Topics in analytic number theory, Lent 2013.
Lecture 18: Large sieve II
Bob Hough
February 28, 2013
Reference for this lecture: Iwaniec and Kowalski, Sections 7.3 7.4, also
Bombieri, pp. 2426.
Recall from last lecture the analytic form of the large sieve.
Theorem (Analytic large sieve). Let N 1 and > 0. Let x1 , ..., xR R/Z be
+N
-well-spaced points. For any M Z and for any complex numbers (ak )M
k=M +1
we have
M
+N
R
X
X
2
2
(N + 2/)
|ak |2 .
|ak e(kxj )|
4
j=1
k=M +1
We introduce arithmetic into this analytic statement by choosing special
points xj .
Definition 18.1. Let Q 1. The Farey fractions of level Q are the set
FQ = {
a
: (a, q) = 1, 0 a < q, q Q}.
q
The following lemma is easy, but important.
Lemma 18.1. The set FQ is at least
Proof. Given two Farey fractions
a
q
6=
1
Q2 -well-spaced.
a0
q0
we have
0
a a0
aq a0 q
=
.
q
q 0
R/Z
qq 0
R/Z
Now aq 0 a0 q 6 0 mod qq 0 so this distance is at least
1
qq 0
1
Q2 .
As a corollary, we have the following additive form of the large sieve.
Theorem 18.2 (Large sieve, additive form). Let N, Q 1 and M an arbitrary
+N
integer. For any complex numbers (bk )M
k=M +1 we have
2
+N
M
+N
X X MX
X
2
ka
bk e( )
(N + 2Q2 )
|bk |2 .
q
4
qQ a mod q k=M +1
(a,q)=1
k=M +1
Our goal in this lecture is to pass from the additive characters to multiplicative ones. The classical statement of the multiplicative large sieve is as
follows.
Theorem 18.3 (Large sieve, multiplicative form). Let N, Q 1 and M an
+N
arbitrary integer. For any complex numbers (bk )M
k=M +1 we have
X X
qQ mod q
2
+N
M
+N
MX
X
2
bk (k)
(N + 2Q2 )
|bk |2 .
4
k=M +1
k=M +1
The indicates that only primitive characters are to be included.
Remark. This is a significant result, in that it asserts an approximate orthogonality between the primitive characters of differing moduli.
It is necessary to exclude the imprimitive characters as they contain redundant information, and indeed the bound would be false with imprimitive
characters included. To see this, note that the real character 4 will induce
Q/4 imprimitive characters. If we set bk = 4 (k) then the left hand side will
have size at least QN 2 /4 QN kbk2 which is substantially larger than the
right hand side if Q is a bit smaller than N .
For Linniks Theorem we will need the following slight variant of the multiplicative large sieve.
Theorem 18.4 (Multiplicative large sieve, variant). Let N, Q 1 and M
M +N
an arbitrary integer. Let (bk )k=M
+1 be complex numbers which satisfy bk = 0
whenever k has a prime factor less than or equal to Q. We have
X
qQ
Q X
log
q
mod q
2
+N
M
+N
MX
X
2
(N + 2Q2 )
|bk |2 .
bk (k)
4
k=M +1
k=M +1
Remark. At the expense of constraining the coefficients bk to vanish when k
has a small factor, we gain a logarithmic weight on the characters of small
conductor. This seemingly minor gain is sometimes decisive, as it will be in
Bombieris proof of Linniks theorem. Well see another such application at the
end of the lecture.
Before giving the proof of Theorem 18.4 we recall two properties of Gauss
sums from Lecture 4. Let mod q be a character, with Gauss sum
X
a
() =
(a)e( ).
q
a mod q
First, if (b, q) = 1 then for all characters ,
X
(a)e(
a mod q
Also, if is primitive then | ()| =
mod q then either
ab
) = (b) ().
q
q, while if is imprimitive, induced by
1. (q , q/q ) = 1 and q/q is squarefree, in which case | ()| = | ( )| =
2. Otherwise, () = 0.
Proof of variant of multiplicative large sieve. From the additive version of the
large sieve, we already know that the RHS is a bound for
2
+N
X X MX
ka
bk e( )
(1)
q
qQ a mod q k=M +1
(a,q)=1
so we now work to express this sum in terms of characters. Opening the square
and moving the sum over a inside, we write (1) as
M
+N
X
bk 1 bk 2
qQ k1 ,k2 =M +1
e(
a mod q
(a,q)=1
k1 a k2 a
).
q
In order to separate the variables, we introduce a second variable a2 and employ
Dirichlet characters:
X
M
+N
X
bk 1 bk 2
qQ k1 ,k2 =M +1
e(
X
a1 k1 a2 k2 1
)
(a1 )(a2 ),
q
(q)
mod q
a1 ,a2 mod q
(a1 a2 ,q)=1
the inner sum selecting for a1 = a2 . Then rearranging the sum, we arrive at
X
qQ
X
qQ
X
1
| ()|2
(q)
mod q
bk 1 bk 2
(a1 )e(
a1 mod q
(a1 ,q)=1
mod q k1 ,k2 =M +1
qQ
M
+N
X
X
1
(q)
M
+N
X
a1 k1
)
q
X
a2 mod q
(a2 ,q)=1
(a2 )e(
a2 k2
)
q
bk1 bk2 (k1 )(k2 )
k1 ,k2 =M +1
2
+N
MX
X
1
| ()|2
bk (k) .
(q)
mod q
k=M +1
Let mod q be the primitive character inducing mod q, and note that for
all k, bk (k) = bk (k), since the characters may differ only when (k, q) > 1,
but then bk = 0. Thus we obtain
2
+N
MX
X 1 X X
2
| ()|
bk (k) ,
(q)
qQ
q |q mod q
k=M +1
the on the sum over characters indicating that only primitive characters are
to be taken.
Set d = q/q . Now | ()|2 = q if both (d, q ) = 1 and d squarefree, and
vanishes otherwise. Arranging the sum over q as a sum over q and d, we arrive
at
2
+N
X MX
X
X
1
(k) .
b
q
k
(q d)
Q
q Q
d q
(d,q )=1
squarefree
mod q
k=M +1
Thus each primitive character mod q is counted with weight
q
(q )
X
dQ/q
(d,q )=1
squarefree
Y
1
1
=
(d)
1 1/p
p|q
X
dQ/q
(d,q )=1
squarefree
1Y
1
.
d
1 1/p
p|d
Expanding out the products, we may write the last expression as
X
nA(q ,Q)
1
,
n
and we note that A(q , Q) contains all n Q/q since the factor
accounts for all products of powers of primes dividing q , while
X
dQ/q
(d,q )=1
squarefree
1
p|q 11/p
1
1Y
d
1 1/p
p|d
accounts for all n < Q/q that are coprime to q . Thus mod q is counted
R Q/q dx
P
Q
with weight at least nQ/q n1 1
x = log q .
As an application of this variant of the multiplicative large sieve, we prove
the following result of Brun and Titchmarsh, which gives an upper bound for
the number of primes in a short interval.
Theorem 18.5 (Brun-Titchmarsh theorem). Let 1 < N < M and let (M, N )
denote the number of primes in the interval [M +1, M +N ]. We have (M, N )
N
log N .
Remark. The Brun-Titchmarsh theorem gives an upper bound of the expected
order of magnitude whenever N is at least a power of M .
Montgomery and Vaughn have shown the inequality
(M, N ) 2
N
.
log N
valid for all M 1, N 3.
+N
Proof. (See Bombieri, pp. 26-27) Define (ak )M
k=M +1 by ak = 1 if k is prime and
ak = 0 otherwise, and let Q = logNN . Thus the conditions of Theorem 18.4 are
satisfied. Taking only the term q = 1 on the LHS and discarding the rest, we
find
2
log Q (M, N )2
(N + 2Q2 )(M, N ),
4
so
N
2 N
+ o(
).
(M, N )
2 log N
log N
Remark. Had we used the optimal constant (N + 1 1) in the analytic large
2N
sieve, this would yield (M, N ) log
N + o(N/ log N ).
