small gaps between primes - boun.edu.tr · 1 small gaps between primes 2 3 d. a. goldston, j....
TRANSCRIPT
Small gaps between primes1
2
D. A. Goldston, J. Pintz, and C. Y. Yıldırım3
Abstract. This paper describes the authors’ joint research on small gaps between primes in the last4
decade and how their methods were developed further independently by Zhang, Maynard, and Tao to5
prove stunning new results on primes. We now know that there are infinitely many primes differing by6
at most 246, and that one can find k primes a bounded distance (depending on k) apart infinitely often.7
These results confirm important approximations to the Hardy–Littlewood Prime Tuples Conjecture.8
Mathematics Subject Classification (2010). Primary 11N05, 11N36; Secondary 11N35.9
Keywords. Hardy–Littlewood prime tuples conjecture, prime numbers, sieves, gaps between primes,10
twin primes.11
1. History12
The twin prime conjecture that n and n + 2 are both primes for infinitely many positive13
integers n, may have been conceived around the time of Euclid, more than two thousand years14
ago. Among as yet unsolved problems in mathematics it is one of the oldest. The purpose of15
the present article is to give an overview of the progress in the last nine years in this subject,16
in particular, of the results of the authors.17
As a young boy Gauss observed in 1792 or 1793 that the primes around x have an average18
distance log x which led him to conjecture that19
π(x) :=∑p≤xp: prime
1 ∼ li x :=
x∫2
dt
log t∼ x
log x(x→∞). (1.1)
This conjecture was proved in 1896 (independently) by Hadamard and de la Vallée Poussin,20
and is now called the Prime Number Theorem.21
A relevant quantity in the study of small gaps between primes is22
∆ := lim infn→∞
dnlog n
= lim infn→∞
pn+1 − pnlog n
, (1.2)
where pi∞i=1 =: P is the set of primes sequenced in increasing order and dn := pn+1 −23
pn. The Prime Number Theorem, (1.1), immediately implies ∆ 6 1, so the first task24
concerning an upper estimation of ∆ was to show an estimate of the type ∆ < 1. During the25
twentieth century there were many papers on upper estimates for ∆. First, in 1926, Hardy and26
Proceedings of International Congress of Mathematicians, 2014, Seoul
2 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
Littlewood (unpublished, see [32]) succeeded in showing, assuming the Generalized Riemann27
Hypothesis (GRH), that28
∆ 6 2/3. (1.3)
The first unconditional bound29
∆ 6 1− c1, (1.4)
with an unspecified but explicitly calculable c1 > 0, was shown by Erdos in 1940 [5] who30
used Brun’s sieve. The next big step was made by Bombieri and Davenport [2] who removed31
the assumption of GRH in Hardy and Littlewood’s method by using Bombieri’s work [1] on32
the large sieve and showed that33
∆ 6(
2 +√
3)/8 = 0.466 . . . . (1.5)
Their method gave ∆ 6 1/2 but they were also able to combine this with an explicit version34
of Erdos’s [5] proof which led them to (1.5). After several smaller improvements (Huxley and35
others), Maier [23] succeeded in combining the matrix method he developed with the ideas36
of Bombieri–Davenport, Erdos and Huxley, making it possible to multiply the best known37
bound by e−γ (γ is Euler’s constant) and reach38
∆ 6 0.248 . . . . (1.6)
In 2005 the authors proved (see [14]; or for a brief account §2, §3 below)39
∆ = 0. (1.7)
2. Ideas behind the proofs of some results concerning small gaps between con-40
secutive primes41
We begin by recounting a number of conjectures related to the twin prime conjecture and42
more generally to small gaps between consecutive primes. Some of them have been known43
for a long time, some of them were introduced by us.44
Conjecture 2.1 (Twin Prime Conjecture). dn = 2 infinitely often.45
A generalization of this was formulated in 1849 by de Polignac.46
Conjecture 2.2 (De Polignac’s Conjecture [29]). For every given positive even integer h,47
dn = h infinitely often.48
For a further generalization we need the notion of admissible k-tuples.49
Definition 2.3. H = hiki=1 (0 6 h1 < h2 · · · < hk, hk ∈ Z) is admissible if the hi’s do50
not cover all residue classes mod p for any prime p.51
This is clearly a necessary condition that n+ hi ∈ P for all integers 1 ≤ i ≤ k holds for52
infinitely many numbers n.53
Dickson formulated in 1904 the conjecture that this condition was also sufficient. Although54
his conjecture included linear forms of type ain+ bi (ai, bi ∈ Z) we will consider the special55
case ai = 1 for all i ∈ [1, k].56
Small gaps between primes 3
Conjecture 2.4 (Dickson’s Conjecture [3]). If H is admissible, then n + hi ∈ P for all57
i ∈ [1, k] holds for infinitely many values of n.58
About twenty years later, in 1923, Hardy and Littlewood formulated this in a quantitative59
form as60
Conjecture 2.5 (Hardy–Littlewood Prime-Tuples Conjecture [20]). If H is an admissible61
k-tuple, then62 ∑n6x
n+hiki=1∈Pk
1 ∼ S(H)x
logk x, (2.1)
where63
S(H) :=∏p
(1− νH(p)
p
)(1− 1
p
)−k> 0, (2.2)
and νH(p) denotes the number of distinct residue classes modulo p occupied by the elements64
ofH.65
Note that the relation S(H) > 0 is equivalent toH being admissible.66
Until now the conjectures were listed in increasing strength. We introduced a weaker67
form of Dickson’s Conjecture:68
Conjecture 2.6 (Conjecture DHL(k, 2)). IfH is an admissible k-tuple, then n+H contains69
at least two primes infinitely often.70
If the above conjecture is true for at least one admissible k-tuple, then it implies another71
conjecture which is a good approximation to the Twin Prime Conjecture. This we called the72
Conjecture 2.7 (Bounded Gaps Conjecture). There exists an absolute constant C such that73
dn = pn+1 − pn 6 C for infinitely many n.74
A still weaker form of the Bounded Gap Conjecture is75
Conjecture 2.8 (Small Gaps Conjecture). ∆ = lim infn→∞
pn+1 − pnlog pn
= 0.76
Within the scope of our work the existence of small or bounded gaps between consecutive77
primes is intimately connected with the distribution of primes in arithmetic progressions. The78
following definition of an admissible level ϑ of primes was already known and used in sieve79
theory.80
Definition 2.9. ϑ is called an admissible level of distribution of primes if for any ε > 0,81
A > 0 we have for any X > 282 ∑q6Xϑ−ε
maxa
(a,q)=1
∣∣∣∣ ∑p≡a(q)p6X
log p− X
ϕ(q)
∣∣∣∣ 6 C(A, ε)X
(logX)A, (2.3)
where C(A, ε) is an ineffective constant depending on A and ε.83
The largest known level ϑ = 1/2 is the celebrated Bombieri–Vinogradov [1, 38] Theorem.84
The strongest possibility, ϑ = 1, is the Elliott–Halberstam [4] Conjecture, and more generally85
one can introduce86
4 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
Conjecture 2.10 (Conjecture EH(ϑ)). (2.3) is true for a fixed ϑ ∈ ( 12 , 1].87
We succeeded in showing in 2005 the following result.88
Theorem 2.11 ([14]). If EH(ϑ) is true for some fixed ϑ > 1/2, then DHL(k, 2) is true for89
k > k0(ϑ) and consequently the Bounded Gaps Conjecture is true, i.e. lim infn→∞
dn <∞.90
Theorem 2.12 ([14]). The Small Gaps Conjecture is true, i.e. ∆ = 0.91
We improved this somewhat later to92
Theorem 2.13 ([15]). lim infn→∞
dn(log n)1/2(log log n)2
<∞.93
Concerning the frequency of small gaps we showed94
Theorem 2.14 ([17, 18]). Given any fixed η > 0 the relation95
dn = pn+1 − pn < η log n (2.4)
holds for a positive proportion of all gaps.96
One of the important ideas which yielded a proof of the Small Gaps Conjecture in [14] and97
which – along with the work of Y. Motohashi and J. Pintz [25] – represented an important step98
in the first proof of the Bounded Gaps Conjecture by Y. Zhang [39] was to attack, among the99
listed seven conjectures, particularly DHL(k, 2). The idea was to find suitable non-negative100
weights an for n ∈ [N, 2N) to be abbreviated later as n ∼ N , such that an should be101
relatively large compared with S =∑n∼N
an > 0 if the set102
n+Hk = n+ hiki=1 (2.5)
contains some (possibly several) primes. A good quantitative formulation is to consider (and103
try to maximize) the ratio104
Ej =SjS∗
:=
∑n∼N
anχP(n+ hj) log(n+ hj)∑n∼N
an log 3N, (2.6)
where χP(m) denotes the characteristic function of primes, that is, χP(m) = 1 if m is prime105
and 0 otherwise.106
The quantity107
α(Hk) =
k∑j=1
Ej (2.7)
describes the (weighted) average number of primes in n+Hk if n runs between N and 2N ,108
i.e. n ∼ N . If we succeed in obtaining for a k-tuple H = Hk a lower bound greater than 1109
for the quantity in (2.7), then DHL(k, 2) is proved (at least for a singleH = Hk), and from110
this the Bounded Gaps Conjecture follows immediately.111
(i) If we start with the simple uniform choice an ≡ 1 we obtain112
α(Hk) ∼ k
logNas N →∞, (2.8)
which clearly tends to 0.113
Small gaps between primes 5
(ii) Choosing an = 1 if n+ hiki=1 ∈ Pk and 0 otherwise, we can seemingly reach the114
optimal value115
α(Hk) = k unless S =∑n∼N
an = 0. (2.9)
Unfortunately, to exclude the possibility S = S(N) = 0 for N > N0 is equivalent to116
the proof of Dickson’s Conjecture, so we arrive at a tautology.117
In the followingH = Hk will always be an admissible k-tuple, but to simplify notation118
we often write simplyH instead ofHk.119
(iii) An essentially equivalent formulation of the above is to use the generalized von Man-120
goldt function121
an = Λk(PH(n)
):=
∑d|PH(n)
µ(d)
(log
PH(n)
d
)k, PH(n) =
k∏i=1
(n+ hi) (2.10)
which vanishes if PH(n) has more than k distinct prime factors. However, in this case122
a direct evaluation of S seems to be hopeless, since d can be as large as Nk.123
(iv) It was an idea of Selberg to approximate (2.10) with the divisors cut at R = N c and124
accordingly use125 ∑d|PH(n)d6R
µ(d) logkR
d. (2.11)
However, this might be negative.126
(v) So the next idea is the weight used in the so-called k-dimensional Selberg sieve, i.e.,127
simply the square of (2.11), namely,128
an,k =
( ∑d6R,d|PH(n)
µ(d) logkR
d
)2
. (2.12)
In this case choosing R 6 N1/2L−A, L = logN , A > A0(k), S can be readily129
evaluated. Assuming EH(ϑ), the unconditional case being EH(1/2) (the Bombieri–130
Vinogradov Theorem), the more difficult sum Sj can also be evaluated, but only under131
the stronger constraint132
R 6 N (ϑ−ε)/2. (2.13)
This yields for the crucial quantity α(Hk) in (2.7)133
α(Hk) = ϑ− ε+O
(1
k
)(2.14)
primes on average, which is unfortunately still less than 1 even under the strongest134
hypothesis ϑ = 1, the original Elliott–Halberstam Conjecture.135
(vi) The winning choice is if we are more modest and instead of Dickson’s Conjecture136
approximate the situation whenk∏i=1
(n+ hi) has at most k + ` different prime factors137
where ` > 0 is a free parameter. (The choice ` = 1 was used earlier by Heath-Brown138
6 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
[22], however, not to localize primes in n+H but to find n values where all components139
n+ hi are almost primes). This means that we use (2.12) with k + ` instead of k, i.e.140
our choice in [14] was141
an,k+` =
( ∑d6R,d|PH(n)
µ(d) logk+`R
d
)2
. (2.15)
This yielded under the condition (2.13) a gain of a factor 2, rather surprisingly. More142
precisely we got143
α(Hk) = 2(ϑ− ε) +O
(`
k
)+O
(1
`
). (2.16)
Under the optimal choice ` =[√
k/2]
this meant144
α(Hk) = 2(ϑ− ε) +O
(1√k
). (2.17)
Consequently if EH(ϑ) is true for some ϑ > 1/2 we obtain α(Hk) > 1 primes on145
average if k > C/(ϑ− 1/2)2.146
In the unconditional case ϑ = 1/2, this yielded Theorem 2.12 but missed the goal147
DHL(k, 2) by a hairbreadth.148
The way to see how this argument could lead to a proof of the Small Gaps Conjecture149
begins by observing that on average only(
2ε+c1√k
)primes were “missing” to obtain150
more than one prime on average. Using all numbers of the form151
n+ h, h ∈ [1, H], H = η logN (2.18)
with an arbitrarily small but fixed η > 0 instead of only152
n+ hi, hi ∈ Hk (2.19)
we could pick up more primes so as to fill the missing part.153
If in case of h ∈ [1, H]\Hk we expect heuristically n+h to be prime with a probability154
1/ logN , we can hope to collect155
η > 2ε+c1√k
+O
(k
logN
)(2.20)
primes among n+ h on average if n ∼ N , h ∈ [1, H] \ Hk.156
The condition (2.20) is clearly satisfied if157
ε <η
3, k > C2η
−2, N > N0(k, ε, η). (2.21)
Small gaps between primes 7
In the original work [14] we used a result of Gallagher [11] and an averaging procedure158
over allHk ⊂ [1, H] to show that the above sketched heuristic works in practice. In the next159
section we use a simpler way, which avoids Gallagher’s Theorem and uses a single, suitably160
chosen k-tupleHk for all k.161
We will not sketch the rather complicated procedure to show Theorem 2.13. We just162
mention here that it needs the investigation of k-tuples with163
k (logN)1/2
(log logN)2, `
√k. (2.22)
In the work [27] it was shown that using a suitable polynomial P (x) instead of the simple164
xk+` in (2.15) (x = log(R/d)) one can improve Theorem 2.13 further to165
Theorem 2.15 ([27]). lim infn→∞
dn(logN)3/7(log logN)4/7
<∞.166
One can raise the more general question of finding the optimal polynomial, or more167
generally the optimal function P (x). B. J. Conrey calculated the optimal weight function,168
actually a Bessel-type function. Later in the work [10] an exact analysis confirmed the169
optimality of the Bessel-type function and the fact that it yielded instead of (2.17) the sharper170
estimate171
α(Hk) = 2(ϑ− ε) +O(k−2/3). (2.23)
This was, however, the same strength as the polynomial in [27] and [10] apart from the implicit172
constant in the above O symbol. Therefore the result in Theorem 2.15 can be considered as173
the limit of the original GPY method.174
Concerning Theorem 2.14 the crucial idea is the fact, discovered by the second named175
author ([26]), and independently by Friedlander and Iwaniec [9] that the weights an are176
strongly concentrated on numbers n where all components n+ hi are almost prime, more177
precisely for numbers n with178
P−( k∏i=1
(n+ hi)
)> Nδ, n ∼ N, (2.24)
where δ is an arbitrarily small fixed positive constant and P−(m) denotes the smallest prime179
factor of n. In fact, it was proved in [26] that180 ∑n∼N
P−(PH(n))6Nδ
an 6 Cδ∑n∼N
an (2.25)
with a constant C = C(k). (The factor C(k)δ was improved to C ′k3δ2 with an absolute181
constant C ′ in [17]).182
3. Sketch of the proof of Theorems 2.11 and 2.12183
In the following we consider a general sieve situation when the number of residues sieved out184
mod p satisfies185
ΩH(p) = Ω(p) = k for p - ∆(H) :=∏i>j
(hi − hj), k fixed (3.1)
8 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
and let Ω(n) be extended multiplicatively for all squarefree values of n. Actually we have186
Ω(p) = ΩH(p) = νH(p). There are three possibilities:187
(i) to work analytically with two complex variables (cf. [14]);188
(ii) to work elementarily (cf. [13] using pure sieve methods beyond (2.3));189
(iii) to work partially elementarily and partially analytically with one complex variable.190
Here we will pursue the third possibility, worked out in an unpublished note of K.191
Soundararajan [37].192
We use a somewhat more general weight function: a polynomial P (y) but note that the193
argument would work the same for a function P (y) analytic on [0, 1], if P (y) has at least a194
kth order zero at 0.195
First we evaluate the sum of the weights an, where in the following we will define196
an =
∑d6R
µ(d)P
(log(R/d)
logR
)2
, (3.2)
197
S =∑n∼N
an ∼ N∑′
d,e6R
µ(d)µ(e)Ω([d, e])
[d, e]P
(log(R/d)
logR
)P
(log(R/e)
logR
)(3.3)
(we ignored a negligible error of size O(R2+ε)) and∑′ will always denote summation over198
squarefree variables).199
Introducing the notation (d, e) = u, d = um, e = un, (m,n) = 1 we obtain200
S ∼N∑′
u6R
∑′
m,n6R/u(m,n)=1
(m,u)=(n,u)=1
µ(m)µ(n)Ω(u)Ω(m)Ω(n)
umnP
(log(R/um)
logR
)P
(logR/un
logR
).
(3.4)We can rewrite the condition (m,n) = 1 using the relation201
∑β|m,β|n
µ(β) =
1 if (m,n) = 1,
0 otherwise(3.5)
as202
S ∼ N∑′
u6R
∑′
β6R/u
µ(β)Ω(u)Ω2(β)
uβ2
( ∑′
m′6R/uβ(m′,u)=1
µ(βm′)Ω(m′)
m′P
(log(R/uβm′)
logR
))2
.
(3.6)Grouping terms with the same value of uβ =: γ with notation m = m′ we have203
S ∼ N∑′
γ6R
Ω(γ)
γ
(∑′
β|γ
µ(β)Ω(β)
β
)( ∑′
m6R/γ(m,γ)=1
µ(m)Ω(m)
mP
(log(R/γm)
logR
))2
. (3.7)
Small gaps between primes 9
Let us denote the inner sum by J(γ, Rγ
)where the first variable refers to the condition204
(m, γ) = 1, the second to m 6 R/γ. Further let for a squarefree γ205
G(s+ 1, γ) :=∑′
m(m,γ)=1
µ(m)Ω(m)
ms+1=: ζ(s+ 1)−kF (s+ 1, γ). (3.8)
Here we have for Re s > 0206
F (s+ 1, γ) =∏p
(1− Ω(p)
ps+1
)(1− 1
ps+1
)−k∏p|γ
(1− Ω(p)
ps+1
)−1. (3.9)
Using the Taylor expansion207
P (x) =
∞∑j=k
P (j)(0)xj
j!(3.10)
and Perron’s formula (c > 0, arbitrary)208
1
2πi
∫(c)
xs
sj+1ds =
(log x)j
j! if x > 1,
0 if 0 6 x 6 1(j ∈ Z+) (3.11)
we can rewrite J(γ, Rγ
)as
J
(γ,R
γ
)=
∞∑j=k
P (j)(0)
(logR)j
∑′
m6R/γ(m,γ)=1
µ(m)Ω(m)
m
1
j!
(log
R/γ
m
)j(3.12)
=
∞∑j=k
P (j)(0)
(logR)j· 1
2πi
∫(c)
∞∑m=1
(m,γ)=1
µ(m)Ω(m)
ms+1
(R
γ
)sds
sj+1
=
∞∑j=k
P (j)(0)
(logR)j· 1
2πi
∫(c)
F (s+ 1, γ)ζ(s+ 1)−k(R
γ
)sds
sj+1.
Since F (s+ 1, γ) is regular for σ > − 12 we can transform the line inside the zero-free
region of ζ(s+ 1), that is, to σ > 1− c/(log(|t|+ 2)), |t| 6 exp(√
logR). The integral is
negligible on the new contour and so we obtain by the residue at s = 0
J
(γ,R
γ
)∼∞∑j=k
P (j)(0)
(logR)jF (1, γ)
(logR/γ)j−k
(j − k)!(3.13)
=F (1, γ)
(logR)k
∞∑ν=0
P (ν+k)(0)
ν!
(log(R/γ)
logR
)ν=
F (1, γ)
(logR)kP (k)
(logR/γ
logR
).
10 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
We remark that although this argument does not work if R/γ is not large enough, thatpart can be shown to be negligible directly from (3.7). So we obtain
S ∼ N
(logR)2k
∑′
γ6R
Ω(γ)
γ
∏p|γ
(1− Ω(p)
p
)· F (1, γ)2
(P (k)
(logR/γ
logR
))2
(3.14)
∼ N
(logR)2kS2(H)
∑′
γ6R
Ω(γ)
γ
∏p|γ
(1− Ω(p)
p
)−1(P (k)
(logR/γ
logR
))2
.
Since apart from finitely many primes, for which209
p | ∆(H) :=∏i>j
(hi − hj) (3.15)
we have Ω(p) = k, the behaviour of Ω(n) is similar to that of the generalized divisor function210
211
τk(n) =∑
n1n2...nk=n
1. (3.16)
This implies (for the details see Lemma 11 of [13])212
∑′
γ6x
Ω(γ)
γ
∏p|γ
(1− Ω(p)
p
)−1∼ S(H)−1
(log x)k
k!. (3.17)
The sum in (3.14) can be evaluated from (3.17) by partial summation, and we obtain213
S ∼ S(H)N
(logR)k(k − 1)!
1∫0
yk−1(P (k)(1− y)
)2dy. (3.18)
Let us consider now the quantity214
Sj =∑′
n∼NanχP(n+ hj) log n, hj ∈ H. (3.19)
In this case (if R < N ) the two conditions215
n+ hj ∈ P, d |k∏i=1
(n+ hi), d 6 R (3.20)
and216
n+ hj ∈ P, d |k∏i=1i 6=j
(n+ hi), d 6 R (3.21)
are equivalent. So the situation is similar to (3.3) if217
R 6 N (ϑ−ε)/2 (3.22)
Small gaps between primes 11
since it is easy to see that by the condition (2.3) (which is unconditionally true with ϑ = 1/2218
by the Bombieri–Vinogradov Theorem) we can substitute χP(n+ hj) log n by 1. Thus we219
have220
Sj ∼∑′
d,e6R
µ(d)µ(e)Ωj([d, e])
[d, e]P
(log(R/d)
logR
)P
(log(R/e)
logR
)(3.23)
with the only difference that we have now Ωj(p) = Ω(p)− 1 = k − 1 if p - ∆. The singularseries Sj(H) is accordingly
Sj(H) =∏p
(1− νH(p)− 1
p− 1
)(1− 1
p
)−(k−1)(3.24)
=∏p
(1− νp(H)
p
)(1− 1
p
)−k= S(H).
So we obtain for all j ∈ [1, k] under the stronger condition (3.22) now analogously to (3.18)221
Sj ∼S(H)N
(logR)k−1(k − 2)!
1∫0
yk−2(P (k−1)(1− y)
)2dy (3.25)
and this gives in total for R = N (ϑ−ε)/2, P (k−1)(x) = Q(x)222
k∑j=1
Sj
S log 3N∼ logR
logNk(k − 1)M(Q) ∼ k(k − 1)(ϑ− ε)
2M(Q) (3.26)
primes on average in n + hiki=1 if n runs between N and 2N and the numbers n are223
weighted by an log n, where224
M(Q) =
1∫0
yk−2(Q(1− y)
)2dy
1∫0
yk−1(Q′(1− y)
)2dy
. (3.27)
In case of the simple choice225
P (x) = xk+`, ` =[√
k/2]⇔ Q(x) = C(k, `)x`+1 (3.28)
we obtain
M(Q) =
1∫0
yk−2(1− y)2`+2dy
(`+ 1)21∫0
yk−1(1− y)2`dy
=(k − 2)!(2`+ 2)!/(k + 2`+ 1)!
(`+ 1)2(k − 1)!(2`)!/(k + 2`)!(3.29)
=4(
1− 12(`+1)
)(k + 2`+ 1)(k − 1)
∼4(
1−O(
1√k
))k2
.
12 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
By (3.26) this yields on the weighted average226
2(ϑ− ε)(
1−O(
1√k
))(3.30)
primes in n+H if n ∼ N .227
The quantity above is clearly greater than 1 if228
ϑ > 1/2, k > k0(ϑ), (3.31)
which proves Theorem 2.11.229
Suppose now h0 /∈ H, let H0 = H ∪ h0, and Ω0(p) = ΩH0(p) is defined as in (3.1)230
with k + 1 in place of k,231
S0 =∑n∼N
anχP(n+ h0) log n. (3.32)
In case of νH0(p) = νH(p) we have Ω0(p) = νH(p) − 1 residue classes in the sieve232
mod p (Ω0 is defined as in (3.1)); if νH0(p) = νH(p) + 1, then Ω0(p) = νH(p). So we have233
in both cases Ω0(p) = νH0(p)− 1 and Ω0(p) = k if p - ∆(H0).234
This yields an analogous asymptotic to (3.18) for S0, withH replaced byH0:235
S0 ∼S(H0)N
(logR)k(k − 1)!
1∫0
yk−1(P (k)(1− y)
)2dy (3.33)
and consequently236
S0
S∼ S (H ∪ h0)
S(H)(as N →∞). (3.34)
This relation helps us to obtain Theorem 2.12 unconditionally. Let us consider an interval237
of length238
H = η logN, (3.35)
where η is an arbitrarily small fixed positive constant. Let us suppose that we can find for any239
k an admissible k-tupleH = Hk such that with a fixed absolute constant c0 > 0240
S (Hk ∪ h0) > c0S(H) for any even h0. (3.36)
In this case using only ϑ = 1/2, that is, the Bombieri–Vinogradov Theorem, we obtain on241
average242
H∑h=1
∑n∼N
anχP(n+ h) log n∑n∼N
an log 3N> (1− 2ε)
(1−O
(1√k
))+c0η
2− k
log 3N> 1 (3.37)
primes between n and n+H if243
k > k0(η), ε < ε0(η), N > N0(η, k, ε). (3.38)
Small gaps between primes 13
In order to show the existence ofHk with (3.36) we can just choose244
H = Hk =i∏p62k
pki=1
. (3.39)
Then we have for any even h with νp = νH(p)
S(H ∪ h)
S(H)> 2
∏2<p62k
1− 2/p
(1− 1/p)2
∏p>2k
1− (νp + 1)/p
1− (νp + 1)/p+ νp/p2(3.40)
> c1∏p>2k
(1 +O
(k
p2
))> c0.
In such a way we obtain Theorem 2.12. We remark that the above proof avoids Gallagher’s245
Theorem [11]. Another proof, also avoiding Gallagher’s Theorem is given in [16] which yields246
some other results, like small gaps between consecutive primes in arithmetic progressions247
and improved upper estimates for the quantity248
∆r = lim infn→∞
pn+r − pnlog pn
. (3.41)
4. Sketch of the proof of Theorem 2.14249
The most crucial idea in the proof of Theorem 2.14 is that we will change the weights and250
instead of the original normalized weights (cf. (2.15)).251
an =
( ∑d≤R,d|PH(n)
µ(d)
(log(R/d)
logR
)k+`)2
, PH(n) =
k∏i=1
(n+ hi), ` =
[√k
2
](4.1)
we will work with the new weight (n ∼ N )252
a′n =
an if P−
(PH(n)
)> Nδ,
0 otherwise,(4.2)
where δ will be a fixed small positive constant with ε < ε0(η), k > k0(η, ε), δ < δ0(k, η, ε),253
R = N (ϑ−ε)/2 and we consider primes in intervals of length254
H = η logN (4.3)
as indicated in (2.2).255
As mentioned at the end of Section 2 the sum of weights a∗n,H with PH(n) having at least256
one small prime divisor not exceeding Nδ is negligible and we have (2.25) with a constant257
C = C(k), i.e.258
0 ≤∑n∼N
(an − a′n) =∑n∼N
P−(PH(n))≤Nδ
an ≤ Cδ∑n∼N
an,
∑n∼N
P−(PH(n))≤Nδ
anχP(n+ h) log(n+ h) ≤ Cδ∑n∼N
anχP(n+ h) log(n+ h).(4.4)
14 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
These are Lemmas 4 and 5 of [26].259
The other tool is Gallagher’s Theorem [11], according to which for k fixed, H →∞260 ∑H⊂[1,H]|H|=k
S(H) ∼ Hk
k!. (4.5)
Let further (for a more detailed proof see [17] and [18])
π(n,H) := π(n+H)− π(N), Θ(n) :=
log n if n ∈ P,0 otherwise,
(4.6)
Θ(n,H) :=
H∑h=1
Θ(n+ h)
261
M :=∑pj∼N
pj+1−pj≤H
1, Q(N,H) :=∑n∼N
π(n,H)>1
1 ≤ HM +O(Ne−c
√logN
), (4.7)
and consider now instead of (3.19) the modified quantity262
S′(h,H) =∑n∼N
a′nΘ(n+ h). (4.8)
The substitution of an by a′n will just slightly change the corresponding value of S′(H)263
and S′(h,H) respectively, to264
S′(H) =∑n∼N
a′n = (1 +O(δ))S(H), (4.9)
265
S′(h,H) =∑n∼N
a′nΘ(n+ h) = (1 +O(δ))S(h,H) (4.10)
compared with266
S(h,H) :=∑n∼N
anΘ(n+ h), (4.11)
where the asymptotics for the quantity (4.11) are given in (3.25) and (3.33) respectively, and267
P (x) = xk+` in this section.268
The crucial change is that in case of a′n,H > 0 all the prime divisors of PH(n) are at least269
Nδ with a fixed small δ, so by (4.1) we have a trivial estimate for it:270
a′n 6 2ω(PH(n)) 6 22k2/δ k,δ 1. (4.12)
On the other hand, in this case we cannot use the simplification of Section 3, that is, to271
work with a suitably chosen single Hk. Averaging over all H ⊆ [1, H], |H| = k, with the272
abbreviations (we take the unconditional case ϑ = 1/2 from now on)273
H
logR=
η(12 − ε
)/2
= η′,∑(k)
H=
∑H⊂[1,H]|H|=k
(4.13)
Small gaps between primes 15
we obtain from (3.18), using (3.28)–(3.29) and (4.5)274 ∑(k)
HS′(H) ∼ (1 +O(δ))
(η′)kNC(k, `)(2`)!
k!(`+ 1)2(k + 2`)!=: (1 +O(δ))B. (4.14)
On the other hand, we have by (3.33) and (4.5)∑(k)
H
∑n∼N
h∈[1,H]\H
a′nΘ(n+ h) (4.15)
∼ (k + 1)∑(k+1)
HS(H)
NC(k, `)(2`)!(1+O(δ))
(logR)k(`+ 1)2(k + 2`)!:= (1 +O(δ))Bη logN.
Finally, we have by (3.25) and (4.5) with ` =[√
k/2], ϑ = 1/2
∑(k)
H
∑h∈H
∑n∼N
a′nΘ(n+ h) (4.16)
∼ (1 +O(δ))kη′kNC(k, `)(2`+ 2)!
k!(k + 2`+ 1)!logR
∼ (1 +O(δ))B
(1− 1
2(`+ 1)
)(1− 2`+ 1
k + 2`+ 1
)(1− 2ε) logN.
Adding (4.15), (4.16) and subtracting from it (4.14) multiplied by log 3N we obtain∑(k)
H
∑n∼N
a′n(Θ(n,H)− log 3N
)(4.17)
> B logN
(1− 2ε)
(1− C√
k
)+ η − 1 +O(δ)
>η
2B logN
if, as stated in the introduction of Section 4 (between (4.2) and (4.3)) we fix ε, k, δ with275
ε < ε0(η), k > k0(η, ε), δ < δ0(k, η, ε). (4.18)
Consequently, if (4.18) holds, which we will always assume in the following, then276
η
2B logN < (1 + o(1)) logN
∑n∼N
π(n,H)>1
π(n,H)∑(k)
Ha′n. (4.19)
Introducing the notation277
T (n,H) :=∑(k)
HP−(PH(n)
)>Nδ
1 (4.20)
we have by (4.6)–(4.7), (4.12) and Cauchy’s inequality
ηB ( ∑
n∼Nπ(n,H)>1
1
)1/2(∑n∼N
π2(n,H)T (n,H)2)1/2
(4.21)
16 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
(
(HM)1/2 +O(N1/2e−c
√logN/2
))(∑n∼N
π2(n,H)T (n,H)2)1/2
.
Further, we have by Selberg’s sieve (Theorem 5.1 of [21] or Theorem 2 in § 2.2.2 of [19]) for278
any setH and δ < 1/2279 ∑n∼N
P−(PH(n)
)>Rδ
1 6|H|!S(H)
(logRδ)|H|N(1 + o(1)) (R,N →∞). (4.22)
This implies by Gallagher’s Theorem (4.5)∑n∼N
π(n,H)2T (n,H)2 ∑
1≤h,h′≤H
∑(k)
H1
∑(k)
H2
∑n∼N
P−(H1∪H2∪h∪h′)>Nδ
1 (4.23)
k N
2k+2∑r=k
∑(r)
H0
S(H0)
(logRδ)rk,δ N
2k+2∑r=k
(H
logR
)rk,δ (η′)kN.
Taking into account the definition of B in (4.14) we obtain from (4.21) and (4.23)280
η(η′)k/2 k,δ
((HM
N
)1/2
+ e−c√logN/2
). (4.24)
Consequently,281
HM
Nk,δ,η 1. (4.25)
Hence,282
M k,δ,ηN
logNk,δ,η π(2N), (4.26)
which proves Theorem 2.14.283
It may be shown (see Theorem 2 of [18]) that this is sharp in the sense that the assertion284
does not remain true if H = o(logN). The proof uses the Selberg sieve upper bound for285
prime tuples and Gallagher’s result (4.5).286
5. Bounded gaps between primes. Zhang’s theorem287
We recall that in our original work (Theorem 2.11 in Section 2) we showed that EH(ϑ) for288
any ϑ > 1/2 implies DHL(k, 2) for k > k0(ϑ), consequently the Bounded Gaps Conjecture.289
From the proof it is trivial that the condition290
maxa,(a,q)=1
(5.1)
in (2.3) can be weakened to291
maxa,(a,q)=1,PH(a)≡0(q)
(5.2)
Small gaps between primes 17
if we want to show for a specificH that n+H contains at least two primes infinitely often.292
However, in 2008 in a joint work of Y. Motohashi and J. Pintz the following stronger form of293
Theorem 2.11 was proved, in which the summation in (2.3) can be reduced to smooth moduli.294
P+(n) will denote the largest prime factor of n.295
Theorem 5.1 ([25]). If there exist δ > 0, ϑ > 1/2 and an admissible k-tuple H with296
k > k0(δ, ϑ) such that for any ε > 0, A > 0297 ∑q≤Nϑ−εP+(q)≤Nδ
maxa
(a,q)=1,q|PH(a)
∣∣∣∣ ∑p≡a(q)p∼N
log p− N
ϕ(q)
∣∣∣∣ ≤ C(A, ε)N
logAN(5.3)
holds for N > N0(H, ϑ, δ), then n+H contains at least two primes for some n ∼ N .298
Remark 5.2. Zhang proved a version of this result, and it appeared with a different proof in299
his work [39]. Zhang proved condition (5.3) with the explicit values300
ϑ =1
2+
1
584, δ =
1
1168, (5.4)
which finally led to301
Theorem 5.3 ([39]). DHL(k, 2) is true for k > 3.5 · 106 and consequently lim infn→∞
(pn+1 −302
pn) 6 C = 7 · 107.303
His proof of (5.4) uses several deep works of Fouvry, Fouvry–Iwaniec, Bombieri–Fried304
lander–Iwaniec, Friedlander–Iwaniec, Heath-Brown, which are based on ideas and works of305
Linnik, Weil, Deligne and Birch–Bombieri concerning the estimate of Kloostermann sums.306
The Polymath 8a project of T. Tao [30] introduced many improvements into this procedure307
(for example to apply instead of the simple weight function P (x) = xk+` the optimal308
Bessel function first used by Conrey, later analyzed in details in [10] together with many309
improvements in both the Motohashi–Pintz Theorem and in the estimation of Kloostermann310
sums) and obtained distribution estimates up to level 1/2 + 7/300, and thus reached311
Theorem 5.4 (Polymath 8a). DHL(k, 2) is true for k > 632 and consequently
lim infn→∞
(pn+1 − pn) 6 4680.
6. Bounded gaps between primes: The Maynard–Tao theorem312
About half a year after the manuscript of Zhang [39], simultaneously and independently, J.313
Maynard [24] and in his Polymath blogs T. Tao [31] introduced another idea which led to a314
new, more efficient proof of the Bounded Gaps Conjecture. The main results of Maynard [24]315
were the following.316
Theorem 6.1 (Maynard [24]). DHL(k, 2) is true for k ≥ 105, consequently
lim infn→∞
(pn+1 − pn) 6 600.
18 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
Theorem 6.2 (Maynard [24]). Assuming the Elliott–Halberstam Conjecture, DHL(k, 2) istrue for k ≥ 5, consequently
lim infn→∞
(pn+1 − pn) 6 12.
The two surprising aspects of the Maynard–Tao method were that it produced not only317
pairs but arbitrarily long (finite) blocks of primes in bounded intervals, and for this knowing318
that (2.3) holds with any fixed ϑ > 0 (however small) would suffice.319
The earlier known strongest result of somewhat similar nature was the much weaker one320
in our work [16]. It asserted for any r > 0321
∆r := lim infn→∞
pn+r − pnlog pn
≤ e−γ(√r − 1
)2. (6.1)
Further, under the very deep Elliott–Halberstam Conjecture (see (2.3) with ϑ = 1) we could322
show [14]323
∆2 = 0. (6.2)
Theorem 6.3 (Maynard–Tao [24]). We have for any r324
lim infn→∞
(pn+r − pn) r3e4r. (6.3)
The main idea of Maynard and Tao is that the weights are defined instead of325
an =
( ∑d6R
d|PH(n)
µ(d)P
(logR/d
logR
))2
, PH(n) =
k∏i=1
(n+ hi) (6.4)
in the more general form326
an =
( ∑d1...dk6Rdi|n+hi
µ(d)P
(log d1logR
, . . . ,log dklogR
))2
, (6.5)
where P (t1, . . . , tk) := Rk → R is a fixed piecewise differentiable function with support327
on t1 + t2 + · · ·+ tk 6 1. The idea of the use of these more general weights goes back to328
Selberg ([36], p. 245). Similar type of weights were used by Goldston and Yıldırım [12],329
but due to the special choice of P (t1, . . . , tk) =k∏i=1
(1− kti), ti 6 1/k, this led only to the330
result331
∆ = lim infn→∞
pn+1 − pnlog pn
61
4. (6.6)
We remark here that the general choice of P(
logR/dlogR
)in Section 3 corresponds to the332
special case of the above with333
P (t1, t2, . . . , , tk) = P (t1 + t2 + · · ·+ tk). (6.7)
Another very interesting remark is that in order to show bounded intervals with arbitrarily334
long finite blocks of primes (with a bound e2r/ϑ in place of e4r) we do not need the value335
Small gaps between primes 19
ϑ = 1/2, that is, the Bombieri–Vinogradov Theorem, just any value ϑ > 0. So we obtain a336
numerically slightly weaker form of the existence of arbitrarily long (finite) blocks of primes337
in bounded invervals even by the use of the first theorem establishing a positive admissible338
level ϑ for the distribution of primes, due to A. Rényi [33, 34] reached in 1947–48, by the339
large sieve of Linnik.340
Upon further work on the Maynard–Tao method in the Polymath 8b project of Tao,341
Theorem 6.1 has been improved to342
Theorem 6.4 (Polymath 8b project). DHL(k, 2) is true for k > 50, consequently
lim infn→∞
(pn+1 − pn) ≤ 246.
7. De Polignac numbers and some conjectures of Erdos on gaps between con-343
secutive primes344
There are various 60–70 years old conjectures of Erdos on which a sharpened version of345
Zhang’s Theorem (or that of Maynard and Tao) combined with other arguments of the second346
named author can give an answer. Below we give a list of them without proofs which can be347
found in [28]. The numerical values reflect the stage at the end of Polymath 8A.348
Using an argument of the second named author (Lemma 4 in [26]) together with a more349
general form of the arguments of Theorem 3 of Zhang and its improvement by Tao’s project,350
the following strengthening of Theorem 3 of Zhang can be shown. (Let P−(n) be the smallest351
prime factor of n.)352
Theorem 7.1 ([28]). Let k ≥ 632,H an admissible k-tuple, hi logN , N > N0(k). Then353
there are at least354
c1(k,H)N
logkN
numbers n ∈ [N, 2N) such that n+H contains at least two primes and almost primes in all355
other components satisfying P−(n+ hi) > N c2(k) for i = 1, 2, . . . , k.356
Remark 7.2. A similar version to the above-mentioned crucial Lemma 4 of [26] appears in357
the book Opera de Cribro of Friedlander–Iwaniec [9] published also in 2010.358
Whereas the original Theorem 3 of Zhang yields only one de Polignac number, by the aid359
of Theorem 7.1 we can show360
Theorem 7.3 ([28]). There are infinitely many de Polignac numbers. In fact, they have a361
positive lower density > 10−7.362
Theorem 7.4 ([28]). There exists an ineffective C such that we have always at least one363
de Polignac number between X and X + C for any X . (All gaps between consecutive de364
Polignac numbers are uniformly bounded.)365
Erdos [6] proved in 1948 the inequality366
lim infn→∞
dn+1
dn≤ 1− c0 < 1 + c0 ≤ lim sup
dn+1
dn(7.1)
with a very small positive value c0 and conjectured that the lim inf = 0 and the lim sup =∞.367
20 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
Theorem 7.5 ([28]). lim infn→∞
dn+1
dn= 0, lim sup
dn+1
dn=∞.368
Further, we have even369
lim infn→∞
dn+1 log n
dn<∞, lim sup
n→∞
dn+1
dn log n> 0. (7.2)
In general it is difficult to show anything for three consecutive differences. However, we370
can show371
Theorem 7.6 ([28]). lim supn→∞
min(dn−1, dn+1)
dn(log n)c=∞ with c = 1/632.372
Since the Prime Number Theorem implies373
1
N
N∑n=1
dnlog n
= 1, (7.3)
it is interesting to investigate the normalized distribution of the sequence dn, dn/ log n. Erdos374
conjectured 60 years ago that the set of limit points,375
J =
dn
log n
′= [0,∞], (7.4)
but no finite limit point was known until 2005, when we showed 0 ∈ J . (We denote by G′376
the set of limit points of the set G.) This was rather strange since in 1955 Erdos [7] and377
simultaneously Ricci [35] proved that J has positive Lebesgue measure. A partial answer to378
the conjecture of Erdos is379
Theorem 7.7 ([28]). There is an (ineffective) constant c∗ such that380
[0, c∗] ⊂ J. (7.5)
The above result raises the question whether considering a finer distribution dn/f(n) with381
a monotonically increasing function f(n) ≤ log n, f(n)→∞ the same phenomenon is still382
true. The answer is yes.383
Theorem 7.8 ([28]). Let f(n) ≤ log n, f(n)→∞ be an increasing function,384
Jf =
dnf(n)
′. (7.6)
Then there is an (ineffective) constant c∗f such that385
[0, c∗f ] ⊂ Jf . (7.7)
Zhang’s theorem shows the existence of infinitely many generalized twin prime pairs386
with a difference at most 7 · 107, while the theorem of Green and Tao shows the existence387
of arbitrarily long (finite) arithmetic progressions in the sequence of primes. A common388
generalization of these two results is given below. (Let p′ denote the prime following p.)389
Theorem 7.9 ([28]). There exists an even d ≤ 4680 with the following property. For any k390
there is a k-term arithmetic progression of primes such that p′ = p+ d for all elements of the391
progression.392
Small gaps between primes 21
Acknowledgements. The first author was supported in part by NSF Grant DMS-1104434.393
The second author was supported by OTKA Grants NK104183, K100291 and ERC-AdG.394
321104.395
References396
[1] Bombieri, E., On the large sieve, Mathematika 12 (1965), 201–225.397
[2] E. Bombieri and H. Davenport, Small differences between prime numbers, Proc. Roy.398
Soc. Ser. A 293 (1966), 1–18.399
[3] L. E. Dickson, An extension of Dirichlet’s theorem on prime numbers, Messenger of400
Mathematics 33 (1904), 155–161.401
[4] P. D. T. A. Elliott, H. Halberstam, A conjecture in prime number theory, Symposia402
Mathematica 4 (1968), 59–72.403
[5] P. Erdos, The difference of consecutive primes, Duke Math. J. 6 (1940), 438–441.404
[6] P. Erdos, On the difference of consecutive primes, Bull. Amer. Math. Soc. 54 (1948),405
885–889.406
[7] P. Erdos, Some problems on the distribution of prime numbers, Teoria dei Numeri, Math407
Congr. Varenna, (1954), 8 pp., 1955.408
[8] P. Erdos, P. Turán, On some new questions on the distribution of prime numbers, Bull.409
Amer. Math. Soc. 54 (1948), 371–378.410
[9] J. Friedlander, H. Iwaniec, Opera de cribro, American Mathematical Society Colloquium411
Publications, 57, American Mathematical Society, Providence, RI, 2010.412
[10] B. Farkas, J. Pintz, S. Révész, On the optimal weight function in the Goldston–Pintz–413
Yıldırım method for finding small gaps between consecutive primes, in: Paul Turán414
Memorial Volume: Number Theory, Analysis and Combinatorics, pp. 75–104, de415
Gruyter, Berlin, 2014.416
[11] P. X. Gallagher, On the distribution of primes in short intervals, Mathematika 23 (1976),417
4–9.418
[12] D. A. Goldston, C. Y. Yıldırım, Higher correlations of divisor sums related to primes,419
III. Small gaps between primes, Proc. London Math. Soc. (3) 96 (2007), 653–686.420
[13] D. A. Goldston, S. W. Graham, J. Pintz, C. Y. Yıldırım, Small gaps between primes or421
almost primes, Trans. Amer. Math. Soc. 361 (2009), no. 10, 5285–5330.422
[14] D. A. Goldston, J. Pintz, C. Yıldırım, Primes in Tuples I, Annals of Math. 170 (2009),423
819–862.424
[15] D. A. Goldston, J. Pintz, C. Yıldırım, Primes in Tuples II, Acta Math. 204 (2010), 1–47.425
[16] D. A. Goldston, J. Pintz, C. Yıldırım, Primes in Tuples III, Functiones et Approximat.426
35 (2006), 76–89.427
22 D. A. Goldston, J. Pintz, and C. Y. Yıldırım
[17] D. A. Goldston, J. Pintz, C. Yıldırım, Primes in Tuples IV: Density of small gaps428
between consecutive primes, Acta Arith. 155 (2012), No. 4, 395–417.429
[18] D. A. Goldston, J. Pintz, C. Yıldırım, Positive Proportion of Small Gaps Between430
Consecutive Primes, Publ. Math. Debrecen 79 (2011), no. 3-4, 433–444.431
[19] G. Greaves, Sieves in Number Theory. Springer, Berlin, Heidelberg, New York, 2001.432
[20] G. H. Hardy, J. E. Littlewood, Some problems of “Partitio Numerorum”, III: On the433
expression of a number as a sum of primes, Acta Math. 44 (1923), 1–70.434
[21] H. Halberstam, H.-E. Richert, Sieve Methods, Academic Press, London, 1974.435
[22] D. R. Heath-Brown, Almost-prime k-tuples, Mathematika 44 (1997), 245–266.436
[23] H. Maier, Small differences between prime number, Michigan Math. J. 35 (1988),437
323–344.438
[24] J. Maynard, Small gaps between primes, preprint.439
[25] Y. Motohashi, J. Pintz A smoothed GPY sieve, Bull. Lond. Math. Soc. 40 (2008), no. 2,440
298–310.441
[26] J. Pintz, Are there arbitrarily long arithmetic progressions in the sequence of twin442
primes? in: An Irregular Mind. Szemerédi is 70, Bolyai Soc. Math Studies, Vol. 21,443
Eds.: I. Bárány, J. Solymosi, pp. 525–559, Springer, 2010.444
[27] J. Pintz, Some new results on gaps between consecutive primes, in: Paul Turán Memorial445
Volume: Number Theory, Analysis and Combinatorics, pp. 261–278, de Gruyter, Berlin,446
2014.447
[28] J. Pintz, Polignac Numbers, Conjectures of Erdos on Gaps between Primes, Arithmetic448
Progression in Primes, and the Bounded Gap Conjecture, Preprint, arXiv:1305.6289,449
2013.450
[29] A. de Polignac, Six propositions arithmologiques déduites du crible d’Ératosthène,451
Nouv. Ann. Math. 8 (1849), 423–429.452
[30] D. H. J. Polymath, New equidistribution estimates of Zhang type, and bounded gaps453
between primes, Preprint, arXiv:1402.0811v1, 2014.454
[31] D. H. J. Polymath, (in preparation); see the Polymath8b webpages.455
[32] R. A. Rankin, The difference between consecutive prime numbers. II, Proc. Cambridge456
Philos. Soc. 36 (1940), 255–266.457
[33] A. Rényi, On the representation of an even number as the sum of a single prime and a458
single almost-prime number, Dokl. Akad. Nauk SSSR 56 (1947), 455–458 (Russian).459
[34] A. Rényi, On the representation of an even number as the sum of a single prime and a460
single almost-prime number, Izv. Akad. Nauk SSSR 12 (1948), 57–78 (Russian).461
[35] G. Ricci, Sull’andamento della differenza di numeri primi consecutivi. Riv. Mat. Univ.462
Parma 5 (1954), 3–54.463
Small gaps between primes 23
[36] A. Selberg, Collected Papers, Vol. II. With a foreword of K. Chandresekharan. Springer,464
Berlin, 1991.465
[37] K. Soundararajan, Notes on Goldston–Pintz–Yıldırım (unpublished).466
[38] A. I. Vinogradov, The density hypothesis for Dirichlet L-series, Izv. Akad. Nauk SSSR467
Ser. Mat. 29 (1965), 903–934 (Russian).468
[39] Y. Zhang, Bounded gaps between primes, Ann. of Math. (2) 179 (2014), No. 3, 1121–469
1174.470
Department of Mathematics and Statistics, San Jose State University, San Jose, CA 95192, U.S.A.E-mail: [email protected]
471
Rényi Mathematical Institute of the Hungarian Academy of Sciences, Budapest, Reáltanoda u.13–15, H-1053 HungaryE-mail: [email protected]
472
Bogaziçi University, Department of Mathematics, Bebek, Istanbul 34342 TurkeyE-mail: [email protected]
473