Random Matrix Theories and Chaotic Dynamics: Oriol Bohigas

RANDOM MATRIX THEORIES
AND CHAOTIC DYNAMICS
ORIOL BOHIGAS
Division de Physique Thorique1

Institut de Physique Nuclaire
F-91406 Orsay CEDEX, France
IPNO/TH
1
Unit de Recherche des Universits Paris XI et Paris VI associe au CNRS
M.-J. Giannoni, A. VQTOB and J. Zinn-Juatin, eda.

Les Bouches, Session LII, 1989
Chaos et Physique Quantique /Chaos and Quantum Physics
Elsevier Science Publishers B. V., 1991
Contents
1. Introduction 4
2. The smooth part of the spectrum 7
3. Characterisation of fluctuations 14
4. Spectral fluctuations of the Wigner-Dyson ensembles 22
4.1. Wigner's Gaussian ensembles 25
4.2. Generalisations 35
4.3. Fluctuation measures 37
4.4. The Coulomb gas 43
4.5. Dyson's circular ensembles 45
4.6. The Weyl-Balian ensembles 48
5. Theory vs. experiment, dynamical basis of RMT spectral fluctuations,
universalities 51
5.1. RMT predictions and nuclear resonance data 51
5.2. Dynamical basis of spectral fluctuations; universalities 55
5.3. More about symmetries 63
5.4. Particular cases 66
5.5. Breaking of the universality regime 72
5.6. New experiments 78
6. Breaking symmetries and induced spectral fluctuation transitions . 83
6.1. The GOE - GUE and GSE - GUE transitions 85
6.2. Dynamical basis of the GOE -* GUE transition 89
6.3. More about magnetic fields (and rotations) 93
6.4. Quantum-mechanical aspects of mixed phase space systems . . 99
A Nearest neighbour spacing distribution for Gaussian ensembles (N 2)
B n-level cluster functions (large N limit) 105
C Two-level cluster function YZ and related quantities (large N limit) 106
D Superposition of several independent spectra 108
E GOE - GUE and GSE - GUE transitions (large N limit) . . . . 110
Re{erervC5 . . . . . - - - . - . -
1. Introduction
The aim of these lectures is to present a rsum of some of the main
ideas, assumptions and results of the Wigner-Dyson type random matrix
theories (RMT) which are relevant in the general context of 'Chaos and
Quantum Physics', the subject of the present course. The beauty and
depth of RMT has been recognised and appreciated almost since their
birth ("... it is the one-dimensional theory par excellence*. Not only does
it have immediate usefulness and validity for real physical systems but,
from the mathematical point of view, it has given rise to profound results
and makes use of the deepest theorems of analysis. One might almost
say that it is also a new branch of mathematics itself." [I]). However,
until some years ago, it has remained the almost secret garden of a few
theoretical physicists with high mathematical skills. The situation now is
changing and a broader community is interested in RMT. This is partly
because RMT are providing interesting and unexpected clues to connect
classical dynamics with quantum phenomena. It is this aspect which
will be emphasised in these lectures and, concerning the main body of
RMT, we will restrict ourselves to a minimum, just in order to keep
these lectures relatively self-contained. However, emphasis will be put
on some generalisations of the "canonical" random matrix ensembles
that increase their flexibility, rendering the incorporation of relevant
physical constraints possible.
Concerning RMT, most of the information can be found in a collection
of reprints as well as in Porter's leading article in Porter's book [2], in the
review article [3], and in a new revised and enlarged edition of Mehta's
book [4]. Some aspects of the connections with chaotic dynamics are
discussed in reviews'and proceedings of conferences [5-11], in this volume
[12-17], and in three recent books by Ozorio de Almeida [18], Tabor [19]
and Gutzwiller [2O].
All along these lectures it should be kept in mind that i) simplicity may
emerge in situations which initially appear as desperately complicated
and random, and ii) probabilistic and statistical arguments may be fruit-
fully applied to systems where no external randomness nor stochasticity
has been introduced from the outside. Let us illustrate these points with
Random matrices and chaotic dynamics 5
two very simple examples.
Example 1. Consider the transcendental number TT. Its decimal (or

binary) expansion contains any given sequence of decimals (or binaries)
infinitely often. Also, the expansion passes every conceivable test of
randomness. The regular continued fraction expansion of I/TT is
1 1 1 1 1 1 1 1 1 1 1 1 1 1 (1 i}
TT ~ 3+ 7+ 15+ 1+ 292+ 1+ 1+ 1+ 2+ 1+ 3+ 1+ 14+
Any regularity in this expansion is hardly visible. Change just a little,

consider not 1/ir but four times it (minus one) and don't restrict to a
regular (numerator unity) continued fraction expansion. One has
(2n +
K 2+ 2+ 2+ 2+
*)' '
Simplicity emerges!
Example 2 [21]. n real numbers W1, . . . , wn are called linearly indepen-

dent or incommensurable if the only solution (Tn1, . . . , m n ) in integers of
the equation.
m iii = Tn1Wi H ---- + "InWn = O (1.3)
is TTi1 = mj = . . . = mn = O. The most famous set of linearly indepen-
dent numbers is any sequence
In pi, InP1, ... , In Pn , (1.4)
of logarithms of primes p,- (pt = 2, pi = 3, . . .). In this case linear

independence is equivalent to the unique factorisation theorem for an
integer.
Take now U1 , . . . , W n linearly independent numbers and construct the
functions
COSW 1 E, COSW 2 E, ... , COSWnE . (1.5)
One can show that they are statistically independent, i.e. for all real
ai,. -,Qf n one has
I COSW1E < Ct1, . . . , C O S W n E < On) = JJ (iR{t

t=l
(1.6)
6 Oriol Bohigaa
In (1.6), MR{A} is the relative measure of a set A of real numbers, defined

by
(1.7)
where A n (-T, T) is the subset of A which lies in the interval ( T, T).
In words, the fraction of time t during which
COSW1J < O1, cosu 2 t < O 2 , , cosu n < CLn (1-8)
simultaneously hold is equal to the product of the fractions of time

during which the inequalities are separately fulfilled. This means that
the fractions of time during which the inequalities hold behave as if they
were probabilities of independent events. The purely arithmetic concept
of linear independence may therefore be closely related to the concept
of statistical independence!
Define now xn(t) as the superposition of n vibrations with incommen-
surable frequencies
and let A n (O 11 O 2 ) be the relative time that xn(t) spends between O1 and
O2. One has that
Um /Ia(A n (O 11 O 2 )) = -L= Pe-'/Mz , (1.10)

" \/27T Ja 1
which is the normal law. As M. Kac phrases it, "that we are led here to
the normal law, usually associated with random phenomena, is perhaps
an indication that the deterministic and probabilistic points of view are
not as irreconcilable as they may appear at first sight" . And the bon mot
of Poincar concerning the normal law ( " Tout It monde y croit pare que
les mathmaticiens s'imaginent que c'est un fait d'observation, et les
observateurs que c'est un thorme de mathmatiques7'} appears now
transparent: the normal law proves to be both "un fait d'observation"
and "un thorme de mathmatiques" .
In what follows we shall mainly discuss the description of the statisti-
cal laws governing a sequence of points 1, 2, . . - , Cn distributed on the
real axis. We shall denote by {,} the ordered sequence, |, < |1+1, and
by {&} the sequence when the ordering (labelling) is disregarded. The
axis may be the energy, the points corresponding to the discrete energy
levels of a quantum system (an atomic nucleus, an atom or a molecule);
or it may be the frequency axis, the points corresponding, for instance,
to the normal frequencies of a vibrating membrane, or the eigenfrequen-

cies of a microwave cavity, or the eigenfrequencies of a small metallic
block. Before the study of the fluctuations one of our main goals
can be attt mpted, one must make a separation between the smooth (or
average) behaviour and the fluctuating (or oscillating) part. This is the
subject of the next Section.
2. The smooth part of the spectrum

We shall now consider the spectrum of the Schrodinger operator H
in d dimensions. The spectrum is determined by the eigenvalue equation
(2.2)
Let the spectrum of H be discrete, let
EI < Et < < < (2.3)
be the eigenvalues of (2.2) and let us introduce the staircase or counting

function
N(E) = B(E - Ef) = Yl I = KO(E-A) (2-4)
i= 1 Bi<B
and its derivative, which gives the density of states
AN
P(E) = aJ =' S(E ~ Ei) = tr S(E - H) . (2.5)
What we want is to extract from N-(E) its smoothed or average part;

the rest is what will be called the fluctuating or oscillating part
N(E) =< N(E) > +NUX(E), (2.6)
the average level density < p > being given by

8 Oriol Bohigaa
The simplest estimate of < N(E) > can be obtained by applying the
semiclassical rule that each quantum state is associated with a phase-
apace volume hd where k is Planck's constant. One has therefore
If the potential V(q) is homogeneous of degree p,
V(aq) = aV(q), (2.9)
one can easily see from (2.8) that

/1+ J A *
< N(E) > (^ . (2.10)
Emphasis will be often put on homogeneous potentials because the clas-

sical dynamics governed by them scales trivially with energy, and when
going to the semiclassical limit the dynamics is then kept unaltered.
Consider now the case of a free particle enclosed in a box of volume
V. The q integration can be performed in (2.8) and one obtains
which is consistent with (2.10) for /i - oo. For a plane two-dimensional

domain D (d = 2) with boundary 9D1 the problem of the free quantum
particle with Dirichlet boundary conditions is equivalent to the problem
of a vibrating membrane (Schrodinger equation = Helrnholtz equation),
a drum whose boundary is 3D. Classically it corresponds to a billiard
problem, namely free motion with elastic reflection when the point par-
ticle bounces at the boundary. Eq. (2.11) reads then as
E1 (2.12)
where 5 is the area of D ; or, using units h - 2m = 1 (E -+ (2m/h2)E) ,
<N(E)>^SE. (2.13)
4ir
This estimate of < N(E) > is called in the literature the Thomas-
Fermi estimate or, for boxes, the Weyl term of an expansion of < N(E) >.
It is unaltered if one takes Neumann instead of Dirichlet boundary con-

ditions. The Thomas-Fermi estimate is the first term of an asymptotic
series in powers of ft of < N(E) >. To generate the expansion one applies
the Wigner transform to calculate average level densities (see Balaza and
Jennings [22] for a general discussion) . For an operator the transform
and its inverse are given by
A(q,p) = [A|.(q,p) = / dx e-**/* < q + ||q - | > (2.14)
dp
One desirable feature from a. semiclassical point of view is the following
exact association of the trace of an operator with a (pseudo) phase-apace
integral
[AB]w(q,p) = [Al w (q l ,p')e lA ' lk [J3] w (q ,p ) (2.17)

P-=P=P"
where A. is the Moyal bracket [23]
=Vq-Vp- -Vq-Vp. - (2.18)
One then considers the Laplace transform of the density of states
Z(P) = /J0 e-'*p(E) (LB = tr (e^) = -^ J dqdp [e-"ff]w(qi p)

(2.19)
and computes the Wigner transform of e~^, integrates over "phase
space" and performs the inverse Laplace transform to recover the level
density. With the exception of the harmonic oscillator for which the
exact result is easily found, one must usually settle for an ft-expansion
of [e~fl*]w(q,p) which finally leads to an ft-expansion of the average level
density.
When the method just sketched is applied, for instance, to two coupled
anisotropic quartic oscillators
3=(p\+Pl) + ti + l>ti + ltqi$ (2-20)

10 Oriol Bohigaa
one gets
pS/2
< N(E) > = tt(0 Lj- + .j(, 6) + 0(AE- 3 / 2 ) , (2.21)
ft
where the functions a() and "/(,&} can be explicitly determined [24).
The leading term is, as it should be, consistent with (2.10).
Because of discontinuities, the Wigner-Weyl calculus cannot be
naively applied to problems with hard walls, and the methods and re-
sults obtained in this case to go beyond (2.12) are reviewed 'in [25]. For
the billiard case one has
<N(E)>~ [SE^tVE+ K] (2.22)

4JT
where is the perimeter of the boundary 3D. The minus (plus) sign
applies when Dirichlet (Neumann) boundary conditions are used. XT is a
constant term containing information on the geometrical and topological
properties of the domain D.
The geometrical features contributing to the constant term are:
i) Curvature contribution
(2.23)
where /c() denotes the local curvature; for the circle, for instance, this
contribution is 1/6.
ii) Corner contribution; for a square (or a rectangle), it is 4 x (1/48).
The topological features concern the connectivity of the surface; for a
multiply connected domain containing r smooth holes, the contribution
to the constant term is (l - r) x (1/6) [26].
There are interesting cases for which the above methods fail. Consider,
for instance, the domain enclosed by the positive 1- and 2-axes and the
hyperbola q^qz 1 in the first quadrant,
D = {(ft, C2) I 9i > O n fe > O n qiqi < 1} . (2.24)
Through the hyperbolic shaped channels (horns), the particle in such a

billiard can escape to infinity. Despite this, the quantum spectrum is
proven to be purely discrete [27] but, of course, (2.13) does not apply
since 5 diverges and the relationship between the density of states and
the phase space volume does not hold. Nevertheless, a modification
of the standard application of Wigner transform techniques leads to
a method of calculating the leading behaviour, first derived by Simon

[27]. To go beyond the leading term, methods to deal with "horn-shaped
regions" have been devised. The main idea is that far out in the narrow
channel regions there exista an adiabatic separation of the degrees of
freedom along and perpendicular to the "axes" of the channels. The
result for this hyperbolic shaped billiard, as derived by Steiner et al.
[28] is
(2.25)
47T
where
a = 2(-r-ln2:r) , 6 = 47r s /Vr 2 (l/4) . (2.26)
Notice that if in (2.24) one takes qrfi < 2, then the leading term in (2.25)
coincides with the leading term of the asymptotic expansion for the
number of non-trivial zeros of the Riemann zeta function with imaginary
part less than E [12|.
A similar problem has been studied by Tomsovic [29], but now with
a smooth potential instead of hard walls, namely
B= 2^(P + P:) + Will (2-27)
for which again the quantum spectrum is purely discrete [27]. If one
takes the symmetry properties into account, one can easily see that
(2.27) is a particular case of (2.20) for b - 1 and f = -1. This Hamilto-
nian has received particular attention for its relevance to certain simple
field theoretical models and for its dynamical properties. This last point
deserves a. short digression.
Examples of systems where global ergodicity and instability has been
rigorously proven (hard chaos in Gutzwiller's terminology) are billiards
and free motion on Riemann surfaces of negative curvature [16,20,30,31].
The question whether there exist analytic Hamiltonians in Euclidean
space with this property has focused much attention but still remains
open. There have been for a long tune two candidates for that (both for
d = 2). One is the Hamiltonian corresponding to the anisotropic Kepler
problem [20]
H = A- + A --- -L= (2.28)
2Tn1 2m, ^+ qz
which appears when studying impurity levels in a semiconductor, and

the other is the Hamiltonian (2.27). For (2.28), it has been thought for a
long time that when the mass ratio mi/m-i exceeds 9/8, the system shows
12 Oriol Bohigas
hard chaos. But Broucke discovered a few years ago, for Jn^m2 = 3/2, a
very small stability island which disappears when the mass ratio attains
the value 2. The challenge is now to find for mi/mj > 2 a stability
island, or to prove that there is none. A similar fate occurred with the
Hamittonian (2.27), which was thought to have global ergodicity but for
which recently Dahlqvist and Russberg [32) have found the presence of
at least one small stable island, which occupies a fraction of the order
of 5 x 10~5 of the surface of section area. Although these questions
are interesting from a mathematical point of view, as a physicist one is
still tempted to qualify the dynamics of Hamiltonians (2.27) or (2.28) as
showing hard chaos, especially in their transition to quantum mechanics.
Except,, of course, if the phenomenon under consideration would have
an ^resolution of the order of the dimension of the stability islands, an
unlikely circumstance.
Adapting the method used for the case (2.24), Tomsovic finds for
(2.27) [29)
<N(E}>~ (E^lnE + a'E5!2) (2.29)
2ir
where
a' = (2/3) (27 + 5 In 2 -4) (2.30)
4 1/s
and energy is measured in units of (fe x/16m ) . Due to the presence
of the "horn-shaped regions", (2.29) differs from (2.21) in a similar way
as (2.25) differs from (2.22), namely by the presence of a logarithmic
term,
It is often necessary to take explicitly into account the exact symme-
tries of the problem and one is then Interested in decomposing < N[E] >
accordingly,
, (2.31)
where v labels a symmetry class. This can be done in some cases in a

very simple way. Consider for instance the Bunimovich stadium, a do-
main bounded by two opposite sides of a rectangle and two semicircles
closing the box. with Dirichlet boundary conditions (see Fig. 11). The
system has the symmetry of the rectangle, the spectrum shows no degen-
eracies and the states can be labelled by it = (C1, C1) (E, = 1, = 1, 2)
according to their symmetric or antisymmetric character with respect
to the :-th axis. For instance, an even-odd state t/i + _(q) is symmetric
(antisymmetric) when reflected with respect to the 1- (2-) axis. Let
N(E), N\(E), N^(E) and N+(E) denote the staircase functions, when
Dirichlet boundary conditions are used, corresponding to the complete
stadium, to the right (or left) half of the stadium, to the upper (or lower)
half of the stadium and to the quarter of the stadium respectively. One
has
N(E) = Nt+(E) + jV_ + () + N+.(E) + N_.(E)
JV 1 (JS)= N,.(E)+N._(E)
l
N-(E)= N.+ (E) +N-(E) '
N+(E)= N-(E).
For the functions < N(E) >, < N\(E) >, < N-(E) >, < N+(E) >,
(2.22) is valid. Therefore, by inverting (2.32) and taking the smooth
parts, one obtains < NV(E) > [33]. The leading term (surface term) of
< NV(E) > is just one fourth of the corresponding term of < N(E) >
and is independent of v. The !/-dependence comes only through the
perimeter and constant terms in (2.22).
Analogously, the hyperbola billiard (2.24) is invariant under reflection
about the straight line 91 = qt and, correspondingly, states can be clas-
sified according to their parity = 1. The average symmetry-resolved
counting function < N e ( E ) > reads then as [28]
j
with bc = b + 62V2, see (2.26).

It is also possible, when using Weyl-Wigner calculus, to incorporate
symmetry decompositions. For the coupled anisotropic quartic oscilla-
tors (2.20) one has also the (E1, E2) labelling as for the stadium, and use
must be made of the Wigner transforms of the parity operators which
are given by
[*M1, P) = ** 6(9i)6(pt) ('" = 1, 2) (2-34)
One obtains, for instance, for the staircase function [24]
lf, ,
4 71 ft 4
(2.35)
As for the stadium case, the leading term is just one fourth of the leading
term of < N(E) > (see (2.21)), but the counting function now contains a
next-to-leading term E3t*/h analogous to the perimeter term in (2.22),
which cancels out when summing over the four symmetry classes.
14 Oril Bohigaa
For homogeneous potentials of degree ft, the semiclassical expansion

cf < N(E) > is in general given in powers of
k/E>+* (2.36)
and the leading term is given by (2.10). This is consistent with the
billiard results for /* > oo, although one should notice that the limit of
the potential V(q) = c2n, n -^ oo, is not the "one-dimensional" billiard
with the wall located at q = 1. Indeed, the boundary conditions are
different. In the semiclassical limit, the difference of phase between the
incoming and outcoming waves at turning points is jr/2 for the potential,
whereas it is TT (O) for the billiard with Dirichlet (Neumann) boundary
conditions.
In summary, semiclassical methods allow to perform the separation
(2.6) in a systematic way. On Fig. 1 are reproduced two examples where
the exact staircase function is compared to the smooth part. It is a
remarkable fact that the ft-expansions of < N(E) > we have been dis-
cussing, although asymptotic, work extremely well down to the ground
state, even when only the first few terms are included in the expansion.
Before we deal with fluctuations, let us describe the operation of un-
folding the spectrum. The aim is to get rid of < N(E) > when com-
paring the fluctuation patterns of different systems labelled w, whose
corresponding smooth behaviours are not the same. Let {Et} be the
original spectrum corresponding to a system u with average counting
function < N^ (E) > ; one "unfolds" it (we are using this terminology
waiting for a better one; Voros suggests "rectifying the spectrum") by
the mapping [8]
As a result, the ordered sequences {,} have the same average behaviour
<N(x)> = x, (2.38)
namely a constant average density equal to unity, independent of w. The

fluctuating or oscillating part of {,} corresponds then to departures
from uniformity. As a function of i, the deviation of the i-th "level" Xj
from the average behaviour
c t \ I O fO QQ^
1
~ 2J - I ' I
will fluctuate around zero. How to characterise the behaviour of {<} or
of (Ai) is the subject of the next Section.
10 20 30 40 50 60 _ 70
0.0 5 O 10.0 15.0 20.0

E
Fig. 1. Comparison of the exact counting function N (E) with its semiclassical
average < N(E) > : (a) for a stadium with straight line segment and radius
equal to one, with (2.22) (taken from [34]). (b) For the "horn-shaped" potential
(2.27) to within an arbitrary additive constant: dashed continuous line, with
leading term of (2.29); full continuous line, with next-to-leading term included
(taken from [29]).
16 Oriol Bohigas
3. Characterisation of fluctuations
Let
I1, xj,..., XN (3.1)
be the positions of N points on the real axis, with average density unity
and let
PN(Xl,...txN)dXl---dxN (3.2)
be the probability regardless of labelling of having one point at X1, an-
other at Z1, ..., another at x# within each of the intervals [x,-, x, + ir,].
The statistical properties of the sequence {i,} are characterised, by
the set of n-lcvcl correlation functions Rn(Xi,* . . , Xn)
N\ r
Rn(Z1,..., xn) = --JdX n + 1 - --dxtt PN(Xl,..., XN) . (3.3)
From the definition it follows that by integrating /Zn+ 1 once, one obtains
n+1 (Xi,... , x n +i) dxn+1 = (N -B) JZ n (X 1 ,..., Xn) . (3.4)
It is convenient to introduce the n-levcl cluster functions Yn which are

obtained from Rn by subtracting out the lower-order correlation terms
Y n (X 1 ,... , xn) = (-)"->i- I)Jn^(X,, with tin G,-) . (3.5)

7= 1
Here G stands for any division of the indices [I1 2, . . . , n] into subgroups
[Gi, G1, ... Gm\. For instance,
n=l m = l [(I)] , .
{
therefore Y1(X1) = B1(X1) = 1 ; >
n =2 m= 1 [(1,2)]
m = 2 [(1)(2)1 (3.7)
therefore Y2(X1, X2) = -JZ2(X1, X2) +
n=3 m = l [(1,2,3)1
therefore Y3(X1, X2, xs) = JZ5(X1, X2, xs)

-(JZ 1 (x l )-ff2(i2,X3)+JZ 1 (x 2 )JZj(x ll x 3 )+JZ 1 (x s )JZ 2 (x 1 ,x 2 )]+2JZ 1 (x 1 )JZ 1 (x 2 )JZ 1 (xs),
etc. The advantage of the cluster functions over the correlation functions
is that they have the property of vanishing when any one (or several) of
the separations (x, - x,\ becomes large. The function Fn describes the
correlation properties of a single cluster of re levels, isolated from more
trivial effects of lower order correlations.
What we are doing is just transposing what is done in statistical me-
chanics to the study of points (3.1) on the real axis (a one-dimensional
system). To get a better insight on the different correlation functions
and on what physical information they carry, let us, for the sake of anal-
ogy and comparison in later Sections, make some comments related to
the study of equilibrium properties of classical fluids.
The probability distribution for observing the system at the configu-
ration space point r lt . . . , r^ is
...., FJV)J
(3 9)
Krfr,...,^)! '
where U is, for instance, a pair potential
"KO. (3-10)
rij = JFj r,-|, and ft 1/fcT is the inverse temperature. Notice that
when studying the positions of points on the real line (the spectrum) we
have not yet specified how to determine a probability PH in (3.2) and
that, for the tune being, nothing like a pair potential v(r,-y) is at our
disposal.
One also introduces the n-point correlation functions
Nl f
Pn(F1,..., Tn)= / j y j x , J dr n + 1 ---dr A r P f t r (ri,...,rAf) (3.11)
as we did in (3.3). One of the goals in statistical mechanics is to com-

pute the correlation functions and other interesting physical quantities
in the thennodynamic limit once v(r,-,-) is known, say the Leonard-Jones
potential. The use of the correlation functions is convenient because
it introduces a hierarchy. As we shall illustrate Ln a moment, many
interesting properties of the system are already determined from the
knowledge of PI(FI, TJ) alone. Keeping the notation which is used in the
theory of fluids, one introduces
ff(ri,r,) = Pi(F1, F2)^p1(F1)P1(F2)] (3.12)

Oriol Bokigas
- r
a \fla \/3a
Fig. 2. Schematic plot of the radial distribution function. Left: for a simple
Quid. Right: angle average of g(r) for a highly ordered solid. (Taken from
[35]).
which corresponds to R2(xltx2) in (3.3). For an isotropic fluid

PI(I-I) -P = const. , (3.13)
(the unfolding is unnecessary), and g(ri,ri) depends only on the relative
distance r t j = r. The function g(r) is called the radial distribution
function. To measure departures from the uncorrelated system one also
introduces
Mr) = 9(r) ~ 1 (3-14)
which corresponds to YI (see (3.7)) and which vanishes for an ideal gas.
On Fig. 2 we give a schematic view of the radial distribution function
of a gas, a liquid and a highly ordered solid. For a liquid or a solid,
g(r) can be measured by X-ray or neutron diffraction. The maxima of
g(r) are located at no for a liquid (\Jna for a solid), where a ia a length
which measures the distance from a given atom to its first neighbours.
In the dilute gas phase, the range of correlations is just the range of the
pair potential (r,-y) and there is no layering. For a liquid, g(T) shows an
oscillatory form which persists beyond the range of the pair potential.
The angle average of g(T) for a solid, compared to a liquid, shows much
lower values of g(r] in-between the maxima.
From the knowledge of the radial distribution function alone, and if
the potential is pair decomposable (3.10), the internal energy E is given
by
" 3, _ 1
and the equation of state of the fluid is

where p is the pressure.

What precedes suggests that a possible strategy in the study of clas-
sical Quids may be to concentrate on g(r) and forget about higher than
two-point correlation functions, at least 'initially. But literally this is not
possible because the different order correlation functions satisfy the set
of coupled partial differential equations
-Pj p n + i(ri,..., IWu)V 1 W(Fi n+1) drn+l

(3.17)
from which it is clear that pn cannot be determined without the knowl-
edge (or some approximation) of /Jn+ 1. Introducing w(r) through
g(r) = e-*"(r> , (3.18)
(3.17) reads, for n = 2, as
Ps(rir ra)
- - / Ps(ri;-r" ra) V1U(F13) drs . (3.19)
PJ PZI IZ
Using the fact that
(3.20)
ft
gives the force between particles 1 and 2 averaged over the equilibrium
distribution of all the other particles, (3.19) has the following physical
interpretation. The force on particle 1 when another is distant F12 from
it has two contributions: one is the direct interaction between particles
1 and 2, and the other is the interaction between 1 and 3 at rs, the
force being weighted with the factor ps/pp2 which gives the probability
of finding 3 at r3, given that there are particles at T1 and F2.
Going now back to the spectrum (3.1), let us briefly describe fluctua-
tion or correlation measures which are in common use, their usefulness
and applicability depending upon the length of the spectrum and upon
which features of the correlation functions one wants to emphasise. We
will often assume that the fluctuations don't depend on the location on
the spectrum, i.e. that the spectrum is stationary, like for an isotrcpic
fluid. In this case the two-level cluster function Y2 depends only on the
relative distance |i12| = |it ij|. One also uses its Fourier transform
b(k) called the two-level form factor
b(k) = f+a y2(x)e'2*fcldz . (3.21)

J-OO
20 Oriol Bohigaa
Ideally, to compare with predictions, the best procedure is first to com-

pute YI, afterwards Yj, etc... .
Let us denote by p(k;s) the density distribution of spacings s =
i-t-fc+i - ; between two levels Xi and i,-+i+ 1 having k levels in between
(Jt = 0, 1, . . .). The nearest neighbour spacing distribution is p(0; s) which
is usually denoted by p(s). The densities p(k;s) are admixtures of all
fc-point correlation functions, but their sum is a two-point function
(3.22)
Ie=Q
Most of the time, p(s) is used to characterise short range level correla-
tions. However, this information is also contained in measures definable
in terms of pure ra-potnt functions [36 j, like
Rn(L) = - Rn(X1, . . . , Zn) dz, din . (3.23)

JO JO
For small L, Rn(L)InI gives the probability that an interval of length L

contains n levels.
Sometimes it is useful to consider the variable n(L), "the number of
levels contained in an interval of length L", or in short, the number
statistic (37,4). By construction, its average is
n(L) = L , (3.24)
and its variance 2 () is a pure two-point measure related to Yj amd b

by [4]
= (n(L) - L)* = L-2f\L-r)Y,(r)dr =

(3.25)
^S2(r) = -2y,(r) . (3.26)
From (3.25) one sees that E2 (L) starts linearly at the origin, with slope
unity independent of the correlations. S7 is connected to RI through
1, z,)dz!dz, - 2 [L(L-s)R2(s)ds = Z*(L)+L(L-l)

JO
(3.27)
To go beyond 2-point measures, one may consider /Zs(L) and

which are pure 3- and 4-point measures, or the sJkeuiness 1Td(L) and the
excess Tz(L) of n(L) [36] which read as
= (n(L)-L)* , }
= s(L)-3(L-l)E2(L)-L(L-l)(L-2) l
'
T,(L)-E 4 (L) = (n(L) - L)* - 3E<(L)

= JZ4(L) -(4L -6)- Ti(L) -E S (L)-3E*(L)
-(6L2 - 18L + 11) E 2 (L) - L(L - I)(L - 2)(L - 3)
(3.29)
Aa can be seen, TI and TS derive from the (2+3)-, and (2+3+4)-level
functions, respectively.
Finally, another convenient variable is the Dyson-hfchta least square
deviation A(L) of the staircase function N(X) from the best straight
line fitting it in an interval of length L [37,4 j. Even if the original
spectrum has been unfolded (see (2.38)), the best straight line fitting
N(X) when considering a given interval [a, a + L] will not just be y = x
but another (presumably close-lying) straight line. The average of A(L)
can be expressed in terms of E2 or YI [4,38,39]
A(L) = j (L3 - 2L2r + r s )E 2 (r)dr

(3.30)
and is therefore also a pure two-point measure which starts linearly at

the origin with slope 1/15 independent of the correlations.
In summary, when analysing sequences (3.1) to be compared to theo-
retical predictions, one should attempt i) to determine as well as possible
Y1 and even higher-order level cluster functions, ii) to avoid performing
redundant analyses, like evaluating quantities which are not indepen-
dent. However, the latter may be difficult to decide when going beyond
two-level functions because even if three- and higher-level correlations
are not determined by the two-level function, lower-order functions im-
pose complicated constraints which are largely unknown or hard to ex-
ploit on the higher-order ones (see (3.17)).
To study correlated spectra it is useful to keep in mind, for the sake
of comparison, the limiting case of an uncorrelated or Poissonian spec-
trum {,}. It can be generated by taking a random variable s whose
22 Oriai Bohigas
probability density is e~* and by constructing the ordered sequence {,-}

from
I1 = O , Xi+I = it+ S t , i = l,2,... (3.31)
where s,- are outcomes of independent trials of the variable s. For a
Poisson spectrum one has for the correlation functions:
Rn(X1,..., xn) = l
Y1(X1) = 1 (3.32)
Yn(X1,..., Xn) = O n > 2 ;
and for the spacing distribution, Rn, the number variance, the Dyson-
Mehta statistic, the skewness and the excess:
p(3) = e - , Rn(L) = L " , E*(L) = L , , ,

A(L) = L/15 , T1(L) = L-1/2 , -Y2(L) = L~l . v
' '
The system which corresponds to a Poisson spectrum in the case of

classical Quids is the ideal gas, for which there are no correlations and
for which, using notation (3.11),
P(ri,...,T n ) = pn . (3.34)
4. Spectral fluctuations of the Wigner-Dyson en-

sembles
In this Section are described the "non-trivial" theories of spectral fluc-
tuations, namely fluctuations of eigenvalues of random matrices [2,4].
The random matrix theories we are concerned with have their origin in
the following physical problem. In the low energy region of the excita-
tion spectrum of an atomic nucleus, the average level density < p(E) >
is small and one expects to describe most of the states in a detailed way
using nuclear models. However, the average level density increases very
rapidly with the excitation energy E :
where c, A and a are constants for a given nucleus. Therefore, by

the time one reaches the region, for instance, of the neutron threshold
(~ 6 MeV), the number of levels is so high that one must give up a
description of microscopic detail, a description of the individual levels
20 30 UO SO 60 70 30 90 100 "O 120
210
Pig. 3. Total ciosa section for the reaction n +2M Th as a function of the
neutron energy En. The peaks show the resonances of the compound nucleus
253
Th at an excitation energy S 4-.En, where S is the neutron separation energy
of "3Th (~ 5.1 MeV). (Taken from the compilation Neutron Cross Sections
(1964) [4O]).
24 Oriol Bohigaa
(see Fig. 3). The aim of nuclear models at this and higher excitation
energies is rather to describe special states, like giant resonances and
other collective states, which have a peculiar structure. But the detailed
description of the sea of background states around the collective ones
must and should be abandoned. Almost thirty years ago, Dyson made
an eloquent rsum of the situation [41] :
"Recent theoretical analyses have had impressive success in interpret-
ing the detailed structure of the low-lying excited states of complex
nuclei. Still, there must be a point beyond which such analyses of in-
dividual levels cannot usefully go. For example, observations of levels
of heavy nuclei in the neutron-capture region give precise information
concerning a stretch of levels from number N to number (N + n), where
N is an integer of the order of 108. It is improbable that level as-
signments based on shell structure and collective or individual-particle
quantum numbers can ever be pushed as far as the millionth level. It
is therefore reasonable to inquire whether the highly excited states may
be understood from the diametrically opposite point of view, assuming
as working hypothesis that all shell structure is washed out and that no
quantum numbers other than spin and parity remain good. The results
of such an inquiry will be a statistical theory of energy levels. The sta-
tistical theory will not predict the detailed sequence of levels 'in any one
nucleus, but it will describe the general appearance and the degree of
irregularity of the level structure that is expected to occur in any nucleus
which is too complicated to be understood in detail.
"In ordinary statistical mechanics a comparable renunciation of exact
knowledge is made. By assuming all states of a very large ensemble to
be equally probable, one obtains useful information about the over-all
behavior of a complex system when the observation of the states in all its
detail is impossible. This type of statistical mechanics is clearly 'inade-
quate for the discussion of nuclear energy levels. We wish to make state-
ments about the fine detail of the level structure, and such statements
cannot be made in terms of an ensemble of states. What is required is
a new kind of statistical mechanics, in which we renounce exact knowl-
edge not of the state of a system but of the nature of the system itself.
We picture a complex nucleus as a "black box" in which a large number
of particles are interacting according to unknown laws. The problem is
then to define in a mathematically precise way an ensemble of systems
in which all possible laws of interaction are equally probable." (The un-
derlining is ours). This programme, Initiated by Wigner, has been to a
large extent successfully accomplished. And partial justification of the
theory has been only recently found in concepts of chaotic dynamics.

The appropriate language to define an ensemble of systems is provided
by random matrix theory. One considers the Hamiltonian matrix H as
an N x N stochastic matrix (its matrix elements are random variables)
and the question is to specify the probability density Pu(H)AH. In or-
der to proceed one must answer the following questions: i) what are the
admissible Hamiltonians and what is the proper measure in the ensem-
ble of these Hamiltonians? ii) given the ensemble of Hamiltonians, are
the properties in whicn we are interested common to the vast majority
of them? Question i) is answered on the basis of genera/ symmetry prin-
ciples as well as of physical plausibility. Question ii) should be answered
by the affirmative and one can then perform ensemble averages which
are equivalent to averages over a generic matrix of the ensemble (ergodic
property).
4-1. Wigner's Gaussian ensembles

The systems we shall deal with are characterised by their Hamiltonians
which can be represented by Hennitian matrices. When there are some
exact quantum numbers corresponding to exact integrals of motion, like
angular momentum and parity ( Jr), and if the basis states are labelled
by these exact quantum numbers, the Hamiltonian matrix splits into
blocks, and the matrix elements connecting different blocks vanish. We
shall assume that such a basis has already been chosen and restrict our
attention to one of the diagonal blocks, an jV x TV Hennitian matrix in
which N is a large integer.
The underlying space-time symmetries obeyed by the system put im-
portant restrictions on the admissible matrix ensembles (see Table l). If
the Hamiltonian is time-reversal invariant and invariant under rotations,
the Hamiltonian matrices can be chosen real symmetric. If the Hamil-
tonian is not time-reversal invariant then, irrespective of its behaviour
under rotations, the Hamiltonian matrices are complex Hermitian. Fi-
nally, if the system is time-reversal invariant but not invariant under
rotations, and if it has half-odd-integer total angular momentum, the
matrices are "quaternion real*. The latter may deserve some explana-
tion.
The structure of a "quaternion real" Hermitian 2N x 2N matrix is a
sum of direct-product matrices
H = H0C0 + H1Ci + #2e2 + #363 (4.2)

26 Orial Bohigas
Tims Angular Rotation Hanultonian Canonical Number of

Reversal Momentum Invariance Matrix Group P Independent
Invariance Real Param.
Integer
Yes Real Symmetric Orthogonal i N(N + l)/2
Half-Odd-Integer Yes J
Half-Odd-Integer No Quaternion Real Symplectic 4 N(ZN-I).
Complex
No --
Hermitian
Unitary 2 N2
Table 1. Invariance properties of the Hamiltonian associated with the isotropy

and homogeneity of space- time and corresponding structure of the Hamiltoniac.
The sign -- means that the corresponding symmetry is disregarded.
where are the 2 x 2 matrices
( I 0\ 0 -1 \ O
* = o i)' 1 oj ' -i
(4.3)
H0 is a real symmetric NxN matrix and HI, H?, Hy are real antisymmet-
ric NxN matrices. Notice that e; (t = 1,2,3) are essentially the Pauli
matrices OV- In the same way that the Hermitian property is preserved
by a unitary transformation or the property of reality and symmetry
by a real orthogonal transformation, the property of being "quaternion
real" is preserved by the 2N x 2N unitary matrices B which satisfy
Z = BZB* , (4.4)
where Z is the antisymmetric unitary matrix

O -1 O O \
1 O O O
O O O -1 O O ...
Z= O O 1 Q O O ... (4.5)
O O O -1 ...
O O 1 O ...
and BT is the transpose of B. The matrices B form the jV-dimensional

symplectic group Sp(#) (see Table 1). The structure of the matrix
(4.2) ensures that all energy levels are doubly degenerate (Kramers'
degeneracy).
In Table 1, JV is the dimensionality of the matrix of sites and {3 counts

the number of real parameters associated with each matrix site. For
0 = 4, the matrix is 2JV X 2JV and there are JV different eigenvalues.
The number of independent real parameters is given by
JV + 0N(N-I)/2 . (4.6)
The structure of the Hamiltonian matrix in the three cases (0 = 1,2,4)
can be clearly seen by writing it down explicitly for JV = 2 :
' o O x y
Q a -y' x'
P - 4 ' x' -y c Q
\y' I 0 c
(4.7)
in (4.7), a, b, c are real numbers and i,y complex numbers.
The classification of matrix ensembles (/3 1, 2 and 4) displayed in
Table 1 is, in a precise way, complete. Indeed, Dyson has shown that
the most general kind of matrix ensemble, defined with a symmetry
group which may be completely arbitrary, reduces to a direct product of
independent irreducible ensembles each of which belongs to one of the
three known types (Dyson's threefold way [42]). The root of that is to
be found in Frobenius' theorem: over the real number field there exist
precisely three associative division algebras, namely the real numbers,
the complex numbers and the real quaternions.
Due to its importance in physical applications and also to be more
specific, let us now concentrate on the case /3 = 1. In order to introduce
a proper measure iff 'in the space of matrices one can proceed as follows.
A metric is defined in the matrix space to which H belongs by expressing
da2 = tr 6HSH+ , (4.8)
in terms of the independent variables X11 of H, as
where M is the number of independent variables (matrix elements). In

(4.8), H+ is the Hennitian conjugate of H. The basic measure dJ is
then the measure induced by this metric, namely
M
dH = (det Q11
28 Oriol Bohigaa
If one considers real symmetric matrices H ( H' = If*) there are

N(N + l)/2 independent variables and one has
ds*= (6Hrf + 2 (6F0)2 (4.11)
so that
dH = 2"'"-1'/4 H dfl* n dHu . (4.12)
l<t<W l<i<i<N
Any automorphism of the studied matrix space which leaves the metric
(4.8) invariant will leave the associated measure dH invariant. For in-
stance, in the case of real symmetric matrices, the invariance of (4.8),
(4.11) under a real orthogonal transformation implies the invariance of
(4.12).
The Gaussian Orthogonal Ensemble (GOE) in the space of N x JV real
symmetric matrices is defined by requiring that
1. The ensemble is invariant under every orthogonal transformation
H' = W* HW , (4.13)
where W is any real orthogonal matrix, i.e., the probability P(H)dH

that a matrix H will be in the volume element dH (4.12) is invariant
under orthogonal transformations,
Pn(HâH1 = PN(H)dH . (4.14)
From (4.14) and from the invariance of the measure dH, one has that
Pfi (H) must also be invariant under orthogonal transformations.
2. The various matrix elements ,H1, (t < j) are independent random
variables.
The first requirement is made to avoid that any given state should
play a particular role: all basis states, and therefore all states, should
behave in the same way. The second requirement is put for the sake of
simplicity and leads to a mathematically solvable problem.
These two requirements determine uniquely the ensemble. The func-
tion PU (H) , which is also invariant under the corresponding automor-
phism, can be written as
= KN exp{- tr(ff z )/4t; 2 } , (4.15)
where KN is a normalisation constant,
tr H* = Hl + 2 Hi , (4.16)
and u defines the scale. For the GOE one therefore has that the matrix
elements are statistically independent and that each matrix element JT1-,-
is zero-centered Gaussian distributed, with zero mean
Wj = O i < j (4.17)
and the same variance v 2 , except for the diagonal elements which have
twice that variance
5?= (! + *>* i < j . (4.18)
We use the notation W for the ensemble average of W'.

What we want now is to derive from (4.15) the joint probability dis-
tribution for the spectrum. Let O be the orthogonal matrix which diag-
onalises H,
O1HO = E (4.19)
where E is diagonal, and let p,- ( = !,..., N(N- 1)/2) be the parameters
which specify the set of eigenvectors. When going from the matrix ele-
ments to the eigenvalue-eigenvector representation, the volume element
dH transforms like
= JN dEl
(4.20)
where JN is the Jacobian
j _ 9(Hu, Sn, -, HNN, Ha, , HN-I-N) / . ,%
d(EitEj,- - , EN, Pi, PJ,- --,Pw(W-I)/!)
Because
fli,-=*j-*. (4-22)
t
the first N columns of the determinant of the Jacobian are independent
of the eigenvalues whereas the remaining ones are linear. Consequently,
in the expansion of the determinant, each term contains N(N l)/2
factors, each of which is linear in the eigenvalues. This means that
the Jacobian is a polynomial of degree N(N l)/2 in the eigenvalues.
Furthermore this polynomial vanishes whenever two eigenvalues, say Ei
and EJ, are equal, as can be easily seen. Indeed, assume the opposite,
namely that JN ^ Q when J?, = Ej. Then, there would exist a one-to-one
correspondence between the matrix elements JT,-,- and the eigenvalues
and eigenvectors, i.e., the latter would be completely determined by
30 Oriol Bohigas
the matrix elements. But we know that when 2?, = E,, the matrix
elements define only an eigenplane and not the individual eigenvectors,
and therefore JN = O. We then conclude that (4.21) is a polynomial
of degree N(N - l)/2 in the eigenvalues, which vanishes whenever two
eigenvalues are equal, leading to
-Jw = II lî - Ei\ MPi. --PN(If-I)It) (4.23)
Integrating over the variables p,-, one anally obtains, from (4.15) and
(4.23), the joint probability distribution PJV(EI,. . . , E N ) for the spec-
trum,
(4.24) is the equivalent of (3.9) and contains all the information con-
cerning the distribution of eigenvalues and the correlations among them.
The last factor in (4.24), which has its origin in the Jacobian (4.21), is
particularly important and exhibits the phenomenon of level repulsion,
namely that small spacings are ruled out.
Once Pff (El, ..., Eff) is specified, one can follow the scheme outlined
in the previous Section. The first step is, by integrating (4.24) over
all the variables except one, to determine the ensemble averaged level
density p(E). In fact, there are many ways to determine p ( E ) , which we
shall not describe. The result, in the large-JV limit, is
f (21)-1 - (4JW - E*Y for \E\ <
P{ l
' (Q for \E\ > 2V^JV , '
where p(E) is normalised to JV. Notice that the level density at E O
increases like JV1/2, the mean level spacing D(E) = 1/p decreases like
JV~1/2 and the spectrum span increases like N1/*, implying that the ratio
of the mean level spacing at the center of the semicircle (4.25) to the
spectrum span decreases like 1/W (see Fig. 4a).
By measuring energies E in appropriate units, namely
x = -A= , (4.26)
Vw*77
and normalising p to unity, one has
>"> .NO (,27)

PIE
P( E)
Fig. 4. (a): the level density pfi(E], approximated by the semicircular level
density (4.25), of eigenvalues of random matrices for N = 10 (dashed line) and
50 (full line), (b): the density of real roots of a random polynomial (4.38) for
N = 10 (dashed line) and 50 (full line).
32 Oriol Bohigaa
which is known in the literature as Wigncr's semicircle law, an improper

name, because the two axes do not even have the same dimension. A
more proper denomination would be semiellipse law. The ingredient
leading to (4.27) can be traced back to the first factor on the r.h.s. of
(4.24). Not surprisingly, the one-dimensional harmonic oscillator wave
functions <PJ(X) may play a role and, in fact, in the limit of large N, it
is the expression
ZX(*) (4-28)
i=i
which produces the semicircle law in one of its derivations.
Similarly to the GOE, one defines the Gaussian Unitary Ensemble
(GUE) of Hcrmitian matrices or the Gaussian Symplectic Ensemble
(GSE) of quaternion real matrices by requiring: i) the ensemble to
be invariant under unitary or unitary symplectic transformations,
ii) the various linearly independent components of H to be statistically
independent.
Equations (4.15), (4.24), (4.25), (4.26) can be extended and take the
general form
/4tr)} (4.29)
(4.30)
_ f \E\ <
~\ O \E\>
x = E/Jpv'N (4.32)
and the semicircle law (4.27) is still valid. In (4.29), (4.30), (4.31), (4.32),
P =1, 2, 4 for GOE, GUE and GSE respectively. From (4.29) one sees
that the matrix elements of the components of H are:
0 = 1, H=S (4.33)
/3=2, H = S + iA (4.34)
/3 = 4 (see (4.2)) , H = 50eo + A1C1 + A2e: -t- A3C3 . (4.35)
In (4.33), (4.34), (4.35), S (A) denotes a real symmetric (antisymmetric)
matrix. The independent matrix elements of the symmetric matrices are
zero-centered normally distributed like (4.17), (4.18), and those of the
antisymmetric matrices are zero-centered normally distributed with the
same variance v2 except for the diagonal matrix elements which are zero.
Some remarks concerning Wigner's semicircle law are in ordsr. Firstly,

it holds under more general conditions than the ones fulfilled by the
Gaussian ensembles. For real symmetric matrices, for instance, one only
needs the statistical independence of the matrix elements plus (4.17),
(4.18) for almost all the matrix elements. In particular, the semicircle
law applies for an ensemble of real symmetric matrices with matrix el-
ements +1 or 1 taken at random. Secondly, one should notice that it
has no resemblance with the level density (4.1) or with other experimen-
tal densities. As we shall see, this should not be considered as a serious
drawback of the theory.
One may ask if it is possible, instead of following the route which
starts from PN(H) and leads to Pu(Ei,...,EN) and PN(E), to follow
the one which starting from PN(H) leads to the study of the character-
istic polynomial, which has random coefficients, and then try to proceed
further. Which kind of things are known on random polynomials and
what can we expect? In fact, one only knows answers to relatively simple
questions [43].
Following Kac [44], consider, for instance, a random polynomial of
degree N with real coefficients
1
+ Oi JS"' + + + (4.36)
and let the coefficients at (t 0,...,N) be components of the (N + I)-
dimensional vector a of unit length,
N
IaI 2 = a* = 1 , (4.37)
which has equal probabilities for directions within equal solid angles.
This is equivalent to assume that the coefficients a; are statistically
independent and that each is normally distributed with zero mean and
variance 1. One of the simplest problems is to determine the density
of real roots of (4.36). The answer is
/1
(4.38)
E = l
where
1 1
l JT\ f AT I -t\E*Nfi
^ wl-fi) = (Jv + L)EI [L litE ZX/!
)(i hE 2JV+2\) 1 . I jt Of\\
(4.Jy)
The total number of real roots, given by the integral of PN(E), tends
asymptotically (large W) to
-hijV (4.40)
TT
34 Oriol Bohigas
which means that there are few real roots. In fact, the fraction of real
roots tends to zero. The function ptf(E), displayed en Fig. 4b, shows
that the real roots tend to concentrate very strongly on +1 and 1.
More precisely, from (4.38) one sees that
(4.41)
That the real roots concentrate on 1 can be guessed directly from

(4.36) because on the average the a/'s are roughly of the same order of
magnitude and, consequently, for (4.36) to be zero, E must be close to
+1 or 1. Otherwise, cancellation becomes difficult.
The density of real zeros PN(E), see (4.38), is very different from the
semicircle p(E), see (4.25) or (4.27) and Fig. 4. Of course, there is no
surprise in that, because, for the characteristic polynomial
det \H - EI\ (4.42)
corresponding to a random matrix H, the coefficients are not distributed
independently and normally. One rather has, hi terms of the matrix
elements of H (see, for instance, [45])

a-o - (-1)
( i\J"
(4.43)
'3
HJi
ji
= det H .
In other words, choose the x principal sub-matrices of H in all possible
ways. The sum of their determinants, to within a sign, is at. (Definition:
a principal sub-matrix of an N x N matrix H is obtained from H by
suppressing symmetrically (N t) rows and columns. There are
(N\ Nl
(4.44)
\t ) ~
principal x sub-matrices of H).
To proceed further, one should work out, assuming that the matrix
elements of H are distributed independently according to (4.17), (4.18),
the joint probability density of the a<'s, etc. This line of inquiry, as far
as we know, has not been explored.
In connection with random polynomials, let us finally mention that
there is, as for the semicircle law, a certain insensitivity of the results to
the precise assumptions concerning the distribution of coefficients. For
instance, (4.40) holds also if the coefficients at are taken at random 1
with equal weights.
4.2. Generalisations
The frequent appearance of the semicircle law or related densities as
well as of joint probability distributions of eigenvalues with a structure
similar to (4.30) can be further illustrated in a more general context.
Following Ginibre [46] [4], consider now an ensemble of complex (in
general not Hermitian) NxN matrices J. The eigenvalues will now be
complex. Define the ensemble by
PN(J)dJ (4.45)
where
(4.46)
j,k
and Pff(J) is the Gaussian invariant measure
PN(J) = exp{-tr JJ + ) . (4.47)
In (4.46) the superscripts denote the real and imaginary part. The en-
semble (4.45) is invariant under all unitary transformations and (4.47)
implies that the real and imaginary parts of the matrix elements of J are
all independent, zero-centered normally distributed with the same vari-
ance. The joint probability distribution PN(ZI, , 2jv) for the spectrum
is
,-2,| 2 (4-48)
where z,- are the (generally complex) eigenvalues. The density of eigen-
values PN(Z) is, in the large- N limit, constant in the complex plane
within a circle of radius oc JV1/2 centered at the origin, and zero outside
the circle.
36 Oriol Bohigaa
Recently, partly motivated by studies of the dynamics of neural net-

works, other ensembles of N x JV asymmetric random matrices, for which
correlations among matrix elements are allowed, have been studied [47].
An ensemble of real matrices has been defined by
T
PN(J) oc exp - - t r ( J 7 - rJJ) J . (4.49)
The measure (4.49) implies
J,,=0, Jg = l + r , J = 1, JaJa = T . (4.50)
It is the last relation (4.50) which shows, through the parameter 1 <
T < 1, that correlations among the matrix elements are allowed. One
recovers for T = 1 the case of symmetric matrices (GOE) and for T = 1
the case of antisymmetric matrices. In the N * oo limit, the eigenvalue
density PK(Z) in the complex plane is, similarly to the case (4.47), con-
stant within an ellipse with semiaxes a (real direction) and b (imaginary
direction) given, by choosing appropriate units proportional to JV1/2, by
In the case r = O (fully asymmetric matrices without correlations among

matrix elements), the ellipse degenerates into a circle. The projection of
Pff (z) on the real axis leads to a generalised semicircle law
pz,N(x) = (Tra2)-1 - 2 (a2 - x2)1/2 \x\ < a , (4.52)
with a similar law along the imaginary axis. In the limit of symmetric
matrices (r 1), the eigenvalues concentrate on a (real) line and
PN(Z) = S(y)ptiN(x) , (4.53)
where Ac1W(2O reduces to Wigner's semicircular law (4.27), as it should.

Similarly to (4.49), ensembles of complex asymmetric matrices with
the invariant Gaussian measure
PN(J) oc exp ! - * tr(JJ + - 2 Re T JJ) j , (4.54)
have been discussed [47]. The average density PN(Z) depends only on
N =2 N-xx Ratio TV -> oo/ AT = 2
TT2
0 =1 IT
2 S
T5 1.047...
/3 = 2 5*
TT 2
1.015...
ol8
J
^
16T4
pa A1
4
-4 0.995...
36TT3" 135 "
Table 2. Small-s behaviour of the spacing distribution p(s) for the Gaussian
Orthogonal (/3 = 1), Unitary (0 = 2) and Symplectic (/3 = 4) Ensembles.
Fu-st column: 2 x 2 matrices. Second column: asymptotic. Third column:
ratio of the first non-vanishing derivative at the origin of the asymptotic over
the 2 x 2 case.
Asymptotically, it is constant within an ellipse in the complex plane

centered at zero with semiaxes a = 1+ \T\ in the direction d and b = I\T\
in the direction T? + jr/2, and is zero outside the ellipse. For \T\ = 0, no
correlations, one recovers Ginibre's result for PN(Z). Notice that, for the
ensembles (4.49} and (4.54), the probability distribution of the spectrum,
like (4.24) or (4.30), is in general not known.
4.3. Fluctuation measures

We shall now enumerate and comment on the spectral fluctuation re-
sults, as described in Section 3, for the Gaussian ensembles (4.29). They
have been derived from (4.30) by integrating over all except 1, 2,... or n
variables. This is a difficult task, in fact a tour de force, and, in Wigner's
words, the derivations contain "beautiful, though old-fashioned, mathe-
matics" [48]. The compactness of the results is remarkable but we will
not even try to give a flavour or to sketch the derivations. We rather
recommend the interested reader to consult Refs. [2,4].
Oriol Bohigas
0.25
o.oo
Fig. 5. (a): Spacing distribution p(s) for the 2 x 2 Gaussian ensembles, see
(A.1), (A.2), (A.3). (b): Rz(L), see (3.27), for the Gaussian ensembles (asymp-
totic) as well as for a Poissonian spectrum.
A few words concerning the spacing distribution. For 2 x 2 matrices

the result can be easily worked out and, for completeness, it is repro-
duced in Appendix A and displayed on Fig. 5a. The asymptotic result,
for large JV, is known, difficult to derive, and cannot be written in a
simple form. It is remarkable that the N 2 result is very close to
the asymptotic one (see Table 2). The distribution (A.l) ( Wigner's sur-
mise) is a standard one in statistics, known as Rayleigh's law. It results
from considering two random variables (X, Y) statistically independent
and normally distributed with probability density
p(x,y) = (22)-1 exP{-(*2 + j/2)/2a2} (4.56)
and introducing new random variables (R, (f>) through X = R cos (j>, Y =
R sin <. The probability density of the new variables is
p(r,p) = (r/22) exp(-r2/2a2) , (4.57)
which is of the form (A.l). As previously discussed, to magnify the
short range behaviour, one can, instead of p(-s), use Ry(L) (3.27), which
is pure two-point (see (C.14), (C.15), (C.16) and Fig. 5b).
The n-level cluster functions Yn(xlt...,xn) are known for the three
ensembles, thefinite-TVas well as the asymptotic expressions [49,4]. The
latter are reproduced in Appendix B. For the two-level cluster function
Y2 (3.7), (3.22), the two-level form factor 6(Jk) (3.21), the number vari-
ance E2 (3.25) and the Dyson-Mehta statistic A (3.30), see Appendix
C and Figs. 6 and 7.
If one compares YZ to 1 g(r) of a liquid or a solid (see (3.14) and
Figs. 2 and 6a), one sees that it makes sense to qualify the structure
of the eigenvalues of random matrices as quasicrystalline. The two-
level cluster function tends slowly to zero, with an (inverse) power law
(see (C.4), (C.5), (C.8)). This implies spectral rigidity, which manifests
itself in a small (logarithmic) fluctuation of n(L) (see (C.20), (C.21),
(C.22) and Fig. 7a) or a slow (logarithmic) increase of A (see (C.23),
(C.24), (C.25) and Fig. 7b), as compared to a linear behaviour for a
Poissonian spectrum (see (3.32), (3.33)). More generally, the cluster
functions satisfy [50,4].
r Yn(X1, ...,Xn) dxn = (n - 1) Yn^1(X1, ...,Xn) (4.58)

J-OO
and, in particular
+
"y,(i)dx=l, (4.59)
Ort'o/ Bohigas
-0-25 r
b(k) r
1.0
-05
-1 O
0.5
Pig. 6. Orthogonal (/? = 1), Unitary (/3 = 2) and Symplectic (ft 4)

ensemble results for: (a) the two-level cluster function, see (3.7), (3.22), (C.2),
(C.5), (C.6) ; (b) the two-level form factor, see (3.21), (C.9), (C.12), (C.13).
0-30 !<
0.25
0-20
000
Fig. 7. Orthogonal (/3 = 1), Unitary (ft 2) and Symplectic (/? = 4) Ensem-
ble results for: (a) the number variance E2, see (3.25), (G.17), (C.18), (G.19) ;
(b) the Dyson-Mehta statistic A, see (3.30).
42 Oriol Bohigas
which express that the spectrum is "incompressible". (Equivalently, in

terms of the two-level form factor (3.21), (4.59) reads b(k = O) = 1).
This means that the velocity of propagation of "sound waves" becomes
infinite as the frequency tends to zero. Relations (4.58) don't follow from
the definitions (3.3), (3.5), they are obviously not satisfied for an ideal
gas (see (3.34)), or a Poissonian spectrum (see (3.32)). They constitute
an example of situation for which the subtle distinction between N and
N-I becomes essential.
Let us also notice that the two-level form factor, at k = 1, is con-
tinuous with the first two derivatives also continuous for /3 = 1, has
discontinuous first derivative for /0 = 2 and has a singularity for ft = 4
(see (C.9), (C.12), (C.13) and Fig. 6b). This is connected with the fact
that the oscillating term in Y2(x) is, for large x, of order aT4, a;"2 and
I"1 for /3 = 1, 2 and 4 respectively (see (C.4), (C.5), (C.8) and Fig. 6a).
One may observe that the results for several fluctuation measures ap-
pear as ordered when going from / = l to/3 = 2 to/3 = 4 (the level
repulsion, as well as the spectral rigidity, for instance, are increasing).
This looks strange if one remembers (see Table l) that / 3 = 2 corresponds
to the case with the least symmetry. The reason for this anomaly is that
in the conventional discussion of GSE one considers the distinct eigen-
values, forgetting Kramers' degeneracy. Would one take into account
the twofold degeneracy, the two-point functions, for instance, would be
modified
RZ(X) -+R2(x/2) + 6(x) (4.60)
Y2(x) -f Y2(x/2) - 6(x) (4.61)

and the number variance, for instance, would, for large L, be the smallest
for (3 = 2, whereas the / 3 = 1 and /3 = 4 values would differ only by a
small additive constant.
Finally, let us remark that for the ensembles (4.47), (4.48) of complex
matrices with complex eigenvalues, the n-point correlation functions of
eigenvalues have also been determined [46]. In the asymptotic limit, the
two-point cluster function takes the simple form [51,4]
Y2(Z1, Z2) = CXpI-Tr)Z1 - z2\2} . (4.62)
Notice that, in contrast to the canonical ensembles (0 = 1,2,4), the

falloff of the correlation function is not as an (inverse) power-law but it
decays much faster, like a Gaussian.
4-4- The, Coulomb gas

To have some physical insight on the properties of eigenvalues of random
matrices, it will be useful to describe the correspondence between the
properties of the latter and those of a Coulomb gas. For that purpose,
let us recall some properties of electrostatics. The potential created
by a charge distribution /i(x) satisfies Poisson's equation
d Q2U
Au E = E - S = -w,, M(X), (4.63)
where Wj is the surface area of the unit sphere hi Rrf. It follows that the
potential V between two point charges which are |x| apart is
d=1 V oc |z| (4.64)
d=2 V oc ln(|x|/u) (4.65)

d =3 ^ oc l/|x| . (4.66)
(4.66) is the familiar Coulomb interaction in three dimensions; in (4.65),
a fixes the scale. (4.64) can be interpreted in three dimensions as the
interaction energy between two infinite parallel planes with a constant
surface charge density and (4.65) as the interaction energy between two
infinite parallel rectilinear wires with constant linear charge density.
Following Dyson [41,4], consider now in a two-dimensional universe a
gas of JV point charges with positions E\,.. .,EN on the straight line
oc < E < +00 in a confining external harmonic potential. The poten-
tial energy U of such a gas is
ff=l>?-I>|--y|, (4-67)
* i >
where the first term represents the confining potential which attracts
the charges towards the origin and the second one the electrostatic re-
pulsion between each pair of charges. Let this gas be in thermodynamic
equilibrium at a temperature T. The probability density of the positions
of the N charges is given by
Pn(E1, ...,EN) = AN exp{-U/kT} , (4.68)
which coincides with (4.30) if we choose appropriate units and make the
identification of ft in (4.30) with 1/kT in (4.68). We therefore have that
44 Oriol Bohigaa
the distribution of eigenvalues in (4.30) (ft = 1,2,4) is equivalent to

the distribution of positions of point charges in thermodynamic equilib-
rium at a temperature kT = 1//?. The correspondence with statistical
mechanics is complete, and (4.68) has the structure (3.9) with a pair
potential (3.10) given by (4.65) (plus an external harmonic potential).
The Coulomb gas exhibits both short range repulsion and long range
rigidity, showing that these properties have a common origin and that
the expression "repulsion of levels" is not merely an image: energy levels
do indeed behave like charges, repelling each other. When the random
matrix model is interpreted as a Coulomb gas, Y2 (see Fig. 6a) defines the
shape of the neutralizing charge cloud induced by each particle around
itself. The orthogonal, unitary and symplectic ensembles correspond to
a temperature T = I, 1/2 and 1/4 respectively, but for the Coulomb gas
the temperature may vary continuously. The Coulomb gas is a single
phase system. Indeed, the thermodynamic functions can be calculated
for any /3 and are analytic over the whole range O < /J < oo with no
thermodynamic transition at any finite temperature. Because the sys-
tem possesses long-range order, it might be more appropriate to call
it a "crystal" rather than a gas. At low temperature the charges are
regularly spaced in a crystalline lattice arrangement and at high tem-
peratures the system behaves like a gas of independent particles, each
particle carrying with it a neutralizing "charge cloud". In contrast to
thermodynamic functions, the -level cluster functions are known only
for /? = 1, 2 and 4.
An electrostatic analogue exists not only, as we have seen, for the
spectra of the canonical ensembles (4.30), but also for the distribution
(4.48) of complex eigenvalues of a complex random matrix. Following
Jancovici [51], consider a one-component plasma confined in a disk of
radius R in a two-dimensional universe. The plasma is a system of AT
identical particles of the same charge embedded in a uniform neutralizing
background of density /i, /zjrfi2 = JV. The logarithmic potential (4.65)
created by the background is
V(x) = -7TM [R2 InR- (R2 - |x|2)/2J |x| < R . (4.69)
Apart from an uninteresting additive constant, the background creates

a harmonic potential inside the disk, and the potential energy of the
system is
- In(Ix,-- Xj |/tf) . (4.70)
Kj
By choosing appropriate units, namely
x, - N1I2XiJR , (4.71)
(4.70) takes the form
-ln|z,--*|, (4.72)
where the complex number representation Zj = |0j|exp(itf,-) of X; has

been used. Let the system be in thennodynamic equilibrium. The
probability density of the positions of the N charges PN(ZI,..., ZN)
will be given by (4.48) for the particular (inverse) temperature / = 2
(p = 1/JfcT). For that particular temperature the two-point cluster func-
tion is known and given by (4.62).
For other values of the temperature one does not know how to de-
termine two-point and higher-order correlation functions. That ft = 2
plays a particular role in deriving results for correlations can be easily
seen. One encounters integrations over the angles of terms of the type
\Zi Zjf. For any /? there is no simple way to perform them. But for
/3 = 2, using the fact that the functions {exp(tmd)} form a complete set
of orthonormal functions, the angular integrations can be performed.
At the temperature T = 1/2 the correlation function g(r) (see (3.12)
and Fig. 2) is typical of a gas, showing neither long-range nor quasi-long-
range order. However, it is somewhat surprising that the correlations
have a Gaussian falloff rather than the exponential one which is found in
the high-temperature approximation [51]. Let us finally notice that when
discussing the Coulomb gas for charges on a line (Coulor>.b gas version of
the canonical ensembles of eigenvalues of Hennitian random matrices),
one has introduced rather artificially an external harmonic confining
potential to prevent charges from going to infinity. In contrast, for
charges on the disk (complex eigenvalues of complex random matrices),
the confining potential is quite naturally provided by the neutralizing
background of the one-component plasma.
4.5. Dyson's circular ensembles

Before ending this Section, we shall briefly describe two methods, lead-
ing to Dyson's circular ensembles and to the Weyl-Balian ensembles,
to derive random matrix ensembles. They allow to suppress some ar-
bitrariness in the definition of the Gaussian ensembles as well as to get
46 Oriol Bohigas
rid of the undesirable, from a physical point of view, semicircular level

density.
The original motivation to introduce the circular ensembles of unitary
matrices is to avoid the following unsatisfactory feature of the Gaussian
ensembles: one cannot define a uniform probability distribution on an
infinite range, therefore it is impossible to define an ensemble hi terms of
matrix elements H^ of H in which all interactions are equally probable.
To turn around this difficulty, by a small formal change, Dyson chooses
to specify the system not by the NxN matrix of the Hamiltonian but
by an N x N unitary matrix S with eigenvalues exp(ttfj) (j = 1,..., N)
distributed around the unit circle [41]. In fact, the precise connection
between S and H need not be specified but one can imagine a definite
relation between them, for instance
S = exp{t#t} . (4.73)
One only assumes that a1 is a function of H so that the angles i?y are
functions of Ej and, over a small range of angles, the relation between
i?;- and Ej will be approximately linear. The basic statistical hypothesis
is that the fluctuation properties of n consecutive levels of H are the
same as the ones of n consecutive eigenangles of S. The interest of this
approach is that it leads to ensembles of unitary matrices S for which
the probability is uniform. Remember that (4.73) is only one possible
device defining a relation between S and H and that one can imagine
others.
The general symmetry restrictions on H (see Table 1) can be trans-
lated in terms of unitary matrices and one is led to uniquely define,
corresponding to /? = 1, 2 and 4, three circular ensembles of unitary
matrices.
i) /3 = 1 ; the Circular Orthogonal Ensemble (COE) is the ensem-

ble of symmetric unitary matrices, which is invariant under every
transformation 5 W^SW, where W is any unitary matrix.
u) (3 = 2 ; the Circular Unitary Ensemble (CUE) is the ensemble
of unitary matrices which is invariant under every transformation
S -* USW, where U and W are any two unitary matrices.
iii) /3 = 4 ; the Circular Symplectic Ensemble (CSE) is the ensemble
of self-dual unitary quaternion matrices which is invariant under
every transformation 5 -* WRSW, where W is any unitary matrix
and WR is the dual of W.
The latter deserves some explanation. Given a quaternion matrix Q,

whose elements are quaternions (<7o> Ci 92,^3)1 or
(4.74)
with et given by (4.3), the dual (or time reversed) of Q, denoted by QR,
has quaternion matrix elements
(QR)0 - fr - (4-75)
In (4.75) the conjugate quaternion q of q is (90, 9i, 92 9s) r
s
X ~" /j frs*\
9 = 9oeo ~ / s 9/G< i v.4. (D;
and 9,-y is the (quaternion) matrix element of Q. Quaternion matrices

satisfying QR = Q are called self-dual, and the ones satisfying QR = Q
are called anti-self-dual. The unitary matrices 5 representing physical
systems which are tune reversal invariant are not quaternion real but
self-dual (5R = 5, S+ = 5~l), whereas the quaternion matrices H
which represent a time-reversal invariant Hamiltonian are Hermitian as
well as self-dual (HR = H, H+ H), and therefore quaternion real (see
(4-2)).
The joint probability density of eigenvalues exp(ti?7-) for the circular
ensembles is
a (4.77)
to be compared to (4.30). The main difference between (4.77) and (4.30)

lies in that the first factor in (4.30), insuring that the eigenvalues don't
go to 00, is neither needed nor present for the case of eigenvalues on
the unit circle. Of course, one can also work out a Coulomb gas problem
equivalent to (4.77), this time with the absence of a confining external
potential.
Integrating (4.77) over all except 1, 2, ... or n variables will provide the
correlation functions. There is of course no semicircle law for the average
density of eigenvalues for the circular ensembles. From their construc-
tion it should be clear that the average density is uniform. The n-point
correlation functions for the three different ensembles and finite N are
known [4]. In the limit of large N it has been shown that the cluster
48 Orioi Bohgas
functions Vn(^1, . , &n) of the circular ensembles and Yn(xi, . . . , Xn) of

the corresponding Gaussian ensembles coincide [49]. Therefore, the pre-
vious discussion of spectral fluctuations for Gaussian ensembles applies
equally well to circular ensembles. In what follows, no more mention of
the circular ensembles will be made. But it should be noticed that they
play a prominent role in the discussion of periodically driven systems
(see lectures by Chirikov [13]) and of chaotic scattering (see lectures by
Smilansky [17]).
4.6. The Weyl-Balian ensembles

Balian has devised an alternative and elegant way to derive the Gaussian
(and circular) ensembles. More importantly, the method used is well
adapted for generalisations, and for the incorporation of suitable physical
constraints in the definition of the matrix ensembles. Let us briefly
sketch Balian 's approach [52].
Borrowed from information theory, let us define the amount of infor-
mation J associated to the probability P(H)
J{P(H)} = f P(H) In P(H)dH . (4.78)
One looks for the function P(H) that minimises J, which is equivalent to
assuming the least possible knowledge about P (H) . In order to confiae
the eigenvalues to a finite range, one imposes a condition on the square
of its norm tr H2. One has then that P (H) should minimise J subject
to the constraints
I Ci (normalisation), (4.79)
j P(H) tr H2 UH = C2 , (4.80)
which leads to
P(H) = 6Xp(A 1 + A2 tr ff2} . (4.81)
By Inserting (4.81) into (4.79), (4.80), one determines the Lagrange mul-
tipliers A1, A2. (4.81) coincides with the GOE joint probability distribu-
tion (4.29). The only information used here has been the invariant mea-
sure (4.12) and the constraints (4.79), (4.80). The statistical indepen-
dence of the matrix elements, instead of being assumed as in Wigner's
original derivation, follows here as a result. Considering the symmetry
properties of the matrix H, one can similarly derive the GUE, GSE as
well as the circular ensembles.
Dalian's approach allows to incorporate other constraints than (4.79),

(4.80). Consider, for instance, that some average properties fa of H are
known. Then
ft = J A(H)P(B)AH = Ci (4.82)
are the constraints, and the probability density is
P(H)= exp C^XJi(H)). (4.83)
i
The Lagrange multipliers A1- are obtained by inserting (4.83) back to

(4.82). Particularly interesting is the case when the average eigenvalue
density p(E) is used as a continuum of constraints
p(E) = tr 6(E - H) . (4.84)
The resulting probability density of the Weyl-Balian ensemble, using
Ozorio's terminology [18], is
P(H) = exp U X(E) tr 6(E - H) dE\ = exp (tr X(H)) (4.85)
or
P(H) = det fi(H) , (4.86)
where
t,(E)=exp{X(E)). (4.87)
By integrating over the parameters specifying the eigenvectors, as for
(4.20), (4.23), one obtains the joint probability for the eigenvalues
P(E1, .... ?N) = exp { A(4)} H \Et - Etf . (4.88)
(4.88) contains the Jacobian factor responsible for the fluctuation prop-
erties discussed above. It differs from (4.30) by the argument of the
exponential, which contains the information concerning p(E). In princi-
ple, the function X(E) can be determined by integrating (4.88) over all
variables except one and equating the result to p(E). A simpler approx-
imate method can be used [52,18] leading to an explicit expression for
X(E),
X(E) ~ -| I dE'p(E') In \E - E'\ . (4.89)
The similarity between the Weyl-Balian ensembles and the Gaussian
ensembles permits, in a small energy range A around E (but large
50 Oriol Bohigas
enough to contain many levels), to fit locally the latter to the former
[18]. Indeed, by choosing
E = E- 2A/A 1 , a2 - -A/(A') 2 (A' = dA/d) , (4.90)
one has
X(E) = -(4CT2)-1 (E - E)2 (4.91)
up to first order in the Taylor expansion about any point inside A.E.
That the orthogonal (or unitary, or symplectic) invariance can be kept
for the Weyl-Balian ensembles can be seen by noticing that, by a proper
choice of a function
f(tr H,tr H*,tr H3,...) (4.92)
multiplying the probability density P(H), any desired density distribu-

tion p(E) of eigenvalues can be obtained. For the Gaussian ensembles,
P(H) contains only tr H, fixing the origin of energies (= O), and tr H2.
The fact that traces of higher powers of H are absent ensures the statis-
tical independence of the matrix elements of H. Stated otherwise, the
use of (4.78) with constraints other than (4.79), (4.80), as for instance
(4.84), induces correlations among the matrix elements.
Could one, instead of (4.78), consider a principle of minimum infor-
mation expressed directly in terms of the probability density of the spec-
trum P(Ei, . . .,EK)1 One should then define J by
J[P(E1,..., Ea)} = p(E1,..., E11)IuP(E1 ..... EN) AE1

(4.93)
Minimising (4.93) with the p (.^-constraint would lead to
P(E1, ...,) oc P(E1)P(E2) - - - p(EN) , (4.94)
which differs from (4.88) and corresponds to an uncorrelated spectrum.

Therefore, the amount of information carried by (4.94) is lower than the
one obtained by inserting (4.88) in (4.93). The additional information
carried by (4.88) (correlations between the Ei) is that {Ef} is not an
arbitrary sequence of random numbers, restricted to fit p(E), but rather
it is a sequence of eigenvalues of a random symmetric (or Hermitian, or
real quaternion) matrix.
5. Theory vs. experiment, dynamical basis of RMT

spectral fluctuations, universalities
5.1. RMT predictions and nuclear resonance data
After having discussed how to characterise spectral fluctuations and the
RMT predictions, let us first summarise the present status of the com-
parison of the theory with the kind of experimental data for which RMT
was originally devised, namely the resonances of the compound nucleus
(see Fig. 3). One purpose of this is al to illustrate that consistency of
theory and experiment is, in the present context, a quantitative matter,
not just a question of qualitative behaviour.
By combining all the high quality data, containing approximately 1700
levels, into a nuclear data snsemble (NDE), a sharp comparison between
RMT predictions and experiment has been performed [36], [52-54]. As
already mentioned towards the beginning of Section 4, we stress again
that for a random matrix ensemble to be useful, the ergodic property of
the ensemble should hold. Specifically, let W(x) be a random function
and denote, as before, by <W(x) > and W(x) the spectral and ensemble
averages respectively. W(x) is said to be ergodic if
< W ( x ) > - > W(x) , (5.1)
Variance (<W(x)>) = <W(x)>* - <W(x)> -> O (5.2)

are fulfilled. We then have
<W(x}> -^W(x) . (5.3)
The limit in (5.1), (5.2), (5.3) means the large-TV limit. Pandey has
shown [38] not only the ergodic properties of the Gaussian and circular
ensembles but, furthermore, that the n-level cluster functions Yn are sta-
tionary, namely they are translation invariant over the entire spectrum.
Would ergodicity or translation invariance be lacking, it would make no
sense, as we have done, to collect data corresponding to different nuclei
at different excitation energies to form a NDE and then to compare with
GOE predictions.
The results of the analysis of the NDE for the spacing distribution
p ( s ) , the Dyson-Mehta statistic A, the variance E2, the skewness 71 and
excess ^2 of the number statistic, and the two-level form factor b, are
displayed on Figs. 8, 9 and 10 (as we know, there are some redundancies).
52 Oriol Bohigaa
Fig. 8. Left: nearest neighbour spacing histogram for the nuclear data ensemble
(NDE) (taken from [53]). Right: Dyson-Mehtastatistic A for NDE (taken from
[54]). GOE and Poisson predictions are plotted for the sake of comparison.
A very direct test of level repulsion (small distances) and spectral

rigidity (large distances), as encompassed by (C. 14), (C.20), can be
performed by making a two-parameter fit to Rz (3.27) and a three-
parameter fit to S2 (3.25) [36]
R2(L) = aL* + /3L3 (small L) (5.4)
(L > 1) . (5.5)
(5.4) and (5.5) are consistent with an independent superposition of a
Poissonian and a GOE spectrum (see Appendix D)1 and in (5.5), a is
given by
+OO
y2(r) dr . (5.6)
/ 00
For long-range order one must have a = O (see (4.59)). The results are
reproduced in Table 3.
To see pure three- and four-body effects, the values of R^ and R^ (see
(3.23)) have also been extracted from the NDE (see Table 4). In Tables
3 and 4, error bars attached to the theoretical value ( one sample error)
are also shown.
The results of these analyses are quite conclusive and go a considerable
distance toward confirming Wigner's suggestion [56] that "the Hamilto-
nian which governs the behavior of a complicated system is a random
symmetric matrix, with no particular properties except for its symmetric
T(L) ' " ' 1
GOE
NDE (experiment)
2 3 4 L
Poisson
NDE (experiment)
Fig. 9. Variance S2, skewness 71 and excess "fa of the number statistic for the
nuclear data ensemble (NDE). GOE and Poisson predictions are plotted for the
sake of comparison. (Taken from [36]).
54 Oriol Bohigaa
Fig. 10. Measure of the two-level form factor 6(fc), see (3.21), from the nuclear
data ensemble (NDE). GOE prediction is plotted for the sake of comparison.
(Taken from [55]).
GOE NDE
a O -0.02
/3 0.55 0.52
a O 0.005 -0.007
6 0.10 0.02 0.11
C O 44 0.02 0.45
Table S. Values of a, ft in (5.4) and a,b,c in (5.5), as determined from the

nuclear data ensemble (NDE). a and /3 are obtained from the values at L =
0.25 and L 0.5 ; a, 6 and c by fitting the number variance E2 in the range
1 < L < 25. GOE VcJues are also included for the sake of comparison. (Taken
from [36]).
R3(X 102) E 4 (X 102)

L GOE NDE GOE NDE
0.5 0.1 0.1 0.1
1.0 5.4 0.8 5.8 0.1 0.1 0.1
1.5 45.3 2.3 48.2 4.2 1.4 3.2
Table 4. Values of R3 and Aj, see (3.23), from the nuclear data ensemble
(NDE) and GOE ( one sample error). (Taken from (36j).
nature". As summarised in [54], it has been "... established an astonish-

ingly good agreement with experiment between a parameter-free theory
(GOE) and the data ... The good agreement with experiment, coupled
with the theoretical understanding which is slowly emerging, reinforces
the belief that the GOE fluctuations are to be found in nature under very
general conditions". In this respect and although we shall not discuss
it here, it is important to mention that fluctuations of some atomic [57]
and molecular [58] spectra have also been analysed and compared with
GOE predictions [14]. The statistical significance is much lower than for
nuclear data but a general consistency between theory and experiment
seems to emerge.
5.2. Dynamical basis of spectral fluctuations; universalities

Once the ability of GOE to predict the fluctuation properties exhibited
by data has been established, one could think that no major question
in this field of RMT remains open. In the rest of these lectures, most
of the effort will be devoted to show that this is not the case and that,
on the contrary, by clarifying the origin of the success of the theory as
well as its domain of validity and its limitations, one can put the whole
subject into a new perspective.
Until now (see for instance Dyson's quotation at the beginning of
Section 4 or Wigner's quotation just above), expressions like "complex
systems", "complicated systems", "large number of interacting parti-
cles" , have remained vague and not well specified. Large number means
~ IQ23, or ~ 102, or ~ 10 or ... ~ 10? We shall now follow the scheme
which renders precise the notion of what is simple and what is compli-
cated (or regular and irregular, or integrable and chaotic) when dealing
with dynamical systems. The hope is that a dynamical basis for RMT
56 Oriol Bohigas
predictions can be found and that fingerprints of the ordered or disor-

dered nature of the classical motion can be identified in the quantum
properties. In fact, Berry and Tabor [59], and Zaslavskii [60] pointed
out, at a very early stage, that the phenomenon of level repulsion, as
for eigenvalues of random matrix ensembles, might provide a clue hi the
distinction between spectra of classically integrable and chaotic systems.
We shall restrict our attention to bounded motion of classically con-
servative Hamiltonian systems with d degrees of freedom. At one ex-
treme, namely for integrable systems, there are, besides the energy, d\
other constants of the motion hi involution. As a result, the energy
surface in phase space has the structure of a d-dimensional torus [18-
20], [61]. At the other extreme, for fully chaotic systems showing hard
chaos, there are no invariant tori and a single generic trajectory cov-
ers uniformly the whole energy shell. We shall adopt Einstein's precept
("Everything should be made as simple as possible, but not simpler")
which, in the present case, means to study two-dimensional systems
whose classical dynamics is extreme, either integrable or fully chaotic.
Indeed, for d 1, there is no distinction between a d-dimensional torus
and a (2d l)-dimensional energy surface, distinction which appears only
from d = 2 onwards. Among the two-dimensional systems, billiards are,
for our purpose, especially suited in several respects:
i) they are, among Hamiltonian systems, the most studied from a math-
ematical point of view, and some important analytical properties are
known;
ii) they show a large variety of behaviours, going from the most regular
(integrable), to the most irregular ones (hard chaos);
iii) the spectrum of the Laplacian with billiard boundaries (Schrdinger
operator for free motion)
(A +A 2 )* -(A +)# = O (5.7)
with Dirichlet boundary conditions, say, is discrete and infinite, and ef-
ficient methods have been devised to compute long sequences of levels.
(When referring to "the spectrum of a billiard", it should be under-
stood the spectrum of (5.7) with Dirichlet boundary conditions on the
boundary 3D of the domain D denning the billiard).
On Fig. 11 are displayed the domains of several billiards that we will
find as the story unfolds. On the left are shown the circular and elliptic
billiards, which are integrable. Indeed, besides the energy, the angular
momentum with respect to the center, for the former, and the product
i
\\\\\\\\v
\\\\\\\\v
Pig. 11. Different billiard domains discussed in the text. On the left, regular
billiards: circle and ellipse. On the right, chaotic billiards: Sinai's billiard,
Bunknovich's stadium and a desyrametrised version of it.
58 Oriol Bohigaa
of the angular momenta with respect to the foci, for the latter, is con-
served, leading thus to two independent conserved quantities. On the
right of Fig. 11 are shown three fully chaotic billiards, namely Sinai's
billiard, whose boundary is a square with a reflecting disk at the center,
the Bunimovich stadium already mentioned in Section 2, and a desym-
metrised version of it, where instead of two semicircles as boundaries
one has one semicircle of radius RI and two quarter circles of radii R2
and RZ. This desymmetrised stadium in the presence of a magnetic
field will be discussed in Section 6, when studying time-reversal break-
ing effects. The origin of the chaotic behaviour of Sinai's billiard is the
negative curvature of the obstacle, whereas for the stadium the instabil-
ity mechanism is somewhat less intuitive. small bundle of trajectories
contracts after reflecting on a circular focusing part of the bound-
ary, but subsequently expands in such a way that the synchronisation
of successive contractions is made impossible: expansions prevail over
contractions, producing the stochastic properties of the billiard. More
technically, both Sinai's and Bunimovich's billiards have positive KoI-
mogorov entropy (K-systems). This entropy provides a measure of the
mixing rate of the system, and is related to the mean rate of exponential
separation of trajectories.
What are the fluctuation properties of the spectra corresponding to
these different billiards? Special care must first be paid to (non-generic)
exact symmetries, as with exact quantum numbers when discussing com-
pound nucleus resonances. For the circular billiard, the eigenfunctions
are
iTi / A\ _ JlL \~imo
m=
(O if
- \\ j.i.,1
M if:t ,.^0 .' 5 = 1,2,... .
(5.8)
In (5.8), Jv is the i/-th Bessel function and kv ,R is the s-th zero of the
j/-th Bessel function
Jv(kVi>R) = O , (5.9)
where R is the radius of the circle. To avoid twofold degeneracies as
implied by (5.8), one considers a semicircular billiard, whose spectrum
is given by the nondegenerate quantised values kv>, with v / O. For
the stadium billiard, as discussed hi Section 2, there are four symme-
try classes and the fluctuation properties, except for the lowest part of
Fig. 13, are analysed considering separately each symmetry class, as im-
plied by the (uncoupled) block structure of the Hamiltonian. For Sinai's
P(X) A,
us
5IHQ I'i. LlIIiOI (J
i /Ii /IJ
i: ,A_
--vj
Ii
Fig. 12. Spectral fluctuations for Sinai's billiard (see Fig. 11). (a) spacing dis-
tribution histogram; (b) Dyson-Mehta statistic. GOE and Poisson predictions
are plotted for the sake of comparison. (Taken from [62]).
billiard, to avoid degeneracies too, one considers the eigenvalues cor-

responding to the symmetry class whose eigenfunctions vanish on the
boundary indicated on the onset of the left part of Fig. 12.
The inspection of Fig. 12, top and middle part of Fig. 13, and Fig. 14,
very strongly suggests that, corresponding to the two limiting cases of
the classical motion, there are characteristic fluctuation patterns of the
quantum spectra. In fact, Berry and Tabor [58] have shown that generi-
cally, for classically integrable systems, the quantal spectral fluctuations
in the semiclassical limit are Poissonian. At the opposite extreme of
classically chaotic motion, the results suggest GOE fluctuations for the
quantum system. It is this kind of extensive numerical evidence which
led the authors of [62], in a Letter entitled "Characterization of chaotic
quantum spectra and universality of level fluctuation laws", to conclude:
"... the basic hypotheses leading to RMT have always been put forward
by invoking the complexity of the system. In other words, it has been
taken as essential that one is dealing with a many-particle system (sys-
tem with many degrees of freedom). Our results indicate that this is by
no means a necessary condition. Indeed, the quantum chaotic system
with two degrees of freedom studied here shows also GOE fluctuations.
The present work should have further developments (for instance, when
tune-reversal invariance does not hold, the adequate model in RMT is
the GUE and one should look for "simple" chaotic systems having GUE
fluctuations). It is an attempt to put in close contact two areas
random matrix physics and the study of chaotic motion that have
remained disconnected so far. It indicates that the methods developed
60 Oriol Bohigaa
1 1 1 1
I I
2 x
T
I i ^ ' I
GOE
05 Stadium
(mixea \
symmetries;
a i 2 *
Fig. 13. Spacing distribution histogram for the semicircular and stadium bil-
liards (see Fig. 11). Top: semicircular billiard. Middle: stadium billiard with
separate analysis for each symmetry class. Bottom: stadium billiard, disregard-
ing symmetries. GOE and Poisson predictions, as well as, for the last case, the
nrediction corresponding to the superposition of four independent GOE spec-
tra (see (D.10), (D.12)-(D.15)), are plotted for the sake of comparison. (Taken
from [33]).
HLJj Poisson
o semi.circle
.stadium
GOE
Fig. 14. Variance E2, skewnesa î and excess "ft of the number statistic for the
semicircular and stadium billiards (see Fig. 11). GOE and Poisson predictions
are plotted for the sake of comparison. (Taken from [33]).
62 Or/ Bohigas
in RMT to study fluctuations provide the adequate tools to characterize

chaotic spectra and that, conversely, the generality of GOE fluctuations
is to be found in properties of chaotic systems. In summary, the ques-
tion at issue is to prove or disprove the following conjecture: Spectra of
time-reversal-invariant systems whose classical analogues are K-systems
show the same fluctuation properties as predicted by GOE (alternative
stronger conjectures that cannot be excluded would apply to less chaotic
systems, provided that they are ergodic). If the conjecture happens to
be true, it will then have been established the universality of the laws
of level fluctuations in quantal spectra already found in nuclei and to a
lesser extent in atoms. Then, they should also be found in other quantal
systems, such as molecules, hadrons, etc...".
At first sight, it may be surprising that the spectral fluctuations of
a regular system are Poissonian (completely uncorrelated) whereas the
ones of chaotic systems are GOE-like (with quasicrystalline order). How-
ever, consider the effect of superposing at random n picket fences (spec-
tra of equally spaced levels). On the interval [0,1] one takes n points
I, (i = l,...,n) at random uniformly distributed and one constructs
an infinite spectrum by attaching to each point x,- a picket fence of
unit spacing. The resulting spectrum, in the limit of large n, is Poisso-
nian. Something similar occurs with the spectra of integrable systems,
the semicircular billiard, for instance [33]. The eigenfrequencies kVi, (i/
fixed, s = 1,2,...) extend from ~ t/ to infinity with a density
(I/TT) (1 - j,2/fc2)1/2 for k > v (5.10)
and are practically fluctuation-free. When examining an interval at high
frequencies containing t ordered eigenfrequencies {fcf} (i' ! ) > ) , suc-
cessive values of i,- correspond to different (and unordered) values of v.
The point now is that zeros of J1, and J1/ with v not too close to v' are
likely to be uncorrelated. Near to AUQO, for instance, the labelling (i/, s)
of successive eigenfrequencies is: (82,4), (70,7), (37,18), (45,15), (5,32),
(32,20), (3,33), (1,34), (18,26), (60,10). We are therefore in a similar
situation as when superposing picket fences, which leads to a Poissonian
spectrum. More generally, when one realises that all the precious infor-
mation encoded hi quantum number assignments of individual levels is
disregarded when considering spectral fluctuations, it is not so surprising
any more that classical ordered motion leads to Poissonian fluctuations.
In some respects, the situation is not very different from when listen-
ing simultaneously to a Mozart, a Beethoven and the Debussy string
quatuors: the result is not far from noise!
5.5. More about symmetries

Some further aspects of the role of exact symmetries are worth men-
tioning. One expects that, if GOE fluctuations apply to eigenvalues of a
chaotic system corresponding to states of a given irreducible representa-
tion of a unitary symmetry of the Hamiltonian, the result of disregarding
the symmetry will result in a superposition of independent sequences,
each component corresponding to one representation, the relative weight
being given by the relative dimensionality (see Appendix D). As illus-
trated at the lower part of Fig. 13, for the stadium billiard, for instance,
one has four representations of equal weights; and the spectral fluctua-
tions do indeed agree with this prediction. Notice that p(s O) = 3/4,
in agreement with (D.14).
To take into account non-generic features due to particular symme-
tries, the general scheme summarised in Table 1 should be implemented,
as pointed out by Robnik and Berry [63]. Let T be an antiunitary sym-
metry such that
T2 = 1 and THT l
=H , (5.11)
(an example of antiunitary symmetry is time reversal). The cases T2 =

+1 and T2 = 1 arise for systems with, respectively, integer and half-
odd-integer values of the total angular momentum; the latter case is
only possible if the number of fermions in the system is odd. T2 = 1
implies Kramers' degeneracy. The association of symmetries with GOE,
GUE and GSE can be summarised as follows [64] :
GOE applies in either of two situations:
i) one or several antiunitary symmetries exist, all with T2 = +1. Only
!evels within a subspace containing states of a given geometrical unitary
symmetry must be considered;
ii) there is an antiunitary symmetry T with T2 = 1 in addition to
sufficiently high geometric symmetries (like two anticommuting parities
both squaring to 1, or full rotational symmetry).
GUE applies in either of two situations:
i) there is no antiunitary symmetry at all
ii) there is one antiunitary symmetry with T2 = 1 in addition to one
geometrical symmetry (one parity) S with S2 = 1, [T, S] = O.
GSE applies for systems with Kramers' degeneracy, T2 1, lacking
any geometric symmetry.
As a consequence, the hydrogen atom in a strong uniform magnetic
field corresponds to the case /3 = 1 and not /8 = 2, even if there is
64 Oriol Bohigaa
time-reversal breaking [14]. Similarly, to induce a transition from GOE

to GUE (see next Section) by applying a uniform magnetic field to a
particle moving in Bunimovich's stadium, the stadium must be desym-
metrised (see Fig. 11), otherwise one would remain in the GOE regime.
Caurier and Grammaticos [65] have studied the motion of a spin- 1/2
particle governed by the Hamiltonian
H = (Px + Pl + P)/2 +W2 + 24/10 + 12zV + Ux2Z2 + 16zV

+f(r).a + r*z(ax + by) ;
(5.12)
in (5.12) , t and s are the orbital and spin angular momenta respectively.
The values of the parameters are chosen in order to ensure that the
classical dynamics is ergodic. By playing with the symmetries of (5.12),
according to the preceding discussion: i) when the last term is absent
(a = b = O), one expects GOE fluctuations, ii) when either a or b is
different from zero, one expects GUE fluctuations, and iii) when both a
and b are different from zero, one expects GSE fluctuations. One should
therefore recover the three cases ft 1, 2 and 4 and this is indeed what
is found numerically, SLS shown on Fig. 15.
Before leaving problems related to symmetries, let us mention two very
intriguing theorems which relate in a very direct way the correlations of
eigenvalues of the orthogonal, unitary and symplectic ensembles.
Let {,-} and {(-} be two independent ordered sequences which have
the correlations of eigenvalues of the orthogonal ensemble with the same
average spacing.
Theorem 1 [66,4]. Construct another sequence {,- '} from the previous
ones by the following operation: superimpose the two sequences and
then pick alternate values from the mixed sequence. In view of a certain
biological analogy, the superscript M may here be considered to denote
the word "meiosis". The correlations of the sequence {x\ '} are those
c.* eigenvalues of the unitary ensemble.
Theorem S [50,4]. Construct from {,-} (or from {[}) a sequence by
picking alternate values. The resulting sequence has the same correla-
tions as eigenvalues of the symplectic ensemble.
The two theorems together imply that the statistical properties of all
three ensembles can be reduced to properties of the orthogonal ensemble
alone. It is amazing that two such similar theorems should exist without
any obvious or logical connection between them, and that the proofs of
the theorems are mere verifications, giving no hint of deeper causes nor
of physical principles that should be at their bases.
LINEAR REPULSION
1.2
QUADRATIC REPULSION
1.0
o.a
^0.6
0.4
0.2
0.0
QUARTIC REPULSION
Fig. 15. Spacing distribution histogram for the spectrum of the Hamiltonian
(5.12). Top: without the last term in (5.12). Middle: with a ^ O or b ^ O in
(5.12). Bottom: with a ^ O and b + O. The GOE, GUE and GSE predictions
(see Fig. 5 and (A.1), (A.2), (A.3)) are also plotted for the sake of comparison.
(Taken from [65]).
66 Oriol Bohigas
From a dynamical point of view, the content of these theorems, in

particular of Theorem 1, seems to offer interesting possibilities which
have not been explored so far. For instance, instead of proceeding in
the way we did when considering the mixed sequences of the stadium
billiard, it seems much more appealing to proceed in a way close to The-
orem 1. By disregarding symmetries, instead of going towards a Poisson
sequence, one may go further away from it. Notice also that Theorem 1
can be viewed as an intrinsic definition of an exact symmetry in terms
of spectral fluctuations, with no reference to explicit geometric features,
etc... Assume, for instance, a parity conserving Hamiltonian leading to
chaotic motion, with the same average quantum eigenvalue density for
both parities. Just by picking alternate levels of the complete spectrum
one should observe correlations as for eigenvalues of the unitary ensem-
ble, indicating that an exact symmetry is operating. On the other hand,
in connection with Theorem 2, one should consider the possibility of
performing an analysis of the nuclear data ensemble (NDE), hopefully
enlarged, by picking alternate values. Symplectic ensemble properties
should be observed and, in particular, instead of the smooth behaviour
of the two-level form factor b(k), see Fig. 10, the singularity at ft = 1
should show up, see Fig. 6b.
5.^. Particular cases

Before we further proceed, one should mention two important non-
generic systems which do not fit into the general scheme of universality
classes outlined above. One concerns regular systems, the other chaotic
systems.
Berry and Tabor [59] pointed out that a system of two uncoupled
harmonic oscillators, an innocent system in appearance and somehow
the integrable system par excellence, is not generic (the energy contours
in action space are flat) and that the spacing distribution p(s) is not
Poissonian but it rather shows level repulsion. In fact, this needs qual-
ifications as will be now shortly discussed, and, as shown by Pandey et
al. [67], the problem can be completely solved.
Without loss of generality, consider the spectrum
E11V = Ii+ OLV n, v = 0,1,2,... (5.13)
with a irrational such that O < a < I. The smoothed density is
< p(E) > =- E/a. (5.14)

and, as usual, we measure spacings in units of the local mean spacing

(the spectrum is unfolded or rectified).
It is easy to convince oneself that the spectrum in the interval
[M, M + 1] contains NM levels with
NM = - Int [-(M + l)/a] = M/a (5.15)
and that the position xn of the levels is given by
M + Frac(rea) for n = 0,1,... ,NM - 1 (5 \6\

M + 1 for n NM
In (5.15), (5.16) the usual notation x = Int(i) + Frac(i) is used. Notice

that the xn are unordered. If one plots on a circle of unit length the
values of the arc lengths Xn starting from an arbitrary point taken as
the origin, for M oo [Nu ~* oo) the points Xn cover uniformly the
circle ((5.16) is ergodic on the circle). However, there are, for given a
and M, at most three different spacings in the interval [M, M +1]. This
means that, locally, the spacing distribution is
3
/ * \~^* cl \ / c -i tj\
*P IY
MM . , (\ S t/ = '
/ * WiO[S
1
V 1 Si /I . \( O . l /il
t=l
From the continued fraction expansion of a, it is possible to work out

analytical expressions for to,- and s, in (5.17) and to prove that the fluc-
tuation properties are not stationary but rather depend on the energy.
Furthermore, when M oo (NM * oo), there is no limit for PNM(S)-
Is there a limiting distribution by defining an average p(s,Nmax) with
the same weight for all spacings,
p(s,Nmax)=^NPN(S)/N^N ? (5.18)
i / i
The answer is in the negative. However, by making specific assumptions

on a, one can derive some precise results. The discussion becomes tech-
nical and we refer the interested reader to [67]. Let us just show some
"monstrous" spacing distributions p(s) for a = (-\/5 l)/2, which is the
simplest case (see Fig. 16). The four spacing distributions displayed and
labelled (1,1,0), (1,1,1/4), (1,1,1/2), (1,1,3/4) correspond to the first ~
21000, 22500, 28800 and 36000 first levels, respectively. By increasing
the number of levels, exactly the same patterns would reappear. For
Oriol Bohigas
Z.Q
Fig. 16. Spacing distributions p(-*} for two uncoupled harmonic oscillators
(5.13), with frequency ratio eqir.. t-~ the golden mean. See text for Turther
explanation. (Taken from [67]}.
instance, by including the first ~ 43800 levels, the spacing distribution

would be the same as the one obtained with the first ~ 21000 levels.
This recurrent behaviour is analytically understood. It is related to the
periodicity of the continued fraction expansion of quadratic irrationals
(irrational numbers which are roots of a second degree algebraic equa-
tion with integer coefficients). One can also show that, if a is a quadratic
irrational, there is strong level repulsion, namely p(s) = O for s < S0(a),
S0 (a) being the largest for a equal to the golden mean. Let us finally
point out that the problem discussed here may have some relevance for
studies of ventricular arrhythmias in the heart, where models with two
separate rhythms are proposed [68].
A numerical experiment, which raises interesting questions, is worth
mentioning at this point [69]. It concerns a one-dimensional, and there-
fore integrable, system. The experiment can be summarised as follows.
Generate a set of eigenvalues {,-} from an N x N GOE matrix and
then unfold the spectrum, {.,-} * {&}. A numerical algorithm can be
devised, by a minimisation procedure, to construct a one-dimensional
potential V(x) whose eigenvalues are {x,-}. Results are displayed on
Fig. 17, The smooth density of {ij} being constant, V(x) must be, on
the average, a quadratic potential (see upper part of Fig. 17). And lo-
cally (see the lower part), complicated fluctuations are superimposed to
V(x)E_
7.0 :-
21.0 -
3.0 7.0 !i.Q 21.J 23.0Y25J

/\
35.0
V(X)E
23.0 E-
200 L
21.0 22.0 23.C Zi.D x 25.G 25.3
Fig. 17- Upper part: fitted potential V(x) to eigenvalues of a 500 X 500 GOE
matrix. Lower part: V(x) in the interval 21 < ar < 26, to show the details of
the potential. (Taken from [69]).
70 OrM Bohigaa
the smooth behaviour of V (x). The experiment has been performed for
JV = 500. Would the algorithm converge in the limit of large N ? In
the affirmative, how singular would be the resulting V(x) ? This exper-
iment, literally, seems to provide a counterexample to the universality
conjecture. Clearly, a better understanding, in particular of the above
questions, is needed.
Concerning non-generic chaotic systems not fitting the general scheme,
we mention tiling billiards on surfaces of negative curvature [34], [7O].
Despite being chaotic, these systems show Poissonian fluctuations, as
discussed in detail by Schmit [16]. It is important to notice that these
tiling billiards are also non-generic, among a large class of chaotic bil-
liards, from two further points of view. Namely, they constitute the
only case for which the Gutzwiller trace formula is exact, and they are
most probably the only systems of bounded geodesic motion that can
be exactly coded with both finite alphabet and grammar rules [70,71].
These three exceptional features are certainly not disconnected.
A very special and interesting case, discussed in detail in Berry's
lectures [12], deserves here a brief comment. The Riemann Hypoth-
esis states that all the non-trivial zeros (zeros different from s =
2, 4, 6,...) of Riemann's f-function, defined by
COO=E1K, Rez>l, (5.19)

n
and by analytic continuation elsewhere in the z-plane, are of the form
s = 1/2 + iEn , n 1,2,... . Consider the sequence (En], En > O. Its
average counting function is known and given by
< N(E) > ~ (E/2n) ln(E/27re) + 7/8 , (5.20)

(see (2.25)). Montgomery conjectured, and there is very strong numer-
ical evidence supporting it [72,73], that the two-level cluster function
corresponding to {En}, once it is unfolded, behaves like the two-level
cluster function for eigenvalues of random matrices of the unitary ensem-
ble (GUE). There is evidence supporting the extension of Montgomery's
conjecture: zeros of f (z) behave like eigenvalues of the GUE (GUE hy-
pothesis, in Odlyzko's terminology [73]). As already emphasised, it is
important to avoid redundant analyses and to try to test separately
the hierarchy of the fc-level cluster functions. Results for the skewness
and excess of the number variance, giving support to the extension of
Montgomery's conjecture, are presented on Fig. 18. Notice that one
can see the short range oscillating structures coming from "data", as
GUE
Y 2 (L)
2.0 GUE
1.5
1.0
0.5
-0.5
-1.0
-1.5
L 3
Fig. 18. Skewness î and excess -72 of the number variance corresponding to
the imaginary part En of the zeros Sn = 1/2 + iEn of the Riemann f function,
compared to the result of eigenvalues of random matrices of the unitary ensem-
ble. The data, provided by A.M. Odlyzko [72], correspond to the analysis of a
stretch of 10s consecutive zeros starting at the (2 X 1011 + l)-th zero.
72 OnW Bohigas
encountered already in the two-level cluster functions of the random

matrix ensembles (see Fig. 6 and (C.2), (C.3), (C.6)) and seen in the
YZ function extracted from the "data" [72,73]. Odlyzko's most recent
calculations allow to make significantly sharper tests of the GUE hy-
pothesis than the one illustrated on Fig. 18 [73]. In particular, it would
be interesting to extract from the data the pure Appoint function Rk, see
(3.27), for small values of L, for k 3 and 4.
5.5. Breaking of the universality regime
Until now, with the exception of some non-generic cases, we have ex-
clusively given evidence which is consistent with the general scheme of
universality classes. We shall now give evidences which show departures
from it. For that purpose, it is convenient to study the level deviation ,
see (2.39). On Fig. 19 is compared the behaviour of a Poissonian spec-
trum and of an integrable (separable) system; on Fig. 20 the behaviour
of eigenvalues of a random matrix and of eigenvalues of a quantum sys-
tem whose classical analogue is chaotic. Notice the difference in scale
for the ordinates in Figs. 19 and 20. With no need of sophisticated
statistical measures, the mere inspection of the figures reveals that the
n 6n plot does not fit the previous general scheme. The long excursions
of 6n far from zero for a Poissonian spectrum are typical of Brownian
motion. In contrast, eigenvalues of the integrable system are not allowed
to wander far from zero and rather exhibit periodicities with military
precision. Similarly, although less visible, the deviations from zero for
the eigenvalues of the chaotic system show periodic components which
are not present in the spectrum of the random matrix; notice also that
the amplitude of the deviations Sn in Fig. 20 are smaller than in Fig. 19.
Before we discuss the origin of this phenomenon, let us mention that
the n Sn plot, as first pointed out by Schmit et al. [74] when studying
the spectrum of a billiard in the presence of a magnetic field, provides
an efficient detector of missing or spurious levels especially suited for
chaotic systems. This may be important from an experimental as well
as from a computational point of view. Observe first that the levels
deviate with a small amplitude from zero. Assume, for instance, that
the 1400-th level has been missed. As a consequence, from n =: 1400
onwards, the computed value of Sn would not fluctuate around zero,
as it should, but rather around 1, and this would be clearly visible on
the plot. Analogously, would there be around 1400 a spurious level, Sn
would not fluctuate from n ~ 1400 onwards around zero, as it should,
-40
1000 1200 1400 1600 IBOO 2000
n
40
I
30
-
i
t ..,,..,
OO UOO 1400
. i
1600
, , I
1800
.1
-1
3000
n
Fig. 19. Fluctuating part of the counting function N(E) as expressed by the
level deviation Sn, see (2.39), for a slice [1000-2000] of 1000 levels, (a) from a
Poiflsonian spectrum with the centroid adjusted, (b) from the two uncoupled
anisotropic quartic oscillators b = 7T/4, ^ = O in (2.20) for levels with
reflection symmetry (+,+); the unfolding is performed using (2.35). (Taken
from [24]).
74 Oriol Bohigaa
.000 1200 1400 1600 iaoo 2000
-n
-'JOO UOO 1400 1600 iaoo 2000
Fig. 20. Fluctuating part of the counting function N(E) as expressed by the
level deviation 6n, see (2.39), for a slice [1000-2000] of 1000 levels, (a) from
a 5000-dimensional GOE matrix; the spectrum is unfolded analytically using
Wigner's semicircle law (4.30). (b) from the two coupled anisotropic quartic
oscillators 6 7T/4, g=-0.60 in (2.20) for levels with reflection symmetry
(+, +); the unfolding is performed using (2.35). (Taken from [24]).
but rather around 1. Although very precise in detecting the number

of missing (or spurious) levels if there are only a few, the n 6n plot
gives only an approximate location of the missing (or spurious) levels.
We will now discuss some of the departures from universalities, as
described by RMT or Poisson fluctuations. We will only sketch some of
the ingredients and results, as this subject is covered in Berry's lectures
[12].
An adequate framework is provided by the semiclassical Gutzwiller
trace formula, which relates the trace of the quantum mechanical
Green's function to the sum over classical periodic orbits, as discussed in
Gutzwiller's lectures [75] . Following Berry, there are, at least, two scales
in the problem. One is the inner scale, provided by the mean level spac-
ing <D> <p>~* ~ hd (see (2.7), (2.8)), where h is Planck's constant
and d the number of degrees of freedom. The other scale, called outer
scale, is provided by h. The spectral properties with inner-scale resolu-
tion are determined by the very long classical orbits whereas the ones
with outer-scale resolution are determined by the short ones. Clearly,
the properties of the latter depend on each particular system and give
rise to non-universal features of the spectral fluctuations, features which,
of course, are not contained in RMT descriptions. Notice that all the
discussions we have made so far concerned the inner scale. Thus, for
instance, when previously referring to short- and long-range correlation
effects, it should be understood short and long range at the inner scale,
but both short range at the outer scale.
Let us illustrate non-universal features with a couple of examples.
From the periodic orbit formalism one expects that the power spectrum
I /+< 2
= / p(E)dE (5.21)
I J-OO
where p(E) is the level density of the (scaled) energy spectrum, will
show distinct peaks, which can be identified with the periods of the short
periodic orbits whose instability is relatively weak. Two examples are
given in Fig. 21. The upper part of the figure corresponds to the stadium
billiard. For the particular case of billiards, the periods are, of course,
directly proportional to the length of the periodic orbit. The first and
most prominent peak can be identified with the family of non-isolated
two-bounce orbits which hit the straight-line boundary perpendicularly
(see Fig. 11). The subsequent peaks correspond to longer (and isolated)
orbits. Similar features can be seen for the power spectrum of the two
coupled quartic oscillators displayed on the lower part of the figure.
76 Oriol Bohigaa
35
30 -
25
20
15
10
O
10 15
1.0
0.9 -
0.8 -
0.7 -
0.6 -
0.5 -
0.4 -
0.3 -
0.2 -
I
0.1 -
0.0
Fig. 21. Power spectra corresponding to two chaotic systems, illustrating the
role of periodic orbits. Upper part: for the stadium billiard (taken from [76]).
Lower part: for the Hamiltonian (2.27), from the lowest 60 states of a given
symmetry class (taken from [77]).
The periodicities visible in the n 6n plots, see lower parts of Figs. 19

and 20, can also be accounted for by the semiclassical periodic orbit
theory and are also due to the presence of (non-universal) short periodic
orbits. These are also responsible for "saturation effects" in, for instance,
the Dyson-Mehta statistic [78], or the number variance [79] : from a
value
Lraax - h < D > /T^n ~fc-l*-'), (5.22)
onwards, where T1nJn is the period of the shortest closed orbit, A as a
function of L ceases to increase but rather saturates, whereas E2 ceases
to increase but rather oscillates. These saturation effects, present in
chaotic as well as in regular systems, have been observed in several
computations, since the first observations made by Seligman et al. [80]
for a chaotic system and by Casati et al. [81] for a regular system
(rectangular billiard).
Very pronounced non-universal "saturation effects" have also been
observed for the zeros of Riemann's function. It should be reminded
that there is a mechanical analogy for the zeros of C [78,2O]. If one
interprets the positive semiaxis Re z = 1/2 as the energy axis and the
positions of the zeros as discrete quantum energies of a classical system,
the semiclassical theory indicates that the action integrals of primitive
orbits should be identified to hi p, where p runs through the set of prime
numbers. The "saturation effects" axe therefore related to prime number
properties. Their effect, for instance, on the number variance, is in
spectacular agreement with Berry's semiclassical theory. Comparison of
the theory with Odlyzko's data can be found in Rf. [79] for a range
O < L < 1000 around the 1012-th zero, and in Rf. [73] for a range
O < L < 5 x 105 around the 1020-th zero. The agreement between
computed and predicted values s almost perfect. TKe non-universal
regime extends until a few units of a mean spacing, and this is also
understood.
Despite common wisdom saying that the semiclassical periodic orbit
theory accounts also for the observed spectral fluctuation universalities,
as predicted by RMT, the situation, at the inner scale, is less conclusive.
One should first observe that the main ingredient used to derive spectral
fluctuations at the inner scale, dominated by very long classical orbits, is
Hannay and Ozorio's uniformity principle [82], [18], [2O]. This principle
is a conjecture which, roughly, says that most of the very long periodic
orbits are almost indistinguishable from the ordinary (non-periodic) tra-
jectories. It gives rise to sum rules involving the amplitudes in the peri-
odic orbit summation formula, sum rules which have distinct behaviours
78 Oriol Bokigas
for regular and chaotic systems [82]. The information contained in these
sum rules can be translated in terms of some of the two-point fluctua-
tion measures discussed above. For instance, the large-r (at the inner
scale) behaviour OfV 2 (O 3^ predicted by RMT, see (C.4), (C.5), can be
recovered for chaotic systems with or without time reversal invariance
[82], [18]. Concerning the two-level form factor 6(A:), see (3.21), Berry
obtains [78]
, ... .... ( 1- k non time-reversal - .
for small A; b(k)
v = < , . .. , v(5.23)
' ^ 1 - 2fc time-reversal '
for large k b(k) -> O , (5.24)
in agreement with GUE and GOE predictions (see (C.1O)-(C.12)). It
should be understood that small k does not mean k close to zero but
rather k > k', where fc* is estimated from the period of the shortest
periodic orbit (see Berry's lectures [12]).
In summary, the semiclassical Gutzwiller trace formula has been very
successfully used to predict spectral fluctuations at the outer-scale en-
ergy resolution. To achieve inner-scale resolution, problems remain. Fur-
ther progress made to better understand the validity of the uniformity
principle may indicate, for instance, how to interpolate between the
small- and large-fc regime for 6(A;), and to investigate if higher than two-
point functions can be semiclassically derived. It is worth noticing that
level repulsion, i.e. small inner-scale behaviour, has resisted until now
semiclassical analysis. And it may turn out that the periodic orbit for-
malism is not the ideal probe for the study of such fine scales of the
spectrum [18].
5.6. New experiments

With the exception of Subsection 5.1, where a detailed comparison of
RMT spectral fluctuation predictions and nuclear resonance data was
presented, very little about experiments has been said. This could leave
the impression that the whole subject is presently exclusively dominated
by theory, applied mathematics and numerical simulations, with no room
for real experiments. The rest of this Section is meant to show that this
impression should be corrected.
We first mention the Kepler diamagnetic problem, which offers a
unique opportunity for studying chaotic motion from several points of
view: i) classical as well as quantum properties can be computed accu-
rately, ii) there is an interesting transition from regular to chaotic motion
when the magnetic field strength is increased, iii) it is an interesting sys-

tem from the point of view of symmetries (dynamical symmetries), iv)
last, but not least, it can be studied experimentally. This subject is
important and a whole chapter of the present volume is devoted to it
(Delande's lectures [14]).
In this Section, particular emphasis has been put to "quantum bil-
liards". Mathematical properties as well as results of numerical simula-
tions have been discussed. But one is naturally led to ask whether these
systems can also be experimentally investigated. And in this respect,
the recent work by Stckmann and Stem deserves special emphasis [83].
These authors have studied, to simulate "quantum billiards", microwave
resonators. In general, different boundary conditions apply to quantum
billiards and to electromagnetic cavities. For the former, the wave func-
tion * vanishes at the boundary (Dirichlet), see Eq. (5.7), whereas for
the latter the tangential component of the electric field and the normal
component of the magnetic field vanish at the boundary. However, if the
cavity is built in such a way that the top and the bottom face are parallel
to each other at a distance d, for frequencies v < t/max = c/2d, where
c is the velocity of light, only transverse magnetic modes are possible
(Bf = O). The electric field is directed in the direction perpendicular to
the top and bottom faces of the resonator (z-direction) and it satisfies
the same wave equation as the quantum-mechanical particle in a box.
One of the important parameters characterising a resonator is its qual-
ity factor Q in a range of frequencies v. It is defined as Q = J//AJ/, where
v is a frequency in that range and Az/ is its typical width. The phys-
ical origin of Ai/ is the attenuation due to the fact that the walls of
the resonator are not perfectly conducting (power dissipated in ohmic
losses). Two successive eigenfrequencies can be just resolved if they are
separated by at least Ai/, leading to the estimate that the total number
of eigenfrequencies that can be resolved is of the order of Q. With a
moderate effort, Q of the order of several thousands can be achieved,
but with further effort much higher values can be attained (~ 104, and
even ~ 105 is not ruled out).
In the experiment of Rf. [83], d ~ 1 cm, J/max ~ 20 GHz and Q ~ 103.
The microwaves are transmitted to the resonator cavity through a mi-
crowave cable and the reflected microwave power is measured as a func-
tion of frequency in a range 0.01 GHz < v < 20 GHz. On Fig. 22 are
reproduced some of the results. On Fig. 22a each resonance corresponds
to an eigenfrequency of the stadium shaped cavity, on Fig. 22b the spac-
ing distribution of a Sinai billiard shaped cavity is reproduced, and on
80 Oriol Bohigas
U
(U i-
ii
?:.
U >
Oj O
ft
V
UO IO 20 JU
Fig. 22. Resonances of microwave cavities, (a) Part of the spectrum of a sta-
dium shaped cavity, the abscissa corresponds to the frequency in GHz. (b)
Histogram of the spacings between successive eigenfrequencies for a Sinai
billiard shaped cavity, shown in the inset of (c); a total of 1002 eigenfrequen-
cies has been registered hi the range 1-18 GHz. The dotted line corresponds
to a Wigner distribution, Eq. (A.I). Inset: histogram of spacings of eigenfre-
quencies for a rectangular shaped cavity (integrable), compared to a Poisson
distribution, Eq. (3.33); 314 eigenfrequencies below 11 GHz are included. The
loss of close-lying eigenfrequencies can be experimentally understood, (c) Power
spectrum of the Sinai billiard shaped cavity, shown at the inset. Each peak in
the power spectrum corresponds to a classical periodic orbit. (Taken from [83]).
Fig. 22c the measured power spectrum is displayed, showing the role of
the classical periodic orbits (compare, for instance, to Figs. 12a and 21).
This experiment should not only be viewed as confirming what we
already knew from numerical computations, but rather as opening the
door to measurements which can go far beyond present computing abil-
ities. For instance, resonance frequencies of cavities with complicated
shapes, "amoeba shapes", can be probably measured and the evolution
of the eigenfrequencies when the shape changes is attainable. Com-
plicated shapes, reminiscent of the succession of shapes of an atomic
nucleus when undergoing fission, for instance, should be investigated.
One should also remark that it is much easier to go from two to three-
dimensional systems experimentally than computationally. In fact, this

type of experimental set-up can be considered as an analogue computer,
illustrating the great superiority of an analogue over a digital computer
when dealing with some particularly difficult numerical problems. Con-
trary to the main evolution during the last decades, where digital com-
puters have almost eliminated analogue computers, some of the prob-
lems we are discussing, as well as other difficult problems in non-linear
dynamics, may give impetus for a renaissance of specialised analogue
computers.
Can one also study experimentally classical billiards? The problem
is easy to state but difficult, in fact impossible hi some sense, to deal
with numerically: as we know, the instabilities prevent the calculation
of long orbits (which does not mean that calculations of "pseudo-orbits"
are worthless!). Let us briefly mention an interesting experimental pos-
sibility, not realised in practice so far. Ultracold neutrons (UCN) are
neutrons with typical energies of the order of E ~ 10~7 eV, which cor-
respond to velocities v ~ 10 ms"1 and wavelengths A ~ 10s A [84]. For
such a long-wavelength particle contained in a vessel, the wall acts es-
sentially as a continuous barrier and the probability for the neutron to
penetrate the wall is very small. Indeed, when a neutron strikes a sur-
face, there exist three possible mechanisms of interaction with the wall:
elastic scattering, inelastic scattering and absorption by a nucleus. At
energies typical of UCN, elastic scattering is by far the most probable
process and despite each individual nucleus attracting a neutron, the
UCN can undergo total reflections from the surface of most materials.
This occurs when the component of their velocity normal to the surface
does not exceed a certain limiting value wlim characteristic of the given
material (typical values are 10 ms"1). Therefore, a neutron will be to-
tally reflected from a material at any angle of incidence if its velocity v
is less than viim. As a result, such neutrons can be stored in a bottle.
(Remember that the mean life of a neutron is ~ 1000 s.) UCN confined
in material walls and especially in "fluid walls" constitute practically a
dissipation-free system of moving mass-points within totally reflecting
boundaries; typically each neutron undergoes ~ 105 elastic reflections.
UCN then provide a unique tool to investigate the transition from classi-
cal to wave-mechanical (quantum) regime by decreasing the typical size
of obstacles inside the vessel, from macroscopic to ~ 103 A "nanostruc-
tures", which can today be produced with ~ 100 A precision [86].
Let us finally mention a possible application of RMT in a different con-
text. It concerns the study of elastodynamic eigenfrequencies recently
82 Oriol Bokigas
a)
OI
O 5. 000
CL
en
o
J
IO. 000
OJ
0 (1KHz)
Fig. 23. Power spectral densities of ultrasound acoustic emission of a small
alumin um block, a) 20-fcHz span of the spectrum, b) 20-kHz span of the
spectrum of the same block, with a different orientation of the structure. (Taken
from [87]).
investigated experimentally from a RMT point of view. Weaver studies

th natural frequencies of small alumin um blocks, of typical sizes of
the order of a few centimetres [87]. The experimental devices include a
support of the blocks, an acoustic source and acoustic detection. The
ultrasound is generated by allowing a small diameter (~1 cm) steel ball
to fall upon the blocks from heights of about 3 cm, and is detected with
a small diameter (2.5 mm) pinlike piezoelectric transducer. The natu-
ral frequencies detected lie in the range 0-250 klls. An example of the
measured power spectral densities is given in Fig. 23. To check that
the distorsion introduced by the support does not alter the measured
natural frequencies, the block is studied under two different orientations
(compare upper and lower part of the Figure).
It is not a priori clear that the correspondence between classical mo-
tion and discrete frequency spectra, established essentially for the scalar
Helmholtz, or Schrdinger, equation, will also hold in the elastodynami-
cal case. Indeed, for the former the equation is of second order, whereas
the basic elastodynamic equations are tensorial and effectively of sixth
order when expressed in terms of a single strain component. Results
for the spacing distribution p(s) and the number variance E2 are repro-
duced on Fig. 24. They strongly suggest that RMT fluctuations apply.
However, it should be stressed that the shapes of the blocks used are
such that the geometrical acoustic limit of the block does not correspond
to a chaotic but to a pseudointegrable system [88]. That, despite this
fact, GOE fluctuations are observed can be attributed to finite wave-
length effects which introduce a coarse-graining in the relevant classical
dynamics [89,9O]. At i.ny rate, Weaver's results seem to confirm that the
GOE spectrum, initially studied for the simple scalar Helmholtz equa-
tion, may also be valid in the more complicated case of elastodynamics,
and they call for further, more systematic and extensive, investigations.
6. Breaking symmetries and induced spectral fluc-

tuation transitions
We have seen that the three universality classes ft = 1,2 and 4 of RMT,
dynamically characterised by classical chaotic dynamics, stem from gen-
eral symmetry considerations. On the other hand, we have also seen
that Poissonian spectral fluctuations at the inner scale hold when the
underlying classical dynamics is regular. What happens when symme-
tries are gradually broken, or when, by varying a parameter, the system
84 Orio/ Bohigaa
a)
L
Fig. 24. Statistical properties of natural frequencies of alumin urn blocks, (a)
Histogramme of 313 spacings between successive eigenfrequencies obtained from
the combined spaciaga of two blocks; GOE (solid curve), GUE and Poisson
(dotted curves) predictions are displayed for the sake of comparison, see Fig. 5a.
(b) Number variance E2 from the eigenfrequencies of one alumin um block;
GOE and Poisson predictions are displayed for the sake of comparison, see
Fig. 7a. See text for further explanation. (Taken from [87]).
undergoes a gradual transition from regular towards chaotic dynamics?

These questions, with special emphasis on spectral fluctuations, will be
discussed in this Section and partial answers will be provided.
6.J. The GOE -+ GUE and GSE -* GUE transitions

Assume that the initial Hamiltonian H0 is time-reversal invariant (TRI)
and that the angular momentum is good or integral (/3=1, see Table 1).
Add to the Hamiltonian a time-reversal breaking part //break
H = HO + -fffaraak (6-1)
For non time-reversal invariant (NTRI) systems, we know that /3 = 2

applies. In RMT language, the problem is then to investigate the random
matrix ensemble of the form (6.1) with
H0 = S(vz) (6.2)
Ht,** = iaA(v*) , (6.3)

where O < a < 1 and S (A) denotes a real symmetric (antisymmetric)
random matrix like in (4.33), (4.34). Remember that v provides the
scale for the size of the matrix elements. For a = O, (6.1-6.3) coincides
with GOE, and for a = 1 with GUE. (6.l)-(6.3) defines an ensemble,
depending on the parameter a, interpolating between GOE and GUE.
The properties of this ensemble have been completely worked out by
Pandey and Mehta [91,4]. We refer the reader to the original literature
and will only reproduce some of the main results.
The joint probability of eigenvalues is given by
. (6.4)
In (6.4), i = 1,2, . . .,2N, and A and f ( E ) are given by

A = U(E, :-E,) (6.5)
(6.4) is valid when the dimensionality of the matrix is even, but the
result can be extended for the odd case. It is convenient to use the
normalisation 2u2(l + a2) = 1.
86 Or to I Bohigas
Notice that the structure of (6.4) is similar to (4.30). The last factor
in (6.4), in the a > O limit, is equal to one, whereas in the a * 1 limit, it
is proportional to |A|. Therefore, one recovers the /3 = 1 and 2 !uniting
cases, see (4.30). Notice also that (6.4) has not the structure corre-
sponding to a Coulomb gas with an inverse temperature intermediate
between /3 = 1 and /3 = 2, see (4.30), (4.68).
As for the GOE and GUE case, the density of eigenvalues of (6.4), in
the large-JV limit, is a semicircle. The n-level correlation functions Rn
are known for arbitrary a, n and W.
What is important, physically, is to identify the transition parameter X
which interpolates the n-level cluster functions between GOE and GUE.
As first derived by French and collaborators [92]-[94], it is given by the
ratio of the root mean square of a typical symmetry-breaking matrix
element over the mean level spacing D
RMS (gbreak)
A=
D(E)
Remember, in (6.7) Hbreak means that part of H which breaks the sym-
metry of HO, see (6.1). In terms of the parameters denning the ensemble,
see (6.1)-(6.3), the transition parameter is given by
(6 8)
-
The expressions for the n-level cluster functions Yn(Xi, . . . , Xn; A) for the
large- N limit of the unfolded spectrum are reproduced in Appendix E.
As can be checked, for A = O one recovers the GOE expressions, and the
A > oo limit gives the GUE expressions.
For fixed a and v, the local mean spacing D(E) = l/p(E) decreases
like JV"1/2, hence A is proportional to i/N and tends to infinity for
large JV. Therefore, in the large-JV limit, when TRI is broken (a ^ O),
the ensemble (6.l)-(6.3) undergoes an instantaneous transition to GUE
fluctuations. However, when going to the large JV limit, but keeping
the ratio av/D finite, the transition is not abrupt. For instance, on
Fig. 25 is plotted the two-level cluster function, see (E.6), for two values
of A. The results corresponding to /3 = 1 (A = O) and ,0 = 2 (A -* oo),
see Fig. 6a, are smoothly interpolated. Notice that the effects of the
symmetry breaking tire, for a fixed value of A, much stronger for small
than for large values of r (remember that r is measured in units of the
mean spacing D). In fact, in terms of A, the transition in the slope of V2
\=0 1
: 15 2
r
i -ti
Y,(r) - \
\ \=0.3
03 *
OO
15
Fig. 25. The two-level cluster function for two values of the transition parame-
ter A describing the GOE > GUE transition (see (E.6)). Upper part: A = 0.1 ;
lower part: A = 0.3.
88 Oriol Bokigaa
ffo flbreak HO + #break

real imaginary
GOE -> GUE
symmetric antisymmetric
Hermitian
self-dual anti-self-dual
GSE - GUE
q , quaternion real q , quaternion imaginary
Table 5. Structure of the Hamiltonian interpolating between GOE and GUE,

and between GSE and GUE, when a time-reversal breaking part is add d to
H0.
at the origin IE instantaneous: for A = O, Yi behaves linearly, whereas for

A ^ O whatever small, it behaves quadratically. This can also be seen
from the small-s behaviour of the spacing distribution p(s) of (6.1)-(6.3)
for the 2 X 2 case (see (E.10).
Other transitions of similar types as for GOE GUE are possible. For
-^r0 "/hen adding a time-reversal-breaking part to a TRI Hamilto-
iTesponding to a non-rotatioi.î-invariant system containing an
odd. number of fermions, one is led to study a GSE (/? = 4) GUE
(/? = 2) transition. The proper random matrix ensemble can again be
written in the form (6.1), with [91,94,4]
H0 = S0(v2)e0 + A1(Vê1 + A2(t>2)e2 + A3(w2)es (6.9)
2 2 2 2
tfbreak = m[A0vt; )eo + Si(V Je1 + S2(v )e2 + S3(t; )e3l . (6.10)
In (6.9), (6.10), 5,- (Ai) (i = 0,1,2,3) denote independent real symmetric
(antisymmetric) random matrices like in (4.33), (4.44), and e,- are given
by (4.3). As summarised in Table 5, m complete analogy with (6.1)-
(6.3) where H0 is the real part of H and .ffbreak is the imaginary part
of H, for the ensemble (6.1), (6.9), (6.10), H0 is the self-dual part of H
(matrix sites with real quaternions, qt real hi (4.74)) and //break is the
anti-self-dual part of H (matrix sites with purely imaginary <fc in (4.74)).
The jouit probability of the eigenvalues has the structure (6.4), but
with f ( E ) given by [91,4]
f(E) = Eexp ^L (6.11)
The ensemble (6.1), (3.9), (6.10) interpolates between GSE (a = O)

and GUE (a ~ 1). For a O, the energy levels are two-fold degenerate
(Kramers' degeneracy) and the degeneracy is broken when a ^ O. Again,

the average level density is semicircular and the n-level correlation fcuc -
tions are known for arbitrary a, v and N. The transition parameter
which interpolates between GSE and GUE is also given by (6.7) . The
n-level cluster functions Vn(Zi, ,EX! A) are given in Appendix E and
the values A = O and - oo correspond to GSE and GUE respectively.
6.2. Dynamical basis of the GOE > GUE transition

Effects due to TRI of the Hamiltonian on spectral fluctuations were
first investigated, from a dynamical point of view, by Seligman and
Verbaarschot [95]. These authors studied the motion of a particle in a
combination of inhomogeneous magnetic and scalar fields. By choosing a
Hamiltonian whose classical limit was chaotic, GUE spectral fluctuations
were exhibited. Since then, ether examples have been studied and one
was given in Subsection 5.3.
However, what we want here is to investigate the transition regime
GOE > GUE. For that purpose, we shall consider a charged particle
moving in a plane two-dimensional domain D (billiard) under the action
of a constant magnetic field B perpendicular to the plane of the billiard.
In the absence of the magnetic field, a trajectory is a succession of straight
line segments with specular reflections on the boundary. When the
magnetic field is applied, a trajectory is a succession of arcs of circles
with specular reflection!-.. The radius R of the circles (Lannor radius) is
constant and given by
- = . <*)
where m, u and e are the mass, the velocity and the charge of the par-
ticle, respectively. If R is less than the typical linear dimension t of the
system, there will be trajectories which are complete circles not hitting
the boundary. Even if hi the absence of the magnetic field the motion is
chaotic, the phase space structure will now be of mixed character (pres-
ence of both chaotic regions and tori, corresponding to regular islands) .
However, if R> I, the magnetic field does not generate regions of regu-
larity and the chaotic nature of the dynamics in the absence of the field
will be preserved when the field is applied.
The quantum-mechanical problem corresponding to the above classi-
cal one consists in solving the tune-independent Schrodinger equation
(p -eA) 2 *(q) = *(q) for * g D

*(q) = O for *(q) e 3D
90 Oriol Bohigas
Fig. 26. Some periodic orbits in Bunimovich's stadium. Except for the top-
right orbit, which is self-retracing, the others are non-self-retracing. (Taken
from [97]).
In (6.13), p is the canonical momentum and A is the vector potential,

which, in the symmetric gauge, is given by A (Bq2/2,Bq1/2). The
Schrdinger equation reads then as
(6.14)
These types of billiards have simple scaling properties. If one scales
the size of the billiard Ln such a way that the surface 5 is multiplied
by 7, S' = T/S, the spectrum changes by E\ = r\~lEi provided that the
magnetic field is scaled by B' i]~lB. The parameter characterising the
system is therefore the magnetic flux $ = eBS which has dimensions
of an action and which will be measured in units of h. Keeping $ < 1
ensures that, even for the low energy part of the spectrum, see (6.12),
the circular (regular) orbits will not affect the quantum spectrum.
In Berry's semiclassical approach [12,78,79], the difference between
GOE and GUE arises from the general existence of pairs of periodic or-
bits. Indeed, as shown on Fig. 26, in the absence of magnetic field, there
are two types of periodic orbits, self-retracing and non-self-retracing
ones. Most of the periodic orbits are pairs of non-self-retracing ones and
they contribute in phase to the periodic orbit sum. The factor of two
difference in the slope of the form factor b(k) at the origin, see (5.23), be-
tween TRI and NTRI systems, can be traced back to the following fact:
when TRI is broken, the phase coherence among orbit pairs is destroyed.
However, when a weak magnetic field is applied, one expects that the
breaking of phase coherence among orbit pairs /ill occur smoothly and
therefore that a GOE > GUE transition will take place.

Eq. (6.14) has been solved numerically for the desymmetrised stadium
(see Fig. 11) [34,74,96]. Remember that if one would instead consider
the symmetric Bunimovich stadium, one would not induce a transition
towards GUE fluctuations, as explained in Subsection 5.3. For several
values of the magnetic flux parameter $, the first 500 eigenvalues of
(6.14) have been determined. To further proceed, what we want is a
link between the transition parameter A, see (6.8), describing the GOE
> GUE transition, and the classical dynamics of the system. This can
be obtained by treating the effect of the magnetic field as a perturbation
and observing that the difference in action of a pair of non-self-retracing
orbits is given by
AS = B f L(t) dt + O(B2) . (6.15)
In (6.15), L(t) is the orbital angular momentum and the integral is taken
over a full period of the unperturbed orbit. The effect of the perturba-
tion on the two-level cluster function can be estimated semiclassically.
One is led to an expression which is identical to the one describing the
GOE > GUE transition, namely Eq. (E.7), if one identifies the transition
parameter A with the magnetic flux $, to within a multiplicative factor
[96]. Results for the Dyson-Mehta statistic are displayed on Fig. 27. As
can be seen, the GOE > GUE transition seems to reproduce the calcu-
lated values very well and A is indeed proportional to $, as predicted by
the semiclassical theory. However, it should be noticed that at scales of
one mean spacing or less, as seen for instance in Ri(L), there seem to
be some inconsistencies which are not fully understood.
The previous discussion differs, in several respects, from Berry and
Robnik [98] who found, and explained, some GUE fluctuation properties
for the case of a billiard perturbed by a single magnetic flux line
the Aharonov-Bohm billiard. In Rf. [98], what was considered is the
semiclassical limit of magnetic flux fixed and h > O, in which case the
transition GOE > GUE is abrupt, as when taking a ^ O in (6.3).
What we have considered here is h small (but fixed) and a small flux $,
corresponding to A finite in (6.8).
Let us finally point out that for the problem just discussed, what mat-
ters is the (linear) paramagnetic term in (6.14). Indeed, both periodic
orbits of a degenerate pair have equal and opposite angular momenta so
that, for a pair, the paramagnetic term generates a difference in action
(6.15). However, both periodic orbits have the same q coordinates, so
92 Oriol Bohigas
Fig. 27. Illustration of the GOE > GUE transition. Dyson-Mehta statistic
A(L) from the spectrum of a charged particle in a deaymmetrised stadium,
see Fig. 11, with a constant magnetic field. The small circles denote values
obtained from the computed spectrum for different values of the magnetic &\ix:
(a) $ = O ; (b) $ = 0.05 ; (c) $ = 0.10 ; (d) $ = 0.20. The thin continuous
Une corresponds to different valu . of the transition parameter A (see (E.6) and
(3.3O)) : (a) A = O ; (b) A = 0.75 ; (c) A = 0.15 ; (d) A = 0.30. Poisson
(P), GOE (0 = 1) and GUE (ft = 2) curves are displayed for the sake of
comparison. (Taken from [96]).
that the (quadratic) diamagnetic term does not break the degeneracy in
action and is uninteresting from the point of view of TRI breaking. In
contrast, when studying the regular towards chaotic transition for the
diamagnetic Kepler problem, the term generating chaotic dynamics is
the diamagnetic ona, quadratic in the field, whereas the paramagnetic
term is uninteresting from the point of view of that transition [14] .
6.3. More about magnetic fields (and rotations)
There is a whole class of problems, involving magnetic fields (or rota-

tions), for which the regular or chaotic nature of the underlying clas-
sical dynamics may be very relevant from a quantum-mechanical point
of view. This should be considered as an entire subject in which some
pioneering work has been done and which will be strongly developed in
the future, for conceptual as well as for practical reasons.
As discussed, for instance, by Mhlschlegel [15], electrons in small
metallic particles constitute an interesting physical system, where fast
progress is being made from an experimental point of view. The simplest
model of a metal is the one of electrons moving independently and the
simplest model of a nucleus is also the one of nuclons moving indepen-
dently in an average field. These independent-particle models, despite
their naive aspect, are quite successful in explaining a large body of
empirical facts. It is then natural to ask whether the regular or chaotic
nature of the underlying classical motion of electrons (or nuclons) in av-
erage potentials, or hi boxes, influences the quantum properties. In the
spirit of snowing developing areas, rather than discussing in detail works
reaching final conclusions, let us give a couple of illustrative examples.
Consider the motion of n electrons in a billiarc (a small metallic par-
ticle!) under the influence of a constant magnetic field B perpendicular
to the billiard, as we did in the previous Subsection. At zero tempera-
ture, the energy E of the system, when neglecting interactions among
the electrons, is given by
(6.16)
i=l
In (6.16), Ei(B) are the single-particle energies of the occupied states,

solutions of (6.14). Often, what is physically interesting is the sensitivity
of energies to changes of the parameters. Here, an interesting quantity
94 Oriol Bohigas
O. 20. B 40. O. 20. g 40.

Fig. 28. Magnetic field dependence of the spectrum of a particle in a billiard:
(a) circular billiard; (b) elliptic billiard. (Taken from [99]).
is the magnetic susceptibility x>
(6.17)
dB2
The question now is whether the value of x is affected by the regular or

chaotic nature of the classical motion of the electrons. This problem has
been 'investigated by Nakamura and Thomas [99]. They have considered
elliptic and circular billiards. As we have already seen, both billiards
give rise to regular motion. However, when the magnetic field is applied,
integrability is preserved for circular billiards., whereas for elliptic bil-
liards chaotic motion is generated [10O]. On Fig. 28 are reproduced the
single-particle energies as a function of the strength for both shapes. For
the circular shape, there is no level repulsion, reflecting the integrability
of the problem, whereas for the elliptic shape one can see many avoided
crossings, indicating that chaos is present in the underlying classical dy-
namics. On Fig. 29 the negative of x '^ displayed, as a function of B.
For the circular billiards, x is found to retain the essential features of
Landau diamagnetism in two-dimensional systems: x takes the largest
value in the vicinity of B = O and decreases monotonically with increas-
ing B. In contrast, for the elliptic billiards, x is greatly reduced close
to B = O compared to the circular case, and it increases on the average
with B. This increase is accompanied by large fluctuations which can
O.
Fig. 29. Negative of diamagnetic susceptibility x as a function of the magnetic

field strength B for n = 100 occupied levels. The upper part corresponds to a
circular billiard and the lower part to an elliptic billiard. (Taken from [99]).
be associated with avoided crossings (level repulsion). Anomalous dips

in x occur for B-values where the Fermi energy Ef, energy of the last
occupied single-particle state, lies at an avoided crossing.
One should however notice that these conclusions on the problem of
size and shape dependence of the free electron diamagnetism, problem
solved by Landau for infinite systems, are controversial, and that re-
cently Nmeth has questioned the role played by the integrability or non-
integrability of the previous model concerning the susceptibility [101 j.
Another interesting physical problem, related to particles hi a mag-
netic field, is the study of rotating objects. What happens to the motion
of the particles when, instead of applying a magnetic field, the billiard
96 Oriol Bohigaa
b)
1-00
C),
Fig. 30. Poincar surfaces of section showing the effect of rotations on billiards:
(a) elliptic billiard, without rotation (taken from [106]) ; (b) stadium billiard ro-
tating around its center (taken from [104]) ; (c) elliptic billiard rotating around
its center (taken from [104]).
rotates? This is not an academic problem for, as is well known, nuclei

exhibit rotational motion and definite nuclear states can be viewed as
corresponding to the motion of independent particles in a rotating mean
field [102]. That rotational motion and motion in a magnetic field are
closely related can be considered as a consequence of Larmor's theorem:
the motion of a system of charges in a weak external uniform magnetic
field coincides with its motion relative to a reference frame rotating with
a constant angular velocity. It should therefore be possible to identify
rotational effects analogous to susceptibility effects when dealing with
magnetic fields.
Several authors have investigated rotating billiards, classically as well
as quantum-mechanically [103,104,105]. In correspondence with the
magnetic field case, where the field generates chaotic regions for reg-
ular billiards (ellipse) and regular regions for chaotic billiards (see Sub-
section 6,2), rotations have a similar effect, as illustrated on Fig. 30.
When rotating, the (regular) elliptic billiard generates chaos whereas
the (chaotic) stadium billiard generates order.
To take into account the effect of the rotation quantum-mechanically,
(a)
-100-
10 15 20
OJ
210
20or\_
10
Fig. 31. Angular velocity dependence of the spectrum of a particle in an elliptic

billiard rotating around its center. The dashed line indicates the Fermi level
(n - 27), see Fig. 32. (b) shows details of (a). (Taken from [105]).
98 Oriol Bohigas
10,-
L
ai
Fig. 32. Moment of inertia as a function of the angular velocity for 27 particles
in an elliptic billiard rotating around its center. (Taken from [105]).
one must solve the time-independent Schrodinger equation by making

the replacement
H -+ H - wLr , (6.18)
where u is the angular velocity and L1 is the z-component of the orbital

angular momentum (rotation around the z-axis) [102]. The replacement
(6.18) accounts for the Coriolis and centrifugal forces. The effect of the
rotation on the spectrum of the elliptic billiard is illustrated on Fig. 31.
Again, one can see avoided crossings indicating the presence of chaos.
The quantity corresponding to the magnetic susceptibility x is now the
moment of inertia J,
The angular velocity dependence of the moment of inertia is illustrated

on Fig. 32. Fast variations of the moment of inertia can be associated
with values of the angular velocity such that the Fermi energy is close
to an avoided crossing. This phenomenon is probably related to nuclear
backbending [107], though more realistic studies should be performed to
be really conclusive.
Fig. 33. Poincar surfaces of section corresponding to two coupled anisotropic

quartic oscillators, see Eq. (2.20), for b = 7T/4 and = -0.35. Left: C1 = O
section. Right: q2 = O section. (Taken from [24,109]).
6.4. Quantum-mechanical aspects of mixed phase space systems
With the exception of the preceding Subsection, most of the emphasis

until now has been given to systems which are either fully regular or
fully chaotic. However, generic systems, like most atoms and molecules,
are neither fully chaotic nor fully regular but rather the phase space
has a mixed character. In this Subsection, to incorporate mixed phase
space effects, some of the previous discussions on spectral fluctuations
will be generalised and new aspects of the chaotic dynamics manifesting
themselves in the quantum regime will be mentioned.
For illustrative purposes it is convenient to discuss the system of two
coupled anisotropic quartic oscillators, see Eq. (2.20). When the cou-
pling parameter is varied from O to 1, the motion changes from
integrable ( = O , uncoupled oscillators) towards practically completely
chaotic ( = l). For intermediate values of C, the motion has mixed
phase space. Let us consider the values b = Tr/4, = 0.35 (see Fig. 33).
The Poincar sections display a phase space structure typical of the
mixed systems we are interested in: a dominant "chaotic sea", repre-
senting here 88% of the energy-shell volume, in which are embedded
KAM islands where the motion is mostly regular. As first pointed out
by Percival [108], to the sharply contrasted regular /chaotic classical mo-
tions should in principle correspond two distinct sorts of quantum levels.
With the regular islands, for which the Einstein-Brillouin-Keller (EBK)
100 Oriol Bohigas
quantisation scheme should apply, should be associated a 'regular spec-

trum', and the remaining levels should be called 'chaotic levels' since
they are associated with the chaotic part of phase space. One good thing
with the quartic oscillators is that, because of the discrete symmetries
and the dynamics of the system, this separation can be explicitly per-
formed for a wide range of the parameter [24,109]. This stems from the
fact that for < * ^. 0.20, all invariant tori admit at least one sym-
metric but distinct duplicate under the action of PI, PZ, or PiPj. Here
Pj are the parity operators: 9, > g,- (t = 1,2). Since the EBK quanti-
sation rules, if verified for an invariant torus, hold automatically for its
twin symmetric brother, all regular levels appear as quasi-degenerate
doublets, belonging to different representations which are determined by
the remaining symmetry of the tori. This, together with the fact that
the regular levels appear locally in nearly uniform (harmonic-oscillator-
like) sequences, allow to filter out the regular levels from the rest of the
spectrum.
For the values of the parameters for which this procedure has been
explicitly done, the separation postulated by Percival works quite well.
All but a few percent of the levels can be assigned to either the regu-
lar or the chaotic class. In a trace formula formalism, this implies the
non-trivial result that there is nearly no interference between the sum
taken over closed orbits inside the regular regions and the one over the
chaotic closed orbits. Sume features however complicate the simplest
semiclassical interpretation. Indeed, some levels, corresponding to the
regular/chaotic interface, are rather of an intermediate status. More-
over, an island may appear quantum mechanically enlarged or reduced
depending on the nature of the interface.
Having disposed of the regular levels, let us focus on the remaining
chaotic ones. We shall see that certain features of the classical motion,
namely transport properties, have counterparts in the quantum spectrum
[24,11O]. As we know, the chaotic spectrum is naturally best interpreted
in terms of statistical quantities modelled by ensembles of random ma-
trices. The universality conjecture discussed in Section 5, viewed from
the correspondence principle, implies that RMT applies to a spectrum
associated with a chaotic region, independently of any peculiar feature
of the classical chaotic regime. It is not true however that even the most
chaotic system has no internal structure. For instance, if one considers
two strongly chaotic billiards placed side by side and connected through
a small hole, neither the classical system nor the quantum one can be
considered as featureless , although the Lyapunov exponents can be
very high. In the limit where the hole is closed, the two billiards act
independently and the resulting spectral fluctuations can be interpreted
by superposing two GOE spectra, weighted by the relative density of
levels (i.e. at the semiclassical approximation the energy-shell Liouville
measure) of each billiard. The independent superposition of two GOE
spectra is of course equivalent to an ensemble of block diagonal GOE
matrices, each block having a size proportional to the relative density
of levels of the billiard it is associated to. At the opposite, if the hole
is very large, a full GOE matrix will be appropriate. In-between, it is
natural to look for a transition ensemble [94], namely an ensemble of
block diagonal GOE matrices connected via a weak GOE-like coupling
provided by the off-diagonal blocks.
As shown by French et al. [94], once the relative dimensionalities
of the diagonal blocks are fixed (here by the energy-shell volume of
each billiard), the only parameter entering the transition theory is the
transition parameter, given by (6.7), where the numerator is the root
mean square of the off-diagonal block elements and D the total mean
spacing. In other words, once A is fixed, all the statistical quantities
are determined, being those of two uncoupled GOE for A = O and of a
full GOE in the limit A > oo. When a chaotic region is separated by
two slightly connected regions, it can be shown [110] that the transition
parameter A is related to the flux <p (i.e. the energy-shell volume per
unit time) exchanged from one region to the other. Namely, one has to
leading order in k
where d is the number of degrees of freedom and // (J = 1, 2) the relative

phase space volume of region /. Here, as for the quartic oscillators,
<t--(E) increases with E, implying that as the energy goes to infinity,
the system eventually behaves as a single GOE. When more than two
slightly connected regions exist, these results generalise by introducing
a set of transition parameters A/j associated in the same manner with
the fluxes tpjj connecting region / to region J.
Returning now to the mixed systems, and in particular to the quartic
oscillators, the existence of partial barriers, and thus of slightly con-
nected regions, will be the typical rather than the exceptional situation.
As shown by MacKay et al. [ill], partial barriers can be formed either
by cantori or by the stable and unstable manifolds originating from hy-
perbolic closed orbits. For the (C, 6) = (-0.35, ?r/4) case of the quartic
oscillators, see (2.20), the stable and unstable manifolds play the ma-
102 Oriol Bohigaa
jor role, giving an effective separation of the chaotic phase space in five
slightly connected subrogions. By measuring the energy-shell volume of
these regions and the flux connecting them, it is possible to predict for
instance the number variance S 2 (r) and the two-point function Ri(f)
which are compared to the actual data on Fig. 34. As can be seen, the
statistical theory correctly reproduces the results of the actual compu-
tation.
Besides spectral statistics, there are some consequences of classical
transport having direct physical interest, such as the localisation of wave-
functions. Indeed, for small enough fluxes, the eigenfunctions will be
localised in one region. The square norm of the projection of a function
<&/ localised in region / onto the subspace associated with region J is
on the average
(6.2!)
valid for Xu <C 1. In Eq. (6.21), Pj is the projection operator onto

the J-th subspace, and the double bracket means local smoothing at
E. Eq. (6.21) implies in particular that a wave packet started in some
region remains mostly localised in it even for infinite time, instead of
spreading over the whole phase space as classically expected.
These two examples (spectral statistics and localisation of wavefunc-
tions) show that, as far as the correspondence principle is concerned,
there is more to be seen in the chaotic part of the phase space of a
mixed system (and presumably of a purely chaotic system as well) than
just its chaotic character. The universality conjecture appears therefore
rather as an, asymptotic result (limit of zero structure). For a large
class of purely chaotic systems it may be however that from a practical
point of view it remains a very good approximation, hardly distinguish-
able from the actual results. For the chaotic part of mixed systems, we
are rather inclined to think that this will rarely be the case.
Epilogue
The Random Matrix Theories we have discussed were invented by
Wigner in the SO's to describe properties of the compound nucleus res-
onances. These theories have permeated a considerable amount of at
first sight unrelated physical phenomena as well as of purely mathe-
matical objects. They have, in particular, led to discover new properties
0.0 1.0 3.0 40 5.0
R 2 (r)i
OO
Fig. 34. Results for the spectral fluctuations and comparison with different
models. Top: number variance S2(r). Bottom: the measure RZ(T) (see
Eq. (3.27)); notice the difference in abscissa, (a) one GOE; (b) five GOE blocks
weighted according to the fractions of total phase space volume, the blocks are
decoupled; (c) from the theory discussed in Subsection 6.4, namely like (b) but
within blocks coupled by the XH as given by the classical dynamics; (d) from
the 16000-th to the 22000-th eigenenergies of the coupled quartic oscillators,
case (f, 6) = (-0.35, JT/4) the regular leveb (~ 12%) have been subtracted
out; (e) from a Poissonian spectrum, for the sake of comparison. See text for
further explanation. (Taken from [24,110]).
104 Oriol Bofiigaa
Fig. 35. Picture illustrating the compound nucleus idea, as presented by N.

Bohr in 1936. In a neutron-nucleus collision the constituent nuclons are viewed
as billiard balls and the nuclear binding as a shallow basin. (Taken from [112]).
of "chaotic billiards". It is remarkable that Niels Bohr, who introduced

the idea of the compound nucleus in 1936, forced by the spectacular
sharpness of the neutron rapture resonances [112], did show in a lec-
ture presenting the new views of nuclear structure to the Chemical and
Physical Society of University College, London, a picture reproduced in
Fig. 35: a shallow basin with perfectly elastic balls, i.e., a billiard with
smooth walls!
Acknowledgements
My sincere gratitude goes to the organisers of the School, M.-J.
Giannoni and A. Voros, for inviting me to participate, for the stimu-
lating atmosphere they created at Les Houches and, last but not least,
for their kindness and patience. Part of the work reported here is the
outcome of collaborations with, in chronological order, R.U. Haq and A.
Pandey, M.-J. Giannoni and C. Schmit, and S. Tomsovic and D. Ullmo.
In preparing the manuscript I have benefited from help and discussions
with P. Leboeuf. A. Fercoq has done the typing with efficiency and care.
I would like to thank all of them.
A Nearest neighbour spacing distribution for Gaus-

sian ensembles (N = 2}
Orthogonal (ft = l) p(s) = -s exp(--s2) (A.I)
I
19 A
Unitary (ft = 2) P(B) = ^s2 exp(--s2) (A.2)
TP TT
218 64
Symplectic (ft = 4) p() = 5 <exp(- s2) (A.3)
B n-level cluster functions (large JV limit)

The expressions of the n-level cluster fonctions are reproduced for the
infinite-dimensional case ((3 = 1,2,4) [49,4]. Let
.M--:= (B.!)
DB(X) = tf(x) = -J^B(X) (B.2)
IB(X) = J* s(t) t (B.3)

JB(X) = Is(x) - E(Z) (LA)
where
O i=0 (B.5)
<0
The n-level cluster functions Yn(xi,..., Xn) are given by
Vn(X1, . . . , X n ) = tr X) k(in)<r(*) ^(î)l (B.6)

^ p
where
I0- - x, - x, (B.7)
and SP denotes a sum over the (n 1) ! distinct cyclic permutations of
the indices (1, 2, . . . , n) . In (B.6) ,CT(X)are 2 x 2 matrices which read as:
For the orthogonal ensemble (/? = l)
, . / s(x) DS(X
"U = ( J s ( X ) S(X
For the unitary ensemble (ft = 2)
For the symplectic ensemble (j3 = 4)

106 Oriol Bohigaa
C Two-level cluster function F2 and related quan-

tities (large N limit)
Let
Si() = f ^d, , Ci(x) = 7 + In. + f îîdy (c.i)

Jo y Jo y
where f is Euler's constant, 7 = 0.5772..., and s(x), e(x) as defined in
(B.I) and (B.5).
The two-level cluster function Y2
See Appendix B. Yj is an even function of its argument.
0 1 Vl \ I \ fo-l \ I \\
0 = 1 F 2 (X) = .'(X) - (S1(TTX) - TTE(X))
(C.2)
Small-a; behaviour
7 2 7 4 3 4
v I T T 3 T 4 ir> \
Large-x behaviour
yfW= * _!+!=! + ... (c.4)

(TTX)'' (Tri)4
/? = 2 Y2(x) = f!H^*\ . (C.5)
Small-x behaviour
Large-x behaviour
(7T/2)sin2rx l + cos22rx
Two-level form factor b(k)

6(fc) is an even function of its argument.
0<fc<l
^j fc >x
(C.9)
2
SmalW: behaviour 6(fc) = 1 - 2fc + 2fc + (C.10)
Large-fc behaviour b(k) = (l/12fc2) + (l/80fc4) + - (C.ll)
(C.12)
Small-L behaviour of R2[L)
ft = 2 (L) - (
^ -4 B2(L) = (163T4/2025)L6 + - - - (C.16)
Number variance S(L
0 =2 E| = ^?[ln(2rL)+-+l-cos(27rL)-Ci(27rL)l+L[l--Si(27rL)l
TT* TT
(C.17)
0=1 E^(L)- 2S^(L) + (Si(TrL) /Tr)2 -Si(7rL)/7T (C.18)
^=4 E2(L) = (1/2)E2(2L) + (Si(27rL)/27r)2 (C.19)
Large-L behaviour
/3-1 E? = -^[ba(27rL) + -y + 1 - 5r2/8] + 0(L-1) (C.20)
| = JL[ln(47rL) + 7 + 1 + 7r2/8] + 0(ZT1) (G.22)

108 Oriol Botiigas
Dyson-Mehta least square deviation A(L)

Largc-L behaviour
/3 = 1 A(L) = -j[ln(27rL) + -7 - 5/4 - jr2/8] + (L~l) (c-23)
0=2 A(L) = jlln^TrL) + -7 - 5/4] + O(L-1) (C.24)
/3 = 4 A(L) = -^T[ln(47rL) + -7 - 5/4 + 7r2/8] + O(L-1) (C.25)
D Superposition of several independent spectra

Let /, be the fraction of levels belonging to the t'-th component spectrum,
whose n-level cluster function is Y(i)in (XI, XZ, ..., In) (i = !,...,). The
n-level cluster function resulting from the superposition is [38]
t
f2
1^(Ux1JiXt, ..., /,xn) (D.I)
=i
with
= l. (D.2)

For the two-level form factor 6(/b),
For the number statistic S2(L) and the Dyson-Mehta least square devi-
ation A(L),
(D.4)
(D.5)
i
K /, = l/, from (D.I), (D.3), (D.4), (D.5),
Y*(*) = \ Yw(x/Q (D.6)
(D.7)
(D.8)
(D.9)
The nearest neighbour spacing distribution p ( s ) , in terms of P(J)(S),
is [4]
(0.10)
where
* (D.ll)
^w= n ()(/")
t
In particular, if the i-th component is GOE (approximated by the N = 2

value), one has (see A.l)
l - l f o ( a ) = e '/ (D.12)
%(s) = erfc (v/irs/2) , (D.13)
and the behaviour at the origin of p(s) resulting from the independent
superposition of I GOE-components is
p(0) ^ O for > 1 . (D.15)

From (D.9), for the particular case = 2, with component 1 being
Poissonian, P(I)(S) = exp( s), with fraction /, and component 2 being
GOE, approximated by the (JV = 2) value (see (A.I)), one obtains
p(s) = /zexp(-/S) erfc ( - fa) + (2//+ /) exp(-/5 -

(D.16)
where / + / = 1.
In (D.13) and (D.16), as usual,
erfc(i) = 1 - erf(i) = 1 - -= e^dt . (D.17)

y T
110 Orj'o/ Bohigaa
E GOE - GUE and GSE -> GUE transitions (large

N limit)
The n-level cluster functions F n (Z 1 ,..., xn; A) for the infinite-dimensional
case, are given by (B.6) with
D(x; A)
s(x)
where A is the transition parameter and s(z) as defined in (B.I) [91,4].

In (E.I), for GOE - GUE (A = O, GOE; A -> oo, GUE),
D(X; A) = -i r ksnkxe2*k*dk (E.2)

TT Jo
V 1 e-2A**3dA; ; (E.3)
' TT JT Jb
and, for GSE - GUE (A = O, GSE; A-KJO, GUE),
1 f* sin fci 2Aata
;; A) = -ir!êâd)fc (E.4)
TT Jo K
fc . (E.5)
TT Ji
For GOE -* GUE,
sin JTI \ a
vrz /
J_
Tz
rsê-^dt
JT Jo J, t
(E.6)
fc a
- -- i- [k sintee*f* /') dJbx
jr'a;2 Jo
f jCT ff J 7 =^- rî

L2 V2V5A/ Jo <
For i > 1,
(x, A)., /sin TZ\ 2 = z2 cos2 TTI 16Tr2A4 sin2

2 TTI+ .
(E .7)
> - - - -
Spacing distribution for 2 x 2 matrices [93,94]

H = S(w2) + iaA(t>2) ; (E.8)
for a O, one recovers GOE; for a = 1, one recovers GUE (see (6.1)-
(6.3)).
(E 9)
-
1 2 1 2 1 2
g = (8/Tr) ^t/ [(I - a )- /2arctan{(l - a ) / /"} + ] (E-IO)
References
[1| E.H. Lieb and D.C. Mattis, Mathematical Physics in One Dimension,
Academic Press, 1966
[2] C.E. Porter, Statistical Theories of Spectra: Fluctuations, Academic
Press, 1965
[3] T.A. Brody, J. Flores, J.B. French, P.A. MeUo, A. Pandey, S.S.M. Wong,
Rev. Mod. Phys. 53 (1983) 385
[4] M.L. Mehta, Random Matrices and the Statistical Theory of Energy
Levels, Academic Press, 1967; new revised and enlarged edition, 1990
[5] G.M. Zaslavsky, Phys. Reports 80 (1981) 157
[6] M.V. Berry, in Chaotic Behaviour of Deterministic Systems (ed. by
R.H.G. Helleman, G. loosa and R. Stora), Les Houches Summer School
Lectures XXXVI, 1981, North-Holland, Amsterdam, 1983, p.171
[7] G. Casati, d., Chaotic Behavior in Quantum Systems, Proceedings
of the Como Conference on Quantum Chaos, 1983, Plenum Press, New
York, 1985
[8] O. Bohigas and M.- J. Giannoni, in Mathematical and Computational
Methods in Nuclear Physics (ed. by J.S. Dehesa, J.M.G. Gomez and A.
Polls), Proceedings of the 6th Granada Workshop, Spain, 1983, Lecture
Notes in Physics 209 (1984) 1, Springer Verlag
[9] T.H. Seligman and H. Nishioka, eds.. Quantum Chaos and Statistical
Nuclear Physics, Proceedings of the 2nd Conference on Quantum Chaos
and 4th Colloquium on Statistical Nuclear Physics, Cuernavaca, 1986,
Lecture Notes in Physics 263 (1986), Springer Verlag
112 Oriol Bohigaa
[10] B. Eckhardt, Phys. Reports 163 (1988) 205
[Uj O. Bohigas and H.A. Weidenmiiller, Ann. Rev. NUL - . Part. Sci. 38 (1988)
421
[12] M V. Berry, these proceedings
[13] B.V. Chirikov, these proceedings
[14] D. Delande, these proceedings
[15] B. Muhlschlegel, these proceedings
[16] C. Schmit, these proceedings
[17] U. Smilansky, these proceedings

[18] A. Ozorio de Almeida, Hamiltoniun Systems: Chaos and Quantiza-
tion, Cambridge University Press, 1988
[19] M. Tabor, Chaos and Integrability in Nonlinear Dynamics, An In-

troduction, Wiley, New York, 1989
[20] M.C. Gutzwiller, Chaos in Classical and Quantum Mechanics,

Springer Verlag, New York, 1990
[21] M. Kac, Statistical Independence in Probability, Analysis and Num-

ber Theory, Wiley, 1959; Enigmas of Chance, University of California
Press, 1987
[22] N.L. Balazs and B.K. Jennings, Phys. Reports 104 (1984) 347
[23] J.E. Moyal, Proc. Camb. Phil. Soc. 45 (1949) 99
[24] O. Bohigas, S. Tomsovic and D. Ullmo, "Some New Aspects of the Quan-
tum Mechanics of Mixed Phase Space Systems", preprint IPNO Orsay,
1990
[25] H. Baltes and E.R. HiIf, Spectra of Finite Systems, Bibliographisches

Institut, Mannheim, 1976
[26] M. Kac, Am. Math. Monthly 73 (196R) 1
[27] B. Simon, Ann. Phys. (N.Y.) 146 (1983) 209; J. Fund. Anal. 53 (1983)
84
[28] M. Sieber and F. Steiner, Physica D44 (1990) 248; Phya. Lett. A148
(1990) 415; F. Steiner and J.P. TriUenberg, J. Math. Phya. 31 (1990)
1670
[29] S. Tomaovic, "The Level Density of a Special Quartic Oscillator", preprint

University of Washington, Seattle, 1990, submitted to J. Phys. A, and
private communication
[30] Y. Colin de Verdire, these proceedings
[31] N.L. Balazs and A. Voros, Phys. Reports 143 (1986) 109
[32] P. Dahlqvist and G. Russberg, Phys. Rev. Lett. 65 (1990) 2837
[33] O. Bohigas, M.-J. Giannoni and C. Schmit, J. Physique Lett. 45 (1984)

L-1015
[34] O. Bohigas, M.-J. Giannoni and C. Schmit, in [9] p.18
[35] D. Chandler, Introduction to Modern Statistical Mechanics, Oxford

University Press, 1987
[36] O. Bohigas, R.U. Haq and A. Pandey, Phys. Rev. Lett. 54 (1985) 1645
[37] F.J. Dyson and M.L. Mehta, J. Math. Phys. 4 (1963) 701, reprinted in
[2]
[38] A. Pandey, Ann. Phys. (N.Y.) 119 (1979) 170
[39] C. Schmit, private communication
[40] A. Bohr and B.R. Mottelson, Nuclear Structure Vol. I, Benjamin, New
York, 1969, p. 178
[41] F.J. Dyson, J. Math. Phys. 3 (1962) 140, reprinted in [2]
[42] F.J Dyson, J. Math. Phys. 3 (1962) 1199, reprinted in [2]
[43] A.T. Bharucha-Reid and M. Sambandham, Random Polynomials, Aca-

demic Press, 1986
[44] M. Kac, Probability and Related Topics in Physical Sciences, 2nd

printing by American Mathematical Society, Providence, 1976
114 Oriol Bohigas
[45] M.L. Mehta, Matrix Theory. Selected Topics and Useful Results,
Hindustan Publishing Corporation and Les Editions de Physique, Lea
Ulis, France, 1989
[46] J. Ginibre, J. Math. Phya. 6 (1965) 440
[47] H.J. Sommera, A. Crisanti, H. Sompolinsky and Y. Stein, Phys. Rev.

Lett. 60 (1988) 1895
[48] E.P. Wigner, SIAM Review 9 (1967) 1
[49] F.J. Dyson, Commun. Math. Phya. 19 (1970) 235; M.L. Mehta, Commun.
Math. Phys. 20 (1971) 245
[50] M.L. Mehta and F.J. Dyson, J. Math. Phys. 4 (1963) 713, reprinted in
[2]
[51] B. Jancovici, Phya. Rev. Lett. 46 (1981) 386; A. Alastuey and B. Jan-
covici, J. Physique 42 (1981) 1
[52] R. Balian, Nuov. Cim. B57 (1968) 183
[53] O. Bohigas, R.U. Haq and A. Pandey, in Nuclear Data for Science and
Technology (ed. by K.H. Bockhoff), Reidel, Dordrecht 1983, p.809
[54] R.U. Haq, A. Pandey and O. Bohigas, Phys. Rev. Lett. 48 (1982) 1086
[55] L. Leviandier, M. Lombardi, R. Jost and J.P. Pique, Phys. Rev. Lett.
56 (1986) 2449
[56] E.P. Wigner, in Symmetries and Reflections, ed. by E.P. Wigner, Ox

Bow, Woodbridge, Connecticut 1979, p.203
[57] H.S. Camarda and P.D. Georgopulos, Phys. Rev. Lett. 50 (1983) 492
[58] Th. Zimmermann, H. Kppel, L.S. Cederbaum, G. Persch and W.

Demtroder, Phys. Rev. Lett. 61 (1988) 3
[59] M.V. Berry and M. Tabor, Proc. R. Soc. Lond. A356 (1977) 375
[60] G.M. Zaslavskii, JETP 46 (1977) 1094
[61] V.I. Arnold, Mathematical Methods of Classical Mechanics, Springer

Verlag, New York, 1978
[62] O. Bohigaa, M.-J. Giannoni and C. Schmit, Phys. Rev. Lett. 52 (1984)
1
[63] M. Robnik and M.V. Berry, J. Phys. A19 (1986) 669; M. Robnik, in [9]
p.120
[64] R. Scharf, B. Dietz, M. Kus, F. Haake and M.V. Berry, Europhya. Lett.
5 (1988) 383
[65] E. Caurier and B. Grammaticos, Phys. Lett. A136 (1989) 387
[66] F.J. Dyson, J. Math. Phys. 3 (1962) 166, reprinted in [2]
[67] A. Pandey, O. Bohigas and M.-J. Giannoni, J. Phys. A22 (1989) 4083
[68] L. Glass, A.L. Goldberger and J. Belair, Am. J. Physiol. 251 (1986)
H841
[69] H. Wu, M. Vallires, D.H. Feng and D.W.L. Sprung, Phys. Rev. A42
(1990) 1027
[70] O. Bohigas, M.-J. Giannoni, C. Schmit, S. Tomsovic and D. Ullmo, Com-

ments At. MoI. Phys. 25 (1990) 31
[71] M.-J. Giannoni and D. Ullmo, Physica D41 (1990) 371; and in prepara-
tion
[72] A.M. Odlyzko, Math, of Comp. 48 (1987) 273
[73] A.M. Odlyzko, "The 1020-th Zero of the Riemann Zeta Function and 70
Million of its Neighbors", preprint AT&T Bell Laboratories, Murray Hill,
1988, and in preparation
[74] C. Schmit, M.-J. Giannoni and O. Bohigas, in Phase Space Approach

to Nuclear Dynamics (ed. by M. Di Toro, W. Norenberg, M. Rosina
and S. Stringari), Proceedings of Topical Meeting, Trieste, 1985, World
Scientific, Singapore, 1986, p.530
[75] M.C. Gutzwiller, these proceedings
[76] S.C. Creagh and R.G. Littlejohn, private communication
[77] B. Eckhardt, G. Hose and E. Pollack, Phys. Rev. A39 (1989) 3776
[78] M.V. Berry, Proc. R. Soc. Lond. A400 (1985) 229; Proc. R. Soc. Lond.
A413 (1987) 183
116 Oriol Bohigaa
[79] M.V. Berry, in Stochastic Processes in Classical and Quantum Sys-

tems (ed. by S. Albeverio, G. Casati and D. Merlini), Proceedings of the
1st Ascona-Como International Conference, Ascona, Switzerland, 1985,
Lecture Notes in Physics 262 (1986) 47, Springer Verlag; Nonlinearity 1
(1988) 399
[80] T.H. Seligman, J.J.M. Verbaarschot and M.R. Zirnbauer, Phys. Rev.
Lett. 53 (1984) 215
[81) G. Casati, B.V. Chirikov and I. Guarneri, Phys. Rev. Lett. 54 (1985)
1350
[82] J.H. Hannay and A.M. Ozorio de Almeida, J. Phys. A17 (1984) 3429
[83] H.-J. Stockmann and J. Stein, Phys. Rev. Lett. 64 (1990) 2215; J. Stein,
Diplornarbeit, Universitt Marburg, 1990; and private communication
[84] V.K. Ignatovich, The Physics of Ultracold Neutrons, Clarendon Press,

Oxford, 1990
[85] P. Ageron, W. Mampe, J.C. Bates and J.M. Pendlebury, Nucl. Inst. Meth.
A249 (1986) 261
[86] P. El-Muzeini, H.J. Roth, W. Mampe, D. Dubbers, O. Bohigas and P.

Leboeuf, in preparation
[87] R.L. Weaver, J. Acoust. Soc. Am. 85 (1989) 1005
[88] P.J. Richens and M.V. Berry, Physica 2D (1981) 495
[89] T. Cheon and T.D. Cohen, Phys. Rev. Lett. 62 (1989) 2769
[90] O. Bohigas, O. Legrand, C. Schmit and D. Sornette, "Comment on spec-

tral statistics in Elastodynamics", J. Acoust. Soc. Am., in press
[91] A. Pandey and M.L. Mehta, Commun. Math. Phys. 87 (1983) 449; M.L.
Mehta and A. Pandey, J. Phys. A16 (1983) 2655 and L-601
[92] A. Pandey, Ann. Phya. (N.Y.) 134 (1981) 110
[93] J.B. French and V.K.B. Kota, Ann. Rev. Nucl. Part. Sci. 32 (1982) 35;
J.B. French, V.K.B. Kota, A. Pandey and S. Tomsovic, Phys. Rev. Lett.
54 (1985) 2313
[94] J B French, V.K.B. Kota, A. Pandey and S. Tomsovic, Ann. Phya. (N.Y.)
181 (1988) 198
(95) T.H. Seiigman and J.J.M. Verbaarschot, Phys. Lett. 108A (1985) 183
[96] O. Bohigas, M.-J. Giannoni, A. Ozorio de Almeida and C. Schtnit, Ia

preparation
[97] S. Tomsovic, private communication
[98J M.V. Berry and M. Robnik, J. Phys. A19 (1986) 649
[99] K. Nakamura and H. Thomas, Phys. Rev. Lett. 61 (1988) 247
[100] M. Robnik and M.V. Berry, J. Phys. A18 (1985) 1361; M. Robnik, in
Nonlinear Phenomena and Chaos (ed. by S. Sarkar), Adam Hilger,
Bristol, 1986
[101] R. Nmeth, Zeit. Phys. B81 (1990) 89
[102] A. Bohr and B R. Mottelson, Nuclear Structure, Vol. II, Benjamin,

London, 1975
[103] D.B. Fairlie and O.K. Siegwart, J. Phya. A21 (1988) 1157; D.K. Sieg-
wart, J. Phys. A22 (1989) 3537
[104] H. Frisk and R. Arvieu, J. Phys. A22 (1989) 1765
[105] A.J.S. Traiber, A.J. Fendrik and M. Bernath, J. Phys. A23 (1990) L305
[106] M.V. Berry, Eur. J. Phys. 2 (1981) 91
[107] P. Ring and P. Schuck, The Nuclear Many-Body P:,)blem, Springer-

Verlag, New York, 1980
[108] LC. Percival, J. Phys. B6 (1973) L229
[109] O. Bohigas, S. Tomsovic and D. UUmo, Phys. Rev. Lett. 64 (1990) 1479
[110] O. Bohigas, S. Tomsovic and D. Ullmo, Phys. Rev. Lett. 65 (1990) 5
[111] R.S. MacKay, J.D. Meiss and I.C. Percival, Physica 13D (1984); 27D
(1987) 1
[112] N. Bohr, Nature 137 (1936) 344; 351

Random Matrix Theories and Chaotic Dynamics: Oriol Bohigas

Uploaded by

Copyright:

Available Formats

Random Matrix Theories and Chaotic Dynamics: Oriol Bohigas

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Random Matrix Theories and Chaotic Dynamics: Oriol Bohigas

Uploaded by

Copyright:

Available Formats

RANDOM MATRIX THEORIES

AND CHAOTIC DYNAMICS

Division de Physique Thorique1

M.-J. Giannoni, A. VQTOB and J. Zinn-Juatin, eda.

two very simple examples.

Example 1. Consider the transcendental number TT. Its decimal (or

Any regularity in this expansion is hardly visible. Change just a little,

Example 2 [21]. n real numbers W1, . . . , wn are called linearly indepen-

In pi, InP1, ... , In Pn , (1.4)

of logarithms of primes p,- (pt = 2, pi = 3, . . .). In this case linear

I COSW1E < Ct1, . . . , C O S W n E < On) = JJ (iR{t

In (1.6), MR{A} is the relative measure of a set A of real numbers, defined

simultaneously hold is equal to the product of the fractions of time

Um /Ia(A n (O 11 O 2 )) = -L= Pe-'/Mz , (1.10)

to the normal frequencies of a vibrating membrane, or the eigenfrequen-

2. The smooth part of the spectrum

in d dimensions. The spectrum is determined by the eigenvalue equation

Let the spectrum of H be discrete, let

EI < Et < < < (2.3)

be the eigenvalues of (2.2) and let us introduce the staircase or counting

What we want is to extract from N-(E) its smoothed or average part;

N(E) =< N(E) > +NUX(E), (2.6)

the average level density < p > being given by

If the potential V(q) is homogeneous of degree p,

V(aq) = aV(q), (2.9)

one can easily see from (2.8) that

Emphasis will be often put on homogeneous potentials because the clas-

which is consistent with (2.10) for /i - oo. For a plane two-dimensional

where 5 is the area of D ; or, using units h - 2m = 1 (E -+ (2m/h2)E) ,

It is unaltered if one takes Neumann instead of Dirichlet boundary con-

A(q,p) = [A|.(q,p) = / dx e-**/* < q + ||q - | > (2.14)

[AB]w(q,p) = [Al w (q l ,p')e lA ' lk [J3] w (q ,p ) (2.17)

=Vq-Vp- -Vq-Vp. - (2.18)

One then considers the Laplace transform of the density of states

Z(P) = /J0 e-'*p(E) (LB = tr (e^) = -^ J dqdp [e-"ff]w(qi p)

3=(p\+Pl) + ti + l>ti + ltqi$ (2-20)

<N(E)>~ [SE^tVE+ K] (2.22)

D = {(ft, C2) I 9i > O n fe > O n qiqi < 1} . (2.24)

Through the hyperbolic shaped channels (horns), the particle in such a

a method of calculating the leading behaviour, first derived by Simon

B= 2^(P + P:) + Will (2-27)

which appears when studying impurity levels in a semiconductor, and

where v labels a symmetry class. This can be done in some cases in a

with bc = b + 62V2, see (2.26).

[*M1, P) = ** 6(9i)6(pt) ('" = 1, 2) (2-34)

One obtains, for instance, for the staircase function [24]

For homogeneous potentials of degree ft, the semiclassical expansion

namely a constant average density equal to unity, independent of w. The

0.0 5 O 10.0 15.0 20.0

n+1 (Xi,... , x n +i) dxn+1 = (N -B) JZ n (X 1 ,..., Xn) . (3.4)

It is convenient to introduce the n-levcl cluster functions Yn which are

Y n (X 1 ,... , xn) = (-)"->i- I)Jn^(X,, with tin G,-) . (3.5)

therefore Y3(X1, X2, xs) = JZ5(X1, X2, xs)

as we did in (3.3). One of the goals in statistical mechanics is to com-

ff(ri,r,) = Pi(F1, F2)^p1(F1)P1(F2)] (3.12)

which corresponds to R2(xltx2) in (3.3). For an isotropic fluid

and the equation of state of the fluid is

where p is the pressure.

-Pj p n + i(ri,..., IWu)V 1 W(Fi n+1) drn+l

5?= (! + > i < j . (4.18)