ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it)

arXiv:2006.04503v2 [math.NT] 21 Jan 2021

E. C. BAILEY AND J. P. KEATING

A BSTRACT. Taking t at random, uniformly from [0, T ], we consider the kth moment, with respect to t, of the
random variable corresponding to the 2βth moment of ζ(1/2 + ix) over the interval x ∈ (t, t + 1], where
ζ(s) is the Riemann zeta function. We call these the ‘moments of moments’ of the Riemann zeta function, and
present a conjecture for their asymptotics, when T → ∞, for integer k, β. This is motivated by comparisons
with results for the moments of moments of the characteristic polynomials of random unitary matrices and is
shown to follow from a conjecture for the shifted moments of ζ(s) due to Conrey, Farmer, Keating, Rubinstein,
and Snaith [1]. Specifically, we prove that a function which, the shifted-moment conjecture of [1] implies, is
a close approximation to the moments of moments of the zeta function does satisfy the asymptotic formula
that we conjecture. We motivate as well similar conjectures for the moments of moments for other families of
primitive L-functions.

1. I NTRODUCTION
1.1. Moments of moments. Moments of L-functions play a central role in analytic number theory. It is a
long-standing conjecture for the Riemann zeta function, ζ(s), that for the 2βth moment on the critical line
Re(s) = 1/2, as T → ∞

  2
T β
Z T
1 2β
(1) Mβ (T ) := |ζ( 21 + it)| dt ∼ aβ cβ log ,
T 0 2π

where aβ is an arithmetic factor given in terms of an Euler product and cβ is discussed below (see also,
for example, [2]). Hardy and Littlewood [3] proved this asymptotic for M1 (T ), and Ingham [4] proved it
for M2 (T ). Their formulae, together with heuristic calculations by Conrey and Ghosh [5] and Conrey and
Gonek [6] for M3 (T ) and M4 (T ) respectively, suggest that cβ · (β 2 )! ∈ N for β ∈ N. Beyond β = 2 there
are no rigorous asymptotic results for Mβ (T ). Ramachandra [7] and Heath-Brown [8] have established the
T β2
lower bound Mβ (T ) ≫ (log 2π ) for positive, rational β, and Soundararajan and Radziwiłł [9] extended
the result to all β ≥ 1. Upper bounds of the correct size are known, conditional on the Riemann hypothesis,
due to arguments of Soundararajan [10] and Harper [11].
There is now a relatively well developed understanding of the conjectural connections between the Rie-
mann zeta function and the characteristic polynomials of random unitary matrices (cf. [1, 12]). The notation
we use for such a characteristic polynomial is
 
(2) PN (A, θ) = det I − Ae−iθ ,

for A ∈ U (N ), the group of N × N unitary matrices. By comparing densities of eigenvalues and zeros, one
T
identifies the matrix size N with a height log 2π up the critical line. Under this analogy, the random matrix
1
2 E. C. BAILEY AND J. P. KEATING

average corresponding to (1) is

Z
(3) Mβ (N ) := |PN (A, θ)|2β dA,
U (N )
where the measure dA is the Haar measure on U (N ). Mβ (N ) can be computed when Re(β) > −1/2,
giving the following result.
Theorem (Keating-Snaith [12]). For Re(β) > −1/2,
N
Y Γ(j)Γ(j + 2β)
Mβ (N ) = .
(Γ(j + β))2
j=1
It follows that asymptotically, as N → ∞,
G 2 (1 + β) β 2
(4) Mβ (N ) ∼ N
G(1 + 2β)
T
where G(s) is the Barnes G-function. One sees that by identifying N with log 2π , Mβ (N ) grows asymp-
totically like Mβ (T ), leading to the conjecture [12] that
G 2 (1 + β)
(5) cβ = .
G(1 + 2β)
This matches the known values when β = 1 [3] and β = 2 [4], as well as the values previously conjectured
for β = 3 [5] and β = 4 [6] using heuristic number-theoretic methods. Recently, these heuristic number-
theoretic methods have been extended to calculate all integer moments of the zeta function on its critical
line, giving results that agree precisely with (5) and with its extension to include lower order terms [1] –
see [13–17]. (The correctness of the formula in the function field setting follows, in the appropriate limit,
from equidistribution; cf. [18].)
In (1), the moments are defined with respect to an average over an asymptotically long stretch of the
critical line. Recently there has been interest in the fluctuations in the values of moments defined with
respect to averages over random short intervals, e.g. intervals of constant length. The moments of these
fluctuations are therefore the moments of moments. These moments of moments have played a central
role in our understanding of the extreme values taken by the zeta function over short intervals [19–22],
and a precise conjecture for the local maximum is due to Fyodorov and Keating [20]. Recent work of
Najnudel [23] and Arguin et al. [24] has led to a verification to leading order of this conjecture, with an
almost sharp upper bound, including the predicted subleading order, due to Harper [22]. At time of writing,
the strongest result is due to Arguin et al. [25], whose work settles the upper bound part of the conjecture
including the predicted tail. Extreme values and moments over mesoscopic and macroscopic intervals, as
well as the transition between the two, are also of interest and have been studied by Arguin et al. [26].
Precisely, we define the moments of moments here as follows.
Definition. For T > 0 and Re(β) > −1/2
Z T Z t+1 k
1 2β
(6) MoMζT (k, β) := |ζ( 12 + ih)| dh dt.
T 0 t
Choosing to average over intervals of length 1 is simply for notational convenience1 ; the results stated in
section 1.2 hold for any interval that is O(1) as T → ∞.
1See also [19, 20] where the moments of moments were introduced; there the intervals are instead of length 2π.
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 3

The corresponding ‘moments of moments’ for the characteristic polynomials of random unitary matrices
may similarly be defined by
Z  Z 2π k
1 2β
(7) MoMU (N ) (k, β) := |PN (A, θ)| dθ dA.
U (N ) 2π 0
These have recently been computed for integer values of k, β in [27, 28]. This suggests a conjecture for
MoMζT (k, β), which we state in section 1.2. We show that this conjecture follows from another conjecture
due to Conrey et al. [1] concerning the shifted moments of the zeta function. Our proof of this follows
similar lines to the corresponding proof in [27], which is based on an expression derived in [29] for the
shifted moments of characteristic polynomials of random matrices. Moreover, by appealing to results of
Assiotis, Bailey, and Keating [30] for the other classical compact matrix groups, in section 1.3 we give
conjectures of moments of moments for other families of L-functions.
1.2. Statement of Result. In order to develop a conjecture for the asymptotics of MoMζT (k, β) for large
T , we use the large N behaviour of MoMU (N ) (k, β). Fyodorov, Hiary and Keating [19], and Fyodorov and
Keating [20] conjectured the asymptotic form2 of MoMU (N ) (k, β) as matrix size N → ∞,
 k
G 2 (1+β) 2
Γ(1 − kβ 2 )N kβ , if k < 1/β 2 ,

G(1+2β)Γ(1−β )2
(8) MoMU (N ) (k, β) ∼
c N k2 β 2 −k+1 , if k > 1/β 2 ,
k,β

where again G(s) is the Barnes G-function and ck,β is an unspecified function of k and β.
For integer k, β, Bailey and Keating [27] determined the asymptotic growth of MoMU (N ) (k, β):
Theorem (Bailey-Keating [27]). For k, β ∈ N,
2 β 2 −k+1

(9) MoMU (N ) (k, β) = ck,β N k (1 + O 1
N ),
where the leading order coefficient ck,β is explicitly given in terms of an integral over a certain simplex.
Furthermore, MoMU (N ) (k, β) is a polynomial in N .
Via an alternative representation (involving Toeplitz determinants and a Riemann-Hilbert analysis) Claeys
and Krasovsky [31] calculated MoMU (N ) (2, β) for β > −1/4, and determined an alternative expression for
the coefficient c2,β . This calculation was recently extended by Fahs [32] to all k ∈ N and non-negative,
real β, though without an explicit expression for ck,β . For integer values of k and β, Assiotis and Keating
showed how one of the approaches developed in [27] leads to a connection with lattice point counts and
gave yet another expression for ck,β , as a volume of a certain region involving continuous Gelfand-Tsetlin
patterns with constraints. In a recent paper [33], the present authors computed the analogue of (7) for a
different logarithmically correlated process - branching random walks - with k ∈ N and β ∈ R. The
resulting expression agrees asymptotically with (9).
By making the natural exchanges in (9), we therefore expect the following to hold.
Conjecture. Let k, β ∈ N. Then,

k2 β 2 −k+1


(10) T
MoMζT (k, β) = αk,β ck,β log 2π 1 + Ok,β log−1 T ,
where ck,β is the same coefficient appearing in (9), and αk,β contains the arithmetic information and takes
the form of an Euler product over primes (cf. (1)).
2Here and throughout, we write A(t) ∼ B(t) as t → ∞ for A(t)/B(t) → 1 as t → ∞.
4 E. C. BAILEY AND J. P. KEATING

A similar conjecture, as well as further motivation for studying moments of moments, can be found in
Fyodorov and Keating [20].
Furthermore, we expect
that the moments
of moments, when suitably smoothed, will have the general
T −δ
form Polyk2 β 2 −k+1 log 2π + Ok,β T , for some δ > 0, where Polyn (x) is a polynomial of degree n
in the variable x.
For integer k we can rewrite MoMζT (k, β) by expanding and switching the order of integration,
Z 1 Z 1Z T k
1 Y
(11) MoMζT (k, β) = ··· |ζ( 12 + i(t + hj ))|2β dtdh1 · · · dhk .
T 0 0 0 j=1

We now apply a conjecture of Conrey et al. [1] concerning shifted moments of the zeta function3 such
as those appearing in the inner integral in (11). Specialising this conjecture to our situation (and to our
notation) we have the following.
Conjecture (Conrey et al. [1]). Take k, β ∈ N, and h = (h1 , . . . , hk ) with hj ∈ R. Then
Z T k Z T
1 Y 1 2β
(12) |ζ( 12 + i(t + hj ))| dt = Pk,β (log t
2π , h)dt + O(T −δ ),
T 0 T 0
j=1

for some δ > 0, where

I I
(−1)kβ 1 G(z1 , . . . , z2kβ )∆(z1 , . . . , z2kβ )2
(13) Pk,β (x; h) := ··· x Pkβ
dz1 · · · dz2kβ
(kβ)!2 (2πi)2kβ e 2 j=1 zkβ+j −zj 2kβ
Q Qk
(z − ih )2β
j=1 l=1 j l
Q
in which ∆(x1 , . . . , xn ) = i<j (xj − xi ) is the Vandermonde determinant, and the contours are small
circles surrounding the poles. Additionally,
Y
(14) G(z1 , . . . , z2kβ ) := Akβ (z1 , . . . , z2kβ ) ζ(1 + zi − zj )
1≤i≤kβ<j≤2kβ

which includes the Euler product4

Z kβ
!−1 !−1
1Y
Y Y e(θ) e(−θ)
(15) Akβ (z) := (1 − pzm −zl −1 ) 1− 1 1− 1 dθ.
p 1≤l≤kβ<m≤2kβ 0 j=1 p 2 +zj p 2 −zkβ+j

In light of this, we define

Z 1 Z 1Z T
1 t
(16) MoMPk,β (T ) := ··· Pk,β (log 2π , h)dtdh1 · · · dhk .
T 0 0 0

Under the conjecture of Conrey et al., MoMPk,β (T ) should approximate MoMζT (k, β) up to a power saving
in T . In (16), the length of the integral over h1 , . . . , hk has been chosen to be 1 in order to be explicit. It is
readily apparent from the proof that the techniques extend to any O(1) interval (with respect to T → ∞).
Our main result is the following.

3The conjecture in [1] is more general in that it allows for different shifts, as well as integrating against different weight functions.
4Throughout we write e(θ) = exp(2πiθ).
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 5

Theorem 1.1. For k, β ∈ N, one has that as T → ∞

2 2
T k β −k+1


(17) MoMPk,β (T ) = αk,β γk,β log 2π (1 + O log−1 T ),
where the term αk,β = Akβ (0, . . . , 0) contains the arithmetic information and Akβ is as defined in (15), and
γk,β denotes the remaining, non-arithmetic, contribution (further details can be found within the proof, see
in particular (40) and (70)).
Therefore MoMPk,β (T ), which, according to the conjecture for the shifted moments of [1] is a close
approximation to MoMζT (k, β), does satisfy the same asymptotic formula as we conjecture holds for the
moments of moments of the zeta function.

1.3. Conjectures for other families of L-functions. Following Katz and Sarnak [18] (see also [1, 2, 34]),
we can also consider other families of L-functions. These families are split by ‘symmetry type’, indicating
which random matrix ensemble (unitary, symplectic, or orthogonal) should be used as comparison.
Here we give an example from each family of the relevant ‘moments of moments’, and the corresponding
conjecture for their growth. For each symmetry type, the moments of moments comprise an average through
the family and an average over small shifts near a symmetry point. The conjectures for the symplectic and
orthogonal families are derived from the relevant random matrix calculations of Assiotis et al. [30], which
we briefly recap here.
Let G (N ) ∈ {U (N ), Sp(2N ), SO(2N )}, where Sp(2N ) is the group of 2N × 2N unitary symplectic
matrices, and SO(2N ) the group of 2N × 2N orthogonal matrices with unit determinant. A generalisation
of the definition given by (7) is the following.
Z  Z 2π k
1 2β
(18) MoMG (N ) (k, β) := |PN (A, θ)| dθ dA,
G (N ) 2π 0

where the external average is over the relevant Haar measure for the chosen matrix group.
Assiotis et al. [30] computed integer moments of moments for G (N ) = Sp(2N ) and G (N ) = SO(2N ),
and their results are summarised here.
Theorem (Assiotis et al. [30]). Let G (N ) = Sp(2N ). Let k, β ∈ N. Then,



(19) MoMSp(2N ) (k, β) = ck,β N kβ(2kβ+1)−k 1 + Ok,β 1
N ,
where the leading order term coefficient ck,β is the volume of a certain convex region5. Moreover,
MoMSp(2N ) (k, β) is a polynomial function in N .
Theorem (Assiotis et al. [30]). Let G (N ) = SO(2N ). Let k, β ∈ N. Then,
(20) MoMSO(2N ) (1, 1) = 2(N + 1)
otherwise,



(21) MoMSO(2N ) (k, β) = c̃k,β N kβ(2kβ−1)−k 1 + Ok,β 1
N ,
where the leading order term coefficient c̃k,β is given as a sum of volumes of certain convex regions6 . More-
over, MoMSO(2N ) (k, β) is a polynomial function in N .

5Differing from the coefficient appearing in (9).

6Again, differing from c
k,β in (9), as well as ck,β given in (19).
6 E. C. BAILEY AND J. P. KEATING

Note that these results differ from the unitary case, where the degree of the polynomial is k2 β 2 − k + 1,
see (9).
Returning to the families of L-functions, the archetypal example of a unitary L-function is the Riemann
zeta function, as discussed in section 1.1. We rewrite the family here for context,
(22) {ζ( 12 + it) | t ≥ 0}.
To each family there is an associated ‘height’ by which the elements are ordered. For the family given by
(22), one orders by the height up the critical line, t. The moments of moments are defined by (6), and their
conjectured asymptotic is given by (10).
Quadratic Dirichlet L-functions provide an example of a symplectic family:


(23) {L(s, χd ) | d a fundamental discriminant, χd (n) = nd },
where the family is ordered by |d|. For Re(s) > 1
∞
X χd (n)
(24) L(s, χd ) := ,
ns
n=1
and otherwise L(s, χd ) is defined by its analytic continuation. The moments of moments are defined as
Z 1 k
1 X∗ 1 2β
(25) MoMLχd (k, β) := ∗ L( 2 + it, χd ) dt ,
D 0
|d|≤D

where the sum is only over fundamental discriminants d, and D ∗ denotes the number of terms in the sum.
We hence conjecture that for k, β ∈ N, as D → ∞,



(26) MoMLχd (k, β) = ηk,β ck,β (log D)kβ(2kβ+1)−k 1 + Ok,β log−1 D ,
where ck,β corresponds to the leading order coefficient in (19), and ηk,β contains the arithmetic information
(see for example [1]).
For an orthogonal example we take the family of twisted elliptic curve L-functions,


(27) {LE (s, χd ) | d a fundamental discriminant, χd (n) = nd },
which is ordered by |d| and where the L-function is defined by
∞
X an χd (n)
(28) LE (s, χd ) :=
ns
n=1
for Re(s) > 1 and by analytic continuation otherwise. The coefficients an are the weightings from the
L-function associated to E, normalised so that the critical line is shifted to Re(s) = 1/2. The moments of
moments are then defined as
Z 1 k
1 X∗ 1 2β
(29) MoMLE (k, β) := ∗ LE ( 2 + it, χd ) dt ,
D 0
|d|≤D

where again the sum is only over fundamental discriminants d, and D ∗ denotes the number of terms in the
sum. We hence conjecture that for k, β ∈ N and k 6= 1 6= β, as D → ∞,



(30) MoMLE (k, β) = ξk,β c̃k,β (log D)kβ(2kβ−1)−k 1 + Ok,β log−1 D ,
where c̃k,β corresponds to the leading order coefficient in (21), and ξk,β contains the arithmetic information
(again, see for example [1]).
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 7

We note that for both the symplectic and the orthogonal families defined above one could also prove
theorems akin to theorem 1.1 using the techniques described in sections 2 and 3. This is due to the fact
that [1] also provides conjectures for the shifted moments of the L-functions in question. However, such
calculations follow near verbatim those for the Riemann zeta function, so we do not report the details.

2. S TRUCTURE OF PROOF OF THEOREM 1.1

Within this section, we outline the proof of theorem 1.1. We require a series of lemmas whose proofs are
presented in section 3.
To contextualise the proof, we note that the general expression for shifted moments of the zeta function
conjectured by Conrey et al. [1] (cf. (12)) has an exact correspondence to characteristic polynomials in the
random matrix case [29] (quoted as theorem 1.5.2. in [1]). A generalised version of the random matrix
theorem was used by the present authors when proving (9). Hence, the approach outlined in this section
follows similar lines to the proof of theorem 1.2 in [27]. However, in several places we will need different
steps to account for the arithmetic factors involved.
We first recall the definition of Pk,β (x; h) from (13),
I I
(−1)kβ 1 G(z1 , . . . , z2kβ )∆(z1 , . . . , z2kβ )2
(31) Pk,β (x; h) = ··· Pkβ dz1 · · · dz2kβ
(kβ)! (2πi)2kβ
2 x
e2 j=1 zkβ+j −zj
Q2kβ Qk
(z − ih )2β
j=1 l=1 j l

where one integrates over small circles surrounding the poles and
Y
(32) G(z1 , . . . , z2kβ ) = Akβ (z1 , . . . , z2kβ ) ζ(1 + zi − zj )
1≤i≤kβ<j≤2kβ

which includes the Euler product

Z kβ
!−1 !−1
1Y
Y Y
zm −zl −1 e(θ) e(−θ)
(33) Akβ (z) = (1 − p ) 1− 1 1− 1 dθ.
p 1≤l≤kβ<m≤2kβ 0 j=1 p 2 +zj p 2 −zkβ+j

Since it will be useful, we also here record an alternative form for Akβ (·) also due to Conrey et al. [1] (a
special case of corollary 2.6.2 in their paper, found there at (2.6.15)),
Qkβ  1

kβ Y 1 −
YX j=1 1+zj −zkβ+n
p
(34) Akβ (z1 , . . . , z2kβ ) = z −zkβ+m ,
p
1−p kβ+n
m=1 n6=m

where additionally they show that each local factor is actually a polynomial in p−1 , p−zj , and pzkβ+j for
j = 1, . . . , kβ.
We begin by decomposing the multiple contour integral Pk,β (x, h) in to a sum of multiple contour in-
tegrals, and apply a variable change so that each contour shifts to small circles surrounding the origin.
Eventually (see (12)–(13)), we will want to set x = log(t/2π).
8 E. C. BAILEY AND J. P. KEATING

Lemma 2.1. Asymptotically for large x,

2β
cl (k, β) ix Pkβ
X  x |Sk,β;l|
Pk,β (x; h) ∼ 2 2kβ
Akβ (iµ1 , . . . , iµ2kβ )e 2 j=1 µj −µkβ+j
(kβ)! (2πi) 2
l1 ,...,lk−1 =0
Z Z Y   2kβ
Y
2(vm −vn )
(35) × ··· ζ 1+ x + i(µm − µn ) f (v; l) dvm ,
Γ0 Γ0 m≤kβ<n m=1
µn 6=µm

where cl (k, β) is a product of binomial coefficients determined within the proof and
Pkβ Q
e j=1 vj −vkβ+j
(vn − vm )2
m<n
µm =µn
(36) f (v; l) := Q Q2kβ 2β .
m≤kβ<n (vm − vn ) m=1 vm
µn =µm

The contours denoted by Γ0 are small circles around the origin, and the µi variables are related to the hj
variables - for the exact definition see (50). Further, the following sets are useful to define, and the former
appears in the statement of the lemma,

(37) Sk,β;l := {(m, n) : 1 ≤ m ≤ kβ < n ≤ 2kβ, µm = µn },

(38) Tk,β;l := {(m, n) : 1 ≤ m ≤ kβ < n ≤ 2kβ, µm 6= µn }.

The next lemma relies on lemma 2.1 and determines the remaining x ≡ log(t/2π) dependence in the
integrand of Pk,β (x; h) after calculating the contribution from the integration over the h1 , . . . , hk .

Lemma 2.2. As T → ∞,

2 2
T k β −k+1
(39) MoMPk,β (T ) ∼ αk,β γk,β log 2π ,

where αk,β = Akβ (0, . . . , 0),

2β Z Z 2kβ
X cl (k, β) Y
(40) γk,β = ··· f (v; l)Ψk,β (v; l) dvm ,
(kβ)!2 (2πi)2kβ Γ0 Γ0
l1 ,...,lk−1 =0 m=1

cl (k, β) and f (v; l) are given by (49) and (36), and Ψk,β (v; l) is a multiple integral defined within the proof,
see (70).

The final lemma establishes non-vanishing of the leading order coefficient.

Lemma 2.3. For k, β ∈ N, the coefficient αk,β γk,β is non-zero.

By combining the statements of lemma 2.2 and lemma 2.3, we arrive at the statement of theorem 1.1

2 2
T k β −k+1



(41) MoMPk,β (T ) = αk,β γk,β log 2π 1 + O log−1 T
2π ,

for k, β ∈ N.
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 9

3. P ROOFS OF LEMMAS 2.1–2.3

In this section we give the proofs of lemma 2.1, lemma 2.2, and lemma 2.3 from section 2.
Proof of lemma 2.1. We recall the statement of lemma 2.1.
Lemma. Asymptotically for large x,
2β
cl (k, β) ix Pkβ
X  x |Sk,β;l|
Pk,β (x; h) ∼ A (iµ , . . . , iµ )e 2 j=1 µj −µkβ+j
kβ 1 2kβ
(kβ)!2 (2πi)2kβ 2
l1 ,...,lk−1 =0
Z Z Y   2kβ
Y
2(vm −vn )
(42) × ··· ζ 1+ x + i(µm − µn ) f (v; l) dvm ,
Γ0 Γ0 m≤kβ<n m=1
µn 6=µm

where cl (k, β) is a product of binomial coefficients determined in the proof and

Pkβ Q
e j=1 vj −vkβ+j m<n (vn − vm )
2
µm =µn
(43) f (v; l) = Q Q2kβ 2β .
m≤kβ<n (vm − vn ) m=1 vm
µn =µm

Recall the definition of Pk,β (x, h) from (13),

I I
(−1)kβ 1 G(z1 , . . . , z2kβ )∆(z1 , . . . , z2kβ )2
(44) Pk,β (x; h) = · · · P dz1 · · · dz2kβ
(kβ)!2 (2πi)2kβ x kβ
e 2 j=1 zkβ+j −zj 2kβ
Q Qk
(z − ih )2β
j=1 l=1 j l

where the contours are small circles around the poles, and G() is given by (14).
We start by decomposing the 2kβ contours in (44) so that each contour consists of k small circles en-
compassing the poles at ih1 , . . . , ihk with connecting straight lines whose contributions cancel out. Denote
a small circular contour encircling ihj by Γihj . Then,
(−1)kβ 1 X
(45) Pk,β (x, h) = Qk,β (x, h; ε1 , . . . , ε2kβ ),
(kβ)! (2πi)2kβ
2
εj ∈{1,...,k}

where
(46) Z Z
G(z1 , . . . , z2kβ )∆(z1 , . . . , z2kβ )2
Qk,β (x, h; ε1 , . . . , ε2kβ ) = ··· x Pkβ Q Qk dz1 · · · dz2kβ
Γihε
1
Γihε
2kβ e 2 j=1 zkβ+j −zj 2kβ j=1 l=1 j(z − ih l )2β

is the multiple contour integral Pk,β (x, h) with the 2kβ contours each specialised around one of the k poles
determined by the vector ε = (ε1 , . . . , ε2kβ ).
Due to the symmetric nature of the integrand, many of the summands evaluate to zero. Such a fact was
established in lemma 3.2 in [27], and is paraphrased here.
Lemma. Write ε = (ε1 , . . . , ε2kβ ) for a choice of contours appearing as a summand in (45). Then the only
choices of ε leading to a non-zero summand are those where each pole is equally represented in ε. That is:
if ni counts the number of occurrences of i in ε, for i ∈ {1, . . . , k}, then the summand corresponding to the
choice ε is identically zero unless n1 = · · · = nk = 2β.
10 E. C. BAILEY AND J. P. KEATING

Remark. The above lemma was formulated in [27] for a slightly different integral, namely
Z Z
e−N (zkβ+1 +···+z2kβ ) ∆(z1 , . . . , z2kβ )2 dz1 · · · dz2kβ
(47) ··· Q Q2kβ Qk .
Γihε Γihε2kβ (1 − e zn −zm ) (z − ih )2β
1 m≤kβ<n m=1 n=1 m n

However, the structure of (47) only differs trivially to that of (46). The key difference is between the terms
of the form (1 − exp(zn − zm ))−1 in (47) versus ζ(1 + zm − zn ) in (46). However, they share the same
analytic structure since both have a simple pole at zn = zm . Therefore the proof presented in [27] holds as
well for (46).

Hence (45) becomes

2β 2β
(−1)kβ 1 X X
(48) Pk,β (x, h) = · · · cl (k, β)Qk,β (x, h; l),
(kβ)!2 (2πi)2kβ
l1 =0 lk−1 =0

where l = (l1 , . . . , lk−1 ) and Qk,β (x, h; l) is the integral Qk,β (x, h; ε) with contours given by

l l lk−1 2β 2β−lk−1 2β−l

z }| 1
{ z }| 2
{ z }| { z }| { z }| { z }| 1 {
ε = (1, . . . , 1, 2, . . . , 2, . . . , k − 1, . . . , k − 1, k, . . . , k, k − 1, . . . , k − 1, . . . , 1, . . . , 1),

and the coefficient cl (k, β) is a product of binomial coefficients:

     Pk−2 
kβ
kβ − l1 kβ − (l1 + l2 ) kβ − m=1 lm
cl (k, β) = ···
l2 l1 l3 lk−1
   Pk−2 
kβ (k − 2)β + l1 kβ − m=1 (2β − lm )
(49) × ··· .
2β − l1 2β − l2 2β − lk−1

We now apply the following change of variables so to shift all the contours to be small circles around the
origin
2vn
zn = + iµn ,
x
where
(50) 
h1 ,


if n ∈ {1, . . . , l1 } ∪ {2(k − 1)β + 1 + l1 , . . . , 2kβ}

h2 , if n ∈ {l1 + 1, . . . , l1 + l2 } ∪ {2(k − 2)β + 1 + l1 + l2 , . . . , 2(k − 1)β + l1 }


. ..
µn = .. .

 Pk−2 Pk−1 Pk−1 Pk−2


 hk−1 , if n ∈ { m=1 lm + 1, . . . , m=1 lm } ∪ {2β + 1 + m=1 lm , . . . , 4β + m=1 lm }

 Pk−1 Pk−1
hk , if n ∈ { m=1 lm + 1, . . . , m=1 lm + 2β}.

Note that µn is dependent on the choice of l (i.e. the exact ordering of the contours), but this is suppressed
from the notation for simplicity.
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 11

Using the fact that the ζ-function has a simple pole at 1 with residue 1, the resulting integrand of
Qk,β (x, h; l) as x → ∞ is


2v
ix Pkβ Pkβ
1+O 1
x Akβ + iµ1 , . . . , x2kβ + iµ2kβ e 2 j=1 µj −µkβ+j e j=1 vj −vkβ+j
2v1
x
Y Y

× (iµn − iµn )2 ζ 1 + 2(vmx−vn ) + i(µm − µn )
m<n 1≤m≤kβ<n≤2kβ
µm 6=µn µn 6=µm
 
Q 2(vn −vm ) 2
m<n 2kβ
µm =µn x Y
2dvm
×Q  Q
2β x
2(vm −vn ) 2kβ Qk 2vm
1≤m≤kβ<n≤2kβ x m=1 n=1 x − i(hn − µm ) m=1
µn =µm



ix Pkβ Pkβ
x −2kβ j=1 µj −µkβ+j e j=1 vj −vkβ+j
= 1 + O x1 2 A kβ (iµ 1 , . . . , iµ 2kβ )e 2
Y Y

× (iµm − iµn )2 ζ 1 + 2(vmx−vn ) + i(µm − µn )
m<n m≤kβ<n
µm 6=µn µn 6=µm
 2 Q

Q 2(vn −vm ) 2kβ 2vm −2β 2kβ
m<n x m=1 x Y
µm =µn
(51) ×Q  Q dvm
2(vm −vn )
m≤kβ<n x m<n (iµm − iµn )2 m=1
µn =µm µm 6=µn



2 ix Pkβ Pkβ
x 4kβ −2kβ
= 1+O 1
x 2 Akβ (iµ1 , . . . , iµ2kβ )e 2 j=1 µj −µkβ+j e j=1 vj −vkβ+j
 
Q 2(vn −vm ) 2
m<n 2kβ
Y
µ =µ x Y
(52) × ζ 1 + 2(vmx−vn ) + i(µm − µn ) Qm n 2(v −v ) dvm
2β .
m n vm
m≤kβ<n m≤kβ<n x m=1
µn 6=µm µn =µm

The power of x coming from the terms that originated from the Vandermonde (i.e. the numerator of
the fraction in (52)) is calculated to be 2kβ(2β − 1). We now extract the x dependence remaining in the
denominator of the integrand (52). First, define the following sets,

(53) Sk,β;l := {(m, n) : 1 ≤ m ≤ kβ < n ≤ 2kβ, µm = µn }

(54) Tk,β;l := {(m, n) : 1 ≤ m ≤ kβ < n ≤ 2kβ, µm 6= µn }.

Thus, |Sk,β;l |+|Tk,β;l | = k2 β 2 , and the pairs (m, n) ∈ Sk,β;l are precisely those involved in the denominator
of (52). As it will be useful, we also briefly record that
Pk
(−1)|Sk,β;l | = (−1) j=1 lj (2β−lj ) = (−1)kβ .

Hence the integrand of Qk,β (x, h; l) is

ix Pkβ Pkβ
x |Sk,β;l | µj −µkβ+j j=1 vj −vkβ+j
1+O 1
x 2 (−1)kβ Akβ (iµ1 , . . . , iµ2kβ )e 2 j=1 e
Q 2
Y   m<n (vn − vm ) 2kβ
Y
2(vm −vn ) µm =µn dvm
(55) × ζ 1+ x + i(µm − µn ) Q 2β .
m≤kβ<n (vm − vn ) m=1
vm
(m,n)∈Tk,β;l µn =µm
12 E. C. BAILEY AND J. P. KEATING

Thus, as x → ∞,
2β Pkβ
X cl (k, β)

x |Sk,β;l | ix
µj −µkβ+j
Pk,β (x, h) ∼ 2 Akβ (iµ1 , . . . , iµ2kβ )e 2 j=1
(kβ)!2 (2πi)2kβ
l1 ,...,lk−1 =0
Z Z Y   2kβ
Y
2(vm −vn )
(56) × ··· ζ 1+ x + i(µm − µn ) f (v; l) dvm ,
Γ0 Γ0 (m,n)∈T m=1
k,β;l

where the terms in the integrand with no h dependence have been collectively denoted by f (v; l), so
Pkβ Q
e j=1 vj −vkβ+j
(vn − vm )2
m<n
µm =µn
(57) f (v; l) := Q Q2kβ 2β ,
m≤kβ<n (vm − vn ) m=1 vm
µn =µm

which concludes the proof.



Proof of lemma 2.2. We recall the statement of lemma 2.2.

Lemma.
T

k2 β 2 −k+1
MoMPk,β (T ) ∼ αk,β γk,β log 2π ,

where αk,β := Akβ (0, . . . , 0),

2β Z Z 2kβ
X cl (k, β) Y
γk,β := ··· f (v; l)Ψk,β (v; l) dvm ,
(kβ)!2 (2πi)2kβ Γ0 Γ0 m=1
l1 ,...,lk−1 =0

cl (k, β) and f (v; l) are given by (49) and (36) and Ψk,β (v; l) is a multiple integral defined within the proof,
see (70).

Set x := log(t/2π). By lemma 2.1, we have that

Z 1 Z 1Z T
1 t
MoMPk,β (T ) = ··· Pk,β (log 2π , h)dtdh1 · · · dhk
T 0 0 0
2β
X cl (k, β)
∼
T (kβ)!2 (2πi)2kβ
l1 ,...,lk−1 =0

ix Pkβ
Z 1 Z 1Z T


x |Sk,β;l |
× ··· Akβ (iµ1 , . . . , iµ2kβ )e 2 j=1 µj −µkβ+j 2
0 0 0

2kβ k
Z Z Y Y Y
2(vm −vn )
(58) × ··· f (v; l) ζ 1+ x + i(µm − µn ) dvm dt dhn .
Γ0 Γ0 (m,n)∈Tk,β;l m=1 n=1
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 13

We begin by firstly rewriting the product of zeta functions in the innermost integrand of (58) using the
definition of µ1 , . . . , µ2kβ , see (50),
Y  
ζ 1 + 2(vmx−vn ) + i(µm − µn )
(m,n)∈Tk,β;l
Y Y   Y  
2(vm −vn ) 2(vm −vn )
(59) = ζ 1+ x + i(hσ − hτ ) ζ 1+ x − i(hσ − hτ ) .
+
1≤σ<τ ≤k (m,n)∈Vσ,τ (m,n)∈Vσ,τ
−

+ , V − partition T
where the sets Vσ,τ σ,τ k,β;l ,

+
(60) Vσ,τ := {(m, n) ∈ Tk,β;l : µm − µn = hσ − hτ },
−
(61) Vσ,τ := {(m, n) ∈ Tk,β;l : µm − µn = hτ − hσ }.
Combining (58) and (59) we find
2β
X cl (k, β)
MoMPk,β (T ) ∼
T (kβ)!2 (2πi)2kβ
l1 ,...,lk−1 =0

ix Pkβ
Z 1 Z 1Z T

|S |
× ··· Akβ (iµ1 , . . . , iµ2kβ )e 2 j=1 µj −µkβ+j x2 k,β;l
0 0
Z Z 0 Y Y  
2(vm −vn )
× ··· f (v; l) ζ 1+ x + i(hσ − hτ )
Γ0 Γ0 +
1≤σ<τ ≤k (m,n)∈Vσ,τ

Y   2kβ
Y k
Y
2(vm −vn )
(62) × ζ 1+ x − i(hσ − hτ ) dvm dt dhn .
(m,n)∈Vσ,τ
− m=1 n=1

Next we switch the order of integration

2β Z T Z Z
X cl (k, β)

x |Sk,β;l |
MoMPk,β (T ) ∼ · · · f (v; l)
T (kβ)!2 (2πi)2kβ 0 2 Γ0 Γ0
l1 ,...,lk−1 =0
Z 1 Z 1 Y Y  
2(vm −vn )
× ··· Akβ (iµ1 , . . . , iµ2kβ ) ζ 1+ x + i(hσ − hτ )
0 0 +
1≤σ<τ ≤k (m,n)∈Vσ,τ

Y   ix Pkβ k
Y 2kβ
Y
2(vm −vn ) µj −µkβ+j
(63) × ζ 1+ x − i(hσ − hτ ) e 2 j=1 dhn dvm dt,
(m,n)∈Vσ,τ
− n=1 m=1

and focus on the inner integrals over h1 , . . . , hk .

We use the structure of (µ1 , . . . , µ2kβ ) in order to write the exponential term in the innermost integrand
of (63) explicitly in terms of h1 , . . . , h2kβ ,
   
kβ k−1
ix X X
(64) exp  µj − µkβ+j  = exp ix (lj − β)(hj − hk ) .
2
j=1 j=1
14 E. C. BAILEY AND J. P. KEATING

Then, we use the Laurent expansion for the zeta function around its pole,

∞
1 X j
(65) ζ(1 + s) = + cj s ,
s
j=0

where cj = (−1)j sj /j! and sj are the Stieltjes constants (so c0 is the Euler-Mascheroni constant).
In this calculation, we will be taking s ∼ 1/x, so the main contribution to (63) comes from just taking the
first term in the Laurent series for each zeta function in the integrand. Hence the integral over the h1 , . . . , hk
becomes

Z 1 Z 1 Y Y  
2(vm −vn )
··· Akβ (iµ1 , . . . , iµ2kβ ) ζ 1+ x + i(hσ − hτ )
0 0 +
1≤σ<τ ≤k (m,n)∈Vσ,τ
Y   ix Pkβ
2(vm −vn )
× ζ 1+ x − i(hσ − hτ ) e 2 j=1 µj −µkβ+j dh1 · · · dhk
(m,n)∈Vσ,τ
−

(66)
Pk−1
Akβ (iµ1 , . . . , iµ2kβ )eix j=1 (lj −β)(hj −hk ) dh1 · · · dhk
Z 1 Z 1
∼ ··· Q Q  Q  .
2(vm −vn ) 2(vm −vn )
0 0
1≤σ<τ ≤k +
(m,n)∈Vσ,τ x + i(hσ − hτ ) (m,n)∈Vσ,τ−
x − i(hσ − hτ )

Observe that each term in the integrand of (66) is a function of the differences δj := hj − hk , j = 1, . . . , k.
For the exponential term this is immediate. For Akβ (·) we use (34), writing Akβ;p (·) for each local factor,

Qkβ  1

kβ
X Y j=1 1 − p1+i(µj −hk ) p
−i(µkβ+n −hk )
(67) Akβ;p (iµ1 , . . . , iµ2kβ ) = .
m=1 n6=m
1− pi(µkβ+n −hk ) p−i(µkβ+m −hk )

Similarly for the terms in the denominator, we write hσ − hτ = (hσ − hk ) − (hτ − hk ). We now perform a
change of variables 2δj = x(hj − hk ) and use the fact that Akβ is analytic in a neighbourhood of (0, . . . , 0).
Hence
Pk−1
Akβ (iµ1 , . . . , iµ2kβ )eix j=1 (lj −β)(hj −hk ) dh1 · · · dhk
Z 1 Z 1
··· Q Q  Q  
2(vm −vn ) 2(vm −vn )
0 0
1≤σ<τ ≤k +
(m,n)∈Vσ,τ x + i(hσ − hτ ) (m,n)∈Vσ,τ−
x − i(hσ − hτ )
(68)

Pk−1
x x x |Tk,β;l |−k−1
(0, . . . , 0) e2i j=1 (lj −β)δj dδ1 · · · dδk−1
Z Z
2 2
2 Akβ
∼ ··· Q Q Q
0 0 1≤σ<τ ≤k +
(m,n)∈Vσ,τ (vm − vn + i(δσ − δτ )) (m,n)∈Vσ,τ − (vm − vn − i(δσ − δτ ))

(69)
 x |Tk,β;l|−k−1
∼ Akβ (0, . . . , 0) Ψk,β (v; l),
2
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 15

where
Z Z Pk−1
∞ ∞
e2i j=1 (lj −β)δj
Ψk,β (v; l) := ··· Q Q
0 0 1≤σ<τ ≤k +
(m,n)∈Vσ,τ (vm − vn + i(δσ − δτ ))
dδ1 · · · dδk−1
(70) ×Q Q .
− (vm − vn − i(δσ − δτ ))
1≤σ<τ ≤k (m,n)∈Vσ,τ

Then Ψk,β (v; l) defined by (70) no longer depends on x. We will further analyse its contribution to
MoMPk,β (T ) in the proof of lemma 2.3 below.
Incorporating (70) in to (63), and recalling that |Sk,β;l | + |Tk,β;l | = k2 β 2 and x = log(t/2π), we find that
(71)
2β Z T  x k2 β 2 −k+1 Z Z 2kβ
X cl (k, β)Akβ (0, . . . , 0) Y
MoMPk,β (T ) ∼ ··· f (v; l)Ψk,β (v; l) dvm dt
T (kβ)!2 (2πi)2kβ 0 2 Γ0 Γ0
l1 ,...,lk−1 =0 m=1

2 2
T k β −k+1



(72) = αk,β γk,β log 2π 1 + O log−1 T
2π ,
where αk,β := Akβ (0, . . . , 0) and
2β Z Z 2kβ
X cl (k, β) Y
γk,β := ··· f (v; l)Ψk,β (v; l) dvm .
(kβ)!2 (2πi)2kβ Γ0 Γ0
l1 ,...,lk−1 =0 m=1


Proof of lemma 2.3. We recall the statement of lemma 2.3.
Lemma. The coefficient αk,β γk,β appearing in the statement of lemma 2.2 is non-zero.
From lemma 2.2, we have that αk,β = Akβ (0, . . . , 0). The value of An (0, . . . , 0) was calculated by
Conrey et al. [1]. Hence, using their formula, we have

2 kβ−1 
Y  
1 (kβ−1) X kβ − 1 2 −m
(73) Akβ (0, . . . , 0) = 1− p .
p
p m=0
m

Thus, all that remains to justify is that γk,β = 6 0. To do this we appeal to the residue theorem7. Fix a
choice of l1 , . . . , lk−1 and notice that
cl (k, β)
> 0.
(kβ)!2
To conclude, we compare to the random matrix case. In previous work [27], recall (cf. (9)) that the
moments of the moments of unitary characteristic polynomials
Z  Z 2π k
1 2β
(74) MoMU (N ) (k, β) := |PN (A, θ)| dθ dA
U (N ) 2π 0
were shown to satisfy
2 β 2 −k+1
(75) MoMU (N ) (k, β) = ck,β N k (1 + O( N1 )),
7The initial part of the proof borrows calculations found in the proof of lemma 3.7 in [27].
16 E. C. BAILEY AND J. P. KEATING

and ck,β is non-zero. The exact form8 of ck,β is found by evaluating

2β Z Z 2kβ
X cl (k, β) Y
ck,β = ··· f (v; l)Ωk,β (v; l) dvm ,
((kβ)!)2 (2πi)2kβ Γ0 Γ0 m=1
l1 ,...,lk−1 =0

where cl (k, β) and f (v; l) are the same functions used throughout the present work (see (49) and (57)),
Z 2π Z 2π Pk−1
1 eiN j=1 (lj −β)(θj −θk )
(76) Ωk,β (v; l) = lim ··· dθ1 · · · dθk
N →∞ (2π)k N Tk,β;l −k+1 0
Q vn −vm
0 (1 − e N ei(µn −µm ) )
(m,n)∈Tk,β;l
Z 2π Z 2π Pk−1
1 eiN j=1 (lj −β)(θj −θk )
= lim ···  
N →∞ (2π)k N Tk,β;l −k+1 0 vn −vm
Q Q i(θ −θ )
0
1≤σ<τ ≤k +
(m,n)∈Vσ,τ 1−e N e τ σ

dθ · · · dθk
(77) ×Q  1 vm −v n
.
(m,n)∈Vσ,τ
− 1 − e N ei(θσ −θτ )
+ , and V − are as defined in (38), (60), and (61) respectively and µ matches the definition
The sets Tk,β;l , Vσ,τ σ,τ
given by (50).
In [27], the denominator in the integrand of Ωk,β was expanded as a geometric series, and then the
integral was explicitly calculated. However, one could proceed in a similar way to the proof of lemma 2.2,
i.e. noticing that the integrand is simply a function of the differences δj := θj − θk and then use the Laurent
expansion of the denominator to pull out the power of N . Proceeding in this way we find
Z ∞ Z ∞ Pk−1
1 ei j=1 (lj −β)δj
Ωk,β (v; l) ∼ ··· Q Q
(2π)k 0 0 1≤σ<τ ≤k + (vn − vm + i(δτ − δσ ))
(m,n)∈Vσ,τ
dδ1 · · · dδk−1
(78) ×Q
(m,n)∈Vσ,τ
− (vn − vm − i(δτ − δσ ))
Z Z Pk−1
(−1)|Tk,β;l | ∞ ∞
ei j=1 (lj −β)δj
∼ ··· Q Q
(2π)k 0 0 1≤σ<τ ≤k +
(m,n)∈Vσ,τ (vm − vn + i(δσ − δτ ))
dδ1 · · · dδk−1
(79) ×Q .
− (vm − vn − i(δσ − δτ ))
(m,n)∈Vσ,τ
P
Thus, since |Sk,β;l | + |Tk,β;l | = k2 β 2 , and |Sk,β;l | = kj=1 lj (2β − lj ), we have

(−1)|Tk,β;l | = (−1)kβ(kβ−1) = 1.
So Ωk,β (v; l) ∼ κΨk,β (v; l) for some positive constant κ. Hence γk,β 6= 0 by comparison with ck,β . 

4. ACKNOWLEDGEMENTS
We would like to thank Valeriya Kovaleva for helpful comments and discussions. We are grateful to the
anonymous referee for their careful reading of the paper and suggestions. ECB would like to thank the
8In [27], the multiple contour integral that they treat differs only from the one we consider here by a simple transformation. For
simplicity, the results presented henceforth have been translated so to be consistent with the notation within this article.
 ON THE MOMENTS OF THE MOMENTS OF ζ(1/2 + it) 17

Heilbronn Institute for Mathematical Research for support. JPK is pleased to acknowledge support from
ERC Advanced Grant 740900 (LogCorRM).

R EFERENCES
[1] J. B. Conrey, D. W. Farmer, J. P. Keating, M. O. Rubinstein, N. C. Snaith, Integral moments of l-functions, Proceedings of the
London Mathematical Society 91 (1) (2005) 33–104.
[2] J. B. Conrey, D. W. Farmer, Mean values of L-functions and symmetry, International Mathematics Research Notices 2000 (17)
(2000) 883–908.
[3] G. H. Hardy, J. E. Littlewood, Contributions to the theory of the Riemann zeta-function and the theory of the distribution of
primes, Acta Mathematica 41 (1918) 119–196.
[4] A. E. Ingham, Mean-value theorems in the theory of the Riemann zeta-function, Proceedings of the London Mathematical
Society 2 (1) (1926) 273–300.
[5] J. B. Conrey, A. Ghosh, Mean values of the zeta-function, III, in: Proceedings of the Amalfi Conference on Analytic Number
Theory, Università di Salerno, 1992.
[6] J. B. Conrey, S. M. Gonek, High moments of the Riemann zeta-function, Duke Mathematical Journal 107 (3) (2001) 577–604.
[7] K. Ramachandra, Some remarks on the mean value of the Riemann zeta-function and other Dirichlet series-II, Hardy-
Ramanujan Journal 3.
[8] D. R. Heath-Brown, Fractional Moments of the Riemann Zeta-Function, Journal of the London Mathematical Society 2 (1)
(1981) 65–78.
[9] M. Radziwiłł, K. Soundararajan, Continuous lower bounds for moments of zeta and l-functions, Mathematika 59 (1) (2013)
119–128.
[10] K. Soundararajan, Moments of the Riemann zeta function, Annals of Mathematics 170 (2009) 981–993.
[11] A. J. Harper, Sharp conditional bounds for moments of the Riemann zeta function, arXiv preprint arXiv:1305.4618.
[12] J. P. Keating, N. C. Snaith, Random matrix theory and ζ(1/2 + it), Communications in Mathematical Physics 214 (1) (2000)
57–89.
[13] B. Conrey, J. P. Keating, Moments of zeta and correlations of divisor-sums: I, Philosophical Transactions of the Royal Society
A: Mathematical, Physical and Engineering Sciences 373 (2040) (2015) 20140313.
[14] B. Conrey, J. P. Keating, Moments of zeta and correlations of divisor-sums: II, in: Advances in the Theory of Numbers,
Springer, 2015, pp. 75–85.
[15] B. Conrey, J. P. Keating, Moments of zeta and correlations of divisor-sums: III, Indagationes Mathematicae 26 (5) (2015)
736–747.
[16] B. Conrey, J. P. Keating, Moments of zeta and correlations of divisor-sums: IV, Research in Number Theory 2 (1) (2016) 24.
[17] B. Conrey, J. P. Keating, Moments of zeta and correlations of divisor-sums: V, Proceedings of the London Mathematical
Society 118 (4) (2019) 729–752.
[18] N. Katz, P. Sarnak, Zeroes of zeta functions and symmetry, Bulletin of the American Mathematical Society 36 (1) (1999)
1–26.
[19] Y. V. Fyodorov, G. A. Hiary, J. P. Keating, Freezing transition, characteristic polynomials of random matrices, and the Riemann
zeta function, Physical Review Letters 108 (17) (2012) 170601.
[20] Y. V. Fyodorov, J. P. Keating, Freezing transitions and extreme values: random matrix theory, and disordered landscapes,
Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 372 (2007) (2014)
20120503.
[21] A. J. Harper, The Riemann zeta function in short intervals [after Najnudel, and Arguin, Belius, Bourgade, Radziwiłł, and
Soundararajan], arXiv preprint arXiv:1904.08204.
[22] A. J. Harper, On the partition function of the Riemann zeta function, and the Fyodorov–Hiary–Keating conjecture, arXiv
preprint arXiv:1906.05783.
[23] J. Najnudel, On the extreme values of the Riemann zeta function on random intervals of the critical line, Probability Theory
and Related Fields 172 (1-2) (2018) 387–452.
[24] L.-P. Arguin, D. Belius, P. Bourgade, M. Radziwiłł, K. Soundararajan, Maximum of the Riemann zeta function on a short
interval of the critical line, Communications on Pure and Applied Mathematics 72 (3) (2019) 500–535.
[25] L.-P. Arguin, P. Bourgade, M. Radziwiłł, The Fyodorov–Hiary–Keating Conjecture. I., arXiv preprint arXiv:2007.00988.
[26] L.-P. Arguin, F. Ouimet, M. Radziwiłł, Moments of the Riemann zeta function on short intervals of the critical line, arXiv
preprint arXiv:1901.04061.
18 E. C. BAILEY AND J. P. KEATING

[27] E. C. Bailey, J. P. Keating, On the moments of the moments of the characteristic polynomials of random unitary matrices,
Communications in Mathematical Physics 371 (2) (2019) 689–726.
[28] T. Assiotis, J. P. Keating, Moments of moments of characteristic polynomials of random unitary matrices and lattice point
counts, Random Matrices: Theory and Applications (to appear).
[29] J. B. Conrey, D. W. Farmer, J. P. Keating, M. O. Rubinstein, N. C. Snaith, Autocorrelation of random matrix polynomials,
Communications in Mathematical Physics 237 (3) (2003) 365–395.
[30] T. Assiotis, E. C. Bailey, J. P. Keating, On the moments of the moments of the characteristic polynomials of Haar distributed
symplectic and orthogonal matrices, Annales de l’Institut Henri Poincaré D (to appear).
[31] T. Claeys, I. Krasovsky, Toeplitz determinants with merging singularities, Duke Mathematical Journal 164 (15) (2015) 2897–
2987.
[32] B. Fahs, Uniform asymptotics of Toeplitz determinants with fisher-hartwig singularities, arXiv preprint arXiv:1909.07362.
[33] E. C. Bailey, J. P. Keating, Moments of Moments and Branching Random Walks, Journal of Statistical Physics 182 (1) (2021)
1–24.
[34] J. P. Keating, N. C. Snaith, Random matrix theory and L-functions at s = 1/2, Communications in Mathematical Physics
214 (1) (2000) 91–100.

S CHOOL OF M ATHEMATICS , U NIVERSITY OF B RISTOL , B RISTOL , BS8 1UG, U NITED K INGDOM

Email address: e.c.bailey@bristol.ac.uk

M ATHEMATICAL I NSTITUTE , U NIVERSITY OF OXFORD , OXFORD , OX2 6GG, U NITED K INGDOM

Email address: jon.keating@maths.ox.ac.uk