On the Universality

of the Riemann zeta-function

Jorn Steuding
Summer 2002
Johann Wolfgang Goethe-Universit at Frankfurt
1
This course gives an introduction to the value distribution theory of the Riemann
zeta-function (s) with a special emphasis on its remarkable universality property. Our
main goal is to derive Voronins celebrated universality theorem which states, roughly
speaking, that any(!) non-vanishing analytic function on a suciently small disc can
be uniformly approximated by (s) . For that aim we go along Voronins original proof;
the modern approach via limit theorems for weak convergent probability measures is
beyond the scope of this course. We state some consequences as there is the classical
result due to H. Bohr that the set of values of (s) taken on vertical lines in the right
half of the critical strip lies dense in the complex plane, and an answer to Hilberts
question whether the zeta-function satises an algebraic dierential equation. Finally,
we shall discuss some open questions in this eld (the problem of eectivization and
Bagchis reformulation of Riemanns hypothesis) and the recent progress done towards
their solution.
This course bases mainly on the nicely written monography [25] on the Riemann
zeta-function. However, we also have to use some results from other elds dierent than
analytic number theory, namely approximation theory of numbers and of functions, the
theory of linear operators in Hilbert spaces and complex analysis, for which we refer to
[14] and [55], respectively. Unfortunately, we cannot give a detailed proof of all facts
which are required to prove Voronins theorem (as for example the seperation theorem
from functional analysis). However,the motivated reader can ll by the given references
all these gaps.
I am very grateful to Rasa

Slezevicien e, Ramunas Garunkstis and Antanas Lau-
rincikas (Vilnius University) for many fruitful discussions, helpful remarks and unlim-
ited interest.
Jorn Steuding, Frankfurt 10/20/2002.
Contents
1 Introduction 1
2 The Riemann zeta-function 4
3 The approximate functional equation 7
4 Density theorems 14
5 A zero-free region 22
6 The prime number theorem 24
7 Diophantine approximation 29
8 Conditionally convergent series 34
9 Finite Euler products 38
10 Voronins universality theorem 46
11 Functional independence 51
12 Self-similarity and the Riemann hypothesis 54
13 Eective bounds 59
14 Other zeta-functions 62
Bibliography 64
1 Introduction
In this introduction we give a brief historical overview on the phenomenon of univer-
sality; more details can be found in [21].
The rst in the mathematical literature appearing universal object was discov-
ered by Fekete in 1914/15 (see [42]); he proved that there exists a real power series

n=1
a
n
x
n
with the property that for any continuous function on the interval [1, 1]
with f(0) = 0 there is a sequence of positive integers (m
k
) such that the partial sums

1nm
k
a
n
x
n
converge to f(x) uniformly on [1, 1] . We shall prove the following
variant due to Luh [38]:
Theorem 1.1 There exists a real power series

n=0
a
n
x
n
with the property that for any interval [a, b] with 0 , [a, b] and any continuous function
f(x) on [a, b] there exists a sequence of positive integers (m
k
) such that
m
k

n=0
a
n
x
n
f(x) uniformly on [a, b]
as k .
The proof relies mainly on Weierstrass approximation theorem, which states that each
continuous function f(x) on a closed interval is the limit of a uniformly convergent
sequence of polynomials; for a proof see [14].
Proof. Denote by (Q
n
) the sequence of polynomials with rational coecients; this is
obviously a countable set. We construct a sequence of polynomials (P
n
) as follows: let
P
0
= Q
0
, and assume that for n N the polynomials P
0
, . . . , P
n1
are known. Let d
n
be the degree of P
n1
. Further, let
n
(x) be a continuous function on [n, n] such
that

n
(x) =
_
Q
n
(x)
n1

=0
P

(x)
_
x
dn1
for x J
n
:=
_
n,
1
n
_

_
1
n
, n
_
.
Then, by Weierstrass approximation theorem, there exists a polynomial F
n
with
max
x[n,n]
[F
n
(x)
n
(x)[ < n
dn2
.
Setting P
n
(x) = F
n
(x)x
dn+1
, the sequence (P
n
) is constructed, and P
n
satises
max
xIn

k=0
P
k
(x) Q
n
(x)

= max
xIn
[(F
n
(x)
n
(x))x
dn+1
[ <
1
n
.
1
Obviously, distinct P
n
have no powers in common. Thus we can rearrange formally
the polynomial series into a power series:

n=0
a
n
x
n
:=

n=0
P
n
(x)
(note that this would be impossible if innitely many P
n
would have a term x
n
in
common).
Again by Weiertstrass approximation theorem, to any continuous function f(x)
on [a, b] there exists a sequence of positive integers (n
k
) , tending to innity, such that
max
x[1,1]
[f(x) Q
n
k
(x)[ <
1
k
.
For suciently large k the interval [a, b] is contained in J
n
k
. In view of the above
estimates we obtain
max
x[a,b]

f(x)
dn
k
+1

n=0
a
n
x
n

= max
x[a,b]

f(x)
n
k

n=0
P
n
(x)

max
x[a,b]
[f(x) Q
n
k
(x)[ + max
x[a,b]

Q
n
k
(x)
n
k

n=0
P
n
(x)

<
1
k
+
1
n
k
,
which tends to zero as k . Thus, putting m
k
= d
n
k
+ 1, the assertion of the
theorem follows.
In the years after Feketes discovery many of such universal approximations were
found. Here we have to mention Birkho [6] who proved the existence of an entire
function f(z) with the property that to any given entire function g(z) there exists a
sequence (a
n
) such that
f(z +a
n
) g(z) locally uniformly in C.
It was Marcinkiewicz [39] in 1935 who was the rst to use the notion universality
when he proved the existence of a continuous function whose dierence quotients can
approximate any measurable function in the sense of convergence almost everywhere.
In all these examples there are two characteristic aspects of universality, namely
the existence of a single object which
is maximal divergent, and
(via a countable process) allows to approximate a maximal class of objects.
This observation led to understand universality as a phenomenon which occurs quite
naturally in limiting processes which diverge or behave irregularly in some cases. Mean-
while it turned out that the phenomenon of universality is anything but a rare event
in analysis! See [21] for an interesting survey on more or less all known types of uni-
versalities and an approach to unify them all.
2
However, for a long time no explicit example of a universal object was found. Sur-
prisingly, Voronin discovered in 1975 that a very famous function has a remarkable
universal property.
Let s = + it with i
2
= 1 and , t R be a complex variable (this mixture
of latin and greek letters is tradition in number theory). Then the Riemann zeta-
function is given by the series
(s) =

n=1
1
n
s
for > 1. (1.1)
Since [n
s
[ = n

, it follows easily (by the integral test) that the series converges ab-
solutely in the half-plane > 1. Series of the type

n
a(n)
n
s
are called Dirichlet series;
see [54], Section IX for details. However, the true analytic character of the Riemann
zeta-function becomes only visible by continuing (s) to the left of = 1. Later we
will prove the identity
(s) =
s
s 1
+s
_

1
[x] x
x
s+1
dx for > 0; (1.2)
here [x] := maxz Z : z x is the integer part function. Since the appearing
integral converges, we have found an analytic continuation for (s) to the half-plane
> 0 except for a simple pole at s = 1 with residue 1. As we shall see later on,
the value distribution of the Riemann zeta-function in the so-called critical strip
0 1 is of special interest out of dierent points of view.
Now, since we have an expression for (s) in the critical strip, we are in the position
to formulate Voronins result [57].
Theorem 1.2 Let 0 < r <
1
4
and let f(s) be a non-vanishing continuous function on
the disc [s[ r , which is analytic in the interior. Then, for any > 0 there exists a
> 0 such that
max
|s|r

_
3
4
+s +i
_
f(s)

< .
Thus, (s) approximates any such function with any precision somewhere in
1
2
< <
1. Interpreting the absolute value of an analytic funtion as an analytic landscape over
the complex plane, we see that any analytic landscape can be found in the analytic
landscape of (s) (up to an arbitrarily small error). Therefore, in German we say:
Wer die Zetafunktion kennt, kennt die Welt!
However, we do not know the zeta-function good enough; we even have not proved the
analytic continuation (1.2).
3
2 The Riemann zeta-function
In this section we shall prove rst properties of (s) which also give a rst impression
on the leading role of the zeta-function in multiplicative number theory.
In the introduction we have seen that the series in (1.1) converges absolutely for
> 1. Since, for
0
> 1,

n=1
1
n
s

n=1
1
n

0
1 +

n=2
_
n
n1
du
u

0
(2.1)
= 1 +
_

1
u

0
du = 1 +
1

0
1
,
the series in question converges uniformly in any half-plane >
0
with
0
> 1.
A well-known theorem of Weierstrass states that the limit of a uniformly convergent
sequence of holomorphic functions is holomorphic; see [54], 2.8. Thus
Theorem 2.1 (s) is analytic for > 1 and satises in this half-plane the identity
(s) =

n=1
1
n
s
=

p
_
1
1
p
s
_
1
. (2.2)
Here and in the sequel p denotes always a prime number. The product in (2.2) is
taken over all primes; it is called Euler product since it was discovered by Euler in
1737. As we shall see in the proof below it can be regarded as the analytic version of
the unique prime factorization of the integers.
Proof. It remains to show the identity between the series and the product in (2.2).
In view of the geometric series expansion and the unique prime factorization of the
integers,

px
_
1
1
p
s
_
1
=

px
_
1 +
1
p
s
+
1
p
2s
+. . .
_
=

n
p|npx
1
n
s
;
as usual, we write d[n if the integer d divides the integer n, and d,[n otherwise. Since

n=1
1
n
s

n
p|npx
1
n
s

n>x
1
n

_

x
u

du =
x
1
1
tends to zero as x , we get identity (2.2) by sending x . The theorem is
proved.
We shall see later on that the Euler product (2.2) is the key to prove Voronins
universality theorem (in spite of the fact that the product does not converge in the
region of universality). However, the representation (2.2) gives also a rst glance on
the close connection between (s) and the distribution of prime numbers. A rst im-
mediate consequence is Eulers proof of the innitude of the prime numbers: assuming
4
that there are only nitely many primes, the product in (2.2) is nite, and therefore
convergent throughout the whole complex plane, contradicting the fact that the (s)
dening Dirichlet series reduces to the divergent harmonic series as s 1+. Hence,
there are innitely many prime numbers. This fact is well known since Euclids elemen-
tary proof, but the analytic access gives much deeper knowledge on the distribution of
prime numbers.
It was the young Gauss who conjectured in 1792 for the number of primes under
a given magnitide:
(x) := p x Li(x), (2.3)
where the logarithmic integral is dened by
Li(x) =
_
x
2
du
log u
;
here and in the sequel we write f(x) = O(g(x)) , resp. f(x) g(x) , with a positive
function g(x) if
limsup
x
[f(x)[
g(x)
is bounded, and we write f(x) g(x) if the latter quantity is actually a limit and
equals one. By partial integration it is easily seen that
Li(x) =
_
x
2
du
log u
=
n

k=1
x(k 1)!
(log x)
k
+O
_
x
(log x)
n+1
_
.
Thus, Gauss conjecture states that, in rst approximation, the number of primes x
is asymptotically
x
log x
.

Cebyshev proved in 1852 by elementary methods that for
suciently large x
0.921 . . . (x)
log x
x
1.055 . . . ,
and additionally that if
lim
x
(x)
log x
x
exists, then the limit has to be equal to one, which supports (2.3).
However, as Riemann showed Eulers analytic access, i.e. the link between the
zeta-function and the primes by (2.2), encodes much more arithmetic information than

Chebyshevs elemenatry approach. In his only but outstanding paper [49] on number
theory Riemann noticed the importance of studying (s) as a function of a complex
variable (Euler dealt only with real s). Riemann proved:
the function
(s)
1
s 1
is entire; consequently, (s) has an anlytic continuation throughout the whole
complex plane except for s = 1 where (s) has a simple pole with residue 1.
5
(s) satises the functional equation

s
2

_
s
2
_
(s) =

1s
2

_
1 s
2
_
(1 s).
Note that the Gamma-function, dened by
(z) =
_

0
u
z1
exp(u) du for Re z > 0,
plays an important role in the theory of the zeta-function; see [54], 1.86 and 4.41,
for a collection of its most important properties.
In view of the Euler product (2.2) it is easily seen that (s) has no zeros in the
half-plane > 1. Using the functional equation, it turns out that (s) vanishes in
< 0 exactly at the so-called trivial zeros
(2n) = 0 for n N;
this is caused by the simple poles of the Gamma-function at the non-positive integers.
All other zeros of (s) are said to be nontrivial, and it comes out that they are all
non-real, and that there location is in fact a nontrivial task. We denote the nontrivial
zeros by = + i . Obviously, they have to lie in the critical strip 0 1 The
functional equation, in addition with the reection principle
(s) = (s), (2.4)
show some symmetries of (s) . In particular, the nontrivial zeros of (s) have to be
distributed symmetrically with respect to the real axis and the so-called critical line
=
1
2
. It was Riemanns ingenious contribution to number theory to point out how
the distribtuion of these nontrivial zeros is linked to the distribution of prime numbers.
Riemann conjectured:
the number N(T) of nontrivial zeros = + i with 0 T (counted
according multiplicities) satises
N(T)
T
2
log
T
2e
;
this was proved in 1895/1905 by von Mangoldt, who found more precisely
N(T) =
T
2
log
T
2e
+O(log T). (2.5)
all nontrivial zeros lie on the critical line =
1
2
, or equivalently,
(s) ,= 0 for >
1
2
; (2.6)
6
this is the famous, yet unproved Riemann hypothesis. Riemann worked with
(
1
2
+ it) and wrote ...und es ist sehr wahrscheinlich, dass alle Wurzeln reell sind.
Hiervon w are allerdings ein strenger Beweis zu w unschen; ich habe indess die Auf-
suchung desselben nach einigen uchtigen vergeblichen Versuchen vorlaug bei Seite
gelassen.... Note that Riemann also calculated the rst zeros: for example, the rst
is =
1
2
+i 14.134 . . . . Further, Riemann conjectured
there exist some constants A, B such that
1
2
s(s 1)

s
2

_
s
2
_
(s) = exp(A+Bs)

_
1
s

_
exp
_
s

_
;
the explicit formula: for any x 2
(x) +

n=2
(x
1/n
)
n
= Li(x)

=+i
>0
_
Li(x

) + Li(x
1
)
_
(2.7)
+
_

x
du
u(u
2
1) log u
log 2;
this was proved in 1895 by von Mangoldt whereas the last but one conjecture was
proved by Hadamard. The explicit formula follows from both product representations
of (s) , the Euler product on one side and the Hadamard product on the other side.
Riemanns ideas led to the rst proof of Gauss conjecture (2.3), the celebrated
prime number theorem, by Hadamard and de La Vallee Poussin (independendly) in
1896.
3 The approximate functional equation
In this section we shall derive not only an analytic continuation for (s) to the half-
plane > 0 but also a quite good approximation which will be very useful later on.
Unfortunately, the proof is rather technical; we follow [55], IV.
What happens on the abscissa of convergence = 1? At s = 1 the zeta-function
dening series reduces to the harmonic series. To obtain an analytic continuation for
(s) we have to seperate this singularity. For that purpose we need Abels partial
summation:
Lemma 3.1 Let
1
<
2
< . . . be a divergent sequence of real numbers, dene for

n
C the function A(u) =

nu

n
, and let F : [
1
, ) C be a continuous
dierentiable function. Then

nx

n
F(
n
) = A(x)F(x)
_
x

1
A(u)F

(u) du.
7
For those who are familiar with the Riemann-Stieltjes integral there is nearly nothing
to show. Anyway,
Proof. We have
A(x)F(x)

nx

n
F(
n
) =

nx

n
(F(x) F(
n
)) =

nx
_
x
n

n
F

(u) du.
Since
1

n
u x, interchanging integration and summation yields the assertion.

Now we apply partial summation to nite pieces of the series (1.1). Let N < M be
positive integers and > 1. Then, application of Lemma 3.1 with F(u) = u
s
,
n
= 1
and
n
= n yields

N<nM
1
n
s
= M
1s
N
1s
+s
_
M
N
[u]
u
s+1
du
= M
1s
N
1s
+s
_
M
N
[u] u
u
s+1
du +s
_
M
N
u
s
du
=
1
s 1
(N
1s
M
1s
) +s
_
M
N
[u] u
u
s+1
du.
Sending M we obtain
Theorem 3.2 For > 0,
(s) =

nN
1
n
s
+
N
1s
s 1
+s
_

N
[u] u
u
s+1
du.
Thus, (s) has an analytic continuation to the half-plane > 0 except for one simple
pole at s = 1 with residue 1.
Putting N = 1 in the formula of Theorem 3.2, we obtain the analytic continuation
(1.2) for (s) from the introduction. But this integral representation gives also a very
useful approximation of (s) in the critical strip. By the periodicity of the function
[u] u we expect that the contribution of the integral is small. In order to give a
rigorous proof of this idea we have to do some preliminary observations.
Let f(u) be any function with continuous derivative on the interval [a, b] . Using
the lemma on partial summation with
n
= 1 if n (a, b] , and
n
= 0 otherwise, we
get

a<nb
f(n) = ([b] [a])f(b)
_
b
a
([u] [a])f

(u) du = [b]f(b) [a]f(a)

_
b
a
[u]f

(u) du.
Obviously,

_
b
a
[u]f

(u) du =
_
b
a
_
u [u]
1
2
_
f

(u) du
_
b
a
_
u
1
2
_
f

(u) du.
Applying partial integration to the last integral on the right hand side, we deduce
8
Lemma 3.3 (Eulers summation formula) Assume that f : [a, b] R has a con-
tinuous derivative. Then

a<nb
f(n) =
_
b
a
f(u) du +
_
b
a
_
u [u]
1
2
_
f

(u) du
+
_
a [a]
1
2
_
f(a)
_
b [b]
1
2
_
f(b).
An easy application is the well-known asymptotic formula

nx
1
n
= log n + +O
_
1
x
_
as x , (3.1)
where = 0.577 . . . is the Euler-Mascheroni constant. This formula describes very
precisely the rate of divergence of the harmonic series. Our application is more dicult.
First, we replace in Eulers summation formula the function u [u]
1
2
by its Fourier
series expansion.
Lemma 3.4 For u , Z,

u
1
2

|m|M
m=0
exp(2imu)
2im

1
2M(u [u])
,
and, for u R,

m=0
exp(2imu)
2im
=
_
u [u]
1
2
if u , Z,
0 if u Z,
where the terms with m have to be added together; the partial sums are uniformly
bounded in and M .
Proof. By symmetry and periodicity it suces to consider only the case 0 < u
1
2
.
Since
_ 1
2
u
exp(2imx) dx =
(1)
m+1
+ exp(2imu)
2im
for 0 ,= m Z,
we obtain

|m|M
m=0
exp(2imu)
2im
u +
1
2
=
_ 1
2
u

|m|M
exp(2imx) dx
=
_ 1
2
u
sin((2M + 1)x)
sin(x)
dx. (3.2)
By the mean-value theorem there exists a (u,
1
2
) such that the latter integral equals
_

u
sin((2M + 1)x)
sin(u)
dx.
9
This implies immediately both formulas of the lemma. It remains to show that the
partial sums of the Fourier series are uniformly bounded in u and M. Substituting
y = (2M + 1)x in (3.2), we get
_ 1
2
u
sin((2M + 1)x)
sin(x)
dx =
_ 1
2
u
sin((2M + 1)x)
x
dx
+
_ 1
2
u
sin((2M + 1)x)
_
1
sin(x)

1
x
_
dx

_

0
sin(y)
y
dy +
_ 1
2
0

1
sin(x)

1
x

dx
with an implicit constant not depending on and M ; obviously both integrals exist,
which gives the uniform boundedness.
Further, we will make use of the following estimate of exponential integrals.
Lemma 3.5 Assume that F : [a, b] R has a continuous non-vanishing derivative
and that G : [a, b] R is continuous. If
G
F

is monotonic on [a, b] , then

_
b
a
G(u) exp(iF(u)) du

G
F

(a)

+ 4

G
F

(b)

.
Proof. First, we assume that F

(u) > 0 for a u b. Since (F

1
(v))

=
F

(F
1
(v))
1
, substituting u = F
1
(v) leads to
_
b
a
G(u) exp(iF(u)) du =
_
F(b)
F(a)
G(F
1
(v))
F

(F
1
(v))
exp(iv) dv.
Application of the mean-value theorem gives, in case of a monotonically increasing
G
F

,
Re
_
_
F(b)
F(a)
G(F
1
(v))
F

(F
1
(v))
exp(iv) dv
_
=
G
F

(F(a))
_

F(a)
cos v dv +
G
F

(F(b))
_
F(b

cos v dv
with some (a, b) . This gives the desired estimate. The same idea applies to the
imaginary part. Furthermore, the case F

(u) < 0 can be treated analogously. The

lemma is proved.
Now we are in the position to prove van der Corputs summation formula.
Theorem 3.6 To any given > 0 there exists a positive constant C = C() only
depending on with the following property: assume that f : [a, b] R is a function
with continuous derivative, g : [a, b] [0, ) is dierentiable function, and that f

, g
and [g

[ are all monotically decreasing. Then

a<nb
g(n) exp(2if(n)) =

(a)<m<f

(b)+
_
b
a
g(u) exp(2i(f(u) mu)) du +S,
where
[S[ C() ([g

(a) +g(a) log([f

(a)[ +[f

(b)[ + 2)) .
10
Van der Corputs summation formula looks rather technical but the idea is simple as
we will shortly explain. The integral
_
b
a
g(u) exp(2i(f(u) mu)) du
is (up to a constant factor) the Fourier transform of g(u) exp(2if(u)) at u = m
(therefore, we may interpret Theorem 3.6 as an approximate version of Poissons sum-
mation formula).
Proof of Theorem 3.6. Using Eulers summation formula with F(u) =
g(u) exp(2if(u)) and the Fourier series expansion of Lemma 3.4, we get

a<nb
g(n) exp(2if(n)) =
_
b
a
g(u) exp(2if(u)) du +O(g(a))
+
_
b
a

m=0
exp(2imu)
2im
d
du
(g(u) exp(2if(u))) du.
Since the series on the right hand side converges uniformly on each compact subset,
which is free of integers, and since its partial sums are uniformly bounded, we may
interchange summation and integration. This yields

a<nb
g(n) exp(2if(n)) =
_
b
a
g(u) exp(2if(u)) du (3.3)
+

m=0
1
m
_
J
1
(m) +
1
2i
J
2
(m)
_
+O(g(a),
where
J
1
(m) :=
_
b
a
f

(u)g(u) exp(2i(f(u) mu)) du,

J
2
(m) :=
_
b
a
g

(u) exp(2i(f(u) mu)) du.

Partial integration gives
J
1
(m) =
_
exp(2i(f(u) mu))g(u)
2i
_
b
u=a

_
b
a
exp(2if(u))
2i
d
du
g(u) exp(2imu) du,
= O(g(a))
1
2i
J
2
(m) +m
_
b
a
g(u) exp(2i(f(u) mu)) du.
Thus,

(a)<m<f

(b)+
m=0
1
m
_
J
1
(m) +
1
2i
J
2
(m)
_
11
=

(a)<m<f

(b)+
m=0
_
b
a
g(u) exp(2i(f(u) hu)) du
+O
_
_
_
_

(a)<m<f

(b)+
m=0
g(a)
[m[
_
_
_
_
.
Now assume that m > f

(a) + and f

(b) > 0. Then f

(u) > 0 for a u b. Using

Lemma 3.5 with F(u) = 2(f(u) mu) and G = gf

, we nd
J
1
(m)

g(a)f

(a)
f

(a) m

.
Hence,

m>f

(a)+
m=0

J
1
(m)
m

g(a)

0<m2|f

(a)|
1
m
+g(a)

m>|f

(a)|
[f

(a)[
m
2
.
The contribution arising from m < f

(b) can be treated similarly. This gives

m[f

(b),f

(a)+]
m=0

J
1
(m)
m

g(a) log([f

(a)[ +[f

(b)[ + 2).
Now assume m > f

(a) + and m ,= 0. Then, by the mean-value theorem,

Re J
2
(m) =
_
b
a
[g

(u)[ cos 2(f(u) mu) du = g

(a)
_

a
cos 2(f(u) mu) du
with some (a, b) . Partial integration yields
_

a
cos 2(f(u) mu) du =
_
Re
exp(2i(f(u) mu)
2im
_

u=a
+Re
1
m
_

a
f

(u) exp(2i(f(u) mu)) du

1
[m[
_
1 +
[f

(a)[
[f

(a) m[
_
.
Therefore,

m>f

(a)+

Re J
2
(m)
m

(a).
With slight modiactions this method applies also to the cases Im J
2
(m) and m
f

(b) . Further, if 0 , [f

(b) , f

(a) +] , then Lemma 3.5 gives

_
b
a
g(u) exp(2if(u)) du g(a).
12
In view of (3.3) the theorem follows from the above estimates under the condition
f

(b) > 0. If this condition is not fullled, then argue with f(u) ku, where k :=
1 [f

(b)] , instead of f(u) .

Now we apply van der Corputs summation formula to the zeta-function. Let > 0.
By Theorem 3.2 we have
(s) =

nx
1
n
s
+

x<nN
exp(it log n)
n

+
N
1s
s 1
+s
_

N
[u] u
u
s+1
du.
Setting g(u) = u

and f(u) =
t
2
log u, we get f

(u) =
t
2u
. Assume that
[t[ 4x, then [f

(u)[
7
8
. With the choice =
1
10
the interval (f

(b) , f

(a) + )
contains only the integer m = 0. Thus Theorem 3.6 yields

x<nN
exp(it log n)
n

=
_
N
x
u
s
du +O(x

) =
N
1s
x
1s
1 s
+O(x

).
In addition with
s
_

N
[u] u
u
s+1
du [s[N

we deduce
Theorem 3.7 We have, uniformly for
0
> 0, [t[ 4x,
(s) =

nx
1
n
s
+
x
1s
s 1
+O
_
x

_
.
This so-valled approximate functional equation was found by Hardy and Little-
wood [22] in 1921 (the name comes from the appearing quantities s and 1 s as
in the functional equation) but was also known by Riemann (see Siegels paper [50]
on Riemanns unpublished papers on (s) ). Meanwhile there are better approximate
functional equations known, that means an approximation by shorter sums with a
smaller error term. However, to indicate the power of this simple approximation we
note

_
1
2
+it
_
t
1
2
as t (3.4)
for any > 0. Also here much better estimates known. For example, using the
functional equation and the Phragmen-Lindelof principle (that is a kind of maximum
principle for unbounded domains), one can obtain
1
4
+ instead of
1
2
for any positive
; Huxley [23] holds the record with the exponent
89
570
+. The yet unproved Lindel of
hypothesis states

_
1
2
+it
_
t

as t ; (3.5)
note that the truth of the Riemann hypothesis would imply this estimate but not vice
versa.
13
4 Density theorems
By use of the approximate functional equation, we shall rst derive a mean-square
formula for (s) in the half-plane >
1
2
. Such mean-square formulae are an important
tool in the theory of the Riemann zeta-function; in particular, they give information
on the number of zeros as we shall see below. We follow [55], VII and IX.
Theorem 4.1 For >
1
2
,
_
T
1
[( +it)[
2
dt = (2)T +O(T
22
log T).
Proof. By the approximate functional equation,
( +it) =

n<t
1
n
+it
+O(t

).
Thus, by the reection principle (2.4),
_
T
1

n<t
1
n
+it

2
dt =
_
T
1

m,n<t
1
n
+it
m
it
dt =

m,n<T
1
(mn)

_
T

_
m
n
_
it
dt
with := maxm, n. The diagonal terms m = n give the contribution

n<T
T n
n
2
= T
_
_
(2)

nT
1
n
2
_
_

n<T
1
n
21
= (2)T +O(T
22
)
by the trick (2.1). The non-diagonal terms m ,= n contribute

m,n<T
m=n
1
(mn)

_
m
n
_
iT

_
m
n
_
i
i log
n
m

0<m<n<T
1
(mn)

log
n
m
.
If m <
n
2
then log
n
m
> log 2 > 0, and hence

n<T

m<
n
2
1
(mn)

log
n
m

n<T
1
n

_
2
T
22
.
If m
n
2
we write n = m+ r with 1 r
n
2
. By the Taylor series expansion of the
logarithm,
log
n
m
= log
_
1
r
n
_
>
r
n
.
This gives, in view of (3.1),

n<T

r
n
2
1
(mn)

log
n
m

n<T
n
12

r
n
2
1
r
T
22
log T.
14
Collecting together, the assertion of the theorem follows.
Obviously, with regard to the simple pole of the zeta-function, the mean-square
formula above cannot hold on the critical line: (2) is unbounded as
1
2
+.
We can derive from the above theorem some information on the zero distribution
of (s) . This observation dates back to H. Bohr and Landau [11], resp. Littlewood
[37]. For that purpose we have to use some facts from classical function theory.
The theorem of residues states: let ( be a closed path in the complex plane without
double points, and let f(s) be meromorphic with poles at s
1
, . . . , s
m
inside ( , then
1
2i
_
C
f(s) ds =
m

j=1
Res
s=s
j
f(s), (4.1)
where
Res
s=s
j
f(s) :=
1
2i
_
|ss
j
|=
f(s) ds
is the residue of f(s) at s = s
j
(which coincides with the coecient a
1
in the
Laurent expansion of f(s) at s
j
, i.e. f(s) =

m=
a
m
(s s
j
)
m
). The special case
of a logarithmic derivative gives the possibility to count zeros and poles. If f(s) and
( satises the conditions above, and if N(0) denotes the number of zeros, N() the
number of poles of f(s) (according multiplicities) inside ( , then
1
2i
_
C
f

f
(s) ds = N(0) N(). (4.2)
In particular, we obtain the so-called argument principle, which states that if f(s) is
analytic, then the change in the argument of arg f(s) as s varies on ( (in positive
direction) equals the number of zeros of f(s) inside ( .
All these statements are well-known facts in complex analysis; see [54], Section III.
Further, we need Littlewoods lemma, i.e. the following integrated version of (4.2):
Lemma 4.2 Let f(s) be regular in and upon the boundary of the rectangle T with
vertices b, b +iT, a + iT, a, and not zero on = b. Denote by n(, T) the number of
zeros = +i of f(s) inside the rectangle with > including those with = T
but not = 0. Then
_
R
log f(s) ds = 2i
_
a
b
n(, T) d.
Since the complex logarithm is a multi-valued function, we have to be careful! Obvi-
ously, f(s) is non-vanishing in the neighbourhood of = b, and thus we dene log f(s)
here starting with any one value of the logarithm, and for other points s of the rec-
tangle by analytic continuation along the polygon with corners b + it, s = + it ,
provided that the path does not cross a zero or pole of f(s) ; if it does, put
log f(s) = lim
0+
log f( +it +i) .
15
Proof. Without loss of generality we may assume that the lines t = 0 and t = T are
free o zeros and poles of f(s) . Obviously,
_
R
log f(s) ds =
_
b
a
log f() d
_
b
a
log f( +iT) d
+
_
T
0
(log f(b +it) + log f(a +it)) i dt. (4.3)
The last integral equals
_
T
0
i
_
b
a
f

f
( +it) d dt =
_
b
a
_
+iT

f
(s) ds d.
By (4.2),
_
+iT

f
(s) ds =
_
_
b

+
_
b+iT
b

_
b+iT
+iT
_
f

f
(s) ds 2in(, T).
Substituting this in formula (4.3) proves the lemma.
Note that Littlewoods lemma can be used, in addition with Stirlings formula and
some facts about entire functions, to prove the Riemann-von Mangoldt formula (2.5)
(see [5]).
We nish our short excursion to function theory and continue with our investigations
on the zeros of the Riemann zeta-function. Let N(, T) denote the number of zeros
= +i of (s) with > , 0 < T (counting multiplicities). Then, application
of Littlewoods lemma with xed b =
0
>
1
2
yields
2
_
1

0
N(, T) d =
_
T
0
log [(
0
+it)[ dt
_
T
0
log [(2 +it)[ dt (4.4)
+
_

0
2
arg ( +iT) d
_

0
2
arg ()) d.
The main contribution comes from the rst integral on the right hand side. The last
integral does not depend on T and so it is bounded. Since (s) has an Euler product
representation (2.2), the logarithm has a Dirichlet series representation:
log (s) =

p
log
_
1
1
p
s
_
=

p,k
1
kp
ks
for > 1, (4.5)
where k runs through the positive integers; here we choose that branch of the logarithm
which is real on the positive real axis. Hence we obtain
_
T
0
log [(2 +it)[ dt = Re
_
_
_

p,k
1
kp
2k
_
T
0
exp(itk log p) dt
_
_
_

n=2
1
n
2
1.
It remains to estimate arg ( + iT) . We may assume that T is not the ordinate
of zero. Since arg (2) = 0 and
arg (s) = arctan
_
Im (s)
Re (s)
_
,
16
where
Re (2 +it) =

n=1
cos(it log n)
n
2
1

n=2
1
n
2
> 1
_

1
du
u
2
= 0,
we have by the argument principle
[ arg (2 +iT)[

2
.
Now assume that Re ( + iT) vanishes q times as
1
2
2. Devide the interval
[
1
2
+iT, 2 +iT] into q +1 parts, throughout each of which Re (s) is of constant sign.
Hence, again by the argument principle, in each part the variation of arg (s) does not
exceed . This gives
[ arg (s)[
_
q +
3
2
_
for
1
2
.
Further, q is the number of zeros of the function
g(z) =
1
2
((z +iT) +(z iT))
for Im z = 0 and
1
2
Re z 2. Thus, q n(
3
2
) , where n(r) is the number of zeros
of (s) for [z 2[ r . Obviously,
_
2
0
n(r)
r
dr
_
2
3
2
n(r)
r
dr n
_
3
2
__
2
3
2
dr
r
= n
_
3
2
_
log
4
3
.
Jensens formula states that if f(s) is an analytic function for [s[ R with zeros
s
1
, . . . , s
m
(according their multiplicities) and f(0) ,= 0, then
1
2
_
2
0
log [f(r exp(i))[ d = log
r
m
[f(0)[
[s
1
. . . s
m
[
for r < R (in a sense, this is nothing else than Poissons integral formula; see [54],
3.61). This gives here
_
2
0
n(r)
r
dr =
1
2
_
2
0
log [(2 + r exp(i))[ d log [(2)[.
In view of (3.4) we obtain
q n
_
3
2
_

1
log
4
3
_
2
0
n(r)
r
dr log T.
This yields
arg ( +iT) log T uniformly for
1
2
,
17
and, consequently, the same bound holds by integration with respect to
1
2
2.
The restriction that T has not to be an imaginary part of a zero of (s) can be removed
from considerations of continuity. Therefore, we may replace (4.4) by
_
1

0
N(, T) d =
1
2
_
T
0
log [(
0
+it)[ dt +O(log T). (4.6)
Now we need a further analytic fact due to Jensen: Jensens inequality states that for
any continuous function f(u) on [a, b] ,
1
b a
_
b
a
log f(u) du log
_
1
b a
_
b
a
f(u) du
_
(for instance, this can be deduced from the arithmetic-geometric mean inequality, or
see [54], 9.623).
Hence, we obtain for any xed
0
>
1
2
_
T
0
log [( +it)[ dt
T
2
log
_
1
T
_
T
0
[( +it)[
2
dt
_
T
by applying Theorem 4.1. Thus,
_
1

0
N(, T) d T.
Let
1
=
1
2
+
1
2
(
0

1
2
) , then we get
N(
0
, T)
1

1
_

0

1
N(, T) d
2

1
2
_
1

1
N(, T) T.
With view to (4.6) we have proved
Theorem 4.3 For any xed >
1
2
,
N(, T) T.
The theorem above is a rst density theorem. In view of the Riemann-von Mangoldt
formula (2.5) we see that, supporting Riemanns hypothesis, all but an innitesimal
proportion of the zeros of (s) lie in the strip
1
2
< <
1
2
+, however small may
be!
However, for later applications we need a stronger result.
Theorem 4.4 For any xed in
1
2
< < 1,
N(, T) T
4(1)
(log T)
10
.
18
Proof. For 2 V T let N
1
(, V ) count the zeros = +i of (s) with
and
1
2
V < V . Taking x = V in Theorem 3.7
(s) =

kV
1
k
s
+
V
1s
s 1
+O
_
V

_
for
1
2
V < t V and
1
2
1. Multiplying this with the Dirichlet polynomial
M
X
(s) :=

mX
(m)
m
s
,
where X = V
21
, gives
(s)M
X
(s) = P(s) +R(s),
where
P(s) :=

mX
(m)
m
s

kV
1
k
s
=

nXV
a(n)
n
s
with
a(n) :=

m|n
mX,nmV
(m) =
_
1 if m = 1,
0 if 1 < n X,
(4.7)
and
R(s) [M
X
(s)[V

.
Note that M
X
(s) , as the truncated Dirichlet series of the reciprocal of (s) , mollies
1
(s)
. We shall use P(s) as a zero-detector. Let s = be a zero of the zeta-function
with
1
2
V < V . Then,
1

X<nXV
a(n)
n

+O([M
X
()[V

),
1

X<nXV
a(n)
n

2
+O([M
X
()[
2
V
2
).
Then, summing up both sides of the latter inequality over all such N zeros leads to
N
1
(V )

1
1
2
V <V
_
_
_

X<nXV
a(n)
n

2
+[M
X
()[
2
V
2
_
_
_. (4.8)
Now we divide the interval [
1
2
V, V ] into subintervals of length 1 of the form [2m+n
1, 2m +n] , where n = 1, 2 and
1
4
V 1 m
1
2
V . Then, we may write

1
1
2
V <V

1
4
V 1m
1
2
V
2

n=1

2m+n1<2m+n
2 max
1n2

1
4
V 1m
1
2
V

2m+n1<2m+n
.
19
In view of the Riemann-von Mangoldt formula (2.5) there are only log V many
zeros with 2m + n 1 < 2m + n. Now let

denote the largest of the related

sums according to 2m +n 1 < 2m +n. Then

1
1
2
V <V
log V

,
resp. in (4.8)
N
1
(V ) log V

_
_
_

X<nXV
a(n)
n

2
+

mX
(m)
m

2
V
2
_
_
_. (4.9)
First of all we shall give a bound for
S(Y ) :=

Y <nU
b(n)
n

2
,
where U 2Y and V Y 1 and
b(n)

d|n
1 =:
0
(n). (4.10)
By partial summation, for xed = +i ,

Y <nU
b(n)
n

=
_
U
Y
C(u) du

with C(u) :=

Y <nu
b(n)
n
i
.
Applying the Cauchy-Schwarz inequality we obtain

Y <nU
b(n)
n

Y
1
_
U
Y
[C(u)[ du +Y

[C(U)[,

Y <nU
b(n)
n

2
Y
21
_
U
Y
[C(u)[
2
du +Y
2
[C(U)[
2
.
This leads to
S(Y ) Y
2

Y <nW
b(n)
n
i

2
,
where W U . Since the distance of the imaginary parts of counted zeros
r
=
r
+i
r
is 1, we can nd

Y <nW
b(n)n
i
r+1

_

r+1
r

Y <nW
b(n)n
it

2
dt
+2
_

r+1
r

Y <nW
b(n)n
it

Y <mW
b(m) log m m
it

dt.
20
Summation over r and application of Cauchy-Schwarz yields
S(Y ) Y
2
(I
1
+
_
I
1
I
2
),
where
I
1
:=
_
V
1
2
V

Y <nW
b(n)n
it

2
dt , I
2
:=
_
V
1
2
V

Y <nW
b(n) log n n
it

2
dt.
Taking (4.7) into account, [a(n)[ satises condition (4.10) on b(n) . By elementary
estimates one can show that

nx

k
0
(n) x(log x)
k
,
where the implicit constant depends only on k; a proof can be found in [25]. This
yields
I
1
(V +Y ) log V

Y <n2Y

2
0
(n) (V Y +Y
2
)(log V )
5
,
I
2
(V Y +Y
2
)(log V )
7
.
Now dividing the rst sum on the right hand side of (4.9) into log V sums, appli-
cation of the latter estimates yields
log V

X<nV X
a(n)
n

2
(V X
12
+ (V X)
22
)(log V )
9
.
Similarly, we get for the second term
V
2
(log T)
2

mX
(m)
m

2
V
2
(V +X
22
)(log V )
9
.
Substituting this in (4.9) with regard to X = V
21
, we obtain
N
1
(V ) V
4(1)
(log V )
9
.
Using this with V = T
1n
and summing up over all n N, proves the theorem.
Theorem 4.4 is due to Bohr and Landau [12]. There are stronger estimates known.
For instance, the strongest unconditional estimate which holds throughout the right
half of the critical strip is
N(, T) T
2.4(1)
(log T)
18.2
due to Gritsenko [20]. The density hypothesis states that
N(, T) T
(2+)(1)
for all > 0; one can show that the Lindelof hypothesis (3.5) implies the density
hypothesis.
21
5 A zero-free region
In order to prove Gauss conjecture we need further knowledge on the zero distribution
of (s) . We shall establish a zero-free region for (s) which covers the abscissa of ab-
solute convergence = 1. In this delicate problem we follow (with slight modications)
the ideas of de La Vallee Poussin; see [55], Section 3.
In the sequel we only argue for s = +it from the upper half-plane; with regard
to (2.4) all estimates below can be reected with respect to the real axis.
Lemma 5.1 We have, for t 8, 1
1
2
(log t)
1
2,
(s) log t and

(s) (log t)
2
.
Proof. Let 1 (log t)
1
3. If n t , then
[n
s
[ = n

n
1(logt)
1
= exp
__
1
1
log t
_
log n
_
n.
Thus, the approximate functional equation (Theorem 3.7) in addition with (3.1) implies
(s)

nt
1
n
+t
1
log t.
The estimate for

(s) follows immediately from Cauchys formula (which is actually

a consequence of the theorem of residues (4.1))

(s) =
1
2i
_
(z)
(z s)
2
dz,
and standard estimates of integrals, or alternatively, by dierentiation of the formula
of Theorem 3.2.
In view of the Euler product (2.2) we have for > 1
[( +it)[ = exp(Re log (s)) = exp
_
_

p,k
cos(kt log p)
kp
k
_
_
.
Since
17 + 24 cos + 8 cos(2) = (3 + 4 cos )
2
0,
it follows that
()
17
[( +it)[
24
[( + 2it)[
8
1. (5.1)
This inequality is the main idea for our following observations. By the approximate
functional equation, Theorem 3.7, we have for small > 1
()
1
1
.
22
Assuming that (1 +it) has a zero for t = t
0
,= 0, it would follow that
[( +it
0
)[ 1,
leading to
lim
1+
()
17
[( +it
0
)[
24
= 0,
which contradicts (5.1). Thus
(1 +it) ,= 0.
It can be shown that the non-vanishing of (1 +it) is equivalent to Gauss conjecture
(2.3), i.e. the prime number theorem without error term; see [55], 3.7. But we
are interested in a prime number theorem with error term. A simple renement of
the argument allows a lower estimate for the modulus of (1 + it) : for t 1 and
1 < < 2, we deduce from (5.1) and Lemma 5.1
1
[( +it)[
()
17
24
[( + 2it)[
1
3
( 1)

17
24
(log t)
1
3
.
Furthermore, with Lemma 5.1,
(1 +it) ( +it) =
_

1

(u +it) du [ 1[(log t)
2
. (5.2)
Hence
[(1 +it)[ [( +it)[ c
1
( 1)(log t)
2
c
2
( 1)
17
24
(log t)

1
3
c
1
( 1)(log t)
2
,
where c
1
, c
2
are certain positive constants. Chosing a constant B > 0 such that
A := c
2
B
17
24
c
1
B > 0 and putting = 1 +B(log t)
8
, we obtain
[(1 +it)[
A
(log t)
6
. (5.3)
This gives an estimate on the left of the line = 1.
Lemma 5.2 There exists a positive constant such that
(s) ,= 0 for 1 min1, (log t)
8
.
Proof. In view of Lemma 5.1 the estimate (5.2) holds for 1 (log t)
8
1.
Using (5.3), it follows that
[( +it)[
Ac
1

(log t)
6
,
where the right hand side is positive for suciently small . This yields the zero-free
region of Lemma 5.2.
The largest known zero-free region for the zeta-function was found by Vinogradov
[56] and Korobov [26]. Using Vinogradovs ingenious method for exponential sums,
they proved
(s) ,= 0 in 1
c
(log [t[)
1
3
(log log [t[)
2
3
(5.4)
for some positive constant c and suciently large [t[ ; see [25], IV.3.
23
6 The prime number theorem
The aim of this section is to prove Gauss conjecture (2.3), the celebrated prime number
theorem.
Out of technical reasons we work with the logarithmic derivative of (s) (instead
of log (s) as Riemann did). Logarithmic dierentiation of the Euler product (2.2),
resp. dierentiation of (4.5), gives for > 1

(s) =

n=1
(n)
n
s
, where (n) :=
_
log p if n = p
k
,
0 otherwise,
is the von Mangoldt -function. Since (s) does not vanish in the half-plane
> 1, the logarithmic derivative is analytic for > 1. As we shall see below all
information on (x) is encoded in
(x) :=

nx
(n) =

px
log p +O
_
x
1
2
_
. (6.1)
The idea of proof is ingenious but simple. Partial summation gives

(s) = s
_

1
(x)
dx
x
s+1
.
If we can now transform this into a formula in which (x) is isolated, we may hope to
nd an asymptotic formula for (x) by contour integration methods. The rst step of
this program can be done by a type of Fourier transformation.
Lemma 6.1 Let c and y be positive and real. Then
1
2i
_
c+i
ci
y
s
s
ds =
_

_
0 if 0 < y < 1,
1
2
if y = 1,
1 if y > 1.
Proof. If y = 1, then the integral in question equals
1
2
_

dt
c +it
=
1

lim
T
_
T
0
c
c
2
+t
2
dt =
1

lim
T
arctan
T
c
=
1
2
,
by well-known properties of the arctan-function. Now assume that 0 < y < 1 and
r > c. Since the integrand is analytic in > 0, Cauchys theorem (resp. the theorem
of residues (4.1)) implies, for T > 0,
_
c+iT
ciT
y
s
s
ds =
_
_
riT
ciT
+
_
r+iT
riT
+
_
c+iT
r+iT
_
y
s
s
ds.
It is easily seen that
_
ciT
riT
y
s
s
ds
1
T
_
c
r
y

d
y
c
T[ log y[
,
_
r+iT
riT
y
s
s
ds
y
r
r
+y
r
_
T
1
dt
t
y
r
_
1
r
+ log T
_
.
24
Sending now r and then T to innity, the rst case follows. Finally, if y > 1,
then we bound the corresponding integrals over the rectangular contour with corners
c iT, r iT , analogously. Now the pole of the integrand at s = 0 with residue
Res
s=0
y
s
s
= lim
s0
y
s
s
s = 1
gives the values 2i for the integral in this case.
We apply this to the logarithmic derivative of the zeta-function and obtain for x , Z
and c > 1
_
c+i
ci

n=1
(n)
n
s
x
s
s
ds =

n=1
(n)
_
c+i
ci
_
x
n
_
s
ds
s
;
here interchanging integration and summation is allowed by the absolute convergence
of the series. In view of Lemma 6.1 it follows that

nx
(n) =
1
2i
_
c+i
ci

n=1
(n)
n
s
x
s
s
ds,
resp.
(x) =
1
2i
_
c+i
ci
_

(s)
_
x
s
s
ds;
this is Perrons formula. Since
_
ci
ciT
y
s
s
ds =
y
s
s log y

ci
s=ciT
+
1
log y
_
ci
ciT
y
s
s
2
ds
y
c
T[ log y[
for 0 < y ,= 1 and T > 0, it follows that
_
ci
ciT
_

n=2
(n)
n
s
_
x
s
s
ds
x
c
T

n=2
(n)
n
c

log
x
n

x
c
T

(c)

+
x(log x)
2
T
+ log x.
This yields
(x) =
1
2i
_
c+iT
ciT

(s)
x
s
s
ds +O
_
x
c
T

(c)

+
x(log x)
2
T
+ log x
_
, (6.2)
which holds for arbitrary x. To nd an asymptotic formula for the integral above we
move the path of integration to the left, excluding s = 0. By the theorem of residues
we expect contributions to the main term from the poles of the integrand, i.e.
the nontrivial zeros of (s) ,
the pole of (s) at s = 1.
25
However, for our purpose it is sucient to exclude the zeros of the zeta-function. In
view of the zero-free region of Lemma 5.2 we put c = 1+ with = (log t)
8
, where
is given by Lemma 5.2, and integrate over the boundary of the rectangle T given by
the corners 1 iT . By this choice (s) does not vanish in and on the boundary
of T. The theorem of residues (4.1) implies
_
c+iT
ciT
_

(s)
_
x
s
s
ds =
_
_
1iT
1+iT
+
_
1+iT
1iT
+
_
1+iT
1+iT
__

(s)
_
x
s
s
ds
+2iRes
s=1
_

(s)
_
x
s
s
.
For the logarithmic derivative of (s) we have

(s) =
d
ds
log (s) =
1
s 1
+O(1)
as s 1. Thus, we obtain for the residue
Res
s=1
_

(s)
_
x
s
s
= lim
s1
(s 1)
_
1
s 1
+O(1)
_
x
s
s
= x.
It remains to bound the integrals. For the horizontal integrals we nd with regard to
Lemma 5.2
_
1+iT
1iT
_

(s)
_
x
s
s
ds
x
1+
T
.
Further, for the vertical integral,
_
1++iT
1iT
_

(s)
_
x
s
s
ds x
1
(log T)
9
.
Collecting together, we dedcue from (6.2)
(x) = x +O
_
x
1+
T
+x
1
(log T)
9
+
x(log x)
2
T
+ log x
_
.
Choosing T = exp(
1
10
(log x)
1
9
) , we arrive at
(x) = x +O
_
xexp(c(log x)
1
9
)
_
.
Setting
(x) :=

px
log p,
it follows from (6.1) that also
(x) = x +O
_
xexp(c(log x)
1
9
)
_
.
26
Applying now partial summation, we nd
(x) =

px
log p
1
log p
=
(x)
log x

_
x
2
(u)
d
du
1
log u
du
=
x
log x

_
x
2
u
d
du
1
log u
du +O
_
xexp
_
c(log x)
1
9
__
.
Partial integration leads to prime number theorem:
Theorem 6.2 There exists a positive constant c such that for x 2
(x) = Li (x) +O
_
xexp
_
c(log x)
1
9
__
.
Thus, the simple pole of the zeta-function is not only the key in Eulers proof of the
innitude of primes (Section 2) but gives also the main term of the asymptotic formula
in the prime number theorem.
We see that the prime numbers, which - on a rst look - seem to be randomly
distributed among the positive integers, satisfy a strong distribution law! For example,
the prime number theorem implies that, if p
n
is the n-th prime number, then
p
n
nlog n.
In view of the largest known zero-free region (5.4) one can obtain
(x) = Li(x) +O
_
_
xexp
_
_
c
(log x)
3
5
(log log x)
1
5
_
_
_
_
.
With a little bit more eort and a little bit more facts from the theory of functions
it is possible to prove the analogue of Riemanns explicit formula (2.7). Integrating
over the full complex plane one can show for x ,= p
k
the exact(!) explicit formula
(x) = x

1
2
log
_
1
1
x
2
_
log(2),
resp. its truncated version
(x) = x

||T
x

+O
_
x
T
(log(xT))
2
_
. (6.3)
This shows a deep relation between the error term in the prime number theorem and
the distribution of the nontrivial zeros of the zeta-function.
Theorem 6.3 For xed [
1
2
, 1) ,
(x) x x
+
(s) ,= 0 for > .
27
With regard to known zeros of (s) on the critical line it turns out that an error term
with <
1
2
is impossible; see [24] for more details.
Unfortunately, for the proof of one implication of Theorem 6.3 we have to use some
facts we have not proved; the interested reader may nd all missing details in [25].
Proof. By partial summation we obtain for > 1

(s) =
s
s 1
+s
_

1
(u) u
u
s+1
du.
If (x) x x
+
, then the integral above converges for > , giving an analytic
continuation for

(s)
1
s 1
to the half-plane > , and, in particular, (s) does not vanish there.
Conversely, if all nontrivial zeros = + i satisfy , then it follows from
(6.3) that
(x) x x

||T
1
[[
+
x
T
(log(xT))
2
. (6.4)
In view of the Riemann-von Mangoldt-Formel (2.5) we have
N(T + 1) N(T) log T,
and therefore

||T
1
[[

[T]+1

m=1
log m
m
(log T)
2
.
Substituting this in (6.4) leads to
(x) x x

(log T)
2
+
x
T
(log(xT))
2
.
Now the choice T = x
1
nishes the proof of this implication.
Theorem 6.3 shows that the Riemanns hypothesis (2.6) is true if and only if
(x) = x +O
_
x
1
2
+
_
,
and since the latter estimate is best possible (there are zeros on the critical line),
Riemanns hypothesis states that the prime numbers are as uniformly distributed as
possible!
In order to prove Voronins universality theorem we have to do some preliminaries.
28
7 Diophantine approximation
In the theory of diophantine approximations one investigates how good an irrational
number can be approximated by rational numbers; this has a plenty of applications in
various elds of mathematics and natural sciences. We follow [25], A.8.
For abbreviation we denote vectors of R
N
by x = (x
1
, . . . , x
N
) R
N
, we write
x = (x
1
, . . . , x
N
) for R and x y = x
1
y
1
+ . . . + x
N
y
N
. Further, for x R
N
and R
N
we write x mod 1 if there exists y Z
N
such that x y .
Moreover, we have to introduce the notion of the Jordan volume of a region R
N
.
Therefore, we consider the sets of parallelepipeds
1
and
2
with sides parallel to the
axes and of volume
1
and
2
with
2

2
; if there are
1
and
2
such that
limsup

1

1
coincides with liminf

2

2
, then has the Jordan volume
= limsup

1
= limsup

2
.
The Jordan sense of volume is more restrictive than the one of Lebesgue, but if the
Jordan volume exists it is also dened in the sense of Lebesgue and equal to it.
Weyl [59] proved
Theorem 7.1 Let a
1
, . . . , a
N
R be linearly independent over the eld of rational
numbers, and let be a subregion of the N-dimensional unit cube with Jordan volume
. Let a = (a
1
, . . . , a
N
) , then
lim
T
1
T
meas (0, T) : a mod 1 = .
Proof. From the denition of the Jordan measure it follows that for any > 0 there
exist two nite sets of open parallelepipeds

j
and

+
j
inside the unit cube such
that
_

j
int
_
+

j
and meas
_
_
_
+

j
_
_
< ; (7.1)
here, as usual, M denotes the closure of the set M, and int M its interior. Denote
by

the characteristic function of

j
, i.e.

(x) =
_
1 if x

j
,
0 if x ,

j
.
Further, let be the characteristic function of mod 1. Consequently,
0

(x) (x)
+
(x) 1,
and _
[0,1]
N
(
+
(x)

(x)) dx < ,
29
where the integral is N -dimensional with dx = dx
1
dx
N
. Dene
(x) =
_
_
_
0 if [x[
1
2
,
c exp
_

_
1
x+
1
2
+
1
x
1
2
__
if [x[ <
1
2
,
where c is given by
_ 1
2

1
2
(x) dx = 1.
Consequently, (x) is an inntely dierentiable function, and hence the functions,
dened by

(x) =
N
_
[0,1]
N

(y)
_
x
1
y
1

_

_
x
N
y
N

_
dy.
for 0 < < 1, are innitely dierentiable functions, too. From (7.1) it follows that for
suciently small we have
0

(x) (x)
+
(x) 1,
and
0
_
[0,1]
N
(
+
(x)

(x)) dx < 2. (7.2)

We conclude
_
T
0

(a) d meas ( (0, T) : a mod 1

_
T
0

+
(a) d (7.3)
and
0
_
T
0

+
(a) d
_
T
0

(a) d 2T.
Both integrands above are innitely dierentiable functions which are 1-periodic in
each variable. Thus, we have the Fourier expansion

(x) =

nZ
N
c

n
exp(2in x),
where
c

n
=
_
[0,1]
N

(x) exp(2in x) dx.

Note that c

0
is the volume of

j
. Integration by parts gives
c

n

N

j=1
([n
j
[ + 1)
k
for k = 1, 2, . . . ,
where the implicit constant depends only on k. This shows that the Fourier series
converges absolutely, and hence, for every > 0 there exists a nite set / Z
N
such
that

(x) =

nM
c

n
exp(2in x) +R(x) with [R(x)[ < .
30
This yields
1
T
_
T
0

(a) d =
1
T
_
T
0

nM
c

n
exp(2in x)) d +
with [[ < 1. Consequently,
1
T
_
T
0

(a) d = c

0
+

0=nM
c

n
T
_
T
0
exp(2in a) d +.
Since the a
n
are linearly independent over Q, we have n a ,= 0 for n ,= 0. It follows
for such n that
_
T
0
exp(2in a)) d 1.
Since > 0 is arbitrary, we obtain
lim
T
1
T
_
T
0

(a) d = c

0
.
This gives with regard to (7.3)
c

0
liminf
T
1
T
meas ( (0, T) : a mod 1
limsup
T
1
T
meas ( (0, T) : a mod 1 c
+
0
+
for any positive . From (7.2) it follows that 0 c
+
0
c

0
2. Now sending 0,
the theorem is proved.
As an immediate consequence of Theorem 7.1 we get the classical inhomogeneous Kro-
necker approximation theorem:
Corollary 7.2 Let
1
, . . . ,
N
R be linearly independent over the eld of rationals,
let
1
, . . . ,
N
be arbitrary real numbers, and let q be a positive number. Then there
exists a number > 0 and integers x
1
, . . . , x
N
such that
[
n

n
x
n
[ <
1
q
for 1 n N.
We give an application to the value distribution of the zeta-function. As we have seen
above (s) is in the half-plane > 1 given by an absolute convergent Dirichlet series.
However, the value distribution of the zeta-function in that region is anything but
boring. Answering a question of Hilbert, H. Bohr and Landau [10], [13] showed that
[(s)[ takes arbitrarily large and arbitrarily small values in > 1 - in spite of the
absence of zeros!
31
Theorem 7.3 For any > 0 and any real there exists an innite sequence of
s = +it with 1+ and t such that
Re exp(i) log ( +it) (1 ) log () +O(1).
In particular,
liminf
>1,t1
[(s)[ = 0 and limsup
>1,t1
[(s)[ = .
Our proof diers slightly from the original one.
Proof. By (4.5) one easily nds
log (s) =

p
1
p
s
+O(1) for > 1. (7.4)
Thus, we have for any t 1 and x 2
Re exp(i) log ( +it)

px
cos(t log p )
p

p>x
1
p

+O(1). (7.5)
Here we use diophantine approximation. By the unique (!) prime factorization of the
integers the logarithms of the prime numbers are linearly independent. Denote by p
n
the n-th prime, then Kroneckers approximation theorem implies that for any given
integers q, N the existence of some real number > 0 and integers x
1
, . . . , x
N
with

log p
n
2


2
x
n

<
1
q
for n = 1, . . . , N.
Obviously, we get with q innitely many with the above property. Setting
N = (x) , we obtain
cos ( log p ) cos
_
2
q
_
for all p x,
provided that q 4. Therefore, we deduce from (7.5)
Re exp(i) log ( +i) cos
_
2
q
_

px
1
p

p>x
1
p

+O(1),
resp.
Re exp(i) log ( +i) cos
_
2
q
_
log () 2

p>x
1
p

+O(1) (7.6)
in view of (7.4). Sending q, x we obtain the estimate of the theorem. In view of
the simple pole of (s) at s = 1 we get with = 0, resp. = , by sending 1+
the further assertions follow.
32
It is even possible to give quantitative estimates for the rate of divergence; see [55],
8.6, and [51].
We conclude with the notion of uniform distribution modulo 1. Let () be a
continuous function with domain of denition [0, ) and range R
N
. Then the curve
() is said to be uniformly distributed mod 1 in R
N
if for every parallelepiped

= [
1
,
1
] . . . [
N
,
N
] with 0
j
<
j
1 for 1 j N
lim
T
1
T
meas ( (0, T) : ()

mod 1 =
N

j=1
(
j

1
).
In a sense, a curve is uniformly distributed mod 1 if the right proportion of values lies
in a given subset of the unit cube.
Since in questions about uniform distribution mod 1 one is interested in the frac-
tional part only, we dene for a curve () in R
N
() = (
1
() [
1
()], . . . ,
N
() [
N
()]);
recall that [x] is the integral part of x R.
Theorem 7.4 Suppose that the curve () is uniformly distributed mod1 in R
N
.
Let D be a closed and Jordan measurable subregion of the unit cube in R
N
and let
be a family of complex-valued continuous functions dened on D. If is uniformly
bounded and equicontinuous, then
lim
T
1
T
_
T
0
f(())

D
() d =
_
D
f(x) dx
uniformly with respect to f , where

D
() is equal to 1 if () D mod 1, and
zero otherwise.
Proof. By the denition of the Riemann-integral as a limit of Riemann sums, we have
for any Riemann integrable function F on the unit cube in R
N
lim
T
1
T
_
T
0
F(()) d =
_
[0,1]
N
F(x) dx. (7.7)
By the assumptions on , for every > 0 there exist f
1
, . . . , f
n
such that for
every f there is an f
j
with 1 j n, and
sup
xD
[f(x) f
j
(x)[ < .
By (7.7) there exists T
0
such that for any T > T
0
and for each function f
1
, . . . , f
n
one
has

_
D
f
j
(x) dx
1
T
_
T
0
f
j
(())

D
() d

< .
33
Now, for any f ,

_
D
f(x) dx
1
T
_
T
0
f(())

D
() d

_
D
f
j
(x) dx
1
T
_
T
0
f
j
(())

D
() d

_
D
(f(x) f
j
(x)) dx

+
1
T

_
T
0
(f(()) f
j
(()))

D
() d

By the estimates above, this is bounded by 3. Since > 0 is arbitrary, the assertion
of the theorem follows.
In the next section we shall meet the heart of Voronins universality theorem.
8 Conditionally convergent series
A series

n
a
n
of real numbers a
n
is called conditionally convergent if

n
[a
n
[
is divergent but

n
a
n
is convergent for an appropiate rearrangement of the terms
a
n
. Riemann proved that any conditionally convergent series can be rearranged such
that its sum converges to an arbitrary preassigned real number; see [2]; hence, every
convergent series, which does not converge absolutely, is conditionally convergent. For
instance, to any given c R there exists a permutation of N (i.e. a one-to-one
mapping on N) such that

n=1
(1)
(n)
= c;
here and in what follows do not confuse the permutation with the prime counting
function. Thus, all conditionally convergent series are in a certain sense universal with
respect to R!
It is the aim of this section to extended Riemanns rearrangement theoremto Hilbert
spaces; recall that a complete normed linear space with inner product is called Hilbert
space. We shall give an example which will be of interest later on. Let R be a positive
real number, then the Hardy space ]
R
2
is the set of functions f(s) which are analytic
for [s[ < R and for which
|f| := lim
rR
_ _
|s|<r
[f(s)[ d dt < .
We dene on ]
R
2
an inner product by
f, g = Re
_ _
|s|R
f(s)g(s) d dt. (8.1)
This makes ]
R
2
into a real Hilbert space.
Pechersky [45] proved
34
Theorem 8.1 Suppose that a series

n
u
n
of vectors in a real Hilbert space ] satis-
es the condition

n=1
|u
n
|
2
< ,
and for any e ] with |e| = 1 the series

n
u
n
, e converges conditionally. Then
for any v ] there is a permutation of N such that

n=1
u
(n)
= v
in the norm of ].
It is obvious how the notion of conditionally convergent series has to be extended to
real Hilbert spaces.
The proof is a bit more complicated than the one for Riemanns rearrangement
theorem but the idea behind is still the same. We start with
Lemma 8.2 Under the assumptions of Theorem 8.1, for any v ] and any > 0
there exist a positive integer N and numbers
1
, . . . ,
N
, equal to 0 or 1, such that
_
_
_s
N

n=1

n
u
n
_
_
_ < .
Proof. We choose an integer m so that

n=m
|u
n
|
2
<

2
9
.
Denote by P
m
the set of all linear combinations
N

n=m

n
u
n
with
n
[0, 1] and N = m, m+ 1, m + 2, . . . ;
obviously, P
m
is convex. Let P
m
be the closure of P
m
with respect to the norm of ];
consequently P
m
is a closed convex set. Now we shall show that P
m
coincides with
].
The seperation theorem for linear operators states that if X is a normed linear
space and D is a convex subset of X which is closed in the norm of X, then for any
s X D there exist > 0 and a linear functional F on X such that
F(x) F(s) for all x D.
The proof follows from the well-known theorem of Hahn-Banach, which relates linear
functionals to convex sets; see [14], V.2.7. A simple consequence is that for any proper
convex subset D of real Hilbert space ], which is closed in the norm of ], there exists
a vector e ] with |e| = 1 such that
sup
xD
x, e < .
35
We return to our problem: suppose that P
m
,= ], then, by the above reasoning,
there exists e ] with |e| = 1 such that sup
xPm
x, e < . Since, by assumption,
the series

nm
u
n
, e converges conditionally with some arrangement of the terms,
the series of the positive terms of the series only is divergent. Thus, for any C there
exist an N and a sequence
m
, . . . ,
N
, equal to 0 or 1, such that
N

n=m

n
u
n
, e > C.
Since

N
m

n
u
n
P
m
, it follows that sup
xPm
x, e = , giving the contradiction.
So we have shown P
m
= ]. Consequently, there exist N m and
m
, . . . ,
N

[0, 1] such that
_
_
_v
N

n=m

n
u
n
_
_
_ <

3
.
By induction we can construct
m
, . . . ,
N
, equal to 0 or 1, such that for any M with
m M N the inequality
_
_
_
M

n=m

n
u
n

n=m

n
u
n
_
_
_
M

n=m
|u
n
|
2
holds. Therefore, we may set
m
= 1 and suppose that
m
, . . . ,
M
have been chosen
so that the last inequality is fullled. With
M+1
, equal to 0 or 1, satisfying
(
M+1

M+1
)
_
M

n=m
(
n

n
)u
n
, u
n
_
0,
we get
_
_
_
M+1

n=m

n
u
n

M+1

n=m

n
u
n
_
_
_
2

_
_
_
M

n=m
(
n

n
)u
n
_
_
_
2
+|u
M+1
|
2

M+1

n=m
|u
n
|
2
.
In particular,
_
_
_
N

n=m

n
u
n

n=m

n
u
n
_
_
_
2

n=m
|u
n
|
2
<

2
9
,
which proves the lemma.
The next step is
Lemma 8.3 Under the assumptions of Theorem 8.1, there exists a permutation n
k

of N such that some subsequence of the partial sums of the series

k
u
n
k
converges to
v in the norm of ].
Proof. Let n
1
= 1. Applying Lemma 8.2 to the series

n2
u
n
, yields the existence
of a nite set T
1
2, 3, . . . such that
_
_
_v u
1

nT
1
u
n
_
_
_ <
1
2
.
36
Now write the indices in T
1
in an arbitrary order. If 2 , T
1
, then write 2. Denote by
T
2
the set of all indices we have so far, and dene N
1
= maxn T
2
. Applying (8.2)
to the series

n=N
1
+1
u
n
, shows that there exists a nite set T
3
N
1
+1, N
1
+2, . . .
such that
_
_
_v

nT
2
u
n

nT
3
u
n
_
_
_ <
1
4
.
Continuing this process, the assertion of the lemma follows. .
Further, we have to prove
Lemma 8.4 Let v
1
, . . . , v
N
be arbitrary elements in a real Hilbert space ]. Then
there exists a permutation of the set 1, . . . , N such that
max
1mN
_
_
_
N

n=1
v
(n)
_
_
_
_
N

n=1
|v
n
|
2
_
1
2
+ 2
_
_
_
N

n=1
v
n
_
_
_.
Proof. First, suppose that
N

n=1
v
n
= 0.
Then we shall construct by induction a permutation n
1
, . . . , n
N
of 1, . . . , N such
that
max
1mN
_
_
_
m

k=1
v
n
k
_
_
_
_
N

n=1
|v
n
|
2
_
1
2
. (8.2)
Therefore, set n
1
= 1 and suppose that n
1
, . . . , n
j
with 1 j N 1 have been
chosen, satisfying
max
1mj
_
_
_
m

k=1
v
n
k
_
_
_
2

n=1
|v
n
|
2
.
Then we choose n
j+1
from the remaining numbers such that

k=1
v
n
k
, v
n
j+1
0;
obviously, such an n
j+1
exists since otherwise

j=n
k
_
j

k=1
v
n
k
, v
j
_
=
_
j

k=1
v
n
k
,
j

k=1
v
n
k
_
> 0.
Hence,
_
_
_
j+1

k=1
v
n
k
_
_
_
2

j+1

k=1
|v
n
k
|
2
.
This yields the existence of a permutation such that (8.2) holds under the assumption

N
n=1
v
n
= 0.
37
For arbitrary v
1
, . . . , v
N
dene
v
N+1
=
N

n=1
v
n
,
and apply the already proved fact; obviously this leads to the additional term
|v
N+1
|
2
=
_
_
_
N

n=1
v
n
_
_
_
2
,
multiplied by 2, in (8.2). The lemma is shown.
Now we are in the position for the
Proof of Theorem 8.1. Without loss of generality we may assume, by Lemma 8.3,
that some subsequence of the partial sums of the series

k
u
k
converges to v in the
norm of ]; we dene
U
n
=
n

k=1
u
k
,
and suppose that the sequence of the U
n
j
converges to v . For each j N there is a
permutation of the set of vectors U
n
j
+1
, . . . , U
n
j+1
in such a way that the value of
m
j
:= max
1mn
j+1
n
j
_
_
_
n
j
+m

n=n
j
+1
u
(n)
_
_
_
is minimal. By Lemma 8.4 it follows that
m
j

_
_

n=n
j
+1
|u
n
|
2
_
_
1
2
+ 2|U
n
j+1
U
n
j
|,
which obviously tends to zero as j . Hence, the corresponding series converges
to v in the norm of ]. Theorem 8.1 is proved.
In the following section we shall return to the zeta-function and start with the proof
of Voronins universality theorem.
9 Finite Euler products
As we have seen in the beginning the Euler product (2.2) does not represent the zeta-
function inside the critical strip. However, as Bohr [7] discovered in his investigations on
the value distribution of (s) , an appropriate truncated Euler product (2.2) converges
almost everywhere inside the critical strip to the zeta-function; see (10.5) below. This
important observation can be used for our approximation problem: if we are able to
approximate a given function by a nite Euler product, then we nally have only to
switch from the nite Euler product to the zeta-function itself!
38
Let denote the set of all sequences of real numbers indexed by the primes, that
are all innite vectors of the form := (
2
,
3
, . . .) with
p
R. Then we dene for
any nite subset M of the set of all primes, any and all complex s

M
(s, ) =

pM
_
1
exp(2i
p
)
p
s
_
1
.
Obviously,
M
(s, ) is an analytic function in s without zeros in the half-plane > 0.
Consequently, its logarithm exists and equals
log
M
(s, ) =

pM
log
_
1
exp(2i
p
)
p
s
_
(out of technical reasons we prefer to consider series than products); here as for log (s)
we may take the principal branch of the logarithm.
The rst step in the proof of Voronins theorem is to show
Theorem 9.1 Let 0 < r <
1
4
and suppose that g(s) is continuous on [s[ r and
analytic in the interior. Further, let
0
=
_
1
4
,
2
4
,
3
4
, . . .
_
. Then for any > 0 and any
y > 0 there exists a nite set M of prime numbers, containing at least all primes
p y , such that
max
|s|r

log
M
_
s +
3
4
,
0
_
g(s)

< .
Unfortunately, the proof makes use of some classical results from analysis, which proofs
are beyond the scope of this course.
Proof. Since g(s) is continuous for [s[ r , there exists > 1 such that
2
r <
1
4
and
max
|s|r

g
_
s

2
_
g(s)

<

2
. (9.1)
The function g
_
s

2
_
is bounded on the disc [s[ r =: R, and thus belongs to the
Hardy space ]
R
2
.
Denote by p
k
the k-th prime number. We consider the series

k=1
u
k
(s) with u
k
(s) := log
_
_
1
exp(2i
p
k
)
p
s+
3
4
k
_
_
1
.
First, we shall prove that for every v ]
R
2
there exists a rearrangement of the series

u
k
(s) for which

k=1
u
j
k
(s) = v(s).
In view of the Taylor expansion of the logarithm the series

k
u
k
(s) diers from

k=1

k
(s) =

k=1
exp
_

2ik
4
_
p
s
3
4
k
39
by an absolute convergent series. Therefore, it is sucient to verify the conditions of
the rearrangement theorem 8.1 for the series

k

k
(s) . Since R <
1
4
,

k=1
|
k
|
2

p
1
p
3
2
< .
Further, we have to check that for any ]
R
2
with ||
2
= 1 the series

k=1

k
, (9.2)
is conditionally convergent for some rearrangement of its terms. With view to the
Cauchy-Schwarz inequality,

k=1

k
,
_
_
_

k=1

1
2
|

1
2
=
_
_
_

k=1

1
2
< ,
it suces to show that there exist two subseries of (9.2), where one is diverging to +
and the other one to .
By (8.1),

k
, = Re
_ _
|s|R
exp
_

2ik
4
_
p
s
3
4
k
(s) d dt
= Re
_
exp
_

2ik
4
_
_ _
|s|R
p
s
3
4
k
(s) d dt
_
. (9.3)
This shows
lim
k

k
, = 0.
Now dene
(x) =
_ _
|s|R
exp
_
x
_
s +
3
4
__
(s) d dt,
then the integral appearing on the right handside of (9.3) equals (log p
k
) . Further,
let (s) =

m=0

m
s
m
, then we may express (x) in terms of the Taylor coecients

m
as follows: obviously,
(x) = exp
_

3x
4
__ _
|s|R
exp(sx)(s) d dt
= exp
_

3x
4
__ _
|s|R

n=0
(sx)
n
n!

m=0

m
s
m
d dt
= exp
_

3x
4
_

m=0

n=0
(1)
n
x
n
n!

m
_ _
|s|R
s
m
s
n
d dt.
Using polar coordinates,
_ _
|s|R
s
m
s
n
d dt =
_
R
0
_
2
0

m+n
exp(i(n m)) d d =
_
2
R
2m+2
2m+2
if m = n,
0 if m ,= n.
40
This yields
(x) = R
2
exp
_

3x
4
_

m=0
(1)
m
x
m

m
m!(m+ 1)
R
2m
= R
2
exp
_

3x
4
_

m=0

m
m!
(xR)
m
, (9.4)
where
m
= (1)
mmR
m
m+1
. Since || = 1, we get
1 =
_ _
|s|R
[(s)[
2
d dt =

m=0
[
m
[
2
_ _
|s|R
[s[
2m
d dt = R
2

m=0
[
m
[
2
m + 1
R
2m
.
Hence,
0 <

m=0
[
m
[
2
1, (9.5)
which implies [
m
[ 1. The function F(z) , given by
F(z) =

m=0

m
m!
z
m
,
denes an entire function in z .
Now we shall show that for any > 0 there exists a sequence of positive real
numbers z
j
, tending to +, for which
[F(z
j
)[ > exp((1 + 2)z
j
). (9.6)
Suppose the contrary. Then there is a (0, 1) and a constant B such that [F(z)[ <
Bexp((1 + 2)z) for any z 0. Consequently,
[ exp((1 +)z)F(z)[ < Bexp((1 +)z) for z 0; (9.7)
Since [
m
[ 1, this estimate even holds for z < 0 by a suitable change of the constant
B.
Here we have to apply the theorem of Paley-Wiener [44], Theorem X in Section 1,
which states that if > 0, then the identity
G(z) =
_

g() exp(iz) d
holds for some function g() if and only if
_

[G(z)[
2
dz <
and G(z) has an analytic continuation throughout the complex plane satisfying G(z)
exp(( + )z) for any > 0, and where the implicit constant may depend on (this
characterizes all transcendent functions of xed exponential type ). Further, we
41
have to make use of Plancherels theorem [46] which states that under the assumptions
on G(z) in the theorem of Paley-Wiener,
g() =
1
2
_

G(z) exp(iz) dz
almost everywhere in R. The proofs can be found in [44] and [1]; they rely essentially
on Fourier theory.
Application of the theorem of Paley-Wiener in our case with G(z) = exp(3z)F(z)
yields, with regard to (9.7), the representation
exp((1 +)z)F(z) =
_
3
3
f() exp(iz) d,
where f() is a square integrable function with support on the interval [3, 3] . Fur-
ther, Plancherels theorem implies
f() =
1
2
_

F(z) exp((1 +)z iz) dz

almost everywhere. Hence, f() is analytic in a strip covering the real axis. Since the
support of f() lies inside a compact interval, the integral above has to be zero outside
this interval. Hence, F(z) has to vanish identically, contradicting the existence of a
sequence of positive real numbers z
j
with (9.6).
Let x
j
=
z
j
R
. Then it follows from (9.4) that
[(x
j
)[ > R
2
exp
_

3x
j
4
_
F(x
j
R) R
2
exp
_
x
j
_
3
4
+R(1 + 2)
__
.
Thus, with suciently small

> 0 we obtain the existence of a sequence of positive

real numbers x
j
, tending to +, for which
[(x
j
)[ > exp((1

)x
j
). (9.8)
Now we shall approximate F and by polynomials. Let N
j
= [x
j
] +1 and assume
that x
j
1 x x
j
+ 1. Since [
m
[ 1,

m=N
2
j
+1

m
m!
(xR)
m

(xR)
N
2
j
(N
2
j
)!

n=0
(xR)
n
n!

N
N
2
j
j
exp(N
j
)
(N
2
j
)!
exp(2x
j
),
by Stirlings formula n!

2nn
n
exp(n) . Similarly,

m=N
2
j
+1
1
m!
_

3x
4
_
m
exp(2x
j
).
Therefore,
F(xR) =
_

_
N
2
j

m=0
+

m=N
2
j
+1
_

m
m!
(xR)
m
= P
j
(x) +O(exp(2x
j
))
42
and analogously
exp
_

3x
4
_
=

P
j
(x) +O(exp(2x
j
)),
where P
j
and

P
j
are polynomials of degree N
2
j
. This yields in view of (9.4)
(x) = Q
j
(x) +O(exp(x
j
)) for x
j
1 x x
j
+ 1,
where Q
j
= P
j

P
j
is a polynomial of degree N
4
j
.
In order to nd lower bounds for Q
j
(x) we have to apply a classical theorem of A.A.
Markov [43] which states that if Q is a polynomial of degree N with real coecients
which satises the inequality
max
1x1
[Q(x)[ 1,
then
max
1x1
[Q

(x)[ N
2
;
for a proof see [43].
We return to the proof of Theorem 9.1. In view of (9.8) suppose that
:= max
x
j
1xx
j
+1
[Re Q
j
(x)[ >
1
2
exp((1

)x
j
), (9.9)
then there exists a [x
j
1, x
j
+ 1] such that Re Q
j
() = . In view of the
mean-value theorem from real analysis there exists a in between x and for which
[Re Q
j
() Re Q
j
(x)[ = [Re Q

j
()( x)[.
Dene = N
8
j
[ x[ . Markovs theorem, applied to Q(x) =
1

Re Q
j
(x x
j
) , implies
[Re Q
j
() Re Q
j
(x)[ .
If
1
2
, then
[Re Q
j
(x)[

2

1
16
exp((1

)x
j
)
for [ x[
1
2
N
8
. If (9.9) does not hold, we may argue analogously with Im Q
j
. In
any case it follows that for suciently large x
j
the intervals [x
j
1, x
j
+ 1] contains
intervals [, +] of length
1
200
N
8
j
all of whose points satisfy at least one of the
inequalities
[Re (x)[ >
1
200
exp((1

)x) , [Im (x)[ >

1
200
exp((1

)x);(9.10)
in particular
x
j
1 x
j
+ 1
1
2x
8
j
and
1
2x
8
j
2.
In order to prove the divergence of a subseries of (9.2) we note that one of the in-
equalities in (9.10) is satised innitely many often as x ; we may assume
43
that it is the one with the real part. By the prime number theorem 6.2, the inter-
val [exp(), exp( +)] contains
_
exp(+)
exp()
du
log u
+O
_
exp
_
c( +)
1
9
__
=
exp( +)
+

exp()

+O
_
exp
_
c( +)
1
9
__
=
exp()

_
exp() 1 +O
_
exp
_
c( +)
1
9
___
many primes, where c > 0 is some absolute constant. An easy computation shows
(exp( +)) (exp())
exp(x
j
)
x
9
j
.
Under these prime numbers p
k
[exp(), exp( + )] we choose those with k
0 mod 4. Since
p
k
=
k
4
, we get with view to (9.3)

k0 mod 4
log p
k
+

k
, =

k0 mod 4
log p
k
+
Re (log p
k
) exp((1

)x
j
)
exp(x
j
)
x
9
j
=
exp(

x
j
)
x
9
j
,
which diverges with x
j
. Analogously, one can create a subseries of (9.2) which
diverges to .
Thus, we have shown that the series (9.2) satsies the conditions of Theorem 8.1.
Hence, there exists a rearrangement of the series (9.2) for which

k=1
u
j
k
(s) = g
_
s

2
_
. (9.11)
Before we can nish the proof of Theorem 9.1 we have to prove the following
Lemma 9.2 Suppose that F(s) is continuous on [s[ R. Suppose that there is a
sequence of analytic functions f
n
(s) for which
lim
n
_ _
|s|r
[F(s) f
n
(s)[
2
d dt = 0,
then for any > 0 there is an integer m such that for any xed r (0, R) and any
n m
max
|s|r
[F(s) f(s)[ < .
Proof. Dene G
n
(s) = F(s) f
n
(s) . By Cauchys formula,
G
n
(s)
2
=
1
2i
_
|sz|=
G
n
(z)
2
z s
dz =
1
2
_
2
0
G
2
n
(s + exp(i)) d
44
for any R. Fix 0 < r < R. Taking the absolute modulus and integrating with
0 < Rr , we arrive at
[G
n
(s)[
2
_
Rr
0
d =
1
2
_
Rr
0
_
2
0
[G
n
(s + exp(i))[
2
d d
=
1
2
_ _
|s|R
[G
n
(s)[
2
d dt.
This yields
[G
n
(s)[
2

1
2(R r)
2
_ _
|s|R
[G
n
(s)[
2
d dt for [s[ R < r.
Now the assumption on the limit implies the estimate of the lemma.
We return to the proof of Theorem 9.1. According to (9.11),
lim
n
n

k=1
u
j
k
(s) = g
_
s

2
_
in the norm of ]
R
2
. This implies
lim
n
_ _
|s|R

g
_
s

2
_

k=1
u
j
k
(s)

2
d dt = 0
uniformly on [s[ R. Thus, application of Lemma 9.2 shows that for suciently large
m
max
|s|R

g
_
s

2
_

k=1
u
j
k
(s)

<

2
.
Now, by denition, there exists a nite set M , containing without loss of generality
all primes p y , such that
log
M
_
s +
3
4
,
0
_
:=
m

k=1
u
j
k
(s).
Hence, in view of (9.1) it follows that
max
|s|r

log
M
_
s +
3
4
,
0
_
g(s)

max
|s|r

log
M
_
s +
3
4
,
0
_
g
_
s

2
_

+ max
|s|r

g
_
s

2
_
g(s)

< .
This nishes the proof of Theorem 9.1.
45
10 Voronins universality theorem
The next and main step in the proof of the universality theorem 1.2 is to switch from
log
M
to the logarithm of the zeta-function.
Theorem 10.1 Let 0 < r <
1
4
and suppose that g(s) is continuous on [s[ r and
analytic in the interior. Then, for any > 0,
liminf
T
1
T
meas
_
[0, T] : max
sK

log
_
s +
3
4
+i
_
g(s)

<
_
> 0.
Note that log (s) has singularities at the zeros of (s) . The truth of Riemanns
hypothesis would imply that all such singularities lie to the left of the strip of univer-
sality
1
2
< < 1. However, unconditionally the set of such singularities has in view of
the density theorem 4.4 zero density. Therefore the existence of these singularities is
negligeable for our observations.
Proof. We choose > 1 and (0, 1) such that r <
1
4
and
max
|s|r

g
_
s

_
g(s)

<
1
.
Set Q = p z and let S = s : r < 2, 1 t t. We shall estimate
J :=
_
2T
T
_ _
E

1
Q
_
s +
3
4
+i, 0
_

_
s +
3
4
+i
_
1

2
d dt d,
where 0 = (0, 0, . . .) . By Theorem 3.7,
(s +i) =

nT
1
n
s+i
+O(T

).
This gives
J =
_ _
E+
3
4
_
2T
T
[
1
Q
(s +i, 0)(s +i) 1[
2
d d dt

_ _
E+
3
4
_
2T
T

1
Q
(s +i, 0)

nT
1
n
s+i
1

2
d d dt
+
_ _
E+
3
4
_
2T
T
T

[
1
Q
(s +i, 0)[
2
d d dt, (10.1)
where S +
3
4
is the set of all s with s
3
4
S . By denition,

1
Q
(s, 0) =

pQ
_
1
1
p
s
_
=

m=1
p|mpQ
(m)
m
s
.
46
Obviously, we may bound the second term appearing on the right hand side of (10.1)
by
T
2(
3
4
r)
max
sE+
3
4
_
2T
T
[
1
Q
(s +i, 0)[
2
d T
2r
1
2

1
Q
_
3
4
r, 0
_

2
.
Furthermore, for T > z a simple computation gives

1
Q
(s, 0)

nT
1
n
s
= 1 +

z<kz
z
T
b
k
k
s
with b
k
=

m|k
p|mpQ;kmT
1.
By a classical estimate for the divisor function from elementary number theory

0
(n) n

(10.2)
(for a proof see [25]). Thus, similarly as in (4.7) we have
[b
k
[
0
(k) k

for any > 0. (10.3)

Hence, for T > z
_
2T
T

1
Q
(s +i, 0)

nT
1
n
s+i
1

2
d =
_
2T
T

z<kz
z
T
b
k
k
s

2
d
= T

z<kz
z
T
[b
k
[
2
k
2
+O
_
_

0<l<kz
z
T
[b
k
b
l
[
(kl)

_
2T
T
_
k
l
_
i
d

_
_
,
Using estimate (10.3) with =

1
2
, the above is bounded by
T

k>z

2
0
(k)
k
2
+

0<l<kz
z
T

0
(k)
0
(l)
(kl)

log
k
l
Tz
12+
1
+ (z
z
T)

0<l<kz
z
T
1
(kl)

log
k
l
.
The appearing sum can be estimated by ((z
z
T)
22
+ 1) log
2
(z
z
T) as we did in the
proof of Theorem 4.1. Thus, we nally arrive at
_ _
E+
3
4
_
2T
T

1
Q
(s +i, 0)

nT
1
n
s+i
1

2
d d dt

_ _
E+
3
4
_
Tz
12+
1
+ (z
z
T)

1
((z
z
T)
22
+ 1) log
2
(z
z
T)
_
d dt
z
2r+
1

1
2
T.
In view of (10.1) we conclude that for any
2
> 0 there exists z
0
such that
J
4
2
T, (10.4)
provided that z > z
0
and T suciently large, say T > T
0
, depending on
2
and z .
Dene
/
T
=
_
[T, 2T] :
_ _
E+
3
4
[
1
Q
(s +i, 0)(s +i) 1[
2
d dt
2
2
_
.
47
Then it follows from (10.1) and (10.4) that for suciently large z and T
meas /
T
> (1
2
)T, (10.5)
which is surprisingly large; this idea goes back to Bohr [7]. Application of Lemma 9.2
gives for /
T
max
|s|r
[
1
Q
(s +i, 0)(s +i) 1[ < C
2
,
where C is a positive constant, depending only on . For suciently small
2
we
deduce
max
|s|r

log
_
s +
3
4
+i
_
log
Q
_
s +
3
4
+i, 0
_

< 2C
2
, (10.6)
provided /
T
; here we used a truncated Taylor expansion of the exponential
function expz = 1 +z +O([z[
2
) .
By Theorem 9.1 there exists a sequence of nite sets of prime numbers M
1
M
2

. . . such that

k=1
M
k
contains all primes and
lim
k
max
|s|r

log
M
k
_
s +
3
4
,
0
_
g
_
s

= 0. (10.7)
Let

= (

2
,

3
, . . .) . By the continuity of log
M
_
s +
3
4
,
0
_
, for any
1
> 0 there
exists a positive for which, whenever
|
p

p
| < for all p M
k
, (10.8)
then
max
|s|r

log
M
k
_
s +
3
4
,
0
_
log
M
k
_
s +
3
4
,

< . (10.9)
Setting
B
T
=
_
[T, 2T] :
_
_
_
log p
2

p
_
_
_ <
_
,
we get
1
T
_
B
_ _
|s|r

log
Q
_
s +
3
4
+i, 0
_
log
M
k
_
s +
3
4
+i, 0
_

2
d dt d
=
_ _
|s|r
1
T
_
B
T

log
Q
_
s +
3
4
+i, 0
_
log
M
k
_
s +
3
4
+i, 0
_

2
d d dt.
Putting () =
_

log 2
2
,
log3
2
, . . .
_
, we may rewrite the inner integral as
_
B
T

log
Q
_
s +
3
4
, ()
_
log
M
k
_
s +
3
4
+i, ()
_

2
d.
48
Application of Theorem 7.1 to the curve () =
_

log 2
2
,
log3
2
, . . . ,
logp
N
2
_
(the loga-
rithms of the prime numbers are linearly independent as we have seen in the proof of
Theorem 7.3) yields
lim
T
1
T
_
B
T

log
Q
_
s +
3
4
, ()
_
log
M
k
_
s +
3
4
+i, ()
_

2
d
=
_
D

log
Q
_
s +
3
4
,
_
log
M
k
_
s +
3
4
+i,
_

2
d
uniformly in s for [s[ r , where T is the subregion of the unit cube in R
N
given by
the inequalities (10.8) and d is the Lebesgue measure. By the denition of
M
(s, )
it follows that for M
k
Q

Q
(s, ) =
M
k
(s, )
Q\M
k
(s, ),
and thus
_
D

log
Q
_
s +
3
4
,
_
log
M
k
_
s +
3
4
+i,
_

2
d
=
_
D

log
Q\M
k
_
s +
3
4
,
_

2
d = meas T
_
[0,1]
N

log
Q\M
k
_
s +
3
4
,
_

2
d.
Since
log
Q\M
k
_
s +
3
4
,
_
=

pQ\M
k

n=1
exp(2i
p
)
np
n(s+
3
4
)
,
we obtain
_
[0,1]
N

log
Q\M
k
_
s +
3
4
,
_

2
d =

pQ\M
k

n=1
1
n
2
p
2+
3n
2
.
If M
k
contains all primes y
k
, then

pQ\M
k

n=1
exp(2i
p
)
n
2
p
2n+
3n
2
y
2r
1
2
k
.
Hence, we nally get
1
T
_
B
T
_ _
|s|r

log
Q
_
s +
3
4
+i, 0
_
log
M
k
_
s +
3
4
+i, 0
_

2
d dt d
y
2r
1
2
k
meas T.
A further application of Theorem 7.1 shows
lim
T
1
T
meas B
T
= meas T,
which implies for y
k
suciently large
meas
_
B
T
:
_ _
|s|r

log
Q
_
s +
3
4
+i, 0
_
log
M
k
_
s +
3
4
+i, 0
_

2
d dt
< y
r
1
4
k
_
>
meas T
2
T.
49
The curve () is uniformly distributed mod1. Thus, application of Theorem 7.4
yields
meas
_
B
T
: max
|s|r

log
Q
_
s +
3
4
+i, 0
_
log
M
k
_
s +
3
4
+i, 0
_

< y
1
5
(r
1
4
)
k
_
>
meas T
2
T. (10.10)
If we now take 0 <
2
<
1
2
meas T, then (10.5) implies
meas /
T
B
T
> 0.
Thus, in view of (10.7) we may approximate g
_
s

_
by log
M
k
_
s +
3
4
, 0
_
(independent
on ), with (10.9) and (10.10) the latter function by log
Q
_
s +
3
4
, 0
_
, and nally
with regard to (10.6) by log
_
s +
3
4
+i
_
on a set of with positive measure. The
theorem is proved.
It is obvious what we have to do to get rid of the logarithm in Theorem 10.1. Let
f(s) = exp(g(s)) ; it is fundamental in complex analysis that every analytic function
f(s) without zeros has an analytic logarithm g(s) . Obviously,
f(s)
_
s +
3
4
_
= f(s)
_
1 exp
_
log
_
s +
3
4
_
g(s)
__
.
Let > 0 and > 1 such that r <
1
4
and
max
|s|r

f(s) f
_
s

< .
Then, since expz = 1 z +O([z[
2
) , we obtain
max
|s|r

_
s +
3
4
+i
_
f
_
s

= max
|s|r

_
s +
3
4
+i
_
exp
_
g
_
s

__

max
|s|r

f
_
s

max
|s|r

exp
_
log
M
_
s +
3
4
,
0
_
g
_
s

__
1

max
|s|r

f
_
s

max
|s|r

log
M
_
s +
3
4
,
0
_
g
_
s

,
which can be made suciently small by Theorem 10.1. Thus, we nally have proved
Voronins theorem:
Corollary 10.2 Let 0 < r <
1
4
and suppose that f(s) is a continuous non-vanishing
function on [s[ r which is analytic in the interior. Then, for any > 0,
liminf
T
1
T
meas
_
[0, T] : max
|s|r

_
s +
3
4
+i
_
f(s)

<
_
> 0.
50
Surprisingly, the set of translates on which (s) approximate a given function f(s)
with arbitrary precision has a positive lower density! This improves Theorem 1.2 from
the introduction signicantly.
Bagchi [3] extended Voronins result signicantly in dierent ways after some rst
progress due to Reich [47]. By that it was possible to replace the discs by arbitrary
compact subsets of the strip
1
2
< < 1 with connected complement. Actually, Bagchi
found a new and very transparent proof by using limit theorems for weakly convergent
probability measures. This approach was completed and extended by Laurincikas in
various details and directions; see [28]. The strongest version of Voronins universality
theorem is
Theorem 10.3 Let / be a compact subset of
1
2
< < 1 with connected complement
and suppose that f(s) is a continuous non-vanishing function on / which is analytic
in the interior. Then, for any > 0,
liminf
T
1
T
meas
_
[0, T] : max
sK
[(s +i) f(s)[ <
_
> 0.
11 Functional independence
We state some consequences of the universality property of the zeta-function. We start
with a classical result due to Bohr [7], namely that the set of values taken by (s) on
a vertical line (
1
2
, 1) lies dense in the complex plane. This can be extended to
Theorem 11.1 Let
1
2
< < 1 be xed, then the sets
(log (s), (log (s))

, . . . , (log (s))
(n1)
) : t R
and
((s),

(s), . . . ,
(n1)
(s)) : t R
lie everywhere dense in C
n
.
Proof. Suppose that we are given a vector (b
0
, b
1
, . . . , b
n1
) C
n
. Let
r =
1
4

1
2
min
_

1
2
, 1
_
and dene
g(s) =
n1

k=0
b
k
k!
s
k
.
Obviously, g
(k)
(0) = b
k
for k = 0, 1, . . . , n 1. By Cauchys formula, we have further
f
(k)
(0) =
k!
2i
_
|s|=
f(s)
s
k+1
ds (11.1)
51
for any > 0. With view to Theorem 10.1 the function g(s) can be approximated to
arbitrary precision on the disc [s[ r by log
_
s +
3
4
+i
_
for some . Hence, taking
f(s) = g(s) log
_
s +
3
4
+i
_
and < r in (11.1), shows that (log (s), log

(s), . . . , log (s)

(n1)
) with
1
2
< < 1
lies somewhere arbtrarily close to (g(0), g

(0), . . . , g
(n1)
(0)) = (b
0
, b
1
, . . . , b
n1
) . This
implies the statement for the rst set.
We use induction on m to prove that for any (m+1) -tuple (a
0
, a
1
, . . . , a
m
) C
m+1
,
where a
0
,= 0, there exists (b
0
, b
1
, . . . , b
m
) C
m+1
for which
exp
_
m

k=0
b
k
s
k
_

k=0
a
k
k!
s
k
mod s
m+1
.
For m = 0 one only has to choose b
0
= log a
0
. By the induction assumption, we may
assume that with some
exp
_
m

k=0
b
k
s
k
_

k=0
a
k
k!
s
k
+s
m+1
mod s
m+2
.
Thus,
exp
_
m

k=0
b
k
s
k
+s
m+1
_
(1 +s
m+1
)
_
m

k=0
a
k
k!
s
k
+s
m+1
_
mod s
m+2
.
Hence, let b
m+1
= be the solution of the equation
a
0
+ =
a
m+1
(m + 1)!
,
which exists by the restriction on a
0
. This shows
exp
_
m+1

k=0
b
k
s
k
_

m+1

k=0
a
k
k!
s
k
mod s
m+2
,
proving the claim.
Now
f(s) := exp
_
n1

k=0
b
k
s
k
_

m+1

k=0
a
k
k!
s
k
mod s
n
.
By Voronins universality theorem, Corollary 10.2, there exists a sequence
j
, tending
with j to innity, such that
lim
j
max
|s|r

_
s +
3
4
+i
j
_

= 0
for some r (0,
1
4
) . In view of (11.1) we obtain
lim
j
max
|s|r

(k)
_
s +
3
4
+i
j
_
f
(k)
(s)

= 0
for k = 1, . . . , n 1 and any (0, r) . Arguing as above, this proves the theorem.
Further, the universality result implies functional independence:
52
Theorem 11.2 Let z = (z
0
, z
1
, . . . , z
n1
) C. Suppose that F
0
(z), F
1
(z), . . . , F
N
(z)
are continuous functions, not all identically zero, then
N

k=0
s
k
F
k
((s), (s)

, . . . , (s)
(n1)
) ,= 0
for some s C,
In particular, we see that the zeta-function does not satisfy any algebraic functional
equation. This solves one of Hilberts famous problems which he posed at the Inter-
national Congress of Mathematicians in Paris 1900. The rst proof of this result was
given by Ostrowski [41].
Proof. First, we shall show that if F(z) is a continuous function and
F((s), (s)

, . . . , (s)
(n1)
)) = 0
identically in s C, then F vanishes identically.
Suppose the contrary, i.e. F(z) , 0. Then there exists a C
n
for which F(a) ,= 0.
Since F is continuous, there exist a neighbourhood U of a and a positive such that
[F(z)[ > for z U.
Choosing an arbitrary (
1
2
, 1) , application of Theorem 11.1 yields the existence of
some t for which
((s), (s)

, . . . , (s)
(n1)
) U,
which contradicts our assumption. This proves our claim, resp. the assertion of the
theorem with N = 0.
Without loss of generality we may assume that F
0
(z) is not identically zero. As
above there exist an open bounded set U and a positive such that
[F
0
(z)[ > for z U.
Denote by M the maximum of all indices m for which
sup
zU
[F
m
(z)[ , = 0.
If M = 0, then the assertion of the theorem follows from the result proved above.
Otherwise, we may take a subset V U such that
inf
zV
[F
M
(z)[ >
for some positive . By Theorem 11.1, there exists a sequence t
j
, tending with j to
innity, such that
(( +it
j
), ( +it
j
)

, . . . , ( +it
j
)
(n1)
) V.
This implies
lim
j

k=0
( +it
j
)
k
F
k
(( +it
j
), ( +it
j
)

, . . . , ( +it
j
)
(n1)
)

= .
This proves the theorem.
53
12 Self-similarity and the Riemann hypothesis
In view of Voronins universality theorem a natural question arises: is it possible to
approximate functions with zeros? The answer is more or less negative but of special
interest. We give an heuristic argument which however can be made waterproof with
a bit more eort and the same techniques which we shall use later on.
In order to see that the zeta-function cannot approximate uniformly a function with
zeros recall Rouches theorem, which states that if f(s) and g(s) are analytic inside
and on a contour ( , and [f(s)[ < [g(s)[ on ( , then g(s) and f(s) + g(s) have the
same number of zeros inside ( ; for a proof see [54], 3.42.
Now assume that f(s) is an analytic function on [s[ r , where 0 < r <
1
4
,
which has a zero with [[ < r but which is non-vanishing on the boundary. For
> 0 suciently small we may assume that max
|s|=r
[f(s)[ > . Hence, whenever the
inequality
max
|s|r

_
s +
3
4
+i
_
(s)

< < min

|s|r
[(s)[; (12.1)
holds,
_
s +
3
4
+i
_
has to have a zero inside [s[ r (since by the maximum principle
the maximum on the left hand side is actually taken on the boundary). Note that the
second inequality in (12.1) holds for suciently small (since the zeros of an analytic
function lie discrete or the function vanishes identically). Therefore, if for any > 0
liminf
T
1
T
meas
_
[0, T] : max
|s|r

_
s +
3
4
+i
_
f(s)

<
_
> 0,
then we expect T many complex zeros of (s) in the strip
3
4
r < <
3
4
+r (for
a rigorous proof one has to consider exactly the densities of values satisfying (12.1);
this can be done along the lines of the proof of Theorem 12.1 below). This contradicts
the density theorem 4.4, which gives
N
_
3
4
r, T
_
= o(T).
Thus, an approximation of a function with a zero on a suciently rich set cannot be
done!
The above reasoning shows that the location of the complex zeros of the zeta-
function is closely connected with its universality property. We can go a little bit
further.
Theorem 12.1 The Riemann hypothesis is true if and only if for any compact subset
/ of
1
2
< < 1 with connected complement and any > 0
liminf
T
1
T
meas
_
[0, T] : max
sK
[(s +i) (s)[ <
_
> 0. (12.2)
54
This theorem is due to Bagchi [3]; in [4] he generalized this result in various directions.
Since Bagchis proof relies mainly on the theory of topolgical dynamics he speaks about
the property (12.2) as strong recurrence. However, we call it self-similarity. It should be
noted that Bohr [8] detected in 1922 a similar result. Therefore, we have to introduce
an important class of zeta-functions.
For a character mod q (i.e. a non-trivial group homomorphism on the prime
residue class mod q ) the Dirichlet L-function is for > 1 given by
L(s, ) =

p
_
1
(p)
p
s
_
1
=

n=1
(n)
n
s
.
Dirichlet L-functions have similar behaviour and properties as the Riemann zeta-
function; actually, (s) may be regarded as Dirichlet L-function associated to the
principal character
0
mod 1. As for the Riemann zeta-function it is conjectured
that L(s, ) does not vanish in >
1
2
, this is the so-called Generalized Riemann
hypothesis. Harald Bohr introduced the fruitful notion of almost periodicity into
analysis. We say that a function L(s) is almost periodic in / if for all > there
exists a sequence of values . . . ,
1
< 0 <
1
<
2
< . . . ... with
liminf
m
(
m+1

m
) > 0 and limsup
m

m
[m[
<
for which
[L(s +i
m
) L(s)[ < for all s /.
Bohr proved that Dirichlet series are almost periodic in their half-plane of absolute
convergence. Moreover, he discovered an interesting relation between Riemanns hy-
pothesis and almost periodicity: if is a non-trivial character, then the Riemann
hypothesis for L(s, ) is true if and only if L(s, ) is almost periodic for >
1
2
. Note
that self-similarity implies almost periodicity (but not vice versa). The condition on
the character seems somehow unnatural but Bohrs argument does not apply to (s) .
However, by Voronins universality theorem this gap can be lled. We shall give a
simple proof of Bagchis theorem which actually combines Bohrs idea with the one of
Bagchi.
Proof. If Riemanns hypothesis is true, then we can apply the universality theorem
10.3 with g(s) = (s) , which implies the self-similarity. The idea for the proof of the
other implication is that if there is at least one proof to the right of the critical line,
then the self-similarity property implies the existence of many zeros, too many with
regard to well-known density theorems).
Suppose that the Riemann hypothesis is not true, then there exists a zero of (s)
with Re >
1
2
. Further, we We have to show that there exists a disc [s[ r <
1
4
and
> 0 such that
lim
T
1
T
meas
_
[0, T] : max
sK
[(s +i
_
(s)[ < = 0. (12.3)
55
Locally, the zeta-function has the expansion
(s) = c(s )
m
+O
_
[s [
m+1
[
_
(12.4)
with some non-zero c C and m N. Now assume for a neighbourhood /

:= s
C : [s of that
max
sK

[(s +i) (s)[ < < min

|s|=
[(s)[; (12.5)
this formula should be compared with (12.1). Then Rouches theorem implies the
existence of a zero of (s + i) in /

. We may say that the zero of (s)

generates the zero of (s +i) . With regard to (12.4) and (12.5) the zeros and
are intimately related, more precisely:
> [() ( i)[ = [( i)[
[c[ [ i [
m
+O
_
[ i [
m+1
_
.
Hence,
[ i [
_

[c[
_ 1
m
+O
_

1+
1
m
_
,
and in particular
1
2
< Re 2
_

[c[
_ 1
m
< Re < 1,
[Im ( + Im )[ < 2
_

[c[
_ 1
m
for suciently small = o(
m+1
) . It may happen that dierent values of for which
(12.5) hold lead to the same zero . Therefore, we have to consider the densities for
such . If we now write
J :=
_
j
J
j
:= [0, T] : max
sK

[(s +i) (s)[ < ,

where the J
j
are disjoint intervals, then it follows that there are

_
_
1
4
_
[c[

_ 1
m
meas J
j
_
_
+ 1 >
1
4
_
[c[

_ 1
m
meas J
j
many zeros according to J
j
, and thus
N
_
_
Re 2
_

[c[
_ 1
m
, T + Im + 2
_

[c[
_ 1
m
_
_

1
4
_
[c[

_ 1
m
meas J;
recall that the zero-counting function N(, T) was dened in Section 4. By the density
theorem 4.4 we obtain meas J = o(T) , which implies (12.3). The theorem is proved.

56
The critical line is a natural borderline for self-similarity since Levinson [36] proved
that the frequency of c-values of the zeta-function, i.e. solutions of (s) = c, close
to the critical line increases rapidly with increasing imaginary part; similar as the
Riemann von-Mangold formula (2.5) in connection with Theorem 4.3 show for the
particular case of zeros.
Assuming Riemanns hypothesis the self-similarity property (12.2) has an interest-
ing interpretation. The amplitude of light waves is a physical bound for the size of
objects which human beings can see (even with microscopes), or take the Planck con-
stant 10
33
which is the smallest size of objects in quantum mechanics. Thus, if we
assume that is less than this quantity, then we cannot distinguish between (s) and
(s +i) whenever
max
sK
[(s +i) (s)[ < .
This shows that even if one would have all knowledge on the zeta-function, one could
not decide wherever one actually is in the analytic landscape of (s) above the right
half of the critical strip without moving to the boundary. The zeta-function as an
amazing maze!
Theorem 12.1 oers an interesting approach towards Riemanns hypothesis. How-
ever, we shall only prove, following Bohrs argument, that the zeta-function is self-
similar in the half-plane of absolute convergence.
Theorem 12.2 Let / be any compact subset in the half-plane > 1. Then, for any
> 0,
liminf
T
1
T
meas
_
[0, T] : max
sK
[(s +i) (s)[ <
_
> 0.
Proof. Since (s) is regular and zero-free in > 1, we may dene the logarithm
(by choosing any one of the values of the logarithm). In view of the Euler product
representation it is easily shown that
log (s) =

p,k
1
kp
ks
for > 1,
where the sum is taken over all prime numbers p and all positive integers k. Hence,
log (s) log (s +i) =

p,k1
1
kp
ks
_
1
1
p
ik
_
. (12.6)
We shall use diophantine approximation to nd values of for which p
ik
lies suf-
ciently close to 1. We apply Theorem 7.1 with a
n
=
1
2
log p
n
where, as usual, p
n
denotes the n-th prime number. Further, we choose
=
_
(z
1
, . . . , z
N
) R
N
: |z
n
| <
1

for 1 n N
_
,
where the parameter > 0 will be chosen later and where |z| denotes the distance
of z to the next integer. Then we nd about
= T
_
4

_
N
57
many values of [0, T] such that
| log p
n
| <
2

for 1 n N
as T tends to innity. Consequently,
cos(k log p) = 1 +O
_
k
2

2
_
, sin(k log p) = O
_
k

_
,
resp.
[1 p
ik
[
2
= (1 cos(k log p))
2
+ sin
2
(k log p)
k
2

2
for p p
N
, provided that > 4k; here and in the sequel all implicit constants are
absolute. This yields

px,ky
1
p
s
_
1
1
p
i
_

xy

,
where x = p
N
. Furthermore, we have

px,k>y
1
kp
ks
_
1
1
p
ik
_

1
y

px
1
p
y

x
y2
y
,
and

p>x,k1
1
kp
ks
_
1
1
p
ik
_

n>x
1
n

x
1
1
.
In view of (12.6) we obtain
log (s) log (s +i)
xy
2

+ +
x
y2
y
+
x
1
1
(12.7)
for a set of values of [0, T] with positive lower density as T and any > 1;
note that the estimates can easily be improved. Since / is compact, there exists
:= min : s /. Then
max
sK
[(s)[ ().
Now for any given

> 0 we can nd values x, y and such that the right hand side
of (12.7) is <

. It remains to get rid of the logarithm. Obviously,

(s) (s +i) = (s)(1 explog (s +i) log (s)).
Hence,
max
sK
[(s) (s +i)[ () max
sK
[ log (s +i) log (s)[ <

().
The choice

=

2M
proves the theorem.
Taking Theorems 12.1 and 12.2 into account, we might conjecture that the zeta-
function has the above self-similarity property for >
1
2
.
58
13 Eective bounds
However, the known proofs of Voronins universality theorem are ineective, giving
neither an estimate for the rst approximating translate nor lower bounds for the
positive lower density. This is caused by Pecherskys ineective theorem 8.1 on the
rearrangement of series. In this section we shall prove for a suciently large class of
functions upper bounds for the upper density; for more on this topic see [52].
Denote by B
r
the closed disc of radius r > 0 with center in the origin. For our
purpose we consider analytic isomorphisms g : B
r
B
1
, i.e. the inverse g
1
exists and is analytic. Obviously, such an analytic isomorphisms g has exactly one
simple zero in the interior of B
r
; moreover, it can be shown (by the Schwarz lemma,
see [54], 7.2) that g /
r
has the representation
g(s) = r exp(i)
s
r
2
s
with R, [[ < r. (13.1)
Denote by /
r
the class of analytic isomorphisms from B
r
onto the unit disc. Further,
dene for an analytic isomorphisms g /
r
with xed r
_
0,
1
4
_
and a positive the
upper density
d(, g) := limsup
T
1
T
meas
_
[0, T] : max
|s|r

log
_
s +
3
4
+i
_
g(s)

<
_
and the lower density d(, g) analogously; note that ( d(, g) > 0 implies the univer-
sality property of Corollary 10.2 with respect to g . Finally, let N(
1
,
2
, T) count the
number of zeros of log (s) in
1
2
<
1
< <
2
< 1 (according multiplicities). Then
Theorem 13.1 Suppose that g /
r
and d(, g) > 0 for all > 0. Then, for any

_
0,
1
2r
_
1
4
+ Re [[
__
,
d(, g)
8r
4

r
2
[[
2
limsup
T
1
T
N
_
3
4
+ Re 2r,
3
4
+ Re + 2r, T
_
(13.2)
= o().
Therefore, the decay of d(, g) with 0 is more than linear in .
Proof. The idea of proof is (as in the proof of Theorem 12.1) that the zero of g
generates some zeros of log (s) in
1
2
< < 1. Since g maps the boundary of B
r
onto
the unit circle, Rouches theorem implies the existence of one simple zero of log (z)
in
/

:=
_
z =
3
4
+s +i : s B
r
_
whenever (12.1) holds.
Universality (as self-similarity) is a phenomenon that happens in intervalls. Now
we have to prove an upper bound for the distance of dierent translates generating the
59
same zero of log (s) : suppose that a zero of log (s) , generated by , lies in two
dierent translates /

1
and /

2
, then
[
1

2
[ <
8r
4

r
2
[[
2
. (13.3)
Suppose that there exist
s
j
= Re
3
4
+it
j
B
r
, and
j
R with log
_
s
j
+
3
4
+i
j
_
= 0,
for j = 1, 2, such that
= s
1
+
3
4
+i
1
= s
2
+
3
4
+i
2
.
In view of (13.1),
[g(s
2
) g(s
1
)[ =
r
2
[[
2
[r
2
s
1
[[r
2
s
2
[
[s
2
s
1
[.
We deduce from (12.1) that [g(s
j
)[ < for j = 1, 2, and therefore
[
1

2
[ = [t
2
t
1
[
4r
4
r
2
[[
2
[g(s
2
) g(s
1
)[ <
8r
4

r
2
[[
2
,
which proves estimate (13.3).
Now, denote by J
j
(T) the disjoint intervalls in [0, T] such that (12.1) is valid
exactly for

_
j
J
j
(T) =: J(T).
Using (13.3), in every intervall J
j
(T) lie at least
1 +
_
r
2
[[
2
8r
4

meas J
j
(T)
_

r
2
[[
2
8r
4

meas J
j
(T)
zeros of log (s) in the strip
1
2
< < 1. Therefore, the number A(T) of such zeros
satises the estimate
8r
4

r
2
[[
2
A(T) meas J(T). (13.4)
The next step is to replace A(T) by the zero counting function appearing in the
theorem.
Obviously, the value distribution of log (z) in /

is ruled by that of g(s) in B

r
.
As we shall see below, this gives a restriction on the real parts of zeros . Let s B
r
.
If [g(s)[ , then, in view of (12.1),

log
_
s +
3
4
+i
_

[g(s)[

g(s) log
_
s +
3
4
+i
_

> 0.
60
Since (13.1) implies
[g(s)[
[ s[
2r
,
we obtain the estimate

Re
3
4
Re

< 2r (13.5)
by taking the real parts.
In view of (13.5) we nd
A(T) N
_
3
4
+ Re 2r,
3
4
+ Re + 2r, T
_
. (13.6)
On the other side, since d(, g) > 0, there exists an incresing sequence (T
k
) with
lim
k
T
k
= such that for any > 0
meas J(T
k
) (d(, g) )T
k
.
Consequently, this together with (13.6), leads in (13.4) to
8r
4

r
2
[[
2
N
_
3
4
+ Re 2r,
3
4
+ Re + 2r, T
k
_
(d(, g) )T
k
.
Sending 0, yields the estimate (13.2) of the theorem. As it was shown by Bohr
and Jessen the limit
lim
T
1
T
N
_
3
4
+ Re ,
3
4
+ Re +, T
_
exists, and tends with to zero (see Hilfssatz 6, [9]). Further, one has
max
sBr

_
s +
3
4
+i
_
expg(s)

max
sBr
[ expg(s)[ max
sBr

exp
_
log
_
s +
3
4
+i
_
g(s)
_
1

e max
sBr

_
s +
3
4
+i
_
expg(s)

.
The Theorem is shown.
Recently, Garunkstis [17] proved a rst eective universality theorem along the lines
of Voronins proof in addition with some old ideas due to Good [19] and new insights.
In particular, his remarkable result shows that if f(s) is analytic in [s[ 0.05 with
max
|s|0.06
[f(s)[ 1, then for any 0 < <
1
2
there exists a
exp
_
exp
_
10
13
__
such that
liminf
T
1
T
meas
_
[0, T] : max
|s|0.0001

_
s +
3
4
+i
_
f(s)

<
_
exp
_

13
_
.
Another approach via the rate of convergence of weak convergent probability measures
is due to Laurincikas [31].
61
14 Other zeta-functions
It is natural to ask whether other functions with similar properties as the Riemann
zeta-function are universal. Meanwhile, it is known that there exists a rich zoo of
Dirichlet series having the universality property; we mention only some further signif-
icant examples:
Joint universality for Dirichlet L-functions, i.e. for a collection of Dirichlet L-
functions attached to pairwise non-equivalent characters mod q (nontrivial
group homomorphisms on the group of prime residue classes), proved by Voronin
[58];
Dirichlet L-functions with respect to the modulus of the characters, by Eminyan
[15];
Dedekind zeta-functions associated to a number eld K over Q

K
(s) =

A
1
(N/)
s
=

P
_
1
1
(NT)
s
_
1
,
where the sum is taken over all non-zero integral ideals, the product is taken over
all prime ideals of the ring of integers of K and where N/ is the norm of the
ideal /, obtained by Reich [48];
Dirichlet series with multiplicative coecients, by Laurincikas [27], Laurincikas
and

Slezevicien e [35];
Matsumoto zeta-functions, obtained by Laurincikas [30];
L-functions associated to cusp forms, resp. new forms (or elliptic curves by Wiles
et al. celebrated proof of Fermats last theorem), by Laurincikas and Matsumoto
[33], resp. Laurincikas, Matsumoto and Steuding [34];
Hurwitz zeta-functions
(s, ) =

m=0
1
(m +)
s
(14.1)
with parameter (0, 1] where is rational ,=
1
2
, 1 or transcendental, proved
by Gonek [18], Bagchi [3];
Lerch zeta-functions, proved by Laurincikas [29];
This list could be continued with Hecke L-functions, Artin L-functions, Rankin-
Selberg convolution L-functions to mention only some more important examples.
Some of the examples given above, e.g. Hurwitz zeta-functions, have the strong
universality property, i.e. that they can approximate functions with zeros. It seems
62
that the restriction to a conditional universality property is intimately linked to the
property of the Dirichlet series in question to have an Euler product, and by that, to
satisfy Riemanns hypothesis. It should be noted that the strong universality leads by
the arguments used in the proof of Theorem 12.1 to the existence of many zeros o the
critical line; see [16] for more details.
All known proofs of universality for zeta-functions depend on a certain indepen-
dence, namely that the logarithms of the prime numbers are linearly independent (as
we have used it in the proof of Voronins universality theorem when we applied Kro-
neckers approximation theorem 7.1), resp. that the numbers log(n+) for n N are
linearly independent if is transcendental (which is necessary to deal with (14.1)).
The Linnik-Ibragimov conjecture states that all functions given by Dirichlet series
and analytically continuable to the left of the half plane of absolute convergence are
universal. However, this is better to understand as a program than a conjecture. For
example, dene a(n) = 1 if n = 2
k
, k N 0, and a(n) = 0 otherwise, then
A(s) =

n=1
a(n)
n
s
=

k=0
1
2
sk
=
1
1 2
s
,
and obviously, this function is far away from beeing universal. However, in [53] a
universality theorem for the so-called Selberg class, which covers all from a number
theoretical point of view interesting Dirichlet series known so far (i.e. with Euler
product), was proved.
More detailed surveys on the value distribution of zeta-functions are the excellent
written papers [32] and [40].
63
References
[1] N.I. Achieser, Lectures on the theory of approximations, New York: Ungar Pub.
Co. 1956
[2] T.M. Apostol, Mathematical Analysis, Addison-Wesley 1974, 2nd ed.
[3] B. Bagchi, The statistical behaviour and universality properties of the Riemann
zeta-function and other allied Dirichlet series, Ph.D.Thesis, Calcutta, Indian Sta-
tistical Institute, 1981
[4] B. Bagchi, Recurrence in topological dynamics and the Riemann hypothesis, Acta
Math. Hungar. 50 (1987), 227-240
[5] B.C. Berndt, The number of zeros of
(k)
(s) , J. London Math. Soc. 2 (1970),
577-580
[6] G.D. Birkhoff, Demonstartion dun theor`eme elementaire sur le fonctions
entieres, C. R. Acad. Sci. Paris 189 (1929), 473-475
[7] H. Bohr, Zur Theorie der Riemannschen Zetafunktion im kritischen Streifen,
Acta Math. 40 (1915), 67-100
[8] H. Bohr,

Uber eine quasi-periodische Eigenschaft Dirichletscher Reihen mit An-
wendung auf die Dirichletschen L-Funktionen, Math. Ann. 85 (1922), 115-122
[9] H. Bohr, B. Jessen,

Uber die Werteverteilung der Riemannschen Zetafunktion,
Acta Math. 58 (1932), 1-55
[10] H. Bohr, E. Landau,

Uber das Verhalten von (s) und
(k)
(s) in der N ahe der
Geraden = 1, Nachr. Ges. Wiss. Gottingen Math. Phys. Kl. (1910), 303-330
[11] H. Bohr, E. Landau, Ein Satz uber Dirichletsche Reihen mit Anwendung auf
die -Funktion und die L-Funktionen, Rend. di Palermo 37 (1914), 269-272
[12] H. Bohr, E. Landau, Sur les zeros de la fonction (s) de Riemann, C.R. 158
(1914), 106-110
[13] H. Bohr, E. Landau, Nachtrag zu unseren Abhandlungen aus den Jahren 1910
und 1923, Nachr. Ges. Wiss. Gottingen Math. Phys. Kl. (1924), 168-172
[14] N. Dunford, J.T. Schwartz, Linear Operators, New York Interscience 1958
[15] K.M. Eminyan, -universality of the Dirichlet L-function, Mat. Zametki 47
(1990), 132-137 (Russian); translation in Math. Notes 47 (1990), 618-622
[16] R. Garunkstis, A. Laurin cikas, The Lerch zeta-function, Kluwer Academic
Publishers, Dordrecht (to appear)
64
[17] R. Garunkstis, The eective universality theorem for the Riemann zeta-
function, preprint .
[18] S.M. Gonek, Analytic properties of zeta and L-functions, Ph.D. thesis, Univer-
sity of Michigan 1979
[19] A. Good, On the distribution of the values of Riemanns zeta-function, Acta
Arith. 38 (1981), 347-388
[20] S.A. Gritsenko, Sharpening a constant in a density theorem, Mat. Zametki 55
(1994), 59-61 (Russian)
[21] K.G. Grosse-Erdmann, Universal families and hypercyclic operators, Bull.
A.M.S. 36 (1999), 345-381
[22] G.H. Hardy, J.E. Littlewood, The zeros of Riemanns zeta-function on the
critical line, Math. Zeit. 10 (1921), 283-317
[23] M.N. Huxley, Exponential sums and the Riemann zeta-function IV, Proc. Lon-
don Math. Soc. 66 (1993), 1-40
[24] A.E. Ingham, The distribution of prime numbers, Cambridge University Press
1932
[25] A.A. Karatsuba, S.M. Voronin, The Riemann zeta-function, de Gruyter 1992
[26] N.M. Korobov, Estimates of trigonometric sums and their applications, Uspehi
Mat. Nauk 13 (1958), 185-192
[27] A. Laurin cikas, Distributions des valeurs de certaines series de Dirichlet, C.R.
289 (1979), 43-45
[28] A. Laurin cikas, Limit theorems for the Riemann zeta-function, Kluwer Acad-
emic Publishers, Dordrecht 1996
[29] A. Laurin cikas, The universality of the Lerch zeta-function, Liet. Matem. Rink.
37 (1997), 367-375 (in Russian); Lith. Math. J. 38 (1997), 275-280
[30] A. Laurin cikas, On the Matsumoto zeta-function, Acta Arith. 84 (1998), 1-16
[31] A. Laurin cikas, On the eectivization of the universality theorem for the Lerch
zeta-function, Liet. Matem. Rink. 40 (2000), 172-178 (in Russian); Lith. Math. J.
40 (2000), 135-139
[32] A. Laurin cikas, The universality of zeta-functions, preprint 2002-26, Vilnius
University
[33] A. Laurin cikas, K. Matsumoto, The universality of zeta-functions attached
to certain cusp forms, Acta Arith. 98 (2001), 345-359
65
[34] A. Laurin cikas, K. Matsumoto, J. Steuding, The universality of L-
functions associated to new forms, preprint 2002-5, Vilnius University
[35] A. Laurin cikas, R.

Slezevicien e, The universality of zeta-functions with mul-
tiplicative coecients, Integral Transforms and Special Functions 13 (2002), 243-
257
[36] N. Levinson, Almost all roots of (s) = a are arbitrarily close to = 1/2, Proc.
Nat. Acad. Sci. U.S.A. 72 (1975), 1322-1324
[37] J.E. Littlewood, On the zeros of the Riemann zeta-function, Proc. Cambridge
Phil. Soc. 22 (1924), 295-318
[38] W. Luh, Universalfunktionen in einfach zusammenhangenden Gebieten, Aequa-
tiones Mathematicae 19 (1979). 183-193
[39] J. Marcinkiewicz, Sur les nombres derives, Fund. Math. 24 (1935), 305-308
[40] K. Matsumoto, Probabilistic value-distribution theory of zeta-functions, Sugaku
53 (2001), 279-296 (in Japanese)
[41] A. Ostrowski,

Uber Dirichletsche Reihen und algebraische Dierentialgleichun-
gen, Math. Z. 8 (1920), 241-298
[42] J. P al, Zwei kleine Bemerkungen, Tohoku Math. J. 6 (1914/15), 42-43
[43] G. P olya, G. Szeg o, Problems and theorems in analysis, Vol. II, Springer 1976
[44] R. Paley, N. Wiener, Fourier transforms in the complex domain, Amer. Math.
Soc. Colloq. Publ. 1934
[45] D.V. Pechersky, On the permutation of the terms of functional series, Dokl.
Akad. Nauk SSSR 209 (1973), 1285-1287
[46] M. Plancherel, Sur les formules de reciprocite du type de Fourier, J. London
Math. Soc. 8 (1933), 220-226
[47] A. Reich, Universelle Wertverteilung von Eulerprodukten, Nach. Akad. Wiss.
Gottingen, Math.-Phys. Kl. (1977), 1-17
[48] A. Reich, , Arch. Math.
[49] B. Riemann,

Uber die Anzahl der Primzahlen unterhalb einer gegebenen Grosse,
(1859)
[50] C.L. Siegel,

Uber Riemanns Nachlass zur analytischen Zahlentheorie, Quellen
und Studien zur Geschichte der Mathematik, Astronomie und Physik, Abt. B:
Studien 2 (1932), 45-80
66
[51] J. Steuding, Extremal values of Dirichlet L-functions in the half-plane of ab-
solute convergence, preprint www.math.uni-frankfurt.de/steuding/steuding.shtml
[52] J. Steuding, Upper bounds for the density of universality, Acta Arith. (to appear)
[53] J. Steuding, Universality in the Selberg class, preprint www.math.uni-
frankfurt.de/steuding/steuding.shtml
[54] E.C. Titchmarsh, The theory of functions, Oxford University Press, 2nd ed.
1939
[55] E.C. Titchmarsh, The theory of the Riemann zeta-function, Oxford University
Press 1951
[56] I.M. Vinogradov, A new estimate for the function (1 +it) , Izv. Akad. Nauk
SSSR, Ser. Mat. 22 (1958), 161-164
[57] S.M. Voronin, Theorem on the universality of the Riemann zeta-function, Izv.
Akad. Nauk SSSR, Ser. Matem., 39 (1975 (in Russian); Math. USSR Izv. 9 (1975),
443-445
[58] S.M. Voronin, On the functional independence of Dirichlet L-functions, Acta
Arith. 27 (1975), 493-503 (in Russian)
[59] H. Weyl,

Uber ein Problem aus dem Gebiete der diophantischen Approximation,
Gottinger Nachrichten (1914), 234-244
67