A MODIFICATION OF THE LINEAR SIEVE, AND THE

COUNT OF TWIN PRIMES

JARED DUKER LICHTMAN

Abstract. We introduce a modification of the linear sieve whose weights satisfy strong
arXiv:2109.02851v1 [math.NT] 7 Sep 2021

factorization properties, and consequently equidistribute primes up to size x in arithmetic

progressions to moduli up to x10/17 . This surpasses the level of distribution x4/7 with the
linear sieve weights from well-known work of Bombieri, Friedlander, and Iwaniec, and which
was recently extended to x7/12 by Maynard. As an application, we obtain a new upper
bound on the count of twin primes, improving a 2004 result of Wu by about 3%.

1. Introduction
Given a finite set A of positive integers, sieve methods offer a broad framework for es-
timating the number of elements in A all whose prime factors exceed z, denoted S(A, z),
in terms of the approximateQdensity g(d)
of multiples of d in A, denoted Ad . Note one
often expects S(A, z) ≈ |A| p<z 1 − g(p) . Combinatorial sieves may be viewed as refine-
ments of the basic inclusion-exclusion principle, and are described by a sequence of weights
λ(d) ∈ {−1, 0, 1} supported on integers up to some level D >1. In particular, the upper
bound weights λ+ (d) for the linear sieve satisfy
Y
 log D
 X

(1.1) S(A, z) 6 |A| 1 − g(p) F log z + o(1) + λ+ (d) |Ad | − |A|g(d)
p<z d 6D
p|d⇒p<z

as D → ∞, under some mild conditions. Here the function F : R >1 → R >1 is defined by a
delay-differential equation, as in (2.5). For sets A sufficiently equidistributed in arithmetic
progressions the secondQsum over d 6D
in (1.1) contributes negligibly, in which case the main
term is S(A, z) . A| p<z 1 − g(p) F (s), where z = D 1/s . In fact, F (s) → 1 as s → ∞
so the main term confirms the naı̈ve expectation in this case. Moreover, F is optimal in the
sense that the bound (1.1) is attained sharply for a particular set A.
The linear sieve is powerful when combined with equidistribution estimates which make
the final sum in (1.1) small. For example, the Bombieri–Vinogradov Theorem shows that
1
for every ε, A > 0, letting Q = x 2 −ε we have
X π(x) x
(1.2) sup π(x; q, a) − ≪ε,A A
.
q 6Q (a,q)=1 ϕ(q) (log x)

So by taking D = Q, (1.1) can give a good upper bound when the set A is related to the
primes, such as when A = {p + 2 : p 6x}, in which case (1.1) gives an upper bound for the
count of twin primes.

Date: September 7, 2021.

2010 Mathematics Subject Classification. Primary 11N35, 11N36; Secondary 11N05.
1
 The estimate (1.2) may be viewed as an assertion of the Generalized Riemann Hypothesis
1
on average over moduli up to Q = x 2 −ε . It remains an important open problem to extend
1
the range to Q = x 2 +δ for some fixed δ > 0. Indeed, Elliott and Halberstam [6] conjectured
such an extension up to Q = x1−ε for any ε > 0.
In some contexts it suffices to relax the setup in (1.2) in order to raise the level of dis-
tribution. In particular, in the case of a fixed residue class a ∈ Z, and the absolute values
replaced by well-factorable weights λ(q) (c.f. Def. 2.1), the celebrated result of Bombieri–
4
Friedlander–Iwaniec [2] raised the level up to Q = x 7 −ε ,
X  π(x)  x
(1.3) λ(q) π(x; q, a) − ≪a,A,ε A
.
q 6Q
ϕ(q) (log x)
(q,a)=1

While the linear sieve weights are not themselves well-factorable, Iwaniec [13] constructed
a well-factorable variant e
λ+ of the weights λ+ (and so (1.3) holds with λ = λ e+ ), which are
+
only slightly altered from λ so that λ e enjoys an analogous linear sieve bound as in (1.1),
+

notably with an identical form of the main term,

Y
 log D
 X

(1.4) S(A, z) 6 |A| 1 − g(p) F log z + o(1) + e+ (d) |Ad | − |A|g(d) .
λ
p<z d 6D
p|d⇒p<z

The bound (1.3) stood for several decades, but quite recently Maynard [15] managed to
7
e+ . Given the
extend the level in (1.3) further to Q = x 12 −ε in the case of the weights λ = λ
7
currently available equidistribution estimates for primes, we note the level x 12 is a natural
barrier for these weights.
In this paper, we modify the technical construction of the linear sieve weights to avoid
this barrier, and thereby produce new sieve weights that induce stronger equidistribution
estimates for primes.
10
Theorem 1.1. Let D = x 17 −ε . There exists a sequence λe∗ (d) ∈ {−1, 0, 1} satisfying:
(1) Equidistribution for primes: for any fixed a ∈ Z, A, ε > 0, we have
X  
e∗ (d) π(x; d, a) − π(x) ≪a,A,ε
λ
x
.
ϕ(d) (log x) A
d 6D
(d,a)=1

(2) Sieve upper bound: for s >1, z = D 1/s , we have

Y

X

S(A, z) 6 |A| 1 − g(p) F ∗ (s) + o(1) + e∗ (d) |Ad | − |A|g(d) ,
λ
p<z d 6D
p|d⇒p<z

where F ∗ (s) 6 1.000081 F (s) when 1 6s 63, for the linear sieve function F as in (2.5).
10
The key feature of Theorem 1.1 is to obtain equidistribution up to level x 17 at the cost of
only a tiny loss in the main term. See Theorem 2.11 and Proposition 5.4 for full technical
statements and additional variations that may be of independent interest.
2
1.1. Application to twin primes. We expect that Theorem 1.1 should give numerous
improvements to sieve bounds related to the primes. As proof of concept in this direction,
we give a new upper bound for the count of twin primes up to x, denoted π2 (x). Recall
Hardy and Littlewood [12] conjectured the asymptotic formula
2x Y 1 − 2/p
(1.5) π2 (x) ∼ =: Π(x).
(log x)2 p>2 (1 − 1/p)2

Theorem 1.2. As x tends to infinity, we have

π2 (x) . 3.29956 Π(x).
Theorem 1.2 gives a 2.94% refinement from the previous record bound of Wu [19]. For
reference, this gives the largest percentage improvement since the work of Bombieri, Friedlan-
der, and Iwaniec [2]. See below for a chronology of the known upper bounds on π2 (x)/Π(x).
Also see Riesel–Vaughan [16, Lemma 5] for a numerically explicit form of Selberg’s bound
[17].
Year Author(s) π2 (x)/Π(x) .
1947 Selberg [17] 8
1964 Pan [11] 6
1966 Bombieri–Davenport [1] 4
1978 Chen [5] 3.9171
1983 Fouvry–Iwaniec [8] 3.7777 · · · = 34/9
1984 Fouvry [7] 3.7647 · · · = 64/17
1986 Bombieri–Friedlander–Iwaniec [2] 3.5
1986 Fouvry–Grupp [9] 3.454
1990 Wu [18] 3.418
2003 Cai–Lu [3] 3.406
2004 Wu [19] 3.39951
The main ingredients for these results come from applying sieve bounds to the set A =
{p + 2 : p 6x}, and using equidistribution of primes in arithmetic progressions to handle
remainder terms. Bombieri–Davenport obtained π2 (x)/Π(x) . 4 as a consequence of the
1
Bombieri–Vinogradov theorem (1.2) and a standard sieve upper bound of level x 2 −ε . More
generally, if one proves level of distribution xθ−ε then one immediately obtains π2 (x)/Π(x) .
2/θ. Bombieri–Friedlander–Iwaniec proved π2 (x)/Π(x) . 7/2 by the well-factorable variant
4
(1.3) level of distribution x 7 −ε , together with the linear sieve with well-factorable remainder
(1.4).
The other key ingredient to subsequent improvements is the switching principle, introduced
in Chen’s celebrated result that there are infinitely many primes p such that p + 2 has at
most two prime factors [4]. The basic insight is to use a weighted sieve inequality to split
the problem into multiple cases, apply sieve bounds to A = {p + 2 : p 6x} in certain
cases, and then reinterpret the remaining cases as new sieving problems for switched sets
B = {m − 2 6x} where the numbers m are constructed from A (as prescribed depending on
the case).
1.2. Outline of main ideas in Theorem 1.1. Maynard’s new equidistribution results
show equidistribution of the primes with sieve weights λ(d) provided d = p1 · · · pr is restricted
3
to suitably well-factorable integers. Unfortunately, the original linear sieve weights only
partially satisfy these well-factorable conditions. In particular for η > 0, when looking at
7
the linear sieve of level x 12 +η , some integers d in its support do not satisfy the conditions,
7
which means that x 12 is the limit for the linear sieve given our current equidistribution
technology. Nevertheless, the key observation here is that only a few exceptional d fail
10
to satisfy these conditions. Moreover up to level x 17 , i.e. η < 1/204, the anatomy of
exceptional d may be precisely characterized in terms of η (given specifically as P4 , P6 in
(3.6)). In particular, as η > 0 grows the family of exceptional integers contribute O(η 5) to
the sieve bound. However, we note this characterization breaks down when η >1/204, and
the contribution becomes considerably larger and more complicated.
As such we carefully revise the construction of the linear sieve, altering a few particular
inclusion-exclusion steps in order to avoid the exceptional integers d with bad factorizations.
Once these terms no longer contribute to the sieve, this produces a worse and more compli-
cated main term, but since there are only a very small number of such terms the resulting
loss is small. And since these modified weights now satisfy stronger factorization properties
in their support, we can now leverage the full strength of Maynard’s equidistribution results.

2. Technical setup and results

2.1. Factorization of weights and their level of distribution.
Definition 2.1 (well-factorable). Let Q ∈ R >1 . A sequence λ(q) is well-factorable of
level Q, if for every factorization Q = Q1 Q2 into Q1 , Q2 ∈ R >1 , there exist sequences γ1 , γ2
such that
(1) |γ1(q1 )|, |γ2(q2 )| 61 for all q1 , q2 ∈ N,
(2) γi (q) = 0 if q ∈ / [1, Qi ] for i = 1, 2,
(3) We have λ = γ1 ∗ γ2 , i.e.,
X
λ(q) = γ1 (q1 )γ2 (q2 ).
q=q1 q2

4
In [2, Theorem 10], Bombieri–Friedlander–Iwaniec established level of distribution x 7 −ε
with well-factorable weights.
Theorem 2.2 (Bombieri–Friedlander–Iwaniec [2]). Fix any a ∈ Z and let A, ε > 0. For any
4
well-factorable sequence λ of level Q 6x 7 −ε , we have
X  π(x)  x
λ(q) π(x; q, a) − ≪a,A,ε A
.
q 6Q
ϕ(q) (log x)
(q,a)=1

Maynard [15] considered a natural strengthening of well-factorable sequences.

Definition 2.3 (triply well-factorable). Let Q ∈ R >1 . A sequence λ(q) is triply well-
factorable of level Q, if for every factorization Q = Q1 Q2 Q3 into Q1 , Q2 , Q3 ∈ R >1 , there
exist sequences γ1 , γ2, γ3 such that
(1) |γ1(q1 )|, |γ2(q2 )|, |γ3(q3 )| 61 for all q1 , q2 , q3 ∈ N,
(2) γi (q) = 0 if q ∈ / [1, Qi ] for i = 1, 2, 3,
4
 (3) We have λ = γ1 ∗ γ2 ∗ γ3 , i.e.,
X
λ(q) = γ1 (q1 )γ2 (q2 )γ3 (q3 ).
q=q1 q2 q3

The definitions of well-factorable and triply well-factorable sequences are quite natural
and relatively simple from a conceptual standpoint. In [15, Theorem 1.1], Maynard obtains
powerful equidistribution results for triply well-factorability that are beyond the scope of
well-factorability. Unfortunately, triply well-factorability is too restrictive a condition for
us to produce Theorem 1.1. As such we are forced to identify the precise mechanism that
enables Maynard’s equidistribution results, and extract the following technical definition
that is implicit in [15].1
Definition 2.4 (programmably factorable). Let 0 < δ < 10−5 . For x ∈ R>1 , a sequence λ(q)
1
is programmably factorable of level Q (relative to x, δ), if for every N ∈ [x2δ , x 3 +δ/2 ]
there exists a factorization Q = Q1 Q2 Q3 with Q1 , Q2 , Q3 ∈ R >1 , satisfying the system
Q1 6 Nx−δ ,
(2.1) N 2 Q2 Q23 6 x1−δ ,
N 2 Q1 Q42 Q33 6 x2−δ ,
NQ1 Q52 Q23 6 x2−δ .
And for every such factorization Q = Q1 Q2 Q3 there exist sequences γ1 , γ2 , γ3 such that
(1) |γ1(q1 )|, |γ2(q2 )|, |γ3(q3 )| 61 for all q1 , q2 , q3 ∈ N,
(2) γi (q) = 0 if q ∈ / [1, Qi ] for i = 1, 2, 3,
(3) We have λ = γ1 ∗ γ2 ∗ γ3 , i.e.,
X
λ(q) = γ1 (q1 )γ2 (q2 )γ3 (q3 ).
q=q1 q2 q3

Programmable factorability is the key technical definition in this paper. It is named in

allusion to the linear programming-type system of inequalities (2.1) that the factors satisfy.
The diagram below displays the various implications among the definitions.

λ is triply well-factorable of level Q. =⇒ λ is well-factorable of level Q.

w
w


λ is programmably factorable of level Q (relative to x, δ).

1Indeed, the definition of programmably factorable in the special case Q3 = 1 gives the implicit condition
(which is implied by well-factorable) that enables Bombieri–Friedlander–Iwaniec to get equidistribution (1.3).
Also see Lemma 5 in [9].
5
 In the key result [15, Theorem 1.1], Maynard extended the level of distribution up to
3 3
Q < x 5 for programmably factorable weights. Note that level x 5 is the natural barrier for
(2.1) to admit a solution.
Theorem 2.5 (Maynard [15]). Fix any a ∈ Z and let A, ε > 0. For any programmably
3
factorable sequence λ of level Q 6x 5 −ε (relative to x, ε/50), we have
X  π(x)  x
λ(q) π(x; q, a) − ≪a,A,ε .
q 6Q
ϕ(q) (log x)A
(q,a)=1

Remark 2.6. [15, Theorem 1.1] was stated for triply-factorable sequences, but its proof in
fact gives the result for programmably factorable sequences.
Note the weights λe+ are composed of well-factorable—but not neccessarily programmably
factorable—sequences of given level D. Nevertheless, Maynard showed the upper bound
weights λe+ of sieve level D = x 127 −ε are programmably factorable of level D 6Q = x 35 −ε
(relative to x, ε/50). By Theorem 2.5 this gives [15, Theorem 1.2] below.
Corollary 2.7 (Maynard [15]). For any fixed a ∈ Z and A, ε > 0, the weights λ e+ from (2.4)
7
of sieve level D = x 12 −ε satisfy
X  π(x)  x
e +
λ (d) π(x; d, a) − ≪a,A,ε .
d 6D
ϕ(d) (log x)A
(d,a)=1

Later, in Proposition 5.4, we shall obtain technical improvements of Corollary 2.7 for
Iwaniec’s weights eλ± (both upper and lower), in special cases where equidistribution is
restricted to moduli which are smooth, or otherwise amenable to programmable factorization.
We may summarize the definitions and results of the section up to this point as follows:

λ is triply well-factorable of level Q. λ is well-factorable of level Q.

3 4
• equidistributed for Q < x .5 • equidistributed for Q < x 7 .
e+ for D 6 Q 32 .
• can take λ = λ e+ for D 6 Q.
• can take λ = λ

λ is programmably factorable of level Q (relative to x, δ).

3
• equidistributed for Q < x 5 .
e+ for D 6 Q < x 127 .
• can take λ = λ

For each type of sequence, we have outlined their corresponding levels of distribution, and
the levels at which the type is satisfied by the upper bound weights for the linear sieve. Ob-
serve well-factorability is flexible enough to accommodate the linear sieve to any level, but
has weaker equidistribution. On the other hand, triple well-factorability has stronger equidis-
tribution, but is too rigid to accommodate the linear sieve (at nontrivial levels). Finally,
programmable factorability also has strong equidistribution in addition to (nontrivially) ac-
commodating the linear sieve, though at the cost of conceptual technicality.
6
2.2. Sieve theory setup and bounds. We recall the standard sieve-theoretic notation.
Given a finite set A ⊂ N, set of primes P, and a threshold z > 0, we define Ad = {n ∈ A :
d | n} and remainder rA via
|Ad| = g(d)|A| + rA (d),
where g is a multiplicative Q < 1 for p ∈ P (we
function, with 0 6g(p) Q
assume g(p) = 0 if
p∈/ P). Also define P (z) = p<z,p∈P p and V (z) = p|P (z) 1 − g(p) . The central object of
interest is the sifted sum
X
(2.2) S(A, z) = S(A, P, z) = 1(n,P (z))=1 .
n∈A

Later for our application of interest, we will set g(d) = 1/ϕ(d). For now, it suffices for us
to assume for all 2 6w 6z,
Y   1 
V (w)
log z
(2.3) = 1 − g(p) = 1+O .
V (z) w 6p<z
log w log w
p∈P

Remark 2.8. The proof of the upper bound for the standard linear sieve only requires a
one-sided inequality for V (w)/V (z), whereas our modification requires the above two-sided
condition (2.3).
The basic result which we shall adapt is the linear sieve with well-factorable remainder,
as in [10, Theorem 12.20].
Theorem 2.9 (Iwaniec [10]). Let ε > 0 and D > 1 be sufficiently small and large, respec-
tively. Then for s >1 and z = D 1/s , we have

X
S(A, z) 6 |A|V (z) F (s) + O(ε) + e+ (d) rA (d),
λ
d|P (z)

X
S(A, z) > |A|V (z) f (s) + O(ε) − e− (d) rA (d),
λ
d|P (z)

e± are
where the implied constant only depends that of (2.3). Here the weights λ
X
(2.4) e± (d) =
λ λ±
j (d)
j 6 exp(ε−3 )

for some well-factorable sequences λ± +

j of level D. The functions F, f : R → R satisfy the
system of delay-differential equations

′
sF (s) = 2eγ [s 63] sF (s) = f (s − 1),

′
(2.5) sf (s) = 0 [s 62] sf (s) = F (s − 1).
Remark 2.10. See Iwaniec [13, Theorem 1] for an alternate formulation and proof, which
gives sharper quantitative bounds than O(ε). However, it is more technical than necessary
for our purposes.
The main result of this article is the following modification of the linear sieve with pro-
grammably factorable remainder.
7
Theorem 2.11. Let A be a finite set of positive integers with density function g(d) satisfying
(2.3), and F (s) the function defined by the system (2.5). Let ε > 0 and x > 1 be sufficiently
7
small and large, respectively. Then for η >0, D = x 12 +η , s >1, and z = D 1/s , we have

X
S(A, z) 6 |A|V (z) F ∗ (s) + O(ε) + e∗ (d) rA (d),
λ
d|P (z)

e∗ are
where the implied constant only depends that of (2.3). Here the weights λ
X
(2.6) e∗ (d) =
λ λ∗j (d)
j 6 exp(ε−3 )

1
for some programmably factorable sequences λ∗j of level D (relative to x, ε/50). For η < 204
1
we have F ∗ (s) = F (s) + O(η 5 ), and F ∗ (s) 6 1.000081 F (s) for 1 6s 63, η = 204 .
Note Theorem 2.5, applied to each λ = λ∗j above, immediately implies the following.

Corollary 2.12. Given any fixed a ∈ Z and A. For η < 204 1 e∗ as in (2.6) of
, the weights λ
7
level D = x 12 +η satisfy
X  
e∗ (d) π(x; d, a) − π(x) ≪a,A,η
λ
x
.
ϕ(d) (log x) A
d 6D
(d,a)=1

Acknowledgments
The author is grateful to James Maynard for suggesting the problem and for many valuable
discussions. The author also thanks Carl Pomerance and Kyle Pratt for helpful feedback.
The author is supported by a Clarendon Scholarship at the University of Oxford.

Notation
We use the Vinogradov ≪ and ≫ asymptotic notation, and the big oh O(·) and o(·)
asymptotic notation. We use f ∼ g, f . g, and f & g to denote f = (1 + o(1))g, f 6(1 +
o(1))g, and f >(1 + o(1))g, respectively. Dependence on a parameter will be denoted by a
subscript.
The letter p will always be reserved to denote a prime number, π(x) is the prime counting
function, and π(x; d, a) is the count of primes up to x congruent to a (mod d). We use ϕ
to denote the Euler totient function, µ the Möbius function, and e(x) := e2πix the complex
exponential. We use 1 to denote the indicator function of a statement. For example, for a
set A denote
( (
1, if a ∈ A, 1, if a1 , . . . , ai ∈ A and a0 ∈
/ A,
1a∈A = , 1aa10 ,...,a
∈A
/
i ∈A
=
0, else. 0, else.

3. Programmably factorable support

Recall the upper and lower bound weights λ± for the linear sieve of level D are defined by
λ± (d) = 1d∈D± µ(d),
8
where D ± = D ± (D) are the standard support sets
(3.1) D + (D) = {p1 · · · pr : D 1/2 >p1 > · · · >pr , and p1 · · · pl−1 p3l 6D for each odd l 6r},
D − (D) = {p1 · · · pr : D 1/2 >p1 > · · · >pr , and p1 · · · pl−1 p3l 6D for each even l 6r}.
Observe that both sets satisfy the containment D ± (D) ⊂ D well (D), where
(3.2) D well (D) = {p1 · · · pr : D 1/2 >p1 > · · · >pr , and p1 · · · pl−1 p2l 6D for each l 6r}.
We shall return to this observation later in the section.
In [15], Maynard deduces Corollary 2.7 for λ e+ from the general Theorem 2.5 by means of
the following key result [15, Proposition 9.1] (along with a construction of Iwaniec we shall
address in later sections), which programmably factorizes elements of the support D + .
7 1
Proposition 3.1 (Maynard [15]). Let 0 < δ < 10−3 and let D = x 12 −50δ , N ∈ [x2δ , x 3 +δ/2 ].
Then every d ∈ D + (D) has a factorization d = d1 d2 d3 such that d1 6 Nx−δ and
N 2 d2 d23 6 x1−δ ,
(3.3) N 2 d1 d42 d33 6 x2−δ ,
Nd1 d52 d23 6 x2−δ .
7
Remark 3.2. The level x 12 is sharp in this construction. Indeed, roughly speaking the linear
7
sieve weights are not programmably factorable of level D = x 12 +η for any η > 0, because the
support set contains obstructing (families of) elements d ∈ D + (D) of the form d = p1 · · · pr
1 2 1
where p1 ≈ · · · ≈ p6 ≈ D 7 , or where p1 ≈ p2 ≈ D 7 and p3 ≈ p4 ≈ D 7 . These obstructions
suggest how we should restrict the support set in order to increase the level.
7
For η > 0, level D = x 12 +η , we define the modified weights λ∗ = λ∗η ,
(3.4) λ∗ (d) = 1d∈D∗ µ(d),
for the support set D ∗ ,
7 7
(3.5) D ∗ = D + (x 12 ) ∪ {p1 · · · pr ∈ D + (x 12 +η ) : p1 · · · pi ∈
/ Pi , i 6r, i ∈ {4, 6}}.
7
Here P4 and P6 = P6,1 ∪ P6,2 are exceptional subsets of D + (x 12 +η ), given by
1 1
P4 = {p1 · · · p4 : p1 < x 6 +2η and p2 p4 > x 4 −3η },
1 1 1
(3.6) P6,1 = {p1 · · · p6 : p1 p2 < x 6 +2η and p2 p3 p4 > x 4 −3η and p6 > x 12 −5η },
1 1 1
P6,2 = {p1 · · · p6 : p1 , p2 p3 < x 6 +2η and p1 p4 , p2 p3 p4 > x 4 −3η and p6 > x 12 −5η }.
The modified support set D ∗ = Dη∗ is understood to depend on η > 0 (as do P4 , P6 ), but
we will suppress this for notational convenience.
In this section, we establish a programmable factorization of the elements of the support
10
∗ 1
D provided D < x 17 , i.e. η < 204 . This will serve as the key technical input for the proof
of Theorem 2.11.
Proposition 3.3 (Factorization of elements of D ∗ ). Let 0 < δ < 10−5, and take 0 < η <
1 7
1
204
− 3δ and A ∈ [xδ , x 3 −δ/2 ]. If d ∈ D ∗ for D = x 12 +η−50δ , then we may factor d = abc such
9
that a 6A and
A2 bc2 6 x1−3δ ,
(3.7) A2 ab4 c3 6 x2−3δ ,
Aab5 c2 6 x2−2δ .

For notational ease we are using A = Nx−δ . Also, on the first attempt working through
technicalities, we encourage the reader to set δ = 0 in order to better view the key features.
Before proving the proposition, we need some lemmas. The first gives a general-purpose
criterion to factor an integer d.
7 1 1
Lemma 3.4. Let D = x 12 +η−50δ for − 84 <η< 60
. A factorization d = abc satisfies (3.7),
provided a, b, c >1 satisfy
1 1
(3.8) a 6A, c 6D/Ab, b ∈ [x 6 +2η , x 4 −3η ].
Proof. By (3.8), Ac 6D/b and so
7 1
A2 bc2 6 D 2 /b 6x2( 12 +η−50δ)−( 6 +2η) = x1−3δ ,
7 1
A2 ab4 c3 6 D 3 b 6x3( 12 +η−50δ)+( 4 −3η) = x2−3δ ,
7 1 23
Aab5 c2 6 D 2 b3 6x2( 12 +η−50δ)+3( 4 −3η) = x 12 −7η−100δ < x2−2δ ,
1 1
using η ∈ (− 84 , 60 ). This gives (3.7). 
The above criterion implies factorizations in the following special cases.
7 1 1
Lemma 3.5. Let D = x 12 +η−50δ for η < 60 . For r >4, let x 6 +2η > p1 > · · · >pr be primes for
+
which d = p1 · · · pr ∈ D (D). Suppose b is one of the subproducts in {p1 p4 , p2 p3 , p2 p4 , p2 p3 p4 }.
Then d has a factorization d = abc satisfying (3.7), provided
1 1
b ∈ [x 6 +2η , x 4 −3η ].
Proof. Let C = D/Ab and note either p1 6A or p1 6C, since
2 2 7 1
p21 6D 3 = x 3 ( 12 +η−50δ) < x 3 +4η−50δ 6D/b = AC.
We proceed by induction on r >4. As the base case r = 4, by Lemma 3.4 it suffices for
each b to factor p1 · · · p4 /b = ac for a 6A, c 6C. Indeed, this holds when b = p2 p3 p4 since
p21 6AC, and similarly:
• If b = p2 p4 then p23 6D/p1 p2 p3 6AC/p1 implies p1 p3 = ac for some a 6A, c 6C.
• If b = p1 p4 then p23 6D/p1 p2 p3 6AC/p2 implies p2 p3 = ac for some a 6A, c 6C.
• If b = p2 p3 then p24 6D/p1 p2 p3 6AC/p1 implies p1 p4 = ac for some a 6A, c 6C.
Now for r >5, we inductively assume a factorization p1 · · · pr−1 = abc with a 6A,c 6C.
Then p2r 6D/p1 · · · pr−1 = AC/(ac) so either apr 6A or cpr 6C, extending the factorization.
Hence Lemma 3.4 applies again, and completes the proof. 
Finally, if the primes dividing d are small enough, we may use the greedy algorithm to
factor d as follows.
10
 7 1
1
Lemma 3.6. Let D = x 12 +η−50δ for η < 60 . For r >4, let x 6 +2η > p1 > · · · >pr be primes
1
for which d = p1 · · · pr ∈ D + (D), and p6 < x 12 −5η if r >6. Then d has a factorization d = abc
satisfying (3.7), provided there is a factorization p1 p2 p3 p4 = d1 d2 d3 satisfying
1
(3.9) d1 6A, d3 6x1−2δ /DA, d2 6D 2 /x = x 6 +2η .

Proof. Let D1 = A, D2 = D 2 /x, D3 = x1−2δ /(DA), so that di 6Di by assumption.

We now greedily append primes to di while preserving di 6Di for all i, i.e. where at the jth
step we replace di 7→ di pj (for one of i = 1, 2, 3) provided di pj 6Di . Starting from j = 5, we
stop either when we have exhausted all primes (i.e. j = r), or di pj > Di for each i = 1, 2, 3.
In the former case, we have d1 d2 d3 = d = p1 · · · pr and di 6Di so we easily get

D1 = A,
A2 D2 D32 = x1−4δ ,
3
A2 D1 D24 D33 = D 5 x−1−3δ 6x5· 5 −1−256δ < x2−3δ ,
3
AD1 D25 D32 = D 8 x−3−2δ 6x8· 5 −3−404δ < x2−2δ .

Thus letting a = d1 , b = d2 , c = d3 gives the desired factorisation.

In the latter case, there exists a terminal index j < r for which di pj > Di for all i = 1, 2, 3.
Note if j is odd, then di pj 6Di for some i, since

D D1 D2 D3
p3j 6 = .
p1 · · · pj−1 d1 d2 d3
1
So the terminal j is even with j >6. By assumption pj 6p6 6x 12 −5η is smaller than the width
1 1 1
of the interval [x 6 +2η , x 4 −3η ]. And since d2 6D2 = x 6 +2η < d2 pj , we deduce e2 := d2 pj lies
1 1
in the interval e2 ∈ [x 6 +2η , x 4 −3η ].
Thus letting E3 := D2 D3 /e2 , for each l > j in turn we shall greedily append the prime pl
onto either d1 or d3 while preserving d1 < D1 and d3 < D3 . Indeed, for all l > j,
D D 1 D 2 D3 D1 E3
p2l 6 = 6 ,
p1 · · · pl−1 d1 d2 d3 pj · · · pl−1 d1 d3 pj+1 · · · pl−1

so there is a factorization ac = d1 d3 pj+1 · · · pl with a 6D1 = A and c 6E3 = D/(Ae2 x2δ ).

Hence the result now follows by Lemma 3.4 with b = e2 . 

Proof of Proposition 3.3. We shall consider 3 cases, depending on the sizes of p1 and p2 p3
1 1
compared to the endpoints of the key interval [x 6 +2η , x 4 −3η ].
1
CASE 1: p1 >D 2 /x = x 6 +2η .
2
Let b := p1 , C := D/Ab. Note C = D/Ab >D 3 /A >1.
Next D >p31 >p1 p22 implies p22 6D/p1 = AC, so either p2 6A or p2 6C. Similarly, since
AC
p1 · · · pj−1p2j 6D for all j 6r, we get p2j 6 p2 ···p j−1
for 3 6j 6r. As such, we may factor
p2 · · · pr = ac for a 6A, c 6C. Hence by Lemma 3.4 p1 · · · pr = abc satisfies (3.7).
11
 1
In the remaining cases, we assume p1 < x 6 +2η . By Lemma 3.5, it remains to consider
1 1
p2 p3 > x 4 −3η or p2 p3 < x 6 +2η . Note
1 1 1 1 1 1 1 7 1
(3.10) p1 p3 6p13 (p1 p2 p33 ) 3 6(x 6 +2η ) 3 D 3 = x 3 ( 6 + 12 +3η−50η) < x 4 +η−16δ .

1 1
CASE 2: p2 p3 > x 4 −3η and p1 < x 6 +2η .
1 1
The proof follows by Lemma 3.5 if p2 p4 ∈ [x 6 +2η , x 4 −3η ]. Thus by definition of P4 , in this
case we may assume
1
(3.11) p2 p4 < x 6 +2η .
Hence we have p4 < x12η+50δ , since
1 1
(3.12) p2 > p2 (p1 p2 p33 )/D > (p2 p3 )3 /D > x3( 4 −3η) /D = x 6 −10η+50δ .
1 1
If p1 p4 > x 6 +2η , then the proof follows by Lemma 3.5 with b = p1 p4 < x( 6 +2η)+12η+50δ <
1 1
x 4 −3η , since η < 204 − 3δ.
1
Else p1 p4 < x 6 +2η . We shall apply Lemma 3.6 with d2 = p1 p4 .
1
If either A or x1−2δ /DA is greater than x 4 +η−16δ >p2 p3 , by (3.10), then Lemma 3.6 com-
pletes the proof with (d1 , d3 ) = (p2 p3 , 1) or (1, p2 p3 ), respectively. Otherwise, A, x1−2δ /DA ∈
1 1 5 1
[x 6 −2η−64δ , x 4 +η−16δ ], since x/D = x 12 −η+50δ . But then, using η < 108 ,
1 5/2 1
max(A, x1−2δ /DA) > (x1−2δ /D) 2 = x 12 −η/2+24δ > x 6 +2η > p2 ,
1 1
(3.13) min(A, x1−2δ /DA) > x 6 −2η−64δ > x 8 +η/2−8δ >(p1 p3 )1/2 >p3 ,
1
by (3.10), which suffices again for Lemma 3.6. Note p6 < x12η+50δ < x 12 −5η when r >6, using
1
η < 204 − 3δ.
1 1
CASE 3: p2 p3 < x 6 +2η and p1 < x 6 +2η .
1 1
By Lemma 3.5, it suffices to consider either p1 p4 < x 6 +2η or p1 p4 > x 4 −3η .
1
Subcase 3.1: p1 p4 < x 6 +2η .
1
Suppose we can show p6 < x 12 −5η (when r >6). Then since x1−2δ /D > D 4 /x2 , either A or
x1−2δ /DA is greater than D 2 /x. Thus Lemma 3.6 will complete the proof, with (d1 , d2 , d3 ) =
(p1 p4 , p2 p3 , 1) or (1, p2 p3 , p1 p4 ).
1
If p1 p3 > x 4 −3η then in this subcase
1 1 p1 p3 1
x 12 +η > (p2 p3 ) 2 >p3 = p4 > p4 x 12 −5η ,
p1 p4
1
1
so p4 < x6η . Hence p6 6p4 < x 12 −5η since η < 108 , which completes the proof.
1 1 1
Else p1 p3 < x 4 . By Lemma 3.5, it suffices p1 p3 < x 6 +2η . Then p3 > x 12 −5η implies
−3η
1 1
p1 < x 12 +7η . If further p1 p2 > x 4 −3η , then similarly
1 p1 p2 1
x 12 +7η > p1 >p2 = p4 > p4 x 12 −5η ,
p1 p4
1
1
so p3 < x12η . Hence p6 6p3 < x 12 −5η since η < 204
, which completes the proof.
12
 1 1
Else p1 p2 < x 4 −3η . By Lemma 3.5, we may assume p1 p2 < x 6 +2η .
1 1
Similarly, suppose p2 p3 p4 < x 4 −3η . By Lemma 3.5 we may assume p2 p3 p4 < x 6 +2η , and so
1 (2/3) 1
+(2/3)η
p6 6(p2 p3 p4 ) 3 6x 12 6x 12 −5η ,
1
using η < 204 , which completes the proof.
1 1
Thus we may assume p2 p3 p4 > x 4 −3η . But unless p6 < x 12 −5η , this subcase will contradict
the definition of P6,1 in (3.6), hence completing the proof.
1
Subcase 3.2: p1 p4 > x 4 −3η .
1
If d2 = p2 p3 p4 < x 6 +2η , then Lemma 3.6 completes the proof with (d1 , d3 ) = (p1 , 1) or
(1, p1 ), since
1 1 1 1
p6 6(p2 p3 p4 ) 3 6x 3 ( 6 +2η) 6x 12 −5η
1 1 1
for η < 204 . And if p2 p3 p4 ∈ [x 6 +2η , x 4 −3η ] the proof follows by Lemma 3.5.
1 1 1 1
Else p2 p3 p4 > x 4 −3η . Note p4 < x 12 +η and p1 = p1 p4 /p4 > x 6 −4η and p2 p3 p4 < x 4 +3η .
Also note we may factor p1 p4 = d1 d3 for d1 6A, d3 6x1−2δ /DA (Indeed this follows if A
1 1 1
or x1−2δ /DA exceeds x 4 +3η >p1 p4 . Else A, x1−2δ /DA ∈ [x 6 −4η−2δ , x 4 +3η ], which also works
1 1 1 5
1
similarly as with (3.13), since p4 < x 12 +η < x 6 −4η−2δ and p1 < x 6 +2η < x 12 −η−δ by η < 60 )
1
−5η
If further p6 > x 12 , then this subcase contradicts the definition of P6,2 in (3.6). Hence
1
we have p6 6x 12 −5η , and so by the above paragraph Lemma 3.6 completes the proof with
d 2 = p2 p3 .
Combining all cases completes the proof of Proposition 3.3. 
3.1. Refined factorization of D well . Proposition 3.3 (programmably) factorizes each d ∈
D ∗ ⊂ D + , and forms the key step to prove the weights λ∗ are programmably factorable. With
applications in mind to twin primes, we shall similarly (programmably) factorize certain
subsets of the well-factorable support D well , as in (3.2).
4 3
In the following result, we factorize d ∈ D well (D) for variable level D ∈ (x 7 , x 5 ), depending
on the anatomy of d. As D ± ⊂ D well , this has implications to both upper and lower bounds
for the standard linear sieve.
7
1 1
Proposition 3.7. Let D well (D) as in (3.2) for D = x 12 +η−50δ and − 84 < η < 60 − 30δ. Let
1
−3η well
x4 >p1 > · · · >pr be primes for which d = p1 · · · pr ∈ D (D). Then d has factorization
1
d = abc satisfying (3.7) if p3 6 x 12 −5η , or if
1 1
b ∈ [x 6 +2η , x 4 −3η ] with b | p1 p2 p3 , b 6= p3 .
1 1
Proof. For i = 1, 2, 3, suppose b = p1 · · · pi lies [x 6 +2η , x 4 −3η ], and let C = D/Ab. Since
p1 · · · pj−1p2j 6D for all i < j 6r, we get p2j 6 pi+1AC
···pj−1
for i < j 6r. As such, we may factor
pi+1 · · · pr = ac for a 6A, c 6C. Hence by Lemma 3.4 p1 · · · pr = abc satisfies (3.7).
1 1
Else, by assumption p1 < x 4 −3η so we may assume further p1 < x 6 +2η . In particular this
gives p21 6D/b. For the remaining b | p1 p2 p3 :
• If b = p2 p3 then p21 6D/b implies p1 6A or p1 6D/Ab.
• If b = p1 p3 then p22 6D/b implies p2 6A or p2 6D/Ab.
13
 • If b = p2 then p23 6D/p1 b implies a factorization p1 p3 = ac for a 6A, c 6D/Ab.
For each b above, we factored p1 p2 p3 = abc for a 6A, c 6D/Ab. Since p1 · · · pj−1 p2j 6D for
all j 6r, by induction we may factor p1 · · · pr = abc for a 6A, c 6D/Ab. By Lemma 3.4
p1 · · · pr = abc satisfies (3.7).
1 1 1
Finally, suppose p3 6 x 12 −5η is smaller than the width of the interval [x 6 +2η , x 4 −3η ]. Since
1 1
p1 < x 6 +2η , we have p1 p3 < x 4 −3η so by the above argument we may assume d2 := p1 p3 <
1
x 6 +2η . Then p32 6p1 p22 6D implies
2 7 x1−2δ
5
p22 6x 3 ( 12 +η) < x 12 −η+48δ =
,
D
1
since η < 60 − 30δ. Thus p2 6A or p2 6x/DA, so there is a factorization p1 p2 p3 = d1 d2 d3
satisfying (3.9). Hence the same greedy argument as in Lemma 3.6 completes the proof,
with p3 playing the role of p6 . 
Taking the maximum valid η as above, we may re-express the above factorization of level
7
θ
x , θ = 12 + η, as follows. Note the maximum θ for which t ∈ [ 16 + 2η, 14 − 3η] is given by
(
2−t
3
if t > 15 ,
(3.14) θ(t) = 1+t
2
if t 6 15 .
7 1
Similarly the maximum θ = 12
+ η for which t 6 12 − 5η is (3 − t)/5.
Corollary 3.8. Let p1 > · · · >pr be primes and write pi = xti . If d = p1 · · · pr ∈ D well (xθ−50δ ),
then there is a factorization d = abc satisfying (3.7) provided
θ 6 θ(t1 ),
for θ(t) as in (3.14). Moreover if t1 6 51 , then it suffices that
n3−t
3
(3.15) θ 6 θ(t1 , t2 , t3 ) := max , θ(t1 ), θ(t2 ), θ(t1 + t2 + t3 ),
5 o
θ(t1 + t2 ), θ(t1 + t3 ), θ(t2 + t3 ) .

4. Modification of the linear sieve

In this section we shall bound the modified linear sieve, analogous to the bounds for the
linear sieve (sometimes called the Jurkat–Richert theorem). This bound will form the basis
for our final result the next section, in which we modify the construction of Iwaniec’s weights.
1
Proposition 4.1. Let ε > 0 be sufficiently small. For η 6 204 , the modified weights λ∗ as in
7
(3.4) of level D = x 12 +η−ε satisfy
  X
∗ log D
S(A, z) 6 |A|V (z) F ( log z ) + o(1) + λ∗ (d) rA (d),
d|P (z)

where F ∗ = Fη∗ is a function satisfying F ∗ (s) = F (s) + O(η 5) for F as in (2.5).

Remark 4.2. It suffices for our purposes to obtain qualitative error o(1) in the factor ac-
companying F ∗ . Though as with the Jurkat–Richert theorem, with greater care one should
obtain a quantitative refinement, e.g. O((log D)−1/6 ). See (12.4)–(12.8) in [10].
14
 7 7
We now adapt the proof. Let D = x 12 +η and D0 = x 12 . For n >1, primes p1 > · · · >pn ,
/ D + (D) then there exists a minimal index l 6n such that p1 · · · pl ∈
if p1 · · · pn ∈ / D + (D).
3
By definition such minimal l is odd (explicitly, this occurs when p1 · · · pl−1 pl > D but
p1 · · · pm−1 p3m 6D for all odd m < l.) Similarly, if p1 · · · pn ∈ / D ∗ there exists a minimal index
∗
l 6n such that p1 · · · pl ∈ / D , which is also odd.
Indeed, to show this let l 6n be the minimal index such that p1 · · · pl ∈ / D ∗ . If (p1 , . . . , pj ) ∈
/
Pj for all j 6l, j ∈ {4, 6}, then clearly l > j must be odd, as with D + (D). On the other
hand, if (p1 , . . . , pj ) ∈ Pj for some j 6l, j ∈ {4, 6}, a priori one might expect l could be even.
6
However, the key point in this case is that p1 · · · pj ∈ D + (D0 ) ⊂ D ∗ (since p1 · · · pj ≈ D 7 by
definition of Pj ). Thus l > j is the minimal index such that p1 · · · pl ∈ / D + (D0 ), and hence
must be odd as claimed.
Using this minimal index, we show the following lemma.
Lemma 4.3. Let h be a multiplicative function with 0 6h(p) 61 for all primes p. Then we
have
Y X
(1 − h(p)) 6 λ∗ (d)h(d).
p|n d|n

Proof. Note if h = 1 identically, we interpret the product as 1n=1 . Now by definition,

X Y X X X
λ∗ (d)h(d) − (1 − h(p)) = µ(d)h(d) − µ(d)h(d) = µ(d)h(d).
d|n p|n d|n d|n d|n
d∈D ∗ / ∗
d∈D

/ D ∗ by its minimal index,

Then splitting up d ∈
X X X X
µ(d)h(d) = − h(p1 · · · pl ) µ(b)h(b) 6 0,
d|n odd l pl <···<p1 <z p1 ···pl b|n
/ ∗
d∈D p1 ···pl−1 ∈D ∗ b|P (pl )
/ ∗
p1 ···pl ∈D
Q
since h >0 and the inner sum on b factors as p|(P (pl ),n) (1 − h(p)) >0, since h(p) 61. 

By Lemma 4.3 with h(d) = 1, we have

X
(4.1) 1n=1 6 λ∗ (d),
d|n

in which case we obtain

X X X X
S(A, P, z) = 1(n,P (z))=1 6 λ∗ (d) = λ∗ (d)|Ad |
n∈A n∈A d|(n,P (z)) d|P (z)
X X
∗
(4.2) =X λ (d) g(d) + λ (d) rA(d) =: XV ∗ (D, z) + RA
∗ ∗
(D, z).
d|P (z) d|P (z)

Following Lemma 4.3 with h = g, we have the identity

X X X
(4.3) V ∗ (D, z) := µ(d)g(d) = V (z) + g(p1 · · · pn )V (pn ),
d|n odd n pn <···<p1 <z
/ ∗
d∈D p1 ···pn−1 ∈D ∗
/ ∗
p1 ···pn ∈D
15
and similarly
X X X
(4.4) V + (D, z) = V (z) + g(p1 · · · pn )V (pn ) =: V (z) + Vn (D, z).
odd n pn <···<p1 <z odd n
p1 ···pn−1 ∈D + (D)
/ + (D)
p1 ···pn ∈D

Then the difference of V ∗ and V + is

X X
(4.5) V ∗ (D, z) − V + (D, z) = g(p1 · · · pn )V (pn ) ∆,
odd n pn <···<p1 <z

where ∆ is the difference of indicator functions,

∗ / + (D)
p ···p ∈D p ···p ∈D + (D)\D ∗ / + (D)
p ···p ∈D
∆ := 1pp11 ···p n ∈D
/ 1 n 1 n
···pn−1 ∈D ∗ − 1p1 ···pn−1 ∈D + (D) = 1p1 ···pn−1 ∈D ∗
n
− 1p11 ···pn−1 ∈D + (D)\D ∗ ,

recalling D ∗ ⊂ D + (D). Note if a point is (p1 , .., p6 ) ∈ P6 then its projection is (p1 , .., p4 ) ∈
/ P4 .
∗ +
So by definitions of D , D (D) from (3.5), (3.1), for odd n we have the identities,
p ···p ∈D + (D)\D ∗
X p ···pn ∈D + (D)\D + (D0 )
(4.6) 1p11 ···pnn−1 ∈D∗ = 1(p1 ,..,pj )∈Pj · 1p11 ···pn−1 ∈D + (D0 ) ,
j∈{4,6}
j<n
/ + (D)
p ···p ∈D
X / + (D)
p ···p ∈D
n n
(4.7) 1p11 ···pn−1 ∈D + (D)\D ∗ = 1(p1 ,..,pj )∈Pj · 1p11 ···pn−1 ∈D + (D)\D + (D0 ) .
j∈{4,6}
j<n

We may plug (4.6) and (4.7) into ∆. In addition, we strategically add and subtract the
indicator function of {p1 · · · pn ∈ / D + (D), p1 · · · pn−1 ∈ D + (D0 )}, which together give
X  
p1 ···pn ∈D + (D)\D + (D0 ) / + (D)
p1 ···pn ∈D
∆ = 1(p1 ,..,pj )∈Pj 1p1 ···pn−1 ∈D+ (D0 ) − 1p1 ···pn−1 ∈D+ (D)\D+ (D0 )
j∈{4,6}
j<n
X
= 1(p1 ,..,pj )∈Pj
j∈{4,6}
j<n
 
p ···pn ∈D + (D)\D + (D0 ) / + (D)
p ···pn ∈D / + (D)
p ···pn ∈D p ···p ∈D / + (D)
· 1p11 ···pn−1 ∈D + (D0 ) + 1p11 ···pn−1 ∈D + (D0 ) − 1p11 ···pn−1 ∈D + (D)\D + (D0 ) − 1p11 ···pnn−1 ∈D+ (D0 )
X  
/ + (D0 )
p1 ···pn ∈D / + (D)
p1 ···pn ∈D
= 1(p1 ,..,pj )∈Pj 1p1 ···pn−1 ∈D+ (D0 ) − 1p1 ···pn−1 ∈D+ (D) .
j∈{4,6}
j<n

Thus plugging ∆ back into (4.5) gives

V ∗ (D, z) − V + (D, z) =
X X X X  
/ + (D0 )
p1 ···pn ∈D / + (D)
p1 ···pn ∈D
g(p1 · · · pn )V (pn ) 1p1 ···pn−1 ∈D+ (D0 ) − 1p1 ···pn−1 ∈D+ (D) .
j∈{4,6} pj <···<p1 <z odd n>j pn <···pj+1 <pj
(p1 ,..,pj )∈Pj
16
Recalling the definition of Vn (D, z) in (4.4), since g is multiplicative we have
X X X 


V ∗ (D, z) − V + (D, z) = g(p1 · · · pj ) Vn−j p1D···p
0
j
, p j − Vn−j D
p1 ···pj
, pj
j∈{4,6} (p1 ,..,pj )∈Pj odd n>j
X X 


(4.8) = g(p1 · · · pj ) V + D0
p1 ···pj
, pj −V+ D
p1 ···pj
, pj .
j∈{4,6} (p1 ,..,pj )∈Pj
P
as V + (D, z) − V + (D ′ , z) = odd n [Vn (D, z) − Vn (D ′ , z)].
Now assuming the two-sided condition (2.3) for g, the proof of [10, Theorem 11.12] (c.f.
(12.4)–(12.8)) gives asymptotic equality,
n
o
+ log D −1/6
(4.9) V (D, z) = V (z) F ( log z ) + O (log D) (z 6D),

so that (4.8) becomes

n
o
V ∗ (D, z) = V (z) F ( log D
log z
) + O (log D) −1/6

X X n
−1/6
o
log D0 /p1 ···pj log D/p ···p D
+ g(p1 · · · pj )V (pj ) F ( log pj
) − F ( log p1j j ) + O log p1 ···pj
.
j∈{4,6} (p1 ,..,pj )∈Pj

Hence by the prime number theorem,

n
o
(4.10) V ∗ (D, z) = V (z) F ∗ ( log D
log z
) + O (log D) −1/6
(z 6D),

where the function F ∗ satisfies

sF ∗ (s) − sF (s) =
X Z  
7 dx1 · · · dxj  127 −x1 −···xj  7
+η−x 1 −···x j
(4.11) ( 12 + η) · F − F 12 xj .
(x1 ,..,xj )∈Pj x1 · · · x2j xj
j∈{4,6}

Here P4 ,P6 = P6,1 ∪ P6,2 are the polytopes in Euclidean space induced from P4 , P6 ,
P4 = {(x1 , ..., x4 ) ∈ D+ ( 12
7
+ η) : x1 < 1
6
+ 2η and x2 + x4 > 1
4
− 3η},
(4.12) P6,1 = {(x1 , ..., x6 ) ∈ D+ ( 12
7
+ η) : x1 + x2 < 1
6
+ 2η and x6 > 1
12
− 5η
1
and x2 + x3 + x4 > 4
− 3η},
P6,2 = {(x1 , ..., x6 ) ∈ D+ ( 12
7
+ η) : x1 , x2 + x3 < 1
6
+ 2η and x6 > 1
12
− 5η
1
and x1 + x4 , x2 + x3 + x4 > 4
− 3η},
and D+ is the induced set from D + ,
D+ (τ ) = {(x1 , . . . , xr ) : x1 > · · · > xr > 0, and x1 + · · · xl−1 + 3xl < τ for each odd l 6r}.
Hence Proposition 4.1 follows.

4.1. Sieve function computation. We now compute F ∗ in terms of F .

1
Proposition 4.4. Let η = 204
. Then for 1 6 s 6 3, we have
(4.13) F ∗ (s) 6 1.000081 F (s).
17
Proof. From (4.11) we have


sF ∗ (s) = sF (s) + 7
12
+ η · 2eγ η(J4 + J6 ),
for integrals Jj , j ∈ {4, 6},
Z  
1 dx1 · · · dxj  127 −x1 −···xj  7
+η−x 1 −···x j
Jj := F − F 12 xj
2eγ η (x1 ,..,xj )∈Pj x1 · · · x2j xj
Z
dx1 · · · dxj h 7 7
i−1
= ( 12 − x1 − · · · xj )( 12 + η − x1 − · · · xj ) ,
(x1 ,..,xj )∈Pj x1 · · · xj

since sF (s) = 2eγ for s ∈ [1, 3]. In particular |Pj | = O(η j ) implies Jj = O(η j+1), and so
F ∗ (s) = F (s) + O(η 5).
1
For η = 204 , we compute directly that
(4.14) J4 6 0.016896.
Next we bound J6 . For (x1 , .., x6 ) ∈ P6 we have x4 < 21 (x2 + x3 ) < 12 1
+ η and 7
12
+η−
x1 − · · · x6 > x5 so
Z Z
dx1 dx2 dx3 dx4 dx5 dx6
J6 6 2
,
P6 x1 x2 x3 1
12
1
−5η<x6 <x5 <x4 < 12 +η x4 x5 x6 (x5 − η)

where P6 = P6,1 ∪ P6,2 is the (3-dimensional) projection of P6 , given explicitly by

P6,1 = {(x1 , x2 , x3 ) ∈ D+ ( 12
7
+ η) : x1 + x2 < 1
6
+ 2η and x3 > 1
12
− 5η
1
and x2 + 2x3 > 4
− 3η},
+ 7 1 1
P6,2 = {(x1 , x2 , x3 ) ∈ D ( 12 + η) : x1 , x2 + x3 < 6
+ 2η and x3 > 12
− 5η
1
and x1 + x3 , x2 + 2x3 > 4
− 3η}.
1
For η = 204
, we compute J6 6 (J6,1 + J6,2 )J6,0 where
Z
dx4 dx5 dx6
J6,0 = 2
6 2.33838,
1
−5η<x 6 <x 5 <x 4 < 1
+η x4 x5 x6 (x5 − η)
Z 12 12
Z
dx1 dx2 dx3 dx1 dx2 dx3
J6,1 = 6 0.000806853, and J6,2 = 6 0.00397946.
P6,1 x1 x2 x3 P6,2 x1 x2 x3
Hence combining with (4.14), for s ∈ [1, 3] we conclude
F ∗ (s) 7


6 1 + ( 12 + η) · η J4 + (J6,1 + J6,2 )J6,0 6 1.000081.
F (s)


5. Programmably factorable remainder, after Iwaniec

In Theorem 2.9, Iwaniec constructed a well-factorable variant λ e± of the weights λ± from the
(Jurkat–Richert) linear sieve. In this section, we prove Theorem 2.11 for the programmably
factorable variant e
λ∗ by adapting Iwaniec’s construction, similarly building on the Jurkat–
Richert type Proposition 4.1 that we obtained in the previous section. We shall also prove a
technical variation on this result, with a variable level depending on anatomy of the moduli.
To set up the construction, we first adapt [10, Proposition 12.18].
18
 7
Proposition 5.1. Let η > 0, and D = x( 12 +η)/(1+ε+τ ) for ε > 0 sufficiently small. Let D∗r
be defined by (5.8). Let λ(r) be the standard (upper and lower, for r odd and even, resp.)
2
weights for the linear sieve of level D ε . Then for u = D ε , τ = ε9 ,

S(A, z) 6 |A|V (z) F ∗ (s) + O(ε5 )
X X (−1)r X X
(5.1) + λ(r) (b) rA (bp1 · · · pr ).
−2 ∗
γ(D1 , . . . , Dr )
0 6r 6ε (D1 ,..,Dr )∈Dr p1 ···pr |P (z)/P (u) b|P (u)
ε
Dj <pj 6Dj1+τ b 6D

Proof. First we write

X
(5.2) S(A, z) 6 S ∗ (A, z) − Sn (A, z)
odd n 6N

for any N >1, where

X X X
S ∗ (A, z) := (−1)r |Ap1···pr |, Sn (A, z) := S(Ap1···pn , pn ).
0 6r 6ε−2 u 6pr <···p1 <z pn <···<p1 <z
p1 ···pr ∈D ∗ p1 ···pn−1 ∈D ∗
/ ∗
p1 ···pn ∈D

We apply the inequality (5.2), not for A = (an ) itself but rather for the subsequence à =
2
(an 1(n,P (u))=1 ). Here we take u = D ε , and then return to A by means of the Fundamental
Lemma.
Let z = D 1/s with 2 6s 6ε−1 . Since z > u, the only change to the above bound (5.2)
when passing to à is the term S ∗ (Ã, z), provided that N is not too large in terms of ε.
Specifically, we require the lower bound for pn (by induction, the linear sieve conditions
−m
imply p1 · · · pm < D 1−2 )
1−N /3
(5.3) pn >(D/p1 · · · pn−1 )1/3 >D 2
2
to be larger than u = D ε , which certainly holds provided
1 1
(5.4) N 6 log .
2 ε
Now it remains to evaluate S ∗ (Ã, z),
X X
(5.5) S ∗ (Ã, z) = (−1)r |Ãp1···pr |.
0 6r 6ε−2 u 6pr <···p1 <z
p1 ···pr ∈D ∗

For each r, we break the range in the inner sum into boxes. Namely, we let D1 , ..., Dr run
over numbers of form
2 (1+τ )j
(5.6) Dε , j = 0, 1, 2, . . .
√
with τ = ε9 . We denote by D+ r = D +
r (D) the set of r-tuples (D 1 , ..., D r ) with D r 6 · · · 6D 1 6 D,
such that
(
3
{(D1 , ..., Dr ) : D1 · · · Dm <D for all odd m 6r} if r even,
D+r = 3 1/(1+τ )
{(D1 , ..., Dr ) : D1 · · · Dm < D for all odd m 6r} if r odd.
19
We note, for ε > 0 sufficiently small, the cardinalities of the D+
r are bounded by
X

(5.7) |D+r | 6 exp ε
−3
.
0 6r 6ε−2
7
Hereafter let D = x( 12 +η)/(1+τ +ε) , and define
(5.8) D∗r = {(D1 , ..., Dr ) ∈ D+
r (D) : (D1 , ..., Di ) ∈
/ Pi,r for i 6r, i ∈ {4, 6}}.
where P4,r , P6,r are (τ -enlarged, for even r) analogues of the polytopes P4 , P6 in (3.6), e.g.
( 1 1
{(D1 , . . . , D4 ) : D11+τ < x 6 +2η and (D2 D4 )1/(1+τ ) > x 4 −3η } if r even,
P4,r = 1 1
{(D1 , . . . , D4 ) : D1 < x 6 +2η and D2 D4 > x 4 −3η } if r odd.

Observe each integer p1 · · · pr has aSunique vector (D1 , . . . , Dr ) such that pi ∈ (Di , Di1+τ ]
for all i 6r, inducing a map ν : N → r D+ r . As a convention ν(1) = () is the empty vector.
By construction, for even r if p1 · · · p4 ∈/ P4 then ν(p1 · · · p4 ) ∈
/ P4,r , and if (D1 , . . . , D4 ) ∈
/ P4,r
then ν −1 (D1 , . . . , D4 ) ∩ P4 = ∅ for odd r. Continuing this argument, we deduce
(5.9) p1 · · · pr ∈ D ∗ =⇒ ν(p1 · · · pr ) ∈ D∗r if r even,
(D1 , . . . , Dr ) ∈ D∗r =⇒ ν −1
(D1 , . . . , Dr ) ⊂ D ∗
if r odd.
Without loss, we may restrict D∗r to vectors with nonempty preimage in D ∗ . Hence by
construction, (5.5) becomes2
X X (−1)r X
(5.10) S ∗ (Ã, z) 6 |Ãp1 ···pr |,
−2 ∗
γ(D1 , . . . , Dr )
0 6r 6ε (D1 ,..,Dr )∈Dr p1 ···pr |P (z)
Dj <pj 6Dj1+τ

where γ(D1 , . . . , Dr ) = k1 ! · · · kℓ ! for the corresponding multiplicities ki >1 (i.e. we have

r = k1 + · · · + kℓ and D1 = · · · = Dk1 < Dk1 +1 = · · · = Dk2 < · · · = Dr .). Note the term
r = 0 corresponds to |Ã| with p1 = · · · = pr = 1.
Now by the Fundamental Lemma [10, Theorem 6.9], we have upper (and lower) bounds
|Ãp1···pr | = S(Ap1···pr , u)
X
6 g(p1 · · · pr )|A|V (u){1 + O(e−1/ε )} + λ(r) (b) rA (bp1 · · · pr ),
b 6D ε

(with 6 replaced by > for the lower bound) where λ(r) is the upper (lower) bound beta-sieve
of level D ε when r is even (odd). Plugging back into (5.10), we get
X X (−1)r
S(A, z) 6 S ∗ (Ã, z) 6
γ(D1 , . . . , Dr )
0 6r 6ε−2 (D1 ,..,Dr )∈D∗r
X  X 
−1/ε (r)
(5.11) × g(p1 · · · pr )|A|V (u){1 + O(e )} + λ (b) rA (bp1 · · · pr ) .
p1 ···pr |P (z) b 6D ε
Dj <pj 6Dj1+τ

2Indeed, we have reverse engineered the definition of D∗r just so that (5.10) holds.
20
 We now compare the main term above to that of the modified linear sieve, as in (4.3)–(4.10)
from the proof of Proposition 4.1, namely,
X 
V ∗ (D, z) := λ∗ (d)g(d) = V (z) F ∗ (s) + o(1) .
d|P (z)/P (u)

The difference between these main terms is accounted for by those d with two close prime
factors, within a ratio D τ , and those d near the boundary. The former contribution is
X X Y
g(d) 6 g(pp′ ) · (1 + g(p)),
d|P (z)/P (u) u 6p<z u 6p<z
p 6p′ <pD τ p 6p′ <pD τ
pp′ |d

and the latter contribution is

X X
(5.12) g(p1 · · · pr ),
r u<pr <···p1 <z
D 1/(1+τ ) <p1 ···p3m <D

where m is the first index (m 6r) for which this occurs. Both contributions may be shown
to be O(ε5), see [10, pp. 254-255]. Hence the Proposition follows. 
Remark 5.2. We make a minor technical point. Namely, at an admissible cost O(ε5) we
may assume Proposition 5.1 holds, where D∗r is further restricted to vectors (D1 , . . . , Dr )
satisfying
1 1
(5.13) ν −1 (D11+τ , . . . , Dr1+τ ) ⊂ D ∗ ,
regardless of parity of r. To show this, note by definitions of P4,r , P6,r the integers p1 · · · pr
1/(1+τ )
with pj ∈ [Dj , Dj ], j 6r, that lie outside D ∗ must satisfy
1 1
x 6 +2η /p1 or x 12 −5η /p6 ∈ [x−2τ , x2τ ].
1 1
Then for B1 = x 6 +2η+2τ , B6 = x 12 −5η+2τ , we have
X log Bi log x4τ τ
g(pi ) ≪ log ≪ ≪ ,
log max(Bi /x4τ , u) log u ε2
Bi >pi > max(B/x4τ ,u)

and so the contribution of such integers to the main term of (5.11) is

X X τ Y
τ log z
≪ g(p1 · · · pr ) ≪ 2 1 + g(p) ≪ 2 ≪ ε5 .
r u<pr <···p <z
ε u<p<z
ε log u
1
Bi >pi >max(Bi /x4τ ,u), i∈{1,6}

Hence (5.13) follows.

We now proceed to Theorem 2.11. For each vector (D1 , .., Dr ) ∈ Dr we define the weight
λ(D1 ,..,Dr ) supported on d in ν −1 (D1 , .., Dr ), namely,

(5.14) λ(D1 ,..,Dr ) (d) := 1d=p 1 ···pr

D <p 6 D 1+τ ∀j
.
j j j

21
Next, we may decompose an integer d into its u-smooth and rough components, d =
b(p1 · · · pr ). Recall for b | P (u) we have λ(r) (b) = µ(b) if b ∈ D (r) , and λ(r) (b) = 0 else.
(r)
Thus the convolution λ(D1 ,..,Dr ) := λ(D1 ,..,Dr ) ∗ λ(r) equals

 (r) ε ε2
µ(b) if d = bp1 · · · pr for b ∈ D (D ), b | P (D ),
(r)
(5.15) λ(D1 ,..,Dr ) (d) = and Dj < pj 6 Dj1+τ ∀j 6r.

0 else.

Hence the remainder in (5.1) equals

X X (−1)r X (r) X
λ(D1 ,..,Dr ) (d) rA(d) = e∗ (d) rA (d),
λ
γ(D1 , . . . , Dr )
0 6r 6ε−2 (D1 ,..,Dr )∈D∗r d|P (z) d|P (z)

for the weights

X X (−1)r (r)
(5.16) e∗ =
λ λ(D1 ,..,Dr ) .
γ(D1 , . . . , Dr )
0 6r 6ε−2 (D1 ,..,Dr )∈D∗r

(r)
Recalling the cardinality of D∗r ⊂ D+
r from (5.7), it suffices to show the weights λ(D1 ,..,Dr ) is
programmably factorable for each vector in D∗r . To this we have the following.
Lemma 5.3. For an integer d denote the vector ν(d) = (D1 , . . . , Dr ) from (5.9). If d
(r)
has a factorization as in (3.7) at level D, then the corresponding weights λν(d) and λν(d) ,
as in (5.14) and (5.15), resp., are programmably factorable sequences of levels D 1+τ and
7
D 1+τ +ε = x 12 +η , resp. (relative to x, ε/50).
Proof. By assumption for each N ∈ [1, x1/3 ], thereQ is a factorization d = d1 d2 d3 satisfying
the system (3.7). For j = 1, 2, 3, write dj = i∈Ij pi for the induced partition of indices
Q
{1, ..., r} = I1 ∪ I2 ∪ I3 . Thus letting Qj = i∈Ij Di , the factorization D1 · · · Dr = Q1 Q2 Q3
satisfies (2.1), since Di < pi .
Further, writing the corresponding subvectors (Di )i∈Ij for j = 1, 2, 3, the weights λ(Di )i∈Ij
are 1-bounded, supported on [1, Q1+τ j ], and give the desired triple convolution,
λ(D1 ,...,Dr ) = λ(Di )i∈I1 ∗ λ(Di )i∈I2 ∗ λ(Di )i∈I3 .
(r)
Hence λ(D1 ,...,Dr ) is programmably factorable of level D 1+τ as claimed. Similarly λ(D1 ,...,Dr ) =
(r)
λ(Di )i∈I1 ∗ λ(Di )i∈I2 ∗ λ(D3 )i∈I is programmably factorable of level D 1+τ +ε . 
1

Now for each vector (D1 , . . . , Dr ) ∈ D∗r , by (5.13) there exists d = p1 · · · pr ∈ D ∗ for some
1/(1+τ )
primes pi ∈ (Di , Di ]. Then for all N ∈ [1, x1/3 ] Proposition 3.3 gives a factorization of
(r) 7
d as in (3.7), and so by Lemma 5.3 λ(D1 ,...,Dr ) is programmably factorable of level x 12 +η .
This completes the proof of Theorem 2.11.

5.1. Variable level of distribution for the linear sieve weights. We now return to
e± for the (upper and lower) linear sieve, given explicitly
Iwaniec’s well-factorable weights λ
22
from (5.15) as the weighted sum,
X X (−1)r (r)
(5.17) e± =
λ λ(D1 ,..,Dr ) .
γ(D1 , . . . , Dr )
0 6r 6ε−2 (D1 ,..,Dr )∈D±
r

We introduce the analogous set of well-factorable vectors Dwell

r = Dwell
r (D),

(5.18) Dwell
r
2
= {(D1 , ..., Dr ) : D1 · · · Dm−1 Dm <D for all m 6r}.
Note D± well
r ⊂ Dr , having dropped parity conditions on the indices m 6r.
We have the following technical variation on Theorem 2.11 for the original linear sieve.
Proposition 5.4. Let (D1 , ..., Dr ) ∈ Dwell θ
r (D) and write D = x , Di = x
ti
for i 6r. If
θ 6θ(t1 ) − ε as in (3.14), then
X X  
(5.19) e± (d) π(x; d, a) − π(x) ≪a,A,ε
λ
x
.
ϕ(d) (log x)A
b=p1 ···pr θ
b|d, d 6x
Di <pi 6Di1+τ d/b|P (pr )
(d,a)=1

Moreover if t1 6 51 and r >3, then (5.19) holds provided that θ 6 θ(t1 , t2 , t3 ) − ε as in (3.15).
If t1 6 15 and r 62, then provided θ 6 3−u
5
− ε,
X X  π(x)  x
e ±
λ (d) π(x; d, a) − ≪a,A,ε .
b=p ···p θ
ϕ(d) (log x)A
1 r b|d, d 6x
Di <pi 6Di1+τ d/b|P (xu )
(d,a)=1

In particular for r = 0 (i.e. the empty vector), θ 6 3−u

5
− ε, this simplifies as
X  
e± (d) π(x; d, a) − π(x) ≪a,A,ε
λ
x
.
θ
ϕ(d) (log x)A
d 6x
d|P (xu )
(d,a)=1

Proof. Given (D1 , ..., Dr ), take an integer b = p1 · · · pr with Di < pi 6Di1+τ . Then for all
(s)
multiples d of b with d/b | P (pr ), the weight λ(D′ ,..,Ds′ ) (d) = 0 unless the vector (D1′ , .., Ds′ )
1
extends (D1 , ..., Dr ). That is, Di′ = Di for all i 6r. Conversely, given such a vector (D1′ , .., Ds′ )
(s)
we have λ(D′ ,..,Ds′ ) (d′ ) = 0 unless d′ = p1 · · · ps with Di < pi 6Di1+τ , i 6r. So by definition of
1
e± as in (5.17), we have
λ
X X X X (−1)s X (s)
(5.20) e± (d) =
λ λ(D′ ,..,Ds′ ) (d).
γ(D1′ , . . . , Ds′ ) 1
b=p1 ···pr b|d, d 6xθ r 6s 6ε−2 (D1′ ,...,Ds′ )∈D±
s
θ
d 6x
Di <pi 6Di1+τ d/b|P (pr ) Di′ =Di , i 6r (d,a)=1
(d,a)=1

Here we have extended (by zero) the inner sum to all d 6xθ , (d, a) = 1.
Next, take such a vector (D1′ , ..., Ds′ ) ∈ D± θ ′
s (x ) with Di = Di for i 6r. Each integer
′ ′ 1+τ
d = p1 · · · ps with Di < pi 6(Di ) lies in d ∈ D (x ). In particular p1 6D11+τ 6xt1 +τ so
well θ+τ

by Corollary 3.8, d has a factorization as in (3.7) at level xθ(t1 +τ ) . Since θ(t) is continuous (in
fact, piecewise linear), for τ > 0 sufficiently small θ(t1 + τ ) > θ(t1 ) − ε > θ. Thus by Lemma
23
 (s)
5.3 the weights λ(D1 ,..,Ds) are programmably factorable sequences of level xθ . Hence for each
such vector, by Theorem 2.5 we have
X (s) x
(5.21) λ(D′ ,..,Ds′ ) (d) ≪a,A,ε .
θ
1 (log x)A
d 6x
(d,a)=1

Plugging (5.21) back into (5.20) gives the bound (5.19), as claimed.
Moreover, if t1 6 15 and r >3 then proceeding as in the above paragraph, by Corollary 3.8
d will factor as in (3.7) to level xθ(t1 +τ,t2 +τ,τ3 +τ ) . Again θ(t, u, v) is continuous, so for τ > 0
sufficiently small θ(t1 + τ, t2 + τ, τ3 + τ ) > θ(t1 , t2 , t3 ) − ε > θ. Hence (5.19) also follows in this
case.
Similarly if t1 6 51 and r 62, proceeding as above with d = p1 · · · ps , the assumption d/b |
P (xu ) implies s 62 or p3 6D31+τ 6xu+τ . Thus by Corollary 3.8 d will factor as in (3.7) to
level x(3−u−τ )/5 >xθ . Hence (5.19) follows in this case as well. 

5.2. Equidistribution for products of primes. We use an extension of Theorem 2.5 to

products of k primes. This is the analogue in the programmably factorable setting of Lemma
7 [18] extending Theorem 2.2 of Bombieri–Friedlander–Iwaniec.
3
Proposition 5.5. Let ε > 0 and λ be a programmably factorableP sequence of level D 6x 5 −ε
(relative to x, ε/50). Take real numbers ε1 , . . . , εk >ε such that i 6k εi = 1. Then for any
fixed integer a ∈ Z, A, B > 0, letting ∆ = 1 + (log x)−B we have
X  X X 
1 x
(5.22) λ(d) 1 − 1 ≪a,ε,A,B .
d 6D
ϕ(d) (log x)A
p1 ···pk ≡a (mod d) (p1 ···pk ,d)=1
(d,a)=1 xεi /∆<pi 6xεi ∀i 6k xεi /∆<pi 6xεi ∀i 6k

Proof. This follows by the same proof method as in Theorem 2.5 (i.e. Maynard’s [15, Theo-
rem 1.1]). Indeed, Maynard just uses the Heath–Brown identity to decompose the indicator
function of primes into Type I/II sums. A similar decomposition holds for products of k
primes, after which we may apply the same Type I/II estimates in Propositions 5.1 and 5.2
of [15]. 
In addition, by replacing Theorem 2.5 with Proposition 5.5 in the proof, we obtain ana-
e± in the case of products
logues of Proposition 5.4 and 5.6 for the linear sieve weights λ = λ
of k primes.
Corollary 5.6. Let (D1 , ..., Dr ) ∈ Dwellr (D)
θ ti
P and write D = x , Di = x for i 6r. Let ε > 0
and real numbers ε1 , . . . , εk >ε such that i 6k εi = 1. Fix an integer a ∈ Z, A, B > 0, and
let ∆ = 1 + (log x)−B . If θ 6θ(t1 ) − ε as in (3.14),
(5.23)
X X  X X 
e± 1 x
λ (d) 1 − 1 ≪a,ε,A,B .
ϕ(d) (log x)A
b=p′1 ···p′r b|d, d 6xθ p1 ···pk ≡a (mod d) (p1 ···pk ,d)=1
Di <p′i 6Di1+τ d/b|P (p′r ) xεi /∆<pi 6xεi ∀i 6k xεi /∆<pi 6xεi ∀i 6k
(d,a)=1

Moreover if t1 6 51 , r >3, then (5.23) holds provided that θ 6 θ(t1 , t2 , t3 ) − ε as in (3.15).

24
 In addition, if r 62, u 6tr , t1 6 15 , and θ 6 3−u
5
− ε, then
(5.24)
X X  X X 
e± 1 x
λ (d) 1 − 1 ≪a,ε,A,B .
ϕ(d) (log x)A
b=p′1 ···p′r b|d, d 6xθ p1 ···pk ≡a (mod d) (p1 ···pk ,d)=1
Di <p′i 6Di1+τ d/b|P (xu ) xεi /∆<pi 6xεi ∀i 6k xεi /∆<pi 6xεi ∀i 6k
(d,a)=1

6. Upper bound for twin primes

Now we shall apply the modified sieve to the set of twin primes
A := {p + 2 : p 6x}.

Q sieve notation specializes as P = {p > 2} and g(d) = 1/ϕ(d)

In this case the
γ
for odd d, so
that V (z) = 2<p<z (1 − 1/ϕ(p)). Recall by Mertens theorem V (z) ∼ S2 /e log z.
We begin in the spirit of Fouvry–Grupp [9], and apply a weighted sieve inequality. To each
non-switched term, we apply the Buchstab identity in order to lower the sieve threshhold
down to z = xǫ for some tiny ǫ > 0. By Proposition 5.4, such smooth sums will satisfy level
of distribution 3−ǫ
5
. Combined with variations on a theme, which identify programmably-
factorable weights in certain cases, the consequent increase in level will be sufficient to obtain
the bound in Theorem 1.2. For a final bit of savings, we use refinements obtained by Wu’s
iteration method [19].
Before moving on, we note that sieve methods and the switching principle has also yielded
progress on the Goldbach problem. Indeed, the upper bound in Wu [19, Theorem 3] for
twin primes is obtained by using the same formulae as in [19, Theorem 1] for the Goldbach
problem, except for altering the level from 12 to 47 (this amounts to replacing factors of 4
with 7/2 in a few instances). Importantly, the quantitative upper bounds for twin primes are
much stronger than those of the Goldbach problem. This is because the latter relies on level
1
2
from Bombieri–Vinogradov for a growing residue a = N, while the former may appeal to
level of distribution 74 from Bombieri–Friedlander–Iwaniec for the fixed residue a = 2 (and
now the subsequent improvements of Maynard). As such our methods have nothing new to
say for the Goldbach problem.
6.1. Lemmas. We begin with a standard
lemma for x1/u -rough numbers in terms of the
Buchstab function ω(u) = f (u) + F (u) /(2eγ ). Alternatively, ω is directly defined via
1
ω(u) = for 1 6u 62,
u

′
u ω(u) = ω(u − 1) for 2 6u.
Lemma 6.1. Let x >2 and y = x1/u . Then we have
X x  x 
1 = ω(u) +O 2
.
n 6x
log y (log y)
p|n⇒p >y

Proof. This is [18, Lemma 12]. 

The argument of Wu makes essential use of weighted sieve inequalities, as in Lemmas 4.1
and 4.2 [19]. We shall employ the latter inequality in the special case d = 1, σ = 1.
25
Lemma 6.2. For 3/10 >ρ >τ3 > τ2 > τ1 >ρ′ >1/20, we have
X
5S(A, xρ ) . Γn ,
1 6n 621

where

′
Γ1 := 4S(A, xρ ) + S(A, xτ1 ), XXX
X ′ Γ9 := S(Ap1p2 p3 , p2 ),
Γ2 := − S(Ap , xρ ), x τ1 6p1 <p2 <p3 <xτ3
xρ′ 6p<xρ XXX
X ′
Γ10 := S(Ap1 p2 p3 , p2 ),
Γ3 := − S(Ap , xρ ), xτ1 6p1 <p2 <xτ2 6p3 <xρ
xρ′ 6p<xτ2
XXX
X Γ11 := S(Ap1p2 p3 , p2 ),
ρ′
Γ4 := − S(Ap , x ), x τ1 6p1 <xτ2 6p2 <p3 <xτ3
XXX
x ρ′ 6p<xτ3 Γ12 := S(Ap1p2 p3 , p2 ),
XX
ρ′
Γ5 := S(Ap1p2 , x ), xρ
′
6p1 <p2 <xτ1
xτ3 6p3 <xρ
x ρ′ 6p1 <p2 <xτ2 XXX
XX Γ13 := S(Ap1p2 p3 , p2 ),
ρ′
Γ6 := S(Ap1 p2 , x ), ′
′ xρ 6p1 <xτ1 6p2 <xτ2 6p3 <xρ
xρ 6p1 <xτ1 XXX
xτ2 6p2 <xτ3
XX Γ14 := S(Ap1 p2 p3 , p2 ),
Γ7 := S(Ap1p2 , p1 ), ρ′
x 6p1 <xτ1
′ xτ2 6p2 <p3 <xρ
xρ 6p1 <p2 <xτ1 XXX
XX Γ15 := S(Ap1p2 p3 , p2 ),
Γ8 := S(Ap1p2 , p1 ),
x τ1 6p1 <xτ2 6p2 <xτ3 6p3 <xρ
′
xρ 6p1 <xτ1 6p2 <xτ2

XXXX
Γ16 := S(Ap1p2 p3 p4 , p3 ),
x τ2 6p1 <p2 <p3 <p4 <xτ3
XXXX
Γ17 := S(Ap1p2 p3 p4 , p3 ),
xτ2 6p1 <p2 <p3 <xτ3 6p4 <xρ
XXXX
Γ18 := S(Ap1p2 p3 p4 , p3 ),
x τ2 6p1 <p2 <xτ3 6p3 <p4 <xρ
XXXX
Γ19 := S(Ap1p2 p3 p4 , p3 ),
xτ1 6p1 xτ2
xτ3 6p2 <p3 <p4 <xρ
XXXXX
Γ20 := S(Ap1p2 p3 p4 p5 , p4 ),
xτ2 6p1 <xτ3 6p2 <p3 <p4 <p5 <xρ
XXXXXX
Γ21 := S(Ap1p2 p3 p4 p5 p6 , p5 ).
x τ3 6p1 <p2 <p3 <p4 <p5 <p6 <xρ

Proof. This is [19, Lemma 4.2] from Wu with d = 1, σ = 1. Here we simplify Wu’s notation
′ ′
slightly, using (d1/s , d1/κ3 , d1/κ2 , d1/κ1 , d1/s ) = (xρ , xτ3 , xτ2 , xτ1 , xρ ). The basic proof idea is
to iterate the Buchstab identity and to strategically neglect some terms by positivity. 
26
6.2. Computations. Given 0.1 6ρ′ 6τ1 < 0.2 6τ2 < τ3 6ρ 60.3., we define integrals In =
In (ρ, ρ′ , τ1 , τ2 , τ3 ) by
Z  1 − t − u − v  dt du dv
In = ω (9 6n 615),
Dn u tu2 v
Z  1 − t − u − v − w  dt du dv dw
(6.1) In = ω (16 6n 619),
Dn v tuv 2 w
Z  1 − t − u − v − w − x  dt du dv dw dx
I20 = ω ,
D20 w tuvw 2x
Z  1 − t − u − v − w − x − y  dt du dv dw dx dy
I21 = ω ,
D21 x tuvwx2 y
where ω is the Buchstab function, and where the domains Dn are
D9 = {(t, u, v) : τ1 < t < u < v < τ3 },
D10 = {(t, u, v) : τ1 < t < u < τ2 < v < ρ},
D11 = {(t, u, v) : τ1 < t < τ2 < u < v < τ3 },
D12 = {(t, u, v) : ρ′ < t < u < τ1 , τ3 < v < ρ},
D13 = {(t, u, v) : ρ′ < t < τ1 < u < τ2 < v < ρ},
D14 = {(t, u, v) : ρ′ < t < τ1 , τ2 < u < v < ρ},
D15 = {(t, u, v) : τ1 < t < τ2 < u < τ3 < v < ρ},
D16 = {(t, u, v, w) : τ2 < t < u < v < w < τ3 },
D17 = {(t, u, v, w) : τ2 < t < u < v < τ3 < w < ρ},
D18 = {(t, u, v, w) : τ2 < t < u < τ3 < v < w < ρ},
D19 = {(t, u, v, w) : τ1 < t < τ2 , τ3 < u < v < w < ρ},
D20 = {(t, u, v, w, x) : τ2 < t < τ3 < u < v < w < x < ρ},
D21 = {(t, u, v, w, x, y) : τ3 < t < u < v < w < x < y < ρ}.
Recall the definitions (3.14) and (3.15),
(
2−t
3
if t > 15 ,
θ(t) = 1+t
2
if t 6 15 .

and
n3−v
θ(t, u, v) := max , θ(t), θ(u), θ(t + u + v),
5 o
θ(t + u), θ(t + v), θ(u + v) .

We also define
(6.2) G1 = 4G(ρ′ ) + G(τ1 ), G3 = G0 + G(τ2 ),
G2 = G0 + G(ρ), G4 = G0 + G(τ3 ),
27
where for c 61/5,
Z Z Z
1
1 c dt
1 c t dt du

(6.3) G(c) = F θǫ /ǫ − f (θǫ − t)/ǫ + F (θǫ − t − u)/ǫ
ǫ ǫ ǫ t ǫ ǫ ǫ tu
Z cZ tZ u
dt du dv

− f (θ(t, u, v) − t − u − v)/v ,
ǫ ǫ ǫ tuv 2
and for c > 1/5,
Z c Z c Z ρ′
1 dt
dt du

(6.4) G(c) = − f (θ(t) − t)/ǫ + F (θ(t) − t − u)/u
ǫ 1/5 t 1/5 ǫ tu2
as well as
Z Z Z ′
1 1/5 dt
1 1/5 ρ dt du

(6.5) G0 = − f (θǫ − t)/ǫ + F (θǫ − t − u)/ǫ
ǫ ρ′ t ǫ ρ′ ǫ tu
Z 1/5 Z ρ′ Z u
dt du dv

− f (θ(t, u, v) − t − u − v)/v .
ρ′ ǫ ǫ tuv 2
We similarly let
Z Z Z Z
1 1/5 t dt du
1 τ2 t dt du

G5 = F (θǫ − t − u)/ǫ + ′ F (θ(t) − t − u)/ρ′
ǫ ρ′ ρ′ tu ρ 1/5 ρ′ tu
Z 1/5 Z t Z ρ′
dt du dv

− 2
f (θ(t, u, v) − t − u − v)/v ,
ρ′ ρ′ ǫ tuv
Z τ3 Z τ1
1 dt du

(6.6) G6 = ′ F (θ(t) − t − u)/ρ′ ,
ρ τ2 ρ′ tu
Z τ1 Z t
1 dt du

G7 = F (θǫ − t − u)/ǫ
ǫ ρ′ ρ′ tu
Z τ1 Z t Z u
dt du dv

− 2
f (θ(t, u, v) − t − u − v)/v ,
ρ′ ρ′ ǫ tuv
Z 1/5 Z τ1 Z τ2 Z τ1
1 dt du
dt du

G8 = F (θǫ − t − u)/ǫ + 2
F (θ(t) − t − u)/u
ǫ τ1 ρ′ tu 1/5 ρ′ tu
Z 1/5 Z τ1 Z u
dt du dv

− 2
f (θ(t, u, v) − t − u − v)/v .
τ1 ρ′ ǫ tuv
Recall the sieve functions F, f satisfy F (s) = 2eγ /s for s ∈ [1, 3], f (s) = 2eγ log(s − 1)/s
R s−1
for s ∈ [2, 4] and F (s) = 2eγ /s · [1 + 2 f (t) dt] for all s >1.
The main bound is the following.
3−ǫ
Proposition 6.3. Let 0 < ǫ 60.1 6ρ′ 6τ1 < 0.2 6τ2 < τ3 6ρ 60.3. and θǫ = 5
. Then for
In , Gn , and G(c) as in (6.1),(6.2),(6.6), and (6.3), we have
 8 21 
ρ Π(x) X 1
X
(6.7) S(A, x ) . Gn + G( 5 ) In .
5eγ n=1 n=9
28
Proof. We first bound S(A, xc ) for c ∈ [ǫ, 1/5]. By the Buchstab identity,
X
S(A, xc ) = S(A, xǫ ) − S(Ap , p).
xǫ 6p<xc

Iterating twice more, we obtain

X
(6.8) S(A, xc ) = S(A, xǫ ) − S(Ap1 , xǫ )
xǫ 6p1 <xc
X X
+ S(Ap1 p2 , xǫ ) − S(Ap1 p2 p3 , p3 ).
xǫ 6p2 <p1 <xc xǫ 6p3 <p2 <p1 <xc

To each term S(Ad , xǫ ) above, we apply the linear sieve of level xθǫ for θǫ = 3−ǫ 5
, as in
Theorem 2.9. And to each term S(Ap1 p2 p3 , p3 ), we apply the linear sieve of level xθ for
θ = θ(t1 , t2 , t3 ), where pi = xti .
To handle the corresponding error terms, note for primes xǫ 6p2 < p1 < xc 6x1/5 and
d ∈ {1, p1 , p1 p2 }, the prime factors of q/d above are bounded by xǫ so that the sets Aq are
equidistributed to level θǫ . Hence for each xǫ 6p2 < p1 < x1/5 , by Proposition 5.4 with u = ǫ,
b = 1, p1 , p1 p2 , we have
X   X  
e+ (q) |Aq | − |A| =
λ e+ (q) π(x; q, −2) − π(x) ≪A
λ
x
,
θǫ
ϕ(q) θǫ
ϕ(q) (log x)A
q 6x q 6x
q|P (xǫ ) q|P (xǫ )

and
X X  |A|  x
e
λ− (q) |Aq | − ≪A ,
p1 p1 |q, q 6x θǫ
ϕ(q) (log x)A
q/p1 |P (xǫ )

and
X X  |A|  x
e +
λ (q) |Aq | − ≪A .
p2 ,p1 p1 p2 |q, q 6x θǫ
ϕ(q) (log x)A
q/p1 p2 |P (xǫ )

In addition for each p3 < p2 < p1 < x1/5 , pi = xti , letting θ = θ(t1 , t2 , t3 ), by Proposition 5.4
with b = p1 p2 p3 ,
X X  
e− (q) |Aq | − |A|
λ ≪A
x
.
p ,p ,p θ
ϕ(q) (log x)A
3 2 1 p1 p2 p3 |q, q 6x
q/p1 p2 p3 |P (p3 )

Thus for c 6 15 , the linear sieve bounds give

(6.9) S(A, xǫ ) . |A|V (xǫ )F θǫ
ǫ
,
and
X X

(6.10) S(Ap1 , xǫ ) & |A|g(p1)V (xǫ )f θǫ −t1
ǫ
,
xǫ 6p1 <xc xǫ 6p1 <xc
29
and
X X

(6.11) S(Ap1 p2 , xǫ ) . |A|g(p1p2 )V (xǫ )F θǫ −t1 −t2
ǫ
,
xǫ 6p2 <p1 <xc xǫ 6p2 <p1 <xc

and
X X

θ−t1 −t2 −t3
(6.12) S(Ap1p2 p3 , p3 ) & |A|g(p1p2 p3 )V (p3 )f t3
.
xǫ 6p3 <p2 <p1 <xc xǫ 6p3 <p2 <p1 <xc

Hence by (6.9), (6.10), (6.11), (6.12), we observe that (6.8) becomes

X

(6.13) S(A, xc ) . − |A| g(p1p2 p3 )V (p3 )f θ−t1 −t
ǫ
2 −t3

xǫ 6p3 <p2 <p1 <xc

 X X 
ǫ θǫ

θǫ −t1

θǫ −t1 −t2


+ |A|V (x ) F ǫ
− g(p1 )f ǫ
+ g(p1 p2 )F ǫ
.
xǫ 6p1 <xc xǫ 6p2 <p1 <xc

Recall V (z) ∼ S2 /eγ log z by Mertens theorem. Thus by partial summation and the prime
number theorem, we obtain
Π(x)
(6.14) S(A, xc ) .
G(c),
eγ
for G(c) as in (6.3). Hence for c = ρ′ , τ1 we have c ∈ [ǫ, 1/5], so we bound Γ1 as
′ Π(x)   Π(x)
Γ1 = 4S(A, xρ ) + S(A, xτ1 ) . 4G(ρ ′
) + G(τ1 ) = γ G1 .
eγ e
′
Now consider c, c ∈ [1/5, 2/7]. We shall apply the linear sieve of level θ(t1 ), as in (3.14).
In general, for p1 = xt1 and θ(t1 ) as in (3.14), Proposition 5.4 gives
X X  |A|  x
e −
λ (q) |Aq | − ≪A ,
p θ(t )
ϕ(q) (log x)A
1 p1 |q, q 6x 1
q/p1 |P (p1 )

so that for c, c′ ∈ [1/5, 2/7], the linear sieve of level θ(t1 ) gives
X X θ(t1 )−t1

(6.15) S(Ap1 , p1 ) & |A|g(p1)V (p1 )f t1
.
xc′ 6p1 <xc xc′ 6p1 <xc

Thus by partial summation and the prime number theorem,

X ′ Π(x)
Π(x)
Γ2 = S(Ap , xρ ) . γ G + G(ρ) = γ G2 .
ρ′ ρ
e e
x 6p<x

Similarly, we obtain
Π(x)
(6.16) Γn . Gn for 1 6n 68.
eγ
Finally, for the remaining Γn , we apply the switching principle. Namely, for Γ9 we have
XXX
(6.17) Γ9 = S(Ap1p2 p3 , p2 ) = S(B, x1/2 ) + O(x1/2 )
xτ1 6p1 <p2 <p3 <xτ3
30
for the set
(6.18) B = {p1 p2 p3 m − 2 6x : xτ1 6p1 < p2 < p3 < xτ3 ,
p′ | m ⇒ p′ >p2 or p′ = p1 }.
Note since p2 , p1 > xτ1 > x.1 , each m above has at most 7 prime factors.
Now by a standard subdivision argument, B is similarly equidstributed S in arithmetic
progressions as is A. Indeed, the basic idea is to partition B = r 67 B(r) , where B(r) is the
subset corresponding to integers m with r prime factors. Then we cover the prime tuples
(p1 , . . . , pr ) into hypercubes of the form
 l1 l1 +1


∆ ,∆ × · · · × ∆lr , ∆lr +1 ,
for ∆ = 1 + (log x)−B with B > 0 sufficiently large, and apply Corollary 5.6 with u = ǫ,
b = 1. This gives
X  |B|  x
e +
λ (q) |Bq | − ≪A .
θǫ
ϕ(q) (log x)A
q 6x
q|P (xǫ )

Similarly for each xε < p′3 < p′2 < p′1 < x1/5 , by Corollary 5.6 with u = ǫ and b = p′1 , p′1 p′2 ,
we have
X X  
e− (q) |Bq | − |B|
λ ≪A
x
,
′ ′ θǫ
ϕ(q) (log x)A
p1 p1 |q, q 6x
q/p′1 |P (xǫ )

and
X X  
e+ (q) |Bq | − |B|
λ ≪A
x
.
ϕ(q) (log x) A
p′2 ,p′1 p′1 p′2 |q, q 6xθǫ

q/p′1 p′2 |P (xǫ )

In addition for each p′3 < p′2 < p′1 < x1/5 , p′i = xti , letting θ = θ(t1 , t2 , t3 ), by Corollary 5.6
with b = p′1 p′2 p′3 , we have
X X  |B|  x
e −
λ (q) |Bq | − ≪A .
′ ′ ′ ′ ′ ′ θ
ϕ(q) (log x)A
p3 ,p2 ,p1 p1 p2 p3 |q, q 6x
q/p′1 p′2 p′3 |P (p′3 )

Iterating the Buchstab identity, we have

X
S(B, x1/2 ) 6S(B, x1/5 ) = S(B, xǫ ) − S(Bp′1 , xǫ )
xǫ 6p′1 <x1/5
X X
+ S(Bp′1 p′2 , xǫ ) − S(Bp′1 p′2 p′3 , p′3 ),
xǫ 6p′2 <p′1 <x1/5 xǫ 6p′3 <p′2 <p′1 <x1/5

and hence by the linear sieve bounds we obtain

G( 51 )
(6.19) Γ9 . S(B, x1/2 ) 6S(B, x1/5 ) . e−γ S2 |B|.
log x
31
 Now to compute |B|, we may drop the condition “or p′ = p1 ” in (6.18) at admissible cost,
so that by Lemma 6.1 we have
Z
x dt1 dt2 dt3  1 − t1 − t2 − t3  x
|B| ∼ 2
ω = · I9 .
log x τ1 <t1 <t2 <t3 <τ2 t1 t2 t3 t2 log x
Thus we have Γ9 . e−γ G( 15 )Π(x)I9 . Similarly, we obtain
Π(x)
(6.20) Γn . G( 15 ) In (for 9 6n 621).
eγ
Hence plugging (6.16), (6.20) into Lemma 6.2 completes the proof. 
Let
ρ = 0.27195, τ3 = 0.24589,
(6.21) ρ′ = 0.12313, τ2 = 0.20867,
ǫ = 0.002, τ1 = 0.16288.
For such choices of parameters, we compute the following integrals from Proposition 6.3,
X X
In 6 0.174404, Gn 6 28.34581, G( 15 ) 6 5.99237.
9 6n 621 1 6n 68

Thus by Proposition 6.3, we obtain the bound

 8 21 
ρ Π(x) X 1
X
(6.22) π2 (x) . S(A, x ) . Gn + G( 5 ) In . 3.30042 Π(x).
5eγ n=1 n=9

We also record the individual bounds (see table at end of article).

6.3. Completing the proof of Theorem 1.2. We shall refine our argument in certain
cases for which Wu’s iteration method [19] applies directly without modification. As such
we have chosen simplicity over full optimization.
In the lemma below, we consider the cases of sets Ap1p2 , where p1 , p2 lie in a prescribed
range, and such that for all multiples bp1 p2 , the sets Abp1 p2 are equidistributed to level xθ
(we also need level xθ for corresponding switched sets B of integers mbp1 p2 − 2).
Lemma 6.4. Let θ ∈ [1/2, 1], s ∈ [2, 3], and A = {p + 2 : p 6x}. There is a function
Hθ (s), monotonically increasing in θ for fixed s and decreasing in s for fixed θ, such that the
following holds: For each (D1 , D2 ) ∈ Dwell θ
2 (x ), we have
X Π(x) X log x 2eγ

(6.23) S(Ap1 p2 , z) . F (s) − H θ (s) ,
1+τ
eγ 1+τ
ϕ(p1 p2 ) log z s
D1 <p1 <D1 D1 <p1 <D1
D2 <p2 <D21+τ D2 <p2 <D21+τ

where z s = xθ /p1 p2 , provided (5.22) holds for λ = λ e+ at level D = xθ with (xε1 , xε2 ) =
(D1 , D2 ), and provided for all vectors (D1 , . . . , Dr ) ∈ Dwell θ
r (x ) extending (D1 , D2 ),
X X  
e± (d) π(x; d, a) − π(x) ≪a,ε,A
λ
x
.
b=p ···p θ
ϕ(d) (log x)A
1 r b|d, d 6x
Di <pi 6Di1+τ (d,a)=1
32
Proof. Wu iterates the weighted sieve inequality (Lemmas 4.1 and 4.2 [19]), on the subset
of (non-switched) terms whose sieving parameter s lies in the interval s ∈ [2, 3]. This
yields a recurrence relation for a function Hθ (s), which encodes the percent savings over the
(normalized) linear sieve sF (s)/(2eγ ).
Starting from a term S(Ap1 p2 , z), each successive iteration of the weighted sieve inequality
is composed of terms of the form S(Abp1 p2 , z ′ ) for some multiple bp1 p2 corresponding to some
vector extending (D1 , D2 ). By assumption, all such sets are equidistributed to level xθ , when
weighted by the upper/lower linear sieve. Similarly, the switched sets are also equidistributed
to level xθ (here we only need the upper bound weights λ e+ ). Finally, the savings function
Hθ (s) inherits the stated monotonicity properties by construction of the iteration. 

The function Hθ depends on the known level of distribution xθ (i.e. Wu used θ = 12 for
Goldbach, and θ = 74 for twin primes). For parameters as in Tables 1 and 2 [19, pp.30–32],

  0.0287118 if 2.0 6 t 6 2.1,
0.0223939 if 2.0 < t 6 2.2, 


 
 0.0280509 if 2.1 < t 6 2.2,

0.0217196 if 2.2 < t 6 2.3, 


 


 
 0.0264697 if 2.2 < t 6 2.3,

0.0202876 if 2.3 < t 6 2.4, 


 
 0.0241936 if 2.3 < t 6 2.4,

0.0181433 if 2.4 < t 6 2.5, 


 

0.0158644 if 2.5 < t 6 2.6, 0.0214619 if 2.4 < t 6 2.5,
H1/2 (t) > and H4/7 (t) > 0.0183875 if 2.5 < t 6 2.6,

0.0129923 if 2.6 < t 6 2.7, 


 
 0.0149960 if 2.6 < t 6 2.7,

0.0100686 if 2.7 < t 6 2.8, 


 
 0.0117724 if 2.7 < t 6 2.8,

0.0078162 if 2.8 < t 6 2.9, 


 


 
 0.0094724 if 2.8 < t 6 2.9,

0.0072943 if 2.9 < t 6 3.0, 


 
 0.0090024 if 2.9 < t 6 3.0,
0 else, 

0 else.

As such, in [19, Theorem 3] Wu obtained π2 (x)/Π(x) . (7/2)(1 − H4/7 (2.1)) 63.39951.

7
To complete our proof of Theorem 1.2 we apply Lemma 6.4, now valid up to level x 12 by
Corollary 2.7 and Proposition 5.5 for λ e+ . Note when the largest integration variable is t > 1 ,
5
2−t 7
we have θ(t) = 3 6 12 iff t > 41 . A key feature we use to satisfy the conditions of Lemma
6.4 is that the level xθ(t1 ) persists, since the largest prime p1 = xt1 is preserved through
successive iterations.
Thus in practice, Lemma 6.4 simply amounts to modifying the integral in G2 by substitut-
γ
ing F (s) − 2es Hθ (s) in for F (s), s = (θ(t) − t − u)/u, when t > 41 (the only parameter > 41 is
ρ, so we only refine G2 ). Denote this as GWu 2 . For ease we also use Hθ (s) >H4/7 (s), by mono-
tonicity in θ. Doing so, with the same parameter choices (6.21), we obtain GWu 2 6 −5.598667
and hence
 X X 
Π(x) Wu 1
(6.24) π2 (x) . G2 + Gn + G( 5 ) In . 3.299552 Π(x).
5eγ 9 6 n 6 21
1 6n6=2 68

This completes the proof of Theorem 1.2.

33
 For slight numerical gains, one may compute Hθ (s) when θ ∈ [ 47 , 12 7
], by tweaking the
formulae in [19]. More substantially, Wu defined a lower bound savings hθ (s), for a substi-
γ
tution of f (s) by f (s) + 2es hθ (s). But in practice, to compute h would require derivations
(analogous to H) of as yet undetermined formulae. We leave these to the reader.

n G n n In n In
1 39.00163 9 0.0332157 17 0.000315
2 −5.591009 10 0.0228322 18 0.000269
3 −3.986553 11 0.0092564 19 0.000164
4 −5.060499 12 0.0150101 20 6 2.70 · 10−6
5 1.864133 13 0.0547244 21 6 5.50 · 10−9
6 0.741181 14 0.0260202
7 0.453663 15 0.0124636
8 0.923736 16 0.0001314
References
1. E. Bombieri, H. Davenport, Small differences between prime numbers, Proc. Roy. Soc. Ser. A 239 (1966),
1–18.
2. E. Bombieri, J. Friedlander, H. Iwaniec, Primes in arithmetic progressions to large moduli, Acta Math.
156 (1986), 203–251.
3. Y.C. Cai, M.G. Lu, On the upper bound for π2 (x), Acta Arith. 110 (2003), 275–298.
4. J.R. Chen, On the representation of a larger even integer as the sum of a prime and the product of at
most two primes, Sci. Sinica 16 (1973), 157–176.
5. J.R. Chen, On the Goldbach’s problem and the sieve methods, Sci. Sinica 21 (1978), 701–739.
6. P. D. T. A. Elliott, H. Halberstam, A conjecture in prime number theory, Symposia Mathematica, Vol.
IV, Academic Press, London, (1970), 59–72.
7. E. Fouvry, Autour du théorème de Bombieri–Vinogradov, Acta Math. 152 (1984), 219–244.
8. E. Fouvry, H. Iwaniec. Primes in arithmetic progressions, Acta Arith., 42 (1983), 197–218
9. E. Fouvry, F. Grupp, On the switching principle in sieve theory, J. reine angew. Math. 370 (1986),
101–125.
10. J. Friedlander, H. Iwaniec, Opera de Cribro Amer. Math. Soc. Colloquium Publications, 57 (2010).
11. C. D. Pan, A new application of the Yu. V. Linnik large sieve method, Chinese Math. Acta 5 (1964),
642–652.
12. G. H. Hardy, J. E. Littlewood, Some Problems of ‘Partitio Numerorum.’ III. On the Expression of a
Number as a Sum of Primes, Acta Math. 44 (1923), 1–70.
13. H. Iwaniec. A new form of the error term in the linear sieve, Acta Arith., 37 (1980), 307–320.
14. H. Iwaniec, E. Kowalski. Analytic number theory, Amer. Math. Soc. Colloquium Publications, 53 (2004).
15. J. Maynard, Primes in arithmetic progressions to large moduli II: well-factorable estimates, preprint
(2020) arXiv:2006.07088
16. H. Riesel, R.C. Vaughan, On sums of primes, Ark. Mat. 21 (1983), 45–74.
17. A. Selberg, On elementary methods in prime number theory and their limitations, in: 11 Skand. Mat.
kongr., Trondheim 1949, 13–22.
18. J. Wu, Sur la suite des nombres premiers jumeaux, Acta Arith. 55 (1990), 365–394.
19. J. Wu, Chen’s double sieve, Goldbach’s conjecture and the twin prime problem, Acta Arith. 114 (2004),
215–273.

Mathematical Institute, University of Oxford, Oxford, OX2 6GG, UK

Email address: [email protected]

34