Eugene P. Wignert: Vol. 9, No. 1, January, 1967
Eugene P. Wignert: Vol. 9, No. 1, January, 1967
Eugene P. Wignert: Vol. 9, No. 1, January, 1967
EUGENE P. WIGNERt
Introduction. It has been observed repeatedly that von iNeumann made im-
portant contributions to almost all parts of mathematics with the exception of
number theory. He had a particular interest in those parts of mathematics which
formed cornerstones of other, more empirical sciences, such as physics or eco-
nomics. A whole new discipline grew out of his theory of games, and it is hard to
conjure up a picture of modern United States industry without the computing
machines which he espoused. The subject about which I wish to talk to you today
is at the crossroads of two of von Neumann's principal interests: it deals with
matrices of very large dimensions in which he became interested in connection
with his development of computers, and it resembles statistical mechanics, to
which he contributed most among the physical theories. This closeness of my
subject to von Neumann's interests is also the reason for my choosing it for
today's discussion. This discussion will contain very little that is new. As for
earlier reviews, there are at least two very good ones: one by Charles Porter,
forming an introduction and summary to a collection of papers on the role of
random matrices in physics [1], the second a more elaborate one by M. L. Mehta,
based on his lectures at the Indian Institute of Technology in Kanpur [2].
There is another reason for my choice of subject. The theory of random
matrices, though initiated by mathematicians and in particular statisticians
[3], [4], [5], [6], [7], has made large strides in the hands of physicists. The names of
Mehta, Gaudin and Dyson come to one's mind most easily. Reading these papers
gave me much pleasure-they contain beautiful, though old-fashioned, mathe-
matics. I would like to share some of this pleasure with you. Second, however, a
number of problems has turned up, apparently too difficult for us amateur mathe-
maticians. I would like to share these problems with you also.
Origin of the problem. For reasons which I hope will become evident in the
course of the discussion, I will proceed in my review pretty much in the anti-
historic order. However, the reason for the interest of physicists in random
matrices should be stated first. This was articulated, most eloquently, by
Dyson [8]:
"Recent theoretical analyses have had impressive success in interpreting
the detailed structure of the low-lying excited states of complex nuclei.
Still, there must come a point beyond which such analyses of individual
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
* The seventh John von Neumann Lecture delivered at the 1966 SIAM Summer Meeting
on August 31, 1966, in New Brunswick, New Jersey. Received by the editors September 22,
1966.
t Department of Physics, Princeton University, Princeton, New Jersey.
1
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
2 EUGENE P. WIGNER
I rt+T
(1) lim - f(qilr), q2(T), pl(T), P2(T), * * *) dr.
T??C Tt
Since the coordinates and momenta are, by Newton's equations of motion, com-
pletely determined as functions of time by their initial values, the averaging
process is an entirely definite one. That, apart from rare exceptions, the average
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 3
exists, that it is a function only of the constants of motion, such as energy, but
independent of the other initial conditions, is a theorem first proved by von
Neumann, Koopman and G. D. Birkhoff [9], [10], [11], [12], [13]. What I wish
to emphasize, however, is only that the averaging process is entirely definite;
it is a time average, and equal time intervals have equal weights.
In contrast to this, the averaging process in the physical applications of the
theory of random processes is not defined. One again deals with a specific system,
with its proper (though in many cases unknown) Hamiltonian, yet pretends
that one deals with a multitude of systems, all with their own Hamiltonians,
and averages over the properties of these systems. Evidently, such a procedure
can be meaningful only if it turns out that the properties in which one is inter-
ested are the same for the vast majority of the admissible Hamiltonians. The
first question, then, is what are the admissible Hamiltonians, and what is the
proper measure in the ensemble of these Hamiltonians. The second question
is, of course, whether, given the ensemble of admissible Hamiltonians with a
proper measure, the properties in which we are interested are common for the
vast majority of them.
The experimental situation which prompted the interest of physicists in
random matrices is illustrated in Figs. 1 and 2. The first of these shows the
situation which did not prompt it. This is a level scheme of three nuclei with
mass number 10. The horizontal lines represent stationary states, the vertical
position is the characteristic value of the Hamilton operator to which the charac-
teristic vector belongs. It is the energy of the stationary state. As the figure
shows, some of the energy values are common to all three nuclei; some are
present only in the Hamiltonian of one of the nuclei, that at the center. J is
the angular momentum quantum number of the state. The diagram of Fig. 1
was obtained experimentally; a physicist interested in these matters knows
most of the numbers of this table, and their relations to each other, by heart.
The diagram, to repeat it, shows the low, i.e., small, characteristic values of
the three Hamiltonians, those of Be'0, B'0 and C'0. The energy difference between
the two lowest states of B10 is 0.717 Mev.
Fig. 2 is the level diagram of another nucleus, U239, in the energy interval
J J
2+ 2- 2+
I+
2+
1+
0+_____
I+
Belo 3+ C 10
FIO
FIG. 1
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
4 EUGENE P. WIGNEli
J
+2
J. L. ROSEN
+ J. S. DESuARDINS
/2 J. RAINWATER
W. W. HAVENS
'2
/2
/2
I/2+
1/2+
FIG. 2
between 4.7834 Mev from the lowest level and 4.7835 Mev therefrom. Some
physicists, myself included, know a few of these energy values by heart because
they happen to play an important role in nuclear chain reactors. However, no
one is familiar with the levels in the next interval of the same width. As Dyson
said in the passage quoted above, only the statistical properties of these levels
are of interest.
What are these statistical properties? First, one would like to know how many
energy levels there are per unit energy interval, that is, the density of the char-
acteristic values of the Hamiltonian. Second, one would like to be able to de-
scribe what might be called subtleties of the arrangement of the characteristic
values: the probability of a given distance to the nearest neighbor, the prob-
ability of two consecutive distances to assume definite values, and so on. Third,
though this is not suggested by the diagram, one would like to obtain the prob-
ability of a transition rate to assume a definite value. The most important
transition is the emission of a neutron by the nucleus, and it is on this that
experimental data are available most abundantly. Altogether, the choice of
questions is strongly influenced by the possibility of experimental study-a
situation not unusual in physics. However, at least the first two questions, the
average spacing of the characteristic values as function of the characteristic
value itself and the distribution of the spacings around their average, are ques-
tions which everyone would naturally ask. The three questions now will be
considered in succession.
The density of characteristic values. It must be admitted at the outset that
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES '5
it is in this case that the independence of the result from the ensemble of Hamil-
tonians, and from the measure chosen therefor, is most nearly demonstrated, and
that it is also the case in which the conclusion is least satisfactory, that is, agrees
least with experimental observation.
It seems to be firmly established that the Hamiltonians of physics are, in the
usual coordinate systems in Hilbert space, real. Hence, the admissible Hamil-
tonians are real symmetric matrices. An N-dimensional real symmetric matrix
can be characterized by N(N - 1)/2 real numbers Hik with i < k, and the
measure in ensemble space is, therefore, a positive real-valued function of these
variables,
which gives the number of matrices in the ensemble, of which the matrix ele-
ments are within the unit interval around the corresponding argument of P. This
P function, therefore, defines the ensemble, and every positive-real-valued P func-
tion defines an ensemble. Needless to say, the problem of spectral density has
not been solved for the general ensemble, characterized by an arbitrary P.
Rather, in every case the independence of the distribution of some of the vari-
ables of P is assumed. If these variables are the matrix elements Hi1, H12,
H13, * of H themselves, P assumes the form
In this case, and if the average values of all Hik are zero, their second moments
equal
and all higher moments exist, the density of the characteristic values for very
large N is given by the so-called semicircle law [14], [15],
This distribution is very different from that of the real roots of an algebraic
equation of order N. The ensemble in this case is obtained by considering the
coefficients to be components of a vector of definite length which has equal
probabilities for directions within equal solid angles [16]. Fig. 3 is a histogram
of o(X), due to N. Rosenzweig, obtained by diagonalizing 20 by 20 matrices,
selected at random from an ensemble which I will discuss later. Not very sur-
prisingly, the distribution approached a semiellipse-semicircle is actually a
misnomer; the two axes do not even have the same dimension.
What is distressing about this distribution is that it shows no similarity to
the observed distribution in spectra. The behavior at large positive X is not
relevant-what is known and what could be hoped to be reproduced by the
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
6 EUGENE P. WIGNER
20- 120
100 X_100
80 so
60 -6
40 -4
20 20
FIG. 3
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOMI MATRICES 7
By a proper choice of this function, any desired density distribution of the levels
can be obtained.
In order to reduce the freedom in the choice of the ensemble, one may recall
that the Hamiltonian operator is unbounded above but bounded from below.
This is the second characteristic I referred to. If the spectrum of the Hamil-
tonian were not bounded from below, there would be no lowest characteristic
value. It should not matter much where the lower bound is; if it is assumed at
zero, one is tempted to substitute
The last statement means that there are no statistical correlations between the
matrix elements of A and that they all have the same Gaussian distribution.
However, the ma-trix elements of H are not statistically independent in this
case. Unfortunately, it turns out that the characteristic values of this ensemble,
all positive, are distributed according to a quarter-circle law so that the density
is quite large at the lower bound. I mention this unsuccessful attempt because
the use of the aforementioned distribution is suggestive and because I do not
recall having read an evaluation thereof before. Bronk has considered a similar
problem and obtained the density of levels as function of energy-it was grossly
unsatisfactory also [17].
This is a disappointing situation. However, perhaps we should not have
expected otherwise. Operators in Hilbert space have properties which no finite-
dimensional matrix has-in particular, a continuous spectrum. The true Hamil-
tonians all have a continuous spectrum, in addition to the discrete one, but the
Hamiltonians considered in the theory of virtual levels are modified and have
only a discrete spectrum. Somehow, these properties should play a role in the
ensemble one chooses: this should contain only matrices which converge, in the
1 Mehta [27] quotes S. N. Roy, Sankhya (Dec. 1943), but without further details.
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
8 EUGENE P. WIGNER
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 9
can all be obtained by integration with respect to the last variable from the
next one of the series and can also be obtained by repeated integration from
the joint distribution function (7) of all the characteristic values. However, as
we well know from the similar problem in classical statistical mechanics, this is
by no means an easy procedure. The most important among the spacing func-
tions is the first one, probably because it requires least data to check it. All
experimental and most theoretical work is directed toward the determination of
this function, T' .
It was assumed, from the beginning, that the aforementioned distributions
depend only on a crude characterization of the underlying matrix ensemble,
that it will be the same for all "reasonable" ensembles of real symmetric ma-
trices. It will also be the same for all reasonable ensembles of general hermitian
matrices, though the T for these will be different from the T for ensembles of
real symmetric matrices. When Dyson drew attention to quaternion matrices,
it was natural to assume the same for these. Similarly, it was assumed that, if
the actual spacings are measured in units of the average spacing, the distribu-
tion will be independent of the average spacing. In this sense, the functions
FtI, 'F2, ... for real symmetric matrices are definite functions, and the same
applies for these functions for hermitian matrices, and so on. One can probably
even define such functions for the real, complex, and quaternion matrix ensem-
bles without symmetry restrictions, except that the variables S should be com-
plex numbers.
The postulate of the existence of an average spacing already implies that
the energy range in which the average spacing of the levels is essentially con-
stant is much larger than the average spacing itself. The same applies to the
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
10 EUGENE P. WIGNER
reciprocal of the average spacing, the density. We have seen before that, in all
ensembles considered, the deiisity is proportional to the square root of the
dimension of the matrix wherever the density is not zero. Since the logarithmic
derivative of the spacing or of the density is inversely proportional to this
square root, the condition of the existence of an average spacing will be satisfied
for ensembles of matrices of high dimensionality wherever the density of the
levels is not zero. The assumption mentioned then implies also that the T are
independent of energy if their variables are measured in terms of the local
average spacing.
What is the evidence for the uniqueness of the functions T? Certainly, the
independence of T' from the matrix ensemble chosen has not been proved for
ensembles of as great generality as the semicircle law. The most general result,
outside of the realm of AWishart-like distributions, refers to the boundary condi-
tion of the truncated Hamiltonian and shows, indeed, that T' is invariant with
respect to this boundary condition even though all characteristic values are
changed by replacing the most usual boundary condition-zero derivative of
the wave function at the nuclear surface by postulating an arbitrary but fixed
real ratio be-tween the value and the derivative of the wave function [25]. This
theorem, as stated, applied to a one-dimensional problem but it can be gener-
alized to many dimensions by introducing a closed surface in Schr6dinger's
configuration space. On this surface, one can introduce a complete orthonormal
set xv and specify boundary conditions for each of them:
2 Incidentally, the exponents of I xi-xj i in the expressions for the joint distribution
of the roots of hermitian and quaternion matrices can be understood in the same way:
three real constraints must be satisfied for two roots of a general hermitian matrix to coiIn-
cide, and five such constraints in the case of a quaternion matrix.
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 11
VN, hence very small. In fact, Dyson [8] replaced the ensemble of real sym-
metric matrices with the ensemble of unitary symmetric matrices. In itself, this
appears harmless because a one-to-one correspondence can be established be-
tween real symmetric and unitary symmetric matrices S. However, he also
assumed that the eiisemble is invariant under the transformation S -*> WTSW
(the T denotes transpose), where W is any unitary matrix. This condition, if
translated into a measure for the real symmetric matrices, is not the most
natural one. Nevertheless, his joint distribution function for all roots is so
similar to that of the Wishart-like ensemble that it can surely substitute for
the latter.
Before embarking on a more serious discussion of the calculation of T1, let us
first make a crude guess. If the location of the roots were independent, the prob-
ability v that there is no root within the distance S from a given root would
obey the equation
for very small h, where c is the probability of a root in the unit interval. This
leads to the differential equation
(I Oa) dv = -cv,
and hence v exp (-cS) from which TJ = -dvIdS = c exp (-cS), i.e., the
simple exponential law would follow. Actually, we know that the probability of
a root right next to another one is proportional to the distance therefrom. This
suggests, instead of (10),
dv -S/
(Ila) - = -cSv V =e
dS
WrS -,,S2/4D2
(12) e1(S) = e
2D
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
12 EUGENE P. WIGNER
0.8 _
..I . I I .
p 0.6 -
0.4
0 2-
2 3
S/D
FIG. 4
80-
FIG. 5
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 13
120-
80
40_ I
FIG. 6
out two of the important calculations concerning level spacings in detail. The
present one concerns the joint distribution function of the roots of real symmetric
matrices of high dimensionality. The ensemble considered is the same in the
present and the next calculation: the probability function P of (2) is
(15)
(15) H2i + 2~~2
E Hi2k 22i#k 2=-1i2k=_
=- 2 +- Hi2k 2 X22
2 i i<k 2 i 2 i 3k 2 2
so that the exponent in (13) does not depend on R or the parameters p. Hence,
all we have to do is to calculate the Jacobian
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
14 EUGENE P. WIGNER
and are independent of the X. The last N(N - 1)/2 rows consist of
They are linear in the X. The Jacobian determinant is, therefore, a homogeneous
polynomial of the X, of order N(N - 1)/2. In order to prove that it has the
product form given before, we have to show only that it vanishes if any two X
coincide. We shall show that it vanishes for X, = X2. Before proceeding with
the calculation, we note that such vanishing is independent of the choice of the
parameters p in the region in which both sets are unique, because the transition
from one set of parameters to a new set merely multiplies J with the Jacobian
of the old parameters with respect to the new ones. Hence, we can choose for
one of the parameters, pi, the angle of the rotation in the plane of the first two
coordinates, such that multiplication of R with this rotation renders R12 = 0.
If we denote this rotation by R12(pl), the rotation R will assume the form
Integration over the p now gives the joint distributioni function of the roots
except for its normalization factor C:
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 15
obtain v(S), one has to carry out many integrations, and such multiple integra-
tions are quite difficult. It was, therefore, a considerable accomplishment on the
part of M\Iehta [27] and Gaudin [28] to calculate T'i. Actually, their calculation
proceeds by evaluating, first, the probability p(z9) that there is no root between
-z9 and z9. They assume that this is also the probability that there is no root
in an interval of the length 2z9 centered at random. The connection between TI
and p then can be obtained from the observation that an interval S between
two roots does not permit any interval of length 26 to be free of roots if S < 2z9;
it gives an interval S - 2z9 for the center of a root-free interval of length 2z9,
if S > 2z9. Since the number of intervals between S and S + dS in a long stretch
L is LoT41(S) dS, where o- is the density of the roots, we have
u(X) = 1 if X > o,
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
16 EUGENE P. WIGNER
the integration could be extended over all values of the variables. This would
be possible if Px were a symmetric function of its variables, but it is not.
Mehta and Gaudin overcome this difficulty by first integrating over all odd
variables in the proper interval. The first such integration, over Xi, must be
extended from - oo to X2 , and if we write
it simply replaces the qj(X1) of the first column by Q,(X2). Integration over X3
will replace the q,(X3) of the third column by Q (X4) - Q (X2)-the limits of
integration are X2 and X4. Adding the first column to this gives Q,(X4). Simi-
larly, the (2j - 1)th column will change, as a result of the integration over X2j-1,
into a column of Q (X2j). One avoids some unessential complications by assum-
ing that the dimension of the original matrix, N = 2m, is even. Then, the inte-
grand becomes, after integration over Xi, X3 , * 2m-1 ,
Q2m-1 (X2) q2m-1 (X2) Q2n-1 (X4) q2m-1 (X4) ... q2m-1 (X2m)
The integrand is now symmetric in the remaining variables and the integration
can be extended over all of them from - o to co if one also divides by m !
Let us now look at the integral of the subdeterminants of the first two columns:
If , and v are both odd, Q, and Qv will be even, and since q, and qv are odd,
I, will vanish. The same is true, however, also if 4 and v are both even: in this
case we can write
Ql(X) = i, + Q,O(X),
(27a) qi, = | (X) dX, QA0 (X) = f (X) dX'.
00~~~~~~~~~~~~
QAo and the similar Qvo are odd. Hence, for even,u and v,
00O 00
The last integral vanishes because the Q therein are odd, the q even. The first
integral is iy2iv - i2i = 0. Hence, I, is different from zero only if ,u is even, v
odd, or conversely. Since, furthermore, I, = -IV, we can restrict ourselves
to the case that the first index is odd, the second even. We note, finally, that in
the only interesting case, in which one of the indices is even, the other odd, the
two terms for I in (27) are equal: since qA(X) = Q,,'(X), we have
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 17
We shall carry the calculation one step further, restricting ourselves from now
on explicitly to the u(X) of (23a). For this purpose, we have to make a choice
of the polynomials a,(X) which appear in (24a). The simplest choice is
ao(X) = 1,
Then
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
18 EUGENE P. WIGNER
(, odd, v even), except that I>o = 26,1 o . This last integral can be decomposed
into one, Io ,from 0 to oo from which the integral Ilv from 0 to v~ must be sub-
tracted: I,> = IoV- I',. The former integral is a well-known one:
2 -21I2 0 0 0
112 1/2
0 4 -6 0 ... 0
II1%1 = Ci 0 0 81/2 10112 *.. o
0 0 0 0 * (4m 4)112
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 19
as long as T '- N-112. Their aggregate effect is, of course, large nevertheless.
Mehta and Gaudin evaluate the determinant by first subtracting suitable
multiples of every column from the next one so as to render 1,I diagonal,
then multiplying the rows or columns with the corresponding factors of C1 so
that 11 I', 11 is replaced by the unit matrix. These operations also alter 11 I',
However, Gaudin succeeded in diagonalizing the modified P, If its charac-
teristic values are denoted by qi, the value of the determinant (30) becomes
This was evaluated numerically and TJ obtained by (22a). This is the origin
of the curve in Fig. 4 which is so close to that given by (12).
I have perhaps described the calculation of T, in unnecessarily much detail.
However, I do consider it a major accomplishment. At the same time, it is
evident that there are several steps in the calculation which are not carried out
in rigorous detail. For this reason, at least, one would like to see a simpler der-
ivation.
We nlow turni to the last of our three items, the statistical distribution of the
matrix elements.
Matrix elements. Mathematically, the calculation of the statistics of the
matrix elements is the easiest of our three problems-in fact, it is an easy prob-
lem-but inhibitions prevented its discovery longer than now appears reasonable.
It was, therefore, a major breakthrough when Scott [30] and Porter and Thomas
[31] proposed the now generally accepted rule, the "Porter-Thomas distribution",
actually essentially without any proof. According to this, the matrix elements
in complex spectra show a Gaussian distribution. This is also well confirmed
experimentally.
We have to do here with two self-adjoint operators, the Hamiltonian H, which
defines the coordinate system, and the other operator M representing the physical
quantity, such as dipole moment, in the matrix elements of which one is inter-
ested. We shall assume, first, that both are real. This means, more precisely, that
there is a coordinate system in which all permissible operators of the physical
quantity in which we are interested have real matrix elements. If we choose the
coordinate axes in such a way that the states represented by them are time-
inversion invariant, the matrix elements of the Hamiltonian will be real. As to
the physical quantity in which we are interested, the matrix elements of its
operator will also be real if the quantity-such as electric dipole moment-is also
time-inversion invariant. This means that it retains its value if all the velocities
are reversed. Most physically important quantities do have this property, or the
opposite one, of reversing their signs if the directions of all the velocities are
reversed. The operator of these is, naturally, also hermitian but purely imaginary
and hence skew-symmetric. The operator of the magnetic moment is of this
nature. However, though this will not be explained in detail, the calculation of
the distribution finction of the matrix elements of time-reversal invariant opera-
tors can be modified so as to be applicable for anti-invariant operators also.
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
20 EUGENE P. WIGNER
Hence, the same distribution function can be expected for the matrix elements of
these as for real, time-inversion invariant operators. It will be assumed, further,
that the density of the characteristic values of the physical quantity in question
is an even function of the characteristic value-a condition naturally fulfilled
for operators which are anti-invariant with respect to time inversion. However,
all known time-inversion invariant transition operators also share this property.
The following calculation uses the coordinate system whose axes are the char-
acteristic functions of M, rather than of H. The characteristic values of this will
be denoted by M. The matrix element in question is, then,
(36) m = Z pixiyi,
where xi and yi are the coordinates of the states between which the matrix ele-
ment is taken. These are characteristic vectors of H, hence xi and yi can be
assumed to be real and m is real. The problem is to calculate the distribution
of m for the Hamiltonians of the ensemble chosen. If this is rotationally invariant
in Hilbert space, it amounts to calculating the distribution of the above expres-
sion for m when the vectors x and y move over a sphere but remain perpendicu-
lar to each other:
This calculation can be carried out in detail, but since the individual terms of
m are of the order I X I/N, where N is the dimension of the space of the matrices,
and since they have similar orders of magnitude and alternate in sign, one can
infer already from the central limit theorem that the distribution is Gaussian:
2-7)(2r; )-1/2e-rP2;;2
Fig. 7 shows that this is quite well confirmed experimentally in the case of
neutron emission. It has also been applied to various moments, not in nuclear
but in atomic physics, and the agreement is satisfactory in these cases also.
(Porter's book [1] contains articles applying the Porter-Thomas distribution to
various types of transition rates.)
It may be worthwhile to add a few words about the basis and limitations of
this formula. First, there is, of course, the same reservation which we had to
make in all other cases: we used a finite-dimensional space, though a space of
high dimensionality, rather than a Hilbert space. It appears that this problem
is not quite as serious in this case as in the case of level densities, and Rosen-
zweig made it at least plausible that the replacement of Hilbert space by a
space of high dimensionality is justified in this case [32].
The other assumption, that of the real nature of the matrices considered, is,
however, relevant. If we had chosen an ensemble of complex hermitean rather
than real symmetric matrices, the result would have been, for the transition
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 21
(j6O
(. ---; SOLID CURVE: EXPERIMENTAL DATA
x
2:30 -
w
j 15 _-
U-
O (6)
Z
z '--L=
- b
>30 1 2 3 4 7 6 7 8 9 10
"30 nl''fAv
FIG. 7
(38) (>jy1PI2I2
This would be easily distinguishable from the former distribution. One can
derive the latter formula by decomposing H into a real symmetric and an imagi-
nary skew-symmetric part. Both, separately, give a Porter-Thomas distribu-
tion, and the distribution for the sum of these is as given above.
This last point is, I believe, significant. There are experiments which indicate
that the actual Hamiltonian is not strictly time-inversion invariant, that it has
a small anti-invariant part. Nowhere would this manifest itself in the character
of the wave functions as strongly as in the region where the levels are close to
each other. Hence, an experimental check on the Porter-Thomas distribution
may give an indication of the magnitude, or at least an upper limit, of the not-
time-inversion invariant part of the Hamiltonian. Of course, the physical opera-
tor M, the matrix elements of which we consider, must be carefully chosen,
and it must be ascertained that the observed transitions are due entirely either
to an invariant or to an anti-invariant operator with respect to time inversion.
I might have mentioned that there is an effect in the level-spacing distribution
similar to the one discussed here: the not-time-inversion invariant part of the
Hamiltonian would manifest itself in this case in an added repulsion of the
levels.
This last observation will conclude the review of the recent contributions of
physicists to the theory of random matrices and of the role of random matrices
in the statistical theory of spectra. There remains, in the solution of almost
every problem which we have tackled, a good deal that should be clarified, and
almost all our derivations should be made more precise. More important than
this would be the clarification of a number of rather general questions, some of
which I wish to enumerate.
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
22 EUGENE P. WIGNER
REFERENCES
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms
RANDOM MATRICES 23
[16] M. KAC, Probability and Related Topics in Physical Sciences, vol. I, Interscience, New
York, 1959, P. 11.
[17] B. V. BRONK, Accuracy of the semicircle approximation for the density of eigenvalues of
random matrices, J. Mathematical Phys., 5 (1964), pp. 215-220. (Also Princeton
dissertation, 1963.)
[18] P. L. Hsu, On the distribution of roots of certain determinantal equations, Ann. Eugenics
(now Ann. Human Genetics), 9 (1939), pp. 250-258.
[19] D. N. NANDA, Distribution of a root of a determinantal equation, Ann. Math. Statist.,
19 (1948), pp. 47-57.
[20] C. E. PORTER AND N. ROSENZWEIG, Statistical properties of atomtic and nuclear spectra,
Ann. Acad. Sci. Fenn. Ser. A VI, no. 44 (1960). (Also reprinted in [1] above.)
[21] J. VON NEUMANN, Charakterisierung des Spectrums eines Integraloperators, Actualites
Sci. Indust. (229), no. 13 (1935). (Reprinted in John von Neumann Collected
Works, vol. IV, A. H. Taub, ed., Pergamon Press, New York, 1962.)
[22] J. M. BLATT AND V. F. WEISSKOPF, Theoretical Nuclear Physics, John Wiley, New York,
1952, ?VIII-6.
[23] T. D. NEWTON, Shell effects on the spacing of nuclear levels, Canad. J. Phys., 34 (1956),
pp. 804-829.
[24] J. GINIBRE, Statistical ensembles of complex, quaternion, and real matrices, J. Mathe-
matical Phys., 6 (1965), pp. 440-449.
[25] E. P. WIGNER, On a class of analytic functions from the quantum theory of collisions,
Ann. of Math., 53 (1951), pp. 36-67.
[26] J. VON NEUMANN AND E. P. WIGNER, Behavior of eigenvalues in adiabatic processes,
Phys. Zeitschr., 30 (1929), pp. 467-470.
[27] M. L. MEHTA, On the statistical properties of the level-spacings in nuclear spectra, Nu-
clear Phys., 18 (1960), pp. 395-419.
[28] M. GAUDIN, Sur la loi limite de l'espacement des valeurs propres d'une matrice aleatoire,
Ibid., 25 (1961), pp. 447-458.
[29] G. SZEGO, Orthogonal Polynomials, American Mathematical Society Colloquium
Publications, vol. 23, American Mathematical Society, New York, 1959, pp.
104-105.
[30] J. M. C. SCOTT, Neutron widths and the density of nuclear levels, Philos. Mag. (7), 45
(1954), pp. 1322-1331.
[31] C. E. PORTER AND R. G. THOMAS, Fluctuations of nuclear reaction widths, Phys. Rev.,
104 (1956), pp. 483-491.
[321 N. ROSENZWEIG, Anomalous statistics of partial radiation widths, Phys. Lett., 6 (1963),
pp. 123-125.
This content downloaded from 139.124.126.14 on Fri, 20 Jan 2017 09:42:16 UTC
All use subject to http://about.jstor.org/terms