired by Regges
work
782
of 1961. Explicit calculations are often easier in such discretized models than in the continuum version.
The dual graph of a discretization of a surface is dened by connecting the centers of the covering n-gons with
lines. These dual graphs are typical Feynman graphs of given genus h of a zero-dimensional
n
eld theory. Thus, the
partition functions Z can be approximated in terms of a model of the
n
type. A triangulated torus and the dual of a
piece of a triangulated surface are shown in Fig. 45. However, certain constraints must be observed. Most importantly,
the partition functions Z
h
(A) in Eq. (7.34) show a double scaling relation. It turns out that only Hermitean N N
matrices M meet all the requirements. The resulting matrix model is of the type
Z

=
_
d[] exp (NtrV (M))
V (M) =
1
2
M
2
+

n>2
(g
n

(n2)/2
)M
n
. (7.35)
The g
n
are coupling constants and is associated with the cosmological constant. As in the previous section, an
asymptotic expansion in inverse powers of the dimension N eectively yields a diagrammatic genus expansion. The
coecient functions have the desired scaling properties and are identied with the functions Z
h
(A). The expansion
parameter 1/N plays the role of the string coupling constant
0
in the discretized version of Eq. (7.34).
155
VIII. UNIVERSALITY
The overwhelming majority of investigations involving invariant random matrix ensembles deals with the Gaussian
case (GRMT). However, the requirement that the ensemble should be invariant under appropriate (unitary, orthogonal,
or symplectic) transformations in matrix space, admits a much wider class of matrix ensembles with a probability
density P
N
(H) of the form
P
N
(H) = Z
1
exp (trV (H)) . (8.1)
Here, Z
1
is the normalization constant or inverse partition function. The potential V (H) is arbitrary provided the
existence of Z is guaranteed. The Gaussian case corresponds to the choice
V (H) =

4v
2
H
2
, (8.2)
where v and the elements of the N N matrix H have the dimension energy. For other forms of the Gaussian
potential typically used in the literature we refer to the discussion at the end of Sec. III A2. It has been mentioned in
the Introduction of this review that the choice (8.2) is dictated by mathematical convenience rather than by physical
principles. Under these circumstances it is necessary and important to ask which statistical properties are generic,
i.e. common to all or very many ensembles dened in Eq. (8.1), and which are specic for the Gaussian case. This is
the issue referred to as universality.
First convincing numerical evidence in favor of universality was provided by Porter and Rosenzweig
783
, see Appendix
A.1 of Mehtas book
23
. These authors calculated the eigenvalue density and the nearest neighbor spacing distribution
for the Gaussian probability density and for the densities P(H
ij
) = (1/2)(1[H
ij
[) and P(H
ij
) = (1/2)((H
ij
+1)+
(H
ij
1)). In all these cases they found convergence to the semicircle law (eigenvalue densities) and to a curve close
to the Wigner surmise (spacing distribution), respectively. For obtaining the semicircle law, statistical independence
of the matrix elements is the essentil ingredient. This investigation greatly enhanced the condence in the usefulness
of RMT. From the mathematical point of view, an interesting connection to the theory of free random variables was
found by Voiculescu
784
.
Orthogonal polynomials
23
provide a powerful analytical tool to investigate Gaussian as well as nonGaussian en-
sembles of random matrices. We briey outline this method. Upon diagonalization of the matrices H one obtains the
joint probability density function for the eigenvalues,
P
(E)
N
(E
1
, . . . , E
N
) = Z
1

i<j
[E
i
E
j
[

exp
_

i
V (E
i
)
_
, (8.3)
where Z is now dened without the integration over angles. We have chosen the energy eigenvalues E
i
as arguments
instead of the dimensionless quantities x
i
, see Sec. III A2. Equation (8.3) involves the Vandermonde determinant

i<j
(E
i
E
j
) = det[(E
i
)
j1
]. The value of this determinant remains unchanged if the monomials (E
i
)
j1
are
replaced by arbitrary (but linearly independent) polynomials p
n
(E) with degrees n = 0, 1, . . . , N 1. If one chooses
these polynomials to be orthogonal in the following sense,
_

dE exp (V (E)) p
n
(E)p
m
(E) =
nm
, (8.4)
and if one denes the functions

n
(E) = p
n
(E) exp(V (E)/2) (8.5)
the joint probability density in the case = 2, to which we restrict ourselves here, has the form
P
(E)
N2
(E
1
, . . . , E
N
) = det[K(E
i
, E
j
)]. (8.6)
We have dened the kernel
K(E, E

) =
N1

i=0

i
(E)
i
(E

) . (8.7)
In the Gaussian case Eq. (8.4) denes the Hermite polynomials. All correlation functions
156
R
k
(E
1
, . . . , E
k
)
_
dE
k+1
. . . dE
N
P
(E)
N
(E
1
, . . . , E
N
) (8.8)
can now be expressed in terms of this kernel. For the mean density of eigenvalues (E) R
(=2)1
(E) and for the
connected twopoint correlation function T
2
(E, E

) = (E)(E

) R
2
(E, E

) we have, in particular,
(E) = K(E, E)
T
(=2)2
= [K(E, E

)]
2
. (8.9)
Fox and Kahn
785
, Le
786
, and Bronk
787
investigated certain classical cases of nonGaussian potentials V (E),
see also chapter 19 in Ref. 23. The cases studied explicitly include the potentials corresponding via Eq. (8.4) to the
Legendre
786
, Laguerre
785,787
, and Jacobi
785,786
polynomials. In these particular examples the kernel K(E, E

) was
shown to be universally given by
K(E, E

) =
1

sin((

E)(E E

))
E E

. (8.10)
in the largeN limit and for = 2. Performing the limit N was made possible by the wellknown Christoel
Darboux formula, see Eq. (8.16). Given

E, Eq. (8.10) holds for values of E and E

where the mean level density is

essentially constant, (

E) (E) (E

). The result (8.10) implied that after proper rescaling (unfolding) of the
energy variable, two and higherpoint correlation functions were identical for the classical ensembles, in contrast to
the mean level density. As a result, the conjecture
786
that correlations should be relatively insensitive to the form
of the potential V (E), became highly plausible.
Further signicant progress in the universality problem had to wait until 1990. Motivated by problems in two
dimensional quantum gravity (see Sec. VII C), Ambjrn, Jurkiewicz, and Makeenko
788
then discovered a very general
and farreaching result. (We follow the summary in Ref. 789). We consider a unitary random matrix ensemble with
a general polynomial potential
V (H) =

k
g
k
k
H
k
. (8.11)
The coecients g
k
are essentially arbitrary except that for N the mean level density should be nonzero in
only one nite interval. Using an expansion of the resolvents in powers of H, i.e. a perturbative method, and keeping
only terms of leading order in N
1
, Ambjrn et al. calculated the connected part G(E, E

) of the twopoint Green

function in the largeN limit. We remind the reader that a 1/N expansion is equivalent to a topological expansion
in the genus of the surfaces needed to draw the diagrams of the eld theory dened by Eq. (8.11), see Sec. VII C2.
While Brezin et al.
746
had found that the mean level density (E) does depend on the details of the potential V (H)
it was now shown by Ambjrn et al. that G(E, E

) depends on the potential only through the spectral end points a

and b. In the special case of a symmetric potential V (H) = V (H) with b = a, G(E, E

) is given by
788
G(E, E

) = tr(E H)
1
tr(E

H)
1
tr(E H)
1
tr(E

H)
1
=
1
4(E E

)
2
_
2 +
(E
2
a
2
) + (E
2
a
2
)
_
(E
2
a
2
)(E
2
a
2
)
_

1
4
1
_
(E
2
a
2
)(E
2
a
2
)
.
(8.12)
The general formula for nonsymmetric potentials and an extension to complex rather than Hermitean matrices is
also given in Ref. 788. The general connected npoint Green functions
G(E
1
, . . . , E
n
) = N
2n2
N
1
tr(E
1
H)
1
. . . N
1
tr(E
n
H)
1
c
(8.13)
were shown
788
to be also universal in the sense stated above. In addition to the results in leading order, a systematic
1/N
2
expansion of the npoint correlators and of the free energy F = lnZ/N
2
was developed in Ref. 790,
G(E
1
, . . . , E
n
) =

h=0
1
N
2h
G
h
(E
1
, . . . , E
n
)
F =

h=0
1
N
2h
F
h
. (8.14)
157
It was shown that these 1/N
2
corrections are again universal. For symmetric potentials, the corrections to the Green
functions are of the form
G
h
(E
1
, . . . , E
n
) =
R
h
_
E
2
i
a
2
, M
(l)
a
, a
_
_

n
i=1
(E
2
i
a
2
)
. (8.15)
Here, the R
h
are rational functions, and the M
(l)
a
are additional parameters, cf. Ref. 789. Analogous results have also
been derived for nonsymmetric potentials
790
and complex matrices
791
. Moreover, the statements made above for the
twopoint Green functions could be generalized
789,792
to the case where, in the largeN limit, the level density (E)
has support in s distinct intervals. For each s a new universality class arises in which the twopoint correlators are
universal and depend on the potential V only through the end points of the s intervals.
For small distances [E E

[, the twopoint correlator of Eq. (8.12) does not yield the universal form (8.10) of
the kernel K(E, E

). The reason is that the two expressions apply in two dierent limits. This became particularly
transparent in a paper by Brezin and Zee
793
who considered potentials V (H) in the form of even polynomials in H.
They made an ansatz for the asymptotic (in n) form of the associated orthogonal polynomials p
n
(E) and, hence,
for
n
(E). Knowledge of this asymptotic form is helpful. Indeed, by means of the ChristoelDarboux identity the
kernel K(E, E

) dened in Eq. (8.7) can be written as

K(E, E

) = c
N

N
(E)
N1
(E

)
N1
(E)
N
(E

)
E E

(8.16)
with an Ndependent constant c
N
. In the largeN limit the ansatz in Ref. 793 and the ensuing expression for
K(E, E

) become exact. Two limiting cases were considered in Ref. 793. (i) The dierence EE

is of order 1/N and

both energy arguments are at a nite distance from the end points of the spectrum. Then, the universal form (8.10)
holds. Local universality (which of course applies equally to all higher correlation functions) was thus proven for all
symmetric potentials of polynomial form. For an extension of this result to more complicated potentials see Ref. 794.
(ii) The dierence [E E

[ is so large that the number of levels in the interval E E

is of order N. Because of this

large number, the kernel K(E, E

) oscillates heavily with changes of E or E

. Averaging over a nite number of such

oscillations in the vicinity of both E and E

, Brezin and Zee obtained

T
smooth
(=2)2
(E, E

) =
1
2N
2

2
1
(E E

)
2
a
2
EE

[(a
2
E
2
)(a
2
E
2
)]
1/2
. (8.17)
Thus, they rediscovered the universal correlation function (8.12) of Ambjrn et al.
788
. (The relation between G(E, E

)
and T
(=2)2
(E, E

) is simple and explained in Ref. 789). Apparently, the perturbative calculation of Ref. 788 cannot
account for the oscillatory behavior of the level correlations. In Ref. 186, this fact was attributed to an interchange
of limits. In a perturbative calculation N is sent to innity before the poles in the resolvents are placed on the real
axis. Therefore the individual poles merge to form a cut and the ne structure is lost. In the method of orthogonal
polynomials and in the supersymmetry method (see below), on the other hand, no such problem arises.
We digress and mention some related work although it is not directly related to the subject of this section. In
Ref. 186, time dependent correlations in a matrix ensemble dened by
P[H] = Z
1
exp
_

_
T
T
dt tr
_
1
2
_
dH
dt
_
2
+V [H]
__
(8.18)
were calculated with diagrammatic techniques. Here, H = H(t) is explicitly a function of time. In particular the
(smoothed) currentcurrent correlator J(E, t)J(E

, 0)) for arbitrary t was derived, thus generalizing earlier results by

Szafer and Altshuler
183
and by Beenakker
185
. The current operator was dened as J(E, t) = N
1

i
(dE
i
/dt)(E
E
i
(t)). In subsequent work
795
these timedependent correlations were related to the cross correlations between the
eigenvalues of two separate (but not independent) Gaussian ensembles. The relationship is very similar to the one
described in Sec. VII A1 between parametric correlations and a continuous matrix model. In Ref. 796 the interesting
relation
G(E, E

) =

2
EE

ln
__
G(E) G(E

)
E E

_
+. . .
_
(8.19)
was established for arbitrary polynomial V [H]. The dots indicate terms that do not contribute to the connected
twopoint level density correlation function T
2
(E, E

). The result (8.19) is rather surprising since it connects a

158
universal quantity like G(E, E

) with the nonuniversal onepoint function G(E). The validity of Eq. (8.19) could be
extended
796
to the case where H = H
1
+H
2
with P[H] = P
1
[H
1
]P
2
[H
2
].
Beenakker
797
gave a particularly simple and elegant derivation of the form (8.17) of the universal wide correlator.
At the same time, he could generalize the result to the orthogonal and symplectic cases, and to nonpolynomial
potentials. Using the functional derivative method which he had developed
652
and applied to the case of the positive
spectrum of the transfer matrix, he calculated directly the connected twopoint correlation function T
2
(E, E

). We
briey summarize the main steps of the derivation. The twopoint function is given by the functional derivative of
the eigenvalue density with respect to the conning potential,
T
2
(E, E

) =
1

(E)
V (E

)
. (8.20)
Here the potential V is arbitrary provided the support of the spectrum remains compact as N . In this limit, V
and are connected by the integral relation
P
_
b
a
dE
(E)
E

E
=
d
dE

V (E

) (8.21)
where Pdenotes the principal value and a and b are the end points of the spectrum. Variation of Eq. (8.21) with respect
to V and inversion of the result together with Eq. (8.20) immediately yields a universal expression for T
2
(E, E

).
The result is valid for all a and b, for = 1, 2, 4, and for all V . For a = b and = 2 it reduces to the known case
given by Eq. (8.17). Extending the methods used in Ref. 790, Itoi
798
also derived explicit expressions for twopoint
correlators in the orthogonal and symplectic symmetry classes and established the universality of the multipoint
correlators.
Hackenbroich and Weidenm uller
799
showed that all local correlation functions (i.e., functions on the scale of the
mean level spacing) are independent of the conning potential V (H). This statement holds under the following
two provisos: The support of the spectrum must remain nite in the limit N , and the arguments of the
correlators must be scaled correctly (energy dierences are to be expressed in units of the local mean level spacing).
The proof applies to all three symmetry classes, and to all types of correlation functions involving level densities, S
matrix elements, and/or parametric correlations of arbitrary order. Starting point is the supersymmetric generating
functional
I =
_
d[]
_
d[H]P
N
(H) exp
_
i
2

L
1/2
(H E +M)L
1/2

_
(8.22)
for a product of Green functions of H. Energy dierences, source terms and, if present, the coupling to external
channels are all contained in the matrix M
799
. From the fact that M is of order N
1
it follows that to leading
order in N
1
, I is a function of the unitary invariants A

(L
1/2
)

L
1/2
)

only. Before averaging, these

invariants are introduced via an additional double integration over a delta function. In a sense, this step replaces
the HubbardStratonovich transformation which applies in the Gaussian case only. Then, the integration over the
variables is carried out. For large N, all remaining integrations (including the ensemble average, i.e. the integration
over the distribution of H) are then evaluated with the help of the saddlepoint approximation. Except for the
occurrence of the local mean level density (E), the nal expression for I coincides with the result for the Gaussian
case. This proof derives its elegance from the fact that one never leaves the level of the generating functional. At the
same time this is its drawback: The proof reduces the form of the correlators to the Gaussian case but is of no help
in actual calculations of the latter.
Freilikher, Kanzieper, and Yurkevich
800,801
addressed the issue of universality for wide classes of conning potentials
(for which the spectrum is bounded). The authors used orthogonal polynomials and mathematical results
802805
which
partly are quite recent. In the ensemble
806,807
with V (E) [E[

, the orthogonal polynomials are dened

800
with
respect to the Freud weights
802
. Rigorous results for the kernel K(E, E

), see Eq. (8.7), led to the following

statements. As approaches unity from above, the mean level density (E) develops a sharp peak in the center of
the spectrum. In contradistinction, all twopoint correlators, local as well as smoothed ones, retain their universal
form down to and including the value = 1. In Ref. 801 the rigorous treatment was extended to the Erdos
type of conning potentials which at innity grow faster than any polynomial, and to the presence of a hard edge
in the spectrum of the Hamiltonian. This last condition is motivated by applications involving the (nonnegative)
eigenparameters of random transfer matrices, see Sec. VI C2. Again, in all cases considered, the universality of both
global (smoothed) and local twopoint correlators could be shown.
Explicit and interesting examples for potentials V (H) leading to nonuniversal statistics were given by Nagao
and Slevin
808
, and by Muttalib and coworkers
809,810
. In the spectrum of the generalized Gaussian and Laguerre
159
ensembles (which involve an additional logarithmic term in the potential V (H)), there exist domains with non
universal correlations
808
. In Ref. 810, the qdependent potential
V (H) =

n=0
ln[1 + 2q
n+1
cosh(2) +q
2n+2
], H = sinh (8.23)
with q [0, 1] was considered. The spectrum is not bounded (the eigenvalues range from to +). As a function
of q, the nearest neighbor spacing distribution interpolates between WignerDyson (WD) and Poisson statistics
810
.
The orthogonal polynomials corresponding to the choice (8.23) are the q Hermite polynomials. Potentials related
to the q Laguerre polynomials exhibit similar spectral properties
809
. It was conjectured
810
that these properties are
shared by the entire family of potentials related to q polynomials
811
.
Since they interpolate between WD and Poisson statistics, potentials of the type (8.23) might be important for
the description of the crossover from metallic to localized behavior in mesoscopic systems, see Sec. VI A2. Another
possibility to describe this transition is provided by random matrix ensembles which break the rotational invariance
in Hilbert space assumed in Eq. (8.1). Examples are ensembles of random band matrices (see Sec. VI D) and other
ensembles
812,142
with a preferential basis.
Why do rotationally invariant potentials of the type (8.23) deviate from universal behavior? Canali and Kravtsov
813
gave an interesting answer: As one goes from the ensemble ( 1) to potentials like (8.23) which grow only
logarithmically with H, the U(N) invariance is spontaneously broken. This seems the common feature of all approaches
which describe the localization transition in terms of random matrix ensembles.
IX. COMMON CONCEPTS
In previous sections of this review, we have encountered numerous examples of successful applications of RMT to
physical systems. We have also listed many open problems. Perhaps the most impressive aspect of this discussion
is the wide range of topics involved. It suggests that RMT is a universal tool. Indeed, RMT correctly describes
uctuation properties of nonrelativistic and relativistic quantum systems. It applies to chaotic quantum systems
with few degrees of freedom, to disordered quantum systems, to classical wave propagation in disordered media, and
to manybody quantum systems. RMT also describes spectral uctuation properties and parametric correlations of
several very dierent wave equations. And RMT connects to quantum gravity and to the CalogeroSutherland type
models in one dimension. In this nal section, we focus attention on the ubiquity of RMT.
To this end we rst recall the foundation, success and limitations of RMT as well as some of the open problems in
the various elds of application.
A. Synopsis
(i) Manybody systems. RMT started with Wigners attempt to formulate a generic model for the spectral uctu-
ation properties of complex manybody systems. The Hamiltonian of every such system is assumed to be a member
of an ensemble of random matrices. The ensemble is characterized completely by symmetry properties. Apart from
appropriate mean values, no detailed knowledge of the system at hand enters the denition of the ensemble. Thus,
systemspecic features (beyond the mean values needed to dene RMT properties) cannot be reproduced and are
assumed to be irrelevant for uctuation properties. This model has been extremely successful in the description
of spectral uctuation properties and of crosssection uctuations in complex atoms, molecules, and atomic nuclei.
This is true in spite of the fact that the interactions between the constituents in these systems are vastly dierent:
Nuclei are bound by the shortrange nuclear force which is repulsive at very small distances, atoms and molecules are
governed by longrange Coulomb forces. Only states near the ground state seem to display systemspecic features.
At higher excitation energies, collective and/or singleparticle properties may still be present but seem to appear only
as gross features. For instance, the giant dipole resonance in nuclei appears as a broad resonance in the absorption
cross section. The spreading width of the resonance is very much larger than the mean level spacing of the manybody
system. The resonance does not correspond to a single eigenstate of the nucleus. Similar statements apply to other
collective modes. Isobaric analogue resonances in heavy nuclei provide a good example. Hence it now appears that,
within each set of conserved quantum numbers and suciently far above the ground state, any strongly interacting
quantum system with many degrees of freedom must be expected to display uctuation properties of RMT type of
the appropriate symmetry.
Does an energy scale exist in such systems (analogous to the Thouless energy in disordered systems) beyond which
RMT ceases to apply? Can we expect to nd the analog of localization in the eigenfunctions? As long as the origin
160
of stochastic behavior in such manybody systems is unclear, questions such as these have no easy answer. It is
conceivable that stochasticity arises from classical chaos. This is the case in systems with few degrees of freedom.
In that case, limitations of RMT similar to the ones to be mentioned under (ii) are expected to apply. It is equally
conceivable, however, that stochasticity is an emergent property of strongly interacting systems with many degrees
of freedom irrespective of whether the underlying classical dynamics is fully or only partially chaotic. After all, an
assumption of similar type (equal a priori occupation probability of all accessible parts of phase space or Hilbert
space) has been used with amazing success as the foundation of classical and quantum statistical mechanics. In this
case, possible limitations of RMT would have a dierent origin, and would probably be of dierent form. The available
experimental evidence (covering sequences of up to several hundred levels) does not point to any such limitations at
present.
(ii) Classically chaotic systems with few degrees of freedom. The fundamental property of these systems is the
instability of the classical trajectories. It leads to stochastic behavior on both the classical and the quantum level.
For fully chaotic systems with a single intrinsic time scale, the situation is fairly well understood. We have discussed
three attempts, based on semiclassics, the nonlinear model, and structural invariance, respectively, to rigorously
establish the Bohigas conjecture. Although a complete proof does not yet exist, the conditions for the occurrence
of spectral uctuations of RMT type are known in their essentials. The energy interval within which uctuations
of RMT type occur, is determined by the shortest periodic orbits. Fully chaotic systems with several intrinsic time
scales, such as, e.g., a chain of coupled chaotic billiards, have not been considered in detail in this review. They show
uctuations similar to those of disordered quasi onedimensional mesoscopic systems and are described by random
band matrices. Here, localization phenomena may occur. Generic systems with mixed phase space pose a major
challenge. The idea of chaosassisted tunneling has been very successful. Closer inspection has shown, however, that
the structure of classical phase space is quite complex, especially at the interface between regular and chaotic domains.
This fact does inuence the uctuation properties. We have emphasized the conceptual importance of model systems
(like two coupled harmonic oscillators) for the investigation of many of these questions. Another challenge consists in
extending the semiclassical results obtained for systems with two or three degrees of freedom to higher dimensions.
(iii) Disordered systems. Both the origin of stochastic behavior and the limitations of RMT are quite clear in
disordered systems. The random distribution of impurity scatterers together with a suitable ergodic theorem furnish
an immediate justication for modeling these systems in a stochastic fashion. Such a justication is much more subtle
in the cases discussed under (i) and (ii), where stochasticity appears as an emergent property of the manybody
dynamics or the classically chaotic dynamics, respectively. For disordered systems, spectral uctuations and the
uctuations of transport coecients are within known limits of classical RMT type. The applicability of classical
RMT is limited to the diusive regime (if we disregard the extreme ballistic limit where the system becomes a quantum
billiard) and, within this regime, to an energy interval given by the Thouless energy. Outside of this interval, and in
the crossover to either the ballistic or the localized regime as well as at the mobility edge, other universal laws govern
the uctuation properties. Except for the behavior at the mobility edge, these laws can be derived from suitable
generalizations of classical RMT. In this sense, RMT has been found to apply universally to disordered systems,
too. We have not been able to dwell on the passage of classical light waves, or of radar waves, through a medium
with a randomly varying index of refraction. However, much of what has been said about disordered mesoscopic
systems applies here, too. In particular, weak localization occurs and is thus seen to be not a quantum, but a wave
phenomenon.
The phenomenon of localization is not limited to disordered systems. As dynamical localization, it occurs likewise
in chaotic quantum systems. The kicked rotor is the standard example. This system can be mapped onto a one
dimensional nonlinear model. This shows the connection to quasi onedimensional disordered wires and to RMT.
(iv) Wave equations. Spectral uctuation properties of RMT type are not linked to a specic linear secondorder
dierential equation (the Schrodinger equation). The spectra of acoustic and elastomechanical waves in irregularly
shaped solids show RMT uctuations, and so do the eigenmodes of Maxwells equations in threedimensional cavities.
These phenomena are governed by rather dierent linear wave equations. In the case of elastomechanical waves
even the boundary conditions (vanishing stress tensor) are completely dierent from those for Schrodinger waves.
Another example of a linear wave equation, the Dirac equation, will be mentioned under (v). Moreover, very dierent
interaction potentials in the Schrodinger equation also give rise to RMT uctuations. We are not aware of any
attempts to explain this general applicability of RMT to various wave phenomena in a generic fashion.
(v) Field theory. We have considered two examples, QCD and a
4
theory involving matrix elds . In general,
local uctuation properties are of little interest in these eld theories. Nevertheless, it is noteworthy that the spectral
uctuations of the Dirac operator in QCD coincide with RMT predictions, for two reasons. First, here is another
example of universality: A rstorder dierential operator coupled to gauge elds possesses RMT uctuation proper-
ties. Second, the existence of RMT uctuations shows that the data generated by lattice gauge calculations contain
generic features. This fact poses a challenge: To separate the generic features from the physical content, in order both
to obtain a better understanding of the problem, and to simplify the calculations.
161
Certain global features of RMT are also universal. This is true, for instance, for the shape of the global RMT
spectrum near the end points, or for the structure of diagrams generated through a loop expansion of the above
mentioned
4
theory. This is why RMT applies to twodimensional gravity and thereby establishes a conncection
with conformal eld theory, and why RMT can be used to model the chiral phase transition of QCD.
B. Discussion
This tour dhorizon shows that most applications of RMT to physical systems predict local uctuations in terms
of a systemspecic input. The input consists of suitable mean values: Spectral uctuations require the mean
level spacing as input. For cross section uctuations, the average scattering matrix serves the same purpose. For
conductance uctuations, the mean conductance takes this role. In this sense, RMT predictions are not parameter
free: They relate uctuations to mean values. This statement can be given a dierent and more interesting form.
The mean values can be absorbed into properly rescaled local variables, and RMT predictions for uctuations and
correlation functions acquire universal form. This applies not only to level correlations, but very generally to the
dependence of correlation functions on any external parameter (parametric correlations). We have presented many
examples where such universal RMT predictions have been tested successfully.
By construction, RMT predictions are based on ensemble averages. On the other hand, variances or correlation
functions extracted from a data set are almost always obtained from a running average over a given system. The
equality of the results of both averaging procedures is guaranteed by ergodicity. Although not proved in full generality,
ergodicity has been shown to hold in a number of cases.
During the last fteen years, the nonlinear model has become an invaluable tool in RMT. Without this model,
the universal features of RMT might never have been discovered to their full extent. Through the medium of eld
theory, the common statistical properties of chaotic and of disordered systems, and the universality of parametric
correlations, have been displayed, and have been cast into compact form. The model has been indispensable also in
other contexts. For instance, many aspects of stochastic scattering would not have been accessible without it. And
the insight that local RMT uctuation properties do not depend on the Gaussian distribution of matrix elements
nds perhaps the simplest and most general proof through the nonlinear model. The need to calculate some RMT
correlation functions has led to a number of extensions of supersymmetry. We think here of new and unconventional
saddle points, of Fourier analysis on graded spaces, and of extensions of the ItzyksonZuber integral. Due to lack of
space we could not do justice to some of these developments. These examples show that modern applications of RMT
have led to mathematical research which has meanwhile become a eld in its own right. The situation is reminiscent of
the early days of (classical) RMT where progress was eventually possible due to mathematical input from the theory
of orthogonal polynomials.
Are the success and the ubiquity of RMT really surprising? One might argue that RMT is trivial because it only
displays the consequences of von NeumannWigner level repulsion, i.e. of the simple fact that two states connected
by a nonvanishing matrix element repel each other. But this argument fails to take into account all other predictions
of RMT which go far beyond simple level repulsion. These predictions derive from the symmetry of the Hamiltonian,
and from the postulated invariance properties of RMT which in turn embody the generic aspects of the entire
approach. The form of the Vandermonde determinant appearing in the invariant measure of RMT is entirely due to
the postulated rotational invariance of the random matrix ensemble in Hilbert space. It is true, of course, that this
determinant implies (a particular form of) level repulsion. The degree of repulsion is determined by the symmetry of
the Hamiltonian. A simple counting argument leads directly to the exponent = 1, 2, 4 in the typical factor [E

in the Vandermonde determinant. But the forms of the twopoint function Y

2
and of higher correlation functions, of
the
3
statistic measuring the spectral stiness, and of the nearest neighbor spacing distribution are not determined
by level repulsion alone and require for their derivation the full rotational invariance of the ensemble. The same is true
of the universal form of correlations of observables which depend upon an external parameter. With regard to these
observables, RMT possesses genuine predictive power and leads to highly nontrivial results. The results are generic,
not based on dynamical principles, and derive only from the underlying symmetry and invariance requirements. Even
on a global scale (as opposed to the local scale of the mean level spacing), spectral correlations still display some degree
of universality. The extension of RMT to random band matrices and random transfer matrices, important for the
understanding of extended systems and localization, breaks rotational invariance. The transition to localization and
the ensuing correlation functions are determined by the way in which the inuence of the Vandermonde determinant in
the invariant measure is gradually neutralized. But even in this domain of application of RMT, generic nondynamical
features remain and are of central interest. They are embodied in the eective Lagrangian of the nonlinear model.
In the end, the universal features of RMT appear to correspond remarkably well to generic properties of quantum
(and certain classical) systems. We recall Balians proof that among all rotation invariant ensembles Gaussian RMT
162
maximizes the information entropy. These facts seem to imply that in the absence of any further knowledge about
the system, Gaussian RMT yields the best approximation to the physical properties of that system.
In spite of this argument, we feel that the universal applicability of RMT remains an amazing fact. It signals the
emergence of a theory of wave phenomena, both classical and quantal, which does not relate to dynamical properties.
RMT is valid essentially on the local scale dened by the mean level spacing of the system. On larger scales, correlation
functions are often used to determine systemspecic properties. The validity of RMT on the local scale shows that
as the resolution is improved and the local scale is reached, such information is lost, and is replaced by universal
features.
To us, the universal validity of RMT is reminiscent of the success of Thermodynamics in the last century. Thermo-
dynamics was built on a few universal principles without recourse to any dynamical theory. With the help of a few
systemspecic parameters, the thermodynamic behavior of widely dierent systems could be described successfully.
The universal applicability of RMT to local uctuation properties of systems governed by a wave equation seems to
mirror this development.
1
O. Bohigas, R.U. Haq, and A. Pandey, in Nuclear Data for Science and Technology, K.H. Bochho (ed.), Reidel, Dordrecht
(1983), p.809.
2
J.B. Garg, J. Rainwater, J.S. Petersen and W.W. Havens, Jr., Phys. Rev. 134 (1964) B985.
3
O. Hausser, A. Richter, and W. von Witsch, Nucl. Phys. A 109 (1968) 329.
4
T. Ericson, Ann. of Phys. (N.Y.) 23 (1963) 390.
5
O. Bohigas, M.J. Giannoni and C. Schmit, Phys. Rev. Lett. 52 (1984) 1.
6
D. Wintgen and H. Friedrich, Phys. Rev. A35 (1987) 1464.
7
C. Ellegaard, T. Guhr, K. Lindemann, J. Nygard and M. Oxborrow, Phys. Rev. Lett. 77 (1996) 4918.
8
S. Washburn and R.A. Webb, Adv. Phys. 35 (1986) 375.
9
C.E. Porter, Statistical Theories of Spectra: Fluctuations, (Academic Press, New York, 1965).
10
N. Bohr, Nature 137 (1936) 344.
11
Nature 137 (1936) 351.
12
E. P. Wigner, Group Theory and its Application to the Quantum Mechanics of Atomic Spectra, Academic Press, New York
(1959).
13
R. Gade, Nucl. Phys. B 398 (1993) 499.
14
A. V. Andreev, B. D. Simons and N. Taniguchi, Nucl. Phys. B 432 (1994) 487.
15
A. Altland and M. R. Zirnbauer, Phys. Rev. Lett. 76 (1996) 3420.
16
E. V. Shuryak and J. J. M. Verbaarschot, Nucl. Phys. A 560 (1993) 306.
17
R. Balian, Nuov. Cim. 57 (1968) 183.
18
T.A. Brody, J. Flores, J.B. French, P.A. Mello, A. Pandey and S.S.M. Wong, Rev. Mod. Phys. 53 (1981) 385.
19
L. K. Hua, Harmonic Analysis of Functions of Several Complex Variables in the Classical Domains, American Physical
Society, Providence (1963).
20
E. Cartan, Abh. Math. Sem. Univ. Hamburg 11 (1935) 116.
21
J. von Neumann and E. P. Wigner, Phys. Zeitschr. 30 (1929) 467.
22
L.A. Pastur, Teor. Mat. Fis. 10 (1972) 102.
23
M.L. Mehta, Random Matrices (Academic, New York, 1991), 2nd ed.
24
T. O. E. Ericson and T. MayerKuckuk, Ann. Rev. Nucl. Science 16 (1966) 183.
25
C. Mahaux and H. A. Weidenm uller, Ann. Rev. Nucl. Part. Science 29 (1979) 1.
26
R. V. Haq, A. Pandey and O. Bohigas, Phys. Rev. Lett. 48 (1982) 1086.
27
C. Mahaux and H.A. Weidenm uller, ShellModel Approach to Nuclear Reactions (NorthHolland, Amsterdam, 1969).
28
A. Pandey and M.L. Mehta, Commun. Math. Phys. 87 (1983) 449.
29
M. L. Mehta and A. Pandey, J. Phys. A: Math. Gen. 16 (1983) 2655, L601.
30
H. L. Harney, A. Richter, and H. A. Weidenm uller, Rev. Mod. Phys. 58 (1986) 607.
31
P.A. Lee and T.V. Ramakrishnan, Rev. Mod. Phys. 57 (1985) 287.
32
P.W. Anderson, Phys. Rev. 109 (1958) 1492.
33
N. F. Mott and W. D. Twose, Adv. Phys. 10 (1961) 107.
34
D.J. Thouless, Phys. Rep. 13 (1974) 93.
35
S.F. Edwards and P.W. Anderson, J. Phys. F 5 (1975) 965.
36
J.J.M. Verbaarschot and M.R. Zirnbauer, J. Phys. A 17 (1985) 1093.
37
F. J. Wegner, Phys. Rev. B 19 (1979) 783.
163
38
R. Oppermann and F. Wegner, Z. Physik B 34 (1979) 327.
39
L. Schafer and F. Wegner, Z. Phys. B 38 (1980) 113.
40
L.P. Gorkov and G.M. Eliashberg, Sov. Phys. JETP 21 (1965) 940.
41
H. A. Weidenm uller, Ann. Phys. (N.Y.) 158 (1984) 120.
42
A. J. MacKane, Phys. Lett. A 76 (1980) 33.
43
G. Parisi and N. Sourlas, J. Physique (Paris) 41 (1981) L403.
44
K.B. Efetov, Adv. in Phys. 32 (1983) 53.
45
K. B. Efetov, Supersymmetry in Disorder and Chaos, Cambridge University Press (1997).
46
J.J.M. Verbaarschot, H.A. Weidenm uller, and M.R. Zirnbauer, Phys. Rep. 129 (1985) 367.
47
O. Bohigas and M.J. Giannoni, Lecture Notes in Physics vol. 209 (Springer, Berlin 1984).
48
S.W. McDonald and A.N. Kaufman, Phys. Rev. Lett. 42 (1979) 1189.
49
G. Casati, F. Valz-Gris and I. Guarneri, Lett. Nuov. Cim. 28 (1980) 279.
50
M.V. Berry, Ann. Phys. (NY) 131 (1981) 163.
51
G.M. Zaslavsky, Phys. Rep. 80 (1981) 157.
52
B. A. Muzykantskii and D. E. Khmelnitzkii, JETP Lett. 62 (1995) 76, Pisma Zh. Eksp. Teor. Fiz. 62 (1995) 68.
53
A. V. Andreev, O. Agam, B. D. Simons, and B. L. Altshuler, Phys. Rev. Lett. 76 (1996) 3947 ; A. V. Andreev, B. D.
Simons, O. Agam, and B. L. Altshuler, Nucl. Phys. B482 (1996) 536.
54
F. Leyvraz and T.H. Seligman, Phys. Lett. A168 (1992) 348.
55
F. Leyvraz and T.H. Seligman, Structural Invariance: a Link between Chaos and Random Matrices, Proceedings of the IV
Wigner-Symposium in Guadalajara, Mexico, ed. N.M. Atakishiyev, T. Seligman and K.B. Wolf (World Scientic, Singapore,
1996), p. 350.
56
M.V. Berry, Proc. R. Soc. London, A400 (1985) 229.
57
F. Haake, Quantum Signatures of Chaos (Springer, Berlin, 1991).
58
F.J. Dyson, J. Math. Phys. 3 (1962) 140.
59
F.J. Dyson, J. Math. Phys. 3 (1962) 157.
60
F.J. Dyson, J. Math. Phys. 3 (1962) 166.
61
E.P. Wigner, Nachr. Akad. Wiss. Gottingen, Math. physik. Kl. (1932) 546.
62
H.A. Kramer, Proc. Acad. Sci. Amsterdam 33 (1930) 959.
63
A. Bohr and B. Mottelson, Nuclear Structure I, (Benjamin, New York, 1969).
64
E.P. Wigner, in Conference on Neutron Physics by TimeofFlight, Gatlinburg, Tennessee 1956.
65
T.A. Brody, Lett. Nuov. Cim. 7 (1973) 482.
66
F.M. Izrailev, Phys. Lett. A134 (1988) 13.
67
F.M. Izrailev, Phys. Rep. 196 (1990) 299.
68
E. Caurier, B. Grammaticos and A. Ramani, J. Phys. A23 (1990) 4903.
69
G. Lenz and F. Haake, Phys. Rev. Lett. 67 (1991) 1.
70
F.J. Dyson and M.L. Mehta, J. Math. Phys. 4 (1963) 701.
71
O. Bohigas and M.J. Gianonni, Ann. Phys. (NY) 89 (1975) 393.
72
J.F. Shriner, Jr. and G.E. Mitchell, Z. Phys. A342 (1992) 53.
73
A. Pandey, Ann. Phys. (NY) 119 (1979) 170.
74
L. Leviandier, M. Lombardi, R. Jost, and J.P. Pique, Phys. Rev. Lett. 56 (1986) 2449.
75
A. Delon, R. Jost, and M. Lombardi, J. Chem. Phys. 95 (1991) 5701.
76
H. Alt, H.D. Graf, T. Guhr, H.L. Harney, R. Hoerbert, H. Rehfeldt, A. Richter and P. Schardt, Phys. Rev. E55 (1997)
6674.
77
T. Guhr and H.A. Weidenm uller, Chem. Phys. 146 (1990) 21.
78
M. Lombardi, J.P. Pique, P. Labastie, M. Broyer and T. Seligman, Comments. At. Mol. Phys. 25 (1991) 345.
79
M. Lombardi and T.H. Seligman, Phys. Rev. A47 (1993) 3571.
80
U. Hartmann, H.A. Weidenm uller, and T. Guhr, Chem. Phys. 150 (1991) 311.
81
Y. Alhassid and N. Whelan, Phys. Rev. Lett. 70 (1993) 572.
82
N. Rosenzweig and C.E. Porter, Phys. Rev. 120 (1960) 1698.
83
M.L. Mehta and F.J. Dyson, J. Math. Phys. 4 (1963) 713.
84
M. Lombardi, O. Bohigas, and T.H. Seligman, Phys. Lett. B324 (1994) 263.
85
J.B. French, P.A. Mello and A. Pandey, Phys. Lett. B80 (1978) 17.
86
A.M. Yaglom, An Introduction to the Theory of Stationary Random Functions, transl. by R.A. Silverman (PrenticeHall,
Englewood Clis, NJ, 1962).
87
F. A. Berezin, Introduction to supermanifolds, Reidel, Dordrecht 1987.
88
M. J. Rothstein, Trans. Am. Math. Soc. 299 (1987) 387.
89
Z. Pluhar, H.A. Weidenm uller, J.A. Zuk, C.H. Lewenkopf and F.J. Wegner, Ann. Phys. (NY) 243 (1995) 1.
90
J. A. Zuk, preprint (1995), to appear in Rev. Mod. Phys.
91
G. Wess and B. Zumino, Nucl. Phys. B 70 (1974) 39.
92
T. Guhr, J. Math. Phys. 32 (1991) 336.
164
93
M.R. Zirnbauer, J. Phys. A29 (1996) 7113.
94
B. Eckhardt, Phys. Rep. 163 (1988) 205.
95
B. Eckhardt, Physica D33 (1988) 89.
96
U. Smilansky, in Chaos and Quantum Physics edited by M.J. Giannoni, A. Voros and J. ZinnJustin (Elsevier Science, New
York, 1990).
97
U. Smilansky, in Proceedings of the Les Houches Summer School on Chaos and Quantum Physics (North-Holland, Amster-
dam, 1991), p. 371.
98
E. Doron, U. Smilansky and A. Frenkel, Physica D50 (1991) 367.
99
C. Lewenkopf and H.A. Weidenm uller, Ann. Phys. (NY) 212 (1991) 53.
100
C. Jung and T. Seligman, Phys. Rep. 285 (1997) 77.
101
C.W.J. Beenakker, to appear in Rev. Mod. Phys. 69 (1997).
102
E.P. Wigner and L. Eisenbud, Phys. REv. 72 (1947) 29.
103
W. Hauser and H. Feshbach, Phys. Rev. 87 (1952) 366.
104
Y.V. Fyodorov and H.J. Sommers, Phys. Rev. Lett. 76 (1996) 4709.
105
Y.V. Fyodorov and H.J. Sommers, JETP Lett. 63 (1996) 1026.
106
Y.V. Fyodorov and H.J. Sommers, J. Math. Phys. 38 (1997) 1918.
107
J. Ginibre, J. Math. Phys. 6 (1965) 440.
108
M.C. Gutzwiller, Chaos in Classical and Quantum Mechanics (Springer, New York, 1990).
109
S. Tomsovic, Phys. Rev. Lett. 77 (1996) 4168.
110
N. Ullah, J. Math. Phys. 8 (1967) 1095.
111
T.E. Krieger and C.E. Porter, J. Math. Phys. 4 (1963) 1272.
112
N. Ullah, J. Math. Phys. 4 (1963) 1279.
113
N. Ullah, Matrix Ensembles in the ManyNucleon Problem (Clarendon Press, Oxford, 1987).
114
Y. Alhassid and C.H. Lewenkopf, Phys. Rev. Lett. 75 (1995) 3922.
115
E.R. Mucciolo, V.N. Prigodin and B.L. Altshuler, Phys. Rev. B51 (1995) 1714.
116
V.N. Prigodin, K.B. Efetov and S. Iida, Phys. Rev. Lett. 71 (1993) 1230.
117
K.B. Efetov and V.N. Prigodin, Phys. Rev. Lett. 70 (1993) 1315.
118
M.V. Berry, J. Phys. A10 (1977) 2083.
119
M.V. Berry, in Chaos and Quantum Physics, edited by M.J. Gianonni, A. Voros and J. ZinnJustin (Elsevier, Amsterdam,
1990), p. 251.
120
V.N. Prigodin, Phys. Rev. Lett. 74 (1995) 1566.
121
V.N. Prigodin, B.L. Altshuler, K.B. Efetov and S. Iida, Phys. Rev. Lett. 72 (1994) 546.
122
F. Wegner, Z. Phys. B 36 (1980) 209.
123
B.L. Altshuler and V.N. Prigodin, Zh. Eksp. Teor. Fiz. 95, 348 (1989) (Sov. Phys. JETP 68 1989)) 198 .
124
V.N. Prigodin, N. Taniguchi, A. Kudrolli, V. Kidambi and S. Sridhar, Phys. Rev. Lett. 75 (1995) 2392.
125
M. Srednicki, Phys. Rev. E54 (1996) 954.
126
F.J. Dyson, J. Math. Phys. 3 (1962) 1191.
127
F.J. Dyson, J. Math. Phys. 13 (1972) 90.
128
G. Lenz and F. Haake, Phys. Rev. Lett. 65 (1990) 2325.
129
Harish-Chandra, Amer. J. Math. 80 (1958) 241.
130
C. Itzykson and J.B. Zuber, J. Math. Phys. 21 (1980) 411.
131
A. Pandey, Chaos, Solitons and Fractals 5 (1995) 1275.
132
A. Pandey, Ann. Phys. (NY) 134 (1981) 110.
133
J.B. French, V.K.B. Kota, A. Pandey and S. Tomsovic, Ann. Phys. (NY) 181 (1988) 198.
134
A. Pandey and P. Shukla, J. Phys. A24 (1991) 3907.
135
T. Guhr, Ann. Phys. (NY) 250 (1996) 145.
136
T. Guhr, Phys. Rev. Lett. 76 (1996) 2258.
137
T. Guhr and H.A. Weidenm uller, Ann. Phys. (NY) 199 (1990) 412.
138
M.V. Berry and M. Robnik, J. Phys. A17 (1984) 2413.
139
Y.V. Fyodorov and A.D. Mirlin, Phys. Rev. Lett. 67 (1991) 2405.
140
S. Iida, H.A. Weidenm uller, and J.A. Zuk, Phys. Rev. Lett. 64 (1990) 583; Ann. Phys. 200 (1990) 219.
141
M.R. Zirnbauer, Phys. Rev. Lett. 69 (1992) 1584.
142
M. Moshe, H. Neuberger, and B. Shapiro, Phys. Rev. Lett. 73 (1994) 1497.
143
T. Guhr and H.A. Weidenm uller, Ann. Phys. (NY) 193 (1989) 472.
144
P. Persson and S.

Aberg, Phys. Rev. E52 (1995) 148.
145
S. Mizutori and S.

Aberg, preprint LundMPh95/15.
146
D. Weinmann and J.L. Pichard, Phys. Rev. Lett. 77 (1996) 1556.
147
F. Leyvraz and T.H. Seligman, J. Phys. A23 (1990) 1555.
148
F. Leyvraz, J. Phys. A26 (1993) 6541.
149
G. Lenz, Dissertation, Universitat Essen, 1992.
165
150
T. Guhr and T. Papenbrock, unpublished.
151
J.B. French, V.K.B. Kota, A. Pandey and S. Tomsovic, Ann. Phys. (NY) 181 (1988) 235.
152
A. Altland, S. Iida, and K.B. Efetov, J. Phys. A 26 (1993) 3545.
153
D.M. Leitner, Phys. Rev. E48 (1993) 2536.
154
O. Bohigas, S. Tomsovic and D. Ullmo, Phys. Rep. 223 (1993) 43.
155
K. Zyczkowski and G. Lenz, Z. Phys. B82 (1991) 299.
156
H.J. Sommers and S. Iida, Phys. Rev. E49 (1994) R2513.
157
V.I. Falko and K.B. Efetov, Phys. Rev. B50 (1994) 11267.
158
V.I. Falko and K.B. Efetov, Phys. Rev. Lett. 77 (1996) 912.
159
S.A. van Langen, P.W. Brouwer and C.W.J. Beenakker, Phys. Rev. E55 (1997) R1.
160
M. Wilkinson, J. Phys. A21 (1988) 4021.
161
M. Wilkinson, Phys. Rev. A41 (1990) 4645.
162
J. Goldberg, U. Smilansky, M.V. Berry, W. Schweitzer, G. Wunner and G. Zeller, Nonlinearity 4 (1991) 1.
163
C.W.J. Beenakker and B. Rejai, Physica A203 (1994) 61.
164
M. Wilkinson, J. Phys. A22 (1989) 2795.
165
M. Wilkinson and E.J. Austin, J. Phys. A28 (1995) 2277.
166
A. Bulgac, G. Do Dang and D. Kusnezov, Phys. Rev. E54 (1996) 3468.
167
N. Pomphrey, J. Phys. B7 (1974) 1909.
168
D.W. Noid, M.L. Koszykowski, M. Tabor and R.A. Marcus, J. Chem. Phys. 72 (1980) 6169.
169
J. Brickmann and R.D. Levine, Chem. Phys. Lett. 120 (1985) 252.
170
P. Gaspard, S.A. Rice and K. Nakamura, Phys Rev. Lett. 63 (1989) 930; ibid. 63 (1989) 2540(E).
171
P. Gaspard, S.A. Rice, H.J. Mikeska and K. Nakamura, Phys. Rev. A42 (1990) 4014.
172
P. Pechukas, Phys. Rev. Lett. 51 (1983) 943.
173
J. Gibbons, T. Hermsen and S. Wojciechowski, Phys. Lett. A94 (1983) 251.
174
K. Nakamura and M. Lakshmanan, Phys. Rev. Lett. 57 (1986) 1661.
175
D. Saher, F. Haake and P. Gaspard, Phys. Rev. A44 (1991) 7841.
176
T. Takami and H. Hasegawa, Phys. Rev. Lett. 68 (1992) 419.
177
J. Zakrzewski and D. Delande, Phys. Rev. E47 (1993) 1650.
178
D. Braun and G. Montambaux, Phys. Rev. B50 (1994) 7776.
179
F. von Oppen, Phys. Rev. Lett. 73 (1994) 789.
180
F. von Oppen, Phys. Rev. E51 (1995) 2647(BR).
181
B. Li and M. Robnik, J. Phys. A29 (1996) 4384.
182
I.V. Yurkevich and V.E. Kravtsov, Phys. Rev. Lett. 78 (1997) 701.
183
A. Szafer and B.L. Altshuler, Phys. Rev. Lett. 70 (1993) 587.
184
B.L. Altshuler and B.I. Shklovskii, Zh. Eksp. Teor. Fiz. 91 (1986) 220 [Sov. Phys. JETP 64 (1986) 127].
185
C.W.J. Beenakker, Phys. Rev. Lett. 70 (1993) 4126.
186
E. Brezin and A. Zee, Phys. Rev. E 49 (1994) 2588.
187
B.D. Simons and B.L. Altshuler, Phys. Rev. Lett. 70 (1993) 4063.
188
B.D. Simons and B.L. Altshuler, Phys. Rev. B48 (1993) 5422.
189
B.D. Simons, P.A. Lee and B.L. Altshuler, Phys. Rev. B48 (1993) 11450(RC).
190
E.R. Mucciolo, private communication (1996).
191
H. Bruus, C.H. Lewenkopf and E.R. Mucciolo, Phys. Rev. B53 (1996) 9968.
192
B.D. Simons, A. Szafer and B.L. Altshuler, Zh. Eksp. Teor. Fiz. 57 (1993) 268.
193
H. Attias and Y. Alhassid, Phys. Rev. E52 (1995) 4776.
194
E. Mucciolo, B.D. Simons, A.V. Andreev and V.N. Progodin, Phys. Rev. Lett. 75 (1995) 1360.
195
C.M. Ko, H.J. Pirner and H.A. Weidenm uller, Phys. Lett. 62B (1976) 248.
196
D.M. Brink, J. Neto and H.A. Weidenm uller, Phys. Lett. 80B (1979) 170.
197
Y. Alhassid and H. Attias, Phys. Rev. Lett. 74 (1995) 4635.
198
G.E. Uhlenbeck and L.S. Ornstein, Phys. Rev. 36 (1930) 823.
199
A. Bohr and B. Mottelson, Nuclear Structure II (Benjamin, New York, 1975).
200
J.B. Garg, ed., Statistical Properties of Nuclei (Plenum, New York, 1972).
201
H.A. Weidenm uller, Prog. Part. Nucl. Phys. 3 (1980) 49.
202
J.B. French and V.K.B. Kota, Ann. Rev. Nucl. Part. Sci. 32 (1982) 35.
203
T.H. Seligman and H. Nishioka, ed., Quantum Chaos and Statistical Nuclear Physics (Springer, Berlin, 1986).
204
O. Bohigas and H.A. Weidenm uller, Ann. Rev. Nucl. Part. Sci. 38 (1988) 421.
205
V. Zelevinsky, Ann. Rev. Nucl. Part. Sci. 46 (1996) 237.
206
M. GoeppertMayer, J.H.D Jensen, Elementary Theory of Nuclear Shell Structure (Wiley, New York, 1955).
207
F. Iachello and A. Arima, The Interacting Boson Model (Cambridge University Press, Cambridge, 1987).
208
A.A. Katsanos, J.R. Huizenga and H.K. Vonach, Phys. Rev. 141 (1966) 1053.
209
J.R. Huizenga and A.A. Katsanos, Nucl. Phys. A98 (1967) 614.
166
210
H.I. Liou, H.S. Camarda, S. Wynchank, M. Slagowitz, G. Hacken, F. Rahn and J. Rainwater, Phys. Rev. C5 (1972) 974.
211
H.I. Liou, J. Rainwater, G. Hacken and U.N. singh, Phys. Rev. C11 (1975) 462.
212
A.B. Jain and J. Blons, Nucl. Phys. A242 (1975) 45.
213
W.A. Watson III, E.G. Bilpuch and G.E. Mitchell, Z. Phys. A300 (1981) 89.
214
G. Hacken, R. Werbin and J. Rainwater, Phys. Rev. C17 (1978) 43.
215
R. Chrien, Phys. Rep. 64 (1980) 337.
216
E.G. Bilpuch, A.M. Lane, G.E. Mitchell and J.D. Moses, Phys. Rep. C28 (1976) 145.
217
O. Bohigas, R.V. Haq and A. Pandey, Phys. Rev. Lett. 54 (1985) 1645.
218
T.A. Brody, E. Cota, J. Flores and P.A. Mello, Nucl. Phys. A259 (1976) 87.
219
T. von Egidy, A.N. Behkami and H.H. Schmidt, Nucl. Phys. A454 (1986) 109.
220
T. von Egidy, H.H. Schmidt and A.N. Behkami, Nucl. Phys. A481 (1988) 189.
221
A.Y. AbulMagd and H.A. Weidenm uller, Phys. Lett. B162 (1985) 223.
222
J.F. Shriner, Jr., G.E. Mitchell and T. von Egidy, Z. Phys. A338 (1991) 309.
223
A.Y. AbulMagd and M.H. Simbel, J. Phys. G22 (1996) 1043.
224
J.D. Garrett, in Proceedings of the 8th International Symposium on Capture Gamma-Ray Spectroscopy, J. Kern, ed. (World
Scientic, Singapore, 1994), p. 494.
225
J.D. Garrett, J.R. German and J.M. Espino, in Proceedings of the V La Rabida International Summer School: Response of
the Nuclear System to External Force, J.M. Arias, M.I. Gallardo and M. Lozano, eds. (Springer, Berlin, 1994), p. 263.
226
J.D. Garrett, J.Q. Robinson, A.J. Foglia and H.Q. Jin, Phys. Lett. B392 (1997) 24.
227
G.E. Mitchell, E.G. Bilpuch, P.M. Endt and F.J. Shriner, Phys. Rev. Lett. 61 (1988), 1473.
228
J.F. Shriner, Jr.., E.G. Bilpuch, P.M. Endt and G.E. Mitchell, Z. Phys. A335 (1990) 393.
229
S. Raman, T.A. Walkiewicz, S. Kahane, E.T. Jurney, J. Sa, Z. Gasci, J.L. Weil, K. Allaart, G. Bonsignori and J.F. Shriner,
Jr., Phys. Rev. C43 (1991) 521.
230
C.R. Bybee, G.E. Mitchell and J.F. Shriner, Jr., Z. Phys. A335 (1996) 327.
231
G.E. Mitchell, E.G. Bilpuch, J.F. Shriner, Jr. and A.M. Lane, Phys. Rep. 117 (1985) 1.
232
H.L. Harney, Z. Phys. A316 (1984) 177.
233
J.F. Shriner, Jr., G.E. Mitchell and E.G. Bilpuch, Phys. Rev. Lett. 59 (1987) 435; 59 (1987) 1492.
234
J.F. Shriner, Jr., G.E. Mitchell and E.G. Bilpuch, Z. Phys. A332 (1989) 45.
235
Proceedings of the International Conference on Nuclear Data for Science and Technology, in J ulich, Germany, ed. S.M.
Qaim (Springer, Berlin 1992), see especially Part X, Nuclear Models and Evaluation Methodology.
236
H. Nishioka and H.A. Weidenm uller, Phys. Lett. B157 (1985) 101.
237
D. Agassi, H.A. Weidenm uller and G. Mantzouranis, Phys. Lett. C22 (1975) 145.
238
H. Feshbach, A.K. Kerman and S. Koonin, Ann. Phys. (NY) 125 (1980) 429.
239
H. Nishioka, J.J.M. Verbaarschot, H.A. Weidenm uller and S. Yoshida, Ann. Phys. (NY) 172 (1986) 67.
240
H.A. Weidenm uller, Ann. Phys. (NY) 158 (1984) 120.
241
M. Soyeur and A. Zuker, Phys. Lett. B41 (1972) 135.
242
S.S.M. Wong, Nucl. Phys. A159 (1970) 235.
243
W.E. Ormand and R.A. Broglia, Phys. Rev. C46 (1992) 1710.
244
S. Dro zd z, S. Nishizaki, J. Speth and J. Wambach, Phys. Rev. C49 (1994) 867.
245
V. Zelevinsky, B.A. Brown, N. Frazier and M. Horoi, Phys. Rep. 276 (1996) 85.
246
R.R. Whitehead, A. Watt, D. Kelvin and A. Conkie, Phys. Lett. B76 (1978) 149.
247
J.J.M. Verbaarschot and P.J. Brussaard, Phys. Lett. B87 (1979) 155.
248
B.A. Brown and G. Bertsch, Phys. Lett. B184 (1984) 5.
249
V. Zelevinsky, M. Horoi and B.A. Brown, Phys. Lett. B350 (1995) 141.
250
F.M. Dittes, H.L. Harney and A. M uller, Phys. Rev. A45 (1992) 701.
251
H.L. Harney, F.H. Dittes and A. M uller, Ann. Phys. (NY) 220 (1992) 159.
252
S. Dro zd z, Trellakis and J. Wambach Phys. Rev. Lett. 76 (1996) 4891.
253
E. Persson, T. Gorin and I. Rotter, Phys. Rev. E54 (1996) 3339.
254
S.

Aberg, Prog. Part. Nucl. Phys. 28 (1992) 11.
255
T. Dssing, B. Herskind, S. Leoni, A. Bracco, R.A. Broglia, M. Matsuo and E. Vigezzi, Phys. Rep. 268 (1996) 1.
256
G.A. Leander, Phys. Rev. C25 (1982) 2780.
257
B. Lauritzen, T. Dssing and R. Broglia, Nucl. Phys. A457 (1986) 61.
258
B. Herskind, T. Dssing, D. Jerrestam, K. Schier, S. Leoni, J. Lisle, R. Chapman, F. Khazaie and J.N. Mo, Phys. Lett. B
276 (1992) 4.
259
B. Herskind, A. Bracco, R.A. Broglia, T. Dssing, A.. Ikeda, S. Leoni, J. Lisle, M. Matsuo and E. Vigezzi, Phys. Rev. Lett.
68 (1992) 3008.
260
S.

Aberg, Phys. Rev. Lett. 64 (1990) 3119.
261
T. Guhr and A. M ullerGroeling, J. Math. Phys. 38 (1997) 1870.
262
T. Dssing and M. Matsuo, submitted to Nuclear Physics A.
263
R. Arvieu, F. Brut, J. Carbonell and J. Touchard, Phys. Rev. A35 (1987) 2389.
167
264
R. Arvieu, P. Rozmej and M. Ploszajczak, Phase Space Structure of Simple Nuclear Models in PREDEAL Summer School:
New Trends in Theoretical and Experimental Nuclear Physics (World Scientic, Singapore, 1992).
265
B. Milek, W. Norenberg and P. Rozmej, Z. Phys. A334 (1989) 233.
266
W.D. Heiss, R.G. Nazmitdinov and S. Radu, Phys. Rev. Lett. 72 (1994) 2351.
267
W.D. Heiss, R.G. Nazmitdinov and S. Radu, Phys. Rev. C52 (1995) 3032.
268
W.D. Heiss and R.G. Nazmitdinov, Phys. Rev. Lett. 73 (1994) 1235.
269
V.M. Strutinsky, Nucl. Phys. A95 (1967) 420.
270
M. Brack, J. Damgaard, A.S. Jensen, H.C. Pauli, V.M. Strutinsky and C.Y. Wong, Rev. Mod. Phys. 44 (1972) 320.
271
V.M. Strutinsky and A.G. Magner, Sov. J. Part. Nucl. 7 (1976) 138.
272
Y. Alhassid, A. Novoselsky and N. Whelan, Phys. Rev. Lett. 65 (1990) 2971.
273
Y. Alhassid and N. Whelan, Phys. Rev. C43 (1991) 2637.
274
Y. Alhassid and N. Whelan, Phys. Rev. Lett. 67 (1991) 816.
275
V. Lopac, S. Brant and V. Paar, Z. Phys. A337 (1990) 131.
276
J.B. French, V.K.B. Kota, A. Pandey and S. Tomsovic, Phys. Rev. Lett. 54 (1985) 2313.
277
E. Blanke, H. Driller, W Glockle, H. Genz, A. Richter and G. Schrieder, Phys. Rev. Lett. 51 (1983) 355.
278
D. Boose, H.L. Harney and H.A. Weidenm uller, Phys. Rev. Lett. 56 (1986) 2012.
279
D. Boose, H.L. Harney and H.A. Weidenm uller, Z. Phys. A325 (1986) 363.
280
P.M. Endt, P. de Wit and C. Alderliesten, Nucl. Phys. A459 (1986) 61.
281
P.M. Endt, P. de Wit and C. Alderliesten, Nucl. Phys. A476 (1988) 333.
282
P.M. Endt, P. de Wit, C. Alderliesten and B.H. Wildenthal, Nucl. Phys. A487 (1988) 221.
283
A.A. Adams, G.E. Mitchell, W.E. Ormand and J.F. Shriner, Jr., Phys. Lett. B392 (1997) 1.
284
O. Sushkov and V.V. Flambaum, JETP Pisma 32 (1980) 377 (English translation JETP Lett. 32 (1983) 352).
285
V.E. Bunakov and V.P. Gudkov, Nucl. Phys. A401 (1983) 93.
286
V.P. Almenkov, S.B. Borzakov, V. Van Thuan, Yu.D. Mareev, L.B. Pikelner, A.S. Khrykin and E.I. Sharapov, Nucl. Phys.
A398 (1983) 93.
287
J.D. Bowman, C.D. Bowman, J.E. Bush, P.P.J. Delheij, C.M. Frankle, C.R. Gould, D.G. Haase, J.N. Knudson, G.E.
Mitchell, S. Penttila, H. Postma, N.R. Roberson, S.J. Seestrom, S.J. Szymanski, V.W. Yuan and X. Zhu, Phys. Rev. Lett.
65 (1990) 1192.
288
C.M. Frankle, J.D. Bowman, J.E. Bush, P.P.J. Delheij, C.R. Gould, D.G. Haase, J.N. Knudson, G.E. Mitchell, S. Penttila,
H. Postma, N.R. Roberson, S.J. Seestrom, S.J. Szymanski, V.W. Yuan and X. Zhu, Phys. Rev. Lett. 67 (1991) 564.
289
V.V. Flambaum and G.F. Gribakin, Prog. Part.. Nucl. Phys. 35 (1995) 423.
290
R.E. Trees, Phys. Rev. 123 (1961) 1293.
291
H.S. Camarda and P.D. Georgopulos, Phys. Rev. Lett. 50 (1983) 492.
292
V.V. Flambaum, A.A. Gribakina, G.F. Gribakin and M.G. Kozlov, Phys. Rev. A50 (1994) 267.
293
E. Haller, H. Koppel and L.S. Cederbaum, Chem. Phys. Lett. 101 (1983) 215.
294
T. Zimmermann, H. Koppel, L.S. Cederbaum, G. Persch and W. Demtroder, Phys. Rev. Lett. 61 (1988) 3.
295
T. Zimmermann, H. Koppel and L.S. Cederbaum, J. Chem. Phys. 91 (1989) 3934.
296
D.M. Leitner, H. Koppel and L.S. Cederbaum, J. Chem. Phys. 104 (1996) 434.
297
S. Mukamel, J. Sue and A. Pandey, Chem. Phys. Lett. 105 (1984) 134.
298
R.L. Sundberg, E. Abramson, J.L. Kinsey and R.W. Field, J. Chem. Phys. 83 (1985) 466.
299
R. Georges, A. Delon and R. Jost, J. Chem. Phys. 103 (1995) 1732.
300
J. Nygard, A. Delon and R. Jost, in preparation.
301
R. Jost and M. Lombardi, Lecture Notes in Physics 263 (Springer, Berlin, 1986), p. 72.
302
J.P. Pique, Y. Chen, R.W. Field and J.L. Kinsey, Phys. Rev. Lett. 58 (1987) 475.
303
W.H. Miller, R. Hernandez, C.B. Moore and W.F. Polik J. Chem. Phys. 93 (1990) 5657.
304
U. Peskin, H. Reisler and W.H. Miller, J. Chem. Phys. 101 (1994) 9672.
305
M.J. Davis, J. Chem. Phys. 98 (1993) 2614.
306
J.B. French and S.S.M. Wong, Phys. Lett. B33 (1970) 449.
307
O. Bohigas and J. Flores, Phys. Lett. B34 (1971) 261.
308
K.K. Mon and J.B. French, Ann. Phys. (NY) 95 (1975) 90.
309
A. M uller and H.L. Harney, Phys. Rev. C35 (1987) 1228.
310
A. M uller and H.L. Harney, Phys. Rev. C37 (1988) 2435.
311
M. Granzow, H.L. Harney and H. Kalka, Phys. Rev. C51 (1995) 3026.
312
J. Moser Stable and Random Motion in Dynamical Systems (Princeton, University Press, 1973).
313
A.J. Lichtenberg and M.A. Liebermann, Regular and Stochastic Motion (Springer, Berlin, 1983).
314
H.G. Schuster, Deterministic Chaos (Physik-Verlag, Weinheim, 1984).
315
M. Robnik, J. Phys. A17 (1984) 1049.
316
E. Teller, J. Phys. Chem. 41 (1937) 109.
317
I.C. Percival, J. Phys. 6 (1973) L229.
318
Ramaswamy and Marcus, J. Chem. Phys. 74 (1981) 1379.
168
319
M.V. Berry, in Proceedings of the Les Houches Summer School on Chaotic Behavior of Deterministic Systems, (North-
Holland, Amsterdam, 1981).
320
R.A. Marcus, in Horizons in Quantum Chemistry, edited by K. Fukui and B. Pullmann (Reidel, Dordrecht, 1980), p. 107.
321
R.A. Marcus, Ann. (NY) Acad. Sci. 357 (1980) 159.
322
G.M. Zaslavsky, Zh. Eksp. Teor. Fiz. 73 (1977) 2089.
323
G.M. Zaslavsky, Sov. Phys. Usp. 22 (1979) 788.
324
M.V. Berry, Phil. Trans. R. Soc. A287 (1977) 237.
325
M.V. Berry, J. Phys. A10 (1977) 2083.
326
M.V. Berry and M. Tabor, Proc. R. Soc. A356 (1977) 375.
327
O. Bohigas, M.J. Giannoni and C. Schmit, J. Physique Lett. 45 (1984) L1015.
328
D.W. Noid, M.L. Koszykowski and R.A. Marcus, J. Chem. Phys. 71 (1979) 2864.
329
T. Yukawa, Phys. Rev. Lett. 54 (1985) 1883.
330
S. Fishman, D.R. Grempel, and R.E. Prange, Phys. Rev. Lett. 49 (1982) 509.
331
D. Grempel, S. Fishman and R. Prange, Phys. Rev. Lett. 49 (1982) 833.
332
D. Grempel, S. Fishman and R. Prange, Phys. Rev. A82 (1984) 1639.
333
A. Altland and M.R. Zirnbauer, Phys. Rev. Lett. 76 (1996) 3420.
334
G. Casati and B.V. Chirikov, Physica D86 (1995) 220.
335
G. Casati and B.V. Chirikov, Quantum Chaos (Cambridge University Press, Cambridge, 1995).
336
W.D. Heiss (ed.), Chaos and Quantum Chaos, Proceedings, Blydepoort (South Africa) (SpringerVerlag, Heidelberg, 1992).
337
A.I. Shnirelman, Usp. Mat. Nauk. 29, (1974) 181.
338
J.E. Bayeld and P.M. Koch, Phys. Rev. Lett. 33 (1974) 258.
339
G. Casati, B.V. Chirikov and D.L. Shepelyansky, Phys. Rev. Lett. 53 (1984) 2525.
340
J.E. Bayeld, G. Casati, L. Guarneri and D.W. Sokol, Phys. Rev. Lett. 63 (1989) 364.
341
F. Benvenuto, G. Casati and D.L. Shepelyansky, Phys. Rev. A55 (1997) 1732.
342
Y.G. Sinai, Funct. Anal. Appl. 2 (1968) 61 and 245.
343
L.A. Bunimovich, Funct. Anal. Appl. 8 (1974) 254.
344
Y.G. Sinai and L.A. Bunimovich, Commun. Math. Phys. 78 (1980) 247.
345
A. Katok and J.M. Strelcyn, Invariant Manifolds, Entropy and Billiards, Smooth Maps with Singularities, Lec. Not. in
Math. 1222 (Springer, Berlin, 1986).
346
H. Weyl, Math. Ann. 71 (1911) 441; Journal f ur die reine und angewandte Mathematik, 141 (1912) 1; Journal f ur die reine
und angewandte Mathematik, 143 (1913) 177.
347
M. Kac, Am. Math. Monthly 73 (1966) 1.
348
R.B. Balian and C. Bloch, Ann. Phys. (NY) 60 (1970) 401; Ann. Phys. (NY) 63 (1971) 592; Ann. Phys. (NY) 64 (1971)
271.
349
H.P. Baltes and E.R. Hilf, Spectra of Finite Systems (Bibliographisches Institut, Mannheim, 1975).
350
O. Bohigas, in Proceedings of the Les Houches Summer School on Chaos and Quantum Physics, (North-Holland, Amsterdam,
1991), p. 89.
351
G. Casati, B.V. Chirikov and I. Guarneri, Phys. Rev. Lett. 54 (1985) 1350.
352
M. Feingold, Phys. Rev. Lett. 55 (1985) 2626.
353
G. Casati, I. Guarneri and F. Valz-Gris, Phys. Rev. A30 (1984) 1586.
354
M. Robnik, J. Phys. A16 (1983) 3971.
355
T. Ishikawa and T. Yukawa, Phys. Rev. Lett. 54 (1985) 1617.
356
M. Sieber, U. Smilansky, S.C. Creagh and R.G. Littlejohn, J. Phys. A26 (1993) 6217.
357
J.P. Keating and M.V. Berry, J. Phys. A20 (1987) L1139.
358
O. Bohigas, D. Boose, R. Egydio de Carvalho and V. Marvulle, Nucl. Phys. A560 (1993) 197.
359
E. Doron and S. Frischat, Phys. Rev. Lett. 75 (1995) 3661.
360
T. Dittrich and U. Smilansky, Nonlinearity 4 (1991) 59.
361
T. Dittrich and U. Smilansky, Nonlinearity 4 (1991) 85.
362
T. Dittrich, E. Doron and U. Smilansky, J. Phys. A27 (1994) 79.
363
T. Dittrich, Phys. Rep. 271 (1996) 268.
364
M. Sieber, H. Primack, U. Smilansky, I. Ussishinski and H. Schanz, J. Phys. A28 (1995) 5041.
365
H. Schanz and U. Smilansky, Chaos, Solitons and Fractals 5 (1995) 1289.
366
U. Smilansky, in Proceedings of the Les Houches Summer School on Mesoscopic Quantum Physics (North-Holland, Ams-
terdam, 1995), p. 373.
367
H. Primack and U. Smilansky, Phys. Rev. Lett. 74 (1995) 4831.
368
A.N. Zemlyakov and A.B. Kato, Math. Notes 18 (1975) 291.
369
P.J. Richens and M.V. Berry, Physica D2 (1981) 495.
370
T. Cheon and T.D. Cohen, Phys. Rev. Lett. 62 (1989) 2769.
371
A. Shudo, Y. Shimizu, P. Seba, J. Stein, H.J. Stockmann and K.

Zyczkowski, Phys. Rev. E49 (1994) 3748.
372
N.L. Balasz, R. Chatterjee and A.D. Jackson, Phys. Rev. E52 (1995) 3608.
169
373
R. Chatterjee, A.D. Jackson and N.L. Balazs, Phys. Rev. E53 (1996) 5670.
374
C. Schmit, in Proceedings of the Les Houches Summer School on Chaos and Quantum Physics, (North-Holland, Amsterdam,
1991), p. 331.
375
E.J. Heller, in Proceedings of the Les Houches Summer School on Chaos and Quantum Physics, (North-Holland, Amsterdam,
1991), p. 547.
376
R.E. Prange, E. Ott, T.M. Antonsen,Jr., B. Georgeot and R. Bl umel, Phys. Rev. E53 (1996) 207.
377
K.M. Frahm, P.W. Brouwer, J.A. Melsen, and C.W.J. Beenakker, Phys. Rev. Lett. 76 (1996) 2981.
378
J.A. Melsen, P.W. Brouwer, K.M. Frahm, and C.W.J. Beenakker, Europhys. Lett. 35 (1996) 7.
379
A. Altland and M.R. Zirnbauer, Phys. Rev. B 55 (1997) 1142.
380
F. Borgonovi, G. Casati and B. Li, Phys. Rev. Lett. 77 (1996) 4744.
381
K.M. Frahm and D.L. Shepelyansky, Phys. Rev. Lett. 78 (1997) 1440.
382
K.M. Frahm and D.L. Shepelyansky, cond-mat/9702135.
383
M.R. Schroder, Acoustica 4, 456 (1954), in German, English translation in J. Audio Eng. Soc. 35 (1987) 307.
384
M.R. Schroder, Number Theory in Science and Communication, with Applications in Cryptography, Physics Digital Infor-
mation, Computing and Self-similarity, 2nd enlarged edition (Springer, Berlin, 1986).
385
H.J. Stockmann and J. Stein, Phys. Rev. Lett. 64 (1990) 2215.
386
E. Doron, U. Smilansky and A. Frenkel, Phys. Rev. Lett. 65 (1990) 3072.
387
R. Bl umel and U. Smilansky, Phys. Rev. Lett. 60 (1988) 477.
388
C. Lewenkopf, A. M uller and E. Doron, Phys. Rev. A45 (1992) 2635.
389
S. Schultz et al., Bull. Am. Phys. Soc. 36 (1991) 358.
390
S. Sridhar, Phys. Rev. Lett. 67 (1991) 785.
391
S. Sridhar and E.J. Heller, Phys. Rev. A46 (1992) 1728.
392
S. Sridhar, D.O. Hogenboom and B.A. Willemsen, J. Stat. Phys. 68 (1992) 239.
393
J. Stein and H.J. Stockmann, Phys. Rev. Lett. 68 (1992) 2867.
394
J. Stein, H.J. Stockmann and U. Storegen, Phys. Rev. Lett. 75 (1995) 53.
395
S. Tomsovic and E.J. Heller, Phys. Rev. Lett. 67 (1991) 664.
396
S. Tomsovic and E.J. Heller, Phys. Rev. E47 (1993) 282.
397
A. Kudrolli, V. Kidambi and S. Sridhar, Phys. Rev. Lett. 75 (1995) 822.
398
A.D. Mirlin and Y.V. Fyodorov, J. Phys. A: Math.Gen. 26 (1993) L551.
399
H.D. Graf, H.L. Harney, H. Lengeler, C.H. Lewenkopf, C. Rangacharyulu, A. Richter, P. Schardt and H.A. Weidenm uller,
Phys. Rev. Lett. 69 (1992) 1296.
400
H. Alt, H.D. Graf, H.L. Harney, R. Hoerbert, H. Lengeler, C. Rangacharyulu, A. Richter and P. Schardt, Phys. Rev. E50
(1994) R1.
401
H. Alt, P. von Brentano, H.D. Graf, R.D. Herzberg, M. Phillip, A. Richter and P. Schardt, Nucl. Phys. A560 (1993) 293.
402
H. Alt, H.D. Graf, H.L. Harney, R. Hoerbert, H. Lengeler, A. Richter, P. Schardt and H.A. Weidenm uller, Phys. Rev.
Lett. 74 (1995) 62.
403
H. Alt, H.D. Graf, H.L. Harney, R. Hoerbert, H. Rehfeld, A. Richter and P. Schardt, Phys. Rev. E53 (1996) 2217.
404
P. So, S.M. Anlage, E. Ott and R.N. Oerter, Phys. Rev. Lett. 74 (1995) 2662.
405
U. Storegen, J. Stein, H.J. Stockmann, M. Kus and F. Haake, Phys. Rev. Lett. 74 (1995) 2666.
406
H.J. Stockmann, J. Stein and M. Kollmann, in Quantum Chaos, edited by G. Casati and B. Chirikov (Cambridge University
Press, New York 1994).
407
F. Haake, M. Kus, P. Seba, H.J. Stockmann and U. Storegen, J. Phys. A29 (1996) 5745.
408
M. Kollmann, J. Stein, U. Storegen, H.J. Stockmann and B. Eckhardt, Phys. Rev. E49 (1994) R1.
409
H.J. Stockmann, U. Storegen and M. Kollmann, J. Phys. A29 (1996) 1.
410
A. Richter, Playing Billiard with Microwaves Quantum Manifestation of Classical Chaos, to appear in Proceedings of the
Workshop on Emerging Applications of Number Theory, University of Minnesota, Minneapolis, 1996.
411
A. Kudrolli, S. Sridhar, A. Pandey and R. Ramaswamy, Phys. Rev. E49 (1994) R11.
412
F. Haake, G. Lenz, P. Seba, J. Stein, H.J. Stockmann and K.

Zyczkowski, Phys. Rev. A44 (1991) R6161.
413
S. Sridhar, D. Hogenboom and A. Kudrolli, preprint.
414
S. Sridhar and A. Kudrolli, Phys. Rev. Lett. 72 (1994) 2175.
415
S. Deus, P.M. Koch and L. Sirko, Phys. Rev. E52 (1995) 1146.
416
H. Alt, H.D. Graf, R. Hoerbert, C. Rangacharyulu, H. Rehfeld, A. Richter, P. Schardt and A. Wirzba, Phys. Rev. E54
(1996) 2303.
417
H. Alt, C. Dembowski, H.D. Graf, R. Hoerbert, H. Rehfeld, A. Richter, R. Schuhmann and T. Weiland, Physical Review
Letters, in press.
418
U. Dorr, H.J. Stockmann, M. Barth and U. Kuhl, preprint, 1997.
419
R.L. Weaver, J. Acoust. Soc. Am. 85 (1989) 1005.
420
L.D. Landau and E.M. Lifshitz, Theory of Elasticity, (Pergamon Press, Oxford, 1959).
421
O. Bohigas, O. Legrand, C. Schmidt and D. Sornette, J. Acoust. Soc. Am. 89 (1991) 1456.
422
D. Delande, D. Sornette and R.L. Weaver, J. Acoust. Soc. Am. 96 (1994) 1873.
170
423
R.L. Weaver and D. Sornette, Phys. Rev. E52 (1995) 3341.
424
C. Ellegaard, T. Guhr, K. Lindemann, H.Q. Lorensen, J. Nygard and M. Oxborrow, Phys. Rev. Lett. 75 (1995) 1546.
425
K.J. Ebeling, in Physical Acoustics, edited by W.P. Mason (Academic, New York, 1984), vol. 17, p. 233.
426
R.H. Lyon, Statistical Energy Analysis of Dynamical Systems (MIT, Cambridge, MA, 1975).
427
B. Sapoval, T. Gobron and A. Margolina, Phys. Rev. Lett. 67 (1991) 2974.
428
O. Legrand, C. Schmit and D. Sornette, Europhys. Lett. 18 (1992) 101.
429
E. Bogomolny and E. Hugues, preprint IPNO/TH 96-46, Orsay (1996). J. Phys. A26 (1993) 2371.
430
M. Dupuis, R. Mazo and L. Onsager, J. Chem. Phys. 33 (1960) 1452.
431
L.P. Kouwenhoven, C.M. Marcus, P.L. McEuen, S. Tarucha, R.M. Westervelt and N.S. Wingreen, Electron Transport in
Quantum Dots, to be published in the proceedings of the summer school on Mesoscopic Electron Transport (Kluwer 1997).
432
C. M. Marcus, S. R. Patel, A. G. Huibers, S. M. Cronenwett, M. Switkes, I. H. Chan, R. M. Clarke, J. A.Folk, S. F.
Godijn, K. Campman and A. C. Gossard, Quantum Chaos in Open versus Closed Quantum Dots: Signatures of Interacting
Particles, to appear in Chaos, Solitons and Fractals (July 1997), cond-mat/9703038.
433
R.A. Jalabert, H.U. Baranger and A.D. Stone, Phys. Rev. Lett. 65 (1990) 2442.
434
M.A. Reed and W.P. Kirk (eds.), Nanostructure Physics and Fabrication (Academic Press, New York, 1989).
435
H.U. Baranger and A.D. Stone, Phys. Rev. Lett. 63 (1989) 414.
436
C.W.J. Beenakker and H. van Houten, Phys. Rev. Lett. 63 (1989) 1857.
437
B. Eckhardt and D. Wintgen, J. Phys. A23 (1990) 335.
438
K.T. Hansen, Phys. Rev. E52 (1995) 2388.
439
P. Cvitanovic and B. Eckhardt, Phys. Rev. Lett. 63 (1989) 823.
440
R.V. Jensen, Chaos 1 (1991) 101.
441
R.B.S. Oakeshott and A. MacKinnon, Superlattices Microstr. 11 (1992) 145.
442
C.M. Marcus, A.J. Rimberg, R. M. Westervelt, P.F. Hopkins and A.C. Gossard, Phys. Rev. Lett. 69 (1992) 506.
443
H.U. Baranger, R.A. Jalabert and A.D. Stone, Phys. Rev. Lett. 70 (1993) 3876.
444
M.J. Berry, J.A. Katine, C.M. Marcus, R.M. Westervelt and A.C. Gossard, Surf. Sci. 305 (1994) 495.
445
M.W. Keller, O. Millo, A. Mittal, D.E. Prober and R.N. Sacks, Surf. Sci. 305 (1994) 501.
446
Z. Pluhar, H.A. Weidenm uller, J.A. Zuk and C.H. Lewenkopf, Phys. Rev. Lett. 73 (1994) 2115.
447
U. Gerland and H.A. Weidenm uller, Europhys. Lett. 35 (1996) 701.
448
K. Frahm and J.L. Pichard, J. Phys. I France 5 (1995) 847.
449
J. Rau, Phys. Rev. B 51 (1995) 7734.
450
K. Frahm, Europhys. Lett. 30 (1995) 457.
451
P.B. Gossiaux, Z. Pluhar and H.A. Weidenm uller, to be published.
452
C.M. Marcus, R.M. Westervelt, P.F. Hopkins and A.C. Gossard, Surf. Sci. 305 (1994) 480.
453
H.U. Baranger and P.A. Mello, Phys. Rev. Lett. 73 (1994) 142.
454
R.A. Jalabert, J.L. Pichard and C.W.J. Beenakker, Europhys. Lett. 27 (1994) 255.
455
P.W. Brouwer and C.W.J. Beenakker, Phys. Rev. B50 (1994) R11263.
456
U. Meirav, M.A. Kastner and S.J. Wind, Phys. Rev. Lett. 65 (1990) 771.
457
L.P. Kouwenhoven, N.C. van der Vaart, A.T. Johnson, W. Kool, C.J.P.M. Harmans, J.G. Williamson, A.A.M. Staring and
C.T. Foxon, Z. Phys. B85 (1991) 367, preprint.
458
P.L. McEuen, E.B. Foxman, U. Meirav, M.A. Kastner, Y. Meir, N.S. Wingreen and S.J. Wind, Phys. Rev. Lett. 66 (1991)
1926.
459
J. Weis, R.J. Haug, K. v. Klitzing, K. Ploog, Phys. Rev. B46 (1992) 12837.
460
E.B. Foxman, P.L. McEuen, U. Meirav, N.S. Wingreen, Y. Meir, P.A. Belk, N.R. Belk, M.A. Kastner and S.J. Wind, Phys.
Rev. B47 (1993) 10020(RC).
461
M.A. Kastner, Rev. Mod. Phys. 64 (1992) 849.
462
C.W.J. Beenakker, Phys. Rev. B44 (1991) 1646.
463
Y. Meir, N. Wingreen and P.A. Lee, Phys. Rev. Lett. 66 (1991) 3048.
464
R.A. Jalabert, A.D. Stone and Y. Alhassid, Phys. Rev. Lett. 68 (1992) 3468.
465
A.D. Stone and H. Bruus, Surf. Sci. 305 (1994) 490.
466
H. Bruus and A.D. Stone, Phys. Rev. B50 (1994) 18275.
467
A.M. Chang, H.U. Baranger, L.N. Pfeier, K.W. West and T.Y. Chang, Phys. Rev. Lett. 76 (1996) 1695.
468
J.A. Folk, S.R. Patel, S.F. Godijn, A.G. Huibers, S.M. Cronenwett, C.M. Marcus, K. Campman and A.C. Gossard, Phys.
Rev. Lett. 76 (1996) 1699.
469
Y. Alhassid and H. Attias, Phys. Rev. Lett. 76 (1996) 1711.
470
H. Friedrich and D. Wintgen, Phys. Rep. 183 (1989) 37.
471
H. Hasegawa, M. Robnik and G. Wunner, Prog. Theor. Phys. Suppl. 98 (1989) 198.
472
C. Neumann, Skalierte Spektroskopie am Wasserstoatom im Magnetfeld und in senkrechten elektrischen und magnetischen
Feldern, thesis (Universitat Bielefeld, 1996).
473
A. Holle, G. Wiebusch, J. Main, K.H. Welge, G. Zeller, G. Wunner, T. Ertl and H. Ruder, Z. Phys. D5 (1987) 279.
474
M. Robnik, J. de Physique Colloque 43 (1982) 45.
171
475
D. Wintgen, Phys. Rev. Lett. 58 (1987) 1589.
476
D. Wintgen and H. Friedrich, Phys. Rev. A36 (1987) 131.
477
A. Holle, J. Main, G. Wiebusch, H. Rottke and K.H. Welge, Phys. Rev. Lett. 61 (1988) 161.
478
J. Main, A. Holle, G. Wiebusch and K.H. Welge, Z. Phys. D6 (1987) 295.
479
D. Wintgen, Phys. Rev. Lett. 61 (1988) 1803.
480
D. Delande and J.C. Gay, J. Phys. B16 (1984) L335.
481
B.D. Simons, A. Hashimoto, M. Courtney, D. Kleppner and B.L. Altshuler, Phys. Rev. Lett. 71 (1993) 2899.
482
J. Zakrzewski, K. Dupret and D. Delande, Phys. Rev. Lett. 74 (1995) 522.
483
R.A. Pullen and A.R. Edmonds, J. Phys. A14 (1981) L477.
484
B. Dorizzi, B. Grammaticos and A. Ramani, J. Math. Phys. 24 (1983) 2282.
485
B. Grammaticos, B. Dorizzi and A. Ramani, J. Math. Phys. 24 (1983) 2289.
486
E. Haller, H. Koppel and L.S. Cederbaum, Phys. Rev. Lett. 52 (1984) 1665.
487
H.D. Meyer, E. Haller, H. Koppel and L.S. Cederbaum, J. Phys. A17 (1984) L831.
488
T.H. Seligman, J.J.M. Verbaarschot, and M.R. Zirnbauer, Phys. Rev. Lett. 53 (1984) 215; J. Phys. A: Math. Gen. 18
(1985) 2751.
489
T. Zimmermann, H.D. Meyer, H. Koppel and L.S. Cederbaum, Phys. Rev. A33 (1986) 4334.
490
O. Bohigas, S. Tomsovic and D. Ullmo, Phys. Rev. Lett. 64 (1990) 1479.
491
O. Bohigas, S. Tomsovic and D. Ullmo, Phys. Rev. Lett. 65 (1990) 5.
492
U. Smilansky, S. Tomsovic and O. Bohigas, J. Phys. A25 (1992) 3261.
493
M.C. Gutzwiller, J. Math. Phys. 11 (1970) 1791.
494
M.C. Gutzwiller, J. Math. Phys. 12 (1971) 343.
495
M.C. Gutzwiller, Phys. Rev. Lett. 45 (1980) 150; Physica D5 (1982) 183.
496
D. Wintgen and H. Marxer, Phys. Rev. Lett. 60 (1988) 971.
497
M.C. Gutzwiller, J. Math. Phys. 8 (1967) 1979.
498
M.C. Gutzwiller, J. Math. Phys. 10 (1969) 1004.
499
M.C. Gutzwiller, J. Math. Phys. 8 (1970) 1791.
500
M.C. Gutzwiller, in Path Integrals and their Applications in Quantum, Statistical and Solidstate Physics, ed. G.J. Pa-
padopoulos and J.T. Devreese (Plenum Press, New York, 1978), p. 163.
501
R. Balian and C. Bloch, Ann. Phys. (NY) 69 (1972) 76.
502
R. Balian and C. Bloch, Ann. Phys. (NY) 85 (1972) 514.
503
J.H. Hannay and A.M. Ozorio de Almeida, J. Phys. A17 (1984) 3429.
504
J.P. Keating, Nonlinearity 4 (1991) 309.
505
W. Luo and P. Sarnak, Commun. Math. Phys. 161 (1994) 419.
506
E. Bogomolny, F. Leyvraz and C. Schmit, Commun. Math. Phys. 176 (1996) 577.
507
J.P. Keating, in Quantum Chaos, ed. G. Casati, I. Guaneri and U. Smilansky (NorthHolland, Amsterdam, 1995), p. 145.
508
E. Bogomolny and J.P. Keating, Nonlinearity 8 (1995) 1115.
509
E. Bogomolny and J.P. Keating, Phys. Rev. Lett. 77 (1996) 1472.
510
A.V. Andreev and B.L. Altshuler, Phys. Rev. Lett. 75 (1995) 902.
511
B.A. Muzykantskii and D.E. Khmelnitskii, Phys. Rev. B 51 (1995) 5480.
512
B.D. Simons, O. Agam and A.V. Andreev, J. Math. Phys. 38 (1997) 1982.
513
P.A.M. Dirac, The Principles of Quantum Mechanics (Oxford University Press, Oxford, 1947).
514
F. Leyvraz, C. Schmit and T.H. Seligman, J. Phys. A29 (1996) L575.
515
J.P. Keating and J.M. Robbins, J. Phys. A30 (1997) L177.
516
S. Washburn, in Mesoscopic Phenomena in Solids, B.L. Altshuler, P.A. Lee, and R.A. Webb eds., Elsevier (1991).
517
D.J. Thouless, Phys. Rev. Lett. 39 (1977) 1167.
518
D. Braun and G. Montambaux, Phys. Rev. 52 (1995) 13903.
519
V.E. Kravtsov and A.D. Mirlin, Pisma Zh. Eksp. Teor. Fiz. 60 (1994) 656 [JETP Lett. 90 (1994) 656].
520
Y.V. Fyodorov and A.D. Mirlin, Phys. Rev. B 51 (1995) 13403.
521
V.E. Kravtsov and I.V. Lerner, Phys. Rev. Lett. 74 (1995) 2563.
522
E. Akkermans and G. Montambaux, Phys. Rev. Lett. 68 (1992) 642.
523
J.T. Edwards and D.J. Thouless, J. Phys. C 5 (1972) 807; D.J. Thouless, Phys. Rep. 13 (1974) 93.
524
N. Dupuis and G. Montambaux, Phys. Rev. B 43 (1991) 14390.
525
A. Altland and Y. Gefen, Phys. Rev. Lett. 71 (1993) 3339; Phys. Rev. B 51 (1995) 10671.
526
A. Altland, Y. Gefen, and G. Montambaux, Phys. Rev. Lett. 76 (1996) 1130.
527
R. Oppermann, Physica A 167 (1990) 301.
528
M.R. Zirnbauer, J. Math. Phys. 37 (1996) 4986.
529
P.W. Brouwer and C.W.J. Beenakker, Phys. Rev. B 52 (1995) R3868.
530
P.W. Brouwer and C.W.J. Beenakker, Phys. Rev. B 52 (1995) 16772.
531
U. Sivan and Y. Imry, Phys. Rev. B 35 (1987) 6074.
532
L.P. Gorkov, O.N. Dorokhov, and F.V. Prigara, Zh. Eksp. Teor. Fiz. 84 (1983) 1440 [Sov. Phys. JETP 57 (1983) 838].
172
533
N.F. Mott, Philos. Mag. 22 (1970) 7.
534
A. Altland and D. Fuchs, Phys. Rev. Lett. 74 (1995) 4269.
535
B.L. Altshuler, I.Kh. Zharekeshev, S.A. Kotochigova, and B.I. Shklovskii, Zh. Eksp. Teor. Fiz. 94 (1988) 343 [Sov. Phys.
JETP 67 (1988) 625.
536
B.I. Shklovskii, B. Shapiro, B.R. Sears, P. Lambrianides, and H.B. Shore, Phys. Rev. B 47 (1993) 11487.
537
E. Hofstetter and M. Schreiber, Phys. Rev. B 48 (1993) 16979.
538
E. Hofstetter and M. Schreiber, Phys. Rev. B 49 (1994) 14726.
539
I.Kh. Zharekeshev and B. Kramer, Phys. Rev. B 51 (1995) 17239.
540
B. Kramer and A. MacKinnon, Rep. Progr. Phys. 56 (1993) 1469.
541
R. Berkovits and Y. Avishai, Phys. Rev. B 53 (1996) R16125.
542
V.E. Kravtsov, I.V. Lerner, B.L. Altshuler, and A.G. Aronov, Phys. Rev. Lett. 72 (1994) 888.
543
A.G. Aronov, V.E. Kravtsov, and I.V. Lerner, Pisma Zh. Eksp. Teor. Fiz. 59 (1994) 40 [JETP Lett. 59 (1994) 39].
544
V.E. Kravtsov and I.V. Lerner, J. Phys. A: Math. Gen. 28 (1995) 3623.
545
B.R. Sears and H.B. Shore, unpublished.
546
A.G. Aronov and A.D. Mirlin, Phys. Rev. B 51 (1995) 6131.
547
A.G. Aronov, V.E. Kravtsov, and I.V. Lerner, Phys. Rev. Lett. 74 (1995) 1174.
548
N. Argaman, Y. Imry, and U. Smilansky, Phys. Rev. B 47 (1993) 4440.
549
S.N. Evangelou, Phys. Rev. B 49 (1994) 16805.
550
I. Varga, E. Hofstetter, M. Schreiber, and J. Pipek, Phys. Rev. B 52 (1995) 7783.
551
I.Kh. Zharekeshev and B. Kramer, Jpn. J. Appl. Phys. 34 (1995) 4361.
552
M. Feingold, Y. Avishai, and R. Berkovits, Phys. Rev. B 52 (1995) 8400.
553
E. Hofstetter and M. Schreiber, Phys. Rev. Lett. 73 (1994) 3137.
554
M. Batsch, L. Schweitzer, I.Kh. Zharekeshev, and B. Kramer, Phys. Rev. Lett. 77 (1996) 1552.
555
L. Schweitzer and I.Kh. Zharekeshev, J. Phys.: Condens. Matt. 7 (1995) L377.
556
S.N. Evangelou, Phys. Rev. Lett. 75 (1995) 2550.
557
J.T. Chalker, I.V. Lerner, and R.A. Smith, Phys. Rev. Lett. 77 (1996) 554.
558
J.T. Chalker, V.E. Kravtsov, and I.V. Lerner, JETP Letters 64 (1996) 386.
559
G. Montambaux, in Proceedings of the Les Houches Summer School on Mesoscopic Quantum Physics (North-Holland,
Amsterdam, 1995), p. 211.
560
F. London, J. de Phys. 8 (1937) 397.
561
F. Hund, Ann. Phys. 32 (1938) 102.
562
N. Byers and C.N. Yang, Phys. Rev. Lett. 7 (1961) 46.
563
F. Bloch, Phys. Rep. 137 (1965) A787; Phys. Rev. B 2 (1970) 109.
564
M. B uttiker, Y. Imry and R. Landauer, Phys. Rev. A 96 (1983) 365.
565
M. B uttiker, Phys. Rev. B 32 (1985) 1846; R. Landauer and M. B uttiker, Phys. Rev. Lett. 54 (1985) 2049.
566
H.F. Cheung, Y. Gefen, E.K. Riedel, and W.H. Shih, Phys. Rev. B 37 (1988) 6050.
567
H.F. Cheung, Y. Gefen, and E.K. Riedel, IBM J. Res. Develop. 32 (1988) 359; E.K. Riedel, H.F. Cheung, and Y. Gefen,
Physica Scripta T25 (1989) 357.
568
H.F. Cheung, E.K. Riedel, and Y. Gefen, Phys. Rev. Lett. 62 (1989) 587.
569
O. EntinWohlmann and Y. Gefen, Europhys. Lett. 5 (1989) 447.
570
H. Bouchiat and G. Montambaux, J. Phys. France 50 (1989) 2695; G. Montambaux, H. Bouchiat, D. Sigeti, and R. Friesner,
Phys. Rev. B 42 (1990) 7647.
571
L.P. Levy, G. Dolan, J. Dunsmuir, and H. Bouchiat, Phys. Rev. Lett. 64 (1990) 2074.
572
V. Chandrasekhar et al., Phys. Rev. Lett. 67 (1991) 3578.
573
D. Mailly, C. Chapelier, and A. Benoit, Phys. Rev. Lett. 70 (1993) 2020.
574
A. Schmid, Phys. Rev. Lett. 66 (1991) 80.
575
F. von Oppen and E.K. Riedel, Phys. Rev. Lett. 66 (1991) 84.
576
B.L. Altshuler, Y. Gefen, and Y. Imry, Phys. Rev. Lett. 66 (1991) 88.
577
A. Altland, S. Iida, A. M ullerGroeling, and H.A. Weidenm uller, Europhys. Lett. 20 (1992) 155; Ann. Phys. 219 (1992)
148.
578
K.B. Efetov, Phys. Rev. Lett. 66 (1991) 2794; K.B. Efetov and S. Iida, Phys. Rev. B 47 (1993) 15794.
579
A. Kamenev and Y. Gefen, Phys. Rev. Lett. 70 (1993) 1976.
580
A. Kamenev and Y. Gefen, Int. J. Mod. Phys. B9 (1995) 751.
581
A. Kamenev and Y. Gefen, Europhys. Lett. 29 (1995) 413.
582
V. Ambegaokar and U. Eckern, Phys. Rev. Lett. 65 (1990) 381.
583
V. Ambegaokar and U. Eckern, Phys. Rev. Lett. 67 (1991) 3192.
584
U. Eckern, Z. Phys. B 82 (1991) 393.
585
U. Eckern and A. Schmid, Europhys. Lett. 18 (1992) 457; Ann. Phys. (Leipzig) 2 (1992) 180.
586
R.A. Smith and V. Ambegaokar, Europhys. Lett. 20 (1992) 161.
587
N. Argaman and Y. Imry, Phys. Scr. T49 (1993) 333.
173
588
G. Montambaux, J. Phys. I France 6 (1996) 1.
589
D. Mattis, Phys. Rev. Lett. 32 (1974) 992.
590
D. Loss, Phys. Rev. Lett. 69 (1992) 343.
591
M. Schick, Phys. Rev. 166 (1968) 404.
592
F. Kusmartsev, Phys. Lett. A161 (1992) 433.
593
A.O. Gogolin and N.V. Prokofev, Phys. Rev. B 50 (1994) 4921.
594
H. Mori, Phys. Rev. B 51 (1995) 12946.
595
T. Giamarchi and S. Shastry, Phys. Rev. B 51 (1995) 10915.
596
T. Giamarchi and H. Schulz, Europhys. Lett. 3 (1987) 1287; Phys. Rev. B 37 (1988) 325.
597
M. Kamal, Z.H. Musslimani, and A. Auerbach, J. Phys. I France 5 (1995) 1487.
598
W. Hausler, Physica B 222 (1996) 43.
599
G. Chiappe, J.A. Verges, and E. Louis, Solid State Commun. 99 (1996) 717.
600
H. Mori and M. Hamada, Phys. Rev. B 53 (1996) 4850.
601
P. Kopietz, Phys. Rev. Lett. 70 (1993) 3123.
602
G. Vignale, Phys. Rev. Lett. 72 (1994) 433; A. Altland and Y. Gefen, Phys. Rev. Lett. 72 (1994) 2973.
603
A. M ullerGroeling, H.A. Weidenm uller, and C.H. Lewenkopf, Europhys. Lett. 22 (1993) 193; A. M ullerGroeling and H.A.
Weidenm uller, Phys. Rev. B 49 (1994) 4752.
604
M. Abraham and R. Berkovits, Phys. Rev. Lett. 70 (1993) 1509.
605
R. Berkovits, Phys. Rev. B 48 (1993) 14381.
606
G. Bouzerar, D. Poilblanc, and G. Montambaux, Phys. Rev. B 49 (1994) 8258.
607
G. Bouzerar and D. Poilblanc, J. Phys. I 4 (1994) 1699.
608
H. Kato and D. Yoshioka, Phys. Rev. B 50 (1994) 4943.
609
P. Schmitteckert and U. Eckern, Phys. Rev. B 53 (1996) 15397.
610
M. Ramin, B. Reulet, and H. Bouchiat, Phys. Rev. B 51 (1995) 5582.
611
G. Bouzerar and D. Poilblanc, Phys. Rev. B 52 (1995) 10772.
612
H. Kato and D. Yoshioka, Physica B 212 (1996) 251.
613
R. Berkovits and Y. Avishai, Europhys. Lett. 29 (1995) 475.
614
G. Bouzerar and D. Poilblanc, preprint (1995), condmat/9512153; in Correlated Fermions and Transport in Mesoscopic
Systems, T. Martin, G. Montambaux, J. Tran Thanh Van (eds.), Editions Frontieres, GifsurYvette (1996), p. 149.
615
R. Berkovits and Y. Avishai, Phys. Rev. Lett. 76 (1996) 291.
616
A.D. Stone and A. Szafer, IBM J. Res. Develop. 32 (1988) 384.
617
R. Kubo, Can. J. Phys. 34 (1956) 1274.
618
R. Landauer, IBM J. Res. Develop. 1 (1957) 233.
619
O.N. Dorokhov, Pisma Zh. Eksp. Teor. Fiz. 36 (1982) 259 [JETP Lett. 36 (1982) 318].
620
P.A. Mello, P. Pereyra, and N. Kumar, Ann. Phys. (NY) 181 (1988) 290.
621
C.P. Umbach, S. Washburn, R.B. Laibowitz, and R.A. Webb, Phys. Rev. B 30 (1984) 4048.
622
G. Blonder, Bull. Am. Phys. Soc. 29 (1984) 535.
623
R.A. Webb, S. Washburn, C.P. Umbach, and R.B. Laibowitz, Phys. Rev. Lett. 54 (1985) 2696.
624
S. Washburn, C.P. Umbach, R.B. Laibowitz, and R.A. Webb, Phys. Rev. B 32 (1985) 4789.
625
V. Chandrasekhar, M.J. Rooks, S. Wind, and D.E. Prober, Phys. Rev. Lett. 55 (1985) 1610.
626
C.P. Umbach, C. van Haesendonck, R.B. Laibowitz, S. Washburn, and R.A. Webb, Phys. Rev. Lett. 56 (1986) 386.
627
R.G. Wheeler, K.K. Choi, A. Goel, R. Wisnie, and D.E. Prober, Phys. Rev. Lett. 49 (1982) 1674.
628
J.C. Licini, D.J. Bishop, M.A. Kastner, and J. Melngailis, Phys. Rev. Lett. 55 (1985) 2987.
629
R.E. Howard, L.D. Jackel, P.M. Mankiewich, and W.J. Skocpol, Science 231 (1986) 346.
630
S.B. Kaplan and A. Hartstein, Phys. Rev. Lett. 56 (1985) 2403.
631
A.G. Aronov and Yu.V. Sharvin, Rev. Mod. Phys. 59 (1987) 755.
632
P.A. Lee, A.D. Stone, and H. Fukuyama, Phys. Rev. B 35 (1987) 1039.
633
A.D. Stone, Phys. Rev. Lett. 54 (1985) 2692.
634
B.L. Altshuler, Pisma Zh. Eksp. Teor. Fiz. 41 (1985) 530 [JETP Lett. 41 (1985) 648].
635
P.A. Lee and A.D. Stone, Phys. Rev. Lett. 55 (1985) 1622.
636
B.L. Altshuler and D.E. Khmelnitskii, Pisma Zh. Eksp. Teor. Fiz. 42 (1985) 291 [JETP Lett. 42 (1986) 359].
637
Y. Imry, Europhys. Lett. 1 (1986) 249.
638
D.S. Fisher and P.A. Lee, Phys. Rev. B 23 (1981) 6851.
639
R. Landauer, Philos. Mag. 21 (1970) 863.
640
M. B uttiker, Phys. Rev. Lett. 57 (1986) 1761.
641
H.U. Baranger and A.D. Stone, Phys. Rev. B 40 (1989) 8169.
642
A.A. Abrikosov, L.P. Gorkov, and I.E. Dzyaloshinski, Methods of Quantum Field Theory in Statistical Physics (Prentice
Hall, Englewood Clis, 1963).
643
K.A. Muttalib, J.L. Pichard, and A.D. Stone, Phys. Rev. Lett. 59 (1987) 2475.
644
J.L. Pichard, N. Zanon, Y. Imry, and A.D. Stone, J. Phys. (Paris) 51 (1990) 587.
174
645
J.L. Pichard, M. Sanquer, K. Slevin, and P. Debray, Phys. Rev. Lett. 65 (1990) 1812.
646
A.D. Stone, P.A. Mello, K.A. Muttalib, and J.L. Pichard, in Mesoscopic Phenomena in Solids, B.L. Altshuler, P.A. Lee,
and R.A. Webb eds., Elsevier (1991).
647
K. Frahm, private communication (1995).
648
K. Frahm and J.L. Pichard, J. Phys. I France 5 (1995) 877.
649
A. H umann, J. Phys. A 23 (1990) 5733.
650
C.W.J. Beenakker and B. Rejaei, Phys. Rev. Lett. 71 (1993) 3689.
651
C.W.J. Beenakker and B. Rejaei, Phys. Rev. B 49 7499.
652
C.W.J. Beenakker, Phys. Rev. Lett. 70 (1993) 1155; Phys. Rev. B 47 (1993) 15763.
653
K. Frahm, Phys. Rev. Lett. 74 (1995) 4706.
654
K. Frahm and A. M ullerGroeling, J. Phys. A: Math. Gen. 29 (1996) 5313.
655
A. Altland, Z. Phys. B 82 (1991) 105.
656
A. Altland, Z. Phys. B 86 (1992) 101.
657
J.A. Zuk, Phys. Rev. B 45 (1992) 8952.
658
A. M ullerGroeling, Phys. Rev. B 47 (1993) 6480.
659
S. Iida, Phys. Rev. B 43 (1991) 6459.
660
S. Iida and A. M ullerGroeling, Phys. Rev. B 44 (1991) 8097.
661
B. Schmidt and A. M ullerGroeling, Phys. Rev. B 47 (1993) 12732.
662
M.R. Zirnbauer, Commun. Math. Phys. 141 (1991) 503.
663
A.D. Mirlin, A. M ullerGroeling, and M.R. Zirnbauer, Ann. Phys. (NY) 236 (1994) 325.
664
P.W. Brouwer and K. Frahm, Phys. Rev. B 53 (1996) 1490.
665
B. Rejaei, Phys. Rev. B 53 (1996) R13235.
666
G. Casati, L. Molinari, and F.M. Izrailev, Phys. Rev. Lett. 64 (1990) 16; J. Phys. A: Math. Gen. 24 (1991) 4755.
667
G. Casati, I. Guarnery, F.M. Izrailev, and R. Scharf, Phys. Rev. Lett. 64 (1990) 5; R. Scharf, J. Phys. A 22 (1989) 4223;
B.V. Chirikov, F.M. Izrailev, and D.L. Shepelyansky, Physica (Amsterdam) 33D (1988) 77.
668
G. Casati, L. Molinari, and F.M. Izrailev, Phys. Rev. Lett. 64 (1990) 1851; S.N. Evangelou and E.N. Economou, Phys.
Lett. A 151 (1990) 345.
669
M. Wilkinson, M. Feingold, and D. Leitner, J. Phys. A 24 (1991) 175; M. Feingold, D.M. Leitner, and M. Wilkinson, Phys.
Rev. Lett. 66 (1991) 986.
670
A. Altland and M.R. Zirnbauer, Phys. Rev. Lett. 77 (1996) 4536.
671
G. Casati, F.M. Izrailev, and V.V. Sokolov, preprint (1997), condmat/9703106; A. Altland and M.R. Zirnbauer, preprint
(1997), condmat/9704129.
672
E. Abrahams, P.W. Anderson, D.C. Licciardello, and T.V. Ramakrishnan, Phys. Rev. Lett. 42 (1979) 673.
673
Y.V. Fyodorov and A.D. Mirlin, Phys. Rev. Lett. 69 (1992) 1093.
674
Y.V. Fyodorov and A.D. Mirlin, Phys. Rev. Lett. 71 (1993) 412.
675
Y.V. Fyodorov and A.D. Mirlin, Pisma Zh. Eksp. Teor. Fiz. 58 (1993) 636 [JETP Lett. 58 (1993) 615].
676
Y.V. Fyodorov and A.D. Mirlin, Int. J. Mod. Phys. B 8 (1994) 3795.
677
K.B. Efetov, Zh. Eksp. Teor. Fiz. 88 (1985) 1032; 92 (1987) 638; 93 (1987) 1125 [Sov. Phys. JETP 61 (1985) 606; 65
(1987) 360; 66 (1987) 634; M.R. Zirnbauer, Phys. Rev. B 34 (1986) 6394; J.J.M. Verbaarschot, Nucl. Phys. B 300 (1988)
263.
678
P.W. Anderson, E. Abrahams, and T.V. Ramakrishnan, Phys. Rev. Lett. 43 (1979) 718.
679
E. Abrahams and T.V. Ramakrishnan, J. NonCryst. Solids 35 (1980) 15.
680
L.P. Gorkov, A.I. Larkin, and D.E. Khmelnitskii, Pisma Zh. Eksp. Teor. Fiz. 30 (1979) 248 [JETP Lett. 30 (1979) 248].
681
F.J. Wegner, Z. Phys. B 35 (1979) 207.
682
K.B. Efetov, A.I. Larkin, and D.E. Khmelnitskii, Zh. Eksp. Teor. Fiz. 79 (1980) 1120 [Sov. Phys. JETP 52 (1980) 568].
683
A. Houghton, A. Jevicki, R.D. Conway, and A.M.M. Pruisken, Phys. Rev. Lett. 45 (1980) 394.
684
S. Hikami, Phys. Rev. B 24 (1981) 2671.
685
L. Schafer and A.M.M. Pruisken, Nucl. Phys. B 200 (1982) 20.
686
K.B. Efetov, Zh. Eksp. Teor. Fiz. 82 (1982) 872 [Sov. Phys. JETP 55 (1982) 514].
687
B.L. Altshuler, V.E. Kravtsov, and I.V. Lerner, Pisma Zh. Eksp. Teor. Fiz. 43 (1986) 342 [JETP Lett. 43 (1986) 441]; Zh.
Eksp. Teor. Fiz. 91 (1986) 2276 [Sov. Phys. JETP 64 (1986) 1352].
688
B.L. Altshuler, V.E. Kravtsov, and I.V. Lerner, Springer Proc. Phys. 28 (1988) 300; Zh. Eksp. Teor. Fiz. 94 (1988) 258
[Sov. Phys. JETP 67 (1988) 795].
689
B.L. Altshuler, V.E. Kravtsov, and I.V. Lerner, in: Mesoscopic phenomena in solids, eds. B.L. Altshuler, P.A. Lee, and
R.A. Webb (Elsevier, Amsterdam 1991) p. 449.
690
Vi. I. Melnikov, Fiz. Tverd. Tela 23 (1981) 782 [Sov. Phys. Solid State 23 (1981) 444].
691
A.A. Abrikosov, Sol. St. Commun. 37 (1981) 997.
692
B.L. Altshuler and V.N. Prigodin, JETP Lett. 45 (1987) 687.
693
V.E. Kravtsov and I.V. Lerner, Zh. Eksp. Teor. Fiz. 88 (1985) 1281 [Sov. Phys. JETP 61 (1985) 758].
694
V.E. Kravtsov and I.V. Lerner, Sov. Phys. JETP 59 (1984) 778.
175
695
V.E. Kravtsov and I.V. Lerner, Solid State Commun. 52 (1984) 593.
696
V.E. Kravtsov, I.V. Lerner, and V.I. Yudson, Sov. Phys. JETP 67 (1988) 1441.
697
V.E. Kravtsov, I.V. Lerner, and V.I. Yudson, Phys. Lett. A 134 (1989) 245.
698
N. Kumar and A.M. Jayannavar, J. Phys. C 19 (1986) L85.
699
B. Shapiro, Phys. Rev. B 34 (1986) 1519.
700
B. Shapiro, Philos. Mag. B 56 (1987) 1031.
701
B.L. Altshuler, V.E. Kravtsov, and I.V. Lerner, JETP Lett. 45 (1987) 199.
702
B.L. Altshuler, V.E. Kravtsov, and I.V. Lerner, in Mesoscopic Phenomena in Solids, B.L. Altshuler, P.A. Lee, and R.A.
Webb eds., Elsevier (1991).
703
I.V. Lerner, Phys. Lett. A 133 (1988) 253.
704
E. McCann and I.V. Lerner, Phys. Lett. A 205 (1995) 393.
705
C. Castellani and L. Peliti, J. Phys. A 19 (1986) L429.
706
H. Eilenberger, Z. Phys. 44 (1968) 1288; K.D. Usadell, Phys. Rev. Lett. 25 (1970) 507; A.I. Larkin and Yu. N. Ovchinnikov,
Zh. Eksp. Teor. Fiz. 73 (1977) 299 [Sov. Phys. JETP 46 (1977) 155].
707
A.D. Mirlin, JETP Lett. 62 (1995) 603.
708
A.D. Mirlin and Y.V. Fyodorov, Phys. Rev. Lett. 72 (1994) 526; J. Phys. I France 4 (1994) 665.
709
A.D. Mirlin, J. Math. Phys. 38 (1997) 1888.
710
V.I. Falko and K.B. Efetov, Europhys. Lett. 32 (1995) 627; Phys. Rev. B 52 (1995) 17413.
711
M. Schreiber, J. Phys. C 18 (1985) 2493; Phys. Rev. B 31 (1985) 6146; M. Schreiber and H. Grussbach, Phys. Rev. Lett.
67 (1991) 607; H. Grussbach and M. Schreiber, Phys. Rev. B 48 (1993) 6650; 51 (1995) 663.
712
W. Pook and M. Janssen, Z. Phys. B 82 (1991) 295; U. Fastenrath, M. Janssen, and W. Pook, Physica A 191 (1992) 401;
M. Janssen, O. Viehweger, U. Fastenrath, and J. Hajdu, Introduction to the Theory of the Integer Quantum Hall Eect
(VCH, Weinheim, 1994).
713
A.D. Mirlin, Phys. Rev. B 53 (1996) 1186.
714
I.E. Smolyarenko and B.L. Altshuler, Phys. Rev. B 55 (1997) 10451.
715
V.E. Kravtsov and I.V. Yurkevich, Phys. Rev. Lett. 78 (1997) 3354.
716
O.N. Dorokhov, Zh. Eksp. Teor. Fiz. 98 (1990) 646 [Sov. Phys. JETP 71 (1990) 360].
717
D.L. Shepelyansky, Phys. Rev. Lett 73 (1994) 2707.
718
D.J. Thouless, J. Phys. C 5 (1972) 77.
719
G. Czycholl, B. Kramer, and A. MacKinnon, Z. Phys. B 45 (1981) 5.
720
J.L. Pichard, J. Phys. C 19 (1986) 1519.
721
Y. Imry, Europhys. Lett. 30 (1995) 405.
722
K. Frahm, A. M ullerGroeling, J.L. Pichard and D. Weinmann, Europhys. Lett. 31 (1995) 169.
723
D. Weinmann, A. M ullerGroeling, J.L. Pichard and K. Frahm, Phys. Rev. Lett. 75 (1995) 1598.
724
F. v. Oppen, T. Wettig, and J. M uller, Phys. Rev. Lett. 76 (1996) 491.
725
F. Borgonovi and D.L. Shepelyansky, Nonlinearity 8 (1995) 877.
726
F. Borgonovi and D.L. Shepelyansky, J. Phys. I France 6 (1996) 287.
727
Ph. Jacquod, D.L. Shepelyansky, and O.P. Sushkov, Phys. Rev. Lett. 78 (1997) 923.
728
Ph. Jacquod and D.L. Shepelyansky, Phys. Rev. Lett. 75 (1995) 3501.
729
Y.V. Fyodorov and A.D. Mirlin, Phys. Rev. B 52 (1995) R11580.
730
K. Frahm and A. M ullerGroeling Europhys. Lett. 32 (1995) 385.
731
K. Frahm, A. M ullerGroeling and J.L. Pichard, Phys. Rev. Lett. 76 (1996) 1509.
732
K. Frahm, A. M ullerGroeling and J.L. Pichard, Z. Phys. B 102 (1997) 261.
733
F. v. Oppen and T. Wettig, Europhys. Lett. 32 (1995) 741.
734
N.F. Mott, Commun. Phys. 1 (1976) 203.
735
I.V. Ponomarev and P.G. Silvestrov, preprint (1996), condmat/9610202.
736
Ph. Jacquod and D.L. Shepelyansky, preprint (1996), condmat/9612006.
737
R. Romer and M. Schreiber, Phys. Rev. Lett. 78 (1997) 515.
738
K. Frahm, A. M ullerGroeling, J.L. Pichard, and D. Weinmann, Phys. Rev. Lett. 78 (1997) 4889; R.A. Romer and M.
Schreiber, Phys. Rev. Lett. 78 (1997) 4890.
739
V.E. Kravtsov and K.A. Muttalib, preprint (1997), condmat/9703167.
740
F. Calogero, J. Math. Phys. 10 (1969) 2191; ibid. 10 (1969) 2197.
741
B. Sutherland, J. Math. Phys. 12 (1971) 246; ibid. 12 (1971) 251; Phys. Rev. A 4 (1971) 2019; ibid. 5 (1972) 1372.
742
B.D. Simons, P.A. Lee and B.L. Altshuler, Phys. Rev. Lett. 70 (1993) 4122.
743
B.D. Simons, P.A. Lee and B.L. Altshuler, Phys. Rev. Lett. 72 (1994) 64.
744
B.D. Simons, P.A. Lee and B.L. Altshuler, Nucl. Phys. B409 (1993) 487.
745
O. Narayan and B.S. Shastry, Phys. Rev. Lett. 71 (1993) 2106.
746
E. Brezin, C. Itzykson, G. Parisi, and J. Zuber, Commun. Math. Phys. 59 (1978) 35.
747
G. Montambaux, D. Poilblanc, J. Bellissard, and C. Sire, Phys. Rev. Lett. 70 (1993) 497; M. Faas, B.D. Simons, X. Zotos,
and B.L. Altshuler, Phys. Rev. B 48 (1993) 5439.
176
748
F.D.M. Haldane, in Proceedings of the International Colloquium on Modern Field Theory, (Tata Institute, Bombay, India,
1994), G. Mandal and S. Das, eds., World Scientic 1995.
749
J. Minahan and A.P. Polychronakos, Phys. Rev. B 50 (1994) 4236.
750
F.D.M. Haldane and M.R. Zirnbauer, Phys. Rev. Lett. 71 (1993) 4055.
751
M.R. Zirnbauer and F.D.M. Haldane, Phys. Rev. B 52 (1995) 8729.
752
Z.N.C. Ha, Phys. Rev. Lett. 73 (1994) 1574; Nucl. Phys. B435 (1995) 604.
753
R.P. Stanley, Adv. Math. 77 (1989) 76.
754
F.D.M. Haldane, in Proceedings of the 16th Taniguchi Symposium, A. Okiji and N. Kawakami, eds., SpringerVerlag 1994.
755
P.J. Forrester, Phys. Lett. A179 (1993) 127.
756
P.J. Forrester, J. Math. Phys. 36 (1995) 86.
757
J.J.M. Verbaarschot, Minneapolis 1996: Continuous Advances in QCD, p. 325; Nucl. Phys. B (Proc. Suppl.) 53 (1997) 88.
758
T. Wettig, A. Schafer and H. A. Weidenm uller, Nucl. Phys. A 610 (1996) 492c.
759
M. A. Halasz and J. J. M. Verbaarschot, Phys. Rev. Lett. 75 (1995) 3920; M. A. Halasz, T. Kalkreuter and J. J. M.
Verbaarschot, Nucl. Phys. B (Proc. Suppl.) 53 (1997) 266.
760
T. Kalkreuter, Phys. Rev. D 51 (1995) 1305; Nucl. Phys. B (Proc. Suppl.) 49 (1996) 168.
761
J. J. M. Verbaarschot, Phys. Lett. B 368 (1996) 137.
762
H. Leutwyler and A. Smilga, Phys. Rev. D 46 (1992) 5607.
763
M. E. BerbenniBitsch, S. Meyer, A. Schafer, J.J.M. Verbaarschot, and T. Wettig, heplat/9704018.
764
J. J. M. Verbaarschot, Phys. Rev. Lett. 72 (1994) 2531.
765
A. D. Jackson and J. J. M. Verbaarschot, Phys. Rev. D 53 (1996) 7223.
766
R.A. Janik, M.A. Novak, and I. Zahed, Phys. Lett. B 392 (1997) 155.
767
T. Wettig, T. Guhr, and H.A. Weidenm uller, Proceedings of Hirschegg 97: QCD Phase Transitions, H. Feldmeier et al.,
editors, GSI Darmstadt (1997) p.69.
768
T. Wettig, A. Schafer and H. A. Weidenm uller, Phys. Lett. B 367 (1996) 28.
769
M.A. Stephanov, Phys. Lett. B 375 (1996) 249.
770
T. Guhr and T. Wettig, preprint (1997), hepth/9704055; A.D. Jackson, M.K. Sener, J.J.M. Verbaarschot, hepth/9704056.
771
J.Jurkiewicz, M.A. Nowak, and I. Zahed, Nucl. Phys. B 478 (1996) 605.
772
M.A. Stephanov, Phys. Rev. Lett. 76 (1996) 4472.
773
R. Janik, M.A. Nowak, G. Papp, and I. Zahed, Phys. Rev. Lett. 77 (1996) 4876.
774
M.A. Halasz, A. Jackson, and J.J.M. Verbaarschot, Phys. Lett. B 395 (1997) 293; and heplat/03087.
775
M.A. Halasz, J.C. Osborn, and J.J.M. Verbaarschot, heplat/04007.
776
G. t Hooft, Nuclear Physics B72 (1974) 461.
777
E. Abdalla, M.C.B. Abdalla, D. Dalmazi and A. Zadra, 2D-Gravity in Non-Critical Strings, Lecture Notes in Physics m20,
Springer, Berlin 1994.
778
A.M. Polyakov, Phys. Lett. 103B (1981) 211.
779
F. David, Nucl. Phys. B257 (1985) 543.
780
V.A. Kazakov, Phys. Lett. 150B (1985) 282.
781
J. Ambjrn, B. Durhuus and J. Frohlich, Nucl. Phys. B259 (1985) 433.
782
T. Regge, Nuovo Cim. 19 (1961) 558.
783
C.E. Porter and N. Rosenzweig, Ann. Acad. Sci. Fennicae, Serie A, VI Physica 44 (1960) 1.
784
D. Voiculescu, in Lecture Notes in Mathematics No. 1132, Springer, Berlin (1985) p. 556; D. Voiculescu, K. Dykema, and
A. Nica, Free Random Variables, American Mathematical Society, Providence, Rhode Island (1992).
785
D.F. Fox and P.B. Kahn, Phys. Rev. 134 B (1964) 1151.
786
H.S. Le, J. Math. Phys. 5 (1964) 763.
787
B.V. Bronk, J. Math. Phys. 6 (1965) 228.
788
J. Ambjrn, J. Jurkiewicz, and Yu.M. Makeenko, Phys. Lett. B 251 (1990) 517.
789
G. Akemann and J. Ambjrn , J. Phys. A 29 (1996) L555.
790
J. Ambjrn, L. Chekov, C.F. Kristjansen, and Yu. Makeenko, Nucl. Phys. B 404 (1993) 127.
791
J. Ambjrn, C.F. Kristjansen, and Yu. Makeenko, Mod. Phys. Lett. A 7 (1992) 3187.
792
G. Akemann, Nucl. Phys. B 482 (1996) 403.
793
E. Brezin and A. Zee, Nucl. Phys. B 402 (1993) 613.
794
E. Brezin and A. Zee, C. R. Acad. Sci 17 (1993) 735.
795
E. Brezin and A. Zee, Nucl. Phys. B 424 (1994) 435; J. DAnna, E. Brezin, and A. Zee, Nucl. Phys. B 443 (1995) 433.
796
E. Brezin and A. Zee, Nucl. Phys. B 453 (1995) 531.
797
C.W.J. Beenakker, Nucl. Phys. B 422 (1994) 515.
798
C. Itoi, Nucl. Phys. B 493 (1997) 651.
799
G. Hackenbroich and H.A. Weidenm uller, Phys. Rev. Lett. 74 (1995) 4118.
800
V. Freilikher, E. Kanzieper, and I. Yurkevich, Phys. Rev. E 53 (1996) 2200.
801
V. Freilikher, E. Kanzieper, and I. Yurkevich, Phys. Rev. E 54 (1996) 210.
802
D.S. Lubinsky, Acta Appl. Math. 33 (1993) 121.
177
803
G. Szego, Math. Annalen 84 (1921) 232.
804
G. Szego, Orthogonal Polynomials (American Mathematical Society, Providence, 1967).
805
D.S. Lubinsky, Pitman Research Notes in Mathematics 202 (Longman, Harlow, Essex, 1989).
806
Y. Chen and S.M. Manning, J. Phys. Condens. Matter 6 (1994) 3039; Y. Chen and K. Eriksen, Int. J. Mod. Phys. B 9
(1995) 1205.
807
C.M. Canali, M. Wallin, and V.E. Kravtsov, Phys. Rev. B 51 (1995) 2831.
808
T. Nagao and K. Slevin, J. Math. Phys. 34 (1993) 2075; ibid 34 (1993) 2317.
809
Y. Chen, M.E.H. Ismail, and K.A. Muttalib, J. Phys. Condens. Matter 4 (1992) L417; ibid 5 (1993) 177.
810
K.A. Muttalib, Y. Chen, M.E.H. Ismail, and V.N. Nicopoulos, Phys. Rev. Lett. 71 (1993) 471; C. Blecken, Y. Chen, and
K.A. Muttalib, J. Phys. A 27 (1994) L563.
811
G. Gasper and M. Rahman, Basic Hypergeometric Series (Cambridge University Press, Cambridge, 1990).
812
J.L. Pichard and B. Shapiro, J. Phys. 4 (1994) 623.
813
C.M. Canali and V.E. Kravtsov, Phys. Rev. E 51 (1995) R5185.
178