1 - arXiv

arXiv:1502.05658v1 [math.NT] 19 Feb 2015
Tail bounds for counts of zeros and eigenvalues, and
an application to ratios
Brad Rodgers
A BSTRACT. Let t be random and uniformly distributed in the interval [T, 2T ], and
consider the quantity N (t + 1/ log T ) − N (t), a count of zeros of the Riemann
zeta function in a box of height 1/ log T . Conditioned on the Riemann hypothesis,
we show that the probability this count is greater than x decays at least as quickly
as e−Cx log x , uniformly in T . We also prove analogous results for the eigenvalues
of a random unitary matrix.
We use results of this sort to show on the Riemann hypothesis that the averages
Z 2T ζ 1 + α + it m
1
2 log T
dt
T T ζ 1 + β + it 2
log T
remain bounded as T → ∞, for α, β complex numbers with β 6= 0. Moreover we
show rigorously that the local distribution of zeros asymptotically controls ratio
averages like the above; that is, the GUE Conjecture implies a (first-order) ratio
conjecture.
1. Introduction
1.1. This paper is comprised of two parts. In the first part we prove, conditioned on the Riemann hypothesis (RH), that local linear statistics of the zeros of
the Riemann zeta function have uniformly sub-exponential tails. More precisely,
label the non-trivial zeros of the zeta function 1/2 + iγ, with γ ∈ R. We prove the
following theorem.
T HEOREM 1.1. Assume RH. Define Q(ξ) := 1/(1 + ξ 2). Then for all x ≥ 2 and all
T ≥ 2.
n
o
X log T
1
meas t ∈ [T, 2T ] :
Q
(γ − t) ≥ x ≪ e−Cx log x ,
T
2π
γ
1
2
BRAD RODGERS
where the constant C and the implicit constant are absolute.
Here and in what follows, zeros are counted with multiplicity (in the unlikely event
that some zero is not simple).
To elaborate on the meaning of this result: the ordinates γ have density log T /2π
T
near a height T , and for t ∈ [T, 2T ], the points { log
(γ − t)} are spaced so as to
2π
have a density of roughly 1, at least for γ near t. Theorem 1.1 therefore bounds the
frequency with which these respaced zeros can occur in large clumps. The theorem
is only of interest when x is large.
Plainly Theorem 1.1 also implies the same estimate when Q is replaced by any
function η that decays quadratically (with constants depending on η). Letting η =
1[0,1/2π] , and defining as usual N(T ) := #{γ : γ ∈ (0, T )}, we obtain a corollary
that may be easier to understand at a glance.
C OROLLARY 1.2. Assume RH. For all x ≥ 2 and all T ≥ 2,
1
meas t ∈ [T, 2T ] : N(t + 1/ log T ) − N(t) ≥ x ≪ e−Cx log x
T
where the constant C and the implicit constant are absolute.
We note that without assuming RH, it is possible to prove an upper bound e−cx ,
where c is an absolute constant.
We apply Theorem 1.1 to consider averages of ratios of the zeta function. We
develop an upper bound for these averages.
T HEOREM 1.3. Assume RH. For any α, β ∈ C with ℜ β 6= 0, and for any m ≥ 0,
uniformly for T ≥ 2,
Z 2T ζ 1 + α + it m
1
2 log T
dt ≪α,β,m 1.
T T ζ 1 + β + it 2
log T
1.2. The second part of the paper requires some knowledge from random matrix theory. Before all else, we develop bounds for counts of eigenvalues of random
unitary matrices analogous to Theorem 1.1. These bounds are apparently new in
the literature.
Moreover, we show rigorously that the asymptotic evaluation of averages of the
sort considered Theorem 1.3 follow from knowing the local distribution of zeros
of the zeta function. Recall the following well-known conjecture about the local
distribution of zeros.
TAIL BOUNDS FOR COUNTS OF ZEROS
3
C ONJECTURE 1.4 (GUE Conjecture). Assume RH. For all fixed k and continuous
and quadratically decaying1 test functions η : Rk → R,
1
T
Z
2T
T
X
γ1 ,...,γk
distinct
η
log T
2π
(γ1 − t), ...,
log T
2π
Z
(γk − t) dt ∼
Rk
η(x) det K(xi − xj ) dk x,
k×k
as T → ∞, where the ij th entry of the k × k determinant is given by K(xi − xj ) =
sin π(xi − xj )/π(xi − xj ).
We also recall a conjecture for the first order asymptotics of ratios of the zeta function.
C ONJECTURE 1.5 (Local Ratios Conjecture with real translations). Assume RH.
For all fixed k ≥ 1 and all fixed collections of numbers α1 , ..., αm , β1 , ..., βm ∈ R,
with βℓ 6= 0 for all ℓ, and2 αi 6= βj for all i, j, we have
E(αi ,βj )
Z 2T Y
m ζ 1 + αℓ + it
det
2
log T
αi −βj
1
,
dt ∼
(1)
1
T T ℓ=1 ζ 1 + βℓ + it
det
2
where
log T
(
e−α+β
E(α, β) :=
1
αi −βj
ℜβ < 0
ℜ β > 0.
As an application of the techniques above, we show that the first of these claims
implies the second.
T HEOREM 1.6. The GUE Conjecture implies the Local Ratios Conjecture with real
translations.
There is a seemingly more general conjecture than Conjecture 1.5 in which α1 , ..., αm ,
β1 , ..., βm are allowed to lie in C, with ℜ βℓ 6= 0 for all ℓ. Such a conjecture may be
called just the Local Ratios Conjecture.
This increase in generality is really only apparent. It is possible using similar methods to see that the GUE Conjecture also implies the Local Ratios Conjecture, for
general α and β. The proof of this claim requires a somewhat more lengthy technical argument, so we will not prove it here. We will instead say only a few words
about what modifications in the proof of Theorem 1.6 are necessary for it at the end
of this paper.
1
1
. A purist may object that it is
By quadratically decaying, we mean η(x) = O 1+x
2 · · · 1+x2
1
k
more natural to make this conjecture for only compactly supported η, but these two versions of this
conjecture may be seen without too much effort to be equivalent.
2
We clearly lose no generality from this restriction.
1
4
BRAD RODGERS
1.3. The study of the average of ratios of the zeta function has a long history.
Conjecture 1.5 was first put forward in the case m = 2 by Farmer [13], who understood it was closely connected with the local distribution of zeros of the Riemann
zeta function. Farmer showed that the m = 2 case of (a uniform version of) what
we have called the Local Ratios Conjecture implies the k = 2 case (pair correlation) of the GUE Conjecture [14], and later produced similar implications for the
m = 3, k = 3 case, while even higher correlations may be obtained from [10]. To
our knowledge the present paper is the first rigorous work in the opposite direction.
More recently, a flurry of work has centered around the average of such ratios when
the translations are not within a distance of O(1/ log T ) of the critical axis, but
instead are up to a distance of O(1) away. In this case great deal of effort has been
put into not only producing asymptotic formulas, but extracting all relevant lower
order terms [6], which have many interesting implications [9]. (We have called
Conjecture 1.5 a ‘Local Ratios Conjecture’ to distinguish it from this expanded
set of conjectures.) Indeed, it is worth noting at this point that the formula in (1)
is not the usual way to write the ratio conjecture; instead one usually insists that
ℜ βl , βℓ′ > 0 and conjectures that
1
T →∞ T
lim
Z
T
m
2T Y
l=1
ζ
ζ
1
2
+
αl
log T
1
2
+
βl
log T
m′ ζ 1 +
+ it Y
2
+ it ℓ=1 ζ 12 +
α′ℓ
log T
βℓ′
log T
+ it
dt
+ it
(2)
is predicted accurately to first order by a random matrix analogue. The expression
for this limit is somewhat more complicated to write down than the formula on
the right hand side of (1) (see for instance [8, 7, 4]). Nonetheless, in spite of the
simplicity of (1), there does not appear to be any way to write down the more
precise lower-order Ratio Conjectures in a way reminiscent of it. It would still be
interesting to see if such a combinatorial formalism can be found.
In any case, an asymptotic formula for the left hand side of (1) implies an asymptotic formula for the left hand side of (2), and vice-versa. This may be seen most
easily by applying the zeta function’s functional equation. We will have nothing to
say about lower order terms however.
Similarly to Farmer’s papers above, some previous work has studied the connections of the GUE Conjecture to averages of the logarithmic derivative of the zeta
function [17, 15, 27].
We note also the concurrent work [5], which considers some similar questions to
those we consider here, but replaces the zeta function with a probabilistic construction called the limiting characteristic polynomial.
TAIL BOUNDS FOR COUNTS OF ZEROS
5
1.4. We turn to a quick conceptual sketch of some of our methods. Both Theorems 1.3 and 1.6 are critically dependent on the tail bound, Theorem 1.1. The
strategy in each case is to write
1
α
h
1
1
i
ζ 2 + log T + it
α
β
= exp Log ζ
+
+ it − Log ζ
+
+ it ,
2 log T
2 log T
ζ 21 + logβ T + it
{z
}
|
:=Lt
(3)
(ignoring for the moment all issues with branch cuts, which end up being minor).
We show from the Hadamard product
representation
for the zeta function that Lt is
P log T
‘very close’ to a linear statistic η 2π (γ − t) , for some function η of quadratic
decay. This is not literally true: Lt , if written as a sum of zeros, must contain
an extra term in the summand that decays very slowly. This term does not decay
quadratically – in fact its sum converges only because of the symmetry of zeros –
but it may be shown that on average this extra term does not much affect the size of
Lt . (This will be the content of Theorem 2.6.)
Thus it is that we see that we can approximate the ratio (3) by the exponential of a
linear statistic of zeros. It is just these linear statistics whose size we have controlled
in Theorem 1.1, and it is in this way that Theorem 1.3 is proved. For Theorem 1.6,
on the other hand, we note that we are able to asymptotically control the moments
of such linear statistics by using the GUE Conjecture and a standard combinatorial
procedure. This asymptotic control on the moments of linear statistics is not ipso
facto enough to pass to the Local Ratios Conjecture however. It is not the case,
that is, that Theorem 1.6 is just a matter of combinatorial manipulation in random
matrix theory.
For instance, instead of Lt , consider the random variables Xn which take the value
0 with probability 1 − e−n and n2 with probability e−n . Then Xn tends to 0 both in
distribution and in the sense of moments: for any fixed k ≥ 0,
E Xnk → 0.
Yet
Xn
so it is not true E e
E eXn = (1 − e−n ) + en
∼ E e0 .
2 −n
→ ∞,
This sort of a pathology is eliminated by Theorem 1.1 and related bounds, and it is
this control that is necessary to show that the average of ratios in (1) converges to a
random matrix limit on the GUE Conjecture.
Our proof of Theorem 1.1 is not long provided certain computational lemmas are
taken on faith, so we will not sketch it here. We mention only that our proof depends
6
BRAD RODGERS
on an application of Markov’s inequality and a smoothing trick. It is, in this sense,
reminiscent of Soundararajan’s approach [30] to bounding the moments of ζ(1/2 +
it) (see also Harper’s refinement [19]).
Finally, we note that in the case that ℜα ≤ ℜβ and ℑα = ℑβ, there is an easier
proof of the bound in Theorem 1.3. In this case one has for all t, T ≥ 2 a pointwise
bound
1
1
α
β
ζ
+
+ it ≪α,β ζ
+
+ it .
2 log T
2 log T
This is a consequence of Lemma 1 of [25]. Nonetheless, such an inequality does
not hold for other ranges of α and β, and Theorem 1.3 cannot in general be reduced
to a pointwise estimate of this sort.
Notation: We follow standard conventions of analytic number theory, so that the
notations f (x) ≪ g(x) and f (x) = O(g(x)) are interchangeable, with both meaning that |f (x)| ≤ Cg(x) for all x, for a constant C. f (x) ≪A g(x) and f (x) =
OA (g(x)) both mean the constantRC may depend on A. The Fourier transform of a
function f is defined by fˆ(ξ) := e−i2πxξ f (x) dx.
In what follows we will assume the Riemann hypothesis, without further statement
of this assumption in Theorems, Lemmas, etc.
1.5. Acknowledgments: I thank Sandro Bettin, Alexei Borodin, Reda Chhaibi,
Brian Conrey, Chris Hughes, Jon Keating, and Kurt Johansson for informative and
encouraging discussions related to this work.
2. Bounding counts of zeros: a proof of Theorem 1.1 and related bounds
2.1. As in many studies of the zeros of the zeta function, a principal tool is the
explicit formula, due in stages to Riemann, Guinand, and Weil [26, 18, 32], relating
the distribution of zeros to primes. A proof may be found in, for instance, [24, pp.
410-416] or [21, pp. 108-109].
T HEOREM 2.1 (The explicit formula). For a compactly supported function g, piecewise continuous with finitely many discontinuities, such that g(x) = 12 (g(x− ) +
g(x+ )) for all x and g(0) = 12 (g(x) + g(−x)) + O(|x|), we have,
Z ∞
X γ Z V ξ Ω(ξ)
−
gˆ
dξ =
(g(x)+g(−x))e−x/2 d ex −ψ(ex ) ,
lim
gˆ
V →∞
2π
2π 2π
−V
−∞
|γ|<V
TAIL BOUNDS FOR COUNTS OF ZEROS
where
ψ(x) :=
X
7
Λ(n),
n≤x
with Λ the von Mangoldt function, and
1 Γ′ 1
ξ 1 Γ′ 1
ξ
Ω(ξ) :=
+
− log π.
+i
−i
2Γ 4
2
2Γ 4
2
Using Stirling’s formula for the digamma function [1, Cor. 1.4.5], one may verify
that,
1 log (|ξ| + 2)/2π
Ω(ξ)
.
(4)
=
+O
2π
2π
|ξ| + 2
This term in the explicit formula therefore corresponds to an approximation of the
density of zeros near height ξ. On the other hand,
Z ∞
Z ∞
∞
X
g(log t)
g(log n)
−x/2
x
x
√
√
g(x)e
d e − ψ(e ) =
dt −
Λ(n),
n
t
−∞
0
n=1
√
R
P
√
and here the term g(log t)/ t dt serves as an approximation to g(log n)Λ(n)/ n.
Motivated by the explicit formula, we adopt the following notation, for a function
η of quadratic decay:
X log T
(γ − t) ,
hη, Zi = hη, ZT (t)i :=
η
2π
γ
Z ∞ Ω(ξ)
log T
o
o
hη, Z i = hη, ZT (t)i :=
η
(ξ − t)
dξ,
2π
2π
−∞
e = hη, ZeT (t)i := hη, Zi − hη, Z o i.
hη, Zi
Note that there is no question about the convergence of the sums or integrals in
these definitions. We will later generalize this notation slightly, but we need not
worry about this generalization for the moment. Note that for typographical reasons
we will sometimes write Z or ZT in place of ZT (t). Unless otherwise indicated,
e
Z = ZT = ZT (t), and likewise for Z o and Z.
2.2.
tion
Let B0 be an absolute constant to be defined shortly. We define the funcG(ξ) := B0
h sin π(ξ + 1/4) 2
π(ξ + 1/4)
sin π(ξ − 1/4) 2 i
,
+
π(ξ − 1/4)
with Fourier transform,
ˆ
G(x)
= B0 (1 − |x|)+ eiπx/2 + e−iπx/2 ,
(5)
(6)
8
BRAD RODGERS
where B0 is an absolute constant chosen so that
Q(ξ) ≤ G(ξ)
∀ξ ∈ R.
(7)
(In fact, B0 may be chosen to be 2π 2 , but we only need to know such a constant
exists, which is apparent from examining G(ξ)/Q(ξ).) There is nothing very special
about this test function G; we have chosen it to satisfy (7) and
ˆ ⊆ [−1, 1].
supp G
(8)
As a consequence of (7), writing Gk (ξ) := G(ξ/k), we see that for all k ≥ 1,
Q(ξ) ≤ Gk (ξ),
∀x ∈ R.
(9)
Moreover,
ˆ k ⊆ [−1/k, 1/k],
supp G
with
ˆ k (x)| ≤ 2B0 k(1 − |kx|)+ .
|G
(10)
To make for a cleaner presentation, we work with notation from elementary probability, letting t be a random variable uniformly distributed on the interval [T, 2T ].
Theorem 1.1 then becomes the claim that uniformly for x ≥ 2 and T ≥ 1,
P(hQ, Zi ≥ x) ≤ e−Cx log x .
The reason we have defined Gk is that the size of hQ, Zi can be controlled by
e It is easy
hGk , Zi, and that this in turn can be controlled by hGk , Z o i and hGk , Zi.
o
to control hGk , Z i, since the measure defining this quantity is very regular. On the
e can be well-controlled up to the k-th moment, with hGk , Zi
e in
other hand hGk , Zi
e More exactly, we prove the following
general not being much larger then hG, Zi.
estimates.
L EMMA 2.2. For an absolute constant B1 , uniformly for T ≥ 2,
hQ, ZT (t)i ≤ B1 log T,
∀t ∈ [T, 2T ].
L EMMA 2.3. For an absolute constant B2 , uniformly for T ≥ 2 and 2ℓ ≤ k, we
have
e 2ℓ ≤ (B2 ℓ)ℓ .
E|hGk , Zi|
√
L EMMA 2.4. For an absolute constant B3 , uniformly for T ≥ 2 and k ≤ T ,
hGk , ZTo (t)i ≤ B3 k,
∀t ∈ [T, 2T ].
These lemmas have standard proofs, which we turn to at the end of this section, but
before doing so, we show that with these computational estimates in hand, Theorem
1.1 follows quickly.
TAIL BOUNDS FOR COUNTS OF ZEROS
9
P ROOF OF T HEOREM 1.1. Note first that in the case that x > B1 log T , Lemma
2.2 implies that
P(hQ, Zi ≥ x) = 0.
We may therefore assume x ≤ B1 log T . Lemma 2.3 allows us to see from Markov’s
inequality that for even integers k and positive y,
k/2
e ≥ y) ≤ 1 E|hGk , Zi|
e k ≪ (B2 k) .
P(hGk , Zi
yk
yk
Yet
Thus,
e + hGk , Z o i.
hQ, Zi ≤ hGk , Zi = hGk , Zi
(11)
(12)
e + hGk , Z o i ≥ x)
P(hQ, Zi ≥ x) ≤ P(hGk , Zi
e ≥ x − B3 k ,
≤ P hGk , Zi
(13)
√
for all even k ≤ T , with the last line following from Lemma 2.4. With no loss of
generality, we may assume x ≥ 4B3 , and consider k defined to be the positive even
integer satisfying
x
x
−2 <k ≤
2B3
2B3
so that in particular
x − B3 k ≥ x/2.
√
√
As long as T is large enough that B1 /2B2 log T ≤ T , then certainly k ≤ T
(since we are considering the case x ≤ B1 log T ). Thus from (11) and (13),
(B2 · x/2B3 )x/4B3
P(hQ, Zi ≥ x) ≤
≪ e−Cx log x ,
x/2B
−2
3
(x/2)
(14)
for an absolute constant C.3
In
√ remains to verify our claim in the case in which T is small enough that B1 /2B2 log T >
T . But this bounded range of T can at most alter the implicit constant in (14). Remark: There is a slightly different approach to this theorem which some readers
may prefer. Instead of the inequality (12), we may make use of a mollification
formula of Selberg [28, Th. 1], which approximates the classical function S(t) by
a Dirichlet polynomial with error terms whose size depends on the length of the
Dirichlet polynomial. One may then compute moments of, say, S(t + 1/ log T ) −
S(t) in the same way we have here, with the Dirichlet polynomial replacing the
e We have taken the route that we have because we will make use
quantity hGk , Zi.
3An argument with more bookkeeping, though still one which makes no attempt at optimization,
shows that one may take any constant C < 1/16π 2 , for instance.
10
BRAD RODGERS
of the same formalism elsewhere in this paper; it applies almost without change to
study the eigenvalues of the unitary group, for instance.
On the other hand, Selberg [29, Th. 2] also proves an uncondtional variant of his
approximation for S(t), and this has been used by Fujii [16, p. 245] to compute
moment bounds for S(t + 1/ log T ) − S(t) unconditionally. Bounds that can be
obtained unconditionally in this way are slightly worse than what we have derived
1
assuming RH. Unconditionally, using the
technique,
one
can
prove
meas
t ∈
T
−cx
[T, 2T ] : N(t + 1/ log T ) − N(t) ≥ x ≪ e , where c is an absolute constant,
but seemingly no better. It would be interesting to see if this could be improved.
Remark: Probably Theorem 1.1 and Corollary 1.2, while sufficient for our purposes, are not optimal. The bounds here would correspond to the ‘right answer’
were the zeros were modeled by a Poisson process, but since zeros of the zeta function tend to repel each other one might guess that the counts are sub-gaussian in
Theorem 1.1 and Corollary 1.2. Seemingly this is a harder statement to prove.
2.3. There is another result similar to Theorem 1.1 that we will require, but
which is somewhat more technical in its statement and proof. We generalize the
e to a wider class of functions than it was applied to before. In parnotation hη, Zi
ticular, we let
Z V log T
Ω(ξ)
X log T
e
e
hη, Zi = hη, ZT (t)i := lim
η
(γ−t) −
(ξ−t)
dξ,
η
V →∞
2π
2π
2π
−V
|γ|<V
where η, T, and t are such that the limit exists. This is consistent with our previous
use of this notation. Likewise, when the limit exists,
X log T
(γ − t) ,
hη, Zi = hη, ZT (t)i := lim
η
V →∞
2π
|γ|<V
o
hη, Z i =
hη, ZTo (t)i
:= lim
V →∞
Z
V
−V
η
log T
2π
(ξ − t)
Ω(ξ)
2π
dξ.
e exists whenever η(ξ) = fˆ(ξ),
By the explicit formula, it may be verified that hη, Zi
for a function f that is (i) compactly supported, (ii) piecewise continuous with
finitely many discontinuities, (iii) satisfying f (x) = 21 (f (x+ ) + f (x− )), and (iv)
with f odd. A more specific example of such a limit existing where the sums and
integral do not absolutely converge is furnished by the function
J(ξ) :=
2πξ
.
1 + (2πξ)2
TAIL BOUNDS FOR COUNTS OF ZEROS
11
In this case, J(ξ) = fˆ(ξ), for the function
f (x) := −sgn(x)e−|x| /2i,
(15)
so one may see by the above discussion that hJ, ZeT (t)i is well defined for all T
e
and t. Alternatively, one may see rather more simply that the limit defining hJ, Zi
converges by exploiting the symmetry of the zeros γ and the function Ω. Indeed, let
us verify this (and prove a little more) for hJ, Z o i, in a lemma we will need later.
L EMMA 2.5. Uniformly for T ≥ 2,
hJ, ZTo (t)i = O(1/ log T ),
∀t ∈ [T, 2T ].
P ROOF. By the symmetry of J,
Z V
h 2πy 2πy i dy
o
hJ, ZT (t)i = lim
J(y) Ω t +
−Ω t−
V →∞ 0
log T
log T
log T
Z ∞
i
h
2
2
2
1
y
1
t log T
dy,
≪
J(y) min 2 2 ,
+
O
t − 2πy + 2
log T 0
y2
t log T
log T
where in the second step in approximating Ω, we have used Stirling’s formula (4)
and then simple Taylor series estimates for the logarithm function. (Note that in the
first line the integrand is positive, so the integral converges absolutely or not at all.)
It is now slightly tedious but straightforward to verify that the integral is O(1) and
therefore the entire expression is O(1/ log T ).
The analogue of Theorem 1.1 that we require is the following.
T HEOREM 2.6. For all x ≥ 2 and all T ≥ 2,
e ≥ x ) ≪ e−Cx log x ,
P( |hJ, Zi|
where the constant C and the implicit constant are absolute.
Applying Lemma 2.5 here, we see likewise:
C OROLLARY 2.7. For all x ≥ 2 and all T ≥ 2,
P( |hJ, Zi| ≥ x ) ≪ e−Cx log x ,
where the constant C and the implicit constant are absolute.
Our proof of Theorem 2.6 is similar to the proof of Theorem 1.1. Again we require
a series of lemmas, to be proved later.
L EMMA 2.8. For an absolute constant B1′ , uniformly for T ≥ 2,
|hJ, ZeT (t)i| ≤ B ′ log T, ∀t ∈ [T, 2T ].
1
12
BRAD RODGERS
For the next two lemmas we define
W (ǫ) (x) :=
sgn(x) −|x|
e (1 − |x|/ǫ)+ .
−2i
(16)
We have defined W (ǫ) so that (W (1/k) )ˆ, for k ≥ 1, plays the role of something like
a smooth approximation to the function
J(ξ)1|ξ|≥k .
More exactly, a computation reveals that,
1 − exp 1+i2πz 1 − exp 1−i2πz k
k
k
−
.
(W (1/k) )ˆ(z) = J(z) +
2i
(1 + i2πz)2
(1 − i2πz)2
|
{z
} |
{z
}
K(z), say
(17)
K(−z)
(We have written z instead of ξ here, because we will later need this expression for
complex values of z as well.) We will see from a Taylor expansion, (W (1/k) )ˆ(ξ) is
small when |ξ| ≤ k, and the terms K(ξ) may be thought of as an error term when
|ξ| is large. A more exact statement of this is as follows:
L EMMA 2.9. For all k ≥ 1,
|J(ξ) − (W (1/k) )ˆ(ξ)| ≤ A Gk (ξ),
∀ξ ∈ R,
where A is an absolute constant.
We also have the moments of (W (1/k) )ˆare very small when k is large.
L EMMA 2.10. For an absolute constant B2′ , uniformly for T ≥ 2 and 2ℓ ≤ k ≤
√
T , we have
e 2ℓ ≤ (B2′ ℓ)ℓ k −2ℓ .
E|h(W (1/k) )ˆ, Zi|
As before, we momentarily delay the proof of these lemmas. Assuming them, we
see that a proof of Theorem 2.6 proceeds in the same manner as that of Theorem
1.1.
P ROOF
OF
T HEOREM 2.6. If x > B1′ log T , then by Lemma 2.8,
e ≥ x) = 0.
P(|hJ, Zi|
So as before we may treat the case that x ≤ B1′ log T . By applying Lemma 2.9, for
all k ≥ 1,
e = h(W (1/k) )ˆ, Zi
e + O(|hGk , Zi|)
e + O(hGk , Z o i),
hJ, Zi
TAIL BOUNDS FOR COUNTS OF ZEROS
13
where the implicit constant in the first error term
√ may be taken as A, and the implicit
constant in the second 2A. As long as k ≤ T , Lemma 2.4 allows us to bound the
second of these error terms: hGk , Z o i ≤ B3 k. Hence using a union bound,
e ≥ x) ≤ P |h(W (1/k) )ˆ, Zi|
e ≥ x/2 + P A|hGk , Zi|
e + AB3 k ≥ x/2 .
P(hJ, Zi|
A choice of k and bound for both probabilities then proceeds as in the proof of
Theorem 1.1, replacing Lemma 2.3 by Lemma 2.10 to bound the first of these
terms.
There is finally one last result of this sort that we will use below.
L EMMA 2.11. For any ǫ ≥ 0 and k ≥ 1/ǫ2 , for T = T (ǫ) sufficiently large,
P(|h(W (1/k) )ˆ, ZT i| ≥ ǫ) ≪ ǫ2 .
2.4.
P ROOF
We finally turn to proofs of the lemmas above.
OF
L EMMA 2.2. We recall the estimate (see [24, Cor. 14.3]),
N(t + 1) − N(t) ≪ log(|t| + 2),
∀t ∈ R.
By inspection, it is easy to verify that log(|u + v| + 2) ≪ log(|u| + 2) + log(|v| + 2)
for all u, v ∈ R.
Now note that for t ∈ [T, 2T ],
hQ, ZT (t)i ≪
∞
X
[N(t + k + 1) − N(t + k)] ·
k=−∞
1
1 + log2 T
k2
∞
X
1
1
≪ log(T + 2)
+
log(|k| + 2)
2
2
2
1 + k log T k=−∞ 1 + k log2 T
k=−∞
≪ log(T ).
∞
X
P ROOF OF L EMMA 2.3. This is a more or less standard computation of moments.
However, some added care is necessary since an estimate is required that is uniform
14
BRAD RODGERS
as moments vary. We note that from the explicit formula,
Z ∞ k
ˆ kx (e−ixt + eixt )e−x/2 d ex − ψ(ex )
e
G
hGk , ZT (t)i =
log T −∞
log T
log p
X
X kr
k
ˆ
=I − 2ℜ
G
log p r(1/2−it) ,
log T r≥1 p
log T
p
|
{z
}
P
r
where
sr , say
log T
log T
(i/2 − t) + G
(i/2 + t) .
2πk
2πk
Here G(x+iy) is, of course, the analytic continuation of the function defined before
in (5). One may check that
log T
exp(log T /4k)
1
G
(i/2 ± t) ≪
≪ 7/4 ≪ 1,
2
2πk
t
T
for k ≥ 1 and t ∈ [T, 2T ].
I := G
Hence from H¨older’s inequality,
1/2ℓ
e 2ℓ
EhGk , Zi
≪1+
1/2ℓ
k X
E|sr |2ℓ
.
log T r≥1
(18)
ˆ ⊆ [−1, 1], a standard argument dating back to Selberg (see [30,
Because supp G
Lem. 3], for a modern treatment that applies directly) reveals4 that for 2ℓ ≤ k
X
2 ℓ
log2 p ˆ kr
2ℓ
E|sr | ≪ ℓ! ·
G
log p
.
r
p
log
T
p
ˆ this quantity is null for all r ≥ 1, when k > log T / log 2.
By the support of G,
In the case that k ≤ log T / log 2 we need a little more work. When r = 1,
2
log T 2
X log2 p
X log2 p kr
ˆ
G
log p ≪
≪
,
p
log T
p
k
1/k
p
p≤T
by Chebyshev (see [24, Ch. 2.2]). When r ≥ 2,
Z ∞
2
X log2 p kr
1
log2 t
ˆ
G
log
p
dt ≪ r .
≪
r
r
p
log T
t
2
2
p
4
In fact, the argument shows that up to twice our range of ℓ may be admitted.
TAIL BOUNDS FOR COUNTS OF ZEROS
15
Returning to (18), we see that
e 2ℓ
EhGk , Zi
1/2ℓ
≪1+
≪ ℓ1/2 ,
log T X 1 k
(ℓ!)1/2ℓ
+
log T
k
2r/2
r≥2
as k ≤ log T / log 2. We have used Stirling’s formula [1, Th. 1.4.1] to bound the
factorial. Exponentiating by 2ℓ gives the lemma.
P ROOF
L EMMA 2.4. We have
Z ∞
2πy 1
o
dy
Gk (y)Ω t +
hG, ZT (t)i =
log T −∞
log T
Z
Z
2πy 1
+
Gk (y)Ω t +
dy.
=
log T
log T
|y|≤T log T
|y|>T log T
OF
By our application of Stirling’s formula (4), this quantity is
Z
Z
1
k2
≪
Gk (y) log T dy +
log y dy
2
log T
|y|≤T log T
|y|>T log T y
≪k + k 2 /T,
which yields the estimate.
P ROOF
OF
L EMMA 2.8. It is easy to verify for ℜs > 1/2 that
Z ∞
ζ′ 1
1
−sx −x/2
x
x
e e
d(ψ(e ) − e ) = −
+s −
.
ζ 2
s − 1/2
0
On RH, by analytic continuation, this identity remains true for ℜs > 0. Making use
of this identity, the Fourier transform expression (15), and the explicit formula, one
may thus verify that
ζ′ 1
1
1
e
ℑ
+
+ it
hJ, ZT (t)i =
log T ζ 2 log T
1 t
t
+
−
2
1
log T 1 − 1 2 + t2
+ log1 T + t2
2
log T
2
1 ζ′ 1
1
1
.
ℑ
+
+ it + O
=
log T ζ 2 log T
log T
(19)
16
BRAD RODGERS
From Lemma 12.1 of [24], we see that for t ∈ [T, 2T ],
X
ζ′ 1
1
1
+
+ it =O(1) +
ζ 2 log T
1/ log T − i(γ − t)
|γ−t|≤1
=O(1) + O(log2 T )
=O(log2 T ),
with the second to last line following from the fact that N(t + 1) − N(t) =
O(log(|t| + 2). Combining this estimate with (19) yields the lemma.
P ROOF OF L EMMA 2.9. A Taylor expansion of the exponential function in (17)
shows that for |ξ| ≤ k (throughout this proof, ξ is real),
(W (1/k) )ˆ(ξ) ≪ 1/k.
(20)
In the same range, plainly J(ξ) ≪ 1. Hence, for |ξ| ≤ k,
|(W (1/k) )ˆ(ξ) − J(ξ)| ≪ 1 ≪ Gk (ξ).
On the other hand, for |ξ| > k, by (17),
|(W (1/k) )ˆ(ξ) − J(ξ)| ≪
k
k2
≪
≪ Gk (ξ).
ξ2
ξ2
(21)
P ROOF OF L EMMA 2.10. Our proof proceeds along the same lines as that of Lemma
2.3. From the explicit formula,
1 XX
r log p log p
(1/k)
′
e
,
h(W
)ˆ, ZT (t)i = I + ℑ
1−k
log T r≥1 p pr(1/2+1/ log T −it)
log T +
{z
}
|
P
r
where
σr , say
log T
log T
(1/k)
I := (W
)ˆ
(i/2 − t) + (W
)ˆ
(−i/2 − t) .
2π
2π
To bound I ′ , we recall (17). It is simple to verify that for t ∈ [T, 2T ],
log T
1
1
J
(±i/2 − t) ≪
≪
2π
T log T
k
√
in the range that k ≤ T . On the other hand, a bit more tediously,
log T
log T k
k
1
K ±
≪ 7/4 ≪ ,
(±i/2 − t) ≪
exp
2
2
2π
(1 − log T /2) + t
2k
T
k
′
(1/k)
TAIL BOUNDS FOR COUNTS OF ZEROS
again for k ≤
√
17
T . This shows that
1
.
k
Thus, as in the H¨older inequality (18) of the proof of Lemma 2.3,
1/2ℓ
1/2ℓ
1 X
1
e 2ℓ
E|σr |2ℓ
.
Eh(W (1/k) )ˆ, Zi
≪ +
k log T r≥1
I′ ≪
(22)
But also as in that proof, for k > log T / log 2,
σr = 0,
∀r ≥ 1.
Otherwise, for k ≤ log T / log 2, the right hand side of (22) is likewise bound by
log T X 1 1
1
≪ +
(ℓ!)1/2ℓ
+
k log T
k
2r/2
r≥2
≪ℓ1/2 /k.
This proves the lemma.
P ROOF OF L EMMA 2.11. We begin by considering the case that ǫ > 1/2. In this
case, the lemma is tautological:
P |h(W (1/k) )ˆ, Zi| ≥ ǫ ≤ 1 ≪ ǫ2 .
We may therefore suppose ǫ ∈ (0, 1/2). From Lemma 2.10, we see (noting that this
condition on ǫ imposes k ≥ 2),
e ≥ ǫ ≤ 1 E |h(W (1/k) )ˆ, Zi|
e 2
P |h(W (1/k) )ˆ, Zi|
ǫ2
1
≤ ǫ2 .
≪
(ǫk)2
On the other hand, from (20) and (21), for all k ≥ 1, we have
(W (1/k) )ˆ(ξ) ≪ J(ξ),
∀ξ ≥ 0.
(23)
Hence, using the symmetry of (W (1/k) )ˆin the first line below,
Z V
h 2πy 2πy i dy
(1/k)
(1/k)
o
(W
)ˆ(y) Ω t +
h(W
)ˆ, ZT (t)i = lim
−Ω t−
V →∞ 0
log T
log T
log T
1
.
≪hJ, ZTo (t)i ≪
log T
We are justified in applying the bound (23) in passing to the second line because,
as in the proof of Lemma 2.5, Ω(t + 2πy/ log T ) − Ω(t − 2πy/ log T ) ≥ 0 for all
y ≥ 0.
18
BRAD RODGERS
Thus for sufficiently large T (such that 1/ log T is small in comparison to ǫ),
e ≥ ǫ/2 ≪ ǫ2 ,
P |h(W (1/k) )ˆ, Zi| ≥ ǫ ≤ P |h(W (1/k) )ˆ, Zi|
as claimed.
3. Ratio bounds
With the bounds of Theorems 1.1 and 2.6 in place, it is a simple matter to bound
moments of ratios of the zeta function.
P ROOF OF T HEOREM 1.3. In this proof we assume ℜ β 6= 0 throughout. Using the
Hadamard product representation for the zeta function [24, Th. 10.12] and Stirling’s
formula for the Gamma function [1, Cor. 1.4.3], it is straightforward to verify that
for fixed α, β ∈ C,
ζ 21 + logα T + it
Y α − i log T (γ − t)
2π
2π
= (1 + o(1)) lim
,
(24)
log T
β
β
V →∞
1
−
i
(γ
−
t)
ζ 2 + log T + it
2π
|γ|≤V 2π
where because ℜβ 6= 0 the product converges to a finite number (on RH). Here o(1)
is a quantity that tends to 0 uniformly for t ∈ [T, 2T ] as T → ∞.
Using
Log(z) := log |z| + iArg(z),
with Arg(z) ∈ (−π, π] for all z ∈ C, and defining
α − iξ Lα,β (ξ) := Log 2π
,
β
−
iξ
2π
one sees that the expression (24) is equal to
(1 + o(1)) exp hLα,β , ZT (t)i
(25)
(A little care must be taken, of course, whenever taking the logarithm of a complex number, but here, due to the exponential, no problems arise. One must check
that the sum defining hLα,β , ZT (t)i converges, but this is straightforward using the
symmetry of γ.)
Note that for |ξ| > max(2π|α|, 2π|β|)
1 − α i2πξ
Lα,β (ξ) = Log
= i(α − β)J(ξ) + Oα,β (Q(ξ)),
β
1 − i2πξ
TAIL BOUNDS FOR COUNTS OF ZEROS
19
while, as long as ℜβ 6= 0, we have for |ξ| < max(2π|α|, 2π|β|),
α
− iξ
2π
=
exp
O
Q(ξ)
,
α,β
β
− iξ
2π
since for this region of ξ, the left hand side is bounded above, and the right hand
side is bounded from below. Hence for all ξ ∈ R,
Lα,β (ξ) = i(α − β)J(ξ) + Oα,β (Q(ξ)).
Thus for fixed α, β, m, with ℜ β 6= 0,
1
ζ 2 + logαT + it m
=(1 + o(1)) exp mℜ hLα,β , Zi
ζ 1 + β + it 2
log T
=(1 + o(1)) exp O hJ, Zi + O hQ, Zi .
Now the theorem at hand follows from Theorem 1.1 and Corollary 2.7.
(26)
Remark: There is an alternative to the identity (25) that is more exact algebraically.
Under RH, it may be seen (for instance, with [31, Eq. (14.10.5)] as a starting point)
that for ℜ α, β > 0,
ζ 21 + logα T + it
e .
= exp hLα,β , Zi
(27)
ζ 21 + logβ T + it
To use this identity in the proof above to treat those values of α or β with negative
real part, the functional equation must be made use of.
4. A random matrix interlude
4.1. In this section, we develop analogues for the unitary group of our tail
bound for linear statistics (for zeta zeros this was Theorem 1.1), the determinantal
evaluation of correlation functions (for the zeta zeros, Conjecture 1.4), the evaluation of ratios of the zeta function (Conjecture 1.5), a uniform upper bound on
moments of ratios (Theorem 1.3) and also the more technical tail bound for oscillatory linear statistics (Lemma 2.11). We will make use of these estimates in the next
section.
The unitary group U(N) is the group of N × N complex matrices g satisfying
g ∗ g = I. In what follows we endow this group with Haar probability measure. Any
such unitary matrix g has N eigenvalues that lie on the unit circle, which we write
as {ei2πθ1 , ..., ei2πθN } with θi ∈ [−1/2, 1/2) for all i.
20
BRAD RODGERS
The k level correlations of eigenvalues are in this case known exactly [3, Eq.
(39.12)].
T HEOREM 4.1 (The Weyl-Gaudin-Dyson integration formula). For k ≤ N and any
integrable function η : [−N/2, N/2)k → C,
Z
X
η(Nθj1 , ..., Nθjk ) =
η(x) det KN (xi − xj ) dk x,
EU (N )
where KN (x) :=
k×k
[−N/2,N/2)k
j1 ,...,jk
distinct
sin(πx)
.
N sin(πx/N )
This implies that for any integrable function η : Rk → C,
Z
X
η(Nθj1 , ..., Nθjk ) ∼
η(x) det K(xi − xj ) dk x.
EU (N )
k×k
Rk
j1 ,...,jk
distinct
T
(γ −t)}
This formula of course mirrors the GUE Conjecture, so that the points { log
2π
may be modeled by the random points {Nθi }.
In fact, instead of the collection of points {Nθ1 , ..., NθN }, it will be even more
natural to work with these
S points pulled back to have period N; that is we consider
the collection of points ν∈Z {N(θ1 + ν), ..., N(θN + ν)}. The reader may check
that here too we have,
Z
X X
η(N(θj1 + ν), ..., N(θjk + ν)) =
η(x) det KN (xi − xj ) dk x
EU (N )
k×k
Rk
j1 ,...,jk ν∈Zk
distinct
∼
Z
Rk
η(x) det K(xi − xj ) dk x.
k×k
(28)
We label the characteristic polynomial of a random unitary matrix g in the following
way:
Λ(A) := det(1 − e−A g),
(29)
where A may be any complex number.
Note that
N
Λ(α/N) Y sin π(θi − α/2πN)
=
Λ(β/N) i=1 sin π(θi − β/2πN)
N Y
V
Y
α
2π
β
V →∞
i=1 ν=−V 2π
= lim
− N(θi + ν)
− N(θi + ν)
,
(30)
TAIL BOUNDS FOR COUNTS OF ZEROS
21
where in passing to the last line we have made use of the classical identity
∞ Y
z2 1− 2 .
sin πz = πz
ℓ
ℓ=1
Aside from being useful later on, by comparison with (24), the identity (30) makes
transparent the similarity between ratios of characteristic polynomials and ratios
of the zeta function. For these ratios, we note a formula that, in effect, is due to
Borodin, Olshanksi, and Strahov [2].
T HEOREM 4.2. For complex numbers A1 , ..., Am and B1 , ..., Bm with ℜ Bℓ 6= 0 for
all ℓ and Ai 6= Bj for all i, j,
E(N Ai ,N Bj )
m
det
Y Λ(Aℓ )
Bj
Ai
e −e EU (N )
=
1
Λ(Bℓ )
det
ℓ=1
eAi −eBj
Recall that the function E is defined in Conjecture 1.5.
In fact, the authors in [2] do not prove exactly Theorem 4.2, but rather a somewhat
more general statement which may be seen with a little work to imply it. An account
of this short derivation from [2] to Theorem 4.2 will be found in section 5.4 of the
forthcoming paper [5]. There is also another proof, based on supersymmetry, in the
paper [23]. This paper uses a rather different notation, but Theorem 4.2 is in fact a
specialization of identity (4.35) there.
As a simple corollary,
C OROLLARY 4.3. For complex numbers α1 , ..., αm and β1 , ..., βm with ℜ βℓ 6= 0
for all ℓ, and αi 6= βj for all i, j,
E(αi ,βj )
m
det αi −βj
Y Λ(αℓ /N)
,
EU (N )
∼
1
Λ(βℓ /N)
det
ℓ=1
αi −βj
as N → ∞.
Furthermore, with a little more work,
C OROLLARY 4.4. For complex numbers α, β with ℜ β 6= 0, and for any m ≥ 0,
uniformly in N
Λ(α/N) m
EU (N ) ≪α,β,m 1.
Λ(β/N)
22
BRAD RODGERS
P ROOF. From H¨older’s inequality, if 2k is an even integer larger than m
Λ(α/N) m Λ(α/N) 2k m/2k
EU (N ) .
≤ EU (N ) Λ(β/N)
Λ(β/N)
Let A := α/N and B := β/N, and note that for a unitary matrix g,
det(1 − e−A g) 2k
det(1 − e−A g)k det(1 − e−A g −1 )k
=
det(1 − e−B g)
det(1 − e−B g)k det(1 − e−B g −1 )k
=
det(1 − e−A g)k det(1 − eA g)k
.
det(1 − e−B g)k det(1 − eB g)k
As long as A 6= B, the average of this quantity can be computed exactly and seen to
be uniformly bounded using Theorem 4.2. And if A = B the corollary is trivial. 4.2. We also have results that mirror Theorem 1.1 and Lemma 2.11 for the
linear statistics of (pulled-back) eigenvalues. In analogy with our discussion of zeta
zeros, for a matrix g ∈ U(N) with eigenangles {θi } as before, we use the notation
hη, Ei = hη, EN (g)i := lim
V →∞
hη, E o i := lim
V →∞
N X
V
X
i=1 ν=−V
Z
V
η N(θi + ν) ,
η(x) dx,
−V
e = hη, EeN (g)i := hη, Ei − hη, E o i,
hη, Ei
when these limits exist. Clearly if η decays quadratically the limits exist, for any
unitary matrix g. As before, we sometime substitute E or EN for EN (g).
For η = fˆ with f ∈ L1 (R) and of bounded variation, the integral defining hη, E oi
may be seen to converge to (f (0+) + f (0−))/2 (see [22, Th. 4.3.4] for instance).
Likewise, by the Poisson summation formula (see [24, Th. D.3] for instance) it may
be seen for such η that the sum defining hη, Ei converges also. Indeed, in this latter
case the Poisson summation formula tells us that
1 X
Tr(g j )F (−j/N),
(31)
hη, EN (g)i =
N j∈Z
where for typographical reasons we write F (x) := (f (x+) + f (x−))/2. Hence
also,
1 X
hη, EeN (g)i =
Tr(g j )F (−j/N).
(32)
N
j6=0
We prove, in analogy with Theorem 1.1,
TAIL BOUNDS FOR COUNTS OF ZEROS
23
T HEOREM 4.5. For Q defined as in Theorem 1.1, for all N ≥ 1 and x ≥ 2,
P hQ, EN i ≥ x ≪ e−Cx log x ,
where the constant C and the implicit constant are absolute.
Likewise, in analogy with Lemma 2.11,
L EMMA 4.6. For any ǫ ≥ 0 and k ≥ 1/ǫ2 , for all N ≥ 1,
P(|h(W (1/k) )ˆ, EN i| ≥ ǫ) ≪ ǫ2 .
Indeed, these results are proved in much the same way, except that we will replace
analytic number theory with a random matrix result of Diaconis and Shashahani
[12]5.
T HEOREM 4.7 (Diaconis-Shahshahani). Consider a = (a1 , ..., ak ) and b = (b1 , ..., bk )
P
P
with a1 , a2 , ..., b1 , b2 , ... ∈ N≥0 . If kj=1 jaj + kj=1 jbj ≤ 2N, then
EU (N )
k
Y
j aj
Tr(g )
j=1
Tr(g j )bj
dg = δab
k
Y
j aj aj !
(33)
j=1
As Diaconis and Shahshahani note, if C1 , C2 , ... are independent standard normal
law
complex variables (that is Cj = X + iY with X and Y independent and identically distributed NR (0, 1/2) variables), then the right hand side of (33) may also be
written
k
Y
p
p
E ( jCj )aj ( jCj )bj .
(34)
j=1
For convenience, by anology with Tr(g −j ) = Tr(g j ), we also define the random
variables C−j := Cj , so that small moments of the traces Tr(g j ) may be identified
with small moments of p
gaussians. (Though a caution: this identification between
j
moments of Tr(g ) and |j|Cj holds only for small moments as in the theorem!)
We are now in a position to prove an analogue of Lemma 2.3.
L EMMA 4.8. For an absolute constant B2′ , uniformly for N ≥ 1 and 2ℓ ≤ k, we
have
E|hGk , EeN i|2ℓ ≤ (B2′ ℓ)ℓ .
5Though
note in this source there is a minor mistake in the statement of the result. This is
corrected in, for instance, [11].
24
BRAD RODGERS
P ROOF. From (32),
1 X
ˆ k (−j/N)
hGk , EeN (g)i =
Tr(g j )G
N j6=0
1 X
ˆ k (−j/N),
Tr(g j )G
=
N
|j|≤N/k
j6=0
ˆ k ⊂ [−1/k, 1/k], as in (10).
with the second line following because supp G
We have then
2ℓ
1 X p
2ℓ
e
ˆ
E|hGk , EN (g)i| = E
|j|Cj Gk (−j/N) ,
N
|j|≤N/k
j6=0
Q
Q
because one may see that any product Tr(g j )aj Tr(g j )bj that would occur in
P
ˆ k (−j/N)|2ℓ must have P jaj + P jbj ≤ 2ℓ · N/k ≤
the expansion of | Tr(g j )G
ˆ k is even, so
N, which is certainly less than 2N. Yet, recalling (6), we see that G
that
Xp
1 Xp
ˆ k (j/N)
ˆ k (−j/N) = 1
|j|Cj G
j(2ℜCj )G
N j6=0
N j>0
2 X
law
ˆ k (j/N)2 ,
= NR 0, 2
jG
N j≥0
with the last reduction because the random variables 2ℜCj are i.i.d real gaussians
with mean 0 and of variance 2.
Therefore
eN (g)i|2ℓ = (2ℓ − 1)!!
E|hGk , E
2 X
ℓ
2
ˆ
j Gk (j/N) ,
N 2 j≥0
ˆ k (x)| ≤
with (2ℓ − 1)!! := (2ℓ − 1) · (2ℓ − 3) · · · 3 · 1. From (10), we know |G
k(1 − |kx|)+ , so
2 X ˆ
k2
2
j
G
(j/N)
≪
k
N 2 j≥0
N2
X
0<j<N/k
j(1 − jk/N)2+ ≪ 1.
Using Stirling’s formula to bound (2ℓ−1)!! = (2ℓ)!/2ℓ ℓ!, we obtain the lemma. Likewise we have an analogue of Lemma 2.4.
TAIL BOUNDS FOR COUNTS OF ZEROS
25
L EMMA 4.9. For an absolute constant B3′ ,
hGk , E o i = B3′ k.
P ROOF. This is evident from the definition of hGk , E o i.
We now prove Theorem 4.5 in the same manner that we proved Theorem 1.1.
P ROOF
OF
T HEOREM 1.1. For even integers k, and all positive y,
′
k/2
e ≥ y) ≤ 1 E|hGk , Ei|
e K ≤ (B2 k) ,
P(hGk , Ei
yk
yk
yet
e + hGk , Ei ≥ x)
P(hQ, Ei ≥ x) ≤ P(hGk , Ei
e ≥ x − B ′ k).
= P(hGk , Ei
3
4B3′
With no loss of generality, we may assume x ≥
and take k to be the positive
even integer satisfying x/2B3′ − 2 ≤ k ≤ x/2B3′ . In particular, we have x − B3′ k ≥
x/2 and the theorem follows, as before by combining the two lines above.
Our proof of Lemma 4.6 is likewise parallel to that of Lemma 2.11.
P ROOF
OF
L EMMA 4.6. From the Poisson summation formula (31),
1 X
Tr(g j )W (1/k) (−j/N).
h(W (1/k) )ˆEN i =
N j6=0
(Note that (W (1/k) (0+) + W (1/k) (0−))/2 = 0. This enables us to dispense with the
j = 0 term of the summand.)
As |W (1/k) (x)| ≪ 1 for all x ∈ R and W (1/k) (x) = 0 for |x| ≥ 1/k, we see from
Theorem 4.7 of Diaconis and Shashahani, as long as k ≥ 2,
1 X
E|h(W (1/k) )ˆ, EN i|2 = 2
|j| · W (1/k) (−j/N)2
N j6=0
1 X
|j|
≪ 2
N
|j|≤N/k
1
.
k2
Now, as in the proof of Lemma 2.11, for ǫ > 1/2, trivially,
P |h(W (1/k) )ˆ, Ei| ≥ ǫ ≤ 1 ≪ ǫ2 .
≪
26
BRAD RODGERS
On the other hand, if ǫ ≤ 1/2, then the conditions of the lemma at hand force that
k ≥ 2, so that
1
P |h(W (1/k) )ˆ, Ei| ≥ ǫ ≤ 2 E |h(W (1/k) )ˆ, Ei|2
ǫ
1
≤ ǫ2 .
≪
2
(ǫk)
Remark: Our proof above has differed from that of Lemma 2.11 in one respect;
unlike with zeta zeros, there is no need to do any work bounding h(W (1/k) )ˆ, E oi.
In the random matrix setting, this quantity is always 0.
5. The average of ratios: a proof of Theorem 1.6
5.1.
sition.
We begin our proof of Theorem 1.6 by demonstrating the following propo-
P ROPOSITION 5.1. Assume the GUE Conjecture. Then for any continuous and
quadratically decaying function η : R → R,
lim E ehη,ZT i = lim E ehη,EN i ,
T →∞
N →∞
with both limits existing.
P ROOF. We note in the first place that the GUE Conjecture and the implication (28)
of the Weyl-Gaudin-Dyson integration formula imply for any non-negative integer
ℓ and continuous and quadratically decaying function η,
lim Ehη, ZT iℓ = lim Ehη, EN iℓ .
T →∞
N →∞
(35)
This is because both hη, ZT iℓ and hη, EN iℓ can respectively be written as a linear
combination of correlation sums,
log T
X log T
∆j (f1 , ..., fj ) :=
f1
(γ1 − t) · · · fj
(γj − t) ,
2π
2π
γ1 ,..,γk
distinct
and
Dj (f1 , ..., fj ) :=
X X
ν∈Z i1 ,...,ij
distinct
f1 (N(θi1 + ν)) · · · fj (N(θij + ν)),
and on the GUE Conjecture ∆j and Dj have the same average as T, N → ∞. For
instance,
hη, Zi = ∆1 (η),
TAIL BOUNDS FOR COUNTS OF ZEROS
27
hη, Zi2 = ∆1 (η 2 ) + ∆2 (η, η),
hη, Zi3 = ∆1 (η 3 ) + 3∆2 (η, η 2 ) + ∆3 (η, η, η),
and so on, and likewise for hη, Ei.
Now, we note that for x ≥ 0 and arbitrary k ≥ 0,
x
0≤e −
as
ex −
k
X
xℓ
ℓ=0
ℓ!
=
k
X
xℓ
ℓ=0
ℓ!
≤
xk+1 x
e ,
(k + 1)!
(36)
∞
∞
X
xℓ
xk+1 X xj
≤
,
ℓ!
(k
+
1)!
j!
j=0
ℓ=k+1
1
1
1
with the inequality following from the relation (k+1+j)!
≤ (k+1)!
. Hence,
j!
k
X
1
hη, Ziℓ k+1 hη,Zi
E ehη,Zi −
≤ (k + 1)! E hη, Zi e
ℓ!
ℓ=0
∞
X
1
≤
(r + 1)k+1 er P hη, Zi ∈ [r, r + 1)
(k + 1)! r=0
(37)
Now, for r ≥ 0, by Theorem 1.1,
P hη, Zi ∈ [r, r + 1) ≪ e−Cr log(r+2) ,
where the constant C and the implicit constant depend on η. More trivially, from
the Taylor expansion of ex ,
(r + 1)k+1 ≤ k! (r + 1)er+1 .
Applying these estimates to (37),
k
∞
ℓ X
X
hη,
Z
i
T
hη,Z
i
E e T −
≪ 1
(r + 1)e2r+1 e−Cr log(r+2)
ℓ!
k + 1 r=0
ℓ=0
≪
1
,
k+1
uniformly in T .
By the same reasoning (replacing Theorem 1.1 with its random matrix analogue
Theorem 4.5),
k
ℓ X
hη,
E
i
N
hη,E
i
E e N −
≪ 1 .
ℓ!
k+1
ℓ=0
uniformly in N.
28
BRAD RODGERS
Hence, applying (35) to the above, we see that as T → ∞,
1 E ehη,ZT i = lim E ehη,EN i + O
+ o(1).
N →∞
k+1
As k may be chosen arbitrarily, the proposition follows.
Remark: This theorem is only a slight modification of a standard theorem in probability theory: that the distribution of a point process is controlled by its correlation
functions, provided the point process has rapidly decaying tails (c.f. [20, Lemma
4.2.6]). In our context, convergence in distribution translates to the claim that if F
is bounded and continuous, limT →∞ F (hη, ZT i) = limN →∞ F (hη, EN i). The fact
that ex is unbounded entailed additional difficulties over the usual proof.
5.2. We are finally in a position to use the GUE Conjecture to evaluate the
average of ratios of the zeta function.
P ROOF. Throughout this proof we take β, βℓ 6= 0, and regard m, and α, β, α1 , ..., αm ,
β1 , ..., βm to be fixed, with αi 6= βj for all i, j. By (25),
1
α
ζ 2 + log T + it
,
exp(hLα,β , ZT (t)i) = (1 + o(1)) ζ 21 + logβ T + it
uniformly for t ∈ [T, 2T ]. From this and the bound of powers of ratios, in Theorem
1.3, one sees that
Z 2T Y
m
m ζ 1 + αℓ + it
X
2
log T
1
dt = E exp(h
Lαℓ ,βℓ , ZT i) + o(1).
(38)
T T ℓ=1 ζ 1 + βℓ + it
ℓ=1
2
log T
We record the observation, also following from (25), that
ζ 1 + α + it 2 log T
.
exp(ℜ hLα,β , ZT (t)i) = (1 + o(1)) ζ 21 + logβ T + it
(39)
This implies, of course, that the left hand side of (39) has a uniformly bounded m-th
moments for fixed α, β, and m, with β 6= 0, by Theorem 1.3.
We define
(1/k)
(1/k)
Lα,β (ξ) := Lα,β (ξ) − i(α − β)(W (1/k) )ˆ(ξ).
Intuitively, Lα,β should be thought of as an approximation to the function Lα,β (ξ)1|ξ|≤k .
In particular, from (26) and Lemma 2.9 – which demonstrate that both Lα,β and
TAIL BOUNDS FOR COUNTS OF ZEROS
29
(W (1/k) )ˆ may be decomposed into a linear combination of the function J and a
function that decays quadratically – we see that
(1/k)
Lα,β (ξ) ≪k Q(ξ).
(40)
Because (W (1/k) )ˆis real valued, we have that for α, β ∈ R, with β 6= 0,
(1/k)
exp(ℜ hLα,β , ZT (t)i) = exp(ℜ hLα,β , ZT (t)i)
(41)
and so the left hand side of (41) also has a uniformly bounded m-th moments for
fixed α, β and m, with β 6= 0.
In the proof that follows we let ǫ > 0 be arbitrary but small, and choose k ≥ 1/ǫ2 .
Defining
m
X
A :=
(αℓ − βℓ ),
ℓ=1
and returning to (38), we have
E exp(h
m
X
ℓ=1
Lαℓ ,βℓ , Zi) = E exp(h
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ + iA(W (1/k) )ˆ, Zi).
(42)
We split this average into two parts, writing
H≥ǫ := {t ∈ [T, 2T ] : |h(W (1/k) )ˆ, Zi ≥ ǫ},
H<ǫ := {t ∈ [T, 2T ] : |h(W (1/k) )ˆ, Zi < ǫ}.
Then (42) is equal to
E 1H≥ǫ · exp(h
|
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ + iA(W (1/k) )ˆ, Zi)
{z
}
:=M
+ E 1H<ǫ · exp(h
|
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ + iA(W (1/k) )ˆ, Zi) .
{z
}
:=N
For sufficiently large T (depending on ǫ), by Cauchy-Schwarz,
v
u
m
q
X
u
(1/k)
t
|M| ≤ P(H≥ǫ ) E exp(2ℜ h
Lαℓ ,βℓ , Zi)
ℓ=1
≪ ǫ,
with the last line following from Lemma 2.11 to bound P(H≥ǫ ) and (41) to bound
the other term.
30
BRAD RODGERS
On the other hand,
N = E 1H<ǫ exp(h
= E 1H<ǫ exp(h
m
X
ℓ=1
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , Zi + O(ǫ))
(1/k)
Lαℓ ,βℓ , Zi) + O ǫ · E 1H<ǫ exp(ℜ h
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , Zi),
as for small ǫ, we have eO(ǫ) = 1 + O(ǫ). Using (41), we see that
E 1H<ǫ exp(ℜ h
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , Zi) ≪ 1,
so that
N = E 1H<ǫ exp(h
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , Zi)
+ O(ǫ)
m
m
X
X
(1/k)
(1/k)
Lαℓ ,βℓ , Zi) − E 1H≥ǫ exp(h
Lαℓ ,βℓ , Zi) + O(ǫ).
= E exp(h
ℓ=1
ℓ=1
And as before, for sufficiently large T , by Cauchy-Schwarz,6
v
u
m
m
q
X
X
u
(1/k)
(1/k)
E 1H≥ǫ exp(h
Lαℓ ,βℓ , Zi) ≤ P(H≥ǫ )tE exp(2ℜ h
Lαℓ ,βℓ , Zi)
ℓ=1
ℓ=1
≪ ǫ.
Putting everything together, we have that
E exp(h
m
X
ℓ=1
Lαℓ ,βℓ , ZT i) = E exp(h
uniformly for sufficiently large T .
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , ZT i) + O(ǫ),
(43)
In exactly the same manner, this argument may be repeated for eigenvalues of the
unitary group, using the results of section 4. We see that
EU (N )
m
Y
Λ(αℓ /N)
ℓ=1
6
Λ(βℓ /N)
= E exp(h
m
X
ℓ=1
Lαℓ ,βℓ , EN i) + o(1),
(44)
Note that it is really only in the inequalities that follow that we have exploited the assumption
Pm
(1/k)
that α, β are real. It is from this assumption that we can easily bound E exp(h ℓ=1 ℜ Lαℓ ,βℓ , Zi)
uniformly in k, by using that fact that ℜiA(W (1/k) )ˆ = 0.
TAIL BOUNDS FOR COUNTS OF ZEROS
31
in analogy to (38), and
E exp(h
m
X
ℓ=1
Lαℓ ,βℓ , EN i) = E exp(h
uniformly for all N, in analogy with (43).
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , EN i) + O(ǫ),
(45)
Using (38) and (43), we see that
Z 2T Y
m ζ 1 + αℓ + it
m
X
2
log T
1
(1/k)
dt = E exp(h
Lαℓ ,βℓ , ZT i) + O(ǫ) + o(1),
T T ℓ=1 ζ 1 + βℓ + it
ℓ=1
2
log T
as T → ∞. Likewise, passing from (44) to (45),
EU (N )
m
Y
Λ(αℓ /N)
ℓ=1
as N → ∞.
Λ(βℓ /N)
= E exp(h
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , EN i) + O(ǫ) + o(1),
Proposition 5.1 implies that the main terms of the right hand sides of these identities
are asymptotically equal:
lim E exp(h
T →∞
ℓ=1
Hence,
1
T
Z
T
m
2T Y
ℓ=1
m
X
ζ
ζ
1
2
+
αℓ
log T
1
2
+
βℓ
log T
(1/k)
Lαℓ ,βℓ , ZT i)
= lim E exp(h
N →∞
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , EN i).
m
+ it
Y
Λ(αℓ /N)
dt = lim EU (N )
+ O(ǫ) + o(1).
N →∞
Λ(βℓ /N)
+ it
ℓ=1
Because ǫ is arbitrary, our theorem now follows from the evaluation in Corollary
4.3.
5.3. We have said that similar methods may be used to show that the GUE
Conjecture implies not only the Local Ratio Conjecture with real translations, but
in fact the Local Ratio Conjecture in general. We conclude by giving a very brief
sketch of how this may be done. We note that in the above argument, the only place
we have used the assumption that α1 , .., αm , β1 , ..., βm are real is in exploiting the
fact that then ℜiA(W (1/k) )ˆ = 0. We do note really need for this term to be 0
though; we need only for its exponential moments to be uniformly bounded in k.
That is, if one shows that uniformly for large k,
P(|h(W (1/k) )ˆ, Zi| ≥ x) ≪ e−Cx log x ,
(46)
32
BRAD RODGERS
this is enough to bound the terms
E exp(2ℜ h
m
X
ℓ=1
(1/k)
Lαℓ ,βℓ , Zi)
uniformly in k, and the proof proceeds as before. (46) in turn may be proven in
much the same way as Theorem 1.1, Theorem 2.6 and Corollary 2.7.
We note the converse implication, that the Local Ratios Conjecture implies the GUE
Conjecture, may be derived from the combinatorial work of Conrey and Snaith [10,
Th. 8], along with a uniform bound like Theorem 1.3. Indeed, using a Tauberian
argument, it should be possible to show that just the Local Ratio Conjecture with
real translations also implies the GUE Conjecture, but we do not treat the matter
here.
References
[1] Andrews, G.E., Askey, R., and Roy, R. Special Functions. Vol. 71 Encyclopedia of Mathematics
and its Applications. Cambridge University Press, 1999.
[2] Borodin, A. Olshanki, G., and Strahov, E. ‘Giambelli compatible point processes.’ Advances in
Applied Mathematics. 37.2 (2006): 209-248.
[3] Bump, D. Lie groups. Vol. 225 Graduate Texts in Mathematics. Springer, 2004.
[4] Bump, D., and Gamburd, A. ‘On the averages of characteristic polynomials from classical
groups.’ Communications in mathematical physics 265.1 (2006): 227-274.
[5] Chhaibi, R., Najnudel, J., and Nikeghbali, A. ‘The Circular Unitary Ensemble and the Riemann
Zeta Function: the Microscopic Landscape.’ Preprint.
[6] Conrey, J. B., Farmer, D. W., and Zirnbauer, M. R. ‘Autocorrelation of ratios of L-functions.’
Comm. Number Theory and Physics 2.3 (2008): 593-636.
[7] Conrey, J. B., Farmer, D. W., and Zirnbauer, M. R. ‘Howe pairs, supersymmetry, and ratios of
random characteristic polynomials for the unitary groups U (N).’ Preprint.
[8] Conrey, J. B., Forrester, P.J., and Snaith, N.C. ‘Averages of ratios of characteristic polynomials
for the compact classical groups.’ Int. Math. Res. Not. IMRN (2005): 397-431.
[9] Conrey, J. B., and Snaith, N.C. ”Applications of the L-functions ratios conjectures.” Proceedings
of the London Mathematical Society 94.3 (2007): 594-646.
[10] Conrey, J.B., and Snaith, N.C. ‘Correlations of eigenvalues and Riemann zeros’, Comm. Number Theory and Physics 2.3 (2008): 477-536.
[11] Diaconis, P., and Evans, S. ‘Linear functionals of eigenvalues of random matrices.’ Trans.
Amer. Math. Soc. 353 (2001): 2615-2633.
[12] Diaconis, P., and Shahshahani, M. ‘On the eigenvalues of random matrices.’ Journal of Applied
Probability 31A (1994): 49-62.
[13] Farmer, D.W. ‘Long mollifiers of the Riemann zeta-function.’ Mathematika 40.01 (1993): 7187.
[14] Farmer, D.W. ‘Mean values of ζ ′ /ζ and the GUE hypothesis’ Int. Math. Res. Not. (1995):
71-82.
[15] Farmer, D. W., Gonek, S. M., Lee, Y., and Lester, S. J. ‘Mean values of ζ ′ /ζ(s), correlations
of zeros and the distribution of almost primes.’ Quart. J. of Math. 64 (2014): 1057 - 89.
TAIL BOUNDS FOR COUNTS OF ZEROS
33
[16] Fujii, A. ‘Explicit formulas and oscillations,’ in Emerging Applications of Number Theory, ed.
D. Hejhal, J. Friedman, M. Gutzwiller, A. Odlyzko (Springer, 1999), pp. 219–267.
[17] Goldston, D. A., Gonek, S.M. and Montgomery, H.L. ‘Mean values of the logarithmic derivative of the Riemann zeta-function with applications to primes in short intervals.’ J. reine angewa.
Math (2001): 105 - 126.
[18] Guinand, A. P. ‘A summation formula in the theory of prime numbers.’ Proc. London Math.
Soc 2.50 (1948): 107 - 119.
[19] Harper, A. J. ‘Sharp conditional bounds for moments of the Riemann zeta function.’ Preprint.
[20] Hough J.B., Krishnapur M., Peres Y., and Vir´ag B., Zeros of Gaussian analytic functions and
determinantal point processes. Vol. 51. Amer Mathematical Society, 2009.
[21] Iwaniec, H., and Emmanuel, K. Analytic number theory. Vol. 53. American Mathematical
Society, 2004.
[22] Kawata, T. Fourier Analysis in Probability Theory. Academic Press, 1972.
[23] Kieburg, M., and Guhr, T. ‘Derivation of determinantal structures for random matrix ensembles
in a new way.’ J. Phys. A: Math. Theor. 43 (2010): 31pp.
[24] Montgomery, H. L., and Vaughan, R.C. Multiplicative number theory I: Classical theory. Vol.
97. Cambridge University Press, 2006.
[25] Radziwiłł, M. ‘The 4.36th Moment of the Riemann Zeta-Function.’ Int. Math. Res. Not. IMRN
(2012): 4245 - 4259.
[26] Riemann, B. ‘Ueber die Anzahl der Primzahlen unter einer gegebenen Grosse.’ Ges. Math.
Werke und Wissenschaftlicher Nachlass 2 (1859): 145 - 155.
[27] Rodgers, B. ‘Arithmetic consequences of the GUE Conjecture for zeta zeros.’ Preprint.
[28] Selberg, A. ‘On the remainder in the formula for N (T ), the number of zeroes of ζ(s) in the
strip 0 < t < T .’ Avh. Norske Vid. Akad. Oslo. I. 1 (1944), 27pp.
[29] A. Selberg, Contributions to the theory of the Riemann zeta-function, Arch. Mat. Naturvid.
48.5 (1946) 89–155.
[30] Soundararajan, K. ‘Moments of the Riemann zeta function.’ Annals of Math. 170 (2009): 981–
993.
[31] Titchmarsh, E. C., and Heath-Brown, D. R. The theory of the Riemann zeta-function. Oxford
University Press, 1986.
[32] Weil, A. ‘Sur les “formules explicites” de la theorie des nombres premiers’, Comm. Sem. Math.
Univ. Lund (1952): 252 - 265.
I NSTITUT
¨
Z URICH
¨
F UR
¨ Z URICH
¨
M ATHEMATIK , U NIVERSIT AT
, W INTERTHURERSTR . 190, CH-8057
E-mail address: brad.rodgers@math.uzh.ch