Topics in analytic number theory, Lent 2013.
Lecture 20: Conclusion of Bombieris theorem
Bob Hough
March 6, 2013
Reference for this lecture: Bombieri Section 6.
Recall the hybrid version of the large sieve.
TheoremP
(Hybrid large sieve). Let T 2. Let {an }
n=1 be complex numbers
satisfying
|an | < and the additional condition that an = 0 if n has a factor
T . Then
2
Z
X
X
T X T X
it
(|an |2 (n + T 3 )).
an (n)n dt
log
q
T
n
n
qT
mod q
[Note: it may be that the RHS = .]
Also, the statement of Bombieris Theorem.
Theorem (Bombieri). There exists c > 0 such that, for all T > 2, and all
1/2 < 1,
X X
N (, T ; ) = O(T c(1) ).
qT mod q
If there exists a real primitive character 1 mod q1 with q1 < T , such that
(s, 0 ) has zero 1 > 1 log T , then we have the improved estimate
X X
N 0 (, T ; ) = O(T c(1) (1 1 ) log T ).
qT mod q
Here the
cluded.
on N 0 (, T ; ) indicates that the single exceptional zero is to be ex-
Completion of proof of Bombieris theorem. Recall that we have set r = C(1
) with C 2 and 1/ log T < r < 1012 .
At the end of last lecture, we had reached the bound
2
Z T +r/2 Z xA X
dy
(log x)3 Cr 0
a
(p)
p
dw
x
N (, T ; ).
1+iw y
log T
T r/2 x
xp<y p
We rearrange and sum over to obtain
X X
qT mod q
cr
x log T X X
N 0 (, T ; )
(log x)3
qT mod q
T +r/2
T r/2
xA
2
X
ap (p) dy
dw
1+iw y
xp<y p
At this point, Bombieri makes a move which would be very wasteful, except that
we are working quite close to the line <(s) = 1; he includes the set of characters
mod q with q T into the set with q T 2 . Thus the above is bounded by
2
2
Z xA X
2 X Z T X
cr
x
T
dy
ap (p)
log
dw
1+iw
(log x)3 x
q
p
y
2
T
2
qT
xp<y
mod q
Assume that x > T 2 . Then by the hybrid large sieve we obtain the bound
xcr
(log x)3
xA
Z xA X
X |ap |2
|ap |2 dy
xcr
6 dy
(p
+
T
)
,
2
3
p
y
(log x) x
p y
xpy
xpy
the last estimate being satisfied if x > T 6 , which we now assume. Finally, we
exchange sum and integral to obtain the bound
xcr
(log x)3
X
xpxA
|ap |2
p
xA
dy
xcr
y
(log x)2
X
xpxA
|ap |2
.
p
Since r = C(1 ) we complete the proof of Bombieris theorem by proving the
following Lemma.
Lemma 20.1 (Bombieri Lemma C). We have the bound
X |ap |2
(log x)2
non-exceptional case
3
(1
)(log
x)
When 1 is exceptional for L(s, 1 )
p
1
A
xpx
Proof. Recall that ap = log p in the non-exceptional case, and ap = (1 +
1 (p)
) log p in the exceptional case.
p11
In the non-exceptional case, we may bound
X
xpxA
(log p)2
A log x
p
X
xpxA
log p
A2 (log x)2 ,
p
e.g. by partial summation against the prime number theorem. Thus in the
exceptional case we may assume that (1 1 ) 100A1log x , or else the bound is
already covered by the non-exceptional case. Assuming this, then
ap = (1 + 1 (p)) log p + 1 (p)(p1 1 1) log p
= (1 + 1 (p)) log p + O((1 1 )(log p)2 ).
Thus for x p xA , using (1 1 ) log p < 1,
a2p (1 + 1 (p))(log p)2 + O((1 1 )(log p)3 ).
Now
X
xpxA
(1 1 )(log p)3
A2 (11 )(log x)2
p
X
xpxA
log p
A3 (11 )(log x)3 ,
p
while
X
xpxA
(1 + 1 (p))(log p)2 )
A2 (log x)2
p
X
xpxA
1 + 1 (p)
p
so for the lemma it will suffice to prove that
X
xpxA
1 + 1 (p)
(1 1 ) log x.
p
(1)
Before proving this we introduce a bit of notation. Let
X
1 (n) =
1 (d).
d|n
Thus 1 (p) = 1 + 1 (p). The coefficients 1 are multiplicative, if (m, n) = 1
then
1 (mn) = 1 (m)1 (n),
and positive. The positivity may be checked at prime powers. Alternatively it
follows from the fact that
F (s) = (s)L(s, 1 )
is the Dedekind zeta function of the quadratic field with character 1 .
Thus it follows that
X
X
X (n) X (p)
1 (np)
1 (n)
1
1
=
.
n
p
np
n
A
A+1
A
n<x
n<x,xpx
xpx
xnx
The bound (1) now follows from the following sub-lemma.
Lemma 20.2. Let x > T 5 and A 1. We have the bounds
X
xnxA+1
and
1 (n)
L(1, 1 ) log x
n
X (n)
L(1, 1 )
1
.
n
(1 1 )
1nx
Proof. For the first statement, note that
X
xnxA+1
X (n)
A+1
1 (n)
1
(en/(x ) en/(x/2) )
n
n
n
since the introduced quantity is 1 throughout the length of the original sum.
Recall the identity
Z 2+i
1
(s)xs ds = e1/x .
2i 2i
Thus the latter sum may be expressed as
Z 2+i
1
(s + 1)L(s + 1, 1 )[x(A+1)s (x/2)s ](s)ds.
2i 2i
3
Shifting the contour to the line <(s) = 1/2 we pass a pole at 0. Note that this
would be a double pole, coming from and , but the factor [x(A+1)s (x/2)s ]
vanishes at 0. The residue at 0 is thus
L(1, 1 )(A log x + log 2).
Meanwhile, the integral is bounded by
Z 1/2+i
(s + 1)L(s + 1, 1 )[x(A+1)s (x/2)s ](s)ds x1/2 T,
1/2i
by putting in the trivial bounds (1/2+it) (1+|t|), L(1/2+it, ) T (1+|t|)
and using exponential decay of in vertical strips. Since L(1, 1 ) 1T by the
class number formula, we have an asymptotic formula if x T 3 , which proves
the first claim,
X
1 (n)
L(1, 1 ) log x.
n
A+1
xnx
If we tried to evaluate the sum in the second claim by standard Perron
summation we would arrive at
Z 1+iT
X (n)
1
ds
1
(s)L(s, 1 )xs .
n
2i 1iT
s
nx
The right hand side has a double pole at 0, expected to give the main contribution. This residue is
L(1, 1 ) log x + L0 (1, 1 ).
Since L0 (1, 1 ) is expected to be negative (there is a nearby 0), this main term is
difficult to estimate. This issue may be avoided by a trick similar to the one in
the proof of Siegels Theorem (we got this from Friedlander and Iwaniec, Opera
de Cribro).
We have
X (n)
X (n)
X (n)
n
1
1
1
x1 1
x1 1 (1 )
1
1
n
n
n
x
nx
nx
nx
The last factor is introduced for convergence in the resulting Perron integral.
Indeed, we observe the integral identity ( > 0)
Z +i
1
ds
1 1/x
x>1
s
x
=
;
0
otherwise
2i i
s(s + 1)
the proof is familiar shift the contour left if x > 1 and right otherwise.
Thus the last sum may be written
Z 11 +1/ log x+i
1
ds
(s + 1 )L(s + 1 , 1 )xs+1 1
.
2i 11 +1/ log xi
s(s + 1)
To conclude we shift the contour to the line <(s) = 1/2 1 passing a pole at
s = 1 1 with residue
L(1, 1 )
.
(1 1 )(2 1 )
4
Shifting the integral is justified by the convexity bound which we will prove
next class: for 0 < 1 (and |t| 2 if = 1)
|( + it)| (1 + |t|)(1)/2+ ,
|L( + it, 1 )| (q1 + |t|)(1)/2+
Thus the integral on the line <(s) = 1/2 1 is bounded by
1
2i
1/21 +i
(s + 1 )L(s + 1 , 1 )xs1+1
1/21 i
1/4+ 1/2
q1
ds
s(s + 1)
T 1/4+ x1/2 .
Again using L(1, 1 ) 1T , we deduce that for x T 3/2+3 the residue term
dominates the error integral. Thus
X 1 (n)
L(1, 1 )
n
(1 1 )
nx
as desired.
