Primes in arithmetic progressions on average I

Tomos Parry

Abstract
Let
q
X x X X ′
Ex (q, a) := log p − M (Q) := φ(q) Ex (q, a)3 .
φ(q) a=1
p≤x q≤Q
arXiv:2406.06450v1 [math.NT] 10 Jun 2024

p≡a(q)

We show that if Q is close to x then M (Q) ≪ Q3 (x/Q)4/3 p, thereby showing that sign-changes in the
average gives power-saving cancellation past the expected x/q heuristic.

1 Introduction
Let X x
Ex (q, a) := log p −
p≤x
φ(q)
p≡a(q)

be the error term when counting

p primes in arithmetic progressions,and which (ignoring log factors) is famously
suspected to be around x/q in size. When averaging, however, additional cancellation may occur, as was
demonstrated by Hooley [2] in the form
q
X X ′
M (Q) := φ(q) Ex (q, a)3
q≤Q a=1
 3/2
x √ x3
≪ Q3 e−c log x
+
Q (log x)A

Hooley remarks further “The assumption of the extended Riemann hypothesis for Dirichlet’s L-functions
enables one to obtain a greatly improved form of our result in which a sharper version of the above bound is
valid for smaller values of Q, thus shedding further light on the distribution of Ex (q, a) for smaller values of
q. But the investigation of this development must await a later paper in this series.”. It seems Hooley never
did manage to return to this development but following his proof we see that he ultimately reduces M (Q)
to an integral looking something like
Z
ζ(s)ζ(s + 1)h(s)(x/Q)s+2 ds
Q3
(−1/2) ζ(2s + 2)s(s + 1)(s + 2)

with h(s) holomorphic and bounded for σ > −1, so we’d guess he meant that under RH the above integral
may be moved to the left (as well as stronger bounds then being available for the circle method part of his
argument) to imply something like
 5/4
x
M (Q) ≪ Q3 + x5/2
Q
although, in respect to the wealth of his contributions, we should also allow for the possibility that he meant
something more.

In any case, as far as we know, there p is no trace of anything in the literature suggesting additional can-
cellation of Ex (q, a) past the expected x/q heuristic further than Hooley’s above result, and certainly no
power-saving without RH. In this paper, we reveal an additional cancellation within the dispersion method
which allows us to (unconditionally) prove

Theorem.  4/3
3 x x3
M (Q) ≪ Q +
Q (log x)A

1
 Let’s give a few words as to the proof. Opening up M (Q) we see it’s
q
X X X X ′ X 1 X 1
φ(q) log p − 3x Ex (q, a)2 − 3x2 Ex (1, 0) − x3
φ(q) φ(q)2
q≤Q p1 ,p2 ,p3 ≤x q≤Q a=1 q≤Q q≤Q
p1 ≡p2 ≡p3 (q)

where the first term is “the difficult part” and where let’s call the second term “the variance part”. Hooley
finds a precise asymptotic for the difficult part and compares it to the main terms of the remaining parts,
including the variance part. We prove our theorem following this argument, but instead of bounding the
integral mentioned above, we further compare it to the error term in the variance part, and show they
completely negate each other. This requires, and what is most likely the most important contribution of our
proof, keeping better track of the main terms than in [2], which in (117) contains a large term impossible
to control without RH. This laying-out of the main terms in a more transparent manner so as to isolate the
correct main term is essential if we wish to pass to the study of the fourth moment with a dispersion method,
and it is indeed that subject which is the main motivation for the present work. We return to this in the
sequel to this article, but let’s just briefly indicate now what we mean: Opening up the fourth moment as
was done above for the third moment, and using (e.g.) a weight φ(q)2 , we get a “difficult part” plus parts
q q
X X ′ X X ′
3 2
−4x φ(q) Ex (q, a) +6x Ex (q, a)2 + ...
q≤Q a=1 q≤Q a=1
| {z }
=M(Q)

so that if we want a bound for the fourth moment of size ≪ Q4 (x/Q)2+δ then at the very least we need a
bound M (Q) ≪ Q3 (x/Q)1+δ .

As a final but ultimately just vague comment, we mention that error terms in integrals negating each other
when otherwise needing to call on RH is something we believe should maybe not be glossed over.

We present our proof in two sections. In the first we prove our theorem subject to Lemma 1 - here we are just
re-writing the first half of [2]. In the second we state and prove Lemma 1, where we find the new cancellation.
In particular this lemma gives us a contour integral with integrand containing a factor 1/s2 (s + 1) instead of
the 1/s(s+1)(s+2) in [2]. This seemingly completely uninteresting observation proves to be just the opposite.

We particularly thank Professor Ahmet Güloğlu for many helpful discussions in writing this paper.

2 Proof of theorem, subject to Lemma 1

We retrace initially Hooley’s route exactly, however we give all the details since there are a few inaccuracies1
R 1
R c+1/100+i∞
in [2]. Notation: By (c) we mean 2πi c+1/100−i∞ .

Let A > 0 and P := x/(log x)A+100 . By (Gallagher’s form of) the Barban-Davenport-Halberstam theo-
rem q
X X ′ x2
Ex (q, a)2 ≪
a=1
(log x)A
q≤P

1 Essentially §4 needs to be re-written, the ultimate problem being that smoothness of sums involved is overlooked. (22) to

(23) misses a large term, causing a sharp cut-off, and the proof of Lemma 2 is given as a reference to another lemma but where
the sum is smoother than in Lemma 2. This is all fixable and of course doesn’t detract from the main point, but it probably
does mean that we’d be wise to write things out again to clear up any possible blurriness.

2
so, with the prime number theorem,
q  
X X ′ x3
M3∗ (Q) = φ(q) Ex (q, a)3 + O
a=1
(log x)A
P <q≤Q
q
X X X X ′
= φ(q) log p − 3x Ex (q, a)2
P <q≤Q p≤x P <q≤Q a=1
p1 ≡p2 ≡p̃3 (q)
 
1 X X 1 x3
− 3x2 Ex (1, 0) − x3 +O
φ(q) φ(q) (log x)A
P <q≤Q P <q≤Q
   
x3 3 x3
=: S1 (Q) − 3xV (Q) + O − x U (Q) + O (1)
(log x)A (log x)A

and by (2), (5), (6), (8), (9), (10) and (11) of [1]
 
  x2
V (Q) Qx log x − Qx − x2 U (Q) + x2 J(x/P ) − J(x/Q) + O
=
(log x)A
Z
log X + C2 F1 (s + 1)(x/Q)s ds
where J(X) := Z log X + C1 + +2
X (−5/4) s(s + 1)(s + 2)
∞
X n/φ(n)
F1 (s) :=
n=1
ns
ζ(2)ζ(3) ζ ′ (0) X log p
Z := C2 := + γ0 + ; (2)
ζ(6) ζ(0) p
p(p − 1)

here we get the value of C2 from (10) of [2]. From Lemma 1 of [1]
 
log P
U (Q) = Z log(Q/P ) + O (3)
P

and of course J(x/P ) ≪ P/x so (2) becomes

Z
2 F1 (s)(x/Q)s+1 ds
V (Q) = Qx log Q − (C2 + 1)Qx + 2Q +O (P x log x) . (4)
(−1/4) (s − 1)s(s + 1)
| {z }
=:I(x/Q)

and with (3) then (1) becomes, writing X = x/Q,

 
  x3
M3∗ (Q) = S1 (Q) − 3Q3 X 2 log Q − X 2 (C2 + 1) + 2XI(X) − 3Zx3 log(Q/P ) + O . (5)
(log x)A

Let
X φ(q) X
J1 (Q) := log p
q
Q<q≤x p≤x
p1 ≡p2 ≡p3 (q)

so that
  Z Q  
S1 (Q) = Q J1 (P ) − J1 (Q) − J1 (P ) − J1 (t) dt. (6)
P

Then, using φ ⋆ 1 = id and where f (x) is some quantity independent of Q,

(
X µ(d) X X X X
J1 (Q) = 6 log p + 3 (log p)2 log p′
d Q<q≤x p<p′ ≤x
d≤x Q<q≤x p <p <p ≤x 1 2 3
d|q p1 ≡p2 ≡p3 (q) d|q p≡p′ (q)
)
X X X φ(q) X
+ 3 log p(log p′ )2 + (log p)3
Q<q≤x p<p′ ≤x
q
Q<q≤x p≤x
d|q p≡p′ (q)
 

=: 6J4 (Q) + 3 J5 (Q) + J6 (Q) + f (x) + O xQ(log x)3 (7)

3
so we need to calculate the Ji ’s - first we do J5 and J6 . Let
Y  ∞
1 X ψ1 (l)
ψ1 (l) := 1+ F (s) := so F (s) = ζ(s)ζ(s + 1) F ∗ (s) =: ζ(s)h(s)
p − 1 − 1/p ls | {z }
p|l l=1
≪ 1 for σ > −1/2
′
α1 = h(1) β1 = h(1)(γ0 − 1) + h (1)
 
h(1) γ0 3 h′ (1)
α2 = β2 = h(1) − +
2 2 4 2
so that, from the usual Perron type formulas and moving the integrals a bit past σ = −1 with the simplest
of bounds for ζ(s),
Z
X ψ1 (l) F (s + 1)X s+1 ds
(X − l) = = X (α1 log X + β1 ) + O(1)
l (0) s(s + 1)
l<X
Z
X
2 ψ1 (l) F (s + 1)X s+2 ds
(X − l) = 2 = 2X 2 (α2 log X + β2 ) + O(X)
l (0) s(s + 1)(s + 2)
l<X
Z
X F (s + 1)X s+2 ds
(X − l)ψ1 (l) = = c1 X 2 + O(X)
(0) (s + 1)(s + 2)
l<X
Z
X F (s + 1)X s+2 ds
log(X/l)lψ1(l) = = c2 X 2 + O(X) for some constants c1 , c2 .
(0) (s + 2)2
l<X

Let
(lQ)2 log(X/l) x log x 1 3 log x
M1 (l) := Q2 (X − l) (γX + δl) − X := γ := − δ= −
2 Q 2 4 4 2
Z x−Ql
M2 (l) := log t (x − Ql − t)dt
0
and note that
X µ(d) 1   Y 
C3 ψ1 (l) log x 1
= +O where C3 = 1− (8)
d φ(dl) l x p
p(p − 1)
d≤x

so the integral representations above give (again for some f (x) independent of Q)
X M1 (l)ψ1 (l)

= γα1 x2 log X + f (x) + O Q2 X
l
l≤X
Z X !
X M2 (l)ψ1 (l) Q2 log Q X 2 ψ1 (l) 2
X ψ1 (l)
= (X − l) +Q log(X − t) (t − l) dt
l 2 l 0 l
l≤X l≤X l<t
 
= Q2 X 2 log Q α2 log X + β2
Z X n o

+ Q2 t log(X − t) α1 log t + β1 dtO Q2 X log X
 0 
2
= x log Q α2 log X + β2
   
2 α1 (log X)2 β1
x + − α1 log X + f (x) + O(Q2 X log X); (9)
2 2
if this last integral proves tricky to check we can always refer to the proof of Lemma 2 in [2] and read from
line -6 to the end. Recall Q ≥ x/(log x)A+100 and put L = (log x)A+10 x/Q. We have from the Siegel-Walfisz
theorem
 ′  x   2+ǫ 
X X
′ 2
X X
′ 2 p − Ql x
log p(log p ) = (log p ) +O +O
p,p′ ≤x ′
φ(dl) L Q
l<x/Q l<x/Q Ql<p ≤x
p≡p′ (dl)
p′ −p
Q<
l
(Z  )  
x
X 1 x2 x2
= (t − Ql) log t dt + O +O
φ(dl) Ql (log x)A+2 (log x)A+1
l<x/Q
 
X M1 (l) x2
= +O
φ(dl) (log x)A+1
l<x/Q

4
and similarly
  x   2+ǫ 
X X X X x − Ql − p x
(log p)2 log p′ = (log p)2 +O +O
p,p′ ≤x
φ(dl) L Q
l<x/Q l<x/Q p≤x−Ql
p≡p′ (dl)
p′ −p
Q<
l
    
X 1 x2 x2
= M2 (l) + O +O
φ(dl) (log x)A+2 (log x)A+1
l<x/Q

so, re-writing the congruence condition, J5 (Q) + J6 (Q) is

 ! 
X µ(d)  X  2

1 x
M1 (l) + M2 (l) + O
d  φ(dl) (log x)A+1 
d≤x l<x/Q

and we conclude from (8) and (9)

   
α1 (log X)2 β1
J5 (Q) + J6 (Q) = C3 x2 + γα1 + − α1 log X + α2 log Q log X + β2 log Q
2 2
 
2 x2
+ f (x) + O Q X log X +
(log x)A
     
2 3α1 5α1 β1 x2
= C3 x log Q α2 − log x + − + β2 + f (x) + O Qx log x +
2 4 2 (log x)A
   2

x
= x2 log Q − log x + 1 + f (x) + O Qx log x + (because C3 h(1) = 1)
(log x)A
 
x2
=: B(Q) + f (x) + O Qx log x + . (10)
(log x)A

Now to J4 . Whenever the letters l, l′ are in context write l̃ = l + l′ . Let

 
1 1 if p > 2 Y Y
r := Φ(∆) = (1 + 2r) Γ∆ (l) := (1 + r)
p − 2 − p1 0 if p = 2
p|∆ p|l
p∤∆
  Y  I(p) 
−r(1 + 3r + r2 ) if p is odd if ∆ is squarefree
I(p) := I(∆) :=
2 if p = 2 0 if not
p|∆
X I(∆)(X − ∆l̃)2 Γ∆ (l)Γ∆ (l̃) X
J (X) := J ∗ (X) := I(∆)(X − ∆l̃)2 ∆l̃Γ∆ (l)Γ∆ (l̃)
∆l̃
∆l̃≤X ∆l̃≤X
Y 1
 Y 
1
 Y 2r

C5 := 1− C 6 := 1 − C 8 := 1 +
p>2
(p − 1)2 p>2
p(p − 2) p>2
p
 
Y Y
U(s) := 1+ σ>0 Y := p−s
p
p −1
Y Y

= ζ(s + 1) 1+ (1 − Y /p) σ > −1/2
p
p−1
2Γ(3) 1 2Γ(3)ζ(0) 1 ζ ′ (0) 13
a := = b= =− Z := + γ0 −
Γ(6) 30 Γ(5) 12 ζ(0) 12
!
X log p ζ(0)
α := a U(1) = β := b U(0) γ := β Z + +
p
p(p − 1) 12
Z
ζ(s)U(s)X s+4 ds
M (X) := αX 5 + βX 4 log X + γX 4 E(X) := 2 (11)
(−1/2) s(s + 3)(s + 4)

and denote by Γ(s) the Gamma function. Hooley deals with J4 using the circle method, this original idea
being the main point of the work. There are no inaccuracies here as far as we can tell, so we simply report
that  
C5 C6 Q2 x2
J4 (X) = J (X) + O
2 (log x)A

5
which (referring to [2]) we can read off from (13), (18), (54), (59), (60), (63), (64) and the sentence following
(64), and we get from (7) and (10)
 
2 2 x2
J1 (Q) = 3C5 C6 Q J (X) + 3B(Q) + f (x) + O xQ(log x) + . (12)
(log x)A

for some f (x) independent of Q. We now need

Lemma 1. Assume the notation (11). For δ = 1/6

   
J ∗ (X) = C8 M (X) + E(X) + O X 7/2−δ .

This should be compared with (117) of [2], where the main term can’t be controlled without RH and
where the error term already contains the usual “zero-free region error”.

We prove this lemma in the next section. Assuming it for the moment, write
Z X Z X
f (X) f (t)dt 2 f (t)dt
Cf (X) = − 6X + 12X
X2 0 t 4
0 t5

so that, for some f (x) independent of Q,

CM (X) = 10αX 3 + 6βX 2 (log X)2 + (12γ − 5β)X 2 log X + X 2 (6β − 5γ)
 
 
U(1)X 3 U(0)X 2 (log X)2 
 ζ ′ (0) X log p   2
= − + U(0) 1 − − γ0 −  X log X
3 2  ζ(0) p
p(p − 1) 
 
| {z }
=−C2 as in (2)
3 2 2 2
=: AX + BX (log X) + CX log X
 
= AX 3 − X 2 log Q(2B log x + C) − B(log Q)2 + f (x)
=: M ∗ (Q)

and
Z Z
ζ(s)U(s) ζ(s)U(s)X s+2
CE (X) = Cx7→xs (X)ds = ds =: E ∗ (X)
(−1/2) s(s + 3)(s + 4) (−1/2) s2 (s + 1)

to conclude
 3/2−∆ !
  x
P CJ ∗ (x/P ) − Q CJ ∗ (x/Q) = C8 P M (x/P ) − Q M (x/Q) + P E (x/P ) − Q E (x/Q) + O Q2
2 2 2 ∗ 2 ∗ 2 ∗ 2 ∗
Q
   3/2−∆ !
2 x
=: C8 A(Q) + E(Q) + O Q .
Q

From Lemma 3 of [2] and (12) this means

 3/2−∆ !
 
2 x x2
J1 (P ) − J1 (Q) = 3 A(Q) + B(Q) + E(Q) + O Q + (C5 C6 C8 = 1)
Q (log x)A
 3  !
x x3 2
 
2 2
= 3A − + 3x (2B + 1) log x + C − 1 log(P/Q) + B(log P ) − B(log Q)
P Q
 3/2−∆ !
2 x
+ 3E(Q) + O Q (from (10); note also U(0) = 1)
Q
 3   3/2−∆ !
x x3 2

2 2

2 x
= 3A − + 3x C2 log(P/Q) + B(log P ) − B(log Q) + 3E(Q) + O Q
P Q Q
 3/2−∆ !
2 ∗∗ 2 x
= 3x M (Q) + 3E(Q) + O Q
Q

6
so by (6) and the above definition for E(Q)
( Z Q Z Z !)
Q
2 ∗∗ ∗∗ ζ(s)U(s)xs −s −s −s −s
S1 (Q) = 3x QM (Q) − M (t)dt + 2 Q(P −Q )− (P −t .
P (−1/2) s2 (s + 1) P
| {z }
s
=Q1−s s−1 +O(xP −s )

The first term in curly brackets is

Z Q  
Ax S 2B log t
= t − − dt
P t2 t t
 
= Ax log(Q/P ) − Q 2B log Q − 2B + C2 + O(P log P )
= Ax log(Q/P ) + Q (log Q − 1 − C2 )
and from (4) the second term gives a total contribution
 
x2
6xQ2 I(x/Q) + O
(log x)A
and our theorem now follows from (5) since U(1) = Z and U(s) = F1 (s) (these were defined in (2)). All
we’re left to do is prove Lemma 1.

3 Proof of Lemma 1
Here we prove Lemma 1. We suspect the power saving may be improved (although it’s not clear to us yet
whether δ = 1/2 is possible, which would be the best input for the fourth moment) but this needs a new
argument in the proof of (27) below. For now in this article we’re content with the simplest saving, the main
point being the cancellation between integrals.

Let r be as in (11) so that

r = 1/p + O(1/p2 ) (13)
Let Γ∆ (l) be as in (11) and for σ > 1 and a Dirichlet character χ let
Y 1 r2 U (p) Y
U (∆) := K(p) := 1 + K∆ (l) := K(p)
1 + 2r
p
p p|l
p|∆
p∤∆

∞ ∞
χ
X Γ∆ (l)χ(l) X K∆ (l)
G∆ (s) := Kd,∆ (s) := . (14)
ls ls
l=1 l=1
d|l

We have2 , writing Y = χ(p)p−s ,

holomorphic and ≪ 1 for σ > −1/2 by (13)
z  }|  Y  {
χ
Y Y Y
G∆ (s) = Lχ (s)Lχ (s + 1) (1 + rY ) 1 − 1− (15)
p p
p∤∆ p|2∆

and in particular, assuming (d, ∆) = 1 and writing Y = p−s ,

   
χ0
Y Y Y Y
G∆ = Lχ0 (s)Lχ0 (s + 1) (1 + rY ) 1 − 1− =: Lχ0 (s)Lχ0 (s + 1)Rd,∆ (s)
p p
p∤d∆ p|∆
Y
= Lχ0 (s) (1 + rY ) =: Lχ0 (s)Qd∆ (s) (16)
p∤d∆

and we have, using K∆ (ld) = K∆ (d)Kd∆ (l),

 

ζ(s)K∆ (d) Y 
 r2 U (p)Y 

Kd (s) = 1 +  (17)
ds  p 
p∤d∆ | {z }
≪1/p for σ>−2

2 G χ (s) is obviously extendable to σ > −1 but we don’t need it

∆

7
so
χ
for σ > −1/2 G∆ (s) ≪ |Lχ (s)Lχ (s + 1)|

≪ |ζ(s)ζ(s + 1)| max d−σ , 1 if χ = χ0
n o
≪ max (dt)1/2−σ , 1 always
ζ(s)
for σ > −2 Kd (s) ≪ . (18)
ds
Let, where Γ is the Gamma function, C8 is as in (11), and Γ∆ , Qd∆ , R1,∆ are as above,

2Γ(3)Qd∆ (1)2 X Γ∆ (e)2

A∆ (d) := Lχ (d/e) (1)2
Γ(6) φ(d/e)e2 0
e|d

2Γ(3)ζ(0)
c :=
Γ(5)
c
B∆ (d) := Qd∆ (1)Qd∆ (0)Γ∆ (d)2
d
∗ c X log p
B∆ (d) := Qd∆ (1)Qd∆ (0)Γ∆ (d)2
d Γ∆ (p)2
p|d
′ ′
ζ (0) Γ (5)
Z := − +1
ζ(0) Γ(5)
R′1,∆ (0)
Z∆ := +Z
R1,∆ (0)
 
Md,∆ (X) := A∆ (d)X 5 + B∆ (d) log X/d + Zd∆ X 4 + B∆ ∗
(d)X 4
 
Z X Γ∆ (e)2
Qd∆ (s)
Id,∆ (X) := Qd∆ (1)  Lχ (d/e) (1)Lχ0 (d/e) (s) X s+4 ds
(−1/2) s(s + 2)(s + 3)(s + 4) φ(d/e)es+1 0
e|d
Z s+4
Kd,∆ (s)X ds
Jd,∆ (X) := C8 U (∆) (19)
(−1/2) (s + 2)(s + 3)(s + 4)
and note in particular, keeping one integral where it is and moving the other to the left,
X4
(d ≤ X) Id,∆ (X) ≪ Jd,∆ (X) ≪ d2 X 2 . (20)
d
Whenever the letters l, l′ are in context write ˜l = l + l′. With r, Γ∆ as in (11) define multiplicative g∆ through
Γ∆ = g∆ ⋆ 1 so that

g1 (d) if (d, ∆) = 1
g∆ (d) = and g1 (p) = r (21)
0 if (d, ∆) > 1 or if d is squarefree
and so that
X X X
(X − l̃)2 l̃Γ∆ (l)Γ∆ (l′ )Γ∆ (l̃) = g∆ (d) (X − l̃)2 ˜lΓ∆ (l)Γ∆ (l′ )
l+l′ ≤X d≤X l+l′ ≤X
l̃≡0(d)
X
=: g∆ (d)Ld,∆ (X). (22)
d≤X

Let C8 , U(s) be as in (11) and let

Y Y 2r rY


Vs (∆) := Vs (p) := 1+ + p + 3r + r2 Y := p−s (23)
p p
p|∆ p|∆

so that
Y
V∆ (s) := Vs (p) σ>0
p∤∆

ζ(s + 1) Y
= Vs (p)(1 − Y /p) σ > −1/2 (by (13)) (24)
Vs (∆) p

8
so that with (11)
∞  
X I(∆)V∆ (s) Y Y I(p)
= Vs (p) + = C8 U(s) σ>0 (25)
∆s+1 p
p
∆=1

We will prove that

  
Ld,∆ (X) = Md,∆ (X) + 2Γ(3)Id,∆ (X) + O X ǫ X 3 + d1/2 X 5/2 + d2 X 2 (26)

and
 
C8 U (∆)Kd (1) 5 X4
Ld,∆ (X) = X + 2Jd,∆ (X) + O (27)
30 d

which from (20) imply

  
X4
Ld,∆ (X) = Md,∆ (X) + 2Γ(3)Id,∆ (X) + 2Jd,∆(X) + O X ǫ min X 3 + d1/2 X 5/2 + X 2 d2 ,
d

and therefore
∞
X ∞
X ∞
X    
g∆ (d)Ld,∆ (X) = g∆ (d)Md,∆ (X) + g∆ (d) 2Γ(3)Id,∆(X) + 2Jd,∆(X) + O X 10/3 (28)
d=1 d=1 d=1

so with (25) we get Lemma 1 if we prove (26), (27) and

∞ Z
X ζ(s)V∆ (s)X s+4 ds
g∆ (d)Id,∆ (X) = (29)
(−1/2) s(s + 2)(s + 3)(s + 4)
d=1

∞ Z
X ζ(s)V∆ (s)X s+4 ds
g∆ (d)Jd,∆ (X) = (30)
(0) (s + 2)(s + 3)(s + 4)
d=1

∞
X
∆3 I(∆)g∆ (d)Md,∆ (X/∆) = M (X). (31)
∆,d=1

3.1 Proof of (26)

Recall that the Gamma function is holomorphic for σ > −1/2 except for a simple pole with residue 1 at
s = 0 and that Lχ0 (s + 1) has a simple pole at s = 0.

For a squarefree character of order d, we know that Lχ0 (s) has a zero at s = 0 of order at least 2 if d
has more than one prime factor, and a simple zero if d = p (and no zero if d = 1), so (recalling the definitions
in (16))
 
Γ(s) + Γ(s + 1) s
Ress=0 Lχ0 (s)Lχ0 (s + 1)Rd,∆ (s) X
Γ(s + 5)
 
ζ(s)ζ(s + 1)R1,∆ (s)Γ(s)X s ζ(0)R1,∆ (0)
= Ress=0 +
Γ(s + 5) Γ(5)
| {z }
d=1
Rd,∆ (0)
+ Ress=0 {Lχ0 (s)Lχ0 (s + 1)Γ(s)}
Γ(5)
| {z }
d=p
 
ζ(0)Q∆ (0) ′
ζ (0) R′1,∆ (0) Γ′ (5) ζ(0) log d Qd∆ (0)
= + − + log X + 1 +
Γ(5) ζ(0) R1,∆ (0) Γ(5) Γ(5)
| {z } | {z }
d=1 d=p
∗ ∗∗
=: C∆ (X) + Cd,∆ . (32)
| {z } | {z }
d=1 d=p

9
Write Γ∆ (l) = Γ∆ (l)Γ∆ (l′ ). Splitting the sum according to the value of e = (l, d) and using Γ∆ (el) =
Γ∆ (e)Γ∆e (l) we have
X X X
(X − ˜
l)2 lΓ∆ (l) = e3 Γ∆ (e)2 (X/e − ˜l)2 lΓ∆e (l)
l+l′ ≤X e|d l+l′ l≤X/e
l̃≡0(d) l̃≡0(d/e)
(l,d/e)=1

X e3 Γ∆ (e)2 X X
= χ(−1) (X/e − l̃)2 lΓ∆e (l)χ(l)χ(l′ )
φ(d/e)
e|d χ(d/e) l+l′ ≤X/e
X e3 Γ∆ (e)2 X χ(d/e)
=: χ(−1)L∆e (X/e); (33)
φ(d/e)
e|d χ(d/e)

here and throughout a “χ(q)” in the superscript is just to remind us of the character’s modulus. With the
usual Perron type representations and recalling (14)
Z Z χ(d) χ(d)
χ(d) G∆ (s)G∆ (w)Γ(s)Γ(w + 1)X s+w+3 dwds χ(d)
L∆ (X) = Γ(3) =: Γ(3)D∆ (X) (34)
(1) (1) Γ(s + w + 4)
then, dropping the indices on G to avoid clutter and reading G(1) = 0 for χ 6= χ0 (recall (15)),
Z Z !
χ(d) G(1)X s+4 G (w)Γ(w + 1)X s+w+3 dw
D∆ (X) = G (s)Γ(s) + ds
(1) Γ(s + 5) (0) Γ(s + w + 4)
Z
G(1)2 X 5 G (s)Γ(s)X s+4 ds
= + G(1)
Γ(6) (0) Γ(s + 5)
Z Z !
G(1)X w+4 G (s)Γ(s)X s+w+3 ds
+ G (w)Γ(w + 1) + dw
(0) Γ(w + 5) (0) Γ(s + w + 4)
Z
G(1)2 X 5 Γ(s) + Γ(s + 1) s+4
= + G(1) G (s) X ds
Γ(6) (0) Γ(s + 5)
| {z }
=:g(s)
Z Z
G (s)G (w)Γ(s)Γ(w + 1)X s+w+3 dwds
+ . (35)
(0) (0) Γ(s + w + 4)
Remembering (16) in (32) the second term is
Z !
n o
χ (d) χ (d)
G(1) Ress=0 G∆0 (s)g(s)X s + G∆0 (s)g(s)X s ds X4
(−1/2)
 
 ∗ ∗∗ χ (d)
+T∆0 (X) X 4

=: G(1) C∆ (X) + Cd,∆ (36)
| {z } | {z }
d=1 d=p

whilst when summed over χ the third term is (not forgetting (18) and reading G(0) = 0 for χ 6= χ0 )
!
X Z G(0)X s+3
Z
G (w)Γ(w + 1)X s+w+3 dw
≤ G (s)Γ(s) + ds
χ (0) Γ(s + 4) (−1/2) Γ(s + w + 4)
  Z
G (s)Γ(s)X s+3 G (s)Γ(s)X s+3 ds
≤ G(0)Ress=0 + G(0)
Γ(s + 4) (−1/2) Γ(s + 4)
X Z  
G(s)Γ(s)X s+w+3 ds
+ G(w)Γ(w + 1) Ress=0 dw
χ (−1/2) Γ(s + w + 4)
Z Z
G (s)G (w)Γ(s)Γ(w + 1)X s+w+3 dwds
+
(−1/2) (−1/2) Γ(s + w + 4)
Z
Γ(s)
≪ Lχ0 (0)2 X 3 log X + Lχ0 (0)Lχ0 (s)Lχ0 (s + 1) X s+3 ds
(−1/2) Γ(s + 4)
XZ Z
Γ(s)Γ(w + 1) s+w+3
+ Lχ (s)Lχ (s + 1)Lχ (w)Lχ (w + 1) X dwds
χ (−1/2) (−1/2) Γ(s + w + 4)
 
≪ X ǫ X 3 + d3/2 X 5/2 + d3 X 2

10
(here we have comfortably bounded the double integral with a well-known mean value estimate) so from
(22), (33) and (34)
 
!
χ (d/e)
X e3 Γ∆ (e)2 G∆e0 (1)2 (X/e)5 χ (d/e)  ∗ χ (d/e) 
Ld,∆ (X) = 2Γ(3) + G∆e0 (1) 
C (X/e) + C ∗∗ +T∆e0 (X/e)
 (X/e)
4
φ(d/e) Γ(6) | ∆e{z } | d/e,∆e
{z }
e|d
d/e=1 d/e=p
 
X e3 Γ∆ (e)2  
+ O X ǫ (X/e)3 + (d/e)3/2 (X/e)5/2 + (d/e)3 (X/e)2 
φ(d/e)
e|d

and (26) follows from the definitions in (16), (19), (32), (35), (36) and (21).

3.2 Proof of (27)

Recall the definition of g∆ from just before (21) and of Ld,∆ (X) from (22). As

X ˜l
1= +O(1)
[D, D′ ]
l+l′ =l̃ | {z }
D|l
(D,D′ )|l̃

we have
X X X
Ld,∆ (X) = g∆ (D) (X − ˜l)2 ˜l 1
D l+l′ ≤X l+l′ =l̃
l̃≡0(d) D|l
 
X X g∆ (D) X4
= (X − ˜l)2 ˜l2 +O
D
[D, D′ ] d
l̃≤X
l̃≡0(d) (D,D′ )|l̃

and then after a straightforward calculation the D-sum is (C8 is defined in 11 and U (∆) in (14))
 
Y 
2g∆(p) Y  g∆ (p)2
1+ 1+    = C8 U (∆)K∆ (l̃)
p
p p 1 + 2g∆ (p)
p|l̃ p

from (21) so, with Kd (s) as in (14), a standard Perron representation gives
Z  4
Kd (s)X s+4 ds X
Ld,∆ (X) = 2C8 U (∆) +O
(1) (s + 2)(s + 3)(s + 4) d
5 Z  4
C8 U (∆)Kd (1)X Kd (s)X s+4 ds X
= + 2C8 U (∆) +O
30 (−1/2) (s + 2)(s + 3)(s + 4) d

after (17) and this becomes (27) after (19).

3.3 Proofs of (29), (30) and (31)

With Γ∆ as in (11) define
Y Y X ∆s (d/e)Γ∆ (e)2
qs (n) := qs (p) := (1 + rY ) Lχ0 (d) (s) =: ζ(s)∆s (d) τ (d) := (37)
es
p|n p|n e|d

so that from (16)

Y
Q∆ (s) = qs (p) σ>0
p∤∆
Y Y
Q∆ (s) = ζ(s + 1) qs (p)(1 − Y /p) (1 − Y /p) σ > −1 Y := p−s
p∤∆ p|∆

Q∆ (s)
Qd∆ (s) = (38)
qs (d)

11
and

τ (p) = 1 + rY (2 + r) for p ∤ ∆. (39)

Let Vs (p), V∆ (s) be as in (23), (24) so that (37) gives

r(1 + r)2
V0 (p) = q1 (p)q0 (p) + . (40)
p

From (38) and (39)

 
rτ (p) r r + r2 Y (2 + r)
q1 (p)qs (p) + = 1+ (1 + rY ) + = Vs (p) for p ∤ ∆
p p p

so from (38) and (21)

∞ ∞
X X Γ∆ (e)2 X g∆ (d)τ (d)
g∆ (d)Qd∆ (1)Qd∆ (s) Lχ (d/e) (1)Lχ0 (d/e) (s) = Q∆ (1)Q∆ (s)ζ(s)
φ(d/e)es+1 0 dq1 (d)qs (d)
d=1 e|d d=1
 
Y rτ (p)
(initially only for σ > 0) = ζ(s) q1 (p)qs (p) +
p
p∤∆

= ζ(s)V∆ (s) (41)

so since (Id,∆ (X) is defined in (19))

∞ Z
X X s+4  
g∆ (d)Id,∆ (X) = LHS of (41) ds (42)
(−1/2) s(s + 2)(s + 3)(s + 4)
d=1

we get (29). Writing

r2 (1 + r)U (p)
l=r+
p
we have from (11) and (14)

l if p ∤ ∆
K∆ (p)Γ∆ (p) − 1 =
0 if p|∆

so from (14) and the definition of g∆ (this was just before (21))
∞ ∞
X X K∆ (l)Γ∆ (l) Y Y 1
g∆ (d)Kd,∆ (s) = = ζ(s) (1 + lY )
ls p
1 + lY
d=1 l=1 p|∆

and (with Vs (p) as in (23))

1 + lY
= Vs (p)
U (p)

so (with C8 as in (11) and V∆ (s) as in (24))

∞
X
C8 U (∆) g∆ (d)Kd,∆ (s) = ζ(s)V∆ (s)
d=1

so from (19) we get

∞ Z
X ζ(s)V∆ (s)X s+4 ds
g∆ (d)Jd,∆ (X) =
(−1/2) (s + 2)(s + 3)(s + 4)
d=1

which is (30). Recall the definitions of A(d), B∆ (d) from (19). Putting s = 1 in (41) and using (25) gives
∞
X I(∆)g∆ (d)A∆ (d) 2Γ(3)C8
= U(1) (43)
∆2 Γ(6)
∆,d=1

12
whilst with (38) and (21)
∞  
X g∆ (d)Qd∆ (1)Qd∆ (0) Y Y r(1 + r)2 e−s log p
Γ∆ (d)2 e−s log d = (1 − 1/p) (1 − 1/p) q1 (p)q0 (p) + (44)
d p
d=1 p|∆ p∤∆

so that in particular we get from (40) (and definitions (19) and (24))
∞
X g∆ (d)Qd∆ (1)Qd∆ (0)
Γ∆ (d)2 = V∆ (0). (45)
d
d=1

so from (25)
∞
X I(∆)g∆ (d)B∆ (d)
= cC8 U(0) (46)
∆
∆,d=1

Now define
r 1 1
θ1 (p) := −1+ θ2 (p) :=
1+r (1 + r)2 (1 + r)2
X X
A1 (n) := θ1 (p) log p A2 (n) := θ2 (p) log p
p|n p|n

1 X I(∆)
M1 (n) := I(∆)g∆ (d)Γ∆ (d)2 M2 (∆) :=
nq1 (n)q0 (n) ∆V0 (∆)
d∆=n
M1∗ (n) := cQ1 (1)Q1 (0)M1 (n) M2∗ (∆) := cV∆ (0)M2 (∆)
X∞ X∞
A1 (s) := M1∗ (n)esA1 (n) A2 (s) := M2∗ (n)esA1 (n)
n=1 n=1

and recall the values of I, Γ∆ , g∆ , qs from (11), (21), (37). We have for p ∤ ∆

M1 (p) M2 (p) I(p) + g1 (p)Γ1 (p)2 I

− = 2
−
1 + M1 (p) 1 + M2 (p) p(1 + r)(1 + r/p) + I(p) + g1 (p)Γ1 (p) Q+G
| {z }
G
=: Q+G

and therefore, writing R = 1/(1 + r),

M1 (p)θ1 (p) M2 (p)θ2 (p) RG R2 (G − I) r(1 + 2r)

− = − + =
1 + M1 (p) 1 + M2 (p) Q+G Q+G (1 + r)(p + pr + r)
so
1 r M1 (p)θ1 (p) M2 (p)θ2 (p) p(1 − pr + 3r) 1
− + − = =
p−1 1+r 1 + M1 (p) 1 + M2 (p) (p − 1)(p + pr + r) p(p − 1)

from (11). Also A2 (0) = A1 (0) = cC8 U(0) from (45) and (25). From (16)

R′1,d∆ (0) X 1 r
 X r X r
= − log p + log p =: S + log p for (d, ∆) = 1
R1,d∆ (0) p
p−1 1+r 1+r 1+r
p|d∆ p|d∆

13
so from (19) (and not forgetting that (11) and (21) cause the d∆, n variables below to be all squarefree)
∞
!
X I(∆)g∆ (d) ∗
− B∆ (d) log(d∆) + B∆ (d)Zd∆ + B∆ (d)
∆
∆,d=1
∞
X I(∆)g∆ (d)Qd∆ (1)Qd∆ (0)
= Γ∆ (d)2
d∆
∆,d=1
 
 X r log p X log p X log p 
× Z + S + − log(d∆) + 2
− 
 1+r | {z }
P
(1 + r) (1 + r)2 
p|d∆ p|d∆ p|∆
= p|d∆ log p
∞
X ∞
X ∞
X
= (Z + S) M1∗ (n) + M1∗ (n)A1 (n) − V1 (0) M2∗ (∆)A2 (∆) (from (38))
n=1 n=1 ∆=1
= (Z + S)A1 (0) + A′1 (0) − A′2 (0)
) !
X 1 r M1 (p)θ1 (p) M2 (p)θ2 (p)
= cC8 U(0) Z + − + − log p
p
p−1 1+r 1 + M1 (p) 1 + M2 (p)
!
X log p
= cC8 U(0) Z + .
p
p(p − 1)

which with (43) and (46) is (31).

References
[1] C. Hooley - On the Barban-Davenport-Halberstam theorem I - Journal für die reine und angewandte
Mathematik, 274/275 (1975)
[2] C. Hooley - On the Barban-Davenport-Halberstam theorem VIII - Journal für die reine und angewandte
Mathematik, 499 (1998)

Tomos Parry
Bilkent University, Ankara, Turkey
[email protected]

14