Professional Documents
Culture Documents
WINSTON HEAP
arXiv:2102.02092v1 [math.NT] 3 Feb 2021
Abstract. We show that the splitting conjecture in the hybrid model of Gonek–
Hughes–Keating holds to order on the Riemann hypothesis. Our results are valid
in a larger range of the parameter X which mediates between the partial Euler and
Hadamard products. We also show that the asymptotic splitting conjecture holds
for this larger range of X in the cases of the second and fourth moments.
1. Introduction
The moments of the Riemann zeta function have been the subject of several con-
jectural methods in recent years. Since the second and fourth moments of Hardy–
Littlewood [23] and Ingham [33], it is only relatively recently that a full conjecture
for all moments was given. This began with the work of Keating–Snaith [34] who
used the now famous connection with random matrix theory to conjecture that for
real k > −1/2,
1 2T
Z
2
|ζ( 12 + it)|2k dt ∼ a(k)g(k)(log T )k
T T
where
k2 X
Y 1 dk (pm )2
(1) a(k) = 1−
p
p m>0
pm
and
G(k + 1)2
(2) g(k) =
G(2k + 1)
where G is the Barnes G-function. This was preceded by conjectures for the 6th and
8th moments due to Conrey–Ghosh [14] and Conrey–Gonek [15], respectively, using
number theoretic methods. The Keating–Snaith conjecture has since been derived
with various different approaches [13, 18, 22].
A drawback of Keating and Snaith’s method was that the arithmetic factor a(k)
had to be incorporated in an ad-hoc way since there was no input from primes in
their random matrix theory model. This was remedied in the method of Gonek–
Hughes–Keating (G–H–K) [22] which forms the main focus of this paper.
1
THE SPLITTING CONJECTURE 2
The first step of G–H–K’s method was to express the zeta function as the product
of partial Euler and Hadamard products. Precisely, Theorem 1 of [22] states that for
2 6 X 6 t1/3 and large t,
1
1 1 1
(3) ζ( 2 + it) = PX ( 2 + it)ZX ( 2 + it) 1 + O
log X
where
X
Λ(n) X
PX (s) = exp , ZX (s) = exp − U((s − ρ) log X)
n6X
ns log n ρ
and Z ∞
U(z) = u(x)E1 (z log x)dx
0
R∞
with E1 (z) = z e−w dw/w and u(x) a smooth, non-negative function of mass 1 with
support in [e1−1/X , e]. To give a rough idea of these objects, note that from the
support conditions on u we have U(z) ≈ E1 (z). This has mass concentrated in the
region z ≪ 1 whereQwe have the approximation E1 (z) ≈ −γ − log z. Thus, roughly
speaking, ZX (s) ≈ |ℑ(s)−ℑ(ρ)|≪1/ log X ((s − ρ)eγ log X). Also, from the definition of
the von Mangoldt function and the Taylor series for the logarithm we find PX (s) ≈
−s −1
Q
p6X (1 − p ) . Therefore, we can indeed view PX (s) and ZX (s) as partial Euler
and Hadamard products.
G–H–K then proceeded to compute the moments of the Euler product, showing
that for X ≪ (log T )2−ǫ ,
1 2T
Z
2
(4) |PX ( 21 + it)|2k dt ∼ a(k)(eγ log X)k , k ∈ R.
T T
They conjectured with random matrix theory that
1 2T log T k2
Z
1 2k
(5) |ZX ( 2 + it)| dt ∼ g(k) γ , k > −1/2
T T e log X
and then proved this in the cases k = 1, 2 for X ≪ (log T )2−ǫ . In order to recover
the Keating–Snaith conjecture they assumed that the moments of the product of PX
and ZX should split as the product of moments.
Conjecture 1 (Splitting conjecture, [22]). Let X, T → ∞ with X ≪ (log T )2−ǫ .
Then for fixed k > −1/2 we have
1 2T 1 2T 1 2T
Z Z Z
2k 2k
1 1
|PX ( 2 + it)ZX ( 2 + it)| dt ∼ 1
|PX ( 2 + it)| dt · |ZX ( 21 + it)|2k dt.
T T T T T T
THE SPLITTING CONJECTURE 3
Their reasoning behind this conjecture was that since PX and ZX oscillate at
different scales (1/ log X vs. 1/ log T ), their contributions should act independently
and hence the moment should split to leading order. They verified this in the cases
k = 1, 2 for X ≪ (log T )2−ǫ . The methodology of the hybrid model has since been
used in various different settings to acquire conjectures for all sorts of L-functions
[1, 8, 9, 10, 11, 19, 25]. In all cases, an equivalent version of the splitting conjecture
plays a key role.
In this paper we prove that the splitting conjecture holds to order on the Riemann
hypothesis (RH). Furthermore, we can extend the range of X past (log T )2−ǫ .
Theorem 1. Assume RH.
θk −ǫ
p Let ǫ, k > 0 be fixed and suppose X, T → ∞ with X 6
(log T ) where θk = 2 1 + 1/2|k|. Then
Z 2T
1 1 2T 1 2T
Z Z
2k 2k
1 1
|PX ( 2 + it)ZX ( 2 + it)| dt ≍ 1
|PX ( 2 + it)| dt · |ZX ( 21 + it)|2k dt.
T T T T T T
As mentioned, this holds in a range of X larger than originally conjectured. We can
also extend the range of X in the asymptotic results (4) and (5), both unconditionally
and on RH. This gives the following.
Theorem 2. The Splitting conjecture holds for k = 1, 2 in the range
X6 1
104
(log T )2 (log2 T )2 .
Assuming RH, we may take
( √
(log T )√6−ǫ when k = 1,
X6
(log T ) 5−ǫ when k = 2.
Our proofs utilise the recent developments in the theory of moments of L-functions
due to Soundararajan [43], Harper [24] and Radziwill–Soundararajan [38]. These
techniques were originally geared for upper bounds although they can be brought to
bear on lower bounds too [27]. We highlight three main ideas.
The first is an innovation of Soundararajan [43]. This was to note that log |ζ( 21 +it)|
can be bounded from above by a sum over primes alone since the zeros contribute
negatively to this quantity (see Lemma 6 below and c.f. formula (3)). With this,
|ζ( 12 + it)| can be bounded from above by an Euler product of flexible length.
The second idea can be found in a paper of Radziwill [37] and features heavily in
the later works of Harper [24] and Radziwill–Soundararajan [38]. It allows one to
compute moments of Euler products provided one can restrict to a certain subset of
[T, 2T ]. For the purposes of this discussion we consider the example
X
−1/2−it
exp p
p6Y
THE SPLITTING CONJECTURE 4
2
with Y = T 1/(log log T ) . On the face of it, this is a very longPDirichlet polynomial.
However, if we can restrict t to a subset of [T, 2T ] on which | p6Y p−1/2−it | 6 V for
a given V , then we can truncate the exponential series effectively using the fact that
10V
z
X zj
(6) e ∼
j=0
j!
for |z| 6 V P and large V . The choice of V is naturally dictated by the variance:
setting V = p6Y p−1 ∼ log log T we get a Dirichlet polynomial of length Y 10V =
T 10/ log log T . This is now short and so the mean square is easily computed. Also, the
exceptional set in this case is of small measure.
The final main input in the arguments of Harper and Radziwill–Soundararajan
allows one to push the length of the prime sum up to Y = T θ , for some fixed θ > 0.
This involves breaking the sum it into subsums of progressively smaller variance. A
similar splitting has appeared in the work of Brun on the pure sieve (see Hooley’s
refinement [29]).
This circle of ideas has been used in a wide variety of different contexts recently.
These include; short interval maxima of the Riemann zeta function [2, 3, 4], uncondi-
tional bounds for the moments of zeta and L-functions [21, 26, 27], value distribution
of L-functions [16, 30, 39], sign changes in Fourier coefficients of modular forms [35],
non-vanishing of central values of L-functions [17] and equidistribution of lattice
points on the sphere [32]. In our case, we use these ideas to prove the following.
2T
1
Z
2
|PX ( 12 + it)|2k dt ∼ a(k)(eγ log X)k
T T
a(k) is given by (1). Assuming RH, this holds for X 6 (log T )θk −ǫ with
where p
θk = 2 1 + 1/2|k|.
1
Proposition 2. Suppose X 6 104
(log T )2 (log2 T )2 . Then for k = 1, 2 we have
2T k2
1 log T
Z
2k
|ZX ( 21 + it)| dt ∼ g(k) γ
T T e log X
√
6−ǫ
where g(k) is given
√
by (2). Assuming RH we may take X 6 (log T ) when k = 1
5−ǫ
and X 6 (log T ) when k = 2.
THE SPLITTING CONJECTURE 5
Proposition 3. Assume RH and let ǫ, k > 0 be fixed. Suppose X 6 (log T )θk −ǫ with
θk as above. Then
k 2
1 2T
log T
Z
2k
|ZX ( 21 + it)| dt ≍ .
T T log X
Remark. The lower bound in Proposition 3 can be made unconditional provided
X 6 ηk (log T )2 (log2 T )2 . We say more on this in section 7.
R 2T 2 2
Since T |ζ( 12 + it)|2k dt is ≪ T (log T )k on RH [24] and ≫ T (log T )k uncondi-
tionally [27], Theorem 1 follows from Propositions 1 and 3 when combined with (3).
Likewise, Theorem 2 follows on combining Propositions 1 and 2 and (3).
Using the case of PX as an example, we describe how the range of X can be
increased past (log T )2−ǫ . First of all, note that since
X Λ(n) 2X 1/2
(7) 6 (1 + o(1)) ,
n1/2+it log n log X
n6X
1/2
we can approximate PX (1/2+it)k with a Dirichlet polynomial of length X 20|k|X / log X
by using (6) to truncate the exponential. If X ≪ (log T )2−ǫ then this is T o(1) and
so we have a short Dirichlet polynomial. Note this holds for all t ∈ [T, 2T ] since the
bound (7) is pointwise. G–H–K computed a Dirichlet polynomial approximation in
a slightly different way, although in order for it to be short they required the same
bound on X, perhaps unsurprisingly.
If X is larger, then in order to have a short Dirichlet polynomial we must restrict
to a subset of [T, 2T ] and in this case we need good bounds on the exceptional set.
Typically, one would expect Gaussian bounds of the shape
1 n X Λ(n) o V2
(8) µ t ∈ [T, 2T ] : ℜ >V ≪ exp −
T n6X
n1/2+it log n log log T
p
in a wide range of V . In practice we are limited to V ≪ (log T )(log2 T )/ log X
which may be much smaller than the maximum 2X 1/2 / log X. For the remaining
range of V one must settle for weaker bounds. For example, in [43] it is shown that
the tails of log |ζ(1/2 + it)| can be bounded by e−V log V when V ≫ log2 T log3 T . We
can show that the tails of our sum satisfy the same bound in the range log2 T log3 T 6
V 6 2X 1/2 / log X provided X ≪ (log T )2 . However, for our purposes the weaker
bound of e−AV with large A is sufficient and this affords us slightly more room in
the size of X.
Another avenue for improvement is to reduce the trivial bound in (7). This be-
comes a manageable task under RH and thus we are able to make further gains in
the size of X under this assumption. We shall prove the following.
THE SPLITTING CONJECTURE 6
1In fact, anything of the form e−AV with large A would be sufficient.
THE SPLITTING CONJECTURE 7
where
log3 T
X 1/2 (log X + log X
) log T 1 X C/log2 T
E(X, Y, T ) ≪ + + 1X>(log T )A
Y log2 T Y 1/2 Y
and A is a large constant.
Proof. By Perron’s formula (Lemma 3.19, [44]) we have
Z 1/2+1/ log X+iY X 1/2 log X
X Λ(n) 1 1 z dz
1/2+it log n
= log ζ(z + 2
+ it)X + O .
n6X
n 2πi 1/2+1/ log X−iY z Y
We shift the contour to the line with real part ℜ(z) = 1/ log X. Restricting Y 6 T /2
we don’t encounter any poles. In the region 1/2 + C/ log log τ 6 σ 6 1, τ ≫ 1, we
have
(log τ )2−2σ log3 τ
log ζ(σ + iτ ) ≪ .
log2 τ
This follows from (14.14.5) of [44], the Phragmen–Lindelöf principle and the bound
log ζ(σ + iτ ) ≪ log3 τ , σ > 1 (the latter can be deduced from the proof of Theorem
14.8 of [44]). From Theorem 14.14 (B) of [44] we have
log τ 2
1
log ζ(σ + iτ ) ≪ log , 2
< σ 6 12 + C/ log2 τ.
log2 τ (σ − 1/2) log2 τ
Therefore, the horizontal contours contribute
(9)
log X
X C/log2 T log( log ) log T X 1/2 log3 T X 1/2 log3 T
2T
≪ 1X>(log T )A + 1/2
+ .
Y log2 T log X Y log2 T max(log( X 2 ), 1) Y log X
(log T )
since S(τ ) ≪ log τ / log2 τ . In the integral over the remaining region, the error term
of (12) contributes
log T log Y log T
≪ +
δY log X log2 T Y log X log2 T
after considering the regions δ < |Y − |t − y|| 6 1 and 1 < |Y − |t − y|| separately.
Therefore, on choosing δ = 1/(Y 1/2 log X) we find that the integral on the right of
(10) is
1 − X −i(t−y)
1 log T
Z
i S(y) 1|t−y|6Y + 1|t−y|>Y dy+O 1/2 .
t/2−Y 6y62t+Y i(t − y) i(t − y) Y log2 T
|t−y±Y |>δ
Integrating by parts along with the bound S1 (τ ) ≪ log τ /(log2 τ )2 , we find that the
second term in this integral is ≪ log T /(Y (log2 T )2 ) which can be absorbed into the
error term immediately above. The range of integration of the first term can be
extended to |t − y| 6 Y at the cost of an error ≪ log T /(Y log2 T ). Combining this
in (10) along with the error terms of (9) the result follows.
and define the set relating to the imaginary part Sℑ (V ) similarly. Let
S(V ) = Sℜ (V ) ∩ Sℑ (V )
and denote
V0 = log2 T log3 T, S = S(V0 ).
Define the complementary sets by
Eℜ (V ) = Sℜ (V ), Eℑ (V ) = Sℑ (V ), E(V ) = S(V ), E = S
where A = [T, 2T ]\A for a given set A. Also, let
X
Λ(n) X Λ(n)
Vmax = max ℜ , ℑ .
t∈[T,2T ] n1/2+it log n n1/2+it log n
n6X n6X
Note that unconditionally Vmax 6 (2+o(1))X 1/2 / log X and that on the Riemann hy-
pothesis Vmax 6 (1/2 + o(1))(log(X 1/2 / log T ) + O(log2 X)) log T / log2 T by Theorem
3. The reason we work with the real and imaginary parts (as opposed to working
with the modulus directly) is so that we have slightly better conditional bounds
for Vmax whilst still maintaining control over the modulus (which is important for
Lemma 5 below). This gives better exponents for our logarithms in the conditional
results but entails slightly more work.
To the estimate the measure of the complementary sets we use the following.
Lemma 2 ([43]). Let T be large and let 2 6 x 6 T . Let m be a natural number such
that xm 6 T / log T . Then for any complex numbers a(p) we have
2m m
1 2T X a(p) |a(p)|2
Z X
dt ≪ m! .
T T p6x p1/2+it p6x
p
1
Lemma 3. Let ǫ, κ > 0. Then for X 6 ηκ (log T )2 (log2 T )2 with ηκ = 16κ2
− ǫ we
have
µ(Eℜ (V )) ≪ T e−(2κ+o(1))V , V0 6 V 6 Vmax
where µ denotes Lebesgue measure. Assuming RH, the same bound holds provided
X 6 (log T )θκ −ǫ where
q
1
(13) θκ = 2 1 + 2κ .
The same results hold for Eℑ (V ) also.
Proof. We first prove the unconditional result. Write
X Λ(n) X 1 X 1
1/2+it
= 1/2+it
+ 1+2it
+ O(1).
n6X
n log n p6X
p √ 2p
p6 X
THE SPLITTING CONJECTURE 11
Then from Jensen’s inequality in the form (a + b + c)2m 6 32m−1 (a2m + b2m + c2m )
with m > 1, we have
Z 2T X Z 2T X
32m−1
1 2m 1 2m 2m
µ(Eℜ (V )) 6 2m dt + dt + O(T C ) .
V p 1/2+it 1+2it
√ 2p
T p6X T
p6 X
p6X
p
and hence we may concentrate on the integral of |D(t, k)|2 over the full set [T, 2T ].
Applying the Montgomery–Vaughan mean value theorem again gives
1 2T X αk (n)2
Z
|D(t, k)|2dt = (1 + o(1)) .
T T n
n∈S(X)
Since |αk (n)| 6 d|k|(n), the sum over terms with Ω(n) > W0 is, for any 1 < r < 2,
X αk (n)2 X d|k| (n)2 r Ω(n) 2
(25) ≪ r −W0 ≪ r −W0 (log X)rk = o(1)
n n
n∈S(X) n∈S(X)
Ω(n)>W0
where in the first inequality we have applied Rankin’s trick in the form r Ω(n)−W0 > 1.
Since αk (n) = βk (n) if Ω(n) 6 W0 the main term is
X αk (n)2 X βk (n)2 X
|βk (n)|2
= +O .
n n n
n∈S(X) n∈S(X) n∈S(X)
Ω(n)6W0 Ω(n)>W0
THE SPLITTING CONJECTURE 15
The bound |βk (n)| 6 d|k|(n) and the same analysis as in (25) shows that this error
term is o(1). From the properties of βk (n) we get
X βk (n)2 Y X dk (pm )2 X βk (pm )2
= m
+
n p6X m
m: p 6X
p m: pm >X
pm
n∈S(X)
Y X dk (pm )2 Y X d (pm )2
|k|
= m
1+O m
.
p6X m>0
p p6X m
m: p >X
p
√
We split the second product at X and apply the bound dk (n) ≪ nǫ to find that it
is
Y 1 Y 1
(26) 1+O 1−ǫ
1 + O 2−ǫ
= 1 + O(X −1/2+ǫ )
√ X √ p
p6 X X<p6X
m
∼ a(k)(eγ log X)k
p6X m>0
p
and thus it suffices to consider the second and fourth moment of the object on the
right. Our aim is to first replace PX by its Dirichlet polynomial approximation and
then apply formulas for the twisted second and fourth moments of the zeta function.
4.1. The second moment. As before, we decompose the integral as
1 2T 1 1
Z Z Z
2 2
(27) 1
|ZX ( 2 + it)| dt = 1
|ZX ( 2 + it)| dt + |ZX ( 12 + it)|2 dt.
T T T S T E
Working unconditionally first, we apply the Cauchy–Schwarz inequality to the inte-
gral over E to find it is
Z 1/2
2 1 1 −1 4
≪ (log T ) |PX ( 2 + it) | dt
T E
using Ingham’s asymptotic for the fourth moment. Since X 6 1014 (log T )2 (log2 T )2
and 1/104 < η2 we find that this is ≪ (log T )2 e−δV0 = o(1) by Lemma 4.
THE SPLITTING CONJECTURE 16
for some constant B. On applying the bound α−1 (n) ≪ d1 (n) ≪ 1 and following the
argument in the proof of Theorem 3, pg. 530, of [22] we easily see that
X α−1 (m)α−1 (n)(m, n) B(m, n)2
log ≪ (log X)10
mn mn
m,n∈S(X)
for any 1 < r < 2. A short computation shows this last sum is
Y
1 + (2r + 1)p−1 + O(p−2) ≪ (log X)2r+1
p6X
and so the terms with Ω(m), Ω(n) > W0 contribute an error of size o(1).
In the main term we replace α−1 (n) with β−1 (n) and then re-extend the sum to
include those integers for which Ω(m), Ω(n) > W0 . By the bounds β−1 (n) ≪ d1 (n) ≪
1, the same argument shows that this introduces an error of o(1). Thus,
X β−1 (m)β−1 (n)(m, n)
I1 = log T + O((log X)10 ).
mn
m,n∈S(X)
To
P estimate the −m second product we note that the sum in the error term
−⌈log X/ log p⌉
√ is ≪
pm >X (m + 1)p ≪ (log X)p and then split the product at X, as
−1/2+ǫ
before. In this way we find it is 1 + O(X ) and therefore by Mertens’ Theorem
log T
I1 ∼
eγlog X
as desired.
4.2. The fourth moment: Initial clearing. Not surprisingly, the fourth moment
requires more work in both the initial stages and the arithmetic computations. Our
aim is to show that
4
1 2T
1 log T
Z
1 4 1 −4
I2 := |ζ( 2 + it)| |PX ( 2 + it)| dt ∼ .
T T 12 eγ log X
In this subsection the goal is to replace PX (1/2 + it)−2 by D(t, −2).
THE SPLITTING CONJECTURE 18
Splitting the integral as in (27) we see that our first task is to bound
1
Z
(29) |ζ( 12 + it)|4 |PX ( 21 + it)|−4 dt.
T E
On RH we can deal with this by applying Hölder’s
√ inequality as in (28). Following
the same argument and using the fact that 5 − ǫ′ 6 θ2(1+ǫ) − ǫ′ /2 shows that this
is o(1).
To bound this unconditionally requires more work. First note that since
′
n X Λ(n) o
E ⊂ E := t ∈ [T, 2T ] : > V0
n6X
n1/2+it log n
we can upper bound by the integral over E ′ . Let Vj = ej V0 and define J to be the
maximal j such that VJ 6 Vmax . Let
n X Λ(n) o
Ej = t ∈ [T, 2T ] : Vj 6 6 V
j+1
n1/2+it log n
n6X
so that
E ′ = ∪Jj=0 Ej .
Then
J
1 1
Z X Z
4 −4 4Vj+1
|ζ( 12 + it)| |PX ( 12 + it)| dt 6 e |ζ( 12 + it)|4 dt
T E′ j=0
T Ej
(30) J 2T
1 Λ(n)
Z X 2rj
−2r
X
4Vj+1
6 e Vj j |ζ( 12 + it)|4 dt
T n1/2+it log n
j=0 T n6X
for any given integer rj > 0. The combinatorics are simplified if we focus on the
prime sums so we apply Jensen’s inequality in the form
X
X Λ(n) 2rj
rj 1 2rj X 1 2rj 2rj
69 + + O(C ) .
n1/2+it log n p1/2+it √ 2p
1+2it
n6X p6X p6 X
It will be clear after the computations that the first sum here gives the dominant
contribution and so we focus on this. Note that by the multinomial theorem,
X 1 rj X g(n)
= r j !
p6X
p1/2+it rj
n1/2+it
n6X
Ω(n)=rj
THE SPLITTING CONJECTURE 19
where g(n) is the multiplicative function satisfying g(pα ) = 1/α!. Accordingly, (30)
is
J
1 2T
Z X g(n) 2
4Vj+1 rj −2rj
X
1 4
(31) ≪ e 9 Vj · |ζ( 2 + it)| rj ! 1/2+it
dt
j=0
T T rj n
n6X
Ω(n)=rj
The twisted fourth moment of the zeta function has been computed before [6, 31]
and has been applied in similar situations [26]. Provided X rj 6 T 1/4−ǫ i.e.
log T
(32) rj 6 (1/4 − ǫ) ,
log X
Proposition 4 and formula (8) of [26] (see section 6 there) give
1 2T rj !2 g(n)g(m)
Z X g(n) 2 X
1 4 4 rj
|ζ( 2 + it)| rj ! 1/2+it
dt ≪ (log T ) 9 .
T T rj n [n, m]
rj
n6X m,n6X
Ω(n)=rj Ω(m)=Ω(n)=rj
Following the arguments of [26] which lead to formula (9) there, we find that this is
X 1 rj X 1
1/2 9rj log log X
rj
(33) ≪ (log T )4 9rj rj ! exp ≪ (log T )5 rj .
p6X
p p6X
p e
We choose
12Vj 12Vmax 48X 1/2 24 log T
rj = 6 . 2
. ∀j
log Vj log Vmax (log X) 100 log X
since X 6 1014 (log T )2 (log2 T )2 . Clearly, this satisfies (32). Then applying (33) in
(31) we find that
1
Z
|ζ( 21 + it)|4 |PX ( 21 + it)|−4 dt
T E′
J
4Vj+1 C log log X
X 12Vj / log Vj
5
≪(log T ) e
j=0
Vj log Vj
J J
j V (12−4e+o(1))
X X
≪(log T )5 e4Vj+1 −(12−o(1))Vj ≪ (log T )5 e−e 0
= o(1).
j=0 j=0
on applying Lemma 5 in the integral over S. After extending to the full range of
integration [T, 2T ] it remains to estimate
1
Z
|ζ( 12 + it)|4 |D(t, −2)|2 dt.
T E
However, from the definition of D(t, k) we have
W0
X |k|j X Λ(n) j X Λ(n)
|D(t, k)| 6 6 exp k .
j! n1/2+it log n n1/2+it log n
j=0 n6X n6X
and so we can apply the same argument as above to acquire the bound o(1) for this
integral. Thus we have
I2 ∼ J 2
where
2T
1
Z
J2 := |ζ( 12 + it)|4 |D(t, −2)|2dt.
T T
4.3. The fourth moment: Arithmetic computations. In this section our aim
is to show that 4
1 log T
J2 ∼ .
12 eγ log X
The formulas for the twisted fourth moment of the zeta function given in the liter-
ature [6, 31] apply to smoothed integrals and accordingly we must first smooth J2 .
Let Φ− , Φ+ be smooth approximations of compact support satisfying
(34) Φ− (t) 6 1t∈[T,2T ] 6 Φ+ (t)
(j)
with derivatives Φ± (t) ≪ T ǫ . For example, we may take Φ− to be compactly
supported on [1, 2] and equal to 1 on the interval [1 + T −ǫ , 2 − T −ǫ ] with smooth,
monotonic decay to zero at each endpoint. Then, on letting Φ be either Φ− or Φ+
we consider the smoothed integral
1 t
Z
J2,Φ := Φ |ζ( 12 + it)|4 |D(t, −2)|2dt.
T R T
We note that the error incurred from these approximations will be ≪ T 1−ǫ which is
tolerable given the asymptotic we seek.
Theorem 4 (Theorem 1.2 of [6]). Let Φ be as above and α1 , α2 , α3 , α4 ≪ 1/ log T . Let
Ξ be the subgroup of S4 consisting of the identity, those permutations which swap just
one element of {1, 2} with {3, 4} and the permutation satisfying τ (1) = 3, τ (2) = 4.
THE SPLITTING CONJECTURE 21
Then for any Dirichlet polynomial n6y a(n)n−s satisfying y 6 T 1/4−ǫ and a(n) ≪ nǫ
P
we have
Z X a(m) 2 t
ζ( 12 + α1 + it)ζ( 21 + α2 + it)ζ( 21 − α3 − it)ζ( 12 − α4 − it) Φ dt
m 1/2+it T
R n6y
Z X P4 α −α
X t t j=1 τ (j) j
= a(m1 )a(m2 ) Φ Zτ (α1 ),τ (α2 ),τ (α3 ),τ (α4 ),m1 ,m2 dt
m ,m 6y R T τ ∈Ξ 2π
1 2
+ O(T 1−ǫ )
where
X 1 n n n n
1 2 3 4
Zα1 ,α2 ,α3 ,α4 ,m1 ,m2 = 1/2+α1 1/2+α2 1/2−α3 1/2−α4
V
m1 n1 n2 =m2 n3 n4
1/2
(m1 m2 ) n1 n2 n3 n4 t2
and
c+i∞
1 G(s)
Z
V (x) = (4π 2 x)−s ds, c>0
2πi c−i∞ s
with G(s) an even function of rapid decay in vertical strips satisfying G(0) = 1.
Remark. We remark that the choice of function G(s) is flexible and it can be
prescribed to have zeros at linear combinations of the shifts. This is fairly typical
and is used to cancel unnecessary poles later on. We will take G(s) = Qα (s) exp(s2 )
where Qα (s) is an even polynomial which is 1 at s = 0 and zero at 2s = α3 − α1 , α4 −
α1 , α3 − α2 and α4 − α2 . Note these conditions on Q imply that for fixed ℜ(s),
2
(35) G(s) ≪ (log T )4 e−ℑ(s)
since αj ≪ 1/ log T .
Let us compute term corresponding to the identity: τ = id. Denote this by
X
K = Kα (t, X) := α−2 (m1 )α−2 (m2 )Zα1 ,α2 ,α3 ,α4 ,m1 ,m2
m1 ,m2 ∈S(X)
X α−2 (m1 )α−2 (m2 ) n n n n
1 2 3 4
= 1/2+α1 1/2+α2 1/2−α3 1/2−α4
V .
m1 n1 n2 =m2 n3 n4 (m1 m2 )1/2 n1 n2 n3 n4 t2
mj ∈S(X)
By shifting the contour in V to the either the left or right depending on whether
x ≪ 1 or x ≫ 1, respectively, we find that V (x) ≪ (1 + |x|)−A for any A > 0.
Accordingly, on applying the bounds αk (m) ≪ mǫ · 1m6T δ we may restrict the above
sum to those nj satisfying n1 n2 n3 n4 ≪ t2+ǫ ≪ T 2+ǫ at the cost of an error of size
THE SPLITTING CONJECTURE 22
o(1). Then the contribution from those m1 with Ω(m1 ) > W0 is for 1 < r < 2
X r Ω(n) d4 (m)2
≪ r −W0 ≪ r −W0 (log T )16r = o(1)
2+ǫ
m
m6T
where for the first inequality we have applied Rankin’s trick in the form r Ω(m)−W0 > 1
along with the bound α−2 (n) ≪ d(n). The same bound holds for the sum over
Ω(m2 ) > W0 . Then on replacing α−2 (m) with β−2 (m) and re-extending the sums
(which by the same arguments incurs an error of o(1)) we have
X β−2 (m1 )β−2 (m2 ) n n n n
1 2 3 4
K= 1/2+α 1/2+α 1/2−α 1/2−α
V + o(1).
m1 n1 n2 =m2 n3 n4 (m1 m2 )
1/2 n
1
1
n2 2
n3 3
n4 4 t2
mj ∈S(X)
Unfolding the integral for V (x) and pushing the sum through we find
Z c+i∞ 2s
1 G(s) t
(36) K= Fα,X (s) ds + o(1)
2πi c−i∞ s 2π
where
X β−2 (m1 )β−2 (m2 )
Fα,X (s) = 1/2+α1 +s 1/2+α2 +s 1/2−α3 +s 1/2−α4 +s
1/2 n
m1 n1 n2 =m2 n3 n4 (m1 m2 ) 1 n2 n3 n4
mj ∈S(X)
X β−2 (m1 )β−2 (m2 )σα1 ,α2 (n1 )σ−α3 ,−α4 (n2 )
=
m1 n1 =m2 n2 (m1 m2 )1/2 (n1 n2 )1/2+s
mj ∈S(X)
d−u −v
P
with σu,v (n) = d1 d2 =n 1 d2 . Expressing this as an Euler product we have
Shifting the line of integration in (36) to ℜ(s) = −1/ log X we pick up a sim-
ple pole only at s = 0 (the poles of Aα (s) being cancelled by the zeros of G(s)).
Since β−2 (n) ≪ d(n) and σαi ,αj (pn ) ≪ pn/ log T d(pn ) we find that on the new line of
integration
Gα,X (s) ≪ (log X)O(1) .
Therefore, on combining this with the bound for G(s) given in (35) we see the integral
over the new line is bounded by
≪ t−2/ log X (log T )O(1) = o(1)
since t ≍ T . Hence
Kα = Aα (0)Gα,X (0) + o(1).
We have satisfactorily computed the contribution from a single Z term and thus it
remains to find the combinatorial sum of these which appears in Theorem 4. Using
the results of [13] we can express this sum as a multiple contour integral. Precisely,
Lemma 2.5.1 there gives
X t P4j=1 ατ (j) −αj
Kτ (α)
τ ∈Ξ
2π
1 Az1 ,z2 ,z3 ,z4 (0)Gz1 ,z2 ,z3 ,z4 ,X (0)∆(z1 , z2 , z3 , z4 )2
Z
= Q4
4(2πi)4 |zj |=3j / log T i,j=1 (zi − αj )
16j64
t P2j=1 (zj+2 −zj )/2
× dz + o(1)
2π
where ∆(z) denotes the vandermonde determinant. A short calculation shows that
∂ X log p
Gz,X (0) ≪ G0,X (0) ≪ G0,X (0) log X
∂zj z=0 p6X
p
From section 2.7 of [13] we know that c4 = g(2) = 1/12 where g(k) is given by (2).
Furthermore,
Y X β−2 (pm1 )β−2 (pm2 )d(pn1 )d(pn2 ) X d(pn )2
G0,X (0) = 1 / .
p6X m +n p 2 (m1 +m2 +n1 +n2 ) n>0
pn
1 1
=m2 +n2
a(p)p−s , suppose
P
Given a general Dirichlet polynomial of the form D(s) = p6Y
t ∈ [T, 2T ] is such that |kD(s)| 6 Z. Then by (17) we have
X (kD(s))j
kD(s)
(37) − 6 e−10Z .
e
06j610Z
j!
where we recall g is the multiplicative function such that g(pα ) = 1/α!, and a(n) is the
completely multiplicative extension of a(p). Observe this is a Dirichlet polynomial
of length Y 10Z .
Our remaining observations relate to mean values of Dirichlet polynomials. We
first state the mean value theorem of Montgomery and Vaughan [36] which gives for
any complex coefficients a(n),
2T
1 X a(n) 2
Z X
(39) dt = (1 + O(N/T )) |a(n)|2 .
T nit
T n6N n6N
where the R
Q
j=1 nj 6 N = o(T ) for all nj ∈ Sj i.e. the product of the Aj (s) is short.
Then the Montgomery–Vaughan mean value theorem readily implies
2T R
1
Z X 2
Aj (it)2 dt ∼
Y X
a1 (n1 ) · · · aR (nR )
T
T j=1 n6N n=n1 ···nR
nj ∈Sj
X
(40) = a1 (n1 ) · · · aR (nR )a1 (nR+1 ) · · · aR (n2R ).
n1 ···nR =nR+1 ···n2R
nj ∈Sj
Suppose in addition that for any j1 , j2 with j1 6= j2 the elements of Sj1 are all coprime
to the elements of Sj2 . Then there is at most one way to write n = R
Q
j=1 nj with
THE SPLITTING CONJECTURE 26
Also, write
X k Ω(n) wj (n)g(n)
Ni,j (t, k) =
n1/2+it
Ω(n)610ℓi
p|n =⇒ Ti−1 <p6Ti
and note that on the set of t ∈ [T, 2T ] such that |kPi,j (t)| 6 ℓi we have
2
exp(2kℜPi,j (t)) = (1 + O(e−9ℓi ))Ni,j (t, k)
(44)
by (37) and (38). Accordingly, if t is such that |kPi,j (t)| 6 ℓi for all 0 6 i 6 j then
on applying (43) we have
j
X wj (p)
Ni,j (t, k)2
Y
(45) exp 2kℜ 1/2+it
≪
p6T
p i=0
j
Pj
since i=0 e−9ℓi is a rapidly converging series. Note this is a Dirichlet polynomial of
length
J PJ
Y 1/4 J/4 /(log log T )1/2
(46) 6 Ti10ℓi = T 10 i=0 θj 6 T 20e 6 T 1/50 .
i=0
THE SPLITTING CONJECTURE 28
We can now state an upper bound for the zeta function in terms of these short
Dirichlet polynomials.
Lemma 7. Assume RH. Then either
|kP0,j (t)| > ℓ0
for some 0 6 j 6 J or
J
Y 2k |kP 2sj Yj
Ni,J (t, k)2 + j+1,l (t)| 2
X
|ζ( 12 +it)|2k
≪ exp Ni,j (t, k)
i=0 06j6J−1
θj ℓ j+1 i=0
j+16l6J
× |M(t, k)|2
for any positive integers sj where
X (k/2)Ω(n) g(n)
M(t, k) = .
n1+2it
Ω(n)610k(log2 T )2
p|n =⇒ p6log T
where
|kPi,l (t)| 6 ℓi ∀1 6 i 6 j, ∀j 6 l 6 J;
Sl (j) = t ∈ [T, 2T ] : .
|kPj+1,l (t)| > ℓj+1
We apply Lemma 6 to each zeta function on the right hand side of (47). If t ∈ Sl (j)
then we take x = Tj to give
X wj (p)
1 2k
X 1 2k
|ζ( 2 + it)| ≪ exp 2kℜ 1/2+it
+ 2kℜ 1+2it
+ .
p6T
p p6log T
2p θj
j
THE SPLITTING CONJECTURE 29
For the first sum overPprimes in the exponential we apply (45). For the second sum
1
we note that, since | p6log T p1+2it | 6 2 log3 T 6 (log log T )2 , we have
2
(k/2)Ω(n) g(n)
X 1 −9k(log2 T )2
X
exp 2kℜ 1+2it
= 1 + O(e )
p6log T
2p 2
n1+2it
Ω(n)610k(log log T )
p|n =⇒ p6log T
by (37) and (38). This is ≪ |M(t, k)|2. Finally, to capture the small size of the set,
we multiply by
2sj
|kPj+1,l (t)|
> 1.
ℓj+1
If t ∈ S(J) then we omit this last step since of course there is no such PJ+1,l (t).
6.3. Proof of the upper bound in Proposition 3. From (42) we are required to
show that k2
1 2T
log T
Z
1 2k 2
|ζ( 2 + it)| |D(t, −k)| dt ≪ .
T T log X
To apply Lemma 7 we must consider two cases; that where |kP0,j (t)| > ℓ0 and
otherwise. We consider the former case first since this is simpler.
So, let E ⊂ [T, 2T ] be the subset on which |kP0,j (t)| > ℓ0 , that is, when
(log log T )3/2
X w j (p) >
.
p1/2+it k
p6T 1/(log log T )2
provided m 6 (1 − o(1))(log log T )2 where in the last line we have used |wj (p)| 6 1
for all j. By Stirling’s formula and Mertens’ theorem this is
m
k2m
1/2 2
(48) µ(E) ≪ T m 2
6 T e−c(log log T )
e(log log T )
for some c > 1 on choosing m = ⌊min(1, k12 )(log log T )2 ⌋. Therefore by Hölder’s
inequality, Harper’s bound for the moments of zeta and (24) we have
1
Z
2
|ζ( 12 + it)|2k |D(t, −k)|2 dt ≪ e−C(log log T ) (log T )O(1) = o(1).
T E
THE SPLITTING CONJECTURE 30
We may now consider the second case where |kP0,j (t)| > ℓ0 and accordingly con-
centrate on the integral
1
Z
|ζ( 21 + it)|2k |D(t, −k)|2 dt.
T [T,2T ]\E
By the second part of Lemma 7 this is
J 2sj Yj
1 2T Y 2k |kP
j+1,l (t)|
Z
2 X 2
≪ Ni,J (t, k) +
exp Ni,j (t, k)
T T i=0 06j6J−1
θj ℓj+1 i=0
j+16l6J
(49)
× |M(t, k)|2|D(t, −k)|2 dt.
To compute the resultant integrals we apply the following lemma.
Lemma 8. For 0 6 sj 6 1/(10θj ) we have
j k 2
1 2T
log Tj
Z
2 2 2sj
Y 2 sj
|D(t, −k)| |M(t, k)| |Pj+1,l (t)| Ni,j (t, k) dt ≪ sj !Pj+1
T T i=0
log X
where
X 1
Pj+1 = .
Tj <p6Tj+1
p
Proof. We write the integrand as a multiple sum. First off, by the multinomial
theorem we have
s
wl (p) j
X
sj
X wl (n)g(n)
Pj+1,l (t) = 1/2+it
= s j ! .
T <p6T
p n1/2+it
j j+1 Ω(n)=sj
p|n =⇒ Tj <p6Tj+1
Therefore
j
Y
sj
D(t, −k)M(t, k)Pj+1,l (t) Ni,j (t, k)
i=0
X α−k (n) X (k/2)Ω(n) g(n) X wl (n)g(n)
=sj !
n1/2+it n1+2it n1/2+it
n∈S(X) Ω(n)610k(log2 T )2 Ω(n)=sj
p|n =⇒ p6log T p|n =⇒ Tj <p6Tj+1
j
k Ω(n) wj (n)g(n)
Y X
× .
i=0
n1/2+it
Ω(n)610ℓi
p|n =⇒ Ti−1 <p6Ti
THE SPLITTING CONJECTURE 31
Since X 6 T0 we may group together all the sums over T0 -smooth numbers as a
single sum and write the above as
j
k Ω(n) wj (n)g(n)
X γ(n)
X wl (n)g(n) Y X
sj !
n
n1/2+it n1/2+it i=1 n1/2+it
Ω(n)=sj Ω(n)610ℓi
p|n =⇒ Tj <p6Tj+1 p|n =⇒ Ti−1 <p6Ti
for some coefficients γ(n) where the product is empty if j = 0. Since this is a short
Dirichlet polynomial we find by (41) that
2T j
1
Z
Ni,j (t, k)2 dt
Y
2 2 2sj
(50) |D(t, −k)| |M(t, k)| |Pj+1,l (t)|
T T i=0
X γ(n)2 X wl (n)2 g(n)2
≪ (sj !)2
n
n n
Ω(n)=sj
p|n =⇒ Tj 6p<Tj+1
j
k Ω(n) wj (n)2 g(n)2
Y X
.
i=1
n
Ω(n)610ℓi
p|n =⇒ Ti−1 <p6Ti
where
W (n) = wj (n2 )wj (n5 )g(n2 )g(n3 )g(n5 )g(n6 )
and the ′ in the sum denotes that n1 , n3 ∈ S(X) and
Ω(n2 ), Ω(n5 ) 6 10ℓ0 Ω(n3 ), Ω(n6 ) 6 10k(log2 T )2
p|n2 , n5 =⇒ p < T0 p|n3 , n6 =⇒ p 6 log T.
We first estimate the terms with Ω(n1 ), Ω(n4 ) > W0 . By the usual arguments, for
1 < r < 2 these terms are bounded by,
−W0
X r Ω(n1 ) dk (n1 )dk (n4 )k Ω(n2 n3 n5 n6 ) W (n)
≪r
2
(n1 n2 n4 n5 )1/2 n3 n6
n1 n2 n3 =
n4 n5 n26
p|nj =⇒ p6T
Y (2r + 2)k 2
−W0 2
=r 1+ + O(p ) ≪ e−W0 (log T )6k = o(1).
−2
p6T
p
THE SPLITTING CONJECTURE 32
We then replace α−k (n) with β−k (n) in the remaining sum. The usual arguments
also allow us to remove the restrictions on all the Ω(nj ) at the cost of an error of size
o(1). Expressing the resultant sum as an Euler product we get
X γ(n)2 Y Y
k2
−2 −2
= 1 + O(p ) 1+ + O(p ) + o(1)
n
n p6X X<p6T0
p
k2
log T0
≍
log X
since β−k (n) is supported on X-smooth numbers where it satisfies β−k (p) = −k and
β−k (pm ) ≪ dk (pm ) for m > 2.
The second sum in (50) is
s
wl (n)2 g(n)2 1 j
X
2
X X sj !g(n)
(sj !) 6 sj ! = sj ! ,
n n T 6p<T
p
Ω(n)=sj Ω(n)=sj j j+1
p|n =⇒ Tj 6p<Tj+1 p|n =⇒ Tj 6p<Tj+1
On noting that
log Tj+1 −3/2
J − j ≪ log(1/θj ), Pj+1 = log + o(1) 6 2, ℓ2j+1 = θj
log Tj
THE SPLITTING CONJECTURE 33
for some constants c, C > 0. Since this last series is bounded the result follows.
where the Ti and ℓi are as before for 0 6 i 6 J. As remarked earlier, in this section
we take T−1 = A with A sufficiently large to be chosen later. We then form the
product
J
Y X γk (n)
(51) N (t, k) := Ni (t, k) =
i=0 n6Y
n1/2+it
for some coefficients γk (n) where from (46) we have Y = T 1/50 . We can think of
this as an approximation to ζ(1/2 + it)k ; It possesses several nice features akin to an
Euler product whilst also being a short Dirichlet polynomial.
We acquire our lower bound by applying Hölder’s inequality, the form of which
will depend on whether 0 < k 6 1 or k > 1. The latter case is somewhat simpler so
we give details for the case 0 < k 6 1 first. By Hölder’s inequality, we have
1 Z
1 2
ζ( + it)N (t, k − 1)N (t, k)|D(t, −k)| dt
T S 2
1 Z 21 1 Z 2T 1−k
2
2k 2
≪ 1
|ζ( 2 + it)| |D(t, −k)| dt |ζ( 12 + it)N (t, k − 1)|2 |D(t, −k)|2 dt
T S T T
1 Z 2T 2
k2
2 2
× |N (t, k)| |N (t, k − 1)| |D(t, −k)| dt .
k
T T
Since D(t, −k) ∼ PX (1/2 + it)−k for t ∈ S the first integral on the right hand side
R 2T
is ≪ T1 T |ZX (1/2 + it)|2k dt.
THE SPLITTING CONJECTURE 34
Remark. Note also that in this argument we can modify the definition of D to be
X αk (n)
D(t, k) =
n1/2+it
p|n =⇒ A<p6X
leads to a bounded multiplicative factor which can be absorbed into the ≪ sign.
Also, to save space in the future we may absorb the condition p|n =⇒ A < p 6 X
into the coefficients αk (n).
The lower bound in the case 0 < k 6 1 now follows from the subsequent Proposi-
tions.
Proposition 4. Suppose X 6 ηk (log T )2 (log2 T )2 . Then
k 2
1 log T
Z
1 2
ζ( + it)N (t, k − 1)N (t, k)|D(t, −k)| dt > C
T S 2 log X
for some C > 0. Assuming RH we may take X 6 (log T )θk −ǫ .
2
Proposition 5. For X 6 T 1/(log2 T ) we have
k2
1 2T
log T
Z
1 2 2
|ζ( 2 + it)N (t, k − 1)| |D(t, −k)| dt ≪ .
T T log X
2
Proposition 6. For X 6 T 1/(log2 T ) we have
k2
1 2T
log T
Z
2
2 2
|N (t, k)| |N (t, k − 1)| |D(t, −k)| dt ≪
k .
T T log X
Note that our argument works unconditionally provided X 6 ηk (log T )2 (log2 T )2
as claimed in the introduction.
7.1. Proof of Proposition 4. Our first job is to extend the range of integration
to the full set [T, 2T ]. By the usual argument involving Hölder’s inequality the
integral over E is o(1). Indeed, from the conditions on X and Lemma 3 we have
T −1 µ(E) ≪ e−2|k|V0 which is enough to kill any power of log T . We also have the
second moment bound for the zeta function, and by (24) the mth moment of D(t, −k)
is also (log T )O(1) . The only new ingredient required is a moment bound for N (t, k)
THE SPLITTING CONJECTURE 35
but by the Montgomery–Vaughan mean value theorem this is, for m 6 50,
1 2T γk (n1 ) · · · γk (n2m )
Z X
|N (t, k)|2m dt ≪
T T n1 ···nm =nm+1 ···n2m
(n1 · · · n2m )1/2
nj 6T 1/50
Y m2 k 2
2 k2
= 1+ + O(p−2 ) ≪ (log T )m .
p6T
p
Therefore,
1
Z
ζ( 12 + it)N (t, k − 1)N (t, k)|D(t, −k)|2 dt = I3 + o(1)
T S
where
1 2T 1
Z
I3 = ζ( 2 + it)N (t, k − 1)N (t, k)|D(t, −k)|2 dt.
T T
To lower bound I3 we apply the approximation
X 1 1
ζ( 12 + it) = 1/2+it
+ O 1/2
.
n6T
n T
The other terms of the integrand satisfy pointwise bounds of the form
X α−k (n)
N (t, k) ≪ Y 1/2+ǫ ≪ T 1/100+ǫ , 1/2+it
≪ X W0 (1/2+ǫ) ≪ T 1/100 ;
n
n
since k Ω(n) has average order (log n)k−1 and α−k (n) ≪ dk (n) ≪ nǫ . We then see that
the error term in the approximation for zeta leads to an error of size o(1).
Therefore
1 2T γk−1(n2 )γk (n3 )α−k (n4 )α−k (n5 ) n3 n5 it
Z X
I3 = dt + o(1).
T T n 6T, n ,n 6Y (n1 n2 n3 n4 n5 )1/2 n1 n2 n4
1 2 3
Now, since α−k (n) is supported on prime powers pm with A < p 6 X we may
write our sum as
X X X J
Y X
(52) = · ·
n1 n2 n4 =n3 n5 n1 n2 n4 =n3 n5 n1 n2 =n3 i=1 n1 n2 =n3
p|nj =⇒ A<p6X p|nj =⇒ X<p6T0 p|nj =⇒ Ti−1 <p6Ti
and note that we have essentially taken n4 = n5 = 1 in the second two sums on the
right. Unfolding the coefficients using (51) gives the first sum as
X (k − 1)Ω(n2 ) k Ω(n3 ) g(n2 )g(n3 )α−k (n4 )α−k (n5 )
(53)
n1 n2 n4 =n3 n5 (n1 n2 n3 n4 n5 )1/2
p|nj =⇒ A<p6X
Ω(n2 ),Ω(n3 )610ℓ0
The usual arguments now allow us to replace α−k (nj ) with β−k (nj ) and remove
the restrictions on Ω(n2 ), Ω(n3 ) at the cost of a term of size o(1). We then express
the resultant sum as an Euler product. Since β−k (p) = −k we find that the leading
terms cancel and that (53) is
Y
−2
(54) 1 + Ok (p ) + o(1).
A<p6X
On taking A sufficiently large we can guarantee that the term Ok (p−2 ) is always < 1
in modulus and hence the above product is > c for some constant c > 0.
With a similar computation the second sum in (52) is
X (k − 1)Ω(n2 ) k Ω(n3 ) g(n2 )g(n3 )
n1 n2 =n3 (n1 n2 n3 )1/2
p|n2 ,n3 =⇒ X<p6T0
Ω(n2 ),Ω(n3 )610ℓ0
X (k − 1)Ω(n2 ) k Ω(n3 ) g(n2 )g(n3 )
(55) = + O(e−10ℓ0 ( log T0 O(1)
log X
) )
n1 n2 =n3 (n1 n2 n3 )1/2
p|n2 ,n3 =⇒ X<p6T0
k2
Y k2
−2 log T0
= 1+ + O(p ) + o(1) > c .
X<p6T0
p log X
For the sums inside the product in (52), we must be a little more careful. The
sums in question are given by
X (k − 1)Ω(n2 ) k Ω(n3 ) g(n2 )g(n3 )
(56) .
n1 n2 =n3 (n1 n2 n3 )1/2
p|n2 ,n3 =⇒ Ti−1 <p6Ti
Ω(n2 ),Ω(n3 )610ℓi
THE SPLITTING CONJECTURE 37
Since 0 < k 6 1 and d(n) 6 2Ω(n) 6 eΩ(n) , the error incurred from dropping the
condition on Ω(n2 ) is, in absolute value,
X eΩ(n2 ) g(n2 )g(n3 ) X e2Ω(n) g(n)
6e−10ℓi 6 e−10ℓi
n1 n2 =n3 (n1 n2 n3 )1/2 n
p|n =⇒ Ti−1 <p6Ti
p|n2 ,n3 =⇒ Ti−1 <p6Ti
2
X 1
log Ti
= exp − 10ℓi + e 2
6 exp − 10ℓi + e log( log Ti−1 ) + o(1) 6 e−9ℓi
Ti−1 <p6Ti
p
log Ti
since log( log Ti−1
) 6 2 and ℓi > 109 . Doing the same for n3 gives an error of the same
size and hence the sum in (56) is
Y k2
> 1+ + O(p ) − e−8ℓi
−2
Ti−1 <p6Ti
p
(57)
k2
−7ℓi
Y −2
> 1−e 1+ + O(p ) .
T <p6T
p
i−1 i
PJ
since, again, i=1 e−7ℓi is a rapidly converging series. This completes the proof of
Proposition 4.
As in [27], we write
BT (m, n)2 1
Z BT (m, n)2 z dz
log = ,
mn 2πi |z|=1/ log T mn z2
so that the main term in I2 becomes
1
Z X hk (m)hk (n)(m, n) BT (m, n)2 z dz
.
2πi |z|=1/ log T m,n mn mn z2
2 −1
and thus we are required to show this last sum is ≪ log1X (log T / log X)k .
Unfolding the coefficients, the above sum becomes
X γk−1(m1 )γk−1(n1 )α−k (m2 )α−k (n2 )(m1 m2 , n1 n2 )1+2z
.
m1 ,m2 ,n1 ,n2
(m1 m2 n1 n2 )1+z
Estimating the terms with Ω(m2 ), Ω(n2 ) > W0 in the usual way we may replace
α−k (n) by β−k (n) and then re-extend the sums at the cost of o(1). Then by multi-
plicativity we can express the resultant sum as
X X J
Y X
· ·
m1 ,m2 ,n1 ,n2 m1 ,n1 i=1 m1 ,n1
p|mj nj =⇒ A<p6X p|m1 n1 =⇒ X<p6T0 p|m1 n1 =⇒ Ti−1 <p6Ti
where, again, we have taken m2 = n2 = 1 in the second two sums since the functions
βk (n) are supported on X-smooth numbers. As usual, we drop the conditions on
Ω(mj ), Ω(nj ) in these sums. Considering, for the moment, the first sum without
these conditions we get
Y (k − 1)m1 +n1 (−k)m2 +n2 (pm1 +m2 , pn1 +n2 )1+2z
X
−2
1+ (m1 +m2 +n1 +n2 )(1+z)
+ O(p )
A<p6X 06m ,n 61
p
j j
m1 +m2 +n1 +n2 6=0
Y 2(k − 1) − 2k (k − 1)2 + k 2 − 2k(k − 1)
= 1+ 1+z
+ + O(p−2 )
A<p6X
p p
Y 1 1 log p 1
−2
= 1 − + O(p ) + O ≪
A<p6X
p log T p log X
THE SPLITTING CONJECTURE 39
where
1
P we have used β−k (n) ≪ dk (n) and |z| 6 1/ log T along with the bound
log T p6X (log p)/p ≪ 1. As usual, for the error term we apply Rankin’s trick along
with similar Euler product computations to give an error of size
≪ e−10ℓ0 +O(log log X)
which is o(1/ log X).
For the second sum we get, with a similar argument,
X γk−1 (m1 )γk−1 (n1 )(m1 , n1 )1+2z
m1 ,n1 (m1 n1 )1+z
p|m1 n1 =⇒ X<p6T0
Y k2 − 1
= 1+ + O(p ) + O(e−10ℓ0 +O(log(log T0 / log X)) )
−2
X<p6T0
p
2 −1
which is ≪ ( log T0 k
log X
) . Finally, for the product of sums we have
J
Y X γk−1(m1 )γk−1 (n1 )(m1 , n1 )1+2z
i=1 m1 ,n1 (m1 n1 )1+z
p|m1 n1 =⇒ Ti−1 <p6Ti
J
k2 − 1
Y Y
−2 −10ℓi +O(log(log Ti / log Ti−1 ))
= 1+ + O(p ) + O(e )
i=1 Ti−1 <p6Ti
p
J
k2 − 1
Y Y
−10ℓi −2
= 1 + O(e ) 1+ + O(p )
i=1 Ti−1 <p6Ti
p
log T k −1 2
since log Ti / log Ti−1 6 2. This last term is ≪ ( log T0
) and so combining these
bounds in (58) we get
k2 −1 k2 −1 k2
1 log T0 log T log T
I4 ≪ log T · · · ≪
log X log X log T0 log X
which completes the proof of Proposition 5.
7.3. Proof of Proposition 6. We begin with the following lemma from [27].
Lemma 9. Let
X 1
Pj (t) := .
Tj−1 <p6Tj
p1/2+it
Then for 0 6 j 6 J
1
|Nj (t, k − 1)Nj (t, k) k |2 6 |Nj (t, k)|2 (1 + O(e−9ℓj )) + O Qj (t) ,
THE SPLITTING CONJECTURE 40
Proof. This is essentially Lemma 1 of [27]. Our sequence Tj is defined slightly differ-
ently but one can check that this makes no difference to the end result.
Lemma 10. With the above notation
1 2T
Z
2
Q0 (t)|D(t, −k)|2 dt ≪ e−10ℓ0 (log X)2k
T T
and for 1 6 j 6 J
2T
1
Z
Qj (t)dt ≪ e−10ℓj .
T T
Proof. We prove the first bound since this is new, the second bound follows similarly
(and is essentially Lemma 2 of [27]). Let L = 10ℓ0 . From the definition of Q0 (t) we
have
1 2T
Z
Q0 (t)|D(t, −k)|2 dt
T T
L/k
12 2L X 2e 2r 1 2T
Z
= · |P0 (t)|2L+2r |D(t, −k)|2 dt
L r=0
r + 1 T T
the Montgomery–Vaughan mean value theorem gives that our original integral is
L/k 1/2
g(n)2
2k 2 12 2e 2r
2L X X
2
≪(log X) · (2L + 2r)!
L r=0
r + 1 n
Ω(n)=2L+2r
(59) p|n =⇒ p6T0
L/k
12 2L X X L+r
2k 2 2e 2r 1/2 1
6(log X) · (2L + 2r)!
L r=0
r+1 p6T
p
0
THE SPLITTING CONJECTURE 41
since g(n)2 6 g(n). Letting P = p6T0 p−1 we find by Stirling’s formula that the
P
summand is
≪ (2/e)L (8e)r (r + 1)−2r (L + r)L+r+1/4 P L+r
which is maximised at the solution of r 2 = 8P √(L + r)(1 + O(1/r)). Since P 6
2 log log T = o(L) this solution r = r0 satisfies 2 2(P L) 6 r0 6 3(P L)1/2 . There-
1/2
fore, (59) is
2
c 2L L
≪ (log X)2k · · (3L)!1/2 P L+r0 r0−2r0
L k
since 2r0 6 L. This is then
2
≪ (log X)2k L−L/2+o(1) ≪ e−10ℓ0
and the result follows.
By Lemma 9 and (41) we find that
1 2T
Z
2
I5 := |N (t, k)| k |N (t, k − 1)|2 |D(t, −k)|2 dt
T T
1 2T
Z
2 −10ℓ0
≪ |N0 (t, k)| (1 + O(e )) + O Q0 (t) |D(t, −k)|2 dt
T T
J+1
Y 1 Z 2T
× |Nj (t, k)|2 (1 + O(e−10ℓj )) + O Qj (t) dt
j=1
T T
since J ≍ log log log T . By Lemma 10 we have
J
1 2T
Z
Y
|Nj (t, k)|2 (1 + O(e−10ℓj )) + O Qj (t) dt
j=1
T T
J
k 2Ω(n) g(n)2
Y X
−10ℓj −10ℓj
= (1 + O(e )) + O(e )
(60) n
j=1 Ω(n)610ℓj
p|n =⇒ Ti−1 <p6Ti
J k 2
k2
Y Y log T
≪ 1+ + O(p−2 ) ≪ .
j=1 Ti−1 <p6Ti
p log T0
The usual arguments allow us to remove the conditions on Ω(mj ) and replace α−k (n)
by β−k (n) at the cost of o(1). We then find that the resultant sum is
k2
k2
Y Y
−2 −2
log T0
1 + O(p ) 1+ + O(p ) ≪
A<p6X X<p6T
p log X
0
since the leading terms cancel over A < p 6 X. Combining this with (60) gives
k2
log T
I5 ≪
log X
thus completing the proof of Proposition 6.
8.1. Modifying the proof of Proposition 4. We see that we can arrive at (52)
in exactly the same way. When dealing with (53), the errors incurred from dropping
the conditions on Ω(n2 ) etc. are now
2
≪ e−10ℓ0 (log X)Ck
whichPof course is still o(1) since k is fixed. Note that on writing (53) as a single
sum n f (n)/n, the coefficients are supported on X-smooth numbers and satisfy
the bounds |f (n)| 6 k 2Ω(n) d3 (n)2 6 (3k)2Ω(n) . Then we find that the equivalent of
(53) is
Y k 4
1+O 2 + o(1) > C
2
p
Bk <p6X
for some C on taking B large enough. In a similar way we find that the equivalent
of (55) is
k 2
Y k2 k 4
log T0
1+ +O 2 + o(1) > C .
X<p6T
p p log X
0
in the current context. The error from removing the condition on Ω(n2 ) in the sums
is
X (ek)Ω(n2 ) k Ω(n3 ) g(n2 )g(n3 )
6e−10ℓi
n1 n2 =n3 (n1 n2 n3 )1/2
p|n2 ,n3 =⇒ Ti−1 <p6Ti
−10ℓi
X (ek)2Ω(n) g(n)
2 2
X 1
6e = exp − 10ℓi + e k
n Ti−1 <p6Ti
p
p|n =⇒ Ti−1 <p6Ti
log Ti
6 exp − 10ℓi + e2 k 2 log( log Ti−1
) + o(1) 6 e−9ℓi
J k 2
k2 k 4
Y
−7ℓi
Y log T
> 1−e 1+ +O 2 >C .
i=1 Ti−1 <p6Ti
p p log T0
THE SPLITTING CONJECTURE 44
Proof. We prove the first formula since the second follows similarly. By the Cauchy–
Schwarz inequality and (23) we have
1/2
1 2T
Z 2T
1
Z
40ℓ0 2 2k 2 40ℓ0 2
|P0 (t)| |D(t, −k)| dt ≪ (log X) |P0 (t) | dt .
T T T T
Letting L = 10ℓ0 = 10(log log T )3/2 the last integral is
4L
g(n)2
X
2
X 1
≪ (4L)! 6 (4L)! 6 (4L)!(log log T )4L
n T <p6T
p
Ω(n)=4L −1 0
p|n =⇒ T−1 <p6T0
References
[1] J. Andrade, A. Shamesaldeen, Hybrid Euler-Hadamard Product for Dirichlet L-functions with
Prime conductors over Function Fields, preprint, arxiv.1909.08953.
[2] L.-P. Arguin, D. Belius, P. Bourgade, M.Radziwill, K. Soundararajan, Maximum of the Riemann
zeta function on a short interval of the critical line Commun. Pure. Appl. Math 72 no. 3 (2019),
500–535.
[3] L.-P. Arguin, P. Bourgade, M. Radziwill, The Fyodorov-Hiary-Keating Conjecture. I, preprint,
arxiv.2007.00988.
[4] L.-P. Arguin, F. Ouimet, M. Radziwill, Moments of the Riemann zeta function on short intervals
of the critical line, preprint, arxiv.1901.04061.
[5] R. Balasubramanian, J. B. Conrey, D. R. Heath-Brown, Asymptotic mean square of the product
of the Riemann zeta-function and a Dirichlet polynomial, J. Reine Angew. Math. 357 (1985),
161–181.
[6] S. Bettin, H. M. Bui, X. Li, M. Radziwill, A quadratic divisor problem and moments of the
Riemann zeta-function, preprint, arXiv.1609.02539.
[7] S. Bettin, V. Chandee, M. Radziwill, The mean square of the product of ζ(s) with Dirichlet
polynomials, J. Reine Angew. Math, 729 (2017), 51–79.
[8] H. Bui, J. Keating, On the mean values of Dirichlet L-functions, Proc. London Math. Soc. 95
(2007), 273–298.
THE SPLITTING CONJECTURE 46
[9] H. Bui, J. Keating, On the mean values of L-functions in orthogonal and symplectic families,
Proc. London Math. Soc. 96 (2008), 335–366.
[10] H. Bui, S. Gonek, M. Milinovich, A hybrid Euler-Hadamard product and moments of ζ ′ (ρ),
Forum Math. 27 (2015), 1799–1828.
[11] H. Bui, A. Florea, Hybrid Euler-Hadamard product for quadratic Dirichlet L-functions in func-
tion fields, Proc. London Math. Soc. 117 (2018), 65–99.
[12] E. Carneiro, V. Chandee, M. B. Milinovich, Bounding S(t) and S1 (t) on the Riemann hypoth-
esis, Math. Ann. 356 no. 3 (2013), 939–968.
[13] J. B. Conrey, D. W. Farmer, J. P. Keating, M. O. Rubinstein, N. C. Snaith, Integral moments
of L-functions, Proc. London Math. Soc. 91 no. 3 (2005), 33–104
[14] J. B. Conrey, A. Ghosh, A conjecture for the sixth power moment of the Riemann zeta-function,
Int. Math. Res. Not. 15 (1998) 775–780.
[15] J. B. Conrey, S. Gonek, High moments of the Riemann zeta function, Duke Math. J. 107
(2001) 577–604.
[16] M. Das, Selberg’s central limit theorem for L-functions of level aspect, preprint,
arxiv.2012.10766.
[17] C. David, A. Florea, M. Lalin, Non-vanishing for cubic L-functions, preprint, arxiv.2006.15661.
[18] A. Diaconu, D. Goldfeld, and J. Hoffstein, Multiple Dirichlet series and moments of zeta and
L-functions, Compositio Math. 139 (2003), 297–360.
[19] G. Djanković, Euler-Hadamard products and power moments of symmetric square L-functions,
Int. J. Number Theory 9 (2013) 621–639.
[20] D. W. Farmer, S. Gonek, C.P. Hughes, The maximum size of L-functions, J. Reine Angew.
Math. 609, (2007) 215–236.
[21] P. Gao, Sharp bounds for moments of quadratic Dirichlet L-functions, preprint,
arxiv.2101.08483.
[22] S.M. Gonek, C.P. Hughes, J.P. Keating, A hybrid Euler-Hadamard product for the Riemann
zeta function, Duke Math. J. 136 no. 3 (2007), 507-549.
[23] G.H. Hardy, J.E. Littlewood, Contributions to the theory of the Riemann zeta-function and
the theory of the distribution of primes, Acta Arith. 41 (1918), 119–196.
[24] A. Harper, Sharp conditional bounds for moments of the Riemann zeta function. preprint,
arXiv.1305.4618.
[25] W. Heap, Moments of the Dedekind zeta function and other non-primitive L-functions, Math.
Proc. Cam. Phil. Soc., doi:10.1017/S030500411900046X.
[26] W. Heap, M. Radziwill, K. Soundararajan, Sharp upper bounds for fractional moments of the
Riemann zeta function, Quarterly J. Math. 70 no. 4 (2019), 1387–1396.
[27] W. Heap, K. Soundararajan, Lower bounds for moments of zeta and L-functions revisited,
preprint, arxiv.2007.13154.
[28] D. R. Heath-Brown, Fractional moments of the Riemann zeta function, J. London Math. Soc.,
24, no. 1 (1981), 65–78.
[29] C. Hooley, On an almost-pure sieve, Acta Arith. 66 no. 4 (1994), 359–368.
[30] P.-H. Hsu, P.-J. Wong, On Selbergs Central Limit Theorem for Dirichlet L-functions, Journal
de Theorie des Nombres de Bordeaux, 32 no. 3 (2020), 685–710.
[31] C. P. Hughes, M. P. Young, The twisted fourth moment of the Riemann zeta function, J. Reine
Angew. Math. 641 (2010), 203–236.
[32] P. Humphries, M. Radziwill, Optimal Small Scale Equidistribution of Lattice Points on the
Sphere, Heegner Points, and Closed Geodesics, preprint, arxiv.1910.01360.
THE SPLITTING CONJECTURE 47
[33] A. E. Ingham, Mean-value theorems in the theory of the Riemann zeta function, Proc. London
Math. Soc. 27 (1926), 273–300.
[34] J. P. Keating, N. C. Snaith, Random matrix theory and ζ(1/2 + it), Comm. Math. Phys. 214
(2000) 57–89.
[35] S. Lester, M Radziwill, Signs of Fourier coefficients of half-integral weight modular forms,
preprint, arxiv.1903.05811.
[36] H. Montgomery, R. Vaughan, Hilbert’s inequality, J. London Math. Soc. 8 no. 2 (1974) 73–82.
[37] M. Radziwill, Large deviations in Selberg’s central limit theorem, preprint, arxiv.1108.5092.
[38] M. Radziwill, K. Soundararajan, Moments and distribution of central L-values of quadratic
twists of elliptic curves, Invent. Math. 202 no. 3 (2015) 1029–1068.
[39] M. Radziwill, K. Soundararajan Selberg’s central limit theorem for log |ζ( 12 + it)|. Enseign.
Math. 63 (2017), 1–19.
[40] K. Ramachandra, Some remarks on the mean value of the Riemann zeta function and other
Dirichlet series, Ann. Acad. Sci. Fennicae, 5 (1980), 145–158.
[41] K. Ramachandra. Some remarks on the mean value of the Riemann zeta function and other
Dirichlet series. II Hardy- Ramanujan J., 3 (1980), 1-24.
[42] A. Selberg, Contributions to the theory of the Riemann zeta-function, Arch. Math. Naturvid.
48 (1946), 89–155.
[43] K. Soundararajan, Moments of the Riemann zeta function, Annals of Math. 170 no. 2 (2009),
981–993.
[44] E. C. Titchmarsh, The Theory of the Riemann Zeta-Function, Second Edition, The Clarendon
Press, Oxford University Press, 1986.
[45] K. M. Tsang, Some Ω-theorems for the Riemann zeta-function, Acta Arith. 46 (1986), 369–395.
Email address: winstonheap@gmail.com