Escolar Documentos
Profissional Documentos
Cultura Documentos
H. A. HELFGOTT
Abstract. The ternary Goldbach conjecture states that every odd number
n
P 7 is the sum of three primes. The estimation of the Fourier series
px e(p) and related sums has been central to the study of the problem
since Hardy and Littlewood (1923).
Here we show how to estimate such Fourier series for in the so-called
major arcs, i.e., for close to a rational of small denominator. This is part of
the authors proof of the ternary Goldbach conjecture.
In contrast to most previous work on the subject, we will rely on a finite
verification of the Generalized Riemann Hypothesis up to a bounded conductor and bounded height, rather than on zero-free regions. We apply a rigorous
verification due to D. Platt; the results we obtain are both rigorous and unconditional.
The main point of the paper will be the development of estimates on parabolic cylinder functions that make it possible to use smoothing functions based
on the Gaussian. The generality of our explicit formulas will allow us to work
with a wide variety of such functions.
Contents
1. Introduction
1.1. Results
1.2. Main ideas
1.3. Acknowledgments
2. Preliminaries
2.1. Notation
2.2. Dirichlet characters and L functions
2.3. Mellin transforms
3. The Mellin transform of the twisted Gaussian
3.1. How to choose a smoothing function?
3.2. The Mellin transform of the twisted Gaussian
3.3. General approach and situation
3.4. Setup
3.5. The saddle point
3.6. The contour
3.7. The integral over the main contour segment C2
3.8. The integral over the rest of the contour
3.9. Conclusion
4. Explicit formulas
4.1. A general explicit formula
2
4.2. Sums and decay for the Gaussian: (t) = et /2
2
4.3. The case of (t) = t2 et /2 and (t)
4.4. The case of + (t)
4.5. A sum for + (t)2
4.6. A verification of zeros and its consequences
1
2
3
5
6
7
7
7
8
9
11
12
13
13
13
17
18
24
27
30
31
40
44
49
52
56
H. A. HELFGOTT
60
68
69
72
73
74
76
78
1. Introduction
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. Hardy
and Littlewood (1923) were the first to treat the problem by means of the circle
method, i.e., Fourier analysis over Z. (Fourier transforms of functions on Z live
on the circle R/Z.)
I. M. Vinogradov [Vin37] showed in 1937 that the ternary Goldbach conjecture
is true for all n above a large constant C. The main innovation in his work
consisted in the estimation of sums of the form
X
(1.1)
(n)e(n)
nN
for outside the so-called major arcs these being a union of short intervals
in R/Z around the rationals with small denominator. (Here (n) is the von
Mangoldt function, defined as (n) = log p for n a power of a prime p and
(n) = 0 for n having at least two prime factors, whereas e(t) = e2it .)
The estimation of such sums for in the major arcs is also important, and
goes back to Hardy and Littlewood [HL23]. In some ways, their work is rather
modern in particular, it studies a version of (1.1) with smooth truncation:
X
(n)e(n)(n/x),
S (, x) =
n
R+
where :
C is a smooth function; in [HL23], (t) = et .
We will show how to estimate sums such as S (, x) for in the major arcs.
We will see how we can obtain good estimates by using smooth functions based
2
on the Gaussian et /2 . This will involve proving new, fully explicit bounds for
the Mellin transform1 of the twisted Gaussian, or, what is the same, parabolic
cylindrical functions in certain ranges. It will also require explicit formulae that
are general and strong enough, even for moderate values of x.
Any estimate on S (,
P x) for in the major arcs relies on the properties of
L-functions L(s, ) = n (n)ns , where : (Z/qZ) C is a multiplicative
character. In particular, what is key is the location of the zeroes of L(s, ) in
the critical strip 0 (s) 1 (a region in which L(s, ) can be defined by
analytic continuation). In contrast to most previous work, we will not use zerofree regions, which are too narrow for our purposes. Rather, we use a verification
of the Generalized Riemann Hypothesis up to bounded height for all conductors
q 300000 (due to D. Platt [Plab]).
1See 2.3 for definitions.
The bounds we will obtain have shown themselves to be strong enough to prove
the ternary Goldbach conjecture. See [Helb]. A key feature of the present work
is that it allows one to mimic a wide variety of smoothing functions by means
of estimates on the Mellin transform of a single smoothing function here, the
2
Gaussian et /2 .
2
1.1. Results. Write (t) = et /2 . Let us first give a bound for exponential
sums on the primes using as the smooth weight.
Theorem 1.1. Let x be a real number 108 . Let be a primitive character
mod q, 1 q r, where r = 300000.
Then, for any R with || 4r/q,
X
(n/x)2
n e 2 = Iq=1 c
(n)(n)e
() x + E x,
x
n=1
x
q
R
We normalize the Fourier transform fb as follows: fb(t) = e(xt)f (x)dx.
2 2
2e2 .
Of course, c
() is just
As it turns out, smooth weights based on the Gaussian are often better in
applications than the Gaussian itself. Let us give a bound based on (t) =
t2 (t).
2
q
q
x
The advantage of (t) = t2 (t) over is that it vanishes at the origin (to
second order); as we shall see, this makes it is easier to estimate exponential
sums with the smoothing M g, where M is a Mellin convolution and g is
nearly arbitrary. Here is a good example that is used, crucially, in [Helc].
2
q
x
q
H. A. HELFGOTT
(1.3)
if t [0, 2],
otherwise.
sin(H log y)
,
log y
hH (t) = (h M FH )(y) =
dy
y
1
6.18 1012 1.14 1010
+
+
|E|
q
q
x
499100
+ 52 .
251100
.
x
X
2
+
(t) log xt dt + E1 x log x
(n)(log n) 2 (n/x) = x
0
n=1
where
310.84
|E2 | 2 106 +
310.84
.
x
X
(n)(n)e(n/x)(n/x).
(1.6)
S (, x) =
n=1
X
(n)(n)e(n/x)(n/x)
(1.8)
S (/x, x) =
n=1
for varying among all Dirichlet characters modulo q r0 and for || cr0 /q,
i.e., || small. Sums such as (1.8) are estimated using Dirichlet L-functions L(s, )
(see 2.2). An explicit formula gives an expression
X
(1.9)
S, (/x, x) = Iq=1 b()x
F ()x + small error,
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 2.3).
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].
H. A. HELFGOTT
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 (1.9) 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.
Once we have fully explicit estimates for the Mellin transform of the twisted
Gaussian, we are able to use essentially any smoothing function based on the
2
Gaussian (t) = et /2 . As we already saw, we can and will consider smoothing
functions obtained by convolving the twisted Gaussian with another function
and also functions obtained by multiplying the twisted Gaussian with another
function. All we need to do is use an explicit formula of the right kind that
is, a formula that does not assume too much about the smoothing function or
the region of holomorphy of its Mellin transform, but still gives very good error
terms, with simple expressions.
All results here will be based on a single, general explicit formula (Lem. 4.1)
valid for all our purposes. The contribution of the zeros in the critical trip can
be handled in a unified way (Lemmas 4.3 and 4.4). All that has to be done for
each smoothing function is to bound a simple integral (in (4.24)). We then apply
a finite verification of GRH and are done.
1.3. 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. Warm thanks
are 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. Gratitude is also felt towards A. Booker, B. Green, R.
Heath-Brown, H. Kadiri, O. Ramare, T. Tao and M. Watkins, for discussions
on Goldbachs problem and related issues. 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
R
If
is in 1 with respect to dt (i.e., 0 |(x)|x1 dx < ), then the
Mellin transform is defined on the line + iR. Moreover, if (x)x1 is in 1 for
= 1 and for = 2 , where 2 > 1 , then it is easy to see that it is also in
1 for all (1 , 2 ), and that, moreover, the Mellin transform is holomorphic
on {s : 1 < (s) < 2 }. We then say that {s : 1 < (s) < 2 } is a strip of
holomorphy for the Mellin transform.
The Mellin transform becomes a Fourier transform (of (e2v )e2v ) by
means of the change of variables x = e2v . We thus obtain, for example, that
the Mellin transform is an isometry, in the sense that
Z
Z
1
2 2 dx
=
|M f ( + it)|2 dt.
|f (x)| x
(2.2)
x
2
0
(x)x1
Recall that, in the case of the Fourier transform, for |fb|2 = |f |2 to hold, it is
enough that f be in 1 2 . This gives us that, for (2.2) to hold, it is enough that
f (x)x1 be in 1 and f (x)x1/2 be in 2 (again, with respect to dt, in both
cases).
We write f M g for the multiplicative, or Mellin, convolution of f and g:
Z
x dw
.
f (w)g
(2.3)
(f M g)(x) =
w w
0
In general,
(2.4)
M (f M g) = M f M g
and
(2.5)
M (f g)(s) =
1
2i
+i
i
[GR00, 17.32]
provided that z and sz are within the strips on which M f and M g (respectively)
are well-defined.
We also have several useful transformation rules, just as for the Fourier transform. For example,
M (f (t))(s) = (s 1) M f (s 1),
(2.6)
M 2 (s) =
1 2s
s
2
bs a s
,
s
M 4 (s) =
1 2s
s
4
where the next-to-last step holds by contour integration, and the last step holds
by the definition of the Gamma function (s).
3. The Mellin transform of the twisted Gaussian
Our aim in this section is to give fully explicit, yet relatively simple bounds for
2
the Mellin transform F () of e(t) (t), where (t) = et /2 and is arbitrary.
The rapid decay that results will establish that the Gaussian is a very good
choice for a smoothing, particularly when the smoothing has to be twisted by an
additive character e(t).
Gaussian smoothing has been used before in number theory; see, notably,
Heath-Browns well-known paper on the fourth power moment of the Riemann
zeta function [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 and elsewhere from now on.
(There has also been work using the Gaussian after a logarithmic change of
variables; see, in particular, [Leh66]. In that case, the Mellin transform is simply
a Gaussian (as in, e.g., [MV07, Ex. XII.2.9]). However, for non-zero, the Mellin
2
transform of a twist e(t)e(log t) /2 decays very slowly, and thus would not be in
general useful.)
2
|F (s)| C0,, e
| |
()2
0.4798| |
+ C1, e
| |
2 25
min 81 (
) , 32 | |
+ C2,, e
10
H. A. HELFGOTT
where
(3.2)
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|| ,
|F (s)| C,
e 4 | | ,
(3.3)
where
C,
e/2 | |/2
2
1 + 23/2 ||
(/2)
| |
11
result to larger half-planes for a is not too hard integration by parts can be
used to push a to the right.
As we shall see in 3.2, the literature on parabolic cylinder functions is rich
and varied. However, it stopped short of giving fully explicit expressions for a
general and z imaginary. That precluded, for instance, the use of the Gaussian
in explicit work on exponential sums in number theory. Such work will now be
possible.
3.1. How to choose a smoothing function? The method of stationary phase
([Olv74, 4.11], [Won01, II.3])) suggests that the main contribution to the integral
Z
dt
e(t)(t)ts
(3.5)
F (t) =
t
0
should come when the phase has derivative 0. The phase part of (3.5) is
e(t) = t(s) = e2it+ log t
(where we write s = + i ); clearly,
(2t + log t) = 2 +
=0
t
+i 1
ei
(3.6)
(t)e(t)ts1 =
2
2
multiplied by a width approximately equal to a constant divided by
p
|(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 .
12
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
(3.8)
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.)
13
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.
3.3. 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.
3.4. Setup. We write
(3.9)
(u) =
u2
(2i)u i log u
2
e(u) u
du
.
u
u 2i
i
= 0,
u
i.e.,
u2 + iu i = 0,
i 2 + 4i
(3.11)
u0 =
.
2
The second derivative at u0 is
(3.12)
(u0 ) =
1
1
u20 + i = 2 (iu0 + 2i ).
2
u0
u0
Assign the names u0,+ , u0, to the roots in (3.11) 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.
3.5. The saddle point. Let us start by estimating
s y/2
(3.13)
u0,+ e = |u0,+ | e arg(u0,+ ) ey/2 ,
14
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
1 + 2 .
=
i + 2 + 4i
(3.14)
y=
2
2
2
2
2
1+j()
1
(3.16)
2
= arcsin s q
q
1+j()
2 1+j()
1
2
2
v
u
u1
= arcsin t
2
2
= 1 arccos
1 + j()
2
2
1 + j()
s
!!
r
y
2
2
j() + 1
arg(u0,+ ) + = arccos
1
2
1 + j() 2
2
2
(3.17)
2(() 1)
1
,
= arccos
()
2
p
where () = (1 + j())/2.
It is clear that
1
2(() 1)
(3.18)
lim arccos
()
2
whereas
2(() 1)
=
(3.19)
arccos
()
2 4
4
as 0+ .
1
We are still missing a factor of |u0,+ | (from (3.13)),
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 (3.12), this is
|u0,+ |1
p
=
| (u0,+ )|
|u0,+ |1
|u0,+ |
p
p
=
.
1
|iu0,+ + 2i |
|iu0,+ + 2i |
|u0,+ |
15
(3.20)
+
+12
=
2
2
2
2
s
r
p
1 + j()
=
=
1 + j() 2
()2 ().
2
2
2
Proceeding as in (3.14), we obtain that
!
r
i
4i
|iu0,+ + 2i | =
i + 1 + 2 + 2i
2
r
r
2
2 j() + 1 i2 j() 1
= + 2i +
2
2
2
2
2
v
(3.21)
!2
!2
u
r
r
2 u
j() + 1
j() 1
t
1 +
=
+
2
2
2
s
r
r
j() + 1
j() 1
2
2
j() + + 1 2
=
2
.
2
2
2
p
p
Since j() 1 = / j() + 1, this means that
(3.22)
v
s
u
2
u
2
(j() + 1 + 2 )
|iu0,+ + 2i | = tj() + 2 + 1
2
j() + 1
2 p
j() + j()2 (())1 (j() + j()2 )
2
p
2 j() p
2 p
Hence
|
|u0,+
p
=
|iu0,+ + 2i |
p
2 ()
()
1
2
= 1
1
1/4
(j())
2 ())1/4
2 2 4 j() 4
(()
1/4
2
3
1
2 2 4
2
41
2
2
(()2 ()) 2 4
.
1
1 (() ())
2 j() 4
It remains to determine the direction of steepest descent at the saddle-point
u0,+ . Let v C point in that direction. Then, by definition, v 2 (u0,+ ) is real
and positive, where is as in 3.9. Thus arg(v) = arg( (u0,+ ))/2. By (3.12),
arg( (u0,+ )) = arg(iu0,+ + 2i ) 2 arg(u0,+ ).
arg(iu0,+ + 2i ) = arctan
1 +
j1
2
j+1
2
16
H. A. HELFGOTT
and
(3.23)
1 +
j1
2
j+1
2
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 (3.16),
2( + j)
arccos ()1 .
arg( (u0,+ ))
1
2v(v + j)
= arg(u0,+ ) arctan
2
2
= arg(u0,+ ) arctan ,
arg(v) =
where
(3.25)
= tan
1
2v(v + j)
arctan
.
2
Since
1
1
tan =
=
2
sin tan
we see that
(3.26)
1+
1
1
,
tan2 tan
1+ 2
2
4 ( + j)
2( + j)
cos =
2
1 + cos
,
2
sin =
2
1 cos
.
2
Hence, if we let
(3.27)
0 = arg(u0,+ ) =
1
1
arccos
,
2
()
we get that
cos 0 = cos
(3.28)
sin 0 = sin
1
1
arccos
2
()
1
1
arccos
2
()
=
=
arctan > 0 ,
1
1
+
,
2 2()
1
1
.
2 2()
17
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.3 In order for the curve to go through the point u0,+ , we must have
c1 r0
(3.31)
+ c0 = sin 0 ,
where
p
r0 = |u0,+ | =
()2 (),
(3.32)
2
and 0 and sin 0 are as in (3.27) and (3.28). 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
3Because c 1, by (A.14).
0
d
c1 r0
=
,
dr
cos 0
18
H. A. HELFGOTT
c1 =
cos 0
,
r0
.
(3.34)
c1 =
2
2 2
2 2
Incidentally, we have also shown that the arc-length infinitesimal is
s
v
r
u
d 2
(c1 r/)2
r2
u1 +
(3.35) |du| = 1 + r
dr = 1 +
dr
=
2 dr.
t
dr
cos2
c0
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.
3.7. 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
(3.36)
(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.
For u = x + iy (or (r, ) in polar coordinates), (3.36) gives us
(3.37)
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
(3.38)
r r0
,
0 =
r0
.
19
(3.39)
and so
(3.40)
c0 c1 0 = sin 0 .
[0 , 1 ]
1 sin 0 sin 0
,
.
c1
c1
(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
(3.42)
( + 0 )2
2 cos 20
(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
(cos 0 )2k1
k!
k=1
c1
1
(c1 )2 sin 0 2
0
+
+
+ ... ,
=
4 4 cos 0
2(cos 0 )3
2
dr
d
2
2
2
du
du
1
| (u0,+ )| |r=r0 = | (u0,+ )| |r=r0 .
dr
2
dr
j() 2
2
2
2 (
2j()
,
2
20
H. A. HELFGOTT
(3.43)
v
u
|du|
du
u
|r=r =
|
=
t1 +
r=r
0
0
dr
|dr|
=
=
Thus,
1
a2 =
2
2
c21
c2
1 + 12
r02
c0
c1
r0
2
2
2 ( )
1
1
2 + 2
2 =
1+
1 + c21
r02
c21
2 1 sin2 0
2
1 + 1/
1 + 2 .
2j()
(1 + 2 ),
2
where is as in (3.26).
Let us simplify our expression for 0 () somewhat. We can replace the third
series in (3.42) by a truncated Taylor series ending at k = 2, namely,
c1
0
1
(c1 )2 sin 1 2
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 a quadratic with negative leading coefficient, its minimum within [0 , 1 ] (see
(3.41)) is bounded from below by min(R((1 sin 0 )/(c1 )), R((sin 0 )/(c1 ))).
We compare
sin 0
= 2c3 sin 0 sin2 0 ,
R
c1
where c3 = sin 0 c1 0 , and
1 sin 0
R
= 2c3 (1 sin 0 ) (1 sin 0 )2
c1
2
=
2
1 + 1/
,
2
2
21
We conclude that
(3.44)
0 ()
((u0,+ ))
+ 2 ,
2 2
where
(3.45)
1 (c1 )2 sin 0
+ sin2 0 (sin 20 ) .
4 2(cos 0 )3
= a2
1 + 1/
1 (c1
0
2
= 2
3
4 2(cos 0 )
8
=
1
2
1
1 3/2
+
2
2
2
1
2
p
1 1/
2
p
=
2
4 1 + 1/
2
2
q
=
= p
= q
4 2 1
4 j+1 j1
4 2 /4
2
2
1
1
2 1
2 =
sin 20 = 2 sin 0 cos 0 = 2
4 4
2
/2
1
=
=
.
=
2
(j + 1)/2
j+1
2
(3.46)
=
(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
(3.47)
Z 1
1 du
2 2 2
((u0,+ ))
( + 0 )
e
() e
dr d.
0
(We are using (3.37) and (3.44).) Since u0,+ is a solution to equation (3.10), we
see from (3.9) that
!
u20,+
((u0,+ )) =
+ iu0,+ i log u0,+
2
iu0,+ i
1
=
+
+ arg(u0,+ ) = (iu0,+ ) + arg(u0 , +).
2
2
2
arccos
=e
2(1)
1
22
H. A. HELFGOTT
If 1, we can bound
1
(3.48)
( + 0 )
(
01
(0 + 0 )1
if 0,
if < 0,
(1 + 0 )1 if > 0.
By (3.35),
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 (3.48): 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
c
c20
0
= 1+
1
+
=
,
dr
1 (sin 0 c1 1 )2
cos2
where is the value of when = 1 .
Finally, we complete the integral in (3.47), we split it in two (depending on
whether 0 or < 0) and use
Z
Z
/2
1
2
2 2 2
e d = .
e
d
0
0
Therefore,
(3.49)
Z
C2
c20
/2
1+
= () e
01 + (j + 0 )1
2
cos
2(1)
1 ! s
arccos 1
2
2
j
1
c0 e
r0
1+ 1+
=
1+
,
2
0
cos2
arccos
2(1)
1
0 + 0 =
(c0 sin ) c1
c0 sin
r
=
=
.
c1
0 =
.
2
Definition (3.25) implies immediately that 1. Thus, by (3.34),
p
(3.51)
c1 0 = 2(1 + 1/) 2 2,
while, by (A.2),
(3.52)
c1 0 =
23
,
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 A. First
of all, by (A.14), c0 > 1 for all > 0; this assures us that c0 sin > 0, and
so the last expression in (3.53) is well defined. By (3.50), 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
c20
1
,
1+
(3.54)
cos2
(3.55)
= .
4
Thus, by (A.29),
s
s
c20
1
1 + 2c20 p
1+
2
cos
We also get
2
2
7.82843.
c0 sin
1 1/ 2
Finally, by (A.32),
(
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
2(1)
arccos 1
((u))
1
2
e
|u|
|du| = C, e
(3.56)
,
C2
where
(3.57)
1
3.3
3/2
C, = min 2,
1 + max 7.831 , 1.631
min( /, )
for all > 0, > 0 and all . By reflection on the x-axis, the same bound holds
for < 0, < 0 and all . Lastly, (3.56) is also valid for = 0, provided we
replace (3.57) and the exponent of (3.56) by their limits as 0+ .
24
H. A. HELFGOTT
3.8. 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
2
and so
1
(3.58)
r =
c0
.
c1
2
C1
= + ,
i + iwt i log t + i
(3.60)
((tw)) =
2
4
4
2
and, by (A.33),
r
+ 0.076392 + =
0.30556 > 0.4798.
4
4
4
2
Consider first the case 1. Then
Z r
Z r
r
1
2 4
((tw)) 1
e
.
e
t
dt r
e 2 4 dt r
0
(3.61)
Hence, for 1,
(3.62)
C1
u2 /2
e(u)u
/2
s1
du /2 e0.4798 .
By (3.60),
r
u2 /2
r
us wr
((wr )) r
4
e
2
e(u)
|
e
=
e
0
s
|s|
25
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 (3.61),
+2
r
(1 + 2)r
2 +1 + (1 + 2) 2
+
.
We conclude that
2 +1 + (1 + 2) 2 r +
u2 /2
s1
e
e(u)u du
e 2 4
C1
!
1+ 2
2 e0.4798
1+
dx
dx
dt
2 /2
2
2 /2
x
s
t
s
=
e
e(x)x
ex /2 x .
e
e(t)t
(3.64)
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
x
x2 /2
e
x2 + ( 2)x4
2
x
+ ( 2)x4 + ( 2)( 4)x6
and so on. By (A.36),
c0
min
c1
,
4
if 0 2,
if 2 4.
if 4 6,
We conclude that
Z
1 2 25
2 /2
dt
,
min
(
)
t
s
2
32
P min
,
,
e
e
e(t)t
t
4
C3
26
H. A. HELFGOTT
We have left the case < 0 for the very end. In this case, we can afford to use
a straight ray from the origin as our contour of integration. Let C be the ray at
angle /4 from the origin, i.e., y = (tan(/4 ))x, x > 0, where > 0 is
small. Write v = e(/4)i . The integral to be estimated is
Z
2
eu /2 e(u)us1 du.
I=
C
Let us try = 0 first. Much as in (3.59) and (3.60), we obtain, for < 0,
Z t
Z
dt
+ 4 1
4
2
t
dt = e
|I|
e
e||t/ 2 t
t
0
0
! Z
!
(3.65)
2
2
dt
= e 4
et t =
() e 4
||
t
||
0
for > 0. Recall that () 1 for 0 < < 1 (because () = ( + 1) and
() 1 for all [1, 2]; the inequality () 1 for [1, 2] can in turn be
proven by (1) = (2) = 1, (1) < 0 < (2) and the convexity of ()). We
see that, while (3.65) is very good in most cases, it poses problems when either
or is close to 0.
Let us first deal with the issue of small. For general and 0,
Z t2
sin 2t cos( 4 )+( 4 ) 1
t
dt
|I|
e 2
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
e
/2
4
.
2
(/2)e
=
y
2(sin 2)/2
Here we can choose = (arcsin 2/ )/2 (for 2). Then 2 (/2) (2/ ) =
/ , and so
(3.66)
|I|
e 2
e/2 /2
/2
4
(/2) e 4 .
2
(/2)e
2
2(2/ )/2
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
us v
2
u2 /2
eu /2 e(u)
|I| e
e(u) |0
du
s
s
C
Z
2
us
(u + i)eu /2iu du
=
(3.67)
s
C
Z
Z
2
i
2
1
eu /2 e(u)us+1 du +
eu /2 e(u)us du.
=
s C
s C
27
2
||
2
1
( + 2) e 4 +
( + 1) e 4
|I| =
|s| ||
|s| ||
!
4
2
1
+ 2 e 4 .
2
||
1 e/2
|| e/2
(( + 2)/2) (+2)/2 e 4 +
(( + 1)/2) (+1)/2 e 4
|s| 2
|s| 2
e/2 /2
||
1+
e 4
2
for [0, 1], where we are using the facts that (s) for s [1/2, 1] and
(s) 1 for s [1, 2].
|I|
3.9. Conclusion. Summing (3.56) with the bounds obtained in 3.8, we obtain
our final estimate. Recall that we can reduce the case < 0 to the case > 0 by
reflection. This finishes the proof of Theorem 3.1. Let us now extract its main
corollary in effect, a simplified restatement of the theorem.
Proof of Corollary 3.2. Let E() be as in (3.2). Let
(
0.1598
if 1.5,
(3.68)
L() =
0.1065 if < 1.5.
Note that 0.1598 E(1.5), whereas 0.1065 E(1.5)/1.5. We claim that E()
L(): this is so for 1.5 because E() is increasing on , and for 1.19
because of (3.79), and for [1.19, 1.5] by the bisection method (with 20 iterations).
By Thm. 3.1, for 0 k 2 and s = + i with [0, 1] and | |
max(4 2 ||, 100),
|F (s + k)| + |F (k + 1 s)|
is at most
E
(3.69) C0,, e
| |
()2
0.4798
+C1, e
2 25
min 81 (
) , 32 | |
+C2,, e
+C e 4 | | ,
| p
|
|
min |
,
||
| | p
2 (1 + 1.63)
, | | if k = 1,
3.507 min
3/2
||
2
!
| p
2
|
|
min |
,
||
3.251 min
, | |
if k = 2,
2 (1 + 1.632 )
3/2
||
28
H. A. HELFGOTT
and where
(3.70)
C1,
C,
!
k+1
1+ 2
1+
| | 2 ,
| |
e/2 | |
k+1
2
C2,,
||
1
+
| |
1
| | 5 p
min
if k = 0,
,
|
|
|| 4
1
if k = 1,
min | | , 5 p| | + 1 if k = 2,
|| 4
for k = 0,
for k = 1,
for k = 2.
e/2
||
1+ 2
+
max 1 + p , e( 4 0.4798)| |
1+
| |
2
| |
!
(3.71)
| |1/2
1.0242 + 2.406 max 1 + , e( 4 0.4798)| |
2
k+1
2
e0.4798| | .
for k 2.
Yet again by | | max(4 2 ||, 100),
(3.74)
2
, | |
min
0.055((2)1 + (2)2 ) 0.0102
||
if k = 0,
if k = 1,
if k = 2.
| |
()2
| |
2
max e0.1066( ) , e0.1598| | ,
whether | |/()2 is greater than 1.5 or not. We conclude that (3.69) is at most
2)
2)
| |)eL(| |/()
29
if k = 1, and at most
2)
if k = 2. (We see that the error terms coming from (3.73) and the second line
of (3.74) are being absorbed by the least significant digits of the bound from
(3.75).) We simplify matters further by using min((| |/)2 , | |) = (| |/||)2 for
| | < 1.5()2 , and bounding min((| |/)2 , | |) | | for | | 1.5()2 .
Let us take a second look at Thm. 3.1. While, in the present paper, we will
use it only through Corollary 3.2, it seems worthwhile to say a few words more
on the behavior of its bounds.
2
The terms in (3.1) other than C0,, eE(| |/() )| | are usually very small. In
practice, Thm. 3.1 should be applied when | |/2|| is larger than a moderate
constant (say 8) and | | is larger than a somewhat larger constant (say 100). The
assumptions of Cor. 3.2 are thus typical.
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
dt
2
et /2 t = 2/21 (/2)
(3.76)
|F (s)| F0 () =
t
0
for (0, 1]. Alternatively, we could use integration by parts (much as in (3.67)),
followed by the trivial bound:
Z
us
F (s + 2) 2i
2
du =
F (s + 1),
eu /2 e(u)
(3.77)
F (s) =
s
s
s
0
and so
(3.78)
|F (s)|
+2
1
2
+2
2
+1
+2 2
|s|
|2|
+1
2
1 + 2||
2
|s|
8
384
for all > 0. We can also write
(3.79)
(3.80)
E() =
sin 2
,
4
2
4(1 + sin )
30
H. A. HELFGOTT
=
2
4
cos
cos cos2
2 1 cos2
2 sin
=
=
.
cos2
cos2
Thus
(3.83)
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 (3.79).
To obtain (3.80), simply define = /2 ; the desired inequality follows
from the last two steps of (3.83).
4. 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 (t)e(t) 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 choice(s)
of will be made so that F (s) decays rapidly in both cases.)
We will be able to base all of our work on a single, general explicit formula,
namely, Lemma 4.1. This explicit formula has simple error terms given purely in
terms of a few norms of the given smoothing function . We also give a common
framework for estimating the contribution of zeros on the critical strip (Lemmas
4.3 and 4.4).
2
The first example we work out is that of the Gaussian smoothing (t) = et /2 .
We actually do this in part for didactic purposes and in part because of its
likely applicability elsewhere; for our applications, we will always use smoothing
2
2
functions based on tet /2 and t2 et /2 , generally in combination with something
2
else. Since (t) = et /2 does not vanish at t = 0, its Mellin transform has a pole
at s = 0 something that requires some additional work (Lemma 4.2; see also
the proof of Lemma 4.1).
Other than that, for each function (t), all that has to be done is to bound
an integral (from Lemma 4.3) and bound a few norms. Still, both for and
for + , we find a few interesting complications. Since + is defined in terms of
a truncation of a Mellin transform (or, alternatively, in terms of a multiplicative
convolution with a Dirichlet kernel, as in (1.3) and (1.5)), bounding the norms
takes a little work. We leave this to Appendix B. The effect of
of + and +
31
the convolution is then just to delay the decay a shift, in that a rapidly decaying
function f ( ) will get replaced by f ( H), H a constant.
2
The smoothing function is defined as a multiplicative convolution of t2 et /2
2
with something else. Given that we have an explicit formula for t2 et /2 , we
obtain an explicit formula for by what amounts to just exchanging the order
of a sum and an integral; this is an idea valid in general (see (4.54)).
4.1. A general explicit formula. We will prove an explicit formula valid whenever the smoothing and its derivative satisfy rather mild assumptions they
will be assumed to be L2 -integrable and to have strips of definition containing
{s : 1/2 (s) 3/2}, though any strip of the form {s : (s) 1 + }
would do just as well.
(For explicit formulas with different sets of assumptions, see, e.g., [IK04, 5.5]
and [MV07, Ch. 12].)
The main idea in deriving any explicit formula is to start with an expression
giving a sum as integral over a vertical line with an integrand involving a Mellin
transform (here, G (s)) and an L-function (here, L(s, )). We then shift the line
of integration to the left. If stronger assumptions were made (as in Exercise 5 in
[IK04, 5.5]), we could shift the integral all the way to (s) = ; the integral
would then disappear, replaced entirely by a sum over zeros (or even, as in the
same Exercise 5, by a particularly simple integral). Another possibility is to shift
the line only to (s) = 1/2 + for some > 0 but this gives a weaker result,
and at any rate the factor L (s, )/L(s, ) can be large and messy to estimate
within the critical strip 0 < (s) < 1.
Instead, we will shift the line to s = 1/2. We can do this because the
assumptions on and are enough to continue G (s) analytically up to there
(with a possible pole at s = 0). The factor L (s, )/L(s, ) is easy to estimate
for s < 0 and s = 0 (by the functional equation), and the part of the integral
on s = 1/2 coming from G (s) can be estimated easily using the fact that the
Mellin transform is an isometry.
1
+
Lemma 4.1. Let : R+
0 R be in C . Let x R , R. Let be a primitive
character mod q, q 1.
Write G (s) for the Mellin transform of (t)e(t). Assume that (t) and (t)
are in 2 (with respect to the measure dt) and that (t)t1 and (t)t1 are in
1 (again with respect to dt) for all in an open interval containing [1/2, 3/2].
Then
X
X
n=1
1/2
where
if q = 1,
if q 6= 1,
L (1, )
2
+
+ O (c0 )
R = (0) log
q
L(1, )
Iq=1
(4.2)
(
1
=
0
(t)
2 (t)
.
(4.3)
c0 = O + (t) t + 2|| + |(t) t|1
3
1
t
t
1
32
H. A. HELFGOTT
Proof. Since (a) (t)t1 is in 1 for in an open interval containing 3/2 and (b)
(t)e(t) has bounded variation (since , 1 , implying that the derivative of
(t)e(t) is also in 1 ), the Mellin inversion formula (as in, e.g., [IK04, 4.106])
holds:
Z 3 +i
2
1
G (s)xs ns ds.
(n/x)e(n/x) =
2i 3 i
2
P
Since G (s) is bounded for (s) = 3/2 (by (t)t3/21 1 ) and n (n)n3/2
is bounded as well, we can change the order of summation and integration as
follows:
Z 3 +i
X
X
2
1
(n)(n)
(n)(n)e(n/x)(n/x) =
G (s)xs ns ds
3
2i
i
n=1
n=1
2
Z 3 +i X
2
1
(4.4)
(n)(n)G (s)xs ns ds
=
2i 3 i n=1
2
Z 3 +i
2
L (s, )
1
G (s)xs ds.
=
2i 3 i
L(s, )
2
(This is the way the procedure always starts: see, for instance, [HL23, Lemma
1] or, to look at a recent standard reference, [MV07, p. 144]. We are being very
scrupulous about integration because we are working with general .)
The first question we should ask ourselves is: up to where can we extend
G (s)? Since (t)t1 is in 1 for in an open interval I containing [1/2, 3/2],
the transform G (s) is defined for (s) in the same interval I. However, we
also know that the transformation rule M (tf (t))(s) = s M f (s) (see (2.6); by
integration by parts) is valid when s is in the holomorphy strip for both M (tf (t))
and M f . In our case (f (t) = (t)e(t)), this happens when (s) (I 1) I (so
that both sides of the equation in the rule are defined). Hence s G (s) (which
equals s M f (s)) can be analytically continued to (s) in (I 1) I, which is an
open interval containing [1/2, 3/2]. This implies immediately that G (s) can
be analytically continued to the same region, with a possible pole at s = 0.
When does G (s) have a pole at s = 0? This happens when sG (s) is non-zero
at s = 0, i.e., when M (tf (t))(0) 6= 0 for f (t) = (t)e(t). Now
Z
M (tf (t))(0) =
f (t)dt = lim f (t) f (0).
0
f (t)
Here we were able to exchange the limit and the integral because f (t)t is in 1
for in a neighborhood of 0; in turn, this is true because f (t) = (t) + 2i(t)
and (t)t and (t)t are both in 1 for in a neighborhood of 0. In fact, we will
33
use the easy bounds |(t) log t| (2/3)(|(t)t1/2 |1 + |(t)t1/2 |1 ), | (t) log t|
(2/3)(| (t)t1/2 |1 + | (t)t1/2 |1 ), resulting from the inequality
1
2 1
t 2 + t 2 | log t|,
(4.5)
3
valid for all t > 0.
We conclude that the Laurent expansion of G (s) at s = 0 is
(4.6)
G (s) =
(0)
+ c0 + c1 s + . . . ,
s
where
c0 = O (|f (t) log t|1 )
(t)
2 (t)
= O + (t) t + 2 + |(t) t|1
.
3
1
t
t
1
2i 2i
L(s, )
(4.7)
Z 1/2+i
L (s, )
1
G (s)xs ds,
2i 1/2i L(s, )
where
R = Ress=0
L (s, )
G (s).
L(s, )
Of course,
G (1) = M ((t)e(t))(1) =
(t)e(t)dt = b().
(4.8)
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
(4.9)
2
2
2
2
2
Hence, unless q = 1, the Laurent expansion of L (s, )/L(s, ) at s = 0 is
1
2
L (1, )
a1 a2
+ log
(1)
+ 2 + ... .
+
s
q
L(1, )
s
s
34
H. A. HELFGOTT
(4.10)
= log z
+ O
2z 9
e2t 1
0
8
1
1
+ O
(2)(2)
= log z
2z 9
(2)2
1
1
10
+O
= log z + O
.
= log z
2z
27
27
e 4 2 + e 4 i+ 2
(s
+
)
cot
=
e 4 i 2 e 4 i+ 2 = 1.
2
2
2
(4.12)
9
q 1
+ 4.1396.
log + log 2 +
2
4
Recall that we must bound the integral on the right side of (4.7). The absolute
value of the integral is at most x1/2 times
Z 1 +i
L (s, )
2
1
ds.
G
(s)
(4.13)
2 1 i L(s, )
2
35
u 1
u 1
L (s, ) 1
2
2
t
t
|G (s)s|2 |ds|
|ds|
1
1
2 i L(s, ) s
2 i
2
By (4.12),
v
v
uZ 1 +i
uZ 1 +i
2
u
u
L (s, ) 1
log q 2
2
2
t
t
L(s, ) s |ds|
s |ds|
1
1
2 i
2 i
v
uZ
u 1 log 2 + 9 + 4.1396 + log 2
t
2
4
+
d
1
2
4 +
2 log q + 226.844,
u 1
2
2
t
|2ite(t)(t) te(t) (t)|2 t2 dt
|G (s)s| |ds| =
2 1 i
0
2
sZ
sZ
(4.14)
= 2||
226.844
2
|(t)|2 dt +
| |2 + 2||||2 .
| (t)|2 dt.
Lemma 4.1 leaves us with three tasks: bounding the sum of G ()x over all
non-trivial zeroes with small imaginary part, bounding the sum of G ()x over
all non-trivial zeroes with large imaginary part, and bounding L (1, )/L(1, ).
Let us start with the last task: while, in a narrow sense, it is optional in
that, in most of our applications, we will have (0) = 0, thus making the term
L (1, )/L(1, ) disappear it is also very easy and can be dealt with quickly.
Since we will be using a finite GRH check in all later applications, we might
as well use it here.
Lemma 4.2. Let be a primitive character mod q, q > 1. Assume that all
non-trivial zeroes = + it of L(s, ) with |t| 5/8 satisfy () = 1/2. Then
L (1, ) 5
L(1, ) 2 log M (q) + c,
36
H. A. HELFGOTT
P
where M (q) = maxn mn (m) and
2 3
c = 5 log
= 15.07016 . . . .
(9/4)/(9/8)
Proof. 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 :
|s s0 | R} of radius R > 0 around s0 satisfies
f (s)
2R log M/|f (s0 )|
=O
(4.16)
f (s)
(R r)2
for all s with |s s0 | r, where 0 < r < R and M is the maximum of |f (z)|
on Cs0 ,R . Assuming L(s, ) has no non-trivial zeros off the critical line with
|(s)| H, where H > 1/2, we set s0 = 1/2 + H, r = H 1/2, and let R H .
We obtain
maxsCs0 ,H |L(s, )|
L (1, )
(4.17)
= O 8H log
.
L(1, )
|L(s0 , )|
Now
|L(s0 , )|
Y
Y (1 p2s0 )1
(2s0 )
.
=
(1 + ps0 )1 =
s
1
0
(1 p )
(s0 )
p
p
Since s0 = 1/2 + H, Cs0 ,H is contained in {s C : (s) > 1/2} for any value of
H. We choose (somewhat arbitrarily) H = 5/8.
By partial summation, for s = + it with 1/2 < 1 and any N Z+ ,
X
X
(m) (N + 1)s
(m)ns
L(s, ) =
nN
(4.18)
nN +1
mN
mn
!
N 11/2
1
,
+N
+ M (q)N
=O
1 1/2
P
where M (q) = maxn mn (m). We set N = M (q)/3, and obtain
p
(4.19)
|L(s, )| 2M (q)N 1/2 = 2 3 M (q).
Let M (q) be as in the statement of Lem. 4.2. Since the sum of (n) ( mod q,
q > 1) over any interval of length q is 0, it is easy to see that M (q) q/2. We
also have the following explicit version of the P
olya-Vinogradov inequality:
(
2
q log q + 42 q log log q + 23 q if (1) = 1,
2
(4.20)
M (q) 1
1
if (1) = 1.
2 q log q + q log log q + q
Taken together with M (q) q/2, this implies that
(4.21)
M (q) q 4/5
37
(4.22)
for all q 1.
Notice, lastly, that
2
log
log q + log e 2
+
q
32
for all q 3. (There are no primitive characters modulo 2, so we can omit q = 2.)
We conclude that, for primitive and non-trivial,
(1, )
2
L
log
log e 2 + log q + 5 log q 54 + 15.07017
+
q
L(1, )
32
2
3 log q + 15.289.
Obviously, 15.289 is more than log 2, the bound for trivial. Hence, the absolute
value of the quantity R in the statement of Lemma 4.1 is at most
(4.23)
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, ). This is little more than partial summation, given a (classical)
bound for the number of zeroes N (T, ) of L(s, ) with |(s)| T . The error
term becomes particularly simple if f is real-valued and decreasing; the statement
is then practically identical to that of [Leh66, Lemma 1] (for principal), except
for the fact that the error term is improved here.
Lemma 4.3. Let f : R+ C be piecewise C 1 . Assume limt f (t)t log t = 0.
Let be a primitive character mod q, q 1; let denote the non-trivial zeros
of L(s, ). Then, for any y 1,
Z
X
1
qT
f (()) =
dT
f (T ) log
2 y
2
(4.24)
where
(4.25)
non-trivial
()>y
Z
1
f (T ) g (T )dT ,
+ O |f (y)|g (y) +
2
y
g (T ) = 0.5 log qT + 17.7
38
H. A. HELFGOTT
Proof. Write N (T, ) for the number of non-trivial zeros of L(s, ) with |(s)|
T . Write N + (T, ) for the number of (necessarily non-trivial) zeros of L(s, )
with 0 < (s) T . Then, for any f : R+ C with f piecewise differentiable
and limt f (t)N (T, ) = 0,
Z
X
f (T ) dN + (T, )
f (()) =
y
:()>y
1
=
2
Now, by [Ros41, Thms. 1719] and [McC84, Thm. 2.1] (see also [Tru, Thm. 1]),
qT
T
+ O (g (T ))
(4.26)
N (T, ) = log
2e
for T 1, where g (T ) is as in (4.25). (This is a classical formula; the references
serve to prove the explicit form (4.25) 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
(4.27)
Z
1
+ O |f (y)|g (y) +
f (T ) g (T )dT .
2
y
Here
(4.28)
f (T )
T
qT
y
qy
log
log
2e
2e
dT =
1
2
f (T ) log
qT
dT.
2
= 0.5
since
g (T )
f (T )
dT,
T
39
is at most
(4.29)
p
p
dt
s dt
(log t)(t)t1
|G (s)| =
|(log
t)(t)|t
=
(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 (4.26) and (4.25),
q
1
+ O (0.5 log q + 17.7) = O (0.819 log q + 16.8).
N (1, ) = log
2e
Therefore,
X
|G ()| (t)t1/2 (0.819 log q + 16.8).
1
non-trivial
|()|1
Let us now consider zeros with |()| > 1. Apply Lemma 4.3 with y = 1 and
(
|G (1/2 + it)| if t T0 ,
f (t) =
0
if t > T0 .
:1<|()|T0
T0
qT
dT
2
1
Z
+ O |f (1)|g (1) +
|f (T )| g (T ) dT ,
f (()) =
f (T ) log
Now
2
Z
Z
Z
1
1
1 T0
2
dT
G
|e(t)(t)|2 dt = ||22
|f (T )| dT
+
iT
1
2
2
0
by Plancherel (as in (2.2)). We also have
Z T0
Z qT0
2
qT 2
2
(log t)2 dt
log
dT
2
q 0
1
Hence
1
T0
qT
dT
f (T ) log
2
s
qT0
log
2e
2
qT0
log
2e
2
+ 1 ||2
+1
T0 .
T0 .
40
H. A. HELFGOTT
Again by Cauchy-Schwarz,
Z
|f (T )| g (T ) dT
1
2
|f (T )|2 dT
T0
1
|g (T )|2 dT .
T0
Summing, we obtain
Z
Z
1 T0
qT
|f (T )| g (T ) dT
f (T ) log
dT +
1
2
1
p
1
qT0
log qT0
T0
+
+ 17.21 |(t)(log t)|2
log
||2 +
2e 2
2
Finally, by (4.30) and (4.25),
|f (1)|g (1) (t)/ t (0.5 log q + 17.7).
1
By (4.32) 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
4.2. Sums and decay for the Gaussian: (t) = et /2 . It is now time to
derive our bounds for the Gaussian smoothing. Thanks to the general work we
have donePso far, there is really only one thing left to do, namely, an estimate for
the sum |F ()| over non-trivial zeros with |()| > T0 .
2
41
non-trivial
|()|>T0
|F ()| =
non-trivial
()>T0
by the functional equation (which implies that non-trivial zeros come in pairs ,
1 ). Hence, by Cor. 3.2,
X
X
f (()),
|F ()|
(4.33)
non-trivial
|()|>T0
where
(4.34)
non-trivial
()>T0
(
2
e0.1065( )
f ( ) = 4.226
e0.1598| |
if | | < 23 ()2 ,
if | | 32 ()2 .
2 T0
t
2c T0
1 1
q
1
+
+
ecT0
log T0 + log
=
2c
2
c
2 T0
1.0242 log
qT0 cT0
e
.
2
42
H. A. HELFGOTT
Now let us deal with the Gaussian term. (It appears only if T0 < (3/2)()2 ,
as otherwise | | (3/2)()2 holds whenever | | T0 .) For any y e, c 0,
Z
(4.37)
ct2
(4.38)
(4.39)
1
dt =
c
ect
dt =
t
ct2
(log t)e
dt
t2
cy
cy 2
1
dt
cy
t2
te
cy
ecy
dt
,
2cy
et
E1 (cy 2 )
ecy
dt =
,
2t
2
2cy 2
log t 1
log t +
2ct2
ect dt =
log y cy2
.
e
2cy
Hence
Z
T0
=
(4.40)
T0
||
1
qT
1
log
+
dT
2
2
4T
q||t
1
||
0.1065t2
log
+
e
dt
2
2
4t
T
0.1065(
)
||
2
T0
log ||
T0
2c ||
||
2
q||
2
log
T0
2c ||
+
8c
1
T0
||
c
2 e
T0
||
2
0.0057 log
8T0
400
2
2
Thus, the last line of (4.40) is less than
(4.41)
1.0057
||
2
0
log qT
2
2cT0
||
T0
||
2
2 c
= 1.0057
e
4cT0
T0
||
2
0.1598T0
4.329e
T0 2
0.1065 ||
+ 0.8015||e
log
qT0
.
2
2
43
t2
2
,
e dt =
| |2 =
2
0
t2 /2 2
2
,
(te
) dt =
| |2 =
4
0
Z
2
2
et (log t)2 dt
| log |2 =
0
2
=
+ 2 2 + 8 log 2 + 8(log 2)2 1.94753,
16
Z t2 /2
e
(1/4)
dt = 3/4 2.15581
| (t)/ t|1 =
2
t
Z0
t2
(3/4)
We can now state what is really our main result for the Gaussian smoothing.
(The version in the introduction will, as we shall later see, follow from this, given
numerical inputs.)
2
X
b()x + O (err, (, x)) x if q = 1,
n
=
n
(n)(n)e
x
x
O (err, (, x)) x
if q > 1,
n=1
where
T0 2
qT0
0.1065 ||
0.1598T0
err, (, x) = log
4.329e
+ 0.802||e
2
p
p
1
+ (2.337 T0 log qT0 + 21.817 T0 + 2.85 log q + 74.38)x 2
+ (3 log q + 14|| + 17)x1 + (log q + 6) (1 + 5||) x3/2 .
Proof. Let F (s) be the Mellin transform of (t)e(t). By Lemmas 4.4 (with
G = F ) and 4.5,
X
F ()x
non-trivial
44
H. A. HELFGOTT
Let us now apply Lemma 4.1. We saw that the value of R in Lemma 4.1 is
bounded by (4.23). We know that (0) = 1. Again by (4.42), we get from (4.3)
that c0 1.4056 + 13.3466||. Hence
|R| 3 log q + 13.347|| + 16.695.
Lastly,
Clearly
2 /2
(t) =
2 /2
t2 et
0
if t 0,
if t < 0.
The fact that this vanishes at t = 0 actually makes it easier to work with at
several levels.
Its Mellin transform is just a shift of that of the Gaussian. Write
t2
F (s) = (M (e 2 e(t)))(s),
(4.44)
G (s) = (M ((t)e(t)))(s).
non-trivial
()>T0
where we are using G () = F ( + 2) and the fact that non-trivial zeros come in
pairs , 1 .
By Cor. 3.2 with k = 2,
X
X
f (()),
|G ()|
non-trivial
|()|>T0
non-trivial
()>T0
where
(4.45)
2
2
|
0.1065 |
|
e
| |e0.1598| | + 41 |
f ( ) = c2
| |e0.1598| |
45
if | | < 23 ()2 ,
if | | 32 ()2 ,
where c2 = 3.262. We are including the term c2 | |e0.1598| | in both cases in part
because we will not bother to take it out (just as we did not bother in the proof
of Lem. 4.5) and in part to ensure that f ( ) is a decreasing function of for
T0 .
We can now apply Lemma 4.3. We obtain, again,
Z
X
1
qT
1
log
+
f (())
f (T )
(4.46)
dT.
2
2
4T
T0
non-trivial
()>T0
+ 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.
Now let us examine the Gaussian term. First of all when does it arise? If
T0 (3/2)()2 , then | | (3/2)()2 holds whenever | | T0 , and so (4.45)
does not give us a Gaussian term. Recall that T0 4 2 ||, which means that
|| 8/3 implies that T0 (3/2)()2 . We can thus assume from now on that
|| > 8/3, since otherwise there is no Gaussian term to treat.
For any y 1, c, c1 > 0,
Z
Z
y
1
1
2
ct2
2 ct2
2
+ 2
dt =
ecy ,
dt <
t e
t + 2 2 e
4c t
2c 4c y
y
y
46
H. A. HELFGOTT
ct2
(t log t + c1 t) e
where
a=
at log et log et
a
dt
t log t +
2
2c
2c
4c t
y
(2cy + a) log y + a cy2
,
e
=
4c2
log ey
2c
y log ey
1
2c
4c2 y
c1 y +
ect dt
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
3
+ 40.106512 (4)2
4 16
1
a
0.092.
+ 4 log 4e
1
4
2
20.1065
40.1065 4
is at most e
(4.49)
times
q||
T0
eT0
(0.61T0 + 0.716||) log
+ 0.61T0 log
+ 0.827|| log
2
||
||
!
1
1 + log T0 /||
qT0
qT0
0.64T0 log
,
0.61 + 0.828
T0 log
T0 /||
2
2
21/4 (1/4)
3
|(t)/ t|1 =
1.07791,
|(t) t|1 = 23/4 (3/4) 1.54568,
4
4
Z 2
Z
2
2
t
t
| (t)/ t|1 =
t3/2 e 2 dt t3/2 e 2 dt 1.48469,
0
2
47
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
(4.52)
!
T02
qT0
0.1065
()2
3.5e0.1598T0 + 0.64e
err, (, x) = T0 log
2
p
p
+ 1.22 T0 log qT0 + 5.053 T0 + 1.423 log q + 37.19 x1/2
+ (3 + 11||)x1 + (log q + 6) (1 + 6||) x3/2 .
Proof. We proceed as in the proof of Prop. 4.6. The contribution of Lemma 4.7
is
!
T2
qT0
0.1065 0 2
0.1598T0
()
x,
3.5e
+ 0.64e
T0 log
2
whereas the contribution of Lemma 4.4 is at most
p
p
Now that we have Prop. 4.8, we can derive from it similar bounds for a smoothing defined as the multiplicative convolution of with something else. In general,
for 1 , 2 : [0, ) C, if we know how to bound sums of the form
X
(4.53)
Sf,1 (x) =
f (n)1 (n/x),
n
we can bound sums of the form Sf,1 M 2 , simply by changing the order of summation and integration:
n
X
Sf,1 M 2 =
f (n)(1 M 2 )
x
n
Z X
Z
(4.54)
n
dw
dw
f (n)1
=
2 (w)
Sf,1 (wx)2 (w) .
=
wx
w
w
0
0
n
48
H. A. HELFGOTT
2
where
(4.56)
!
T2
qT0
0.1065 0 2
0.1598T0
()
3.5e
+ 0.0019 e
err, (, x) = T0 log
2
p
p
1
+ 1.675 T0 log qT0 + 6.936 T0 + 1.954 log q + 51.047 x 2
+ (6 + 22||)x1 + (log q + 6) (3 + 17||) 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. 4.8, 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)dt 2 (w)dw = x
w
w
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, wx) wx 2 (w)
err, (w, wx)2 (w)dw.
(4.57)
1
1
w
4
Since
and
Z
w
2 (w)dw = 1,
w
0.1065(4)2
e
w
1
w2
1
2 (w)dw 0.002866,
49
(4.58)
2
T
q
X
0.1065 0
log qT0 ,
|G ()| 9.462 T0 e0.1598T0 + 11.287||e
2
non-trivial
|()|>T0
where T0 = T0 H.
Proof. As usual,
non-trivial
|()|>T0
|G ()| =
non-trivial
()>T0
50
H. A. HELFGOTT
where c1 = 3.516. (As in the proof of Lemma 4.7, we are really putting in extra
terms so as to simplify our integrals.)
From (4.59), we conclude that
|G ()| + |G (1 )| f ( ),
for = + i , > 0, where
|M h(ir)|1
g( H)
2
is decreasing for T0 (because g( ) is decreasing for T0 H). By (B.17),
|M h(ir)|1 16.193918.
We apply Lemma 4.3, and get that
Z
X
1
qT
1
|G ()|
f (T )
log
+
dT
2
2
4T
T0
non-trivial
|()|>T0
(4.61)
Z
|M h(ir)|1
qT
1
1
=
log
+
dT.
g(T H)
2
2
2
4T
T0
f ( ) =
Now we just need to estimate some integrals. For any y e2 , c > 0 and
, 1 0,
Z
ct
y
1
te dt
+ 2
ecy ,
c
2c
y
y
Z
y
1
a
ct
t log(t + ) + e dt
+ 2
log(y + )ecy ,
c
c
y
t
y
where
a=
1 + c1
1
+
.
2 log(y + )
+
e
,
log
2
0.1598
2
0.15982 T0 H
51
Proceeding just as before, we see that the contribution of the Gaussian term in
(4.60) to (4.61) is at most
(4.63)
Z
c1 |M h(ir)|1 1
qT
1 T H 0.1065( T H )2
log
+
e
dT
2
2
2
4T
2||
T0
Z
q|| /2
2
||
H
+
+ log
T e0.1065T dT
9.06194
log T +
4 T0 H
||
2
T
||
||
H 2
+
qT0 0.1065 T0
||
2
20.1065(T0 H)
e
1+
,
log
9.06194
T0 H
T0
8 0.1065
2
|| log ||
qT0 0.1065
e
11.287|| log
2
T0 H
2
.
X
c
+ ()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
(4.65)
q
0.1065
T0
2
log qT0
2
p
p
+ (1.631 T0 log qT0 + 12.42 T0 + 1.321 log q + 34.51)x1/2 ,
+ (9 + 11||)x1 + (log q)(11 + 6||)x3/2 ,
where T0 = T0 H.
Proof. We can apply Lemmas 4.1 and Lemma 4.4 because + (t), (log t)+ (t) and
(t) are in (by (B.25), (B.28) and (B.32)) and (t)t1 and (t)t1 are in
+
2
+
+
1 for in an open interval containing [1/2, 3/2] (by (B.30) and (B.33)). (Because
of (4.5), the fact that + (t)t1/2 and + (t)t1/2 are in 1 implies that + (t) log t is
also in 1 , as is required by Lemma 4.4.)
We apply Lemmas 4.1, 4.4 and 4.10. We bound the norms involving + using
the estimates in B.3 and B.4. Since + (0) = 0 (by the definition (B.3) of + ),
the term R in (4.2) is at most c0 , where c0 is as in (4.3). We bound
p
p
p
p
2
2.922875
(1/2) + (3/2) + 1.062319
(5/2) + (7/2)
c0
3
p
p
4
(3/2) + (5/2) 6.536232 + 9.319578||
+
|| 1.062319
3
using (B.30) and (B.33). By (B.25), (B.32) and the assumption H 50,
|+ |2 0.80044,
|+
|2 10.845789.
52
H. A. HELFGOTT
where we estimate the norms |+ |2 , | log |2 and |(t)/ t|1 by (B.25), (B.28)
and (B.30).
4.5. A sum for + (t)2 . Using a smoothing function sometimes leads to considering sums involving the square of the smoothing function. In particular, [Hela]
2 something that could be slightly challenging to
requires a result involving +
prove, given the way in which + is defined. Fortunately, we have bounds on
|+ | and other -norms (see Appendix B.5). Our task will also be made easier
by the fact that we do not have a phase e(n/x) this time. All in all, this will be
yet another demonstration of the generality of the framework developed in 4.1.
Proposition 4.12. Let = + be as in (4.58), H 50. Let x 108 . Assume
that all non-trivial zeros of the Riemann zeta function (s) with |()| T0 lie
on the critical line, where T0 2H + max(H/4, 50).
Then
Z
X
2
2
(t) log xt dt + O (err2 ,+ ) x log x,
+
(n)(log n)+ (n/x) = x
(4.67)
0
n=1
where
(4.68)
err2 ,+
p
(T0 2H)
(log T0 )2
4
+ 1.21(log T0 )
T0 e
= 0.607
log x
p
+ (2.06 T0 log T0 + 43.87) x1/2
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) +
n=1
n=1
2 /2
(n)+,1 (n/x).
n=1
2
+,1 (r)
+,0 (r) = h2H (t)tet ,
2
(log x)+,0 + +,1 = + (t) log xt.
Let +,2 =
We wish to apply Lemma 4.1. For this, we must first check that some norms
are finite. Clearly,
2
2
+,2 (t) = +
(t) log x + +
(t) log t
Thus, we see that +,2 (t) is in 2 because + (t) is in 2 and + (t), + (t) log t are
both in (see (B.25), (B.38), (B.40)):
2
2
|+,2 (t)|2 +
(t)2 log x + +
(t) log t2
(4.69)
|+ | |+ |2 log x + |+ (t) log t| |+ |2 .
53
log x + 2+ (t)+
+ +
+,2 (t) 2+ (t)+
(t)
(t)
log
t
(t)/t
2
2
2
2
2 |+ | + log x + 2 |+ (t) log t| + + |+ (t)/t| |+ | .
(t)t1
1
+ (t)t
(B.33), and + (t), + (t) log t, + (t)/t are all in ).
We now apply Lemma 4.1 with q = 1, = 0. Since +,2 (0) = 0, the residue
term R equals c0 , which, in turn, is at most 2/3 times
(|+ | log x + |+ (t) log t| ) + (t)/ t + + (t) t
1
1
+ 2 (|+ | log x + |+ (t) log t| ) + (t)/ t + + (t) t
1
1
+ |+ (t)/t| + (t)/ t + + (t) t .
1
Using the bounds (B.38), (B.40), (B.41) (with the assumption H 50), (B.30)
and (B.33), we get that this means that
c0 18.15014 log x + 7.84532.
Since q = 1 and = 0, we get from (4.70) that
1/2
+
2||
|
|
(log q + 6.01) +,2
+,2 2 x
2
1/2
x
= 6.01 +,2
(162.56 log x + 59.325)x1/2 .
2
We now apply Lemma 4.4 as we may, because of the finiteness of the norms
we have already checked, together with
(4.72)
2
2
|+,2 (t) log t|2 +
(t) log t2 log x + +
(t)(log t)2 2
|+ (t) log t| (|+ (t)|2 log x + |+ (t) log t|2 )
(by (B.40), (B.25) and (B.28); use the assumption H 50). We will also need
the bounds
(4.73)
(from (4.69), by the norm bounds (B.38), (B.40) and (B.25), all with H 50)
and
+,2 (t)/ t (|+ (t)| log x + |+ (t) log t| ) + (t)/ t
1
1
(4.74)
1.24703 log x + 0.40745
(by
P (B.38), (B.40) (again with H 50) and (B.30)). We obtain that the sum
|G0 ()|x (where G0 () = M +,2 ()) over all non-trivial zeros with |()|
54
H. A. HELFGOTT
p
p
(1.3221 log x + 0.6621) T0 log T0 + (3.2419 log x + 5.0188) T0
+ 43.1 log x + 14.1,
Note that the term 18.539 log x from (4.71) is at most 0.002x1/2 log x.
It remains to bound the sum of M +,2 () over non-trivial zeros with |()| >
T0 . This we will do, as usual, by Lemma 4.3. For that, we will need to bound
M +,2 () for in the critical strip.
2
2
The Mellin transform of et is (s/2)/2, and so the Mellin transform of tet
2
is ((s + 1)/2)/2. By (2.6), this implies that the Mellin transform of (log t)tet
is ((s + 1)/2)/4. Hence, by (2.5),
Z
1
M (h2H )(ir) Fx (s ir) dr,
(4.76)
M +,2 (s) =
4
where
(4.77)
Fx (s) = (log x)
s+1
2
1
+
2
s+1
2
Moreover,
(4.78)
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 41.73727 by (B.17).
Hence
Z
1
|M +,2 (s)|
|M (h2H )(ir)|dr max |Fx (s ir)|
4
|r|2H
(4.79)
41.73727
(4.80)
|(s)| 2|s|(s)1/2 e|(s)|/2 e1/6|z| ,
and so
(4.81)
p
|(s)| 2.526 |(s)|e|(s)|/2
for s C with 0 < (s) 1 and |(s)| 25. Moreover, by [OLBC10, 5.11.2]
and the remarks at the beginning of [OLBC10, 5.11(ii)],
1
1
1
(s)
= log s
+ O
(s)
2s
12|s|2 cos3 /2
55
for | arg(s)| < ( (, )). Again, for s = + i with 0 < 1 and | | 25,
this gives us
p
| |2 + 1
1
+O
+ O
| |
2| |
1
O (0.236)
1
+
= log | | + O
+
2| |2 2| |
| |2
0.53
.
= log | | + O
| |
(s)
= log | | + log
(s)
12| |2 cos3
1
arctan | |
2
1
1
0.53
s+1
2
2
2
| /2|
2
p
1
2.526((log x) + log | | 0.335) | |e| |/2 .
2
(4.82)
1
f (T ) = 8.39 log x + log T
2
(4.83)
r
(| |2H)
| |
4
H e
.
2
The functions t 7 tet/2 and t 7 (log t) tet/2 are decreasing for t 3/;
setting t = T /2 H, we see that the right side of (4.83) is a decreasing function
of T for T T0 , since T0 /2 H 25 > 3/.
We can now apply Lemma 4.3, and get that
X
(4.84)
non-trivial
|()|>T0
|M +,2 ()|
f (T )
T0
1
T
1
log
+
2
2 4T
dT.
Since T T0 150 > 2, we know that ((1/2) log(T /2)+1/4T ) (1/2) log T .
Hence, the right side of (4.84) is at most
8.39
2
(4.85)
T0
(log T )2
(log x)(log T ) +
2
In general, for T0 e2 ,
r
T (T 2H)
4
dT.
e
2
T0
4 p
2/
T0 (log T0 )2 + (log e2 T0 )2 e 4 ,
T0
T
p
Z0
T0
4
2/
2
T /4
T (log T )e
dT
T0 (log T0 ) + log e T0 e 4 ;
T0
T0
T (log T )2 eT /4 dT
56
H. A. HELFGOTT
for T0 150,
the
quantities
on
the
right
are
at
most
1.284
T0 (log T0 )2 eT0 /4
and 1.281 T0 (log T0 )eT0 /4 , respectively. Thus, (4.84) and (4.85) give us that
X
|M +,2 ()|
non-trivial
|()|>T0
p
(T0 2H)
1.284
2
4
(log T0 ) + 1.281(log x)(log T0 )
T0 e
2
p
(T0 2H)
4
(0.607(log T0 )2 + 1.21(log x)(log T0 )) T0 e
.
8.39
2 2
The method used was rigorous; its implementation uses interval arithmetic.
Let us see what this verification gives us when used as an input to Prop. 4.6.
We are interested in bounds on | err, (, x)| for q r and || 0 r/2q. We set
r = 3 105 and 0 = 8, and so || 4r/q. (We will not be using the verification
for q even with 3 105 < q 4 105 .)
We let T0 = 108 /q. Thus,
(4.86)
1000
108
,
=
5
3 10
3
108 /q
1000
T0
=
||
4r/q
12
T0
and so
4.329e0.1598T0 3.184 1023 ,
0.1065
0.802e
T02
()2
4.3166 1033 .
Since || 4r/q 1.2 106 /q 1.2 106 and qT0 108 , this gives us
!
T2
qT0
8.59 1026
0.1065 0 2
0.1598T0
()
log
5.28 1022 +
4.329e
+ 0.802||e
2
q
5.281 1022 .
Again by T0 = 108 /q,
p
p
2.337 T0 log qT0 + 21.817 T0 + 2.85 log q + 74.38
is at most
648662
+ 111,
57
and
1.7 107
,
q
1.2 108
.
(log q + 6) (1 + 5||) 19 +
q
Hence, assuming x 108 to simplify, we see that Prop. 4.6 gives us that
3 log q + 14|| + 17 55 +
22
648662
+ 111
55 +
1.7107
q
+
x
x
1
650400
5.281 1022 +
+ 112
x
q
err, (, x) 5.281 10
1.2108
q
3/2
x
19 +
3.5e0.1598T0 + 0.64e
.
T0 log
2
q
We use the same bound when we have 0.0019 instead of 0.64 on the left side,
as in (4.56). (The coefficient affects what is by far the smaller term, so we are
wasting nothing.) Again by T0 = 108 /q and q r,
p
p
275263
1.22 T0 log qT0 + 5.053 T0 + 1.423 log q + 37.19
+ 55.2
q
p
p
377907
+ 75.7.
1.675 T0 log qT0 + 6.936 T0 + 1.954 log q + 51.047
q
For x 108 , we use || 4r/q 1.2 106 /q to bound
1322
1
3/2
x1/2 .
(3 + 11||)x + (log q + 6) (1 + 6||) x
0.0004 +
q
2644
1
3/2
(6 + 22||)x + (log q + 6) (3 + 17||) x
0.0007 +
x1/2 .
q
Summing, we obtain
1
276600
4.269 1014
+
err,
+ 56
q
x
q
2 /2
for (t) = t2 et
and
err,
1
4.269 1014
+
q
x
380600
+ 76
for (t) = t2 et /2 M 2 (t). This proves Theorem 1.2 and Corollary 1.3.
Now let us work with the smoothing weight + . This time around, set r =
150000 if q is odd, and r = 300000 if q is even. As before, we assume
q r,
|| 4r/q.
We can see that Platts verification [Plab], mentioned before, allows us to take
250r
,
H = 200,
T0 = H +
q
58
H. A. HELFGOTT
since Tq is always at least this (Tq = 108 /q > 200 + 3.75 107 /q for q 150000
odd, Tq 200 + 7.5 107 /q for q 300000 even).
Thus,
250r
T0 H
= 250,
r
250r
250
T0 H
= 19.89436 . . .
q
4
and also
T0 200 + 250 150000 3.751 107 ,
Hence
(T H)2
p
0.1065 0 2
0.1598(T0 H)
()
9.462 T0 He
+ 11.287||e
r
250r 4r
4.2259 1017
+
5.57888 1018
q
q
6.6947 1012
3.6598 1013
+
.
q
q
3.6598 1013
6.6947 1012
+
q
q
s
!
1.35 108
1
8
+ 1.321 log 300000 + 34.51 x 2
+ 1.631 log 1.35 10 + 12.42
q
1.2 106
1.2 106
1
+ 9 + 11
x + (log 300000) 11 + 6
x3/2
q
q
1.14 1010
6.18 1012
+
q
q
1
499076
1.32 106
139
9
9.1 107
+
51.17
+
+
+
+
+
q
qx
x
q x
x
x
1.35 108
err+ , (, x) log
6.18 1012
1.14 1010
+
err+ , (, x)
+
q
q
499100
1
+ 52 .
q
x
4r
Hence
(T H)2
p
0.1065 0 2
()
9.462 T0 He0.1598(T0 H) + 11.287||e
2
101300000 + 6773000e0.1065t .
59
err+ ,T (, x) log
The value t = 20 seems good enough; we choose it because it is not far from
optimal in the range 1027 x 1030 . We get that T0 = 12000000 + 200; since
T0 < 108 , we are within the range of the computations in [Plab] (or for that
matter [Wed03] or [Plaa]). We obtain
err+ ,T (, x) 3.34 1011 +
251100
.
x
Lastly, let us look at the sum estimated in (4.67). Here it will be enough to go
up to just T0 = 2H + max(50, H/4) = 450, where, as before, H = 200. Of course,
the verification of the zeros of the Riemann zeta function does go that far; as
we already said, it goes until 108 (or rather more: see [Wed03] and [Plaa]). We
make, again, the assumption x 1012 . We look at (4.68) and obtain
err2 ,+
(4.87)
(log 450)2
450e 4 50
+
1.21
log
450
0.607
12
log 10
1
+ 2.06 450 log 450 + 43.87 x 2
1.536 1015 +
310.84
.
x
where the
R integrals were computed rigorously using VNODE-LP [Ned06]. (The
integral 0 2 (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 )
274.86
(0.80013 log x + 0.214)
H 7/2
1.944 106 log x + 5.2 107 ,
60
H. A. HELFGOTT
where we are using (B.23) and evaluate | log |2 rigorously as above. By (B.23)
and (B.24),
Z
27428
274.86 2
2
log x +
(+ (t) (t)) log xt dt
7/2
H7
H
0
5.903 1012 log x + 2.143 1012 .
We conclude that
Z
2
+
(t) log xt dt
(4.88)
0
= (0.640206 + O (1.95 106 )) log x 0.021095 + O (5.3 107 )
We add to this the error term 1.536 1015 + 310.84/ x from (4.87), and simplify
using the assumption x 1012 . We obtain:
(4.89)
n=1
2
(n)(log n)+
(n/x) = 0.640206x log x 0.021095x
= ,
3/2
2()(() + j())
2
2
6Implemented by the author from the description in [Tuc11, p. 8788], using D. Platts
61
and so
(A.1)
() 1
1 .
2()(() + j())
2
,
2 2
2
2
2
c0 ()
0.7983 1
+ > 1.06
2
2
(A.4)
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
=
,
,
=
=
2(1 + x)3/2
(1 + x)3/2
(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
(A.5)
1
1
3
1
1 x
1 x + x2 .
2
2
8
1+x
Hence, for 0,
(A.6)
1 + 2 /2 4 /8 j() 1 + 2 /2,
and so
() 1 + 2 /8 54 /128
(A.7)
(A.8)
1
=q
()
1
1+
1
1+
j()1
2
j() 1
2
2
74
3 4
1
+
,
8 16
8
128
2
1 j() 1
1 .
1
2
2
8
4
16
1
1
2
1
=q
()
j()1
2
2
1 j() 1 3
+
1
2
2
8
+
2
62
H. A. HELFGOTT
Hence
sin 0 =
(A.9)
sin 0
1
1
2 2()
2
74
=
16 256
4
7 2
,
16
2
= ,
16
4
while
(A.10)
cos 0 =
1
1
+
2 2()
2
1 ,
16
cos 0
2
74
+
,
16 256
1 2 /8
2
2( + j)
2 2 + 5 /8
2
(A.11)
Assuming 0 1,
1
1+
52
16
94
64
52 94
+
+
1
16
64
1 32
2
32
52 94
16
64
2 !
33
.
4
64
52 614
+
,
16
256
7
1 8 + 128
2( + j)
2 2 + 52 214
8 2 128 4
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
s
(A.12)
2
() = 1 +
2( + j)
2( + j)
2
2
73 355
1 33
1 2
1+
+
2 4
64
8 16
4
64
512
2
3
5
3
7
1
96
35
+
+
+ 13
1 +
4
4 32
64
256 2048
512
2
2
3
4
7
165
+
,
1 +
4 32
64
2048
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
63
Hence
(A.13)
r
2
2
4
74
2
74
cos 0 1
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
73
165 4
4
2
2
+
3/2
1
1 +
4 32
64
2048
32 3 256
7 3
49
+
5
3/2
4 128
3 1024 !
!
!
3
7 3 3
7
167
3
3
1+
+
+
+
+
4
16
2304 2048
2048 192
147456
3
0.09494 ,
16
where we are again using 1. We conclude that, for all (0, 1/2],
1+
c0 () > 1.
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
(A.15),
q
2
1
+
1
1
2 q =
q
q
2
2
1
2
1
1
+
1 + 2
1 + 2
2
2
2
2
(A.16)
q
2
2
1
+
2
8
q + 3/2
=
2
+ 12
1 2 +
1+ 2
64
H. A. HELFGOTT
j
2+
2
(A.17)
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 (A.16) (and hence (A.17)) holds for all 1. By
(A.15),
q
2( + j)
1
1
2 1 + 12
1
+
+
+
2
2
2
(A.18)
1
1
1
1
1
2 1 + 12 + 1
2 2
23/2
for 16. (Again, (A.18) is also true for 1 16 by the bisection method; it
is trivially true for [0, 1], since the last term of (A.18) is then negative.) We
also have the easy upper bound
1
1
1
1
q
q
+
(A.19)
=
2( + j)
2 + 1
2 2 (2)3/2
2 ( + )
2
= 1+
2( + j)
2( + j)
2
1
1
1
1
1
1
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
(A.20)
1
1
1 +
2
1 +
2+
2
2
2
2
1
1
1
1
+
1 +
2
2 + 1
2
!
1
1
2
1
2
5
1
(A.21)
1+
2 +
1 +
2
2
2 4
2 2
1
1
1
1
+ 1 1 + 2 1 +
2
2
2
1
1
1
1
1
9
37
1+
2+
3 1+
2 4 8
2 16
25/2
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
65
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 (A.12), we get the simpler
bound
(A.22)
33
2
+
+
1
4 32
32
4
234
2 113
+
+
2
8
64
1024
1 2
1+
1+
2 2( + j)
2( + j)
2 4
4
64
(A.23)
2
3
2
3
5
1 +
+
1 +
4 32
64
4
64
for 1. (This immediately implies the easy bound 1, which follows
anyhow from (3.25) for all 0.)
By (A.6),
2
1 + 2 /2 4 /8
8
1 + 2 /2 4 /8
j
4
2
2
2
2
2
1 + 8
1 + 8
8 + 64
2 +
1 +
2
2
8
64
128
2
4
64
2 2 2
2
2
3
2
4
11
3
1
4 3 152 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 bisection method (5 iterations).
We conclude that, for all > 0,
4 3
(A.24)
max 1,
.
2
This bound has the right asymptotics for 0+ and +.
Let us now bound c0 from above. By (A.13) and (A.23),
52
74
2
c0 () = () cos 0 + sin 0 1 +
+
1
+
4
64
32 512
4
(A.25)
3
4
5
6
2
2
23
7
35
3
+
+
+ 15 1 +
1+
64
128 2048 2048
2
15
for 1. Since 1 and 0 [0, /4] [0, /2], the bound
(A.26)
c0 () cos 0 + sin 0 2
66
H. A. HELFGOTT
1
1
1 2
9
1 +
+1 2 1 +
.
2
2 2 16
From (A.24) and (A.26), we obtain that
(A.28)
1
1 + 2c20 1 (1 + 2 2) = 5
for all 0. At the same time, (A.24) and (A.25) imply that
4 3 1
42 24
1
2
+ 2
1 + 2c0
3+
2
15
15
1
2
4
3
3
4
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
(1 + 2c20 )/ 0.86 also holds for 0.29 6; for > 6, the same inequality
holds by (A.28).
We have thus shown that
(A.29)
1 + 2c20
min(5, 0.86)
2 1 +
8
128
8
(A.30)
2
52
2 54
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 (A.15) and 2 = (1 + j)/2 (1 + )/2,
v
q
u
r
u +1 1 + 1
r
s
2
1
2 t 2
2
1
2
1+
1+
2
2
2
2
(A.31)
s
r
r
1
1
2
1
1
3/2
1
=
2
2
2
2
2
2
p
for 4. We check by the bisection method (20 iterations) that ( 2 )/2
We conclude that
(A.32)
(
/6
1
23/2
67
if 4,
for all .
We still have a few other inequalities to check. Let us first derive an easy lower
bound on c1 () for large: by (A.1), (A.16) and (A.5),
r
r
s
1
1 + 1/
1
2
8
1
1
+
c1 () =
1
2
2
3/2
2
2
r
r
2
2
1
1
3
1
1+
1
1
=
4
2 4
2
9
1 + 9
2
1 12 + 8
2
16
1
c0 1/ 2
2 2
,
3
2 2 1 3
2c1
2 2 1 4
4
+
=
8
8
32
8
32
for any > 0. Hence, by (A.8) and (A.22),
s
r
4
5 2
1 + 1/
2 2 /8
,
1
c1 () =
4
2
2
4
4
4
+ 9
8
32
whereas, for 1,
by (A.13). 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
(A.33)
1
c0 1/ 2
.
sup
2
1
0
In the same way, we see that
1
c0
1
3 0.171
c1
1 4
for 36 and
1 + 4
c0
0.267
c1
4 1 54 2 1 4
68
H. A. HELFGOTT
for 0.1. The bisection method applied to [0.1, 36] with 24 iterations (including
12 initial iterations) now gives
c0
0.29888.
(A.34)
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 (A.14). By (A.8), (A.23) and (A.30),
s
r
2
74
2 8 + 128
52
1 + 1/
1 +
c1 () =
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.
Now consider the case of large . By and 1,
p
( 2 )/
1 1 1/ 2
c0
1/
p
.
(A.35)
p
q
1+1/
c1
2 1 + 1/
1 + 1/
(This is off from optimal by a factor of about 2.) For 200, (A.35) implies
5 sin3
sin
sin 2
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
(A.37) for [0, 1]. The two sides of (A.37) 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 ).
Appendix B. Norms of smoothing functions
Our aim here is to give bounds on the norms of some smoothing functions
and, in particular, on several norms of a smoothing function + : [0, ) R
2
based on the Gaussian (t) = et /2 .
69
As before, we write
(B.1)
(
t2 (2 t)3 et1/2
h : t 7
0
if t [0, 2],
otherwise
+ (t) = hH (t)tet
(B.3)
where
FH (t) =
(B.4)
sin(H log y)
,
log y
hH (t) = (h M FH )(y) =
(B.6)
1
(B.5)
M FH (i ) = 1/2
0
Thus,
dy
y
By (2.4), M hH = M h M FH . Now
of variables; using this, we get that,
if | | < H,
if | | = H,
if | | > H.
M h(i )
M hH (i ) = 12 M h(i )
if | | < H,
if | | = H,
if | | > H.
s+2
s+1
x
x
(3) (x)
dx = lim
dx
(x)
(B.7) = lim+
+
s(s + 1)
s(s + 1)(s + 2)
t0
t0
t
t
Z
xs+3
dx,
= lim
(4) (x)
s(s + 1)(s + 2)(s + 3)
t0+ t
70
H. A. HELFGOTT
where (4) (x) is understood in the sense of distributions at the finitely many
points where it is not well-defined as a function.
R
Let s = it, = h. Let Ck = limt0+ t |h(k) (x)|xk1 dx for 0 k 4. Then
(B.7) gives us that
(B.8)
C2
C3
C4
C1
,
,
,
.
M h(it) min C0 ,
|t| |t||t + i| |t||t + i||t + 2i| |t||t + i||t + 2i||t + 3i|
h (2) = h (2) = 0, and h(x), h (x) and h (x) are all continuous. The function
h has a discontinuity at t = 2. As we said,
we understand h(4) in the sense of
R 2+
distributions at t = 2; for example, lim0 2 h(4) (t)dt = lim0 (h(3) (2 + )
h(3) (2 )).
Symbolic integration easily gives that
Z 2
t(2 t)3 et1/2 dt = 92e1/2 12e3/2 = 2.02055184 . . .
(B.9)
C0 =
0
The situation with (x2 (2 x)3 ex1/2 ) is similar: it has zeros at the roots of
H2 (x) = 0, where H2 (x) = H1 (x) + H1 (x) (and, in general, Hk+1 (x) = Hk (x) +
Hk (x)). This time, we will prefer to find the roots numerically. It is enough to
find (candidates for) the roots using any available tool7 and then check rigorously
that the sign does change around the purported roots. In this way, we check that
H2 (x) = 0 has two roots 2,1 , 2,2 in the interval (0, 2), another root at 2, and
two more roots outside [0, 2]; moreover,
2,1 = 0.48756597185712 . . . ,
2,2 = 1.48777169309489 . . . ,
(B.11)
where we verify the root using interval arithmetic. The sign of H2 (x) (and hence
of h (x)) is first +, then , then +. Write 2,0 = 0, 2,3 = 2. By integration by
parts,
(B.12)
Z 2,2
Z 2
Z 2,1
Z
h (x)x dx
h (x)x dx +
h (x)x dx
|h (x)|x dx =
C2 =
=
(1)j+1
j=1
=2
2,1
3
X
h (x)x|2,j
2,j1
2,j
2,j1
h (x) dx
2,2
2
X
(1)j+1 h (2,j )2,j h(2,j ) = 10.79195821037 . . . .
j=1
71
2
(1)
h (x)x2 dx
|h (x)|x dx =
C3 =
0
3
X
(1)j
=
j=1
(B.13)
3,j1
j=1
h (x)x2 |3,j
3,j1
3,j
3,j1
h (x) 2x dx
3
X
(1)j h (x)x2 h (x) 2x + 2h(x) |3,j
=
3,j1
j=1
2
X
(1)j (h (3,j )23,j 2h (3,j )3,j + 2h(3,j ))
=2
j=1
C3 = 75.1295251672 . . .
The treatment of the integral in C4 is very similar, at least as first. There are
two roots of H4 (x) = 0 in the interval (0, 2), namely,
4,1 = 0.45839599852663 . . .
4,2 = 1.54626346975533 . . .
The sign of H4 (x) on the interval [0, 2] is first , +, then . Using integration
by parts as before, we obtain
Z 2
(4) 3
h (x) x dx
0+
=2
4,1
(4)
(x)x dx +
0+
4,2
(4)
4,1
(x)x dx
h(4) (x)x3 dx
4,1
2
X
(1)j h(3) (4,j )34,j 3h(2) (4,j )24,j + 6h (4,j )4,j 6h(4,j )
j=1
since limt0+ h(k) (t)tk = 0 for 0 k 3, limt2 h(k) (t) = 0 for 0 k 2 and
limt2 h(3) (t) = 24e3/2 . Now
Z
|h(4) (x)x3 |dx = lim |h(3) (2 + ) h(3) (2 )| 23 = 23 24e3/2 ,
0+
Hence
(B.15)
C4 =
0+
(4) 3
h
(x)
x dx + 24e3/2 23 = 2013.18185012 . . .
72
H. A. HELFGOTT
We will, however, find it easier to deal with M h by means of the bound (B.8), in
part because (B.16) amounts to an invitation to numerical instability.
For instance, it is easy to use (B.8) to give a bound for the 1 -norm of M h(it).
Since C4 /C3 > C3 /C2 > C2 /C1 > C1 /C0 ,
Z
M h(it)dt
|M h(it)|1 = 2
0
!
Z C3 /C2
Z C2 /C1
Z C4 /C3
Z
dt
dt
dt
dt
C1
+ C2
+ C1
+ C3
+ C4
2 C0
2
3
4
C0
C2 /C1 t
C1 /C0 t
C3 /C2 t
C4 /C3 t
2
C32
C4 C33
C1 C2
C2 C0
C3 C2
+
C
+
,
= 2 C1 + C1 log
+
2
C2 C3
2 C32 C42
3 C43
C12
and so
(B.17)
|M h(it)|1 16.1939176.
Z C3
Z C4
Z
C2 dt
C3 dt
dt
i
dt + C2
+ C3
+ C4
|t + i| dt + C1
C0
2
3
C
C1
C
C
4 t
2
3
t
t
t
0
C0
C1
C2
C3
s
!!
!
C2
p
2+11
C0
C14 C12
t
C
1 C1
2
| C11
=
+ C1
t + 1 + log
+ 2 + sinh
4
2
C
t
C0
C0
0
C0
2
C4 C3
C3 C1
C2 C3
+
+ C2 log
+ C3
,
2
C3 C4
2 C42
C2
Z
C1
C0
C2
C1
1 +
and so
(B.18)
73
By (B.8),
(B.21)
Hence
(B.22)
|M h(it)|2 dt
C42
C42
dt
.
t8
7H 7
C42
dt
.
t
7H 7
This is at most
Now
t0
Z
t | log t|
t2 3
max e
t2 3
max e
t0
t | log t| = max
t2 3
max e
t[0,1]
dt
.
t
t2 3
= 0.14882234545 . . .
t[1,5]
t log t
where we find the maximum by the bisection method with 40 iterations.8 Hence,
by (B.22),
Z
C2
(+ (t) (t))2 | log t|dt 0.148822346 4
7
0
(B.24)
27427.502
165.61251 2
.
H7
H 7/2
B.3. Norms involving + . Let us now bound some 1 - and 2 -norms involving
+ . Relatively crude bounds will suffice in most cases.
First, by (B.23),
274.8569
(B.25)
|+ |2 | |2 + |+ |2 0.800129 +
,
H 7/2
where we obtain
(B.26)
| |2 = 0.640205997 . . . = 0.8001287 . . .
by symbolic integration.
8The bisection method (as described in, e.g., [Tuc11, 5.2]) can be used to show that the
minimum (or maximum) of f on a compact interval I lies within an interval (usually a very
small one). Here, the bisection method was carried rigorously, using interval arithmetic. The
method was implemented by the author from the description in [Tuc11, p. 8788], using D.
Platts interval arithmetic package.
74
H. A. HELFGOTT
) (1/2
Now, (M +
+ it) equals 1/2 times the additive convolution of M hH (it)
2
log
|
(B.27)
32
0.3220301,
|+ log |2
we get that
(B.28)
|+ log |2 0.8299818.
+3/2 t2 /2
1+ t2 /2
e
e
|+ (t)t |1 = hH (t)t
t
|hH (t)/ t|2
1
2
s Z
sZ
H
dt
2
1
+3/2 t2 /2
|hH (t)|2 = t+3/2 et /2
|M h(ir)|2 dr
e
= t
2
t
2 H
2
0
s Z
1
2
2
t+3/2 et /2
|M h(ir)|2 dr = t+3/2 et /2 |h(t)/ t|2 .
2
2
2
Since
+3/2 t2 /2
e
t
=
2
|h(t)/ t|2 =
we conclude that
sZ
r
et2 t2+3 dt
for > 2.
( + 2)
,
2
31989 585e3
1.5023459,
8e
8
|+ (t)t |1 1.062319
(B.30)
( + 2)
2
M (+ )(s) ds =
|s M + (s)|2 ds.
|+ |2 =
1
1
2i i
2i i
2
75
Recall that + (t) = hH (t) (t), where (t) = tet /2 . Thus, by (2.5), the
function 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
M h(ir)M (s ir)dr
|s| |M + (s)| =
2 H
Z H
3
(B.31)
|ir 1||M h(ir)| |s ir||M (s ir)|dr
2 H
3
=
(f g)(t),
2
where f (t) = |it 1||M h(it)| and g(t) = | 1/2 + it||M (1/2 + it)|. (Since
|(1/2 + i(t r)) + (1 + ir)| = |1/2 + it| = |s|, 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 .
By (B.18),
|f |1 = |(r + i)M h(ir)|1 27.8622803.
Yet again by Plancherel,
Z
Z 1 +i
2
|s|2 |M (s)|2 ds =
|g|22 =
21 i
1
+i
2
1
i
2
2
|(M (
))(s)|2 ds = 2|
|2 =
Hence
(B.32)
1
3
1 3
|+
|2
|f g|2
27.8622803
2
2
2 2
3 2
.
4
3 2
10.845789.
4
t hH (t)tet /2 + hH (t)(1 t2 )et /2 t
|+ (t)t |1 = hH (t)te
1
1
2
hH (t)t+1 et /2 + |+ (t)t1 |1 + |+ (t)t+1 |1 .
1
We can bound the last two terms by (B.30). Much as in (B.29), we note that
2
2
hH (t)t+1 et /2 t+1/2 et /2 |hH (t) t|2 ,
1
1
2
1
2
Z
1
dt =
|M (hH )(1 + ir)|2 dr
2
0
s Z
Z
H
1
|(ir)M hH (ir)|2 dr =
|(ir)M h(ir)|2 dr
2 H
s Z
Z H
1
|M (h )(1 + ir)|2 dr
|M (h )(1 + ir)|2 dr = |h (t) t|2 ,
2
H
|hH (t)|2 t
( + 1)
103983 1899e3
+1/2 t2 /2
e
, |h (t) t|2 =
= 2.6312226,
t
=
2
16e
16
2
76
H. A. HELFGOTT
we conclude that
(B.33)
|+
(t)t |1
for > 1.
r
p
p
( + 1)
1.062319 ( ( + 1) + ( + 3)) +
2.6312226
2
p
p
2.922875 ( + 1) + 1.062319 ( + 3)
B.5. The -norm of + . Let us now get a bound for |+ | . Recall that
2
+ (t) = hH (t) (t), where (t) = tet /2 . Clearly
(B.34)
| | = (1) = 1,
| (t)t| = 2/e.
sin t
dt
t
0
is defined for all x; it tends to /2 as x + and to /2 as x (see
[AS64, (5.2.25)]). We apply integration by parts to the second integral in (B.35),
and obtain
Z
t
1
d
Si(w)dw h(t)
h w/H
hH (t) h(t) =
H log 2 dw
e
t
Z
1 d
t
=
Si(w)
dw
h w/H
0
dw
2
e
Z
d
1 0
t
Si(w) +
dw.
h w/H
H log 2 dw
2
e
Si(x) =
Now
tew/H
d
t
t
=
t|h | .
h
h
dw
H
ew/H
ew/H Hew/H
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 0 Si(x) x < /2 for
x [0, 1] and /2 < x Si(x) 0 for x [1, 0]. Hence
Z 1
Z w/H !
w/H
2e
2t|h |
e
dw +
dw
|hH (t) h(t)|
H
w
0 2
1
4 E1 (1/H)
1/H
= t|h | (1 e
)+
,
H
where E1 is the exponential integral
E1 (z) =
et
dt.
t
77
By [AS64, (5.1.20)],
log(H + 1)
,
e1/H
and, since log(H +1) = log H +log(1+1/H) < log H +1/H < (log H)(1+1/H) <
(log H)e1/H for H e, we see that this gives us that E1 (1/H) < log H (again
for H e, as is the case). Hence
1 + 4 log H
1
4 log H
|hH (t) h(t)|
H
+
(B.36)
< |h | 1 e
,
< |h |
t
H
H
0 < E1 (1/H) <
1 + 4 log H
2 h(t) hH (t)
2
1+
<
1
+
|h
|
e
t
e
H
|h | = |h (2,2 )| 2.805820379671,
(B.38) |+ | < 1 +
1+
log H
.
H
(B.41)
1+
log H
.
H
Taking derivatives, we see that | (t)t2 | = 33/2 e3/2 . Hence, yet again by
(B.36) and (B.37),
(B.42)
1+
log H
.
H
78
H. A. HELFGOTT
References
[AS64]
[Olv74]
[Plaa]
[Plab]
[Pla11]
[Ros41]
[Tem10]
[Tru]
[Tuc11]
[Tur53]
[TV03]
[Vin37]
[Wed03]
[Whi03]
[Wig20]
[Won01]
79