Chapter 7

The Prime Number Theorem

In this nal chapter we will take advantage of an opportunity to apply many of the
ideas and results from earlier chapters in order to give an analytic proof of the famous
prime number theorem: If (x) is the number of primes less than or equal to x, then
x
1
(x) lnx 1 as x . That is, (x) is asymptotically equal to x/ lnx as x .
(In the sequel, prime will be taken to mean positive prime.)
Perhaps the rst recorded property of (x) is that (x) as x , in other words,
the number of primes is innite. This appears in Euclids Elements. A more precise
result that was established much later by Euler (1737) is that the series of reciprocals of
the prime numbers,
1
2
+
1
3
+
1
5
+
1
7
+
1
11
+ ,
is a divergent series. This can be interpreted in a certain sense as a statement about
how fast (x) as x . Later, near the end of the 18-th century, mathemati-
cians, including Gauss and Legendre, through mainly empirical considerations, put forth
conjectures that are equivalent to the above statement of the prime number theorem
(PNT). However, it was not until nearly 100 years later, after much eort by numerous
19-th century mathematicians, that the theorem was nally established (independently)
by Hadamard and de la Vallee Poussin in 1896. The quest for a proof led Riemann, for
example, to develop complex variable methods to attack the PNT and related questions.
In the process, he made a remarkable and as yet unresolved conjecture known as the
Riemann hypothesis, whose precise statement will be given later. Now it is not clear on
the surface that there is a connection between complex analysis and the distribution of
prime numbers. But in fact, every proof of the PNT dating from Hadamard and de la
Vallee Poussin, up to 1949 when P. Erd os and A.Selberg succeeded in nding elemen-
tary proofs, has used the methods of complex variables in an essential way. In 1980, D.J.
Newman published a new proof of the PNT which, although still using complex analy-
sis, nevertheless represents a signicant simplication of previous proofs. It is Newmans
proof, as modied by J. Korevaar, that we present in this chapter.
There are a number of preliminaries that must be dealt with before Newmans method
can be applied to produce the theorem. The proof remains far from trivial but the steps
1
2 CHAPTER 7. THE PRIME NUMBER THEOREM
along the way are of great interest and importance in themselves. We begin by introducing
the Riemann zeta function, which arises via Eulers product formula and forms a key link
between the sequence of prime numbers and the methods of complex variables.
7.1 The Riemann Zeta function
The Riemann zeta function is dened by
(z) =

n=1
1
n
z
where n
z
= e
z ln n
. Since |n
z
| = n
Re z
, the given series converges absolutely on Re z > 1
and uniformly on {z : Re z 1 + } for every > 0. Let p
1
, p
2
, p
3
, . . . be the sequence
2,3,5, . . . of prime numbers and note that for j = 1, 2, . . . and Re z > 1, we have
1
1 1/p
z
j
= 1 +
1
p
z
j
+
1
p
2z
j
+ .
Now consider the partial product
m

j=1
1
1 p
z
j
=
m

j=1
(1 +
1
p
z
j
+
1
p
2z
j
+ ).
By multiplying the nitely many absolutely convergent series on the right together, rear-
ranging, and applying the fundamental theorem of arithmetic, we nd that the product is
the same as the sum

nPm
1
n
z
, where P
m
consists of 1 along with those positive integers
whose prime factorization uses only primes from the set {p
1
, . . . , p
m
}. Therefore

j=1
1
1 p
z
j
=

n=1
1
n
z
, Re z > 1.
We now state this formally.
7.1.1 Eulers Product formula
For Re z > 1, the Riemann zeta function (z) =

n=1
1/n
z
is given by the product

j=1
_
1
1 p
z
j
_
where {p
j
} is the (increasing) sequence of prime numbers.
The above series and product converge uniformly on compact subsets of Re z > 1,
hence is analytic on Re z > 1. Furthermore, the product representation of shows that
has no zeros in Re z > 1 (Theorem 6.1.7). Our proof of the PNT requires a number
of additional properties of . The rst result is concerned with extending to a region
larger than Re z > 1.
7.1. THE RIEMANN ZETA FUNCTION 3
7.1.2 Extension Theorem for Zeta
The function (z) 1/(z 1) has an analytic extension to the right half plane Re z > 0.
Thus has an analytic extension to {z : Re z > 0, z = 1} and has a simple pole with
residue 1 at z = 1.
Proof. For Re z > 1, apply the summation by parts formula (Problem 2.2.7) with a
n
= n
and b
n
= 1/n
z
to obtain
k1

n=1
n
_
1
(n + 1)
z

1
n
z
_
=
1
k
z1
1
k1

n=1
1
(n + 1)
z
.
Thus
1 +
k1

n=1
1
(n + 1)
z
=
1
k
z1

k1

n=1
n
_
1
(n + 1)
z

1
n
z
_
.
But
n
_
1
(n + 1)
z

1
n
z
_
= nz
_
n+1
n
t
z1
dt = z
_
n+1
n
[t]t
z1
dt
where [t] is the largest integer less than or equal to t. Hence we have
k

n=1
1
n
z
= 1 +
k1

n=1
1
(n + 1)
z
=
1
k
z1
+ z
k1

n=1
_
n+1
n
[t]t
z1
dt
=
1
k
z1
+ z
_
k
1
[t]t
z1
dt.
Letting k , we obtain the integral formula
(z) = z
_

1
[t]t
z1
dt (1)
for Re z > 1. Consider, however, the closely related integral
z
_

1
tt
z1
dt = z
_

1
t
z
dt =
z
z 1
= 1 +
1
z 1
.
Combining this with (1) we can write
(z)
1
z 1
= 1 + z
_

1
([t] t)t
z1
dt.
Now x k > 1 and consider the integral
_
k
1
([t] t)t
z1
dt. By (3.3.3), this integral is an
entire function of z. furthermore, if Re z > 0, then
|
_
k
1
([t] t)t
z1
dt|
_
k
1
t
Re(z+1)
dt
_

1
t
1Re z
dt =
1
Re z
.
4 CHAPTER 7. THE PRIME NUMBER THEOREM
This implies that the sequence f
k
(z) =
_
k
1
([t]t)t
z1
dt of analytic functions on Re z > 0
is uniformly bounded on compact subsets. Hence by Vitalis theorem (5.1.14), the limit
function
f(z) =
_

1
([t] t)t
z1
dt
(as the uniform limit on compact subsets of Re z > 0) is analytic, and thus the function
1 + z
_

1
([t] t)t
z1
dt
is also analytic on Re z > 0. But this function agrees with (z)
1
z1
for Re z > 1, and
consequently provides the required analytic extension of to Re z > 0. This completes
the proof of the theorem.
We have seen that Eulers formula (7.1.1) implies that has no zeros in the half plane
Re z > 1, but how about zeros of (the extension of) in 0 < Re z 1? The next theorem
asserts that has no zeros on the line Re z = 1. This fact is crucial to our proof of the
PNT.
7.1.3 Theorem
The Riemann zeta function has no zeros on Re z = 1, so (z 1)(z) is analytic and
zero-free on a neighborhood of Re z 1.
Proof. Fix a real number y = 0 and consider the auxiliary function
h(x) =
3
(x)
4
(x + iy)(x + i2y)
for x real and x > 1. By Eulers product formula, if Re z > 1 then
ln|(z)| =

j=1
ln|1 p
z
j
| = Re

j=1
Log(1 p
z
j
) = Re

j=1

n=1
1
n
p
nz
j
where we have used the expansion Log(1 w) =

n=1
w
n
/n, valid for |w| < 1. Hence
ln|h(x)| = 3 ln|(x)| + 4 ln|(x + iy)| + ln|(x + i2y)|
= 3 Re

j=1

n=1
1
n
p
nx
j
+ 4 Re

j=1

n=1
1
n
p
nxiny
j
+ Re

j=1

n=1
1
n
p
nxi2ny
j
=

j=1

n=1
1
n
p
nx
j
Re(3 + 4p
iny
j
+ p
i2ny
j
).
But p
iny
j
= e
iny ln pj
and p
i2ny
j
= e
i2ny ln pj
. Thus Re(3 + 4p
iny
j
+ p
i2ny
j
) has the
form
3 + 4 cos + cos 2 = 3 + 4 cos + 2 cos
2
1 = 2(1 + cos )
2
0.
7.1. THE RIEMANN ZETA FUNCTION 5
Therefore ln|h(x)| 0 and consequently
|h(x)| = |
3
(x)||
4
(x + iy)||(x + i2y)| 1.
Thus
|h(x)|
x 1
= |(x 1)(x)|
3

(x + iy)
x 1

4
|(x + i2y)|
1
x 1
.
But if (1+iy) = 0, then the left hand side of this inequality would approach a nite limit
|

(1 +iy)|
4
|(1 +i2y)| as x 1
+
since has a simple pole at 1 with residue 1. However,
the right hand side of the inequality contradicts this. We conclude that (1 + iy) = 0.
Since y is an arbitrary nonzero real number, has no zeros on Re z = 1.
Remark
The ingenious introduction of the auxiliary function h is due to Mertens (1898). We
now have shown that any zeros of in Re z > 0 must lie in the strip 0 < Re z < 1.
The study of the zeros of has long been the subject of intensive investigation by many
mathematicians. Riemann had stated in his seminal 1859 paper that he considered it
very likely that all the zeros of in the above strip, called the critical strip, lie on the
line Re z = 1/2. This assertion is now known as the Riemann hypothesis, and remains
as yet unresolved. However, a great deal is known about the distribution of the zeros
of in the critical strip, and the subject continues to capture the attention of eminent
mathematicians. To state just one such result, G.H. Hardy proved in 1915 that has
innitely many zeros on the line Re z = 1/2. Those interested in learning more about this
fascinating subject may consult, for example, the book Riemanns Zeta Function by H.M.
Edwards. Another source is http://mathworld.wolfram.com/RiemannHypothesis.html.
We turn next to zetas logarithmic derivative

/, which we know is analytic on

Re z > 1. In fact, more is true, for by (7.1.3),

/ is analytic on a neighborhood of
{z : Re z 1 and z = 1}. Since has a simple pole at z = 1, so does

/, with residue
Res(

/, 1) = 1. [See the proof of (4.2.7).] We next obtain an integral representation

for

/ that is similar to the representation (1) above for . [See the proof of (7.1.2).]
But rst, we must introduce the von Mangoldt function , which is dened by
(n) =
_
lnp if n = p
m
for some m,
0 otherwise.
Thus (n) is lnp if n is a power of the prime p, and is 0 if not. Next dene on x 0 by
(x) =

nx
(n). (2)
An equivalent expression for is
(x) =

px
m
p
(x) lnp,
6 CHAPTER 7. THE PRIME NUMBER THEOREM
where the sum is over primes p x and m
p
(x) is the largest integer such that p
mp(x)
x.
(For example, (10.4) = 3 ln2 +2 ln3 +ln5 +ln7.) Note that p
mp(x)
x i m
p
(x) lnp
lnx i m
p
(x)
ln x
ln p
. Thus m
p
(x) =
_
ln x
ln p
_
where as before, [ ] denotes the greatest integer
function. The function will be used to obtain the desired integral representation for

/.
7.1.4 Theorem
For Re z > 1,

(z)
(z)
= z
_

1
(t)t
z1
dt (3)
where is dened as above.
Proof. In the formulas below, p and q range over primes. If Re z > 1, we have (z) =

p
(1 p
z
)
1
by (7.1.1), hence

(z) =

p
p
z
lnp
(1 p
z
)
2

q=p
1
1 q
z
= (z)

p
p
z
lnp
(1 p
z
)
2
(1 p
z
)
= (z)

p
p
z
lnp
1 p
z
.
Thus

(z)
(z)
=

p
p
z
lnp
1 p
z
=

n=1
p
nz
lnp.
The iterated sum is absolutely convergent for Re z > 1, so it can be rearranged as a double
sum

(p,n),n1
(p
n
)
z
lnp =

k
k
z
lnp
where k = p
n
for some n. Consequently,

(z)
(z)
=

k=1
k
z
(k) =

k=1
k
z
((k) (k 1))
by the denitions of and . But using partial summation once again we obtain, with
a
k
= k
z
, b
k+1
= (k), and b
1
= (0) = 0 in Problem 2.2.7,
M

k=1
k
z
((k) (k 1)) = (M)(M + 1)
z
+
M

k=1
(k)(k
z
(k + 1)
z
).
7.2. AN EQUIVALENT VERSION OF THE PRIME NUMBER THEOREM 7
Now from the denition (2) of (x) we have (x) xlnx, so if Re z > 1 we have
(M)(M + 1)
z
0 as M . Moreover, we can write
M

k=1
(k)(k
z
(k + 1)
z
) =
M

k=1
(k)z
_
k+1
k
t
z1
dt
=
M

k=1
z
_
k+1
k
(t)t
z1
dt
= z
_
M
1
(t)t
z1
dt
because is constant on each interval [k, k +1). Taking limits as M , we nally get

(z)
(z)
= z
_

1
(t)t
z1
dt, Re z > 1.
7.2 An Equivalent Version of the Prime Number The-
orem
The function dened in (2) above provides yet another connection, through (3), between
the Riemann zeta function and properties of the prime numbers. The integral that appears
in (3) is called the Mellin transform of and is studied in its own right. We next establish
a reduction, due to Chebyshev, of the prime number theorem to a statement involving
the function .
7.2.1 Theorem
The prime number theorem holds, that is, x
1
(x) lnx 1, i x
1
(x) 1 as x .
Proof. Recall that
(x) =

px
_
lnx
lnp
_
lnp

px
lnx
lnp
lnp (1)
= lnx

px
1
= (lnx)(x).
8 CHAPTER 7. THE PRIME NUMBER THEOREM
However, if 1 < y < x, then
(x) = (y) +

y<px
1
(y) +

y<px
lnp
lny
(2)
< y +
1
lny

y<px
lnp
y +
1
lny
(x).
Now take y = x/(lnx)
2
in (2), and we get
(x)
x
(lnx)
2
+
1
lnx 2 lnlnx
(x).
Thus
(x)
lnx
x

1
lnx
+
lnx
lnx 2 lnlnx
(x)
x
. (3)
It now follows from (1) and (3) that
(x)
x

lnx
x
(x)
1
lnx
+
lnx
lnx 2 lnlnx
(x)
x
and from this we can see that x
1
(x) 1 i x
1
(x) lnx 1 as x .
The goal will now be to show that (x)/x 1 as x . A necessary intermediate
step for our proof is to establish the following weaker estimate on the asymptotic behavior
of (x).
7.2.2 Lemma
There exists C > 0 such that (x) Cx, x > 0. For short, (x) = O(x).
Proof. Again recall that (x) =

px
[
ln x
ln p
] lnp, x > 0. Fix x > 0 and let m be an integer
such that 2
m
< x 2
m+1
. Then
(x) = (2
m
) + (x) (2
m
)
(2
m
) + (2
m+1
) (2
m
)
=

p2
m
_
ln2
m
lnp
_
lnp +

2
m
<p2
m+1
_
ln2
m+1
lnp
_
lnp.
Consider, for any positive integer n,

n<p2n
lnp = ln

n<p2n
p.
7.3. PROOF OF THE PRIME NUMBER THEOREM 9
Now for any prime p such that n < p 2n, p divides (2n)!/n! = n!
_
2n
n
_
. Since such a p
does not divide n!, it follows that p divides
_
2n
n
_
. Hence

n<p2n
p
_
2n
n
_
< (1 + 1)
2n
= 2
2n
,
and we arrive at

n<p2n
lnp < 2nln2.
Therefore

p2
m
lnp =
m

k=1
_
_

2
k1
<p2
k
lnp
_
_
<
m

k=1
2
k
ln2 < 2
m+1
ln2
and

2
m
<p2
m+1
lnp < 2
m+1
ln2.
But if p x is such that
_
ln x
ln p
_
> 1, then
ln x
ln p
2 and hence x p
2
so that

x p.
Thus those terms in the sum

px
_
ln x
ln p
_
lnp where
_
ln x
ln p
_
> 1 occur only when p

x,
and the sum of terms of this form contribute no more than

x
lnx
lnp
lnp = (

x) lnx.
It follows from the above discussion that if 2
m
< x 2
m+1
, then
(x) 2
m+1
ln2 + 2
m+1
ln2 + (

x) lnx
= 2
m+2
ln2 + (

x) lnx
< 4xln2 + (

x) lnx
4xln2 +

xlnx
= (4 ln2 +
1

x
lnx)x.
Since
1

x
lnx 0 as x , we conclude that (x) = O(x), which proves the lemma.
7.3 Proof of the Prime Number Theorem
Our approach to the prime number theorem has been along traditional lines, but at this
stage we will apply D.J. Newmans method (Simple Analytic Proof of the Prime Number
Theorem, American Math. Monthly 87 (1980), 693-696) as modied by J. Korevaar (On
10 CHAPTER 7. THE PRIME NUMBER THEOREM
Newmans Quick Way to the Prime Number Theorem, Math. Intelligencer 4 (1982), 108-
115). Korevaars approach is to apply Newmans ideas to obtain properties of certain
Laplace integrals that lead to the prime number theorem.
Our plan is to deduce the prime number theorem from a Tauberian theorem (7.3.1)
and its corollary (7.3.2). Then we will prove (7.3.1) and (7.3.2).
7.3.1 Auxiliary Tauberian Theorem
Let F be bounded and piecewise continuous on [0, +), so that its Laplace transform
G(z) =
_

0
F(t)e
zt
dt
exists and is analytic on Re z > 0. Assume that G has an analytic extension to a neigh-
borhood of the imaginary axis, Re z = 0. Then
_

0
F(t) dt exists as an improper integral
and is equal to G(0). [In fact,
_

0
F(t)e
iyt
dt converges for every y R to G(iy).]
Results like (7.3.1) are named for A. Tauber, who is credited with proving the rst
theorem of this type near the end of the 19th century. The phrase Tauberian theorem
was coined by G.H. Hardy, who along with J.E. Littlewood made a number of contri-
butions in this area. Generally, Tauberian theorems are those in which some type of
ordinary convergence (e.g., convergence of
_

0
F(t)e
iyt
dt for each y R), is deduced
from some weaker type of convergence (e.g., convergence of
_

0
F(t)e
zt
dt for each z
with Re z > 0) provided additional conditions are satised (e.g., G has an analytic exten-
sion to a neighborhood of each point on the imaginary axis). Taubers original theorem
can be found in The Elements of Real Analysis by R.G. Bartle.
7.3.2 Corollary
Let f be a nonnegative, piecewise continuous and nondecreasing function on [1, ) such
that f(x) = O(x). Then its Mellin transform
g(z) = z
_

1
f(x)x
z1
dx
exists for Re z > 1 and denes an analytic function g. Assume that for some constant c,
the function
g(z)
c
z 1
has an analytic extension to a neighborhood of the line Re z = 1. Then as x ,
f(x)
x
c.
As stated earlier, we are rst going to see how the prime number theorem follows from
(7.3.1) and (7.3.2). To this end, let be as above, namely
(x) =

px
_
lnx
lnp
_
lnp.
7.3. PROOF OF THE PRIME NUMBER THEOREM 11
Then is a nonnegative, piecewise continuous, nondecreasing function on [1, ). Fur-
thermore, by (7.2.2), (x) = O(x), so by (7.3.2) we may take f = and consider the
Mellin transform
g(z) = z
_

1
(x)x
z1
dx.
But by (7.1.4), actually g(z) =

(z)/(z), and by the discussion leading up to the

statement of (7.1.4),

(z)
(z)
+
1
z1
has an analytic extension to a neighborhood of each
point of Re z = 1, hence so does g(z)
1
z1
. Consequently, by (7.3.2), we can conclude
that (x)/x 1, which, by (7.2.1), is equivalent to the PNT. Thus we are left with the
proof of (7.3.1) and its corollary (7.3.2).
Proof of (7.3.1)
Let F be as in the statement of the theorem. Then it follows just as in the proof of (7.1.2),
the extension theorem for zeta, that Fs Laplace transform G is dened and analytic
on Re z > 0. Assume that G has been extended to an analytic function on a region
containing Re z 0. Since F is bounded we may as well assume that |F(t)| 1, t 0.
For 0 < < , dene
G

(z) =
_

0
F(t)e
zt
dt.
By (3.3.3), each function G

is entire, and the conclusion of our theorem may be expressed

as
lim

(0) = G(0).
That is, the improper integral
_

0
F(t) dt exists and converges to G(0). We begin the
analysis by using Cauchys integral formula to get a preliminary estimate of |G

(0)G(0)|.
For each R > 0, let (R) > 0 be so small that G is analytic inside and on the closed path
. .
.
.
-iR
iR

(
R
)
1

-
R

+
R

-
R

+
R

R
=
+
Figure 7.3.1

R
in Figure 7.3.1. (Note that since G is analytic on an open set containing Re z 0,
12 CHAPTER 7. THE PRIME NUMBER THEOREM
such a (R) > 0 must exist, although it may well be the case that (R) 0 as R +.)
Let
+
R
denote that portion of
R
that lies in Re z > 0, and

R
the portion that lies in
Re z < 0. By Cauchys integral formula,
G(0) G

(0) =
1
2i
_

R
(G(z) G

(z))
1
z
dz. (1)
Let us consider the consequences of estimating |G(0) G

(0)| by applying the usual M-

L estimates to the integral on the right hand side of (1) above. First, for z
+
R
and
x = Re z, we have

G(z) G

(z)
z

=
1
R

F(t)e
zt
dt

1
R
_

|F(t)|e
xt
dt

1
R
_

e
xt
dt
=
1
R
e
x
x
(2)

1
R
1
x
=
1
R
1
Re z
.
But 1/ Re z is unbounded on
+
R
, so we see that a more delicate approach is required to
shows that G(0) G

(0) 0 as . Indeed, it is here that Newmans ingenuity

comes to the fore, and provides us with a modication of the above integral representation
for G(0) G

(0). This will furnish the appropriate estimate. Newmans idea is to replace
the factor 1/z by (1/z) +(z/R
2
) in the path integral in (1). Since (G(z) G

(z))z/R
2
is
analytic, the value of the path integral along
R
remains unchanged. We further modify
(1) by replacing G(z) and G

(z) by their respective products with e

z
. Since e
z
is entire
and has the value 1 at z = 0, we can write
G(0) G

(0) =
1
2i
_

R
(G(z) G

(z))e
z
(
1
z
+
z
R
2
) dz.
Note that for |z| = R we have (1/z) + (z/R
2
) = (z/|z|
2
) + (z/R
2
) = (2 Re z)/R
2
, so that
if z
+
R
, (recalling (2) above),
|(G(z) G

(z))e
z
(
1
z
+
z
R
2
)|
1
Re z
e
Re z
e
Re z
2 Re z
R
2
=
2
R
2
.
Consequently,

1
2i
_

+
R
(G(z) G

(z))e
z
(
1
z
+
z
R
2
) dz

1
R
by the M-L theorem. Note that this estimate of the integral along the path
+
R
is inde-
pendent of . Now let us consider the contribution to the integral along
R
of the integral
7.3. PROOF OF THE PRIME NUMBER THEOREM 13
along

R
. First we use the triangle inequality to obtain the estimate

1
2i
_

R
(G(z) G

(z))e
z
(
1
z
+
z
R
2
) dz

1
2i
_

R
G(z)e
z
(
1
z
+
z
R
2
) dz

1
2i
_

R
G

(z)e
z
(
1
z
+
z
R
2
) dz

= |I
1
(R)| +|I
2
(R)|.
First consider I
2
(R). Since G

(z) is an entire function, we can replace the path of inte-

gration

R
by the semicircular path from iR to iR in the left half plane. For z on this
semicircular arc, the modulus of the integrand in I
2
(R) is
|(
_

0
F(t)e
zt
dt)e
z
2 Re z
R
2
)|
1
| Re z|
2| Re z|
R
2
=
2
R
2
.
(Note that |F| 1, we can replace the upper limit of integration by , and e
x
1 for
x 0.) This inequality also holds if Re z = 0 (let z iy). Thus by the M-L theorem we
get |I
2
(R)| (1/2)(2/R
2
)(R) = 1/R, again.
Finally, we consider |I
1
(R)|. This will be the trickiest of all since we only know that on

R
, G is an analytic extension of the explicitly dened G in the right half plane. To deal
with this case, rst choose a constant M(R) > 0 such that |G(z)| M(R) for z

R
.
Choose
1
such that 0 <
1
< (R) and break up the integral dening I
1
(R) into two
parts, corresponding to Re z <
1
and Re z
1
. The rst contribution is bounded in
modulus by
1
2
M(R)e
1
(
1
(R)
+
1
R
)R =
1
2
RM(R)(
1
(R)
+
1
R
)e
1
,
which for xed R and
1
tends to 0 as . On the other hand, the second contribution
is bounded in modulus by
1
2
M(R)(
1
(R)
+
1
R
)2Rarcsin

1
R
,
the last factor arising from summing the lengths of two short circular arcs on the path
of integration. Thus for xed R and (R) we can make the above expression as small
as we please by taking
1
suciently close to 0. So at last we are ready to establish the
conclusion of this theorem. Let > 0 be given. Take R = 4/ and x (R), 0 < (R) < R,
such that G is analytic inside and on
R
. Then as we saw above, for all ,

1
2i
_

+
R
(G(z) G

(z))e
z
(
1
z
+
z
R
2
) dz

1
R
=

4
and also

1
2i
_

R
(G

(z)e
z
(
1
z
+
z
R
2
) dz

1
R
=

4
.
14 CHAPTER 7. THE PRIME NUMBER THEOREM
Now choose
1
such that 0 <
1
< (R) and such that
1
2
M(R)(
1
(R)
+
1
R
)2Rarcsin

1
R
<

4
.
Since
1
2
RM(R)(
1
(R)
+
1
R
)e
1
<

4
for all suciently large, say
0
, it follows that
|G

(0) G(0)| < ,

0
which completes the proof.
Proof of (7.3.2)
Let f(x) and g(z) be as in the statement of the corollary. Dene F on [0, +) by
F(t) = e
t
f(e
t
) c.
Then F satises the rst part of the hypothesis of the auxiliary Tauberian theorem, so
let us consider its Laplace transform,
G(z) =
_

0
(e
t
f(e
t
) c)e
zt
dt,
which via the change of variables x = e
t
becomes
G(z) =
_

1
(
1
x
f(x) c)x
z
dx
x
=
_

1
f(x)x
z2
dx c
_

1
x
z1
dx
=
_

1
f(x)x
z2
dx
c
z
=
g(z + 1)
z + 1

c
z
=
1
z + 1
[g(z + 1)
c
z
c].
It follows from the hypothesis that g(z +1) (c/z) has an analytic extension to a neigh-
borhood of the line Re z = 0, and consequently the same is true of the above function G.
Thus the hypotheses of the auxiliary Tauberian theorem are satised, and we conclude
that the improper integral
_

0
F(t) dt exists and converges to G(0). In terms of f, this
says that
_

0
(e
t
f(e
t
) c) dt exists, or equivalently (via the change of variables x = e
t
once more) that
_

1
(
f(x)
x
c)
dx
x
7.3. PROOF OF THE PRIME NUMBER THEOREM 15
exists. Recalling that f is nondecreasing, we can infer that f(x)/x c as x . For
let > 0 be given, and suppose that for some x
0
> 0, [f(x
0
)/x
0
] c 2. It follows that
f(x) f(x
0
) x
0
(c + 2) x(c + ) for x
0
x
c + 2
c +
x
0
.
Hence,
_ c+2
c+
x0
x0
(
f(x)
x
c)
dx
x

_ c+2
c+
x0
x0

x
dx = ln(
c + 2
c +
).
But
_
x2
x1
(
f(x)
x
c)
dx
x
0 as x
1
, x
2
, because the integral from 1 to is convergent.
Thus for all x
0
suciently large,
_ c+2
c+
x0
x0
(
f(x)
x
c)
dx
x
< ln(
c + 2
c +
).
However, reasoning from the assumption that [f(x
0
)/x
0
]c 2, we have just deduced the
opposite inequality. We must conclude that for all x
0
suciently large, [f(x
0
)/x
0
]c < 2.
Similarly, [f(x
0
)/x
0
] c > 2 for all x
0
suciently large. [Say [f(x
0
)/x
0
] c 2.
The key inequality now becomes
f(x) f(x
0
) x
0
(c 2) x(c ) for (
c 2
c
)x
0
x x
0
and the limits of integration in the next step are from
c2
c
x
0
to x
0
.] Therefore f(x)/x c
as x , completing the proof of both the corollary and the prime number theorem.
The prime number theorem has a long and interesting history. We have mentioned
just a few of the many historical issues related to the PNT in this chapter. There are
several other number theoretic functions related to (x), in addition to the function (x)
that was introduced earlier. A nice discussion of some of these issues can be found in Eric
W. Weisstein, Prime Number Theorem, from MathWorldA Wolfram Web Resource,
http://mathworld.wolfram.com/PrimeNumberTheorem.html. This source also includes a
number of references on PNT related matters.
References
1. Karl Sabbach, The Riemann Hypothesis-The Greatest Unsolved Problem in Mathe-
matics, Farrar, Strass and Giroux, New York, 2002.
2. Julian Havil, Gamma-Exploring Eulers Constant, Princeton University Press, Prince-
ton and Oxford, 2003, Chapter 15.