18.785: Analytic Number Theory, MIT, spring 2007 (K.S.

Kedlaya)
The Bombieri-Vinogradov theorem (proof) (revised 9 May 07)

In this unit, we prove the Bombieri-Vinogradov theorem, in the form stated in the pre­
vious unit.

1 Bounding character sums

For f an arithmetic function, put
⎞ 1 ⎞
Df (x; N, m) = f (n) − f (n);
ψ(N)
n�x,n�m (mod N ) n�x,n≡(Z/N Z)�

that is, Df (x; N, m) measures the deviation between the sum of f on an arithmetic progres­
sion, and the sum on all arithmetic progressions of the same modulus. The following lemma
tells us that bounding this deviation allows us to control the sum of f twisted by a Dirichlet
character.

Lemma 1. Let f be an arithmetic function with support in {1, . . . , x}, and put |f |2 =
( n |f (n)|2 )1/2 . Suppose that for some � � (0, 1], we have
�

|Df (x; N, m)| � x1/2 �9 |f |2 (1)

whenever m � (Z/N Z)� . Then for any nonprincipal character � of modulus r, and any
positive integer s, ⎠ ⎠
⎠ ⎞ ⎠
f (n)�(n)⎠⎠ � x1/2 �3 rχ (s)|f |2 .
⎠ ⎠
⎠
⎠
⎠n≡(Z/sZ)� ⎠

Proof. By Möbius inversion, we can write

⎞ ⎞ ⎞
f (n)�(n) = µ(k) f (n)�(n).
n≡(Z/sZ)� k|s n�0 (k)

We split this sum on k at K = �−6 . We bound the sum for each fixed k > K by Cauchy-
Schwarz; the total is thus dominated by
⎞
|f |2 (x/k)1/2 � |f |2 x1/2 K −1/2 χ (s).
k|s,k>K

For the terms k � K, we write the sum as (using Möbius inversion again)
⎞ ⎞ ⎞
µ(k) µ(δ) f (n)�(n).
k|s,k�K �|k n≡(Z/�Z)�

1
We split the inside sum over classes modulo δr; on each class, we apply (1). Since we are
summing over all residue classes, and � is nonprincipal, the main terms cancel out; the sum
is thus dominated by
⎞ ⎞
|f |x1/2 �9 |µ(δ)|ψ(δr) � |f |2 x1/2 �9 Kψ(r)χ (s).
k|s,k�K �|k

Since K = �−6 , we may add the two bounds to give the desired inequality.
Using the large sieve inequality, we obtain the following.
Theorem 2. There exists an absolute constant c > 0 with the following property. Let f
be an arithmetic function with support in {1, . . . , x} satisfying (1). Let g be an arithmetic
function with support in {1, . . . , y}, and let h = f φ g be the Dirichlet convolution. Then
⎞
max � |Dh (xy; N, m)| � c|f |2 |g|2 (�(xy)1/2 + x1/2 + y 1/2 + Q) log2 Q.
m≡(Z/N Z)
N �Q

Proof. We have
⎛ �⎛ �
1 ⎞ ⎞ ⎞
Dh (xy; N, a) = �(a) f (m)�(m) g(n)�(n) ,
ψ(N ) �∗=� m n
0

with � running over Dirichlet characters of modulus N. Rewriting this as a sum only over
primitive characters (factoring N = rs, where r is the “primitive modulus”), and using the
fact that ψ(rs) � ψ(r)ψ(s) for all r, s, we can bound the left side of the desired inequality by
⎠ ⎠⎠ ⎠
⎞ 1 ⎞ 1 ⎞ ⎠⎠ ⎞ ⎠⎠ ⎞
⎠⎠
⎠
⎠
⎠ f (m)�(m) ⎠ ⎠ g(n)�(n) ⎠,
⎠ (2)
s�Q
ψ(s) 1<r�Q
ψ(r) ⎠
� ⎠
⎠⎠
(m,s)=1 ⎠⎠ (n,s)=1
⎠

with � now running over primitive characters of level r.

We now split the sum over r at R = �−1 . For r � R, we apply Lemma 1; those terms
are dominated by
⎞ χ (s) ⎞
|f ||g|y 1/2�3 r � c|f ||g|y 1/2�3 R2 log2 Q.
s�Q
ψ(s) r�R

(Note: we are not doing anything to the g terms other than bounding the whole sum by
|g| and pulling it out. We apply the lemma to the f terms.) For r > R, we split the sum
further into ranges like P < r � 2P and apply the multiplicative large sieve inequality in
each range. Rather, we apply it twice: once with the f sum to obtain
⎠ ⎠2
⎠ ⎠
⎞ 1 ⎞ ⎞ 1
f (m)�(m)⎠⎠ � (4P 2 + x − 1)|f |22 ,
⎠ ⎠
⎠
P <r�2P
ψ(r) � ⎠
⎠ P
(m,s)=1 ⎠

2
and again with the g sum. Putting together with Cauchy-Schwarz, we get a bound
⎠ ⎠⎠ ⎠
⎠ ⎠⎠ ⎠
⎞ 1 ⎞ ⎞ ⎞ 1
g(n)�(n)⎠ � (4P 2 +x)1/2 (4P 2 +y)1/2 |f |2 |g|2 .
⎠ ⎠ ⎠ ⎠
⎠ f (m)�(m)⎠ ⎠
⎠ ⎠ ⎠
P <r�2P
ψ(r) � ⎠
⎠
⎠ P
m≡(Z/sZ)
� ⎠⎠ n≡(Z/sZ)
�

Now summing, over P = R, 2R, . . . until P > Q, we get a bound on the sum over r in (2) of

c|f |2 |g|2 (Q + x1/2 + y 1/2 + x1/2 y 1/2 R−1 ).

(That R−1 is the reason we had to limit this argument to r large.) The sum over s throws
on another two factors of log Q, yielding the claim.

2 Proof of the theorem

We now proceed to the proof of the Bombieri-Vinogradov theorem. First, we mention an
identity of Vaughan that will be useful: for any y, z � 1 and n > z,
⎞ n ⎞ ⎞
�(n) = µ(b) log − µ(b)�(c) + µ(b)�(c). (3)
b
b�y,b|n b�y,c�z,bc|n b>y,c>z,bc|n

Given x, define the incomplete logarithm

⎞
�(δ) = log δ − �(k);
k�x1/5 ,k|�

then (3) with y = z = x1/5 implies that for x1/5 < n � x,

⎞ ⎞
�(n) = �(δ)µ(m) + �(δ)µ(m). (4)
�m=n,m�x1/5 �m=n,x1/5 <m�x4/5

Let �0 (n) and �1 (n) denote the two sums on the right side of (4). Then

D� (x; N, m) = D�0 (x; N, m) + D�1 (x; N, m) + O(x1/5 log x),

with the error term coming from terms with n < x1/5 .
It is straightforward to prove that
⎞
max � |D�0 (x; N, m)| = O(Qx2/5 log x), (5)
m≡(Z/N Z)
N �Q

so we concentrate on the contribution from �1 . We want to apply Theorem 2, but we cannot

write the sum �1 (n) as a convolution because of the restriction n � x.
To get around this, we cut the interval 1 � n � x into O(λ −1 ) subintervals of the form
y < n � (1 + λ)y, where x1/5 < λ � 1 is a parameter we will set later. We cover the
summation range
δm = n, x1/5 < m � x

3
by ranges
δm = n, L < δ � (1 + λ)L, M < m � (1 + λ)M
with L, M taking values (1 + λ)j . We run L, M over the ranges x1/5 < L, M < x4/5 with
LM = x; the only trouble is that we do not properly cover the areas n < x1/5 and (1+λ)−1 x <
n < (1 + λ)x. The contribution from the error regions is O(λN −1 x log x).
What remains is the sum over L, M of
⎞ 1 ⎞
D(L, M ; N, m) = �(δ)µ(m) − ,
ψ(N)
l,m�m (mod N ) lm≡(Z/N Z)�

where l, m run over L < δ � (1 + λ)L, M < m � (1 + λ)M. For each L, M, we may
apply Theorem 2 with � = (log x)−A ; the hypothesis (1) is satisfied by the Siegel-Walfisz
theorem (the error bound on the prime number theorem in arithmetic progressions). If we
take Q = �x1/2 , we get
⎞
max � |D(L, M; N, m)| = O(λ�x(log x)3 ).
m≡(Z/N Z)
N �Q

Summing over L, M , we obtain

⎞
max � |D�1 (x; N, m)| = O((λ −1 x + �)x(log x)3 .
m≡(Z/N Z)
N �Q

We now choose λ = �1/2 , so this bound becomes �1/2 x(log x)3 . Adding back in (5) gives
⎠ ⎠
⎞ ⎠ �(x) ⎠ = O(�1/2 x(log x)3 ).
⎠
max � ⎠⎠�(x; N, m) −
1/2
m≡(Z/N Z) ψ(N) ⎠
N ��x

Using the prime number theorem with error term, we can take �(x) = x + O(λx). This gives
the Bombieri-Vinogradov theorem with B(A) = 2A + 6.

3 The Barban-Davenport-Halberstam theorem

We leave the proof of the Barban-Davenport-Halberstam theorem to the reader; it is actually
somewhat simpler than Bombieri-Vinogradov. Here is the key step.

Theorem 3. There exists an absolute constant c > 0 with the following property. Let f be
an arithmetic function with support in {1, . . . , x} satisfying (1). Then
⎞ ⎞
|Df (x; N, m)|2 � c|f |2 (�x + Q)(log Q)2 .
N �Q a≡(Z/N Z)�

We note in passing the following corollary.

4
Corollary 4. With conditions as in Theorem 2, for ab ≤= 0, we have
⎠ ⎝ �⎝ �⎠
⎠ ⎠
⎞ ⎠ ⎞ 1 ⎞ ⎞ ⎠
⎠ f (m)g(n) − � f (m)� � g(n)�⎠⎠
⎠ ψ(N)
N �Q,(ab,N )=1 ⎠m,n:am�bn (N ),(mn,N )=1 (m,N )=1 (n,N )=1 ⎠
� c|f ||g|(x + Q)1/2 (�y + Q)1/2 log2 Q.

Exercises
1. Prove (3).

2. Use (3) to deduce (4).

3. Prove (5).

4. Prove Theorem 3, by imitating the proof of Theorem 2.

5. Deduce Corollary 4 from Theorem 3. (Hint: rewrite the difference in terms of Df and
Dg .)