Chapter 9

The functional equation for the

Riemann zeta function

We will eventually deduce a functional equation, relating (s) to (1 s). There are
various methods to derive this functional equation, see E.C. Titchmarsh, The theory
of the Riemann zeta function. We give a proof based on a functional equation for
the Jacobi theta function (z) = m2 z
P
m= e . We start with some preparations.

9.1 Poissons summation formula

We start with a simple result from Fourier analysis. Given a function f : [0, 1] C,
we define the Fourier coefficients of f by
Z 1
cn (f ) := f (t)e2int dt for n Z.
0

Theorem 9.1. Let f be a complex analytic function, defined on an open subset of

C containing the real interval [0, 1]. Then
N
( 1

X
2inx 2
f (0) + f (1) if x = 0 or x = 1,
lim cn (f )e =
N f (x) if 0 < x < 1.
n=N

Remarks 1. The condition that f be analytic on an open subset containing [0, 1] is

much too strong, but it has been inserted first since it is sufficient for our purposes,

125
and second since it considerably simplifies the proof. Dirichlet proved the above
theorem for functions f : [0, 1] C that are differentiable and whose derivative is
piecewise continuous.
2. It may be that a doubly infinite series an = limM,N N
P P
PN n= n=M an di-
verges, while limN n=N an converges. For instance, if an = an for n
P
Z \ {0}, then limN N
P
n=N an = a0 , while n= an may be horribly divergent.

Proof. We first assume that either 0 < x < 1, or that x {0, 1} and f (0) = f (1).
We use the so-called Dirichlet kernel
N
X 2N
X
DN (x) = e2inx = e2iN x e2inx
n=N n=0
2i(2N +1)x
e 1
= e2iN x
e2ix 1
ei(2N +1)x ei(2N +1)x sin(2N + 1)x
= ix ix
= .
e e sin x
Further, we use Z 1 
2int 1 if n = 0,
e dt =
0 0 if n 6= 0.
Using these facts, we obtain
N
X N
X Z 1 
2inx 2int
f (x) cn (f )e = f (x) f (t)e dt e2inx
n=N n=N 0

N
X Z 1  N
X Z 1 
2int 2inx 2int
= f (x)e dt e f (t)e dt e2inx
n=N 0 n=N 0

(the first integral is f (x) if n = 0 and 0 if n 6= 0)

N
X Z 1
f (x) f (t) e2in(tx) dt


=
n=N 0

N
!
Z 1 X
e2in(tx) dt


= f (x) f (t)
0 n=N

1



sin (2N + 1)(t x)
Z
= (f (x) f (t) dt.
0 sin (t x)

126
Fix x and define
f (x) f (z)
g(z) := .
sin (z x)
We show that g is analytic on an open set containing [0, 1]. First, suppose that
0 < x < 1. By assumption, f is analytic on an open set U C containing [0, 1]. By
shrinking U if needed, we may assume that U contains [0, 1] but not x + n for any
non-zero integer n. Then sin (z x) has a simple zero at z = x but is otherwise
non-zero on U . This shows that g(z) is analytic on U \ {x}. But g(z) is also analytic
at z = x, since the simple zero of sin (z x) is cancelled by the zero of f (x) f (z).
In case that x {0, 1} and f (0) = f (1) one proceeds in the same manner.
Using integration by parts, we obtain
N
X Z 1
2inx
f (x) cn (f )e = g(t) sin{(2N + 1)(t x)}dt
n=N 0

1
1
Z
= g(t)d cos{(2N + 1)(t x)}
(2N + 1) 0

1 n
= g(1) cos{(2N + 1)(1 x)} g(0) cos{(2N + 1)x} +
(2N + 1)
Z 1 o
+ g 0 (t) cos{(2N + 1)(t x)}dt .
0

The functions g(t), g 0 (t) are continuous, hence bounded on [0, 1] since g is analytic,
and also the cosine terms are bounded on [0, 1]. It follows that the above expression
converges to 0 as N .
We are left with the case x {0, 1} and f (0) 6= f (1). Let

fe(z) := f (z) + (f (0) f (1))z.

Then fe is analytic on U and fe(0) = fe(1) = f (0). It is easy to check that the
function id : z 7 z has Fourier coefficients c0 (id) = 21 , cn (id) = 1/2in for n 6= 0.
In particular, cn (id) = cn (id) for n 6= 0. Consequently,
N N N
!
X X
X
lim cn (f ) = lim cn (fe) + f (1) f (0) cn (id)
N N
n=N n=N n=N
1 1



= f (0) + 2
f (1) f (0) = 2
f (0) + f (1) .

127
This completes our proof.

Theorem 9.2 (Poissons summation formula for finite sums). Let a, b be integers
with a < b and let f be a complex analytic function, defined on an open set containing
the interval [a, b]. Then
b
X N
X Z b
1
f (t)e2int dt


f (m) = 2
f (a) + f (b) + lim
N a
m=a n=N
Z b Z
X b
1



= 2
f (a) + f (b) + f (t)dt + 2 f (t) cos 2nt dt.
a n=1 a

Proof. Pick m {a, . . . , b 1}. Then by Theorem 9.1,

N
X Z m+1
1
f (t)e2int dt


2
f (m) + f (m + 1) = lim
N m
n=N

Z m+1 N Z
X m+1
f (t) e2int + e2int dt


= f (t)dt + lim
m N m
n=1
Z m+1 Z
X m+1
= f (t)dt + 2 f (t) cos 2nt dt.
m n=1 m

Now take the sum over m = a, a + 1, . . . , b 1.

P
We need a variation on Theorem 9.2, dealing with infinite sums m= f (m).

Theorem 9.3. Let f be a complex function such that:

(i) f (z) is analytic on U () := {z C : |Im z| < } for some > 0;
(ii) there are C > 0, > 0 such that

|f (z)| 6 C (|z| + 1)1 for z U ().

Then

X N
X Z
f (n) = lim f (t)e2int dt.
N
n= n=N

P
The idea is to apply Theorem 9.1 to the function F (z) := m= f (z + m). We
first prove some properties of this function.

128
Lemma 9.4. (i) F (0) = F (1) =
P
m= f (m).
(ii) The function F (z) is analytic on an open set containing [0, 1].
R1 R
(iii) For every n Z we have 0 F (t)e2int dt = f (t)e2int dt.

Proof. (i) Obvious.

(ii) Let U := {z C : < Re z < 1 + , |Im z| < }. Assuming that
is sufficiently small, we have |f (z + m)| 6 C(|m| )1 =: Am for z U ,
m Z \ {0}. All summands f (z + m) are analytic on U 0 , and the series m6=0 Am
P

converges. So by Corollary 2.26, the function F (z) is analytic on U .

(iii) Since |f (t + m)e2int | 6 Am for t [0, 1], m Z \ {0}, and m6=0 Am
P

converges, the series 2int

P
m= f (t + m)e converges uniformly on [0, 1]. Therefore,
we may interchange the integral and the infinite sum, and obtain
Z 1 Z 1
 X 
X Z 1
2int 2int
F (t)e dt = f (t + m) e dt = f (t + m)e2int dt
0 0 m= m= 0

X
X Z m+1
2in(t+m)
= f (t + m)e dt = f (t)e2int dt
m= m= m
Z
= f (t)e2int dt.

R
In the last step we have used that the integral f (t)e2int dt converges, due to
our assumption |f (z)| 6 C(|z| + 1)1 for z U ().

Proof of Theorem 9.3. By combining Theorem 9.1 with Lemma 9.4 we obtain

X N
X Z 1
1
F (t)e2int dt


f (m) = 2
F (0) + F (1) = lim
N 0
m= n=N

N
X Z
= lim f (t)e2int dt.
N
n=N

129
9.2 A functional equation for the theta function
The Jacobi theta function is given by

2z
X
(z) := em (z C, Re z > 0).
m=

Verify yourself that (z) converges and is analytic on {z C : Re z > 0}.

Theorem 9.5. (z 1 ) = z (z) for z C, Re z > 0, where z is chosen such

that | arg z| < 4 .

Remark. Let A := {z C : Re z > 0}. We may choose the argument of z A

such that | arg z| < 2 . Then indeed, we may choose z such that | arg z| < 4 .

Proof. Both (z 1 ) and z(z) are analytic on A. Hence it suffices to prove the
identity in Theorem 9.5 on a subset of A having a limit point in A. For this subset
we take R>0 . Thus, it suffices to prove that

X 2 /x X 2x
em = x em for x > 0.
m= m=

2
We apply Theorem 9.3 to f (z) := ez /x with x > 0 fixed. Verify that f satisfies
all conditions of that Theorem. Thus, for any x > 0,
N Z
m2 /x 2 /x)2int
X X
e = lim e(t dt.
N
m= n=N

We compute the integrals by substituting u = t x. Thus,
Z Z
(t2 /x)2int
2
e dt = x eu 2in xu du

Z
x)2 n2 x
= x e(u+in du

Z
n2 x x)2
= xe e(u+in du.

130
In the lemma below we prove that the last integral is equal to 1. Then it follows
that
N
X
m2 /x
X n2 x X 2
e = lim xe = x en x ,
N
m= n=N n=

since the last series converges. This proves our Theorem.

R 2
Lemma 9.6. Let z C. Then
e(u+z) du = 1.

Proof. The following proof was suggested to me by Michiel Kosters. Let

Z
2
F (z) := e(u+z) du.

We show that this defines an analytic function on C. To this end, we prove that F is
analytic on D(0, R) := {z C : |z| < R} for every R > 0. We apply Theorem 2.29.
2
First, (u, z) 7 e(u+z) is continuous, hence measurable, on R D(0, R). Second,
2
for every fixed u R, z 7 e(u+z) is analytic on D(0, R). Third,

2 2 2 +2uRe z+Re z 2 )
|e(u+z) | = eRe (u+z) = e(u
2 +2Ru+R2 2 +2R2
6 eu = e(uR) ,
R 2 +2R2
and
e(uR) du converges. So by Theorem 2.29, F is analytic on D(0, R).
Knowing that F is analytic on C, in order to prove that F (z) = 1 for z C it
is sufficient to prove, for any set S C with a limit point in C, that F (z) = 1 for
z S. For the set S we take R. For z R we obtain, by substituting v = u + z,
Z Z Z
(u+z)2 v 2 2
F (z) = e du = e dv = 2 ev dv.
0

Now a second substitution t = v 2 yields

Z
1/2
F (z) = et t1/2 dt = 1/2 ( 21 ) = 1.
0

131
9.3 The functional equation for the Riemann zeta
function
Put
(s) := 12 s(s 1) s/2 ( 12 s)(s) = (s 1) s/2 ( 21 s + 1)(s),

where we have used the identity 21 s( 12 s) = ( 21 s + 1).

Theorem 9.7. The function has an analytic continuation to C.

For this continuation we have (1 s) = (s) for s C.

Before proving this, we deduce some consequences.

Corollary 9.8. The function has an analytic continuation to C\{1} with a simple
pole with residue 1 at s = 1.
For this continuation we have

(1 s) = 21s s cos( 21 s)(s) (s) for s C \ {0, 1}.

Proof. We define the analytic continuation of by

(s) s/2 1/( 21 s + 1)

(s) = .
s1

By Corollary 8.5, 1/ is analytic on C, and the other functions in the numerator

are also analytic on C. Hence is analytic on C \ {1}. The analytic continuation
of defined here coincides with the one defined in Theorem 5.2 on {s C : Re s >
0} \ {1} since analytic continuations to connected sets are uniquely determined.
Hence (s) has a simple pole with residue 1 at s = 1.
We derive the functional equation. By Theorem 9.7 we have, for s C \ {0, 1},

(1 s) (s)
(1 s) = 1 1 = 1
2
(1 s)(s) (1s)/2 ( 2 (1 s)) 2
s(s 1) (1s)/2 ( 21 (1 s))
1
2
1) s/2 ( 12 s)
s(s
= 1 (s) = F (s)(s),
2
s(s 1) (1s)/2 ( 21 (1 s))

132
say. Now we have

(1/2)s ( 12 s)( 21 s + 12 )
F (s) = 1 1
( 2 2 s)( 12 + 21 s)

(1/2)s 21s (s)
= (by Corollary 8.12, Theorem 8.3)
/ sin(( 12 12 s))
= s 21s cos( 12 s)(s).
This implies Corollary 9.8.
Corollary 9.9. has simple zeros at s = 2, 4, 6, . . ..
has no other zeros outside the critical strip {s C : 0 < Re s < 1}.

Proof. We first show that (s) 6= 0 if Re s > 1 or Re s 6 0. We use the second

expression for (s). By Corollary 5.4 and Theorem 4.5, we know that (s) 6= 0 for
s C with Re s > 1, s 6= 1. Further, lims1 (s 1)(s) = 1, hence (s 1)(s) 6= 0 if
Re s > 1. By Corollary 8.5, we know that ( 21 s + 1) 6= 0 if Re s > 1. hence (s) 6= 0
if Re s > 1. But then by Theorem 9.7, (s) 6= 0 if Re s 6 0.
We consider (s) for Re s 6 0. For s 6= 2, 4, 6, . . ., the function ( 12 s + 1) is
analytic. Further, for these values of s, we have (s) 6= 0, hence (s) must be 6= 0
as well. The function ( 21 s) has simple poles at s = 2, 4, 6, . . .. To make (s)
analytic and non-zero for these values of s, the function must have simple zeros
at s = 2, 4, 6, . . ..

Proof of Theorem 9.7 (Riemann). Let for the moment, s C, Re s > 1. Recall that
Z
1
( 2 s) = et t(s/2)1 dt.
0
2
Substituting t = n u gives
Z Z
n2 u 2
1
( 2 s) = e 2 (s/2)1
(n u) 2 s/2 s
d(n u) = n en u u(s/2)1 du.
0 0

Hence Z
s/2 2
( 12 s)ns = en u u(s/2)1 du,
0
and so, by summing over n,
Z
2u
X
s/2
( 12 s)(s) = en u(s/2)1 du.
n=1 0

133
We justify that the infinite integral and infinite sum can be interchanged. We use
the following special case of the Fubini-Tonelli theorem: if {fn : (0, ) C} n=1
R
is a sequence of measurable functions such that
P
n=1 P0
|f n (u)|du converges, then
R
all integrals 0 fn (u)du (n > 1) converge, the series n=1 fn (u) converges almost
everywhere on (0, ) and moreover,

Z
!
X Z X
fn (u)du, fn (u) du
n=1 0 0 n=1

converge and are equal. In our situation we have that indeed (putting := Re s)

Z Z
n2 u 2
X X
|e u (s/2)1
|du = en u u(s/2)1 du
n=1 0 n=1 0

X
= /2 ( 21 )n (reversing the above argument)
n=1

= /2 ( 21 )()

converges. Thus, we conclude that for s C with Re s > 1,

Z
2
X
s/2
(9.1) ( 12 s)(s) = (u) u(s/2)1 du, where (u) = en u .
0 n=1

P 2u
Recall that (u) = n= en = 1 + 2(u).
We want to replace the right-hand side of (9.1) by something that converges for
every s C. Obviously, for s C with Re s < 0 there are problems if u 0. To
R R R1 R1
overcome these, we split the integral 0 into 1 + 0 and then transform 0 into
R
an integral 1 by means of a substitution v = u1 . After this substitution, the
integral contains a term (v 1 ). By Theorem 9.5, we have

(v 1 ) = 1
2
((v 1 ) 1) = 12 v 1/2 (v) 21
1 1/2
= 2
v 2(v) + 1) 12 = v 1/2 (v) + 21 v 1/2 12 .

We work out in detail the approach sketched above. We keep for the moment our

134
assumption Re s > 1. Thus,
Z Z
1
s 1
2 ( 2 s)(s) = (u)u(s/2)1
du (v 1 )v 1s/2 dv 1
1 1
Z Z
1s/2 2
= (u)u(s/2)1 du + v 1/2 (v) + 21 v 1/2 1
2
v v dv
1 1
Z Z
(s+1)/2 (s/2)1
1
(v) v (s/2)1 + v (s+1)/2 dv



= 2
v v dv +
1 1

where we have combined the terms without into one integral, and the terms
involving into another integral. Since we are still assuming Re s > 1, the first
integral is equal to
 
1 2 (s1)/2 2 s/2 1 1 1
2
v + v = = .
s1 s 1 s1 s s(s 1)

Hence
Z
s/2 1
( 12 s)(s) (v) v (s/2)1 + v (s+1)/2 dv.


= +
s(s 1) 1

For our function (s) = 21 s(s 1) s/2 ( 21 s)(s) this gives

Z
1 1
(v) v (s/2)1 + v (s+1)/2 dv if Re s > 1.


(9.2) (s) = 2 + 2 s(s 1)
1
R
Assume for the moment that F (s) := 1 (v) v (s/2)1 + v (s+1)/2 dv defines an

analytic function on C. Then we can use the right-hand side of (9.2) to define the
analytic continuation of (s) to C. By substituting 1 s for s in the right-hand side,
we see that (1 s) = (s).
It remains to prove that F (s) defines an analytic function on C. To this end,
it suffices to prove that F (s) is analytic on UA := {s C : |Re s| < A} for every
A > 0.
We apply as usual Theorem 2.29. We check that f (v, s) = (v) v (s/2)1 +
v (s+1)/2 satisfies the conditions of that theorem.

a) f (v, s) is measurable on (1, ) UA . For (v) = n2 v

P
n=1 e is measurable,
being a pointwise convergent series of continuous, hence measurable functions, and
also v (s/2)1 + v (s+1)/2 is measurable, since it is continuous.

135
 b) s 7 (v) v (s/2)1 + v (s+1)/2 is analytic on UA for every fixed v. This is

obvious.
c) There is a measurable function M (v) on (1, ) such that |f (v, s)| 6 M (v) for
R
s UA and 1 M (v)dv < . Indeed, we first have for v (1, )

0 6 (v) 6 ev 1 + e3v + e8v 6 2ev

and second, for v (1, ), s UA

|v (s/2)1 + v (s+1)/2 | 6 v (A/2)1 + v ((A+1)/2 6 2v (A/2)1 .

Hence
|f (v, s)| 6 4ev v (A/2)1 =: M (v).
Further, Z Z
M (v)dv 6 4 ev v (A/2)1) dv 6 4 ( 21 A) < .
1 0

So f (v, s) satisfies all conditions of Theorem 2.29, and it follows that F (s) =
R
1
f (v, s)dv is analytic on UA .

9.4 The functional equations for L-functions

(q)
Let q be an integer > 2 and a Dirichlet character modulo q with 6= 0 . We
give, without proof, a functional equation for L(s, ) in the case that is primitive,
i.e., that it is not induced by a character modulo d for any proper divisor d of q.
Notice that for any character modulo q we have (1)2 = (1) = 1, hence
(1) {1, 1}. A character is called even if (1) = 1, and odd if (1) = 1.
There will be different functional equations for even and odd characters.
In Chapter 4 we defined the Gauss sum related to a character mod q by
q1
X
(1, ) = (a)e2ia/q .
a=0

According to Theorem 4.17, if is primitive then | (1, )| = q.
By we denote the complex conjugate of a character .

136
Theorem 9.10. Let q be an integer with q > 2, and a primitive character mod q.
Put

 q s/2
1 q
(s, ) := ( 2 s)L(s, ), c() := if is even,
(1, )

 q (s+1)/2
1

i q
(s, ) := 2 (s + 1) L(s, ), c() := if is odd.
(1, )

Then (s, ) has an analytic continuation to C, and

(1 s, ) = c()(s, ) for s C.

Remark. We know that |c()| = 1. In general, it is a difficult problem to compute

c().
The proof of Theorem 9.10 is similar to that of that of the functional equation
for L(s, ), but with some additional technicalities, see H. Davenport, Multiplicative
Number Theory, Chapter 9.
We deduce some consequences.

(q)
Corollary 9.11. Let q be an integer > 2 and a character mod q with 6= 0 .
Then L(s, ) has an analytic continuation to C.

Proof. First assume that is primitive and is even. Then

L(s, ) = (s, )(/q)s/2 /( 21 s).

The functions (s, ) and (/q)s/2 are both analytic on C, and according to Corollary
7.5, 1/( 21 s) is analytic on C as well. Hence L(s, ) is analytic on C.
In a completely similar manner one shows that L(s, ) is analytic on C if is
primitive and odd.
Now suppose that is not primitive. Let q 0 be the conductor of . By Corollary
4.13, is induced by a character 0 mod q 0 . Verify yourself that 0 is primitive. We
(q)
have q 0 > 1, since otherwise, would be equal to 0 .
For s C with Re s > 1 we have, noting that (p) = 0 (p) if p is a prime not

137
dividing q and (p) = 0 if p is a prime dividing q,
Y 1 Y 1
L(s, ) = s
=
p
1 (p)p 1 (p)ps
0
p-q

Y 1 Y
0 s

0
Y
0 s


= 1 (p)p = L(s, ) 1 (p)p .
p
1 0 (p)ps
p|q p|q

Now clearly, we can extend L(s, ) analytically to C by defining

Y
L(s, ) = L(s, 0 ) 1 0 (p)ps for s C.


(9.3)
p|q

We consider the zeros of L-functions. Notice that (9.3) implies that if is induced
by a primitive character 0 , then L(s, ) has the same set of zeros as L(s, ), except
for possible zeros of p|q 1 0 (p)ps , which all lie on the line Re s = 0.
Q

We consider henceforth only the zeros of L(s, ) for primitive characters . We

have proved in Chapter 5 that L(s, ) 6= 0 if Re s > 1. The next corollary considers
the zeros s with Re s 6 0.

Corollary 9.12. Let q be an integer > 2 and a primitive character mod q.

(i) If is even, then L(s, ) has simple zeros at s = 0, 2, 4, . . . and L(s, ) 6= 0
if Re s 6 0, s 6 {0, 2, 4, . . .}.
(ii) If is odd, then L(s, ) has simple zeros at s = 1, 3, 5, . . . and L(s, ) 6= 0
if Re s 6 0, s 6 {1, 3, 5, . . .}.

Proof. Exercise.

138