Professional Documents
Culture Documents
Problema de Golbach - Matematicas
Problema de Golbach - Matematicas
Problema de Golbach - Matematicas
H. A. HELFGOTT
Abstract. The ternary Goldbach conjecture, or three-primes problem, asserts that every odd integer n greater than 5 is the sum of three primes. The
present paper proves this conjecture.
Both the ternary Goldbach conjecture and the binary, or strong, Goldbach
conjecture had their origin in an exchange of letters between Euler and Goldbach in 1742. We will follow an approach based on the circle method, the
large sieve and exponential sums, supplemented by rigorous computations, including a verification of zeros of L-functions due to D. Platt. The improved
estimates on exponential sums are proven in a twin paper by the author.
Contents
1. Introduction
1.1. Results
1.2. History
1.3. Main ideas
1.4. Acknowledgments
2. Preliminaries
2.1. Notation
2.2. Dirichlet characters and L functions
2.3. Fourier transforms
2.4. Mellin transforms
3. Preparatory work on major arcs
3.1. Decomposition of S (, x) by characters
3.2. The integral over the major arcs
4. Optimizing and coordinating smoothing functions
4.1. The symmetric smoothing function
4.2. The approximation + to
4.3. The smoothing function
5. Mellin transforms and smoothing functions
5.1. Exponential sums and L functions
5.2. How to choose a smoothing function?
5.3. The Mellin transform of the twisted Gaussian
5.4. Totals
6. Explicit formulas
6.1. A general explicit formula
2
6.2. Sums and decay for (t) = t2 et /2 and (t)
6.3. Sums and decay for + (t)
6.4. A verification of zeros and its consequences
7. The integral of the triple product over the minor arcs
7.1. The L2 norm over arcs: variations on the large sieve for primes
7.2. Bounding the quotient in the large sieve for primes
7.3. Putting together 2 bounds over arcs and bounds
1
2
2
2
4
7
8
8
8
9
9
10
11
12
22
23
24
30
36
36
37
38
53
55
56
61
66
78
81
81
86
92
H. A. HELFGOTT
8. Conclusion
8.1. The 2 norm over the major arcs
8.2. The total major-arcs contribution
8.3. Minor-arc totals
8.4. Conclusion: proof of main theorem
Appendix A. Sums over primes
Appendix B. Sums involving (q)
Appendix C. Validated numerics
C.1. Integrals of a smoothing function
C.2. Extrema via bisection and truncated series
References
102
102
103
108
114
115
117
120
120
122
130
1. Introduction
1.1. Results. The ternary Goldbach conjecture (or three-prime problem) states
that every odd number n greater than 5 can be written as the sum of three
primes. Both the ternary Goldbach conjecture and the (stronger) binary Goldbach conjecture (stating that every even number greater than 2 can be written
as the sum of two primes) have their origin in the correspondence between Euler
and Goldbach (1742).
I. M. Vinogradov [Vin37] showed in 1937 that the ternary Goldbach conjecture
is true for all n above a large constant C. Unfortunately, while the value of C has
been improved several times since then, it has always remained much too large
(C = e3100 , [LW02]) for a mechanical verification up to C to be even remotely
feasible.
The present paper proves the ternary Goldbach conjecture.
Main Theorem. Every odd integer n greater than 5 can be expressed as the sum
of three primes.
The proof given here works for all n C = 1030 . The main theorem has
been checked deterministically by computer for all n < 1030 (and indeed for all
n 8.875 1030 ) [HP].
We are able to set major arcs to be few and narrow because the minor-arc
estimates in [Hel] are very strong; we are forced to take them to be few and
narrow because of the kind of L-function bounds we will rely upon.
At issue are
(1) a fuller use of the close relation between the circle method and the large
sieve;
(2) a combination of different smoothings for different tasks;
(3) the verification of GRH up to a bounded height for all conductors q
150000 and all even conductors q 300000 (due to David Platt [Plab]);
(4) better bounds for exponential sums, as in [Hel].
All major computations including D. Platts work in [Plab] have been
conducted rigorously, using interval arithmetic.
1.2. History. The following brief remarks are here to provide some background;
no claim to completeness is made. Results on exponential sums over the primes
are discussed more specifically in [Hel, 1].
1.2.1. Results towards the ternary Goldbach conjecture. Hardy and Littlewood
[HL23] proved that every odd number larger than a constant C is the sum of three
primes, conditionally on the generalized Riemann Hypothesis. This showed, as
they said, that the problem was not unergreifbar (as it had been called by Landau
in [Lan12]).
Vinogradov [Vin37] made the result unconditional. An explicit value for C
15
(namely, C = 33 ) was first found by Borodzin in 1939. This value was improved
to C = 3.33 1043000 by J.-R. Chen and T. Z. Wang [CW89] and to C = 2 101346
by M.-Ch. Liu and T. Wang [LW02]. (J.-R. Chen had also proven that every
large enough even number is either the sum of two primes or the sum p1 + p2 p3
of a prime p1 and the product p2 p3 of two primes.)
In [DEtRZ97], the ternary Goldbach conjecture was proven for all n conditionally on the generalized Riemann hypothesis.
1.2.2. Checking Goldbach for small n. Numerical verifications of the binary Goldbach conjecture for small n were published already in the late nineteenth century; see [Dic66, Ch. XVIII]. Richstein [Ric01] showed that every even integer
4 n 4 1014 is the sum of two primes. Oliveira e Silva, Herzog and Pardi
[OeSHP13] have proven that every even integer 4 n 4 1018 is the sum of two
primes.
The question is then until what point one can establish the ternary Goldbach
conjecture using [OeSHP13]. Clearly, if one can show that every interval of length
4 1018 within [1, N ] contains a prime, then [OeSHP13] implies that every odd
number between 7 and N can be written as the sum of three primes. This was
used in a first version of [Hel] to show that the best existing result on prime
gaps ([RS03], with [Plaa] as input) implies that every odd number between 7 and
1.23 1027 is the sum of three primes. A more explicit approach to prime gaps
[HP] now shows that every odd integer 7 n 8.875694 1030 is the sum of three
primes.
1.2.3. Work on Schnirelmans constant. Schnirelmans constant is a term for
the smallest k such that every integer n > 1 is the sum of at most k primes.
(Thus, Goldbachs binary and ternary conjecture, taken together, are equivalent
to the statement that Schnirelmans constant is 3.) In 1930, Schnirelman [Sch33]
showed that Schnirelmans constant k is finite, developing in the process some of
the bases of what is now called additive or arithmetic combinatorics.
In 1969, Klimov proved that k 6 109 ; he later improved this result to
k 115 [KPS72]
(with G. Z. Piltay and T. A. Sheptiskaya) and k 55. Results
by Vaughan [Vau77] (k = 27), Deshouillers [Des77] (k = 26) and Riesel-Vaughan
[RV83] (k = 19) then followed.
Ramare showed in 1995 that every even n > 1 is the sum of at most 6 primes
[Ram95]. Recently, Tao [Tao] established that every odd number n > 1 is the
sum of at most 5 primes. These results imply that k 6 and k 5, respectively.
The present paper implies that k 4.
1.2.4. Other approaches. Since [HL23] and [Vin37], the main line of attack on
the problem has gone through exponential sums. There are proofs based on
cancellation in other kinds of sums ([HB85], [IK04, 19]), but they have not
been made to yield practical estimates. The same goes for proofs based on other
principles, such as that of Schnirelmans result or the recent work of X. Shao
H. A. HELFGOTT
[Sha]. (It deserves to be underlined that [Sha] establishes Vinogradovs threeprime result without using L-function estimates at all; the constant C is, however,
extremely large.)
1.3. Main ideas. We will limit the discussion here to the general setup, the
estimates for major arcs and the efficient usage of exponential-sum estimates on
the minor arcs. The development of new exponential-sum estimates is the subject
of [Hel].
In the circle method, the number of representations of a number N as the
sum of three primes is represented as an integral over the circle R/Z, which is
partitioned into major arcs M and minor arcs m = (R/Z) \ M:
Z
X
(S(, x))3 e(N )d
(n1 )(n2 )(n3 ) =
(1.1)
n1 +n2 +n3 =N
R/Z
Px
where S(, x) =
n=1 (n)e(n). The aim is to show that the sum of the
integral over M and the integral over m is positive; this will prove the threeprimes theorem.
The major arcs M = Mr0 consist of intervals (a/q cr0 /qx, a/q + cr0 /qx)
around the rationals a/q, q r0 , where c is a constant. In previous work1, r0
grew with x; in our setup, r0 is a constant. Smoothing changes the left side of
(1.1) into a weighted sum, but, since we aim at an existence result rather than
at an asymptotic for the number of representations p1 + p2 + p3 of N , this is
obviously acceptable.
Typically,
work on major arcs yields rather precise estimates on the integral
R
over M in (1.1), whereas work
on minor arcs gives upper bounds on the absolute
R
value of the integral over m in (1.1). Exponential-sum estimates, such as those
in [Hel], provide upper bounds for maxm |S(, x)|.
1.3.1. Major arc bounds. We will be working with smoothed sums
(1.2)
S (, x) =
(n)(n)e(n/x)(n/x).
n=1
(1.4)
S (/x, x) =
(n)(n)e(n/x)(n/x)
n=1
for varying among all Dirichlet characters modulo q r0 and for || cr0 /q,
i.e., || small. Using estimates on (1.4) efficiently in the estimation of (1.3) is
a delicate task; this is the subject of 3. Let us now focus on how to obtain
estimates on (1.4).
1Ramar
es work [Ram10] is in principle strong enough to allow r0 to be an unspecified large
constant. Taos work [Tao] reaches this standard only for x of moderate size.
Sums such as (1.4) are estimated using Dirichlet L-functions L(s, ). An explicit formula gives an expression
(1.5)
where Iq=1 = 1 if q = 1 and Iq=1 = 0 otherwise. Here runs over the complex
numbers with L(, ) = 0 and 0 < () < 1 (non-trivial zeros). The function
F is the Mellin transform of e(t)(t) (see 2.4).
The questions are then: where are the non-trivial zeros of L(s, )? How fast
does F () decay as () ?
Write = (s), = (s). The belief is, of course, that = 1/2 for every
non-trivial zero (Generalized Riemann Hypothesis), but this is far from proven.
Most work to date has used zero-free regions of the form 1 1/C log q| |, C
a constant. This is a classical zero-free region, going back, qualitatively, to de la
Vallee-Poussin (1899). The best values of C known are due to McCurley [McC84]
and Kadiri [Kad05].
These regions seem too narrow to yield a proof of the three-primes theorem.
What we will use instead is a finite verification of GRH up to Tq , i.e., a computation showing that, for every Dirichlet character of conductor q r0 (r0 a
constant, as above), every non-trivial zero = + i with | | Tq satisfies
() = 1/2. Such verifications go back to Riemann; modern computer-based
methods are descended in part from a paper by Turing [Tur53]. (See the historical article [Boo06].) In his thesis [Pla11], D. Platt gave a rigorous verification for
r0 = 105 , Tq = 108 /q. In coordination with the present work, he has extended
this to
all odd q 3 105 , with Tq = 108 /q,
all even q 4 105 , with Tq = max(108 /q, 200 + 7.5 107 /q).
This was a major computational effort, involving, in particular, a fast implementation of interval arithmetic (used for the sake of rigor).
What remains to discuss, then, is how to choose in such a way F () decreases
fast enough as | | increases, so that (1.5) gives a good estimate. We cannot hope
for F () to start decreasing consistently before | | is at least as large as a multiple
of 2||. Since varies within (cr0 /q, cr0 /q), this explains why Tq is taken
inversely proportional to q in the above. As we will work with r0 150000, we
also see that we have little margin for maneuver: we want F () to be extremely
small already for, say, | | 80||. We also have a Scylla-and-Charybdis situation,
courtesy of the uncertainty principle: roughly speaking, F () cannot decrease
faster than exponentially on | |/|| both for || 1 and for large.
The most delicate case is that of large, since then | |/|| is small. It turns
out we can manage to get decay that is much faster than exponential for large,
while no slower than exponential for small. This we will achieve by working
2
with smoothing functions based on the (one-sided) Gaussian (t) = et /2 .
2
The Mellin transform of the twisted Gaussian e(t)et /2 is a parabolic cylinder
function U (a, z) with z purely imaginary. Since fully explicit estimates for U (a, z),
z imaginary, have not been worked in the literature, we will have to derive them
ourselves. This is the subject of 5.
H. A. HELFGOTT
We still have some freedom in choosing + and , though they will have to be
2
based on (t) = et /2 . The main term in our estimate for (1.3) is of the form
Z Z
N
(t1 + t2 ) dt1 dt2 ,
+ (t1 )+ (t2 )
(1.6)
C0
x
0
0
where C0 is a constant. Our upper bound for the minor-arc integral, on the other
hand, will be proportional to |+ |22 | |1 . The question is then how to make (1.6)
divided by |+ |22 | |1 as large as possible. A little thought will show that it is
best for + to be symmetric, or nearly symmetric, around t = 1 (say), and for
be concentrated on a much shorter interval than + , while x is set to be x/2 or
slightly less.
It is easy to construct a function of the form t 7 h(t) (t) symmetric around
t = 1, with support on [0, 2]. We will define + = hH (t) (t), where hH is an
approximation to h that is band-limited in the Mellin sense. This will mean that
the decay properties of the Mellin transform of e(t)+ (t) will be like those of
e(t) (t), i.e., very good.
How to choose ? The bounds in [Hel] were derived for 2 = (2I[1/2,1] ) M
(2I[1/2,1] ), which is nice to deal with in the context of combinatorially flavored
analytic number theory, but it has a Mellin transform that decays much too
slowly.2 The solution is to use a smoothing that is, so to speak, Janus-faced, viz.,
2
= (2 M )(t), where (t) = t2 et /2 and is a large constant. We estimate
sums of type S (, x) by estimating S2 (, x) if S2 (, x) if lies on a minor arc,
or by estimating S (, x) if lies on a major arc. (The Mellin transform of is
just a shift of that of .) This is possible because 2 has support bounded away
from zero, while is also concentrated away from 0.
1.3.2. Minor arc bounds: exponential sums and the large sieve. Let mr be the
complement of Mr . In particular, m = mr0 is the complement of M = Mr0 . Exponential sum-estimates, such as those in [Hel], give bounds on maxmr |S(, x)|
that decrease with r.
We need to do better than
Z
Z
S(, x)3 e(N ) d (max |S(, x)| )
|S(, x)|2 d
m
m
m
Z
(1.7)
2
2
|S(, x)| d ,
(max |S(, x)| ) |S|2
m
as this inequality involves a loss of a factor of log x (because |S|22 x log x).
Fortunately, minor arc estimates are valid not just for a fixed r0 , but for the
complement of Mr , where r can vary within a broad range. By partial summation,
these estimates can be combined with upper bounds for
Z
Z
2
|S(, x)|2 d.
|S(, x)| d
Mr
M r0
Giving an estimate for the integral over Mr0 (r0 a constant) will be part of our
task over the major arcs. The question is how to give an upper bound for the
integral over Mr that is valid and non-trivial over a broad range of r.
2This parallels the situation in the transition from Hardy and Littlewood [HL23] to Vinogradov [Vin37]. Hardy and Littlewood used the smoothing (t) = et , whereas Vinogradov
used the brusque (non-)smoothing (t) = I[0,1] . Arguably, this is not just a case of technological
decay; I[0,1] has compact support and is otherwise easy to deal with in the minor-arc regime.
The answer lies in the deep relation between the circle method and the large
sieve. (This was obviously not available to Vinogradov in 1937; the large sieve
is a slightly later development (Linnik [Lin41], 1941) that was optimized and
fully understood later still.) A large sieve is, in essence, an inequality giving a
discretized version of Plancherels identity. Large sieves for primes show that the
inequality can be sharpened for sequences of prime support, provided that, on
the Fourier side, the sum over frequencies is shortened. The idea here is that
this kind of improvement can be adapted back to the continuous context, so as
to give upper bounds on the L2 norms of exponential sums with prime support
when is restricted
to special subsets of the circle. Such an L2 norm is nothing
R
other than Mr |S(, x)|2 d.
The first version of [Hel] used an idea of Heath-Browns3 that can indeed be
understood in this framework. In 7.1, we shall prove a better bound, based on a
large sieve for primes due to Ramare [Ram09]. We will re-derive this sieve using
an idea of Selbergs. We will then make it fully explicit in the crucial range (7.2).
(This, incidentally, also gives fully explicit estimates for Ramares large sieve in
its original discrete context, making it the best large sieve for primes in a wide
range.)
R
The outcome is that Mr |S(, x)|2 d is bounded roughly by 2x log r, rather
than by x log x (or by 2e x log r, as was the case when Heath-Browns idea was
used). The lack of a factor of log x makes it possible to work with r0 equal to a
constant, as we have done; the factor of e reduces the need for computations by
more than an order of magnitude.
1.4. Acknowledgments. The author is very thankful to D. Platt, who, working in close coordination with him, provided GRH verifications in the necessary
ranges, and also helped him with the usage of interval arithmetic. He is also
much indebted to O. Ramare for his help and feedback, especially regarding 7
and Appendix B. Special thanks are also due to A. Booker, B. Green, R. HeathBrown, H. Kadiri, T. Tao and M. Watkins for discussions on Goldbachs problem
and related issues.
Warm thanks are also due to A. C
ordoba and J. Cilleruelo, for discussions on
the method of stationary phase, and to V. Blomer and N. Temme, for further help
with parabolic cylinder functions. Additional references were graciously provided
by R. Bryant, S. Huntsman and I. Rezvyakova.
Travel and other expenses were funded in part by the Adams Prize and the
Philip Leverhulme Prize. The authors work on the problem started at the Universite de Montreal (CRM) in 2006; he is grateful to both the Universite de Montreal
H. A. HELFGOTT
if x > 0,
1
sgn(x) = 0
if x = 0,
1 if x < 0.
We write O (R) to mean a quantity at most R in absolute value.
character , all non-trivial zeros of L(s, ) lie on the critical line. Verifiable finite
versions of the generalized Riemann hypothesis generally read: for every Dirichlet
character of modulus q Q, all non-trivial zeros of L(s, ) with |(s)| f (q)
lie on the critical line (where f : Z R+ is some given function).
2.3. Fourier transforms. The Fourier transform on R is normalized as follows:
Z
b
e(xt)f (x)dx
f (t) =
for f : R C.
The trivial bound is |fb| |f |1 . Integration by parts gives that, if f is
differentiable k times outside finitely many points, then
!
!
d
(k) |
(k) |
|
f
|f
1
(2.1)
fb(t) = O
= O
.
2t
(2t)k
It could happen that |f (k) |1 = , in which case (2.1) is trivial (but not false).
In practice, we require f (k) L1 . In a typical situation, f is differentiable k
times except at x1 , x2 , . . . , xk , where it is differentiable only (k 2) times; the
contribution of xi (say) to |f (k) |1 is then | limxx+ f (k1) (x)limxx f (k1) (x)|.
i
In general, M (f M g) = M f M g and
Z +i
1
M f (z)M g(s z)dz
(2.3)
M (f g)(s) =
2i i
[GR00, 17.32]
provided that z and sz are within the strips on which M f and M g (respectively)
are well-defined.
The Mellin transform is an isometry, in the sense that
Z
Z
1
2 2 dt
|f (t)| t
(2.4)
=
|M f ( + it)|2 dt.
t
2
0
bs a s
,
s
we see that
(2.6)
M 2 (s) =
1 2s
s
2
M 4 (s) =
1 2s
s
4
10
H. A. HELFGOTT
where the next-to-last step holds by contour integration, and the last step holds
by the definition of the Gamma function (s).
3. Preparatory work on major arcs
Let
(3.1)
S (, x) =
j
(n)e(n)(n/x),
In the circle method, the set R/Z gets partitioned into the set of major arcs
M and the set of minor arcs m; the contribution of each of the two sets to the
integral (3.3) is evaluated separately.
Our object here is to treat the major arcs: we wish to estimate
Z
S1 (, x)S2 (, x)S3 (, x)e(N )d
(3.4)
M
for M = M0 ,r , where
(3.5)
[ [ a
[
0 r a
0 r
M0 ,r =
, +
q
2qx q
2qx
qr a mod q
q odd (a,q)=1
[ a 0 r a 0 r
, +
q
qx q
qx
q2r a mod q
q even (a,q)=1
11
mod q a mod q
(a,q)=1
X e(ab/q)
X e(ab/q) X
(a1 a0 ) =
(q)
(q)
mod q
a mod q
(a,q)=1
a mod q
(a,q)=1
(a1 a0 ),
mod q
P
where q = q/ gcd(q, a
). Now, mod q (a1 a0 ) = 0 unless a = a0 (in which
0
P
case mod q (a1 a0 ) = (q )). Thus, (3.7) equals
(q )
(q)
(q )
e(ab/q) =
(q)
a mod q
(a,q)=1
aa0 mod q
a
(q )
(q)
0b
k
(k,q/q )=1
k mod q/q
(k,q/q )=1
mod q/q
kb
q/q
(a0 + kq )b
q
(q )
e
(q)
a0 b
q
(q/q )
provided that (b, q) = 1. (We are evaluating a Ramanujan sum in the last step.)
Hence, for = a/q + /x, q x, (a, q) = 1,
X
1 X
(, a)
(n)(n)e(n/x)(n/x)
(q)
n
equals
X ((q, n ))
n
((q, n ))
(n)e(n)(n/x).
12
H. A. HELFGOTT
(n)
.
x
x
n
1
p|q
(n,q)6=1
X
X
X p
1
,
S (, x) =
(a) ()S,
, x + O 2
log p
(q)
x
x
mod q
where
(3.9)
p|q
S, (, x) =
(n)(n)e(n)(n/x).
3 Y
X
j=1 j 6=j
|Sj (, x)|
X
p|q
log p
p
x
We will later see that the integral of (3.11) over S 1 is negligible for our choices
of j , it will, in fact, be of size O(x(log x)A ), A a constant. (In (3.10), we have
reduced our problems to estimating S, (/x, x) for primitive; a more obvious
way of reaching the same goal would have multiplied made (3.11) worse by a
factor of about q. The error term O(x(log x)A ) should be compared to the
main term, which will be of size about a constant times x2 .)
3.2. The integral over the major arcs. We are to estimate the integral (3.4),
where the major arcs M0 ,r are defined as in (3.5). We will use 1 = 2 = + ,
3 (t) = (t), where + and will be set later.
We can write
Z
(t/x)e(t/x)dt + O (err, (, x)) x
S, (/x, x) = S (/x, x) =
(3.12)
0
= b() x + O (err,T (, x)) x
for = T the trivial character, and
(3.13)
for primitive and non-trivial. The estimation of the error terms err will come
later; let us focus on (a) obtaining the contribution of the main term, (b) using
estimates on the error terms efficiently.
The main term: three principal characters. The main contribution will be
given by the term in (3.10) with 1 = 2 = 3 = 0 , where 0 is the principal
character mod q.
13
(0 , n) =
(q/d)d.
d|(q,n)
This simplifies to (q/(q, n))((q, n)) for q square-free. The special case n = 1
gives us that (0 ) = (q).
Thus, the term in (3.10) with 1 = 2 = 3 = 0 equals
e(N a/q)
(q)3 S+ ,0 (/x, x)2 S ,0 (/x, x)e(N/x),
(q)3
(3.15)
where, of course, S,0 (, x) = S (, x) (since 0 is the trivial character). Summing (3.15) for = a/q + /x and a going over all residues mod q coprime to q,
we obtain
q
(q,N
) ((q, N ))
(q)3 S+ ,0 (/x, x)2 S ,0 (/x, x)e(N/x).
(q)3
The integral of (3.15) over all of M = M0 ,r (see (3.5)) thus equals
(3.16)
Z 0 r
X ((q, N ))
2qx
2
S2+ ,0 (, x)S ,0 (, x)e(N )d
(q) ((q, N ))
3
0 r
(q)
qr
q odd
2qx
Z 0
X ((q, N ))
qx
2
+
(q) ((q, N ))
S2+ ,0 (, x)S ,0 (, x)e(N )d.
3
r
(q)
0
r
q2r
q even
qx
qr
q odd
2qx
Z 0
X ((q, N ))
qx
2
(c
+ (x))2 b (x)e(N )d.
+x
(q) ((q, N ))
3
r
(q)
0
r
q2r
q even
qx
We would like to complete both the sum and the integral. Before, we should
say that we will want to be able to use smoothing functions + whose Fourier
transform are not easy to deal with directly. All we want to require is that there
be a smoothing function , easier to deal with, such that be close to + in 2
norm.
Assume, then, that
|+ |2 0 | |,
14
H. A. HELFGOTT
(3)
qr
q odd
2qx
Z 0
X ((q, N ))
qx
2
+x
(q) ((q, N ))
(b (x))2 b (x)e(N )d.
3
r
(q)
0
r
q2r
q even
plus
(3.19)
qx
X (q)2 Z
q
(q)2
|(c
+ ()) (b ()) ||b ()|d .
2
|c
+ () + b ()|2 d
|c
+ () b ()| d
|b ()|
| |1 |c
+ b |2 |c
+ + b |2 = | |1 |+ |2 |+ + |2
| |1 |+ |2 (2| |2 + |+ |2 ) = | |1 | |22 (2 + 0 )0 .
x3
= x3
x3
(b (x))2 b (x)e(N )
X
0 r
q
min 2||x
,r
(q,2)
(q)2 =1
(b (x))2 b (x)e(N )d
(b (x))2 b (x)e(N )
((q, N ))
((q, N ))d
(q)3
X ((q, N ))
q1
(q)3
X
0 r
q
>min 2||x
,r
(q,2)
(q)2 =1
q
>min
(q,2)
0 r
,r
2||
(q)2 ((q, N ))
((q, N ))
((q, N ))d.
(q)3
(q)2
d.
(q)2
1
x2 | |1
|b |2
d + 2x2 | |1
d
3
r
(2)
0 r
0 /2
0 /2
!
(3) 2
|
|
x2
.
| |1 4.310040 | |21 + 0.00113 5 1
r
0
4This is obviously crude, in that we are bounding ((q, N ))/(q) by 1. We are doing so in
15
(q)3
(q)2 ((q, N )) =
Y
p|N
1
(p 1)2
Y
1+
pN
1
(p 1)3
(3.22)
(b (x))2 b (x)e(N )d =
Z Z
1
N
(t1 + t2 ) dt1 dt2 .
(t1 ) (t2 )
x 0
x
0
(This is standard. One rigorous way to obtain (3.22) is to approximate the integral over (, ) by an integral with a smooth weight, at different scales;
as the scale becomes broader, the Fourier transform of the weight approximates
(as a distribution) the function. Apply Plancherel.)
Hence, (3.17) equals
N
(t1 + t2 ) dt1 dt2
(t1 ) (t2 )
x
x
0
0
Y
Y
1
1
1+
1
(p 1)2
(p 1)3
2
(3.23)
p|N
pN
(3.24)
2
2.82643| |2 (2 + 0 ) 0 +
(3)
| |21
05
2
| |1 x
Here (3.23) is just as in the classical case [IK04, (19.10)], except for the fact
that a factor of 1/2 has been replaced by a double integral. We will later see
how to choose our smoothing functions (and x, in terms of N ) so as to make the
double integral as large as possible.
What remains to estimate is the contribution of all the terms of the form
err, (, x) in (3.12) and (3.13). Let us first deal with another matter bounding
the 2 norm of |S (, x)|2 over the major arcs.
The 2 norm. We can always bound the integral of |S (, x)|2 on the whole
circle by Plancherel. If we only want the integral on certain arcs, we use the
bound in Prop. 7.2 (based on work by Ramare). If these arcs are really the major
arcs that is, the arcs on which we have useful analytic estimates then we can
hope to get better bounds using L-functions. This will be useful both to estimate
the error terms in this section and to make the use of Ramares bounds more
efficient later.
16
H. A. HELFGOTT
By (3.8),
X
a mod q
gcd(a,q)=1
2
S a + ,
q
x
1 XX
)
(
()
(q)2
a mod q
gcd(a,q)=1
2
1 X
2
| ()| |S, (/x, x)|2 + Kq,1 (2|S (0, x)| + Kq,1 ),
=
(q)
where
Kq,1 = (1 +
q)(log x)2 || .
() =
q
q
q
q
( ),
q.
a mod q
(a,q)=1
1
(q)
6=T
q
q
q O
| err, (, x)|2 x2 + Kq,1 (2|S (0, x)| + Kq,1 )
|b
()|2 + O (|err,T (, x)(2||1 + err,T (, x))|)
(q)
2
2
+ O q max | err, (, x)| x + Kq,2 x ,
=
2 (q)x2
6=T
17
qr a mod q
q odd (a,q)=1
0 r
a
+ 2qx
q
0 r
a
2qx
q
|S (, x)| d +
X Z
q2r a mod q
q even (a,q)=1
a 0 r
+ qx
q
0 r
a
qx
q
|S (, x)|2 d
0 r
0 r
X 2 (q)x2 Z qx
X 2 (q)x2 Z 2qx
2
|b
(x)| d +
|b
(x)|2 d
=
0 r
0 r
(q)
(q)
qr
q odd
+ O
(3.25)
q2r
q even
2qx
qx
X 2 (q)x2 gcd(q, 2)0 r
ET, 0 r (2||1 + ET, 0 r )
(q)
qx
2
2
q
X 0 rx
O q
+
q
qr
q odd
max
Kq,2
| err, (, x)| +
x
mod q
6=T
||0 r/2q
X 20 rx
Kq,2
2
,
+
O q max | err, (, x)| +
mod
q
q
x
q2r
q even
where
6=T
||0 r/q
and T is the trivial character. If all we want is an upper bound, we can simply
remark that
0
0
X 2 (q) Z qx
X 2 (q) Z 2qx
|b
(x)|2 d + x
|b
(x)|2 d
x
(q) 0 r
(q) 0 r
r
qr
q odd
2qx
q2r
q even
qx
X 2 (q)
X 2 (q)
X 2 (q)
2
2
|b
+
.
|
=
2||
2
(q)
(q) 2
(q)
q2r
q even
qr
q odd
qr
q odd
qr
q odd
2qx
0 r
X 2 (q) Z 2q
2
1
|b ()| d + O
=
log r + 0.85 (|b |22 |b
|22 ).
(q) 0 r
2
qr
q odd
2q
18
H. A. HELFGOTT
Also,
0
0
X 2 (q) Z 2qx
X 2 (q) Z qx
2
|b
(x)| d = x
|b
(x)|2 d.
x
(q) 0 r
(q) 0 r
r
q2r
q even
qr
q odd
qx
2qx
0 r
2q
|b ()| d =
0 r
2q
|b ()| d O
(3)
= | |22 + O
| |21 q 5
5 6 (0 r)5
Hence
X 2 (q)
(q)
qr
q odd
0 r
2q
0 r
2q
|b ()|2 d = | |22
X 2 (q)
(q)
qr
q odd
0 r
2q
!
(3)
| |21
d
(2)6
.
+O
(q) 5 6 (0 r)5
qr
q odd
(3)
X 2 (q) |(3) |2 q 5
1 X 2 (q)q | |21
1
(q) 5 6 (0 r)5
r
(q) 5 6 05
qr
qr
q odd
q odd
(3)
log r 0.425
| |21
+
.
0.64787 +
4r
r
5 6 05
Going back to (3.25), we use (B.2) to bound
X 2 (q)x2 gcd(q, 2)0 r
q
(q)
qx
2.59147 0 rx.
+
q
q
q
q
r 2q
qr
q odd
q2r
q even
qr
q 2
qr
r
log 2e2 r.
2
We have proven the following result.
2 log er log
r(log
2e
r)K
+ O 0 xr 2 +
0
r,2 ,
0
2
where
(3.27)
E,r,0 =
max
mod q
qrgcd(q,2)
||gcd(q,2)0 r/2q
Kr,2 = (1 +
q| err, (, x)|,
19
||s
(3.28)
X 2 (q)
(q)
qr
q odd
and
Lr,0 = 2| |22
(3.29)
X 2 (q)
+ O (log r + 1.7) (|b |22 |b
|22 )
(q)
qr
q odd
(3)
+ O
2| |21
5 6 05
!
log r 0.425
+
.
0.64787 +
4r
r
The error term xrET,0 r will be very small, since it will be estimated using
the Riemann zeta function; the error term involving Kr,2 will be completely
2
negligible. The term involving xr(r + 1)E,r,
; we see that it constrains us to
0
have | err, (x, N )| less than a constant times 1/r if we do not want the main
term in the bound (3.26) to be overwhelmed.
3.2.1. The triple product and its error terms. There are at least two ways we can
evaluate (3.4). One is to substitute (3.10) into (3.4). The disadvantages here
are that (a) this can give rise to pages-long formulae, (b) this gives error terms
proportional to xr| err, (x, N )|, meaning that, to win, we would have to show
that | err, (x, N )| is much smaller than 1/r. What we will do instead is to use
our 2 estimate (3.26) in order to bound
the contribution of non-principal terms.
This will give us a gain of almost r on the error terms; in other words, towin,
it will be enough to show later that | err, (x, N )| is much smaller than 1/ r.
The contribution of the error terms in S3 (, x) (that is, all terms involving
the quantities err, in expressions (3.12) and (3.13)) to (3.4) is
X
X
X 1
3 (a)e(N a/q)
(3 )
(q)
qr
q odd
(3.30)
+
q2r
q even
a mod q
(a,q)=1
3 mod q
1
(q)
0 r
2qx
0 r
2qx
(3 )
3 (a)e(N a/q)
a mod q
(a,q)=1
3 mod q
0 r
qx
0 r
qx
20
H. A. HELFGOTT
We should also remember the terms in (3.11); we can integrate them over all of
R/Z, and obtain that they contribute at most
Z
3 Y
X
X
X p
2
|Sj (, x)| max
log p
j
d
qr
x
R/Z
j=1 j 6=j
p|q
p
2
|Sj (, x)|2 max
log p
j
qr
x
j=1 j 6=j
1
p|q
X
X p
2
2
=2
(n)+ (n/x) log r max
pr
x
n
1
sX
X
X p
2
2
2
2
(n)+ (n/x)
(n) (n/x) log r max
+4
pr
x
n
n
3 Y
X
0
2qx
qr a mod q
q odd (a,q)=1
(3.31)
X Z
q
0 r
qx
0 r
qx
q2r a mod q
q even (a,q)=1
M0 ,r
max
mod q
||0 r/2q
| err , (, x)|
S ( + a/q, x)2 d max | err , (, x)|
+
mod q
||0 r/q
S ()2 d
+
max
mod q
qrgcd(q,2)
||gcd(q,2)0 r/q
q| err , (, x)|.
(3.32)
which is its main term (coming from (3.12)). Instead of having an 2 norm as
in (3.31), we have the square-root
of a product of two squares of 2 norms (by
R
Cauchy-Schwarz), namely, M |S+ ()|2 d and
0
0
X 2 (q) Z 2qx
X 2 (q) Z qx
2
|
b
(x)x|
d
+
|b (x)x|2 d
(q)2 0 r
(q)2 0 r
r
(3.33)
qr
q odd
q2r
q even
2qx
x|b |22
X 2 (q)
q
(q)2
qx
21
+x
where A = (1/x)
max
q | err , (, x)| A
max
q | err+ , (, x)|( A + B+ ) B ,
mod q
qrgcd(q,2)
||gcd(q,2)0 r/q
mod q
qrgcd(q,2)
||gcd(q,2)0 r/q
2
M |S+ (, x)| d
B = 2.82643| |22 ,
(3.35)
B+ = 2.82643|+ |22 .
equals
(3.36)
S+ (, x)2 S (, x)e(N )d
C0 C , x2 + 2.82643| |22 (2 + 0 ) 0 +
(3)
| |21
05
| |1 x2
p
+ O (E ,r,0 A+ + E+ ,r,0 1.6812( A+ + 1.6812|+ |2 )| |2 ) x2
q
+ O 2Z+2 ,2 (x)LS (x, r) x + 4 Z+2 ,2 (x)Z2 ,2 (x)LS+ (x, r) x ,
where
(3.37)
Y
1
1
,
C0 =
1
1+
(p 1)2
(p 1)3
pN
p|N
Z Z
N
(t1 ) (t2 )
C , =
(t1 + t2 ) dt1 dt2 ,
x
0
0
Y
22
H. A. HELFGOTT
(3.38)
E,r,0 =
max
mod q
qgcd(q,2)r
||gcd(q,2)0 r/2q
q | err, (, x)|,
ET,0 r/2
A = L,r,0 ||22 + 5.190 r ET, 0 r |1 | +
2
2
3 log r
2
E,r,
+ 0 rx1 (log 2e2 r)Kr,2
+ 0 r 2 +
0
2
X 2 (q)
L,r,0 2||22
(q)
qr
q odd
Z,k (x) =
(n)(n/x),
LS (x, r) = log r max
,
pr
x n
x
1
23
(
2
t3 (2 t)3 e(t1) /2
(t) = h(t) (t) =
0
if t [0, 2],
otherwise.
Lemma 4.1. Let V be a real vector space with an inner product h, i. Then, for
any v, w V with |w v|2 |v|2 /2,
hv, wi = |v|2 |w|2 + O (2.71|v w|22 ).
x x2
1
1+x=1+ +
max
2
2 0t1 4(1 (tx)2 )3/2
2
x
x
=1+ +O
2
23/2
2
=1++O
+ O 2.71
.
=1+
+ 3/2
2
|v|22
|v|22
2
|v|2 |w|2 = |v|22 + hw v, vi + O 2.71|w v|22 = hv, wi + O 2.71|w v|22 .
24
H. A. HELFGOTT
sZ
sZ
+O
(4.5)
2.71
= | |22 + O
=
| |22
| |22
| (t)|2 dt
+O
2.71
| ((2 ) + t)|2 dt
Z
2.71(2 )
2
+ O (2.71(2 )
(r + t) dr
2 Z
0
2
| |2 ).
2
dt
(r + t)2 dtdr
2 2
=
d
(| |2 + O (2.71(2 ) | |2 ))
x
0
!
Z N
Z N
= | |22
provided that N/x 2. We see that it will be wise to set N/x very slightly larger
than 2. As we said before, will be scaled so that it is concentrated on a small
interval [0, ).
4.2. The approximation + to . We will define + : [0, ) R by
(4.7)
1
0
if |(s)| H,
otherwise.
25
log y
It is easy to check that the Mellin transform of this is indeed identical to [iH,iH]
on the imaginary axis: (4.9) is the Dirichlet kernel under a change of variables.
Now, in general, the Mellin transform of f M g is M f M g. We define
Z
sin(H log y) dy
h(ty 1 )
hH (t) = h M 1 IH (t) =
t
log y
y
2
(4.10)
Z 2
t
sin(H log y) dy
=
h(ty)
log y
y
0
and obtain that the Mellin transform of hH (t), on the imaginary axis, equals the
restriction of M h to the interval [iH, iH]. We may adopt the convention that
hH (t) = h(t) = 0 for t < 0.
4.2.1. The difference + in 2 norm. By (4.3) and (4.7),
Z
2
|hH (t) (t) h(t) (t)|2 dt
|+ |2 =
0
Z
(4.11)
dt
2
|hH (t) h(t)|2 .
max | (t)| t
t0
t
0
H
0
2
The maximum maxt0 | (t)| t is 1/ 2e.
Now, consider any : [0, ) C that (a) has compact support (or fast decay),
(b) satisfies (t) = O(t3 ) near the origin and (c) is quadruply differentiable
outside a finite set of points. We can show by integration by parts (cf. (2.1))
that, for (s) > 2,
Z
Z
Z
xs
xs+1
s dx
(x) dx =
(x)
=
dx
(x)x
M (s) =
x
s
s(s + 1)
0
0
0
Z
Z
xs+2
xs+3
=
(3) (x)
dx = lim
dx,
(4) (x)
s(s + 1)(s + 2)
s(s + 1)(s + 2)(s + 3)
t0+ t
0
where (4) (x) is understood in the sense of distributions at the finitely many
places where it is not well-defined
R as a function.
Let s = it, = h. Let Ck = 0 |h(k) (x)|xk1 dx for 0 k 4. Then
C4
(4.12)
M h(it) = O
.
|t||t + i||t + 2i||t + 3i|
Hence
(4.13)
and so
1
=
|hH (t) h(t)|
t
2 dt
C4
|+ |2
7
1
|M h(it)| dt
1
2e
1/4
1
.
H 7/2
C42
C42
dt
t6
7H 7
26
H. A. HELFGOTT
|+ |2
547.5562
.
H 7/2
This is at most
Z
Z
2
|hH (t) (t) h(t) )|2 | log t|dt
(+ (t) (t)) | log t|dt
0
0
Z
dt
2
max | (t)| t| log t|
|hH (t) h(t)|2 .
t0
t
0
Now
2
max | (t)|2 t| log t| = min
(t)t log t 0.3301223,
t0
t[0,1]
where we find the minimum by the bisection method (carried out rigorously, as
in Appendix C.2, with 30 iterations), Hence, by (4.13),
Z
480.394
(+ (t) (t))2 | log t|dt
(4.15)
.
H 7/2
0
***
As we said before, (M hH )(it) is just the truncation of (M h)(it) to the interval
[H, H]. We can write down M h explicitly:
M h = e1/2 (1)s (8(s + 3, 2) + 12(s + 4, 2) + 6(s + 5, 2) + (s + 6, 2)),
where (s, x) is the (lower) incomplete Gamma function
Z x
et ts1 dt.
(s, x) =
0
,
,
.
(4.16)
M h(it) = O min C0 ,
|t| |t||t + i| |t||t + i||t + 2i|
By (C.2), (C.3), (C.4), (C.6) and (C.7),
C0 1.622284,
C3 131.3399,
C1 3.580004,
C4 3920.882.
C2 15.27957,
(We will compute rigorously far more precise bounds in Appendix C.1, but these
bounds are all we could need.)
4.2.2. Norms involving + . Let us now bound some norms involving + . Relatively crude bounds will suffice in most cases.
First, by (4.14),
(4.17)
|+ |2 | |2 + |+ |2 0.80013 +
where we obtain
(4.18)
| |2 =
547.5562
,
H 7/2
0.640205997 . . . = 0.8001287 . . .
27
by symbolic integration. We can use the same idea to bound |+ (t)tr |2 , r 1/2:
|+ (t)tr |2 | (t)tr |2 + |(+ )(t)tr |2
s
r
max |
| (t)t |2 +
For example,
t0
(t)|2 t2r+1
dt
t
r
C4,1 1
.
| (t)tr |2 + max | (t)|2 t2r+1
t0
7 H 7/2
| (t)t0.7 |2 0.80691,
| (t)t0.3 |2 0.66168,
Thus
511.92
,
7/2
H
(4.19)
644.08
|+ (t)t0.3 |2 0.66168 +
.
H 7/2
equals (s 1)(M )(s 1). Since the Mellin
The Mellin transform of +
+
transform is an isometry in the sense of (2.4),
Z 1 +i
Z 1 +i
2
2
2
1
1
2
M (+ )(s) ds =
|s M + (s)|2 ds.
|+ |2 =
1
1
2i i
2i i
|+ (t)t0.7 |2 0.80691 +
Recall that + (t) = hH (t) (t). Thus, by (2.3), M + (1/2 + it) equals 1/2
times the (additive) convolution of M hH (it) and M (1/2 + it). Therefore, for
s = 1/2 + it,
Z
|s| H
|s| |M + (s)| =
M h(ir)M (s ir)dr
2 H
Z H
3
(4.20)
|ir 1||M h(ir)| |s ir||M (s ir)|dr
2 H
3
(f g)(t),
=
2
where f (t) = |ir 1||M hH (it)| and g(t) = | 1/2 + it||M (1/2 + it)|. (Since
(1/2 + i(t r)) + (1 + ir) = 1/2 + it = s, either either | 1/2 + i(t r)| |s|/3 or
|1+ir| 2|s|/3; hence |sir||ir1| = |1/2+i(tr)||1+ir| |s|/3.) By Youngs
inequality (in a special case that follows from Cauchy-Schwarz), |f g|2 |f |1 |g|2 .
Again by Cauchy-Schwarz and Plancherel (i.e., isometry),
Z
2 Z H
2
2
|f |1 =
|ir 1||M hH (ir)|dr =
|ir 1||M h(ir)|dr
2H
= 4H
|ir 1| |M h(ir)| dr = 2H
H
Z
0
|(th) (t)|2
|M ((th) )(ir)|2 dr
dt
4H 3.79234.
t
1
+i
2
1
i
2
|(M (
))(s)|2 ds
2
|
|2
.
4
28
H. A. HELFGOTT
Hence
3
3/2
1
1/4
3.79234 0.87531 H.
4H
(4.21) |+
|2 |f g|2
2
2 2
2 3/2
t | , > 0, in the same way. The only difference is that the
We can bound |+
2
integral is now taken on (1/2 + i, 1/2 + + i) rather than (1/2
i, 1/2 + i). For example, if = 0.7, then |s||M + (s)| = (6/2)(f g)(t),
where f (t) = |ir 1||M hH (it)| and g(t) = |0.2 + it||M ,1 (0.9 + it)|. Here
Z 1.2i
and so
(4.22)
1
3
0.7
|+
t |2 4H 3.79234 0.55091 1.95201 H.
2
log |22
1
=
2i
1
+i
2
1
|M (+ log)(s)| ds =
2i
2
1
i
2
1
+i
2
1
i
2
Now, (M + ) (1/2 + it) equals 1/2 times the additive convolution of M hH (it)
and (M ) (1/2 + it). Hence, by Youngs inequality, |(M + ) (1/2 + it)|2
(1/2)|M hH (it)|1 |(M ) (1/2 + it)|2 . By the definition of hH , Cauchy-Schwarz
and isometry,
sZ
Z H
H
|M h(ir)|2 dr
|M h(ir)|dr 2H
|M hH (it)|1 =
H
2H
sZ
|M h(ir)|2 dr
2H 2|h(t)/ t|2 .
2| log |2 .
Hence
2
H
2H|h(t)/ t|2 | log |2 = |h(t)/ t|2 | log |2 .
|+ log |2
3/2
(2)
Since
(4.23)
| log |2 1.39554,
we get that
(4.24)
|M hH (ir)|1 1.99301 2H
and
(4.25)
|+ log |2 1.10959 H.
29
dt
= | (t)t+1/2 |2
|hH (t)|2
t
0
s Z
i
1
|M hH (ir)|2 dr
= | (t)t+1/2 |2
2 i
s Z
1
| (t)t+1/2 |2
|M h(ir)|2 dr | (t)t+1/2 |2 |h(t)/ t|2 .
2
Since
| t+1/2 |2 =
sZ
t2
t2+1 dt =
( + 1)
2
|+ |1 0.996505.
Let us now get a bound for |+ | . Recall that + (t) = hH (t) (t). Clearly
(4.27)
| (t)t| = e1/2 .
| | = (1) = 1,
x
0
sin t
dt
t
H log 2 dw
2
e
t
Now
tew/H
d
t
=
h
dw
H
ew/H
h
t|h | .
ew/H
Hew/H
t
30
H. A. HELFGOTT
Integration by parts easily yields the bounds | Si(x) /2| < 2/x for x > 0 and
| Si(x) + /2| < 2/|x| for x < 0; we also know that Si(x) > 0 for x > 0 and
Si(x) < 0 for x < 0. Hence
Z 1
Z w/H !
2e
2t|h |
|hH (t) h(t)|
dw +
dw
H
w
0 2
1
t|h |
4
=
1 + E1 (1/H) ,
H
E1 (z) =
By [AS64],
0 < E1 (1/H) <
et
dt.
t
log(H + 1)
< log H.
e1/H
Hence
1 + 4 log H
|hH (t) h(t)|
< |h |
t
H
(4.29)
and so
4
h(t) hH (t)
< 1 + e1/2 |h | 1 + log H .
|+ | 1 + e
t
H
(4.30)
|h | = |h ( 21 3)| 3.65234.
1/2
We have proven
1 + 4 log H
log H
< 1 + 2.21526
.
H
H
We will need another bound of this kind, namely, for + log t. We start as in
(4.27):
|+ log t| | log t| + |(h(t) hH (t)) (t) log t|
(4.31)
1+
log H
.
H
4.3. The smoothing function . Here the challenge is to define a smoothing
function that is good both for minor-arc estimates and for major-arc estimates.
The two regimes tend to favor different kinds of smoothing function. For minorarc estimates, both [Tao] and [Hel] use
(4.32)
(4.33)
31
(For example, we can give major-arc estimates for + , which is based on (by
2
(4.7)). We will actually use here the function t2 et /2 , whose Mellin transform is
M (s + 2) (by, e.g., [BBO10, Table 11.1]).)
We will follow the simple expedient of convolving the two smoothing functions,
one good for minor arcs, the other one for major arcs. In general, let 1 , 2 :
[0, ) C. It is easy to use bounds on sums of the form
X
(4.34)
Sf,1 (x) =
f (n)1 (n/x)
n
where 1 = 2 I[1/2,1] . Bounding the sums S2 (, x) on the minor arcs was the
main subject of [Hel].
Before we use [Hel, Main Thm.], we need an easy lemma so as to simplify its
statement.
Lemma 4.2. For any q 1 and any r max(3, q),
q
< (r),
(q)
where
(4.37)
2.50637
.
log log r
Proof. Since (r) is increasing for r 27, the statement follows immediately for
q 27 by [RS62, Thm. 15]:
q
< (q) (r).
(q)
For r < 27, it is clear that q/(q) 2 3/(1 2) = 3; it is also easy to see that
(r) > e 2.50637 > 3 for all r > e.
It is time to quote the main theorem in [Hel]. Let x x0 , x0 = 2.16 1020 .
Let 2 = a/q + /x, q Q, gcd(a, q) = 1, |/x| 1/qQ, where Q = (3/4)x2/3 .
Then, if 3 q x1/3 /6, [Hel, Main Thm.] gives us that
||
q x,
(4.38)
|S2 (, x)| gx max 1,
8
32
H. A. HELFGOTT
where
(4.39)
p
(Rx,2r log 2r + 0.5) (r) + 2.5 Lr
+
+ 3.2x1/6 ,
gx (r) =
r
2r
with
Rx,t = 0.27125 log
(4.40)
+ 0.41415
9x1/3
2 log 2.004t
16 80
80
111
,
+ log 2 9 t 9 +
+
9
5
1+
7 13
Lt = (t) log 2 4 t 4
log 4t
(We are using Lemma 4.2 to bound all terms 1/(q) appearing in [Hel, Main
Thm.]; we are also using the obvious fact that, for 0 q fixed and 0 < a < b, 0a q b
is minimal when 0 is minimal.) If q > x1/3 /6, then, again by [Hel, Main Thm.],
(4.41)
where
(4.42)
We will work with x varying within a range, and so we must pay some attention
to the dependence of (4.38) and (4.41) on x.
Lemma 4.3. Let gx (r) be as in (4.39) and h(x) as in (4.42). Then
(
h(x) if x < (6r)3
x 7
gx (r) if x (6r)3
is a decreasing function of x for r 3 fixed and x 21.
Proof. It is clear from the definitions that x 7 h(x) (for x 21) and x 7 gx,0 (r)
are both decreasing. Thus, we simply have to show that h(x1 ) gx1 ,0 (r) for
x1 = (6r)3 . Since x1 (6 11)3 > e12.5 ,
Rx1 ,2r 0.27125 log(0.065 log x1 + 1.056) + 0.41415
Hence
1/3
Rx1 ,2r log 2r + 0.5 0.27215 log log x1 log x1 0.27215 log 12.5 log 3 + 0.5
0.09072 log log x1 log x1 0.255.
At the same time,
1/3
(4.43)
2.50637
x1
+
e log log x1 e log 3 + 1.9521
6
log log r
e log log x1
for r 37, and we also get (r) e log log x1 for r [11, 37] by the bisection
method (carried out rigorously, as in Appendix C.2, with 10 iterations). Hence
p
(Rx1 ,2r log 2r + 0.5) (r) + 2.5
p
(0.09072 log log x1 log x1 0.255) e log log x1 + 2.5
0.1211 log x1 (log log x1 )3/2 + 2,
33
and so
p
(Rx1 ,2r log 2r + 0.5) (r) + 2.5
1/6
1/6
<1
=
,
(log f ) =
f
f2
i.e., a2 + 2at 2ab, and that certainly happens when t b. In our case, b =
3 log(2.004r/9), and so t 3 log 6r implies t b.
34
H. A. HELFGOTT
where
(4.46)
p
(Rx,K,,2r log 2r + 0.5) (r) + 2.5 Lr
(4.48)
1/K
dw
S2 (, wx)(w)
+
w
1/K
S2 (, wx)(w)
dw
.
w
We bound the first integral by the trivial estimate |S2 (, wx)| |S2 (0, wx)|
and Cor. A.3:
Z 1/K
Z 1/K
dw
dw
wx(w)
1.04488
|S2 (0, wx)|(x)
w
w
0
0
Z 1/K
(w)dw.
= 1.04488x
0
If w 1/K, then wx x0 , and we can use (4.38) or (4.41). If q > (x/K)1/3 /6,
then |S2 (, wx)| h(x/K)wx by (4.41); moreover, |S2 (, y)| h(y)y for
x/K y < (6q)3 (by (4.41)) and |S2 (, y)| gy,1 (r) for y (6q)3 (by (4.38)).
Thus, Lemma 4.3 gives us that
Z
Z
dw
dw
|S2 (, wx)|(w)
h(x/K)wx (w)
w
w
1/K
1/K
Z
(w)dw h(x/K)||1 x.
= h(x/K)x
1/K
If q (x/K)1/3 /6, we always use (4.38). We can use the coarse bound
Z
dw
3.2K 1/6 ||1 x5/6
3.2x1/6 wx (w)
w
1/K
Since Lr does not depend on x,
Z
dw
Lr
Lr
wx (w)
||1 x.
w
r
1/K r
35
By Lemma 4.4 and q (x/K)1/3 /6, y 7 Rey ,t is convex-up and decreasing for
y [log(x/K), ). Hence
( log w
log w
Rx,t if w < 1,
R
+
1
1
1
x/K,t
log K
Rwx,t log K
Rx,t
if w 1.
Therefore
Z
dw
Rwx,t wx (w)
w
1/K
!
!
Z 1
Z
log w
log w
Rx,t (w)xdw
Rx,t x(w)dw +
1 Rx/K,t + 1
1
log K
log K
1/K
1
Z
Z 1
x
(w)dw + (Rx/K,t Rx,t )
Rx,t x
(w) log wdw
log K 1/K
1/K
C,2
Rx,t ||1 + (Rx/K,t Rx,t )
x,
log K
where
C,2,K =
we see that it is enough to prove that r 7 Ry,t log 2r log log r/ 2r is decreasing
on r for y = x and y = x/K (under the assumption that r 175).
Looking at (4.40) and at (4.49), it remains only to check that
!r
log 8r
log log r
(4.50)
r 7 log 1 +
9x1/3
r
2 log
(4.49)
r 7
4.008r
1
+ logr8r
r
22
log 8r
2
log 1 +
log 8r
2
+
1
1
< ,
2r log r log log r
2r
9x
where = log 4.008r
. Since r x1/3 /6, log 54/4.008 > 2.6. Thus, it is enough
to ensure that
2/2.6
1
+
< 1.
(4.51)
log 8r
log 8r
log r log log r
1 + 2 log 1 + 2
Since this is true for r = 175 and the left side is decreasing on r, the inequality
is true for all r 175.
36
H. A. HELFGOTT
Lemma 4.7. Let x 1024 . Let : [0, ) [0, ) be continuous and in L1 .
Let gx, (r) and h(x) be as in (4.46) and (4.42), respectively. Then
2 0.275
gx,
x
h(x/ log x).
3
Proof. We can bound gx, (r) from below by
p
(Rx,r log 2r + 0.5) (r) + 2.5
gmx (r) =
2r
0.275
Let r = (2/3)x
. Using the assumption that x 1024 , we see that
8 0.275
log 3 x
+ 0.41415 0.67086.
Rx,r = 0.27125 log 1 +
1
93/2
0.275
3
2 log 2.004 x
Using x 1024 again, we get that
2.50637
5.72864.
log log r
Since log 2r = 0.275 log x + log 4/3, we conclude that
0.44156 log x + 4.1486
p
.
gmx (r)
4/3 x0.1375
(r) = e log log r +
Recall that
0.44156 log x + 2
0.2727(log x)3/2
,
>
x0.1375
x1/6
for x 1024 , we are done.
2.1486
1218 log x
>
0.1375
x
x1/3
R+
0
R+
0
where :
X
1
L (s, )
(n)(n)f (n/x) =
(5.1)
F (s)xs ds,
2i 2i
L(s, )
n=1
37
along the way. (Once the line of integration is far enough to the left, the terms ts
within (5.1) become very small, and so the value of the integral ought to become
very small, too.)
We will assume we know the zeros of L(s, ) up to a certain height H0
meaning, in particular, that we know that non-trivial zeros with (s) H0 lie
on the critical line (s) = 1/2 but we have little control over the zeros above
H0 . (The best zero-free regions available are not by themselves strong enough
for our purposes.) Thus, we want to choose so that the Mellin transform F
decays very rapidly both for = 0 and for non-zero and bounded.
5.2. How to choose a smoothing function? The method of stationary phase
([Olv74, 4.11], [Won01, II.3])) suggests that the main contribution to (5.2)
should come when the phase has derivative 0. The phase part of (5.2) is
e(t) = t(s) = e2it+ log t
(where we write s = + i ); clearly,
(2t + log t) = 2 +
=0
t
(t)e(t)t
s1
+i 1
|(2it + log t) | =
2||
| /t2 | = p .
| |
1
.
2 2
In other words, if sgn( ) 6= sgn() and is not too small, asking that F ( +i )
decay rapidly as | | amounts to asking that (t) decay rapidly as t 0.
Thus, if we ask for F ( + i ) to decay rapidly as | | for all moderate , we
are requesting that
(1) (t) decay rapidly as t ,
(2) the Mellin transform F0 ( + i ) decay rapidly as .
Requirement (2) is there because we also need to consider F ( + it) for very
small, and, in particular, for = 0.
There is clearly an uncertainty-principle issue here; one cannot do arbitrarily
well in both aspects at the same time. Once we are conscious of this, the choice
(t) = et in Hardy-Littlewood actually looks fairly good: obviously, (t) = et
decays exponentially, and its Mellin transform (s + i ) also decays exponentially
as . Moreover, for this choice of , the Mellin transform F (s) can be
written explicitly: F (s) = (s)/(1 2i)s .
38
H. A. HELFGOTT
for (a) > 1/2; this can be extended to all a, z C either by analytic continuation or by other integral representations ([AS64, 19.5], [Tem10, 12.5(i)]).
Hence
1
(i)2
(5.5)
F (s) = e
(s)U s , 2i .
2
[HL23, Lemma 4] does not make all constants explicit. The constants and trivial-zero terms
were fully worked out for q = 1 by [Wig20] (cited in [MV07, Exercise 12.1.1.8(c)]; the sign of
hyp,q (z) there seems to be off). As was pointed out by Landau (see [Har66, p. 628]), [HL23,
Lemma 4] actually has mistaken terms for non-primitive. (The author thanks R. C. Vaughan
for this information and the references.)
39
it would seem that all fully explicit error terms in the literature are either for a
and z real, or for a and z outside our range of interest (see both Olvers work and
[TV03].) The bounds in [Olv61] involve non-explicit constants. Thus, we will
have to find expressions with explicit error bounds ourselves. Our case is that of
a in the critical strip, z purely imaginary.
Gaussian smoothing has been used before in number theory; see, notably,
[HB79]. What is new here is that we will derive fully explicit bounds on the Mellin
transform of the twisted Gaussian. This means that the Gaussian smoothing will
be a real option in explicit work on exponential sums in number theory from now
on.
5.3.1. General approach and situation. We will use the saddle-point method (see,
e.g., [dB81, 5], [Olv74, 4.7], [Won01, II.4]) to obtain bounds with an optimal
leading-order term and small error terms. (We used the stationary-phase method
solely as an exploratory tool.)
What do we expect to obtain? Both the asymptotic expressions in [Olv59] and
the bounds in [Olv61] make clear that, if the sign of = (s) is different from that
of , there will a change in behavior when gets to be of size about (2)2 . This is
unsurprising, given our discussion using stationary phase: for |(a)| smaller than
a constant times |(z)|2 , the term proportional to e(/4)| | = e|(a)|/2 should
be dominant, whereas for |(a)| much larger than a constant times |(z)|2 , the
2
1
term proportional to e 2 ( 2 ) should be dominant.
5.3.2. Setup. We write
(5.6)
(u) =
u2
(2i)u i log u
2
e(u) u
du
.
u
i 2 + 4i
(5.8)
u0 =
.
2
The second derivative at u0 is
1
1
(5.9)
(u0 ) = 2 u20 + i = 2 (iu0 + 2i ).
u0
u0
(5.7)
u 2i
Assign the names u0,+ , u0, to the roots in (5.8) according to the sign in front
of the square-root (where the square-root is defined so as to have argument in
(/2, /2]).
We assume without loss of generality that 0. We shall also assume at first
that 0 (i.e., 0), as the case < 0 is much easier.
40
H. A. HELFGOTT
where y = (iu0 ). (This is the main part of the contribution of the saddle
point, without the factor that depends on the contour.) We have
!
r
2
2
p
4i
i
i
i + 2 + 4i
(5.11)
y=
1 + 2 .
=
2
2
2
2
2
1+j()
1
(5.13)
2
= arcsin s q
q
1+j()
1
2 1+j()
2
2
v
u
u1
= arcsin t
2
s
!
1
2
2
= arccos
1 + j()
2
1 + j()
s
!!
r
2
y
2
j() + 1
1
arg(u0,+ ) + = arccos
2
1 + j() 2
2
2
(5.14)
2(() 1)
1
,
= arccos
()
2
p
where () = (1 + j())/2.
It is clear that
1
2(() 1)
(5.15)
lim arccos
()
2
whereas
(5.16)
as 0+ .
arccos
2(() 1)
=
()
2 4
4
41
1
We are still missing a factor of |u0,+ | (from (5.10)),
p a factor of |u0,+ | (from
the invariant differential du/u) and a factor of 1/ | (u0,+ )| (from the passage
by the saddle-point along a path of steepest descent). By (5.9), this is
|u |1
p 0,+
=
| (u0,+ )|
|u0,+ |1
|u0,+ |
p
p
=
.
1
|iu0,+ + 2i |
|iu0,+ + 2i |
|u0,+ |
=
1 + j() 2
()2 ().
=
2
2
2
r
r
2
2
2
i
j()
+
1
j()
1
= + 2i +
2
2
2
2
2
v
(5.18)
!2
!2
u
r
r
j()
+
1
j()
1
2 u
+
= t 1 +
2
2
2
s
r
r
2
j()
+
1
j() 1
=
2
.
j() + 2 + 1 2
2
2
2
p
p
Since j() 1 = / j() + 1, this means that
(5.19)
v
s
u
2
u
2
|iu0,+ + 2i | = tj() + 2 + 1
(j() + 1 + 2 )
2
j() + 1
2 p
j() + j()2 (())1 (j() + j()2 )
2
p
2 j() p
2 p
2
1
Hence
|
|u0,+
p
=
|iu0,+ + 2i |
p
2 ()
()
1
2
=
1
1
(j())1/4
2 2 4 j() 4
(()2 ())1/4
21/4
3
2 2 4
2
41
2
2
1
2
j() 4
(() ())
(()2 ()) 2 4
42
H. A. HELFGOTT
and positive, where is as in 5.6. Thus arg(v) = arg( (u0,+ ))/2. By (5.9),
arg( (u0,+ )) = arg(iu0,+ + 2i ) 2 arg(u0,+ ).
Starting as in (5.18), we obtain that
and
(5.20)
arg(iu0,+ + 2i ) = arctan
q
1 +
j1
2
j+1
2
j1
2
j+1
2
1 +
j1
2
j+1
2
p
p
2(j 1) + 2(j + 1)
=
=
j1
1 + j+1
2
p
q
2
j 2 1 + (j + 1)
+ j+1
+ 1 ( + (j + 1))
=
=
j1
j1
(j + 1)(1 + j/)
2( + j)
(1 + j/)
=
=
.
=
j1
1+
Hence, by (5.13),
arg( (u0,+ )) = arctan
2( + j)
arccos ()1 .
1
2v(v + j)
arg( (u0,+ ))
= arg(u0,+ ) arctan
2
2
= arg(u0,+ ) arctan ,
arg(v) =
where
(5.22)
= tan
1
2v(v + j)
arctan
.
2
Since
1
1
=
tan =
2
sin tan
we see that
(5.23)
1+
1
1
,
tan2 tan
1+ 2
2
4 ( + j)
2( + j)
cos =
2
1 + cos
,
2
sin =
2
0 = arg(u0,+ ) =
1 cos
.
2
Hence, if we let
(5.24)
1
1
arccos
,
2
()
we get that
cos 0 = cos
(5.25)
sin 0 = sin
1
1
arccos
2
()
1
1
arccos
2
()
=
=
43
1
1
+
,
2 2()
1
1
.
2 2()
arctan > 0 ,
for r [(c0 1)/c1 , c0 /c1 ], where c0 and c1 are parameters to be set later.
The curve goes from (0, (c0 1)/c1 ) to (c0 /c1 , 0), and stays within the first
quadrant.7 In order for the curve to go through the point u0,+ , we must have
c1 r0
+ c0 = sin 0 ,
(5.28)
where
p
()2 (),
r0 = |u0,+ | =
(5.29)
2
7Because c 1, by (C.21).
0
44
H. A. HELFGOTT
and 0 and sin 0 are as in (5.24) and (5.25). We must also check that the curve
C goes through u0,+ in the direction of steepest descent. The argument of the
point (x, y) is
c r
y
1
+ c0 .
= arcsin = arcsin
r
Hence
d arcsin c1r + c0
c1 /
d
c1 r
=
=r
=r
.
r
dr
dr
cos
cos arcsin c1r + c0
This means that, if v is tangent to C at the point u0,+ ,
tan(arg(v) arg(u0,+ )) = r
d
c1 r0
=
,
dr
cos 0
c1 =
cos 0
,
r0
c1 =
1 + 1/
,
2
c0 =
1
1
+
+
2 2
1
1
.
2 2
2
c1
c
1
The contour will be as follows: first we go out of the origin along a straight
radial segment C1 ; then we meet the curve C, and we follow it clockwise for a
segment C2 , with the saddle-point roughly at its midpoint; then we follow another
radial ray C3 up to infinity. For small, C3 will just be the x-axis. Both C1 and
C3 will be contained within the first quadrant; we will specify them later.
5.3.5. The integral over the main contour segment C2 . We recall that
u2
+ iu i log u.
2
Our aim is now to bound the integral
Z
e((u)) u1 du
(5.33)
(u) =
C2
over the main contour segment C2 . We will proceed as follows. First, we will
parameterize the integral using a real variable , with the value = 0 corresponding to the saddle point u = u0,+ . We will bound ((u)) from below by
an expression of the form ((u0,+ )) + 2 . We then bound |u|1 |du/d| from
above by a constant. This procedure will give a bound that is larger than the
true value by at most a (very moderate) constant factor.
45
(5.34)
r 2 2y 2
y
x2 y 2
y + =
y + arcsin
2
2
r
2
4
=
0 () = 2 2 0 (),
((u)) =
( + 0 )2
(1 2(sin 0 c1 )2 )
2
arcsin(sin 0 c1 )
+ 0
(sin 0 c1 ) +
,
0 () =
and
(5.35)
r r0
,
0 =
r0
.
c0 c1 0 = sin 0 .
(Here = (1 sin 0 )/(c1 ) corresponds to the intersection with the y-axis, and
= (sin 0 )/(c1 ) corresponds to the intersection with the x-axis.)
We work out the expansions around 0 of
(5.39)
2 cos 20
( + 0 )2
(1 2(sin 0 c1 )2 ) = 0
+ (0 cos 20 + 202 c1 sin 0 )
2
2
cos 20
2 2 2
+
+ 4c1 0 sin 0 c1 0 2
2
+ 2(0 c21 2 + c1 sin 0 ) 3 c21 2 4 ,
0 sin 0
sin 0
+ 0
(sin 0 c1 ) =
+
+ c1 0 + c1 2 ,
0
1 X Pk (sin 0 ) (c1 )k k
arcsin(sin 0 c1 )
=
+
4
4 4
k!
(cos 0 )2k1
k=1
0
c1
1
(c1 )2 sin 0 2
=
+
+
+ ... ,
4 4 cos 0
2(cos 0 )3
2
where P1 (t) = 1 and P
k+1 (t) = Pk (t)(1t )+(2k1)tP (t) for k 1. (This follows
from (arcsin z) = 1/ 1 z 2 ; it is easy to show that (arcsin z)(k) = Pk (z)(1
z 2 )(k1/2) .)
P
k
We sum these three expressions and obtain a series 0 () =
k ak . We
already know that
(1) a0 equals the value of ((u))/(2 2 ) at the saddle point u0,+ ,
(2) a1 = 0,
46
H. A. HELFGOTT
(3)
1
a2 =
2
2
dr
d
2
2
2
du
du
1
2j()
|
iu
+
2i
|
2
0,+
=
| (u0,+ )| =
=
,
2
2
2
|u0,+ |
2
2 ( )
and, by (5.31) and (5.32),
v
u
|du|
du
u
|r=r =
|
=
t1 +
r=r
0
0
|dr|
dr
(5.40)
=
=
Thus,
1
a2 =
2
2
c21
c2
1 + 12
r02
c0
c1
r0
2
2
2 ( )
1
1
2 + 2
2 =
1+
1 + c21
c21
r02
2 1 sin2 0
2
1 + 1/
1 + 2 .
2j()
(1 + 2 ),
2
where is as in (5.23).
Let us simplify our expression for 0 () somewhat. We can replace the third
series in (5.39) by a truncated Taylor series ending at k = 2, namely,
0
1
(c1 )2 sin 1 2
c1
arcsin(sin 0 c1 )
=
+
+
4
4 4 cos 0
2(cos 1 )3
for some 1 between 0 and . Then 1 [0, /2], and so
0
1 c1
arcsin(sin 0 c1 )
.
4
4 4 cos 0
Since
R() = c21 2 2 + 2(sin 0 c1 0 )c1
is positive. It is:
2
=
2
47
1 + 1/
,
2
2
We conclude that
(5.41)
0 ()
((u0,+ ))
+ 2 ,
2 2
where
(5.42)
= a2
1 (c1 )2 sin 0
+ sin2 0 (sin 20 ) .
4 2(cos 0 )3
1 (c1
1 + 1/
0
2
= 2
4 2(cos 0 )3
8
=
1
2
1
1 3/2
2 + 2
2
p
1 1/
2
p
=
4 2 1 + 1/
2
2
=
= p
= q
4 2 1
4 j+1 j1
4 2 /4
2
2
1
2
1
1
2 1
sin 20 = 2 sin 0 cos 0 = 2
2 =
4 4
/2
1
=
.
=
=
2
(j + 1)/2
j+1
2j
2
(5.43)
=
(1
+
.
2
2
2
2 2 j + 1
Now recall that our task is to bound the integral
Z 1
Z
2 2 0 ()
1 du dr
((u))
1
e
(( + 0 ))
e
|u|
|du| =
dr d d
0
C2
(5.44)
Z 1
((u0,+ ))
2 2 2
1 du
() e
e
( + 0 )
dr d.
0
(We are using (5.34) and (5.41).) Since u0,+ is a solution to equation (5.7), we
see from (5.6) that
!
u20,+
+ iu0,+ i log u0,+
((u0,+ )) =
2
iu0,+ i
1
+
+ arg(u0,+ ) = (iu0,+ ) + arg(u0 , +).
=
2
2
2
48
H. A. HELFGOTT
e((u0 ,+)) = e
If 1, we can bound
( + 0 )1
(5.45)
2(1)
1
(
01
(0 + 0 )1
if 0,
if < 0,
1
(1 + 0 )
if > 0.
By (5.32),
s
r
2
du
(c
r/)
(c1 ( + 0 ))2
1
= 1+
=
1
+
dr
cos2
1 (sin 0 c1 )2
(This diverges as /2; this is main reason why we cannot actually follow the
curve all the way to the y-axis.) Since we are aiming at a bound that is tight
only up to an order of magnitude, we can be quite brutal here, as we were when
using (5.45): we bound (c1 r/)2 from above by its value when the curve meets
the x-axis (i.e., when r = c0 /c1 ). We bound cos2 from below by its value when
= 1 . We obtain
s
s
2
du
c20
c
0
= 1+
1
+
=
,
dr
1 (sin 0 c1 1 )2
cos2
where is the value of when = 1 .
Finally, we complete the integral in (5.44), we split it in two (depending on
whether 0 or < 0) and use
Z
Z
/2
1
2 2 2
2
d
e
e d = .
0
0
Therefore,
(5.46)
Z
e((u)) |u|1 |du|
C2
/2
c20
1
1
+
(
+
1+
= () e
j
0
0
cos2
1 ! s
arccos 1 2(1)
2
2
1
c0 e
j
=
1+
r0
,
1+ 1+
2
2
0
cos
arccos
1
2(1)
(c0 sin ) c1
c0 sin
r
=
=
.
0 + 0 =
c1
49
.
0 =
2
Definition (5.22) implies immediately that 1. Thus, by (5.31),
p
(5.48)
c1 0 = 2(1 + 1/) 2 2,
while, by (C.9),
(5.49)
c1 0 =
,
0
0 + 0
c0 sin
c0 sin
whereas
1/ 2
sin 0
1
1.62628.
1+
=1+
1+
0
c1 0
0.79837 2
We will now use some (rigorous) numerical bounds, proven in Appendix C.2.
First of all, by (C.21), c0 > 1 for all > 0; this assures us that c0 sin > 0,
and so the last expression in (5.50) is well defined. By (5.47), this also shows that
0 + 0 > 0, i.e., the curve C stays within the first quadrant for 0 /2, as
we said before.
We would also like to have an upper bound for
s
1
c20
(5.51)
,
1+
cos2
(5.52)
= .
4
Thus, by (C.36),
s
s
c20
1
1 + 2c20 p
1+
min(5,
0.86)
min(
5, 0.93 ).
2
cos
We also get
2
2
7.82843.
c0 sin
1 1/ 2
Finally, by (C.39),
(
r
/6
if 4,
2
1
1
2
if > 4
2 23/2 1 23/2
2
=
min
, .
r0 23
3
if > 2
3
We conclude that
Z
arccos 1 2(1)
((u))
1
2
,
e
|u|
|du| = C, e
(5.53)
C2
50
H. A. HELFGOTT
where
(5.54)
C,
1
3/2
3.3
1
1
1 + max 7.83
, 1.63
= min 2,
min( /, )
for all > 0, > 0 and all . By reflection on the x-axis, the same bound holds
for < 0, < 0 and all . Lastly, (5.53) is also valid for = 0, provided we
replace (5.54) and the exponent of (5.53) by their limits as 0+ .
5.3.6. The integral over the rest of the contour. It remains to complete the contour. Since we have set = /4, C1 will be a segment of the ray at 45 degrees
from the x-axis, counterclockwise
(i.e., y = x, x 0). The segment will go from
and so
r =
c1
(5.55)
1
c0
2
e((tw)) t1 dt,
C1
t
= + ,
i + iwt i log t + i
(5.57)
((tw)) =
2
4
4
2
and, by (C.40),
r
+ 0.076392 + =
0.30556 > 0.4798.
4
4
4
2
4
2
dt r
e
r
4
2
(5.58)
Hence, for 1,
(5.59)
C1
u2 /2
e(u)u
/2
s1
du /2 e0.4798 .
51
By (5.57),
r
u2 /2
r
us wr
((wr )) r
4
e
2
|
e
=
e
e(u)
0
s
|s|
4 +1
u2 /2
2
t
dt +
e
e(u)
du
e
e 2 4 t dt
s
|s| 0
|s| 0
C1
!Z
+1
r t
r
r
e 2 4 dt
+
|s|
|s|
0
!
+1
r
+ r
r
2
min
, r e 2 4
+2
r
r
2r
e 2 4.
+
By (5.58),
+2
r
(1 + 2)r
2 +1 + (1 + 2) 2
+
.
We conclude that
2 +1 + (1 + 2) 2 r +
2 /2
u
s1
e 2 4
e
e(u)u
du
C1
!
1+ 2
2 e0.4798
1+
2
dx
2
dx
2
dt
x /2
s
t /2
s
=
e
e(x)x
ex /2 x .
e
e(t)t
(5.61)
c0
c0
t
x
x
C3
c1
c1
Now
2
2
2
ex /2 x2 = ex /2 x1 ( 2)ex /2 x3
2
2
2
ex /2 (x2 + ( 2)x4 ) = ex /2 x1 ( 2)( 4)ex /2 x5
52
H. A. HELFGOTT
x
x2 /2
e
x2 + ( 2)x4
2
x
+ ( 2)x4 + ( 2)( 4)x6
c0
min
c1
,
4
if 0 2,
if 2 4.
if 4 6,
We conclude that
Z
2 25
2 /2
dt
min 21 ( ) , 32
t
s
P min , 5
,
e
e(t)t
e
t
4
C3
Let us try = 0 first. Much as in (5.56) and (5.57), we obtain, for < 0,
Z
Z t
dt
+ 4 1
2
t
dt = e 4
e||t/ 2 t
e
|I|
t
0
0
! Z
!
(5.62)
2
2
dt
= e 4
et t =
() e 4
||
t
||
0
( 4 )
e
=
( 4 )
e( 4 )
t
=
t
(sin 2)/2
t2 sin 2 dt
e
2/21
(sin 2)/2
ey y 2
t2
e 2 t
dt
t
dy
=
2/2 (/2)e 4 .
/2
y
2(sin 2)
Here we can choose = (arcsin 2/ )/2 (for 2). Then 2 (/2) (2/ ) =
/ , and so
(5.63)
|I|
e/2 /2
e 2
4
/2
4
(/2)
2
(/2)e
2
2(2/ )/2
53
The only issue that remains is that may be close to 0, in which case (/2)
can be large. We can resolve this, as before, by doing an integration by parts. In
general, for 1 < < 1, s 6= 0:
Z
us
2
eu /2 e(u)
|I| e
du
s
C
Z
us
2
(u + i)eu /2iu du
=
s
C
Z
Z
1
2
i
2
u /2
s+1
=
e
e(u)u du +
eu /2 e(u)us du.
s C
s C
u2 /2
(5.64)
us
e(u) |v
s 0
!+2
2
||
( + 2) e 4 +
||
|s|
!
2
4
+ 2 e 4 .
2
||
!+1
2
( + 1) e 4
||
1 e/2
|| e/2
(( + 2)/2) (+2)/2 e 4 +
(( + 1)/2) (+1)/2 e 4
|s| 2
|s| 2
/2
/2
||
e
1+
e 4
2
|I|
for [0, 1], where we are using the facts that (s)
(s) 1 for s [1, 2].
5.4. Totals. Summing (5.53) with the bounds obtained in 5.3.6, we obtain our
final estimate. Recall that we can reduce the case < 0 to the case > 0 by
reflection. We have proven the following statement.
2 /2
2 /2
et
e(t)ts
dt
,
t
(5.65)
|F (s)| C0,, e
| |
()2
0.4798
+ C1,, e
| |
2 25
min 18 (
) , 32 | |
+ C2,, e
54
where
(5.66)
H. A. HELFGOTT
2(() 1)
1
arccos
,
()
p !
| |
= min 2, 2||
1 + max(7.831 , 1.631 )
1
E() =
2
C0,,
3.3
C1,,
C2,,
min
s
p
!
1+ 2
1 + 1 + 2
2
= 1+
,
,
() =
2
| | 5 p
= P min
,
| |
,
2|| 4
3/2
| | p
| |
2|| ,
e 4 | | ,
|F (s)| C,
(5.67)
where
C,
e/2 /2
2
1 + 23/2 ||
(/2)
| |
The terms in (5.65) other than C0,, eE(| |/() )| | are usually very small.
In practice, we will apply Prop. 5.1 when | |/2|| is larger than a moderate
constant (say 8) and | | is larger than a somewhat larger constant (say 100).
Thus, C0,, will be bounded.
2
For comparison, the Mellin transform of et /2 (i.e., F0 = M f0 ) is 2s/21 (s/2),
which decays like e(/4)| | . For very small (e.g., | | < 2), it can make sense to
use the trivial bound
Z
2/2
2
dt
et /2 t = 2/21 (/2)
(5.68)
|F (s)| F0 () =
t
0
for (0, 1]. Alternatively, we could use integration by parts (much as in (5.64)),
followed by the trivial bound:
Z
us
F (s + 2) 2i
2
eu /2 e(u)
(5.69)
F (s) =
du =
F (s + 1),
s
s
s
0
and so
r
+1
+2
+ 2 2 1 |2| +1
2 2 1 +2
1 + 2||
2
2
(5.70)
|F (s)|
|s|
2
|s|
(5.71)
E()
8
384
for all > 0. We can also write
sin 2
,
(5.72)
E() =
4
2
4(1 + sin )
55
cos2
cos4 cos2
2 sin
2 1 cos2
=
.
=
2
cos
cos2
Thus
(5.75)
2 cos1 1
(1 cos ) cos
(1 cos2 ) cos
=
2E() =
=
2 sin
sin
sin (1 + cos )
cos2
=
sin 2
sin cos
=
.
1 + cos
4 cos2 2
53
,
4 24 8
giving us (5.71).
To obtain (5.72), simply define = /2 ; the desired inequality follows
from the last two steps of (5.75).
Let us end by a remark that may be relevant to applications outside number
theory. By (5.5), Proposition 5.1 gives us bounds on the parabolic cylinder function U (a, z) for z purely imaginary and |(a)| 1/2. The bounds are useful when
|(a)| is at least somewhat larger than |(a)| (i.e., when | | is large compared to
). As we have seen in the above, extending the result to broader bands for a is
not too hard integration by parts can be used to push a to the right.
6. Explicit formulas
An explicit formula is an expression restating a sum such as S, (/x, x) as a
sum of the Mellin transform G (s) over the zeros of the L function L(s, ). More
specifically, for us, G (s) is the Mellin transform of (s)e(s) for some smoothing
function and some R. We want a formula whose error terms are good both
for very close or equal to 0 and for farther away from 0. (Indeed, our choices
of will be made so that F (s) decays rapidly in both cases.)
2
We will derive two explicit formulas: one for (t) = t2 et /(2) , and the other
for = + , where + is defined as in (4.7) for some value of H > 0 to be set
2
later. As we have already discussed, both functions are variants of (t) = et /2 .
Thus, our expressions for G will be based on our study of the Mellin transform
F of e(t) in 5.
56
H. A. HELFGOTT
X
X
n=1
where
Iq=1
(
1
=
0
if q = 1,
if q =
6 1.
and
P the norms ||2 , | |2 are taken with respect to the usual measure dt. The sum
is a sum over all non-trivial zeros of L(s, ).
X
1
L (s, )
(n)(n)e(n/x)(n/x) =
G (s)xs ds
2i 1+i
L(s, )
n=1
(see [HL23, Lemma 1] or, e.g., [MV07, p. 144]). We shift the line of integration
to (s) = 1/2. If q = 1 or (1) = 1, then L(s, ) does not have a zero at
s = 0; if q > 1 and (1) = 1, then L(s, ) has a simple zero at 0 (as discussed
in, e.g., [Dav67, 19]). Thus, we obtain
Z 2+i
X
1
L (s, )
G (s)xs ds = Iq=1 G (1)x
G ()x I1, G (0)
2i 2i
L(s, )
(6.2)
Z 1/2+i
L (s, )
1
G (s)xs ds,
2i 1/2i L(s, )
where I1, = 1 if q > 1 and (1) = 1 and I1, = 0 otherwise. Of course,
Z
(t)e(t)dt = b().
G (1) = M ((t)e(t))(1) =
0
It is time to estimate the integral on the right side of (6.2). By the functional
equation (as in, e.g., [IK04, Thm. 4.15]),
s+
1s+
1
1
L (1 s, )
L (s, )
= log
,
(6.3)
L(s, )
q
2
2
2
2
L(1 s, )
where (s) = (s)/(s) and
(
0
=
1
if (1) = 1
if (1) = 1.
1
.
+
= (1 s) + log 2 + cot
(6.4)
2
2
2
2
2
57
Now, if (z) = 3/2, then |t2 + z 2 | 9/4 for all real t. Hence, by [OLBC10,
(5.9.15)] and [GR00, (3.411.1)],
Z
tdt
1
2
(z) = log z
2
2
2z
(t + z )(e2t 1)
0
!
Z
tdt
1
+ 2 O
= log z
9 2t
2z
1)
0
4 (e
Z
1
tdt
8
(6.5)
= log z
+ O
2z 9
e2t 1
0
8
1
1
+ O
(2)(2)
= log z
2z 9
(2)2
1
1
10
= log z
+O
= log z + O
.
2z
27
27
e 4 2 + e 4 i+ 2
(s
+
)
=
cot
e 4 i 2 e 4 i+ 2 = 1.
2
2
2
(6.7)
q 1
9
log + log 2 +
+ 4.1396.
2
4
Recall that we must bound the integral on the right side of (6.2). The absolute
value of the integral is at most x1/2 times
Z 1 +i
L (s, )
2
1
G (s) ds.
(6.8)
2 1 i L(s, )
2
u
u 1
2
2
L
(s,
)
1
t
|ds| t 1
|G (s)s|2 |ds|
2 1 i L(s, ) s
2 1 i
2
By (6.7),
v
v
uZ 1 +i
uZ 1 +i
(s, ) 1 2
u
u
log q 2
2
2
L
t
t
L(s, ) s |ds|
s |ds|
21 i
12 i
v
uZ
u 1 log 2 + 9 + 4.1396 + log 2
t
2
4
+
d
1
2
4 +
2 log q + 226.844,
58
H. A. HELFGOTT
d(e(t)(t))
= 2ite(t)(t) te(t) (t)
dt
u 1
2
t
|G (s)s|2 |ds| =
|2ite(t)(t) te(t) (t)|2 t2 dt
2 1 i
0
2
sZ
sZ
= 2||
|(t)|2 dt +
| (t)|2 dt.
226.844
2
| |2 + 2||||2 .
Recall that these are sums over the non-trivial zeros of L(s, ).
We first prove a general lemma on sums of values of functions on the non-trivial
zeros of L(s, ).
Lemma 6.2. Let f : R+ C be piecewise C 1 . Assume limt f (t)t log t = 0.
Then, for any y 1,
Z
X
qT
1
dT
f (T ) log
f (()) =
2 y
2
(6.11)
where
(6.12)
non-trivial
()>y
Z
1
+ O |f (y)|g (y) +
f (T ) g (T )dT ,
2
y
g (T ) = 0.5 log qT + 17.7
59
:()>y
1
=
2
Now, by [Ros41, Thms. 1719] and [McC84, Thm. 2.1] (see also [Tru, Thm. 1]),
T
qT
log
+ O (g (T ))
2e
for T 1, where g (T ) is as in (6.12). (This is a classical formula; the references
serve to prove the explicit form (6.12) for the error term g (T ).)
Thus, for y 1,
Z
X
T
qT
y
qy
1
log
log
f (T )
dT
f (()) =
2 y
2e
2e
:()>y
(6.14)
Z
1
f (T ) g (T )dT .
+ O |f (y)|g (y) +
2
y
(6.13)
N (T, ) =
Here
(6.15)
f (T )
T
qT
y
qy
log
log
2e
2e
1
dT =
2
f (T ) log
qT
dT.
2
= 0.5
since
g (T )
f (T )
dT,
T
is at most
p
p
60
H. A. HELFGOTT
s dt
(log t)(t)t1
|(log
t)(t)|t
=
|G (s)| =
(log t)(t)t
1
t
t
0
0
for s = + i .
Let us start by bounding the contribution of very low-lying zeros (|()| 1).
By (6.13) and (6.12),
N (1, ) =
1
q
log
+ O (0.5 log q + 17.7) = O (0.819 log q + 16.8).
2e
Therefore,
X
non-trivial
|()|1
|G ()| (t)t1/2 (0.819 log q + 16.8).
1
Let us now consider zeros with |()| > 1. Apply Lemma 6.2 with y = 1 and
(
|G (1/2 + it)| if t T0 ,
f (t) =
0
if t > T0 .
This gives us that
X
:1<|()|T0
(6.18)
T0
qT
dT
2
1
Z
+ O |f (1)|g (1) +
|f (T )| g (T ) dT ,
f (()) =
f (T ) log
Hence
1
T0
qT
f (T ) log
dT
2
s
qT0
log
2e
2
qT0
log
2e
+ 1 ||2
Again by Cauchy-Schwarz,
Z
|f (T )| g (T ) dT
1
2
|f (T )|2 dT
2
+1
T0
T0 .
T0 .
|g (T )|2 dT .
61
T0
Summing, we obtain
Z
Z
qT
1 T0
|f (T )| g (T ) dT
dT +
f (T ) log
1
2
1
p
1
log qT0
qT0
+
+ 17.21 |(t)(log t)|2
log
||2 +
T0
2e 2
2
Finally, by (6.16) and (6.12),
|f (1)|g (1) (t)/ t (0.5 log q + 17.7).
1
By (6.18) and the assumption that all non-trivial zeros with |()| T0 lie on
the line (s) = 1/2, we conclude that
X
p
|G ()| (||2 + | log |2 ) T0 log qT0
non-trivial
1<|()|T0
3
,
8
2 /2
tet
dt = 1
7
,
16
21/4 (1/4)
1.07791.
|(t)/ t|1 =
4
| |22 =
62
H. A. HELFGOTT
2
non-trivial
|()|>T0
|G ()| =
non-trivial
()>T0
C0,, e
| |
()2
+ C1, e
C0,, 2 (1 + 1.632 )
C1,
0.4798
min
2 25
min 81 (
) , 32 | |
+ C2, e
2
| | p
| |
2|| ,
3/2
3.251 min
+ C e 4 | | ,
2 2
, | |
3
!
3
1+ 2
| | 5 p
1+
| | 2 , C2, min
,
| | + 1,
| |
|| 4
e/2 | |3/2
2
and = 2.
For T0 100,
(6.20)
and so
!
1+ 2
e/2 ( 0.4798)| |
1+
+
e 4
1.025
| |
2
It is clear that this is a decreasing function of | | for | | > 3/(2 0.4798) (and
hence for | | 100).
We bound
(
E(1.5) 0.1598
if 1.5,
(6.21)
E() E(1.5)
1.5 0.1065 if < 1.5.
This holds for 1.5 because E() is increasing on , for 1.19 because of
(5.71), and for [1.19, 1.5] by the bisection method (with 20 iterations).
63
(6.22)
| |
()2
e 8 ( ) 0.055e0.1598 3 ( ) ,
(6.23)
and
2
0.1598 min 23 (
) ,| |
25
2
0.055 1
1
| |
| |
2 | | 2
+1
+
0.0039
0.055
||
4
2 4
3
p
1
5
5 1
+
| | 0.0075| |.
0.055
| | + 1 0.055
4
4 10 100
Hence
(6.24)
C0,, e
| |
()2
2 25
min 81 (
) , 32 | |
+ C2, e
C0,,
2
0.1598 min 32 (
) ,| |
where
2 2
= 3.26 min
, | | .
3
It is clear that the right side of (6.24) is a decreasing function of when
!
2 2
1
min
, | |
3
0.1598
C0,,
non-trivial
|()|>T0
f ( ) =
C0,,
non-trivial
()>T0
2
0.1598 min 23 (
) ,| |
f (()),
is at most
involving C0,,
Z
1
qT
1
3.26
log
+
T e0.1598T dT
(6.26)
2
2
4T
T0
is at most
If || > 4, the contribution of C0,,
Z
qT
1
1
log
+
T e0.1598T dT
3.26
(6.27)
2
4T
max(T0 ,()2 ) 2
64
H. A. HELFGOTT
1
y
1
2
ct2
+
dt
=
e
ecy ,
2 t2
2y
4c
2c
4c
y
y
Z
Z
at log et log et
a
2
2
ct2
2
dt
(t log t + c1 t) e
t log t +
2 ect dt
2c
2c
4c t
y
y
t2 +
where
a=
log ey
2c
y log ey
4c12 y
2c
c1 y +
c1 y + 4c21y2
1
+ y log ey
.
y
12
2c
4c y
2
2
2c|| 4c2 4
T0
T0
2c
+
a
log ||
+ a 0.1065 T0 2
1
||
||
+
e
2
4c2
and
4
1
1
8 + 40.10652 (4)2
+ 4 log 4e
a
0.088.
1
4
20.1065 40.10652 4
T0 2
0.1065 ||
times
Multiplying by 3.26||, we get that (6.28) is at most e
T0
eT0
q||
+ 2.44T0 log
+ 3.2|| log
(2.44T0 + 2.86||) log
2
||
||
!
(6.29)
1
1 + log T0 /||
qT0
qT0
T0 log
2.56T0 log
,
2.44 + 3.2
T0 /||
2
2
where we are using several times the assumption that T0 4 2 ||.
Let us now go back to (6.26). For any y 1, c, c1 > 0,
Z
1
y
ct
+
ecy ,
te dt =
c c2
y
Z
Z
c1 ct
a1
a
1
t log t +
e dt
t+
log t 2 ect dt
t
c
c c t
y
y
y
a cy
+
e
log y,
c c2
where
a=
log y
c1
1
c + c + y
log y
1
c c2 y
65
+ 2 +
+ 2 log T0 e0.1598T0
log
2
2
c
c
c
c
and
a
/2
log T0
1
0.1598 + 0.1598 + T0
log T0
1
0.1598 0.15982 T0
1.235.
(6.30)
3.5T0 log
X
b()x + O (err, (, x)) x if q = 1,
n (n/x) =
(n)(n)e
x
O (err, (, x)) x
if q > 1,
n=1
where
!
T02
qT0
0.1065
()2
3.5e0.1598T0 + 2.56e
err, (, x) = T0 log
2
p
p
+ 1.22 T0 log qT0 + 5.056 T0 + 1.423 log q + 38.19 x1/2
+ (log q + 6.01) (0.89 + 5.13||) x3/2 .
Proof. Immediate from Lemma 6.1, Lemma 6.3 and Lemma 6.4.
Now that we have Prop. 6.5, we can derive from it similar bounds for a smoothing defined as the multiplicative convolution of with something else just as
we discussed at the beginning of 4.3.
2
X
b ()x + O (err , (, x)) x if q = 1,
n (n/x) =
(n)(n)e
x
O (err , (, x)) x
if q > 1,
n=1
66
H. A. HELFGOTT
where
(6.33)
!
T2
qT0
0.1065 0 2
0.1598T0
()
3.5e
+ 0.0074 e
err, (, x) = T0 log
2
p
p
1
+ 1.22 T0 log qT0 + 5.056 T0 + 1.423 log q + 38.19 x 2
+ (log q + 6.01) (0.89 + 2.7||) x3/2 .
dw
n
n
2 (w)
(n)(n)e
x
wx
w
0 n=1
=
1X
1
4
(n)(n)e
n=1
wn
wx
dw
n
2 (w) ,
wx
w
since 2 is supported on [1/4, 1]. By Prop. 6.5, the main term (if q = 1)
contributes
Z
Z 1
dw
b(w)2 (w)dw
=x
b(w)xw 2 (w)
1
w
0
4
Z Z
Z Z
dt
r
e(r) 2 (w)dw
=x
(t)e(wt)dt2 (w)dw = x
w
w
0
0
Z Z
r
dw
=x
2 (w)
e(r)dr = b () x.
w
w
0
The error term is
Z 1
Z 1
dw
=x
err, (w, xw) wx 2 (w)
err, (w, xw)2 (w)dw.
(6.34)
1
1
w
4
4
R
R
Since w 2 (w)dw = 1, w w2 (w)dw = (3/4) log 2 and9
Z
0.1065(4)2 12 1
w
2 (w)dw 0.002866,
e
w
2 /2
67
1
Lemma 6.7. Let : R+
0 R be in C . Let x 1, R. Let be a primitive
character mod q, q 1.
Let H 3/2, T0 > H + 1. Assume that all non-trivial zeros of L(s, ) with
|()| T0 satisfy (s) = 1/2.
Write G (s) for the Mellin transform of (t)e(t). Assume that G is holomorphic on {s : 1/5 (s) 1 + } {s : 1/2 (s) 1/5, |(s)| T0 1}.
Then
X
X
n=1
(6.36)
+O
(2||||2 + | |2 ) x1/2 ,
H
where
Iq=1
(
1
=
0
if q = 1,
if q =
6 1,
H = T0 1
and
P the norms ||2 , | |2 are taken with respect to the usual measure dt. The sum
is a sum over all non-trivial zeros of L(s, ).
Proof. We start just as in the proof of Lem. 6.1, except we shift the integral only
up to (s) = 1/5 in the central interval, and push it up to (s) = 1/2 only in
the tails:
(6.37)
n=1
1
2i
L (s, )
L(s, )
G ()x
G (s)xs ds,
where I1, = 1 if q > 1 and (1) = 1 and I1, = 0 otherwise, and C is the
contour consisting of
(1) a segment C1 = [1/5 iH , 1/5 + iH ], where H = T0 1 > H,
(2) the union C2 of two horizontal segments:
1
1
1
1
C2 = iH , iH + iH , + iH ,
2
5
2
5
1
1
1
1
C3 = i, iH + iH , + i .
2
2
2
2
We can still use (6.3) to estimate L (s, )/L(s, ) (given L (1s, )/L(1s, )).
We estimate (z) for (z) 4/5 much as in (6.5): since |t2 + z 2 | 16/25 for t
68
H. A. HELFGOTT
real,
1
+ 2 O
(z) = log z
2z
1
2z
1
= log z
2z
= log z
tdt
16 2t
25 (e
1)
1
50
+
O
(2)(2)
16
(2)2
25
145
+O
= log z + O
192
192
0
e 10
i 2
10
i+ 2
i 2
10
s
+
e
e
1
+
2
=
cot = i
e 10 2 e 10 i+ 2
2
e 10 i 2 e 10 i+ 2
e 10 i + e 10
i
i
=
3.07769.
cot
e 10 e 10 i
10
(The inequality is clear for 0; the case < 0 follows by symmetry.) Similarly,
(s
+
1)
= cot 3 0.32492.
cot
2
5
e 2
2
2
e 2 H
We now must estimate L (s, )/L(s, ), where (s) 4/5 and |(s)| H .
By a lemma of Landaus (see, e.g., [MV07, Lemma 6.3], where the constants
are easily made explicit) based on the Borel-Caratheodory Lemma (as in [MV07,
Lemma 6.2]), any function f analytic and zero-free on a disc Cs0 ,R = {s : |ss0|
R} of radius R > 0 around s0 satisfies
f (s)
= O
f (s)
for all s with |s s0 | r, where 0 < r < R and M is the maximum of |f (z)|
on Cs0 ,R . We set s0 = 3/2 + i (where = (s)), r = 1/2 + 1/5 = 7/10, and
let R 1 , using the assumption that L(s, ) has no non-trivial zeroes off the
critical line with imaginary part H + 1 = T0 .
We obtain
maxsCs0 ,1 |L(s, )|
200
L (s, )
= O
log
.
(6.38)
L(s, )
9
|L(s0 , )|
Clearly,
|L(s0 , )|
Y
Y (1 p3 )1
(3)
(1 + p3/2 )1 =
0.46013.
=
3/2
1
(3/2)
(1 p
)
p
p
X
X
L(s, ) =
(m)ns
(m) (N + 1)s
nN
nN +1
mN
mn
!
N 11/2
1
= O 3 N + M (q)/ N ,
+N
+ M (q)N
=O
1 1/2
P
where M (q) = maxn mn (m). We set N = M (q)/3, and obtain
p
|L(s, )| 2M (q)N 1/2 = 2 3 M (q).
=O
= O 8 log
log q + 44.8596 .
L(s, )
(3)/(3/2)
9
Therefore, by (6.3) and (6.4),
(6.39)
L (s, )
145
= log log(1 s) + O
L(s, )
q
192
+ O (4 log q + 44.8596)
cot
+ log 2 + O
2
10
= log(1 s) + O (5 log q + 52.2871)
(6.40)
L (s, )
145
= log log(1 s) + O
L(s, )
q
192
H
coth
+ O (4 log q + 44.8596)
+ log 2 + O
2
2
= log(1 s) + O (5 log q + 49.1654).
L (s, )
log(1 s) + O (log q + 5.2844).
L(s, )
(6.41)
Let 1 j 3. By Cauchy-Schwarz,
Z
1
L (s, )
G (s)ds
2 Cj L(s, )
is at most
1
2
Cj
s Z
L (s, ) 1 2
1
L(s, ) s |ds| 2
Cj
|G (s)s|2 |ds|.
where c1,1 = 5, c1,2 = 5, c1,3 = 1, c2,1 = 52.2871, c2,2 = 49.1654, c2,3 = 5.2844.
69
70
H. A. HELFGOTT
For j = 1,
5 log q 2
2
1
2
s |ds| = (5 log q) 10 tan 5H 125(log q) ,
C1
Z H 1
Z
2 + 16 + c 2
| log |1 s|| + c2,j 2
log
2,j
2
25
|ds|
d 42949.3,
1
s
+
2
H
C1
25
Z
where we have computed the last integral numerically. As before, G (s)s is the
Mellin transform of (6.9). Hence
v
s Z
u Z 1 +i
u 1
5
1
2
|G (s)s| |ds| t
|G (s)s|2 |ds|
1
2 C1
2 i
5
sZ
2|||(t)t7/10 |2 + | (t)t7/10 |2 .
Hence,
Z
1
2
is at most
C1
L (s, )
G (s)ds
L(s, )
For j = 3,
Z
log q 2
|ds|
2(log q)2
|ds| = 2(log q)2
<
,
s
2
H
1/2+iH |s|
C3
2
Z 1
Z
9
2
| log |1 s|| + c2,j 2
2 log + 4 + c2,j
|ds| 2
d
1
2
s
H
C3
4 +
2
Z log + log 2 + c2,j
2
d
2
2
H
Z
2(log H + 6.71)2
.
H
C1
log q + log H + 6.71
L (s, )
G (s)ds
(2||||2 + | |2 ).
L(s, )
H
Lastly, for j = 2,
Z
L (s, )
1
|ds|
2 C2 L(s, )
1
2
+ 51
(log |1 s| + O (5 log q + c2,2 ))
.
10
71
qT0
0.1065 0 2
0.1598(T0 H)
()
.
3.44e
+ 0.63||e
|G ()| H log
2
non-trivial
|()|>T0
Proof. Clearly,
non-trivial
|()|>T0
|G ()| =
non-trivial
()>T0
We now proceed much as in the proof of Lem. 6.4. By Prop. 5.1, given s =
+ i , [0, 1], 4 2 max(1, ||),
|F ()| + |F (1 )|
is at most
E
C0,, e
| |
()2
0.4798
+ C1, e
C0,, 2 (1 + 7.831 )
C2, min
1
4
| |
||
2
3/2
2
2 25
min 81 (
) , 32 | |
+ C2, e
1
4.217,
!1
25
,
, | |
16
For T0 H 100,
!
1
1+ 2
e/2 1/2
2 +
1+
2
and so
C1, =
e/2 1/2
2
2 3/2 ||
1+ p
| |
+ C e 4 | | ,
!
1
1+ 2
2,
1+
!
2 3/2 ||
1+ p
.
| |
e( 4 0.4798)| |
72
H. A. HELFGOTT
Thus,
E
C0,, e
| |
()2
2 25
min 81 (
) , 32 | |
+ C2, e
2
0.1598 min 23 (
) ,| |
4.22e
g( ) = 4.22e
+ 0.033| |e0.4798| |
|M hH (ir)|dr 1.99301
H
.
2
f ( ) = 0.7951 H g( H)
is decreasing for T0 (because g( ) is decreasing for T0 H).
We apply Lemma 6.2, and get that
Z
X
qT
1
1
log
+
dT
|G ()|
f (T )
2
2
4T
T0
non-trivial
|()|>T0
= 0.7951 H
T0
g(T H)
qT
1
1
log
+
2
2
4T
dT.
We continue as in the proof of Lemma 6.4, only the integrals are somewhat
simpler this time. For any c > 0,
Z
qT
1
1
c(T H)
e
log
+
dT
2
2
4T
T0
(6.43)
1
1
qT0
1
1
log
+
+
ec(T0 H) .
2
2c
2
2c
4c T0
At the same time,
Z
H 2
1
qT0
1
23 c( T
)
log
+
dT
e
2
2
4T
T0
||
qT0
Z
H 2
2 log 2 +
||
q||t
1
23 c(t
)
=
log
+
e
dt
4 T0 H
T0
2
2
4t
3 c ||
||
4T0
||
ec1 (T H) ,
23 c
T0 H
2
73
we see that
Z
Te
T0
(6.44)
qT
1
1
log
+
dT
2
2
4T
!
1 + cT10
qT0 1
1
ec1 (T0 H)
+
T0 log
+
2c
2
c
4c
!
1 + c11T0
1
qT0
1
+
T0 log
+
ec1 (T0 H) .
2c1
2
c1
4c1
c1 (T H)
c c2
,
c1 c21
1
e(c1 c0 )(T0 H)
T0 +
c1
c
1
3(T0 H) +
e(c1 c0 )(T0 H) 1.3 1012 .
c1
c1
1
1
1 2c
1
2 + 4c
+
0
2c T0 log qT
2
1
2
0
log qT
2 +
4 T0 H
3 c ||
4T0
3.35533
1
2
0.62992 log
3.4303
4T0 log
4
3c
T0
2
qT0
2
qT0
.
2
We conclude that
Z
qT
1
1
log
+
dT
0.7951
g(T H)
2
2
4T
T0
qT0
log
log
0.1065
3.44e0.1598(T0 H) + 0.63||e
(T0 H)2
()2
.
X
c
+ ()x + O err+ , (, x) x if q = 1,
n + (n/x) =
(n)(n)e
err
x
O
(,
x)
x
if q > 1,
,
+
n=1
74
H. A. HELFGOTT
where
(6.46)
!
(T H)2
qT0
0.1065 0 2
0.1598(T0 H)
()
3.44e
+ 0.63||e
err+ , (, x) = H log
2
p
1
+O (((0.641 + 1.11 H) log qT0 + (1.5 + 19.1 H)) T0 + 1.65 log q + 44)x 2
4
Proof. We apply Lemmas 6.7, Lemma 6.3 and Lemma 6.8. We bound the norms
involving + using the estimates in 4.2.2. The error terms in (6.36) total at most
(6.47)
[ 12 , 15 ]
Since x 103 , the last line of (6.47) is easily absorbed into the first line of (6.47)
(by a change in the last significant digits). Much as in the proof of Lem. 6.8, we
bound
Z
1
|G ( + i(T 1))|
|M hH (ir)|dr max |F ( + i )|
T 1H
2
dt
| (t)t+1 |1 + 2||| (t)t |1
t
0
+1
+1
/2
2
+1 +2
=2
.
||
2
2
|sF (s)|
|t + 2i| (t)t+1
Since [1/2, 1/5] and x 103 , a quick verification gives that x 2/2 (/2+1)
and x 2(+1)/2 || (( + 1)/2) are maximal when = 1/5. Hence
|F (s)|x
and so
(6.48)
+ 0.061 H (0.23 log T + 1.12 log q + 11.04) x1/5
75
Finally, let us prove a simple result that will allow us to compute a key 2
norm.
Proposition 6.10. Let = + be as in (6.35), H 50. Let x 106 . Assume
that all non-trivial zeros of the Riemann zeta function (s) with |()| T0 lie
on the critical line, where T0 2H + max(H, 100).
Then
Z
X
2
2
+
(t) log xt dt + O (err2 ,+ ) x log x,
(n)(log n)+
(n/x) = x
(6.49)
0
n=1
where
err2 ,+
(6.50)
p
(log T0 )2
+ 0.224 log T0 H T0 e(T0 2H)/4
= 0.311
log x
p
Proof. We will need to consider two smoothing functions, namely, +,0 (t) =
+ (t)2 and +,1 = + (t)2 log t. Clearly,
2
(n)(log n)+
(n/x) = (log x)
(n)+,0 (n/x) +
Since + (t) = hH
2
(t)et /2 ,
(n)+,1 (n/x).
n=1
n=1
n=1
where
1
Fx (s) = (log x)(s) + (s).
2
(6.52)
Moreover,
(6.53)
M h2H (ir)
1
=
2
and so M h2H (ir) is supported on [2H, 2H]. We also see that |M h2H (ir)|1
|M hH (ir)|21 /2. We know that |M hH (ir)|21 /2 1.993012 H by (4.24).
Hence
(6.54)
Z
1
|M (h2H )(ir)|dr max |Fx ((s ir)/2)|
|M +,0 (s)|
4
|r|2H
1.993012
H max |Fx ((s ir)/2)| 0.3161H max |Fx ((s ir)/2)|.
4
|r|2H
|r|2H
(6.55)
|(s)| 2|s|(s)1/2 e|(s)|/2 e1/6|z| ,
and so
(6.56)
p
|(s)| 2.511 |(s)|e|(s)|/2
76
H. A. HELFGOTT
= log s
+O
(s)
2s
cos3 /2
for | arg(s)| < ( (, )). Again, s = + i with 1 1 and | | 99,
this gives us
(s)
= log | | + O (3.02).
(s)
Hence, under the same conditions on s,
1
log | | + 1.51)(s)
2
p
1
2.511((log x) + log | | + 1.51) | |e| |/2 .
2
(6.57)
Thus, by (6.54),
r
1
| |
He(| |2H)/4
(6.58) |M +,2 ()| 0.7938H((log x) + log | | + 1.17)
2
2
for = + i with | | T0 1 2H + 99 and 1 1.
We now apply Lemma 6.7 with = +,2 , = 0 and trivial. We obtain that
(6.59)
X
(n)+,2 (n/x)
n=1
Z
+ O
+ O
+,2 (t)dt x
(t)t7/10 |2 x1/5
M +,2 ()x + O 82.7|+,2
7 log T0
+ 11.04 max |M +,2 ( + i(T0 1))| x
10
[ 12 , 15 ]
!
log T0 + 6.71
p
|+,2 |2 x1/2
(T0 1)
2 , we can bound
Since +,2 = (log xt)+
2 3/10
|+,2
(t)t7/10 |2 |2+ +
(log xt)t7/10 |2 + |+
t
|2
7/10
2(|+ | log x + |+ (t) log t| )|+
t
|2 + |+ | |+ t3/10 |2
by (4.31), (4.32), (4.22) and (4.19), using the assumption H 50. Similarly,
|+,2
|2 2(|+ | log x + |+ (t) log t| )|+
|2 + |+ | |+ |2
77
|+,2 |2 x(1/2+1/5)
82.7|+,2
(t)t7/10 |2 x1/5 + p
(T0 1)
T0
He(T0 12H)/4 .
2
for 1 1 and x 106 . Thus, the error terms in (6.59) total at most
+ 17.21|+,2 log |2
T0 + 34.5 +,2 (t)/ t
|+ log |2 1.1096 H, |(t)/ t|1 1.3267 from (4.31), (4.32) and (4.17) (with
H 50) and (4.26). (We also use the assumption x 106 .) Since T0 200, this
means that
X
p
78
H. A. HELFGOTT
To bound the contribution of the zeros beyond T0 , we apply Lemma 6.2, and
get that
Z
X
T
1
1
log
+
dT,
|M +,2 ()|
(6.61)
f (T )
2
2 4T
T0
non-trivial
|()|>T0
where
r
1
| | (| |2H)/4
f (T ) = 1.5876H ((log x) + log | | + 1.17)
e
2
2
(see (6.58)). Since T T0 200, we know that ((1/2) log(T /2) + 1/4T )
0.216 log T . In general,
Z
T0
4 p
2/
2
2
2 T /4
2
T (log T ) e
dT
T0 (log T0 ) + (log e T0 ) e 4 ,
T0
T
p
Z0
T0
4
2/
T /4
2
T (log T )e
dT
T0 (log T0 ) + log e T0 e 4 ,
T0
T0
p
Z
T
4
2/
0
T eT /4 dT =
T0 +
e 4 ;
T
0
T0
T0 (log T0 )2 eT0 /4 ,
for T0 200, the quantities on the right
are
at
most
1.281
1.279 T0 (log T0 )eT0 /4 , and 1.278 T0 eT0 /4 , respectively. Thus, (6.61) gives
us that
X
|M +,2 ()| 1.281 (0.216 log T0 ) f (T0 )
non-trivial
|()|>T0
T0 e(T0 2H)/4 .
q | err, (, x)|
(6.62)
E,r,0 =
max
mod q
qgcd(q,2)r
||gcd(q,2)0 r/2q
79
5 107
.
q/ gcd(q, 2)
Thus
T0
1000
5 107
=
,
150000
3
5 107
1000
T0
=
= 26.525823 . . .
4r
12
and so
0.1065
T02
()2
2.575 1023 .
8
8
8
+ 1.22 10 log 10 + 5.056 10 + 1.423 300000 log 300000 + 38.19 300000
1 +
3/2
x 2 + (log 300000 + 6.01) (0.89 300000 + 2.7 4 300000) x
3.75 107
,
q/ gcd(q, 2)
H = 200.
Thus
T0 H
3.75 107
= 250,
150000
3.75 107
750
T0 H
=
= 19.89436 . . .
4r
12
and also
qT0 2r H + 7.5 107 1.35 108 .
Hence
3.44
0.1598(T0 H)
2re
0.1065
+ 0.63 4r e
(T0 H)2
()2
1.953 1013 .
80
H. A. HELFGOTT
1
+ (16.339 log 1.35 108 + 271.62) 1.35 108 + 64.81 2r x+ 2
4
We record our final conclusions: for E,r,0 defined as in (6.62) and r = 150000,
E ,r,8 3.061 108 ,
(6.63)
t.
4r
Hence
0.1598(T0 H)
3.44e
0.1065
+ 0.63||e
(T0 H)2
()2
101300000 + 156000e0.1065t .
4/5
+ O 2.52 108 + 200(0.015 log T0 + 163) x1 .
ET+ ,4r
We choose t = 20; this gives T0 3.77 107 , which is certainly within the checked
range. We obtain
(6.64)
(log 600)2
+ 0.224 log 600 200 600e50
err2 ,+ 0.311
log x1
(6.65)
1/2
4/5
+ 419.3 200x1
+ (6.2 200 + 5.3) 600 log 600 x1
4.8 1065 + 2.17 1011 + 1.13 1020 2.2 1011 .
81
where the
were computed rigorously using VNODE-LP [Ned06]. (The
R integrals
2
integral 0 (t)dt can also be computed symbolically.) By Cauchy-Schwarz and
the triangle inequality,
Z
(+ (t) (t)) (t) log xt dt |+ |2 | (t) log xt|2
0
|+ |2 (| |2 log x + | log |2 )
547.56
(0.80013 log x + 0.04574)
H 7/2
3.873 106 log x + 2.214 107 ,
where we are using (4.14) and evaluate | log |2 rigorously as above. (We are
also using the assumption x x1 to bound 1/ log x.) By (4.14) and (4.15),
Z
547.56
480.394
(+ (t) (t))2 log xt dt
log x +
7/2
H
H 7/2
0
6
4.8398 10 log x + 4.25 106 .
We conclude that
(6.66)
Z
2
+
(t) log xt dt
We add to this the error term 2.2 1011 log x from (6.65), and simplify using the
assumption x x1 . We obtain:
(6.67)
n=1
2
(n/x) = (0.6402 + O (2 105 ))x log x 0.0211x.
(n)(log n)+
82
H. A. HELFGOTT
Heath-Brown to the present author, who communicated it to Tao ([Tao, Lem. 4.6]
and adjoining comments). We will be able to do better than that estimate here.
The role played by Montgomerys inequality in Heath-Browns method is played
here by a result of Ramares ([Ram09, Thm. 2.1]; see also [Ram09, Thm. 5.2]).
The following Proposition is based on Ramares result, or rather on one possible
proof of it. Instead of using the result as stated in [Ram09], we will actually be
using elements of the proof of [Bom74, Thm. 7A], credited to Selberg. Simply
integrating Ramares inequality would give a non-trivial if slightly worse bound.
on the primes. Assume
Proposition 7.1. Let {an }
n=1 , an C, be supported
x.
Let
Q0 1, 0 1 be such
that {an } is in L1 L2 and
that
a
=
0
for
n
n
p
2
that 0 Q0 x/2; set Q = x/20 Q0 . Let
(7.1)
M=
[ a 0 r a 0 r
, +
.
q
qx q
qx
qr a mod q
(a,q)=1
Let S() =
n an e(n)
|S()|2 d
max max
qQ0 sQ0 /q
Gq (Q0 /sq)
Gq (Q/sq)
X
n
|an |2 ,
where
(7.2)
X 2 (r)
.
(r)
Gq (R) =
rR
(r,q)=1
Proof. By (7.1),
(7.3)
|S()| d =
X Z
qQ0
0 Q0
qx
0 Q0
qx
2
X a
d.
S
+
a mod q
(a,q)=1
a mod q
(a,q)=1
X
q |q
(q ,q/q )=1
2 (q/q )=1
2
X X
an (n)
mod q
83
for every q x, where we use the assumption that n is prime and > x (and
thus coprime to q) when an =
6 0. Hence
2
Z 0 Q0
Z
X
X
qx
1 X
2
an e(n)(n) d
q
|S()| d =
0 Q0 (q)
n
qQ0
q
(q )
q
(q )
q Q0
q Q0
q |q
(q ,q/q )=1
2 (q/q )=1
2 (r)
(r)
rQ0 /q
(r,q )=1
0 Q0
q x
0 Q0
q x
qx
2
X X
an e(n)(n) d
0 Q0
q rx
Q
q0 rx0 mod q
0
Q
r q0 min 1, ||x
2 (r)
(r)
2
X X
an e(n)(n) d
mod q
(r,q )=1
q Q0
q
(q )
0 Q0
q x
0 Q0
q x
0
r qQ min 1, ||x
2 (r)
(r)
mod q
(r,q )=1
2
X X
a
e(n)(n)
d
n
2
0 Q
X X
X 2 (r) Z qrx
q
an e(n)(n) d.
(q)
(r) 0 Q
n
rQ/q
(r,q)=1
qrx
mod q
qr
X
(b)e2in qr
b=1
(b,qr)=1
for primitive of modulus q. Here cr (n) stands for the Ramanujan sum
X
cr (n) =
e2nu/r .
u mod r
(u,r)=1
r x coprime to q,
2
2 qr
X
X
b
.
(b)S
an e(n)(n) =
q
n
qr
b=1
(b,qr)=1
q. Hence, for
84
H. A. HELFGOTT
Thus,
=
X X 2 (r)
(rq)
qQ rQ/q
(r,q)=1
1
(q)
XZ
0 Q
qx
0 Q
qx
0 Q
qx
0 Q
qx
2
qr
X X
b
(b)S
d
qr
mod q b=1
(b,qr)=1
0 Q
qrx
Q
0
qrx
2
q
X X
b
(b)S
d
q
mod q b=1
(b,q)=1
2
q
X
S b d.
q
b=1
(b,q)=1
Let us now p
check that the intervals (b/q 0 Q/qx, b/q + 0 Q/qx) do not overlap.
Since Q = x/20 , we see that 0 Q/qx = 1/2qQ. The difference between two
distinct fractions b/q, b /q is at least 1/qq . For q, q Q, 1/qq 1/2qQ +
1/2Qq . Hence the intervals around b/q and b /q do not overlap. We conclude
that
2 X
Z
S b =
|an |2 ,
q
R/Z
n
and so, by (7.4), we are done.
We will actually use Prop. 7.1 in the slightly modified form given by the following statement.
on the primes. Assume
Proposition 7.2. Let {an }
n=1 , an C, be supported
x.
Let
Q0 1, 0 1 be such
n
p
2
that 0 Q0 x/2;
P set Q = x/20 Q0 . Let M = M0 ,Q0 be as in (3.5).
Let S() = n an e(n) for R/Z. Then
Z
Gq (2Q0 /sq) X
q2Q
G
(2Q/sq)
s2Q
/q
0
0
q
M ,Q
n
0
q even
where
(7.5)
X 2 (r)
.
(r)
Gq (R) =
rR
(r,q)=1
Proof. By (3.5),
Z
|S()| d =
X Z
qQ0
q odd
0 Q0
2qx
0 Q0
2qx
X Z
0 Q0
qx
qQ0
q even
0 Q0
qx
2
X a
d
S
+
a mod q
(a,q)=1
2
X a
S
+ d.
q
a mod q
(a,q)=1
85
R
We proceed as in the proof of Prop. 7.1. We still have (7.3). Hence M |S()|2 d
equals
2
Q
X q Z 2q0 x0
X
2 (r) X X
a
e(n)(n)
d
n
)
0 Q0
(q
(r)
Q
q Q0
q odd
q 2Q0
q even
2q x
mod q
0
r q0 min 1, 2||x
(r,2q )=1
q
(q )
0 Q0
q x
0 Q0
q x
2Q0
q
0
min 1, 2||x
2 (r)
(r)
2
X X
an e(n)(n) d.
mod q
(r,q )=1
(The sum with q odd and r even is equal to the first sum; hence the factor of 2
in front.) Therefore,
max
|S()|2 d max
21
(7.6)
q odd
+ max
max
q 2Q0 s2Q0 /q
q even
where
1 =
qQ
q odd
qQ
q odd
2 =
q2Q
q even
q
(q)
rQ/q
(r,2q)=1
2 (r)
(r)
0
2qrx
mod q
qrx
mod q
0
X 2 (r) Z qrx
q
(q)
(r) 0 Q
Q
r2Q/q
(r,q)=1
qrx
2
X X
an e(n)(n) d.
r2Q/q
(r,q)=1
r even
Gq (2Q0 /sq )
2 ,
Gq (2Q/sq )
2
X X
an e(n)(n) d
0 Q
2qrx
0
X 2 (r) Z qrx
q
(q)
(r) 0 Q
2
X X
an e(n)(n) d.
mod q
1 + 2
q2Q
q even
2qx
1
(q)
0 Q
qx
0 Q
qx
b=1
(b,q)=1
2
q
X
S b d.
q
b=1
(b,q)=1
Let us now check that the intervals of integration (b/q 0 Q/2qx, b/q + 0 Q/2qx)
(for q odd), (b/q 0 Q/qx, b/q + 0 Q/qx) (for q even) do not overlap. Recall
that 0 Q/qx = 1/2qQ. The absolute value of the difference between two distinct
fractions b/q, b /q is at least 1/qq . For q, q Q odd, this is larger than 1/4qQ +
86
H. A. HELFGOTT
1/4Qq , and so the intervals do not overlap. For q Q odd and q 2Q even (or
vice versa), 1/qq 1/4qQ + 1/2Qq , and so, again the intervals do not overlap.
If q Q and q Q are both even, then |b/q b /q | is actually 2/qq . Clearly,
2/qq 1/2qQ + 1/2Qq , and so again there is no overlap. We conclude that
2 X
Z
S b =
21 + 2
|an |2 .
q
R/Z
n
7.2. Bounding the quotient in the large sieve for primes. The estimate
given by Proposition 7.1 involves the quotient
Gq (Q0 /sq)
,
(7.7)
max max
qQ0 sQ0 /q Gq (Q/sq)
where Gq is as in (7.2). The appearance of such a quotient (at least for s = 1) is
typical of Ramares version of the large sieve for primes; see, e.g., [Ram09]. We
will see how to bound such a quotient in a way that is essentially optimal, not
just asymptotically, but also in the ranges that are most relevant to us. (This
includes, for example, Q0 106 , Q 1015 .)
As the present work shows, Ramares work gives bounds that are, in some
contexts, better than those of other large sieves for primes by a constant factor
(approaching e = 1.78107 . . . ). Thus, giving a fully explicit and nearly optimal
bound for (7.7) is a task of clear general relevance, besides being needed for our
main goal.
We will obtain bounds for Gq (Q0 /sq)/Gq (Q/sq) when Q0 2 1010 , Q Q20 .
As we shall see, our bounds will be best when s = q = 1 or, sometimes, when
s = 1 and q = 2 instead.
P
Write G(R) for G1 (R) = rR 2 (r)/(r). We will need several estimates for
Gq (R) and G(R). As stated in [Ram95, Lemma 3.4],
(7.8)
G(R) =
(r)
r
p
rR
rR
p|r
(7.10)
X1
X 2 (r) Y X 1
> log R.
=
r
pj
r
rR
p|r j1
rR
X
log p
(d)
log R + cE +
+ O 7.284R1/3 f1 (d)
(7.12)
Gd (R) =
d
p
p|d
f1 (d) =
Y
p|d
and
(7.14)
p1/3 + p2/3
(1 + p2/3 ) 1 +
p(p 1)
cE = +
X
p2
by [RS62, (2.11)].
If R 182, then
(7.15)
87
!1
log p
= 1.3325822 . . .
p(p 1)
where the upper bound is valid for R 120. This is true by (7.12) for R 4 107 ;
we check (7.15) for 120 R 4 107 by a numerical computation.10 Similarly,
for R 200,
log R + 1.661
log R + 1.698
(7.16)
G2 (R)
2
2
by (7.12) for R 1.6108 , and by a numerical computation for 200 R 1.6108 .
Write = (log Q0 )/(log Q) 1. We obtain immediately from (7.16) that
(7.17)
log Q0 + 1.661
G2 (Q0 )
G2 (Q)
log Q + 1.698
for Q, Q0 200.
Let us start by giving an easy bound, off from the truth by a factor of about
e (like some other versions of the large sieve). First, we need a simple explicit
lemma.
Lemma 7.3. Let m 1, q 1. Then
Y
p
(7.18)
e (log(m + log q) + 0.65771).
p1
p|qpm
Q
Proof. Let P = pmp|q p. Then, by [RS75, (5.1)],
P
Y
Pq
p = qe pm log p qe(1+0 )m ,
pm
e log((1 + 0 )m + log q) +
(P)
log(m + log q)
2.50637/e
e log(m + log q) + 0 +
.
log(m + log q)
Thus (7.18) holds when m + log q 8.53, since then 0 + (2.50637/e )/ log(m +
log q) 0.65771. We verify all choices of m, q 1 with m + log q 8.53 computationally; the worst case is that of m = 1, q = 6, which give the value 0.65771
in (7.18).
10Using D. Platts implementation [Pla11] of double-precision interval arithmetic based on
Lambovs [Lam08] ideas.
88
H. A. HELFGOTT
.
Q
Gq (Q/sq)
log Q0 + 1.31
log QQ0 + 1.31
Q
Proof. Let P = pQ0/sqp|q p. Then
and so
(7.19)
.
Gq (Q/sq)
GP (Q/Q0 )
(P)
G(Q/Q0 ).
P
By Lem. 7.3,
P
Q0
e log
+ log q + 0.658 .
(P)
sq
Hence, using (7.9), we get that
log Q0 + log q + 1.172
e
sq
Gq (Q0 /sq)
P/(P)
(7.20)
,
Gq (Q/sq)
G(Q/Q0 )
log Q + 1.31
Q0
We will use Lemma 7.4 when Q0 > 2 1010 , since then the numerical bounds
we will derive are not available. As we will now see, we can also use Lemma 7.4
to obtain a bound that is useful when sq is large compared to Q0 , even when
Q0 2 1010 .
Lemma 7.5. Let Q0 1, Q 200Q0 . Let q Q0 , s Q0 /q, q an even integer.
Let = (log Q0 )/ log Q 2/3. If
2Q0
(1)e
log q,
1.1617 Q0
sq
then
(7.21)
Gq (2Q0 /sq)
log Q0 + 1.698
.
Gq (2Q/sq)
log Q
Proof. Apply Lemma 7.4. By (7.20), we see that (7.21) will hold provided that
log QQ0 + 1.31
2Q0
e log
+ log q + 1.172
(log Q0 + 1.698)
sq
log Q
(7.22)
(log Q0 + 1.698)(log Q0 1.31)
.
log Q0 + 1.698
log Q
89
log Q0 + 1.698
e log
+ log q + 1.172 (1 ) log Q0 + 1.698 + 1.31 1.698.
sq
For all [0, 2/3],
2
1.698 + 1.31 1.698 1.172 0.526 0.388 0.267.
3
Hence it is enough that
2Q0
(1)e
+ log q ee ((1) log Q0 +0.267) = c Q0
,
sq
where c = exp(exp() 0.153) = 1.16172 . . . .
.
Gq (2Q/sq)
log Q
(7.23)
Gq (R) =
(q)
log R + cE +
q
X log p
p|q
+ errq,R .
log Q0 + 1.698
sq
1.661)
.
2
log Q
Note that
log Q0 + 1.698
1.698
1.698
=+
= 1+
.
log Q
log Q
log Q0
Thus, (7.25) will be true provided that
1.698
sq
1.661) + (1.698 cE )
log sq 1 +
(log
log Q0
2
X log p
q
+
errq, 2Q0
p
(q)
sq
p|q
90
H. A. HELFGOTT
p
(q)
sq
p|q
where Q0,min = 105 . Since Q0,min = 105 and 1 0.87, we know that 1 (1 +
1.698/ log Q0,min ) < 1.14751 < 1. Now all that remains to do is to take the
maximum mq,R,1 of errq,R over all R satisfying
R > 1.1617 max(Rq, Q0,min )(11 )e
(7.27)
log q
(since all smaller R are covered by Lemma 7.5) and verify that
(7.28)
mq,R,1
(q)
(q)
q
for
(q) = (1 1.14751 ) log q + 2.701351 + 0.365
X log p
p|q
(Values of s larger than 1 make the conditions stricter and the conclusions weaker,
thus making the task only easier.)
Ramares bound (7.12) implies that
errq,R 7.284R1/3 f1 (q),
The first question is: for which q does (7.27) imply (7.29)? It is easy to see
that (7.27) implies
R1(11 )e
log q
R(11 )e
log q
(1.1617q (11 )e
log q)
1
1(11 )e
1
1(11 )e
Thus, if (q) is greater than the quantity (q) given in (7.29), we are done.
and p (log p)/p are decreasing functions of p for
Now, (p/(p 1)) f1 (p) Q
p 3. Hence, for even q < pp0 p, p0 a prime,
X log p
(7.31)
(q) (1 1.14751 ) log q + 2.271
p
p<p
0
and
(7.32)
(q)
p<p0
Q
7.284 p<p0 f1 (p)
p
P
p 1 (1 1.14751 ) log q + 2.701351 + 0.365 p<p
0
log p
p
!3
51.273
3
66.508
for q <
for q <
for q <
Q
Q
Q
p23 p
= 223092870,
p29 p
= 6469693230,
p31 p
= 200560490130.
91
and so
(q)
1.1617q (10.55)e
1
298.03
1
1(10.55)e
1
1(10.55)e
1.222q 0.338 .
1.1616q (10.55)e
By [RS62, (3.14), (3.24), (3.30)] and a numerical computation for p1 < 106 ,
X log p
Y
X
< log p1 , log
p =
log p > 0.8p1 ,
p
pp1
pp1
pp1
Y p
1
and
log f1
pp1
Hence
p =
log(1 + p
2/3
) + log
p31
pp1
1/3
0.729p1
7.284 0.1489e0.729p1
1.94 log p1
0.3688 0.8p1 log p1
2.025e0.2217p1
p1/3 + p2/3
1+
p(p 1)
!1
+ log 0.1489.
1/3
(q)
!3
1/3 3
1.265e0.729p1
92
H. A. HELFGOTT
Gq (R) depends only on R, we can tell from (7.24) that, since we are taking the
maximum of errq,R , it is enough to check integer values of R. We check all integers
R in [(q), (q)) for all even q < 3.59 107 by an explicit computation.11
Finally, we have the trivial bound
Gq (Q0 /sq)
1,
Gq (Q/sq)
(7.33)
5
{an } is in L1 L2 and that an =
p0 for n x. Let Q0 10 , 0 1 be such
2
that (200Q0 ) x/20 ; set Q = x/20 and = (log Q0 )/ log Q. Let M0 ,Q0 be
as in (3.5). P
Let S() = n an e(n) for R/Z. Then, if 0.55,
Z
Z
log Q0 + 1.698
2
|S()|2 d.
|S()| d
log Q
R/Z
M ,Q
0
P
Here, of course, R/Z |S()|2 d = n |an |2 (Plancherel). If > 0.55, we will
use the trivial bound
Z
Z
|S()|2 d
|S()|2 d
(7.34)
R/Z
M0 ,r
Proof. Immediate from Prop. 7.2, Lem. 7.4 and Prop. 7.6.
7.3. Putting together 2 bounds over arcs and bounds. First, we need a
simple lemma essentially a way to obtain upper bounds by means of summation
by parts.
+
Lemma 7.8. Let f, g : {a, a + 1, . . . , b} R+
0 , where a, b Z . Assume that,
for all x [a, b],
X
(7.35)
f (n) F (x),
anx
n=a
b
X
n=a
Pn
m=a f (m).
nu
b1
X
n=a
11Here, as elsewhere in this section, numerical computations were carried out by the author
93
Let h(x) = maxxnb g(n). Then h is non-increasing. Hence (7.35) and (7.36)
imply that
b
X
n=a
f (n)g(n)
b
X
f (n)h(n)
n=a
S(b)h(b) +
F (b)h(b) +
P
b1
X
n=a
b1
X
n=a
anx n
x
A(u)F (u)du.
(Abel summation)
anx n
= h(a)h(x+
n=a
= (h(a) h(b))F (b 1)
b1
b1
b1
n=a
h(u)F (u)du
h(u)F (u)du,
h(u)F (u)du.
We will now seeRour main application of Lemma 7.8. We have to bound an integral of the form M ,r |S1 ()|2 |S2 ()|d, where M0 ,r is a union of arcs defined
0
R
as in (3.5). Our inputs are (a) a bound on integrals of the form M ,r |S1 ()|2 d,
0
(b) a bound on |S2 ()| for (R/Z) \ M0 ,r . The input of type (a) is what we
derived in 7.1 and 7.2; the input of type (b) is a minor-arcs bound, and as such
is the main subject of [Hel].
P
1
Proposition 7.9. Let S1 () =
n an e(n), an C, {an } in L . Let S2 :
R/Z C be continuous. Define M0 ,r as in (3.5).
Let r0 be a positive integer not greater than r1 . Let H : [r0 , r1 ] R+ be a nondecreasing continuous function, continuous and differentiable almost everywhere,
94
H. A. HELFGOTT
such that
P
(7.37)
1
|an |2
M0 ,r+1
for some 0 x/2r12 and all r [r0 , r1 ]. Assume, moreover, that H(r1 ) = 1. Let
g : [r0 , r1 ] R+ be a non-increasing function such that
(7.38)
max
(R/Z)\M0 ,r
where
1
I0 = P
2
n |an |
(7.40)
M0 ,r0
|S1 ()|2 d.
The condition 0 x/2r12 is there just to ensure that the arcs in the definition
of M0 ,r do not overlap for r r1 .
Proof. For r0 r < r1 , let
Let
1
f (r) = P
2
n |an |
Then, by (7.38),
1
f (r1 ) = P
2
n |an |
1
2
n |an |
(R/Z)\M0 ,r0
M0 ,r+1 \M0 ,r
(R/Z)\M0 ,r1
|S1 ()|2 d.
|S1 ()|2 d.
r1
X
f (r)g(r).
r=r0
By (7.37),
(7.41)
r0 rx
1
f (r) = P
2
n |an |
=
1
P
2
n |an |
M0 ,x+1
0 ,r0
r0 rr1
1
P
2
n |an |
R/Z
|S1 ()|2 d
!
|S1 ()|2 d
I0 H(x) I0
|S1 ()|2
2
|S1 ()|
I0 = 1 I0 = H(r1 ) I0 .
95
r=r0
ru
r0
g(r0 )(H(r0 ) I0 ) +
r1
r0
Theorem 7.10 (Total of minor arcs). Let x 1024 , where 4 1750. Let
X
(7.42)
S (, x) =
(n)e(n)(n/x).
n
Then
Zr 0
where
S=
(7.43)
p> x
||1 x
(M + T ) +
S (0, x) E
!2
2
(log p)2 +
(n/x),
E = (C+ ,0 + C+ ,2 ) log x + (2C+ ,0 + C+ ,1 ) x1/2 ,
C+ ,0 = 0.7131
(7.44)
and
(7.45)
C+ ,1 = 0.7131
1
(sup + (r))2 dt,
t rt
log t
(sup + (r))2 dt,
t rt
C+ ,2 = 0.51941|+ |2 ,
Z
1.04488 1/K
|(w)|dw
C,3 (K) =
||1
0
96
H. A. HELFGOTT
Proof. Let y = x/. Let Q = (3/4)y 2/3 , as in [Hel, Main Thm.] (applied with
y instead of x). Let (R/Z) \ M8,r , where r r0 and y is used instead of x
to define M8,r (see (3.5)). There exists an approximation 2 = a/q + /y with
q Q, ||/y 1/qQ. Thus, = a /q + /2y, where either a /q = a/2q or
a /q = (a + q)/2q holds. (In particular, if q is odd, then q = q; if q is even,
then q may be q or 2q.)
There are three cases:
(1) q r. Then either (a) q is odd and q r or (b) q is even and q 2r.
Since is not in M8,r , then, by definition (3.5), ||/y 0 r/q = 8r/q. In
particular, || 8.
Thus, by Prop. 4.5 and Lemma 4.6,
||
q ||1 y gy, (r) ||1 y.
(7.46)
|S (, x)| = |S2 M (, y)| gy,
8
(2) r < q y 1/3 /6. Then, by Prop. 4.5 and Lemma 4.6,
(7.47)
||
|S (, x)| = |S2 M (, y)| gy, max
, 1 q ||1 y gy, (r) ||1 y.
8
(3) q > y 1/3 /6. Again by Prop. 4.5,
(7.48)
Let
y
|S (, x)| = |S2 M (, y)| h
||1 + C,3 (K) y,
K
where h(x) is as in (4.42). (Note that C,3 (K), as in (7.44), equals
C,0,K /||1 , where C,0,K is as in (4.48).) We set K = log y. Since
y = x/ 1024 , it follows that y/K = y/ log y > 2.16 1020 .
2 0.55
r1 = y 2 ,
3
(
gx, (r)
g(r) =
gx, (r1 )
if r r1 ,
if r > r1 .
Since > 4, we see that r1 (x/16)0.55/2 . By Lemma 4.6, g(r) is a decreasing function; moreover, by Lemma 4.7, gy, (r1 ) h(y/ log y), and so g(r)
h(y/ log y) for all r. Thus, we have shown that
(7.49)
p> x
S2,+ (, x) =
(n)e(n)+ (n/x) +
n non-prime
n> x
(n)e(n)+ (n/x).
n x
2
X
j=1
sZ
(R/Z)\M8,r0
|S (, x)||Sj,+ (, x)|2 d.
Clearly,
Z
(R/Z)\M8,r0
|S (, x)||S2,+ (, x)|2 d
max |S (, x)|
(R/Z)
n=1
97
(n) (n/x)
R/Z
|S2,+ (, x)|2 d
(n)2 + (n/x)2 +
n non-prime
n x
(n)2 + (n/x)2 .
0.7131
2
(n)2
+ (t/x) dt
nt
n non-prime
Z
2
log e t
+ 2
t
n non-prime
(log t) 1.4262 t + 2 (t/x) dt
Z
t
2 + log tx 2
+ (t)dt
x,
dt 0.7131
x
t
0
nr1
n x
2
0.51941|+ | x log x.
This shows that
Z
X
(n) (n/x) E = S (0, x) E,
|S (, x)||S2,+ (, x)|2 d
(R/Z)\M8,r0
n=1
where E is as in (7.43).
It remains to bound
Z
(7.50)
(R/Z)\M8,r0
|S (, x)||S1,+ (, x)|2 d.
We wish to apply Prop. 7.9. Corollary 7.7 gives us an input of type (7.37); we
have just derived a bound (7.43) that provides an input of type (7.38). More
precisely, by Corollary 7.7, (7.37) holds with
( (log(r+1))+1.698
if r < r1 ,
log x/16
H(r) =
1
if r r1 .
Since r1 = (2/3)y 0.275 and 1750,
log((2/3)(x/)0.275 + 1) + 1.698
p
log x/16
!
0.275
0.275 log 23 + 1.698 + 0.275 log 16
p
1
+
= 0.45.
0.5
0.5
log x/16
rr1+
rr1
98
H. A. HELFGOTT
instead of g(r) (by (7.49)). Here (7.51) is a decreasing function of r because g(r)
is, as we already checked. Hence, Prop. 7.9 gives us that (7.50) is at most
(7.52)
Z r1
9g(r1 )
1
g(r)
p
dr +
g(r0 )(H(r0 ) I0 ) + (1 I0 ) C,3 (log x) +
r
+
1
20
log x/16 r0
P
times ||1 y p>x (log p)2 2 (p/x), where
Z
1
(7.53)
I0 = P
|S1,+ (, x)|2 d.
2 2
p> x (log p) + (n/x) M8,r0
|S1,+ (, x)|2 d =
sZ
M8,r0
sZ
M8,r0
M8,r0
|S+ (, x)|2 d
sZ
R/Z
(, x)|2
|S+
As we already showed,
Z
|S2,+ (, x)|2 d =
Thus,
n non-prime
or n x
sZ
M8,r0
sZ
R/Z
|S2,+ (, x)|2 d
|S2,+ (, x)|2 d.
(n)2 + (n/x)2 E.
I0 S ( J E)2 ,
log r
2 log e2 r0
log 2r
2 log 2e2 r0
r+
1
dr
=
dr
=
dr = log
,
,
,
3/2
r0
r0
r0
r 3/2
r0 r
r0
r0 r
Z
Z
(log 2r)3
(log 2r)2
2P2 (log 2r0 )
2P3 (log 2r0 )
,
,
dr
=
dr
=
1/2
r0
r 3/2
r 3/2
r0
r0
r
0
where
(7.55)
P2 (t) = t2 + 4t + 8,
r1
r0
(r) log r
dr,
r2
r1
r0
(r)
dr,
r 3/2
99
Clearly, (r) e log log r = 2.50637/ log log r is decreasing on r. Hence, for
r 105 ,
(r) e log log r + c ,
where c = 1.025742. Let F (t) = e log t + c . Then F (t) = e /t2 < 0. Hence
p
d2 F (t)
F (t)
(F (t))2
p
<0
=
dt2
2 F (t) 4(F (t))3/2
p
for all t > 0. In other words, F (t) is convex-down, and so we can bound
p
p
F (t) from above by F (t0 ) + F (t0 ) (t t0 ), for any t t0 > 0. Hence, for
r r0 105 ,
p
p
p
p
d F (t)
r
|t=log r0 log
(r) F (log r) F (log r0 ) +
dt
r0
r
p
log
e
r0
= F (log r0 ) + p
.
F (log r0 ) 2 log r0
Thus, by (7.54),
(7.57)p
Z
p
(r)
e
log e2 r0
1
e
p
F
(log
r
)
2
dr
0
F (log r0 )
r0
r0
r 3/2
F (log r0 ) log r0
r0
p
2 F (log r0 )
e
=
1
.
r0
F (log r0 ) log r0
The other integrals in (7.56) are easier. Just as in (7.57), we extend the range
of integration to [r0 , ]. Using (7.54), we obtain
Z
Z
c
(r)
F (log r)
log log r0
dr
dr = e
+ E1 (log r0 ) + ,
2
2
r
r
r
r0
0
r
r0
Z 0
c log er0
(1 + log r0 ) log log r0 + 1
(r) log r
dr e
+ E1 (log r0 ) +
,
2
r
r0
r0
r0
where E1 is the exponential integral
E1 (z) =
et
dt.
t
By [OLBC10, (6.8.2)],
1
1
E1 (log r)
.
r(log r + 1)
r log r
Hence
r0
,
r2
r0
r0
log log r + 1
e
0
log r0 + c
(r) log r
dr
log er0 .
r2
r0
100
H. A. HELFGOTT
Finally,
Z
r1
r0
2 log log r
(r)
2c
2c
log r
r1
e
|r0 +
+ 2E1
2
r0
r1
r
r 3/2
1
1
log log r0 log log r1
2e
+ 2c
r0
r1
r0
r1
1
1
.
+ 4e
r0 log r0
r1 (log r1 + 2)
It is time to estimate
Z
(7.58)
r1
r0
Ry,2r log 2r
r 3/2
(r)
dr,
dr.
3/2
3/2
r
r0
r0 r
We have already bounded the second integral. Let us look at Rz,t (defined in
+ 0.41415, where
(4.40)). We can write Rz,t = 0.27125Rz,t
!
log
4t
.
(7.59)
Rz,t
= log 1 +
9z 1/3
2 log 2.004t
Clearly,
Rz,e
1+
t /4 = log
t/2
log
36z 1/3
2.004
c+
ab
at
bt
(b t)2
f (t) =
In our case, a = 1/2, c = 1 and b = log 36z 1/3 log(2.004) > 0. Hence, for t < b,
b 3
b
3
b t > 0,
ab((a 2c)(b 2t) 2ct) =
2t + (b 2t) =
2
2
2 2
et /4 < 9z 1/3 /2.004. It is easy to check that, since we are assuming y 1024 ,
4 0.55
9
2r1 = y 2 <
3
2.004
y
log y
1/3
9z 1/3
.
2.004
(7.60)
2
2
Rz,2r
Rz,2r
log r/r0
log r1 /r
2
+ Rz,2r
.
1
log r1 /r0
log r1 /r0
101
Therefore, by (7.54),
Z r1
(Rz,2r log 2r)2
dr
log r/r0
log r1 /r
2
2
+ Rz,2r1
(log 2r)2 3/2
dr
Rz,2r0
3/2
log
r
/r
log
r
/r
r
r
1 0
1 0
r0
r0
2
2Rz,2r0
P3 (log 2r0 ) P3 (log 2r1 )
P2 (log 2r0 ) P2 (log 2r1 )
log 2r1
=
log rr10
r0
r1
r0
r1
2
2Rz,2r1
P2 (log 2r0 ) P2 (log 2r1 )
P3 (log 2r0 ) P3 (log 2r1 )
+
log
2r
0
log rr10
r0
r1
r0
r1
!
P2 (log 2r0 ) P2 (log 2r1 )
log 2r0 2
2
2
(R
R
)
= 2 Rz,2r0
z,2r1
x,2r0
log rr01
r0
r1
2
2
Rz,2r
Rz,2r
P3 (log 2r0 ) P3 (log 2r1 )
1
0
+2
r1
log r0
r0
r1
P2 (log 2r0 ) P2 (log 2r1 )
2
= 2Rz,2r
0
r0
r1
2
2
Rz,2r1 Rz,2r0 P2 (log 2r0 ) P3 (log 2r1 ) (log 2r0 )P2 (log 2r1 )
+2
,
log rr10
r0
r1
Z
r1
where P2 (t) and P3 (t) are as in (7.55), and P2 (t) = P3 (t)tP2 (t) = 2t2 +16t+48.
Putting all terms together, we conclude that
Z r1
g(r)
(7.61)
dr f0 (r0 , y) + f1 (r0 ) + f2 (r0 , y),
r
r0
where
f0 (r0 , y) =
(7.62)
(1 c )
I0,r0 ,r1 ,y + c
I0,r0 ,r1 ,
y
log y
s
2
I1,r0
r0
F (log r0 )
e
5
1
f1 (r0 ) =
+
F (log r0 ) log r0
2r0
2r0
13
80
1
log er0 + 10.102 Jr0 +
log er0 + 23.433
+
r0
4
9
f2 (r0 , y) = 3.2
(log y)1/6
r1
log ,
1/6
r0
y
r0
r1
2
2
Rz,2r1 Rz,2r0 P2 (log 2r0 ) P3 (log 2r1 ) (log 2r0 )P2 (log 2r1 )
log rr10
r0
r1
Jr = F (log r) +
e
,
log r
I1,r = F (log r) +
2e
,
log r
c =
C,2,log y /||1
log log y
102
H. A. HELFGOTT
8. Conclusion
We now need to gather all results, using the smoothing functions
+ = h200 (t) (t)
(as in (4.7), with H = 200) and
= (2 M )(t)
(as in Thm. 7.10, with 1750 to be set soon). We define (t) = t2 et /2 . Just
2
like before, hH = h200 is as in (4.10), (t) = ,1 (t) = et /2 , and 2 = 1 M 1 ,
where 1 = 2 I[1/2,1/2] .
We fix a value for r0 , namely, r0 = 150000. Our results will have to be valid
for any x x1 , where x1 is fixed. We set x1 = 4.5 1029 , since we want a result
valid for N 1030 , and, as was discussed in (4.1), we will work with x1 slightly
smaller than N/2. (We will later see that N 1030 implies the stronger lower
bound x 4.9 1029 .)
8.1. The 2 norm over the major arcs. We apply Lemma 3.1 with = +
and as in (4.3). Let us first work out the error terms. Recall that 0 = 8. We
use the bound on E+ ,r,0 given in (6.63), and the bound on ET+ ,0 r/2 in (6.64):
E+ ,r,0 2.139 108 ,
We also need to bound a few norms: by (4.17), (4.18), (4.26), (4.31) and (4.30),
|+ |1 0.996505,
547.5562
0.80014,
2007/2
1 + 4 log 200
1.0858.
1 + 2.21526
200
|+ |2 0.80013 +
(8.1)
|+ |
By (3.12),
S+ (0, x) = c
+ (0) x + O err+ ,T (0, x) x
5.190 r ET
0 r
+, 2
and
||1 +
ET
0 r
+, 2
!!
0.031789
3 log r
2
2
0 r
2+
E+ ,r,0 + (log 2e r)Kr,2 4.844 107 .
2
We recall from (4.18) and (4.14) that
(8.2)
and
(8.3)
0.8001287 | |2 0.8001288
|+ |2
547.5562
4.84 106 .
H 7/2
103
(8.4)
(2)
(3)
We bound | |1 using the fact that (as we can tell by taking derivatives) (t)
increases from 0 at t = 0 to a maximum within [0, 1/2], and then decreases to
(2)
(1) = 7, only to increase to a maximum within [3/2, 2] (equal to that within
[0, 1/2]) and then decrease to 0 at t = 2:
(2)
(3)
(8.5)
(2)
(2)
t[0,1/2]
= 4 max
t[0,1/2]
(2)
(t)
+ 14 4 4.6255653 + 14 32.503,
qr
q odd
.
1+
(8.7)
C0 =
1
(p 1)2
(p 1)3
pN
p|N
The only prime that we know does not divide N is 2. Thus, we use the bound
Y
1
C0 2
1
1.3203236.
(p 1)2
p>2
104
H. A. HELFGOTT
2 /2
2 (w)
(2 )()d =
()d =
w
w
0
1/4
0
0
Z
Z 1
|2 |1 ||1
1
=
()ddw.
2 (w)
1/4
N/xw
Now
y 2 /2
() = ye
Z
+ 2
t2
e
y/ 2
()d <
dt <
2
y+
y
1
2 +
ey
e2
2 /2
()d
1/4
0
(8.9)
1
2
||1 2 + 2 e2 .
Let us now focus on the second integral in (8.8). Write N/x = 2 + c1 /. Then
the integral equals
Z 2+c1 /
Z
1
2
(c1 / + ) ()d 3
(u c1 )2 (2 M )(u) du
0
0
Z
Z 1
1
(vw c1 )2 (v)dvdw
2 (w)
= 3
1/4
0
r
r
Z 1
1
2
2
= 3
2 (w) 3
w 2 2c1 w + c1
dw
1/4
2
2
r
r
1 49 9
2
c1 +
c .
= 3
48 2 4
2 1
48 2
3
2 2
We conclude that
C ,
Setting
1
1.9986
1
2
2
2
||1 | |2 | |2 2 + 2 e2
.
3
= 49
C ,
1
(||1 | |22 0.000833).
105
p
Here it is useful to note that ||1 = 2 , and so, by (4.18), ||1 | |22 = 0.80237 . . . .
We have finally chosen x in terms of N :
(8.11)
x=
N
N
N
=
= 0.49724 . . . N.
c1 =
49/36
1
2+
2 + 361 2
2 + 2 49
Thus, we see that, since we are assuming N 1030 , we in fact have x 4.9724 . . .
1029 , and so, in particular,
x
1028 .
(8.12)
x 4.9 1029 ,
E ,r,8
3.061 108
,
where the factor of comes from the scaling in (t) = (2 M )(t) (which in
effect divides x by ). It remains only to bound the more harmless terms of type
Z,2 and LS .
By (3.12),
1X 2
1X
2
Z+2 ,2 =
(n)+
(n/x)
(n)+ (n/x) (+ (n/x) log n)
x n
x n
1X
(8.14)
(n)+ (n/x) (|+ (t) log+ (t/2)| + |+ | log 2x)
x n
(|+ |1 + | err+ ,T (0, x)|) (|+ (t) log+ (t/2)| + |+ | log 2x),
where log+ (t) := max(0, log t). Proceeding as in (4.27), we see that
|+ (t) log+ (t/2)| | (t) log+ (t/2)| + |h hH | | (t) log+ (t/2)|
= |h hH | | (t) log+ (t/2)| .
Since
| (t) log+ (t/2)| max | (t)(t/2 1)| 0.012
t2
and |h hH | < 0.1415 (by (4.29) and (4.30)), we see that |+ (t) log+ (t/2)| <
0.0017. We also have the bounds on |+ |1 and |+ | in (8.1) and the bound
| err+ ,T (0, x)| ET+ ,0 r/2 5.122 109 from (6.64). Hence
Z+2 ,2 0.99651 (0.0017 + 1.0858 log 2x) 0.002 + 1.09 log x.
Similarly,
Z2 ,2 =
1X 2
1X
(n)2 (n/x)
(n) (n/x) ( (n/x) log n)
x n
x n
| | = |2 M |
2 (t)
|| 1.92182 2 1.414.
t 1
e
106
H. A. HELFGOTT
We conclude that
(8.17)
p
Z2 ,2 ( /2/49 + 3.061 108 )(1.46503 + (2/e) log(x/49)) 0.0189 log x.
We have bounds for | | and |+ | . We can also bound
and
(8.18)
X | t|
log x
= (log r) max
||
+
pr log p
p /x
(log r) max
pr
and so
(8.19)
1
p x
| t|
log x
|| +
log p
1 1/p
(log r)(log x)
|| + 2(log r)| t| ,
log 2
LS
2
3/2
log x + 2 (3/e)
log r (1.0615 log x + 2.3189) log r,
e log 2
LS+
1.14146
log x + 2 1.1507 log r (1.6468 log x + 2.3014) log r.
log 2
(3)
| |21
05
is at most
107
| |2 =
q
3
32
3
1.53659
8 3 (log 2)
.
=
3.061 108
1.53659
8
8.739 + 2.139 10 1.6812( 8.739 + 1.6812 0.80014)
6
1.6097 10
.
where we are using the bound A+ 8.739 we obtained in (8.6). (We are also
using the bounds on norms in (8.1).)
Using the bounds (8.14), (8.17) and (8.19), we see that the third line of (3.36)
is at most
2 (0.0189 log x (1.0615 log x + 2.3189) log 150000) x
p
+ 4 (0.002 + 1.14 log x) 0.0189 log x(1.6468 log x + 2.3014)(log 150000)x
13(log x)2 x,
where we just use the very weak assumption x 1010 , though we can by now
assume (8.12)).
We conclude that, for r = 150000, the integral over the major arcs
Z
S+ (, x)2 S (, x)e(N )d
M8,r
is
(8.20)
!
6
/2
1.6097
10
x2 +
x2 + 14(log x)2 x
C0 C0 , x2 + O 0.0001691
0.0002136x2
= C0 C0 , x2 + O
= C0 C0 , x2 + O 4.359 106 x2 ,
1.0594003 0.000833
1.05857
1.3203236||1 | |22 0.000833
=
.
49
Z
M8,r
S+ (, x)2 S (, x)e(N )d
1.05856 2
x ,
108
H. A. HELFGOTT
Z
Similarly,
C+ ,1 0.7131
1
0
2
sup + (r) dt
rt
1.08582
+
t
1.15072
dt
t5/2
2
Z
sup + (r) dt 0.7131
rt
2.31092.
1.15072 log t
dt 0.41965.
t5/2
(8.22)
Hence
S (0, x) E 5.84 1014
(8.23)
We can bound
(8.24)
X
n
x2
.
2
(n)(log n)+
(n/x) = 0.64022x log x 0.0211x
2
we can bound
Z 1
x 1.04488 log x/ 2 t2 /2
0.2779
= p
.
C,3 log
t e
dt
(log x/)3
/2 0
109
and so
x
(S ( J E)2 )
T = C,3 log
0.2779
((0.6402 + O (2 105 ))x log x 0.0211x 8.672999x)
(log x/)3
x log x
x
0.17792
2.41609
3
(log x/)
(log x/)3
x
x
0.17792
1.72365
.
2
(log x/)
(log x/)3
T 3.638 105 x.
(8.26)
1/K
0.093426,
where
again we use VNODE-LP for rigorous numerical integration. Since ||1 =
p
/2, this implies that
C,2 /||1
0.07455
log K
log log y
(8.27)
and so
(8.28)
Ry,K,,t
0.07455
0.07455
R
+
1
Ry,t .
=
log log y y/K,t
log log y
Let t = 2r0 = 300000; we recall that K = log y. Recall from (8.12) that
y = x/ 1028 ; thus, y/K 1.5511 1026 and log log y 4.16624. Going back
to the definition of Rx,t in (4.40), we see that
!
log(8 150000)
+ 0.41415 0.55394,
(8.29) Ry,,2r0 0.27125 log 1 +
9(1028 )1/3
2 log 2.0042150000
(8.30)
Ry/K,2r0 0.27125 log
and so
Ry,K,,2r0
Using
1+
log(8 150000)
2 log
9(1.55111026 )1/3
2.0042150000
+ 0.41415 0.5703,
0.07455
0.07455
0.5703 + 1
0.55394 0.55424.
4.16624
4.16624
(r) = e log log r +
2.50637
5.42506,
log log r
110
H. A. HELFGOTT
log y 1/6
(0.55424 log 300000 + 0.5) 5.42506 + 2.5 394.316
+
+ 3.2
g(r0 ) =
150000
y
2 150000
0.039182.
Using again the bound x 4.9 1029 from (8.12), we obtain
log(150000 + 1) + 1.698
p
S ( J E)2
log x/16
13.6164
1
(0.64022x log x 0.0211x) 8.673x
log
x log 4
2
13.6164 1.33393x 8.673x 9.49046x.
Therefore,
(8.31)
g(r0 )
log(150000 + 1) + 1.698
p
S ( J E)2
log x/16
0.039182 9.49046x
0.37186x.
Ry,2r1
(8.32)
2 0.55
2
log
8
y
3
= 0.27125 log 1 +
+ 0.41415
9y 1/3
2 log
0.55
2.0042 23 y 2
!
16
0.55
2 log y + log 3
+ 0.41415
= 0.27125 log 1 +
27/8
7
log
y
+
2
log
60
1.002
33
0.27125 log 1 +
+ 0.41415 0.74266.
14
2 0.55
2
log
8
y
3
= 0.27125 log
1+
7
60
log y
0.55
16
2 log y + log 3
27/8
23 log log y + 2 log 1.002
+ 0.41415.
Let
f (t) =
16
0.55
2 t + log 3
27/8
2
7
60 t 3 log t + 2 log 1.002
33
+
=
14
2
3
7
60 t
log t c
32 log t + 2 log
27/8
1.002
111
27/8
23 log t + 2 log 1.002
23 log t c
f (t) =
2
27/8
7
2
t
log
t
+
2
log
60
3
1.002
27/8
7
2
7
log
t
(c
+
2/3)
+
c
2
log
90
60
1.002 3t
.
=
27/8 2
7
2
t
log
t
+
2
log
60
3
1.002
2
3t
7
60 t
7
60
2
3t
Hence f (t) < 0 for t 25. It follows that Ry/K,2r1 (with K = log y and r1
varying with y) is decreasing on y for y e25 , and thus, of course, for y 1028 .
We thus obtain from (8.33) that
(8.34)
Ry/K,2r1 R
1028
log 1028
,2r1
0.76994.
0.07455
0.07455
0.76994 + 1
0.74266 0.74315.
4.16624
4.16624
Since r1 = (2/3)y 0.55/2 and (r) is increasing for r 27, we know that
(8.35)
2.50637
(r1 ) (y 0.55/2 ) = e log log y 0.55/2 +
log log y 0.55/2
2.50637
2
= e log log y +
2 e log 0.55 e log log y 1.42763
log log y log 0.55
for y 1028 . Hence, (4.40) gives us that
Lr1
32
80
2 3 44 111
(e log log y 1.42763) log 13 y +
+ log 16 y 9 +
9
5
34
39
3.859 log y log log y + 1.7957 log y + 15.65 log log y + 15.1
25
13
6
Moreover,
p
p
p
1.42763
(r1 ) = e log log y 1.42763 e log log y
2 e log log y
and so
p
4 11
(0.74315 log y 40 + 0.5) (r1 )
3
p
1.42763 (0.2044 log y + 0.7138)
112
H. A. HELFGOTT
Therefore, by (4.46),
gy, (r1 )
2
3y
11
40
y 80
3.2(log y)1/6
y 1/6
r1
r0
g(r)
dr
r
in the expression (7.45) for M . We will use the bound on the integral given
in (7.61). The easiest term to bound there is f1 (r0 ), defined in (7.62), since it
depends only on r0 : for r0 = 150000,
f1 (r0 ) = 0.0161322 . . . .
It is also not hard to bound f2 (r0 , x), also defined in (7.62):
2 11
40
(log y)1/6
(log y)1/6 11
3x
f2 (r0 , y) = 3.2
log
3.2
log
y
+
0.6648
log
r
0 ,
r0
40
y 1/6
y 1/6
and so, since r0 = 150000 and y 1028 ,
f2 (r0 , y) 0.000895.
Let us now look at the terms I1,r , c in (7.63). We already saw in (8.27) that
c =
0.07455
C,2 /||1
0.0179.
log K
log log y
I1,r0 = F (log r0 ) +
2e
= 5.42938 . . .
log r0
It thus remains only to estimate I0,r0 ,r1 ,z for z = y and z = y/K, where K = log y.
We already know that
Ry,2r0 0.55394,
Ry,2r1 0.74266,
Ry/K,2r0 0.5703,
Ry/K,2r1 0.76994
by (8.29), (8.30), (8.32) and (8.34). We also have the trivial bound Rz,t 0.41415
valid for any z and t for which Rz,t is defined.
113
Omitting negative terms from (7.63), we easily get the following bound, crude
but useful enough:
I0,r0 ,r1 ,z
2
Rz,2r
0
2
0.414152 P2 (log 2r0 )
P2 (log 2r0 ) Rz,2r
1
+
,
r0
log rr10
r0
where P2 (t) = t2 + 4t + 8 and P2 (t) = 2t2 + 16t + 48. For z = y and r0 = 150000,
this gives
P2 (log 2r0 ) 0.742662 0.414152 P2 (log 2r0 )
+
0.55
r0
r0
2
log 2y3r0
0.55722
;
0.17232 + 11
40 log y log 225000
(8.38)
11
40
0.61773
.
log y log 225000
q
p
(1 c ) I0,r0 ,r1 ,y + c I0,r0 ,r1 , y
log y
s
0.9821
0.17232 +
+ 0.0179
0.18265 +
11
40
11
40
0.55722
log y log 225000
0.61773
.
log y log 225000
f0 (r0 , y) 0.52514
y
log y
0.52514.
2
5.42939 0.087932.
r0
2S
2(0.64022x log x 0.0211x)
1.33393x.
x
log 16
log x log 16
Hence
(8.39)
2S
x
log 16
r1
r0
g(r)
dr 0.20029x.
r
Putting (8.31), (8.36) and (8.39), we conclude that the quantity M defined in
(7.45) is bounded by
(8.40)
114
H. A. HELFGOTT
Gathering the terms from (8.23), (8.26) and (8.40), we see that Theorem 7.10
states that the minor-arc total
Z
|S (, x)||S+ (, x)|2 d
Zr0 =
(R/Z)\M8,r0
is bounded by
Zr 0
(8.41)
p
||1 x
(M + T ) +
0.83226
S (0, x) E
!2
x
x
+ 5.84 1014
105 )
x2
2
for r0 = 150000, x 4.9 1029 , where we use yet again the fact that ||1 =
This is our total minor-arc bound.
/2.
8.4. Conclusion: proof of main theorem. As we have known from the start,
X
(n1 )(n2 )(n3 )+ (n1 )+ (n2 ) (n3 )
(8.42)
n1 +n2 +n3 =N
R/Z
+O
(R/Z)\M8,r0
x2
x2
x2
1.05856 + O 0.83226
0.2263
for r0 = 150000, where x = N/(2+1/(36 2)), as in (8.11). (We are using (8.21)
2
and (8.41).) Recall that = 49 and (t) = (2 M )(t), where (t) = t2 et /2 .
It only remains to show that the contribution of terms with n1 , n2 or n3 nonprime to the sum in (8.42) is negligible. (Let us take out n1 , n2 , n3 equal to 2 as
well, since some prefer to state the ternary Goldbach conjecture as follows: every
odd number 9 is the sum of three odd primes.) Clearly
X
(n1 )(n2 )(n3 )+ (n1 )+ (n2 ) (n3 )
n1 +n2 +n3 =N
n1 , n2 or n3 even or non-prime
(8.43)
3|+ |2 | |
n1 +n2 +n3 =N
n1 even or non-prime
3|+ |2 | | (log N )
n1 N non-prime
or n1 = 2
(n1 )
n2 N
(n2 ).
115
n2 N non-prime
or n2 = 2
x2
1.0568N 3/2 log N
by = 49 and (8.11). Since 0.00114N 2 > 0, this shows that every odd number
N 1030 can be written as the sum of three odd primes.
Since the ternary Goldbach conjecture has already been checked for all N
30
10 [HP], we conclude that every odd number N > 7 can be written as the
sum of three odd primes, and every odd number N > 5 can be written as the
sum of three primes. The main theorem is hereby proven: the ternary Goldbach
conjecture is true.
Appendix A. Sums over primes
P
Here we treat some sums of the type
n (n)(n), where has compact
support. Since the sums are over all integers (not just an arithmetic progression)
and there is no phase e(n) involved, the treatment is relatively straightforward.
The following is standard.
Lemma A.1 (Explicit formula). Let : [1, ) C be continuous and piecewise
C 1 with L1 ; let it also be of compact support contained in [1, ). Then
Z
X
X
1
(x)dx
(M )(),
(A.1)
(n)(n) =
1
x(x2 1)
1
n
where runs over the non-trivial zeros of (s).
The non-trivial zeros of (s) are, of course, those in the critical strip 0 <
(s) < 1.
Remark. Lemma A.1 appears as exercise 5 in [IK04, 5.5]; the condition there
that be smooth can be relaxed, since already the weaker assumption that
be in L1 implies that the Mellin transform (M )(
P + it) decays quadratically on t
as t , thereby guaranteeing that the sum (M )() converges absolutely.
Lemma A.2. Let x 10. Let 2 be as in (4.33). Assume that all non-trivial
zeros of (s) with |(s)| T0 lie on the critical line.
116
H. A. HELFGOTT
Then
n
0
X
log eT
9/4 6.03
9.7
1/2
2
=
x
+
O
0.135x
+
+
+
x.
(n)
(A.2)
2
2
x
x
T
2
T
0
0
n
The assumption that all non-trivial zeros up to T0 = 3.061 1010 lie on the
critical line was proven rigorously in [Plaa]; higher values of T0 have been reached
elsewhere ([Wed03], [GD04]).
Proof. By Lemma A.1,
Z
n Z t
X
X
2 (t/x)
2
=
dt
(M )(),
dt
(n)2
x
x
t(t2 1)
1
1
where (u) = R2 (u/x) and runs over all non-trivial zeros of (s). Since 2 is
9.61114
dt
=
O
dt
=
O
.
2 2
t(t2 1)
x2
1/4 tx (t 1/100)
1
By (2.6),
X
X
X 1 2 2
(M )() =
M 2 ()x =
x = S1 (x)2S1 (x/2)+S1 (x/4),
where
(A.3)
Sm (x) =
X x
.
m+1
Setting aside the contribution of all with |()| T0 and all with |()| > T0
and (s) 1/2, and using the symmetry provided by the functional equation,
we obtain
X 1
X
1
|Sm (x)| x1/2
+
x
m+1
m+1
||
||
x1/2
1
||m+1
|()|>T0
|()|>1/2
|()|>T0
1
||m+1
We bound the first sum by [Ros41, Lemma 17] and the second sum by [RS03,
Lemma 2]. We obtain
2.68
1
eT0
+
+ m x1/2 ,
(A.4)
|Sm (x)|
x log
2mT0m T0m+1
2
where 1 = 0.0463, 2 = 0.00167 and 3 = 0.0000744.
Hence
X
3
2.68 9x
eT0
1
+ 2
log
+
+ 2 1 x1/2 .
(M )() x
2T0
4
2
2
T0
117
Corollary A.3. Let 2 be as in (4.33). Assume that all non-trivial zeros of (s)
with |(s)| T0 , T0 = 3.061 1010 , lie on the critical line. Then, for all x 1,
n
X
min (1 + )x + 0.2x1/2 , 1.04488x ,
(A.5)
(n)2
x
n
where = 2.73 1010 .
= .
(B.1)
log
1+
p(p 1)
p(p 1) n>r n(n 1)
r
p>r
p>r
A quick computation12 now suffices to give
(B.2)
2.591461
< 2.591463
and so
(B.3)
1.295730
X 2 (q)
< 1.295732,
(q)q
q odd
since the expression bounded in (B.3) is exactly half of that bounded in (B.2).
Again using (B.1), we get that
(B.4)
2.826419
X 2 (q)
q
(q)2
< 2.826421.
In what follows, we will use values for convergent sums obtained in much the
same way an easy tail bound followed by a computation.
12Using D. Platts integer arithmetic package.
118
H. A. HELFGOTT
where
cE = +
X
p
log p
= 1.332582275 + O (109 /3)
p(p 1)
by [RS62, (2.11)]. As we already said in (7.15), this, supplemented by a computation for r 4 107 , gives
X 2 (q)
log r + 1.312
log r + 1.354
(q)
qr
for r 195. (The numerical verification here goes up to 1.38108 ; for r > 3.18108 ,
use B.6.)
Clearly
X 2 (q)
X 2 (q)
=
.
(B.7)
(q)
(q)
qr
q odd
q2r
q even
qr
q odd
X 2 (q)
,
(q)2
qr
q even
pj (p 1)j
,
fj (pk ) = 0
(p 1)j p
We will also find the following estimate useful.
fj (p) =
for k 2.
1+ j
(B.9)
aj
Aj1
p
aA
(a,m)=1
p|m
119
1+ j
C=
(2j)
p
p|m
Let A 2. Since
aA
(a,m)=1
X
2 (a)
=
C
aj
a<A
(a,m)=1
2 (a)
aj
and
C=
2 (a)
<
aj
2 (a)
1
1
+ j +
,
aj
A
(j 1)Aj1
a
(a,m)=1
a<A
(a,m)=1
a<A
(a,m)=1
2 (a)
1
+ j +
j
a
A
1
dt
tj
we obtain
X
aA
(a,m)=1
X
1
Aj1 1
2 (a)
=
C
+
aj
Aj1
Aj1
Aj1 1
C
< j1 +
A
Aj1
C
1
j1 + j1
A
A
a<A
(a,m)=1
2 (a)
aj
Aj1
1
1
+
A j1
1 .
1
1
+
j
A
(j 1)Aj1
1
Aj1
a<A
(a,m)=1
2 (a)
aj
We now obtain easily the estimates we want: by (B.8) and Lemma B.1 (with
j = 2 and m = 1),
(B.10)
X 2 (q)
qr
(q)2
a
q
b
a2
qr ab=q
b1
ar/b
15 Y
2p 1
6.7345
(2)/(4) X
2
1+
.
f2 (b) =
2
r
r p
(p 1) p
r
=
b1
120
H. A. HELFGOTT
(B.11)
X f2 (b) X 2 (a)
X 2 (q)
(2)/(4) 1 X
f2 (b)
=
(q)2
b
a2
1 + 1/22 r
qr
q odd
b1
b odd
b odd
ar/b
a odd
12 1 Y
2p 1
2.15502
= 2
1+
2
r
(p 1) p
r
p>2
X 2 (q)
X 2 (q)
4.31004
=
.
2
2
(q)
(q)
r
(B.12)
qr
q even
qr/2
q odd
Lastly,
(B.13)
X 2 (q)q
X 1 r
X
X 1 X
X 1
2 (q)
=
=
+1
2 (q)
(q)
(d)
(d)
2(d) d
qr
q odd
qr
q odd
dr
d odd
d|q
qr
d|q
q odd
dr
d odd
1 X 1
log r
r X 1
+
0.64787r +
+ 0.425,
2
(d)d 2
(d)
4
d odd
dr
d odd
1.6222831573801406 C0 1.6222831573801515.
2 and
1
( 10 + 2); only the first three of these four roots lie in the interval [0, 2].
14By VNODE-LP [Ned06], running on PROFIL/BIAS [Kn
u99].
15It does not seem possible to use [Ned06] directly on an integrand involving the abs function,
121
( 10 2, 2). Hence
Z
= 2h( 10 2) = 3.58000383169 + O 1012 .
The situation with (x3 (2 x)3 ex1/2 ) is similar: it has zeros at the roots of
h (x)xdx +
h (x)xdx
h (x)xdx
|h (x)|xdx =
C2 =
=h
+h
2,1
(x)x|0 2,1
(x)x|22,2
2,1
0
2
h (x)dx
2,2
(x)x|2,1
2,2
2,2
h (x)dx
2,1
h (x)dx
2,2
To compute C3 , we proceed in the same way, except now we must find the roots
numerically. It is enough to find (candidates for) the roots using any available
tool16 and then check rigorously that the sign does change around the purported
roots. In this way, we check that the roots 3,1 , 3,2 , 3,3 of H3 (x) = 0 lie within
the intervals
[0.366931547524632, 0.366931547524633],
[1.233580882085861, 1.233580882085862],
[1.847147885393624, 1.847147885393625],
respectively. The sign of H3 (x) on the interval [0, 2] is first +, then , then +,
then . Proceeding as before, we obtain that
Z
C3 =
|h (x)|x2 dx
0
(C.5)
=2
3
X
(1)j+1 (h (3,j )23,j 2h (3,j )3,j + 2h(3,j ))
j=1
C3 = 131.3398196149 + O 1010 .
The treatment of the integral in C4 is very similar, at least as first. The roots
4,1 , 4,2 of H4 (x) = 0 lie within the intervals
[0.866114027542349, 0.86611402754235],
[1.640243631518005, 1.640243631518006].
16Routine find root in SAGE was used here.
122
H. A. HELFGOTT
The sign of H4 (x) on the interval [0, 2] is first , then +, then . Using integration
by parts as before, we obtain
Z 2
(4) 3
h (x) x dx = lim h(3) (x)x3 lim h(3) (x)x3
x0+
+2
j=1
Now
x2
(1)j h(3) (4,j )34,j 3h (4,j )24,j + 6h (4,j )4,j 6h(4,j )
2
X
= 2199.91310061863 + O 3 1011 .
Z
Hence
(C.7)
C4, = 48 e3/2 23 = 3920.8817036284 + O 1010 .
= ,
0
3/2
2()(() + j())
2
2
and so
(C.8)
() 1
1
1 .
2()(() + j())
2
(C.9)
,
2 2
2
2
2
123
(C.11)
0.018
It remains to study c0 () for [0, 0.01]. The method we are about to give
actually works for all [0, 1].
Since
1
1
,
1+x =
1+x =
,
4(1 + x)3/2
2 1+x
3/4
1
1/2
1
1
=
,
,
=
=
3/2
3/2
2(1 + x)
(1 + x)
(1 + x)5/2
1+x
1+x
a truncated Taylor expansion gives us that, for x 0,
1
1
1
1 + x x2 1 + x 1 + x
2
8
2
(C.12)
1
3
1
1
1 x + x2 .
1 x
2
2
8
1+x
Hence, for 0,
(C.13)
1 + 2 /2 4 /8 j() 1 + 2 /2,
and so
() 1 + 2 /8 54 /128
(C.14)
(C.15)
Hence
1
=q
()
1 j() 1 3
1
+
2
2
8
j() 1
2
1 + j()1
2
2
4
2
74
1
3 4
1
+
,
1
+
2 4
16
8 16
8
128
1 j() 1
1
1
2
1
=q
1 .
()
2
2
8
1 + j()1
2
sin 0 =
(C.16)
sin 0
while
(C.17)
cos 0 =
s
r
1
1
2 2()
74
=
16 256
4
7 2
,
16
= ,
16
4
1
1
+
2 2()
2
1 ,
16
cos 0
2
74
+
,
16 256
2
1 2 /8
2( + j)
2 2 + 52 /8
2
1 32
2
32
33
.
4
64
124
H. A. HELFGOTT
Assuming 0 1,
1
1+
52
16
94
64
52 94
+
+
1
16
64
52 94
16
64
2 !
52 614
+
,
16
256
7
1 8 + 128
2( + j)
2 2 + 52 214
8 2 128 4
7
52 464
+
+
1
1
4
8
128
16
256
2
35 4
81 6 161 8
7
73 355
+
+ 14
+
.
1
4
16
128
2048
2
4
64
512
Hence, we obtain
2
() = 1 +
2( + j)
2( + j)
2
2
73 355
1 2
1 33
+
1+
2 4
64
8 16
4
64
512
(C.19)
2
3
5
3
7
1
96
35
1 +
+
+
+ 13
4
4 32
64
256 2048
512
2
2
3
4
7
165
+
,
1 +
4 32
64
2048
where, in the last line, we use again the assumption 1.
For x [1/4, 0],
1
x
1
x2
x2
1+x 1+ x
=
1
+
2
2 4(1 1/4)3/2
2 33/2
x2
1
1
1 + x.
1+x 1+ x
2
8
2
Hence
(C.20)
r
2
2
4
74
2
74
1
cos 0 1
3/2
+
1
+
32 3 256
16 256
32 512
49
7 2
4 sin 0
1 3/2
4
32
4
3 256
for 1. Therefore,
c0 () = () cos 0 + sin 0
2
73
165 4
4
2
1 +
+
3/2
1
4 32
64
2048
32 3 256
7 3
49
3/2
+
5
4 128
3 1024 !
!
!
7 3 3
7
167
3
3
3
+
+
+
+
4
1+
16
2304 2048
2048 192
147456
1+
3
0.09494 ,
16
125
where we are again using 1. We conclude that, for all (0, 1/2],
c0 () > 1.
Together with (C.10) and (C.11), this shows that
(C.21)
c0 () > 1
> 0.
2
2
r
r
r r
r
1
1
1
1
1+j
1
1+ + 2
() =
+ +
=
1+
,
2
2 2 4
2
2
2
2
p
p
together with the trivial bounds j() and () j()/2 /2. By
(C.22),
q
1 + 2
1
1
2 q =
2 q
q
2
2
2
1
1
+
1
+
1
+
2
2
2
2
2
(C.23)
q
2
2
1+
8
2
q + 3/2
=
2
1+ 2
1 2 +
+ 1
22
j
2+
2
for 15. In fact, the bisection method (applied with 20 iterations, including 10
initial iterations after which the possibility of finding a minimum within each
interval is tested) shows that (C.23) (and hence (C.24)) holds for all 1. By
(C.22),
(C.25)
2( + j)
2 1 +
q
1
2 1+
2
1
1
1
2 1 + 2 +
1
2
++
1
2
1
1
1
2 2
23/2
for 16. (Again, (C.25) is also true for 1 16 by the bisection method; it
is trivially true for [0, 1], since the last term of (C.25) is then negative.) We
126
H. A. HELFGOTT
1
1
q
q
=
+
(C.26)
2( + j)
2 + 1
2 2 (2)3/2
2 2 ( 2 + )
valid for 1/2.
Hence, by (C.12), (C.25) and (C.26),
s
2
= 1+
2( + j)
2( + j)
1
1
1
1
1
1 2
1
1
+
+
1 +
1+
3/2
2
2 2
2 2
2
2
for 3. Again,
we use the bisection method (with 20 iterations) on [1/2, 3],
and note that 1/ 2 < 1/ for < 1/2; we thus obtain
1
1
(C.27)
1 +
2
for all > 0.
We recall (5.43) and the lower bounds (C.24) and (C.8). We get
s
r
2 !
1
1
8
1
1 2
1
1+ 1
1
+
2+
2
2
2
2
1
1
1
1
+
1 +
2
2 + 1
2
!
1
1
1
2
1
2
5
(C.28)
1+
2 +
1 +
2
2
2 4
2 2
1
1
1
1
+ 1 1 + 2 1 +
2
2
2
1
1
1
1
9
37
1
2+
3 1+
1+
5/2
2 4 8
2 16
2
for 2. This implies that () > 1 for 11. (Since our estimates always
give an error of at most O(1/ ), we also get lim () = 1.) The bisection
method (with 20 iterations, including 6 initial iterations) gives that () > 1 also
holds for 1 11.
Let us now look at what happens for 1. From (C.19), we get the simpler
bound
33
2
+
1
(C.29)
1 +
4 32
32
4
valid for 1, implying that
234
2 113
+
+
2
8
64
1024
for 1. We also have, by (5.23) and (C.18),
2
33
1
1 2
1+
1+
2 2( + j)
2( + j)
2 4
4
64
(C.30)
2
3
2
3
5
+
1 +
1 +
4 32
64
4
64
2 1
127
2
2
4
2
2
2
2
1 + 8
1 + 8
8 + 64
+
2
2
8
64
128
2
4
64
2
2 2
2
2
3
2
2
11
3
1
4 3 15
33
7
1 +
+
1+ +
4
32
64
2
2
32
2
2
64
64
4 3
2
for 1. This implies the bound () > 1 for all 1. Conversely, ()
4/ 3/2 follows from () > 1 for > 8/5. We check () 4/ 3/2 for
[1, 8/5] by the bijection method (5 iterations).
We conclude that, for all > 0,
4 3
(C.31)
max 1,
.
2
This bound has the right asymptotics for 0+ and +.
Let us now bound c0 from above. By (C.20) and (C.30),
74
2
52
+
1
+
c0 () = () cos 0 + sin 0 1 +
4
64
32 512
4
(C.32)
32
3
234
75
356
2
1+
+
+
+ 15 1 +
64
128 2048 2048
2
15
for 1. Since 1 and 0 [0, /4] [0, /2], the bound
(C.33)
c0 () cos 0 + sin 0 2
1
1
1 2
9
1 +
+1 2 1 +
.
2
2 2 16
From (C.31) and (C.33), we obtain that
1
1 + 2c20 1 (1 + 2 2) = 5
(C.35)
for all 0. At the same time, (C.31) and (C.32) imply that
1
4 3 1
42 24
+ 2
1 + 2c20
3+
2
15
15
1
4
3
3
42
3
1+
+
=
1
1+
4
8
45
675
4
2
for 0.4. Hence (1 + 2c20 )/ 0.86 for < 0.29. The bisection method
(20 iterations, starting by splitting the range into 28 equal intervals) shows that
128
H. A. HELFGOTT
(1 + 2c20 )/ 0.86 also holds for 0.29 6; for > 6, the same inequality
holds by (C.35).
We have thus shown that
(C.36)
1 + 2c20
min(5, 0.86)
2 1 +
8
128
8
(C.37)
2
2 54
52
1
2
2 54
4
=1+
1+
4
64
128 8
8
8
64
for 1, and so
r
52
,
1
2
4
8
and this is greater than /6p
for 1/3. The bisection method (20 iterations, 5
initial steps) confirms that ( 2 )/2 > /6 also holds for 2/3 < 4. On
the other hand, by (C.22) and 2 = (1 + j)/2 (1 + )/2,
v
q
u
r
r
u +1 1 + 1
s
2
2
t 2
2
1
1
2
1+
1+
2
2
2
2
(C.38)
s
r
r
1
1
2
1
+
3/2
1
=
1
2
2
2
2
2
2
p
for 4. We check by the bisection method (20 iterations) that ( 2 )/2
s
1
1
1 + 1/
1
2
8
1
1
+
c1 () =
2
2
3/2
2
2
r
r
1
1
3
2
1
2
1+
1
1
=
4
2 4
2
for 1. Together with (C.34), this implies that, for 2,
1
9
1 + 9
2
1 12 + 8
2
16
c0 1/ 2
1
2 2
,
=
q
2 2 1 3
2c1
2 2 1 3
129
+
=
8
8
32
8
32
,
1
c1 () =
1
1
4
2
2
4
4
4
+ 9
8
32
whereas, for 1,
by (C.20). Thus
1 + 4 12
c0 1/ 2
0.0584
2c1
2 4 1 54 2 1 4
for 0.1.
We check the remaining interval [0.1, 25] (or [0.1, 8], if we aim at the
bound 1/ 8) by the bisection method (with 24 iterations, including 12 initial
iterations or 15 iterations and 10 initial iterations, in the case of [0.1, 8]) and
obtain that
c0 1/ 2
0.0763896
0.0763895 max
0
2c1
(C.40)
1
c0 1/ 2
.
sup
2
1
0
In the same way, we see that
1
1
c0
3 0.171
c1
1 4
for 36 and
1 + 4
c0
0.267
c1
4 1 54 2 1 4
for 0.1. The bisection method applied to [0.1, 36] with 24 iterations (including
12 initial iterations) now gives
c0
0.29888.
(C.41)
0.29887 max
>0 c1
We would also like a lower bound for c0 /c1 . For c0 , we can use the lower bound
c0 1 given by (C.21). By (C.15), (C.30) and (C.37),
s
r
2
74
2 8 + 128
1 + 1/
52
c1 () =
1 +
2
2 /8 54 /64
4
64
2
2
4
5
5
4
1+
1 +
<
16
4
64
for 1/4. Thus, c0 /(c1 ) 1/4 for [0, 1/4]. The bisection method (with 20
iterations, including 10 initial iterations) gives us that c0 /(c1 ) 1/4 also holds
for [1/4, 6.2]. Hence
c0
c1
4
for 6.2.
130
H. A. HELFGOTT
( 2 )/
1 1 1/ 2
1/
c0
p
.
(C.42)
p
q
c1
1+1/
2 1 + 1/
1 + 1/
(This is off from optimal by a factor of about 2.) For 200, (C.42) implies
(C.44)
sin
sin 2
5 sin3
4 cos2 2
2 cos2 24 cos6
The left side is positive for all (0, /2], since cos2 /2 1/ 2 and (sin 2)/2
is less than 2/2 = . The right side is negative for > 1 (since it is negative
for = 1, and (sin )/(cos )2 is increasing on ). Hence, it is enough to check
(C.44) for [0, 1]. The two sides of (C.44) are equal for = 0; moreover,
the first four derivatives also match at = 0. We take the fifth derivatives of
both sides; the bisection method (running on [0, 1] with 20 iterations, including
10 initial iterations) gives us that the fifth derivative of the left side minus the
fifth derivative of the right side is always positive on [0, 1] (and minimal at 0,
where it equals 30.5 + O 109 ).
References
[AS64]
[BBO10]
[BM98]
[Bom74]
[Boo06]
[Bou99]
[CW89]
[Dav67]
[dB81]
[Des77]
M. Abramowitz and I. A. Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables, volume 55 of National Bureau of Standards
Applied Mathematics Series. For sale by the Superintendent of Documents, U.S.
Government Printing Office, Washington, D.C., 1964.
J. Bertrand, P. Bertrand, and J.-P. Ovarlez. Mellin transform. In A. D. Poularikas,
editor, Transforms and applications handbook. CRC Press, Boca Raton, FL, 2010.
M. Berz and K. Makino. Verified integration of ODEs and flows using differential algebraic methods on high-order Taylor models. Reliab. Comput., 4(4):361369,
1998.
E. Bombieri. Le grand crible dans la theorie analytique des nombres. Societe
Mathematique de France, Paris, 1974. Avec une sommaire en anglais, Asterisque,
No. 18.
A. R. Booker. Turing and the Riemann hypothesis. Notices Amer. Math. Soc.,
53(10):12081211, 2006.
J. Bourgain. On triples in arithmetic progression. Geom. Funct. Anal., 9(5):968984,
1999.
J. R. Chen and T. Z. Wang. On the Goldbach problem. Acta Math. Sinica,
32(5):702718, 1989.
H. Davenport. Multiplicative number theory, volume 1966 of Lectures given at the
University of Michigan, Winter Term. Markham Publishing Co., Chicago, Ill., 1967.
N. G. de Bruijn. Asymptotic methods in analysis. Dover Publications Inc., New
York, third edition, 1981.
131
[KPS72]
N. I. Klimov, G. Z. Pil tja, and T. A. Septickaja.
An estimate of the absolute
132
H. A. HELFGOTT
[OLBC10] F. W. J. Olver, D. W. Lozier, R. F. Boisvert, and Ch. W. Clark, editors. NIST handbook of mathematical functions. U.S. Department of Commerce National Institute
of Standards and Technology, Washington, DC, 2010. With 1 CD-ROM (Windows,
Macintosh and UNIX).
[Olv58]
F. W. J. Olver. Uniform asymptotic expansions of solutions of linear second-order
differential equations for large values of a parameter. Philos. Trans. Roy. Soc. London. Ser. A, 250:479517, 1958.
[Olv59]
F. W. J. Olver. Uniform asymptotic expansions for Weber parabolic cylinder functions of large orders. J. Res. Nat. Bur. Standards Sect. B, 63B:131169, 1959.
[Olv61]
F. W. J. Olver. Two inequalities for parabolic cylinder functions. Proc. Cambridge
Philos. Soc., 57:811822, 1961.
[Olv65]
F. W. J. Olver. On the asymptotic solution of second-order differential equations
having an irregular singularity of rank one, with an application to Whittaker functions. J. Soc. Indust. Appl. Math. Ser. B Numer. Anal., 2:225243, 1965.
[Olv74]
F. W. J. Olver. Asymptotics and special functions. Academic Press [A subsidiary of
Harcourt Brace Jovanovich, Publishers], New York-London, 1974. Computer Science
and Applied Mathematics.
[Plaa]
D. Platt. Computing (x) analytically. Preprint. Available as arXiv:1203.5712.
[Plab]
D. Platt. Numerical computations concerning grh. Preprint.
[Pla11]
D. Platt. Computing degree 1 L-functions rigorously. PhD thesis, Bristol University,
2011.
[Ram95]
O. Ramare. On Snirel
mans constant. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4),
22(4):645706, 1995.
[Ram09]
O. Ramare. Arithmetical aspects of the large sieve inequality, volume 1 of HarishChandra Research Institute Lecture Notes. Hindustan Book Agency, New Delhi,
2009. With the collaboration of D. S. Ramana.
[Ram10]
O. Ramare. On Bombieris asymptotic sieve. J. Number Theory, 130(5):11551189,
2010.
[Ric01]
J. Richstein. Verifying the Goldbach conjecture up to 4 1014 . Math. Comp.,
70(236):17451749 (electronic), 2001.
[Ros41]
B. Rosser. Explicit bounds for some functions of prime numbers. Amer. J. Math.,
63:211232, 1941.
[RS62]
J. B. Rosser and L. Schoenfeld. Approximate formulas for some functions of prime
numbers. Illinois J. Math., 6:6494, 1962.
[RS75]
J. B. Rosser and L. Schoenfeld. Sharper bounds for the Chebyshev functions (x)
and (x). Math. Comp., 29:243269, 1975. Collection of articles dedicated to Derrick
Henry Lehmer on the occasion of his seventieth birthday.
[RS03]
O. Ramare and Y. Saouter. Short effective intervals containing primes. J. Number
Theory, 98(1):1033, 2003.
[RV83]
H. Riesel and R. C. Vaughan. On sums of primes. Ark. Mat., 21(1):4674, 1983.
[Sch33]
L. Schnirelmann. Uber
additive Eigenschaften von Zahlen. Math. Ann., 107(1):649
690, 1933.
[Sha]
X. Shao. A density version of the Vinogradov three prime theorem. Preprint. Available as arXiv:1206.6139.
[Tao]
T. Tao. Every odd number greater than 1 is the sum of at most five primes. Preprint.
Available as arXiv:1201.6656.
[Tem10]
N. M. Temme. Parabolic cylinder functions. In NIST handbook of mathematical
functions, pages 303319. U.S. Dept. Commerce, Washington, DC, 2010.
[Tru]
T. S. Trudgian. An improved upper bound for the error in the zero-counting formulae
for Dirichlet L-functions and Dedekind zeta-functions. Preprint.
[Tuc11]
W. Tucker. Validated numerics: A short introduction to rigorous computations.
Princeton University Press, Princeton, NJ, 2011.
[Tur53]
A. M. Turing. Some calculations of the Riemann zeta-function. Proc. London Math.
Soc. (3), 3:99117, 1953.
[TV03]
N. M. Temme and R. Vidunas. Parabolic cylinder functions: examples of error
bounds for asymptotic expansions. Anal. Appl. (Singap.), 1(3):265288, 2003.
[Vau77]
R. C. Vaughan. On the estimation of Schnirelmans constant. J. Reine Angew.
Math., 290:93108, 1977.
[Vin37]
[Wed03]
[Whi03]
[Wig20]
[Won01]
133