Chapter 1

Euler’s Product Formula

1.1 The Product Formula

The whole of analytic number theory rests on one marvellous formula due to
Leonhard Euler (1707-1783):
 −1
n−s = 1 − p−s
X Y
.
n∈N, n>0 primes p

Informally, we can understand the formula as follows. By the Funda-

mental Theorem of Arithmetic, each n ≥ 1 is uniquely expressible in the
form
n = 2e2 3e3 5e5 · · · ,
where e2 , e3 , e5 , . . . are non-negative integers (all but a finite number being
0).
Raising this to the power −s,

n−s = 2−e2 s 3−e3 s 5−e5 s · · · .

Adding for n = 1, 2, 3, . . . ,
   
n−s = 1 + 2−s + 2−2s + · · · 1 + 3−s + 3−2s + · · · 1 + 5−s + 5−2s + · · · · · · ,
X

each term on the left arising from just one product on the right. But for each
prime p,  −1
1 + p−s + p−2s + · · · = 1 − p−s ,
and the result follows.
P −s
Euler’s Product Formula equates the Dirichlet series n on the left
with the infinite product on the right.

1–1
1.2. INFINITE PRODUCTS 1–2

To make the formula precise, we must develop the theory of infinite prod-
ucts, which we do in the next Section.
To understand the implications of the formula, we must develop the the-
ory of Dirichlet series, which we do in the next Chapter.

1.2 Infinite products

1.2.1 Definition
Infinite products are less familiar than infinite series, but are no more com-
plicated. Both are examples of limits of sequences.
Definition 1.1. The infinite product
Y
cn
n∈N

is said to converge to ` 6= 0 if the partial products

Y
Pn = cm → ` as n → ∞.
0≤m≤n

We say that the infinite product diverges if either the partial products do
not converge, or else they converge to 0 (as would be the case for example if
any factor were 0).
Q
Proposition 1.1. If cn is convergent then
cn → 1.
Proof I We have
Pn
cn = .
Pn−1
Since Pn → ` and Pn−1 → `, it follows that
`
cn → = 1.
`
J
It is usually more convenient to write the factors in the form
cn = 1 + an .
In these terms, the last Proposition states that
Y
(1 + an ) convergent =⇒ an → 0.
1.2. INFINITE PRODUCTS 1–3

1.2.2 The complex logarithm

The theory of infinite products requires some knowledge of the complex log-
arithmic function.
Suppose z 6= 0. Let
z = reiθ ,
where r > 0. We are interested in solutions w ∈ C of
ew = z.
If w = x + iy then
ex = r, e−iy = e−iθ ,
ie
x = log r, y = θ + 2nπ
for some n ∈ Z.
Just one of these solutions to ew = z satisfies
−π < y = =(w) ≤ π.
We call this value of w the principal logarithm of z, and denote it by Log z.
Thus
eLog z = z, −π < =(z) ≤ π.
The general solution of ew = z is
w = Log z + 2nπi (n ∈ Z).
Now suppose
w1 = Log z1 , w2 = Log z2 .
Then
ew1 +w2 = z1 z2 = eLog(z1 z2 )
It follows that
Log(z1 z2 ) = Log z1 + Log z2 + 2nπi,
where it is easy to see that n = 0, −1 or 1.
If <(z) > 0 then z = reiθ with −π/2 < θ < π/2. It follows that
<(z1 ), <(z2 ) > 0 =⇒ −π/2 < =(Log z1 ), =(Log z2 ) < π/2;
and so
−π < =(Log z1 + Log z2 ) < π.
Thus
<(z1 ), <(z2 ) > 0 =⇒ Log(z1 z2 ) = Log z1 + Log z2 .
In particular, this holds if |z1 |, |z2 | < 1 (Fig 1.1).
1.2. INFINITE PRODUCTS 1–4

θ
1

Figure 1.1: |z − 1| < 1, Log z = log r + iθ

1.2.3 Convergence
Proposition 1.2. Suppose an 6= −1 for n ∈ N. Then
Y X
(1 + an ) converges ⇐⇒ Log(1 + an ) converges.
P
Proof I Suppose Log(1 + an ) converges to S. Let
X
Sn = Log(1 + am ).
m≤n

Then
eSn =
Y
(1 + am ).
m≤n

But
Sn → S =⇒ eSn → eS .
Thus (1 + an ) converges to es .
Q
Q
Conversely, suppose (1 + an ) converges. Let
Y
Pn = (1 + an ).
m≤N

Given  > 0 there exists N such that

Pn
| − 1| < 
Pm
if m, n ≥ N .
1.2. INFINITE PRODUCTS 1–5

It follows that if m, n ≥ N then

Log(Pn /PN ) = Log(Pm /PN ) + Log(Pn /Pm ).

In particular (taking m = n − 1),

Log(Pn /PN ) = Log(Pn−1 /PN ) + Log(1 + an ).

Hence X
Log(Pn /PN ) = Log(1 + am ).
N <m≤m

Since
Pn → ` =⇒ Log(Pn /LN ) → Log(`/PN ),
P
we conclude that n>N (1 + an ) converges to Log(`/PN ); and in particular,
P
n≥0 Log(1 + an ) is convergent. J

Proposition 1.3. Suppose an 6= −1 for n ∈ N. Then

X Y
|an | convergent =⇒ (1 + an ) convergent.

Proof I The function Log(1 + z) is holomorphic in |z| < 1, with Taylor

expansion
Log(1 + z) = z − z 2 /2 + z 3 /3 − · · · .
Thus if |z| < 1/2 then

|Log(1 + z)| ≤ |z| + |z|2 + |z|3 + · · ·

|z|
=
1 − |z|
≤ 2|z|.

|an | converges. Then an → 0; and so

P
Now suppose

|an | ≤ 1/2

for n ≥ N . It follows that

|Log(1 + an )| ≤ 2|an |

for n ≥ N . Hence X
Log(1 + an ) converges.
J
1.3. PROOF OF THE PRODUCT FORMULA 1–6

1.3 Proof of the product formula

Proposition 1.4. For <(s) > 1,
 −1
n−s = 1 − p−s
X Y
,
n∈N, n>0 primes p

in the sense that each side converges to the same value.

Proof I Let σ = <(s). Then
|n−s | = n−σ .
Thus
N N
n−s | ≤ n−σ .
X X
|
M +1 M +1
Now Z n
−σ
n ≤ x−σ dx;
n−1
and so
N Z N
−σ
x−σ dx
X
n ≤
M +1 M

1  −σ 
= M − N −σ
σ
→ 0 as M, N → ∞.
Hence n−s is convergent, by Cauchy’s criterion.
P

On the other hand,

(1 − p−s )
Y

is absolutely convergent, since

|p−s | = p−σ ≤ n−σ ,
X X X

(1 − p−s ) is convergent, by Propo-

Q
which we just saw is convergent. Hence
sition 1.3; and so therefore is
Y −1
1 − p−s .
To see that the two sides are equal note that
0
−s −1
Y  
−s
χ(n)n−s ,
X X
1 − χ(p)p = χ(n)n +
p≤N n≤N

where the second sum on the right extends over those n > N all of whose
prime factors are ≤ N .
As N → ∞, the right-hand side → n−s , since this sum is absolutely
P

convergent; while by definition, the left-hand side → (1 − p−s )−1 . We

Q

conclude that the two sides converge to the same value. J

1.4. EULER’S THEOREM 1–7

1.4 Euler’s Theorem

Proposition 1.5. (Euler’s Theorem)
X 1
= ∞.
primes p p
P
Proof I Suppose 1/p is convergent. Then
!
Y 1
1−
p
is absolutely convergent, and so converges to ` say, by Proposition ?? It
follows that !
1
→ `−1 .
Y
1−
p≤N p
But
N
!
X 1 Y 1
≤ 1− ,
1 n p≤N p
since each n on the left is expressible in the form

n = pe11 · · · perr

with p1 , . . . , pr ≤ N .
P
Hence 1/n is convergent. But
1 Z n+1 dx
> .
n n x
Thus
N Z N +1
X
−1 dx
n ≥ = log(N + 1).
1 1 x
Since log N → ∞ as N → ∞ it follows that 1/n is divergent.
P

Our hypothesis is therefore untenable, and

X 1
diverges.
p
J
P1
This is a remarkably accurate result; p
only just diverges. For it follows
from the Prime Number Theorem,
x
π(x) ∼ ,
log x
1.4. EULER’S THEOREM 1–8

that if pn denotes the nth prime (so that p2 = 3, p5 = 11, etc) then
pn ∼ n log n.
To see that, note that π(pn ) = n (ie the number of primes ≤ pn is n).
Thus setting x = pn in the Prime Number Theorem,
pn
n∼ ,
log pn
ie
pn
→ 1.
n log pn
Taking logarithms,
log pn − log n − log log pn → 0;

hence
log n
→ 1,
log pn
ie

log pn ∼ log n.
We conclude that
pn ∼ n log pn ∼ n log n.
P P
Returning to Euler’s Theorem, we see that 1/p behaves like 1/n log n.
The latter diverges, but only just, as we see by comparison with
Z
dx
= log log x.
x log x
On the other hand,
X 1
p p log p
converges for any  > 0, since
X 1
n n log1+ n
converges by comparison with
Z
dx 1
1+ = − log− x.
x log x 
What is perhaps surprising is that it is so difficult to pass from Euler’s
Theorem to the Prime Number Theorem.
Chapter 2

Dirichlet series

2.1 Definition
Definition 2.1. A Dirichlet series is a series of the form

a1 1−s + a2 2−s + a3 3−s + · · · ,

where ai ∈ C.

Remarks. 1. For n ∈ N we set

n−s = e−s log n ,

taking the usual real-valued logarithm. Thus n−s is uniquely defined

for all s ∈ C. Moreover,
0 0
m−s n−s = (mn)−s , n−s n−s = n−(s+s ) ;

while
1−s = 1
for all s.

2. The use of −s rather than s is simply a matter of tradition. The series

may of course equally well be written
a2 a3
a1 + + + ··· .
2s 3s

3. The term ‘Dirichlet series’ is often applied to the more general series

a0 λ−s −s −s
0 + a1 λ 1 + a2 λ 2 + · · · ,

2–1
2.2. CONVERGENCE 2–2

where
0 < λ0 < λ1 < λ2 < · · · ,
and
λ−s = e−s log λ

Such series often appear in mathematical physics, where the λi might

be, for example, the eigenvalues of an elliptic operator. However, we
shall only make use of Dirichlet series of the more restricted type de-
scribed in the definition above; and we shall always use the term in
that sense, referring to the more general series (if at all) as generalised
Dirichlet series.

4. It is perhaps worth noting that generalised Dirichlet series include

power series
cn xn ,
X
f (x) =
in the sense that if we make the substitution x = e−s then

f (e−s ) = cn e−ns = cn (en )−s .

X X

2.2 Convergence
Proposition 2.1. Suppose

f (s) = a1 1−s + a2 2−s + · · ·

converges for s = s0 . Then it converges for all s with

<(s) > <(s0 ).

Proof I We use a technique that might be called ‘summation by parts’, by

analogy with integration by parts.
Lemma 1. Suppose an , bn (n ∈ N) are two sequences. Let
X X
An = am , Bn = bm .
m≤n m≤n

Then
N
X N
X
an Bn = AN BN +1 − AM −1 BM − An bn+1 .
M M
2.2. CONVERGENCE 2–3

Proof I Substituting an = An − An−1 ,

N
X N
X
an Bn = (An − An−1 )Bn
M M
N
X
= An (Bn − Bn+1 ) + AN BN +1 − AM −1 BM
M
N
X
=− An bn+1 + AN BN +1 − AM −1 BM .
M

J
P P
Lemma 2. Suppose an converges and bn converges absolutely. Then
X
an Bn

converges.
Proof I By the previous Lemma,
N
X N
X
an Bn = AN BN +1 − AM −1 BM − An bn+1
M M
N
X
= AN (BN +1 − BM ) + (AN − AM −1 )BM − An bn+1 .
M
P P
Let an = A, bn = B. The partial sums of both series must be
bounded; say
|An | ≤ C, |Bn | ≤ D.
Then
N
X N
X
| an Bn | ≤ C|BN +1 − BM | + D|AN − AM −1 | + C |bn+1 |.
M M

As M, N → ∞,
N
X
BN +1 − BM → 0, AN − AM −1 → 0, |bn+1 | → 0.
M

Hence
N
X
an Bn → 0
M

as M, N → ∞; and therefore
P
an Bn converges, by Cauchy’s criterion. J
2.2. CONVERGENCE 2–4

Let s0 = s − s0 . Then <(s0 ) > 0. We apply the last Lemma with an n−s0
0
for an , and n−s for Bn . Thus

bn = Bn − Bn−1
0 0
= n−s − (n − 1)−s
Z n
0 dx
= −s0 x−s .
n−1 x
Hence
Z n
0 dx 0
|bn | ≤ |s | |x−s |
n−1 x
Z n
0 dx
= |s0 | x−σ ,
n−1 x
where σ 0 = <(s0 ).
Summing,
N Z N
0 dx
|bn | ≤ |s0 | x−σ
X

M M −1 x
|s0 |  −σ 0 −σ 0

= (M − 1) − N .
σ0
It follows that
N
X
|bn | → 0 as M, N → ∞.
M

Thus |bn | is convergent, and so the conditions of the last Lemma are
P

fulfilled. We conclude that

0 0
an n−s0 n−s = an n−(s0 +s ) = an n−s
X X X

is convergent. J
Corollary 2.1. A Dirichlet series either
1. converges for all s,

2. diverges for all s, or

3. converges for all s to the right of a line

<(s) = σ0 ,

and diverges for all s to the left of this line.

2.3. ABSOLUTE CONVERGENCE 2–5

σ + iT X + iT

σ0 σ X

Figure 2.1: Uniform convergence

Definition 2.2. We call σ0 the abscissa of convergence, setting σ0 = −∞ if

the series always converges, and σ0 = ∞ if the series never converges.
Proposition 2.2. The function
f (s) = a1 1−s + a2 2−s + · · ·
is holomorphic in the half-plane <(s) > σ0 .
Proof I Suppose σ > σ0 . The argument in the proof of the last Proposition
actually shows that an n−s converges uniformly in any rectangle
P

{S = x + it : σ ≤ x ≤ X; −T ≤ t ≤ T }
strictly to the right of <(s) = σ0 (Fig 2.1), since
N
X |s0 | −σ0
|bn | ≤ M
M σ0
|s − s0 | −(σ−σ0 )
≤ M
σ − σ0
in this region.
Thus f (s) is holomorphic in this rectangle. We conclude that f (s) is
holomorphic in the half-plane <(s) > σ0 . J

2.3 Absolute convergence

Absolute convergence is simpler than convergence, since
|an n−s | = |an |n−σ ,
X X

where σ = <(s). Thus a Dirichlet series converges absolutely at all, or none,

of the points on the line <(s) = σ.
2.3. ABSOLUTE CONVERGENCE 2–6

Proposition 2.3. If

f (s) = a1 1−s + a2 2−s + · · ·

converges absolutely for s = s0 then it converges absolutely for all s with

<(s) ≥ <(s0 ).

Proof I This follows at once from the fact that each term

|an n−s | = |an |n−σ

is a decreasing function of σ. J

Corollary 2.2. A Dirichlet series either

1. converges absolutely for all s,

2. does not converge absolutely for any s, or

3. converges absolutely for all s to the right of a line

<(s) = σ1 ,

and does not converge absolutely for any s to the left of this line.

Definition 2.3. We call σ1 the abscissa of absolute convergence, setting

σ1 = −∞ if the series always converges absolutely, and σ1 = ∞ if the series
never converges absolutely.

Proposition 2.4. We have

σ0 ≤ σ1 ≤ σ0 + 1.

Proof I Suppose
<(s) > σ0 .
Then
f (s) = a1 1−s + a2 2−s + · · ·
is convergent. Hence
an n−s → 0 as n → ∞.
In particular, an n−s is bounded, say

|an n−s | ≤ C.
2.4. THE RIEMANN ZETA FUNCTION 2–7

But then
|an n−(s+1+) | ≤ Cn−(1+)
P −(1+)
for any  < 0. Since n converges, it follows that

f (s + 1 + )

converges absolutely. We have shown therefore that

σ > σ0 =⇒ σ + 1 +  ≥ σ1

for any  > 0, from which it follows that

σ0 + 1 ≥ σ1 .

J
Proposition 2.5. If an ≥ 0 then σ1 = σ0 .
Proof I This is immediate, since in this case

|an n−σ | = an n−σ .

X X

2.4 The Riemann zeta function

Although we have already met the function ζ(s), it may be best to give a
formal definition.
Definition 2.4. The Riemann zeta function ζ(s) is defined by the Dirichlet
series
ζ(s) = 1−s + 2−s + · · · .
Remarks. 1. We shall often refer to the Riemann zeta function ζ(s) simply
as the zeta function. This is slightly inaccurate, since the term ‘zeta
function’ is applied to a wide range of related functions. However, the
Riemann zeta function is the only such function we shall use; so it
will cause no confusion if we use the unadorned term ‘zeta function’ to
describe it.

2. For example, there is a zeta function ζk (s) corresponding to each num-

ber field k, defined by

N (a)−s ,
X
ζk (s) =
a
2.4. THE RIEMANN ZETA FUNCTION 2–8

where a runs over the ideals in k (or rather, in the ring of integers
I(k) = k ∩ Z̄), and N (a) is the number of residue classes moda.
Since the unique factorisation theorem holds for ideals, the analogue of
Euler’s product formula holds:
Y −1
ζk (s) = 1 − N (p)−s ,
p

where the product runs over all prime ideals in I(k).

This allows the Prime Number Theorem to be extended to give an
approximate formula for the number of prime ideals p in the number
field k with N (p) ≤ n.

3. In another direction, the zeta function ζE (s) of an elliptic differential

(or pseudo-differential) operator E is defined by

λ−s
X
ζE (s) = n ,

where λn (n = 0, 1, 2, . . . ) are the eigenvalues of E (necessarily positive,

if E is elliptic).
The Riemann zeta function ζ(s) can be interpreted in this sense as the
zeta function of the Laplacian operator on the circle S 1 .

Proposition 2.6. The abscissa of convergence of the Riemann zeta function

is
σ0 = 1.

Proof I This follows at once from the fact that

n−σ < ∞ ⇐⇒ σ > 1.

X

Let us recall how this is established, by comparing the sum with the
R −σ
integral x dx. If n − 1 ≤ x ≤ n,

n−σ ≤ x−σ ≤ (n − 1)−σ .

Integrating, Z n
n−σ < x−σ dx < (n − 1)−σ ,
n−1

Summing from n = M + 1 to N ,
N Z N N +1
−σ
x−σ dx < n−σ ,
X X
n <
M +1 M M
2.4. THE RIEMANN ZETA FUNCTION 2–9

It follows that n−σ and ∞ x−σ dx converge or diverge together.

P R

But we can compute the integral directly: if n = 1 then

Z Y
x−1 dx = log X − log Y,
X
and so the integral diverges; while if n 6= 1 then
Z Y
1
x−σ dx = (M 1−σ − N 1−σ ),
X σ−1
and so the integral converges if σ > 1 and diverges if σ < 1. J
Corollary 2.3. The zeta function ζ(s) is holomorphic in the half-plane
<(s) > 1.
We can continue ζ(s) analytically to the half-plane <(s) > 0 in the fol-
lowing way.
Proposition 2.7. The Dirichlet series
f (s) = 1−s − 2−s + 3−s − · · ·
has abscissa of convergence σ0 = 0, and so defines a holomorphic function
in the half-plane <(s) > 0.
Proof I Suppose σ > 0. Then
f (σ) = 1−σ − 2−σ + 3−σ − · · ·
converges, since the terms alternate in sign and decrease to 0 in absolute
value. It follows, by Proposition 2.1, that f (s) converges for <(s) > 0.
The series certainly does not converge for <(s) < 0, since the terms do
not even → 0. Thus σ0 = 0. J
The abscissa of absolute convergence σ1 of f (s) is 1 since the terms have
the same absolute value as those of ζ(s).
Proposition 2.8. If <(s) > 1 then
f (s) = (1 − 21−s )ζ(s).
Proof I If <(s) > 1 then the Dirichlet series for f (s) converves absolutely,
so we may re-arrange its terms:
f (s) = 1−s − 2−s + 3−s − · · ·
= (1−s + 2−s + 3−s + · · · ) − 2(2−s + 4−s + · · · )
= ζ(s) − 2 · 2−s (1−s + 2−s + · · · )
= ζ(s) − 2 · 2−s ζ(s)
= (1 − 21−s )ζ(s).
J
2.4. THE RIEMANN ZETA FUNCTION 2–10

Proposition 2.9. The zeta function ζ(s) extends to a meromorphic function

in <(s) > 0, with a single simple pole at s = 1 with residue 1.

Proof I We have
1
ζ(s) = f (s)
1 − 21−s
for <(s) > 1. But the right-hand side is meromorphic in <(s) > 0, and so
defines an analytic continuation of ζ(s) (necessarily unique, by the theory of
analytic continuation) to this half-plane.
Since f (s) is holomorphic in this region, any pole of ζ(s) must be a pole
of 1/(1 − 21−s ), ie a zero of 1 − 21−s . But

21−s = e(1−s) log 2 .

Hence
21−s = 1 ⇐⇒ (1 − s) log 2 = 2nπi
for some n ∈ Z. Thus 1/(1 − 21−s ) has poles at
2nπ
s=1+ i (n ∈ Z).
log 2
At first sight this seems to give an infinity of poles of ζ(s) on the line
<(s) = 1. However, the following argument shows that f (s) must vanish at
all these points except s = 1, thus ‘cancelling out’ all the poles of 1/(1−21−s )
except that at s = 1.
Consider

g(s) = 1−s + 2−s − 2 · 3−s + 4−s + 5−s − 2 · 6−s + · · · .

Like f (s), this converges for all σ > 0. For if we group g(σ) in sets of three
terms

g(σ) = (1−σ + 2−σ − 2 · 3−σ ) + (4−σ + 5−σ − 2 · 6−σ ) + · · ·

we see that each set is > 0. Thus the series either converges (to a limit > 0),
or else diverges to +∞.
On the other hand, we can equally well group g(σ) as

g(σ) = 1−σ + 2−σ − (2 · 3−σ − 4−σ − 5−σ ) − (2 · 6−σ − 7−σ − 8−σ ) + · · · .

Now each group is < 0, if we omit the terms 1−σ + 2−σ . Thus g(σ) either
converges (to a limit < 1−σ + 2−σ ), or else diverges to −∞.
We conclude the g(σ) converges (to a limit between 0 and 1−σ + 2−σ ).
2.4. THE RIEMANN ZETA FUNCTION 2–11

Hence g(s) converges for <(s) > 0.

But if <(s) > 1 we can re-write g(s) as
g(s) = (1−s + 2−s + 3−s + · · · ) − 3(3−s + 6−s + 9−s + · · · )
= (1 − 31−s )ζ(s).
Thus
1
ζ(s) = g(s).
1 − 31−s
The right hand side is meromorphic in the half-plane <(s) > 0, giving a
second analytic continuation of ζ(s) to this region, which by the theory of
analytic contination must coincide with the first.
But the poles of 1/(1 − 31−s ) occur where
(1 − s) log 3 = 2mπi,

ie
2πm
s=1+ i.
log 3
Thus ζ(s) can only have a pole where s is expressible in both forms
2πn 2πm
s=1+ i=1+ i (m, n ∈ Z).
log 2 log 3
But this implies that
m log 2 = n log 3,

ie

2m = 3n ,
which is of course impossible (by the Fundamental Theorem of Arithmetic)
unless m = n = 0.
We have therefore eliminated all the poles except s = 1. At s = 1,
1 1 1
f (1) = 1 − + − + · · · = log 2.
2 3 4
(This follows on letting x → 1 from below in log(1 + x) = x − x2 /2 + · · · .)
On the other hand, if we set s = 1 + s0 then
0
1 − 21−s = 1 − e−s log 2

2
= s0 log 2 − s0 /2! log2 2 + · · ·
= s0 log 2(1 − s0 /2 log 2 + · · ·
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–12

Thus
1 1
=
1−2 1−s 1 − 2−s0
1 s0
= (1 + log 2 + · · · )
log 2 s0 2
1
= + h(s),
log 2 s0

where h(s) is holomorphic. Hence

1
ζ(1 + s0 ) = + h(s)f (s).
s0
We conclude that ζ(s) has a simple pole at s = 1 with residue 1. J
In Chapter 7 we shall see that ζ(s) can in fact be analytically continued
to the whole of C. It has no further poles; its only pole is at s = 1.

2.5 The Riemann-Stieltjes integral

It is helpful (although by no means essential) to introduce a technique which
allows us to express sums as integrals, and brings ‘summation by parts’ into
the more familiar guise of integration by parts.
Let us recall the definition of the Riemann integral ab f (x) dx of a contin-
R

uous function f (x) on [a, b]. Note that f (x) is in fact uniformly continuous
on [a, b], ie given  > 0 there exists a δ > 0 such that

|x − y| < δ =⇒ |f (x) − f (y)| < 

for x, y ∈ [a, b].

By a dissection ∆ of [a, b] we mean a sequence

∆ : a = x0 < x1 < · · · < xn = b.

We set
k∆k = max |xi+1 − xi |.
0≤i<n

The dissection ∆ is said to be a refinement of ∆, and we write ∆0 ⊂ ∆ if

0

the set of dissection-points xi of ∆ is a subset of the set of dissection-points

of ∆0 . Evidently
∆0 ⊂ ∆ =⇒ k∆0 k ≤ k∆k.
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–13

Let X
S(f, ∆) = f (xi )(xi+1 − xi ).
0≤i<n

Then S(f, ∆) is convergent as k∆k → 0, ie given  > 0 there exists δ > 0

such that
k∆1 k, k∆2 k < δ =⇒ |S(f, ∆1 ) − S(f, ∆2 )| < .
This follows from 2 lemmas (each more or less immediate).
1. Suppose
|x − y| < δ =⇒ |f (x) − f (y)| < .
Then

k∆1 k < δ, ∆2 ⊂ ∆ =⇒ |S(f, ∆1 ) − S(f, ∆2 )| < (b − a).

2. Given 2 dissections ∆1 , ∆2 of [a, b] we can always find a common re-

finement ∆3 , ie
∆3 ⊂ ∆1 , ∆3 ⊂ ∆2 .

These in turn imply

3.
k∆1 k, k∆2 k < δ =⇒ |S(f, ∆1 ) − S(f, ∆2 )| < 2(b − a).

Thus, by Cauchy’s criterion, S(f, ∆) converges as ∆ → 0, ie there exists an

I ∈ C such that ie given  > 0 there exists δ > 0 such that

|S(f, ∆) − I| <  if k∆k < δ.

Even if f (x) is not continuous, we say that it is Riemann-integrable on

[a, b] with
Z b
f (x) dx = I
a
if
S(f, ∆) → I as k∆k → 0.
Now suppose M (x) is an increasing (but not necessarily strictly increas-
ing) function on [a, b], ie

x ≤ y =⇒ f (x) ≤ f (y).

Then we set
X
SM (f, ∆) = f (xi )(M (xi+1 ) − M (xi )).
0≤i<n
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–14

Proposition 2.10. If f (x) is continuous and M (x) is increasing then

SM (f, ∆) converges as k∆k → 0.

Proof I The result follows in exactly the same way as for the Riemann in-
tegral above, with (1) replaced by

1’. Given  > 0, suppose δ > 0 is such that

|x − y| < δ =⇒ |f (x) − f (y)| < .

Then if k∆1 k, k∆2 k < δ,

|S(f, ∆1 ) − S(f, ∆2 )| < (M (b) − M (a)).

J
Definition 2.5. We call

I = lim SM (f, ∆)
k∆k→0

the Riemann-Stieltjes integral of f (x) with respect to M (x), and write

Z b
f (x) dM = I.
a

2.5.1 Functions of bounded variation

Definition 2.6. A (real- or complex-valued) function f (x) is said to be of
bounded variation on the interval [a, b] if there exists a constant C such that
X
A(f, ∆) = |f (xi ) − f (xi−1 )| ≤ C

for all dissections ∆ of [a, b].

Proposition 2.11. Any linear combination

f (x) = µ1 f1 (x) + · · · + µr fr (x) (µ1 , . . . , µr ∈ C)

of functions f1 (x), . . . , fr (x) of bounded variation is itself of bounded varia-

tion.
Proof I This follows at once from the fact that

A(f, ∆) ≤ |µ1 |A(f1 , ∆) + · · · + |µr |A(fr , ∆).

J
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–15

Proposition 2.12. Any monotone increasing or decreasing function f (x) is

of bounded variation.
Proof I If f (x) is increasing then

|f (xi ) − f (xi−1 | = f (xi ) − f (xi−1 ;

and so
A(f, ∆) = f (b) − f (a).
If f (x) is decreasing then −f (x) is increasing, so the result follows from the
last Proposition. J
Proposition 2.13. A function f (x) of class C 1 [a, b], ie with continuous
derivative f 0 (x) on [a, b], is of bounded variation.
Proof I Since f 0 (x) is continuous, it is bounded: say

|f 0 (x)| ≤ C.

Also, by the Mean Value Theorem,

f (xi ) − f (xi−1 = (xi − xi−1 )f 0 (ξ),

where xi−1 < ξ < xi . Hence

|f (xi ) − f (xi−1 | ≤ C(xi − xi−1 );

and so
A(f, ∆) ≤ C(b − a).
J
Proposition 2.14. A real-valued function f (x) is of bounded variation on
[a, b] if and only if it can be expressed as the difference of two increasing
functions:
f (x) = M (x) − N (x),
where M (x), N (x) are monotone increasing.
Proof I If f (x) is expressible in this form then it is of bounded variation,
by Propositions 2.12 and 2.11.
For the converse, let
X
P (f, ∆) = (f (xi ) − f (xi−1 )),
i:f (xi )≥f (xi−1 )
X
N (f, ∆) = − (f (xi ) − f (xi−1 )).
i:f (xi )<f (xi−1 )
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–16

for each dissection ∆ of [a, b]. Then P (f, ∆), N (f, ∆) ≥ 0; and

P (f, ∆) − N (f, ∆) = f (b) − f (a), P (f, ∆) + N (f, ∆) = A(f, ∆).

It follows that
0 ≤ P (f, ∆), N (f, ∆) ≤ A(f, ∆).
Hence
P (f ) = sup P (f, ∆), N (f ) = sup N (f, ∆)
∆ ∆

are defined.
Lemma 3. We have

P (f ) − N (f ) = f (b) − f (a).

Proof I Given  > 0 we can find dissections ∆1 , ∆2 such that

P (f ) ≥ P (f, ∆1 ) > P (f ) − ,
N (f ) ≥ N (f, ∆2 ) > N (f ) − .

If now ∆ is a common refinement of ∆1 , ∆2 then

P (f ) ≥ P (f, ∆) ≥ P (f, ∆1 ) > P (f ) − ,

N (f ) ≥ N (f, ∆) ≥ N (f, ∆2 ) > N (f ) − .

But
P (f, ∆) − N (f, ∆) = f (b) − f (a).
It follows that

P (f ) − N (f ) −  ≤ f (b) − f (a) ≤ P (f ) − N (f ) + .

Since this is true for all  > 0,

P (f ) − N (f ) = f (b) − f (a).

J
Now suppose a ≤ x ≤ b. We apply the argument above to the interval
[a, x]. Let p(x), n(x) be the functions P (f ), N (f ) for the interval [a, x]. By
the last Lemma,
p(x) − n(x) = f (x) − f (a).
It is easy to see that p(x), n(x) are increasing functions of x. For suppose

a ≤ x < y ≤ b.
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–17

To each dissection
∆ : a = x0 < x 1 < · · · xn = x
of [a, x] we can associate the dissection

∆0 : a = x0 < x1 < · · · xn < xn+1 = y

of [a, y]; and then

P (f, ∆0 ) ≥ P (f, ∆), N (f, ∆0 ) ≥ N (f, ∆).

It follows that
p(y) ≥ p(x), n(y) ≥ n(x),
ie p(x), n(x) are monotone increasing. Since

f (x) = (f (a) + p(x)) − n(x),

this establishes the result. J

Proposition 2.15. The function f (x) is of bounded variation on [a, b] if and

only if it can be expressed as a linear combination of increasing functions:

f (x) = µ1 M1 (x) + · · · + µr Mr (x),

where M1 (x), . . . , Mr (x) are monotone increasing, and µ1 , . . . , µr ∈ C.

Proof I It follows from Propositions 2.11 and 2.12 that a function of this
form is of bounded variation.
For the converse, note that if f (x) is complex-valued then it can be split
into its real and imaginary parts:

f (x) = fR (x) + ifI (x)

where fR (x), fI (x) are real-valued functions. It is easy to see that if f (x) is
of bounded variation then so are fR (x) and fI (x). Hence each is expressible
as a difference of increasing functions, say

fR (x) = MR (x) − NR (x), fI (x) = MI (x) − NI (x).

But then
f (x) = MR (x) − NR (x) + iMI (x) − iNI (x),
which is of the required form. J
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–18

This result allows us to extend the Riemann-Stieltjes integral to functions

of bounded variation.
Suppose U (x) is a function of bounded variation on [a, b]. We set
X
SU (f, ∆) = f (xi )(U (xi+1 ) − U (xi ))

for any dissection ∆ of [a, b].

Proposition 2.16. If f (x) is continuous and U (x) is of bounded variation

then
SU (f, ∆) converges as k∆k → 0.

Proof I By the last Proposition, we can express U (x) as a linear combination

of increasing functions. The result then follows from Proposition 2.10. J

Definition 2.7. We call

I = lim SU (f, ∆)
k∆k→0

the Riemann-Stieltjes integral of f (x) with respect to U (x), and write

Z b
f (x) dU = I.
a

We extend the Riemann-Stieltjes integral to non-continuous functions

f (x) as we do the familiar Riemann integral. Thus if

SU (f, ∆) → I as k∆k → 0

then we say that f (x) is Riemann-Stieltjes integrable on [a, b], with

Z b
f (x) dU = I.
a

Similarly, we extend the Riemann-Stieltjes integral to infinite ranges in

the same was as the Riemann integral. Thus we set
Z ∞ Z X
f (x)dU = lim f (x)dU,
a X→∞ a

if the limit exists.

In one important case the Riemann-Stieltjes integral reduces to the fa-
miliar Riemann integral.
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–19

Proposition 2.17. Suppose U (x) is of class C 1 [a, b], ie U (x) has continuous
derivative U 0 (x) on [a, b]; and suppose f (x)U 0 (x) is Riemann integrable on
[a, b]. Then f (x) is Riemann-Stieltjes integrable, and
Z b Z b
f (x)dU = f (x)U 0 (x)dx.
a a

Proof I Suppose ∆ is a dissection of [a, b]. We compare SU (f, ∆) with

S(f U 0 , ∆).
By the Mean Value Theorem,

U (xi+1 ) − U (xi ) = U 0 (ξi )

where xi < ξi < xi+1 . Moreover, since U 0 (x) is continuous on [a, b], it is
absolutely continuous; so given any  > 0 we can find δ > 0 such that

|U 0 (xi ) − U 0 (ξi )| < 

if xi+1 − xi < δ.
Hence
|SU (f, ∆) − S(f U 0 , ∆)| ≤ max|f |(b − a)
if k∆k < δ, from which the result follows. J

2.5.2 Discontinuities
Proposition 2.18. If f (x) is a function of bounded variation on [a, b] then
the left limit
f (x − 0) = lim f (t)
t→x−0

exists for all x ∈ [a, b); and the right limit

f (x + 0) = lim f (t)
t→x+0

exists for all x ∈ (a, b].

Proof I The result is (almost) immediate if f (x) is increasing. It follows for

any function f (x) of bounded variation by Proposition 2.15, since

f (x) = µ1 M1 (x)+· · ·+µr Mr (x) for all x =⇒ f (x−0) = µ1 M1 (x−0)+· · ·+µr Mr (x−0),

and similarly for the right limit. J

2.5. THE RIEMANN-STIELTJES INTEGRAL 2–20

The function f (x) is continuous at x = ξ if

f (ξ − 0) = f (ξ) = f (ξ + 0).

Otherwise f (x) has a discontinuity at ξ.

Proposition 2.19. The discontinuities of a function f (x) of bounded vari-

ation are enumerable.

Proof I It is sufficient to prove the result for an increasing function; for if

f (x) = µ1 M1 (x) + · · · + µr Mr (x)

then the discontinuities of f (x) lie in the union of the discontinuities of

M1 (x), . . . , Mr (x); and a finite union of enumerable sets is enumerable.
Let us define the ‘jump’ at a discontinuity ξ to be

j(ξ) = f (ξ + 0) − f (ξ − 0).

Note that for an increasing function

f (ξ − 0) ≤ f (ξ) ≤ f (ξ + 0).

Thus f (x) is discontinuous at ξ if and only if j(ξ) > 0.

Lemma 4. Suppose M (x) is increasing on [a, b]; and suppose

a ≤ ξ1 < ξ2 < · · · < ξn ≤ b.

Then X
j(ξi ) ≤ f (b) − f (a).
1≤i≤n

Proof I Choose a dissection x0 , x1 , . . . , xn of [a, b] with

a = x0 ≤ ξ1 < x1 < ξ2 < x2 < · · · < xn−1 < ξn ≤ xn = b.

Then it is easy to see that

f (xi ) − f (xi−1 ) ≥ j(ξi );

and so, on addition, X

f (b) − f (a) ≥ j(ξi ).
J
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–21

Corollary 2.4. Suppose M (x) is increasing on [a, b]; Then the number of
discontinuities with

j(x) = f (x + 0) − f (x − 0) ≥ 2−r

is
≤ 2r (f (b) − f (a)).
Using the Lemma, we can enumerate the discontinuities of M (x) by first
listing those with j(x) ≥ 1, then those with 1 > j(x) ≥ 2−1 , then those with
2−1 > j(x) ≥ 2−2 , and so on. In this way we enumerate all the discontinuities:

ξ0 , ξ1 , ξ2 , . . . .

J
Remarks. 1. Note that we are not claiming that the discontinuities can
be enumerated in increasing order, ie so that

ξ0 < ξ1 < ξ2 < · · · .

That is not so, in general; f (x) could , for example, have a discontinuity
at every rational point.

2. The discontinuity at ξ can be divided into two parts:

f (ξ) − f (ξ − 0) and f (ξ + 0) − f (ξ).

However, if f (x) is right-continuous, ie

f (x + 0) = f (x)

for all x ∈ [a, b), then the second contribution vanishes, and the dis-
continuity is completely determined by

j(ξ) = f (ξ + 0) − f (ξ − 0) = f (ξ) − f (ξ − 0).

To simplify the discussion, the functions we use have all been chosen
to be right-continuous; for example, we set

π(x) = k{p : p ≤ x}k,

although we could equally well have taken the left-continuous function

π1 (x) = k{p : p < x}k.

2.5. THE RIEMANN-STIELTJES INTEGRAL 2–22

(From a theoretical point of view, it might have been preferable to have

imposed the symmetric condition
1
f (x) = (f (x + 0) + f (x − 0)) .
2
However, for our purposes the added complication would outweigh the
theoretical advantage.)

Definition 2.8. The step function Hξ (x) is defined by


0 if x < ξ,
Hξ (x) =
1 if x ≥ ξ.

Proposition 2.20. Suppose U (x) is a right-continuous function of bounded

variation on [a, b]. Then X
j(ξ)
ξ

is absolutely convergent.

Proof I It is sufficient to prove the result when U (x) is increasing, by Propo-

sition 2.15. But in that case j(ξ) > 0, and
X
j(ξ) ≤ f (b) − f (a),
ξ

by Lemma 4. J

Proposition 2.21. Suppose U (x) is a right-continuous function of bounded

variation on [a, b]. Then U (x) can be split into two parts,

U (x) = J(x) + f (x),

where f (x) is continuous, and

X
J(x) = j(ξ)Hξ (x),

the sum extending over all discontinuities ξ of f (x) in [a, b].

Proof I It is sufficient to prove the result in the case where U (x) is increas-
ing, by Proposition 2.15.
Let
f (x) = U (x) − J(x).
We have to show that f (x) is continuous.
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–23

The step function Hξ (x) is right-continuous. Hence J(x) is right-continuous;

and since U (x) is right-continuous by hypothesis, it follows that f (x) is right-
continuous. We have to show that f (x) is also left-continuous.
Suppose x < y. Then
X
J(y) − J(x) = j(ξ)
x<ξ≤y

≤ U (y) − U (x),
by Proposition 4. Thus
f (x) = U (x) − J(x) ≤ U (y) − J(y) = f (y),
ie f (x) is increasing.
Moreover,
0 ≤ f (y) − f (x) ≤ U (y) − U (x).
Hence
0 ≤ f (y) − f (y − 0) ≤ U (y) − U (y − 0).
In particular, if U (x) is left-continuous at y then so is f (x).
Now suppose U (x) has a discontinuity at y. If x < y then
J(y) − J(x) ≥ j(y) = U (y) − U (y − 0).
Hence
J(y) − J(y − 0) ≥ U (y) − U (y − 0),

ie

f (y − 0) = U (y − 0) − J(y − 0) ≥ f (y) = U (y) − J(y).

Since f (x) is increasing, it follows that
f (y − 0) = f (y),
ie f (x) is left-continuous at y. J
Definition 2.9. We call f (x) the continuous part of U (x), and J(x) the
purely discontinous part.
Remarks. 1. This is our own terminology; there do not seem to be stan-
dard terms for these two parts of a function of bounded variation. That
is probably because they are more generally studied through the mea-
sure or distribution dU , with the step function Hξ (x) replaced by the
Dirac delta ‘function’ δξ (x) = dHξ .
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–24

2. Our definition of J(x) entails that J(a) = 0. With that condition, the
splitting of U (x) is unique. If we drop the condition then J(x) and
f (x) are defined up to a constant.

Proposition 2.22. Suppose

X
U (x) = j(ξ)Hξ (x)

is a purely discontinuous (but right-continuous) function of bounded variation

on [a, b]; and suppose f (x) is continuous on [a, b]. Then
Z b X
f (x)dU = j(ξ)f (ξ).
a
P
Proof I Since j(ξ) is absolutely convergent, it is sufficient to prove the
result for a single step function Hξ (x).
Suppose ∆ is a dissection of [a, b]; and suppose

xi < ξ ≤ xi+1 .

Then

SHξ (f, ∆) = f (xi )(Hξ (xi+1 ) − Hξ (xi ))

= f (xi ).

Since
f (xi ) → f (ξ) as k∆k → 0,
the result follows. J
Rb
In practice we shall encounter the Riemann-Stieltjes integral a f (x)dU
in just two cases: the case above, where f (x) is continuous and U (x) is
purely discontinuous; and the case where U (x) ∈ C 1 [a, b], when (as we saw
in Proposition 2.17)
Z Z
f (x)dU = f (x)U 0 (x)dx.

2.5.3 Integration by parts

Proposition 2.23. Suppose U (x), V (x) are of bounded variation on [a, b];
and suppose either U (x) or V (x) is continuous. Then
Z b Z b
U (x) dV + V (x) dU = U (b)V (b) − U (a)V (a).
a a
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–25

R
Proof I We may suppose that U (x) R
is continuous. Then U (x)dV is cer-
tainly defined; we must show that V (x)dU is also defined.
Let
∆ : a = x0 < x 1 < · · · < x n = b
be a dissection of [a, b].
Our proof is based on the formula for ‘summation by parts’ (Lemma 1),
which we may re-write as
n−1
X n−1
X
Ai b i + ai+1 Bi = An Bn−1 − A1 B0 .
i=1 i=1

Adding An bn + a1 B0 to each side, this becomes

n
X n
X
Ai b i + ai Bi−1 = An Bn − A0 B0 .
i=1 i=1

Now let us substitute

Ai = U (xi ), Bi = V (xi ).

The first sum becomes

n−1
X n−1
X
Ai b i = U (xi ) (V (xi ) − V (xi−1 )) .
i=1 i=1

This is almost SU (V, ∆). There is a discrepancy because we are taking the
value U (xi ) at the top of the interval [xi−1 , xi ] rather than at the bottom.
However, U (x) is continuous, and so uniformly continuous, on [a, b]. Thus
given  > 0 we can find δ > 0 such that

|U (xi ) − U (xi−1 | < 

if k∆k < δ. It follows that

X X
| Ai bi − SU (V, ∆)| ≤  |V (xi ) − V (xi−1 )|
= A(V, ∆).

Since V (x) is of bounded variation, there is a constant C > 0 such that

A(V, ∆) ≤ C

for all dissections ∆ of [a, b]. Thus

X
| Ai bi − SU (V, ∆)| ≤ C.
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–26

Turning to the second term,

n
X n
X
ai Bi−1 = (U (xi ) − U (xi−1 )) V (xi−1 )
i=1 i=1
= SU (V, ∆).

Now we know that

Z b
SV (U, ∆) → U (x)dV as k∆k → 0.
a
R
It follows that V dU is also defined, ie V (x) is Riemann-Stieltjes integrable
with respect to U (x) over [a, b], and
Z b Z b
U (x)dV + V (x)dU = U (b)V (b) − U (a)V (a)
a a
= [U (x)V (x)]ba .

2.5.4 The abscissa of convergence revisited

To see how the Riemann-Stieltjes integral can be used, we look again at the
proof of Proposition 2.1. Let

an n−s .
X
f (s) =

We have to show that

f (s) convergent =⇒ f (s + s0 ) convergent

if <(s0 ) > 0.
Let
an n−s .
X
V (x) =
n≤x

Then V (x) has discontinuities at each integer point x = n, with j(n) = n−s .
Thus by Proposition 2.22
N Z N
0) 0
a−(s+s x−s dV
X
n =
M +1 M
Z N
= U (x)dV,
M
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–27

where
0
U (x) = x−s .
Integrating by parts (by Proposition 2.23),
N Z N
0
a−(s+s )
= [U (x)V (x)]N
X
n M − V (x)dU
M +1 M
Z N
= [U (x)V (x)]N
M − V (x)U 0 (x)dx
M
Z N
0 dx
= [U (x)V (x)]N
M −s 0
x−s V (x) ,
M x
0
since U (x) has continuous derivative s0 x−(s +1) .
Since f (s) is convergent, V (x) is bounded, say

V (x) ≤ V.

Thus if σ 0 = <(s0 ),
N Z N
0)

−σ 0 −σ 0
 0 dx
a−(s+s 0
x−σ
X
| n | ≤V M +N + |s |
M +1 M x
0
 0 0
 V |s |  −σ0 −σ 0

= V M −σ + N −σ + M − N
σ0
|s0 | −σ0
!
0 0
≤V M −σ + N −σ + 0M
σ
→ 0 as M, N → ∞.

We conclude that f (s + s0 ) is convergent if σ 0 = <(s0 ) > 0.

2.5.5 Analytically continuing ζ(s): an alternative ap-

proach
As another application of the Riemann-Stieltjes integral, we give an alterna-
tive method of extending ζ(s).
Let
G(x) = [x].
(This function is sometimes called the Gauss function.)
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–28

Suppose <(s) > 1. Then

∞
n−s
X
ζ(s) =
1
Z ∞
= x−s dG
0
Z ∞
h
−s
i∞ dx
= x G(x) +s x−s G(x)
0 0 x
Z ∞
dx
=s x−s G(x) .
1 x
(Note that G(x) = 0 for x ∈ [0, 1); so there is no convergence problem at
x = 0.)
We can write
x = G(x) + F (x),
where F (x) is the ‘fractional part’ of x. Thus

0 ≤ F (x) ≤ 1.

Now
Z ∞ Z ∞
dx
ζ(s) = s x−s dx − s x−s F (x)
"1 1 x
1−s ∞
# Z ∞
x dx
=s −s x−s F (x)
1−s 1 1 x
Z ∞
s dx
= −s x−s F (x) .
s−1 1 x
But the integral on the right converges if <(s) > 0, since
Z Y Z Y
−s dx dx
| x F (x) | ≤ x−σ
X x X x
1  −σ 
= X − Y −σ
σ
→ 0 as X, Y → ∞.

Thus Z ∞
s dx
ζ(s) = + x−s F (x)
s−1 1 x
gives an analytic continuation of ζ(s) to <(s) > 0.
Moreover, since the integral is holomorphic in this region, we see that ζ(s)
has a single simple pole at s = 1 with residue 1 in the half-plane <(s) > 0.
2.5. THE RIEMANN-STIELTJES INTEGRAL 2–29

We can even extend ζ(s) further, to the half-plane <(s) > −1, if we take
a little care. (In Chapter 7 we shall show by an entirely different method
that ζ(s) can be continued analytically to the whole complex place C; so the
present exercise is just that — an exercise.)
Let
1
h(x) = F (x) − .
2
Then Z n+1
h(x)dx = 0.
n
Hence Z x
H(x) = h(t)dt
0
is bounded; in fact
1
|H(x)| ≤ .
4
Suppose <(s) > 0. Integrating by parts (in the usual sense),
Z ∞
dx 1 Z ∞ −s dx Z ∞ −s dx
F (x) = x + x h(x)
1 x 2 "1 # x 1 x
−s ∞ Z ∞
1 x h
−s−1
i ∞
= + x H(x) + x−s−2 H(x)dx
2 −s 1 1 1
Z ∞
1
= + x−s−2 H(x)dx.
2s 1

Thus Z ∞
s 1
ζ(s) = − −s x−s−2 H(x)dx.
s−1 2 1

But the integral on the right converges if <(s) > −1, since
Z Y
−s−2 1 Z Y −σ−2
| x H(x)dx| ≤ x dx
X 4 X
1  
= X −(σ+1) − Y −(σ+1)
4(σ + 1)
→ 0 as X, Y → ∞.

Thus Z ∞
s 1
ζ(s) = − −s x−(s+2) H(x)dx
s−1 2 1

gives an analytic continuation of ζ(s) to <(s) > −1.

2.6. THE RELATION BETWEEN AN AND σ0 2–30

2.6 The relation between An and σ0

Power series are simpler than Dirichlet series, in that the radius of conver-
gence of a power series
cn xn
X

is equal to the radius of absolute convergence, both being given by

r = lim sup|cn |1/n .

We must expect the corresponding result for a Dirichlet series

an n−s
X

to involve the partial sums X

An = an
m≤n

rather than the coefficients an themselves.

Proposition 2.24. Suppose

an n−s
X
f (s) =

has abscissa of convergence σ0 . Then

An = o(nσ )

for any σ > σ0 .

Conversely, if
An = O(nσ )
then σ ≥ σ0 .
Proof I Suppose σ > σ0 . Choose σ 0 with

σ > σ 0 > σ0 .

Then
0
f (σ 0 ) = an n−σ
X

0
is convergent. Hence an n−σ is bounded, say
0
|an n−σ | ≤ C.

Then
0
|an | ≤ Cnσ
2.6. THE RELATION BETWEEN AN AND σ0 2–31

ie
0
an = O(nσ ) = o(nσ ).
Conversely, suppose
An = O(nσ ),
say
|An | ≤ Cnσ ;
and suppose
σ 0 > σ.
Let X
A(x) = an .
n≤x

Then
|A(x)| = |A([x])| ≤ C[x]σ ≤ Cxσ .
Integrating by parts,
N Z N
−σ 0 0
x−σ dA
X
an n =
M +1 M
Z N
h 0
iN 0 dx
= x−σ A(x) + σ0 x−σ A(x) .
M M x
Hence
N Z N
−σ 0 σ−σ 0 σ−σ 0 0 dx
xσ−σ
X
| an n | ≤ C(M +N ) + Cσ
M +1 M x
σ
 
0
≤C 2+ N σ−σ .
σ0 −σ
Thus
N
0
an n−σ | → 0 as M, N → ∞.
X
|
M +1

Hence, by Cauchy’s criterion,

0
an n−σ
X

is convergent; and so
σ 0 ≤ σ0 .
Since this holds for all σ 0 > σ,
σ ≤ σ0 .
J
2.7. DIRICHLET SERIES WITH POSITIVE TERMS 2–32

2.7 Dirichlet series with positive terms

If the Dirichlet series
an n−s
X
f (s) =
has abscissa of convergence σ0 then f (s) is holomorphic in the half-plane
<(s) > σ0 . But the converse is not in general true, ie we may be able to
continue f (s) analytically to a function f (s) holomorphic in the half-plane
<(s) > σ 0 , where σ 0 < σ0 .
For example, the abscissa of convergence of
 
1 − 21−s ζ(s) = 1−s − 2−s + 3−s − 4−s + · · ·

is σ0 = 0. (The terms in the series do not even → 0 for <(s) < 0.) But as we
shall see, this series extends analytically to an entire function, ie a function
holomorphic in the whole of C.
However, the following Proposition shows that if the coefficients an of the
Dirichlet series are positive then the converse does hold — f (s) cannot be
extended holomorphically across the line <(s) = σ0 .
Proposition 2.25. Suppose the Dirichlet series

an n−s
X
f (s) =

has abscissa of convergence σ0 ; and suppose an ≥ 0 for all n. If f (s) can be

extended to a function meromorphic in an open set containing s = σ0 then
f (s) must have a pole at s = σ0 .
Proof I Suppose f (s) is holomorphic in

D0 = D(σ0 , δ) = {z ∈ C : |z − σ0 | < δ}

Let
δ
σ = σ0 + .
4
Then f (s) is holomorphic in
3δ 3δ
D1 = D(σ, ) = {z ∈ C : |z − σ| < } ⊂ D0 .
4 4
It follows by Taylor’s theorem that
1 00
f (s) = f (σ) + f 0 (σ)(s − σ) + f (σ)(s − σ)2
2!
for s ∈ D1 .
2.7. DIRICHLET SERIES WITH POSITIVE TERMS 2–33

D0

D1

D2
σ0 σ0 σ

Figure 2.2: Convergence of Dirichlet series with positive terms

Now
an n−s
X
f (s) =
near s = σ (since this point is in the half-plane of convergence). Moreover
this series converges uniformly and absolutely for s sufficiently close to σ,
say inside
δ
D2 = D(σ, .
8
It follows that we can differentiate term-by-term, as often as we like:

f 0 (s) = − an log n n−s ,

X

f 00 (s) = an log2 n n−s ,

X

f 000 (s) = − an log3 n n−s ,

X

etc. In particular
f (k) (σ) = (−1)k an logk nn−σ ,
X

where f (k) denotes the kth derivative of f (s).

Now let us apply Taylor’s expansion to compute f (σ 0 ), where

δ
σ0 = σ − .
4
2.7. DIRICHLET SERIES WITH POSITIVE TERMS 2–34

We have
1 (k)
f (σ 0 ) = f (σ)(σ 0 − σ)k
X

k k!
!k
1 δ
(−1) f (k) (σ)
k
X
= .
k k! 4

Substituting from above for the f (k) ,

!k
0 1 δ
an logk nn−σ .
X X
f (σ ) =
k k! 4 n

Since all the terms on the right are positive (the two factors (−1)k cancelling
out), the double series is absolutely convergent, and we can invert the order
of the summations:
!k
0 −σ 1 δ
logk n
X X
f (σ ) = an n
n k k! 4

The series on the right may seem complicated, but common-sense tells
us what the sum must be. We could have carried out the whole operation
entirely within the half-plane of convergence, in which case we know that
0
f (σ 0 ) = an n−σ .
X

Clearly this must still be true.

In fact,
!k !k
1 δ 1 δ log n
logk n
X X
=
k k! 4 k k! 4
δ log n/4
=e
= nδ/4
0
= nσ−σ ,

and so
0
f (σ 0 ) = an n−σ nσ−σ
X

n
0
an n−σ .
X
=
n

Thus f (σ 0 ) converges, which is impossible since σ 0 < σ0 . We conclude

that our original assumption is untenable: f (s) cannot be holomorphic in a
neighbourhood of s = σ0 . J
Chapter 3

The Prime Number Theorem

3.1 Statement of the theorem

The Prime Number Theorem asserts that
x
π(x) ∼ .
log x
It is more convenient — and preferable — to express this in a slightly different
form.
Definition 3.1. For x ≥ e we set
Z x
dt
Li(x) = .
e log t
Proposition 3.1. As x → ∞,
x
Li(x) ∼ .
log x
Proof I Integrating by parts,
Z x
dt
Li(x) =
e log t
" #x Z
1 x 1
= t + t dt
log t e e t log2 t
Z x
x dt
= −e+ 2 .
log x e log t

It is clear from this that

Li(x) → ∞ as x → ∞.

3–1
3.1. STATEMENT OF THE THEOREM 3–2

Thus the result will follow if we show that

Z x
dt
= o(Li(x)).
e log2 t
But
Z x Z x1/2 Z x
dt dt dt
2 = 2 +
e log t e log t x 1/2 log2 t
Z x
1 dt
≤ x1/2 +
log(x1/2 ) x1/2 log t
2 Li(x)
≤ x1/2 + .
log x
From above,
x
Li(x) ≥ − e.
log x
Thus
x1/2 = o(Li(x));
and so Z x
dt
= o(Li(x)),
e log2 t
as required. J
Remark. We can extend this result to give an asymptotic expansion of Li(x).
Integrating by parts,
Z x " #x
Z x
dt 1 1
n = t n + t dt
e log t log t e e nt logn+1 t
x 1 Z x dt
= − e + .
logn x n e logn+1 t

It follows that
x x 1 x 1 x x
Li(x) = + 2 + 3 + ··· + n + O( n+1 ).
log x log x 2! log x (n − 1)! log x log x
Corollary 3.1. The Prime Number Theorem can be stated in the form:

π(x) ∼ Li(x).
3.2. PREVIEW OF THE PROOF 3–3

Remark. This is actually a more accurate form of the Prime Number Theo-
rem, in the following sense. It has been shown that

π(x) − Li(x)

changes sign infinitely often, ie however large x gets we find that sometimes
π(x) ≥ Li(x), and sometimes π(x) < Li(x).
On the other hand, it follows from the Remark above that Li(x) is sub-
stantially larger than x/ log x; and it has also been shown that π(x) > x/ log x
for all sufficiently large x.

3.2 Preview of the proof

The proof of the Prime Number Theorem is long and intricate, and divided
into several more or less independent parts. A preview may therefore be
helpful.

1. We start from Euler’s Product Formula

 −1
1 − p−s
Y
ζ(s) = .
primes p

2. Logarithmic differentiation converts this to

ζ 0 (s) X log p p−s
− = −s
ζ(s) p 1−p

an n−s ,
X
=
n

where 
log p if n = pr
an =
0 otherwise.

3. The function ζ 0 (s)/ζ(s) has poles wherever ζ(s) has a pole or zero. It
follows from Euler’s Product Formula that ζ(s) has no zeros in <(s) >
1. Accordingly ζ 0 (s)/ζ(s) has a pole at s = 1 (with residue 1) and no
poles in <(s) > 1.

4. Although this is not essential, our argument is somewhat simplified if

we ‘hive off’ the part of the Dirichlet series corresponding to higher
prime-powers:
ζ 0 (s) X
− = log p p−s + h(s),
ζ(s)
3.2. PREVIEW OF THE PROOF 3–4

where
p−rs .
X X
h(s) = log p
r>1

The function h(s) converges for <(s) > 1/2, as can be seen by com-
parison with ζ(2s). Its partial sums are therefore of order o(n1/2+ ), by
Proposition 2.24. Consequently the contribution of h(s) can be ignored
in our argument.
5. We are left with the function
log p p−s
X
Θ(s) =
p
Z ∞
= x−s dθ,
0

where X
θ(x) = log p.
p≤x

6. A (fairly) simple exercise in summation by parts shows that

x
π(x) ∼ ⇐⇒ θ(x) ∼ x.
log x
Accordingly, the proof of the Prime Number Theorem is reduced to
showing that
θ(x) ∼ x,

ie

θ(x) = x + o(x).

7. The dominant term x in θ(x) arises from the pole of Θ(s) at s = 1, in

the following sense.
Consider the function ζ(s). This has a pole at s = 1 with residue 1,
and it has partial sums
X
A(x) = 1 = [x] = x + O(1).
n≤x

If now we subtract ζ(s) from Θ(s) then we ‘remove’ the pole at s = 1;

and at the same time we subtract x from θ(x). More precisely, let
Ψ(s) = Θ(s) − ζ(s)
an n−s ,
X
=
3.2. PREVIEW OF THE PROOF 3–5

where 
log p − 1 if n = p,
an =
−1 otherwise.
Then Z ∞
Ψ(s) = x−s dψ,
0
where
ψ(x) = θ(x) − [x] = θ(x) − x + O(1).
The Prime Number Theorem, as we have seen, is equivalent to the
statement that
ψ(x) = o(x).

8. Riemann hypothesised — we shall see why in Chapter 7 — that all

the zeros of ζ(s) in the ‘critical strip’ 0 ≤ <(s) ≤ 1 lie on the line
<(s) = 1/2.
If that were so then Ψ(s) would be holomorphic in <(s) > 1/2, and it
would follow from Proposition 2.24 that

ψ(x) = o(x1/2+ )

for any  > 0, which is more than enough to prove the Theorem.
In fact, Riemann showed that with a little more care one can deduce
from his hypothesis that

ψ(x) = O(x1/2 log x),

ie

θ(x) = x + O(x1/2 log x),

from which it follows that

π(x) = Li(x) + O(x1/2 log x).

This — if it could be established — would constitute a remarkably

strong version of the Prime Number Theorem.

9. The Riemann Hypothesis would allow us to push back the abscissa of

convergence of Ψ(s) all the way to σ = 1/2.
3.2. PREVIEW OF THE PROOF 3–6

It would be sufficient for our purposes if we could push it back to any

σ < 1, since this would imply that

ψ(x) = o(xσ+ )

for any  > 0.

Unfortunately, this has never been established. The best that we can
do is to show that ζ(s) has no zeros actually on the line <(s) = 1:

ζ(1 + it) 6= 0

for t ∈ R \ {0}.
This proof of this result is, in a sense, the heart of the proof of the
Prime Number Theorem. The argument we use is rather strange; we
show that if ζ(s) had a zero at s = 1 + it then it would have a pole at
s = 1 + 2it, which we know is not the case.

10. This takes us a tiny step forward; it shows that Ψ(s) is holomorphic in
<(s) ≥ 1.
Proposition 2.24 only tells us that in this case

ψ(x) = o(x1+ )

for any  > 0, which is useless.

We need a much stronger result which tells us that if the Dirichlet series
an n−s is holomorphic everywhere on its critical line <(s) = σ0 (and
P

satisfies some natural auxiliary conditions) then its partial sums satisfy

A(x) = o(xσ0 ).

Results of this kind — relating partial sums of Dirichlet series to the

behaviour on the critical line — are known as Tauberian theorems,
after Alfred Tauber, author of the first such result.
Tauber’s original result used real function theory, and was very diffi-
cult. Fortunately, complex function theory yields a Tauberian theorem
sufficient for our purpose with relative ease.
This allows us to conclude that

ψ(x) = o(x),

which as we have seen is tantamount to the Prime Number Theorem.

3.3. LOGARITHMIC DIFFERENTIATION 3–7

3.3 Logarithmic differentiation

Recall the notion of logarithmic differentiation. Suppose
f (x) = u1 (x) · · · un (x),
where ui (x) is differentiable and ui (x) > 0 for 1 ≤ i ≤ n. Taking logarithms,
X
log f (x) = log ui (x).
Differentiating,
f 0 (x) X u0i (x)
= .
f (x) ui (x)
it is easy to establish this result without using logarithms: on differentiating
the product,
f 0 (x) = u1 (x) · · · u0i (x) · · · un (x);
X

and the result follows on dividing by f (x). This shows that the result holds
without assuming that ui (x) > 0. Indeed, by this argument the result holds
for complex-valued functions: if
Y
f (z) = u(z),
1≤i≤n

where u1 (z), . . . , un (z) are holomorphic in U , then

f 0 (z) X u0i (z)
= ,
f (z) ui (z)
except where z = 0.
We want to extend this to infinite products.
Proposition 3.2. Suppose an (z) (n ∈ N) is a sequence of holomorphic func-
tions on the open set U ⊂ C; and suppose the series
X
|an (z)|
is uniformly convergent on U . Then
Y
f (z) = (1 + an (z))
n

is holomorphic on U ; and
f 0 (z) X a0n (z)
=
f (z) n 1 + an (z)

on U .
3.4. FROM π(X) TO θ(X) 3–8

Proof I The partial products

Y
Pn (z) = (1 + am (z))
m≤n

converge uniformly to f (z) in U :

Pn (z) → f (z).

It follows that
Pn0 (z) → f 0 (z).
Hence
Pn0 (z) f 0 (z)
→ .
Pn (z) f (z)
But
Pn0 (z) X a0m (z)
= .
Pn (z) m≤n 1 + am (z)
We conclude that
X a0m (z) f 0 (z)
= .
n∈N 1 + am (z) f (z)
J

3.4 From π(x) to θ(x)

Definition 3.2. We set X
θ(x) = log p.
p≤x

Thus 


0 for x < 2

log 2for 2 ≤ x < 3

θ(x) = 



log 6 for 3 ≤ x < 5


...

Proposition 3.3. π(x) ∼ Li(x) ⇐⇒ θ(x) ∼ x.

Proof I Suppose
x
π(x) ∼ Li(x) ∼ .
log x
3.4. FROM π(X) TO θ(X) 3–9

Then
X
θ(X) = log p
p≤x
Z X
= log 2 + log x dπ
e
Z X
1
= log 2 + [log x π(x)]X
e −
π(x)dx
e x
Z X
π(x)
= log 2 + π(X) log X − 1 − dx.
e x
Since π(x) ∼ x/ log x,
x
π(x) ≤ C
log x
for some constant C; and so
Z X Z X
π(x) dx
0≤ dx ≤ C = C Li(x) = o(x),
e x e log x

by Proposition 3.1. Thus

x
π(x) ∼ =⇒ π(x) log x ∼ x =⇒ θ(x) ∼ x.
log x
Conversely, suppose
θ(x) ∼ x.
Then
Z X
1
π(X) = 1 + dθ
e log x
" Z X#X
θ(x) θ(x)
=1+ + 2 dx
log x e e x log x
Z X
θ(X) θ(x)
= + (1 − log 2) + dx.
log X e x log x

Now
θ(x) ∼ x =⇒ θ(x) ≤ Cx
for some C; and so
Z X
θ(x)
0≤ 2 dx
e x log x
Z X
dx
≤C 2 .
e log x
3.5. THE ZEROS OF ζ(S) 3–10

Hence Z X
θ(x) θ(x) dx
≥ π(x) ≥ +C 2 .
log x log x e log x

But
θ(x) x
θ(x) ∼ x =⇒ ∼ ∼ Li(x),
log x log x
while Z X
dx
2 = o(Li(x)),
e log x

as we saw in the proof of Proposition 3.1.

We conclude that
x
π(x) ∼ ∼ Li(x).
log x
J
Corollary 3.2. The Prime Number Theorem is equivalent to:

θ(x) ∼ x.

3.5 The zeros of ζ(s)

Proposition 3.4. The Riemann zeta function ζ(s) has no zeros in the half-
plane <(s) > 1.
Proof I This follows at once from Euler’s Product Formula:
Y −1
ζ(s) = 1 − p−s .
p

For the right-hand side converges absolutely for <(s) > 1; and by the defini-
tion of convergence its value is 6= 0. J
We want to show that ζ(s) has no zeros on the line <(s) = 1. This
is equivalent to showing that ζ 0 (s)/ζ(s) has no poles on this line except at
s = 1.
Proposition 3.5. For <(s) > 1,
ζ 0 (z)
an n−s ,
X
=−
ζ(s)
where 
log p if n = pr
an = 
0 otherwise.
3.5. THE ZEROS OF ζ(S) 3–11

Proof I The result follows at once on applying Proposition 3.2 to Euler’s

Product Formula, J
It is convenient to divide the Dirichlet series for ζ 0 (s)/ζ(s) into two parts,
the first corresponding to primes p, and the second to prime-powers pr (r ≥ 2).
Definition 3.3. We set
log p p−s .
X
Θ(s) = −
p

Proposition 3.6. The function Θ(s) is holomorphic in <(s) > 1.

Proof I We know that
n−s
X
ζ(s) =
is uniformly convergent in <(s) ≥ σ for any σ > 1. It follows that we can
differentiate term-by-term:
ζ 0 (s) = log n n−s
X

in <(s) > 1. Since the coefficients are all positive, the convergence is absolute.
But the series for Θ(s) consists of some of the terms of ζ 0 (s), and so also
converges absolutely in <(s) > 1. J
Proposition 3.7. For <(s) > 1,
ζ 0 (z)
= −Θ(s) + h(s),
ζ(s)
where h(s) is holomorphic in <(s) > 1/2.
Proof I We have
p−rs .
X X
h(s) = − log p
p r≥2
−s −σ
If σ = <(s) then |p | = p . Thus
|p−rs | = p−rσ
X X X X
log p log p
p r≥2 p r≥2
−2σ
X p
= log p
p 1 − p−σ
1
log p p−2σ
X
≤ −σ
1−2
1
= Θ(2σ),
1 − 2−σ
which converges for 2σ > 1, ie σ > 1/2, by Proposition 3.6. J
3.5. THE ZEROS OF ζ(S) 3–12

1 + 2it + σ

Figure 3.1: Comparing Θ(s) at three points

Proposition 3.8. The Riemann zeta function ζ(s) has no zeros on the line
<(s) = 1:
ζ(1 + it) 6= 0 (t ∈ R \ {0}).
Proof I We shall show (in effect) that if ζ(s) has a zero at s = 1 + it then
it must have a pole at s = 1 + 2it; but we know that is impossible, since the
only pole of ζ(s) in <(s) > 0 is at s = 1.
We work with Θ(s) rather than ζ(s). If ζ(s) has a zero of multiplicity m
at 1 + it then ζ 0 (s)/ζ(s) has a simple pole with residue m, and so Θ(s) has
a simple pole with residue −m. Similarly, where ζ(s) has a pole of order M ,
Θ(s) has a simple pole with residue M .
We are going to compare

Θ(1 + σ), Θ(1 + it + σ), Θ(1 + 2it + σ)

for small σ > 0 (Fig 3.1).

We have

log p p−(1+σ) ,
X
Θ(1 + σ) =
log p p−(1+σ) p−it ,
X
Θ(1 + it + σ) =
log p p−(1+σ) p−2it .
X
Θ(1 + 2it + σ) =
3.5. THE ZEROS OF ζ(S) 3–13

Note that

p−it = cos(t log p) − i sin(t log p),

p−2it = cos(2t log p) − i sin(2t log p).

Lemma 5. For all θ ∈ R,

cos 2θ + 4 cos θ + 3 ≥ 0.

Proof I For τ ∈ R,
eiτ + e−iτ = 2 cos(τ ) ∈ R.
Raising this to the fourth power,
 4
eiτ + e−iτ = e4iτ + e−4iτ + 4(e2iτ ) + e−2iτ ) + 6 ≥ 0,

ie

cos 4τ + cos 2τ + 3 ≥ 0.

The result follows on setting θ = 2τ . J

Lemma 6. For σ > 0,

< (Θ(1 + 2i + σ) + 4Θ(1 + it + σ) + 3Θ(1 + σ)) ≥ 0.

Proof I We have
 
< p−2it + 4p−it + 3 = cos(2t log p) + 4 cos(t log p) + 3 ≥ 0,

by the last Lemma.

The result follows on multiplying by log p p−(1+σ) and summing. J
Remark. If we had taken squares instead of fourth powers, we would have
found
< (Θ(1 + it + σ) + Θ(1 + σ)) ≥ 0,
which is not quite sufficient for our purposes.
However, higher even powers would have done as well, eg sixth powers
yield

< (Θ(1 + 3it + σ) + 6<(1 + 2it + σ) + 15Θ(1 + it + σ) + 10Θ(1 + σ)) ≥ 0,

which would have done.

3.5. THE ZEROS OF ζ(S) 3–14

Now suppose ζ(s) has a zero of multiplicity m at s = 1 + it, and suppose

it also has a zero of multiplicity M at s = 1 + 2it, where we allow M = 0 if
there is no zero. Then
1
Θ(1 + σ) = + f1 (σ),
σ
m
Θ(1 + it + σ) = − + f2 (σ),
σ
M
Θ(1 + 2it + σ) = − + f3 (σ),
σ
where f1 (σ), f2 (σ), f3 (σ) are all continuous (and so bounded) for small σ.
Adding, and taking the real part,
1 − 4m − 3M
< (Θ(1 + 2i + σ) + 4Θ(1 + it + σ) + 3Θ(1 + σ)) = + f (σ),
σ
where f (σ) is continuous. By the last Lemma, this is ≥ 0 for all σ > 0. It
follows that
1 − 4m − 3M ≥ 0.
But that is impossible, since m, n ∈ N with m > 0. J
Remark. This proof is just a neat way of dressing up the following intuitive
argument.
We know that Θ(s) has a pole at s = 1, with residue 1:
1
log p p−(1+σ) =
X
Θ(1 + σ) = + O(σ).
σ
Note that the terms are all positive.
Now suppose ζ(s) has a zero of multiplicity m at s = 1 + it. Then Θ(s)
has a pole at s = 1 + it with residue −m:
m
log p p−(1+σ) (cos(t log p) + i sin(t log p)) = −
X
Θ(1 + it + σ) = + O(σ).
σ
Comparing this with the formula for Θ(1 + σ), and noting that

−1 ≤ cos(t log p) ≤ 1,

we see that in order to reach −1/σ (let alone −m/σ), cos(t log p) must be
close to −1 for almost all p.
But
cos τ = −1 =⇒ cos 2τ = +1.
3.6. THE TAUBERIAN THEOREM 3–15

Thus follows that cos(2t log p) is close to 1 for almost all p; and that in turn
implies that

log p p−(1+σ (cos(2t log p) + i sin(2t log p))

X
Θ(1 + 2it + σ) =

is close to 1/σ, which means that Θ(s) must have a pole with residue 1 (ie
ζ(s) must have a simple pole) at s = 1 + 2it, which we know is not the case.

3.6 The Tauberian theorem

Proposition 3.9. Suppose the function f : [0, ∞) → C is

1. bounded; and

2. integrable over [0, X] for all X.

Then Z ∞
F (s) = e−xs f (x)dx
0

is defined and holomorphic in <(s) > 0.

Suppose F (s) can be extended analytically to a holomorphic function in
<(s) ≥ 0. Then f (x) is integrable on [0, ∞), and
Z ∞
f (x)dx = F (0).
0

Proof I Suppose
|f (x)| ≤ C.
For each X > 0, Z X
FX (s) = e−xs f (x)dx.
0
is an entire function, ie holomorphic in the whole of the complex plane C.
Suppose σ = <(s) > 0. If X < Y then
Z Y
FY (s) − FX (s) = e−xs f (x)dx.
X

Thus
Z Y
|FY (s) − FX (s)| ≤ C e−xσ dx
X
C −Xσ
= (e − e−Y σ).
σ
3.6. THE TAUBERIAN THEOREM 3–16

Thus
FY (s) − FX (s) → 0 as X, Y → ∞.
Hence Z ∞
F (s) = e−xs f (x)dx
0

converges for <(s) > 0.

Moreover, our argument shows that this convergence is uniform in <(s) ≥
σ for each σ > 0. Hence F (s) is holomorphic in each such half-plane, and so
in <(s) > 0.
We have to show that
Z X
FX (0) = f (x)dx → F (0)
0
R∞
as X → ∞. (This will prove both that 0 f (x)dx converges, and that its
value is F (0).)
By Cauchy’s Theorem,
1 Z ds
FX (0) − F (0) = (FX (s) − F (s))
2πi γ s
around any contour γ surrounding 0 within which F (s) is holomorphic. We
can even introduce a holomorphic factor λ(s) satisfying λ(0) = 1:
1 Z ds
FX (0) − F (0) = (FX (s) − F (s)) λ(s) .
2πi γ s
We choose the contour γ in the following way. Suppose R > 0. (We shall
later let R → ∞.) By hypothesis, F (s) is holomorphic at each point s = it
of the imaginary axis, ie it is holomorphic in some circle centred on s = it. It
follows by a standard compactness argument that we can find a δ = δ(R) > 0
such that F (s) is holomorphic in the rectangle

{s = x + iy : −δ ≤ x ≤ 0; −R ≤ y ≤ R}.

To simplify the later computations we assume — as we evidently may —

that δ ≤ R.
We take γ to be the contour formed by a large semicircle γ1 of radius R
in the positive half-plane, completed by 3 sides γ2 = γ2a + γ2b + γ2c of the
above rectangle in the negative half-plane (Fig 3.2).
We also choose our factor λ(s) (for reasons that will become apparent)
to be
s2
!
Xs
λ(s) = e 1+ 2 .
R
3.6. THE TAUBERIAN THEOREM 3–17

γ2a Ri
−δ + Ri
γ1
γ2b

−δ − Ri γ2c −Ri

Figure 3.2: The contour γ

Note that we are playing with two constants, X and R, both tending to
∞. The interaction between them is subtle. First we fix R, and let X → ∞.
We shall show that there is a constant c such that

|FX (0) − F (0)| ≤ c/R

for sufficiently large X. Since this holds for all R, it will show that

FX (0) → F (0) as X → ∞,

as required.
First we consider
Z
ds
I1 (X, R) = (FX (s) − F (s)) λ(s) .
γ1 s
For σ = <(s) > 0, Z ∞
FX (s) − F (s) = e−xs dx.
X
Thus
Z ∞
|FX (s) − F (s)| ≤ C e−xσ dx
X
C
= e−Xσ .
σ
As to the factor λ(s),
|eXs | = eXσ ;
while if s = Reiθ then
s2
1+ = 1 + e2iθ ,
R2
3.6. THE TAUBERIAN THEOREM 3–18

and so
s2 2σ
|1 + 2
| = eiθ + e−iθ = 2 cos θ = .
R R
Hence
C −Xσ Xσ 2σ
|(FX (s) − F (s)) λ(s)| ≤ e ·e
σ R
2C
= .
R
(We see now how the two parts of λ(s) were chosen to cancel out the
factors e−Xσ and 1/σ.)
Since s = Reiθ ,
ds
= ieiθ dθ;
s
and so
2Cπ
|I1 (X, R)| ≤ .
R
Turning to the part γ2 of the integral in the negative half-plane, we con-
sider FX (s) and F (s) separately:

I2 (X, R) = I20 (X, R) + I200 (X, R),

where
Z
ds
I20 (X, R) = FX (s)λ(s)
γ2 s
Z
ds
I200 (X, R) = F (s)λ(s) .
γ2 s
Since FX (s) is an entire function, we can replace the contour γ2 in the
integral I20 (X, R) by the half-circle γ20 of radius R in the negative half-plane
(Fig 3.3), ie the complementary half-circle to γ1 .
We have Z X
FX (s) = e−xs f (x)dx.
0
Thus if σ = <(s) ≤ 0 then
Z X
|FX (s)| ≤ C e−xσ dx
0
C −Xσ
≤ e .
−σ
As before,
|eXs | = eXσ ;
3.6. THE TAUBERIAN THEOREM 3–19

Ri

γ20 γ2

−Ri

Figure 3.3: From γ2 to gamma02

while
s2
|1 + | = |eiθ + e−iθ |
R2
= 2|cos θ|
−2σ
= .
R
Thus
2Cπ
I20 (X, R) ≤ .
R
It remains to consider
Z
ds
I200 (X, R) = F (s)λ(s) .
γ2 s
We divide the integrand into two parts: the factor

eXs → 0 as X → ∞

for all s ∈ γ2 except for the two end-points ±Ri; while the remaining factor

s2
!
1
F (s) 1 + 2
R s

is holomorphic in and on γ2 , and is therefore bounded there, say

|F (s)| ≤ D.

That is sufficient to show (for a given R) that

I200 (X, R) → 0 as X → ∞.
3.7. PROOF 3–20

More precisely,
Z δ
|I2a (X, R)|, |I2c (X, R)| ≤ D e−Xσ dσ
0
D
≤ ,
X
while
|I2b (X, R)| ≤ 2Re−Xδ .
Thus all three parts of I2 (X, R) tend to 0, and so

I2 (X, R) → 0 as X → ∞

for each R > 0.

Putting all this together, we deduce that
5Cπ
|FX (0) − F (0)| ≤
R
for sufficiently large X. It follows that

FX (0) − F (0) → 0 as X → ∞,

as required. J

3.7 Proof
We now have all the ingredients to complete the proof of the Prime Number
Theorem.
Proof I By Proposition 3.3, it is sufficient to prove that

θ(x) ∼ x.

We need to ‘bootstrap’ this result, by showing first that

θ(x) = O(x).

Lemma 7. There exists a constant C such that

θ(x) ≤ Cx

for all x ≥ 0.
3.7. PROOF 3–21

Proof I Consider the binary coefficient

!
2n (2n)(2n − 1) · · · (n + 1)
= .
n 1 · 2···n
This is of course an integer; and all the primes between n and 2n are factors,
since each divides the top but not the bottom. Thus
!
Y 2n
p≤ .
n<p≤2n n

But !
2n
≤ 22n ,
n
since the binomial coefficient is one term in the expansion of (1 + 1)2n . Thus
p ≤ 22n .
Y

n<p≤2n

Taking logarithms of both sides,

θ(2n) − θ(n) ≤ 2n log 2.
Setting n = 2m−1 , 2m−2 , . . . , successively,
θ(2m ) − θ(2m−1 ) ≤ 2m log 2,
θ(2m−1 ) − θ(2m−2 ) ≤ 2m−1 log 2,
...
θ(2) − θ(1) ≤ 2 log 2.
Adding,
θ(2m ) = θ(2m ) − θ(1) ≤ (2m + 2m−1 + · · · + 2) log 2
≤ 2m+1 log 2.
Now suppose
2m−1 < x ≤ 2m .
Then
θ(x) ≤ θ(2m )
≤ 2m+1 log 2
= (4 log 2)2m−1
≤ (4 log 2)x.
J
3.7. PROOF 3–22

Now let
ψ(x) = θ(x) − x.
We have to show that
ψ(x) = o(x).
For <(s) > 1, let Z ∞
Ψ(s) = x−s dψ.
1
Integrating by parts,
Z X Z X
−s
h
−s
iX dx
x dψ = x ψ(x) +s x−s ψ(x)
1 1 1 x
Z X
dx
= X −s ψ(X) − 1 + s x−s ψ(x) .
1 x
But
X −s ψ(X) → 0 as X → ∞
since
|ψ(X)| ≤ max(θ(X), X) ≤ C 0 X.
Thus
Z ∞
dx
Ψ(x) = 1 + s x−s ψ(x)
1 x
Z ∞
dx
=1+s x−s (θ(x) − x)
1 x
1
 
= 1 + s Θ(s) − .
s−1
Now Θ(s) has a pole at s = 1 with residue 1 (arising from the pole of
ζ(s)). It follows that Ψ(s) is holomorphic at s = 1; and it has no poles
elsewhere on <(s) = 1, since Θ(s) does not. Thus
Z ∞
1 dx
(Ψ(s) − 1) = x−s ψ(x)
s 1 x
is holomorphic in <(s) ≥ 1,
On making the change of variable x = et (we can think of this as passing
from the multiplicative group R+ to the additive group R),
Z ∞
1 dx
(Ψ(s) − 1) = x−s ψ(x)
s Z1 x
∞
= e−ts ψ(et )dt.
0
3.7. PROOF 3–23

We are almost in a position to apply our Tauberian theorem. There is one

last change; the theorem, as we expressed it, assumed that the critical line
was the imaginary axis <(s) = 0. But the critical line of Ψ(s) is <(s) = 1.
We therefore set
s = 1 + s0 .
We have
Z ∞
1 0
Ψ(s) = e−t(1+s ) ψ(et )dt
s Z0 ∞ 0
= e−ts e−t ψ(et )dt.
0

Now we can apply the theorem, since

|e−t ψ(et )| ≤ |e−t θ(et |

≤ e−t Cet
≤ C,

ie e−t ψ(et ) is bounded; while

1
Ψ(1 + s0 )
1 + s0
is holomorphic on <(s0 ) = 0.
We conclude that Z ∞
e−t ψ(et )dt
0

converges to Ψ(1). (We only need the convergence, not the value.)
Changing variables back to x = et , we deduce that
Z ∞
ψ(x) Z ∞
θ(x) − x
2
dx = dx
1 x 1 x2
converges.
It remains to show that this implies that

θ(x) ∼ x.

Suppose that were not so. Then either

lim sup θ(x)
>1
x
or else
lim inf θ(x)
<1
x
3.7. PROOF 3–24

(or both). In other words, there exists a δ > 0 such that either

θ(X) ≥ (1 + δ)X

for arbitrarily large X, or else

θ(X) ≤ (1 − δ)X

for arbitrarily large X.

Suppose
θ(X) ≥ (1 + δ)X.
Since θ(x) is increasing, it follows that

X ≤ x ≤ (1 + δ)X =⇒ θ(x) ≥ θ(X) ≥ (1 + δ)X ≥ x,

ie

θ(x) − x ≥ 0

on the interval [X, (1 + δ)X].

More precisely,
Z (1+δ)X
θ(x) − x Z (1+δ)X
(1 + δ)X − x
2
dx ≥ dx
X x X x2
Z 1+δ
(1 + δ) − y
≥ dy, on setting x = Xy,
1 y2
Z 1+δ
1
≥ (1 + δ − y)dy
(1 + δ)2 1
Z δ
1
≥ u du
(1 + δ)2 0
δ2
= .
2(1 + δ)2

But the fact that there exist such intervals [X, (1 + δ)X] with arbitrarily
large X contradicts the convergence of
Z ∞
θ(x) − x
dx,
x2
which we have already established. We conclude that
θ(x)
lim sup ≤ 1.
x
3.7. PROOF 3–25

Similarly, suppose
θ(X) ≤ (1 − δ)X.
Since θ(x) is increasing, it follows that
(1 − δ)X ≤ x ≤ X =⇒ θ(x) ≤ θ(X) ≤ (1 − δ)X ≤ x,

ie

θ(x) − x ≤ 0
on the interval [(1 − δ)X, X].
More precisely,
Z X
θ(x) − x Z X
x − θ(x)
− 2
dx = dx
(1−δ)X x (1−δ)X x2
Z X
x − (1 − δ)X
≥ dx
(1−δ)X x2
Z 1
y − (1 − δ)
≥ dy
1−δ y2
Z 1
1
≥ (y − 1 + δ)dy
(1 − δ)2 1−δ
Z δ
1
≥ u du
(1 − δ)2 0
δ2
= .
2(1 − δ)2
Again, this contradicts the convergence of
Z ∞
θ(x) − x
dx.
x2
Hence
θ(x)
lim inf ≥ 1.
x
We have shown therefore that
θ(x)
→ 1,
x
ie

θ(x) ∼ x.
The proof of the Prime Number Theorem is complete. J
Chapter 4

The Dirichlet L-functions

4.1 Characters of a finite abelian group

4.1.1 Definition of a character
Definition 4.1. A character of a finite abelian group A is a homomorphism

χ : A → C× .

The character defined by the trivial homomorphism is called the principal

character and is denoted by χ1 :

χ1 (a) = 1

for all a ∈ A.

Remarks. 1. We generally denote abelian groups multiplicatively — con-

trary perhaps to the usual practice — because the groups (Z/m)× to
which we shall apply the theory are multiplicative.

2. For a map χ : A → C× to be a character it is sufficient that

χ(ab) = χ(a)χ(b)

for all a, b ∈ A. For if that is so then

e2 = e =⇒ χ(e)2 = χ(e) =⇒ χ(e) = 1;

and furthermore, if a ∈ A then an = e for some n by Lagrange’s

Theorem, so that
a−1 = an−1 ,

4–1
4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–2

and therefore

χ(a−1 ) = χ(an−1 ) = χ(a)n−1 = χ(a)−1 ,

since
χ(a)n = χ(an ) = χ(e) = 1.
Example. Suppose

A = Cn = {e, g, g 2 , . . . , g n−1 : g n = e}.

Let ω = e2πi/n .
The cyclic group Cn has just n characters, namely

χ(j) : g i → ω ij (0 ≤ j < n).

For these are certainly characters of Cn ; while conversely, if χ is such a

character then

g n+1 = g =⇒ χ(g)n+1 = χ(g n+1 ) = χ(g)

=⇒ χ(g) = ω j for some j ∈ [0, n − 1]
=⇒ χ = χ(j) .

Proposition 4.1. If χ is a character of the finite abelian group A then

|χ(a)| = 1

for all a ∈ A.

Proof I By Lagrange’s Theorem, an = e for some n. Hence

χ(a)n = χ(an ) = χ(e) = 1 =⇒ |χ(a)| = 1.

Proposition 4.2. For any character χ of A,

χ(a−1 ) = χ(a).

Proof I This follows at once from Proposition 4.1, since

|z| = 1 =⇒ z −1 = z̄.

J
4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–3

4.1.2 The dual group A∗

Proposition 4.3. The characters of a finite abelian group A form a group
A∗ under multiplication:

(χχ0 )(a) = χ(a)χ0 (a).

The principal character χ1 is the identity of A∗ ; and the inverse of χ is the

character
χ−1 (a) = χ(a−1 ) = χa.
Proof I The first part follows at once, since

(χχ0 )(ab) = χ(ab)χ0 (ab)

= (χ(a)χ(b))(χ0 (a)χ0 (b))
= (χ(a)χ0 (a))(χ(b)χ0 (b))
= (χχ0 )(a)(χχ0 )(b)

The last two parts are trivial. J

Definition 4.2. The group A∗ of characters is called the dual group of A.
Example. If A = Cn then, as we have seen,

A∗ = {χ(0) , χ(1) , . . . , χ(n−1) },

where χ(j) (g i ) = ω ij . It is easy to see that

0 0
χ(i) χ(i ) = χ(i+i mod n)
.

It follows that the characters can be identified with the group Z mod m;
hence
Cn∗ ∼
= Z/(n) ∼
= Cn .
We may say that the cyclic group Cn is self-dual.
Proposition 4.4. Every finite abelian group A is self-dual, ie

A∗ ∼
= A.

Proof I We know that A is expressible as a product of cyclic groups:

A = Cn1 × · · · × Cnr .

Lemma 8. If A = B × C then

A∗ = B ∗ × C ∗ .
4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–4

Proof I We can identify B, C with the subgroups B × {e}, {e} × C of A.

Thus each character χ of A defines characters χB , χC of B, C by restriction.
Moreover, since
(b, c) = (b, e) · (e, c)
it follows that
χ(b, c) = χB (b)χC (c).
This gives a one-one correspondence
χ ←→ (χB , χC )
between characters of A, and pairs of characters of B and C; and it is straight-
forward to verify that this is an isomorphism. J
It follows from the Lemma that
A∗ = Cn∗1 × · · · × Cn∗r .
But we have seen that
Cn∗ ∼
= Cn
for any cyclic group Cn . It follows that
A∗ ∼
= A.
J
Remark. This isomorphism is non-canonical, in the sense that there is no
natural way of picking out one such isomorphism.
More precisely, the functor
A A∗
is contravariant, ie each homomorphism
α:A→B
gives rise to a homomorphism
α ∗ : B ∗ → A∗
in the opposite direction; and there is no way in general of choosing an
isomorphism θ : A → A∗ such that the diagram
θ
A −−−→ A∗
 x
 
α α∗
y 
θ
A −−−→ A∗
is commutative for all α.
4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–5

If a ∈ A then the map

χ 7→ χ(a)
defines a character of the group A∗ . This gives a (natural) homomorphism

A → A∗∗ .

Proposition 4.5. For any finite abelian group

A∗∗ = A.

Proof I Since
|A∗∗ | = |A∗ | = |A|,
it is sufficient (by the Pigeon-Hole Principle) to show that the map

A → A∗∗

is injective.
Lemma 9. If a 6= e then there exists a character χ such that

χ(a) 6= 1.

Proof I The elements

B = {b : χ(b) = 1 for all χ ∈ A∗ }

form a subgroup B ⊂ A; and every character of A is a character of the

quotient-group A/B, ie
A∗ = (A/B)∗ .
But that is impossible unless B = {e}, since otherwise

|(A/B)∗ | = |A/B| < |A| = |A∗ |.

J
We conclude that the homomorphism

A → A∗∗

is injective, and is therefore an isomorphism. J

4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–6

Remark. The character theory of finite abelian groups is a more-or-less trivial

case of the character theory of locally compact abelian groups.
Each such group A has a dual A∗ , consisting of the characters, ie continu-
ous homomorphisms χ : A → C× such that |χ(a)| = 1 for all a. (For compact
abelian groups this last condition necessarily holds. But in the non-compact
case we must impose it.)
For example, the additive group R is self-dual: R∗ = R. This is the basis
of the Fourier integral.
The dual of the torus T is the additive group of integers: T∗ = Z. That
is the basis of Fourier series.
The character theory of general locally compact abelian groups is some-
times called generalised Fourier analysis.

4.1.3 Sums over elements

Proposition 4.6. Suppose χ is a character of the finite abelian group A.
Then 
|A| if χ = χ ,
X 1
χ(a) =
0 otherwise.
a∈A

Proof I If χ = χ1 , ie χ(a) = 1 for all a ∈ A then the sum is clearly |A|.

Suppose χ 6= χ1 . Then we can find a b ∈ A such that

χ(b) 6= 1.

As a runs over A so does ab. Hence

X X
χ(a) = χ(ab)
a∈A a∈A
X
= χ(b) χ(a).
a∈A

Thus
X
(χ(b) − 1) χ(a) = 0,
a

and so
X
χ(a) = 0.
a

J
4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–7

Proposition 4.7. Suppose χ, χ0 are characters of the finite abelian group A.

Then 
|A| if χ = χ0 ,
0
X
χ(a)χ (a) =
0 otherwise.
a∈A

Proof I By Proposition 4.3,

χ(a)χ0 (a) = χ−1 (a)χ0 (a)
= (χ−1 χ0 )(a).
Hence
χ(a)χ0 (a) = (χ−1 χ0 )(a),
X X

a∈A a∈A
and the result follows from Proposition 4.6, since
χ−1 χ0 = χ1 ⇐⇒ χ = χ0 .
J

4.1.4 Sums over characters

Proposition 4.8. Suppose a ∈ A, where A is a finite abelian group. Then

X |A| if a = e,
χ(a) =
0 otherwise.
χ∈A∗

Proof I If a = e then χ(a) = 1 for all χ ∈ A∗ and the sum is evidently |A|.
Suppose a 6= e. By the Lemma to Proposition 4.5, we can find a χ0 ∈ A∗
such that
χ0 (a) 6= 1.
As χ runs over A∗ so does χ0 χ. Hence
(χ0 χ)(a)
X X
χ(a) =
χ∈A∗ χ∈A∗

= χ0 (a)
X
χ(a).
χ∈Aa st

Thus
(χ0 (a) − 1)
X
χ(a) = 0,
χ

and so
X
χ(a) = 0.
χ

J
4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–8

Proposition 4.9. Suppose ab ∈ A, where A is a finite abelian group. Then


X |A| if a = b,
χ(a)χ(b) =
0 otherwise.
χ∈A∗

Proof I Since

χ(a)χ(b) = χ(a−1 )χ(a)

= χ(a−1 b).

the result follows at once from Proposition 4.8. J

Remark. Alternatively, Propositions 4.8 and 4.9 follow at once from Propo-
sitions 4.6 and 4.7, on applying the latter to the dual group A∗ , and using
the fact that A∗∗ = A.

4.1.5 Functions on a finite abelian group

Suppose A is a finite abelian group. The functions

f :A→C

form a vector space C(A) (over C) of dimension |A|, of which the |A| char-
acters of A are elements.
It is convenient to introduce an inner product in the space C(A) of func-
tions on A.

Definition 4.3. If f (a), g(a) ∈ C(A) we set

1 X
hf gi = f (a)g(a).
|A| a∈A

It is a straightforward matter to verify that this is a positive-definite

hermitian form, ie

1. hg f i = hf gi;

2. hf f i ≥ 0, and hf f i = 0 ⇐⇒ f = 0;

3. hf λ1 g1 + λ2 g2 i = λ1 hf g1 i + λ2 hf g2 i.

Now we can re-state Proposition 4.7 as follows.

4.1. CHARACTERS OF A FINITE ABELIAN GROUP 4–9

Proposition 4.10. The characters of A form an orthonormal set:


1 if χ = χ0 ,
hχ0 χi =
0 otherwise.

Corollary 4.1. The characters are linearly independent.

Proof I Suppose X
λi χi = 0.
i
Then
* +
X
0 = χj λi χ i
i
X
= λi hχj χi i
i
= λj
for all j. J
Corollary 4.2. The characters form a basis for C(A). Explicitly, each func-
tion f : A → C is uniquely expressible as a linear combination of characters:
X
f= λχ χ,
χ

with
1 X
λχ = hχ f i = χ(a)f (a).
|A| a∈A
Proof I The characters must form a basis for C(A), since they are linearly
independent and there are
|A| = dim C(A)
of them.
So certainly X
f= λχ χ,
χ

for some λχ ∈ C. To determine these coefficients, we take the inner-product

with χ0 :
hχ0 f i = hχ0 λχ χi
X

= λχ0 .
J
4.2. MULTIPLICATIVE CHARACTERS MODM 4–10

We shall make use of one particular case of this.

Corollary 4.3. Let cb (x) denote the characterestic function of the element
b, ie 
1 if a = b,
cb (a) =
0 otherwise.

Then hcb χi = χ(b)/|A|, and so

1 X
cb = χ(b)χ.
|A| χ∈A∗

4.2 Multiplicative characters modm

Suppose m ∈ N, m 6= 0. We denote the ring of residue classes modm
by Z/(m). We can identify the classes in Z/(m) with their representatives
r, 0 ≤ r < m.
Recall that we denote by φ(m) the number of residue classes coprime to
m, ie
φ(m) = k{r : 0 ≤ r < m, gcd(r, m) = 1}k.

Proposition 4.11. The φ(m) residue classes coprime to m form a multi-

plicative group.

Proof I If r, s are coprime to m then so is rs. It remains to show that each

such residue class has an inverse s mod m:

rs ≡ 1 mod m.

If gcd(r, m) = 1 then the map

x 7→ rx mod m : Z/(m) → Z/(m)

is injective, since

rx ≡ ry mod m =⇒ m | r(x − y) =⇒ m | (x − y) =⇒ x ≡ y mod m.

It follows (by the Pigeon-Hole Principle) that this map is surjective. In

particular there exists an s such that

rs ≡ 1 mod 1.

J
4.2. MULTIPLICATIVE CHARACTERS MODM 4–11

Definition 4.4. We denote this multiplicative group by Z/m)× .

Example.

(Z/1)× = {1},
(Z/2)× = {1},
(Z/3)× = {1, 2} = {±1} ∼ = C2 ,
(Z/4)× ∼
= {1, 3} = {±1} = C2 ,
(Z/5)× = {1, 2, 3, 4} = {±1, ±2} ∼ = C4 ,
(Z/6)× ∼
= {1, 5} = {±1} = C2 ,
(Z/7)× = {1, 2, 3, 4, 5, 6} = {±1, ±2, ±3} ∼
= C6 ,
(Z/8)× = {1, 3, 5, 7} = {±1, ±3} ∼ = C2 × C2 ,
(Z/9)× = {1, 2, 4, 5, 7, 8} = {±1, ±2, ±4} ∼
= C6 ,

Proposition 4.12. Suppose m = m1 m2 , where gcd(m1 , m2 ) = 1. Then

(Z/m)× = (Z/m1 )× × (Z/m2 )× .

Proof I By the Chinese Remainder Theorem, the ring-homomorphism

Θ : Z/(m) → Z/(m1 ) × Z/(m2 ) : x 7→ (x mod m1 , x mod m2 )

is an isomorphism.
Suppose r ∈ Z/(m). Then

gcd(r, m) = 1 ⇐⇒ gcd(r, m1 ) = 1 = gcd(r, m2 ).

Hence Θ maps (Z/m)× onto (Z/m1 )× ×(Z/m2 )× , which proves the result. J
Example. Since gcd(4, 3) = 1,

(Z/12)× = (Z/4)× × (Z/3)× ,

with the pairings

1 7→ (1, 1),
5 7→ (1, 2),
7 7→ (3, 1),
11 7→ (3, 2).
4.2. MULTIPLICATIVE CHARACTERS MODM 4–12

Corollary 4.4. If gcd(m, n) = 1 then

φ(mn) = φ(m)φ(n).

Corollary 4.5. Suppose

m = pe11 · · · perr ,
where p1 , . . . , pr are distinct primes. Then

(Z/m)× = (Z/pe11 )× · · · × (Z/perr )× .

Thus the structure of the groups (Z/m)× is reduced to the structure of

the groups Z/pr )× . Although we shall not make use of the following results,
it may be helpful to know what these groups look like.
Proposition 4.13. If p is prime then the group (Z/p)× is cyclic.
Proof I We have
(Z/p)× = {1, 2, . . . , p − 1}.
Since each element of the ring Z/(p) except 0 is invertible, Z/(p) is in fact a
field.
By Lagrange’s Theorem, if G is a group of order n then

gn = e

for all g ∈ G. The smallest number e > 0 such that

ge = e

is called the exponent of G. By Lagrange’s Theorem, e | n.

Lemma 10. The exponent of (Z/p)× is p − 1.
Proof I Each element r ∈ (Z/p)∗ satisfies the equation

xe − 1 = 0

over the field Z/(p). But this equation has at most e roots. It follows that

p − 1 ≤ e.

Since e | (p − 1) it follows that

e = p − 1.

J
4.2. MULTIPLICATIVE CHARACTERS MODM 4–13

Lemma 11. Suppose A is a finite abelian group of exponent e. Then A has

an element of order e.
Proof I Let
e = pe11 · · · perr .
For each i there must be an element ai whose order is divisible by pei i ; for
otherwise pi would occur to a lower power in the exponent e. Let

order(ai ) = pei i qi .

Then
bi = aqi i
has order ei .
But if A is a finite abelian group, and a, b ∈ A have orders r, s then

gcd(r, s) = 1 =⇒ order(ab) = rs.

For suppose order(ab) = n. Then

(ab)rs = 1 =⇒ n | rs.

On the other hand, since r, s are coprime we can find x, y ∈ Z such that

rx + sy = 1.

But then
(ab)sy = asy = a1−rx = a.
It follows that r | n. Similarly s | n. Since gcd(r, s) = 1 this implies that

rs | n.

Hence
n = rs.
Applying this to
a = b1 · · · br
we conclude that a has order

pe11 · · · perr = e.

J
By these two Lemmas, we can find an element a ∈ (Z/p)× of order p − 1.
Hence (Z/p)× is cyclic. J
4.2. MULTIPLICATIVE CHARACTERS MODM 4–14

Generators of (Z/p)× are called primitive roots modp.

If a is a primitive root modp then it is easy to see that ar is a primitive
root if and only if gcd(r, p−1) = 1. It follows that there are φ(p−1) primitive
roots modp.
For example, there are just φ(6) = 2 primitive roots mod 7, namely 3 and
5 = 3−1 mod 7.

Proposition 4.14. If p is an odd prime number then the multiplicative group

(Z/pe )×

is cyclic for all e ≥ 1.

Proof I We have proved the result for e = 1. We derive the result for e > 1
in the following way.
The group (Z/pe )× has order

φ(pe ) = pe−1 (p − 1).

By the last Proposition, there exists an element a with

order(a mod p) = p − 1.

Evidently
order(a mod p) | order(a mod pe ).
Thus the order of a mod p is divisible by p − 1. It is therefore sufficient by
Lemma 11 to show that there exists an element of order pe−1 in the group.
The elements of the form x = 1 + py form a subgroup

S = {x ∈ (Z/pe )× : x ≡ 1 mod pe }

of order pe−1 . It suffices therefore to show that this subgroup is cyclic.

That is relatively straightforward, since this group is ‘almost additive’.
Each element of the group has order pj for some j. We have to show that
some element x = 1 + py has order pe−1 , ie
e−2
(1 + py)p 6≡ 1 mod pe .

By the binomial theorem,

pe−2 2 2 pe−2 3 3
! !
pe−2 e−2
(1 + py) =1+p py + py + p y + ··· .
2 3
4.2. MULTIPLICATIVE CHARACTERS MODM 4–15

We claim that all the terms after the first two are divisible by pe , ie

pe−2 r r
!
pe | py
r

for r ≥ 2. To see this, note that

pe−2 r pe−2 (pe−2 − 1) · · · (pe−2 − r + 1) r

!
p = p
r 1 · 2···r
(pe−2 − 1) · · · (pe−2 − r + 1) e−2 pr
= p
1 · 2 · · · (r − 1) r
e−2 r
!
p − 1 e−2 p
= p .
r−1 r

Thus it is sufficient to show that

pr
p2 |
r
for r ≥ 2; and that follows at once from the fact that

pr−1 > r,

eg because
pr−1 > (1 + 1)r−1 ≥ 1 + (r − 1) = r.
Thus any element of the form 1 + py where y is not divisible by p (for
example, 1 + p) must have multiplicative order pe−1 , and so must generate
S. In particular the subgroup S is cyclic, and so (Z/pe )× is cyclic. J
Turning to p = 2, it is evident that (Z/2)× is trivial, while (Z/4)× = C2 .

Proposition 4.15. If e ≥ 3 then

(Z/2e )× ∼
= C2 × C2e−2 .

Proof I Since
φ(2e ) = 2e−1 ,
(Z/2e )× contains 2e−1 elements. By the Structure Theorem for finite abelian
groups, it is sufficient to show that (Z/2e )× has exponent 2e−2 . For then one
of the factors in
(Z/2e )× = C2e1 × · · · × C2er
must be C2e−2 , and the remaining factor must be C2 .
4.2. MULTIPLICATIVE CHARACTERS MODM 4–16

This is certainly true for (Z/8)× = {±1, ±3}, since

(±1)2 = (±3)2 = 1.

It follows that (Z/2e )× cannot be cyclic for e > 3; for if a generated

(Z/2e )× then it would generate Z/8)× . (In effect, (Z/8)× is a quotient group
of (Z/2e )× .) Thus the Proposition will be proved if we can find an element
of order 2e−2 mod 2e .
We argue as we did for odd p, except that now we take x = 1 + 22 y. By
the binomial theorem,

2e−3 4 2 2e−3 6 3
! !
2 2e−3 e−3 2
(1 + 2 y) =1+2 2 y+ 2y + 2 y + ··· .
2 3

As before, all the terms after the first two are divisible by 2e , ie

pe−3 2r r
!
e
2 | 2 y
r

for r ≥ 2. For
2e−3 2r 2e−3 (2e−3 − 1) · · · (2e−3 − r + 1) 2r
!
2 = 2
r 1 · 2···r
(2e−3 − 1) · · · (2e−3 − r + 1) e−3 22r
= 2
1 · 2 · · · (r − 1) r
e−3 2r
!
2 − 1 e−3 2
= 2 .
r−1 r

Thus it is sufficient to show that

22r
23 |
r
for r ≥ 2; and that follows at once from the fact that

22(r−1) > r,

eg because

22(r−1) = (1 + 1)2(r−1) ≥ 1 + 2(r − 1) + 1 = 2r.

Thus any element of the form 1 + 22 y with y odd (for example, 5) must
have multiplicative order 2e−1 , which as we have seen is sufficient to prove
the result. J
4.2. MULTIPLICATIVE CHARACTERS MODM 4–17

4.2.1 Characters and multiplicative functions

Suppose χ is a character of (Z/m)× . Thus in principle χ is a function

χ : (Z/m)∗ → C× .

However, we extend χ to a function

χ : Z/(m) → C,

by setting
χ(r) = 0 if gcd(r, m) > 1.
Now we extend χ to a function

χ : N → Z/(m) → C

by composition. (It should cause no confusion that we use the same symbol
χ for all three functions.)
For example, suppose m = 6. Since φ(6) = 2, there are just 2 multiplica-
tive characters mod6, the principal character χ1 and the character

1 if r ≡ 1 mod 6,
χ(r) =
−1 if r ≡ 5 mod 6.

The corresponding function χ : N → C is given by


0

 if n ≡ 0, 2, 3, 4 mod 6,
χ(n) =

1 if n ≡ 1 mod 6,


−1 if n ≡ 5 mod 6.

Recall that a function

χ:N→C
is said to be multiplicative if

gcd(m, n) = 1 =⇒ χ(mn) = χ(m)χ(n),

and χ(0) = 0, χ(1) = 1. (We include the latter condition to exclude the case
where f (n) = 0 for all n).
We say that χ(n) is strictly multiplicative if

χ(mn) = χ(m)χ(n)

for all m, n ∈ N, and χ(0) = 0 χ(1) = 1.

4.2. MULTIPLICATIVE CHARACTERS MODM 4–18

Proposition 4.16. If χ is a multiplicative character modm then the corre-

sponding function
χ:N→C
is strictly multiplicative.
Proof I This is immediate; for if r or s is not coprime to m then neither is
rs, and so
χ(rs) = 0 = χ(r)χ(s).
J
Let us say that a function f : N → C is modular with modulus m if

f (n + m) = f (n)

for all n.
If is clear that if d | m then any multiplicative character modd defines a
function which is modular with modulus m.
The following result shows that every function f : N → C which is both
strictly multiplicative and modular arises in this way.
Proposition 4.17. Suppose f : N → C is modular modm. Then f (n) is
strictly multiplicative if and only if it is defined by a multiplicative character
modd for some d | m.
Proof I We argue by induction on m.
Suppose
f (d) 6= 0
for some proper divisor d | m, 1 < d < m. Then

r ≡ s mod m/d =⇒ rd ≡ sd mod m

=⇒ f (rd) = f (sd)
=⇒ f (r)f (d) = f (s)f (d)
=⇒ f (r) = f (s).

Thus f (n) is modular modm/d. It follows from our inductive hypothesis

that f (n) is defined by a multiplicative character mode for some e | d | m.
Suppose to the contrary that

d | m, d > 1 =⇒ f (d) = 0;

and suppose d = gcd(r, m) > 1, say

r = dr0 , m = dm0 .
4.3. DIRICHLET’S L-FUNCTIONS 4–19

Then
f (r) = f (d)f (r0 ) = 0.
On the other hand, if gcd(r, m) = 1 then r has a multiplicative inverse
modm, say
rs ≡ 1 mod m;
and so
f (r)f (s) = f (1) = 1 =⇒ f (r) 6= 0.
It follows that f (n) is defined by a function
χ : (Z/m)× → C× ,
which is readily seen to be a multiplicative character modm. J

4.3 Dirichlet’s L-functions

Dirichlet observed that Euler’s Product Formula could be extended to include
mutliplicative factors. Informally, if χ(n) is multiplicative then
χ(n)n−s =
X Y
Fp (s),
primes p

where
Fp (s) = 1 + χ(p)p−s + χ(p2 )p−2s + · · · .
This follows from the fact that if
n = pe11 · · · perr
then
χ(n) = χ(pe11 ) · · · χ(perr ),

and so
   
χ(n)n−s = χ(pe11 )n−e1 s · · · χ(perr )n−er s ,

If χ(n) is strictly multiplicative then

Fp (s) = 1 + χ(p)p−s + χ(p)2 p−2s + · · ·
 −1
= 1 − χ(p)p−s ;
and so Y −1
χ(n)n−s = 1 − χ(p)p−s
X
.
p
4.3. DIRICHLET’S L-FUNCTIONS 4–20

Definition 4.5. Suppose χ is a multiplicative character modm, regarded as

a function χ : N → C. Then the Dirichlet L-function corresponding to χ is
defined by the Dirichlet series

χ(n)n−s .
X
Lχ (s) =
n∈N

Proposition 4.18. Suppose χ is a multiplicative character modm.

If χ 6= χ1 then the Dirichlet series Lχ (s) converges in the half-plane
<(s) > 0, and thus defines a holomorphic function there.
If χ = χ1 then Lχ (s) converges in the half-plane <(s) > 1. However, this
function can be continued analytically to the half-plane <(s) > 0, in which it
has a single simple pole at s = 1, with residue φ(m)/m.
Proof I Let X
S(x) = χ(n).
n≤x

Lemma 12. If χ 6= χ1 then S(x) is bounded. More precisely,

|S(x)| ≤ φ(m).

Proof I By Proposition 4.1,

X
χ(r) = 0.
r∈(Z/m)×

It follows that X
χ(r) = 0,
r∈Z/(m)
P
ie χ(r) vanishes over any complete set of residues. Hence
mq−1
X
S(mq − 1) = χ(n) = 0.
n=0

for any q. Now suppose mq ≤ x < m(q + 1). Then

[x]
X
S(x) = χ(n).
n=mq

This sum contain ≤ m terms, of which at most φ(m) are non-zero. Since
|χ(n)| = 1 for each of these terms, we conclude that

|S(x)| ≤ φ(m).

J
4.3. DIRICHLET’S L-FUNCTIONS 4–21

Remark. In fact it is easy to see that

φ(m)
|S(x)| ≤ .
2
For
[x] m(q+1)−1
X X
S(x) = χ(n) = − χ(n);
n=mq [x]+1

and these two sums together contain φ(m) non-zero terms, so one of them
contains ≤ φ(m)/2 such terms.
Integrating by parts,
N Z N
−s
x−s dS
X
χ(n)n =
M M
Z N
dx
= [x−s S(x)]N
M +s x−s S(x) .
M x
Thus
N Z N
−s −σ −σ dx
x−σ
X
| χ(n)n | ≤ φ(m)(M +N ) + |s|φ(m)
M M x
!
−σ −σ |s|
= φ(m) M +N + (M −σ − N −σ .
σ

Since M −σ , N −σ → 0 as M, N → ∞, it follows that

N
χ(n)n−s | → 0
X
|
M

as M, N → ∞. Hence the series is convergent, by Cauchy’s criterion.

Now suppose χ = χ1 . Let

φ(m) X
h(s) = Lχ (s) − = a(n)n−s ,
m
where 
1 − φ(m)/m if gcd(n, m) = 1
a(n) =
−φ(m)/m if gcd(n, m) > 1
Evidently, X
a(r) = 0,
r∈Z/(m)
4.3. DIRICHLET’S L-FUNCTIONS 4–22

while |a(n)| < 1 for all n ∈ N. It follows by the argument we used above
that the Dirichlet series
a(n)n−s
X
h(s) =
n≥1

converges in <(s) > 0, and so defines a holomorphic function there.

We conclude that
φ(m)
Lχ (s) = ζ(s) + h(s)
m
defines the analytic continuation of Lχ (s) to <(s) > 0, with the only pole
arising from the pole of ζ(s) at s = 1. J

Proposition 4.19. Suppose χ is a multiplicative character modm. Then

 −1
1 − χ(p)p−s
Y
Lχ (s) =
primes p

for <(s) > 1.

Proof I This follows in exactly the same way as for ζ(s). Thus if <(s) > 1
then
0
−s −1
Y  
−s
χ(n)n−s ,
X X
1 − χ(p)p = χ(n)n +
p≤N n≤N

where the second sum on the right extends over those n > N all of whose
prime factors are ≤ N .
The sum
χ(n)n−s
X

n∈N, n6=0

converges absolutely for <(s) > 1, by comparison with ζ(s), since

N N
−s
|n−s |.
X X
|χ(n)n | ≤
M M

It follows that Y  −1

1 − χ(p)p−s → Lχ (s)
p≤N

as N → ∞.
J
Chapter 5

Dirichlet’s Theorem

Definition 5.1. Suppose r, m ∈ N. We denote by πr,m the number of primes

p ≤ x congruent to r mod m:

πr,m (x) = k{p ≤ x : p ≡ r mod m}k.

If we suppose — as we may — that 0 ≤ r < m then πr,m (x) measures the

number of primes ≤ x in the arithmetic sequence

r, r + m, r + 2m, . . . .

If r and m have a factor in common then clearly there is at most one prime
in this sequence, namely its first element r if r is prime:

gcd(r, m) > 1 =⇒ πr,m (x) ≤ 1.

We are not interested in this trivial case.

Proposition 5.1. (Dirichlet’s Theorem) If gcd(r, m) = 1 then

Li(x) 1 x
πr,m ∼ ∼ .
φ(m) φ(m) log x

Remarks. 1. It is not strictly accurate to speak of this as Dirichlet’s The-

orem, since Dirichlet only showed that if gcd(r, m) = 1 then there are
an infinity of primes in the arithmetic sequence r, r + m, r + 2m, . . . .
However, his argument, when combined with the techniques used to
prove the Prime Number Theorem in Chapter 3, immediately yields
the stronger result above; so it is not unreasonable to give Dirichlet’s
name to the theorem.

5–1
5.1. PREVIEW OF THE PROOF 5–2

2. Our proof of Dirichlet’s Theorem closely mirrors our earlier proof of

the Prime Number Theorem; and where the arguments are identical
we refer to the earlier proof for details.
As in the earlier case, we start with a preview, followed by some pre-
liminary results, before giving the proof proper.

5.1 Preview of the proof

This preview should be read in conjunction with our earlier preview (Sec-
tion 3.2) of the proof of the Prime Number Theorem.

1. We start from the analogue to Euler’s Product Formula:

 −1
1 − χ(p)p−s
Y
Lχ (s) = .
primes p

2. Logarithmic differentiation converts this to

L0χ (s) X χ(p) log p p−s
=− −s
Lχ (s) p 1 − χ(p)p

an χ(n)n−s ,
X
=−
n

where 
log p if n = pe
an =
0 otherwise.

3. Now we use the fact that we can pick out a particular residue class by
taking an appropriate linear combination of characters:
1 X L0 (s)
χ(r) χ an n−s ,
X
=
φ(m) χ Lχ (s) n≡r mod m

where the sum on the left runs over all the multiplicative characters
modm.
4. As before, it is convenient to ‘hive off’ the part of the Dirichlet series
on the right corresponding to higher prime-powers:
an n−s = Θr,m (s) + h(s),
X

where
log p p−s ,
X
Θr,m (s) =
p≡r mod
5.1. PREVIEW OF THE PROOF 5–3

while h(s) converges absolutely for <(s) > 1/2, by comparison with
ζ(2s), and so may be ignored in our argument.

5. As before (again!), Z ∞
Θr,m (s) = x−s dθr,m ,
0
where X
θr,m (x) = log p.
p≤x, p≡r mod m

6. The argument by which we showed before that

x
π(x) ∼ ⇐⇒ θ(x) ∼ x
log x
now shows that
1 x x
πr,m (x) ∼ ⇐⇒ θr,m (x) ∼ .
φ(m) log x φ(m)

Accordingly, the proof of Dirichlet’s Theorem is reduced to showing

that
x
θr,m (x) ∼ ,
φ(m)

ie
x
θr,m (x) = + o(x).
φ(m)

7. The function L0χ (s)/Lχ (s) has poles wherever Lχ (s) has a pole or zero.
It follows from the Product Formula that Lχ (s) has no zeros in <(s) >
1. Accordingly

1 X L0χ (s)
Θr,m (s) = − χ(r) + h(s)
φ(m) χ Lχ (s)

is holomorphic in <(s) > 1.

8. As with the Prime Number Theorem, the fundamental problem is to

determine what happens on the line <(s) = 1. The heart of Dirichlet’s
Theorem is the proof that none of the L-functions has a zero on this
line:
<(s) = 1 =⇒ Lχ (s) 6= 0.
5.1. PREVIEW OF THE PROOF 5–4

The proof that Lχ (1 + it) 6= 0 for t 6= 0 is straightforward; in effect, the

proof that ζ(1 + it) 6= 0 carries over unchanged. But now we have to
prove also that
Lχ (1) 6= 0
for χ 6= χ1 ; and this turns out to be a much more formidable task.

9. Having got over this hurdle, it follows that Θr,m (s) has a simple pole
at s = 1, arising from the pole of Lχ1 (s), with residue 1/φ(m), and no
other poles on the line <(s) = 1.

10. The rest of the proof is as before. We ‘remove’ the pole at s = 1 by

subtracting an appropriate multiple of ζ(s). Thus
1
Ψr,m (s) = Θr,m (s) − ζ(s)
φ(m)
is holomorphic in <(s) ≥ 1; and
Z ∞
Ψr,m (s) = x−s dψr,m ,
1

where
1
ψr,m (x) = θr,m (x) − [x]
φ(m)
1
= θr,m (x) − x + O(1).
φ(m)

11. The Tauberian Theorem now shows that

Z ∞
ψr,m (x)
dx
1 x2
converges. (Note that the bootstrap lemma — Lemma 7 — carries over
since
θr,m (x) ≤ θ(x)
for all x.)
From this we deduce, as before, that
x
θr,m (x) ∼ ;
φ(m)
and that, as we have seen, establishes Dirichlet’s Theorem.
5.2. FROM πR,M (X) TO θR,M (X) 5–5

5.2 From πr,m(x) to θr,m(x)

Definition 5.2. For r, m ∈ N we set
X
θr,m (x) = log p.
p≤x, p≡r mod m

Proposition 5.2. If gcd(r, m) = 1 then

Li(x) x
πr,m (x) ∼ ⇐⇒ θr,m (x) ∼ .
φ(m) φ(m)
Proof I This is in effect a re-wording of Proposition 3.3, taking φ(m)πr,m (x)
in place of π(x), and φ(m)θr,m (x) in place of θ(x), J
Corollary 5.1. Dirichlet’s Theorem is equivalent to:
x
θr,m (x) ∼
φ(m)
for gcd(r, m) = 1.

5.3 Picking out the residue class

Definition 5.3. For r, m ∈ N we set

log p p−s .
X
Θr,m (s) =
p≡r mod m

Proposition 5.3. If gcd(r, m) = 1 then

1 X L0 (s)
χ̄(r) χ = −Θr,m (s) + h(s),
φ(m) χ Lχ (s)

where h(s) is holomorphic in <(s) > 1/2.

Proof I If <(s) > 1 then by Proposition 4.19
Y −1
Lχ (s) = 1 − χ(p)p−s .

Differentiating logarithmically,
L0χ (s) X χ(p) log p p−s
=− −s
Lχ (s) p 1 − χ(p)p

= −Θr,m (s) + hr,m (s),

5.4. THE ZEROS OF Lχ (S) 5–6

where
p−es .
X X
hr,m (s) = − log p
p pe ≡r mod m

Since the function hr,m (s) consists of certain terms taken from the corre-
sponding series for h(s) in Proposition 3.7, and since we showed that this
series converges absolutely for <(s) > 1/2, it follows that hr,m (s) also con-
verges absolutely, and so is holomorphic, in <(s) > 1/2. J

5.4 The zeros of Lχ(s)

Proposition 5.4. Suppose χ is a multiplicative character modm. Then
Lχ (s) has no zeros in <(s) > 1.
Proof I This follows at once from the product formula for Lχ (s), like the
corresponding result for ζ(s). J
Proposition 5.5. If t 6= 0 then
Lχ (1 + it) 6= 0.
Proof I Consider
log p p−s
X
Θ1,m (s) =
p≡1 mod m

1 X L0χ (s)
=− + h(s),
φ(m) χ Lχ (s)
where h(s) is holomorphic in <(s) > 1/2 (and so may be ignored).
Each character χ for which Lχ (1 + it) = 0 will contribute to the residue
of Θr,m (s) at s = 1 + it. More precisely, if the multiplicity of this zero is mχ
then
1 X
res1+it (Θr,m ) = − mχ .
φ(m) χ
(If Lχ (1 + it) 6= 0 then we set mχ = 0.) Similarly, if each Lχ (s) has a zero
with multiplicity Mχ at s = 1 + 2it then
1 X
res1+2it (Θr,m ) = − Mχ .
φ(m) χ
We know that Lχ1 (s) has a simple pole at s = 1. Suppose that, for
χ 6= χ1 , Lχ (s) has a zero with multiplicity µχ at s = 1. Then
 
1  X
res1 (Θr,m ) = 1− µχ  .
φ(m) χ6=χ1
5.4. THE ZEROS OF Lχ (S) 5–7

But now, applying Lemma 6 to Θ1,m (s) in exactly the same way that we
applied it to Θ(s),

< (Θ1,m (1 + 2i + σ) + 4Θ1,m (1 + it + σ) + 3Θ1,m (1 + σ)) ≥ 0

for any σ > 0; and from this it follows, as before, that

res1+2i (Θ1,m ) + 4 res1+i (Θ1,m ) + 3 res1 (Θ1,m ) ≥ 0,

ie
X X X
Mχ + 4 mχ + 3 µχ ≤ 3.

Since Mχ , mχ , µχ are all non-negative integers, this implies that

mχ = 0 for all χ.

(For if mχ ≥ 1 for any χ this will already ‘out-vote’ the right-hand side.) In
other words,
Lχ (1 + it) 6= 0.
J
Proof I In the proof above, we lumped all the Lχ (s) together. We can
equally well consider the Lχ (s) separately, by modifying Lemma 6 slightly,
as follows.
Lemma 13. Let
χ(p) log p p−s .
X
Θχ (s) =
Then
< (Θχ (1 + 2i + σ) + 4Θχ (1 + it + σ) + 3Θχ (1 + σ)) ≥ 0
for any σ > 0.
Proof I If χ(p) 6= 0 then |χ(p)| = 1, say

χ(p) = eiθp .

Since χ(n) is strictly multiplicative,

χ2 (p) = (χ(p))2 = e2iθp .

It follows that
 
< χ2 (p)p−2it + 4χ(p)p−it + 3 = cos (2(t log p + θp ))+4 cos(t log p+θp )+3 ≥ 0,

by Lemma 5, with θ = t log p + θp . J

5.4. THE ZEROS OF Lχ (S) 5–8

We deduce, as before, that

res1+2i (Θχ2 ) + 4 res1+i (Θχ ) + 3 ≥ 0,

ie

−Mχ − 4mχ + 3 ≥ 0,

where Mχ , mχ are the multiplicities of the zeros of Lχ2 (s) at s = 1 + 2it and
of Lχ (s) at s = 1 + it. Since Mχ and mχ are both non-negative integers, it
follows that

mχ = 0,

ie

Lχ (1 + it) 6= 0.

J
There is one important difference between the proofs of the Prime Number
Theorem and Dirichlet’s Theorem. In the earlier proof, we knew that ζ(s)
had a simple pole at s = 1. But now, while we know that Lχ1 (s) has a
simple pole at s = 1 we must also consider the behaviour of Lχ (s) at s = 1
for χ 6= χ1 .
Of course, Lχ (s) cannot have a pole at s = 1 if χ 6= χ1 , since we know
by Proposition 4.18 that Lχ (s) is holomorphic in <(s) > 0. However, it
could have a zero at s = 1, and this would affect the residue of Θr,m (s) at
s = 1, and that in turn would affect the number of primes in the arithmetic
sequence.
We must show that this does not in fact occur, ie

Lχ (1) 6= 0

if χ 6= χ1 .
It turns out that there are two very different cases to consider, according
as χ is real or not. For non-real characters, the result follows easily by the
argument used above to show that Lχ (1 + it) 6= 0. However, the real case is
a much harder nut to crack.
Definition 5.4. The multiplicative character χ(n) mod m is said to be real
if

χ̄ = χ,
5.4. THE ZEROS OF Lχ (S) 5–9

ie

χ(n) ∈ R for all n ∈ N.

Proposition 5.6. The character χ is real if and only if

χ(n) ∈ {0, ±1}

for all n ∈ N.
Proof I If χ(n) ∈ {0, ±1} then evidently χ is real.
Conversely, suppose χ is real. If χ(n) 6= 0 then |χ(n)| = 1. Hence
χ(n) = ±1, since these are the only reals on the unit circle in C. J
Corollary 5.2. Suppose χ is a multiplicative character modm. Then

χ real ⇐⇒ χ2 = χ1 .

Proposition 5.7. If χ is non-real then

Lχ (1) 6= 0.

Proof I We have in effect already proved this result, in both of the proofs
of Proposition 5.5.
Thus in the first proof, taking any point s = 1 + it (whether Lχ (s) has a
zero there or not) it follows that
X
µχ ≤ 1.
χ

In other words,
Lχ (1) = 0
for at most one character χ.
But

Lχ (1) = 0 =⇒ Lχ (σ) → 0 as σ → 1 + 0
χ(n)nσ → 0
X
=⇒
χ(n)nσ → 0
X
=⇒
=⇒ Lχ̄ (σ) → 0
=⇒ Lχ̄ (1) = 0.

Thus if Lχ (1) = 0 and χ is non-real then

X
mχ ≥ 2,
χ
5.4. THE ZEROS OF Lχ (S) 5–10

which as we have seen is impossible.

As for the second proof, although we assumed that t 6= 0, our argument
actually shows that

res1+2i (Θχ2 ) + 4 res1+i (Θχ ) + 3 ≥ 0

even if t = 0, ie
res1 (Θχ2 ) + 4 res1 (Θχ ) + 3 ≥ 0
But now if χ is not real then χ2 6= χ1 and so Θχ2 does not have a pole
at s = 1. Hence both residues are negative, and we deduce as before that
Θχ (s) cannot have a zero at s = 1. J
Remark. These proofs might be considered something of overkill. More sim-
ply,
log p p1+σ ≥ 0
X
Θ1,m (1 + σ) =
p≡ mod m

for σ > 0. Hence

res1 (Θ1,m ) ≥ 0,

ie
X
1− µχ ≥ 0,
χ

from which it follows that mχ > 0 for at most one χ.

Proposition 5.8. If χ 6= χ1 is real then

Lχ (1) 6= 0.

Proof I Suppose χ is real; and suppose Lχ (1) = 0. Consider the product

F (s) = ζ(s)Lχ (s).

The putative zero of Lχ (s) at s = 1 cancels out the pole of ζ(s), leaving a
function F (s) holomorphic in <(s) > 0.
The following result on the product of two Dirichlet series is readily es-
tablished.
Lemma 14. Suppose the two Dirichlet series

an n−s , bn n−s
X X
f (s) = g(s) =
5.4. THE ZEROS OF Lχ (S) 5–11

are absolutely convergent in <(s) > σ. Then the product series

cn n−s ,
X
f (s)g(s) =

where X
cn = ad be ,
n=de

is also absolutely convergent in <(s) > σ.

Applying this result to F (s) = ζ(s)Lχ (s), we see that for <(s) > 1

f (n)n−s ,
X
F (s) =

where X
f (n) = χ(d).
d|n

Lemma 15. 1. f (n) is multiplicative;

2. f (n) ≥ 0 for all n;

3. f (n2 ) > 0.
Proof I 1. In general, if χ(n) is multiplicative then so is
X
f (n) = d | nχ(d).

For suppose n = n1 n2 , where gcd(n1 , n2 ) = 1. Then any factor d | n

splits into two coprime factors d = d1 d2 , where d1 | n1 and d2 | n2 . It
follows that
X
f (n) = χ(d)
d|n
X
= χ(d1 d2 )
d1 |n1 , d2 |n2
X X
= χ(d1 ) χ(d2 )
d1 |n1 d2 |n2

= f (n1 )f (n2 ).

2. Suppose
n = pe11 · · · perr .
Since f (n) is multiplicative,

f (n) = f (pe11 ) · · · f (perr .

5.4. THE ZEROS OF Lχ (S) 5–12

But

f (pe ) = χ(1) + χ(p) + · · · + χ(pe )

= χ(1) + χ(p) + · · · + χ(p)e ,

since χ is strictly multiplicative. Recall that χ(n) ∈ {0, ±1}. It follows

that 
1 if χ(p) = 0,


e
f (p ) = e + 1 if χ(p) = 1,

e


(−1) + 1 if χ(p) = −1s.
In particular
f (pe ) ≥ 0
in all cases, and so
f (n) ≥ 0.

3. Each prime factor in n2 occurs to an even power p2e . It follows from

the expression for f (pe ) above that

f (pe ) = 1, 2e + 1 or 1

according as χ(p) = 0, 1 or − 1. In all cases,

f (p2e ) > 0,

and so
f (n2 ) > 0.
J
Now suppose
f (n)n−s
X
F (s) =
has abscissa of convergence σ0 . Since the coefficients are non-negative this is
also the abscissa of absolute convergence.
By Proposition 2.25, since F (s) is holomorphic in <(s) > 0 it follows that

σ0 ≤ 0.

This is amazing; it tells us that

f (n)n−σ < ∞
X

for all σ > 0.

5.5. PROOF OF DIRICHLET’S THEOREM 5–13

But we know that

f (n2 ) ≥ 1.
These terms alone contribute

(n2 )−σ = n−2σ

X X

= ζ(2σ).

But we know that ζ(σ) diverges if σ ≤ 1. It follows that F (s) diverges

for σ ≤ 1/2, contradicting our assertion that σ0 ≤ 0.
Thus our original assumption that Lχ (1) = 0 is untenable:

Lχ (1) 6= 0

for any real character χ 6= χ1 . J

5.5 Proof of Dirichlet’s Theorem

We now have all the ingredients for our proof, which as we have said (many
times) closely imitates that of the Prime Number Theorem.
Proof I Since Lχ (s) has no zeros in <(s) ≥ 1, by Propositions 5.4, 5.5, 5.7
and 5.8, it follows that if χ 6= χ1 then

L0χ (s)
Lχ (s)

is holomorphic in <(s) ≥ 1; while on the other hand, Lχ1 (s) has a simple
pole at s = 1, by Proposition 4.18, and so

L0χ1 (s)
Lχ1 (s)

has a simple pole with residue 1 at s = 1, and no other poles in <(s) ≥ 1.

It follows that
1 X L0χ (s)
χ̄(r)
φ(m) χ Lχ (s)
has a simple pole with residue 1/φ(m) at s = 1, and no other poles in
<(s) ≥ 1. The same is therefore true of Θr,s (s), by Proposition 5.3.
Thus
1
Ψr,m (s) = Θr,m (s) − ζ(s)
φ(m)
5.5. PROOF OF DIRICHLET’S THEOREM 5–14

is holomorphic in <(s) ≥ 1; and since

Z ∞
Ψr,m (s) = x−s dψr,m
1
Z ∞
dx
=s x−s ψr,m (x)
Z1∞ x
=s e−st ψr,m (et )dt,
0

for <(s) > 1, we can apply our Tauberian Theorem, Proposition 3.9, with
1
F (s) = Ψr,m (s + 1)
s+1
and
f (x) = e−x ψr,m (ex ).
(As we noted earlier, the condition that f (x) is bounded follows at once from
the fact that
θr,m (x) ≤ θ(x) ≤ Cx
for some constant C.)
We conclude that
Z ∞ Z ∞
−t t ψr,m (x)
e ψr,m (e )dt = dx
0 1 x2
Z ∞
θr,m (x) − x/φ(m)
= dx
1 x2
converges; and from this we deduce, as before, that
x
θr,m (x) ∼ ,
φ(m)

from which Dirichlet’s Theorem follows, by Corollary 5.1. J

Chapter 6

The gamma function

6.1 Definition
Definition 6.1. For <(s) > 0 we set
Z ∞
dx
Γ(s) = xs e−x
0 x
The integral converges as x → ∞ for all s, since e−x → 0 faster than any
power xn → ∞. It converges at 0 for <(s) > 0 since

|xs−1 e−x | ≤ xσ−1 .

Proposition 6.1. Γ(s) is a holomorphic function for <(s) > 0.

Proof I The finite integral

Z X
dx
xs e−x
0 x
is holomorphic for each X > 0, by one of the standard results of complex
function theory.
Moreover, it is readily verified that if <(s) ≥ σ > 0 then
Z X
dx
xs e−x → Γ(s)
0 x
uniformly as X → ∞.
It follows that Γ(s) is holomorphic. J

6–1
6.2. THE FIRST IDENTITY 6–2

6.2 The first identity

Proposition 6.2. For <(s) > 0,

Γ(s + 1) = sΓ(s).

Proof I Integrating by parts,

Z ∞
Γ(s + 1) = xs e−x dx
0
h i∞ Z ∞
= xs · −e−x +s xs−1 e−x dx
0 0
= sΓ(s).

Corollary 6.1. For n ∈ N,

Γ(n + 1) = n!

Proof I For n = 0,
Z ∞
Γ(1) = e−x dx
0
h i∞
= −e−x
0
= 1.

The result for general n follows on repeated application of the Proposition.

J

6.3 Analytic continuation

Proposition 6.3. Γ(s) can be continued analytically to a meromorphic func-
tion in the whole plane, with simple poles at s = 0, −1, −2, . . . , the pole at
s = −n having residue (−1)n /n!.

Proof I By repeated application of the last Proposition,

1
Γ(s) = Γ(s + n).
s(s + 1) · · · (s + n − 1)

This holds for <(s) > 0. But the right-hand side is defined for <(s) > −n,
and so extends Γ(s) to this region.
6.4. AN ALTERNATIVE APPROACH 6–3

By putting together these extensions for different n (which must coincide

on their overlap by the theory of analytic continuation), we can extend Γ(s)
to the whole complex plane.
If r < n then we see from the formula above that Γ(s) has a simple pole
at s = −r with residue
1 Γ(n − r)
Γ(n − r) = (−1)r
(−r)(−r + 1) · · · (−1)(1)(2) · · · (−r + n − 1) r!(n − r − 1)!
r
(−1)
= .
r!
J

6.4 Analytic continuation: an alternative ap-

proach
There is an entirely different way of extending Γ(s) to the whole plane, which
has special significance for us, since we shall later apply the same method to
extend ζ(s) and Lχ (s) to the whole plane.
Let us ‘cut’ the complex plane along the positive real axis from 0 to +∞.
Then we can define log z holomorphically in the cut plane by setting

log(Reiθ ) = log R + iθ (0 ≤ θ ≤ 2π).

(The cut prevents us encircling 0 and thus passing from one branch of log z
to another.) On the upper edge of the cut θ = 0, and so

log z = log x

at z = x > 0. On the lower edge θ = 2π, and so

log z = log x + 2πi

at z = x > 0.
Passing to
z s = es log z ,
we have
z s = xs
at z = x on the upper edge of the cut, while

z s = e2πis xs
6.4. AN ALTERNATIVE APPROACH 6–4

γ1
γ2
γ3

Figure 6.1: The contour γ = γ1 + γ2 + γ3

at z = x on the lower edge.

Now let us consider the integral
Z
dz
I(s) = z s e−z ,
γ z
around the contour γ = γ1 + γ2 + γ3 (Fig 6.1), which comes in from +∞
to  along the upper edge of the cut (γ1 ), travels around the circle radius 
around 0 in the positive, or anti-clockwise, direction (γ2 ) and then returns
to +∞ along the lower edge of the cut (γ3 ).
Note that by Cauchy’s Theorem I(s) is independent of . For, writing
I (s) temporarily for I(s), the difference
Z
dz
I1 (s) − I2 (s) = z s e−z
C z
where C is the contour shown in Figure 6.2, within which the integrand is
holomorphic. Hence
I1 (s) − I2 (s) = 0,
ie I(s) is independent of .
(Cauchy’s Theorem can be expressed in topological terms as follows. Sup-
pose f (z) is meromorphic in the open set U , with poles at z0 , z1 , . . . . Let us
‘puncture’ U at these points, ie pass to U 0 = U \ {z0 , z1 , . . . }. If now one
contour γ in U 0 can be deformed into another contour γ 0 , without passing
through any poles, then
Z Z
f (z) dz = f (z) dz.
γ γ0

In other words, Z
f (z) dz
γ

depends only on the homotopy class of γ.)

Proposition 6.4. If <(s) > 0,
1 Z
dz
Γ(s) = z s e−z .
e2πis − 1 γ z
6.4. AN ALTERNATIVE APPROACH 6–5

Figure 6.2: The difference I1 (s) − I2 (s)

Proof I As  → 0,
Z ∞
dx
I1 (s) → − xs e−x = −Γ(s).
0 x
Similarly, Z ∞
dx
I3 (s) → e2πis xs e−x = e2πis Γ(s).
0 x
Also, if σ = <(s),

|I2 (s)| ≤ 2π · σ−1

= 2πσ
→ 0.

We conclude that
I(s) → (e2πis − 1)Γ(s)
as  → 0. Since I(s) is in fact independent of , it follows that

I(s) = (e2πis − 1)Γ(s),

ie
1
Γ(s) = I(s)
e2πis −1
for <(s) > 0. J
The integral I(s) converges for all s ∈ C, since the ‘diversion’ round 0
along γ2 avoids the problem of convergence at s = 0; it therefore defines an
entire function.
6.5. Γ(S) AS A LIMIT 6–6

Proposition 6.5. The formula

1 Z
dz
Γ(s) = z s e−z .
e2πis −1 γ z
extends Γ(s)to a meromorphic function in the whole of C, with simple poles
at s = 0, −1, −2, . . . .
Proof I Since I(s) is an entire function, the only poles of Γ(s) must arise
from poles of
1
2πis
.
e −1
But this function has simple poles with residue 1/2πi at each integer point
s = n ∈ Z. That is clear at s = 0, since

e2πis − 1 = 2πis + O(s2 )

in the neighbourhood of s = 0; and the same result holds at s = n since the

function is periodic with period 1.
However, I(s) = 0 if s = n > 0, since the integrand is in fact holomorphic
in the uncut plane. This cancels out the pole; and in any case we know that
Γ(n + 1) = n!.
For n = −n ≤ 0, it is still true that the integrand is holomorphic in
C \ {0}, but now it has a pole of order n + 1 at s = 0. The residue of the
pole is given by the coefficient of z n in e−z . Thus
2πi
I(s) = ;
n!
and so Γ(s) has a simple pole at s = −n with residue 1/n!, as we saw
before. J

6.5 Γ(s) as a limit

Euler originally defined the gamma function as a limit, in the following way.
Definition 6.2. For n ∈ N, we set
n!ns
Γ(s, n) = .
s(s + 1) · · · (s + n)
Proposition 6.6. As n → ∞,

Γ(s, n) → Γ(s).
6.5. Γ(S) AS A LIMIT 6–7

Proof I Recall that

x n
 
1− → e−x
n
as n → ∞. This follows on taking logarithms, since

x2
n !
x x

log 1 − = −n + + ···
n n 2n2
1
= −t + O( ).
n
x 2
In fact, since each term −x, − 2n , . . . increases with n, this argument shows
that (1 − x/n)n increases monotonically to e−x , for each x ≥ 0.
Let 
(1 − x/n)n if 0 ≤ x ≤ n
f (x, n) =
0 if x > n
Then
f (x, n) → e−x
uniformly in any finite range [0, X]; and

0 ≤ f (x, n) ≤ e−x

for all x.
It follows that if <(s) > 0 then
n
Z n
1 dx Z ∞ s dx

s
x 1− = x f (x, n) → Γ(s)
0 x x 0 x
as n → ∞.
6.5. Γ(S) AS A LIMIT 6–8

But we can compute this integral by repeated integration by parts. Thus

Z n n
1 dx

xs 1 − = Γ(s, n)
0 x x Z n
x n
 
= xs−1 1 − dx
0 n
x n n Z n xs
 s
x x n−1
   
= 1− + 1− dx
s n 0 0 s n
n Zn s x n−1
 
= x 1− dx
ns 0 n
n(n − 1) Z n s+1 x n−2
 
= 2 x 1− dx
n s(s + 1) 0 n
= ···
n(n − 1)(n − 2) · · · 2 Z n
x
 
= n−1 xs+n−2 1 − dx
n s(s + 1) · · · (s + n − 2) 0 n
Z n
n!
= n xs+n−1 dx
n s(s + 1) · · · (s + n − 1) 0
n! h in
= n xs+n
n s(s + 1) · · · (s + n) 0

n!
= n ns+n
n s(s + 1) · · · (s + n)
n!ns
=
s(s + 1) · · · (s + n)
= Γ(s, n).

We have therefore established that

Γ(s, n) → Γ(s)

as n → ∞, provided <(s) > 0. We can extend the result to all s (except

s = 0, −1, −2, . . . ) by noting that

nr+s n!
Γ(s + r, n) =
(s + r)(s + r + 1) · · · (s + r + n)
s(s + 1) · · · (s + n)
= nr Γ(s, n).
(s + r)(s + r + 1) · · · (s + r + n)
Thus if n ≥ r,
Γ(s + r, n) nr
Γ(s, n) = .
s(s + 1) · · · (s + r − 1) (s + n + 1) · · · (s + n + r)
6.5. Γ(S) AS A LIMIT 6–9

Now suppose <(s) > −r. From above,

Γ(s + r, n) → Γ(s + r).

Moreover,
nr 1
= s+1 s+r →1
(s + n + 1) · · · (s + n + r) (1 + n
) · · · (1 + n

as n → ∞. It follows that
Γ(s + r)
Γ(s, n) → = Γ(s).
s(s + 1) · · · (s + r − 1)

We have thus extended the result to <(s) > −r, and so to the whole plane
(excluding the poles s = 0, −1, −2, . . . ). J
We can re-write Γ(s, n) as
ns 1
Γ(s, n) = .
s (1 + s)(1 + 2s ) · · · (1 + ns )

Thus −1
s

s
Y
sΓ(s, n) = n 1+ .
1≤m≤n m
We can also re-write n as
23 n
n= ···
12 n−1
1
Y  
= 1+ .
1≤m≤(n−1)
m

Thus s
1

ns =
Y
1+ .
1≤m≤(n−1)
m
Hence −s ( −1  s )
1 s 1
 Y
sΓ(s, n) = 1 + 1+ 1+ .
n 1≤m≤n m m
Since (1 + n1 )s → 1, it follows that
( −1  s )
Y s 1
1+ 1+ → sΓ(s).
1≤m≤n m m
6.6. THE SECOND IDENTITY 6–10

In other words, Γ(s) can be expressed as the infinite product

1 Y
Γ(s) = (1 + am ),
s m≥1
where s
s −1 1
  
1 + am = 1 + 1+ .
m m
This infinite product converges absolutely, since
s −1 1 s
   
1 + am = 1 + 1+
m m
2
! !
s s 1 s s(s − 1) 1
= 1− + + O( 3 1+ + + O( 3
m m2 m m 2m2 m
s(s − 1) 1
=1− 2
+ O( 3 ),
2m m
P −2
and we know of course that m converges.
Since the series |am | is uniformly convergent in any compact (ie closed
P

and bounded) subset C not containing any of the poles, the function de-
fined by the infinite product is holomorphic in C. This gives a third way of
extending Γ(s) holomorphically to the entire plane.

6.6 The second identity

Proposition 6.7. For all s ∈ C \ Z,
π
Γ(s)Γ(1 − s) = .
sin πs
Proof I We have
ns n1−s 1
Γ(s, n)Γ(1 − s, n) = s s s s
s(1 + s)(1 + 2 ) · · · (1 + n ) (1 − s)(1 − 2 ) · · · (1 − n ) 1 − s + n
!−1
1 Y s2 n
= 1− 2
s 1≤m≤n m 1−s+n
But we saw in Chapter 1 that
s2
!
Y
sin πs = πs 1− 2 .
m
It follows that
π
Γ(s, n)Γ(1 − s, n) → ,
sin πs
from which the result follows. J
6.7. THE THIRD IDENTITY 6–11

We shall give another proof of this result below.

√
Proposition 6.8. Γ(1/2) = π.
Proof I Setting s = 1/2 in the identity above,
π
Γ(1/2)2 =
sin π2
= π.

Thus √
Γ(1/2) = ± π.
Since Z ∞
dx
Γ(1/2) = x1/2 e−x > 0,
0 x
it follows that √
Γ(1/2) = π.
J
Corollary 6.2. For each n ∈ N,
1 13 1 1
Γ(n + ) = · · · (n − ) √
2 22 2 π

6.7 The third identity

We can write (2n)! as

(2n)! = (1 · 3 · 5 · · · (2n − 1)) (2 · 4 · 6 · · · (2n))

13 1
 
= 22n · · · (n − ) n!
22 2
1
Γ(n + 2 )
= 22n n!.
Γ( 12 )
Dividing each side by 2n,

2n−1 Γ(n + 12 )Γ(n)

Γ(2n) = 2
Γ( 12 )

ie
√
Γ(n)Γ(n + 12 ) = 21−2n πΓ(2n).

This strongly suggests — but does not establish— the following result.
6.8. THE EULERIAN INTEGRAL 6–12

Proposition 6.9. For all s,

√
Γ(s)Γ(s + 12 ) = 21−2s πΓ(2s).
Proof I We have
1
1 n2s+ 2 (n!)2
Γ(s, n)Γ(s + ) =
2
s(s + 21 )(s + 1)(s + 32 ) · · · (s + n)(s + n + 12 )
1
22n+2 n2s+ 2 (n!)2
=
2s(2s + 1) · · · (2s + 2n)(2s + 2n + 1)
while
(2n)2s (2n)!
Γ(2s, 2n) =
2s(2s + 1) · · · (2s + 2n)
22s n2s (2n)!
= .
2s(2s + 1) · · · (2s + 2n)
Thus
1
22s Γ(s)Γ(s + 21 ) 22n+2 n 2 (n!)2 1
=
Γ(2s) (2n)! 2s + 2n + 1
1
22n n 2 (n − 1)!2 2n
=
(2n − 1)! 2s + 2n + 1
1
22n n 2 Γ(n)2 2n
= .
Γ(2n) 2s + 2n + 1
Note that the right-hand side is independent of s, except for the factor
2n/(2s + 2n + 1), which tends
√to 1 and can thus be ignored. We have to show
that the right-hand side → π as n → ∞, ie
1
22n n 2 Γ(n)2 √
→ π.
Γ(2n)
It follows that if the result holds for one s then it will hold for all s.
But we saw in the introduction to the Proposition that the result holds for
positive integers s = m > 0. We conclude that it holds for all s. J

6.8 The Eulerian integral

Definition 6.3. For <(u) > 0, <(v) > 0, we set
Z 1
B(u, v) = tu−1 (1 − t)v−1 dt.
0
6.8. THE EULERIAN INTEGRAL 6–13

t=0
X =x+y

x
O t=1

Figure 6.3: The double integral for Γ(u)Γ(v)

The integral converges at 0 if <(u) > 0; it converges at 1 if <(v) > 0.

Setting t = sin2 θ, the integral can be written in the form
Z π/2
B(u, v) = 2 sin2u θ cos2v θ dθ.
0

B(u, v) is often called the Eulerian integral of the first kind ; the integral
by which we defined Γ(s) being the Eulerian integral of the second kind.

Proposition 6.10. For <(u) > 0, <(v) > 0,

Γ(u)Γ(v)
B(u, v) = .
Γ(u + v)

Proof I We compute Γ(u)Γ(v) as a double integral:

Z ∞ Z ∞
u−1 −x
Γ(u)Γ(v) = x e dx y v−1 e−y dy
Z0 Z 0
u−1 v−1 −(x+y)
= x y e dx dy,

where the double integral extends over the first quadrant.

Now let us change variables to
x
X = x + y, t = .
x+y
Inversely,
x = Xt, y = X(1 − t).
The Jacobian is !
∂(x, y) t 1−t
= det = −X.
∂(X, t) X −t
6.8. THE EULERIAN INTEGRAL 6–14

Thus the integral becomes

Z ∞Z 1
Γ(u)Γ(v) = X u+v−2 tu−1 (1 − t)v−1 e−X XdX dt
0 0
Z ∞ Z 1
= X u+v−1 e−X dX tu−1 (1 − t)v−1 dt
0 0
= Γ(u + v)B(u, v),

as required. J
This provides an alternative proof of our second identity
π
Γ(s)Γ(1 − s) = .
sin πs
For suppose 0 < <(s) < 1. Then

Γ(s)Γ(1 − s) = Γ(1)B(s, 1 − s)
Z 1
dt
= ts (1 − t)−s
0 t
Z 1 s
t dt
= .
0 1−t x
Let
t
u= .
1−t
As t increases from 0 to 1, u increases from 0 to ∞. Also
u 1
t= =1− ,
u+1 u+1
and so
dt u + 1 du
=
t u (u + 1)2
du
= .
u(u + 1)

Thus Z ∞
us
Γ(s)Γ(1 − s) = du
0 u(u + 1)
Now we can play the same ‘trick’ that we used to continue Γ(s) analyti-
cally:
1 Z
zs
Γ(s)Γ(1 − s) = 2πis dz,
e − 1 γ z(z + 1)
6.8. THE EULERIAN INTEGRAL 6–15

γ10
γ4 γ2
−1
γ30

Figure 6.4: The contour γ = γ1 + γ2 + γ3

where γ is the contour shown in Fig 6.1, with the proviso now that  < 1, to
avoid the pole at s = −1.
But now let us ‘complete’ the contour with a large circle radius R (Fig 6.4).
Let
0
Z
zs
I (s) = dz,
γ 0 z(z + 1)

where γ 0 = γ10 +γ2 +γ30 +γ4 , with corresponding definitions of I10 (s), I2 (s), I30 (s), I4 (s).
As R → ∞,
I10 (s) → I1 (s), I30 (s) → I3 (s),
Also
Rσ
|I4 (s)| ≤ 2πR ;
R(R − 1)
and so
I4 (s) → 0
as R → ∞.
In fact I 0 (s) is independent of R (provided R > 1) by the same argument
that showed I(s) was independent of . Hence

I 0 (s) = I(s).

But now we can compute I 0 (s) by Cauchy’s Theorem. Since we are going
6.8. THE EULERIAN INTEGRAL 6–16

round γ 0 in the ‘wrong way’ (clockwise),

zs
I 0 (s) = −2πi res−1 (
z(z + 1)
(−1)s
= −2πi
−1
πis
= 2πie .

We conclude that
2πieπis
Γ(s)Γ(1 − s) = 2πis
e −1
2i
= πis π
e − e−πis
π
= ,
sin πs
since sin z = (eiz − e−iz )/2i.
Chapter 7

The functional equation for ζ(s)

7.1 Analytic continuation of ζ(s)

Proposition 7.1. For <(s) > 0,
Z ∞
xs dx
Γ(s)ζ(s) = .
0 ex − 1 x
Proof I The rôle of the gamma function in the theory of ζ(s) stems from
the following result.
Lemma 16. If <(s) > 0,
Z ∞
dx
xs e−nx = n−s Γ(s).
0 x
Proof I On making the change of variable y = nx,
Z ∞ Z ∞
s −nx dx dy
xe = n−s y s e−y
0 x 0 y
−s
= n Γ(s).

7–1
7.1. ANALYTIC CONTINUATION OF ζ(S) 7–2

2πi
γ1
γ2
γ3

−2πi

Figure 7.1: The contour γ

Summing this result for n = 1, 2, 3, . . . ,

∞
n−s Γ(s)
X
ζ(s)Γ(s) =
n=1
∞ Z ∞
dx
xs e−nx
X
=
n=1 0 x
Z ∞ ∞
dx
xs e−nx
X
=
0 n=1 x
Z ∞
e−x dx
= xs
0 1 − e−x x
Z ∞
xs dx
= ,
0 ex − 1 x
the interchange of sum and integral being justified by the absolute conver-
gence of the two together. J
Now we can play the same ‘trick’ that we used to analytically continue
the gamma function, integrating around the contour γ = γ1 + γ2 + γ3 in the
cut plane introduced in Proposition 6.1, with the added proviso in this case
that we must take the radius of the small circle  < 2π, to avoid the poles of
1/(ez − 1) at ±2πi (Fig 7.1).

Proposition 7.2. The Riemann zeta function ζ(s) can be analytically con-
tinued to the whole complex plane C through the formula
1 Z
z s dz
Γ(s)ζ(s) = .
e2πis − 1 γ ez − 1 z
Proof I Let Z
z s dz
I(s) = ,
γ ez − 1 z
so that
I(s) = I1 (s) + I2 (s) + I3 (s),
7.1. ANALYTIC CONTINUATION OF ζ(S) 7–3

where I1 (s), I2 (s), I3 (s) denote the corresponding integrals along γ1 , γ2 , γ3 .

As in Section 6.3, I(s) is independent of , by Cauchy’s Theorem. And as
there,
z s = xs = es log x
at z = x on the upper edge of the cut, while

z s = es(log x+2πi) = e2πis xs

at z = x on the lower edge of the cut. Thus

xs dx
Z ∞
I1 (s) + I3 (s) = (e2πis − 1)
 ex − 1 x
2πis
→ (e − 1)ζ(s) as  → 0,

by Proposition 7.1.
On the other hand, the function
z
f (z) =
ez −1
is holomorphic, and so bounded, in |z| ≤ π, say

|f (z)| ≤ C,

ie
1
| | ≤ C|z|−1 .
ez −1
Hence

|I2 (s)| ≤ 2πCσ−1 .

Thus if <(s) > 1 then

I2 (s) → 0 as  → 0.
Since I(s) is independent of , it follows that

I(s) = (e2πis − 1)ζ(s),

ie
1
ζ(s) = I(s).
e2πis −1
J
7.2. THE FUNCTIONAL EQUATION 7–4

The following alternative form of this result is often more convenient.

Corollary 7.1. For all s,

Γ(1 − s) −πis Z z s dz
ζ(s) = e .
2πi γ ez − 1 z

Proof I By Proposition 6.7,

π
Γ(s)Γ(1 − s) = .
sin πs
But
 
e2πis−1 = eπis eπis − e−πis
= 2ieπis sin πs.

Thus
1
ζ(s) = I(s)
Γ(s)(e2πis − 1)
Γ(1 − s) −πis
= e I(s).
2πi
J

Proposition 7.3. The only pole of ζ(s) in the entire complex plane C is the
simple pole (with residue 1) at s = 1.

Proof I T hef unctionΓ(1 − s) has poles at s = 1, 2, 3, . . . , since Γ(s) has

poles at s = 0, −1, −2, . . . . On the other hand, the function I(s) is entire, as
is e−πis .
It follows from the Corollary to the last Proposition that ζ(s) can only
have poles at s = 1, 2, 3, . . . . But we know that ζ(s) is holomorphic for
<(s) > 1. Thus the only possible pole is at s = 1, and we already know that
there is a simple pole there with residue 1. J

7.2 The functional equation

Proposition 7.4. The Riemann zeta function ζ(s) satisfies the functional
equation
ζ(1 − s) = 2 cos πs
2
(2π)−s Γ(s)ζ(s).
7.2. THE FUNCTIONAL EQUATION 7–5

2πi
γ10
γ4 γ2
γ30

−2πi

Figure 7.2: The contour γ 0

Proof I Suppose σ = <(s) < 0. Let

zs 1
F (z) = z ;
e −1z
and let Z
0
I (s) = F (z)dz
γ0

around the clockwise contour

γ 0 = γ10 + γ2 + γ30 + γ4

(Fig 7.2), where γ10 runs from R to  along the upper edge of the cut, γ2 is a
small circle radius  as before, γ30 runs from  to R along the lower edge of the
cut, and γ4 is the circle radius R = (2n+1)π considered above. Let us denote
the corresponding integrals along these contours by I10 (s), I2 (s), I30 (s), I4 (s),
so that
I 0 (s) = I10 (s) + I2 (s) + I30 (s) + I4 (s).
To avoid the poles of 1/(ez − 1) at z = 2nπi let us take

R = (2n + 1)π,

so that the circle γ4 passes mid-way between two poles at the top, and simi-
larly at the bottom.
7.2. THE FUNCTIONAL EQUATION 7–6

As n → ∞ (and so R → ∞),

I10 (s) → I1 (s), I30 (s) → I3 (s),

On the other hand, we shall show that, since <(s) < 0,

I4 (s) → 0

as n → ∞. It will follow that

I 0 (s) → I(s).

The function
1
f (z) =
ez −1
has poles at
z = 2nπi (n ∈ Z).
The following Lemma shows that provided we keep a reasonable distance
away from the poles, the function f (z) will remain reasonably small.
Lemma 17. There is a constant C such that
1
≤ C.
|ez − 1|

provided
|z − 2nπi| ≥ 1
for all n ∈ Z.
Proof I Since f (z) = 1/(ez − 1) is periodic with period 2πi, it is sufficient
to consider its value in the strip

S = {z = x + iy : −π ≤ y ≤ π}

outside the disk

D = {z : |z| ≤ 1}
(Fig 7.3).
The function
z
g(z) = zf (z) =
−1 ez
is holomorphic in S, and is therefore bounded in any finite part of this strip,
say
|g(z)| ≤ c
7.2. THE FUNCTIONAL EQUATION 7–7

πi

R
S

−πi

Figure 7.3: Determining sup|1/(ez − 1)|

in
R = {z : −1 ≤ <(z) ≤ 1}
(Fig 7.3). Thus
|f (z)| ≤ c
in R \ D (since |z| ≥ 1 outside D).
On the other hand, if <(z) ≥ 1 then

|ez − 1| ≥ |ez | − 1 ≥ e − 1;

while if <(z) ≤ −1 then

|ez − 1| ≥ 1 − |ez | ≥ 1 − e−1 .

It follows that
1
≤ C = max(c, 1/(1 − e−1 ).
|ez − 1|
J
By the Lemma,
1
≤C
|ez − 1|
7.2. THE FUNCTIONAL EQUATION 7–8

on the large circle γ4 ; while

|z s | = Rσ
on this circle. Hence

|I4 (s)| ≤ 2πCRσ

→ 0 as n → ∞,

since σ = <(s) < 0.

It follows that
 
I 0 (s) → I(s) = e2πis − 1 Γ(s)ζ(s)

as n → ∞.
But now. since the contour γ 0 is closed, we can compute the integral
I 0 (s) by Cauchy’s Theorem, from the residues of F (z) at its poles within the
contour. Since the contour runs in the ‘wrong’ direction (clockwise rather
than anti-clockwise), we must negate the sum. Thus

I 0 (s) = −2πi
X
(res2mπi (F ) + res−2mπi (F )) .
0<m≤n

In the neighbourhood of z = 0,
1 1 1
f (z) = = = + h(z),
ez −1 2
z + z /2 + · · · z

where h(z) is holomorphic at z = 0. It follows that f (z) has a simple pole

with residue 1 at z = 0. Therefore, since f (z) is periodic with period 2πi, it
has a simple pole with residue 1 at z = 2nπi for each n ∈ Z. Thus
(2nπi)s (−2nπi)s
res2nπi (F ) = , res−2nπi (F ) = .
2nπi −2nπi
We must take care to compute the powers correctly. Recall that if

z = reiθ (0 ≤ θ ≤ 2π)

then we must take

z s = rs eiθs .
Thus
z = 2nπi = 2nπeiπ/2 =⇒ z s = (2nπ)s eπis/2 ,
while
z = −2nπi = 2nπe3πi/2 =⇒ z s = (2nπ)s e3πis/2 .
7.2. THE FUNCTIONAL EQUATION 7–9

It follows that
eπi/2 e3πi/2
!
0 s
X
I (s) = −2πi (2nπ) +
0<m≤n 2nπi −2nπi
 
= (2π)s ns−1 e3πi/2 − eπi/2 .
X

0<m≤n

Since
ns−1 = ζ(1 − s),
X

we conclude that
1
Γ(s)ζ(s) = I(s)
e2πis −1
1
= lim I 0 (s)
e2πis − 1 n→∞
e3πis/2 − eπis/2
= (2π)s ζ(1 − s)
e2πis − 1
eπis/2 − e−πis/2
= (2π)s ζ(1 − s)
eπis − e−πis
1
= (2π)s ζ(1 − s)
eπis/2 + e−πis/2
1
= (2π)s ζ(1 − s),
2 cos(πs/2)
ie
2 cos(πs/2)Γ(s)ζ(s) = (2π)s ζ(1 − s).
All this was on the assumption that <(s) < 0. But now it follows by
analytic continuation that the result holds for all s ∈ C. J
The functional equation can be re-written in various ways, using the
properties of Γ(s) established in Chapter 6. In particular we can express it
in a form invariant under the transformation s → 1 − s. (In geometric terms,
this transformation describes reflection in the point s = 1/2.)

Proposition 7.5. Let

s
 
ξ(s) = s(s − 1)π − 2 Γ s
2
ζ(s).

Then ξ(s) is an entire function; and

ξ(1 − s) = ξ(s).
7.2. THE FUNCTIONAL EQUATION 7–10

Proof I The function Γ(s/2) has poles at s = 0, −2, −4, . . . , while ζ(s) has
zeros at s = −2, −4, . . . . This leaves a pole at s = 0 which is cancelled by
the zero of the factor s, In addition, the pole of ζ(s) at s = 1 is cancelled by
the zero of the factor s − 1. Thus all possible poles of ξ(s) are accounted for,
and this function must be entire.
By the second gamma function identity (Proposition 6.7), with (1 − s)/2
in place of s,

1−s
 
1+s
 π
Γ 2
Γ 2
= π(1−s)
sin 2
π
= ,
cos πs
2

since sin(π/2 − τ ) = cos τ .

By the third gamma function identity (Proposition 6.9), with s/2 in place
of s,
1
   
Γ 2s Γ 1+s
2
= 21−s π 2 Γ(s).
Dividing one relation by the other,

Γ( 2s ) πs 1
1−s = 2
1−s
cos Γ(s)π − 2 .
Γ( 2 ) 2

But the functional equation can be written

ζ(1 − s) πs
= 21−s cos Γ(s)π −s
ζ(s) 2
s
Γ( 2 ) 1 −s
= 1−s π2 .
Γ( 2 )

Thus if we set
s
η(s) = Γ( )ζ(s)
2

then

η(1 − s) 1
= π 2 −s .
η(s)

But now if we set

s
θ(s) = π 2
7.3. THE BEHAVIOUR OF ζ(S) FOR <(S) ≤ 0 7–11

then
θ(1 − s) 1
= π 2 −s .
θ(s)
Hence
η(1 − s) θ(1 − s)
=
η(s) θ(s)

ie

β(1 − s) = β(s),

where

β(s) = η(s)θ(s)
s
 
= π− 2 Γ s
2
ζ(s).

We conclude that, since the function s(s − 1) is invariant under s 7→ 1 − s

(we include it to remove the pole at s = 1),

ξ(s) = s(s − 1)β(s)

s
 
= s(s − 1)π − 2 Γ s
2
ζ(s)

satisfies
ξ(1 − s) = ξ(s).
J

7.3 The behaviour of ζ(s) for <(s) ≤ 0

The functional equation allows us to determine how ζ(s) behaves ‘on the far
side’ of the critical strip 0 ≤ <(s) ≤ 1; for the map

s 7→ 1 − 2

sends the left-hand half-plane <(s) < 0 into the half-plane <(s) > 1, where
ζ(s) is well-behaved.
We already know that ζ(s) has no poles in <(s) ≤ 0, by Proposition 7.3.
It does however have zeros, as we shall see.
Proposition 7.6. The Riemann zeta function ζ(s) has simple zeros at s =
−2, −4, −6, . . . . It has no other zeros (or poles) in <(s) ≤ 0.
7.4. THE VALUES OF ζ(2N ) 7–12

Proof I Since π −s/2 and Γ(s/2) have no zeros anywhere, it follows that any
zero of
s
 
ξ(s) = s(s − 1)π 2 Γ ss ζ(s)
must be a zero of ζ(s), except possibly for s = 0, 1.
At s = 1, ζ(s) has a simple pole which is cancelled out by the zero of
s − 1. Thus ξ(1) 6= 0; and since ξ(0) = ξ(1) by the functional equation
ξ(1 − s) = ξ(s), it follows that ξ(0) 6= 0. Thus

ξ(s) = 0 =⇒ ζ(s) = 0.

Now we know that ζ(s) has no zeros in <(s) ≥ 1 by Propositions 3.4 and
3.8. Hence ξ(s) has no zeros in <(s) ≥ 1. Thus ξ(s) has no zeros in <(s) ≤ 0,
since ξ(1 − s) = ξ(s).
It follows that ζ(s) has zeros in <(s) ≤ 0 just at those points where
s(s − 1)Γ(s/2) has poles. Now Γ(s/2) has simple poles at s = 0, −2, −4, . . . ;
but the pole at s = 0 is cancelled by the zero of s at this point. We conclude
that ζ(s) has simple zeros at s = −2, −4, −6, . . . , and that these are the only
zeros of ζ(s) in <(s) ≤ 0. J

7.4 The values of ζ(2n)

The functional equation allows us to express ζ(2n) in terms of ζ(1 − 2n).
Although at first sight this might seem a step backwards, it turns out that the
latter can be determined with relative ease, using Cauchy’s Residue Theorem.
Interestingly, the argument only works for even values; there seem to be
no simple expressions for

ζ(3), ζ(5), ζ(7), . . . .

7.4.1 The Bernouilli numbers

Our formulae for ζ(2n) involve the Bernouilli numbers, rational numbers
which occur in many mathematical formulae.
Definition 7.1. The Bernouilli numbers Bn (n ∈ N) are defined by
z X zi
= Bi .
ez − 1 n∈N i!

Remarks. 1. Different authors use slightly different notations for the Bernouilli
numbers. As we shall see, the odd Bernouilli numbers all vanish after
the first. What we call B2n is sometimes denoted by Bn .
7.4. THE VALUES OF ζ(2N ) 7–13

Again, it will follow from our formulae for ζ(2n) that B2 , B4 , B6 , . . . are
alternatively positive and negative. Sometimes Bn is used to denote the
absolute value, so that Bn ≥ 0 for all n.
However, we shall stick with the definition above.
2. We can compute the Bernouilli numbers recursively from the identity
  
1 + 12 z + 16 z 2 + 1 3
24
z + 1 4
120
z + ··· B0 + B1 z + 12 B2 z 2 + 61 B3 z 3 + 1
B z4
24 4
+ · · · = 1.

Comparing constant terms,

B0 = 1.

Comparing coefficients of z, z 2 , z 3 , z 4 ,

B1 + 21 B0 =0 =⇒ B1 = − 12 ,
1
B + 12 B1 + 61 B0
2 2
=0 =⇒ B2 = − 16 ,
1
B + 41 B2 + 16 B1 + 24
6 3
1
B0 =0 =⇒ B3 = 0,
1 1 1 1 1 1
B + 12 B3 + 12 B2 + 24 B1 + 24 B0
24 4
=0 =⇒ B4 = − 30 .

Proposition 7.7. The odd Bernouilli numbers after B1 all vanish:

B2n+1 = 0 (n = 1, 2, 3, . . . ).

Proof I Let
z
f (z) = .
ez − 1
Then
−z
f (−z) =
e−z
−1
zez
= z .
e −1
Thus
f (z) − f (−z) = −z.
On the other hand,
X zn
f (z) − f (−z) = 2 Bn .
n odd n!

It follows that
1
B1 = − , B3 = B5 = · · · = 0.
2
J
7.4. THE VALUES OF ζ(2N ) 7–14

7.4.2 Determining ζ(1 − 2n)

Proposition 7.8. For n = 1, 2, 3, . . . ,
B2n
ζ(1 − 2n) = − .
2n
Proof I By the Corollary to Proposition 7.2, setting s = 1 − 2n,

Γ(2n) −πi(1−2n) Z z −2n

ζ(1 − 2n) = e dz.
2πi γ ez − 1

Now the function

z −2n
F (z) =
ez − 1
is meromorphic in the complex plane. In particular, the values of F (z) at
z = x on the upper and lower edges of the cut coincide. It follows that the
integrals of F (z) along γ1 and γ3 cancel out, leaving

Γ(2n)
ζ(1 − 2n) = − I2 ,
2πi
where
Z
I2 = F (z)dz
γ2
= 2πi res0 (F ),

by Cauchy’s Theorem.
But
z −2n
F (z) =
ez − 1
z
= z −2n−1
ez −1
−2n−1
X zr
=z Br
r≥0 r!
X z r−2n−1
= Br .
r≥0 r!

By definition, res0 (F ) is the coefficient of z −1 in this expansion. Thus

B2n
res0 (F ) = .
(2n)!
7.4. THE VALUES OF ζ(2N ) 7–15

Hence
Γ(2n) B2n
ζ(1 − 2n) = (−2πi)
2πi (2n)!
(2n − 1)!
=− B2n
(2n)!
B2n
=− .
2n
J

7.4.3 Determining ζ(2n)

Proposition 7.9. For n = 1, 2, 3, . . . ,
B2n
ζ(2n) = (−1)n−1 22n−1 π 2n .
(2n)!

Proof I By the functorial equation, Proposition 7.4,

ζ(1 − 2n) = 2 cos 2πn

2
(2π)−2n Γ(2n)ζ(2n).

Thus
ζ(1 − 2n)
ζ(2n) = (−1)n 22n−1 π 2n
Γ(2n)
B2n
= (−1)n−1 22n−1 π 2n
2nΓ(2n)
B2n
= (−1)n−1 22n−1 π 2n .
(2n)!
J
For n = 1 this gives

1 1 2 π2
ζ(2) = 1 + + + · · · = π B2 = ,
22 32 6
a result which is probably familiar, and which can be proved in several ways.
For n = 2 it gives

1 1 1 4 π4
ζ(4) = 1 + + + · · · = − 3
π B4 = .
24 34 90
7.5. POSTSCRIPT 7–16

7.5 Postscript
In his seminal paper (Appendix B), Riemann gave a second proof of the
functional equation. Although at first sight this seems more complicated than
his first proof (given above) it has turned out to have far greater significance.
By Lemma 16, with πn2 in place of n,
Z ∞
2x dx
Γ(s)(πn2 )−s = xs e−πn .
0 x
Summing over n, as before,
Z ∞
ψ(x) − 1 dx
Γ(s)π −s ζ(2s) = zs ,
0 2 x
where ∞
2
e−πn x .
X
ψ(x) =
−∞

Some 20 years before Riemann’s work, Jacobi had published a study of

the function ψ(x), in the course of which he showed that ψ(x) satisfies the
functional equation
1
 
ψ x1 = x− 2 ψ(x).
It is a straightforward matter to derive the functional equation for ζ(s) from
this.
It follows from Jacobi’s identity that the theta function
2x
epiin
X
θ(x) = ψ(x/i) =

satisfies the equation   q

1 i
θ x
= x
θ(x).
It is clear that θ(x) is also periodic with period 1:

θ(x + 1) = θ(x).

Now the transformations x 7→ 1/x, x 7→ x+1 generate the modular group

consisting of the transformations
az + b
z 7→ (a, b, c, d ∈ Z, ad − bc = 1).
cz + d
This group can be identified with the group of 2 × 2 matrices
!
a b
SL2 (Z) = { : ad − bc = 1}
c d
7.5. POSTSCRIPT 7–17

The relation between zeta functions and modular functions — functions

invariant, or nearly invariant, under the modular group — has proved remark-
ably fruitful. Andrew Wiles’ proof of Fermat’s Last Theorem, for example,
was based on this correspondence.
Another advantage of this approach is that it leads to a functional equa-
tion for Lχ (s), although one relating Lχ (1 − s) to Lχ̄ (s), where χ̄ is the
conjugate character to χ, given by

χ̄(a) = χ(a) = χ(a−1 ).

This identity in turn suggests the Generalised Riemann Hypothesis, which

asserts that the zeros of Lχ (s) in the critical strip 0 < <(s) < 1 all lie on the
line <(s) = 1/2.
Incidentally, the zeta functions ζk (s) of number fields k, which we briefly
alluded to earlier, can all be expressed in terms of the Riemann zeta function
ζ(s) and the L-functions Lc hi(s); and the Riemann Hypothesis for ζk (s)
would follow from the Generalised Riemann Hypothesis. In that sense, the
Generalised Riemann Hypothesis is as general as one would wish.
Contents

1 Euler’s Product Formula 1–1

1.1 The Product Formula . . . . . . . . . . . . . . . . . . . . . . . 1–1
1.2 Infinite products . . . . . . . . . . . . . . . . . . . . . . . . . 1–2
1.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . 1–2
1.2.2 The complex logarithm . . . . . . . . . . . . . . . . . . 1–3
1.2.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . 1–4
1.3 Proof of the product formula . . . . . . . . . . . . . . . . . . . 1–6
1.4 Euler’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 1–7

2 Dirichlet series 2–1

2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–1
2.2 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–2
2.3 Absolute convergence . . . . . . . . . . . . . . . . . . . . . . . 2–5
2.4 The Riemann zeta function . . . . . . . . . . . . . . . . . . . 2–7
2.5 The Riemann-Stieltjes integral . . . . . . . . . . . . . . . . . . 2–12
2.5.1 Functions of bounded variation . . . . . . . . . . . . . 2–14
2.5.2 Discontinuities . . . . . . . . . . . . . . . . . . . . . . 2–19
2.5.3 Integration by parts . . . . . . . . . . . . . . . . . . . 2–24
2.5.4 The abscissa of convergence revisited . . . . . . . . . . 2–26
2.5.5 Analytically continuing ζ(s) . . . . . . . . . . . . . . . 2–27
2.6 The relation between An and σ0 . . . . . . . . . . . . . . . . . 2–30
2.7 Dirichlet series with positive terms . . . . . . . . . . . . . . . 2–32

3 The Prime Number Theorem 3–1

3.1 Statement of the theorem . . . . . . . . . . . . . . . . . . . . 3–1
3.2 Preview of the proof . . . . . . . . . . . . . . . . . . . . . . . 3–3
3.3 Logarithmic differentiation . . . . . . . . . . . . . . . . . . . . 3–7
3.4 From π(x) to θ(x) . . . . . . . . . . . . . . . . . . . . . . . . . 3–8
3.5 The zeros of ζ(s) . . . . . . . . . . . . . . . . . . . . . . . . . 3–10
3.6 The Tauberian theorem . . . . . . . . . . . . . . . . . . . . . . 3–15
3.7 Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3–20

0–18
CONTENTS 0–19

4 The Dirichlet L-functions 4–1

4.1 Characters of a finite abelian group . . . . . . . . . . . . . . . 4–1
4.1.1 Definition of a character . . . . . . . . . . . . . . . . . 4–1
4.1.2 The dual group A∗ . . . . . . . . . . . . . . . . . . . . 4–3
4.1.3 Sums over elements . . . . . . . . . . . . . . . . . . . . 4–6
4.1.4 Sums over characters . . . . . . . . . . . . . . . . . . . 4–7
4.1.5 Functions on a finite abelian group . . . . . . . . . . . 4–8
4.2 Multiplicative characters modm . . . . . . . . . . . . . . . . . 4–10
4.2.1 Characters and multiplicative functions . . . . . . . . . 4–17
4.3 Dirichlet’s L-functions . . . . . . . . . . . . . . . . . . . . . . 4–19

5 Dirichlet’s Theorem 5–1

5.1 Preview of the proof . . . . . . . . . . . . . . . . . . . . . . . 5–2
5.2 From πr,m (x) to θr,m (x) . . . . . . . . . . . . . . . . . . . . . . 5–5
5.3 Picking out the residue class . . . . . . . . . . . . . . . . . . . 5–5
5.4 The zeros of Lχ (s) . . . . . . . . . . . . . . . . . . . . . . . . 5–6
5.5 Proof of Dirichlet’s Theorem . . . . . . . . . . . . . . . . . . . 5–13

6 The gamma function 6–1

6.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6–1
6.2 The first identity . . . . . . . . . . . . . . . . . . . . . . . . . 6–2
6.3 Analytic continuation . . . . . . . . . . . . . . . . . . . . . . . 6–2
6.4 An alternative approach . . . . . . . . . . . . . . . . . . . . . 6–3
6.5 Γ(s) as a limit . . . . . . . . . . . . . . . . . . . . . . . . . . . 6–6
6.6 The second identity . . . . . . . . . . . . . . . . . . . . . . . . 6–10
6.7 The third identity . . . . . . . . . . . . . . . . . . . . . . . . . 6–11
6.8 The Eulerian integral . . . . . . . . . . . . . . . . . . . . . . . 6–12

7 The functional equation for ζ(s) 7–1

7.1 Analytic continuation of ζ(s) . . . . . . . . . . . . . . . . . . . 7–1
7.2 The functional equation . . . . . . . . . . . . . . . . . . . . . 7–4
7.3 The behaviour of ζ(s) for <(s) ≤ 0 . . . . . . . . . . . . . . . 7–11
7.4 The values of ζ(2n) . . . . . . . . . . . . . . . . . . . . . . . . 7–12
7.4.1 The Bernouilli numbers . . . . . . . . . . . . . . . . . 7–12
7.4.2 Determining ζ(1 − 2n) . . . . . . . . . . . . . . . . . . 7–14
7.4.3 Determining ζ(2n) . . . . . . . . . . . . . . . . . . . . 7–15
7.5 Postscript . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7–16