Random Matrices and L-functions

J. P. Keating
and
N. C. Snaith
School of Mathematics
University of Bristol
Bristol BS8 1TW
UK

November 7, 2002

Abstract
In recent years there has been a growing interest in connections between the statistical
properties of number theoretical L-functions and random matrix theory. We review the history
of these connections, some of the major achievements, and a number of applications.

1 The history in brief

Number theory and random matrix theory met, by chance, over a cup of tea in the common room
at the Institute for Advanced Study in Princeton in the early 1970s. H.L. Montgomery, then a
graduate student, had shown his latest work to F.J. Dyson. His main conclusion was a conjecture
for the two-point correlation function of the zeros of the Riemann zeta function, which he had
managed to prove for a limited range of correlations [57]. Dyson recognized this to be the same
as the two-point correlation function, which he had calculated a decade earlier, for the eigenvalues
of matrices drawn at random from U (N ) (the group of N × N unitary matrices) uniformly with
respect to Haar measure [30].
The Riemann zeta function is extremely important in number theory because it allows
for analytical techniques to be applied to the study of the distribution of prime numbers, and
Montgomery’s conjecture plays a central role in the theory of this distribution.
Although the full proof of Montgomery’s conjecture has still escaped completion, Odlyzko’s
numerical computations have provided very strong evidence in support of it [59]. Working on the
powerful computers of AT&T, in the 1980s Odlyzko calculated batches of the Riemann zeros high
up on the critical line where the Riemann Hypothesis places them, and computed numerically the
two-point correlation function, as well as many other statistics of the zeros. He also computed the
distribution of the values of the zeta function. His numerics provide very convincing evidence that,
as predicted by Montgomery’s conjecture, the two-point correlation function of the Riemann zeros
converges to the random matrix result of Dyson as zeros higher and higher on the critical line are
considered. Odlyzko’s numerical work continues today and he is currently working at the dizzying
height of the 1022 nd Riemann zero.
In the 1980s the Riemann zeta function also took on a new life in mathematical and the-
oretical physics when it became a tool in the field of quantum chaos. The eigenvalues of complex
quantum systems (eg. nuclei [4, 18, 35] and disordered mesoscopic systems [31]) and systems show-
ing chaotic behaviour in the classical limit [12, 16, 17] display the statistics of the three standard
ensembles of Hermitian matrices, the Gaussian Unitary Ensemble (GUE), the Gaussian Orthogonal
Ensemble (GOE) and the Gaussian Symplectic Ensemble (GSE), or equivalently (in the appropri-
ate limit) the three corresponding standard ensembles of unitary matrices: the CUE, COE and
CSE. (Here the “C” stands for “circular”.) However, while the eigenvalue statistics show certain
behaviour which is easily identified and predicted by random matrices, there are further charac-
teristics which are system-specific and are connected to the periodic orbits of the system under
consideration [6]. These relate to the approach to the random-matrix limit as ~ → 0. It was
realized by Berry [7] that very similar contributions result from the primes and behave in the same
way with respect to the statistics of the Riemann zeros as do the periodic orbits to the statistics of
the eigenvalues. In this case, the primes describe the approach to the random-matrix limit as the
height up the critical line increases.
This launched an era of study of the Riemann zeta function in the field of quantum chaos
[2, 6, 8, 9, 10, 11, 47]. Most of the results are reviewed in [11, 48]. In essence, the connection
between the Riemann zeta function and the prime numbers was being used to point the way through
more complicated periodic orbit calculations; and in return the familiarity which arose with the
zeta function enabled physicists to contribute insights from physics to its study. In particular, it
resulted in a return to the question of the statistics of the Riemann zeros when Bogomolny and
Keating [13, 15] in 1995 and 1996 showed, subject to certain conjectures of Hardy and Littlewood
concerning the distribution of primes, that not just the two-point, but the general n-point statistics
of the Riemann zeros are the same as those of the eigenvalues of random unitary matrices in the
limit as one looks at zeros infinitely high up the critical line.
At the same time, first Hejhal [37] with the three-point case, then Rudnick and Sarnak [61]
generalized Montgomery’s theorem by proving for a limited range of correlations, as in the two-
point statistic, that the n-point correlations of the Riemann zeros high on the critical line coincide
with the corresponding random unitary matrix statistics.
At the end of the 1990s, two developments occurred which illustrate how deeply random
matrix theory is intertwined with the Riemann zeta function. Far from there being some accidental
similarity between the zeros of this one function and the eigenvalues of random matrices, it became
apparent that this connection was far more general. The zeta function is but one example of a
broader class of functions known as L-functions. These all satisfy generalizations of the Riemann
Hypothesis. For any individual L-function, it is believed that the zeros high up on the critical line
are distributed like the eigenvalues of random unitary matrices, that is, exactly as in the case of the
Riemann zeta function [57, 61]. More interesting, however, is the fact that it has been conjectured
by Katz and Sarnak [45, 46] that averages over various families of L-functions, with the height up
the critical line of each one fixed, are described not only by averages over the unitary group U (N ),
but by averages over other classical compact groups, for example the orthogonal group O(N ) or
the unitary symplectic group U Sp(2N ), depending upon the family in question. The eigenvalue
statistics of these groups have also been found to occur in disordered superconducting systems [1].
The second of the two significant recent developments was the discovery that calculations
purely within random matrix theory can suggest answers to important questions that number
theorists have been unable using standard techniques to make progress on. In [51, 52] Keating and
Snaith showed that by studying the value distribution and moments of the characteristic polynomial
of a random matrix, one can make predictions about the value distribution and moments of the
Riemann zeta function and other L-functions. The reason for this is clear: the characteristic
polynomial of a random matrix has zeros (the eigenvalues of the matrix) which, conjecturally, show
the same statistical behaviour of the zeros of L-functions. Thus random matrix theory can be put

2
to very practical use in the study of L-functions.
Our aim in the following sections is to expand further on the developments outlined above.
Specifically, we will concentrate on those aspects not covered in previous reviews [11, 48].

2 Pair correlations
We begin with some basic facts about the Riemann zeta function.
The Riemann zeta function is defined by
∞
X 1 Yµ ¶
1 −1
ζ(s) = = 1− s (1)
ns p
p
n=1

for Res > 1, where p labels the primes, and then by analytic continuation to the rest of the complex
plane. It has a pole at s = 1, zeros at s = −2, −4, −6, . . . (the trivial zeros) and infinitely many
zeros, called the non-trivial zeros, in the critical strip 0 < Res < 1. The Riemann Hypothesis
states that all of the non-trivial zeros lie on the critical line Res = 1/2; that is, ζ(1/2 + it) = 0
has non-trivial solutions only when t = tn ∈ R [63]. This is known to be true for at least 40% of
the non-trivial zeros [20], for the first 1.5 × 109 + 1 of them [53], and for batches lying much higher
[59]. (A distributed computing project claimed as of July 31st 2002 that the first 50, 631, 912, 399
non-trivial zeros lie on the line! [65]) The zeta function satisfies a functional equation:

1
s− 2 Γ( 12 − 12 s)
ζ(s) = π ζ(1 − s). (2)
Γ( 12 s)
In the following, for ease of presentation, we will assume the Riemann Hypothesis to be
true, although this is not strictly necessary.
The mean density of the non-trivial zeros increases logarithmically with height t up the
critical line. Specifically, defining unfolded zeros by
1 tn
wn = tn log , (3)
2π 2π
it is known that
1
lim # {wn < W } = 1. (4)
W →∞ W

The question then arises as to the statistical distribution of the unfolded zeros: are they equally
spaced, with unit spacing between neighbours, randomly distributed with unit mean spacing, or do
they have some other limiting distribution? Statistics such as the two-point correlation function
contain information about this distribution. For example, let
1
Fζ (α, β; W ) = #{wn , wm ∈ [0, W ] : α ≤ wn − wm < β}; (5)
W
that is, Fζ measures correlations between pairs of unfolded zeros wn ∈ [0, W ]. The question is,
first, does a limit distribution
Fζ (α, β) = lim Fζ (α, β; W ) (6)
W →∞

exist, and second, if so, what can one say about it?
In 1973 Montgomery provided at least part of the answer to this. The two-point correlation
function R2,ζ (x) of the Riemann zeros is related to Fζ (α, β) by

3
 Z β
Fζ (α, β) = (R2,ζ (x) + δ(x))dx. (7)
α
Using more general test functions, we can write

1 X
R2,ζ (f, W ) = f (wj − wk )
W j6=k
wj ,wk ≤W
Z ∞ X
1
= f (x) δ(x − wj + wk )dx. (8)
−∞ W j6=k
wj ,wk ≤W

Montgomery’s theorem is

Theorem 1 (Montgomery 1973 [57])

For test functions f (x) such that
Z ∞
fˆ(τ ) = f (x)e2πixτ dx
−∞

has support in (−1, 1), the following limit exists:

Z ∞
lim R2,ζ (f, W ) = f (x)R2 (x)dx,
W →∞ −∞
with
µ ¶2
sin(πx)
R2 (x) = 1 − .
πx

Based on this result, Montgomery further conjectured that Theorem 1 is true for fˆ(τ ) of
unrestricted support. This we write as:

Conjecture 1 (Montgomery 1973 [57])

If, for Fζ (α, β) defined as in (6), we write
Z β
Fζ (α, β) = (R2,ζ (x) + δ(x))dx,
α
then

R2,ζ (x) = R2 (x),

with R2 (x) as in Theorem 1.

We are now in a position to explain the connection with random matrices. Random matrix
theory was initiated by Eugene Wigner in the 1950s to describe the statistical distribution of
nuclear energy levels. It was later developed by Dyson, Mehta, Gaudin and others in the 1960s
into a rigorous area of mathematical physics. For a detailed introduction see [54]. One important
result - the two-point correlation function for the eigenvalues of unitary matrices - was proved by
Dyson in 1963. Let A be an N × N unitary matrix; that is, A ∈ U (N ). Denote the eigenvalues of

4
A by exp(iθn ), where 1 ≤ n ≤ N and θn ∈ R (this follows from unitarity). Clearly the eigenphases
θn have mean density N/2π, so the unfolded eigenphases

N
φn = θn , (9)
2π
have unit mean density (i.e. φn ∈ [0, N )). Next let us define, by analogy with (5),
1
F (α, β; A, N ) = #{φn , φm : α ≤ φn − φm < β}. (10)
N
Now, the unitary group U (N ) comes with a natural invariant measure - Haar measure - which
we will denote dA, and so one may compute the average over A of this function, with A taken
uniformly with respect to Haar measure:
Z
FU (α, β; N ) = F (α, β; A, N )dA. (11)
U (N )

Dyson’s theorem is then:

Theorem 2 (Dyson 1963 [30])

The limit distribution
FU (α, β) = lim FU (α, β; N ),
N →∞

(defined as in (10) and (11)) exists and takes the form

Z β
FU (α, β) = [R2,U (x) + δ(x)] dx, (12)
α

where δ(x) is Dirac’s δ-function and

µ ¶2
sin(πx)
R2,U (x) = R2 (x) ≡ 1 − .
πx

The integrand in (12) may be thought of as the two-point correlation function for the
eigenphases of a random unitary matrix, unfolded to have unit mean spacing. The fact that it
is a non-trivial function of correlation distance x means that eigenphases are correlated in a non-
trivial way. Note that in Theorem 2 the definition of R2 (x) is identical to that in Theorem 1, so the
Riemann zeros and the eigenvalues of matrices in U (N ) appear to have exactly the same non-trivial
correlations in the limit of infinite height up the critical line on the one hand, and infinite matrix
size on the other.
The basic idea underlying the method used to prove Montgomery’s theorem is the following
one. The formula (1) expressing ζ(s) as a product over the primes may be used to relate the zeros
tn to sums over the primes. Hence the pair correlation of the zeros may be written as a sum over
pairs of primes p, q. The contribution from the diagonal terms with p = q may be evaluated using
the prime number theorem. The condition on f (x) in Montgomery’s theorem is designed so that
only the diagonal terms contribute. Relaxing the condition on f would require the evaluation of
the off-diagonal terms (p 6= q). The details of these calculations are reviewed in [48].

5
 1

0.8

0.6

0.4

0.2

0.5 1 1.5 2 2.5 3

Figure 1: The two-point correlation function of 106 Riemann zeros around the height of the 1020 th
zero (dots) and the two-point correlation function R2 (x) (smooth curve), see Theorem 2, of the
eigenvalues of matrices in U (N ) in the large-N limit. Figure courtesy of A. M. Odlyzko.

3 More on the zero statistics

There is substantial evidence in support of Montgomery’s conjecture. Among other statistics,
Odlyzko has computed the two-point correlation function numerically for batches of zeros high up
on the critical line (e.g. near to the 1020 th zero) and his results [59] show striking agreement with
R2 (x) (given in Theorem 1), as illustrated by Figure 1.
Odlyzko’s computations show a similarly convincing agreement for the distribution of spac-
ings between adjacent unfolded Riemann zeros and the equivalent random matrix statistic [59].
This nearest-neighbour distribution depends on all of the n-point correlation functions and so sug-
gests that Montgomery’s conjecture generalizes to relate correlations between n-tuples of zeros and
the corresponding correlations between n-tuples of eigenphases. That Theorem 1 also generalizes
was proved by Hejhal for the three-point correlation function [37], and by Rudnick and Sarnak [61]
for all the n-point correlations, as we will now explain.
Rudnick and Sarnak define an n-point correlation sum as follows. For a set BN of N
unfolded Riemann zeros w1 ≤ w2 ≤ · · · ≤ wN and for a test function f which satisfies

f (x) ≡ f (x1 , . . . , xn ) is symmetric, (13a)

f (x + t(1, . . . , 1)) = f (x) for t ∈ R (ie. f (x) is a function of successive (13b)
0
differences of the x s), and
X
f (x) → 0 rapidly as |x| → ∞ in the hyperplane xj = 0, (13c)
j

define

n! X
Rn,ζ (BN , f ) = f (S). (14)
N S⊂B
N
|S|=n

On the random matrix side, the n-point correlation function of the eigenphases of matrices
from U (N ) is defined as

6
 Z 2π Z 2π
N!
Rn (θ1 , . . . , θn ; N ) = ··· P (θ1 , · · · θN )dθn+1 · · · dθN , (15)
(N − n)! 0 0

where the joint probability density function of the eigenphases (derived from Haar measure) is given
by
1 Y
P (θ1 , . . . , θN ) = |eiθn − eiθm |2 ; (16)
N !(2π)N
1≤m<n≤N

that is, P (θ1 , . . . , θN )dθ1 · · · dθN is the probability that a matrix plucked from this ensemble has
eigenphases between θ1 and θ1 + dθ1 , between θ2 and θ2 + dθ2 , and so on. These n-point correlation
functions were evaluated by Dyson [30] and are given by

Rn (θ1 , . . . , θn ; N ) = det[KN (θj − θk )]j,k=1,...,n , (17)

where

1 sin(N θ/2)
KN (θ) = . (18)
2π sin(θ/2)
In the large-N limit, then, the n-point correlation function of the unfolded eigenphases is

Rn (φ1 , . . . , φn ) = det[K(φj − φk )]j,k=1,...,n , (19)

with K(x) = sinπxπx .
With these definitions, we have the theorem

Theorem 3 (Rudnick and Sarnak 1996 [61])

Let the ˆ
P test function f satisfy the conditions (13), and in addition assume that f (ξ) is
supported in j |ξj | < 2. Further, assume the Riemann Hypothesis to be true. Then as N → ∞,
Z µ ¶
x1 + · · · + xn
Rn,ζ (BN , f ) → f (x)Rn (x)δ dx1 · · · dxn .
Rn n

In the above theorem Rudnick and Sarnak make use of the assumption that the Riemann
Hypothesis is true, but in [61] they define a smoothed version of the n-point correlation function
and with this prove a similar result to Theorem 3 without it being necessary for the Riemann zeros
to lie on the critical line. It is also important to note that Rudnick and Sarnak proved Theorem 3
not only for the zeros of the Riemann zeta function, but also for a whole class of other L-functions,
as will be discussed in the following section.
The methods used in proving Theorem 3 are essentially a direct generalization of those
underlying Theorem 1, described in outline at the end of Section 2; that is the result follows from
an evaluation of the diagonal contributions to multiple sums over the primes.
In Theorem 3 as in the case of the two-point correlation function (Theorem 1), there is
a restriction imposed on the support of the test function f (x). No statistics of the Riemann
zeros have been proved to agree with random matrix results outside of this range of correlation.
To do so requires an evaluation of the off-diagonal contributions to multiple sums over the primes.
However, there is a heuristic calculation of Bogomolny and Keating [13, 15] which shows the n-point
correlation function to be, high on the critical line, exactly the same as that predicted by random
matrix theory, without any restrictions on the correlation range. Their method uses a conjecture by

7
Hardy and Littlewood on the correlations between primes to evaluate the off-diagonal contributions
needed. (See [48] for a review of the basic ideas in the case of the two-point correlation function.)
There have recently been other studies of the statistical distribution of the zeros of the
Riemann zeta function. For example, the idea introduced in [52] that statistical properties of the
zeta function at a finite height up the critical line might be modelled by finite-size random matrices,
with height and size related by N ∼ log t, which will be discussed in later sections, has been verified
in a systematic study of the fluctuations in the number of unfolded zeros lying in ranges of a given
length (see the contribution by Coram and Diaconis [27] to this issue).
The conclusion to be drawn is that the statistical distribution of the Riemann zeros, in the
limit as one looks infinitely high up the critical line, coincides with the statistical distribution of
the eigenvalues of random unitary matrices, in the limit of large matrix size. (We note as well that
a great deal is also known about the way in which zero statistics approach the large-height limit
described by random matrix theory - see, for example, [8, 14, 49]. Results concerning the approach
to this limit were reviewed recently by Berry and Keating in [11].)

4 Families of L-functions
As mentioned in the introduction, the connection between the Riemann zeros and random matrix
theory is merely one example of a much more general relationship. The results of Rudnick and
Sarnak hold not just for the Riemann zeta function but for other individual L-functions, as we will
now discuss. Moreover, Katz and Sarnak [45, 46] have proposed a fundamental generalization -
in terms of families of similar L-functions with each family subscribing to a symmetry type, not
just the familiar unitary symmetry of U (N ), but also the symmetry corresponding to O(N ) and
U Sp(2N ).
We will consider first the basic properties of L-functions, and then give a simple example
of a family by way of illustration, before discussing further the results of Katz and Sarnak.
L-functions share the same general structure as the Riemann zeta function in that there
exists for each a Dirichlet series and an Euler product over the primes p like those in (1), for
example
∞
̰ !
X an Y X apk
L(s) = = . (20)
ns p
pks
n=1 k=0

The coefficients an might be Dirichlet characters or Fourier coefficients of an automorphic cusp form.
Whatever the source of the various L-functions, they have in common an analytic continuation
beyond the region in which the series converges, and a functional equation which relates the L-
function in one half of the complex plane with the other half. We will always consider L-functions
to be normalized so that the line of symmetry is Res = 1/2. The Generalized Riemann Hypothesis
then proposes that all the non-trivial zeros of a given L-function lie on this critical line. Theorem 3
of Rudnick and Sarnak applies equally well to any primitive L-function (one which does not factor
into a product of L-functions) providing that the condition
X |apk log p|2
< ∞, (21)
p
pk

holds for any k ≥ 2, where the sum is over the prime numbers (this clearly holds in the case of
the Riemann zeta function). This suggests that for any individual L-function, the distribution of
its zeros high on the critical line will display the same characteristics as the distribution of the

8
eigenvalues of matrices pulled at random from U (N ) (uniformly with respect to Haar measure) for
large N .
Now we turn to families of L-functions. Take, for example, an L-function with coefficients
determined by a real Dirichlet character:
∞
X Y£
χd (n) ¤−1
L(s, χd ) = = 1 − χd (p)p−s (22)
ns p
n=1
¡d¢
where χd (n) = n is Kronecker’s extension of Legendre’s symbol which is defined for p prime,

µ ¶  +1 if p - d and x2 ≡ d(modp) is soluble
d
= 0 if p|d . (23)
p 
−1 if p - d and x2 ≡ d(modp) is not soluble
The character χd exists for all fundamental discriminants d, and the L-functions attached
to these characters are said to form a family as we vary d. The family can be partially ordered
by the conductor |d|. Katz and Sarnak noted that U (N ) and the other circular matrix ensembles
which are standard, as mentioned in the introduction, in physics, COE= U (N )/O(N ) and CSE=
U (2N )/U Sp(2N ), are only three of the dozen or so symmetric spaces characterized by Cartan.
Amongst the others are the compact groups U Sp(2N ) and O(N ). In the context of physical
systems, these other symmetric spaces were realized to be of importance independently by Altland
and Zirnbauer [1]. From Katz and Sarnak we borrow Table 1 of the symmetric spaces which are
important here.

Symmetry Type G Realization of G(N ) as matrices

U , also called CUE U (N ), the compact group of N × N unitary matrices.
SO (even) SO(2N ), the compact group of 2N × 2N
unitary matrices A satisfying At A = I, det A = 1.
SO (odd) SO(2N + 1), the compact group of (2N + 1) × (2N + 1)
unitary matrices A satisfying At A = I, det A = 1.
Sp U Sp(2N ), the compact group of 2N × · 2N unitary¸ matrices A
0 IN
satisfying At JA = J, J = .
−IN 0
COE U (N )/O(N ), symmetric unitary N × N matrices identified with the
above cosets via B → B t B.
CSE U (2N )/U Sp(2N ), 2N × 2N unitary matrices satisfying J t H t J = H
identified by B → BJB t J t .

Table 1: Symmetric spaces G(N ), from [46].

Following the notation of Katz and Sarnak, we let G(N ) stand for any of the ensembles in
Table 1. Since A ∈ G(N ) is a unitary matrix, we write the eigenvalues as eiθ1 (A) , . . . , eiθN (A) , and
label them so that 0 ≤ θ1 (A) ≤ · · · ≤ θN (A) < 2π. The measure, denoted dA, on the compact
groups is always Haar measure, and for the circular ensembles we use the volume form as the
probability measure.
Amongst other local statistics, Katz and Sarnak define the k-th consecutive spacings
N
#{1 ≤ j ≤ N | 2π (θj+k − θj ) ∈ [a, b]}
µk (A)[a, b] = , (24)
N

9
and show that for fixed k ≥ 1, the same limit
Z
lim µk (A)dA = µk (CU E), (25)
N →∞ G(N )

exists irrespective of how G(N ) is chosen from among the first four ensembles in Table 1. (Katz
and Sarnak write µk (GU E), but since the local statistics of the ensemble of Hermitian matrices,
GUE, and the CUE ensemble of unitary matrices are the same in the limit N → ∞, either notation
is suitable.) Further, Katz and Sarnak show that for a typical (in measure) A ∈ G(N ) the statistic
µk (A) approaches µk (CU E) as N → ∞. The same type of result is established also for the n-point
correlations of the eigenvalues, and thus the local statistics of the entire set of N (or 2N ) eigenvalues
of matrices from any of the four compact groups mentioned above tend to the same limit as N
becomes large.
In contrast to this, Katz and Sarnak showed that the statistics of only the first eigenvalue
(or more generally the first few eigenvalues) are specific to the particular ensemble chosen. If we
define the distribution of the k-th eigenvalue of a matrix A varying over G(N ),
θk (A)N
νk (G(N ))[a, b] = meas{A ∈ G(N ) : 2π ∈ [a, b]}, (26)
then Katz and Sarnak show that the limit

lim νk (G(N )) = νk (G) (27)

N →∞

exists, but in contrast to (25), the limit depends on the ensemble G.

Based on this, Katz and Sarnak proposed that while the high zeros of any one L-function
will always show the typical statistics of U (N ) (otherwise known as the CUE) the statistics of the
lowest zeros near to s = 1/2 will show, when their distribution is determined over a suitable family
of L-functions, the statistics of U (N ), O(N ) or U Sp(2N ).
To be more explicit, if we assume the Riemann hypothesis for the L-functions we are
considering, then we write the zeros as 1/2 + iγ (n) and order them

· · · ≤ γ (−2) ≤ γ (−1) ≤ 0 ≤ γ (1) ≤ γ (2) ≤ · · · . (28)

If F denotes a family of L-functions, and an individual L-function within the family is identified
by f and has conductor cf , the zeros near s = 1/2 are normalized to have unit mean spacing by
scaling them in the following way:

γfj log cf
. (29)
2π
Let FX denote the members of the family F with conductor less than X. Then Katz and Sarnak
define the distribution of the jth eigenvalue as
½ (j)
¾
γf log cf
# f ∈ FX : 2π ∈ [a, b]
νj (X, F)[a, b] = . (30)
#FX
It is then expected, and Katz and Sarnak provide analytical and numerical evidence for this, that
νj (X, F) will converge, as X grows large, to νj (G(N )), where G(N ) represents the symmetry type
of the family: U (N ), O(N ) or U Sp(2N ). Similarly, any other statistics of the lowest zeros would
also be expected, upon averaging over the family, to tend to the random matrix statistics of the
correct symmetry type in the same limit as above.

10
 For example, in the case of the family of L-functions with real Dirichlet characters described
at (22), the low-lying zeros appear to show symplectic symmetry. In Figure 2, taken from [60], we
see for j = 1 and j = 2 the good agreement between the numerically calculated distribution of the
jth zero above s = 1/2 on the critical line and the distribution of the jth eigenvalue of the group
U Sp(2N ). These distributions are visibly different from those of the lowest eigenvalues of matrices
from U (N ) where there is no repulsion of the first zero by θ = 0.

1.5 1.5
2nd zero above 0 for L(s,chi_d),
1st zero above 0 for L(s,chi_d), 10E12 < |d| < 10E12+200000, 7243 d’s,
10E12 < |d| < 10E12+200000, 7243 d’s, mean of 2nd zero equals 1.8432,
mean of 1st zero equals 0.8268, renormalized to have mean 1.76
renormalized to have mean 0.7827

1.0 1.0

0.5 0.5

0.0 0.0
0 1 2 3 0 1 2 3 4

Figure 2: The numerically computed histogram of the distribution over d of the height of the first
zero above s = 12 of L(s, χd ) (left) versus ν1 (Sp) and the distribution of the height of the second
zero above s = 12 versus ν2 (Sp) (right). Figure courtesy of M. O. Rubinstein [60].

The question of determining the symmetry type of a given family a priori is in general a
difficult one. The method used by Katz and Sarnak is that for some families of L-functions, a
related family of zeta functions on finite fields can be defined. In the case of these zeta functions
the definition of families is straightforward, the Riemann hypothesis has been proven (in that all
zeros lie on a circle) and the symmetry type is determined by the monodromy of the family (see
[45]). The symmetry type of the related family of L-functions is then assumed to be the same. We
return to this problem of determining the symmetry type of families in a later section.
Further studies of the statistics of low-lying zeros of Dirichlet L-functions have recently
been carried out by Hughes and Rudnick [41], who compared the moments of linear statistics of
scaled zeros around the symmetry point s = 21 with similar moments of statistics of eigenphases
near the point 1. In both cases they found comparable mock-Gaussian behaviour.
The idea of relating zero statistics for L-functions to averages over the classical compact
groups has been extended by Keating, Linden and Rudnick [50] to the exceptional Lie groups.
Specifically, they construct a family of L-functions associated with a finite field in which the relevant
average is over the exceptional group G2 .

5 Random matrices and log ζ(1/2 + it)

In the preceding sections we have described the overwhelming evidence for a connection between
the statistics of the zeros of L-functions and the eigenvalues of ensembles of matrices. The next

11
step is to see what use can be made of that connection. Since the value distribution of a function
is, to some extent, determined by its zeros, in this section and the following one we will describe
how the value distributions of L-functions and of logarithms of L-functions can be probed using
random matrix theory. We start with the logarithm.
At a given height t up the critical line, log ζ(1/2 + it) is a complex number, and one might
ask: how are the real and imaginary parts of it distributed as t varies? In the limit as t → ∞, the
answer to this question is provided by a beautiful theorem due to Selberg [63, 59]
Theorem 4 (Selberg)
For any rectangle B ∈ C,
¯ ¯
¯ ¯¯
1¯ ¯ log ζ(1/2 + it)
lim t : T ≤ t ≤ 2T, q ∈ B ¯¯
T →∞ T ¯¯ 1 ¯
2 log log T
Z Z
1 2 2
= e−(x +y )/2 dx dy.
2π B
That is, in the limit as T , the height up the criticalpline, tends to infinity, the value distributions
of the real and imaginary parts of log ζ(1/2 + iT )/ (1/2) log log T each tend, independently, to
a Gaussian with unit variance and zero mean. Crucially for us, Odlyzko’s computations for these
distributions when T ≈ t1020 show significant systematic deviations from this limiting form [59].
For example, increasing moments of both the real and imaginary parts diverge markedly from
the Gaussian values. There is, of course, no contradiction; this merely suggests that the limiting
Gaussian distribution is approached rather slowly as T → ∞. It does, though, lead to the question
of how to model the statistical properties of log ζ(1/2 + it) when t is large but finite.
Given its success in describing other statistical properties of the zeta function, it is natural
ask whether random matrix theory might be used as the basis of such a model. The question is,
then: what property of a matrix plays the role of the zeta function? The answer is simple: since
the zeros of the zeta function are distributed like the eigenvalues of a random unitary matrix, the
zeta function might be expected to be similar, in respect of its value distribution, to the function
whose zeros are the eigenvalues, that is, to the characteristic polynomial of such a matrix. This
idea was introduced and investigated in detail in [52]. Here we give details of some main results.
The characteristic polynomial of a unitary matrix A may be defined by
Λ(z) ≡ ΛA (z) = det(I − Az). (31)
The moment generating function for Re log Λ(eiθ ), for example, is thus
Z Z
iθ
MU (s; N ) = exp(sRe log ΛA (e ))dA = |ΛA (eiθ )|s dA, (32)
U (N ) U (N )

where the integration is, as before, with respect to Haar measure. Obviously ΛA may be written
in terms of the eigenangles of A:
N ³
Y ´
iθ
ΛA (e ) = 1 − ei(θn +θ) . (33)
n=1
Haar measure on U (N ) may also be expressed in terms of these eigenangles [64], allowing one to
write
Z Z 2π Z 2π ¯N ¯s
1 Y ¯¯ ¯2 ¯ Y ¯
iθm ¯ ¯ i(θn +θ) ¯
|ΛA (eiθ )|s dA = · · · ¯e iθj
− e ¯ ¯ (1 − e )¯ dθ1 · · · dθN .
U (N ) (2π)N N ! 0 0 ¯ ¯
1≤j<m≤N n=1
(34)

12
This N -dimensional integral may then be computed by relating it to an integral evaluated by
Selberg [54], giving
N
Y Γ(j)Γ(j + s)
MU (s; N ) = . (35)
(Γ(j + s/2))2
j=1

All information about the value distribution of Re log Λ on the unit circle is contained within
(35): moments may be computed in terms of the derivatives of MU (s; N ) at s = 0, and the value
distribution itself is the fourier transform of MU (iy; N ). In the same way, information about the
value distribution of Im log Λ, and the joint value distribution of the real and imaginary parts of
log Λ may be computed. This leads to a central limit theorem for log Λ (see also [3, 28]):

Theorem 5 (Keating and Snaith 2000 [52])

 
  Z Z
log Λ 1 1 2 2
lim meas A ∈ U (N ) : q ∈B = e− 2 (x +y ) dxdy
N →∞  1
log N  2π B
2

for rectangles B ∈ C.

This theorem corresponds precisely to Selberg’s for the value distribution of log ζ(1/2 + it),
suggesting that random matrix theory, in the limit as the matrix-size tends to infinity, can indeed
model the value distribution of log ζ(1/2 + it) as t → ∞. The question that remains is whether it
can also model the approach to the limit, that is, the value distribution when t is large but finite.
In order to relate the large-t asymptotics for the zeta function to the large-N asymptotics
for the characteristic polynomials we need a connection between t and N . Note that the scaling in
Theorem 4 and that in Theorem 5 coincide if we set
t
N = log . (36)
2π
Such an identification is natural, because it corresponds to equating the mean density of the Rie-
mann zeros at height t to the mean density of eigenphases for N × N unitary matrices, and these
are the only parameters that appear in the connection between the respective statistics (cf. (3) and
(9)). This therefore prompts the question as to whether the rate of approach to Selberg’s theorem
as t → ∞ is related to that for Theorem 5 as N → ∞ (which can be computed straightforwardly
using (35)) if we make the identification (36).
As already noted above, Odlyzko’s numerical computations of the value distribution of the
zeta function near to the 1020 th zero show significant deviations from the Gaussian limit given in
Theorem 4. The integer closest to log(t1020 /2π) is N = 42 (t1020 ≈ 1.5202 × 1019 ), so in Figure 3 we
plot the value distribution for Re log ζ(1/2 + it), scaled as in Theorem 4, computed by Odlyzko [59],
together with the value distribution for Re log Λ, scaled as in Theorem 5, with respect to matrices
taken from U (42). Also shown is the Gaussian with zero mean and unit variance which represents
the limit distribution in both cases (as t → ∞ and N → ∞ respectively). The negative logarithm
of these curves is plotted in Figure 4, highlighting the behaviour in the tails. In order to quantify
the data, the moments of the three distributions are listed in Table 2.
It is clear that random matrix theory provides an accurate description of the value distri-
bution of Re log ζ(1/2 + it). It also models Im log ζ(1/2 + it) equally well [52]. This then suggests
that, statistically, the zeta function at a large height t up the critical line behaves like a polynomial
of degree N , where t and N are related by (36); and, moreover, that the polynomial in question is
the characteristic polynomial of a random unitary matrix.

13
Figure 3: The value distribution for Re log Λ with respect to matrices taken from U (42), Odlyzko’s
data for the value distribution of Re log ζ(1/2 + it) near the 1020 th zero (taken from [59]), and the
standard Gaussian, all scaled to have unit variance. (Taken from [52].)

Moment ζ a) ζ b) U (42) Normal

1 0.0 0.0 0.0 0
2 1.0 1.0 1.0 1
3 -0.53625 -0.55069 -0.56544 0
4 3.9233 3.9647 3.89354 3
5 -7.6238 -7.8839 -7.76965 0
6 38.434 39.393 38.0233 15
7 -144.78 -148.77 -145.043 0
8 758.57 765.54 758.036 105
9 -4002.5 -3934.7 -4086.92 0
10 24060.5 22722.9 25347.77 945

Table 2: Moments of Re log ζ(1/2 + it), calculated by Odlyzko over two ranges (labelled a and b)
near the 1020 th zero (t ' 1.520 × 1019 ) (taken from [59]), compared with the moments of Re log Λ
for U (42) and the Gaussian (normal) moments, all scaled to have unit variance.

14
 Figure 4: minus the logarithm of the value distributions plotted in Figure 3. (Taken from [52].)

Of course, specific properties of the zeta function would be expected to appear in the
description of its value distribution. The point is that these contribute at lower order in the
asymptotics, with the leading order being given by random matrix theory. For example, it is shown
in [52] that as N → ∞
Z
1 1
(Im log ΛA )2 dA = log N + (γ + 1) + o(1), (37)
U (N ) 2 2
where γ is Euler’s constant, while Goldston [34] has proved, under the assumption of the Riemann
Hypothesis and Montgomery’s conjecture, that as T → ∞
Z
1 T
(Im log ζ(1/2 + it))2 dt (38)
T 0
∞ X
X
1 T 1 (1 − m) 1
= log log + (γ + 1) + + o(1).
2 2π 2 p
m2 pm
m=2

These expressions coincide under the identification (36), except for the sum over primes in (38).
Obviously the primes have their origin in number theory, rather than random matrix theory.
In determining the value distribution of log ΛA (eiθ ) (e.g. as in Theorem 5), the averages
were performed over matrices A taken uniformly with respect to Haar measure on the unitary
group U (N ). It is natural to ask how close this average is to an average with respect to θ when
A is fixed; that is, about ergodicity. It was proved in [40] that indeed the average is ergodic, in
the sense that in the limit as N → ∞, the average over θ equals that over A for all but a set of
matrices of zero measure.
As has been described above, the scaling of log ΛA with respect to 12 log N leads to a central
limit theorem. What about different scalings, characterizing, for example, the large deviations of

15
log ΛA ? These were also computed in [40], and shown to agree with numerical calculations (e.g. the
behaviour seen in the tails in Figure 4) and other results known to hold for the zeta function.

6 Random matrices and moments of L-functions

Determining the value distribution of an L-function is, it turns out, a significantly harder problem
than determining the value distribution of its logarithm. Selberg’s theorem completely characterizes
the limiting distribution of log ζ(1/2 + it), while for ζ(1/2 + it) almost nothing is known.
Regarding the moments of |ζ(1/2 + it)|, there is a long-standing and important conjecture
that f (λ), defined by
Z T
1 1
lim λ 2 |ζ(1/2 + it)|2λ dt = f (λ)a(λ), (39)
T →∞ log T T 0

where ( Ã !)
Y ∞ µ
X ¶
λ2 Γ(λ + m) 2 −m
a(λ) = (1 − 1/p) p , (40)
p
m!Γ(λ)
m=0

exists, and a much-studied problem then to determine the values it takes, in particular for integer
λ (see, for example, [63, 43]). Obviously f (0) = 1. In 1918, Hardy and Littlewood proved that
f (1) = 1 [36], and in 1926 Ingham proved that f (2) = 1/12 [42]. No other values are known. Based
on number-theoretical arguments, Conrey and Ghosh have conjectured that f (3) = 42/9! [24], and
Conrey and Gonek that f (4) = 24024/16! [25].
Given the success of random matrix theory in describing the value distribution of log ζ(1/2+
it), it is natural to ask whether it has anything to contribute on this issue. Invoking the identification
(36), the question for the characteristic polynomials that is analogous to (39) is whether
Z
1
fU (λ) = lim 2 |ΛA (eiθ )|2λ dA (41)
N →∞ N λ U (N )

exists, and, if it does, what values it takes. The answer to this question was given in [52], where
the following theorem is proved.

Theorem 6 (Keating and Snaith 2000 [52])

The coefficient fU , defined as in (41), exists and is given by

G2 (1 + λ)
fU (λ) = ,
G(1 + 2λ)

where G denotes the Barnes G-function [5]. Hence fCU E (0) = 1 (trivial) and
k−1
Y j!
fU (k) =
(j + k)!
j=0

for integers k ≥ 1.

Thus, for example, fU (1) = 1, fU (2) = 1/12, fU (3) = 42/9! and fU (4) = 24024/16!. The fact
that these values coincide with those associated, or believe to be associated, with the zeta function
strongly suggests that

16
Conjecture 2 (Keating and Snaith 2000 [52])
With f (λ) and fU (λ) defined as in (39) and (41), respectively,
f (λ) = fU (λ)
for all Reλ > −1/2.
This conjecture is also supported by Odlyzko’s numerical data for non-integer values of λ between
zero and two [52].
All that has been said so far on the moments of the Riemann zeta function deals only
with the leading order coefficient, which survives the limit in (39). In recent work of Conrey,
Farmer, Keating, Rubinstein and Snaith [23], however, conjectures arrived at through heuristic
number theoretic arguments produce in a concise form all the significant lower-order terms, and
are remarkably similar to the corresponding results for random matrix characteristic polynomials.
Specifically, the 2kth moment of the zeta function and the 2kth moment of ΛA are polynomials of
degree k 2 . In the random matrix case, this polynomial can be written down explicitly using (35).
In the zeta function case, we do not know the analogue of (35). The coefficients in the polynomial
can instead be computed from the expression conjectured below. We show then that (35) can be
written in a form that is analogous to this expression. The expressions in question are written in
terms of contour integrals (in the manner suggested by [19]) and involve the Vandermonde:
Y
∆(z1 , . . . , zm ) = (zj − zi ). (42)
1≤i<j≤m

Conjecture 3 (Conrey, Farmer, Keating, Rubinstein and Snaith 2002 [23])

Z T
ζ( 12 + it + α1 ) · · · ζ( 12 + it + αk )ζ( 21 − it − αk+1 ) · · · ζ( 12 − it − α2k )dt
0
Z T 1
= t
Wk (log 2π ; α1 , . . . , αk ; αk+1 , . . . , α2k )(1 + O(t− 2 +² ))dt,
0
where

Wk (x; α1 , . . . , αk ; αk+1 , . . . , α2k )

I I
x (−1)k x Pk
=e 2
(−α1 −α2 −···−αk +αk+1 +···+α2k )
· · · e 2 j=1 zj −zj+k
k!2 (2πi)2k
G(z1 , . . . , z2k )∆2 (z1 , . . . , z2k )
× Q2k Q2k dz1 . . . dz2k ,
i=1 j=1 (zi − αj )
with the path of integration being small circles surrounding the poles αi . Here
k Y
Y k
G(z1 , . . . , z2k ) = Ak (z1 , . . . , z2k ) ζ(1 + zi − zj+k ),
i=1 j=1
and Ak is the Euler product

k µ
k Y
YY ¶Z k µ
1Y ¶−1 µ ¶−1
1 e(θ) e(−θ)
Ak (z) = 1− 1− 1− dθ,
p i=1 j=1
p1+zi −zj+k 0 j=1 p1/2+zj p1/2−zj+k

with e(θ) = exp(2πiθ).

17
This is to be compared with the following theorem in random matrix theory:

Theorem 7 (Conrey, Farmer, Keating, Rubinstein and Snaith 2002 [22])

Wk (U (N ); α1 , . . . , α2k )
Z
= ΛA (e−α1 ) · · · ΛA (e−αk )ΛA† (eαk+1 ) · · · ΛA† (eα2k )dA
U (N )
I I
N (−1)k N Pk
= e 2 (−α1 −α2 −···−αk +αk+1 +···α2k ) 2k 2
· · · e 2 j=1 zj −zj+k
(2πi) k!
Y ∆2 (z1 , . . . , z2k )
× (1 − ezq −z` )−1 × Q2k Q2k dz1 · · · dz2k .
1≤`≤k i=1 j=1 (zi − αj )
k+1≤q≤2k

The autocorrelation function in Theorem 7 is defined so as to be comparable with that in

Conjecture 3, since in the case that αi is purely imaginary, e∓αi sits on the unit circle, in analogy
with 1/2 + it ± αi lying on the critical line when αi is purely imaginary in the Riemann zeta
version. With Q these
Q definitions the structures of Wk (U (N ); α1 , . . . , α2k ) and Wk (x; α1 , . . . , α2k ) are
identical if k`=1 km=1 (1 − ezm+k −z` )−1 plays the role of G(z1 , . . . , z2k ): crucially, these functions
have poles in the same places. G(z1 , . . . , z2k ), however, contains arithmetic information about the
zeta function which clearly cannot be predicted by random matrix theory. The multiple integral
expressions for Wk (U (N ); α1 , . . . , α2k ) and Wk (x; α1 , . . . , α2k ) resemble those obtained by Brézin
and Hikami [19] for the autocorrelation functions of the characteristic polynomials of Hermitian
matrices in the limit of large matrix size N . In addition, expressions which are exact for finite N
have been derived – by Mehta and Normand [55] in the case of moments and (in this issue) Fyodorov
and Strahov [32] for autocorrelation functions – for quantities involving characteristic polynomials
of matrices from Hermitian ensembles. For the ensembles of unitary matrices autocorrelations of
characteristic polynomials have been calculated independently, using Lie theory, by Nonnenmacher
and Zirnbauer [58].
That Wk (x; 0, 0, . . . , 0) is actually a polynomial of degree k 2 can be seen by considering the
order
Q 2k−1 of the pole at zj = 0. From the numerator of the integrand we extract the coefficient of
zi , a polynomial of degree 2k(2k − 1). The Vandermonde determinant squared is a homoge-
neous polynomial of degree 2k(2k − 1). However, the poles coming from the ζ(1 + zi − zj+k ) cancel
k 2 of the Vandermonde P factors. This requires us, in computing the residue, to take, in the Taylor
expansion of exp( x2 k1 zj − zj+k ), terms up to degree k 2 .
t t 2
The fact that Wk (logR 2π ; 0, 0, . . .2k, 0) is a polynomial in log 2π of degree k 2corresponds nicely
to Wk (U (N ); 0, . . . , 0) = U (N ) |ΛA (1)| dA, which is a polynomial of degree k in N , as can be
seen from (35). As was already mentioned, equating the density of the Riemann zeros at height t
t
with the density of the random matrix eigenvalues yields the equivalence N = log 2π . After some
manipulation one can see that ak in (40) equals Ak (0, . . . , 0) and so the leading term of the integral
expression in Conjecture 3 for Wk (x; 0, . . . , 0) coincides precisely with the leading term conjectured
by Keating and Snaith described above. The expression in Conjecture 3 has recently been shown
to be connected with the analytic properties of multiple Dirichlet series [29].
Thus it appears that random matrix theory, specifically results concerning the characteristic
polynomials of random unitary matrices, leads to a conjectural solution, supported by all available
evidence, to the long-standing problem of calculating the moments of the Riemann zeta function on
its critical line. In a similar way, the ideas of Katz and Sarnak, detailed in Section 4, suggest that
averages over families of L-functions of the value at the symmetry point s = 1/2 should be predicted

18
by matrix ensemble averages over U (N ), O(2N ) or U Sp(2N ) (depending on the symmetry type of
the family of L-functions) of the characteristic polynomial evaluated at the point 1 (the symmetry
point of the eigenvalues). The results of Conrey and Farmer [21] and Keating and Snaith [51]
show that the evidence that the leading order coefficient of such mean values splits into a product
of a random matrix coefficient (calculated in the same way as fU (λ), but for the average over
the appropriate group) and a product over primes specific to the family under consideration is as
strong as that for the moments of the Riemann zeta function high on the critical line described
earlier in this section. For example, for the family of L-functions mentioned at (22) the moment is
conjectured, for large D, to be

Conjecture 4 (Conrey and Farmer 2000 [21], Keating and Snaith 2000 [51])
∗
π2 X
L(1/2, χd )k ∼ f (k)a(k)( 12 log D)k(k+1)/2 ,
6D
|d|≤D
P∗
where is over fundamental discriminants, χd (n) = ( nd ) is the Kronecker symbol, and the sum
is over all real, primitive Dirichlet characters of conductor up to D. Here
k(k+1) Ã !
Y (1 − p1 ) (1 − √1 )−k + (1 + √1 )−k
2
p p 1
a(k) = 1 +
p
1+ p
2 p

and the coefficient f (k) is

Z
1
f (k) = lim |ΛA (1)|k dA
N →∞ N k(k+1)/2 U Sp(2N )
YN
1 Γ(1 + N + j)Γ( 12 + k + j)
= lim 22N k .
N →∞ N k(k+1)/2
j=1
Γ( 12 + j)Γ(1 + k + N + j)

For integer k, this is

k
Y j!
f (k) = 2k(k+1)/2 .
(2j)!
j=1

This conjecture agrees with previous results for k = 1, 2, 3, and the case of k = 4 is almost within
reach of current methods [33, 44, 62]. A similar conjecture exists for families of L-functions with
orthogonal symmetry.
Further, in [23] the same similarities of structure as we saw in Conjecture 3 and Theorem
7 are found between values of L-functions averaged over families and autocorrelations of random
matrix characteristic polynomials. L-functions satisfy the functional equation

L(s) = εXL (s)L(1 − s), (43)

where XL (s) is a Γ-factor similar to that appearing in (2) which is specific to the L-function. ε has
modulus one and appears to provide another method of determining the symmetry type of a given
family. We note that the characteristic polynomial

ΛA (s) = det(I − As) (44)

19
of the unitary matrix A satisfies the functional equation

ΛA (s) = det A(−s)N ΛA ( 1s ). (45)

Here the transformation s → 1s plays the same role relative to the zeros of Λ(s) on the unit circle
as s → 1 − s plays relative to the zeros of L(s) on the line Res = 21 . Similarly, ε takes a role
analogous to det A. If A is drawn at random from U (N ) (with Haar measure) then det A can take
any complex value with modulus one. However, as A varies over O(2N ), the determinant can only
take the values ±1, and for A in U Sp(2N ) we always have det A = +1. In analogy to this, it
seems that when ε varies over a family of L-functions and takes complex values with |ε| = 1, then
the low-lying zeros of that family display eigenvalue statistics of the unitary group. If within a
family ε takes the value both +1 and −1, then the zeros near s = 12 are expected to behave like
the eigenvalues of matrices from O(2N ), while values ε = 1 throughout the family would imply
symplectic symmetry.
It is convenient, however, to define
1 −1
ZL (s) := ε− 2 XL 2 (s)L(s), (46)
which satisfies the functional equation
ZL (s) = ZL (1 − s), (47)
as well as the analogous characteristic polynomial
ZA (s) = e−iφ/2 (−s)−N/2 ΛA (s). (48)
The functional equation
ZA (s) = ZA ( 1s ) (49)
is then the precise analogue of (47).
For the family L(1/2, χd ) determined by the real Dirichlet characters (see (22)), averages
of ZL over the conductor d < D are conjectured to have the form [23]:

Conjecture 5 (Conrey, Farmer, Keating, Rubinstein and Snaith 2002 [23])

Suppose g(u) is a suitable weight function. Then, if F is the family of real Dirichlet L-
functions with fundamental discriminants d < 0 we have
X X∗ |d| 1
ZL ( 21 + α1 ) · · · ZL ( 12 + αk )g(d) = Qk (α, log 2π )g(|d|)(1 + O(|d|− 2 +² )),
L∈F d<0

in which
(−1)k(k−1)/2 2k 1
Qk (α, x) = (50)
k! (2πi)k
I I Q
G− (z1 , . . . , zk )∆(z12 , . . . , zk2 )2 kj=1 zj x Pk zj
× ··· k Y k
e 2 j=1 dz1 . . . dzk ,
Y
(zj − α` )(zj + α` )
`=1 j=1

where the path of integration encloses the ±α’s. Here

k
à z
!1
Y Γ( 34 + 2j )2zj 2 Y
G− (z1 , . . . , zk ) = Ak (z1 , . . . , zk ) z ζ(1 + zi + zj ),
j=1
Γ( 34 − 2j ) 1≤i≤j≤k

20
and Ak is the Euler product, which is absolutely convergent for |Rezj | < 1/2, for j = 1, . . . , k,
defined by
Y Y µ 1
¶
Ak (z1 , . . . , zk ) = 1 − 1+zi +zj
p 1≤i≤j≤k
p
  Ã !−1 Ã !−1  
Yk Yk µ ¶
 1 1 1  1 1 −1
× 1− 1 + 1+ 1 + 1+ .
2 p p
j=1 p 2 +zj j=1 p 2 +zj

There is a similar conjecture for the analogous sum over positive fundamental discriminants.
For this conjecture G− is replaced by G+ , where

k
à z
!1
Y Γ( 41 + 2j )2zj 2 Y
G+ (z1 , . . . , zk ) = Ak (z1 , . . . , zk ) z ζ(1 + zi + zj ),
j=1
Γ( 14 − 2j ) 1≤i≤j≤k

and Ak is as before.

Since the symmetry type of this family is believed to be symplectic, we compare the mean
value above with the following theorem [22]:

Theorem 8 (Conrey,Farmer,Keating,Rubinstein and Snaith 2002 [22])

We have the following integral over the unitary symplectic group with Haar measure:
Z
(−1)k(k−1)/2 2k 1
ZA (e−α1 ) · · · ZA (e−αk )dA =
U Sp(2N ) k! (2πi)k
I I Q
Y
−zm −z` −1
∆(z12 , . . . , zk2 )2 kj=1 zj N Pk zj
× ··· (1 − e ) k Yk
e j=1 dz1 · · · dzk ,
1≤`≤m≤k
Y
(zj − αi )(zj + αi )
i=1 j=1

where the contours of integration enclose the ±α’s.

We notice once again that the structure of the multiple integral Q in Theorem −z 8 is the same
as that in (50), the difference only arising in the replacement of 1≤`≤m≤k (1 − e m −z` )−1 by
G± (z1 , . . . , zk ); these two alternatives having the same set of poles as a result of the product over
zeta functions in G. Equating the density of the L-function zeros near the point s = 1/2 and the
|d|
density of the eigenvalues on the unit circle results in the equivalence N = 12 log 2π . The result in
Conjecture 5 agrees with the Conjecture 4 for the leading order term. Similar results to Theorem
8 and Conjecture 5 are presented in [23] for families displaying orthogonal symmetry. Thus these
results on the one hand lend support to the supposition that low-lying zeros of L-functions in
families follow random matrix statistics in the manner proposed by Katz and Sarnak, while on
the other illustrate the uses of random matrix theory in answering difficult number theoretical
questions.
Conjecturing the value distribution of L-functions at s = 1/2 via random matrix theory
using the techniques described above ties in to other important questions in number theory; for
instance, for a family associated with elliptic curves, the number of L-functions which vanish at
s = 1/2 is connected to the Birch and Swinnerton-Dyer Conjecture. This number can be predicted
using random matrix theory [26]. The basic idea is that the L-functions in question form an

21
orthogonal family and so their value distribution at 12 can be written down using the analogue of
Conjecture 4. It has also been shown that random matrix theory proves equally successful in the
study of the derivative of the Riemann zeta function [38, 39, 56]. It is unlikely that the uses of
random matrix theory in number theory end with the applications discussed in this review.

7 Final Remarks
The obvious question one is left with is: what is the reason for the connection between random
matrices and L-functions? It has long been imagined there might be a spectral interpretation of
the zeros. If the Generalized Riemann Hypothesis is true, such an interpretation could be the
reason why; for example, if the zeros tn of ζ(s) are the eigenvalues of a self-adjoint operator, or the
eigenphases of a unitary operator, then automatically they would all be real. Some speculations
along these lines are reviewed in [11], others have been pursued by Connes and co-workers. If the
zeros are indeed related to the eigenvalues of a self-adjoint or unitary operator, and if that operator
behaves ‘typically’, this would then suggest that the zeros might be distributed like the eigenvalues
of random matrices. Alternatively, the success of random matrix theory in describing properties of
the zeta function might be interpreted as evidence in favour of a spectral interpretation.

Acknowledgements
Many thanks to Andrew Odlyzko and Michael Rubinstein for kindly providing us with the figures
displaying the results of their numerical computations. The second author was supported by a
Royal Society Dorothy Hodgkin Fellowship.

References
[1] A. Altland and M.R. Zirnbauer, Nonstandard symmetry classes in mesoscopic normal-
superconducting hybrid structures, Phys. Rev. B, 55(2):1142–61, 1997.

[2] N. Argaman, F.M. Dittes, E. Doron, J.P. Keating, A.Y. Kitaev, M. Sieber, and U. Smilansky,
Correlations in the actions of periodic orbits derived from quantum chaos, Phys. Rev. Lett.,
71(26):4326–29, 1993.

[3] T.H. Baker and P.J. Forrester, Finite-N fluctuation formulas for random matrices, J. Stat.
Phys., 88:1371–1385, 1997.

[4] R. Balian, Nuov. Cim., B57:183, 1968.

[5] E.W. Barnes, The theory of the G-function, Q. J. Math., 31:264–314, 1900.

[6] M.V. Berry, Semiclassical theory of spectral rigidity, Proc. R. Soc. Lond. A, 400:229–251,
1985.

[7] M.V. Berry, Riemann’s zeta function: A model for quantum chaos?, In Quantum chaos and
statistical nuclear physics; editors, T.H. Seligman and H. Nishioka, pages 1–17. Lecture notes
in physics 263, Springer-Verlag, New York, 1986.

[8] M.V. Berry, Semiclassical formula for the number variance of the Riemann zeros, Nonlinearity,
1:399–407, 1988.

22
 [9] M.V. Berry and J.P. Keating, A rule for quantizing chaos?, J.Phys. A: Math. Gen., 23:4839–
49, 1990.

[10] M.V. Berry and J.P. Keating, A new asymptotic representation for ζ(1/2 + it) and quantum
spectral determinants, Proc. R. Soc. Lond. A, 437:151–73, 1992.

[11] M.V. Berry and J.P. Keating, The Riemann zeros and eigenvalue asymptotics, SIAM Rev.,
41(2):236—266, 1999.

[12] M.V. Berry and M. Tabor, Level clustering in the regular spectrum, Proc. R. Soc. Lond. A.,
356:375–394, 1977.

[13] E.B. Bogomolny and J.P. Keating, Random matrix theory and the Riemann zeros I: three-
and four-point correlations, Nonlinearity, 8:1115–1131, 1995.

[14] E.B. Bogomolny and J.P. Keating, Gutzwiller’s trace formula and spectral statistics: beyond
the diagonal approximation, Phys. Rev. Lett., 77:1472–1475, 1996.

[15] E.B. Bogomolny and J.P. Keating, Random matrix theory and the Riemann zeros II:n-point
correlations, Nonlinearity, 9:911–935, 1996.

[16] O. Bohigas and M.J. Giannoni, In Lecture Notes in Physics 209: Mathematical and computa-
tional methods in nuclear physics; editors, J.M.G.Gomez J.S. Dehesa and A. Polls, pages 1–99.
Springer Verlag, New York, 1984.

[17] O. Bohigas, M.J. Giannoni, and C. Schmit, Characterization of chaotic quantum spectra and
universality of level fluctuation, Phys. Rev. Lett., 52:1–4, 1984.

[18] O. Bohigas, R.U. Haq, and A. Pandey, Higher-order correlations in spectra of complex-systems,
Phys. Rev. Lett., 54(15):1645–48, 1985.

[19] E. Brézin and S. Hikami, Characteristic polynomials of random matrices, Communications in

Mathematical Physics, 214:111–135, 2000, arXiv:math-ph/9910005.

[20] J.B. Conrey, More than 2/5 of the zeros of the Riemann zeta function are on the critical line,
J. Reine. Ang. Math., 399:1–26, 1989.

[21] J.B. Conrey and D.W. Farmer, Mean values of L-functions and symmetry, Int. Math. Res.
Notices, 17:883–908, 2000, arXiv:math.nt/9912107.

[22] J.B. Conrey, D.W. Farmer, J.P. Keating, M.O. Rubinstein, and N.C. Snaith, Autocorrelation
of random matrix polynomials, preprint, 2002, arXiv:math-ph/0208007.

[23] J.B. Conrey, D.W. Farmer, J.P. Keating, M.O. Rubinstein, and N.C. Snaith, Integral moments
of zeta- and L-functions, preprint, 2002, arXiv:math.nt/0206018.

[24] J.B. Conrey and A. Ghosh, On mean values of the zeta-function, iii, Proceedings of the Amalfi
Conference on Analytic Number Theory, Università di Salerno, 1992.

[25] J.B. Conrey and S.M. Gonek, High moments of the Riemann zeta-function, Duke Math. J.,
107:577–604, 2001.

23
[26] J.B. Conrey, J.P. Keating, M.O. Rubinstein, and N.C. Snaith, On the frequency of vanishing
of quadratic twists of modular L-functions, In Number Theory for the Millennium: Proceedings
of the Millennial Conference on Number Theory; editor, B.C. Berndt et al. A K Peters, Ltd,
Boston, 2001.

[27] M. Coram and Persi Diaconis, New tests of the correspondence between unitary eigenvalues
and the zeros of Riemann’s zeta function, Preprint, 2002.

[28] O. Costin and J.L. Lebowitz, Gaussian fluctuation in random matrices, Physical Review
Letters, 75(1):69–72, 1995.

[29] A. Diaconu, D. Goldfeld, and J. Hoffstein, Multiple dirichlet series and moments of zeta- and
L-functions, preprint, 2001, math.nt/0110092.

[30] F.J. Dyson, Statistical theory of the energy levels of complex systems, i, ii and iii, J. Math.
Phys., 3:140–175, 1962.

[31] K. Efetov, Supersymmetry in disorder and chaos, Cambridge University Press, New York,
1997.

[32] Y.V. Fyodorov and E. Strahov, An exact formula for general spectral correlation function of
random Hermitian matrices, preprint, 2002, arXiv:math-ph/0204051.

[33] D. Goldfeld and C. Viola, Mean values of L-functions associated to elliptic, Fermat and other
curves at the centre of the critical strip, S. Chowla Anniversary Issue, J. Number Theory,
11:305–320, 1979.

[34] D.A. Goldston, On the function S(T ) in the theory of the Riemann zeta-function, Journal of
Number Theory, 27:149–177, 1987.

[35] R.U. Haq, A. Pandey, and O. Bohigas, Fluctuation properties of nuclear-energy levels – do
theory and experiment agree, Phys. Rev. Lett., 48(16):1086–89, 1982.

[36] G.H. Hardy and J.E. Littlewood, Contributions to the theory of the Riemann zeta-function
and the theory of the distribution of primes, Acta Mathematica, 41:119–196, 1918.

[37] D.A. Hejhal, On the triple correlation of zeros of the zeta function, Inter. Math. Res. Notices,
7:293–302, 1994.

[38] C.P. Hughes, Random matrix theory and discrete moments of the Riemann zeta function,
preprint, 2002, arXiv:math.nt/0207236.

[39] C.P. Hughes, J.P. Keating, and N. O’Connell, Random matrix theory and the derivative of
the Riemann zeta function, Proc. R. Soc. Lond. A, 456:2611–2627, 2000.

[40] C.P. Hughes, J.P. Keating, and N. O’Connell, On the characteristic polynomial of a random
unitary matrix, Commun. Math. Phys., 220(2):429–451, 2001.

[41] C.P. Hughes and Z. Rudnick, Linear statistics of low-lying zeros of L-functions, preprint, 2002,
arXiv:math.nt/0208230.

[42] A.E. Ingham, Mean-value theorems in the theory of the Riemann zeta-function, Proc. Lond.
Math. Soc., 27:273–300, 1926.

24
[43] A. Ivic, Mean values of the Riemann zeta function, Tata Institute of Fundamental Research,
Bombay, 1991.

[44] M. Jutila, On the mean value of L(1/2, χ) for real characters, Analysis, 1:149–161, 1981.

[45] N.M. Katz and P. Sarnak, Random Matrices, Frobenius Eigenvalues and Monodromy, AMS,
Providence, Rhode Island, 1999.

[46] N.M. Katz and P. Sarnak, Zeros of zeta functions and symmetry, Bull. Amer. Math. Soc.,
36:1–26, 1999.

[47] J.P. Keating, Periodic orbit resummation and the quantization of chaos, Proc. R. Soc. Lond.
A, 436:99–108, 1992.

[48] J.P. Keating, The Riemann zeta function and quantum chaology, In Quantum Chaos; editors,
G. Casati, I Guarneri, and U. Smilansky, pages 145–85. North-Holland, Amsterdam, 1993.

[49] J.P. Keating, Periodic orbits, spectral statistics, and the Riemann zeros, In Supersymmetry and
trace formulae: chaos and disorder; editors, I.V. Lerner, J.P. Keating, and D.E. Khmelnitskii,
pages 1–15. Plenum, New York, 1999.

[50] J.P. Keating, N. Linden, and Z. Rudnick, Random matrix theory, the exceptional lie groups,
and L-functions.

[51] J.P. Keating and N.C. Snaith, Random matrix theory and L-functions at s = 1/2, Commun.
Math. Phys, 214:91–110, 2000.

[52] J.P. Keating and N.C. Snaith, Random matrix theory and ζ(1/2 + it), Commun. Math. Phys.,
214:57–89, 2000.

[53] J. van de Lune, J.J. te Riele, and D.T. Winter, On the zeros of the Riemann zeta function in
the critical strip, IV Math. Comp., 46:667–681, 1986.

[54] M.L. Mehta, Random Matrices, Academic Press, London, second edition, 1991.

[55] M.L. Mehta and J.-M. Normand, Moments of the characteristic polynomial in three ensembles
of random matrices, J. Phys. A, 34:4627–39, 2001.

[56] F. Mezzadri, Random matrix theory and the zeros of ζ 0 (s), preprint, 2002, arXiv:math-
ph/0207044.

[57] H.L. Montgomery, The pair correlation of the zeta function, Proc. Symp. Pure Math, 24:181–
93, 1973.

[58] S. Nonnenmacher and M. Zirnbauer, Personal communication.

[59] A.M. Odlyzko, The 1020 th zero of the Riemann zeta function and 70 million of its neighbors,
Preprint, 1989.

[60] M. Rubinstein, Evidence for a Spectral Interpretation of Zeros of L-functions, PhD thesis,
Princeton University, 1998.

[61] Z. Rudnick and P. Sarnak, Principal L-functions and random matrix theory, Duke Mathemat-
ical Journal, 81(2):269–322, 1996.

25
[62] K. Soundararajan, Non-vanishing of quadratic Dirichlet L-functions at s = 12 , Ann. of Math.,
152(2):447–488, 2000.

[63] E.C. Titchmarsh, The Theory of the Riemann Zeta Function, Clarendon Press, Oxford, second
edition, 1986.

[64] H. Weyl, Classical Groups, Princeton University Press, 1946.

[65] www.hipilib.de/zeta/index.html.

26