Exact Formulas

for
Averages of Secular Coefficients
by

Andean Medjedovic

A thesis
presented to the University of Waterloo
in fulfillment of the
thesis requirement for the degree of
Master of Mathematics
in
Pure Mathematics

Waterloo, Ontario, Canada, 2021

© Andean Medjedovic 2021

 Author’s Declaration

I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis,
including any required final revisions, as accepted by my examiners.

I understand that my thesis may be made electronically available to the public.

ii
 Abstract

We study averages of secular coefficients that frequently appear in random matrix theory.
We obtain exact formulas, identities and new asymptotics for these integrals as well as a
technique to deal with singularities that classically occur in the study of these problems.

iii
 Acknowledgements

It was my great fortune to have Michael O. Rubinstein advise me through the past few
years. Thank you for the discussions, guidance and encouragement you have provided me
with throughout the program.

I would also like to thank the researchers at the American Institute of Mathematics
studying Random Matrix Theory for their insights and lectures on the field, as well as their
general fellowship.

Lastly, I would like to thank my parents, Ed and Sanja, and my brother, Denny, for
their love and support.

iv
 Dedication

Dedicated to my parents, Ed and Sanja Medjedovic.

v
Table of Contents

0.1 List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

0.2 List of Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

1 Introduction 1
1.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The polynomials γk (c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Symmetric Function Theory 8

2.1 Young Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Singularity Removal For Moments . . . . . . . . . . . . . . . . . . . . . . . 10

3 Secular Coefficients of Matrix Groups 12

3.1 The Unitary Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 The Symplectic Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 The Orthogonal and Special Orthogonal Group . . . . . . . . . . . . . . . 18
3.3.1 The Orthogonal Group . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3.2 The Special Orthogonal Group . . . . . . . . . . . . . . . . . . . . 20

4 Asymptotic Behavior of the Unitary Group & Lower Order Terms 21

4.1 Analysis by Minors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1.1 A recursion for FN,k (x) . . . . . . . . . . . . . . . . . . . . . . . . . 24

vi
5 Further Properties 25
5.1 Unimodality of γk (c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

6 Conclusions 28

References 30

APPENDICES 32

A Tables and Plots 33

A.1 Tables of γkG (c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
A.1.1 Unitary Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
A.1.2 Symplectic Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
A.1.3 Orthogonal Group . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
A.2 Plots of γkG (c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

vii
0.1 List of Tables
A.1.1 + Fig. A.1 Unitary
A.1.2 + Fig. A.2 Symplectic
A.1.3 + Fig. A.3, A.4 Orthogonal

viii
0.2 List of Notation
P∞
(i) ζ(s) = 1
n=1 ns is the Riemann zeta function.

(ii) δ(x) is the Dirac delta function.

(iii) dk (n) = n1 n2 ...nk =n 1 is the k-fold divisor function. It is the number of ways to write
P
n as a product of k natural numbers.

(iv) RM T is an abbreviation for Random Matrix Theory.

(v) N T is an abbreviation for Number Theory.

(vi) SSY T is an abbreviation for Semi-Standard Young Tableaux. A construct in parti-

tion theory which we will properly define later.

(vii) λ is a partition. λi is the ith part of the partition. s(λ) is the size of the partition
and λ0 is the conjugate partition to λ.

(viii) U (N ), SP (2N ), O(N ), SO(N ) are the unitary, symplectic, and (special) orthogonal
matrix groups.

(ix) χG is a character on the group G.

ix
Chapter 1

Introduction

The goal of this thesis is to study the random matrix theory analogue of moments of L-
functions. In particular, we develop a theory of averages of powers of determinants over
matrix groups. Certain properties of these determinants have been studied by Keating,
Rogers, Roditty-Gershon and Rudnick [12], Bump and Gamburd [8], as well as one of the
authors [3]. These averages have long been known to be related to conjectures for asymp-
totics of higher moments of the ζ function [10].

1.1 Outline
• We motivate the study of a class of functions and so called “Secular Coefficients”. We
begin by reviewing known results for the unitary case in the rest of the introduction.
We define and generalize the set of polynomials known within the literature as γk (c).
We summarize all the results contained in this thesis.

• In the next section, we briefly review some symmetric function theory and partition
theory. We prove a Lemma the will be invaluable in our investigation that will
allow us the remove certain singularities that classically appear in the study of these
averages of characteristic polynomials of random matrices.

• We apply this Lemma along with results from Bump-Gamburd [8] as well as enumer-
ations coming from the theory of plane partitions to get exact determinant formulas
for averages of determinants of random matrices. We can use these ideas to deal with

1
 a wide case of matrix families, the classical groups. This is the main achievement of
the thesis.

• We then further analyze the Unitary case, obtaining properties of lower order terms
of γk (c).

• We give a short proof of the unimodality of γk (c), which was conjectured by Ze’ev
Rudnick.

• Lastly, we succinctly summarize further relations between the Riemann ζ function

and averages of functions over random matrix groups.

The motivation is that we are trying to understand moments of the zeta function. We
begin with taking powers of ζ, and we have the following identity for the divisor function.
Let dk (n) be the k-th divisor numbers, i.e. the Dirichlet coefficients of the k-th power of
the Riemann zeta function:
∞
X dk (n)
ζ(s)k = , <s > 1. (1.1)
1
ns

The Dirichlet coefficient dk (n) is equal to the number of ways of writing n as a product of
k factors. Define
X
Sk (X) = dk (n). (1.2)
n≤X

The main term in the asymptotics of Sk (x) comes from the pole at s = 1 of ζ k (s). Let
XPk−1 (log X) be the residue, at s = 1 of ζ(s)k X s /s, with Pk−1 (log X) being a polynomial
in log X of degree k − 1. Then

Sk (X) = XPk−1 (log X) + ∆k (X), (1.3)

with ∆k (X) denoting the remainder term. The k-divisor problem asserts that ∆k (x) =
k−1
Ok (x 2k + ). It is this remainder term that needs to be understood further.
The behaviour of ∆k in short intervals was studied by Keating, Rodgers, Roditty-
Gershon, and Rudnick [12]. Let

∆k (x; H) = ∆k (x + H) − ∆k (x) (1.4)

be the remainder term for sums of dk over the interval [x, x + H].

2
 Define ∞  2
Yn 1 2X Γ(k + j) 1o
ak = (1 − )k . (1.5)
p
p j=0
Γ(k)j! pj

the product convergence is seen by expanding the terms with respect to p giving a
product over 1 − pC2 + O( p13 ), where C is a constant in k. By considering the analogous
problem for function fields and related random matrix theory statistics, Keating, Rodgers,
Roditty-Gershon, and Rudnick conjectured [12]:

Conjecture 1. If 0 < α < 1 − 1

k
is fixed, then for H = X α ,
Z 2X
1  2 2
∆k (x, H) dx ∼ ak Pk (α)H(log X)k −1 , X→∞ (1.6)
X X

where Pk (α) is given by

2 −1 1
Pk (α) = (1 − α)k γk ( ). (1.7)
1−α

Here γk (c) is a piecewise polynomial function defined in the next section. Thereby, we
hope to gain a better understanding of the statistics of the k-divisor function by under-
standing the general theory of γk (c) and related constructions.

We briefly touch on the results found by Keating et al. and how they connect not only
RMT and NT, but analogous questions for function fields.

Let U be an N × N matrix. We define the secular coefficients, Scj (U ), to be the

coefficients of the characteristic polynomial of U :
N
X
det(I + xU ) = Scj (U )xj . (1.8)
j=0

Thus Sc0 (U ) = 1, Sc1 (U ) = tr U , ScN (U ) = det U . The secular coefficients are just ele-
mentary symmetric functions in the eigenvalues of U .

Let G be one of the matrix groups U (N ), Sp(2N ), SO(N ) or O(N ). Working with
respect to the natural Haar measure in each case, define, for G = Sp(2N ), SO(N ), or
U (N ),

3
 Z X
IkG (n, N ) := Scj1 (U ) . . . Scjk (U )dU. (1.9)
G j +···+j =n
1 k
0≤j1 ,...,jk ≤N

Unless G = U (N ) where we introduce a conjugate term, squaring the integrand (oth-

erwise the average becomes 0):
Z X
IkG (n, N ) := |Scj1 (U ) . . . Scjk (U )|2 dU. (1.10)
G j +···+j =n
1 k
0≤j1 ,...,jk ≤N

The connection to function field theory needs some additional notation. Let f be
a monic polynomial in Fq and use dk (f ) to denote the number of ways to write f as
f = f1 . . . fk with fi monic. We assume that the index A is a monic polynomial in Fq .
Furthermore, for a monic, define

I(A; h) = {f : ||f − A|| ≤ q h } (1.11)

with ||f || = q deg(f ) and X

N (A; h) := dk (f ) (1.12)
f ∈I(A;h)

to be the divisor sum in function fields. Defining the difference and variance in short
intervals similarly,  
h+1 n + k − 1
∆k (A; h) := N (A; h) − q , (1.13)
k−1
1 X
Var(N ) := n |∆k (A; h)|2 . (1.14)
q
deg(A)=n

We then have the following estimate of the function field variance:

Theorem 1 (KRRR). If 0 ≤ h ≤ min(n − 5, (1 − k1 )n − 2), then as q → ∞

H
Var(N ) = H · IkG (n; n − h − 2) + O √ , (1.15)
q

for H = q h+1 .

4
 In this case H is comparable to the short interval X a in the NT case.
The following result in this direction is the following theorem due to Keating et al [12]
which gives the leading asymptotics of IkG in terms of γk (c).

Theorem 2 (KRRR). Let c := m/N . Then for c ∈ [0, k],

U (n) 2 −1 2 −2
Ik (m, N ) = γk (c)N k + Ok (N k ). (1.16)

1.2 The polynomials γk (c)

The function γk (c), mentioned in Conjecture 1 and Theorem 2, is defined by the following
integral over a slice of the unit hyper-cube:
Z
1 Y
γk (c) = δ(t1 + . . . + tk − c) (ti − tj )2 dt1 . . . dtk , (1.17)
k! G(1 + k)2 [0,1]k i<j

where G is the Barnes G-function, so that for positive integers k, G(1 + k) = 1! · 2! ·

3! · · · (k − 1)!.
The function γk (c) is supported on [0, k] and symmetric around k2 .

γk (c) = γk (k − c) (1.18)

It is also known that

Theorem 3 (KRRR).
X k 2 2 2
γk (c) = (c − `)(k−`) +` −1 gk,` (c − `) (1.19)
0≤`<c
`

where gk,` (c − `) are polynomials in c − `. No explicit form for gk,` is currently known. Note
that the above implies that on each interval [j − 1, j], (for integer j), γk (c) is a polynomial.

While the motivation in studying γk (c) from a number theoretic perspective comes
primarily from the connection to divisor sums, they are of their own interest from the
perspective of random matrix theory. The focus of our thesis is on the underlying random
matrix theory.

5
1.3 Main Results
The main results of this thesis are determinant identities for the generating function of
IkG (n, N ). No exact formulas for these generating functions are known in the literature.
Let G ∈ {U (N ), O(N ), SP (2N ), SO(N )} be a matrix group and consider
∞
X
G
Pk,N (u) = un IkG (n, N ).
n=0

Then if G = U (N )

Theorem 4.
G CN,k 1 − uN +i+j−1
Pk,N (u) = 2 det
(1 − u)k N +i+j−1
with
k
Y (N + k − j − 1)!
CN,k = .
j=1
(j − 1)!2 (N + j − 1)!

If G = SP (2N ) then

Theorem 5.
G
Pk,N (u) =
    
1 j − 1 j−i 2N + 2k + 1 − j 2N +2k+2−j−i
k+1
det u − u .
(1 − u2 )( 2 ) 1≤i,j≤k i−1 i−1

And finally, if G = O(N ) or G = SO(N ) we have

Theorem 6.
G 1 1
Pk,N (u) =
2 (1 − u2 )(k2)
    
j − 1 j−i 2N + 2k − 1 − j 2N +2k−j−i
det u − u
i−1 i−1
    
j − 1 j−i 2N + 2k − 1 − j 2N +2k−j−i
+ det u + u .
i−1 i−1

6
and
G
Pk,N (u) =
    
1 j − 1 j−1 2N + 2k − j − 1 2N +2k−j−i
k
det u + u ,
(1 − u2 )(2) i−1 i−1

respectively.

The secondary results of this thesis are slightly more qualitative results. In Section
U (N )
4 we prove that the lower order terms in the asymptotics for Ik in N have properties
U (N )
similar to γk (c). That is to say, if Ik (cN, N ) ∼ m=0 γk,m (c)N k2 −1−m
then:
P

1. γk,m (c) is symmetric around k/2.

2. γk,m (c) is supported on [0, k] and on each interval [j, j + 1] (for j an integer) it is a
polynomial.

3. Each polynomial piecewise composing γk,m (c) is of degree at most k 2 − m.

4. γk,m (c) is differentiable k 2 − m − 2j(k − j) − 1 times at a transition point c = j.

For example, γk,0 (c) = γk (c) and has exactly the above properties.
In section 5 we prove a conjecture of Ze’ev Rudnick [personal communication], that
γk (c) is unimodal.

7
Chapter 2

Symmetric Function Theory

In this section we introduce some basics of symmetric function theory. The connection
to symmetric function theory was used independently by Conrey, Farmer, Keating, Ru-
binstein and Snaith in CFKRS[9] as well as Bump and Gamburd in BG[8] to determine
moments of characteristic polynomials of the classical compact groups. These results were
used in CFKRS[9] to conjecture the asymptotics of the shifted moments of the ζ-function.
We will describe the relevant symmetric function theory need for our results.

2.1 Young Diagrams

Let λ = (λ1 ≥ λ2 ≥ . . . ≥ λk ) be a partition of n. Then λ1 + λ2 + . . . + λk = n. To
each partition λ we associate to it what is known as a Ferrer’s diagram. The diagram
is a collection of “cells” off length λi across. For example the partition of 14 given by
(5, 4, 2, 2, 1) corresponds to Ferrer’s diagram

8
 We say a Ferrer’s diagram is a semi-standard young tableau when the cells are labeled
by integers less than n in such a way so that the rows are non-decreasing and the columns
are increasing, starting with 1 at the top-right most cell. A young tableau for the above
would be:
1 2 2 3 4
3 4 4 4
4 5
5 7
7

We say such a semi-standard young tableau, T , is of shape λ if the Ferrer’s diagram of

the tableau is the Ferrer’s diagram for λ. In which case we write T ∼ λ.
We should also introduce the Schur polynomials sλ (x1 , . . . , xk ), let ∆(x) be the deter-
minant of the Vandermonde matrix:
Y
∆(x) = det xi−1 j = (xi − xj ). (2.1)
1≤i,j≤k
i6=j

We define the Schur polynomial of λ to be

 λ1 +k−1 λ1 +k−1
. . . xλk 1 +k−1

x1 x2
xλ2 +k−2 xλ2 +k−2 . . . xλ2 +k−2 
 1 2 k
det  .. .. . .. 

 . . . . . 
λk λk λk
x1 x2 ... xk
sλ (x1 , . . . , xk ) = . (2.2)
∆(x)
Notice that sλ is actually a polynomial as the determinant is 0 when xj = xk for any j, k,
canceling with the pole from the Vandermonde factor in the denominator. This definition
of the Schur-functions is concise but unintuitive. An alternate definition follows.
We say T has type a = (a1 , a2 , ...) if T has ai = ai (T ) parts equal to i. The SSYT above
has type (1, 2, 2, 5, 2, 0, 2). It is common to use the notational abbreviation
a (T ) a2 (T )
xT = x1 1 x2 ··· ,
so for the example SSYT above,
xT = x11 x22 x23 x54 x25 x27 .
We finally come to the combinatorial definition of Schur functions.

9
Definition 1. For a partition λ, the Schur function in the variables x1 , ..., xr indexed by
λ is a multivariable polynomial defined by
X a (T )
sλ (x1 , ..., xr ) := x11 · · · xar r (T ) ,
T

where the sum is over all SSYTs T whose entries belong to the set {1, ..., r} (i.e. ai (T ) = 0
for i > r).

For example, the SSYTs of shape (4, 2) whose entries belong to the set {1, 2} are

1 1 1 1 1 1 1 2 1 1 2 2
2 2 2 2 2 2

and so
s(4,2) (x1 , x2 ) = x41 x22 + x31 x32 + x21 x42 .

Nota bene, the value sλ (1, . . . , 1) enumerates the total number of SSYT associated to
the partition λ.

2.2 Singularity Removal For Moments

Consider a polynomial P (x) given by
(2.3)
 ai 
P (x) = det xj−1 ,
1≤i,j≤k

where ai are non-negative integers. Then,

k−1
a
X Y
P (x) = P (x0 , . . . , xk−1 ) = sgn(σ) xi σ(i) . (2.4)
σ∈Sn i=0

This is an alternating polynomial and thus divisible by ∆(x). We are interested in

finding ∆(x)
P (x)
when x0 = x1 = x2 = . . . = xk−1 = u. Taking the limit as x1 → x2 , x2 → x3 ,
etc. and applying L’ Hopital’s rule gives
P (x) 1 ∂ k−1 ∂2 ∂
lim = · · · P (x). (2.5)
x→(u,...,u) ∆(x) 1!2! . . . (k − 1)! ∂xk−1
k−1
∂x2 2 ∂x1 (u,...,u)

10
 We expand P (x) according to its definition taking derivatives and matching i! with the
aσ(i) terms to get binomial coefficients.
k−1
Y aσ(i)  aσ(i) −i
P (x) X
lim = sgn(σ) xi (u,...,u)
. (2.6)
x→(u,...,u) ∆(x) i
σ∈S ni=0

And we have computed the removable singularities of P (x)

∆(x)
to be

k−1
Y 
P (x) X aσ(i) aσ(i) −i
(u,...,u)
= sgn(σ) xi (u,...,u)
(2.7)
∆(x) σ∈Sn i=0
i
  
aj aj −i+1
= det x (2.8)
1≤i,j≤k i − 1 i−1 (u,...,u)
  
aj
= det uaj −i+1 . (2.9)
1≤i,j≤k i−1

We can extend this theorem slightly in the following Lemma.

Lemma 1. Let P (x) = det1≤i,j≤k [pj (xi−1 )] be an alternating polynomial where each pj is
itself a polynomial. Then

1 ∂ i−1
 
P (x)
|(u,...,u) = det pj (u) . (2.10)
∆(x) 1≤i,j≤k i − 1! ∂ui−1

Proof. If each pj is a monomial then the proof is detailed above. In the case that pj are
not monomials we may split up the determinant as a sum of monomials by multi-linearity
and apply the above recipe on each term individually. Adding the terms together by
multi-linearity again yields Lemma 1.

This Lemma will be crucial in removing singularities that appear in expressions for
averages of secular coefficients. This will allow us to get an exact formula for certain
matrix theory integrals that appear in the literature.

11
Chapter 3

Secular Coefficients of Matrix Groups

3.1 The Unitary Group

We will apply the singularity removal technique to equation (2.9) in Autocorrelations of
Random Matrix polynomials [9]. That formula is reproduced below in equation (3.3). Let
G = U (N ) and let U ∈ U (N ). First notice the following relation

N
!k N
!k
X X
det(I − xU )k det(I − yU ∗ )k = Scj (U )(−x)j Sci (U ∗ )(−y)i (3.1)
j=1 i=1

and integrate over the unitary group.

Z X
det(I − xU )k det(I − yU ∗ )k dU = IkG (n, N )(xy)n . (3.2)
G 0≤m≤kN

In the above equation only diagonal terms remain, i.e. the coefficients of the terms of
form xn y m , m 6= n, are 0. Consider the map U 7→ eit U which by the invariance of the Haar
measure does not change the value of the integral. Under this map, U ∗ gets scaled by e−it .
We can absorb the eit terms in x and e−it in y so that the term xn y m in the sum becomes
e(n−m)it xn y m . Since the integral is invariant under this transformation, the sum should be
too, and so the coefficient of any term with n 6= m is indeed 0.

12
 Formula (2.9) of the Autocorrelations paper is copied below:

r
Y Z n
Y m
Y
wlN det(I − wi−1 U ) det(I − wj U ∗ )dU (3.3)
l=m+1 U (N ) i=m+1 j=1

1 w1 w12 · · · w1m−1 w1N +m w1N +m+1 · · · w1N +n−1

1 .. .. .. . . .. .. .. .. ..
=Q . . . . . . . . . .
1≤`<q≤n (wq − w` ) 2 m−1 N +m N +m+1 N +n−1
1 wn wn · · · wn wn wn · · · wn

Specializing to m = k, n = 2k, w1 = w2 = . . . = wk = x and wk+1 = . . . = w2k = 1 and

removing the singularities as in Lemma 1 gives

U (N ) 1 A(x) B(x)
Ik (n, N ) = [xn ] (3.4)
(1 − x)k2 A(1) B(1)

where

 
j − 1 j−i
Aij (x) = x (3.5)
i−1
 
N + 2k + j − 1 N +2k+j−i
Bij (x) = x . (3.6)
i−1

We now are going to perform row reductions on the above. Notice A(x)−1 = A(−x),
as can be verified using the underlying binomial identity

k      k  
X l−1 j−1 j−1 X i−l j − i
i+l
(−1) = (−1) . (3.7)
l=1
i−1 l−1 i − 1 l=1 l−i

If j > i the sum on the right is an alternating

sum of the (j − i) row of Pascal’s
th

triangle and vanishes. If j < i the factor of i−1 infront of the sum is 0. And if i = j
j−1

only one term contributes to the sum, namely l = i, giving 1. Thus, multiplying the block
matrix in 3.4 on the left by the block matrix
 
A(−x) 0
(3.8)
0 A(−1)

13
 gives
    
A(−x) 0 A(x) B(x) I A(−x)B(x)
= , (3.9)
0 A(−1) A(1) B(1) I A(−1)B(1)

and then by multiplying with

 
I 0
(3.10)
−I I

to remove the bottom left I:

    
I 0 I A(−x)B(x) I A(−x)B(x)
= . (3.11)
−I I I A(−1)B(1) 0 A(−1)B(1) − A(−x)B(x)

These multiplications do not change the determinant as both multiplications are by

triangular matrices with 1’s on the diagonal. Therefore the determinant of the matrix in
(3.4) equals the determinant of the lower k × k block above, i.e.

|A(−1)B(1) − A(−x)B(x)|k×k . (3.12)

Next we compute the entries of the above matrix. The i, j entry is

k   
l−i l − 1 N +k+j−1
X
(−1) (1 − xN +k+j−i ) (3.13)
l=1
i − 1 l − 1

but,      
l−1 N +k+j−1 N +k+j−1 N +k+j−i
= . (3.14)
i−1 l−1 i−1 l−i
so that equation 3.13 equals
  k  
N +k+j−1 X
l N +k+j −i
(1 − x N +k+j−i
)(−1)i
(−1) . (3.15)
i−1 l=1
l − i

But the sum above equals

k−i  
X N +k+j−i
(−1) i
(−1) l
. (3.16)
l=0
l

14
 This is an alternating sum of the N + k + j − i row of Pascal’s triangle which so the
above famously equals  
k N +k+j −i−1
(−1) . (3.17)
k−i
Returning to the k × k determinant we see that the i, j entry of the matrix equals
  
k−i N + k + j − 1 N +k+j−i−1
(−1) (1 − xN +k+j−i ). (3.18)
i−1 k−i
This product of binomial coefficients equals
  
N +k+j−1 N +k+j−i−1 (N + k + j − 1)!
= .
i−1 k−i (i − 1)!(k − i)!(N + j − 1)!(N + k + j − i)
(3.19)
(−1)k−i
We can thus pull out from row i of the determinant a factor of ((i−1)!(k−i)!) and a factor of
(N +k+j−1)!
(N +j−1)!
from column j. Therefore, the determinant in (3.4) equals, on collecting these
factors,
k
(−1)k−j (N + k + j − 1)! 1 − xN +k+j−i
Y  
det = (3.20)
j=1
(j − 1)!(k − j)!(N + j − 1)! N + k + j − i k×k
k
1 − xN +i+j−1
 
Y (N + k + j − 1)!
det (3.21)
j=1
(j − 1)!2 (N + j − 1)! N + i + j − 1 k×k
where, in the last equality we have reversed the k rows of the matrix. We have thus arrived
at the formula of Theorem 4:
U (N ) CN,k 1 − xN +i+j−1
Ik (n, N ) = [x ] n
det . (3.22)
(1 − x)k2 N +i+j−1
Here CN,k is a constant depending only on N and k and can be given explicitly in
several ways:
k Q
(N + k + j − 1)! 1≤i,j≤k (N + i + j − 1)
Y
CN,k = 2 (N + j − 1)!
= Q 2
(3.23)
j=1
(j − 1)! 1≤i<j≤k (j − i)
1
CN,k = 1 (3.24)
det1≤i,j≤k [ N +i+j−1 ]
G(N + 2k)G(N )
CN,k = (3.25)
G(N + k)2 G(k)2
where G(m) = 1!2! . . . (m − 1)! is the Barnes G-function.

15
3.2 The Symplectic Group
We move on the symplectic case now. Let G = SP (2N ). We begin with proposition (11)
and equation (43) from Bump-Gamburd [8].

Z k
Y Sp(2k)
det (1 + xi U ) dU = (x1 . . . xk )N χ N k (x±1 , . . . , x±1
k ). (3.26)
Sp(2N ) i=1
h i 1

Sp(2k)
Here χ is a certain irreducible character from the representation theory of GLn (C).
hN k i
A partition is said to be even if all parts of it are even. From section 7.1 of the same paper
we have X
Sp(2k)
(x1 . . . xk )N χ<N k > (x±1
1 , . . . , x ±1
k ) = sλ (x1 , . . . , xk ). (3.27)
λ1 ≤2N
λ even

where the sum is taken over all even partitions.

Let G = Sp(2N ).
Consider the generating function
2kN
X Z
xn IkG (n, N ) = det(1 + xU )k dU. (3.28)
n=0 Sp(2N )

We are trying to extract the [xn ] coefficient of

k
X z }| {
sλ (x, . . . , x) . (3.29)
λ1 ≤2N
λ even

By the combinatorial interpretation of Schur functions the coefficient we desire is

k
X z }| {
sλ (1, . . . , 1) . (3.30)
s(λ)=n
λ1 ≤2N
λ even

k
z }| {
where s(λ) is the size of the partition. One can see that sλ (1, . . . , 1) as the number of semi-
standard young tableaux of type λ. Hook content formula gives sλ (1, ..., 1) = u∈λ n+c(u)
Q
h(u)

16
where c(u) and h(u) are the content and hook of a cell u ∈ λ.

Other identities for partitions of the form described in equation (3.29) are well-known
within literature dealing with plane partitions. A famous example is the Hall-Littlewood
identity [1].
k
X Y 1 Y 1
sλ (x1 , . . . , xk ) = 2
(3.31)
i=1
1 − xi i<j 1 − xi xj
λ even

Note that if n < 2N then the constraint from our formula drops out and the Hall-
Littlewood identity allows us to immediately calculate
 n k+1

 2 +( 2 )−1 , for n even
Sp(2n)
Ik (n, N ) = (k+1
2 )
−1 . (3.32)
0, otherwise

In other domains we must use bounded forms of the Hall-Littlewood identities. For
this we use the Desarmenien-Stembridge-Proctor formula [14], [4] , [5].
k
X 1 Y 1 Y 1 h i
sλ (x1 , . . . , xk ) = det x j−1
− x 2N +2k+1−j
(3.33)
λ1 ≤2N
∆(x) i=1 1 − x2i i<j 1 − xi xj 1≤i,j≤k i i

λ even

where ∆(x) = i<j (xi − xj ) is the Vandermonde determinant. The difficulty here is
Q
singularities appear when all xi are equal. Of course, since we are ultimately dealing with
a finite sum of polynomials , these singularities must be removable.
We now apply the formula derived in Lemma 1 above to the Desarmenien-Stembridge-
Proctor formula.

k
1 Y 1 Y 1 h
j−1 2N +2k+1−j
i
det x − x =
∆(x) i=1 1 − x2i i<j 1 − xi xj 1≤i,j≤k i i (u,...,u)

h i
j−1 2N +2k+1−j
1 det 1≤i,j≤k x i − x i
k+1 (u,...,u)
(3.34)
(1 − u2 )( 2 ) ∆(x)

In this case, since we are not working with monomial terms anymore the determinant
expression gets more complicated but we can decompose it by multi-linearity and then

17
apply the above formula to get rid of the 1
∆(x)
, putting everything back together again with
multi-linearity.
h i
det1≤i,j≤k xj−1
i − x 2N +2k+1−j
i
(u,...,u)
(3.35)
∆(x)
1 X X Y Y
(−1)|S| sgn(σ)
2N +2k+1−σ(i) σ(i)−1
= xi xi (u,...,u)
(3.36)
∆(x) σ∈S i∈S i6∈S
n S⊂{1,...,k}
X X sgn(σ) Y 2N +2k+1−σ(i) Y σ(i)−1
= (−1)|S| x xi (3.37)
σ∈Sn
∆(x) i∈S i i6∈S
(u,...,u)
S⊂{1,...,k}
    
j − 1 j−i 2N + 2k + 1 − j 2N +2k+2−j−i
= det u − u (3.38)
1≤i,j≤k i−1 i−1

To summarize, if we let
2kN
X Sp(2n)
Pk,N (u) = un Ik (n, N ). (3.39)
n=0

Then we have the following formula of Theorem 5:

Pk,N (u) =
    
1 j − 1 j−i 2N + 2k + 1 − j 2N +2k+2−j−i
k+1
det u − u .
(1 − u2 )( 2 ) 1≤i,j≤k i−1 i−1

3.3 The Orthogonal and Special Orthogonal Group

In this section we use similar ideas to the previous section to deal with the G = SO(2N )
and G = O(2N ) case.

3.3.1 The Orthogonal Group

Let G = O(2N ). Our starting point is again
Z X
IkG (n, N ) := Scj1 (U ) . . . Scjk (U )dU (3.40)
G j +···+j =n
1 k
0≤j1 ,...,jk ≤N

18
 for a matrix group G.
Consider the generating function
2kN
X Z
xn IkG (n, N ) = det(1 + xU )k dU. (3.41)
n=0 G

Again, we refer to Bump-Gamburd for the first step. In equation 102, after specializing
to xi = xj for all i, j they give
Z X
det(I + xU )k dU = sλ (x, . . . , x). (3.42)
G λ1 ≤2N
λ0 even

where λ0 is the conjugate partition of λ. As before, if we want IkG (n, N ) we can isolate
the xn term of the above as X
sλ (1, . . . , 1), (3.43)
s(λ)=n
λ1 ≤2N
λ0 even
the total number of SSYT of partitions with even conjugate. Okada [7] gives an enumera-
tion of such sums and we will apply our Lemma 1 to remove the singularities:

X 1 det(xj−1 − xi2N +2k−1−j ) + det(xj−1 + x2N +2k−1−j

)
sλ (x1 , . . . , xk ) = i
Q i i
(3.44)
λ1 ≤2N
2 1≤i<j≤k (xi xj − 1)(xi − xj )
λ0 even

Let
2kN
X
Pk,N (u) = un IkG (n, N )
n=0

be the polynomial whose coefficients enumerate the averages we are after. Setting all
xi = u and using Lemma 1 the resulting sum of determinants gives the first formula of
Theorem 6.
1 1
Pk,N (u) =
2 (1 − u2 )(k2)
     
j − 1 j−i 2N + 2k − 1 − j 2N +2k−j−i
det u − u
i−1 i−1
    
j − 1 j−i 2N + 2k − 1 − j 2N +2k−j−i
+ det u + u .
i−1 i−1

19
3.3.2 The Special Orthogonal Group
Let G = SO(2N ) and keep the same notation as the previous subsection. The special
orthogonal case is a little easier to handle. Equation 71 in Bump-Gamburd gives a relation
for the integral we want in terms of a matrix

k
Z Y
det(I + xj g) = (x1 . . . xk )N χON2kk (x±1 , · · · , x±1
k ) (3.45)
G j=1
h i 1

Where the character χ can be written explicitly as

O(2k) ±1
(x1 . . . xk )N χN k (x±1
1 , · · · , xk ) =

−(N +k−1) −(N +k−2) −(N )

x1N +k−1 + x1 x1N +k−2 − x1 · · · xN
1 − x1
.. .. .. ..
det . . . .
−(N +k−1) −(N +k−2) −(N )
(3.46)
xkN +k−1 − xk xkN +k−2 − xk N
· · · xk − xk

(x1 · · · xk )k+N −1
×Q .
16i<j6k (xi − xj )(xi xj − 1)

If we let
2kN
X
Pk,N (u) = un IkG (n, N )
n=0

the consequently (after an application of Lemma 1) we obtain, the second formula in

Theorem 6:

Pk,N (u) =
    
1 j − 1 j−1 2N + 2k − j − 1 2N +2k−j−i
k
det u + u .
(1 − u2 )(2) i−1 i−1

20
Chapter 4

Asymptotic Behavior of the Unitary

Group & Lower Order Terms

4.1 Analysis by Minors

Let
xN +i+j−1 − 1
 
FN,k (x) := det .
1≤i,j≤k N +i+j−1
This is, up to sign, the determinant that occurs in Theorem 4, the unitary case, though we
prefer here to write the numerator as xN +i+j−1 − 1. Our goal is to get an understanding
of the asymptotic behavior of this determinant so we can get higher order analogues of γk (c).

We expand the above determinant as a sum of its minors. Imagine choosing sets
S, T ⊂ {1, . . . , k} that denote rows/columns where we choose powers of x in our power
series expansion of F and what remains is the minor S c , T c . Each minor is a CauchyP matrix
and there are known formulas for computing these determinants. Let s(S) = a∈S a, the
sum of elements of S.
xN +i+j−1
   
X
s(S)+s(T ) −1
FN,k (x) = (−1) det det .
i∈S,j∈T N + i + j − 1 i∈S c ,j∈T c N + i + j − 1
S,T ⊂{1,...,k}
|S|=|T |
(4.1)
The determinant on the right hand side that is dependent on x is homogeneous. A
more general version of this formula can be found in [13].

21
 FN,k (x) =
   
X
k−|S|+s(S)+s(T ) (N −1)|S|+
P P
i+ j∈T j 1 1
(−1) x i∈S det det .
i∈S,j∈T N +i+j−1 i∈S c ,j∈T c N +i+j−1
S,T ⊂{1,...,k}
|S|=|T |
(4.2)

We now make use of Cauchy’s determinant formula.

Theorem 7 (Cauchy). Let A = {α1 , . . . , αk }, B = {β1 , . . . , βk }. Then

 
∆(A)∆(B) 1
det , =
αi + βj
P (A, B)
where ∆(S) = i<j (si − sj ) and P (S + T ) = s∈S,t∈T (s + t).
Q Q

Let N + S denote the set obtained by adding the integer N to each element of S.
Likewise, let T − 1 be the set obtained by subtracting 1 from each element of T . Applying
this to the product of two minors in our expression for FN,k with A = N + S and B = T − 1
and noticing we can factor out CN,k , using 3.24 yields

∆(S)∆(T )∆(S c )∆(T c )

   
1 1
det det =
i∈S,j∈T N + i + j − 1 i∈S c ,j∈T c N + i + j − 1 P (N + S, T − 1)P ((N + S)c , (T − 1)c
(4.3)
1 P (N + S, (T − 1) )P ((N + S)c , T − 1)
c
= .
CN,k P (S, −S c ) P (T, −T c )
(4.4)

In the first equality we used ∆(N + S) = ∆(S) and ∆(T − 1) = ∆(T ) and likewise for
their complements, S c , T c . In the second equality we factor out the CN,k
1
and are left with
the remaining products. To proceed multiply the polynomial FN,k (x) by the power series
(−1)k C
of (1−x)kN,k
2 . The x
n
coefficient of the resulting polynomial is

n
k2 − 1 + n − m
X  
k
(−1) CN,k [xm ]FN,k (x) (4.5)
m=0
k2 − 1

22
 For given k, if N is sufficiently large, notice that powers in the above polynomial cluster
around jN for an integer j ≤ k. That is, all non-zero terms in FN,k that involve terms xm
for m = jN + l with l being an integer less than k 2 . Let jN ≤ n = cN ≤ (j + 1) N so the
above becomes

c X 2
k  2 
X k − 1 + (c − j) N − l
(4.6)
j=0 l=0
k2 − 1
X |S|+s(S)+s(T ) P (S + N, (T − 1)c )P ((S + N )c , T − 1)
× (−1)
P (S, −S c ) P (T, −T c ) .
S,T ⊂{1,...,k}
P |S|=|T P|
(N −1)|S|+ s∈S s+ t∈T t=jN +l

We can take note of the following properties from the above formula. As c passes
through integers 1, 2, . . . k new terms are added to the above double sum. These terms
are a polynomial in (c − j). Suppose we want to know the polynomials associated to
the N k −m term. This is a generalization of γk (c) which occurs when m = 1. All terms
2

involving (c − j) to some power come from the binomial coefficient. The product of minors
on the right contributes at most terms of order N 2j(k−j) . Therefore, at the transition points
we are adding polynomials which have zeroes of order k 2 − m − 2j(k − j) (assuming this
quantity is positive), coming from the binomial coefficients in the above expression. This
makes the resulting piecewise function very smooth. To be precise,

Theorem 8. The piecewise function of polynomials giving asymptotics for the N k power
2 −m

of N has the following properties:

• It is symmetric around k/2.

• It is supported on [0, k] and on each interval [j, j + 1] (for j an integer) it is a

polynomial.

• Each polynomial is of degree at most k 2 − m.

• It is differentiable k 2 − m − 2j(k − j) − 1 times at a transition point c = j.

U (N )
The first property is a consequence of the functional relation for Ik . The second
U (N )
property comes from 4.6 and noticing that Ik is 0 for c > k. The third property comes
from noticing that in the binomials in 4.6, a factor of c is paired with a factor of N always.

23
The fourth property comes from the previously described differentiability at 0. That is to
say, if
U (N ) 2 2 2
Ik (n, N ) = γk (c)N k −1 + γk,1 (c)N k −2 + γk,2 (c)N k −3 + . . . ,
then γk,m (c) share the same properties as γk (c) in the above way. All of the lower order
terms in N are highly smooth symmetric piecewise polynomials on the domain [0, k].

4.1.1 A recursion for FN,k (x)

Let M be a k × k matrix, Mij then (k − 1) × (k − 1) the matrix obtained by deleting row i

and column j of M and Mi,j
l,m
be the (k − 2) × (k − 2) matrix obtained from M by deleting
rows i and j, and columns l and m.
The Desnanot-Jacobi identity states that
1,k
det(M ) det(M1,k ) = det(M11 ) det(Mkk ) − det(M1k ) det(Mk1 ). (4.7)

Applying this identity to FN,k (x) gives

FN +2,k−1 (x)FN,k−1 (x) − FN +1,k−1 (x)2

FN,k (x) = . (4.8)
FN +2,k−2 (x)

X N +i+j−1 −1
This follows from the observation that the entries of FN,k (x) are of the form N +i+j−1
,
with N + i + j − 1 increasing by 1 as we increment either i or j.
This recursion allows one to determine the polynomial FN,k (x) the from the polynomials
for k − 1 and k − 2.

24
Chapter 5

Further Properties

5.1 Unimodality of γk (c)

We review some more basic properties of γk (c). In the appendix we have plots of γk (c) for
k = 4. On each interval [j − 1, j] for j ≤ 4, an integer, γk (c) is a different polynomial.
These polynomials approximate a Gaussian.

Indeed, the Gaussian behavior suggest that γk (c) is unimodal. This question was raised
by Rudnick during a conference a few years ago. Recently, Rogers remarked that γk (c)
is log-concave and outlined a proof[11]. We give a shorter proof here and show that this
log-concavity implies unimodality.

The Gaussian behaviour was shown explicitly in earlier work due to Basor, Ge and Ru-
binstein [3], at least asymptotically around the center. The following theorem summarizes
the Gaussian nature in the limiting case
Theorem 9 (Basor, Ge, Rubinstein). Let bk = 8(1 − 1/(4k 2 )) and c = k/2 + o(k). Then
r
G(k + 1)2 bk −bk (c−k/2)2
γk (c) ∼ e .
G(2k + 1) π

We move on to the proof of unimodality, log-concavity and some recurrence relations

for γk (c) and related functions.

25
 Let !α !β !γ
k
Y k
Y Y
Pα,β,γ (x) = xi 1 − xi |xi − xj | . (5.1)
i=1 i=1 i6=j

We are interested in the integral

Z k
!
X
yα,β,γ (c) = δ c− xi Pα,β,γ (x) , (5.2)
Ck i=1

with C k being the unit cube and δ being the Dirac delta function which is a general-
ization of the integral that appears in the definition (1.15) of γk (c).

Theorem 10. The functions yα,β,γ (c) are unimodal if α, β, γ > 1 and real.

We first prove unimodality is guaranteed by log-concavity. Let f : [0, 1] → R and

assume f is bounded, continuous and log-concave. Furthermore assume f is positive on
its interior. We prove that f must be unimodal.

Proof. Suppose f 0 (a) = f 0 (b) = 0 for some a 6= b in [0, 1], where a is a global maximum.
Since f is log-concave, log f is a concave function with vanishing derivative at a and b.
Consider the line segment from (a, log f (a)) to (b, log f (b)). WLOG let b < a, so it has
positive slope. Since the derivative of log(f ) at b is 0 there is some neighbourhood to the
right of b contained under the line segment. But this contradicts concavity.

Now it remains to see that yα,β,γ (c) is log-concave. Consider the domain where the
integrand is non-zero, C k ∩ Hc where Hc is the hyperplane ki=1 xi = c This is a convex
P
set, it suffices to show Pα,β,γ (c) is log-concave on this set. This is because taking marginals
of log-concave functions preserves log-concavity [6].

Lemma 2. Pα,β,γ (x) is log-concave on the domain C K ∩ Hc .

Proof. Since a product of log concave functions is log-concave, it suffices to prove log-
concavity of each term separately. That is, we show xαi , (1 − xi )β and |xi − xj |γ are
log-concave. Indeed, it suffices to take the domain of integration to be 0 ≤ xi ≤ xj ≤ 1
for i < j by symmetry (introducing a factor of n!). Taking the log of xαi gives α log(xi )
which is concave on [0, 1]. Similarly, we can substitute u = 1 − xi in the second case, and
u = |xi − xj | = xj − xi in the third. In each case the domain is still within [0, 1].

26
Some Identities

We derive some general identities for the derivative of yα,β,γ (c). Note first that

yα,β,γ (c) = yβ,α,γ (k − c) (5.3)

via the substitution xi 7→ 1 − xi .

Consider the two sets C k ∩ Hc and C k ∩ Hc+ . With the substitution xi 7→ xi + k we
can get a bijection between the two sets, apart from some small section around the border.
Expanding using the definition of derivative:

yα,β,γ (c + ) − yα,β,γ (c)

.


Which yields

Theorem 11.
Z !
0 1 Xα β
yk,α,β,γ (c) = δ(α)yk−1,γ,β,γ (c)−δ(β)yk−1,α,γ,γ (c−1)+ Pk,α,β,γ (x) + .
k C k ∩Hc i
xi 1 − xi

Here we use δ(α) to denote the function that takes on the
value of 1 if α = 0 and 0
otherwise. If we instead consider the substitution xi 7→ 1 + c xi which achieves a similar


effect to the above we can again expand the derivative to get

Theorem 12.
Z !
0
X 1
cyk,α,β,γ (c) = C1 yk,α,β,γ (c) − kδ(β)yk−1,α,γ,γ (c − 1) + β k− Pk,α,β,γ (x).
C k ∩Hc i
1 − xi

With C1 = αk + βk + γ k
being a constant in c.


2

27
Chapter 6

Conclusions

We have established determinant formulae for averages of secular coefficients. In the limit
these random matrix theory averages are conjectured to behave like the number theoretic
integrals over divisor sums. We also showed that the lower order terms of the random
matrix theory averages have a similar behaviour to γk (c). We end the thesis by raising
some further questions for research.

Q1. We know that γk (c) has an integral formulation as

Z X Y
γk (c) = δ( xi − c) (xi − xj )2 dx
[0,1]k i 1≤i<j≤k

and γk (c) is the highest order term (N k −1 ) in the asymptotics of IkG (n, N ) with G = U (N ).
2

Do there exist integral formulations of the cases when G = O(N ) or G = Sp(2N )? What
about the lower order terms?

Q2. We have seen that the divisor function dk (n) in number theory gives rise to
the polynomials γk (c) in random matrix theory through the conjecture due to Keating et
al.[12]. Is there a natural arithmetic function that gives rise to Symplectic and Orthogonal
γk (c)? We suspect that χ(n)dk (n), for real quadratic characters χ and dk (n2 ) gives rise to
Symplectic behaviour.
Q3. Since we have determinant identities for IkG (n, N ), is it possible to derive asymp-
totics from analyzing them? We were able to understand some properties from a general

28
analysis in the previous section but it’s not clear if these determinant identities can give
asymptotics for γkG (c) and lower order terms as k → ∞.
Q4. In the paper of Keating et al. a lattice point calculation for IkG (n, N ) with
U (N )
G = U (N ) is given which is then used to derive some other properties. Ik (m; N ) is
(j)
equal to the count of lattice points x = (xi ) ∈ Zk satisfying the set of relations
2

(j)
1. 0 ≤ xi ≤ N for all 1 ≤ i, j ≤ k
(k) (k−1) (1)
2. x1 + x2 + · · · + xk = kN − m, and

3. x ∈ Ak ,

where Ak is the collection of k × k matrices whose entries satisfy the following system of
inequalities,
(1) (2) (k)
x1 ≤ x1 ≤ · · · ≤ x1
≤

≤
(1) (2) (k)
x2 ≤ x2 ≤ ··· ≤ x2
≤

≤
.. .. .. ..
. . . .
≤

(1) (2) (k)

xk ≤ xk ≤ ··· ≤ xk

Can natural lattice point counting analogues be given for G = Sp(2N ) or O(N )?

29
References

[1] George E Andrews. Plane Partitions (I): The Mac Mahon Conjecture. University of
Wisconsin-Madison, Mathematics Research Center, 1975.

[2] E. C. Bailey, S. Bettin, G. Blower, J. B. Conrey, A. Prokhorov, M. O. Rubinstein,

and N. C. Snaith. Mixed moments of characteristic polynomials of random unitary
matrices. Journal of Mathematical Physics, 60(8):083509, Aug 2019.

[3] Estelle Basor, Fan Ge, and Michael O. Rubinstein. Some multidimensional integrals in
number theory and connections with the painlevé v equation. Journal of Mathematical
Physics, 59(9):091404, Sep 2018.

[4] Edward A Bender and Donald E Knuth. Enumeration of plane partitions. Journal of
Combinatorial Theory, Series A, 13(1):40–54, 1972.

[5] D. Betea and M. Wheeler. Refined cauchy and littlewood identities, plane partitions
and symmetry classes of alternating sign matrices. Journal of Combinatorial Theory,
Series A, 137:126–165, Jan 2016.

[6] Herm Jan Brascamp and Elliott H Lieb. On extensions of the brunn-minkowski and
prékopa-leindler theorems, including inequalities for log concave functions, and with
an application to the diffusion equation. Journal of Functional Analysis, 22(4):366 –
389, 1976.

[7] David Bressoud. Elementary proofs of identities for schur functions and plane parti-
tions. The Ramanujan Journal, 4, 03 2000.

[8] Daniel Bump and Alex Gamburd. On the averages of characteristic polynomials from
classical groups. Communications in Mathematical Physics, 265(1):227–274, Feb 2006.

30
 [9] J.B. Conrey, D.W. Farmer, J.P. Keating, M.O. Rubinstein, and N.C. Snaith. Auto-
correlation of random matrix polynomials. Communications in Mathematical Physics,
237(3):365–395, Jun 2003.

[10] J.B. Conrey, D.W. Farmer, J.P. Keating, M.O. Rubinstein, and N.C. Snaith. Lower
order terms in the full moment conjecture for the riemann zeta function. Journal of
Number Theory, 128(6):1516 – 1554, 2008.

[11] Ofir Gorodetsky and Brad Rodgers. The variance of the number of sums of two squares
in Fq [t] in short intervals, 2020.

[12] J. P. Keating, B. Rodgers, E. Roditty-Gershon, and Z. Rudnick. Sums of divisor

functions in Fq [t] and matrix integrals. Mathematische Zeitschrift, 288(1-2):167–198,
Mar 2017.

[13] Marvin Marcus. Determinants of sums. The College Mathematics Journal, 21(2):130–
135, 1990.

[14] Robert A. Proctor. New symmetric plane partition identities from invariant theory
work of de concini and procesi. European Journal of Combinatorics, 11(3):289 – 300,
1990.

[15] Kannan Soundararajan. Moments of the riemann zeta function. Annals of Mathemat-
ics, pages 981–993, 2009.

31
APPENDICES

32
Appendix A

Tables and Plots

A.1 Tables of γkG(c)
Given a matrix group and integers k, j we give the polynomial defining γkG (c) on c ∈ [j−1, j].

A.1.1 Unitary Group

(k, j) (k 2 − 1)!γk (c)

(2, 1) c3
(2, 2) (2 − c)3
(3, 1) c8
(3, 2) −2c8 + 24c7 252c6 + 1512c5 4830c4 + 8568c3 8484c2 + 4392c927
(3, 3) (c − 3)8
(4, 1) c15
−3c15 + 60c14 − 1680c13 + 29120c12 − 294840c11 + 1873872c10 − 7927920c9
(4, 2) +23268960c8 − 48674340c7 + 73653580c6 − 80912832c5 + 63969360c4
-35497280 c3 + 13131720c2 − 2910240c + 292464
3c15 − 120c14 + 3360c13 − 58240c12 + 644280c11 − 4948944c10 + 28428400c9
(4, 3) -128700000 c8 + 470398500c7 − 1381480100c6 + 3179336160c5 − 5531176560c4
+6950332480 c3 − 5910494520c2 + 3031004640c − 705916304
(4, 4) (4 − c)15

33
A.1.2 Symplectic Group

(k+2)(k−1)
(k, j) 2
!γk (c)
2
(2, 1) c
(2, 2) (c − 2)2
(3, 1) c5
(3, 2) 15c4 − 90c3 + 190c2 − 165c + 51
(3, 3) (3 − c)5
(4, 1) c9
(4, 2) c9 − 36c8 + 576c7 − 3696c6 + 12096c5 − 22680c4 + 25536c3 − 17136c2 + 6336c − 996
(4, 3) −c9 + 1680c6 − 20160c5 + 106344c4 − 307776c3 + 508176c2 − 449856c + 165916
(4, 4) (4 − c)9

A.1.3 Orthogonal Group

The orthogonal group has a slightly different form than the unitary and symplectic groups
for odd k. For odd k, γk (c) is supported on [0, k − 1]. Also, when k = 2 the scaling factor
is 1 in the below table.
(k+1)(k−2)
(k, j) 2
!γk (c)
(2, 1) 1
(2, 2) 1
(3, 1) c2
(3, 2) c2
(3, 3) 0
(4, 1) c5
(4, 2) c5
(4, 3) (4 − c)5
(4, 4) (4 − c)5
(5, 1) c9
(5, 2) c9
(5, 3) −c9 + 3360c6 − 50400c5 + 330624c4 − 1182720c3 + 2396160c2 − 2580480c + 1146880
(5, 4) −c9 + 3360c6 − 50400c5 + 330624c4 − 1182720c3 + 2396160c2 − 2580480c + 1146880
(5, 5) 0

34
A.2 Plots of γkG(c)
To illustrate the gaussian and highly smooth nature of γkG (c) we plot it below for k = 4.

Figure A.1: G = U (N ), k = 4

Figure A.2: G = SP (2N ), k = 4

And for odd k in the case that G = O(N ):

35
Figure A.3: G = O(N ), k = 4

Figure A.4: G = O(N ), k = 5

36