On a numerical upper bound for the

extended Goldbach conjecture

David Quarel

June 2017

A thesis submitted for the degree of Bachelor of Science (Advanced) (Honours)

of the Australian National University
Declaration

The work in this thesis is my own except where otherwise stated.

David Quarel
Acknowledgements

I wish to thank my supervisor Tim Trudgian. His advice and weekly meetings
provided much needed support whenever I’d get stuck on a problem or become
disheartened. I find myself in the most fortunate position of finishing honours in
mathematics with the same person as when I started it, way back in 2013 when
I took MATH1115.1
Having the most engaging lecturer during my first year of undergraduate
helped to foster my interest in mathematics, and I’ve yet to see any other lecturer
at this university take the time to write letters of congratulations to those that
did well in their course.

1
I still remember the hilarious anecdote in class about cutting a cake “coaxially”.

v
Abstract

The Goldbach conjecture states that every even number can be decomposed as the
sum of two primes. Let D(N ) denote the number of such prime decompositions
for an even N . It is known that D(N ) can be bounded above by
 Y 
∗ N Y 1 1
D(N ) ≤ C 1+ 1− = C ∗ Θ(N )
log2 N
p|N
p − 2 p>2 (p − 1)2
p>2

where C ∗ denotes Chen’s constant. It is conjectured [20] that C ∗ = 2. In 2004, Wu

[54] showed that C ∗ ≤ 7.8209. We attempted to replicate his work in computing
Chen’s constant, and in doing so we provide an improved approximation of the
Buchstab function ω(u),

ω(u) = 1/u, (1 ≤ u ≤ 2),

(uω(u))0 = ω(u − 1), (u ≥ 2).

based on work done by Cheer and Goldston [6]. For each interval [j, j + 1], they
expressed ω(u) as a Taylor expansion about u = j + 1. We expanded about the
point u = j + 0.5, so ω(u) was never evaluated more than 0.5 away from the
center of the Taylor expansion. which gave much stronger error bounds.
Issues arose while using this Taylor expansion to compute the required in-
tegrals for Chen’s constant, so we proceeded with solving the above differential
equation to obtain ω(u), and then integrating the result. Although the values
that were obtained undershot Wu’s results, we pressed on and refined Wu’s work
by discretising his integrals with finer granularity. The improvements to Chen’s
constant were negligible (as predicted by Wu). This provides experimental evi-
dence, but not a proof, that were Wu’s integrals computed on smaller intervals
in exact form, the improvement to Chen’s constant would be similarly negligible.
Thus, any substantial improvement on Chen’s constant likely requires a radically
different method to what Wu provided.

vii
Contents

Acknowledgements v

Abstract vii

Notation and terminology xi

1 History of Goldbach problems 1

1.1 Origin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Properties of D(N ) . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Ternary Goldbach conjecture . . . . . . . . . . . . . . . . . . . . . 2
1.4 Extended Goldbach conjecture . . . . . . . . . . . . . . . . . . . . 6
1.5 The Linnik–Goldbach problem . . . . . . . . . . . . . . . . . . . . 8
1.6 History of Romanov’s constant . . . . . . . . . . . . . . . . . . . . 9
1.7 Chen’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Preliminaries for Wu’s paper 13

2.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Preliminaries of sieve theory . . . . . . . . . . . . . . . . . . . . . 13
2.3 Selberg sieve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Examining Wu’s paper 21

3.1 Chen’s method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Wu’s improvement . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 A lower bound for H(s) . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 Discretising the integral . . . . . . . . . . . . . . . . . . . . . . . 28

4 Approximating the Buchstab function 31

4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 A piecewise Taylor expansion . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Errors of Taylor expansion . . . . . . . . . . . . . . . . . . 36

ix
x CONTENTS

4.2.2 Improving Cheer’s Method . . . . . . . . . . . . . . . . . . 37

5 Numerical computations 41
5.1 Justifying Integration Method . . . . . . . . . . . . . . . . . . . . 41
5.2 Interpolating Wu’s data . . . . . . . . . . . . . . . . . . . . . . . 48

6 Summary 51
6.1 Buchstab function . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 Chen’s constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.3 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

A 55
A.1 Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

Bibliography 65
Notation and terminology

In the following, p, q usually denote primes.

Notation

GRH Generalised Riemann hypothesis.

TGC Ternary Goldbach conjecture.

GC Goldbach conjecture.

FTC Fundamental theorem of calculus.

f (x)  g(x) g is an asymptotic upper bound for f , that is, there exists
constants M and c such that for all x > c, |f (x)| ≤ M |g(x)|.

f (x)  g(x) g is an asymptotic lower bound for f , that is, there exists
constants M and c such that for all x > c, |f (x)| ≥ M |g(x)|.
Equivalent to g  f .

f (x) ∼ g(x) g is an asymptotic tight bound for f , that is, lim fg(n)
(n)
= 1.
n→∞
Equivalent to f  g and g  f .

C0 Twin primes constant, §1.10

π(x) Number of primes p less than x.

P Set of all primes p.

P≤n Set of all numbers with at most n prime factors. Also denoted
Pn for brevity.

S(A; P, z) Sifting function, §2.1.

A Set of numbers to be sifted by a sieve S(A; P, z).

xi
xii NOTATION AND TERMINOLOGY

A≤x {a : a ∈ A, a ≤ x}.

Ad {a : a ∈ A≤x , a ≡ 0 mod d} for some square-free d.

1E (x) Characteristic function on a set E. Equals 1 if x ∈ E, 0

otherwise.

P Usually denotes a set of primes.

X Shorthand for X(x), denotes the main term in an approxima-

tion to a sieve.

R The remainder, or error term, for an approximation to a sieve.

C∗ Chen’s constant, §1.8.

ω(u) Buchstab’s function, §4.1.

γ Euler–Mascheroni constant.

µ(x) Möbius function.

ϕ(n) Euler totient function.

K Linnik–Goldbach constant, §1.9.

#A The cardinality of the set A.

(a, b) Greatest common divisor of a and b. Sometimes denoted

gcd(a, b) if (a, b) is ambiguous and could refer to an ordered
pair.

lcm(a, b) Least common multiple of a and b.

f ∗g Dirichlet convolution of f and g, defined as (f ∗ g)(x) =

P
d|n f (d)g(n/d)

a|b a divides b, i.e. there exists some n such that b = na.

a-b a does not divide b

Rx 1
Li(x) Logarithmic integral, Li(x) = 2 log t
dt
Chapter 1

History of Goldbach problems

1.1 Origin
One of the oldest and most difficult unsolved problems in mathematics is the
Goldbach conjecture (sometimes called the Strong Goldbach conjecture), which
originated during correspondence between Christian Goldbach and Leonard Euler
in 1742. The Goldbach conjecture states that “Every even integer N > 2 can be
expressed as the sum of two primes” [17].
For example,

4=2+2 8=5+3 10 = 7 + 3 = 5 + 5.

Evidently, this decomposition need not be unique. The Goldbach Conjecture has
been verified by computer search for all even N ≤ 4 × 1018 [12], but the proof
for all even integers remains open.∗ Some of the biggest steps towards solving
the Goldbach conjecture include Chen’s theorem [44] and the Ternary Goldbach
conjecture [23].

1.2 Properties of D(N )

As the Goldbach conjecture has been unsolved for over 250 years, a lot of work
has gone into solving weaker statements, usually by weakening the statement
that N is the sum of two primes. Stronger statements have also been explored.
We define D(N ) as the number of ways N can be decomposed into the sum of
two primes. The Goldbach conjecture is then equivalent to D(2N ) ≥ 1 for all
∗
Weakening Goldbach to be true for only sufficiently large N is also an open problem.

1
2 CHAPTER 1. HISTORY OF GOLDBACH PROBLEMS

N . As N grows larger, more primes exist beneath N (approximately logNN many,

from the prime number theorem) that could be used to construct a sum for N .
Thus, we expect large numbers to have many prime decompositions. And indeed
empirically this seems to be the case.
Below are some examples of prime decompositions of even numbers.

36 = 31 + 5 66 = 61 + 5 90 = 83 + 7 = 61 + 29
= 29 + 7 = 59 + 7 = 79 + 11 = 59 + 31
= 23 + 13 = 53 + 13 = 73 + 17 = 53 + 37
= 19 + 17 = 47 + 19 = 71 + 13 = 47 + 43
= 43 + 23 = 67 + 23
= 37 + 29

Intuitively, we would expect that as N grows large, D(N ) should likewise grow
large. The extended Goldbach conjecture [20] asks how D(N ) grows asymptot-
ically. As expected, the conjectured formula for D(N ) grows without bound as
N increases (Figure 1.1).
We can see from the plots that the points of D(N ) cluster into bands, and
2Θ(N ) also shares this property (Figure 1.2). The plot of D(N ) called “Gold-
bach’s Comet” [14], which has many interesting properties. We plot D(N ), and
colour each point either red, green or blue if (N/2 mod 3) is 0, 1 or 2 respectively
(Figure 1.3). [14]. We can see the bands of colours are visibly separated. If D(N )
is plotted only for prime multiples of an even number [2], say

N = 12 × 2, 12 × 3, 12 × 5, . . .

the resulting plot is almost a straight line (Figure 1.4).

1.3 Ternary Goldbach conjecture

The Ternary Goldbach conjecture (TGC) states that “every odd integer greater
than 5 can be expressed as the sum of three primes”. This statement is directly
implied by the Goldbach conjecture, as if every even n > 2 can be expressed as
n = p + q, where p, q prime, we can express all odd integers n0 greater than 5 as
three primes
n0 = n + 3 = p + q + 3.
In 1923, Hardy and Littlewood [20] showed that, under the assumption of the
generalised Riemann hypothesis (GRH), the TGC is true for all sufficiently large
1.3. TERNARY GOLDBACH CONJECTURE 3

Figure 1.1: Plot of D(N ) for N ≤ 25000.

Figure 1.2: Plot of 2Θ(N ), the conjectured tight bound to D(N ), for N ≤ 25000.
4 CHAPTER 1. HISTORY OF GOLDBACH PROBLEMS

Figure 1.3: Plot of D(N ) for N ≤ 25000, points coloured based on N/2 mod 3.

Figure 1.4: Plot of D(N ) restricted to N = 12p, 12p ≤ 105

1.3. TERNARY GOLDBACH CONJECTURE 5

odd numbers. They also specified how D3 (n) (the number of ways n can be
decomposed as the sum of three primes) grows asymptotically
n2 Y  
1
D3 (n) ∼ C3 3 1− 2 , (1.1)
log n p − 3p + 3
p|n
p prime

where
Y  1

C3 = 1+ .
p>2
(p − 1)3
p prime

Using the Taylor expansion of exp(x) and the fact that every term in the sum is
positive, we obtain the inequality
∞
x2 X xn
1+x≤1+x+ + ... = = exp(x).
2! n=0
n!

Hence by induction, for all non-negative sequences {a1 , . . . , an }

n n
!
Y X
(1 + ai ) ≤ exp ai . (1.2)
i=1 i=1

Therefore we may note that C3 must converge, as

∞
!
Y 1
 X 1
C3 < 1 + 3 ≤ exp 3
< ∞.
n≥2
n n=2
n

We also note that the product over all primes dividing n in (1.1) grows slowly
compared to the main term, as for all n ≥ 2
1
n2 − 3n + 3 ≥ n
2
1 2
1− 2 ≥1−
n − 3n + 3 n
Y  1
 Y  2 Y  2

1− 2 ≥ 1− ≥ 1− .
p − 3p + 3 p 2<p≤n
p
p|n p|n
p prime p prime p prime

We have from Mertens’ theorem [10] that

Y  e−γ

1
1− ∼ . (1.3)
p≤n
p log n
p prime

Combining this with the following inequality

2 1 1
1− ≥ (1 − )2 for x ≥ 3, (1.4)
x 2 x
6 CHAPTER 1. HISTORY OF GOLDBACH PROBLEMS

we obtain 2
Y  e−2γ
 
2 1 Y 1
1− ≥ 1− ∼ ,
2<p≤n
p 2 2<p≤n p 2 log2 n
p prime p prime

which gives an asymptotic lower bound on D3 (n)

n2
D3 (n)  .
log5 n

This gives a stronger version of the TGC, as it implies that the number of different
ways that an odd integer can be decomposed as the sum of three primes can
be arbitrarily large. In 1937, Vinogradov improved this result by removing the
dependency on GRH [50].
An explicit value for how large N needs to be before the TGC holds was found
15
by Borozdin, who showed that N > 33 is sufficient [16]. In 1997, the TGC was
shown to be conditionally true for all odd N > 5 by Deshouillers et al, [11] under
the assumption of GRH. The unconditional TGC was further improved by Liu
and Wang [34] in 2002, who showed that Borozdin’s constant can be reduced to
N > e3100 . Thus, in principle, all that was needed to prove the TGC was to check
that it held for all N < e3100 ≈ 107000 . Given that the number of particles in
the universe ≈ 1080 , no improvements to computational power would likely help.
Further mathematical work was necessary.
In 2013, Helfgott reduced the constant sufficiently as to allow computer as-
sisted search to fill the gap, obtaining N > 1027 [23]. When combined with
his work numerically verifying the TGC for all N < 8.875 × 1030 , [23] Helfgott
succeeded in proving the TGC.

1.4 Extended Goldbach conjecture

Hardy and Littlewood’s paper [20] discuss how one can directly obtain an asymp-
totic formula (conditional on GRH) for the numbers of ways even N can be
decomposed into the sum of four primes. Let D4 (N ) be the number of such ways.

n3 Y
 
1 1
D4 (N ) ∼ C4 4 1+ , (1.5)
3 log n (p − 2)(p2 − 2p + 2)
p|n
p>2

where
Y 1

C4 = 1− .
p>2
(p − 1)4
1.4. EXTENDED GOLDBACH CONJECTURE 7

The asymptotic formulae for D3 (N ) and D4 (N ) are very similar (checking con-
vergence and the long term behaviour of D4 (N ) is the same proof as for D3 (N )).
Hardy–Littlewood claim that one can easily generalise their work to obtain an
asymptotic formula for Dr (N ) for r > 2. However, for r = 2, the results do not
easily generalise:

“It does not fail in principle, for it leads to a definite result which
appears to be correct; but we cannot overcome the difficulties of the
proof...”[20, p. 32]

The following asymptotic formula for D(N ) was thus conjectured [20]

D(N ) ∼ 2Θ(N ), (1.6)

where
CN N Y 1
 Y 1

Θ(N ) := , CN := C0 1+ , C0 := 1− .
log2 N p|N
p−2 p>2
(p − 1)2
p>2

Now C0 denotes the twin primes constant (see 1.10), and each term in the infinite
product in CN is greater than 1. Thus, CN > C0 , which provides an asymptotic
lower bound of
N
D(N )  , (1.7)
log2 N
Hence, (1.6) implies the Goldbach conjecture (for large N ), which should indicate
the difficulty of the problem.
There has been progress in using (1.6) as a way to construct an upper bound
for D(N ), by seeking the smallest values of C ∗ (hereafter referred to as Chen’s
constant) such that
D(N ) ≤ C ∗ Θ(N )‡ (1.8)
Upper bounds for C ∗ have been improved over the years, but recent improvements
have been small† (see Table 1.1). Recent values have been obtained by various
sieve theory methods by constructing large sets of weighted inequalities (the one
in Wu’s paper has 21 terms!). However, the reductions in C ∗ are minuscule.
One would suspect that any further improvements would require a radically new
method, instead of complicated inequalities with more terms.
Wu Dong Hua [53] claimed to have C ∗ ≤ 7.81565, but some functions he
defined to compute C ∗ failed to have some necessary properties [48].
‡
Some authors instead include the factor of 2 inside C ∗ , so they instead assert D(N ) ∼ Θ(N )
for (1.6), and D(N ) ≤ 2C ∗ Θ(N ) for (1.8).
†
We remind the reader than C ∗ is conjectured to be 2.
8 CHAPTER 1. HISTORY OF GOLDBACH PROBLEMS

C∗ Year Author
16 +  1949 Selberg [45]
12 1964 Pan [37]
8 1966 Bombieri & Davenport [3]
7.8342 1978 Chen [7]
7.8209 2004 Wu [54]

Table 1.1: Improvements of C ∗ in (1.8) over time, conjectured that C ∗ = 2.

1.5 The Linnik–Goldbach problem

The Linnik–Goldbach problem [41] is another weaker form of the GC, which asks
for the smallest values of K such that for all sufficient large even N ,

N = p1 + p2 + 2e1 + 2e2 + . . . + 2er , r ≤ K. (1.9)

That is, we can express N as the sum of two primes and at most K powers of two
(we refer to K as the Linnik–Goldbach constant). In 1953, Linnik [30] proved
that there exists some finite K such that the statement holds, but did not include
an explicit value for K. (See Table 1.2 for historical improvements on K.) The
methods of Heath-Brown and Puchta [22], and of Pintz and Ruzsa [40] show that
K satisfies (1.9) if
C3
λK−2 < ,
(C − 2)C2 + cC0−1 log 2
∗

for particular constants λ, C0 , C ∗ , C2 , C3 , c.

Here, C0 is the twin primes constant, given by
Y  1

C0 = 1− ≈ 0.66016 . . . (1.10)
p>2
(p − 1)2
p prime

The infinite product for C0 is easily verified as convergent, as for all integers n > 2

1
0<1− < 1.
(n − 1)2

So C0 is an infinite product, with each term strictly between 0 and 1. Therefore,

0 ≤ C0 < 1. The value of C0 is easily computed: Wrench provides C0 truncated
to 42 decimal places [52]. As far as lowering the value of K, improvements on
the other constants will be required.
1.6. HISTORY OF ROMANOV’S CONSTANT 9

Now, C2 is defined by
 
∞
X  |µ(2d + 1)| Y 1 
C2 =  ,
 (2d + 1)
d=0
p − 2
p|d
p>2

where
(n) = min{v : 2v ≡ 1 mod d}.
v∈N

and µ(x) is the Möbius function. It has been shown by Khalfalah–Pintz [39] that

1.2783521041 < C2 C0 < 1.2784421041, (1.11)

So since C0 is easily computed, we can convert (1.11) to

1.93642 < C2 < 1.93656. (1.12)

It is remarked by Khalfalah–Pintz that estimating C2 is very difficult, and that

any further progress is unlikely. The remaining constants are defined by similarly
complicated expressions (see [41]). Platt and Trudgian [41] make improvements
on both C3 and λ, by showing that one can take

(C3 , λ) = (3.02858417, 0.8594000)

and obtain unconditionally that K ≥ 11.0953, a near miss for K = 11. Platt–
Trudgian remark that any further improvements in estimating C3 using their
method with more computational power would be limited. Obtaining K = 11 by
improving Chen’s constant (assuming all others constants used by Platt–Trudgian
are the same) would require C ∗ ≤ 7.73196, which is close to the best known value
of C ∗ ≤ 7.8209. This provides a motivation for reducing C ∗ .

1.6 History of Romanov’s constant

In 1849, de Polignac conjectured that every odd n > 3 can be written as a sum
of an odd prime and a power of two. He found a counterexample n = 959, but
an interesting question that follows is how many counterexamples are there? Are
there infinitely many? If so, how common are they among the integers? If we let

A(N ) = #{n : n < N, n = 2m + p} (1.13)

10 CHAPTER 1. HISTORY OF GOLDBACH PROBLEMS

K K assuming GRH Year Author

- <∞ 1951 Linnik [29]
<∞ - 1953 Linnik [30]
54000 - 1998 Liu, Liu and Wang [31]
25000 - 2000 H.Z.Li [28]
- 200 1999 Liu, Liu and Wang [32]
2250 160 1999 T.Z.Wang [51]
1906 - 2001 H.Z.Li [27]
13 7 2002 Heath-Brown and Puchta∗ [22]
12 - 2011 Liu and Lü [33]

Table 1.2: Values of K in (1.9) over the years

A(N ) A(N ) A(N ) †

R− = lim inf R = lim R+ = lim sup (1.14)
N →∞ N N →∞ N N →∞ N
then R denotes the density of these decomposable numbers, and R+ and R−
provide upper and lower bounds respectively.

Romanov [43] proved in 1934 that R− > 0, though he did not provide an
estimate on the value of R. An explicit lower bound on R was not shown un-
til 2004, with Chen–Sun [8] who proved R− ≥ 0.0868. Habsieger–Roblot [18]
improved both bounds, obtaining 0.0933 < R− ≤ R+ < 0.49045. Pintz [38]
obtained R− ≥ 0.09368, but under the assumption of Wu Dong Hua’s value of
Chen’s constant C ∗ ≤ 7.81565, the proof of which is flawed [48].
Since 2n 6= p + 2k for an odd prime p, it is trivial to show that R+ ≤ 1/2. Van
de Corput [49] and Erdős [13] proved in 1950 that R+ < 0.5. In 1983, Romani
[42] computed A(N )/N for all N < 107 and noted that the minima were located
when N was a power of two. He then constructed an approximation to A(N ), by
using an approximation to the prime counting function π(x)

x ∞
(m − 1)!
Z
1 X
π(x) ∼ Li(x) := dt = m x,
2 log t m=1
log x

then using the Taylor expansion of Li(x) to obtain a formula for A(N ) with
†
Some authors denote R = limn→∞ 2A(N )/N (and similar for R+ and R− ), in which case
the trivial bound is R+ ≤ 1.
∗
This was also proved independently by Pintz and Ruzsa [40] in 2003. They also claim to
have K ≤ 8, but this is yet to appear in literature.
1.6. HISTORY OF ROMANOV’S CONSTANT 11

unknown coefficients bi
∞
X x
A(N ) = Rx + bi .
i=1
(log x)i

By using the precomputed values for A(N ), the unknown coefficients may be
estimated, thus extrapolating values for A(N ). Using this method, Romani con-
jectured that R ≈ 0.434.
Bomberi’s [42] probabilistic approach, which randomly generates probable
primes (as they are computationally faster to find than primes, and pseudoprimes‡
are uncommon enough that it does not severely alter the results) in the interval
[1, n], in such a way that the distribution closely matches that of the primes, is
also used to compute R. The values obtained closely match Romani’s work, and
so it is concluded that R ≈ 0.434 is a reasonable estimate.
The lower bound to Romonov’s constant R− , the Linnik–Goldbach constant
K and Chen’s constant C ∗ are all related. There is a very complicated connection
relating R− and K [38], but here we prove a simple property relating R− and K.

Proposition. If R− > 1/4, then K ≤ 2 [38].

Proof. If R− > 1/4 then for any given even N , more than 1/4 of the numbers up
to N can be written as the sum of an odd prime and a power of two. No even n
can be written as n = p + 2k since p is odd. So more than 1/2 of the odd numbers
up to N can be written in this way. If we construct a set of pairs of odd integers
  
N −r N +r
(1, N − 1), (3, N − 3), . . . , , ,
2 2
where r = N mod 4, then by the pigeon-hole principle, there must exist a pair
(k, N − k) such that

k = p1 + 2v1 , N − k = p2 + 2 v2 .

Therefore
N = k + (N − k) = p1 + p2 + 2v1 + 2v2 .

Pintz also proved the following theorems [38].

Theorem 1.6.1. If C ∗ = 2 holds, then R− ≥ 0.310987.

This can be combined with the proposition above to obtain a relationship be-
tween C ∗ and K, although Pintz also shows a weaker estimate on C ∗ is sufficient.
‡
A composite number that a probabilistic primality test incorrectly asserts is prime.
12 CHAPTER 1. HISTORY OF GOLDBACH PROBLEMS

Theorem 1.6.2. If C ∗ ≤ 2.603226, then R− > 1/4 and consequently K ≤ 2.

As an explicit lower bound for R did not appear until 2004, and this bound
is much lower than the expected 0.434 given by Romani, attempting to improve
K by improving R− appears a difficult task.

1.7 Chen’s theorem

Chen’s theorem [26] is another weaker form of Goldbach, which states that every
sufficiently large even integer is the sum of two primes, or a prime and a semiprime
(product of two primes). By letting R(N ) denotes the number of ways N can be
decomposed in this manner, Chen attempts to get a lower bound on the number
of ways N can be decomposed as N = p + P3 , where P3 is some number with no
more than three prime factors, using sieve theory methods. He then removes the
representations where N = p + p1 p2 p3 [19]. By doing so, Chen obtains that there
exists some constant N0 such that if N is even and N > N0 , then

#{p : p ≤ N, n − p ∈ P2 } > 0.67Θ(N ),

where Θ(N ) is the same function used for the conjectured tight bound of D(N )
(see 1.6). We have the asymptotic lower bound (see 1.7)

N
Θ(N )  .
log2 N
so this implies that for all sufficiently large even N

N = p + q for q ∈ P2

In 2015, Yamada [55] proved that letting N0 = exp exp 36 is sufficient for the
above to hold.
Chapter 2

Preliminaries for Wu’s paper

2.1 Summary
Wu [54] proved that C ∗ ≤ 7.8209. He obtained this result by expressing the
Goldbach conjecture as a linear sieve, and finding an upper bound on the size
of the sieved set. The upper bound constructed is very complicated, using a
series of weighted inequalities with 21(!) terms. The approximation to the sieve
is computed using numerical integration, weighting the integrals over judicially
chosen intervals. Few intervals (9), over which to discretise the integration are
used, thereby reducing the problem to a linear optimisation with 9 equations.

“If we divide the interval [2, 3] into more than 9 subintervals, we can
certainly obtain a better result. But the improvement is minuscule.”[54,
p. 253]

The main objective of this paper is to quantify how “minuscule” the improvement
is.

2.2 Preliminaries of sieve theory

A sieve is an algorithm that takes a set of integers and sifts out, or removes,
particular integers in the set based on some properties.
The classic example is the Sieve of Eratosthenes [21, Chpt 1], an algorithm
devised by Erathosthenes in the 3rd century BC to generate all prime numbers
below some bound. The algorithm proceeds as follows: Given the first n primes
P = {p1 , . . . , pn }, list the integers from 2 to (pn + 2)2 − 1. For each prime p in
the set P, strike out all multiples of p in the list starting from 2p. All remaining

13
14 CHAPTER 2. PRELIMINARIES FOR WU’S PAPER

numbers are primes below (pn + 2)2 , as the first number k that is composite
and not struck out by this procedure must share no prime factors with P. So
k = p2n+1 ≥ (pn + 2)2 .
This method of generating a large set of prime numbers is efficient, as by
striking off the multiples of each prime from a list of numbers, the only opera-
tions used is addition and memory lookup. Contrast this with primality testing
via trial division, as integer division is slower than integer addition. O’Neill
[36] showed that to generate all primes up to a bound n, the Sieve of Eratos-
thenes takes O(n log log n) operations whereas repeated trial division would take
3
O(n 2 / log2 n) operations. Moreover, the Sieve of Eratosthenes does not use divi-
sion in computing the primes, only addition.
Sieve theory is concerned with estimating the size of the sifted set. Formally,
given a finite set A ⊂ N, a set of primes P and some number z ≥ 2, we define
the sieve function [21]

S(A; P, z) = #{a ∈ A : ∀p ∈ P with p < z, (a, p) = 1}, (2.1)

i.e. by taking the set A and removing the elements that are divisible by some prime
p ∈ P with p < z, the number of elements left over is S(A; P, z). Removing the
elements of A in this manner is called sifting. Analogous to a sieve that sorts
through a collection of objects and only lets some through, the unsifted numbers
are those in A that do not have a prime factor in P. Many problems in number
theory can be re-expressed as a sieve, and thus attacked. We provide an example
related to the Goldbach conjecture included in [21]. For a given even N , let

A = {n(N − n) : n ∈ N, 2 ≤ n ≤ N − 2}, P = P.

Then
√ √
S(A, P; N ) = #{n(N − n) : n ∈ N, p < N ⇒ (n(N − n), p) = 1}.
√
Now since (n(N − n), p) = 1 ⇒ (n, p) = 1 and (N − n, p) = 1 for all p < N ,
both n and N − n must be prime, as if not, either n or N − n has a prime factor
√
≥ N , which implies n ≥ N (impossible) or N − n ≥ N (also impossible). Hence
√ √
S(A, P; N ) = #{p ∈ P : (N − p) ∈ P, p < N },
√
Now for all primes p, if p < N , then p ≤ N/2, hence
√
S(A, P; N ) ≤ #{p ∈ P : (N − p) ∈ P, p ≤ N/2},
≤ #{p ∈ P : (N − p) ∈ P, p ≤ N − p} = D(N ).
2.2. PRELIMINARIES OF SIEVE THEORY 15

So this particular choice of sieve is a lower bound for D(N ). If we have a particular
(possibly infinite) subset of the integers A, we may wish to take all numbers in
A less that or equal to x, denoted∗ A≤x

A≤x = {a ∈ A : a ≤ x},

and ask how fast #A≤x grows with x. For example, if A were the set of all even
numbers,
A = {n ∈ N : n even} = {0, 2, 4, . . .},
the size of the set of A≤x would be computed as:
jxk
#A≤x = #{a ∈ A : a ≤ x} = + 1.
2
It is usually more difficult to figure out how sets grow. Likewise, an exact formula
is usually impossible. Therefore, the best case is usually an asymptotic tight
bound. For example, if A = P, the set of all primes, then the prime number
theorem [1, p. 9] gives
x
#{p ∈ P : p ≤ n} = π(x) ∼ .
log x
The main aim of sieve theory is to decompose #A≤N as the sum of a main term
X and an error term R, such that X dominates R for large N . We can then
obtain an asymptotic formula for #A≤N [19]. This partially accounts for the
many proofs in number theory that only work for some extremely large N , where
N can be gargantuan (see the Ternary Goldbach conjecture in Chapter 1), as the
main term may only dominate the error term for very large values. In some cases
it can only be shown that the main term eventually dominates the error term, no
explicit value as to how large N needs to be before this occurs is required.
Now, for a given subset A≤x , we choose a main term† X that is hopefully a
good approximation of #A≤x .
We denote
Ad = {a : a ∈ A≤x , a ≡ 0 mod d},
for some square-free number d. For each prime p, we assign a value for the
function ω(p), with the constraint that 0 ≤ ω(p) < p, so that
ω(p)
#Ap ≈ X.
p
∗
In the literature, it is common to use A instead of A≤x , and it is implicitly understood
that A only includes the numbers less than some x.
†
Note that since X depends on x, we would normally write X(x), but we wished to stick to
convention.
16 CHAPTER 2. PRELIMINARIES FOR WU’S PAPER

By defining
Y
ω(1) = 1, ω(d) = ω(p),
p|d

for all square-free d, we ensure that ω(d) is a multiplicative function. We now

define the error term (sometimes called the remainder term):

ω(d)
Rd = #Ad − X.
d
So we have that
ω(d)
#Ad = X + Rd ,
d
ω(d)
where (hopefully) the main term d
X dominates the error term Rd . Given a set
of prime numbers P, we define
Y
P (z) = p.
p∈P,p<z

We can rewrite the sieve using two identities of the Möbius function and
multiplicative functions.

Theorem 2.2.1. [1, Thm 2.1] If n ≥ 1 then


X 1 n = 1,
µ(d) = (2.2)
0 n > 1.
d|n

Proof. The case n = 1 is trivial. For n > 1, write the prime decomposition of n
as n = pe11 . . . penn . Since µ(d) = 0 when d is divisible by a square, we need only
consider divisors d of n of the form d = pa11 . . . pann , where each ai is either zero or
one, (i.e., divisors that are products of distinct primes). Enumerating all possible
products made from {p1 , . . . , pn }, we obtain
X   
µ(d) = µ(1) + µ(p1 ) + . . . + µ(pn ) + µ(p1 p2 )
d|n

+ . . . + µ(pn−1 pn ) + . . . + µ(p1 . . . pn )
     
n n n
=1+ (−1) + (1) + . . . + (−1)n
1 2 n
n  
X n
= (−1)i = (1 − 1)n = 0.
i=0
i
2.2. PRELIMINARIES OF SIEVE THEORY 17

Theorem 2.2.2. [1, Thm 2.18] If f is multiplicative ‡ we have

X Y
µ(d)f (d) = (1 − f (p)). (2.3)
d|n p|n

Proof. Let X
g(n) = µ(d)f (d).
d|n

Then g(n) is multiplicative, as

X
g(pq) = µ(d)f (d) = µ(1)f (1) + µ(p)f (p) + µ(q)f (q) + µ(pq)f (pq),
d|pq

= 1 − f (p) − f (q) + f (p)f (q) = (1 − f (p))(1 − f (q)),

  
X X
= µ(d)f (d)  µ(d)f (d) = g(p)g(q).
d|p d|q

So by letting n = pa11 . . . pann , we have

n
Y n X
Y
g(n) = g(pai i ) = µ(d)f (d).
i=1 i=1 d|pai
i

Now since µ(d) is zero for d = p2 , p3 , . . ., the only non-zero terms that are divisors
of pai i are 1 and pi .
n
Y n
Y Y
g(n) = (µ(1)f (1) + µ(pi )f (pi )) = (1 − f (pi )) = (1 − f (p)).
i=1 i=1 p|n

Now given this, we can rewrite the sieve as

X X X X
S(A, P, x) = 1= µ(d) = µ(d)#Ad , (2.4)
n∈A n∈A d|n d|P (z)
(n,P (z))=1 d|P (z)

as we sum over all n ∈ A such that a is square-free, as the Möbius function is

only non-zero for square-free inputs. Using the approximation to #Ad , we obtain
X µ(d)ω(d) X
S(A, P, x) = X + µ(d)Rd ,
d
d|P (z) d|P (z)
Y   X (2.5)
ω(p)
=X 1− + µ(d)Rd .
p
p|P (z) d|P (z)

‡
A function f is multiplicative if f (m)f (n) = f (mn) whenever gcd(n, m) = 1.
18 CHAPTER 2. PRELIMINARIES FOR WU’S PAPER

Writing
Y  ω(p)

W (z; ω) = 1− ,
p<z,p∈P
p

we obtain
X
|S(A; P, z) − XW (z; ω)| ≤ µ(d)Rd .
d|P (z)

Then, by using the triangle inequality, we obtain the worst case remainder when
all the µ(d) terms in the sum are one
X
|S(A; P, z) − XW (z; ω)| ≤ |Rd |. (2.6)
d|P (z)

This is the Sieve of Eratosthenes–Legendre, which crudely bounds the error be-
tween the actual sieved set and the approximations generated by choosing X
and ω(p). Sometimes, even approximating this sieve proves too difficult, so the
problem is weakened before searching for an upper or lower bound. For example,
see (1.7). For an upper bound, we search for functions µ+ (d) that act as upper
bounds to µ(d), in the sense that

X X 1 (n, P (z)) = 1,
+
µ (d) ≥ µ(d) = (2.7)
0 (n, P (z)) > 1.
d|(n,P (z)) d|(n,P (z))

Then
X µ+ (d)ω(d) X
S(A; P, z) ≤ X + |µ+ (d)||Rd |. (2.8)
d
d|P (z) d|P (z)

Minimising the upper bound while ensuring µ+ (d) satisfies (2.7) and has suffi-
ciently small support (to reduce the size of the summation) is, in general, very
difficult. Nevertheless, this method is used by Chen [26] to construct the follow-
ing upper bounds of sieves that we have related to the Goldbach conjecture and
the twin primes conjecture § .
n
#{(p, p0 ) : p ∈ P, p0 ∈ P≤2 , p + p0 = 2n}  (2.9)
log2 n
x
#{p ∈ P : p ≤ x, p + 2 ∈ P≤2 }  (2.10)
log2 x
where P≤2 is the set of all integers with 1 or 2 prime factors.
§
The twin primes conjecture asserts the existence of infinitely many primes p for which p + 2
is also prime. Examples include 3 and 5, 11 and 13, 41 and 43, . . .
2.3. SELBERG SIEVE 19

The first sieve is a near miss for approximating the Goldbach sieve, and weak-
ens the Goldbach conjecture to allow sums of primes, or a prime and a semiprime¶ .
The second sieve similarly weakens the twin primes conjecture. Thus, Chen’s
sieves imply that there are infinitely many primes p such that p + 2 is either
prime or semiprime, and that every sufficiently large even integer n can be de-
composed into the sum of a prime and a semiprime. However, the asymptotic
bounds above do not show constants that may be present, so “sufficiently large”
could be very large indeed.

2.3 Selberg sieve

Let λ1 = 1 and for d ≥ 2, let λd ≥ 2 be arbitrary real numbers. By (2.4), we can
construct the upper bound
 2
X X X X 
S(A, P, z) = µ(d) ≤ 
  .
λd  (2.11)
n∈A d|n n∈A d|n
d|P (z) d|P (z)

By squaring and rearranging the order of summations [19], we obtain

X X
S(A, P, z) ≤ λd1 λd2 1. (2.12)
dv |P (z) a∈A
v=1,2 a≡0 mod lcm(d1 ,d2 )

This is used to construct an upper bound for the sieve in the same form as (2.8).
X ω(D) X
S(A; P, z) ≤ X λd1 λd2 + |λd1 λd2 RD | = XΣ1 + Σ2 , (2.13)
D
dv |P (z) dv |P (z)
v=1,2 v=1,2

Now one can choose the other remaining constants λd , d ≥ 2 so that the main
term Σ1 is as small an upper bound as possible, while ensuring Σ1 dominates Σ2 .
Since this is in general very difficult even for simple sequences A, Selberg looked
at the more restricted case of setting all the constants

λd = 0 for d ≥ z

and then choosing the remaining λd terms to minimise Σ1 , which is now a

quadratic in λd . Having fewer terms to deal with makes it easier to control,
and hence to bound the size of the remainder term Σ2 .
¶
A semiprime is a product of two primes.
20 CHAPTER 2. PRELIMINARIES FOR WU’S PAPER
Chapter 3

Examining Wu’s paper

In this chapter we look at the section of Wu’s paper needed to compute C ∗ .

3.1 Chen’s method

Chen wished to take the set

A = {N − p : p ≤ N },

and apply a sieve to it, keeping only the integers in A that are prime. This
leaves the set of all primes of the form N − p, which (due to double counting) is
asymptotic to 2D(N ).
The goal is to approximate the sieve S(A; P, z), as described in Chapter 2.
Chen uses the sieve Selberg [45] used to prove C ∗ ≤ 16 + , which has some extra
constraints on the multiplicative function ω(p) in the main term, to make it easier
to estimate. Define
Y ω(p)

V (z) = 1− (3.1)
p<z
p
and suppose there exists a constant K > 1 such that
 
V (z1 ) log z2 K
≤ 1+ for z2 ≥ z1 ≥ 2. (3.2)
V (z2 ) log z1 log z1
Then the Rosser–Iwaniec [25] linear sieve is given by
    X X
log Q
S(A; P, z) ≤ XV (z) F +E + λ+l (q)r(A, q), (3.3)
log z l<L q|P (z)
    X X
log Q
S(A; P, z) ≥ XV (z) f +E − λ−
l (q)r(A, q), (3.4)
log z l<L q|P (z)

21
22 CHAPTER 3. EXAMINING WU’S PAPER

∗
where F and f are the solutions of the following coupled difference equations,

F (u) = 2e−γ /u, f (u) = 0, (0 < u ≤ 2) (3.5)

(uF (u))0 = f (u − 1), (uf (u))0 = F (u − 1), (u ≥ 2) (3.6)

The Rosser–Iwaniec sieve is a more refined version of Selbergs sieve, as the error
− −
terms λ+ +
l and λl have some restrictions that, informally, λl (or λl ) can be
decomposed into the convolution of two other functions λ = λ1 ∗ λ2 . We will be
concerned only with (3.3), as we only need to find an upper bound for the sieve.
Lower bounds on the sieve would imply the Goldbach conjecture, which would
be difficult. Chen improved on the sieve (3.3) by introducing two new functions
H(s) and h(s) such that (3.3) holds with f (s) + h(s) and F (s) − H(s) in place
of f (s) and F (s) respectively.[54].
   
log Q
S(A; P, z) ≤ XV (z) (F (s) − H(s)) + E + error, (3.7)
log z

Chen proved that h(s) > 0 and H(s) > 0 (which is obviously a required property,
as otherwise these functions would make the bound on S(A; P, z) worse) using
three set of complicated inequalities (the largest had 43 terms!).

3.2 Wu’s improvement

Wu followed the same line of reasoning as Chen, but created a new set of in-
equalities that describe the functions H(s) and h(s). Wu’s inequalities are both
simpler (having only 21 terms) and make for a tighter bound on S(A; P, z). We
will not look into the inequalities, but rather a general overview of the method
Wu used. For full details, see [54, p. 233].
Let δ > 0 be a sufficiently small number, k ∈ N and 0 ≤ i ≤ k. Define
1+k
Q = N 1/2−δ , d = Q/d, L = log N, Wk = N d . (3.8)

Let ∆ ∈ R such that 1 + L−4 ≤ ∆ ≤ 1 + 2L−4 .

Define
A(s) = sF (s)/2e−γ , a(s) = sf (s)/2eγ , (3.9)

where F and f are defined by (3.5) and (3.6).

∗
The astute reader will note the similarity between F, f and the Buchstab function ω(u)
(see 4.1)
3.2. WU’S IMPROVEMENT 23

Define
X
Φ(N, σ, s) = σ(d)S(Ad , {p : (p, dN ) = 1}, d1/s ), (3.10)
d

where σ is a arithmetical function that is the Dirchlet convolution of a collection

of characteristic functions
σ = 1E1 ∗ . . . ∗ 1Ei ,

where
Ej = {p : (p, N ) = 1} ∩ [Vj /∆, Vj ),

and V1 , . . . , Vi are real numbers satisfying a set of inequalities[54, p. 244]. Infor-

mally, the inequalities state that the Vi terms are ordered in size, bounded below
by Wk
V1 ≥ . . . ≥ Vi ≥ Wk ,

and that no one Vi term can be too large. Each Vi term is bounded above by
both Q and the previous terms V1 , . . . , Vi−1 .
s
Q
Vi ≤ .
V1 . . . Vi−1

So if any one Vi term is big, the rest of the terms beyond will be constrained to
be small.
Φ can be thought of as breaking up the sieve S(A; P, z) into smaller parts,
where the set that is being sieved over is only those elements in A that are
multiples of d, and the index of summation is given by the σ, as σ(d) will be zero
everywhere expect on its set of support. Beaking up the sieve this was allows
Wu to prove some weighted inequalities reated to Φ, that would ordinarily be too
difficult to prove in general for the entire sieve.
Define
X σ(d)CdN
Θ(N, σ) = 4Li(N ) , (3.11)
d
ϕ(d) log d

where CN is defined in (1.6).

Now for k ∈ N+ , N0 ≥ 2 and s ∈ [1, 10] we defined Hk,N0 (s) and hk,N0 as
the supremum of h such that for all N ≥ N0 and functions σ comprised of the
convolution of no more than k characteristic functions, the following inequalities
hold

Φ(N, σ, s) ≤ {A(s) − h}Θ(N, σ), Φ(N, σ, s) ≥ {a(s) + h}Θ(N, σ). (3.12)

24 CHAPTER 3. EXAMINING WU’S PAPER

Wu shows that both Hk,N0 (s) and hk,N0 are decreasing, and defines
   
H(s) = lim lim Hk,N0 (s) , h(s) = lim lim hk,N0 (s) (3.13)
k→∞ N0 →∞ k→∞ N0 →∞

Now it is very difficult to get an explicit form of H or h, or to even conclude

anything about the behaviour beyond it is decreasing. Wu proceeds by proving
the following integral equations
Z s0 −1 Z s0 −1
0 H(t) 0 h(t)
h(s) ≥ h(s ) + dt, H(s) ≥ H(s ) + dt, (3.14)
s−1 t s−1 t

These integral equations are still difficult to work with, as all Wu proves about h
and H is that H(s) is decreasing on [1, 10], and h(s) is increasing on [1, 2], and
is decreasing on [2, 10].
Wu proves an upper bound for the smaller parts of the sieve Φ
X
5Φ(N, σ, s) ≤ σ(d)(Γ1 − . . . − Γ4 + Γ5 + . . . + Γ21 ) + Oδ,k (N 1−η )
d
X
2Φ(N, σ, s) ≤ σ(d)(Ω1 − Ω2 + Ω3 ) + Oδ,k (N 1−η )
d

where the Γi terms are the dreaded 21 terms in Wu’s weighted inequality [54,
p.233]. Wu uses this weighted inequality

3.3 A lower bound for H(s)

Now to get an upper bound on the sieve, Wu needs to compute H(s). The integral
equation is difficult to resolve, so it is weakened to give a lower bound for H(s),
given the following Wu obtains two lower bounds for H(s),

Proposition 3.3.1. For 2 ≤ s ≤ 3 ≤ s0 ≤ 5 and s0 − s0 /s ≥ 2, we have

Z 3
H(s) ≥ Ψ1 (s) + H(t)Ξ1 (t, s) dt, (3.15)
1

where Ξ(t, s) = Ξ(t, s, s0 ) is given by

(t + 1)2
   
0 σ0 (t) 16 1[α2 ,3] (t)
Ξ1 (t, s, s ) := log + log
2t (s − 1)(s0 − 1) 2t (s − 1)(s0 − 1)
 
1[α3 ,α2 ] (t) t+1
+ log ,
2t (s − 1)(s0 − 1 − t)
(3.16)
3.3. A LOWER BOUND FOR H(S) 25

and where Ψ(s) = Ψ(s, s0 ) is given by

s0 −1 1−1/s0
log(t − 1) log(s0 t − 1)
Z Z
0 1
Ψ1 (s, s ) := dt + dt − I1 (s, s0 ), (3.17)
2 t 2 1−1/s t(1 − t)

where I1 (s, s0 ) is defined as

 
φ−t−u−v
ZZZ
0 dt du dv
I1 (s, s ) = max ω , (3.18)
φ≥2 u tu2 v
1/s0 ≤t≤u≤v≤1/s

and ω(u) is the Buchstab Function, see (4.1).

The equations Ξ(s) and Ψ1 (s) are derived from Wu’s sieve inequalities, and
they provide a way to rewrite the complicated integral equations (3.14) into an
inequality that can be attacked.

Proposition 3.3.2. For 2 ≤ s ≤ 3 ≤ s0 ≤ 5 and s ≤ k3 ≤ k2 ≤ k1 ≤ s0 such that

s0 − s0 /s ≥ 2, 1 ≤ αi ≤ 3 for (1 ≤ i ≤ 9), α1 < α4 , α5 < α8

are all satisfied, then

Z 3
H(s) ≥ Ψ2 (s) + H(s)Ξ2 (t, s) dt, (3.19)
1

where Ψ2 (s) = Ψ2 (s, s0 , k1 , k2 , k3 ) is given by

Z 0
2 s −1 log(t − 1) 2 k1 −1 log(t − 1)
Z
0
Ψ2 (s, s , k1 , k2 , k3 ) = − dt − dt
5 2 t 5 2 t
0
1 k2 −1 log(t − 1) 1 1−1/s log(s0 t − 1)
Z Z
− dt + dt (3.20)
5 2 t 5 1−1/s t(1 − t)
25
1 1−1/k1 log(k1 t − 1)
Z
2X
+ dt − I2,i (s).
5 1−1/k3 t(1 − t) 5 i=9

and where I2,i (s) = I2,i (s, s0 , k1 , k2 , k3 ) is given by

26 CHAPTER 3. EXAMINING WU’S PAPER

 
φ−t−u−v
Z
dt du dv
I2,i (s) = max ω (9 ≤ i ≤ 15),
φ≥2 u tu2 v
D2,i
 
φ − t − u − v − w dt du dv dw
Z
I2,i (s) = max ω (16 ≤ i ≤ 19),
φ≥2 v tuv 2 w
D2,i
  (3.21)
φ − t − u − v − w − x dt du dv dw dx
Z
I2,20 (s) = max ω ,
φ≥2 w tuv 2 x
D2,20
 
φ − t − u − v − w − x − y dt du dv dw dx dy
Z
I2,21 (s) = max ω .
φ≥2 x tuvwx2 y
D2,21

and the domains of integration are

D2,9 = {(t, u, v) : 1/k1 ≤ t ≤ u ≤ v ≤ 1/k3 },

D2,10 = {(t, u, v) : 1/k1 ≤ t ≤ u ≤ 1/k2 ≤ v ≤ 1/s},
D2,11 = {(t, u, v) : 1/k1 ≤ t ≤ 1/k2 ≤ u ≤ v ≤ 1/k3 },
D2,12 = {(t, u, v) : 1/s0 ≤ t ≤ u ≤ 1/k1 , 1/k3 ≤ v ≤ 1/s},
D2,13 = {(t, u, v) : 1/s0 ≤ t ≤ 1/k1 ≤ u ≤ 1/k2 ≤ v ≤ 1/s},
D2,14 = {(t, u, v) : 1/s0 ≤ t ≤ 1/k1 , 1/k2 ≤ u ≤ v ≤ 1/s},
D2,15 = {(t, u, v) : 1/k1 ≤ t ≤ 1/k2 ≤ u ≤ 1/k3 ≤ v ≤ 1/s}, (3.22)
D2,16 = {(t, u, v, w) : 1/k2 ≤ t ≤ u ≤ v ≤ w ≤ 1/k3 },
D2,17 = {(t, u, v, w) : 1/k2 ≤ t ≤ u ≤ v ≤ 1/k3 ≤ w ≤ 1/s},
D2,18 = {(t, u, v, w) : 1/k2 ≤ t ≤ u ≤ 1/k3 ≤ v ≤ w ≤ 1/s},
D2,19 = {(t, u, v, w) : 1/k1 ≤ t ≤ 1/k2 , 1/k3 ≤ u ≤ v ≤ w ≤ 1/s},
D2,20 = {(t, u, v, w, x) : 1/k2 ≤ t ≤ 1/k3 ≤ u ≤ v ≤ w ≤ x ≤ 1/s},
D2,21 = {(t, u, v, w, x, y) : 1/k3 ≤ t ≤ u ≤ v ≤ w ≤ x ≤ y ≤ 1/s}.

This large set of integrals is derived by finding an integral equation that provides
an upper bound for each term of the form
X
σ(d)Γi
d

Taking the sum of all these integrals will give a bound for Φ, and hence for H(s).
3.3. A LOWER BOUND FOR H(S) 27

The αi terms are given by

α1 := k1 − 2, α2 := s0 − 2,
α3 := s0 − s0 /s − 1, α4 := s0 − s0 /k2 − 1,
α5 := s0 − s0 /k3 − 1, α6 := s0 − 2s0 /k2 ,
α7 := s0 − s0 /k1 − s0 /k3 , α8 := s0 − s0 /k1 − s0 /k2
α9 := k1 − k1 /k2 − 1.

The function Ξ2 (t, s) = Ξ2 (t, s, s0 , k1 , k2 , k3 ) is given by

 
σ0 (t) 1024
Ξ2 (t; s) := log
5t (s − 1)(s0 − 1)(k1 − 1)(k2 − 1)(k3 − 1)
(t + 1)5
 
1[α2 ,3] (t)
+ log
5t (s − 1)(s0 − 1)(k1 − 1)(k2 − 1)(k3 − 1)
 
1[α9 ,α1 ] (t) t+1
+ log
5t (k − 1)(k1 − 1 − t)
 2 
1[α5 ,α2 ] (t) t+1
+ log
5t (k − 1)(s0 − 1 − t)
 3 
1[α3 ,α2 ] (t) t+1
+ log
5t (s − 1)(s0 − 1 − t)
(t + 1)2
 
1[α1 ,α2 ] (t)
+ log
5t (k1 − 1)(k2 − 1)
s0 2
 
1[α7 ,α5 ] (t)
+ log
5t(1 + t/s0 ) (k s0 − s0 − k1 t)(k3 s0 − s0 − k3 t)
 01 0 
1[α5 ,α8 ] (t) s (s − 1 − t)
+ log
5t(1 − t/s0 ) k s0 − s0 − k1 t
 1
s0

1[α6 ,α8 ] (t)
+ log
5t(1 − t/s0 ) k2 s0 − s0 − k2 t
1[α8 ,α2 ] (t)
+ log (s0 − 1 − t) .
5t(1 − t/s0 )
(3.23)

Ξ2 , together with the α terms, are derived by a complicated Lemma [54, p. 248]
relating three separate integral inequality equations.
where
Z b
c 1 σ(3, t + 2, t + 1)
σ(a, b, c) := log dt, σ0 (t) := . (3.24)
a t−1t 1 − σ(3, 5, 4)

where
28 CHAPTER 3. EXAMINING WU’S PAPER

All of these complicated integrals are a way of bounding Wu’s 21 term in-
equality. For each term Γi in the inequality, it has a corresponding integral. For
example, here is one term from the inequality,
XXX
Γ10 = S(Adp1 p2 p3 , {p : (p, dN ) = 1}, p2 }) (3.25)
d1/k1 ≤p1 ≤p2 ≤d1/k2 ≤p3 ≤d1/s

which corresponds to
 
φ−t−u−v
Z
dt du dv
I2,10 (s) = max ω
φ≥2 u tu2 v
1/k1 ≤t≤u≤1/k2 ≤v≤1/s

The sieve is related to the function F , as F is part of the upper bound for the
sieve (3.7). Now if we define the Dickman function ρ(u) by

ρ(u) = 1 1 ≤ u ≤ 2,
(3.26)
(u − 1)ρ0 (u) = −ρ(u − 1) u ≥ 2.

then we can actually write F and f in terms of ω and ρ [19].

 
γ ρ(u)
F (u) = e ω(u) + , u > 0, (3.27)
u
 
γ ρ(u)
f (u) = e ω(u) − , u > 0. (3.28)
u
which provides a way to link the Buchstab function back to the sieve.

3.4 Discretising the integral

Wu comments that obtaining an exact solution is very difficult, and provides a
lower bound on H(s) by splitting up the integrals in 3.15 and 3.19 into 9 pieces.
By letting s0 := 1 and si := 2.1 + 0.1i for i = 1, . . . , 9, and the fact that H(s) is
decreasing on the interval [1, 10], Wu obtains
9
X
H(si ) ≥ Ψ2 (si ) + ai,j H(sj ) (3.29)
j=1

where Z sj
ai,j := Ξ2 (t, si )dt i = 1, . . . , 4; j = 1, . . . , 9
sj−1

and
9
X
H(si ) ≥ Ψ1 (si ) + ai,j H(sj ) (3.30)
j=1
3.4. DISCRETISING THE INTEGRAL 29

where Z sj
ai,j := Ξ1 (t, si )dt i = 5, . . . , 9; j = 1, . . . , 9
sj−1

Now H(si ) is given as a linear combination of all the other H(sj ), 1 ≤ j ≤ 9 values,
which simplifies the problem from resolving a complicated integral equation, to a
simple linear optimisation problem. These discretisations of the integrals can be
written as matrix equations.
 
Ψ2 (s1 )
 .. 
     . 
a1,1 . . . a1,9 H(s1 ) 
Ψ (s )

 .. . . . . 2 4
A :=  . . ..  , H :=  ..  , B :=  (3.31)
    

Ψ1 (s5 )
a9,1 . . . a9,9 H(s9 )  .. 
 
 . 
Ψ1 (s9 )

Thus 3.29 and 3.30 can be rewritten as

(I − A)H ≥ B (3.32)

Wu further simplifies the problem by simply changing the inequality to an equal-

ity, and solves the system of simultaneous linear equations. This provides a lower
bound to the linear optimisation problem given. So the equation below is solved
for X, which is easy to do.
(I − A)X = B (3.33)
Thus, we obtain that
H≥X (3.34)
Chen’s constant is then equal to the largest element in the vector X, as that
corresponds to a lower bound for H(s) at some point s. Bounds on H give us
bounds on the sieve, by (3.7). Naturally, we choose the best element in X (i.e.
the largest) to get the best bound on H. It happens to be that the first element
in X is always the biggest, as the function Ψ2 (s) is maximised at approximately
s = 2.2. Now, by taking σ = {1} and s = 2.2 in the sum of sieves Φ(N, σ, s)
(3.10)
X
Φ(N, {1}, 2.2) = σ(d)S(Ad , {p : (p, dN ) = 1}, d1/2.2 ),
d

= S(A, {p : (p, N ) = 1}, d1/2.2 ),

= S(A, {p : (p, N ) = 1}, N (1/2−δ)/2.2 ),
30 CHAPTER 3. EXAMINING WU’S PAPER

By letting
A(s) = sF (s)/2eγ
and from the definition of Θ(N, σ),
X CN
Θ(N, {1}) = 4Li(N )
d
log d
4Li(N )CN
=
log N 1/2−δ

Therefore, [54, p. 253]

4Li(N )CN
Φ(N, {1}, 2.2) ≤ {A(2.2) − Hk,N0 (2.2)} ≤ 8{1 − x1 }Θ(N )
log N 1/2−δ

where x1 is the first element of the vector X (3.34).

Thus, Wu obtains an upper bound for Chen’s constant.
Chapter 4

Approximating the Buchstab

function

In this section we examine how the Buchstab Function ω(u) (which appears in
most of Wu’s integrals) is computed.

4.1 Background
The Buchstab function is defined by the following delay differential equation∗

ω(u) = 1/u 1 ≤ u ≤ 2,
(4.1)
(uω(u))0 = ω(u − 1) u ≥ 2.

From the graph it appears that ω(u) quickly approaches a constant value.
Buchstab [4] showed that
lim ω(u) = e−γ , (4.2)
u→∞

(where γ is the Euler–Mascheroni constant) and that the convergence is faster

than exponential, i.e
|ω(u) − e−γ | = O(e−u ). (4.3)

Hua [24] gave a much stronger bound

|ω(u) − e−γ | ≤ e−u(log u+log log u+(log log u/ log u)−1)+O(u/ log u) . (4.4)
∗
The definition of ω(u) is very similar to the other difference equations F and f (3.5) and
(3.6).

31
32 CHAPTER 4. APPROXIMATING THE BUCHSTAB FUNCTION

Figure 4.1: Plot of the Buchstab Function ω(u) for 1 ≤ u ≤ 4.

A slightly stronger bound was obtained numerically in Mathematica over the

interval of interest,
∀u ∈ [2, 10.5], |ω(u) − e−γ | ≤ 0.38e−1.275u log u .
The Buchstab function is related to rough numbers; numbers whose prime
factors all exceed some value. If we define
Φ(x, y) = #{n ≤ x : p|n ⇒ p ≥ y}, (4.5)
i.e. the number of positive integers with no prime divisors below y, the following
limit holds [9]
lim Φ(y u , y)y −u log y = ω(u). (4.6)
y→∞
If we define
W (u) = ω(u) − e−γ ,
then W (u) mimics a decaying periodic function. The period of the decay is slowly
increasing, and is expected to limit to 1. Also, it has been shown that in every
interval [u, u + 1] of length 1, W (u) has at least one, but no more than two,
zeros. [6] If we let c0 = 1, c1 = 2 and c2 , c3 , . . . denote the critical points of
the Buchstab function, [6] we have that c1 , c3 , c5 , . . . are local minimums that are
strictly increasing, and c2 , c4 , . . . are local maximums that are strictly decreasing,
in the sense that
1
= ω(c1 ) < ω(c3 ) < ω(c5 ) < . . . (4.7)
2
1 = ω(c0 ) > ω(c2 ) > ω(c4 ) > . . . (4.8)
4.2. A PIECEWISE TAYLOR EXPANSION 33

This allows us to obtain the trivial bound

1/2 ≤ ω(u) ≤ 1. (4.9)

Importantly, the Buchstab function decays very quickly to a constant, which

makes approximating it easy for the purposes of computing Chen’s constant. For
1 ≤ u ≤ 2, ω(u) has a closed form, and since it converges so quickly to e−γ , we can
set ω(u) to be a constant when u is sufficiently large. Cheer–Goldston published
a table of the critical points and zeros of W (u), and by using the critical points

W (9.72844) = −9.58 × 10−14 , W (10.52934) = 3.57 × 10−15

we can conclude using (4.7) that

∀u ≥ 10, |W (u)| ≤ 10−13 . (4.10)

For the purposes of numerically integrating ω(u), we only consider the region
2 ≤ u ≤ 10, as beyond that, we can set ω(u) = e−γ for u ≥ 10 and bound the
error by (4.10).

4.2 A piecewise Taylor expansion

A closed form for the Buchstab function can be defined in a piecewise manner for
each interval n ≤ u ≤ n + 1, as the difference equation can be rearranged such
that the value in a particular interval depends on its predecessor.

Theorem 4.2.1. [6] If we define

ωj (u) = ω(u), j ≤u≤j+1

and integrate (4.1) we obtain

Z u−1
uωj+1 (u) = ωj (t)dt + (j + 1)ωj (j + 1) for j + 1 ≤ u ≤ j + 2. (4.11)
j

Proof. ω(u − 1) is defined for all u ≥ 2, so by integrating (4.1),

Z u
uω(u) = ω(t − 1)dt,
2

and then by changing variables

Z u−1
uω(u) = ω(t)dt.
1
34 CHAPTER 4. APPROXIMATING THE BUCHSTAB FUNCTION

Restrict u such that j + 1 ≤ u ≤ j + 2.

Z u−1
uωj+1 (u) = ωj (t)dt
1
Z u−1 Z j
= ωj (t)dt + ω(t)dt
j 1
Z u−1 Z j
uωj+1 (u) − ωj (t)dt = ω(t)dt.
j 1

Rj
Observe that 1
ω(t)dt is a constant. Call this constant K.
Z u−1
∀u ∈ [j + 1, j + 2], K = uωj+1 (u) − ωj (t)dt.
j

Choose u = j + 1.
K = (j + 1)ωj+1 (j + 1).
To force the Buchstab function to be continuous, we assume that the piecewise
splines agree at the knots, that is,

ωj (j + 1) = ωj+1 (j + 1).

So K = (j + 1)ωj (j + 1) and hence

Z u−1
uωj+1 (u) = ωj (t)dt + (j + 1)ωj (j + 1).
j

By definition, ω1 (u) = u−1 . To obtain ω2 (u), apply (4.11).

Z u−1
uω2 (u) = ω1 (t)dt + 2ω1 (2)
2
Z u−1
1 1
= dt + 2
2 t 2
= log(u − 1) − log(2 − 1) + 1
log(u − 1) + 1
ω2 (u) = , 2 ≤ u ≤ 3.
u
All values of ωn (u) can, in principle, be obtained by repeating the above process,
but they cannot be expressed in terms of elementary functions, as
log(u − 1) + 1
Z
du,
u
4.2. A PIECEWISE TAYLOR EXPANSION 35

is a non-elementary integral. The technique Cheer–Goldston [6] used expresses

each ωj (u) as a power series about u = j + 1.
∞
X
ωj (u) = ak (j)(u − (j + 1))k . (4.12)
k=0

For the case of ω2 (u) = log(u−1)+1

u
, both log(u−1) and u−1 are analytic, so we may
write them in terms of their power series about u = 3. Cheer–Goldston shows
that by doing this, one obtains
1 + log 2
a0 (2) = , (4.13)
3
k−1  m !
k+1 1 + log 2 1 X 1 2
ak (2) = (−1) − k+1 + k
. (4.14)
3 3(2 ) m=0 k − m 3

We note that
k−1  m Xk−1  m ∞  m
X 1 2 2 X 2
≤ ≤ = 3.
m=0
k−m 3 m=0
3 m=0
3

Whence it follows that

1 + log 2 1 1
|ak (2)| ≤ − k+1
+ k ≤ k.
3 2 2
To obtain the value of ak (j) in general, we follow the method of Marsaglia et al.
[35] and substitute (4.12) into (4.11) to obtain
X∞ Z u−1 X∞
k
u ak (j +1)(u−(j +2)) = ak (j)(t−(j +1))k dt+(j +1)a0 (j). (4.15)
k=0 j k=0

The Taylor expansion of ω2 (u) can be shown to converge uniformly in the interval
1.5 ≤ u ≤ 4.5, as
∞ ∞ ∞ k ∞  k
X
k
X 1 k
X u−3 X 3
|ω2 (u)| ≤ |ak (2)||u − 3| ≤ k
|u − 3| ≤ ≤ = 4.
k=0 k=0
2 k=0
2 k=0
4

Hence by the Weierstrass M-test[15], the series for ω2 (u) is uniformly convergent
for 1.5 ≤ u ≤ 4.5. This allows us to compute the integral of the Taylor series of
ω2 (u), (and thereby compute ω3 (u)) by swapping the order of the sum and the
integral, by the Fubini-Tonelli theorem [46]. By using (4.11) we can prove by
induction that the Taylor series of ωj (u) converges uniformly for j ≤ u ≤ j + 1.
This allows interchange of the sums and integrals in (4.15). Thus, we obtain
∞ ∞
(u − (j + 2))k+1 − (−1)k+1
X X  
k
u ak (j+1)(u−(j+2)) = (j+1)a0 (j)+ ak (j) .
k=0 k=0
k + 1
36 CHAPTER 4. APPROXIMATING THE BUCHSTAB FUNCTION

By making a substitution x = u − (j + 2) we obtain

∞ ∞ ∞
X
k+1
X
k
X ak−1 (j) k
x + (−1)k−1 .


ak (j+1)x +(j+2) ak (j+1)x = (j+1)a0 (j)+
k=0 k=0 k=1
k
By subsequently equating coefficients of like terms we obtain the following recur-
sive formula for ak (j), for all k ≥ 0, j ≥ 2.
∞
1 X (−1)k
a0 (j) = a0 (j − 1) + ak (j − 1), (4.16)
j + 1 k=1 k + 1
ak−1 (j − 1) − kak−1 (j)
ak (j) = . (4.17)
k(j + 1)
where the base case ak (2) and a0 (2) are given in (4.13). Now one can define the
Buchstab function in a piecewise manner

1/u 1 ≤ u ≤ 2,
ω(u) = (4.18)
ωj (u) j ≤ u ≤ j + 1.

4.2.1 Errors of Taylor expansion

In computing ω(u), we truncate the Taylor expansion to some degree N , and
then compute the coefficients of the resulting polynomials to a given accuracy
[6]. Denote the approximation of ω2 (u) as T2 (u), and define the error to be
E2 (u) = ω2 (u) − T2 (u). Define the worst case error as

E = max |E2 (u)|.

2≤u≤3

By substuting T2 (u) into (4.11), Cheer [6] shows that one obtains a new approx-
imation T3 (u) for ω3 (u), with a new error term of
Z u−1 
1
|E3 (u)| = E2 (t)dt + 3E2 (3)
u 2
Z u−1  (4.19)
1 (u − 3)E + 3E
≤ Edt + 3E = = E.
u 2 u
By repeating this argument and then by induction,

∀j ≥ 2 max |ωj (u) − Tj (u)| = E.

j≤u≤j+1

So the accuracy of ω2 (u) holds for the rest of the ωk (u), excluding computational
errors due to machine precision arithmetic † [6]. (If needed, Mathematica supports
†
Real number arithmetic on a computer is implemented using floating point numbers, which
only have finite accuracy. Hence, every operation introduces rounding errors, and with enough
operations, can cause issues with the final result.
4.2. A PIECEWISE TAYLOR EXPANSION 37

arbitrary precision arithmetic, so the magnitude of the computations errors could

be made arbitrarily small.)
The error E is easily computed, as by the Taylor remainder theorem [47], for
every u ∈ [2, 3] there exists a ξ ∈ [2, 3] such that
N
X
ω2 (u) = ak (2)(u − 3)k + aN +1 (2)(ξ − 3)N +1 . (4.20)
k=0

Then by definition, the error term (for a Taylor expansion to order N ) can be
written as
1
EN = max |aN +1 (2)(ξ − 3)N +1 | ≤ |aN +1 (2)| ≤ N +1 . (4.21)
2≤ξ≤3 2
So if ω2 (u) is approximated with a power series up to order N , then the maximum
error for ω(u) anywhere is 2−(N +1) .

4.2.2 Improving Cheer’s Method

The approximation of the Buchstab function was improved, by computing the
power series for each ωj (u) about j + 0.5. This way, ωj (u) was evaluated at most
0.5 away from the centre of the Taylor expansion. In the previous case, ωj (u)
could be evaluated up to 1 away from the centre. This provided a much lower
error for the same degree Taylor expansion, improving the accuracy of all results
using the Buchstab function. By a similar technique as above, by expanding
about the middle of each interval [j, j + 1] we obtain
∞
X
ωj (u) = ak (j)(u − (j + 1/2))k for j ≤ u ≤ j + 1, (4.22)
k=0

where
3

 k+1 X k−1  m !
1 + log 3 2 1 3
ak (2) = (−1)k+1 − 2
k+1
+ , (4.23)
(5/2) 5 3 m=0
k − m 5
∞  k

1 X ak (j − 1) (−1)
a0 (j) = k
j+ , (4.24)
j + 1/2 k=0 2 2(k + 1)
 
1 ak−1 (j − 1)
ak (j) = − ak−1 (j) . (4.25)
j + 1/2 k
Applying the Taylor remainder theorem, we calculate the error of the Taylor
expansion of ω2 (u), truncated to N terms:
1
E := max |ω2 (u) − T2 (u)| ≤ |aN +1 (2)||(3 − 2.5)N +1 | ≤ |aN +1 (2)|. (4.26)
2≤ξ≤3 2N +1
38 CHAPTER 4. APPROXIMATING THE BUCHSTAB FUNCTION

The coefficient aN +1 (2) can be bounded above, as

k−1  m Xk−1  m ∞  m
X 1 3 3 X 3 5
≤ ≤ = .
m=0
k−m 5 m=0
5 m=0
5 2

So,

 N +2  N +1
1 + log 32


3 2 5 2
|aN +1 (2)| ≤ − N +2
+ ≤ . (4.27)
(5/2) 5 3 2 3

Hence the error bound is improved:

 N +1
1 2 1
E≤ ≤ N +1 . (4.28)
2N +1 3 3

which is better than the old error bound (4.21) by an exponential factor. Again,
by a similar argument to (4.19), this error E can be shown to hold everywhere.
In practice this is a rather weak error bound, as the actual error is much less. By
computing
1
|aN +1 (2)|,
2N +1

for 10 ≤ N ≤ 20, (rejecting the first few N until the points settle out) and
plotting the values on a log plot, we observe the values form a line. By fitting a
curve of the form

log y = mx + c,

to this line, the asymptotic behaviour of the numerical upper bound error is
deduced to be approximately O(3.31−N ). The base case for the above derivations
was done with ω2 (u) instead of ω1 (u) as the Taylor expansion of 1/u converges
slowly. The corresponding error bounds obtained are much weaker. One could, in
principle, use ω3 (u) as the base case and thus obtain a much stronger error bound
for ωj (u), j ≥ 3, but as mentioned above, ω3 (u) is not an elementary function.
So, ω2 (u) is the best we could hope to use. If one used the power series expression
of ω3 (u) to compute the error bounds, each ak (3) would be defined in terms of
4.2. A PIECEWISE TAYLOR EXPANSION 39

ak (2) as shown,
∞
(−1)k
 
1 X
−k
a0 (3) = ak (2)2 j+
3 + 1/2 k=0 2(k + 1)
∞ k−1
 k+1 X  m !
1 + log 23


1 X (−1)k+1 3 2 1 3
= k
− k+1
+
3 + 1/2 k=0 2 (5/2) 5 3 m=0
k−m 5
(−1)k
 
j+ ,
2(k + 1)
 
1 ak−1 (2)
ak (3) = − ak−1 (3) .
3 + 1/2 k

and so, expanded out in full, the formulas for the coefficients would become (even
more) unwieldy. Since the error E decreases exponentially quickly with respect
to the degree of the polynomial, our implementation of the Buchstab function
(given above) was deemed sufficient.
40 CHAPTER 4. APPROXIMATING THE BUCHSTAB FUNCTION
Chapter 5

Numerical computations

We proceeded by implementing the above in Mathematica, using the definition of

the Buchstab function given in Chapter 4. The approximation to the Buchstab
function used is defined as
k
X
−γ
ωT (u) = 1[k+1,∞) e + 1[j,j+1) ωj (u),
j=1

where ωj (u) is the Taylor polynomial approximation to ω(u) around u = j + 1/2,

to some degree N (4.22). The Buchstab function ω(u) was approximated using a
polynomial spline over k intervals, and declared equal to the limit eγ (4.2) beyond
the k th interval.

5.1 Justifying Integration Method

We attempted to compute the first of Wu’s integrals (3.18) without maximising
with respect to φ
 
φ − t − u − v dt du dv
ZZZ
0
I1 (s, s , φ) = ωT , (5.1)
u tu2 v
1/s0 ≤t≤u≤v≤1/s

but ran into several problems. Using Mathematica’s inbuilt integration routine,
the computation would either finish quickly, or never halt, depending on the value
of φ chosen. It was discovered that the function could only be integrated for large
φ, such that ωT (u) would always be in the constant region. These problems were
not resolved. Instead, we integrated the Buchstab function by computing an anti-
derivative of ωT (u), and applying the fundamental theorem of calculus (FTC),
three times.

41
42 CHAPTER 5. NUMERICAL COMPUTATIONS

We can show that FTC validly applies in this instance. As ωT (u) is a piece-
wise spline of polynomials, it is continuous everywhere except at the points
u = 2, 3, . . . , k + 1 where the splines meet. Using two theorems about Lebesgue
integration [46]

Theorem 5.1.1. If g is integrable, and 0 ≤ f ≤ g, then f is integrable.

Theorem 5.1.2. If f is integrable on [a, b], then there exists an absolutely con-
tinuous function F such that F 0 (x) = f (x) almost everywhere, and in fact we
Rx
may take F (x) = a f (y) dy.

The Buchstab function is strictly positive and bounded, as 1/2 ≤ ω(u) ≤ 1

(4.9). The Taylor approximation ωT (u) will also be non-negative and bounded,
as we can easily bound the difference between ωT (u) and ω(u) to be less than
any  > 0 (4.28). Thus if ωT (u) is defined using k many splines, we can bound
ωT (u) above by a constant on the interval [1, k + 1]. Beyond u = k + 1, ωT (u)
is constant, so ωT (u) is integrable on any interval of the form [1, k + 1 + r] for
some r > 0. We can thus obtain FTC as
Z x Z b
F (x) = f (y) dy ⇒ F (b) − F (a) = f (x) dx.
a a

We remove the maximisation over φ, and set I1 to be a function dependant on φ.

We can convert the domain of integration

{(t, u, v) : 1/s0 ≤ t ≤ u ≤ v ≤ 1/s}

into three definite integrals, which is more suitable to apply FTC.

Z v=1/s Z t=v Z u=v  
φ − t − u − v 1
I˜1 (s, s , φ) =
0
ωT du dt dv. (5.2)
v=1/s0 t=1/s0 u=t u tu2 v

it was expected that the FTC could be used 3 times to compute I1 , however the
values computed were nonsense. When attempting to replicate the entry in Wu’s
table (5.1) for i = 6, we expected 0.0094 . . . and obtained a large negative result,
−10000 or so. It is believed that since ωT (u) was defined in a piecewise manner,
integrating ωT φ−t−u−v


u
would involve resolving many inequalities, which would
only increase in complexity after three integrations.
Attempts to numerically integrate ωT (u) were successful, however it was much
faster to use Mathematica to numerically solve the differential delay equation
defining the Buchstab function (4.1) with standard numerical ODE solver rou-
tines, obtain a numerical solution for ω(u), and numerically integrate the result.
 5.1. JUSTIFYING INTEGRATION METHOD 43

i si s0i k1,i k2,i k3,i Ψ1 (si ) (Wu) Ψ1 (si ) Ψ2 (si ) (Wu) Ψ2 (si )
1 2.2 4.54 3.53 2.90 2.44 0.01582635 0.01615180
2 2.3 4.50 3.54 2.88 2.43 0.01224797 0.01547663
3 2.4 4.46 3.57 2.87 2.40 0.01389875 0.01406834
4 2.5 4.12 3.56 2.91 2.50 0.01177605 0.01187935
5 2.6 3.58 0.00940521 0.00947409
6 2.7 3.47 0.00655895 0.00659089
7 2.8 3.34 0.00353675 0.00354796
8 2.9 3.19 0.00105665 0.00105838
9 3.0 3.00 0.00000000 0.00000000

Table 5.1: Table of values for Ψ1 and Ψ2 , compared with Wu’s results [54].

The resulting values were also closer to Wu’s. The following table includes Wu’s
results, compared to the results we obtained.

Given an si , Wu chose the parameters s0i , k1,i , k2,i , k3,i to maximise Ψ1 (si ) or
Ψ2 (si ). All the values we calculated were an overshoot of Wu’s results, so it was
not surprising that we obtained a smaller value for Chen’s constant (3.32). For
i = 5, 6, . . . , 9, the values of s0i Wu gave were verified to maximise our version of
Ψ1 (si ), by trying all values s0i = 3, 3.01, . . . , 5.00, which is given by the constraint
3 ≤ s0i ≤ 5 on (3.17). So even though our Ψ1 did not match Wu’s, both were
maximised at the same points. This indicated that we could use our “poor mans”
Φ1 to investigate the behaviour of Φ1 , but not to compute exact values for it.

When computing the values for this table, we first wanted to see for a fixed s
and s0 , how I1 (s) varied as a function of φ. It was determined that the function
did not grow too quickly near the maximum, and as φ grew large, Ψ1 tended
to a constant, which is consistent with ω(u) quickly tending to e−γ . Thus, we
can apply simple numerical maximisation techniques, without getting trapped in
local maxima, or missing the maximum because the function changes too quickly.
We only need to maximise φ over a small interval φ ∈ [2, 4], and we obtain the
maximal value by computing I1 for a few points in the interval of interest, choosing
the maximum point, and then computing again for a collection of points in a small
neighbourhood about the previous maximum. This is iterated a few times until
the required accuracy is obtained. (Listing 5.1).
44 CHAPTER 5. NUMERICAL COMPUTATIONS

Figure 5.1: Plot of I1 (φ, 2.6, 3.58) for 2 ≤ φ ≤ 4

def I1 (s, s0 ):
φlow := 2
φhigh := 4
 := 0.1
φM AX := 2
while ( ≥ 0.001):
Φ := {φlow , φlow + , φlow + 2, . . . , φhigh }
φM AX := φ ∈ Φ such that I(φ, ˜ s, s0 ) is maximal
φlow := max(φM AX − , 2)
φhigh := φM AX + 
 := /10
return I˜1 (φM AX , s, s0 )
enddef
Listing 5.1: Algorithm to compute I1 (s, s0 )

Initially, it seemed odd that for all the values of s and s0 that were tried,
I˜1 (φ, s, s0 ) appeared to always be maximal when φ = 2. The actual maximum
for I˜1 (φ, s, s0 ) occurs at a point φ < 2, but was being cut off by the constraint
that φ ≥ 2. It seemed unusual to us that Wu would maximise an integral over
φ ≥ 2, when the integral is always maximal when φ = 2. However, for some of
the other integrals in (3.21), the function was maximal at some point φ > 2. This
behaviour is made clear when we plot I2,11 (φ, s = 2.6, s0 = 3.58) (Figure 5.2) as a
function of φ, for a fixed value of s and s0 .
For some of the integrals needed to compute I2 , the numerical integration
5.1. JUSTIFYING INTEGRATION METHOD 45

Figure 5.2: Plot of I2,11 for s = 2.6, s0 = 3.58, 2 ≤ φ ≤ 2.5

did not converge. Whether this was a property of the integrals (i.e. they were
not defined for those particular values of φ) or a symptom of how the numerical
integration routines in Mathematica operate was not determined. To maximise
the integrals with respect to φ, we could no longer assume they were defined
for all φ ≥ 2. Instead, it was found that for each integral, there would be a
corresponding value φlow such that the method of numerical integration converged
for all φ ≥ φlow . This value of φlow was computed by method of bisection (Listing
5.2).
def findphi (i, s, s0 , k1 , k2 , k3 ):
if ( computing I2,i (2, s, s0 , k1 , k2 , k3 ) suceeded ):
return 2
φlow := 2
φhigh := 5
 := 0.001
while (|φhigh − φlow | ≥ ):
φmid := (φhigh + φlow )/2
if ( computing I2,i (φmid , s, s0 , k1 , k2 , k3 ) suceeded ):
φhigh := φmid
else :
φlow := φmid
return φhigh
enddef
Listing 5.2: Algorithm to compute the least φ such that I2,i (φ, s) is defined
46 CHAPTER 5. NUMERICAL COMPUTATIONS

I1 and I2,i can now be computed, and thus Ψ1 (3.17) and Ψ2 (3.20) can be
easily computed using standard numerical integration techniques. This gives the
entries of the vector B (3.31). Computing the matrix A (3.31) is comparatively
easy, as each element ai,j of the matrix A is the integral of either Ξ1 (see 3.16) or
Ξ2 (see 3.16) over an interval, which is easily computed numerically. Thus, the
system of linear equations (I − A)X = B (where I is the identity matrix) can be
solved for X. This is compared to the value for X that Wu obtained.
   
0.0223939 . . . 0.0227656 . . .
   
0.0217196 . . .  0.0219942 . . . 
   
0.0202876 . . .  0.0205028 . . . 
   
0.0181433 . . .  0.0182930 . . . 
   
   
XWu =  0.0158644 . . .  X =  0.0152937 . . . 
   (5.3)
0.0129923 . . .  0.0126404 . . . 
   
   
0.0100686 . . .  0.0099076 . . . 
   
0.0078162 . . . 0.0078020 . . .
   
0.0072943 . . . 0.0073089 . . .

So from these vectors, we use (3.34) to obtain two different values for C ∗ , ours
and Wu’s,
∗
CW u = 8(1 − 0.0223938) ≤ 7.82085,
(5.4)
C ∗ = 8(1 − 0.0227655) ≤ 7.8178752,

which seems to be an improvement on Chen’s constant, but since the calculations

used to obtain this value are numerical in nature and don’t match Wu’s results,
it can be hard to conclude anything useful from them. Nonetheless, the effect on
Chen’s constant of sampling si with finer granularity was investigated. We could
no longer use the table Wu had given (5.1), so we had to numerically optimise over
both φ and s0i . Each evaluation of Ψ2 is very expensive, as we have to compute
each of the integrals given in (3.21). This means it is computationally difficult to
compute Ψ2 (si ) without the corresponding parameters φ, s0i , k1,i , k2,i , k3,i , as for
an interval size of n points, we now need to compute n5 points, which can quickly
become infeasible. We focused on increasing the sample size for Ψ1 , by diving the
interval of [2, 3] into [2, 2.6] ∪ [2.6, 3.0], and increasing the resolution of sampling
in the interval [2.6, 3.0]. By defining

2.1 + 0.1i 1 ≤ i ≤ 4,
s0i =
2.6 + 0.01(i − 5) 5 ≤ i ≤ 45.
5.1. JUSTIFYING INTEGRATION METHOD 47

we use the same values before (See 5.1) for 1 ≤ i ≤ 4 and then compute new
values of Ψ1 for each new value of si , so we evaluate Ψ1 at 40 points instead
of 5. We can then compute Ψ1 by numerically maximising over s0i , in a similar
fashion to computing I1 (See Listing 5.1). Note that Ψ1 depends on I1 , so we are
optimising with respect to two variables, φ and s0i . We also need to recompute
the matrix A given by (3.31) on the new set of intervals. Thus, the component
ai,j of A is now given by
Z s j
Ξ2 (t, si )dt i = 1, . . . , 4; j = 1, . . . , 45




sj−1
ai,j := Z sj (5.5)
Ξ1 (t, si )dt i = 5, . . . , 45; j = 1, . . . , 45




sj−1

So we now have a new matrix A and vector B with which to solve (I − A)X = B.

 
0.0228801 . . .
0.0221067 . . .
   
Ψ2 (s1 )
 
 0.0206136 . . .
 ..  
0.0184024 . . .

   . 
a1,1 . . . a1,45    
0.0184024 . . .
 Ψ (s )   
 .. .. ..   2 4  
A :=  . . .  , B :=  ⇒X=
 
 Ψ1 (s5 )   0.0153999 . . . 

a45,1 . . . a45,45  .. 
  
0.0157765 . . .

 .  



0.0163132 . . .
Ψ1 (s45 )  
0.0170428 . . .
 
..
.

Once that was complete, we redid the calculations again with even finer gran-
ularity for s0i , choosing

2.1 + 0.1i 1 ≤ i ≤ 5,
s0i =
2.6 + 0.001(i − 5) 6 ≤ i ≤ 405.

These calculations were run in parallel on a computer equipped with an i5-

3570K CPU overclocked to 3.80Ghz, and with 6GB of RAM. The time taken to
compute additional intervals increases rapidly, the projected time for 4000 inter-
vals would take longer than a day to compute, with likely minimal improvement
on the constant.
48 CHAPTER 5. NUMERICAL COMPUTATIONS

Points Time B Time A x1 C∗

Wu - - 0.0223938 7.82085
5 1m27s < 1s 0.0227655 7.8178752
40 11m31s 10s 0.0228800 7.81696
400 1h53m 11m12s 0.0229275 7.81658

Table 5.2: Improvements on C ∗ by diving [2.6, 3.0] into more points

5.2 Interpolating Wu’s data

Since Ψ2 is complicated to evaluate, we looked at approximating it by interpola-
tion, so see what effects additional values of Ψ2 would have on Chen’s constant.
We assume the best possible case where possible, to obtain the best possible
value for C ∗ . By demonstrating that even under the best assumptions the im-
provements on C ∗ are minimal, this should indicate that improving C ∗ using Wu’s
integrals will give a similarly small improvement. Ψ2 appears to be decreasing by
Chen’s table, and so too for Ψ1 , so we interpolate the points with Mathematica
in-built interpolation routine. By plotting the interpolant of Ψ2 and Ψ1 , we can
see one overtakes the other at around s = 2.55. This is why Wu had split up his
table in this way, choosing Ψ1 for the lower half and Ψ2 for the upper, to get the
best possible bound.
We define 
Ψ (s) s ≤ m
2
ΨB = (5.6)
Ψ1 (s) s ≥ m

where m is the point where Ψ2 (m) = Ψ1 (m). Thus, ΨB is the best case scenario,
by essentially taking the maximum of the two upper bounds. We need the values
for s0 , k1 , k2 , k3 to compute C ∗ , as we need to compute the matrix A, which in
turn depends on the values of ai,j (see 5.5). So we obtain those variables by
interpolation also.
We then discretise the interval [2, 3] from coarse to fine, and calculate the new
value of C ∗ for each of these discretisations. In doing so, we can get an estimate
as to how much additional intervals of integration impact the value of C ∗ . After
computing C ∗ for a varying number of intervals, from Wu’s original 9 up to 600
(which took 84 minutes, 40 seconds on the same hardware used to compute Table
5.2), we can clearly see that adding more intervals does make a slight difference,
but there are diminishing returns beyond a hundred or so. One would expect not
to make any gains on C ∗ by attempting to compute a million intervals.
5.2. INTERPOLATING WU’S DATA 49

(b) Interpolant of ΨB , the best of Ψ1 and

(a) Interpolant for s0 Ψ2 put together

(c) Interpolant of Ψ2 (Orange) and Ψ1 (d) Interpolants for s0 , k , k , k (top to bot-

1 2 3
(Blue) from Wu’s data tom)

Figure 5.3: Various interpolants of Wu’s data

Figure 5.4: Value of C ∗ versus number of intervals used in integration.

50 CHAPTER 5. NUMERICAL COMPUTATIONS
Chapter 6

Summary

6.1 Buchstab function

We have improved on the approximation of the Buchstab function ω(u) that
Cheer–Goldston [6] gave. We used a spline, where each polynomial in the spline
is the Taylor expansion of ω(u) about the centre u = j + 0.5 of the interval
[j, j + 1], whereas Cheer–Golston expanded about the right endpoint u = j + 1
of this interval. We derived an implicit formula for the coefficients aj (k) for the
power series
∞
X
ωj (u) = aj (k)(u − (j + 0.5))k ,
k=0

where each aj (k) is defined in terms of the previous values an (m) for 2 ≤ n <
j, 0 ≤ m < k, and the values of a2 (k) are derived from the Taylor expansion of
log(u − 1) + 1
ω2 (u) = .
u
The coefficients are more complicated than the ones Cheer–Goldston obtained,
but when the Taylor approximation of ω(u) is truncated to degree N , we show
that the corresponding error is less than 1/3N +1 (4.28) versus Cheer’s error bound
of 1/2N +1 .
Although the spline is not continuous at the knots, the size of the discontinu-
ities can be made arbitrarily small by increasing the degree of the power series.

6.2 Chen’s constant

An attempt was made to use our approximation to the Buchstab function to
evaluate the integrals Wu gave, to place an upper bound on C ∗ . This was not

51
52 CHAPTER 6. SUMMARY

successful, so instead we numerically solved the ODE that defined ω(u), and nu-
merically evaluated Wu’s integrals instead. This gave us approximately the same
results as Wu. Wu resolved a set of functional inequalities using a discretisation of
[2, 3] into 9 pieces, so we replicated the results, and then increased the resolution
for the section [2.6, 3.0] of the interval from 5 points to 40, and then 400. The
corresponding change in C ∗ was as expected, minimal. The relative difference in
C ∗ between sampling 5 points and 400 points in the interval [2.6, 3.0] was only
1.66 × 10−2 %. This leads us to believe that we should expect similar results, had
we computed an exact form of Wu’s integrals.
We interpolated the values of Ψ1 , Ψ2 and the variables required from Wu’s
data, and (under the assumption that these interpolations are a reasonable ap-
proximation to the exact form), we have shown that Wu was indeed right in
asserting that cutting [2, 3] into more subintervals would result in a better value
for C ∗ , but only by a minuscule amount.
Wu gives us that H(s) is decreasing on [1, 10]. The constraint on the param-
eter s is that 2 ≤ s ≤ 3, but Wu’s discretisation starts at s = 2.2. Wu claims
this was chosen as Ψ2 attains its maximum at approximately s = 2.1. How close
this value of s is to the true maximal value is not stated, so this could be an-
other avenue for improvement. One could explore the value of Ψ2 in a small
neighbourhood around s = 2.1, to see what the true maximum is.
Slightly nudging this value left of 2.2 to, say, 2.19 might mean a better upper
bound for H, and thus a better value for Chen’s constant.

6.3 Future work

There is still a lot of work that could be done to decrease Chen’s constant. The
obvious improvement is to find a way to compute Ψ∗ using the approximation
of ω(u) given above. We could also look at attempting to get the current im-
plementation of Ψ running quicker, which would allow sampling more points in
the interval [2, 2.5], and finding the coefficients φ, s0i , k1,i , k2,i , k3,i that make Ψ2
maximal, as the current version of Ψ2 is too slow to optimise over 5 variables
simultaneously. We could look at more clever ways to maximise Ψ using more
advanced numerical maximisation methods. The behaviour of Ψ is not patholog-
ical, in the sense that the functions appear to have some degree of smoothness,

∗
Whenever Ψ is mentioned in this section, take it to mean both Ψ1 and Ψ2 .
6.3. FUTURE WORK 53

and the derivative is not too large. It seems true that Ψ are locally concave †
near the local maximum, so once values for the variables have been found that
bring Ψ near the maximum, one could then run a battery of standard convex
optimisation techniques on the problem.
We could also look at alternative forms of approximating ω(u). Since we are
never concerned with the values of ω(u) at single points, but rather the integral
of ω(u) over intervals, we could look to approximations of ωj (u) on each interval
[j, j + 1] that minimise the L∞ error

||ωj (u) − ω(u)||L∞ [j,j+1] = sup |ωj (u) − ω(u)|,

u∈[j,j+1]

as Taylor expansions are by no means the best polynomial approximation with

respect to L∞ error. Chebyshev ‡ or Hermite § polynomials are ideal for this
purpose, and the speed of computing the approximation is not a concern, as we
only need to compute the polynomial approximation of ω(u) once. So the new
Chebyshev approximation would be identical to the current approximation, but
the polynomials would have different coefficients. We could also look at Bernstein
polynomials. ¶
As ω(u) is continuous, we have by the Weierstrass Approximation Theorem[15]
that for any  > 0 and any interval [1, n] there exists a polynomial p(x) such that
||ω(u)−p(u)||L∞ [1,n] ≤ . Bernstein polynomials provide a constructive way to find
such a polynomial p(x). This might prove beneficial as integrating ω( φ−t−u−v u
)
in I1 (3.18) would prove easier if ω(u) were represented as a single polynomial
instead of a spline, but since ω(u) is not smooth and in fact has a cusp at u = 1,
the degree of a polynomial required to closely match ω(u) to within say 106
everywhere could be massive.
†
A function f is concave if for any two points x1 ≤ x2 , the unique line g that intersects
(x1 , f (x1 )) and (x2 , f (x2 )) satisfies f (z) ≥ g(z) for all x1 ≤ z ≤ x2 .
‡
The Chebyshev polynomials are a set of polynomials defined by the recursive formula
T0 (x) = 1, T1 (x) = x, Tn+1 (x) = 2xTn (x) − Tn−1 (x). They have the property that the error
incurred by using them to approximate a function is more uniformly distributed over the interval
the function is being approximated. Taylor series are exact at a single point, and the error tends
to grow quickly as points move away from the center of the expansion [5].
§
Hermite polynomials are similar to Chebyshev polynomials, but Hermite polynomial ap-
proximation also takes into account the derivative of the function being approximated, and
attempts to match both the values of the function, and it’s derivative [5].
Pn
¶
The Bernstein polynomials are a family of polynomials, given by Bn (x) = k=0 nk xk (1 −

x)n−k f (k/n). This is the nth Bernstein polynomial for the function f . It can be shown that
||f − Bn ||L∞ ([0,1]) → 0 i.e. that Bn converges to f uniformly.[15]
54 CHAPTER 6. SUMMARY

None of these piecewise polynomial implementations of ω(u) would fix the

problem of integrating I1 (3.18), but if that problem were solved, it might permit
Wu’s values to be more accurately replicated.
For more rigorous ways of showing the difficulty of improving C ∗ , we could
compute some additional points for Wu’s table (5.1), and then interpolate the
result. The interpolation method did remarkably well at demonstrating the neg-
ligible improvements to C ∗ by a finer discretisation. However, with so few points
to interpolate, it is difficult to argue that our interpolant functions for Ψ are ac-
tually representative of the behaviour of Ψ. Ideally, we would want to, either by
analytic or numeric methods, create an upper bound for Ψ that is in a sense “too
good”, and then use that to compute C ∗ . If the resulting value for C ∗ was only a
very small change from Wu’s value, then that provides very strong evidence that
sizeable improvements on C ∗ cannot be made by using Wu’s method.
Appendix A

All the code used for the results in this thesis was written in Mathematica

A.1 Code
Implementation of D(N )

d[n_] := If[Mod[n/2, 2] == 0, deven[n], dodd[n]]

deven[n_] := Module[{t = 0, p = 2},
While[p <= n/2, If[PrimeQ[n - p], t++]; p = NextPrime[p]]; t]
dodd[n_] := Module[{t = 0, p = 2},
While[p <= n/2 - 1, If[PrimeQ[n - p], t++]; p = NextPrime[p]];
If[PrimeQ[n/2], t++]; t]

Taylor expansion of the Buchstab function ω(u).

w2*Exact[u_] := (Log[u - 1] + 1)/u;

degreeT = 10;
precisionT = MachinePrecision;
intervalsT = 5;
wT[1, u_] :=
Evaluate[N[Normal[Series[1/u, {u, 1 + 1/2, degreeT}]], precisionT]];
wT[2, u_] :=
Evaluate[N[
Normal[Series[w2Exact[u], {u, 2 + 1/2, degreeT}]],
precisionT]];
aw2 =
N[Evaluate[
CoefficientList[wT[2, u] /. u -> x + 2 + 1/2, x]],

55
56 APPENDIX A.

precisionT];
Clear[aw];
aw[k_, 2] := aw2[[k + 1]]
aw[0, j_] :=
aw[0, j] = (1/(j + 1/2))*
Sum[(aw[k, j - 1]/2^k)*(j + (-1)^k/(2*(k + 1))),
{k, 0, degreeT}]
aw[k_, j_] :=
aw[k, j] = (1/(j + 1/2))*(aw[k - 1, j - 1]/k - aw[k - 1, j])
wT[j_, u_] :=
ReleaseHold[
Hold[Sum[aw[k, j]*(u - (j + 1/2))^k, {k, 0, degreeT}]]]
wLim = Exp[-EulerGamma];
wA[u_] := Boole[u > intervalsT + 1]*wLim +
Sum[Boole[Inequality[k, LessEqual, u, Less, k + 1]]*wT[k,
u], {k, 1, intervalsT}]

Numerical solution to ω(u).

wN[u_] := Evaluate[w[u] /.
NDSolve[{D[u*w[u], u] == w[u - 1], w[u /; u <= 2] == 1/u}, w,
{u, 1, 30}, WorkingPrecision -> 50]];

Defining unsafeIntegral to blindly apply FTC without checking for validity.

unsafeIntegrate[f_] := f;
unsafeIntegrate[f_, var__,
rest___] :=
(unsafeIntegrate[(Integrate[f, var[[1]]] /.
var[[1]] -> var[[3]]), rest]
- unsafeIntegrate[(Integrate[f, var[[1]]] /. var[[1]] -> var[[2]]),
rest])

Numerically evaluating I1 (3.18)

i1phiN[p_, sp_, s_] :=

Quiet[ReleaseHold[NIntegrate[Boole[t <= u <= v]
*wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{u, sp, s}, {t, sp, s}, {v, sp, s}, MaxRecursion -> 10,
A.1. CODE 57

WorkingPrecision -> MachinePrecision, MaxPoints -> 100000,

AccuracyGoal -> 20]]][[1]]

Numerically evaluating I2i (3.21)

i2[i_, p_, s_, sp_, k1_, k2_, k3_] :=

Piecewise[{{NIntegrate[Boole[t <= u <= v]
*wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{t, 1/k1, 1/k3}, {u, 1/k1, 1/k3}, {v, 1/k1, 1/k3},
MaxRecursion -> 10,
WorkingPrecision -> MachinePrecision, MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 9},
{NIntegrate[Boole[t <= u]*wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{t, 1/k1, 1/k2}, {u, 1/k1, 1/k2}, {v, 1/k2, 1/s},
MaxRecursion -> 10,
WorkingPrecision -> MachinePrecision, MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 10},
{NIntegrate[Boole[u <= v]
*wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{t, 1/k1, 1/k2}, {u, 1/k2, 1/k3}, {v, 1/k2, 1/k3},
MaxRecursion -> 10,
WorkingPrecision -> MachinePrecision, MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 11},
{NIntegrate[Boole[t <= u]*wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{t, 1/sp, 1/k1}, {u, 1/sp, 1/k1}, {v, 1/k3, 1/s},
MaxRecursion -> 10,
WorkingPrecision -> MachinePrecision, MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 12},
{NIntegrate[wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{t, 1/sp, 1/k1}, {u, 1/k1, 1/k2}, {v, 1/k2, 1/s},
MaxRecursion -> 10, WorkingPrecision ->
MachinePrecision, MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 13},
58 APPENDIX A.

{NIntegrate[Boole[u <= v]*wN[(p - t - u - v)/u]*(1/(t*u^2*v)),

{t, 1/sp, 1/k1}, {u, 1/k2, 1/s}, {v, 1/k2, 1/s},
MaxRecursion -> 10, WorkingPrecision ->
MachinePrecision, MaxPoints -> 100000, AccuracyGoal -> 20][[1]],
i == 14},
{NIntegrate[wN[(p - t - u - v)/u]*(1/(t*u^2*v)),
{t, 1/k1, 1/k2}, {u, 1/k2, 1/k3}, {v, 1/k3, 1/s},
MaxRecursion -> 10, WorkingPrecision -> MachinePrecision,
MaxPoints -> 100000, AccuracyGoal -> 20][[1]],
i == 15},
{NIntegrate[Boole[t <= u <= v <= w]
*wN[(p - t - u - v - w)/v]*(1/(t*u*v^2*w)),
{t, 1/k2, 1/k3}, {u, 1/k2, 1/k3}, {v, 1/k2, 1/k3},
{w, 1/k2, 1/k3},
MaxRecursion -> 10, WorkingPrecision -> MachinePrecision,
MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 16},
{NIntegrate[Boole[t <= u <= v]
*wN[(p - t - u - v - w)/v]*(1/(t*u*v^2*w)),
{t, 1/k2, 1/k3}, {u, 1/k2, 1/k3}, {v, 1/k2, 1/k3},
{w, 1/k3, 1/s},
MaxRecursion -> 10, WorkingPrecision -> MachinePrecision,
MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 17},
{NIntegrate[Boole[t <= u && v <= w]
*wN[(p - t - u - v - w)/v]*(1/(t*u*v^2*w)),
{t, 1/k2, 1/k3}, {u, 1/k2, 1/k3},
{v, 1/k3, 1/s}, {w, 1/k3, 1/s},
MaxRecursion -> 10, WorkingPrecision -> MachinePrecision,
MaxPoints -> 100000,
AccuracyGoal -> 20][[1]], i == 18},
{NIntegrate[Boole[u <= v <= w]
*wN[(p - t - u - v - w)/v]*(1/(t*u*v^2*w)),
{t, 1/k1, 1/k2}, {u, 1/k3, 1/s}, {v, 1/k3, 1/s},
{w, 1/k3, 1/s}, MaxRecursion -> 10,
A.1. CODE 59

WorkingPrecision -> MachinePrecision,

MaxPoints -> 100000, AccuracyGoal -> 20][[1]],
i == 19},
{NIntegrate[Boole[u <= v <= w]
*wN[(p - t - u - v - w - x)/w]*
(1/(t*u*v*w^2*x)), {t, 1/k2, 1/k3}, {u, 1/k3, 1/s},
{v, 1/k3, 1/s}, {w, 1/k3, 1/s},
{x, 1/k3, 1/s}, MaxRecursion -> 10,
WorkingPrecision -> MachinePrecision,
MaxPoints -> 100000,
AccuracyGoal -> 20][[1]], i == 20},
{NIntegrate[Boole[t <= u <= v <= w <= x <= y]
*wN[(p - t - u - v - w - x - y)/x]*
(1/(t*u*v*w*x^2*y)), {t, 1/k3, 1/s},
{u, 1/k3, 1/s}, {v, 1/k3, 1/s},
{w, 1/k3, 1/s}, {x, 1/k3, 1/s}, {y, 1/k3, 1/s},
MaxRecursion -> 10,
WorkingPrecision -> MachinePrecision, MaxPoints -> 100000,
AccuracyGoal -> 20][[1]],
i == 21}}]

Setting up Wu’s data

k1L[x_] := {3.53, 3.54, 3.57, 3.56}[[x]];

k2L[x_] := {2.9, 2.88, 2.87, 2.91}[[x]];
k3L[x_] := {2.44, 2.43, 2.4, 2.5}[[x]];
sL[i_] := Piecewise[{{1, i == 0}, {2.1 + 0.1*i, i > 0}}];
spL[i_] := {4.54, 4.5, 4.46, 4.12, 3.58, 3.47, 3.34, 3.19, 3.}[[i]];

Defining Ψ2

psi2comp[i_, s_, s2_, k1_, k2_,

k3_] := (-2/5)*Integrate[Log[t - 1]/t, {t, 2, s2 - 1}] -
(2/5)*Integrate[Log[t - 1]/t, {t, 2, k1 - 1}] -
(1/5)*Integrate[Log[t - 1]/t, {t, 2, k2 - 1}] +
(1/5)*
Integrate[Log[s2*t - 1]/(t*(1 - t)), {t, 1 - 1/s, 1 - 1/s2}] +
60 APPENDIX A.

(1/5)*
Integrate[Log[k1*t - 1]/(t*(1 - t)), {t, 1 - 1/k3, 1 - 1/k1}] +
(-2/5)*Sum[i2comp[i, k], {k, 9, 21}]

where i2comp gives the output of (Listing 5.1)

i2comp[i_, k_] := Map[Last, \[Phi]Mout[[i]]][[k - 8]]

Part of Ψ2

i2Sum[i_, s_, sp_, k1_, k2_, k3_] :=

(-2/5)*Sum[i2[k, \[Phi]Max[k, i], sL[i], spL[i],
k1L[i], k2L[i], k3L[i]], {k, 9, 19}];

Implementation of (Listing 5.1), numerical maximisation of I1 with respect to

φ.

i1testMax[i_, tolmax_] := Module[{tol = 0.1, pMax, pL = pLB[i, k], pH = 5},

While[tol > tolmax,
pMax = MaximalBy[monitorParallelTable[Quiet[{p, i2[k, p, sL[i], spL[i], k1L[i],
k2L[i], k3L[i]]}], {p, pL, pH, tol}, 1], Last][[1]][[1]];
pL = Max[pMax - tol, 2]; pH = pMax + tol; tol = tol/10];
{pMax, i2[k, pMax, sL[i], spL[i], k1L[i], k2L[i], k3L[i]]}]

Implementation of (Listing 5.2) to find the smallest φ such that the integral
I2,i is still defined.

i1Find[i_, k_, {p_, pL_, pH_, tol_}] :=

Piecewise[{{2, Quiet[Check[Quiet[i2[k, 2, sL[i], spL[i], k1L[i], k2L[i],
k3L[i]]],
"Null"]] != "Null"}, {i2Find2[i, k, {p, pL, pH, tol}], True}}]
i2Find[i_, k_, {p_, pL_, pH_, tol_}] :=
Piecewise[{{(pH + pL)/2, pH - pL < tol &&
Quiet[Check[Quiet[i2[k, (pL + pH)/2, sL[i], spL[i], k1L[i], k2L[i], k3L[i]]],
"Null"]] != "Null"}, {i2Find2[i, k, {p, pL, (pH + pL)/2, tol}],
Quiet[Check[Quiet[i2[k, (pL + pH)/2, sL[i], spL[i], k1L[i], k2L[i], k3L[i]]],
"Null"]] != "Null"}, {i2Find2[i, k, {p, (pH + pL)/2, pH, tol}],
Quiet[Check[i2[k, (pH + pL)/2, sL[i], spL[i], k1L[i], k2L[i], k3L[i]],
"Null"]] ==
"Null"}}]
A.1. CODE 61

i2Find[i_, k_, {p_, pL_, pH_, tol_}] :=

Piecewise[{{2, Quiet[Check[Quiet[i2[k, 2, sL[i], spL[i], k1L[i], k2L[i],
k3L[i]]],
"Null"]] != "Null"}, {i2Find2[i, k, {p, pL, pH, tol}], True}}]
i2Find2[i_, k_, {p_, pL_, pH_, tol_}] :=
Piecewise[{{(pH + pL)/2, pH - pL < tol &&
Quiet[Check[Quiet[i2[k, (pL + pH)/2, sL[i], spL[i], k1L[i], k2L[i],
k3L[i]]],
"Null"]] != "Null"}, {i2Find2[i, k, {p, pL, (pH + pL)/2, tol}],
Quiet[Check[Quiet[i2[k, (pL + pH)/2, sL[i], spL[i], k1L[i], k2L[i],
k3L[i]]],
"Null"]] != "Null"}, {i2Find2[i, k, {p, (pH + pL)/2, pH, tol}],
Quiet[Check[i2[k, (pH + pL)/2, sL[i], spL[i], k1L[i], k2L[i], k3L[i]],
"Null"]] ==
"Null"}}]

Computing ai,j (5.5).

al[i_, s_, sp_, k1_, k2_, k3_] := Piecewise[{{k1 - 2, i == 1},
{sp - 2, i == 2},
{sp - sp/s - 1, i == 3}, {sp - sp/k2 - 1, i == 4},
{sp - sp/k3 - 1, i == 5},
{sp - (2*sp)/k2, i == 6}, {sp - sp/k1 - sp/k3, i == 7},
{sp - sp/k1 - sp/k2, i == 8},
{k1 - k1/k2 - 1, i == 9}}];
sig0[t_] := sig[3, t + 2, t + 1]/(1 - [3, 5, 4]);
sig[a_, b_, c_] :=
Evaluate[unsafeIntegrate[(1/t)*Log[c/(t - 1)], {t, a, b}]];
Xi1[t_, s_, sp_] := (0[t]/(2*t))*Log[16/((s - 1)*(sp - 1))] +
(Boole[sp - 2 <= t <= 3]/(2*t))*Log[(t + 1)^2/((s - 1)*(sp - 1))] +
(Boole[sp - 2 <= t <= sp - sp/s - 1]/(2*t))*Log[(t + 1)/((s - 1)*
(sp - 1 - t))];
a1Int[a_, b_, s_, sp_] := NIntegrate[Xi1[t, s, sp], {t, a, b}];
a1[i_, j_] := a1Int[sL[j - 1], sL[j], sL[i], spL[i]];

and Ξ2 (3.23)
62 APPENDIX A.

Xi2[t_, s_, sp_, k1_, k2_, k3_] :=

(0[t]/(5*t))*Log[1024/((s - 1)*(sp - 1)*(k1 - 1)*(k2 - 1)
*(k3 - 1))] +
(Boole[sp - 2 <= t <= 3]/(5*t))*Log[(t + 1)^5/((s - 1)*
(sp - 1)*(k1 - 1)*(k2 - 1)*(k3 - 1))] +
(Boole[k1 - k1/k2 - 1 <= t <= k1 - 2]/(5*t))*
Log[(t + 1)/((k2 - 1)*(k1 - 1 - t))] +
(Boole[sp - sp/k3 - 1 <= t <= sp - 2]/(5*t))*
Log[(t + 1)/((k3 - 1)*(sp - 1 - t))] +
(Boole[sp - sp/s - 1 <= t <= sp - 2]/(5*t))*
Log[(t + 1)/((s - 1)*(sp - 1 - t))] +
(Boole[k1 - 2 <= t <= sp - 2]/(5*t))*
Log[(t + 1)^2/((k1 - 1)*(k2 - 1))] +
(Boole[sp - sp/k1 - sp/k3 <= t <= sp - sp/k3 - 1]/
(5*t*(1 - t/sp)))*Log[sp^2/((k1*sp - sp - k1*t)*(k3*sp - sp - k3*t))]
+ (Boole[sp - sp/k3 - 1 <= t <= sp - sp/k1 - sp/k2]/(5*t*(1 - t/sp)))*
Log[(sp*(sp - 1 - t))/(k1*sp - sp - k1*t)] +
(Boole[sp - (2*sp)/k2 <= t <= sp - sp/k1 - sp/k2]/(5*t*(1 - t/sp)))*
Log[sp/(k2*sp - sp - k2*t)] + (Boole[sp - sp/k1 - sp/k2 <= t <= sp - 2]/
(5*t*(1 - t/sp)))*Log[sp - 1 - t];

a2Int[a_, b_, s_, sp_, k1_, k2_, k3_]

:= NIntegrate[Xi2[t, s, sp, k1, k2, k3], {t, a, b}];

a2[i_, j_] := a2Int[sL[j - 1], sL[j], sL[i], spL[i], k1L[i], k2L[i], k3L[i]];

Solving the matrix equations (3.32)

bChen = {0.015826357,
0.015247971,
0.013898757,
0.011776059,
0.009405211,
0.006558950,
0.003536751,
0.001056651,
0};
A.1. CODE 63

aMat = Table[a[i, j], {i, 1, 9}, {j, 1, 9}];

MatrixForm[LinearSolve[IdentityMatrix[9] - aMat, bVec]]

Code to interpolate Wu’s data and integrate over d many intervals.

psi1data = Table[{sL[i], 1[[i - 4]]}, {i, 5, 9}];

psi2data = Table[{sL[i], 2[[i]]}, {i, 1, 4}];

int1 = Interpolation[psi1data];
int2 = Interpolation[psi2data];

root = s /. FindRoot[int1[s] - int2[s], {s, 2.55}];

interpsi[s_] :=
Piecewise[{{int2[2.2], 2 <= s <= 2.2}, {int2[s], 2.2 <= s <= root},
{int1[s], root <= s <= 3}}];

spdata = Table[{sL[i], spL[i]}, {i, 1, 9}]

k1data = Table[{sL[i], k1L[i]}, {i, 1, 4}];
k2data = Table[{sL[i], k2L[i]}, {i, 1, 4}];
k3data = Table[{sL[i], k3L[i]}, {i, 1, 4}];

intk1 = Interpolation[k1data];
intk2 = Interpolation[k2data];
intk3 = Interpolation[k3data];
intSp = Interpolation[spdata]

a1Int[a_, b_, s_, sp_] := NIntegrate[Xi1[t, s, sp], {t, a, b}]

sint[i_, d_] := sint[i, d] = Piecewise[{{1, i == 0},

{Range[22/10, 3, (3 - 22/10)/(d - 1)][[i]], i > 0}}];

a1[i_, j_, d_]

:= a1Int[sint[j - 1, d], sint[j, d], sint[i, d], intSp[sint[i, d]]];

a2Int[a_, b_, s_, sp_, k1_, k2_, k3_]

64 APPENDIX A.

:= NIntegrate[Xi2[t, s, sp, k1, k2, k3], {t, a, b}];

a2[i_, j_, d_]

:= a2Int[sint[j - 1, d], sint[j, d], sint[i, d], intSp[sint[i, d]],
intk1[sint[i, d]], intk2[sint[i, d]], intk3[sint[i, d]]];

a[i_, j_, d_]

:= Piecewise[{{a1[i, j, d], sint[i, d] > root},
{a2[i, j, d], sint[i, d] < root}}];

aMat[d_] := ParallelTable[a[i, j, d], {i, 1, d}, {j, 1, d}];

bsec[i_, d_] := interpsi[sint[i, d]];
bVec[d_] := Table[interpsi[sint[i, d]], {i, 1, d}];
xVec[d_] := LinearSolve[IdentityMatrix[d] - aMat[d], bVec[d]];
chen[d_] := 8*(1 - xVec[d][[1]])

Quiet[Function[x, AbsoluteTiming[{x, chen[x]}]] /@ {200, 300, 400, 500, 600}]

Bibliography

[1] T. Apostol. Introduction to Analytic Number Theory. Undergraduate Texts

in Mathematics. Springer, New York, 1998.

[2] J. Baker. Excel and the Goldbach comet. Spreadsheets in Education (eJSiE),
2(2):2, 2007.

[3] E. Bombieri and H. Davenport. Small differences between prime numbers. In

Proceedings of the Royal Society of London A: Mathematical, Physical and
Engineering Sciences, volume 293, pages 1–18. The Royal Society, 1966.

[4] A. Buchstab. Asymptotic estimate of one general number-theoretic function

(Russian). Mat. Sat., 2 (44):1239–1246, 1937.

[5] R. Burden, J. D. Faires, and A. Reynolds. Numerical analysis (1985).

Prindle, Weber & Schmidt, 1985.

[6] A. Y. Cheer and D. A. Goldston. A differential delay equation arising from

the sieve of Eratosthenes. Mathematics of computation, 55(191):129–141,
1990.

[7] J.-R. CHEN. On the Goldbach’s problem and the sieve methods. Scientia
Sinica, 21(6):701–739, 1978.

[8] Y.-G. Chen and X.-G. Sun. On romanoff’s constant. Journal of Number
Theory, 106(2):275–284, 2004.

[9] N. G. de Bruijn. On the number of uncancelled elements in the sieve of

eratosthenes. Indag. math., 12:247–256, 1950.

[10] J. De Koninck and F. Luca. Analytic Number Theory: Exploring the

Anatomy of Integers. Graduate studies in mathematics. American Math-
ematical Society, 2012.

65
66 BIBLIOGRAPHY

[11] J.-M. Deshouillers, G. Effinger, H. Te Riele, and D. Zinoviev. A complete

Vinogradov 3-primes theorem under the Riemann hypothesis. Electronic
Research Announcements of the American Mathematical Society, 3(15):99–
104, 1997.

[12] T. O. e Silva. Goldbach conjecture verification, 2017. [Online; accessed

31-January-2017].

[13] P. Erdös. On integers of the form 2k + p and some related problems. Summa
Brasil. Math., pages 113–125, 1950.

[14] H. F. Fliegel and D. S. Robertson. Goldbach’s comet: the numbers related

to Goldbach’s conjecture. Journal of Recreational Mathematics, 21(1):1–7,
1989.

[15] R. Goldberg. Methods of real analysis. Blaisdell book in the pure and applied
sciences. Blaisdell Pub. Co., 1964.

[16] S. W. Golomb. The invincible primes. The Sciences, 25(2):50–57, 1985.

[17] R. Guy. Unsolved problems in number theory. Problem books in mathemat-

ics. Springer-Verlag, 1994.

[18] L. Habsieger and X.-F. Roblot. On integers of the form p + 2k . Acta Arith-
metica, 122(1):45–50, 2006.

[19] H. Halberstam and H. Richert. Sieve methods. L.M.S. monographs. Aca-

demic Press, 1974.

[20] G. H. Hardy and J. E. Littlewood. Some problems of partitio numerorum;

iii: On the expression of a number as a sum of primes. Acta Mathematica,
44(1):1–70, 1923.

[21] D. Heath-Brown. Lectures on sieves. arXiv preprint math/0209360, 2002.

[22] D. R. Heath-Brown and J.-C. Puchta. Integers represented as a sum of

primes and powers of two. arXiv preprint math/0201299, 2002.

[23] H. A. Helfgott. The ternary Goldbach conjecture is true. arXiv preprint

arXiv:1312.7748, 2013.

[24] L.-K. Hua. Estimation of an integral. Sci. Sinica, 4:393–402, 1951.

BIBLIOGRAPHY 67

[25] H. Iwaniec. A new form of the error term in the linear sieve. Acta Arith-
metica, 37(1):307–320, 1980.

[26] C. JING-RUN. On the representation of a larger even integer as the sum of

a prime and the product of at most two primes. SCIENCE CHINA Mathe-
matics, 16(2):157, 1973.

[27] H. Li. The number of powers of 2 in a representation of large even integers

by sums of such powers and of two primes (ii). Acta Arithmetica, 96:369–379,
2001.

[28] H. Z. Li. The number of powers of 2 in a representation of large even

integers by sums of such powers and of two primes. ACTA ARITHMETICA-
WARSZAWA-, 92(3):229–237, 2000.

[29] Y. V. Linnik. Prime numbers and powers of two. Trudy Matematicheskogo

Instituta imeni VA Steklova, 38:152–169, 1951.

[30] Y. V. Linnik. Addition of prime numbers with powers of one and the same
number. Matematicheskii Sbornik, 74(1):3–60, 1953.

[31] J. Liu, M. Liu, and T. Wang. The number of powers of 2 in a representation of

large even integers ii. Science in China Series A: Mathematics, 41(12):1255–
1271, 1998.

[32] J. Liu, M.-C. Liu, and T. Wang. On the almost Goldbach problem of linnik.
Journal de théorie des nombres de Bordeaux, 11(1):133–147, 1999.

[33] Z. Liu and G. LÜ. Density of two squares of primes and powers of 2. Inter-
national Journal of Number Theory, 7(05):1317–1329, 2011.

[34] T. Z. W. M. C. Liu. On the Vinogradov bound in the three primes Goldbach

conjecture. Acta Arithmetica, 105(2):133–175, 2002.

[35] G. Marsaglia, A. Zaman, and J. C. Marsaglia. Numerical solution of

some classical differential-difference equations. Mathematics of Computa-
tion, 53(187):191–201, 1989.

[36] M. E. O’NEILL. The genuine sieve of eratosthenes. Journal of Functional

Programming, 19(01):95–106, 2009.
68 BIBLIOGRAPHY

[37] C.-D. Pan. A new application of the yu. v. linnik large sieve method. Chinese
Math. Acta, 5:642–652, 1964.

[38] J. Pintz. A note on romanov’s constant. Acta Mathematica Hungarica,

112(1-2):1–14, 2006.

[39] J. Pintz and A. Khalfalah. On the representation of Goldbach numbers by

a bounded number of powers of two. Proceedings Elementary and Analytic
Number Theory, 2004.

[40] J. Pintz and I. Ruzsa. On linnik’s approximation to Goldbach’s problem, i.

Acta Arithmetica, 109(2):169–194, 2003.

[41] D. Platt and T. Trudgian. Linnik’s approximation to Goldbach’s conjecture,

and other problems. Journal of Number Theory, 153:54–62, 2015.

[42] F. Romani. Computations concerning primes and powers of two. Calcolo,

20(3):319–336, 1983.

[43] N. Romanoff. Über einige sätze der additiven zahlentheorie. Mathematische

Annalen, 109(1):668–678, 1934.

[44] P. Ross. On chen’s theorem that each large even number has the form p1+
p2 or p1+ p2 p3. Journal of the London Mathematical Society, 2(4):500–506,
1975.

[45] A. Selberg. On elementary methods in primenumber-theory and their lim-

itations. Den 11te Skandinaviske Matematikerkongress (Trondheim, 1949),
Johan Grundt Tanums Forlag, Oslo, pages 13–22, 1952.

[46] E. M. Stein and R. Shakarchi. Real analysis: measure theory, integration,

and Hilbert spaces. Princeton University Press, 2009.

[47] J. Stewart. Calculus. Cengage Learning, 2015.

[48] T. Trudgian and J. Wu. Personal correspondance. 2013.

[49] J. G. van der Corput. On de polignac’s conjecture. Simon Stevin, 27:99–105,

1950.

[50] I. M. Vinogradov. Representation of an odd number as a sum of three primes.

Dokl. Akad. Nauk. SSR 15, page 291294, 1937.
BIBLIOGRAPHY 69

[51] T. Wang. On linniks almost Goldbach theorem. Science in China Series A:

Mathematics, 42(11):1155–1172, 1999.

[52] J. W. Wrench Jr. Evaluation of artin’s constant and the twin-prime constant.
Mathematics of Computation, pages 396–398, 1961.

[53] D. H. Wu. An improvement of j. r. chen’s theorem. Shanghai Keji Daxue

Xuebao, pages 94–99, 1997.

[54] J. Wu. Chen’s double sieve, Goldbach’s conjecture and the twin prime prob-
lem. Acta Arithmetica, 114:215–273, 2004.

[55] T. Yamada. Explicit chen’s theorem. arXiv preprint arXiv:1511.03409, 2015.