An introduction to analytic number theory

This expository article aims to convey a taste of analytic number theory via the study
of two problems centering around the Riemann zeta function, namely the prime number
theorem and the Dirichlet theorem on primes in arithmetic progressions. We shall assume
some knowledge on the reader of complex analysis (for instance, infinite products, contour
integrals and analytic continuation) and algebra (limited to the language of group theory).

1 Riemann zeta function

The Riemann zeta function is defined by

∞
X 1
ζ(s) =
n=1
ns

for all complex numbers s with Re s > 1. Its relation to number theory is already seen
in the following Euler’s product formula, which relates it to one of the most primitive
notions in number theory, namely prime numbers:

Lemma 1. For all real numbers s > 1,

Y 1
ζ(s) =
p
1 − p−s

where in the product p runs through all prime numbers.

Proof. Every positive integer n can be written uniquely as the product of prime numbers.
Hence if p1 , p2 , . . . is a listing of all prime numbers, then pe11 pe22 . . . runs through all positive
integers when the exponents e1 , e2 , . . . independently run through all non-negative integers

1
with all but finitely many being non-zero. As a result,
∞
X 1 XX X 1
= lim . . .
n=1
ns k→∞ e ≥0 e ≥0 e ≥0 (p1 p2 . . . pekk )s
e1 e2
1 2
!k ! !
X 1 X 1 X 1
= lim ...
k→∞ pse1
e1 ≥0 1
pse2
e2 ≥0 2
psek
ek ≥0 k
    
1 1 1
= lim ...
k→∞ 1 − p−s1 1 − p−s2 1 − p−s
k
Y 1
=
p
1 − p−s

as desired, at least formally. It is easy to justify the manipulations of the infinite series
and products here, which we leave to the reader.

Such Euler product formula are very useful in general. There are various general-
izations of the above product formula to other zeta functions and their generalizations,
called L-functions (see also Section 2.3 below). They relate the analytic object, namely
the zeta functions, to arithmetic information by means of a product involving one term
at each prime. By a similar token, they are very useful in the study of geometric objects
(like elliptic curves): this belongs to another huge area of mathematics that we shall not
take up here.

As a warm-up we shall use the Euler product formula to prove the well-known fact that
there are infinitely many primes. Indeed we shall prove a stronger quantitative assertion,
namely

Theorem 2. X1
=∞
p
p

where the sum runs over all prime numbers p.

Proof. By the Euler product formula, at least for real values of s > 1,
X
log ζ(s) = − log(1 − p−s )
p

where again the sum runs through all prime numbers p. The Taylor series expansion of
the logarithm says − log(1 − x) = x + O(x2 ) for x close to zero, so for s > 1 but close to
1,
X X 1
log ζ(s) = (p−s + O(p−2s )) = + O(1).
p p
ps

2
Here O(x2 ) is an error term that upon division by x2 remains bounded as x → 0, and
O(1) is a term that remains bounded as s → 1+ . The last equality in the above equation
holds because for every prime p, we can choose the same P constant P
C > 0 so that each of
−2s −2 −2
term O(p ) is bounded by Cp for s close to 1 , and +
p ≤ n−2 converges. We
will use this fact again implicitly later. Letting s → 1+ , we get
X1
=∞
p
p

if we know lims→1+ ζ(s) = ∞. This is indeed the case, since the series ∞ 1
P
n=1 n diverges.

In fact it is readily seen from the definition that ζ is an analytic function of s for
Re s > 1, and that the Euler product formula continues to hold in that range of s. We
shall see later that ζ can be analytically continued so that it has a simple pole at s = 1.
In other words, we shall see that
lim ζ(s) = ∞,
s→1
which is stronger than the assertion lims→1+ ζ(s) = ∞ that we have just used above.

The above proof of the infinitude of primes can be modified to prove the following
Dirichlet’s theorem on the infinitude of primes in any (non-trivial) arithmetic progres-
sions:
Theorem 3. In any arithmetic progression {qk + l : k ∈ N}, where q, l ∈ N are relatively
prime, there exists infinitely many primes.

Clearly for the arithmetic progression {qk + l} to contain infinitely many primes, q
and l have to be relatively prime. The remarkable discovery of Dirichlet is that this is
also sufficient. The goal of section 2 is to prove this theorem; note that the case q = 4,
l = 3 says that there are infinitely many primes of the form 4k + 3, which is well-known
and elementary.

The study of the zeta function shall also lead to the famous prime number theorem:
Theorem 4. Let π(x) be the number of prime numbers not exceeding x. Then
x
π(x) ∼
log x
!
π(x)
as x → ∞, i.e lim x = 1.
x→∞
log x

We shall prove this in section 3 using complex analysis.

The two sections are basically logically independent and can be read in any order.
We hope they can give the reader some idea how analysis helps one solve outstanding
questions in number theory.

3
2 Dirichlet’s theorem on primes in arithmetic pro-
gressions

In this section we fix two positive integers q and l that are relatively prime. To prove
Dirichlet’s theorem, following the proof of the infinitude of primes given above, we shall
prove
X 1
(1) = ∞,
p
p≡l(mod q)

where the sum is only over those primes p that are congruent to l (mod q). (Hereafter the
symbol p is reserved for prime numbers.) To do so we shall need some Fourier analysis
on finite abelian groups, to which we now turn.

2.1 Fourier analysis on finite abelian groups

Let G be a finite abelian group. In application we shall take G = (Z/qZ)× , the multi-
plicative group of units in the ring Z/qZ. In other words, G will be the multiplicative
group of integers modulo q that are relatively prime to q. Since the results below are
fairly general, we shall state it for an arbitrary finite abelian group G.

Definition. A character χ of G is a group homomorphism χ : G → U (1), where U (1)

is the multiplicative group of complex numbers whose moduli are 1. The group of all
characters of G (under pointwise multiplication) will be denoted by Ĝ.

For example, there is the trivial character χ0 , defined by χ0 (a) = 1 for all a ∈ G. It is
often useful to consider this trivial character separately, as we will do in the proof of the
Dirichlet’s theorem.

For readers who know some representation theory, the χ’s in the above definition are
precisely the characters of the complex irreducible representations of G. The key fact
that we shall need is the following version of Peter-Weyl theorem, which says that the
characters of a finite abelian group G can be used to obtain a series expansion of any
complex-valued functions on G that resembles the ordinary Fourier series of L2 functions
on the unit circle:

Theorem 5. Let H be the space of complex-valued functions on G, equipped with the

inner product
1 X
< f, g >= f (a)g(a)
|G| a∈G

4
for any f , g ∈ H. Then Ĝ is an orthonormal basis of H. In particular, for f ∈ H we
define for all χ ∈ Ĝ
1 X
fˆ(χ) =< f, χ >= f (a)χ(a);
|G| a∈G
then X
f (a) = fˆ(χ)χ(a)
χ∈Ĝ

for all a ∈ G.

1
P
Again, the readers familiar with group representations will recognize |G| a∈G δa as
2
the Haar measure on G, and H as the Hilbert space L (G) with respect to this measure.
Hence the above theorem is a special case of the Peter-Weyl theorem in the setting of
finite abelian groups. Since for finite abelian groups G the above theorem admits a much
simpler proof, for the convenience of readers who do not know representation theory, we
shall give the simple proof in this case below (see section 2.5).

Now we specialize to the case where G = (Z/qZ)× : we shall take f to be the function
on (Z/qZ)× defined by
(
1 if a = l (mod q)
f (a) = δl (a) = ,
0 for all other a ∈ (Z/qZ)×

where l is the integer that is fixed at the beginning of this section that is relatively prime
to q. Then since |G| = φ(q), where φ is the Euler phi function,
1
fˆ(χ) = χ(l)
φ(q)

for all characters χ of (Z/qZ)× and thus

1 X
(2) δl (a) = χ(l)χ(a)
φ(q) χ

for all a ∈ (Z/qZ)× . Note that both δl (a) and χ(a) above were only defined for a ∈
(Z/qZ)× , and the equality was asserted only for such a’s. Below we shall extend the
domain of definitions of the δl ’s and the χ’s to the whole Z/qZ and see that (2) continues
to hold. This will lead us to the concept of Dirichlet characters.

2.2 Dirichlet characters

Our goal in this section is to extend (2) so that it holds for all a ∈ Z/qZ, and indicate
how this is related to the proof of (1).

5
 The extension of (2) is trivial; one simply extend the definitions of δl and the χ’s by
setting them to vanish for the non-units in Z/qZ, i.e. we set

/ (Z/qZ)× ,
δl (a) = 0 for a ∈

and for all characters χ of (Z/qZ)× , we set

/ (Z/qZ)× .
χ(a) = 0 for a ∈

The extended χ’s are called the Dirichlet characters modulo q; there are φ(q) of them.
The extension of the trivial character is still denoted by χ0 ; so
(
1 if a ∈ (Z/qZ)×
χ0 (a) =
0 if a ∈ / (Z/qZ)× .

Now (2) obviously continues to hold on all of Z/qZ, i.e.

1 X
(3) δl (a) = χ(l)χ(a)
φ(q) χ

for all a ∈ Z/qZ, where the sum is still over all characters χ of (Z/qZ)× . By abuse of
notation, for a ∈ Z we shall also denote by δl (a) and χ(a) the values of δl and χ evaluated
at the equivalent class a modulo q, and clearly (3) continues to hold for all a ∈ Z. We
shall also call such χ Dirichlet characters modulo q.

With this notation in mind, by (3) we can write the sum in (1) as
X 1 X δl (p)
lim+ = lim
s→1 ps s→1+ p ps
p≡l(mod q)

1 X X χ(p)
= lim+ χ(l) s
;
s→1 φ(q) χ p
p
X
here once again denotes sum over all primes p. Note that we have put in the power
p
s > 1 to make sure that the sums converge absolutely so that it can be rearranged.
Separating the term involving the trivial character χ0 , we get
X 1 1 X 1 1 X X χ(p)
= + χ(l) .
ps φ(q) ps φ(q) χ6=χ p
ps
p≡l(mod q) p6 |q 0

As s → 1+ , the first term diverges to infinity by Theorem 2. If we can show that the
second term remains bounded as s → 1+ , then it follows that (1) holds, and Dirichlet’s
theorem is proved. In fact we shall prove the following:

6
Theorem 6. If χ is a non-trivial character of (Z/qZ)× , then
X χ(p)

p
ps

remains bounded as s → 1+ , where the sum is over all primes p.

This is plausible because if χ is a non-trivial Dirichlet character modulo Pq, then χ

takes on more than one value on the unit circle U (1), and thus in the sum p χ(p) ps
, it
P 1
is likely that there are cancellations that were not present in the sum p ps . This extra
cancellation will lead to the desired boundedness of the sum p χ(p) as s → 1+ .
P
ps

In the next section we prove this theorem, using the Dirichlet L-functions.

2.3 Dirichlet L-functions

To prove Theorem 6, we imitate the proof of Theorem 2. There we made use of the
Euler product formula. It turns out that there is a version of the Euler product formula
involving the Dirichlet characters, namely
Lemma 7. For any real numbers s > 1 and any Dirichlet characters χ modulo q, we have
∞
X χ(n) Y 1
(4) =
n=1
ns p
1 − χ(p)p−s

where in the product p runs through all prime numbers.

The interest of the lemma is of course in the case when χ is a non-trivial Dirichlet
character; this is the case that we will apply the lemma below. The left side of the above
lemma is usually called the Dirichlet L-function. It is usually denoted L(s, χ), i.e.
∞
X χ(n)
L(s, χ) = .
n=1
ns
The sum is convergent for real numbers s > 1. The proof of Lemma 7 makes use of the
multiplicative property of the Dirichlet characters, namely
χ(ab) = χ(a)χ(b)
for all a, b ∈ Z, and is otherwise along the same line as that of the ordinary Euler’s product
formula. We omit its proof.

In fact both sides of the above product formula have natural analytic continuation to
the half-space {Re s > 1} in the complex plane, and the equality continues to hold there.
Since we shall not make use of that, we shall be contented with the above ‘real’ version
of the product formula.

7
Proof of Theorem 6. Following the proof of Theorem 2, we take the logarithm of both
sides of (4). Now both sides of (4) involve complex numbers since they involve non-trivial
Dirichlet characters. As a result we must take the complex logarithm instead of the real
one. The result is, for appropiate branches of the complex logarithm, that
 
X χ(p)
(5) log L(s, χ) = − log 1 − s
p
p

for all s > 1. We shall be more careful about how the branches are chosen, but arguing
formally for the moment, assuming that we have the appropiate power series expansion
of the right hand side of the above identity, we get
X χ(p) X X χ(p)
(6) log L(s, χ) = + O(p−2s ) = + O(1)
p
ps p p
ps

for s > 1. Hence the desired boundedness of the sum

X χ(p)

p
ps

as s → 1+ follows from the following theorem:

Theorem 8. If χ is a non-trivial Dirichlet character modulo q, then L(s, χ) continues to
a continuously differentiable function for s ∈ (0, ∞), and

L(1, χ) 6= 0.

We shall prove this in the next section.

To justify the formal argument given above, first we choose an appropiate branch of
logarithm for the two sides of (5). For the left hand side of (5), we define, for s > 1 and
χ 6= χ0 , Z ∞ 0
L (t, χ)
log L(s, χ) = − dt;
s L(t, χ)
this is legitimate since L(t, χ) = 1 + O(e−ct ) and L0 (t, χ) = O(e−ct ) as t → ∞. For the
right hand side of (5), note that χ(p)p−s has absolute value less than 1 for all primes p
and all s > 1; thus one can simply use the principal branch of logarithm for the terms on
the right hand side of (5). In particular, the power series representation is valid:
 
1 X χ(p)
log = − + O(p−2s ),
1 − χ(p)p−s p
p s

which incidentally justifies the equality (6). To show that the equality in (5) holds, note
that since we are just taking possibly different branches of logarithm, the two sides of the

8
equation differ at most by an integral multiple of 2πi that may depend on s. Write the
difference of the two sides as 2πiM (s) where M (s) is an integer for all s. Then M is a
continuous function of s that takes only integer values, and thus M is constant. Letting
s → ∞, M is identically zero. Hence the two sides of (5) are equal.

It thus remains to prove Theorem 8, which is the goal of the next section.

2.4 Non-vanishing of L(1, χ)

Let χ be a non-trivial Dirichlet character modulo q, fixed throughout this section.

Proposition 9. L(s, χ) can be analytically continued to Re s > 0.

Pn
Proof. We use summation by parts. Let sn = k=1 χ(k) (and s0 = 0). Then for any
positive integer N ,
N N N −1  
X χ(n) X sn − sn−1 X 1 1 sN
(7) = = sn − + .
n=1
ns n=1
ns n=1
ns (n + 1)s Ns

Now sn is a bounded sequence, since

q
X
χ(k) = φ(q) < χ, χ0 >= 0
k=1

by the orthogonality of χ and χ0 asserted in Theorem 5; the same holds as long as the
sum is over a complete residue class modulo q. It follows that |sn | ≤ q for all n ∈ N. Next
for n ∈ N and s with Re s > 0,

1 1 |s|
s
− s
≤ Res+1 .
n (n + 1) n

Hence
N −1  
X 1 1
sn −
n=1
ns (n + 1)s
converges absolutely and locally uniformly for Re s > 0 as N → ∞, and this gives via (7)
the desired analytic continuation of L(s, χ) to Re s > 0.

Proposition 10. L(1, χ) 6= 0.

Proof. We consider two cases: the case where χ takes on some (non-real) complex values,
so that χ is a (non-trivial) Dirichlet character different from χ, and the case where χ takes
on only real values, i.e. when χ = χ.

9
Case 1: χ 6= χ

We need the following lemma:

Lemma 11. If s > 1 then
Y
(8) L(s, χ) ≥ 1
χ

where the product extends over all Dirichlet characters χ modulo q. (In particular the
product is real.)

Proof of lemma. By the Euler product formula (5) for L-functions, if log represents the
principal branch of the complex logarithm, then for s > 1,
!
Y Y X 1
L(s, χ) = exp log
χ χ p
1 − χ(p)p−s
∞
!
k
XXX 1 χ(p)
= exp
χ p k=1
k psk
∞
!
XX 1 X
= exp sk
χ(pk )
p k=1
kp χ
∞
!
XX φ(q)δ1 (pk )
= exp sk
≥1
p k=1
kp

where the last equality follows from the Fourier series expansion (3) of the function δ1 .

With the lemma we now prove L(1, χ) 6= 0 for our fixed non-trivial complex Dirichlet
character χ. In the product in (8), all but the trivial character contribute a factor that
is holomorphic near s = 1. The only factor that blows up corresponds to the trivial
character χ0 , and since by (4),
−s
L(s, χ0 ) = (1 − p−s −s
1 )(1 − p2 ) . . . (1 − pN )ζ(s)

where p1 , p2 , . . . , pN is a listing of all prime factors of q, it follows that the factor L(s, χ0 )
can only blow up like a constant multiple of
1
ζ(s) ∼
s−1
near s = 1. (The fact that ζ has a simple pole at s = 1 was mentioned after the proof
of Theorem 2 and will be proved in section 3.1 below.) Now if for our fixed non-trivial
complex Dirichlet character, we have L(1, χ) = 0, then L(s, χ) vanishes up to order at

10
least 1 near s = 1; the same happens to χ, since then L(1, χ) = L(1, χ) = 0. However
χ is a different (non-trivial) Dirichlet character modulo q from χ. Hence there are two
different terms in the product (8), each vanishing up to order at least 1 near s = 1. This
kills the only possible singularity of Qthe product, namely the one corresponding to the
term L(s, χ0 ), and thus the product χ L(s, χ) tends to 0 as s → 1+ , contradicting (8).
This shows that L(1, χ) cannot be zero, and concludes the proof in this case.

Case 2: χ = χ

Since χ is the same Dirichlet character as χ, the above argument would not work in
this case. We shall take another route: we shall consider the sums
X χ(n)
Sk =
mn≤k
(mn)1/2

where k is a positive integer and the sum is over all ordered pairs (m, n) of positive integers
whose product does not exceed k. The key is the following:
Lemma 12. If χ is a real non-trivial Dirichlet character, then

(a) Sk = 2k 1/2 L(1, χ) + O(1) as k → ∞;

(b) Sk ≥ c log k for some constant c > 0.

It follows that L(1, χ) cannot be zero.

Proof of lemma. (a) The proof depends on two different ways of summing Sk . Note that
the sum is over the lattice points (m, n) that lie in the first quadrant of the (m, n)
plane under the hyperbola mn = k. Draw the (m, n) plane such that m represent
the horizontal coordinate and n the vertical coordinate. First we sum ‘vertically and
horizontally’: we write Sk as
 
 χ(n) = I + II.
X X X X
Sk =  +
1/2 1/2 1/2
(mn)1/2
1≤m<k k <n≤k/m 1≤n≤k 1≤m≤k/n

Now
X 1 X χ(n)
(9) I= ,
m1/2 n1/2
1≤m<k1/2 k1/2 <n≤k/m

and to exploit the cancellation due to the factors χ(n) in the inner sum, we estimate
the inner sum by summation by parts in a way similar to how we analytically continued

11
the L-functions. In fact again denoting nk=1 χ(k) by sn , we have by the boundedness
P
of {sn } that
b b−1   b−1
X χ(n) X 1 1 −1/2
X 1
1/2
= s n 1/2
− 1/2
+ O(a ) ≤ C 3/2
+ O(a−1/2 ).
n=a
n n=a
n (n + 1) n=a
n
Pb−1 1 Rb 1
However, the sum n=a n3/2 can be approximated by the integral a x3/2 dx, and the
Pb−1 1 −1/2
result is that n=a n3/2 = O(a ). Hence applying this to the inner sum in I, we
get
X 1
I ≤ Ck −1/4 .
m1/2
1≤m<k1/2

Once again we can use the estimate 1≤m<k1/2 m11/2 = O(k 1/4 ) that we obtain by
P
R k1/2 1
approximating the sum by the integral 1 x1/2
dx. Hence
I = O(1)
is a bounded term as k → ∞.
Next we estimate II: we write II as
X χ(n) X 1
II = .
n1/2 m1/2
1≤n≤k1/2 1≤m≤k/n

By comparing it to an integral, the inner sum is 2(k/n)1/2 + c + O((k/n)−1/2 ) for some

constant c; in fact
X 1 Z b X Z m+1  1 
1 1
1/2
− 1/2
dx ≤ 1/2
− 1/2 dx
1≤m≤b
m 1 x m m x
1≤m≤bbc
X 1
≤
2m3/2
1≤m≤bbc
∞ ∞
X 1 X 1
= 3/2
−
m=1
2m 2m3/2
m=bbc+1
−1/2
= c + O(b ).
Hence
 
 1/2  −1/2
X χ(n) k X χ(n) X χ(n) k
II = 2 +c + O  
n1/2 n n1/2 n1/2 n
1≤n≤k1/2 1≤n≤k1/2 1/2 1≤n≤k
X χ(n) X χ(n)
= 2k 1/2 +c + O(1)
n n1/2
1≤n≤k1/2 1≤n≤k1/2
X χ(n) X χ(n)
= 2k 1/2 L(1, χ) + 2k 1/2 +c + O(1)
1/2
n n1/2
n>k 1≤n≤k1/2

12
 The two sums can be estimated by first summing by parts and then approximating
by integrals, in very much the same way that we tackled the inner sum in (9). The
result is that the two terms remain bounded as k → ∞, so together with the estimate
for I we get
Sk = 2k 1/2 L(1, χ) + O(1)
as desired.
(b) This time we sum Sk by summing along the hyperbolas; we write
k X k
X χ(n) X 1
Sk = = A
1/2 j
j=1 mn=j
(mn)1/2 j=1
j

where X
Aj = χ(n).
n|j

However,
Aj ≥ 0
for all j, and
Aj ≥ 1
if j is a perfect square. This can be proved as follows: first, since χ is multiplicative,
Aj as a function of j is also multiplicative. It follows that we only have to consider
the case where j = pN is a prime power. Now if j = pN for some prime p, then since
now χ is assumed to take only real values (i.e. 0, 1 and −1 since the values of χ lies
on U (1)), we have


N +1 if χ(p) = 1
N 
X 1 if χ(p) = −1 and N is even
Aj = χ(p)i = .
i=0


0 if χ(p) = −1 and N is odd

1 if χ(p) = 0, i.e. if p divides q

In any case Aj ≥ 0, and Aj ≥ 1 if j = pN with N even. This proves the assertions

about Aj . As a result,
k
X 1 X 1 X 1
Sk = 1/2
Aj ≥ = ≥ c log k,
j a a
j=1 {a∈N : a2 ≤k} 1/2
a≤k

as desired. This completes the proof of the lemma, and thus completes the proof of
the Dirichlet’s theorem.

13
2.5 Representations of finite abelian groups

Finally, we prove the Peter-Weyl theorem as stated in Theorem 5 for finite abelian groups.
There is a more general version of Peter-Weyl theorem for topological groups (not neces-
sarily abelian or finite). Here we shall just give the proof in the simple setting of finite
abelian groups.

In this section G will be a finite abelian group (group law written multiplicatively),
and Ĝ will be the group of characters of G. L2 (G) will denote the complex vector space
of all complex-valued functions on G, with inner product given as in Theorem 5.

First we verify the orthogonality between distinct characters: If χ1 6= χ2 are both

characters of G, then
1 X
< χ1 , χ2 >= χ1 (a)χ2 (a) =< χ1 χ−1
2 , χ0 >
|G| a∈G

where χ0 is the trivial character, and χ1 χ−1

2 is a non-trivial character. Hence it suffices
to verify the orthogonality between the trivial character χ0 and a non-trivial character χ,
i.e. X
χ(a) = 0
a∈G
P
for any non-trivial χ ∈ Ĝ, to which we now turn our attention. Let S = a∈G χ(a). Then
since χ ∈ Ĝ is non-trivial, χ(b) 6= 1 for some b ∈ G. Hence by change of variable,
X X
S= χ(a) = χ(ab) = χ(b)S
a∈G a∈G

which implies that S = 0, as desired.

Next we verify that there are enough characters to span L2 (G). Our proof is based
on a lemma in linear algebra, concerning the simultaneous diagonalization of a family of
commuting linear operators on a finite dimensional complex inner product space.

First note that if χ ∈ Ĝ, then for any b ∈ G, we have

χ(ab) = λb χ(a)

for all a ∈ G, where λb = χ(b). In other words, if Tb : L2 (G) → L2 (G) is the operator given
by translation by b (i.e. given any function f on G, Tb (f ) is defined to be the function
such that Tb (f )(a) = f (ab) for all a ∈ G), then

Tb (χ) = λb χ

for all b ∈ G and all χ ∈ Ĝ. Hence given any b ∈ G, any character χ of G is an eigenvector
of the operator Tb . The converse is to a large extent true: if f ∈ L2 (G) is an eigenvector

14
of the operator Tb for all b ∈ G, then there exists a family of complex constants λb indexed
by b ∈ G such that
f (ab) = λb f (a)
for all a, b ∈ G. It follows that λb has to be f (b)/f (1), i.e.
1
f (ab) = f (a)f (b)
f (1)
for all a, b ∈ G. Hence if in addition f (1) = 1, we have then |f (a)| = 1 for every a ∈ G
and so f is a character of G, i.e. f ∈ Ĝ. This says
Lemma 13. A function f defined on G is a character of G if and only if f (1) = 1 and
f is an eigenvector of Tb for any b ∈ G.

Hence our goal reduces to showing that there are enough simultaneous eigenvectors of
the family of operators {Tb : b ∈ G} such that they span L2 (G). The key observations are
that all these Tb ’s commute, i.e.
Ta Tb = Tb Ta
for all a, b ∈ G, and that the Tb ’s are all unitary operators on L2 (G), i.e.
< Tb (f ), Tb (g) >=< f, g >
for all f, g ∈ H and all b ∈ G, which allow us to invoke the following lemma.
Lemma 14. If {Ti }ki=1 is a commuting family of unitary operators on a finite dimensional
complex inner product space H, then there is a basis {f1 , . . . , fn } of H such that each fj
is an eigenvector of all the Ti ’s, i.e. the Ti ’s can be simultaneously diagonalized.

This lemma applied to the family {Tb : b ∈ G} of operators on H = L2 (G) allows us

to assert that there is an eigenbasis {f1 , . . . , fn } such that each fj is an eigenvector of all
the Tb ’s. Renormalizing the fj ’s such that
f
f˜j := ,
fj (1)

we obtain a basis {f˜1 , . . . , f˜n } of L2 (G) such that each f˜j is a character of G, according
to Lemma 13. This proves that L2 (G) admits a basis consisting of characters of G and
thus concludes the proof of Theorem 5, save that we have to verify that it is possible to
renormalize the fj ’s, i.e. we are left to check that fj (1) 6= 0 for all j. This is easy though:
if fj (1) = 0 then since fj is an eigenvector of all the Tb ’s, we have
fj (b) = Tb (fj )(1) = λj,b fj (1) = 0
for all b ∈ G, where λj,b is the eigenvalue of Tb corresponding to the eigenvector fj ; but
this says fj is identically zero, which is absurd since fj is a basis vector of L2 (G). This
proves that fj (1) 6= 0 for all j, and we are done.

15
Proof of Lemma 14. The proof is by induction on the number k of commuting unitary
operators involved. When k = 1, this is the well-known fact that a unitary operator on a
finite dimensional complex inner product space is diagonalizable. So now assume that the
assertion holds for some k ∈ N and take k + 1 commuting unitary operator on H. First
Tk+1 is unitary, thus H decomposes into eigenspaces corresponding to distinct eigenvalues,
say H = H1 ⊕ H2 ⊕ . . . Hm , with Hj being the eigenspace of Tk+1 with eigenvalue λj . Now
T1 , . . . , Tk preserve each of the Hj ’s, because if v ∈ Hj , then for i = 1, . . . , k,

Tk+1 (Ti v) = Ti Tk+1 v = Ti (λj v) = λj (Ti v).

It follows that {T1 , . . . , Tk } are commuting unitary operators on each of these Hj ’s, thus
are simultaneously diagonalizable on these Hj ’s by induction hypothesis. This gives a
simultaneous diagonalization of the k + 1 operators {T1 , . . . , Tk+1 } on the bigger space H,
and Lemma 14 is proved.

3 Prime number theorem

There are a number of different approaches to the prime number theorem, some more
elementary and others more sophiscated. Since our aim is to illustrate how analysis can
be used to study number theory, we shall present here a classical proof using complex
analysis. We shall derive the prime number theorem from properties of the Riemann zeta
function. Before that, we shall need to take a closer look at the zeta function, and in
particular investigate the zeros of the zeta function on the line {Re s = 1}.

3.1 Analytic continuation of ζ

The zeta function was defined by a series

∞
X 1
ζ(s) =
n=1
ns

which converges for all complex s with Re s > 1. For a deeper understanding of the zeta
function, it is necessary to analytically continue to a larger region in the complex plane.
There are two customary ways of doing that, one via a functional equation that arises
from the connection of the zeta function with a modular form, namely the Jacobi theta
function ϑ(t), and another via comparing the sum with an integral. The latter approach
is more like what we have been doing in proving the non-vanishing of L(1, χ) for real χ,
and despite being more elementary, it is less powerful than the first approach, for we do
not get a functional equation out of it. Below we describe both approaches.

16
3.1.1 Functional equation approach

Here we shall make use of the theta function

∞
2t
X
ϑ(t) = e−πn
n=−∞

(t > 0) and the Poisson summation formula, which says that for a function f defined on
the real line that satisfies suitable decay estimates, e.g.
A A
|f (x)| ≤ and |fˆ(ξ)| ≤ ,
1 + x2 1 + ξ2
we have
∞
X ∞
X
(10) f (x + n) = fˆ(n)e2πinx
n=−∞ n=−∞

as L1 periodic functions of x (x taking values in [0, 1]). In other words, the two ways
of periodizing the function f gives the same result, and if further f is continuous, then
setting x = 0 we get
X∞ X∞
f (n) = fˆ(n).
n=−∞ n=−∞

The proof of the Poisson summation formula is not too difficult; one simply checks
that both sides of (10) are periodic L1 functions of x defined on [0, 1], and that their
Fourier coefficients agree at every n ∈ N. The details are left to the reader.
2t
The point here is that the Poisson summation formula applied to e−πn gives an
elegant functional equation for ϑ, namely

ϑ(t) = t−1/2 ϑ(1/t)

2
for real t > 0. One just has to note that e−πt is its own Fourier transform, and apply the
Poisson summation formula. The details are omitted.

Now what has the zeta function has to do with the theta function? The connection
here is given by the principle of subordination via the Gamma function. Recall that
Z ∞
dt
Γ(s) = e−t ts
0 t

for complex s with Re s > 0. We want to study ζ(s), which is the sum of n−s over n,
using the ϑ function. So we squeeze out the term n−s from this definition of Γ, and at the

17
 2
same time create terms of the form e−πn t : in fact by the customary change of variable,
we have, for Re s > 1, that
s Z ∞ Z ∞
−πn2 t 2 s/2 dt 2 dt
Γ = e (πn t) =π ns/2 s
e−πn t ts/2 ,
2 0 t 0 t
and thus s Z ∞
−s/2 −s 2 dt
π n Γ = e−πn t ts/2 .
2 0 t
Summing over n ∈ N we get
∞
ϑ(t) − 1 s/2 dt
s Z
−s/2
π ζ(s)Γ = t .
2 0 2 t
R1 R∞
Break the integral into two parts, namely 0 + 1 , and using the functional equation for
ϑ along with a change of variable in the first integral, we get
Z ∞
−s/2
s 1 1
ϑ(t) − 1
π ζ(s)Γ = − + t(1−s)/2−1 + ts/2−1 dt
2 s−1 s 1 2
in which the integral converges for all s ∈ C and is invariant under the substitution
s ↔ 1 − s. Hence if we define ξ(s) to be the right hand side of the above equality and
insist that we have s
ξ(s) = π −s/2 ζ(s)Γ ,
2
then we have meromorphically extended ζ to the whole complex plane, and we have the
functional equation
ξ(s) = ξ(1 − s),
which also serves as a functional equation for ζ upon the substitution s ↔ 1 − s.

Note that ξ is holomorphic except at the simple poles s = 0 and s = 1, and from
1
(11) ζ(s) = π s/2 ξ(s)
Γ(s/2)
1
we see that the simple pole of ξ at s = 0 is cancelled by the simple zero of Γ(s/2) there.
1/2
Furthermore, at s = 1, Γ(1/2) = π . As a result, ζ is now holomorphic on the whole
complex plane, with a simple pole at s = 1 where ζ behaves like
1
ζ(s) ∼ .
s−1
This settles the questions about the singularities of ζ.

The zeros of ζ are far more fascinating. From (11) one sees immediately that ζ(s) = 0
1
whenever s is a negative even integer, for Γ(s/2) vanishes there. These are called the trivial
zeros of ζ, and they are the only zeros of ζ off the critical strip {0 ≤ Re s ≤ 1}. This

18
is because ζ is nowhere zero on {Re s > 1}, and thus so is ξ; then by the functional
equation, ξ has no zeros in the region {Re s < 0} as well, so by (11) one sees that the only
1
possible zeros of ζ off the critical strip comes from those of Γ(s/2) , and they are precisely
the negative even integers. The famous Riemann hypothesis asserts that the only
non-trivial zeros of ζ (i.e. those lying in the critical strip) lie on the line Re s = 12 . This
has remained open for more than a hundred years, despite the effort of many first rate
mathematicians. In fact the zeros of ζ inside the critical strip has been of great interest
since Riemann’s famous memoir in 1859. The proof of the prime number theorem that
we shall give involves showing that ζ has no zeros on the line Re s = 1, and in fact it was
known that the prime number theorem is equivalent to the fact that ζ does not vanish on
the line Re s = 1.

3.1.2 Integral approach

Next we describe the integral approach of analytically continuing ζ. Here we compare

the partial sums of ζ(s) to an integral, and continue ζ to the region {Re s > 0}. More
precisely, we write, for Re s > 1 and N ∈ N, that
N −1 Z N N −1
X 1 1 X
(12) − dx = δn (s),
n=1
ns 1 xs n=1

where Z n+1  
1 1
δn (s) = s
− s dx.
n n x
Note that each δn is an entire function of s, and satisfies an estimate
Z n+1
1 1 |s|
|δn (s)| ≤ − dx ≤
n ns xs nRe s+1
by the mean value inequality applied to x 7→ x−s . Hence
∞
X
H(s) := δn (s)
n=1

converges locally uniformly on the region {Re s > 0}, and defines a holomorphic function
there. Letting N → ∞ in (12), we have
1
ζ(s) = + H(s)
s−1
for Re s > 1. However, since the right hand side of this equation is defined and meromor-
phic on {Re s > 0}, it provides an meromorphic continuation of ζ to this larger region,
1
and it is again readily seen that ζ has a simple pole s−1 at the point s = 1, as we have
observed using the functional equation approach.

19
3.2 Estimates of ζ and ζ 0

Our goal now will be to establish some estimates of ζ (and its derivative) on the line
Re s = 1, which is a key to proving the prime number theorem. This will be based on the
integral approach of continuing ζ that we described above.

Lemma 15. For each σ0 ∈ (0, 1] and each ε > 0, there exists a constant cε such that

(a) |ζ(s)| ≤ cε |Im s|1−σ0 +ε for all s with Re s ≥ σ0 and |Im s| ≥ 1.

(b) |ζ 0 (s)| ≤ cε |Im s|ε for all s with Re s ≥ 1 and |Im s| ≥ 1.

Proof. To see that this is true, first observe that the second assertion follows from the first
by Cauchy’s estimates: if Re s ≥ 1 and |Im s| ≥ 1, then given ε > 0, Cauchy’s estimate
for holomorphic functions gives
1
|ζ 0 (s)| ≤ max |ζ(t)|,
ε/2 |t−s|=ε/2

so applying (a) with σ0 ≥ 1 − ε/2 and ε/2 in place of ε we have

cε/2 (|Im s| + ε/2)1−(1−ε/2)+ε/2

|ζ 0 (s)| ≤ = O(|Im s|ε )
ε/2

with an implicit constant depending on ε. For the first assertion it suffices to consider
the case where ε is small, say 1 − σ0 + ε < 1. In this case the inequality follows from two
estimates of δn (s), one being the estimate

|s|
|δn (s)| ≤
nRe s+1
which we already alluded to, the other being
2
|δn (s)| ≤
nRe s
which holds trivially by estimating the integrand of the integral defining δn . Now taking
a geometric mean of these two estimates, for any η ∈ [0, 1], we have
η  1−η
2|s|η

|s| 2
|δn (s)| ≤ ≤ .
nRe s+1 nRe s nσ0 +η

Taking η = 1 − σ0 + ε, we have

|Im s|1−σ0 +ε
|δn (s)| ≤ C .
n1+ε

20
This is because if Re s ≤ 1 + ε, then
p p
|s| ≤ (1 + ε)2 + |Im s|2 ≤ |Im s| (1 + ε)2 + 1

while if Re s ≥ 1 + ε the estimate is trivial:

2 2 2|Im s|1−σ0 +ε
|δn (s)| ≤ ≤ ≤ .
nRe s n1+ε n1+ε
Hence by (12),
∞
1 X 1
|ζ(s)| ≤ + C|Im s|1−σ0 +ε 1+ε
≤ cε |Im s|1−σ0 +ε
s−1 n=1
n

as desired.

We shall need just one more estimate of ζ on the line {Re s = 1}, this time from below.
This is really a quantitative way of saying that ζ has no zeros on Re s = 1 (note that ζ
blows up near the pole s = 1), and in fact if one examines the proof that follows, one will
see that we have actually shown that ζ(s) 6= 0 on the line Re s = 1.

Lemma 16. For every ε > 0, there exists cε such that

1
≤ cε |Im s|ε
ζ(s)

for all s with Re s ≥ 1 and |Im s| ≥ 1.

Proof. First observe that

log |ζ 3 (σ)ζ 4 (σ + it)ζ(σ + 2it)| ≥ 0

for all real σ > 1 and real t. This is because by the Euler product formula, for Re s > 1,
  X X p−ms ∞
X 1 X
log ζ(s) = log = = cn n−s
p
1 − p−s p m
m n=1

where cn = 1/m if n is a prime power with n = pm and cn = 0 otherwise. It follows that

for σ > 1 and t is real, then

log |ζ 3 (σ)ζ 4 (σ + it)ζ(σ + 2it)|

=3Re log ζ(σ) + 4Re log ζ(σ + it) + Re log ζ(σ + 2it)
X
= cn n−σ (3 + 4 cos θn + cos 2θn )
n

21
where θn = t log n. However, cn ≥ 0 for all n, and 3 + 4 cos θ + cos 2θ = 2(1 + cos θ)2 ≥ 0
for all real θ. Thus
log |ζ 3 (σ)ζ 4 (σ + it)ζ(σ + 2it)| ≥ 0
as desired.

From this we prove the lemma: note now for σ > 1 and |t| ≥ 1, we have

|ζ 3 (σ)ζ 4 (σ + it)ζ(σ + 2it)| ≥ 1.

Hence for σ large, using Lemma 15, we have

|ζ(σ + it)| ≥ c−1

ε |ζ(σ)|
−3/4
|2t|−ε/4 ≥ c0ε |t|−ε/4 .

The inequality we want to prove follows easily from this.

For σ close to 1, since zeta function has simple pole at 1, we have

(13) |ζ(σ + it)| ≥ c−1

ε |ζ(σ)|
−3/4
|2t|−ε/4 ≥ c0ε (σ − 1)3/4 |t|−ε/4 .

If σ − 1 ≥ A|t|−5ε for some constant A (the value of A to be specified below), then we are
in good shape, and
|ζ(σ + it)| ≥ c00ε |t|−4ε
as desired. Hence it suffices to consider the case where σ − 1 < A|t|−5ε . However, if this is
the case, then we can first choose an σ̃ such that σ̃ > σ, and σ̃ − 1 = A|t|−5ε . Then since

|ζ(σ + it)| ≥ |ζ(σ̃ + it)| − |ζ(σ + it) − ζ(σ̃ + it)|,

we have by (13) that

|ζ(σ + it)| ≥ c0ε (σ̃ − 1)3/4 |t|−ε/4 − |ζ(σ + it) − ζ(σ̃ + it)|.

Applying mean value theorem, the last term in absolute value is estimated by

|ζ(σ + it) − ζ(σ̃ + it)| ≤ c000 ε 000

ε |σ̃ − σ||t| ≤ cε (σ̃ − 1)|t|
ε

using the estimate for ζ 0 in Lemma 15. Hence

|ζ(σ + it)| ≥ c0ε (σ̃ − 1)3/4 |t|−ε/4 − c000 ε

ε (σ̃ − 1)|t| .

Set now A = (c0ε /(2c000 4

ε )) and using the fact that σ̃ − 1 = A|t|
−5ε
, we have the first term
000 ε
on the right hand side exactly equal to 2cε (σ̃ − 1)|t| . Hence

|ζ(σ + it)| ≥ c000 ε 0000 −4ε

ε (σ̃ − 1)|t| ≥ cε |t|

as was to be shown in this case as well.

22
3.3 Chebyshev’s ψ function

Having understood more the zeta function, now we turn to the proof of the prime number
theorem. We shall need two auxillary functions. The first one is the Chebyshev’s ψ
function, defined for real x by X
ψ(x) = log p
pm ≤x

where the sum is over all prime powers pm not exceeding x. It is readily seen that
X  log x 
ψ(x) = log p
p≤x
log p

where here the sum is over all primes p not exceeding x. The second one is the integral
of ψ, which we denote by ψ1 : Z x
ψ1 (x) = ψ(t)dt.
1
The following notation will be adopted: we write

f (x) ∼ g(x)

to mean
f (x)
lim = 1.
x→∞ g(x)

Hence the prime number theorem says that

x
π(x) ∼ .
log x

To prove the prime number theorem, we make a series of reductions. First we have
x
Theorem 17. π(x) ∼ log x
if and only if ψ(x) ∼ x.

Proof. Note that

X log x
ψ(x) ≤ log p = π(x) log x,
p≤x
log p

and that for all α ∈ (0, 1),

X
ψ(x) ≥ log p ≥ (π(x) − π(xα )) log(xα ) ≥ α(π(x) − xα ) log x,
xα <p≤x

the last inequality following from the fact that π(xα ) ≤ xα . Hence
ψ(x) π(x) 1 ψ(x) log x
≤ x ≤ + 1−α .
x log x
α x x

23
It follows that if ψ(x) ∼ x, then letting x → ∞ in the above estimate, we get

π(x) π(x) 1
1 ≤ lim inf x ≤ lim sup x ≤
x→∞
log x x→∞ log x
α

for all α ∈ (0, 1), and letting α → 1− , we get π(x) ∼ x

log x
. One can prove the converse in
a similar fashion.

Next we note that

Theorem 18. ψ(x) ∼ x if and only if ψ1 (x) ∼ x2 /2.

Proof. This is because for any β > 1,

Z βx
1 1
ψ(x) ≤ ψ(t)dt = (ψ1 (βx) − ψ1 (x)),
(β − 1)x x (β − 1)x

the first inequality following from the fact that ψ is an increasing function of x. It follows
that  
ψ(x) 1 ψ1 (βx) 2 ψ1 (x)
≤ β − ,
x β − 1 (βx)2 x2
and if ψ1 (x) ∼ x2 /2 then letting x → ∞ we get
 
ψ(x) 1 1 2 1 1
lim sup ≤ β − ≤ (β + 1).
x→∞ x β−1 2 2 2

Letting β → 1− we get
ψ(x)
lim sup ≤ 1.
x→∞ x
Similarly, this time using the fact that for 0 < α < 1,
Z x
1 1
ψ(x) ≥ ψ(t)dt = (ψ1 (x) − ψ1 (αx)),
(1 − α)x αx (1 − α)x
we get
ψ(x)
lim inf≥1
x
x→∞

assuming ψ1 (x) ∼ x2 /2. It follows that ψ(x) ∼ x, and the converse can be proved
similarly.

24
3.4 Proof of the prime number theorem

According to the results in the last section, to prove the prime number theorem it suffices
to prove ψ1 (x) ∼ x2 /2. This is what we shall aim at now. Here more complex analysis
comes in. First we relate ψ1 to the zeta function, so that we can make use of our knowledge
of ζ that we gathered in the previous sections. This is done via a magic identity:

Lemma 19. For all real c > 1,

c+i∞
xs+1 ζ 0 (s)
Z
1
(14) ψ1 (x) = − ds,
2πi c−i∞ s(s + 1) ζ(s)

the contour integral being upwards along the vertical line Re s = c.

Proof. We have on one hand the ψ1 function, and on the other the ζ function. To make
the two ends meet, we use the Λ function, defined for n ∈ N by
(
log p if n = pk for some prime p and some k ≥ 1
Λ(n) = .
0 otherwise

First,
X ∞
X
ψ(t) = Λ(n) = Λ(n)fn (t)
1≤n≤t n=1

where fn is the characteristic function of the interval [n, ∞). Hence

Z x ∞
X Z x X
ψ1 (x) = ψ(t)dt = Λ(n) fn (t)dt = Λ(n)(x − n).
1 n=1 1 1≤n≤x

Next, taking the logarithmic derivative of the Euler product formula, we get
∞
ζ 0 (s) X X log p X Λ(n)
− = = .
ζ(s) p m
pms n=1
ns

We would like to plug this into the right hand side of the desired identity and evaluate
the contour integrals. The relevant contour integrals are given by the following lemma.
Lemma 20. If c > 0 then
(
c+i∞
as if a ∈ (0, 1]
Z
1 0
ds = ,
2πi c−i∞ s(s + 1) 1 − 1/a if a ≥ 1

the integral taken along the vertical line Re s = c.

25
 Assuming this for the moment, we see that
c+i∞ ∞ Z c+i∞
xs+1 ζ 0 (s) (x/n)s
Z
1 X 1
− ds = x Λ(n) ds
2πi c−i∞ s(s + 1) ζ(s) n=1
2πi c−i∞ s(s + 1)
X  n
=x Λ(n) 1 −
1≤n≤x
x
= ψ1 (x),

proving Lemma 19.

Hence it suffices now to prove the contour integral in Lemma 20.

Proof of Lemma 20. Indeed the contour integral can be evaluated by residue theorem. If
a ≥ 1, then for T >> 0, replace the contour by ΓT = L1 + C2 + L3 , where L1 is the
vertical line from c − i∞ to c − iT , C2 is the half circle centered at c and of radius T that
lies to the left of Re s = c, and L3 is the vertical line from c + iT to c + i∞. The residue
that one gains this way (at the poles s = 0, −1) is precisely 1 − 1/a, and as T → ∞ the
new contour integral tends to zero (here one uses that a ≥ 1). This proves the second
part of the assertion in Lemma 20. The other part is proved similarly, except that this
time one has to replace the contour by Γ̃T which one obtains by reflecting ΓT along the
line Re s = c. That this is necessary is due to the fact that one needs this to argue that
the contour integral along the new contour tends to zero as T → ∞, and when a ∈ (0, 1]
this can only be achieved by using the contour Γ̃T . The details are left to the reader.

We shall now use the magic identity in Lemma 19 to finish the proof of the prime
number theorem. We will verify that ψ1 (x) ∼ x2 /2 using Lemma 19 and the estimates in
section 3.2. The idea is to shift the contour integral to from Re s = c (c > 1) to Re s = 1,
so that the power of x in the integrand of (14) gets closer to the desired value of 2.

More precisely, we shall do the following: write the integrand in (14) as F (s). Then
for any T > 0, if we denote by γT the contour l1 + l2 + l3 + l4 + l5 , where

• l1 is the vertical line from 1 − i∞ to 1 − iT

• l2 is the horizontal line from 1 − iT to c − iT

• l3 is the vertical line from c − iT to c + iT

• l4 is the horizontal line from c + iT to 1 + iT

• l5 is the vertical line from 1 + iT to 1 + i∞,

26
we see that
Z c+i∞ Z
(15) F (s)ds = F (s)ds.
c−i∞ γT

To see this we make use of the residue theorem. For this we use the estimate
ζ 0 (s)
≤ C|Im s|1/2
ζ(s)

for Re s ≥ 1, |Im s| ≥ 1 that we obtain by Lemma 15 and 16, which imply the decay

|F (s)| ≤ C|Im s|−3/2

for the same s. This shows that F (s) is integrable along the contour (c − i∞, c + i∞) − γT ,
and by residue theorem, since F (s) has no singularity in the region enclosed by this
contour, it follows that its integral along the contour is zero, and (15) is established.

Next we change the contour again: let γT,δ be the contour l1 + l6 + l7 + l8 + l5 , where
l1 , l5 are as above, and

• l6 is the horizontal line from 1 − iT to 1 − δ − iT

• l7 is the vertical line from 1 − δ − iT to 1 − δ + iT

• l8 is the horizontal line from 1 − δ + iT to 1 + iT .

In doing so we have to be careful, and we need to ensure that F (s) and hence 1/ζ(s)
has no singularity on γT,δ so defined. This we do by choosing δ ∈ (0, 1) sufficiently small
(how small it is depend on our choice of T ) so that ζ has no zero on l6 + l7 + l8 , which is
possible since ζ has no zero on the line Re s = 1 (and hence in an open neighborhood of
(1 − iT, 1 + iT ) for each finite T > 0). Now

x2
Z Z
F (s)ds = F (s)ds + (2πi)(− ),
γT γT,δ 2

since the residue of F (s) at the pole s = 1 is −x2 /2. This is because

h(s)
ζ(s) =
s−1
near s = 1 for some non-zero holomorphic function h(s), and carrying out logarithmic
differentiation we see that
ζ 0 (s) 1
=− + h1 (s)
ζ(s) s−1
for some holomorphic function h1 (s) near s = 1.

27
 To summarize what we have got so far, we see that
xs+1 ζ 0 (s)
Z
1 2 1
ψ1 (x) = x − ds,
2 2πi γT,δ s(s + 1) ζ(s)

where T > 0 can be arbitrarily large and δ ∈ (0, 1) is sufficiently small relative to T .
Hence to verify that ψ1 (x) ∼ x2 /2, which has been the goal of this section, we shall prove
that
xs−1 ζ 0 (s)
Z
lim ds = 0.
x→∞ γ
T,δ
s(s + 1) ζ(s)
Now fix η > 0, and we want to see that the integral is smaller than η when x is sufficiently
large. Choosing T sufficiently large, we have
xs−1 ζ 0 (s)
Z
η
ds ≤ .
l1 +l5 s(s + 1) ζ(s) 2

This is because ζ 0 (s)/ζ(s) ≤ C|Im s|1/2 on Re s = 1, again from the estimates we got in
Lemma 15 and 16. Next fixing this T , we observe that
xs−1 ζ 0 (s)
Z
ds ≤ CT xRes−1 = CT x−δ ,
l7 s(s + 1) ζ(s)
and
1
xs−1 ζ 0 (s)
Z Z
1
ds ≤ C xσ−1 dσ ≤ C .
l6 +l8 s(s + 1) ζ(s) 1−δ log x
Hence choosing x large enough, we have
xs−1 ζ 0 (s)
Z
η
ds ≤ .
l6 +l7 +l8 s(s + 1) ζ(s) 2
This proves that
xs−1 ζ 0 (s)
Z
ds ≤ η,
γT,δ s(s + 1) ζ(s)
for large x, and hence ψ1 (x) ∼ x2 /2. This completes the proof of the prime number
theorem.

References
[1] E. Stein and R. Shakarchi, Fourier analysis, Princeton University Press, 2003.
[2] E. Stein and R. Shakarchi, Complex analysis, Princeton University Press, 2003.
[3] T. Apostol, Introduction to analytic number theory, New York: Springer-Verlag, 1976.

28